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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 DECLARE_RWSEM(css_set_rwsem);
87 EXPORT_SYMBOL_GPL(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(css_set_rwsem);
90 static DEFINE_MUTEX(cgroup_mutex);
91 static DECLARE_RWSEM(css_set_rwsem);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
123 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys *cgroup_subsys[] = {
128 #include <linux/cgroup_subsys.h>
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name[] = {
135 #include <linux/cgroup_subsys.h>
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
144 struct cgroup_root cgrp_dfl_root;
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
150 static bool cgrp_dfl_root_visible;
152 /* some controllers are not supported in the default hierarchy */
153 static const unsigned int cgrp_dfl_root_inhibit_ss_mask = 0
154 #ifdef CONFIG_CGROUP_DEBUG
155 | (1 << debug_cgrp_id)
159 /* The list of hierarchy roots */
161 static LIST_HEAD(cgroup_roots);
162 static int cgroup_root_count;
164 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
165 static DEFINE_IDR(cgroup_hierarchy_idr);
168 * Assign a monotonically increasing serial number to csses. It guarantees
169 * cgroups with bigger numbers are newer than those with smaller numbers.
170 * Also, as csses are always appended to the parent's ->children list, it
171 * guarantees that sibling csses are always sorted in the ascending serial
172 * number order on the list. Protected by cgroup_mutex.
174 static u64 css_serial_nr_next = 1;
176 /* This flag indicates whether tasks in the fork and exit paths should
177 * check for fork/exit handlers to call. This avoids us having to do
178 * extra work in the fork/exit path if none of the subsystems need to
181 static int need_forkexit_callback __read_mostly;
183 static struct cftype cgroup_base_files[];
185 static void cgroup_put(struct cgroup *cgrp);
186 static int rebind_subsystems(struct cgroup_root *dst_root,
187 unsigned int ss_mask);
188 static int cgroup_destroy_locked(struct cgroup *cgrp);
189 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss);
190 static void css_release(struct percpu_ref *ref);
191 static void kill_css(struct cgroup_subsys_state *css);
192 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
194 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
196 /* IDR wrappers which synchronize using cgroup_idr_lock */
197 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
202 idr_preload(gfp_mask);
203 spin_lock_bh(&cgroup_idr_lock);
204 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
205 spin_unlock_bh(&cgroup_idr_lock);
210 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
214 spin_lock_bh(&cgroup_idr_lock);
215 ret = idr_replace(idr, ptr, id);
216 spin_unlock_bh(&cgroup_idr_lock);
220 static void cgroup_idr_remove(struct idr *idr, int id)
222 spin_lock_bh(&cgroup_idr_lock);
224 spin_unlock_bh(&cgroup_idr_lock);
227 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
229 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
232 return container_of(parent_css, struct cgroup, self);
237 * cgroup_css - obtain a cgroup's css for the specified subsystem
238 * @cgrp: the cgroup of interest
239 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
241 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
242 * function must be called either under cgroup_mutex or rcu_read_lock() and
243 * the caller is responsible for pinning the returned css if it wants to
244 * keep accessing it outside the said locks. This function may return
245 * %NULL if @cgrp doesn't have @subsys_id enabled.
247 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
248 struct cgroup_subsys *ss)
251 return rcu_dereference_check(cgrp->subsys[ss->id],
252 lockdep_is_held(&cgroup_mutex));
258 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
259 * @cgrp: the cgroup of interest
260 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
262 * Similar to cgroup_css() but returns the effctive css, which is defined
263 * as the matching css of the nearest ancestor including self which has @ss
264 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
265 * function is guaranteed to return non-NULL css.
267 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
268 struct cgroup_subsys *ss)
270 lockdep_assert_held(&cgroup_mutex);
275 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
278 while (cgroup_parent(cgrp) &&
279 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
280 cgrp = cgroup_parent(cgrp);
282 return cgroup_css(cgrp, ss);
285 /* convenient tests for these bits */
286 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
288 return !(cgrp->self.flags & CSS_ONLINE);
291 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
293 struct cgroup *cgrp = of->kn->parent->priv;
294 struct cftype *cft = of_cft(of);
297 * This is open and unprotected implementation of cgroup_css().
298 * seq_css() is only called from a kernfs file operation which has
299 * an active reference on the file. Because all the subsystem
300 * files are drained before a css is disassociated with a cgroup,
301 * the matching css from the cgroup's subsys table is guaranteed to
302 * be and stay valid until the enclosing operation is complete.
305 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
309 EXPORT_SYMBOL_GPL(of_css);
312 * cgroup_is_descendant - test ancestry
313 * @cgrp: the cgroup to be tested
314 * @ancestor: possible ancestor of @cgrp
316 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
317 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
318 * and @ancestor are accessible.
320 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
323 if (cgrp == ancestor)
325 cgrp = cgroup_parent(cgrp);
330 static int cgroup_is_releasable(const struct cgroup *cgrp)
333 (1 << CGRP_RELEASABLE) |
334 (1 << CGRP_NOTIFY_ON_RELEASE);
335 return (cgrp->flags & bits) == bits;
338 static int notify_on_release(const struct cgroup *cgrp)
340 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
344 * for_each_css - iterate all css's of a cgroup
345 * @css: the iteration cursor
346 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
347 * @cgrp: the target cgroup to iterate css's of
349 * Should be called under cgroup_[tree_]mutex.
351 #define for_each_css(css, ssid, cgrp) \
352 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
353 if (!((css) = rcu_dereference_check( \
354 (cgrp)->subsys[(ssid)], \
355 lockdep_is_held(&cgroup_mutex)))) { } \
359 * for_each_e_css - iterate all effective css's of a cgroup
360 * @css: the iteration cursor
361 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
362 * @cgrp: the target cgroup to iterate css's of
364 * Should be called under cgroup_[tree_]mutex.
366 #define for_each_e_css(css, ssid, cgrp) \
367 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
368 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
373 * for_each_subsys - iterate all enabled cgroup subsystems
374 * @ss: the iteration cursor
375 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
377 #define for_each_subsys(ss, ssid) \
378 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
379 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
381 /* iterate across the hierarchies */
382 #define for_each_root(root) \
383 list_for_each_entry((root), &cgroup_roots, root_list)
385 /* iterate over child cgrps, lock should be held throughout iteration */
386 #define cgroup_for_each_live_child(child, cgrp) \
387 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
388 if (({ lockdep_assert_held(&cgroup_mutex); \
389 cgroup_is_dead(child); })) \
393 /* the list of cgroups eligible for automatic release. Protected by
394 * release_list_lock */
395 static LIST_HEAD(release_list);
396 static DEFINE_RAW_SPINLOCK(release_list_lock);
397 static void cgroup_release_agent(struct work_struct *work);
398 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
399 static void check_for_release(struct cgroup *cgrp);
402 * A cgroup can be associated with multiple css_sets as different tasks may
403 * belong to different cgroups on different hierarchies. In the other
404 * direction, a css_set is naturally associated with multiple cgroups.
405 * This M:N relationship is represented by the following link structure
406 * which exists for each association and allows traversing the associations
409 struct cgrp_cset_link {
410 /* the cgroup and css_set this link associates */
412 struct css_set *cset;
414 /* list of cgrp_cset_links anchored at cgrp->cset_links */
415 struct list_head cset_link;
417 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
418 struct list_head cgrp_link;
422 * The default css_set - used by init and its children prior to any
423 * hierarchies being mounted. It contains a pointer to the root state
424 * for each subsystem. Also used to anchor the list of css_sets. Not
425 * reference-counted, to improve performance when child cgroups
426 * haven't been created.
428 struct css_set init_css_set = {
429 .refcount = ATOMIC_INIT(1),
430 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
431 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
432 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
433 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
434 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
437 static int css_set_count = 1; /* 1 for init_css_set */
440 * cgroup_update_populated - updated populated count of a cgroup
441 * @cgrp: the target cgroup
442 * @populated: inc or dec populated count
444 * @cgrp is either getting the first task (css_set) or losing the last.
445 * Update @cgrp->populated_cnt accordingly. The count is propagated
446 * towards root so that a given cgroup's populated_cnt is zero iff the
447 * cgroup and all its descendants are empty.
449 * @cgrp's interface file "cgroup.populated" is zero if
450 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
451 * changes from or to zero, userland is notified that the content of the
452 * interface file has changed. This can be used to detect when @cgrp and
453 * its descendants become populated or empty.
455 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
457 lockdep_assert_held(&css_set_rwsem);
463 trigger = !cgrp->populated_cnt++;
465 trigger = !--cgrp->populated_cnt;
470 if (cgrp->populated_kn)
471 kernfs_notify(cgrp->populated_kn);
472 cgrp = cgroup_parent(cgrp);
477 * hash table for cgroup groups. This improves the performance to find
478 * an existing css_set. This hash doesn't (currently) take into
479 * account cgroups in empty hierarchies.
481 #define CSS_SET_HASH_BITS 7
482 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
484 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
486 unsigned long key = 0UL;
487 struct cgroup_subsys *ss;
490 for_each_subsys(ss, i)
491 key += (unsigned long)css[i];
492 key = (key >> 16) ^ key;
497 static void put_css_set_locked(struct css_set *cset, bool taskexit)
499 struct cgrp_cset_link *link, *tmp_link;
500 struct cgroup_subsys *ss;
503 lockdep_assert_held(&css_set_rwsem);
505 if (!atomic_dec_and_test(&cset->refcount))
508 /* This css_set is dead. unlink it and release cgroup refcounts */
509 for_each_subsys(ss, ssid)
510 list_del(&cset->e_cset_node[ssid]);
511 hash_del(&cset->hlist);
514 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
515 struct cgroup *cgrp = link->cgrp;
517 list_del(&link->cset_link);
518 list_del(&link->cgrp_link);
520 /* @cgrp can't go away while we're holding css_set_rwsem */
521 if (list_empty(&cgrp->cset_links)) {
522 cgroup_update_populated(cgrp, false);
523 if (notify_on_release(cgrp)) {
525 set_bit(CGRP_RELEASABLE, &cgrp->flags);
526 check_for_release(cgrp);
533 kfree_rcu(cset, rcu_head);
536 static void put_css_set(struct css_set *cset, bool taskexit)
539 * Ensure that the refcount doesn't hit zero while any readers
540 * can see it. Similar to atomic_dec_and_lock(), but for an
543 if (atomic_add_unless(&cset->refcount, -1, 1))
546 down_write(&css_set_rwsem);
547 put_css_set_locked(cset, taskexit);
548 up_write(&css_set_rwsem);
552 * refcounted get/put for css_set objects
554 static inline void get_css_set(struct css_set *cset)
556 atomic_inc(&cset->refcount);
560 * compare_css_sets - helper function for find_existing_css_set().
