2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #include <linux/cgroup.h>
30 #include <linux/cred.h>
31 #include <linux/ctype.h>
32 #include <linux/errno.h>
33 #include <linux/init_task.h>
34 #include <linux/kernel.h>
35 #include <linux/list.h>
37 #include <linux/mutex.h>
38 #include <linux/mount.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/spinlock.h>
45 #include <linux/rwsem.h>
46 #include <linux/string.h>
47 #include <linux/sort.h>
48 #include <linux/kmod.h>
49 #include <linux/delayacct.h>
50 #include <linux/cgroupstats.h>
51 #include <linux/hashtable.h>
52 #include <linux/pid_namespace.h>
53 #include <linux/idr.h>
54 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
55 #include <linux/flex_array.h> /* used in cgroup_attach_task */
56 #include <linux/kthread.h>
57 #include <linux/delay.h>
59 #include <linux/atomic.h>
62 * pidlists linger the following amount before being destroyed. The goal
63 * is avoiding frequent destruction in the middle of consecutive read calls
64 * Expiring in the middle is a performance problem not a correctness one.
65 * 1 sec should be enough.
67 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
69 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
73 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
74 * creation/removal and hierarchy changing operations including cgroup
75 * creation, removal, css association and controller rebinding. This outer
76 * lock is needed mainly to resolve the circular dependency between kernfs
77 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
79 static DEFINE_MUTEX(cgroup_tree_mutex);
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 EXPORT_SYMBOL_GPL(cgroup_mutex); /* only for lockdep */
89 static DEFINE_MUTEX(cgroup_mutex);
93 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
94 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
96 static DEFINE_SPINLOCK(release_agent_path_lock);
98 #define cgroup_assert_mutexes_or_rcu_locked() \
99 rcu_lockdep_assert(rcu_read_lock_held() || \
100 lockdep_is_held(&cgroup_tree_mutex) || \
101 lockdep_is_held(&cgroup_mutex), \
102 "cgroup_[tree_]mutex or RCU read lock required");
105 * cgroup destruction makes heavy use of work items and there can be a lot
106 * of concurrent destructions. Use a separate workqueue so that cgroup
107 * destruction work items don't end up filling up max_active of system_wq
108 * which may lead to deadlock.
110 static struct workqueue_struct *cgroup_destroy_wq;
113 * pidlist destructions need to be flushed on cgroup destruction. Use a
114 * separate workqueue as flush domain.
116 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
118 /* generate an array of cgroup subsystem pointers */
119 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
120 static struct cgroup_subsys *cgroup_subsys[] = {
121 #include <linux/cgroup_subsys.h>
125 /* array of cgroup subsystem names */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
127 static const char *cgroup_subsys_name[] = {
128 #include <linux/cgroup_subsys.h>
133 * The dummy hierarchy, reserved for the subsystems that are otherwise
134 * unattached - it never has more than a single cgroup, and all tasks are
135 * part of that cgroup.
137 static struct cgroupfs_root cgroup_dummy_root;
139 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
140 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
142 /* The list of hierarchy roots */
144 static LIST_HEAD(cgroup_roots);
145 static int cgroup_root_count;
147 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
148 static DEFINE_IDR(cgroup_hierarchy_idr);
151 * Assign a monotonically increasing serial number to cgroups. It
152 * guarantees cgroups with bigger numbers are newer than those with smaller
153 * numbers. Also, as cgroups are always appended to the parent's
154 * ->children list, it guarantees that sibling cgroups are always sorted in
155 * the ascending serial number order on the list. Protected by
158 static u64 cgroup_serial_nr_next = 1;
160 /* This flag indicates whether tasks in the fork and exit paths should
161 * check for fork/exit handlers to call. This avoids us having to do
162 * extra work in the fork/exit path if none of the subsystems need to
165 static int need_forkexit_callback __read_mostly;
167 static struct cftype cgroup_base_files[];
169 static void cgroup_put(struct cgroup *cgrp);
170 static int rebind_subsystems(struct cgroupfs_root *root,
171 unsigned long added_mask, unsigned removed_mask);
172 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
173 static int cgroup_destroy_locked(struct cgroup *cgrp);
174 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
176 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
179 * cgroup_css - obtain a cgroup's css for the specified subsystem
180 * @cgrp: the cgroup of interest
181 * @ss: the subsystem of interest (%NULL returns the dummy_css)
183 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
184 * function must be called either under cgroup_mutex or rcu_read_lock() and
185 * the caller is responsible for pinning the returned css if it wants to
186 * keep accessing it outside the said locks. This function may return
187 * %NULL if @cgrp doesn't have @subsys_id enabled.
189 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
190 struct cgroup_subsys *ss)
193 return rcu_dereference_check(cgrp->subsys[ss->id],
194 lockdep_is_held(&cgroup_tree_mutex) ||
195 lockdep_is_held(&cgroup_mutex));
197 return &cgrp->dummy_css;
200 /* convenient tests for these bits */
201 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
203 return test_bit(CGRP_DEAD, &cgrp->flags);
206 struct cgroup_subsys_state *seq_css(struct seq_file *seq)
208 struct kernfs_open_file *of = seq->private;
209 struct cgroup *cgrp = of->kn->parent->priv;
210 struct cftype *cft = seq_cft(seq);
213 * This is open and unprotected implementation of cgroup_css().
214 * seq_css() is only called from a kernfs file operation which has
215 * an active reference on the file. Because all the subsystem
216 * files are drained before a css is disassociated with a cgroup,
217 * the matching css from the cgroup's subsys table is guaranteed to
218 * be and stay valid until the enclosing operation is complete.
221 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
223 return &cgrp->dummy_css;
225 EXPORT_SYMBOL_GPL(seq_css);
228 * cgroup_is_descendant - test ancestry
229 * @cgrp: the cgroup to be tested
230 * @ancestor: possible ancestor of @cgrp
232 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
233 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
234 * and @ancestor are accessible.
236 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
239 if (cgrp == ancestor)
245 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
247 static int cgroup_is_releasable(const struct cgroup *cgrp)
250 (1 << CGRP_RELEASABLE) |
251 (1 << CGRP_NOTIFY_ON_RELEASE);
252 return (cgrp->flags & bits) == bits;
255 static int notify_on_release(const struct cgroup *cgrp)
257 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
261 * for_each_css - iterate all css's of a cgroup
262 * @css: the iteration cursor
263 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
264 * @cgrp: the target cgroup to iterate css's of
266 * Should be called under cgroup_mutex.
268 #define for_each_css(css, ssid, cgrp) \
269 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
270 if (!((css) = rcu_dereference_check( \
271 (cgrp)->subsys[(ssid)], \
272 lockdep_is_held(&cgroup_tree_mutex) || \
273 lockdep_is_held(&cgroup_mutex)))) { } \
277 * for_each_subsys - iterate all enabled cgroup subsystems
278 * @ss: the iteration cursor
279 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
281 #define for_each_subsys(ss, ssid) \
282 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
283 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
285 /* iterate across the active hierarchies */
286 #define for_each_active_root(root) \
287 list_for_each_entry((root), &cgroup_roots, root_list)
290 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
291 * @cgrp: the cgroup to be checked for liveness
293 * On success, returns true; the mutex should be later unlocked. On
294 * failure returns false with no lock held.
296 static bool cgroup_lock_live_group(struct cgroup *cgrp)
298 mutex_lock(&cgroup_mutex);
299 if (cgroup_is_dead(cgrp)) {
300 mutex_unlock(&cgroup_mutex);
306 /* the list of cgroups eligible for automatic release. Protected by
307 * release_list_lock */
308 static LIST_HEAD(release_list);
309 static DEFINE_RAW_SPINLOCK(release_list_lock);
310 static void cgroup_release_agent(struct work_struct *work);
311 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
312 static void check_for_release(struct cgroup *cgrp);
315 * A cgroup can be associated with multiple css_sets as different tasks may
316 * belong to different cgroups on different hierarchies. In the other
317 * direction, a css_set is naturally associated with multiple cgroups.
318 * This M:N relationship is represented by the following link structure
319 * which exists for each association and allows traversing the associations
322 struct cgrp_cset_link {
323 /* the cgroup and css_set this link associates */
325 struct css_set *cset;
327 /* list of cgrp_cset_links anchored at cgrp->cset_links */
328 struct list_head cset_link;
330 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
331 struct list_head cgrp_link;
334 /* The default css_set - used by init and its children prior to any
335 * hierarchies being mounted. It contains a pointer to the root state
336 * for each subsystem. Also used to anchor the list of css_sets. Not
337 * reference-counted, to improve performance when child cgroups
338 * haven't been created.
341 static struct css_set init_css_set;
342 static struct cgrp_cset_link init_cgrp_cset_link;
345 * css_set_rwsem protects the list of css_set objects, and the chain of
346 * tasks off each css_set.
348 static DECLARE_RWSEM(css_set_rwsem);
349 static int css_set_count;
352 * hash table for cgroup groups. This improves the performance to find
353 * an existing css_set. This hash doesn't (currently) take into
354 * account cgroups in empty hierarchies.
356 #define CSS_SET_HASH_BITS 7
357 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
359 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
361 unsigned long key = 0UL;
362 struct cgroup_subsys *ss;
365 for_each_subsys(ss, i)
366 key += (unsigned long)css[i];
367 key = (key >> 16) ^ key;
372 static void put_css_set_locked(struct css_set *cset, bool taskexit)
374 struct cgrp_cset_link *link, *tmp_link;
376 lockdep_assert_held(&css_set_rwsem);
378 if (!atomic_dec_and_test(&cset->refcount))
381 /* This css_set is dead. unlink it and release cgroup refcounts */
382 hash_del(&cset->hlist);
385 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
386 struct cgroup *cgrp = link->cgrp;
388 list_del(&link->cset_link);
389 list_del(&link->cgrp_link);
391 /* @cgrp can't go away while we're holding css_set_rwsem */
392 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
394 set_bit(CGRP_RELEASABLE, &cgrp->flags);
395 check_for_release(cgrp);
401 kfree_rcu(cset, rcu_head);
404 static void put_css_set(struct css_set *cset, bool taskexit)
407 * Ensure that the refcount doesn't hit zero while any readers
408 * can see it. Similar to atomic_dec_and_lock(), but for an
411 if (atomic_add_unless(&cset->refcount, -1, 1))
414 down_write(&css_set_rwsem);
415 put_css_set_locked(cset, taskexit);
416 up_write(&css_set_rwsem);
420 * refcounted get/put for css_set objects
422 static inline void get_css_set(struct css_set *cset)
424 atomic_inc(&cset->refcount);
428 * compare_css_sets - helper function for find_existing_css_set().
429 * @cset: candidate css_set being tested
430 * @old_cset: existing css_set for a task
431 * @new_cgrp: cgroup that's being entered by the task
432 * @template: desired set of css pointers in css_set (pre-calculated)
434 * Returns true if "cset" matches "old_cset" except for the hierarchy
435 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
437 static bool compare_css_sets(struct css_set *cset,
438 struct css_set *old_cset,
439 struct cgroup *new_cgrp,
440 struct cgroup_subsys_state *template[])
442 struct list_head *l1, *l2;
444 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
445 /* Not all subsystems matched */
450 * Compare cgroup pointers in order to distinguish between
451 * different cgroups in heirarchies with no subsystems. We
452 * could get by with just this check alone (and skip the
453 * memcmp above) but on most setups the memcmp check will
454 * avoid the need for this more expensive check on almost all
458 l1 = &cset->cgrp_links;
459 l2 = &old_cset->cgrp_links;
461 struct cgrp_cset_link *link1, *link2;
462 struct cgroup *cgrp1, *cgrp2;
466 /* See if we reached the end - both lists are equal length. */
467 if (l1 == &cset->cgrp_links) {
468 BUG_ON(l2 != &old_cset->cgrp_links);
471 BUG_ON(l2 == &old_cset->cgrp_links);
473 /* Locate the cgroups associated with these links. */
474 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
475 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
478 /* Hierarchies should be linked in the same order. */
479 BUG_ON(cgrp1->root != cgrp2->root);
482 * If this hierarchy is the hierarchy of the cgroup
483 * that's changing, then we need to check that this
484 * css_set points to the new cgroup; if it's any other
485 * hierarchy, then this css_set should point to the
486 * same cgroup as the old css_set.
488 if (cgrp1->root == new_cgrp->root) {
489 if (cgrp1 != new_cgrp)
500 * find_existing_css_set - init css array and find the matching css_set
501 * @old_cset: the css_set that we're using before the cgroup transition
502 * @cgrp: the cgroup that we're moving into
503 * @template: out param for the new set of csses, should be clear on entry
505 static struct css_set *find_existing_css_set(struct css_set *old_cset,
507 struct cgroup_subsys_state *template[])
509 struct cgroupfs_root *root = cgrp->root;
510 struct cgroup_subsys *ss;
511 struct css_set *cset;
516 * Build the set of subsystem state objects that we want to see in the
517 * new css_set. while subsystems can change globally, the entries here
518 * won't change, so no need for locking.
520 for_each_subsys(ss, i) {
521 if (root->subsys_mask & (1UL << i)) {
522 /* Subsystem is in this hierarchy. So we want
523 * the subsystem state from the new
525 template[i] = cgroup_css(cgrp, ss);
527 /* Subsystem is not in this hierarchy, so we
528 * don't want to change the subsystem state */
529 template[i] = old_cset->subsys[i];
533 key = css_set_hash(template);
534 hash_for_each_possible(css_set_table, cset, hlist, key) {
535 if (!compare_css_sets(cset, old_cset, cgrp, template))
538 /* This css_set matches what we need */
542 /* No existing cgroup group matched */
546 static void free_cgrp_cset_links(struct list_head *links_to_free)
548 struct cgrp_cset_link *link, *tmp_link;
550 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
551 list_del(&link->cset_link);
557 * allocate_cgrp_cset_links - allocate cgrp_cset_links
558 * @count: the number of links to allocate
559 * @tmp_links: list_head the allocated links are put on
561 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
562 * through ->cset_link. Returns 0 on success or -errno.
564 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
566 struct cgrp_cset_link *link;
569 INIT_LIST_HEAD(tmp_links);
571 for (i = 0; i < count; i++) {
572 link = kzalloc(sizeof(*link), GFP_KERNEL);
574 free_cgrp_cset_links(tmp_links);
577 list_add(&link->cset_link, tmp_links);
583 * link_css_set - a helper function to link a css_set to a cgroup
584 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
585 * @cset: the css_set to be linked
586 * @cgrp: the destination cgroup
588 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
591 struct cgrp_cset_link *link;
593 BUG_ON(list_empty(tmp_links));
594 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
597 list_move(&link->cset_link, &cgrp->cset_links);
599 * Always add links to the tail of the list so that the list
600 * is sorted by order of hierarchy creation
602 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
606 * find_css_set - return a new css_set with one cgroup updated
607 * @old_cset: the baseline css_set
608 * @cgrp: the cgroup to be updated
610 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
611 * substituted into the appropriate hierarchy.
613 static struct css_set *find_css_set(struct css_set *old_cset,
616 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
617 struct css_set *cset;
618 struct list_head tmp_links;
619 struct cgrp_cset_link *link;
622 lockdep_assert_held(&cgroup_mutex);
624 /* First see if we already have a cgroup group that matches
626 down_read(&css_set_rwsem);
627 cset = find_existing_css_set(old_cset, cgrp, template);
630 up_read(&css_set_rwsem);
635 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
639 /* Allocate all the cgrp_cset_link objects that we'll need */
640 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
645 atomic_set(&cset->refcount, 1);
646 INIT_LIST_HEAD(&cset->cgrp_links);
647 INIT_LIST_HEAD(&cset->tasks);
648 INIT_HLIST_NODE(&cset->hlist);
650 /* Copy the set of subsystem state objects generated in
651 * find_existing_css_set() */
652 memcpy(cset->subsys, template, sizeof(cset->subsys));
654 down_write(&css_set_rwsem);
655 /* Add reference counts and links from the new css_set. */
656 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
657 struct cgroup *c = link->cgrp;
659 if (c->root == cgrp->root)
661 link_css_set(&tmp_links, cset, c);
664 BUG_ON(!list_empty(&tmp_links));
668 /* Add this cgroup group to the hash table */
669 key = css_set_hash(cset->subsys);
670 hash_add(css_set_table, &cset->hlist, key);
672 up_write(&css_set_rwsem);
677 static struct cgroupfs_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
679 struct cgroup *top_cgrp = kf_root->kn->priv;
681 return top_cgrp->root;
684 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
688 lockdep_assert_held(&cgroup_mutex);
690 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
695 root->hierarchy_id = id;
699 static void cgroup_exit_root_id(struct cgroupfs_root *root)
701 lockdep_assert_held(&cgroup_mutex);
703 if (root->hierarchy_id) {
704 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
705 root->hierarchy_id = 0;
709 static void cgroup_free_root(struct cgroupfs_root *root)
712 /* hierarhcy ID shoulid already have been released */
713 WARN_ON_ONCE(root->hierarchy_id);
715 idr_destroy(&root->cgroup_idr);
720 static void cgroup_destroy_root(struct cgroupfs_root *root)
722 struct cgroup *cgrp = &root->top_cgroup;
723 struct cgrp_cset_link *link, *tmp_link;
725 mutex_lock(&cgroup_tree_mutex);
726 mutex_lock(&cgroup_mutex);
728 BUG_ON(atomic_read(&root->nr_cgrps));
729 BUG_ON(!list_empty(&cgrp->children));
731 /* Rebind all subsystems back to the default hierarchy */
732 WARN_ON(rebind_subsystems(root, 0, root->subsys_mask));
735 * Release all the links from cset_links to this hierarchy's
738 down_write(&css_set_rwsem);
740 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
741 list_del(&link->cset_link);
742 list_del(&link->cgrp_link);
745 up_write(&css_set_rwsem);
747 if (!list_empty(&root->root_list)) {
748 list_del(&root->root_list);
752 cgroup_exit_root_id(root);
754 mutex_unlock(&cgroup_mutex);
755 mutex_unlock(&cgroup_tree_mutex);
757 kernfs_destroy_root(root->kf_root);
758 cgroup_free_root(root);
762 * Return the cgroup for "task" from the given hierarchy. Must be
763 * called with cgroup_mutex and css_set_rwsem held.
765 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
766 struct cgroupfs_root *root)
768 struct css_set *cset;
769 struct cgroup *res = NULL;
771 lockdep_assert_held(&cgroup_mutex);
772 lockdep_assert_held(&css_set_rwsem);
775 * No need to lock the task - since we hold cgroup_mutex the
776 * task can't change groups, so the only thing that can happen
777 * is that it exits and its css is set back to init_css_set.
779 cset = task_css_set(task);
780 if (cset == &init_css_set) {
781 res = &root->top_cgroup;
783 struct cgrp_cset_link *link;
785 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
786 struct cgroup *c = link->cgrp;
788 if (c->root == root) {
800 * There is one global cgroup mutex. We also require taking
801 * task_lock() when dereferencing a task's cgroup subsys pointers.
802 * See "The task_lock() exception", at the end of this comment.
804 * A task must hold cgroup_mutex to modify cgroups.
806 * Any task can increment and decrement the count field without lock.
807 * So in general, code holding cgroup_mutex can't rely on the count
808 * field not changing. However, if the count goes to zero, then only
809 * cgroup_attach_task() can increment it again. Because a count of zero
810 * means that no tasks are currently attached, therefore there is no
811 * way a task attached to that cgroup can fork (the other way to
812 * increment the count). So code holding cgroup_mutex can safely
813 * assume that if the count is zero, it will stay zero. Similarly, if
814 * a task holds cgroup_mutex on a cgroup with zero count, it
815 * knows that the cgroup won't be removed, as cgroup_rmdir()
818 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
819 * (usually) take cgroup_mutex. These are the two most performance
820 * critical pieces of code here. The exception occurs on cgroup_exit(),
821 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
822 * is taken, and if the cgroup count is zero, a usermode call made
823 * to the release agent with the name of the cgroup (path relative to
824 * the root of cgroup file system) as the argument.
826 * A cgroup can only be deleted if both its 'count' of using tasks
827 * is zero, and its list of 'children' cgroups is empty. Since all
828 * tasks in the system use _some_ cgroup, and since there is always at
829 * least one task in the system (init, pid == 1), therefore, top_cgroup
830 * always has either children cgroups and/or using tasks. So we don't
831 * need a special hack to ensure that top_cgroup cannot be deleted.
833 * The task_lock() exception
835 * The need for this exception arises from the action of
836 * cgroup_attach_task(), which overwrites one task's cgroup pointer with
837 * another. It does so using cgroup_mutex, however there are
838 * several performance critical places that need to reference
839 * task->cgroup without the expense of grabbing a system global
840 * mutex. Therefore except as noted below, when dereferencing or, as
841 * in cgroup_attach_task(), modifying a task's cgroup pointer we use
842 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
843 * the task_struct routinely used for such matters.
845 * P.S. One more locking exception. RCU is used to guard the
846 * update of a tasks cgroup pointer by cgroup_attach_task()
849 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
850 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
851 static const struct file_operations proc_cgroupstats_operations;
853 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
856 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
857 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
858 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
859 cft->ss->name, cft->name);
861 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
866 * cgroup_file_mode - deduce file mode of a control file
867 * @cft: the control file in question
869 * returns cft->mode if ->mode is not 0
870 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
871 * returns S_IRUGO if it has only a read handler
872 * returns S_IWUSR if it has only a write hander
874 static umode_t cgroup_file_mode(const struct cftype *cft)
881 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
884 if (cft->write_u64 || cft->write_s64 || cft->write_string ||
891 static void cgroup_free_fn(struct work_struct *work)
893 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
895 atomic_dec(&cgrp->root->nr_cgrps);
896 cgroup_pidlist_destroy_all(cgrp);
900 * We get a ref to the parent, and put the ref when this
901 * cgroup is being freed, so it's guaranteed that the
902 * parent won't be destroyed before its children.
904 cgroup_put(cgrp->parent);
905 kernfs_put(cgrp->kn);
909 * This is top cgroup's refcnt reaching zero, which
910 * indicates that the root should be released.
912 cgroup_destroy_root(cgrp->root);
916 static void cgroup_free_rcu(struct rcu_head *head)
918 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
920 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
921 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
924 static void cgroup_get(struct cgroup *cgrp)
926 WARN_ON_ONCE(cgroup_is_dead(cgrp));
927 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
928 atomic_inc(&cgrp->refcnt);
931 static void cgroup_put(struct cgroup *cgrp)
933 if (!atomic_dec_and_test(&cgrp->refcnt))
935 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
939 * XXX: cgrp->id is only used to look up css's. As cgroup and
940 * css's lifetimes will be decoupled, it should be made
941 * per-subsystem and moved to css->id so that lookups are
942 * successful until the target css is released.
944 mutex_lock(&cgroup_mutex);
945 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
946 mutex_unlock(&cgroup_mutex);
949 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
952 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
954 char name[CGROUP_FILE_NAME_MAX];
956 lockdep_assert_held(&cgroup_tree_mutex);
957 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
961 * cgroup_clear_dir - remove subsys files in a cgroup directory
962 * @cgrp: target cgroup
963 * @subsys_mask: mask of the subsystem ids whose files should be removed
965 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
967 struct cgroup_subsys *ss;
970 for_each_subsys(ss, i) {
973 if (!test_bit(i, &subsys_mask))
975 list_for_each_entry(cfts, &ss->cfts, node)
976 cgroup_addrm_files(cgrp, cfts, false);
980 static int rebind_subsystems(struct cgroupfs_root *root,
981 unsigned long added_mask, unsigned removed_mask)
983 struct cgroup *cgrp = &root->top_cgroup;
984 struct cgroup_subsys *ss;
987 lockdep_assert_held(&cgroup_tree_mutex);
988 lockdep_assert_held(&cgroup_mutex);
990 /* Check that any added subsystems are currently free */
991 for_each_subsys(ss, i)
992 if ((added_mask & (1 << i)) && ss->root != &cgroup_dummy_root)
995 ret = cgroup_populate_dir(cgrp, added_mask);
1000 * Nothing can fail from this point on. Remove files for the
1001 * removed subsystems and rebind each subsystem.
1003 mutex_unlock(&cgroup_mutex);
1004 cgroup_clear_dir(cgrp, removed_mask);
1005 mutex_lock(&cgroup_mutex);
1007 for_each_subsys(ss, i) {
1008 unsigned long bit = 1UL << i;
1010 if (bit & added_mask) {
1011 /* We're binding this subsystem to this hierarchy */
1012 BUG_ON(cgroup_css(cgrp, ss));
1013 BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1014 BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1016 rcu_assign_pointer(cgrp->subsys[i],
1017 cgroup_css(cgroup_dummy_top, ss));
1018 cgroup_css(cgrp, ss)->cgroup = cgrp;
1022 ss->bind(cgroup_css(cgrp, ss));
1024 /* refcount was already taken, and we're keeping it */
1025 root->subsys_mask |= bit;
1026 } else if (bit & removed_mask) {
1027 /* We're removing this subsystem */
1028 BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1029 BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1032 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1034 cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1035 RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1037 cgroup_subsys[i]->root = &cgroup_dummy_root;
1038 root->subsys_mask &= ~bit;
1042 kernfs_activate(cgrp->kn);
1046 static int cgroup_show_options(struct seq_file *seq,
1047 struct kernfs_root *kf_root)
1049 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1050 struct cgroup_subsys *ss;
1053 for_each_subsys(ss, ssid)
1054 if (root->subsys_mask & (1 << ssid))
1055 seq_printf(seq, ",%s", ss->name);
1056 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1057 seq_puts(seq, ",sane_behavior");
1058 if (root->flags & CGRP_ROOT_NOPREFIX)
1059 seq_puts(seq, ",noprefix");
1060 if (root->flags & CGRP_ROOT_XATTR)
1061 seq_puts(seq, ",xattr");
1063 spin_lock(&release_agent_path_lock);
1064 if (strlen(root->release_agent_path))
1065 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1066 spin_unlock(&release_agent_path_lock);
1068 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1069 seq_puts(seq, ",clone_children");
1070 if (strlen(root->name))
1071 seq_printf(seq, ",name=%s", root->name);
1075 struct cgroup_sb_opts {
1076 unsigned long subsys_mask;
1077 unsigned long flags;
1078 char *release_agent;
1079 bool cpuset_clone_children;
1081 /* User explicitly requested empty subsystem */
1086 * Convert a hierarchy specifier into a bitmask of subsystems and
1087 * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1088 * array. This function takes refcounts on subsystems to be used, unless it
1089 * returns error, in which case no refcounts are taken.