561 * @cset: candidate css_set being tested
562 * @old_cset: existing css_set for a task
563 * @new_cgrp: cgroup that's being entered by the task
564 * @template: desired set of css pointers in css_set (pre-calculated)
566 * Returns true if "cset" matches "old_cset" except for the hierarchy
567 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
569 static bool compare_css_sets(struct css_set *cset,
570 struct css_set *old_cset,
571 struct cgroup *new_cgrp,
572 struct cgroup_subsys_state *template[])
574 struct list_head *l1, *l2;
577 * On the default hierarchy, there can be csets which are
578 * associated with the same set of cgroups but different csses.
579 * Let's first ensure that csses match.
581 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
585 * Compare cgroup pointers in order to distinguish between
586 * different cgroups in hierarchies. As different cgroups may
587 * share the same effective css, this comparison is always
590 l1 = &cset->cgrp_links;
591 l2 = &old_cset->cgrp_links;
593 struct cgrp_cset_link *link1, *link2;
594 struct cgroup *cgrp1, *cgrp2;
598 /* See if we reached the end - both lists are equal length. */
599 if (l1 == &cset->cgrp_links) {
600 BUG_ON(l2 != &old_cset->cgrp_links);
603 BUG_ON(l2 == &old_cset->cgrp_links);
605 /* Locate the cgroups associated with these links. */
606 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
607 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
610 /* Hierarchies should be linked in the same order. */
611 BUG_ON(cgrp1->root != cgrp2->root);
614 * If this hierarchy is the hierarchy of the cgroup
615 * that's changing, then we need to check that this
616 * css_set points to the new cgroup; if it's any other
617 * hierarchy, then this css_set should point to the
618 * same cgroup as the old css_set.
620 if (cgrp1->root == new_cgrp->root) {
621 if (cgrp1 != new_cgrp)
632 * find_existing_css_set - init css array and find the matching css_set
633 * @old_cset: the css_set that we're using before the cgroup transition
634 * @cgrp: the cgroup that we're moving into
635 * @template: out param for the new set of csses, should be clear on entry
637 static struct css_set *find_existing_css_set(struct css_set *old_cset,
639 struct cgroup_subsys_state *template[])
641 struct cgroup_root *root = cgrp->root;
642 struct cgroup_subsys *ss;
643 struct css_set *cset;
648 * Build the set of subsystem state objects that we want to see in the
649 * new css_set. while subsystems can change globally, the entries here
650 * won't change, so no need for locking.
652 for_each_subsys(ss, i) {
653 if (root->subsys_mask & (1UL << i)) {
655 * @ss is in this hierarchy, so we want the
656 * effective css from @cgrp.
658 template[i] = cgroup_e_css(cgrp, ss);
661 * @ss is not in this hierarchy, so we don't want
664 template[i] = old_cset->subsys[i];
668 key = css_set_hash(template);
669 hash_for_each_possible(css_set_table, cset, hlist, key) {
670 if (!compare_css_sets(cset, old_cset, cgrp, template))
673 /* This css_set matches what we need */
677 /* No existing cgroup group matched */
681 static void free_cgrp_cset_links(struct list_head *links_to_free)
683 struct cgrp_cset_link *link, *tmp_link;
685 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
686 list_del(&link->cset_link);
692 * allocate_cgrp_cset_links - allocate cgrp_cset_links
693 * @count: the number of links to allocate
694 * @tmp_links: list_head the allocated links are put on
696 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
697 * through ->cset_link. Returns 0 on success or -errno.
699 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
701 struct cgrp_cset_link *link;
704 INIT_LIST_HEAD(tmp_links);
706 for (i = 0; i < count; i++) {
707 link = kzalloc(sizeof(*link), GFP_KERNEL);
709 free_cgrp_cset_links(tmp_links);
712 list_add(&link->cset_link, tmp_links);
718 * link_css_set - a helper function to link a css_set to a cgroup
719 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
720 * @cset: the css_set to be linked
721 * @cgrp: the destination cgroup
723 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
726 struct cgrp_cset_link *link;
728 BUG_ON(list_empty(tmp_links));
730 if (cgroup_on_dfl(cgrp))
731 cset->dfl_cgrp = cgrp;
733 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
737 if (list_empty(&cgrp->cset_links))
738 cgroup_update_populated(cgrp, true);
739 list_move(&link->cset_link, &cgrp->cset_links);
742 * Always add links to the tail of the list so that the list
743 * is sorted by order of hierarchy creation
745 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
749 * find_css_set - return a new css_set with one cgroup updated
750 * @old_cset: the baseline css_set
751 * @cgrp: the cgroup to be updated
753 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
754 * substituted into the appropriate hierarchy.
756 static struct css_set *find_css_set(struct css_set *old_cset,
759 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
760 struct css_set *cset;
761 struct list_head tmp_links;
762 struct cgrp_cset_link *link;
763 struct cgroup_subsys *ss;
767 lockdep_assert_held(&cgroup_mutex);
769 /* First see if we already have a cgroup group that matches
771 down_read(&css_set_rwsem);
772 cset = find_existing_css_set(old_cset, cgrp, template);
775 up_read(&css_set_rwsem);
780 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
784 /* Allocate all the cgrp_cset_link objects that we'll need */
785 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
790 atomic_set(&cset->refcount, 1);
791 INIT_LIST_HEAD(&cset->cgrp_links);
792 INIT_LIST_HEAD(&cset->tasks);
793 INIT_LIST_HEAD(&cset->mg_tasks);
794 INIT_LIST_HEAD(&cset->mg_preload_node);
795 INIT_LIST_HEAD(&cset->mg_node);
796 INIT_HLIST_NODE(&cset->hlist);
798 /* Copy the set of subsystem state objects generated in
799 * find_existing_css_set() */
800 memcpy(cset->subsys, template, sizeof(cset->subsys));
802 down_write(&css_set_rwsem);
803 /* Add reference counts and links from the new css_set. */
804 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
805 struct cgroup *c = link->cgrp;
807 if (c->root == cgrp->root)
809 link_css_set(&tmp_links, cset, c);
812 BUG_ON(!list_empty(&tmp_links));
816 /* Add @cset to the hash table */
817 key = css_set_hash(cset->subsys);
818 hash_add(css_set_table, &cset->hlist, key);
820 for_each_subsys(ss, ssid)
821 list_add_tail(&cset->e_cset_node[ssid],
822 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
824 up_write(&css_set_rwsem);
829 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
831 struct cgroup *root_cgrp = kf_root->kn->priv;
833 return root_cgrp->root;
836 static int cgroup_init_root_id(struct cgroup_root *root)
840 lockdep_assert_held(&cgroup_mutex);
842 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
846 root->hierarchy_id = id;
850 static void cgroup_exit_root_id(struct cgroup_root *root)
852 lockdep_assert_held(&cgroup_mutex);
854 if (root->hierarchy_id) {
855 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
856 root->hierarchy_id = 0;
860 static void cgroup_free_root(struct cgroup_root *root)
863 /* hierarhcy ID shoulid already have been released */
864 WARN_ON_ONCE(root->hierarchy_id);
866 idr_destroy(&root->cgroup_idr);
871 static void cgroup_destroy_root(struct cgroup_root *root)
873 struct cgroup *cgrp = &root->cgrp;
874 struct cgrp_cset_link *link, *tmp_link;
876 mutex_lock(&cgroup_mutex);
878 BUG_ON(atomic_read(&root->nr_cgrps));
879 BUG_ON(!list_empty(&cgrp->self.children));
881 /* Rebind all subsystems back to the default hierarchy */
882 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
885 * Release all the links from cset_links to this hierarchy's
888 down_write(&css_set_rwsem);
890 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
891 list_del(&link->cset_link);
892 list_del(&link->cgrp_link);
895 up_write(&css_set_rwsem);
897 if (!list_empty(&root->root_list)) {
898 list_del(&root->root_list);
902 cgroup_exit_root_id(root);
904 mutex_unlock(&cgroup_mutex);
906 kernfs_destroy_root(root->kf_root);
907 cgroup_free_root(root);
910 /* look up cgroup associated with given css_set on the specified hierarchy */
911 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
912 struct cgroup_root *root)
914 struct cgroup *res = NULL;
916 lockdep_assert_held(&cgroup_mutex);
917 lockdep_assert_held(&css_set_rwsem);
919 if (cset == &init_css_set) {
922 struct cgrp_cset_link *link;
924 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
925 struct cgroup *c = link->cgrp;
927 if (c->root == root) {
939 * Return the cgroup for "task" from the given hierarchy. Must be
940 * called with cgroup_mutex and css_set_rwsem held.
942 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
943 struct cgroup_root *root)
946 * No need to lock the task - since we hold cgroup_mutex the
947 * task can't change groups, so the only thing that can happen
948 * is that it exits and its css is set back to init_css_set.
950 return cset_cgroup_from_root(task_css_set(task), root);
954 * A task must hold cgroup_mutex to modify cgroups.
956 * Any task can increment and decrement the count field without lock.
957 * So in general, code holding cgroup_mutex can't rely on the count
958 * field not changing. However, if the count goes to zero, then only
959 * cgroup_attach_task() can increment it again. Because a count of zero
960 * means that no tasks are currently attached, therefore there is no
961 * way a task attached to that cgroup can fork (the other way to
962 * increment the count). So code holding cgroup_mutex can safely
963 * assume that if the count is zero, it will stay zero. Similarly, if
964 * a task holds cgroup_mutex on a cgroup with zero count, it
965 * knows that the cgroup won't be removed, as cgroup_rmdir()
968 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
969 * (usually) take cgroup_mutex. These are the two most performance
970 * critical pieces of code here. The exception occurs on cgroup_exit(),
971 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
972 * is taken, and if the cgroup count is zero, a usermode call made
973 * to the release agent with the name of the cgroup (path relative to
974 * the root of cgroup file system) as the argument.
976 * A cgroup can only be deleted if both its 'count' of using tasks
977 * is zero, and its list of 'children' cgroups is empty. Since all
978 * tasks in the system use _some_ cgroup, and since there is always at
979 * least one task in the system (init, pid == 1), therefore, root cgroup
980 * always has either children cgroups and/or using tasks. So we don't
981 * need a special hack to ensure that root cgroup cannot be deleted.
983 * P.S. One more locking exception. RCU is used to guard the
984 * update of a tasks cgroup pointer by cgroup_attach_task()
987 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
988 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
989 static const struct file_operations proc_cgroupstats_operations;
991 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
994 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
995 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
996 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
997 cft->ss->name, cft->name);
999 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1004 * cgroup_file_mode - deduce file mode of a control file
1005 * @cft: the control file in question
1007 * returns cft->mode if ->mode is not 0
1008 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1009 * returns S_IRUGO if it has only a read handler
1010 * returns S_IWUSR if it has only a write hander
1012 static umode_t cgroup_file_mode(const struct cftype *cft)
1019 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1022 if (cft->write_u64 || cft->write_s64 || cft->write)
1028 static void cgroup_get(struct cgroup *cgrp)
1030 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1031 css_get(&cgrp->self);
1034 static void cgroup_put(struct cgroup *cgrp)
1036 css_put(&cgrp->self);
1039 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1041 cgrp->child_subsys_mask = cgrp->subtree_control;
1045 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1046 * @kn: the kernfs_node being serviced
1048 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1049 * the method finishes if locking succeeded. Note that once this function
1050 * returns the cgroup returned by cgroup_kn_lock_live() may become
1051 * inaccessible any time. If the caller intends to continue to access the
1052 * cgroup, it should pin it before invoking this function.