1091 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1093 char *token, *o = data;
1094 bool all_ss = false, one_ss = false;
1095 unsigned long mask = (unsigned long)-1;
1096 struct cgroup_subsys *ss;
1099 BUG_ON(!mutex_is_locked(&cgroup_mutex));
1101 #ifdef CONFIG_CPUSETS
1102 mask = ~(1UL << cpuset_cgrp_id);
1105 memset(opts, 0, sizeof(*opts));
1107 while ((token = strsep(&o, ",")) != NULL) {
1110 if (!strcmp(token, "none")) {
1111 /* Explicitly have no subsystems */
1115 if (!strcmp(token, "all")) {
1116 /* Mutually exclusive option 'all' + subsystem name */
1122 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1123 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1126 if (!strcmp(token, "noprefix")) {
1127 opts->flags |= CGRP_ROOT_NOPREFIX;
1130 if (!strcmp(token, "clone_children")) {
1131 opts->cpuset_clone_children = true;
1134 if (!strcmp(token, "xattr")) {
1135 opts->flags |= CGRP_ROOT_XATTR;
1138 if (!strncmp(token, "release_agent=", 14)) {
1139 /* Specifying two release agents is forbidden */
1140 if (opts->release_agent)
1142 opts->release_agent =
1143 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1144 if (!opts->release_agent)
1148 if (!strncmp(token, "name=", 5)) {
1149 const char *name = token + 5;
1150 /* Can't specify an empty name */
1153 /* Must match [\w.-]+ */
1154 for (i = 0; i < strlen(name); i++) {
1158 if ((c == '.') || (c == '-') || (c == '_'))
1162 /* Specifying two names is forbidden */
1165 opts->name = kstrndup(name,
1166 MAX_CGROUP_ROOT_NAMELEN - 1,
1174 for_each_subsys(ss, i) {
1175 if (strcmp(token, ss->name))
1180 /* Mutually exclusive option 'all' + subsystem name */
1183 set_bit(i, &opts->subsys_mask);
1188 if (i == CGROUP_SUBSYS_COUNT)
1193 * If the 'all' option was specified select all the subsystems,
1194 * otherwise if 'none', 'name=' and a subsystem name options
1195 * were not specified, let's default to 'all'
1197 if (all_ss || (!one_ss && !opts->none && !opts->name))
1198 for_each_subsys(ss, i)
1200 set_bit(i, &opts->subsys_mask);
1202 /* Consistency checks */
1204 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1205 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1207 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1208 opts->cpuset_clone_children || opts->release_agent ||
1210 pr_err("cgroup: sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1216 * Option noprefix was introduced just for backward compatibility
1217 * with the old cpuset, so we allow noprefix only if mounting just
1218 * the cpuset subsystem.
1220 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1224 /* Can't specify "none" and some subsystems */
1225 if (opts->subsys_mask && opts->none)
1229 * We either have to specify by name or by subsystems. (So all
1230 * empty hierarchies must have a name).
1232 if (!opts->subsys_mask && !opts->name)
1238 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1241 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1242 struct cgroup_sb_opts opts;
1243 unsigned long added_mask, removed_mask;
1245 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1246 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1250 mutex_lock(&cgroup_tree_mutex);
1251 mutex_lock(&cgroup_mutex);
1253 /* See what subsystems are wanted */
1254 ret = parse_cgroupfs_options(data, &opts);
1258 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1259 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1260 task_tgid_nr(current), current->comm);
1262 added_mask = opts.subsys_mask & ~root->subsys_mask;
1263 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1265 /* Don't allow flags or name to change at remount */
1266 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1267 (opts.name && strcmp(opts.name, root->name))) {
1268 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1269 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1270 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1275 /* remounting is not allowed for populated hierarchies */
1276 if (!list_empty(&root->top_cgroup.children)) {
1281 ret = rebind_subsystems(root, added_mask, removed_mask);
1285 if (opts.release_agent) {
1286 spin_lock(&release_agent_path_lock);
1287 strcpy(root->release_agent_path, opts.release_agent);
1288 spin_unlock(&release_agent_path_lock);
1291 kfree(opts.release_agent);
1293 mutex_unlock(&cgroup_mutex);
1294 mutex_unlock(&cgroup_tree_mutex);
1299 * To reduce the fork() overhead for systems that are not actually using
1300 * their cgroups capability, we don't maintain the lists running through
1301 * each css_set to its tasks until we see the list actually used - in other
1302 * words after the first mount.
1304 static bool use_task_css_set_links __read_mostly;
1306 static void cgroup_enable_task_cg_lists(void)
1308 struct task_struct *p, *g;
1310 down_write(&css_set_rwsem);
1312 if (use_task_css_set_links)
1315 use_task_css_set_links = true;
1318 * We need tasklist_lock because RCU is not safe against
1319 * while_each_thread(). Besides, a forking task that has passed
1320 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1321 * is not guaranteed to have its child immediately visible in the
1322 * tasklist if we walk through it with RCU.
1324 read_lock(&tasklist_lock);
1325 do_each_thread(g, p) {
1328 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1329 task_css_set(p) != &init_css_set);
1332 * We should check if the process is exiting, otherwise
1333 * it will race with cgroup_exit() in that the list
1334 * entry won't be deleted though the process has exited.
1336 if (!(p->flags & PF_EXITING))
1337 list_add(&p->cg_list, &task_css_set(p)->tasks);
1340 } while_each_thread(g, p);
1341 read_unlock(&tasklist_lock);
1343 up_write(&css_set_rwsem);
1346 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1348 atomic_set(&cgrp->refcnt, 1);
1349 INIT_LIST_HEAD(&cgrp->sibling);
1350 INIT_LIST_HEAD(&cgrp->children);
1351 INIT_LIST_HEAD(&cgrp->cset_links);
1352 INIT_LIST_HEAD(&cgrp->release_list);
1353 INIT_LIST_HEAD(&cgrp->pidlists);
1354 mutex_init(&cgrp->pidlist_mutex);
1355 cgrp->dummy_css.cgroup = cgrp;
1358 static void init_cgroup_root(struct cgroupfs_root *root)
1360 struct cgroup *cgrp = &root->top_cgroup;
1362 INIT_LIST_HEAD(&root->root_list);
1363 atomic_set(&root->nr_cgrps, 1);
1365 init_cgroup_housekeeping(cgrp);
1366 idr_init(&root->cgroup_idr);
1369 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1371 struct cgroupfs_root *root;
1373 if (!opts->subsys_mask && !opts->none)
1374 return ERR_PTR(-EINVAL);
1376 root = kzalloc(sizeof(*root), GFP_KERNEL);
1378 return ERR_PTR(-ENOMEM);
1380 init_cgroup_root(root);
1382 root->flags = opts->flags;
1383 if (opts->release_agent)
1384 strcpy(root->release_agent_path, opts->release_agent);
1386 strcpy(root->name, opts->name);
1387 if (opts->cpuset_clone_children)
1388 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1392 static int cgroup_setup_root(struct cgroupfs_root *root, unsigned long ss_mask)
1394 LIST_HEAD(tmp_links);
1395 struct cgroup *root_cgrp = &root->top_cgroup;
1396 struct css_set *cset;
1399 lockdep_assert_held(&cgroup_tree_mutex);
1400 lockdep_assert_held(&cgroup_mutex);
1402 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1405 root_cgrp->id = ret;
1408 * We're accessing css_set_count without locking css_set_rwsem here,
1409 * but that's OK - it can only be increased by someone holding
1410 * cgroup_lock, and that's us. The worst that can happen is that we
1411 * have some link structures left over
1413 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1417 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1418 ret = cgroup_init_root_id(root, 2, 0);
1422 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1423 KERNFS_ROOT_CREATE_DEACTIVATED,
1425 if (IS_ERR(root->kf_root)) {
1426 ret = PTR_ERR(root->kf_root);
1429 root_cgrp->kn = root->kf_root->kn;
1431 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1435 ret = rebind_subsystems(root, ss_mask, 0);
1440 * There must be no failure case after here, since rebinding takes
1441 * care of subsystems' refcounts, which are explicitly dropped in
1442 * the failure exit path.
1444 list_add(&root->root_list, &cgroup_roots);
1445 cgroup_root_count++;
1448 * Link the top cgroup in this hierarchy into all the css_set
1451 down_write(&css_set_rwsem);
1452 hash_for_each(css_set_table, i, cset, hlist)
1453 link_css_set(&tmp_links, cset, root_cgrp);
1454 up_write(&css_set_rwsem);
1456 BUG_ON(!list_empty(&root_cgrp->children));
1457 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1459 kernfs_activate(root_cgrp->kn);
1464 kernfs_destroy_root(root->kf_root);
1465 root->kf_root = NULL;
1467 cgroup_exit_root_id(root);
1469 free_cgrp_cset_links(&tmp_links);
1473 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1474 int flags, const char *unused_dev_name,
1477 struct cgroupfs_root *root;
1478 struct cgroup_sb_opts opts;
1479 struct dentry *dentry;
1483 * The first time anyone tries to mount a cgroup, enable the list
1484 * linking each css_set to its tasks and fix up all existing tasks.
1486 if (!use_task_css_set_links)
1487 cgroup_enable_task_cg_lists();
1489 mutex_lock(&cgroup_tree_mutex);
1490 mutex_lock(&cgroup_mutex);
1492 /* First find the desired set of subsystems */
1493 ret = parse_cgroupfs_options(data, &opts);
1497 /* look for a matching existing root */
1498 for_each_active_root(root) {
1499 bool name_match = false;
1502 * If we asked for a name then it must match. Also, if
1503 * name matches but sybsys_mask doesn't, we should fail.
1504 * Remember whether name matched.
1507 if (strcmp(opts.name, root->name))
1513 * If we asked for subsystems (or explicitly for no
1514 * subsystems) then they must match.
1516 if ((opts.subsys_mask || opts.none) &&
1517 (opts.subsys_mask != root->subsys_mask)) {
1524 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1525 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1526 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1530 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1535 * A root's lifetime is governed by its top cgroup. Zero
1536 * ref indicate that the root is being destroyed. Wait for
1537 * destruction to complete so that the subsystems are free.
1538 * We can use wait_queue for the wait but this path is
1539 * super cold. Let's just sleep for a bit and retry.
1541 if (!atomic_inc_not_zero(&root->top_cgroup.refcnt)) {
1542 mutex_unlock(&cgroup_mutex);
1543 mutex_unlock(&cgroup_tree_mutex);
1552 /* no such thing, create a new one */
1553 root = cgroup_root_from_opts(&opts);
1555 ret = PTR_ERR(root);
1559 ret = cgroup_setup_root(root, opts.subsys_mask);
1561 cgroup_free_root(root);
1564 mutex_unlock(&cgroup_mutex);
1565 mutex_unlock(&cgroup_tree_mutex);
1567 kfree(opts.release_agent);
1571 return ERR_PTR(ret);
1573 dentry = kernfs_mount(fs_type, flags, root->kf_root);
1575 cgroup_put(&root->top_cgroup);
1579 static void cgroup_kill_sb(struct super_block *sb)
1581 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1582 struct cgroupfs_root *root = cgroup_root_from_kf(kf_root);
1584 cgroup_put(&root->top_cgroup);
1588 static struct file_system_type cgroup_fs_type = {
1590 .mount = cgroup_mount,
1591 .kill_sb = cgroup_kill_sb,
1594 static struct kobject *cgroup_kobj;
1597 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1598 * @task: target task
1599 * @buf: the buffer to write the path into
1600 * @buflen: the length of the buffer
1602 * Determine @task's cgroup on the first (the one with the lowest non-zero
1603 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1604 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1605 * cgroup controller callbacks.
1607 * Return value is the same as kernfs_path().
1609 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1611 struct cgroupfs_root *root;
1612 struct cgroup *cgrp;
1613 int hierarchy_id = 1;
1616 mutex_lock(&cgroup_mutex);
1617 down_read(&css_set_rwsem);
1619 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1622 cgrp = task_cgroup_from_root(task, root);
1623 path = cgroup_path(cgrp, buf, buflen);
1625 /* if no hierarchy exists, everyone is in "/" */
1626 if (strlcpy(buf, "/", buflen) < buflen)
1630 up_read(&css_set_rwsem);
1631 mutex_unlock(&cgroup_mutex);
1634 EXPORT_SYMBOL_GPL(task_cgroup_path);
1637 * Control Group taskset
1639 struct task_and_cgroup {
1640 struct task_struct *task;
1641 struct cgroup *cgrp;
1642 struct css_set *cset;
1645 struct cgroup_taskset {
1646 struct task_and_cgroup single;
1647 struct flex_array *tc_array;
1653 * cgroup_taskset_first - reset taskset and return the first task
1654 * @tset: taskset of interest
1656 * @tset iteration is initialized and the first task is returned.
1658 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1660 if (tset->tc_array) {
1662 return cgroup_taskset_next(tset);
1664 return tset->single.task;
1667 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1670 * cgroup_taskset_next - iterate to the next task in taskset
1671 * @tset: taskset of interest
1673 * Return the next task in @tset. Iteration must have been initialized
1674 * with cgroup_taskset_first().
1676 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1678 struct task_and_cgroup *tc;
1680 if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1683 tc = flex_array_get(tset->tc_array, tset->idx++);
1686 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1689 * cgroup_task_migrate - move a task from one cgroup to another.
1690 * @old_cgrp; the cgroup @tsk is being migrated from
1691 * @tsk: the task being migrated
1692 * @new_cset: the new css_set @tsk is being attached to
1694 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1696 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1697 struct task_struct *tsk,
1698 struct css_set *new_cset)
1700 struct css_set *old_cset;
1702 lockdep_assert_held(&cgroup_mutex);
1703 lockdep_assert_held(&css_set_rwsem);
1706 * We are synchronized through threadgroup_lock() against PF_EXITING
1707 * setting such that we can't race against cgroup_exit() changing the
1708 * css_set to init_css_set and dropping the old one.
1710 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1711 old_cset = task_css_set(tsk);
1714 rcu_assign_pointer(tsk->cgroups, new_cset);
1717 list_move(&tsk->cg_list, &new_cset->tasks);
1720 * We just gained a reference on old_cset by taking it from the
1721 * task. As trading it for new_cset is protected by cgroup_mutex,
1722 * we're safe to drop it here; it will be freed under RCU.
1724 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1725 put_css_set_locked(old_cset, false);
1729 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1730 * @cgrp: the cgroup to attach to
1731 * @leader: the task or the leader of the threadgroup to be attached
1732 * @threadgroup: attach the whole threadgroup?
1734 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1735 * task_lock of @tsk or each thread in the threadgroup individually in turn.
1737 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *leader,
1740 int ret, i, group_size;
1741 struct cgroupfs_root *root = cgrp->root;
1742 struct cgroup_subsys_state *css, *failed_css = NULL;
1743 /* threadgroup list cursor and array */
1744 struct task_struct *task;
1745 struct task_and_cgroup *tc;
1746 struct flex_array *group;
1747 struct cgroup_taskset tset = { };
1750 * step 0: in order to do expensive, possibly blocking operations for
1751 * every thread, we cannot iterate the thread group list, since it needs
1752 * rcu or tasklist locked. instead, build an array of all threads in the
1753 * group - group_rwsem prevents new threads from appearing, and if
1754 * threads exit, this will just be an over-estimate.
1757 group_size = get_nr_threads(leader);
1760 /* flex_array supports very large thread-groups better than kmalloc. */
1761 group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1764 /* pre-allocate to guarantee space while iterating in rcu read-side. */
1765 ret = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1767 goto out_free_group_list;
1771 * Prevent freeing of tasks while we take a snapshot. Tasks that are
1772 * already PF_EXITING could be freed from underneath us unless we
1773 * take an rcu_read_lock.
1775 down_read(&css_set_rwsem);
1779 struct task_and_cgroup ent;
1781 /* @task either already exited or can't exit until the end */
1782 if (task->flags & PF_EXITING)
1785 /* as per above, nr_threads may decrease, but not increase. */
1786 BUG_ON(i >= group_size);
1788 ent.cgrp = task_cgroup_from_root(task, root);
1789 /* nothing to do if this task is already in the cgroup */
1790 if (ent.cgrp == cgrp)
1793 * saying GFP_ATOMIC has no effect here because we did prealloc
1794 * earlier, but it's good form to communicate our expectations.
1796 ret = flex_array_put(group, i, &ent, GFP_ATOMIC);
1802 } while_each_thread(leader, task);
1804 up_read(&css_set_rwsem);
1805 /* remember the number of threads in the array for later. */
1807 tset.tc_array = group;
1808 tset.tc_array_len = group_size;
1810 /* methods shouldn't be called if no task is actually migrating */
1813 goto out_free_group_list;
1816 * step 1: check that we can legitimately attach to the cgroup.
1818 for_each_css(css, i, cgrp) {
1819 if (css->ss->can_attach) {
1820 ret = css->ss->can_attach(css, &tset);
1823 goto out_cancel_attach;
1829 * step 2: make sure css_sets exist for all threads to be migrated.
1830 * we use find_css_set, which allocates a new one if necessary.
1832 for (i = 0; i < group_size; i++) {
1833 struct css_set *old_cset;
1835 tc = flex_array_get(group, i);
1836 old_cset = task_css_set(tc->task);
1837 tc->cset = find_css_set(old_cset, cgrp);
1840 goto out_put_css_set_refs;
1845 * step 3: now that we're guaranteed success wrt the css_sets,
1846 * proceed to move all tasks to the new cgroup. There are no
1847 * failure cases after here, so this is the commit point.
1849 down_write(&css_set_rwsem);
1850 for (i = 0; i < group_size; i++) {
1851 tc = flex_array_get(group, i);
1852 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
1854 up_write(&css_set_rwsem);
1855 /* nothing is sensitive to fork() after this point. */
1858 * step 4: do subsystem attach callbacks.
1860 for_each_css(css, i, cgrp)
1861 if (css->ss->attach)
1862 css->ss->attach(css, &tset);
1865 * step 5: success! and cleanup
1868 out_put_css_set_refs:
1870 for (i = 0; i < group_size; i++) {
1871 tc = flex_array_get(group, i);
1874 put_css_set(tc->cset, false);
1879 for_each_css(css, i, cgrp) {
1880 if (css == failed_css)
1882 if (css->ss->cancel_attach)
1883 css->ss->cancel_attach(css, &tset);
1886 out_free_group_list:
1887 flex_array_free(group);
1892 * Find the task_struct of the task to attach by vpid and pass it along to the
1893 * function to attach either it or all tasks in its threadgroup. Will lock
1894 * cgroup_mutex and threadgroup; may take task_lock of task.
1896 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
1898 struct task_struct *tsk;
1899 const struct cred *cred = current_cred(), *tcred;
1902 if (!cgroup_lock_live_group(cgrp))
1908 tsk = find_task_by_vpid(pid);
1912 goto out_unlock_cgroup;
1915 * even if we're attaching all tasks in the thread group, we
1916 * only need to check permissions on one of them.
1918 tcred = __task_cred(tsk);
1919 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
1920 !uid_eq(cred->euid, tcred->uid) &&
1921 !uid_eq(cred->euid, tcred->suid)) {
1924 goto out_unlock_cgroup;
1930 tsk = tsk->group_leader;
1933 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
1934 * trapped in a cpuset, or RT worker may be born in a cgroup
1935 * with no rt_runtime allocated. Just say no.
1937 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
1940 goto out_unlock_cgroup;
1943 get_task_struct(tsk);
1946 threadgroup_lock(tsk);
1948 if (!thread_group_leader(tsk)) {
1950 * a race with de_thread from another thread's exec()
1951 * may strip us of our leadership, if this happens,
1952 * there is no choice but to throw this task away and
1953 * try again; this is
1954 * "double-double-toil-and-trouble-check locking".
1956 threadgroup_unlock(tsk);
1957 put_task_struct(tsk);
1958 goto retry_find_task;
1962 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
1964 threadgroup_unlock(tsk);
1966 put_task_struct(tsk);
1968 mutex_unlock(&cgroup_mutex);
1973 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1974 * @from: attach to all cgroups of a given task
1975 * @tsk: the task to be attached
1977 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1979 struct cgroupfs_root *root;
1982 mutex_lock(&cgroup_mutex);
1983 for_each_active_root(root) {
1984 struct cgroup *from_cgrp;
1986 down_read(&css_set_rwsem);
1987 from_cgrp = task_cgroup_from_root(from, root);
1988 up_read(&css_set_rwsem);
1990 retval = cgroup_attach_task(from_cgrp, tsk, false);
1994 mutex_unlock(&cgroup_mutex);
1998 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2000 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2001 struct cftype *cft, u64 pid)
2003 return attach_task_by_pid(css->cgroup, pid, false);
2006 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2007 struct cftype *cft, u64 tgid)
2009 return attach_task_by_pid(css->cgroup, tgid, true);
2012 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2013 struct cftype *cft, const char *buffer)
2015 struct cgroupfs_root *root = css->cgroup->root;
2017 BUILD_BUG_ON(sizeof(root->release_agent_path) < PATH_MAX);
2018 if (!cgroup_lock_live_group(css->cgroup))
2020 spin_lock(&release_agent_path_lock);
2021 strlcpy(root->release_agent_path, buffer,
2022 sizeof(root->release_agent_path));
2023 spin_unlock(&release_agent_path_lock);
2024 mutex_unlock(&cgroup_mutex);
2028 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2030 struct cgroup *cgrp = seq_css(seq)->cgroup;
2032 if (!cgroup_lock_live_group(cgrp))
2034 seq_puts(seq, cgrp->root->release_agent_path);
2035 seq_putc(seq, '\n');
2036 mutex_unlock(&cgroup_mutex);
2040 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2042 struct cgroup *cgrp = seq_css(seq)->cgroup;
2044 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2048 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2049 size_t nbytes, loff_t off)
2051 struct cgroup *cgrp = of->kn->parent->priv;
2052 struct cftype *cft = of->kn->priv;
2053 struct cgroup_subsys_state *css;
2057 * kernfs guarantees that a file isn't deleted with operations in
2058 * flight, which means that the matching css is and stays alive and
2059 * doesn't need to be pinned. The RCU locking is not necessary
2060 * either. It's just for the convenience of using cgroup_css().
2063 css = cgroup_css(cgrp, cft->ss);
2066 if (cft->write_string) {
2067 ret = cft->write_string(css, cft, strstrip(buf));
2068 } else if (cft->write_u64) {
2069 unsigned long long v;
2070 ret = kstrtoull(buf, 0, &v);
2072 ret = cft->write_u64(css, cft, v);
2073 } else if (cft->write_s64) {
2075 ret = kstrtoll(buf, 0, &v);
2077 ret = cft->write_s64(css, cft, v);
2078 } else if (cft->trigger) {
2079 ret = cft->trigger(css, (unsigned int)cft->private);
2084 return ret ?: nbytes;
2087 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2089 return seq_cft(seq)->seq_start(seq, ppos);
2092 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2094 return seq_cft(seq)->seq_next(seq, v, ppos);
2097 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2099 seq_cft(seq)->seq_stop(seq, v);
2102 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2104 struct cftype *cft = seq_cft(m);
2105 struct cgroup_subsys_state *css = seq_css(m);
2108 return cft->seq_show(m, arg);
2111 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2112 else if (cft->read_s64)
2113 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2119 static struct kernfs_ops cgroup_kf_single_ops = {
2120 .atomic_write_len = PAGE_SIZE,
2121 .write = cgroup_file_write,
2122 .seq_show = cgroup_seqfile_show,
2125 static struct kernfs_ops cgroup_kf_ops = {
2126 .atomic_write_len = PAGE_SIZE,
2127 .write = cgroup_file_write,
2128 .seq_start = cgroup_seqfile_start,
2129 .seq_next = cgroup_seqfile_next,
2130 .seq_stop = cgroup_seqfile_stop,
2131 .seq_show = cgroup_seqfile_show,
2135 * cgroup_rename - Only allow simple rename of directories in place.
2137 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2138 const char *new_name_str)
2140 struct cgroup *cgrp = kn->priv;
2143 if (kernfs_type(kn) != KERNFS_DIR)
2145 if (kn->parent != new_parent)
2149 * This isn't a proper migration and its usefulness is very
2150 * limited. Disallow if sane_behavior.
2152 if (cgroup_sane_behavior(cgrp))
2155 mutex_lock(&cgroup_tree_mutex);
2156 mutex_lock(&cgroup_mutex);
2158 ret = kernfs_rename(kn, new_parent, new_name_str);
2160 mutex_unlock(&cgroup_mutex);
2161 mutex_unlock(&cgroup_tree_mutex);
2165 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2167 char name[CGROUP_FILE_NAME_MAX];
2168 struct kernfs_node *kn;
2169 struct lock_class_key *key = NULL;
2171 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2172 key = &cft->lockdep_key;
2174 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2175 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2183 * cgroup_addrm_files - add or remove files to a cgroup directory
2184 * @cgrp: the target cgroup
2185 * @cfts: array of cftypes to be added
2186 * @is_add: whether to add or remove
2188 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2189 * For removals, this function never fails. If addition fails, this
2190 * function doesn't remove files already added. The caller is responsible
2193 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2199 lockdep_assert_held(&cgroup_tree_mutex);
2201 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2202 /* does cft->flags tell us to skip this file on @cgrp? */
2203 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2205 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2207 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2211 ret = cgroup_add_file(cgrp, cft);
2213 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2218 cgroup_rm_file(cgrp, cft);
2224 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2227 struct cgroup_subsys *ss = cfts[0].ss;
2228 struct cgroup *root = &ss->root->top_cgroup;
2229 struct cgroup_subsys_state *css;
2232 lockdep_assert_held(&cgroup_tree_mutex);
2234 /* don't bother if @ss isn't attached */
2235 if (ss->root == &cgroup_dummy_root)
2238 /* add/rm files for all cgroups created before */
2239 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2240 struct cgroup *cgrp = css->cgroup;
2242 if (cgroup_is_dead(cgrp))
2245 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2251 kernfs_activate(root->kn);
2255 static void cgroup_exit_cftypes(struct cftype *cfts)
2259 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2260 /* free copy for custom atomic_write_len, see init_cftypes() */
2261 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2268 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2272 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2273 struct kernfs_ops *kf_ops;
2275 WARN_ON(cft->ss || cft->kf_ops);
2278 kf_ops = &cgroup_kf_ops;
2280 kf_ops = &cgroup_kf_single_ops;
2283 * Ugh... if @cft wants a custom max_write_len, we need to
2284 * make a copy of kf_ops to set its atomic_write_len.
2286 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
2287 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
2289 cgroup_exit_cftypes(cfts);
2292 kf_ops->atomic_write_len = cft->max_write_len;
2295 cft->kf_ops = kf_ops;
2302 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
2304 lockdep_assert_held(&cgroup_tree_mutex);
2306 if (!cfts || !cfts[0].ss)
2309 list_del(&cfts->node);
2310 cgroup_apply_cftypes(cfts, false);
2311 cgroup_exit_cftypes(cfts);
2316 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2317 * @cfts: zero-length name terminated array of cftypes
2319 * Unregister @cfts. Files described by @cfts are removed from all
2320 * existing cgroups and all future cgroups won't have them either. This
2321 * function can be called anytime whether @cfts' subsys is attached or not.
2323 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2326 int cgroup_rm_cftypes(struct cftype *cfts)
2330 mutex_lock(&cgroup_tree_mutex);
2331 ret = cgroup_rm_cftypes_locked(cfts);
2332 mutex_unlock(&cgroup_tree_mutex);
2337 * cgroup_add_cftypes - add an array of cftypes to a subsystem
2338 * @ss: target cgroup subsystem
2339 * @cfts: zero-length name terminated array of cftypes
2341 * Register @cfts to @ss. Files described by @cfts are created for all
2342 * existing cgroups to which @ss is attached and all future cgroups will
2343 * have them too. This function can be called anytime whether @ss is
2346 * Returns 0 on successful registration, -errno on failure. Note that this
2347 * function currently returns 0 as long as @cfts registration is successful
2348 * even if some file creation attempts on existing cgroups fail.
2350 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2354 ret = cgroup_init_cftypes(ss, cfts);
2358 mutex_lock(&cgroup_tree_mutex);
2360 list_add_tail(&cfts->node, &ss->cfts);
2361 ret = cgroup_apply_cftypes(cfts, true);
2363 cgroup_rm_cftypes_locked(cfts);
2365 mutex_unlock(&cgroup_tree_mutex);
2368 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2371 * cgroup_task_count - count the number of tasks in a cgroup.