1054 static void cgroup_kn_unlock(struct kernfs_node *kn)
1056 struct cgroup *cgrp;
1058 if (kernfs_type(kn) == KERNFS_DIR)
1061 cgrp = kn->parent->priv;
1063 mutex_unlock(&cgroup_mutex);
1065 kernfs_unbreak_active_protection(kn);
1070 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1071 * @kn: the kernfs_node being serviced
1073 * This helper is to be used by a cgroup kernfs method currently servicing
1074 * @kn. It breaks the active protection, performs cgroup locking and
1075 * verifies that the associated cgroup is alive. Returns the cgroup if
1076 * alive; otherwise, %NULL. A successful return should be undone by a
1077 * matching cgroup_kn_unlock() invocation.
1079 * Any cgroup kernfs method implementation which requires locking the
1080 * associated cgroup should use this helper. It avoids nesting cgroup
1081 * locking under kernfs active protection and allows all kernfs operations
1082 * including self-removal.
1084 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1086 struct cgroup *cgrp;
1088 if (kernfs_type(kn) == KERNFS_DIR)
1091 cgrp = kn->parent->priv;
1094 * We're gonna grab cgroup_mutex which nests outside kernfs
1095 * active_ref. cgroup liveliness check alone provides enough
1096 * protection against removal. Ensure @cgrp stays accessible and
1097 * break the active_ref protection.
1100 kernfs_break_active_protection(kn);
1102 mutex_lock(&cgroup_mutex);
1104 if (!cgroup_is_dead(cgrp))
1107 cgroup_kn_unlock(kn);
1111 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1113 char name[CGROUP_FILE_NAME_MAX];
1115 lockdep_assert_held(&cgroup_mutex);
1116 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1120 * cgroup_clear_dir - remove subsys files in a cgroup directory
1121 * @cgrp: target cgroup
1122 * @subsys_mask: mask of the subsystem ids whose files should be removed
1124 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1126 struct cgroup_subsys *ss;
1129 for_each_subsys(ss, i) {
1130 struct cftype *cfts;
1132 if (!(subsys_mask & (1 << i)))
1134 list_for_each_entry(cfts, &ss->cfts, node)
1135 cgroup_addrm_files(cgrp, cfts, false);
1139 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1141 struct cgroup_subsys *ss;
1142 unsigned int tmp_ss_mask;
1145 lockdep_assert_held(&cgroup_mutex);
1147 for_each_subsys(ss, ssid) {
1148 if (!(ss_mask & (1 << ssid)))
1151 /* if @ss has non-root csses attached to it, can't move */
1152 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1155 /* can't move between two non-dummy roots either */
1156 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1160 /* skip creating root files on dfl_root for inhibited subsystems */
1161 tmp_ss_mask = ss_mask;
1162 if (dst_root == &cgrp_dfl_root)
1163 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1165 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1167 if (dst_root != &cgrp_dfl_root)
1171 * Rebinding back to the default root is not allowed to
1172 * fail. Using both default and non-default roots should
1173 * be rare. Moving subsystems back and forth even more so.
1174 * Just warn about it and continue.
1176 if (cgrp_dfl_root_visible) {
1177 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1179 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1184 * Nothing can fail from this point on. Remove files for the
1185 * removed subsystems and rebind each subsystem.
1187 for_each_subsys(ss, ssid)
1188 if (ss_mask & (1 << ssid))
1189 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1191 for_each_subsys(ss, ssid) {
1192 struct cgroup_root *src_root;
1193 struct cgroup_subsys_state *css;
1194 struct css_set *cset;
1196 if (!(ss_mask & (1 << ssid)))
1199 src_root = ss->root;
1200 css = cgroup_css(&src_root->cgrp, ss);
1202 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1204 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1205 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1206 ss->root = dst_root;
1207 css->cgroup = &dst_root->cgrp;
1209 down_write(&css_set_rwsem);
1210 hash_for_each(css_set_table, i, cset, hlist)
1211 list_move_tail(&cset->e_cset_node[ss->id],
1212 &dst_root->cgrp.e_csets[ss->id]);
1213 up_write(&css_set_rwsem);
1215 src_root->subsys_mask &= ~(1 << ssid);
1216 src_root->cgrp.subtree_control &= ~(1 << ssid);
1217 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1219 /* default hierarchy doesn't enable controllers by default */
1220 dst_root->subsys_mask |= 1 << ssid;
1221 if (dst_root != &cgrp_dfl_root) {
1222 dst_root->cgrp.subtree_control |= 1 << ssid;
1223 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1230 kernfs_activate(dst_root->cgrp.kn);
1234 static int cgroup_show_options(struct seq_file *seq,
1235 struct kernfs_root *kf_root)
1237 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1238 struct cgroup_subsys *ss;
1241 for_each_subsys(ss, ssid)
1242 if (root->subsys_mask & (1 << ssid))
1243 seq_printf(seq, ",%s", ss->name);
1244 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1245 seq_puts(seq, ",sane_behavior");
1246 if (root->flags & CGRP_ROOT_NOPREFIX)
1247 seq_puts(seq, ",noprefix");
1248 if (root->flags & CGRP_ROOT_XATTR)
1249 seq_puts(seq, ",xattr");
1251 spin_lock(&release_agent_path_lock);
1252 if (strlen(root->release_agent_path))
1253 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1254 spin_unlock(&release_agent_path_lock);
1256 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1257 seq_puts(seq, ",clone_children");
1258 if (strlen(root->name))
1259 seq_printf(seq, ",name=%s", root->name);
1263 struct cgroup_sb_opts {
1264 unsigned int subsys_mask;
1266 char *release_agent;
1267 bool cpuset_clone_children;
1269 /* User explicitly requested empty subsystem */
1273 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1275 char *token, *o = data;
1276 bool all_ss = false, one_ss = false;
1277 unsigned int mask = -1U;
1278 struct cgroup_subsys *ss;
1281 #ifdef CONFIG_CPUSETS
1282 mask = ~(1U << cpuset_cgrp_id);
1285 memset(opts, 0, sizeof(*opts));
1287 while ((token = strsep(&o, ",")) != NULL) {
1290 if (!strcmp(token, "none")) {
1291 /* Explicitly have no subsystems */
1295 if (!strcmp(token, "all")) {
1296 /* Mutually exclusive option 'all' + subsystem name */
1302 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1303 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1306 if (!strcmp(token, "noprefix")) {
1307 opts->flags |= CGRP_ROOT_NOPREFIX;
1310 if (!strcmp(token, "clone_children")) {
1311 opts->cpuset_clone_children = true;
1314 if (!strcmp(token, "xattr")) {
1315 opts->flags |= CGRP_ROOT_XATTR;
1318 if (!strncmp(token, "release_agent=", 14)) {
1319 /* Specifying two release agents is forbidden */
1320 if (opts->release_agent)
1322 opts->release_agent =
1323 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1324 if (!opts->release_agent)
1328 if (!strncmp(token, "name=", 5)) {
1329 const char *name = token + 5;
1330 /* Can't specify an empty name */
1333 /* Must match [\w.-]+ */
1334 for (i = 0; i < strlen(name); i++) {
1338 if ((c == '.') || (c == '-') || (c == '_'))
1342 /* Specifying two names is forbidden */
1345 opts->name = kstrndup(name,
1346 MAX_CGROUP_ROOT_NAMELEN - 1,
1354 for_each_subsys(ss, i) {
1355 if (strcmp(token, ss->name))
1360 /* Mutually exclusive option 'all' + subsystem name */
1363 opts->subsys_mask |= (1 << i);
1368 if (i == CGROUP_SUBSYS_COUNT)
1372 /* Consistency checks */
1374 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1375 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1377 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1378 opts->cpuset_clone_children || opts->release_agent ||
1380 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1385 * If the 'all' option was specified select all the
1386 * subsystems, otherwise if 'none', 'name=' and a subsystem
1387 * name options were not specified, let's default to 'all'
1389 if (all_ss || (!one_ss && !opts->none && !opts->name))
1390 for_each_subsys(ss, i)
1392 opts->subsys_mask |= (1 << i);
1395 * We either have to specify by name or by subsystems. (So
1396 * all empty hierarchies must have a name).
1398 if (!opts->subsys_mask && !opts->name)
1403 * Option noprefix was introduced just for backward compatibility
1404 * with the old cpuset, so we allow noprefix only if mounting just
1405 * the cpuset subsystem.
1407 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1411 /* Can't specify "none" and some subsystems */
1412 if (opts->subsys_mask && opts->none)
1418 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1421 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1422 struct cgroup_sb_opts opts;
1423 unsigned int added_mask, removed_mask;
1425 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1426 pr_err("sane_behavior: remount is not allowed\n");
1430 mutex_lock(&cgroup_mutex);
1432 /* See what subsystems are wanted */
1433 ret = parse_cgroupfs_options(data, &opts);
1437 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1438 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1439 task_tgid_nr(current), current->comm);
1441 added_mask = opts.subsys_mask & ~root->subsys_mask;
1442 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1444 /* Don't allow flags or name to change at remount */
1445 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1446 (opts.name && strcmp(opts.name, root->name))) {
1447 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1448 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1449 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1454 /* remounting is not allowed for populated hierarchies */
1455 if (!list_empty(&root->cgrp.self.children)) {
1460 ret = rebind_subsystems(root, added_mask);
1464 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1466 if (opts.release_agent) {
1467 spin_lock(&release_agent_path_lock);
1468 strcpy(root->release_agent_path, opts.release_agent);
1469 spin_unlock(&release_agent_path_lock);
1472 kfree(opts.release_agent);
1474 mutex_unlock(&cgroup_mutex);
1479 * To reduce the fork() overhead for systems that are not actually using
1480 * their cgroups capability, we don't maintain the lists running through
1481 * each css_set to its tasks until we see the list actually used - in other
1482 * words after the first mount.
1484 static bool use_task_css_set_links __read_mostly;
1486 static void cgroup_enable_task_cg_lists(void)
1488 struct task_struct *p, *g;
1490 down_write(&css_set_rwsem);
1492 if (use_task_css_set_links)
1495 use_task_css_set_links = true;
1498 * We need tasklist_lock because RCU is not safe against
1499 * while_each_thread(). Besides, a forking task that has passed
1500 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1501 * is not guaranteed to have its child immediately visible in the
1502 * tasklist if we walk through it with RCU.
1504 read_lock(&tasklist_lock);
1505 do_each_thread(g, p) {
1506 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1507 task_css_set(p) != &init_css_set);
1510 * We should check if the process is exiting, otherwise
1511 * it will race with cgroup_exit() in that the list
1512 * entry won't be deleted though the process has exited.
1513 * Do it while holding siglock so that we don't end up
1514 * racing against cgroup_exit().
1516 spin_lock_irq(&p->sighand->siglock);
1517 if (!(p->flags & PF_EXITING)) {
1518 struct css_set *cset = task_css_set(p);
1520 list_add(&p->cg_list, &cset->tasks);
1523 spin_unlock_irq(&p->sighand->siglock);
1524 } while_each_thread(g, p);
1525 read_unlock(&tasklist_lock);
1527 up_write(&css_set_rwsem);
1530 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1532 struct cgroup_subsys *ss;
1535 INIT_LIST_HEAD(&cgrp->self.sibling);
1536 INIT_LIST_HEAD(&cgrp->self.children);
1537 INIT_LIST_HEAD(&cgrp->cset_links);
1538 INIT_LIST_HEAD(&cgrp->release_list);
1539 INIT_LIST_HEAD(&cgrp->pidlists);
1540 mutex_init(&cgrp->pidlist_mutex);
1541 cgrp->self.cgroup = cgrp;
1542 cgrp->self.flags |= CSS_ONLINE;
1544 for_each_subsys(ss, ssid)
1545 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1547 init_waitqueue_head(&cgrp->offline_waitq);
1550 static void init_cgroup_root(struct cgroup_root *root,
1551 struct cgroup_sb_opts *opts)
1553 struct cgroup *cgrp = &root->cgrp;
1555 INIT_LIST_HEAD(&root->root_list);
1556 atomic_set(&root->nr_cgrps, 1);
1558 init_cgroup_housekeeping(cgrp);
1559 idr_init(&root->cgroup_idr);
1561 root->flags = opts->flags;
1562 if (opts->release_agent)
1563 strcpy(root->release_agent_path, opts->release_agent);
1565 strcpy(root->name, opts->name);
1566 if (opts->cpuset_clone_children)
1567 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1570 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1572 LIST_HEAD(tmp_links);
1573 struct cgroup *root_cgrp = &root->cgrp;
1574 struct css_set *cset;
1577 lockdep_assert_held(&cgroup_mutex);
1579 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1582 root_cgrp->id = ret;
1584 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
1589 * We're accessing css_set_count without locking css_set_rwsem here,
1590 * but that's OK - it can only be increased by someone holding
1591 * cgroup_lock, and that's us. The worst that can happen is that we
1592 * have some link structures left over
1594 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1598 ret = cgroup_init_root_id(root);
1602 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1603 KERNFS_ROOT_CREATE_DEACTIVATED,
1605 if (IS_ERR(root->kf_root)) {
1606 ret = PTR_ERR(root->kf_root);
1609 root_cgrp->kn = root->kf_root->kn;
1611 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1615 ret = rebind_subsystems(root, ss_mask);
1620 * There must be no failure case after here, since rebinding takes
1621 * care of subsystems' refcounts, which are explicitly dropped in
1622 * the failure exit path.
1624 list_add(&root->root_list, &cgroup_roots);
1625 cgroup_root_count++;
1628 * Link the root cgroup in this hierarchy into all the css_set
1631 down_write(&css_set_rwsem);
1632 hash_for_each(css_set_table, i, cset, hlist)
1633 link_css_set(&tmp_links, cset, root_cgrp);
1634 up_write(&css_set_rwsem);
1636 BUG_ON(!list_empty(&root_cgrp->self.children));
1637 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1639 kernfs_activate(root_cgrp->kn);
1644 kernfs_destroy_root(root->kf_root);
1645 root->kf_root = NULL;
1647 cgroup_exit_root_id(root);
1649 percpu_ref_cancel_init(&root_cgrp->self.refcnt);
1651 free_cgrp_cset_links(&tmp_links);
1655 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1656 int flags, const char *unused_dev_name,
1659 struct cgroup_root *root;
1660 struct cgroup_sb_opts opts;
1661 struct dentry *dentry;
1666 * The first time anyone tries to mount a cgroup, enable the list
1667 * linking each css_set to its tasks and fix up all existing tasks.
1669 if (!use_task_css_set_links)
1670 cgroup_enable_task_cg_lists();
1672 mutex_lock(&cgroup_mutex);
1674 /* First find the desired set of subsystems */
1675 ret = parse_cgroupfs_options(data, &opts);
1679 /* look for a matching existing root */
1680 if (!opts.subsys_mask && !opts.none && !opts.name) {
1681 cgrp_dfl_root_visible = true;
1682 root = &cgrp_dfl_root;
1683 cgroup_get(&root->cgrp);
1688 for_each_root(root) {
1689 bool name_match = false;
1691 if (root == &cgrp_dfl_root)
1695 * If we asked for a name then it must match. Also, if
1696 * name matches but sybsys_mask doesn't, we should fail.
1697 * Remember whether name matched.
1700 if (strcmp(opts.name, root->name))
1706 * If we asked for subsystems (or explicitly for no
1707 * subsystems) then they must match.
1709 if ((opts.subsys_mask || opts.none) &&
1710 (opts.subsys_mask != root->subsys_mask)) {
1717 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1718 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1719 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1723 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1728 * A root's lifetime is governed by its root cgroup.
1729 * tryget_live failure indicate that the root is being
1730 * destroyed. Wait for destruction to complete so that the
1731 * subsystems are free. We can use wait_queue for the wait
1732 * but this path is super cold. Let's just sleep for a bit
1735 if (!percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1736 mutex_unlock(&cgroup_mutex);
1738 ret = restart_syscall();
1747 * No such thing, create a new one. name= matching without subsys
1748 * specification is allowed for already existing hierarchies but we
1749 * can't create new one without subsys specification.
1751 if (!opts.subsys_mask && !opts.none) {
1756 root = kzalloc(sizeof(*root), GFP_KERNEL);
1762 init_cgroup_root(root, &opts);
1764 ret = cgroup_setup_root(root, opts.subsys_mask);
1766 cgroup_free_root(root);
1769 mutex_unlock(&cgroup_mutex);
1771 kfree(opts.release_agent);
1775 return ERR_PTR(ret);
1777 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1778 CGROUP_SUPER_MAGIC, &new_sb);
1779 if (IS_ERR(dentry) || !new_sb)
1780 cgroup_put(&root->cgrp);
1784 static void cgroup_kill_sb(struct super_block *sb)
1786 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1787 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1790 * If @root doesn't have any mounts or children, start killing it.
1791 * This prevents new mounts by disabling percpu_ref_tryget_live().
1792 * cgroup_mount() may wait for @root's release.
1794 * And don't kill the default root.
1796 if (css_has_online_children(&root->cgrp.self) ||
1797 root == &cgrp_dfl_root)
1798 cgroup_put(&root->cgrp);
1800 percpu_ref_kill(&root->cgrp.self.refcnt);
1805 static struct file_system_type cgroup_fs_type = {
1807 .mount = cgroup_mount,
1808 .kill_sb = cgroup_kill_sb,
1811 static struct kobject *cgroup_kobj;
1814 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1815 * @task: target task
1816 * @buf: the buffer to write the path into
1817 * @buflen: the length of the buffer
1819 * Determine @task's cgroup on the first (the one with the lowest non-zero
1820 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1821 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1822 * cgroup controller callbacks.
1824 * Return value is the same as kernfs_path().
1826 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1828 struct cgroup_root *root;
1829 struct cgroup *cgrp;
1830 int hierarchy_id = 1;
1833 mutex_lock(&cgroup_mutex);
1834 down_read(&css_set_rwsem);
1836 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1839 cgrp = task_cgroup_from_root(task, root);
1840 path = cgroup_path(cgrp, buf, buflen);
1842 /* if no hierarchy exists, everyone is in "/" */
1843 if (strlcpy(buf, "/", buflen) < buflen)
1847 up_read(&css_set_rwsem);
1848 mutex_unlock(&cgroup_mutex);
1851 EXPORT_SYMBOL_GPL(task_cgroup_path);
1853 /* used to track tasks and other necessary states during migration */
1854 struct cgroup_taskset {
1855 /* the src and dst cset list running through cset->mg_node */
1856 struct list_head src_csets;
1857 struct list_head dst_csets;
1860 * Fields for cgroup_taskset_*() iteration.
1862 * Before migration is committed, the target migration tasks are on
1863 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1864 * the csets on ->dst_csets. ->csets point to either ->src_csets
1865 * or ->dst_csets depending on whether migration is committed.
1867 * ->cur_csets and ->cur_task point to the current task position
1870 struct list_head *csets;
1871 struct css_set *cur_cset;
1872 struct task_struct *cur_task;
1876 * cgroup_taskset_first - reset taskset and return the first task
1877 * @tset: taskset of interest
1879 * @tset iteration is initialized and the first task is returned.
1881 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1883 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1884 tset->cur_task = NULL;
1886 return cgroup_taskset_next(tset);
1890 * cgroup_taskset_next - iterate to the next task in taskset
1891 * @tset: taskset of interest
1893 * Return the next task in @tset. Iteration must have been initialized
1894 * with cgroup_taskset_first().
1896 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1898 struct css_set *cset = tset->cur_cset;
1899 struct task_struct *task = tset->cur_task;
1901 while (&cset->mg_node != tset->csets) {
1903 task = list_first_entry(&cset->mg_tasks,
1904 struct task_struct, cg_list);
1906 task = list_next_entry(task, cg_list);
1908 if (&task->cg_list != &cset->mg_tasks) {
1909 tset->cur_cset = cset;
1910 tset->cur_task = task;
1914 cset = list_next_entry(cset, mg_node);
1922 * cgroup_task_migrate - move a task from one cgroup to another.
1923 * @old_cgrp: the cgroup @tsk is being migrated from
1924 * @tsk: the task being migrated
1925 * @new_cset: the new css_set @tsk is being attached to
1927 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1929 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1930 struct task_struct *tsk,
1931 struct css_set *new_cset)
1933 struct css_set *old_cset;
1935 lockdep_assert_held(&cgroup_mutex);
1936 lockdep_assert_held(&css_set_rwsem);
1939 * We are synchronized through threadgroup_lock() against PF_EXITING
1940 * setting such that we can't race against cgroup_exit() changing the
1941 * css_set to init_css_set and dropping the old one.
1943 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1944 old_cset = task_css_set(tsk);
1946 get_css_set(new_cset);
1947 rcu_assign_pointer(tsk->cgroups, new_cset);
1950 * Use move_tail so that cgroup_taskset_first() still returns the
1951 * leader after migration. This works because cgroup_migrate()
1952 * ensures that the dst_cset of the leader is the first on the
1953 * tset's dst_csets list.
1955 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1958 * We just gained a reference on old_cset by taking it from the
1959 * task. As trading it for new_cset is protected by cgroup_mutex,
1960 * we're safe to drop it here; it will be freed under RCU.
1962 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1963 put_css_set_locked(old_cset, false);
1967 * cgroup_migrate_finish - cleanup after attach
1968 * @preloaded_csets: list of preloaded css_sets
1970 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1971 * those functions for details.