2372 * @cgrp: the cgroup in question
2374 * Return the number of tasks in the cgroup.
2376 static int cgroup_task_count(const struct cgroup *cgrp)
2379 struct cgrp_cset_link *link;
2381 down_read(&css_set_rwsem);
2382 list_for_each_entry(link, &cgrp->cset_links, cset_link)
2383 count += atomic_read(&link->cset->refcount);
2384 up_read(&css_set_rwsem);
2389 * css_next_child - find the next child of a given css
2390 * @pos_css: the current position (%NULL to initiate traversal)
2391 * @parent_css: css whose children to walk
2393 * This function returns the next child of @parent_css and should be called
2394 * under either cgroup_mutex or RCU read lock. The only requirement is
2395 * that @parent_css and @pos_css are accessible. The next sibling is
2396 * guaranteed to be returned regardless of their states.
2398 struct cgroup_subsys_state *
2399 css_next_child(struct cgroup_subsys_state *pos_css,
2400 struct cgroup_subsys_state *parent_css)
2402 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2403 struct cgroup *cgrp = parent_css->cgroup;
2404 struct cgroup *next;
2406 cgroup_assert_mutexes_or_rcu_locked();
2409 * @pos could already have been removed. Once a cgroup is removed,
2410 * its ->sibling.next is no longer updated when its next sibling
2411 * changes. As CGRP_DEAD assertion is serialized and happens
2412 * before the cgroup is taken off the ->sibling list, if we see it
2413 * unasserted, it's guaranteed that the next sibling hasn't
2414 * finished its grace period even if it's already removed, and thus
2415 * safe to dereference from this RCU critical section. If
2416 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2417 * to be visible as %true here.
2419 * If @pos is dead, its next pointer can't be dereferenced;
2420 * however, as each cgroup is given a monotonically increasing
2421 * unique serial number and always appended to the sibling list,
2422 * the next one can be found by walking the parent's children until
2423 * we see a cgroup with higher serial number than @pos's. While
2424 * this path can be slower, it's taken only when either the current
2425 * cgroup is removed or iteration and removal race.
2428 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2429 } else if (likely(!cgroup_is_dead(pos))) {
2430 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2432 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2433 if (next->serial_nr > pos->serial_nr)
2437 if (&next->sibling == &cgrp->children)
2440 return cgroup_css(next, parent_css->ss);
2442 EXPORT_SYMBOL_GPL(css_next_child);
2445 * css_next_descendant_pre - find the next descendant for pre-order walk
2446 * @pos: the current position (%NULL to initiate traversal)
2447 * @root: css whose descendants to walk
2449 * To be used by css_for_each_descendant_pre(). Find the next descendant
2450 * to visit for pre-order traversal of @root's descendants. @root is
2451 * included in the iteration and the first node to be visited.
2453 * While this function requires cgroup_mutex or RCU read locking, it
2454 * doesn't require the whole traversal to be contained in a single critical
2455 * section. This function will return the correct next descendant as long
2456 * as both @pos and @root are accessible and @pos is a descendant of @root.
2458 struct cgroup_subsys_state *
2459 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2460 struct cgroup_subsys_state *root)
2462 struct cgroup_subsys_state *next;
2464 cgroup_assert_mutexes_or_rcu_locked();
2466 /* if first iteration, visit @root */
2470 /* visit the first child if exists */
2471 next = css_next_child(NULL, pos);
2475 /* no child, visit my or the closest ancestor's next sibling */
2476 while (pos != root) {
2477 next = css_next_child(pos, css_parent(pos));
2480 pos = css_parent(pos);
2485 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
2488 * css_rightmost_descendant - return the rightmost descendant of a css
2489 * @pos: css of interest
2491 * Return the rightmost descendant of @pos. If there's no descendant, @pos
2492 * is returned. This can be used during pre-order traversal to skip
2495 * While this function requires cgroup_mutex or RCU read locking, it
2496 * doesn't require the whole traversal to be contained in a single critical
2497 * section. This function will return the correct rightmost descendant as
2498 * long as @pos is accessible.
2500 struct cgroup_subsys_state *
2501 css_rightmost_descendant(struct cgroup_subsys_state *pos)
2503 struct cgroup_subsys_state *last, *tmp;
2505 cgroup_assert_mutexes_or_rcu_locked();
2509 /* ->prev isn't RCU safe, walk ->next till the end */
2511 css_for_each_child(tmp, last)
2517 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
2519 static struct cgroup_subsys_state *
2520 css_leftmost_descendant(struct cgroup_subsys_state *pos)
2522 struct cgroup_subsys_state *last;
2526 pos = css_next_child(NULL, pos);
2533 * css_next_descendant_post - find the next descendant for post-order walk
2534 * @pos: the current position (%NULL to initiate traversal)
2535 * @root: css whose descendants to walk
2537 * To be used by css_for_each_descendant_post(). Find the next descendant
2538 * to visit for post-order traversal of @root's descendants. @root is
2539 * included in the iteration and the last node to be visited.
2541 * While this function requires cgroup_mutex or RCU read locking, it
2542 * doesn't require the whole traversal to be contained in a single critical
2543 * section. This function will return the correct next descendant as long
2544 * as both @pos and @cgroup are accessible and @pos is a descendant of
2547 struct cgroup_subsys_state *
2548 css_next_descendant_post(struct cgroup_subsys_state *pos,
2549 struct cgroup_subsys_state *root)
2551 struct cgroup_subsys_state *next;
2553 cgroup_assert_mutexes_or_rcu_locked();
2555 /* if first iteration, visit leftmost descendant which may be @root */
2557 return css_leftmost_descendant(root);
2559 /* if we visited @root, we're done */
2563 /* if there's an unvisited sibling, visit its leftmost descendant */
2564 next = css_next_child(pos, css_parent(pos));
2566 return css_leftmost_descendant(next);
2568 /* no sibling left, visit parent */
2569 return css_parent(pos);
2571 EXPORT_SYMBOL_GPL(css_next_descendant_post);
2574 * css_advance_task_iter - advance a task itererator to the next css_set
2575 * @it: the iterator to advance
2577 * Advance @it to the next css_set to walk.
2579 static void css_advance_task_iter(struct css_task_iter *it)
2581 struct list_head *l = it->cset_link;
2582 struct cgrp_cset_link *link;
2583 struct css_set *cset;
2585 /* Advance to the next non-empty css_set */
2588 if (l == &it->origin_css->cgroup->cset_links) {
2589 it->cset_link = NULL;
2592 link = list_entry(l, struct cgrp_cset_link, cset_link);
2594 } while (list_empty(&cset->tasks));
2596 it->task = cset->tasks.next;
2600 * css_task_iter_start - initiate task iteration
2601 * @css: the css to walk tasks of
2602 * @it: the task iterator to use
2604 * Initiate iteration through the tasks of @css. The caller can call
2605 * css_task_iter_next() to walk through the tasks until the function
2606 * returns NULL. On completion of iteration, css_task_iter_end() must be
2609 * Note that this function acquires a lock which is released when the
2610 * iteration finishes. The caller can't sleep while iteration is in
2613 void css_task_iter_start(struct cgroup_subsys_state *css,
2614 struct css_task_iter *it)
2615 __acquires(css_set_rwsem)
2617 /* no one should try to iterate before mounting cgroups */
2618 WARN_ON_ONCE(!use_task_css_set_links);
2620 down_read(&css_set_rwsem);
2622 it->origin_css = css;
2623 it->cset_link = &css->cgroup->cset_links;
2625 css_advance_task_iter(it);
2629 * css_task_iter_next - return the next task for the iterator
2630 * @it: the task iterator being iterated
2632 * The "next" function for task iteration. @it should have been
2633 * initialized via css_task_iter_start(). Returns NULL when the iteration
2636 struct task_struct *css_task_iter_next(struct css_task_iter *it)
2638 struct task_struct *res;
2639 struct list_head *l = it->task;
2640 struct cgrp_cset_link *link;
2642 /* If the iterator cg is NULL, we have no tasks */
2645 res = list_entry(l, struct task_struct, cg_list);
2646 /* Advance iterator to find next entry */
2648 link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
2649 if (l == &link->cset->tasks) {
2651 * We reached the end of this task list - move on to the
2652 * next cgrp_cset_link.
2654 css_advance_task_iter(it);
2662 * css_task_iter_end - finish task iteration
2663 * @it: the task iterator to finish
2665 * Finish task iteration started by css_task_iter_start().
2667 void css_task_iter_end(struct css_task_iter *it)
2668 __releases(css_set_rwsem)
2670 up_read(&css_set_rwsem);
2674 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
2675 * @to: cgroup to which the tasks will be moved
2676 * @from: cgroup in which the tasks currently reside
2678 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
2680 struct css_task_iter it;
2681 struct task_struct *task;
2685 css_task_iter_start(&from->dummy_css, &it);
2686 task = css_task_iter_next(&it);
2688 get_task_struct(task);
2689 css_task_iter_end(&it);
2692 mutex_lock(&cgroup_mutex);
2693 ret = cgroup_attach_task(to, task, false);
2694 mutex_unlock(&cgroup_mutex);
2695 put_task_struct(task);
2697 } while (task && !ret);
2703 * Stuff for reading the 'tasks'/'procs' files.
2705 * Reading this file can return large amounts of data if a cgroup has
2706 * *lots* of attached tasks. So it may need several calls to read(),
2707 * but we cannot guarantee that the information we produce is correct
2708 * unless we produce it entirely atomically.
2712 /* which pidlist file are we talking about? */
2713 enum cgroup_filetype {
2719 * A pidlist is a list of pids that virtually represents the contents of one
2720 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
2721 * a pair (one each for procs, tasks) for each pid namespace that's relevant
2724 struct cgroup_pidlist {
2726 * used to find which pidlist is wanted. doesn't change as long as
2727 * this particular list stays in the list.
2729 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
2732 /* how many elements the above list has */
2734 /* each of these stored in a list by its cgroup */
2735 struct list_head links;
2736 /* pointer to the cgroup we belong to, for list removal purposes */
2737 struct cgroup *owner;
2738 /* for delayed destruction */
2739 struct delayed_work destroy_dwork;
2743 * The following two functions "fix" the issue where there are more pids
2744 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
2745 * TODO: replace with a kernel-wide solution to this problem
2747 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
2748 static void *pidlist_allocate(int count)
2750 if (PIDLIST_TOO_LARGE(count))
2751 return vmalloc(count * sizeof(pid_t));
2753 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
2756 static void pidlist_free(void *p)
2758 if (is_vmalloc_addr(p))
2765 * Used to destroy all pidlists lingering waiting for destroy timer. None
2766 * should be left afterwards.
2768 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
2770 struct cgroup_pidlist *l, *tmp_l;
2772 mutex_lock(&cgrp->pidlist_mutex);
2773 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
2774 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
2775 mutex_unlock(&cgrp->pidlist_mutex);
2777 flush_workqueue(cgroup_pidlist_destroy_wq);
2778 BUG_ON(!list_empty(&cgrp->pidlists));
2781 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
2783 struct delayed_work *dwork = to_delayed_work(work);
2784 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
2786 struct cgroup_pidlist *tofree = NULL;
2788 mutex_lock(&l->owner->pidlist_mutex);
2791 * Destroy iff we didn't get queued again. The state won't change
2792 * as destroy_dwork can only be queued while locked.
2794 if (!delayed_work_pending(dwork)) {
2795 list_del(&l->links);
2796 pidlist_free(l->list);
2797 put_pid_ns(l->key.ns);
2801 mutex_unlock(&l->owner->pidlist_mutex);
2806 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
2807 * Returns the number of unique elements.
2809 static int pidlist_uniq(pid_t *list, int length)
2814 * we presume the 0th element is unique, so i starts at 1. trivial
2815 * edge cases first; no work needs to be done for either
2817 if (length == 0 || length == 1)
2819 /* src and dest walk down the list; dest counts unique elements */
2820 for (src = 1; src < length; src++) {
2821 /* find next unique element */
2822 while (list[src] == list[src-1]) {
2827 /* dest always points to where the next unique element goes */
2828 list[dest] = list[src];
2836 * The two pid files - task and cgroup.procs - guaranteed that the result
2837 * is sorted, which forced this whole pidlist fiasco. As pid order is
2838 * different per namespace, each namespace needs differently sorted list,
2839 * making it impossible to use, for example, single rbtree of member tasks
2840 * sorted by task pointer. As pidlists can be fairly large, allocating one
2841 * per open file is dangerous, so cgroup had to implement shared pool of
2842 * pidlists keyed by cgroup and namespace.
2844 * All this extra complexity was caused by the original implementation
2845 * committing to an entirely unnecessary property. In the long term, we
2846 * want to do away with it. Explicitly scramble sort order if
2847 * sane_behavior so that no such expectation exists in the new interface.
2849 * Scrambling is done by swapping every two consecutive bits, which is
2850 * non-identity one-to-one mapping which disturbs sort order sufficiently.