1973 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1975 struct css_set *cset, *tmp_cset;
1977 lockdep_assert_held(&cgroup_mutex);
1979 down_write(&css_set_rwsem);
1980 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1981 cset->mg_src_cgrp = NULL;
1982 cset->mg_dst_cset = NULL;
1983 list_del_init(&cset->mg_preload_node);
1984 put_css_set_locked(cset, false);
1986 up_write(&css_set_rwsem);
1990 * cgroup_migrate_add_src - add a migration source css_set
1991 * @src_cset: the source css_set to add
1992 * @dst_cgrp: the destination cgroup
1993 * @preloaded_csets: list of preloaded css_sets
1995 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1996 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1997 * up by cgroup_migrate_finish().
1999 * This function may be called without holding threadgroup_lock even if the
2000 * target is a process. Threads may be created and destroyed but as long
2001 * as cgroup_mutex is not dropped, no new css_set can be put into play and
2002 * the preloaded css_sets are guaranteed to cover all migrations.
2004 static void cgroup_migrate_add_src(struct css_set *src_cset,
2005 struct cgroup *dst_cgrp,
2006 struct list_head *preloaded_csets)
2008 struct cgroup *src_cgrp;
2010 lockdep_assert_held(&cgroup_mutex);
2011 lockdep_assert_held(&css_set_rwsem);
2013 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2015 if (!list_empty(&src_cset->mg_preload_node))
2018 WARN_ON(src_cset->mg_src_cgrp);
2019 WARN_ON(!list_empty(&src_cset->mg_tasks));
2020 WARN_ON(!list_empty(&src_cset->mg_node));
2022 src_cset->mg_src_cgrp = src_cgrp;
2023 get_css_set(src_cset);
2024 list_add(&src_cset->mg_preload_node, preloaded_csets);
2028 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2029 * @dst_cgrp: the destination cgroup (may be %NULL)
2030 * @preloaded_csets: list of preloaded source css_sets
2032 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2033 * have been preloaded to @preloaded_csets. This function looks up and
2034 * pins all destination css_sets, links each to its source, and append them
2035 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2036 * source css_set is assumed to be its cgroup on the default hierarchy.
2038 * This function must be called after cgroup_migrate_add_src() has been
2039 * called on each migration source css_set. After migration is performed
2040 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2043 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2044 struct list_head *preloaded_csets)
2047 struct css_set *src_cset, *tmp_cset;
2049 lockdep_assert_held(&cgroup_mutex);
2052 * Except for the root, child_subsys_mask must be zero for a cgroup
2053 * with tasks so that child cgroups don't compete against tasks.
2055 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2056 dst_cgrp->child_subsys_mask)
2059 /* look up the dst cset for each src cset and link it to src */
2060 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2061 struct css_set *dst_cset;
2063 dst_cset = find_css_set(src_cset,
2064 dst_cgrp ?: src_cset->dfl_cgrp);
2068 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2071 * If src cset equals dst, it's noop. Drop the src.
2072 * cgroup_migrate() will skip the cset too. Note that we
2073 * can't handle src == dst as some nodes are used by both.
2075 if (src_cset == dst_cset) {
2076 src_cset->mg_src_cgrp = NULL;
2077 list_del_init(&src_cset->mg_preload_node);
2078 put_css_set(src_cset, false);
2079 put_css_set(dst_cset, false);
2083 src_cset->mg_dst_cset = dst_cset;
2085 if (list_empty(&dst_cset->mg_preload_node))
2086 list_add(&dst_cset->mg_preload_node, &csets);
2088 put_css_set(dst_cset, false);
2091 list_splice_tail(&csets, preloaded_csets);
2094 cgroup_migrate_finish(&csets);
2099 * cgroup_migrate - migrate a process or task to a cgroup
2100 * @cgrp: the destination cgroup
2101 * @leader: the leader of the process or the task to migrate
2102 * @threadgroup: whether @leader points to the whole process or a single task
2104 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2105 * process, the caller must be holding threadgroup_lock of @leader. The
2106 * caller is also responsible for invoking cgroup_migrate_add_src() and
2107 * cgroup_migrate_prepare_dst() on the targets before invoking this
2108 * function and following up with cgroup_migrate_finish().
2110 * As long as a controller's ->can_attach() doesn't fail, this function is
2111 * guaranteed to succeed. This means that, excluding ->can_attach()
2112 * failure, when migrating multiple targets, the success or failure can be
2113 * decided for all targets by invoking group_migrate_prepare_dst() before
2114 * actually starting migrating.
2116 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2119 struct cgroup_taskset tset = {
2120 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2121 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2122 .csets = &tset.src_csets,
2124 struct cgroup_subsys_state *css, *failed_css = NULL;
2125 struct css_set *cset, *tmp_cset;
2126 struct task_struct *task, *tmp_task;
2130 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2131 * already PF_EXITING could be freed from underneath us unless we
2132 * take an rcu_read_lock.
2134 down_write(&css_set_rwsem);
2138 /* @task either already exited or can't exit until the end */
2139 if (task->flags & PF_EXITING)
2142 /* leave @task alone if post_fork() hasn't linked it yet */
2143 if (list_empty(&task->cg_list))
2146 cset = task_css_set(task);
2147 if (!cset->mg_src_cgrp)
2151 * cgroup_taskset_first() must always return the leader.
2152 * Take care to avoid disturbing the ordering.
2154 list_move_tail(&task->cg_list, &cset->mg_tasks);
2155 if (list_empty(&cset->mg_node))
2156 list_add_tail(&cset->mg_node, &tset.src_csets);
2157 if (list_empty(&cset->mg_dst_cset->mg_node))
2158 list_move_tail(&cset->mg_dst_cset->mg_node,
2163 } while_each_thread(leader, task);
2165 up_write(&css_set_rwsem);
2167 /* methods shouldn't be called if no task is actually migrating */
2168 if (list_empty(&tset.src_csets))
2171 /* check that we can legitimately attach to the cgroup */
2172 for_each_e_css(css, i, cgrp) {
2173 if (css->ss->can_attach) {
2174 ret = css->ss->can_attach(css, &tset);
2177 goto out_cancel_attach;
2183 * Now that we're guaranteed success, proceed to move all tasks to
2184 * the new cgroup. There are no failure cases after here, so this
2185 * is the commit point.
2187 down_write(&css_set_rwsem);
2188 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2189 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2190 cgroup_task_migrate(cset->mg_src_cgrp, task,
2193 up_write(&css_set_rwsem);
2196 * Migration is committed, all target tasks are now on dst_csets.
2197 * Nothing is sensitive to fork() after this point. Notify
2198 * controllers that migration is complete.
2200 tset.csets = &tset.dst_csets;
2202 for_each_e_css(css, i, cgrp)
2203 if (css->ss->attach)
2204 css->ss->attach(css, &tset);
2207 goto out_release_tset;
2210 for_each_e_css(css, i, cgrp) {
2211 if (css == failed_css)
2213 if (css->ss->cancel_attach)
2214 css->ss->cancel_attach(css, &tset);
2217 down_write(&css_set_rwsem);
2218 list_splice_init(&tset.dst_csets, &tset.src_csets);
2219 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2220 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2221 list_del_init(&cset->mg_node);
2223 up_write(&css_set_rwsem);
2228 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2229 * @dst_cgrp: the cgroup to attach to
2230 * @leader: the task or the leader of the threadgroup to be attached
2231 * @threadgroup: attach the whole threadgroup?
2233 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2235 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2236 struct task_struct *leader, bool threadgroup)
2238 LIST_HEAD(preloaded_csets);
2239 struct task_struct *task;
2242 /* look up all src csets */
2243 down_read(&css_set_rwsem);
2247 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2251 } while_each_thread(leader, task);
2253 up_read(&css_set_rwsem);
2255 /* prepare dst csets and commit */
2256 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2258 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2260 cgroup_migrate_finish(&preloaded_csets);
2265 * Find the task_struct of the task to attach by vpid and pass it along to the
2266 * function to attach either it or all tasks in its threadgroup. Will lock
2267 * cgroup_mutex and threadgroup.
2269 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2270 size_t nbytes, loff_t off, bool threadgroup)
2272 struct task_struct *tsk;
2273 const struct cred *cred = current_cred(), *tcred;
2274 struct cgroup *cgrp;
2278 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2281 cgrp = cgroup_kn_lock_live(of->kn);
2288 tsk = find_task_by_vpid(pid);
2292 goto out_unlock_cgroup;
2295 * even if we're attaching all tasks in the thread group, we
2296 * only need to check permissions on one of them.
2298 tcred = __task_cred(tsk);
2299 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2300 !uid_eq(cred->euid, tcred->uid) &&
2301 !uid_eq(cred->euid, tcred->suid)) {
2304 goto out_unlock_cgroup;
2310 tsk = tsk->group_leader;
2313 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2314 * trapped in a cpuset, or RT worker may be born in a cgroup
2315 * with no rt_runtime allocated. Just say no.
2317 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2320 goto out_unlock_cgroup;
2323 get_task_struct(tsk);
2326 threadgroup_lock(tsk);
2328 if (!thread_group_leader(tsk)) {
2330 * a race with de_thread from another thread's exec()
2331 * may strip us of our leadership, if this happens,
2332 * there is no choice but to throw this task away and
2333 * try again; this is
2334 * "double-double-toil-and-trouble-check locking".
2336 threadgroup_unlock(tsk);
2337 put_task_struct(tsk);
2338 goto retry_find_task;
2342 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2344 threadgroup_unlock(tsk);
2346 put_task_struct(tsk);
2348 cgroup_kn_unlock(of->kn);
2349 return ret ?: nbytes;
2353 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2354 * @from: attach to all cgroups of a given task
2355 * @tsk: the task to be attached
2357 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2359 struct cgroup_root *root;
2362 mutex_lock(&cgroup_mutex);
2363 for_each_root(root) {
2364 struct cgroup *from_cgrp;
2366 if (root == &cgrp_dfl_root)
2369 down_read(&css_set_rwsem);
2370 from_cgrp = task_cgroup_from_root(from, root);
2371 up_read(&css_set_rwsem);
2373 retval = cgroup_attach_task(from_cgrp, tsk, false);
2377 mutex_unlock(&cgroup_mutex);
2381 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2383 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2384 char *buf, size_t nbytes, loff_t off)
2386 return __cgroup_procs_write(of, buf, nbytes, off, false);
2389 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2390 char *buf, size_t nbytes, loff_t off)
2392 return __cgroup_procs_write(of, buf, nbytes, off, true);
2395 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2396 char *buf, size_t nbytes, loff_t off)
2398 struct cgroup *cgrp;
2400 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2402 cgrp = cgroup_kn_lock_live(of->kn);
2405 spin_lock(&release_agent_path_lock);
2406 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2407 sizeof(cgrp->root->release_agent_path));
2408 spin_unlock(&release_agent_path_lock);
2409 cgroup_kn_unlock(of->kn);
2413 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2415 struct cgroup *cgrp = seq_css(seq)->cgroup;
2417 spin_lock(&release_agent_path_lock);
2418 seq_puts(seq, cgrp->root->release_agent_path);
2419 spin_unlock(&release_agent_path_lock);
2420 seq_putc(seq, '\n');
2424 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2426 struct cgroup *cgrp = seq_css(seq)->cgroup;
2428 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2432 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2434 struct cgroup_subsys *ss;
2435 bool printed = false;
2438 for_each_subsys(ss, ssid) {
2439 if (ss_mask & (1 << ssid)) {
2442 seq_printf(seq, "%s", ss->name);
2447 seq_putc(seq, '\n');
2450 /* show controllers which are currently attached to the default hierarchy */
2451 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2453 struct cgroup *cgrp = seq_css(seq)->cgroup;
2455 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2456 ~cgrp_dfl_root_inhibit_ss_mask);
2460 /* show controllers which are enabled from the parent */
2461 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2463 struct cgroup *cgrp = seq_css(seq)->cgroup;
2465 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2469 /* show controllers which are enabled for a given cgroup's children */
2470 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2472 struct cgroup *cgrp = seq_css(seq)->cgroup;
2474 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2479 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2480 * @cgrp: root of the subtree to update csses for
2482 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2483 * css associations need to be updated accordingly. This function looks up
2484 * all css_sets which are attached to the subtree, creates the matching
2485 * updated css_sets and migrates the tasks to the new ones.