2852 static pid_t pid_fry(pid_t pid)
2854 unsigned a = pid & 0x55555555;
2855 unsigned b = pid & 0xAAAAAAAA;
2857 return (a << 1) | (b >> 1);
2860 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
2862 if (cgroup_sane_behavior(cgrp))
2863 return pid_fry(pid);
2868 static int cmppid(const void *a, const void *b)
2870 return *(pid_t *)a - *(pid_t *)b;
2873 static int fried_cmppid(const void *a, const void *b)
2875 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
2878 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
2879 enum cgroup_filetype type)
2881 struct cgroup_pidlist *l;
2882 /* don't need task_nsproxy() if we're looking at ourself */
2883 struct pid_namespace *ns = task_active_pid_ns(current);
2885 lockdep_assert_held(&cgrp->pidlist_mutex);
2887 list_for_each_entry(l, &cgrp->pidlists, links)
2888 if (l->key.type == type && l->key.ns == ns)
2894 * find the appropriate pidlist for our purpose (given procs vs tasks)
2895 * returns with the lock on that pidlist already held, and takes care
2896 * of the use count, or returns NULL with no locks held if we're out of
2899 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
2900 enum cgroup_filetype type)
2902 struct cgroup_pidlist *l;
2904 lockdep_assert_held(&cgrp->pidlist_mutex);
2906 l = cgroup_pidlist_find(cgrp, type);
2910 /* entry not found; create a new one */
2911 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
2915 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
2917 /* don't need task_nsproxy() if we're looking at ourself */
2918 l->key.ns = get_pid_ns(task_active_pid_ns(current));
2920 list_add(&l->links, &cgrp->pidlists);
2925 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
2927 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
2928 struct cgroup_pidlist **lp)
2932 int pid, n = 0; /* used for populating the array */
2933 struct css_task_iter it;
2934 struct task_struct *tsk;
2935 struct cgroup_pidlist *l;
2937 lockdep_assert_held(&cgrp->pidlist_mutex);
2940 * If cgroup gets more users after we read count, we won't have
2941 * enough space - tough. This race is indistinguishable to the
2942 * caller from the case that the additional cgroup users didn't
2943 * show up until sometime later on.
2945 length = cgroup_task_count(cgrp);
2946 array = pidlist_allocate(length);
2949 /* now, populate the array */
2950 css_task_iter_start(&cgrp->dummy_css, &it);
2951 while ((tsk = css_task_iter_next(&it))) {
2952 if (unlikely(n == length))
2954 /* get tgid or pid for procs or tasks file respectively */
2955 if (type == CGROUP_FILE_PROCS)
2956 pid = task_tgid_vnr(tsk);
2958 pid = task_pid_vnr(tsk);
2959 if (pid > 0) /* make sure to only use valid results */
2962 css_task_iter_end(&it);
2964 /* now sort & (if procs) strip out duplicates */
2965 if (cgroup_sane_behavior(cgrp))
2966 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
2968 sort(array, length, sizeof(pid_t), cmppid, NULL);
2969 if (type == CGROUP_FILE_PROCS)
2970 length = pidlist_uniq(array, length);
2972 l = cgroup_pidlist_find_create(cgrp, type);
2974 mutex_unlock(&cgrp->pidlist_mutex);
2975 pidlist_free(array);
2979 /* store array, freeing old if necessary */
2980 pidlist_free(l->list);
2988 * cgroupstats_build - build and fill cgroupstats
2989 * @stats: cgroupstats to fill information into
2990 * @dentry: A dentry entry belonging to the cgroup for which stats have
2993 * Build and fill cgroupstats so that taskstats can export it to user
2996 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2998 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
2999 struct cgroup *cgrp;
3000 struct css_task_iter it;
3001 struct task_struct *tsk;
3003 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3004 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3005 kernfs_type(kn) != KERNFS_DIR)
3009 * We aren't being called from kernfs and there's no guarantee on
3010 * @kn->priv's validity. For this and css_tryget_from_dir(),
3011 * @kn->priv is RCU safe. Let's do the RCU dancing.
3014 cgrp = rcu_dereference(kn->priv);
3020 css_task_iter_start(&cgrp->dummy_css, &it);
3021 while ((tsk = css_task_iter_next(&it))) {
3022 switch (tsk->state) {
3024 stats->nr_running++;
3026 case TASK_INTERRUPTIBLE:
3027 stats->nr_sleeping++;
3029 case TASK_UNINTERRUPTIBLE:
3030 stats->nr_uninterruptible++;
3033 stats->nr_stopped++;
3036 if (delayacct_is_task_waiting_on_io(tsk))
3037 stats->nr_io_wait++;
3041 css_task_iter_end(&it);
3049 * seq_file methods for the tasks/procs files. The seq_file position is the
3050 * next pid to display; the seq_file iterator is a pointer to the pid
3051 * in the cgroup->l->list array.
3054 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3057 * Initially we receive a position value that corresponds to
3058 * one more than the last pid shown (or 0 on the first call or
3059 * after a seek to the start). Use a binary-search to find the
3060 * next pid to display, if any
3062 struct kernfs_open_file *of = s->private;
3063 struct cgroup *cgrp = seq_css(s)->cgroup;
3064 struct cgroup_pidlist *l;
3065 enum cgroup_filetype type = seq_cft(s)->private;
3066 int index = 0, pid = *pos;
3069 mutex_lock(&cgrp->pidlist_mutex);
3072 * !NULL @of->priv indicates that this isn't the first start()
3073 * after open. If the matching pidlist is around, we can use that.
3074 * Look for it. Note that @of->priv can't be used directly. It
3075 * could already have been destroyed.
3078 of->priv = cgroup_pidlist_find(cgrp, type);
3081 * Either this is the first start() after open or the matching
3082 * pidlist has been destroyed inbetween. Create a new one.
3085 ret = pidlist_array_load(cgrp, type,
3086 (struct cgroup_pidlist **)&of->priv);
3088 return ERR_PTR(ret);
3093 int end = l->length;
3095 while (index < end) {
3096 int mid = (index + end) / 2;
3097 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3100 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3106 /* If we're off the end of the array, we're done */
3107 if (index >= l->length)
3109 /* Update the abstract position to be the actual pid that we found */
3110 iter = l->list + index;
3111 *pos = cgroup_pid_fry(cgrp, *iter);
3115 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3117 struct kernfs_open_file *of = s->private;
3118 struct cgroup_pidlist *l = of->priv;
3121 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3122 CGROUP_PIDLIST_DESTROY_DELAY);
3123 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3126 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3128 struct kernfs_open_file *of = s->private;
3129 struct cgroup_pidlist *l = of->priv;
3131 pid_t *end = l->list + l->length;
3133 * Advance to the next pid in the array. If this goes off the
3140 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3145 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3147 return seq_printf(s, "%d\n", *(int *)v);
3151 * seq_operations functions for iterating on pidlists through seq_file -
3152 * independent of whether it's tasks or procs
3154 static const struct seq_operations cgroup_pidlist_seq_operations = {
3155 .start = cgroup_pidlist_start,
3156 .stop = cgroup_pidlist_stop,
3157 .next = cgroup_pidlist_next,
3158 .show = cgroup_pidlist_show,
3161 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3164 return notify_on_release(css->cgroup);
3167 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3168 struct cftype *cft, u64 val)
3170 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3172 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3174 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3178 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3181 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3184 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3185 struct cftype *cft, u64 val)
3188 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3190 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3194 static struct cftype cgroup_base_files[] = {
3196 .name = "cgroup.procs",
3197 .seq_start = cgroup_pidlist_start,
3198 .seq_next = cgroup_pidlist_next,
3199 .seq_stop = cgroup_pidlist_stop,
3200 .seq_show = cgroup_pidlist_show,
3201 .private = CGROUP_FILE_PROCS,
3202 .write_u64 = cgroup_procs_write,
3203 .mode = S_IRUGO | S_IWUSR,
3206 .name = "cgroup.clone_children",
3207 .flags = CFTYPE_INSANE,
3208 .read_u64 = cgroup_clone_children_read,
3209 .write_u64 = cgroup_clone_children_write,
3212 .name = "cgroup.sane_behavior",
3213 .flags = CFTYPE_ONLY_ON_ROOT,
3214 .seq_show = cgroup_sane_behavior_show,
3218 * Historical crazy stuff. These don't have "cgroup." prefix and
3219 * don't exist if sane_behavior. If you're depending on these, be
3220 * prepared to be burned.
3224 .flags = CFTYPE_INSANE, /* use "procs" instead */
3225 .seq_start = cgroup_pidlist_start,
3226 .seq_next = cgroup_pidlist_next,
3227 .seq_stop = cgroup_pidlist_stop,
3228 .seq_show = cgroup_pidlist_show,
3229 .private = CGROUP_FILE_TASKS,
3230 .write_u64 = cgroup_tasks_write,
3231 .mode = S_IRUGO | S_IWUSR,
3234 .name = "notify_on_release",
3235 .flags = CFTYPE_INSANE,
3236 .read_u64 = cgroup_read_notify_on_release,
3237 .write_u64 = cgroup_write_notify_on_release,
3240 .name = "release_agent",
3241 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3242 .seq_show = cgroup_release_agent_show,
3243 .write_string = cgroup_release_agent_write,
3244 .max_write_len = PATH_MAX - 1,
3250 * cgroup_populate_dir - create subsys files in a cgroup directory
3251 * @cgrp: target cgroup
3252 * @subsys_mask: mask of the subsystem ids whose files should be added
3254 * On failure, no file is added.
3256 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3258 struct cgroup_subsys *ss;
3261 /* process cftsets of each subsystem */
3262 for_each_subsys(ss, i) {
3263 struct cftype *cfts;
3265 if (!test_bit(i, &subsys_mask))
3268 list_for_each_entry(cfts, &ss->cfts, node) {
3269 ret = cgroup_addrm_files(cgrp, cfts, true);
3276 cgroup_clear_dir(cgrp, subsys_mask);
3281 * css destruction is four-stage process.
3283 * 1. Destruction starts. Killing of the percpu_ref is initiated.
3284 * Implemented in kill_css().
3286 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3287 * and thus css_tryget() is guaranteed to fail, the css can be offlined
3288 * by invoking offline_css(). After offlining, the base ref is put.
3289 * Implemented in css_killed_work_fn().
3291 * 3. When the percpu_ref reaches zero, the only possible remaining
3292 * accessors are inside RCU read sections. css_release() schedules the
3295 * 4. After the grace period, the css can be freed. Implemented in
3296 * css_free_work_fn().
3298 * It is actually hairier because both step 2 and 4 require process context
3299 * and thus involve punting to css->destroy_work adding two additional
3300 * steps to the already complex sequence.
3302 static void css_free_work_fn(struct work_struct *work)
3304 struct cgroup_subsys_state *css =
3305 container_of(work, struct cgroup_subsys_state, destroy_work);
3306 struct cgroup *cgrp = css->cgroup;
3309 css_put(css->parent);
3311 css->ss->css_free(css);
3315 static void css_free_rcu_fn(struct rcu_head *rcu_head)
3317 struct cgroup_subsys_state *css =
3318 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
3320 INIT_WORK(&css->destroy_work, css_free_work_fn);
3321 queue_work(cgroup_destroy_wq, &css->destroy_work);
3324 static void css_release(struct percpu_ref *ref)
3326 struct cgroup_subsys_state *css =
3327 container_of(ref, struct cgroup_subsys_state, refcnt);
3329 rcu_assign_pointer(css->cgroup->subsys[css->ss->id], NULL);
3330 call_rcu(&css->rcu_head, css_free_rcu_fn);
3333 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
3334 struct cgroup *cgrp)
3341 css->parent = cgroup_css(cgrp->parent, ss);
3343 css->flags |= CSS_ROOT;
3345 BUG_ON(cgroup_css(cgrp, ss));
3348 /* invoke ->css_online() on a new CSS and mark it online if successful */
3349 static int online_css(struct cgroup_subsys_state *css)
3351 struct cgroup_subsys *ss = css->ss;
3354 lockdep_assert_held(&cgroup_tree_mutex);
3355 lockdep_assert_held(&cgroup_mutex);
3358 ret = ss->css_online(css);
3360 css->flags |= CSS_ONLINE;
3361 css->cgroup->nr_css++;
3362 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
3367 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
3368 static void offline_css(struct cgroup_subsys_state *css)
3370 struct cgroup_subsys *ss = css->ss;
3372 lockdep_assert_held(&cgroup_tree_mutex);
3373 lockdep_assert_held(&cgroup_mutex);
3375 if (!(css->flags & CSS_ONLINE))
3378 if (ss->css_offline)
3379 ss->css_offline(css);
3381 css->flags &= ~CSS_ONLINE;
3382 css->cgroup->nr_css--;
3383 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], css);
3387 * create_css - create a cgroup_subsys_state
3388 * @cgrp: the cgroup new css will be associated with
3389 * @ss: the subsys of new css
3391 * Create a new css associated with @cgrp - @ss pair. On success, the new
3392 * css is online and installed in @cgrp with all interface files created.
3393 * Returns 0 on success, -errno on failure.
3395 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
3397 struct cgroup *parent = cgrp->parent;
3398 struct cgroup_subsys_state *css;
3401 lockdep_assert_held(&cgroup_mutex);
3403 css = ss->css_alloc(cgroup_css(parent, ss));
3405 return PTR_ERR(css);
3407 err = percpu_ref_init(&css->refcnt, css_release);
3411 init_css(css, ss, cgrp);
3413 err = cgroup_populate_dir(cgrp, 1 << ss->id);
3417 err = online_css(css);
3422 css_get(css->parent);
3424 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
3426 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",
3427 current->comm, current->pid, ss->name);
3428 if (!strcmp(ss->name, "memory"))
3429 pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
3430 ss->warned_broken_hierarchy = true;
3436 percpu_ref_cancel_init(&css->refcnt);
3442 * cgroup_create - create a cgroup
3443 * @parent: cgroup that will be parent of the new cgroup
3444 * @name: name of the new cgroup
3445 * @mode: mode to set on new cgroup
3447 static long cgroup_create(struct cgroup *parent, const char *name,
3450 struct cgroup *cgrp;
3451 struct cgroupfs_root *root = parent->root;
3453 struct cgroup_subsys *ss;
3454 struct kernfs_node *kn;
3456 /* allocate the cgroup and its ID, 0 is reserved for the root */
3457 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3461 mutex_lock(&cgroup_tree_mutex);
3464 * Only live parents can have children. Note that the liveliness
3465 * check isn't strictly necessary because cgroup_mkdir() and
3466 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
3467 * anyway so that locking is contained inside cgroup proper and we
3468 * don't get nasty surprises if we ever grow another caller.
3470 if (!cgroup_lock_live_group(parent)) {
3472 goto err_unlock_tree;
3476 * Temporarily set the pointer to NULL, so idr_find() won't return
3477 * a half-baked cgroup.
3479 cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
3485 init_cgroup_housekeeping(cgrp);
3487 cgrp->parent = parent;
3488 cgrp->dummy_css.parent = &parent->dummy_css;
3489 cgrp->root = parent->root;
3491 if (notify_on_release(parent))
3492 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3494 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
3495 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
3497 /* create the directory */
3498 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
3506 * This extra ref will be put in cgroup_free_fn() and guarantees
3507 * that @cgrp->kn is always accessible.
3511 cgrp->serial_nr = cgroup_serial_nr_next++;
3513 /* allocation complete, commit to creation */
3514 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
3515 atomic_inc(&root->nr_cgrps);
3519 * @cgrp is now fully operational. If something fails after this
3520 * point, it'll be released via the normal destruction path.
3522 idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
3524 err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
3528 /* let's create and online css's */
3529 for_each_subsys(ss, ssid) {
3530 if (root->subsys_mask & (1 << ssid)) {
3531 err = create_css(cgrp, ss);
3537 kernfs_activate(kn);
3539 mutex_unlock(&cgroup_mutex);
3540 mutex_unlock(&cgroup_tree_mutex);
3545 idr_remove(&root->cgroup_idr, cgrp->id);
3547 mutex_unlock(&cgroup_mutex);
3549 mutex_unlock(&cgroup_tree_mutex);
3554 cgroup_destroy_locked(cgrp);
3555 mutex_unlock(&cgroup_mutex);
3556 mutex_unlock(&cgroup_tree_mutex);
3560 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3563 struct cgroup *parent = parent_kn->priv;
3565 return cgroup_create(parent, name, mode);
3569 * This is called when the refcnt of a css is confirmed to be killed.
3570 * css_tryget() is now guaranteed to fail.
3572 static void css_killed_work_fn(struct work_struct *work)
3574 struct cgroup_subsys_state *css =
3575 container_of(work, struct cgroup_subsys_state, destroy_work);
3576 struct cgroup *cgrp = css->cgroup;
3578 mutex_lock(&cgroup_tree_mutex);
3579 mutex_lock(&cgroup_mutex);
3582 * css_tryget() is guaranteed to fail now. Tell subsystems to
3583 * initate destruction.
3588 * If @cgrp is marked dead, it's waiting for refs of all css's to
3589 * be disabled before proceeding to the second phase of cgroup
3590 * destruction. If we are the last one, kick it off.
3592 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
3593 cgroup_destroy_css_killed(cgrp);
3595 mutex_unlock(&cgroup_mutex);
3596 mutex_unlock(&cgroup_tree_mutex);
3599 * Put the css refs from kill_css(). Each css holds an extra
3600 * reference to the cgroup's dentry and cgroup removal proceeds
3601 * regardless of css refs. On the last put of each css, whenever
3602 * that may be, the extra dentry ref is put so that dentry
3603 * destruction happens only after all css's are released.
3608 /* css kill confirmation processing requires process context, bounce */
3609 static void css_killed_ref_fn(struct percpu_ref *ref)
3611 struct cgroup_subsys_state *css =
3612 container_of(ref, struct cgroup_subsys_state, refcnt);
3614 INIT_WORK(&css->destroy_work, css_killed_work_fn);
3615 queue_work(cgroup_destroy_wq, &css->destroy_work);
3619 * kill_css - destroy a css
3620 * @css: css to destroy
3622 * This function initiates destruction of @css by removing cgroup interface
3623 * files and putting its base reference. ->css_offline() will be invoked
3624 * asynchronously once css_tryget() is guaranteed to fail and when the
3625 * reference count reaches zero, @css will be released.
3627 static void kill_css(struct cgroup_subsys_state *css)
3630 * This must happen before css is disassociated with its cgroup.
3631 * See seq_css() for details.
3633 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
3636 * Killing would put the base ref, but we need to keep it alive
3637 * until after ->css_offline().
3642 * cgroup core guarantees that, by the time ->css_offline() is
3643 * invoked, no new css reference will be given out via
3644 * css_tryget(). We can't simply call percpu_ref_kill() and
3645 * proceed to offlining css's because percpu_ref_kill() doesn't
3646 * guarantee that the ref is seen as killed on all CPUs on return.
3648 * Use percpu_ref_kill_and_confirm() to get notifications as each
3649 * css is confirmed to be seen as killed on all CPUs.
3651 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
3655 * cgroup_destroy_locked - the first stage of cgroup destruction
3656 * @cgrp: cgroup to be destroyed
3658 * css's make use of percpu refcnts whose killing latency shouldn't be
3659 * exposed to userland and are RCU protected. Also, cgroup core needs to
3660 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
3661 * invoked. To satisfy all the requirements, destruction is implemented in
3662 * the following two steps.
3664 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
3665 * userland visible parts and start killing the percpu refcnts of
3666 * css's. Set up so that the next stage will be kicked off once all
3667 * the percpu refcnts are confirmed to be killed.
3669 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
3670 * rest of destruction. Once all cgroup references are gone, the
3671 * cgroup is RCU-freed.
3673 * This function implements s1. After this step, @cgrp is gone as far as
3674 * the userland is concerned and a new cgroup with the same name may be
3675 * created. As cgroup doesn't care about the names internally, this
3676 * doesn't cause any problem.
3678 static int cgroup_destroy_locked(struct cgroup *cgrp)
3679 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
3681 struct cgroup *child;
3682 struct cgroup_subsys_state *css;
3686 lockdep_assert_held(&cgroup_tree_mutex);
3687 lockdep_assert_held(&cgroup_mutex);
3690 * css_set_rwsem synchronizes access to ->cset_links and prevents
3691 * @cgrp from being removed while put_css_set() is in progress.
3693 down_read(&css_set_rwsem);
3694 empty = list_empty(&cgrp->cset_links);
3695 up_read(&css_set_rwsem);
3700 * Make sure there's no live children. We can't test ->children
3701 * emptiness as dead children linger on it while being destroyed;
3702 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
3706 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
3707 empty = cgroup_is_dead(child);
3716 * Initiate massacre of all css's. cgroup_destroy_css_killed()
3717 * will be invoked to perform the rest of destruction once the
3718 * percpu refs of all css's are confirmed to be killed. This
3719 * involves removing the subsystem's files, drop cgroup_mutex.
3721 mutex_unlock(&cgroup_mutex);
3722 for_each_css(css, ssid, cgrp)
3724 mutex_lock(&cgroup_mutex);
3727 * Mark @cgrp dead. This prevents further task migration and child
3728 * creation by disabling cgroup_lock_live_group(). Note that
3729 * CGRP_DEAD assertion is depended upon by css_next_child() to
3730 * resume iteration after dropping RCU read lock. See
3731 * css_next_child() for details.
3733 set_bit(CGRP_DEAD, &cgrp->flags);
3735 /* CGRP_DEAD is set, remove from ->release_list for the last time */
3736 raw_spin_lock(&release_list_lock);
3737 if (!list_empty(&cgrp->release_list))
3738 list_del_init(&cgrp->release_list);
3739 raw_spin_unlock(&release_list_lock);
3742 * If @cgrp has css's attached, the second stage of cgroup
3743 * destruction is kicked off from css_killed_work_fn() after the
3744 * refs of all attached css's are killed. If @cgrp doesn't have
3745 * any css, we kick it off here.
3748 cgroup_destroy_css_killed(cgrp);
3750 /* remove @cgrp directory along with the base files */
3751 mutex_unlock(&cgroup_mutex);
3754 * There are two control paths which try to determine cgroup from
3755 * dentry without going through kernfs - cgroupstats_build() and
3756 * css_tryget_from_dir(). Those are supported by RCU protecting
3757 * clearing of cgrp->kn->priv backpointer, which should happen
3758 * after all files under it have been removed.
3760 kernfs_remove(cgrp->kn); /* @cgrp has an extra ref on its kn */
3761 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
3763 mutex_lock(&cgroup_mutex);
3769 * cgroup_destroy_css_killed - the second step of cgroup destruction
3770 * @work: cgroup->destroy_free_work
3772 * This function is invoked from a work item for a cgroup which is being
3773 * destroyed after all css's are offlined and performs the rest of
3774 * destruction. This is the second step of destruction described in the
3775 * comment above cgroup_destroy_locked().
3777 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
3779 struct cgroup *parent = cgrp->parent;
3781 lockdep_assert_held(&cgroup_tree_mutex);
3782 lockdep_assert_held(&cgroup_mutex);
3784 /* delete this cgroup from parent->children */
3785 list_del_rcu(&cgrp->sibling);
3789 set_bit(CGRP_RELEASABLE, &parent->flags);
3790 check_for_release(parent);
3793 static int cgroup_rmdir(struct kernfs_node *kn)
3795 struct cgroup *cgrp = kn->priv;
3799 * This is self-destruction but @kn can't be removed while this
3800 * callback is in progress. Let's break active protection. Once
3801 * the protection is broken, @cgrp can be destroyed at any point.
3802 * Pin it so that it stays accessible.
3805 kernfs_break_active_protection(kn);
3807 mutex_lock(&cgroup_tree_mutex);
3808 mutex_lock(&cgroup_mutex);
3811 * @cgrp might already have been destroyed while we're trying to
3814 if (!cgroup_is_dead(cgrp))
3815 ret = cgroup_destroy_locked(cgrp);
3817 mutex_unlock(&cgroup_mutex);
3818 mutex_unlock(&cgroup_tree_mutex);
3820 kernfs_unbreak_active_protection(kn);
3825 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
3826 .remount_fs = cgroup_remount,
3827 .show_options = cgroup_show_options,
3828 .mkdir = cgroup_mkdir,
3829 .rmdir = cgroup_rmdir,
3830 .rename = cgroup_rename,
3833 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
3835 struct cgroup_subsys_state *css;
3837 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
3839 mutex_lock(&cgroup_tree_mutex);
3840 mutex_lock(&cgroup_mutex);
3842 INIT_LIST_HEAD(&ss->cfts);
3844 /* Create the top cgroup state for this subsystem */
3845 ss->root = &cgroup_dummy_root;
3846 css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
3847 /* We don't handle early failures gracefully */
3848 BUG_ON(IS_ERR(css));
3849 init_css(css, ss, cgroup_dummy_top);
3851 /* Update the init_css_set to contain a subsys
3852 * pointer to this state - since the subsystem is
3853 * newly registered, all tasks and hence the
3854 * init_css_set is in the subsystem's top cgroup. */
3855 init_css_set.subsys[ss->id] = css;
3857 need_forkexit_callback |= ss->fork || ss->exit;
3859 /* At system boot, before all subsystems have been
3860 * registered, no tasks have been forked, so we don't
3861 * need to invoke fork callbacks here. */
3862 BUG_ON(!list_empty(&init_task.tasks));
3864 BUG_ON(online_css(css));
3866 mutex_unlock(&cgroup_mutex);
3867 mutex_unlock(&cgroup_tree_mutex);
3871 * cgroup_init_early - cgroup initialization at system boot
3873 * Initialize cgroups at system boot, and initialize any
3874 * subsystems that request early init.
3876 int __init cgroup_init_early(void)
3878 struct cgroup_subsys *ss;
3881 atomic_set(&init_css_set.refcount, 1);
3882 INIT_LIST_HEAD(&init_css_set.cgrp_links);
3883 INIT_LIST_HEAD(&init_css_set.tasks);
3884 INIT_HLIST_NODE(&init_css_set.hlist);
3886 init_cgroup_root(&cgroup_dummy_root);
3887 cgroup_root_count = 1;
3888 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
3890 init_cgrp_cset_link.cset = &init_css_set;
3891 init_cgrp_cset_link.cgrp = cgroup_dummy_top;
3892 list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
3893 list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
3895 for_each_subsys(ss, i) {
3896 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
3897 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
3898 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
3900 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
3901 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
3904 ss->name = cgroup_subsys_name[i];
3907 cgroup_init_subsys(ss);
3913 * cgroup_init - cgroup initialization
3915 * Register cgroup filesystem and /proc file, and initialize
3916 * any subsystems that didn't request early init.
3918 int __init cgroup_init(void)
3920 struct cgroup_subsys *ss;
3924 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
3926 for_each_subsys(ss, i) {
3927 if (!ss->early_init)
3928 cgroup_init_subsys(ss);
3931 * cftype registration needs kmalloc and can't be done
3932 * during early_init. Register base cftypes separately.
3934 if (ss->base_cftypes)
3935 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
3938 /* allocate id for the dummy hierarchy */
3939 mutex_lock(&cgroup_mutex);
3941 /* Add init_css_set to the hash table */
3942 key = css_set_hash(init_css_set.subsys);
3943 hash_add(css_set_table, &init_css_set.hlist, key);
3945 BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
3947 err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
3951 mutex_unlock(&cgroup_mutex);
3953 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
3957 err = register_filesystem(&cgroup_fs_type);
3959 kobject_put(cgroup_kobj);
3963 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
3967 static int __init cgroup_wq_init(void)
3970 * There isn't much point in executing destruction path in
3971 * parallel. Good chunk is serialized with cgroup_mutex anyway.
3973 * XXX: Must be ordered to make sure parent is offlined after
3974 * children. The ordering requirement is for memcg where a
3975 * parent's offline may wait for a child's leading to deadlock. In
3976 * the long term, this should be fixed from memcg side.
3978 * We would prefer to do this in cgroup_init() above, but that
3979 * is called before init_workqueues(): so leave this until after.
3981 cgroup_destroy_wq = alloc_ordered_workqueue("cgroup_destroy", 0);
3982 BUG_ON(!cgroup_destroy_wq);
3985 * Used to destroy pidlists and separate to serve as flush domain.
3986 * Cap @max_active to 1 too.
3988 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
3990 BUG_ON(!cgroup_pidlist_destroy_wq);
3994 core_initcall(cgroup_wq_init);
3997 * proc_cgroup_show()
3998 * - Print task's cgroup paths into seq_file, one line for each hierarchy
3999 * - Used for /proc/<pid>/cgroup.