2487 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2489 LIST_HEAD(preloaded_csets);
2490 struct cgroup_subsys_state *css;
2491 struct css_set *src_cset;
2494 lockdep_assert_held(&cgroup_mutex);
2496 /* look up all csses currently attached to @cgrp's subtree */
2497 down_read(&css_set_rwsem);
2498 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2499 struct cgrp_cset_link *link;
2501 /* self is not affected by child_subsys_mask change */
2502 if (css->cgroup == cgrp)
2505 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2506 cgroup_migrate_add_src(link->cset, cgrp,
2509 up_read(&css_set_rwsem);
2511 /* NULL dst indicates self on default hierarchy */
2512 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2516 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2517 struct task_struct *last_task = NULL, *task;
2519 /* src_csets precede dst_csets, break on the first dst_cset */
2520 if (!src_cset->mg_src_cgrp)
2524 * All tasks in src_cset need to be migrated to the
2525 * matching dst_cset. Empty it process by process. We
2526 * walk tasks but migrate processes. The leader might even
2527 * belong to a different cset but such src_cset would also
2528 * be among the target src_csets because the default
2529 * hierarchy enforces per-process membership.
2532 down_read(&css_set_rwsem);
2533 task = list_first_entry_or_null(&src_cset->tasks,
2534 struct task_struct, cg_list);
2536 task = task->group_leader;
2537 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2538 get_task_struct(task);
2540 up_read(&css_set_rwsem);
2545 /* guard against possible infinite loop */
2546 if (WARN(last_task == task,
2547 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2551 threadgroup_lock(task);
2552 /* raced against de_thread() from another thread? */
2553 if (!thread_group_leader(task)) {
2554 threadgroup_unlock(task);
2555 put_task_struct(task);
2559 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2561 threadgroup_unlock(task);
2562 put_task_struct(task);
2564 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2570 cgroup_migrate_finish(&preloaded_csets);
2574 /* change the enabled child controllers for a cgroup in the default hierarchy */
2575 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2576 char *buf, size_t nbytes,
2579 unsigned int enable = 0, disable = 0;
2580 struct cgroup *cgrp, *child;
2581 struct cgroup_subsys *ss;
2586 * Parse input - space separated list of subsystem names prefixed
2587 * with either + or -.
2589 buf = strstrip(buf);
2590 while ((tok = strsep(&buf, " "))) {
2593 for_each_subsys(ss, ssid) {
2594 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2595 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2599 enable |= 1 << ssid;
2600 disable &= ~(1 << ssid);
2601 } else if (*tok == '-') {
2602 disable |= 1 << ssid;
2603 enable &= ~(1 << ssid);
2609 if (ssid == CGROUP_SUBSYS_COUNT)
2613 cgrp = cgroup_kn_lock_live(of->kn);
2617 for_each_subsys(ss, ssid) {
2618 if (enable & (1 << ssid)) {
2619 if (cgrp->subtree_control & (1 << ssid)) {
2620 enable &= ~(1 << ssid);
2624 /* unavailable or not enabled on the parent? */
2625 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2626 (cgroup_parent(cgrp) &&
2627 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2633 * Because css offlining is asynchronous, userland
2634 * might try to re-enable the same controller while
2635 * the previous instance is still around. In such
2636 * cases, wait till it's gone using offline_waitq.
2638 cgroup_for_each_live_child(child, cgrp) {
2641 if (!cgroup_css(child, ss))
2645 prepare_to_wait(&child->offline_waitq, &wait,
2646 TASK_UNINTERRUPTIBLE);
2647 cgroup_kn_unlock(of->kn);
2649 finish_wait(&child->offline_waitq, &wait);
2652 return restart_syscall();
2654 } else if (disable & (1 << ssid)) {
2655 if (!(cgrp->subtree_control & (1 << ssid))) {
2656 disable &= ~(1 << ssid);
2660 /* a child has it enabled? */
2661 cgroup_for_each_live_child(child, cgrp) {
2662 if (child->subtree_control & (1 << ssid)) {
2670 if (!enable && !disable) {
2676 * Except for the root, subtree_control must be zero for a cgroup
2677 * with tasks so that child cgroups don't compete against tasks.
2679 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2684 cgrp->subtree_control |= enable;
2685 cgrp->subtree_control &= ~disable;
2686 cgroup_refresh_child_subsys_mask(cgrp);
2688 /* create new csses */
2689 for_each_subsys(ss, ssid) {
2690 if (!(enable & (1 << ssid)))
2693 cgroup_for_each_live_child(child, cgrp) {
2694 ret = create_css(child, ss);
2701 * At this point, cgroup_e_css() results reflect the new csses
2702 * making the following cgroup_update_dfl_csses() properly update
2703 * css associations of all tasks in the subtree.
2705 ret = cgroup_update_dfl_csses(cgrp);
2709 /* all tasks are now migrated away from the old csses, kill them */
2710 for_each_subsys(ss, ssid) {
2711 if (!(disable & (1 << ssid)))
2714 cgroup_for_each_live_child(child, cgrp)
2715 kill_css(cgroup_css(child, ss));
2718 kernfs_activate(cgrp->kn);
2721 cgroup_kn_unlock(of->kn);
2722 return ret ?: nbytes;
2725 cgrp->subtree_control &= ~enable;
2726 cgrp->subtree_control |= disable;
2727 cgroup_refresh_child_subsys_mask(cgrp);
2729 for_each_subsys(ss, ssid) {
2730 if (!(enable & (1 << ssid)))
2733 cgroup_for_each_live_child(child, cgrp) {
2734 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2742 static int cgroup_populated_show(struct seq_file *seq, void *v)
2744 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2748 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2749 size_t nbytes, loff_t off)
2751 struct cgroup *cgrp = of->kn->parent->priv;
2752 struct cftype *cft = of->kn->priv;
2753 struct cgroup_subsys_state *css;
2757 return cft->write(of, buf, nbytes, off);
2760 * kernfs guarantees that a file isn't deleted with operations in
2761 * flight, which means that the matching css is and stays alive and
2762 * doesn't need to be pinned. The RCU locking is not necessary
2763 * either. It's just for the convenience of using cgroup_css().
2766 css = cgroup_css(cgrp, cft->ss);
2769 if (cft->write_u64) {
2770 unsigned long long v;
2771 ret = kstrtoull(buf, 0, &v);
2773 ret = cft->write_u64(css, cft, v);
2774 } else if (cft->write_s64) {
2776 ret = kstrtoll(buf, 0, &v);
2778 ret = cft->write_s64(css, cft, v);
2783 return ret ?: nbytes;
2786 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2788 return seq_cft(seq)->seq_start(seq, ppos);
2791 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2793 return seq_cft(seq)->seq_next(seq, v, ppos);
2796 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2798 seq_cft(seq)->seq_stop(seq, v);
2801 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2803 struct cftype *cft = seq_cft(m);
2804 struct cgroup_subsys_state *css = seq_css(m);
2807 return cft->seq_show(m, arg);
2810 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2811 else if (cft->read_s64)
2812 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2818 static struct kernfs_ops cgroup_kf_single_ops = {
2819 .atomic_write_len = PAGE_SIZE,
2820 .write = cgroup_file_write,
2821 .seq_show = cgroup_seqfile_show,
2824 static struct kernfs_ops cgroup_kf_ops = {
2825 .atomic_write_len = PAGE_SIZE,
2826 .write = cgroup_file_write,
2827 .seq_start = cgroup_seqfile_start,
2828 .seq_next = cgroup_seqfile_next,
2829 .seq_stop = cgroup_seqfile_stop,
2830 .seq_show = cgroup_seqfile_show,
2834 * cgroup_rename - Only allow simple rename of directories in place.
2836 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2837 const char *new_name_str)
2839 struct cgroup *cgrp = kn->priv;
2842 if (kernfs_type(kn) != KERNFS_DIR)
2844 if (kn->parent != new_parent)
2848 * This isn't a proper migration and its usefulness is very
2849 * limited. Disallow if sane_behavior.
2851 if (cgroup_sane_behavior(cgrp))
2855 * We're gonna grab cgroup_mutex which nests outside kernfs
2856 * active_ref. kernfs_rename() doesn't require active_ref
2857 * protection. Break them before grabbing cgroup_mutex.
2859 kernfs_break_active_protection(new_parent);
2860 kernfs_break_active_protection(kn);
2862 mutex_lock(&cgroup_mutex);
2864 ret = kernfs_rename(kn, new_parent, new_name_str);
2866 mutex_unlock(&cgroup_mutex);
2868 kernfs_unbreak_active_protection(kn);
2869 kernfs_unbreak_active_protection(new_parent);
2873 /* set uid and gid of cgroup dirs and files to that of the creator */
2874 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2876 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2877 .ia_uid = current_fsuid(),
2878 .ia_gid = current_fsgid(), };
2880 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2881 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2884 return kernfs_setattr(kn, &iattr);
2887 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2889 char name[CGROUP_FILE_NAME_MAX];
2890 struct kernfs_node *kn;
2891 struct lock_class_key *key = NULL;
2894 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2895 key = &cft->lockdep_key;
2897 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2898 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2903 ret = cgroup_kn_set_ugid(kn);
2909 if (cft->seq_show == cgroup_populated_show)
2910 cgrp->populated_kn = kn;
2915 * cgroup_addrm_files - add or remove files to a cgroup directory
2916 * @cgrp: the target cgroup
2917 * @cfts: array of cftypes to be added
2918 * @is_add: whether to add or remove
2920 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2921 * For removals, this function never fails. If addition fails, this
2922 * function doesn't remove files already added. The caller is responsible
2925 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2931 lockdep_assert_held(&cgroup_mutex);
2933 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2934 /* does cft->flags tell us to skip this file on @cgrp? */
2935 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2937 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2939 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
2941 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
2945 ret = cgroup_add_file(cgrp, cft);
2947 pr_warn("%s: failed to add %s, err=%d\n",
2948 __func__, cft->name, ret);
2952 cgroup_rm_file(cgrp, cft);
2958 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2961 struct cgroup_subsys *ss = cfts[0].ss;
2962 struct cgroup *root = &ss->root->cgrp;
2963 struct cgroup_subsys_state *css;
2966 lockdep_assert_held(&cgroup_mutex);
2968 /* add/rm files for all cgroups created before */
2969 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2970 struct cgroup *cgrp = css->cgroup;
2972 if (cgroup_is_dead(cgrp))
2975 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2981 kernfs_activate(root->kn);
2985 static void cgroup_exit_cftypes(struct cftype *cfts)
2989 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2990 /* free copy for custom atomic_write_len, see init_cftypes() */
2991 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2998 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3002 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3003 struct kernfs_ops *kf_ops;
3005 WARN_ON(cft->ss || cft->kf_ops);
3008 kf_ops = &cgroup_kf_ops;
3010 kf_ops = &cgroup_kf_single_ops;
3013 * Ugh... if @cft wants a custom max_write_len, we need to
3014 * make a copy of kf_ops to set its atomic_write_len.