4000 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4001 * doesn't really matter if tsk->cgroup changes after we read it,
4002 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4003 * anyway. No need to check that tsk->cgroup != NULL, thanks to
4004 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4005 * cgroup to top_cgroup.
4008 /* TODO: Use a proper seq_file iterator */
4009 int proc_cgroup_show(struct seq_file *m, void *v)
4012 struct task_struct *tsk;
4015 struct cgroupfs_root *root;
4018 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4024 tsk = get_pid_task(pid, PIDTYPE_PID);
4030 mutex_lock(&cgroup_mutex);
4031 down_read(&css_set_rwsem);
4033 for_each_active_root(root) {
4034 struct cgroup_subsys *ss;
4035 struct cgroup *cgrp;
4036 int ssid, count = 0;
4038 seq_printf(m, "%d:", root->hierarchy_id);
4039 for_each_subsys(ss, ssid)
4040 if (root->subsys_mask & (1 << ssid))
4041 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4042 if (strlen(root->name))
4043 seq_printf(m, "%sname=%s", count ? "," : "",
4046 cgrp = task_cgroup_from_root(tsk, root);
4047 path = cgroup_path(cgrp, buf, PATH_MAX);
4049 retval = -ENAMETOOLONG;
4057 up_read(&css_set_rwsem);
4058 mutex_unlock(&cgroup_mutex);
4059 put_task_struct(tsk);
4066 /* Display information about each subsystem and each hierarchy */
4067 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4069 struct cgroup_subsys *ss;
4072 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4074 * ideally we don't want subsystems moving around while we do this.
4075 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4076 * subsys/hierarchy state.
4078 mutex_lock(&cgroup_mutex);
4080 for_each_subsys(ss, i)
4081 seq_printf(m, "%s\t%d\t%d\t%d\n",
4082 ss->name, ss->root->hierarchy_id,
4083 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4085 mutex_unlock(&cgroup_mutex);
4089 static int cgroupstats_open(struct inode *inode, struct file *file)
4091 return single_open(file, proc_cgroupstats_show, NULL);
4094 static const struct file_operations proc_cgroupstats_operations = {
4095 .open = cgroupstats_open,
4097 .llseek = seq_lseek,
4098 .release = single_release,
4102 * cgroup_fork - attach newly forked task to its parents cgroup.
4103 * @child: pointer to task_struct of forking parent process.
4105 * Description: A task inherits its parent's cgroup at fork().
4107 * A pointer to the shared css_set was automatically copied in
4108 * fork.c by dup_task_struct(). However, we ignore that copy, since
4109 * it was not made under the protection of RCU or cgroup_mutex, so
4110 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
4111 * have already changed current->cgroups, allowing the previously
4112 * referenced cgroup group to be removed and freed.
4114 * At the point that cgroup_fork() is called, 'current' is the parent
4115 * task, and the passed argument 'child' points to the child task.
4117 void cgroup_fork(struct task_struct *child)
4120 get_css_set(task_css_set(current));
4121 child->cgroups = current->cgroups;
4122 task_unlock(current);
4123 INIT_LIST_HEAD(&child->cg_list);
4127 * cgroup_post_fork - called on a new task after adding it to the task list
4128 * @child: the task in question
4130 * Adds the task to the list running through its css_set if necessary and
4131 * call the subsystem fork() callbacks. Has to be after the task is
4132 * visible on the task list in case we race with the first call to
4133 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4136 void cgroup_post_fork(struct task_struct *child)
4138 struct cgroup_subsys *ss;
4142 * use_task_css_set_links is set to 1 before we walk the tasklist
4143 * under the tasklist_lock and we read it here after we added the child
4144 * to the tasklist under the tasklist_lock as well. If the child wasn't
4145 * yet in the tasklist when we walked through it from
4146 * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
4147 * should be visible now due to the paired locking and barriers implied
4148 * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
4149 * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
4152 if (use_task_css_set_links) {
4153 down_write(&css_set_rwsem);
4155 if (list_empty(&child->cg_list))
4156 list_add(&child->cg_list, &task_css_set(child)->tasks);
4158 up_write(&css_set_rwsem);
4162 * Call ss->fork(). This must happen after @child is linked on
4163 * css_set; otherwise, @child might change state between ->fork()
4164 * and addition to css_set.
4166 if (need_forkexit_callback) {
4167 for_each_subsys(ss, i)
4174 * cgroup_exit - detach cgroup from exiting task
4175 * @tsk: pointer to task_struct of exiting process
4176 * @run_callback: run exit callbacks?
4178 * Description: Detach cgroup from @tsk and release it.
4180 * Note that cgroups marked notify_on_release force every task in
4181 * them to take the global cgroup_mutex mutex when exiting.
4182 * This could impact scaling on very large systems. Be reluctant to
4183 * use notify_on_release cgroups where very high task exit scaling
4184 * is required on large systems.
4186 * the_top_cgroup_hack:
4188 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4190 * We call cgroup_exit() while the task is still competent to
4191 * handle notify_on_release(), then leave the task attached to the
4192 * root cgroup in each hierarchy for the remainder of its exit.
4194 * To do this properly, we would increment the reference count on
4195 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
4196 * code we would add a second cgroup function call, to drop that
4197 * reference. This would just create an unnecessary hot spot on
4198 * the top_cgroup reference count, to no avail.
4200 * Normally, holding a reference to a cgroup without bumping its
4201 * count is unsafe. The cgroup could go away, or someone could
4202 * attach us to a different cgroup, decrementing the count on
4203 * the first cgroup that we never incremented. But in this case,
4204 * top_cgroup isn't going away, and either task has PF_EXITING set,
4205 * which wards off any cgroup_attach_task() attempts, or task is a failed
4206 * fork, never visible to cgroup_attach_task.
4208 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4210 struct cgroup_subsys *ss;
4211 struct css_set *cset;
4215 * Unlink from the css_set task list if necessary. Optimistically
4216 * check cg_list before taking css_set_rwsem.
4218 if (!list_empty(&tsk->cg_list)) {
4219 down_write(&css_set_rwsem);
4220 if (!list_empty(&tsk->cg_list))
4221 list_del_init(&tsk->cg_list);
4222 up_write(&css_set_rwsem);
4225 /* Reassign the task to the init_css_set. */
4227 cset = task_css_set(tsk);
4228 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4230 if (run_callbacks && need_forkexit_callback) {
4231 /* see cgroup_post_fork() for details */
4232 for_each_subsys(ss, i) {
4234 struct cgroup_subsys_state *old_css = cset->subsys[i];
4235 struct cgroup_subsys_state *css = task_css(tsk, i);
4237 ss->exit(css, old_css, tsk);
4243 put_css_set(cset, true);
4246 static void check_for_release(struct cgroup *cgrp)
4248 if (cgroup_is_releasable(cgrp) &&
4249 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
4251 * Control Group is currently removeable. If it's not
4252 * already queued for a userspace notification, queue
4255 int need_schedule_work = 0;
4257 raw_spin_lock(&release_list_lock);
4258 if (!cgroup_is_dead(cgrp) &&
4259 list_empty(&cgrp->release_list)) {
4260 list_add(&cgrp->release_list, &release_list);
4261 need_schedule_work = 1;
4263 raw_spin_unlock(&release_list_lock);
4264 if (need_schedule_work)
4265 schedule_work(&release_agent_work);
4270 * Notify userspace when a cgroup is released, by running the
4271 * configured release agent with the name of the cgroup (path
4272 * relative to the root of cgroup file system) as the argument.
4274 * Most likely, this user command will try to rmdir this cgroup.
4276 * This races with the possibility that some other task will be
4277 * attached to this cgroup before it is removed, or that some other
4278 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
4279 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4280 * unused, and this cgroup will be reprieved from its death sentence,
4281 * to continue to serve a useful existence. Next time it's released,
4282 * we will get notified again, if it still has 'notify_on_release' set.
4284 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4285 * means only wait until the task is successfully execve()'d. The
4286 * separate release agent task is forked by call_usermodehelper(),
4287 * then control in this thread returns here, without waiting for the
4288 * release agent task. We don't bother to wait because the caller of
4289 * this routine has no use for the exit status of the release agent
4290 * task, so no sense holding our caller up for that.
4292 static void cgroup_release_agent(struct work_struct *work)
4294 BUG_ON(work != &release_agent_work);
4295 mutex_lock(&cgroup_mutex);
4296 raw_spin_lock(&release_list_lock);
4297 while (!list_empty(&release_list)) {
4298 char *argv[3], *envp[3];
4300 char *pathbuf = NULL, *agentbuf = NULL, *path;
4301 struct cgroup *cgrp = list_entry(release_list.next,
4304 list_del_init(&cgrp->release_list);
4305 raw_spin_unlock(&release_list_lock);
4306 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
4309 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
4312 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4317 argv[i++] = agentbuf;
4322 /* minimal command environment */
4323 envp[i++] = "HOME=/";
4324 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4327 /* Drop the lock while we invoke the usermode helper,
4328 * since the exec could involve hitting disk and hence
4329 * be a slow process */
4330 mutex_unlock(&cgroup_mutex);
4331 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4332 mutex_lock(&cgroup_mutex);
4336 raw_spin_lock(&release_list_lock);
4338 raw_spin_unlock(&release_list_lock);
4339 mutex_unlock(&cgroup_mutex);
4342 static int __init cgroup_disable(char *str)
4344 struct cgroup_subsys *ss;
4348 while ((token = strsep(&str, ",")) != NULL) {
4352 for_each_subsys(ss, i) {
4353 if (!strcmp(token, ss->name)) {
4355 printk(KERN_INFO "Disabling %s control group"
4356 " subsystem\n", ss->name);
4363 __setup("cgroup_disable=", cgroup_disable);
4366 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
4367 * @dentry: directory dentry of interest
4368 * @ss: subsystem of interest
4370 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
4371 * to get the corresponding css and return it. If such css doesn't exist
4372 * or can't be pinned, an ERR_PTR value is returned.
4374 struct cgroup_subsys_state *css_tryget_from_dir(struct dentry *dentry,
4375 struct cgroup_subsys *ss)
4377 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4378 struct cgroup_subsys_state *css = NULL;
4379 struct cgroup *cgrp;
4381 /* is @dentry a cgroup dir? */
4382 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4383 kernfs_type(kn) != KERNFS_DIR)
4384 return ERR_PTR(-EBADF);
4389 * This path doesn't originate from kernfs and @kn could already
4390 * have been or be removed at any point. @kn->priv is RCU
4391 * protected for this access. See destroy_locked() for details.
4393 cgrp = rcu_dereference(kn->priv);
4395 css = cgroup_css(cgrp, ss);
4397 if (!css || !css_tryget(css))
4398 css = ERR_PTR(-ENOENT);
4405 * css_from_id - lookup css by id
4406 * @id: the cgroup id
4407 * @ss: cgroup subsys to be looked into
4409 * Returns the css if there's valid one with @id, otherwise returns NULL.
4410 * Should be called under rcu_read_lock().
4412 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
4414 struct cgroup *cgrp;
4416 cgroup_assert_mutexes_or_rcu_locked();
4418 cgrp = idr_find(&ss->root->cgroup_idr, id);
4420 return cgroup_css(cgrp, ss);
4424 #ifdef CONFIG_CGROUP_DEBUG
4425 static struct cgroup_subsys_state *
4426 debug_css_alloc(struct cgroup_subsys_state *parent_css)
4428 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
4431 return ERR_PTR(-ENOMEM);
4436 static void debug_css_free(struct cgroup_subsys_state *css)
4441 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
4444 return cgroup_task_count(css->cgroup);
4447 static u64 current_css_set_read(struct cgroup_subsys_state *css,
4450 return (u64)(unsigned long)current->cgroups;
4453 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
4459 count = atomic_read(&task_css_set(current)->refcount);
4464 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
4466 struct cgrp_cset_link *link;
4467 struct css_set *cset;
4470 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
4474 down_read(&css_set_rwsem);
4476 cset = rcu_dereference(current->cgroups);
4477 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
4478 struct cgroup *c = link->cgrp;
4479 const char *name = "?";
4481 if (c != cgroup_dummy_top) {
4482 cgroup_name(c, name_buf, NAME_MAX + 1);
4486 seq_printf(seq, "Root %d group %s\n",
4487 c->root->hierarchy_id, name);
4490 up_read(&css_set_rwsem);
4495 #define MAX_TASKS_SHOWN_PER_CSS 25
4496 static int cgroup_css_links_read(struct seq_file *seq, void *v)
4498 struct cgroup_subsys_state *css = seq_css(seq);
4499 struct cgrp_cset_link *link;
4501 down_read(&css_set_rwsem);
4502 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
4503 struct css_set *cset = link->cset;
4504 struct task_struct *task;
4506 seq_printf(seq, "css_set %p\n", cset);
4507 list_for_each_entry(task, &cset->tasks, cg_list) {
4508 if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
4509 seq_puts(seq, " ...\n");
4512 seq_printf(seq, " task %d\n",
4513 task_pid_vnr(task));
4517 up_read(&css_set_rwsem);
4521 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
4523 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4526 static struct cftype debug_files[] = {
4528 .name = "taskcount",
4529 .read_u64 = debug_taskcount_read,
4533 .name = "current_css_set",
4534 .read_u64 = current_css_set_read,
4538 .name = "current_css_set_refcount",
4539 .read_u64 = current_css_set_refcount_read,
4543 .name = "current_css_set_cg_links",
4544 .seq_show = current_css_set_cg_links_read,
4548 .name = "cgroup_css_links",
4549 .seq_show = cgroup_css_links_read,
4553 .name = "releasable",
4554 .read_u64 = releasable_read,
4560 struct cgroup_subsys debug_cgrp_subsys = {
4561 .css_alloc = debug_css_alloc,
4562 .css_free = debug_css_free,
4563 .base_cftypes = debug_files,
4565 #endif /* CONFIG_CGROUP_DEBUG */