3016 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3017 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3019 cgroup_exit_cftypes(cfts);
3022 kf_ops->atomic_write_len = cft->max_write_len;
3025 cft->kf_ops = kf_ops;
3032 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3034 lockdep_assert_held(&cgroup_mutex);
3036 if (!cfts || !cfts[0].ss)
3039 list_del(&cfts->node);
3040 cgroup_apply_cftypes(cfts, false);
3041 cgroup_exit_cftypes(cfts);
3046 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3047 * @cfts: zero-length name terminated array of cftypes
3049 * Unregister @cfts. Files described by @cfts are removed from all
3050 * existing cgroups and all future cgroups won't have them either. This
3051 * function can be called anytime whether @cfts' subsys is attached or not.
3053 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3056 int cgroup_rm_cftypes(struct cftype *cfts)
3060 mutex_lock(&cgroup_mutex);
3061 ret = cgroup_rm_cftypes_locked(cfts);
3062 mutex_unlock(&cgroup_mutex);
3067 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3068 * @ss: target cgroup subsystem
3069 * @cfts: zero-length name terminated array of cftypes
3071 * Register @cfts to @ss. Files described by @cfts are created for all
3072 * existing cgroups to which @ss is attached and all future cgroups will
3073 * have them too. This function can be called anytime whether @ss is
3076 * Returns 0 on successful registration, -errno on failure. Note that this
3077 * function currently returns 0 as long as @cfts registration is successful
3078 * even if some file creation attempts on existing cgroups fail.
3080 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3087 if (!cfts || cfts[0].name[0] == '\0')
3090 ret = cgroup_init_cftypes(ss, cfts);
3094 mutex_lock(&cgroup_mutex);
3096 list_add_tail(&cfts->node, &ss->cfts);
3097 ret = cgroup_apply_cftypes(cfts, true);
3099 cgroup_rm_cftypes_locked(cfts);
3101 mutex_unlock(&cgroup_mutex);
3106 * cgroup_task_count - count the number of tasks in a cgroup.
3107 * @cgrp: the cgroup in question
3109 * Return the number of tasks in the cgroup.
3111 static int cgroup_task_count(const struct cgroup *cgrp)
3114 struct cgrp_cset_link *link;
3116 down_read(&css_set_rwsem);
3117 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3118 count += atomic_read(&link->cset->refcount);
3119 up_read(&css_set_rwsem);
3124 * css_next_child - find the next child of a given css
3125 * @pos: the current position (%NULL to initiate traversal)
3126 * @parent: css whose children to walk
3128 * This function returns the next child of @parent and should be called
3129 * under either cgroup_mutex or RCU read lock. The only requirement is
3130 * that @parent and @pos are accessible. The next sibling is guaranteed to
3131 * be returned regardless of their states.
3133 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3134 * css which finished ->css_online() is guaranteed to be visible in the
3135 * future iterations and will stay visible until the last reference is put.
3136 * A css which hasn't finished ->css_online() or already finished
3137 * ->css_offline() may show up during traversal. It's each subsystem's
3138 * responsibility to synchronize against on/offlining.
3140 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3141 struct cgroup_subsys_state *parent)
3143 struct cgroup_subsys_state *next;
3145 cgroup_assert_mutex_or_rcu_locked();
3148 * @pos could already have been unlinked from the sibling list.
3149 * Once a cgroup is removed, its ->sibling.next is no longer
3150 * updated when its next sibling changes. CSS_RELEASED is set when
3151 * @pos is taken off list, at which time its next pointer is valid,
3152 * and, as releases are serialized, the one pointed to by the next
3153 * pointer is guaranteed to not have started release yet. This
3154 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3155 * critical section, the one pointed to by its next pointer is
3156 * guaranteed to not have finished its RCU grace period even if we
3157 * have dropped rcu_read_lock() inbetween iterations.
3159 * If @pos has CSS_RELEASED set, its next pointer can't be
3160 * dereferenced; however, as each css is given a monotonically
3161 * increasing unique serial number and always appended to the
3162 * sibling list, the next one can be found by walking the parent's
3163 * children until the first css with higher serial number than
3164 * @pos's. While this path can be slower, it happens iff iteration
3165 * races against release and the race window is very small.
3168 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3169 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3170 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3172 list_for_each_entry_rcu(next, &parent->children, sibling)
3173 if (next->serial_nr > pos->serial_nr)
3178 * @next, if not pointing to the head, can be dereferenced and is
3181 if (&next->sibling != &parent->children)
3187 * css_next_descendant_pre - find the next descendant for pre-order walk
3188 * @pos: the current position (%NULL to initiate traversal)
3189 * @root: css whose descendants to walk
3191 * To be used by css_for_each_descendant_pre(). Find the next descendant
3192 * to visit for pre-order traversal of @root's descendants. @root is
3193 * included in the iteration and the first node to be visited.
3195 * While this function requires cgroup_mutex or RCU read locking, it
3196 * doesn't require the whole traversal to be contained in a single critical
3197 * section. This function will return the correct next descendant as long
3198 * as both @pos and @root are accessible and @pos is a descendant of @root.
3200 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3201 * css which finished ->css_online() is guaranteed to be visible in the
3202 * future iterations and will stay visible until the last reference is put.
3203 * A css which hasn't finished ->css_online() or already finished
3204 * ->css_offline() may show up during traversal. It's each subsystem's
3205 * responsibility to synchronize against on/offlining.
3207 struct cgroup_subsys_state *
3208 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3209 struct cgroup_subsys_state *root)
3211 struct cgroup_subsys_state *next;
3213 cgroup_assert_mutex_or_rcu_locked();
3215 /* if first iteration, visit @root */
3219 /* visit the first child if exists */
3220 next = css_next_child(NULL, pos);
3224 /* no child, visit my or the closest ancestor's next sibling */
3225 while (pos != root) {
3226 next = css_next_child(pos, pos->parent);
3236 * css_rightmost_descendant - return the rightmost descendant of a css
3237 * @pos: css of interest
3239 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3240 * is returned. This can be used during pre-order traversal to skip
3243 * While this function requires cgroup_mutex or RCU read locking, it
3244 * doesn't require the whole traversal to be contained in a single critical
3245 * section. This function will return the correct rightmost descendant as
3246 * long as @pos is accessible.
3248 struct cgroup_subsys_state *
3249 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3251 struct cgroup_subsys_state *last, *tmp;
3253 cgroup_assert_mutex_or_rcu_locked();
3257 /* ->prev isn't RCU safe, walk ->next till the end */
3259 css_for_each_child(tmp, last)
3266 static struct cgroup_subsys_state *
3267 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3269 struct cgroup_subsys_state *last;
3273 pos = css_next_child(NULL, pos);
3280 * css_next_descendant_post - find the next descendant for post-order walk
3281 * @pos: the current position (%NULL to initiate traversal)
3282 * @root: css whose descendants to walk
3284 * To be used by css_for_each_descendant_post(). Find the next descendant
3285 * to visit for post-order traversal of @root's descendants. @root is
3286 * included in the iteration and the last node to be visited.
3288 * While this function requires cgroup_mutex or RCU read locking, it
3289 * doesn't require the whole traversal to be contained in a single critical
3290 * section. This function will return the correct next descendant as long
3291 * as both @pos and @cgroup are accessible and @pos is a descendant of
3294 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3295 * css which finished ->css_online() is guaranteed to be visible in the
3296 * future iterations and will stay visible until the last reference is put.
3297 * A css which hasn't finished ->css_online() or already finished
3298 * ->css_offline() may show up during traversal. It's each subsystem's
3299 * responsibility to synchronize against on/offlining.
3301 struct cgroup_subsys_state *
3302 css_next_descendant_post(struct cgroup_subsys_state *pos,
3303 struct cgroup_subsys_state *root)
3305 struct cgroup_subsys_state *next;
3307 cgroup_assert_mutex_or_rcu_locked();
3309 /* if first iteration, visit leftmost descendant which may be @root */
3311 return css_leftmost_descendant(root);
3313 /* if we visited @root, we're done */
3317 /* if there's an unvisited sibling, visit its leftmost descendant */
3318 next = css_next_child(pos, pos->parent);
3320 return css_leftmost_descendant(next);
3322 /* no sibling left, visit parent */
3327 * css_has_online_children - does a css have online children
3328 * @css: the target css
3330 * Returns %true if @css has any online children; otherwise, %false. This
3331 * function can be called from any context but the caller is responsible
3332 * for synchronizing against on/offlining as necessary.
3334 bool css_has_online_children(struct cgroup_subsys_state *css)
3336 struct cgroup_subsys_state *child;
3340 css_for_each_child(child, css) {
3341 if (css->flags & CSS_ONLINE) {
3351 * css_advance_task_iter - advance a task itererator to the next css_set
3352 * @it: the iterator to advance
3354 * Advance @it to the next css_set to walk.
3356 static void css_advance_task_iter(struct css_task_iter *it)
3358 struct list_head *l = it->cset_pos;
3359 struct cgrp_cset_link *link;
3360 struct css_set *cset;
3362 /* Advance to the next non-empty css_set */
3365 if (l == it->cset_head) {
3366 it->cset_pos = NULL;
3371 cset = container_of(l, struct css_set,
3372 e_cset_node[it->ss->id]);
3374 link = list_entry(l, struct cgrp_cset_link, cset_link);
3377 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3381 if (!list_empty(&cset->tasks))
3382 it->task_pos = cset->tasks.next;
3384 it->task_pos = cset->mg_tasks.next;
3386 it->tasks_head = &cset->tasks;
3387 it->mg_tasks_head = &cset->mg_tasks;
3391 * css_task_iter_start - initiate task iteration
3392 * @css: the css to walk tasks of
3393 * @it: the task iterator to use
3395 * Initiate iteration through the tasks of @css. The caller can call
3396 * css_task_iter_next() to walk through the tasks until the function
3397 * returns NULL. On completion of iteration, css_task_iter_end() must be
3400 * Note that this function acquires a lock which is released when the
3401 * iteration finishes. The caller can't sleep while iteration is in
3404 void css_task_iter_start(struct cgroup_subsys_state *css,
3405 struct css_task_iter *it)
3406 __acquires(css_set_rwsem)
3408 /* no one should try to iterate before mounting cgroups */
3409 WARN_ON_ONCE(!use_task_css_set_links);
3411 down_read(&css_set_rwsem);
3416 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3418 it->cset_pos = &css->cgroup->cset_links;
3420 it->cset_head = it->cset_pos;
3422 css_advance_task_iter(it);
3426 * css_task_iter_next - return the next task for the iterator
3427 * @it: the task iterator being iterated
3429 * The "next" function for task iteration. @it should have been
3430 * initialized via css_task_iter_start(). Returns NULL when the iteration
3433 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3435 struct task_struct *res;
3436 struct list_head *l = it->task_pos;
3438 /* If the iterator cg is NULL, we have no tasks */
3441 res = list_entry(l, struct task_struct, cg_list);
3444 * Advance iterator to find next entry. cset->tasks is consumed
3445 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3450 if (l == it->tasks_head)
3451 l = it->mg_tasks_head->next;
3453 if (l == it->mg_tasks_head)
3454 css_advance_task_iter(it);
3462 * css_task_iter_end - finish task iteration
3463 * @it: the task iterator to finish
3465 * Finish task iteration started by css_task_iter_start().
3467 void css_task_iter_end(struct css_task_iter *it)
3468 __releases(css_set_rwsem)
3470 up_read(&css_set_rwsem);
3474 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3475 * @to: cgroup to which the tasks will be moved
3476 * @from: cgroup in which the tasks currently reside
3478 * Locking rules between cgroup_post_fork() and the migration path
3479 * guarantee that, if a task is forking while being migrated, the new child
3480 * is guaranteed to be either visible in the source cgroup after the
3481 * parent's migration is complete or put into the target cgroup. No task
3482 * can slip out of migration through forking.
3484 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3486 LIST_HEAD(preloaded_csets);
3487 struct cgrp_cset_link *link;
3488 struct css_task_iter it;
3489 struct task_struct *task;
3492 mutex_lock(&cgroup_mutex);
3494 /* all tasks in @from are being moved, all csets are source */
3495 down_read(&css_set_rwsem);
3496 list_for_each_entry(link, &from->cset_links, cset_link)
3497 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3498 up_read(&css_set_rwsem);
3500 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3505 * Migrate tasks one-by-one until @form is empty. This fails iff
3506 * ->can_attach() fails.
3509 css_task_iter_start(&from->self, &it);
3510 task = css_task_iter_next(&it);
3512 get_task_struct(task);
3513 css_task_iter_end(&it);
3516 ret = cgroup_migrate(to, task, false);
3517 put_task_struct(task);
3519 } while (task && !ret);
3521 cgroup_migrate_finish(&preloaded_csets);
3522 mutex_unlock(&cgroup_mutex);
3527 * Stuff for reading the 'tasks'/'procs' files.
3529 * Reading this file can return large amounts of data if a cgroup has
3530 * *lots* of attached tasks. So it may need several calls to read(),
3531 * but we cannot guarantee that the information we produce is correct
3532 * unless we produce it entirely atomically.
3536 /* which pidlist file are we talking about? */
3537 enum cgroup_filetype {
3543 * A pidlist is a list of pids that virtually represents the contents of one
3544 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3545 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3548 struct cgroup_pidlist {
3550 * used to find which pidlist is wanted. doesn't change as long as
3551 * this particular list stays in the list.
3553 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3556 /* how many elements the above list has */
3558 /* each of these stored in a list by its cgroup */
3559 struct list_head links;
3560 /* pointer to the cgroup we belong to, for list removal purposes */
3561 struct cgroup *owner;
3562 /* for delayed destruction */
3563 struct delayed_work destroy_dwork;
3567 * The following two functions "fix" the issue where there are more pids
3568 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3569 * TODO: replace with a kernel-wide solution to this problem
3571 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3572 static void *pidlist_allocate(int count)
3574 if (PIDLIST_TOO_LARGE(count))
3575 return vmalloc(count * sizeof(pid_t));
3577 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3580 static void pidlist_free(void *p)
3582 if (is_vmalloc_addr(p))
3589 * Used to destroy all pidlists lingering waiting for destroy timer. None
3590 * should be left afterwards.
3592 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3594 struct cgroup_pidlist *l, *tmp_l;
3596 mutex_lock(&cgrp->pidlist_mutex);
3597 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3598 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3599 mutex_unlock(&cgrp->pidlist_mutex);
3601 flush_workqueue(cgroup_pidlist_destroy_wq);
3602 BUG_ON(!list_empty(&cgrp->pidlists));
3605 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3607 struct delayed_work *dwork = to_delayed_work(work);
3608 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3610 struct cgroup_pidlist *tofree = NULL;
3612 mutex_lock(&l->owner->pidlist_mutex);
3615 * Destroy iff we didn't get queued again. The state won't change
3616 * as destroy_dwork can only be queued while locked.
3618 if (!delayed_work_pending(dwork)) {
3619 list_del(&l->links);
3620 pidlist_free(l->list);
3621 put_pid_ns(l->key.ns);
3625 mutex_unlock(&l->owner->pidlist_mutex);
3630 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3631 * Returns the number of unique elements.
3633 static int pidlist_uniq(pid_t *list, int length)
3638 * we presume the 0th element is unique, so i starts at 1. trivial
3639 * edge cases first; no work needs to be done for either
3641 if (length == 0 || length == 1)
3643 /* src and dest walk down the list; dest counts unique elements */
3644 for (src = 1; src < length; src++) {
3645 /* find next unique element */
3646 while (list[src] == list[src-1]) {
3651 /* dest always points to where the next unique element goes */
3652 list[dest] = list[src];
3660 * The two pid files - task and cgroup.procs - guaranteed that the result
3661 * is sorted, which forced this whole pidlist fiasco. As pid order is
3662 * different per namespace, each namespace needs differently sorted list,
3663 * making it impossible to use, for example, single rbtree of member tasks
3664 * sorted by task pointer. As pidlists can be fairly large, allocating one
3665 * per open file is dangerous, so cgroup had to implement shared pool of
3666 * pidlists keyed by cgroup and namespace.
3668 * All this extra complexity was caused by the original implementation
3669 * committing to an entirely unnecessary property. In the long term, we
3670 * want to do away with it. Explicitly scramble sort order if
3671 * sane_behavior so that no such expectation exists in the new interface.
3673 * Scrambling is done by swapping every two consecutive bits, which is
3674 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3676 static pid_t pid_fry(pid_t pid)
3678 unsigned a = pid & 0x55555555;
3679 unsigned b = pid & 0xAAAAAAAA;
3681 return (a << 1) | (b >> 1);
3684 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3686 if (cgroup_sane_behavior(cgrp))
3687 return pid_fry(pid);
3692 static int cmppid(const void *a, const void *b)
3694 return *(pid_t *)a - *(pid_t *)b;
3697 static int fried_cmppid(const void *a, const void *b)
3699 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3702 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3703 enum cgroup_filetype type)
3705 struct cgroup_pidlist *l;
3706 /* don't need task_nsproxy() if we're looking at ourself */
3707 struct pid_namespace *ns = task_active_pid_ns(current);
3709 lockdep_assert_held(&cgrp->pidlist_mutex);
3711 list_for_each_entry(l, &cgrp->pidlists, links)
3712 if (l->key.type == type && l->key.ns == ns)
3718 * find the appropriate pidlist for our purpose (given procs vs tasks)
3719 * returns with the lock on that pidlist already held, and takes care
3720 * of the use count, or returns NULL with no locks held if we're out of
3723 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3724 enum cgroup_filetype type)
3726 struct cgroup_pidlist *l;
3728 lockdep_assert_held(&cgrp->pidlist_mutex);
3730 l = cgroup_pidlist_find(cgrp, type);
3734 /* entry not found; create a new one */
3735 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3739 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3741 /* don't need task_nsproxy() if we're looking at ourself */
3742 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3744 list_add(&l->links, &cgrp->pidlists);
3749 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3751 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3752 struct cgroup_pidlist **lp)
3756 int pid, n = 0; /* used for populating the array */
3757 struct css_task_iter it;
3758 struct task_struct *tsk;
3759 struct cgroup_pidlist *l;
3761 lockdep_assert_held(&cgrp->pidlist_mutex);
3764 * If cgroup gets more users after we read count, we won't have
3765 * enough space - tough. This race is indistinguishable to the
3766 * caller from the case that the additional cgroup users didn't
3767 * show up until sometime later on.
3769 length = cgroup_task_count(cgrp);
3770 array = pidlist_allocate(length);
3773 /* now, populate the array */
3774 css_task_iter_start(&cgrp->self, &it);
3775 while ((tsk = css_task_iter_next(&it))) {
3776 if (unlikely(n == length))
3778 /* get tgid or pid for procs or tasks file respectively */
3779 if (type == CGROUP_FILE_PROCS)
3780 pid = task_tgid_vnr(tsk);
3782 pid = task_pid_vnr(tsk);
3783 if (pid > 0) /* make sure to only use valid results */
3786 css_task_iter_end(&it);
3788 /* now sort & (if procs) strip out duplicates */
3789 if (cgroup_sane_behavior(cgrp))
3790 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3792 sort(array, length, sizeof(pid_t), cmppid, NULL);
3793 if (type == CGROUP_FILE_PROCS)
3794 length = pidlist_uniq(array, length);
3796 l = cgroup_pidlist_find_create(cgrp, type);
3798 mutex_unlock(&cgrp->pidlist_mutex);
3799 pidlist_free(array);
3803 /* store array, freeing old if necessary */
3804 pidlist_free(l->list);
3812 * cgroupstats_build - build and fill cgroupstats
3813 * @stats: cgroupstats to fill information into
3814 * @dentry: A dentry entry belonging to the cgroup for which stats have
3817 * Build and fill cgroupstats so that taskstats can export it to user
3820 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3822 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3823 struct cgroup *cgrp;
3824 struct css_task_iter it;
3825 struct task_struct *tsk;
3827 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3828 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3829 kernfs_type(kn) != KERNFS_DIR)
3832 mutex_lock(&cgroup_mutex);
3835 * We aren't being called from kernfs and there's no guarantee on
3836 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3837 * @kn->priv is RCU safe. Let's do the RCU dancing.
3840 cgrp = rcu_dereference(kn->priv);
3841 if (!cgrp || cgroup_is_dead(cgrp)) {
3843 mutex_unlock(&cgroup_mutex);
3848 css_task_iter_start(&cgrp->self, &it);
3849 while ((tsk = css_task_iter_next(&it))) {
3850 switch (tsk->state) {
3852 stats->nr_running++;
3854 case TASK_INTERRUPTIBLE:
3855 stats->nr_sleeping++;
3857 case TASK_UNINTERRUPTIBLE:
3858 stats->nr_uninterruptible++;
3861 stats->nr_stopped++;
3864 if (delayacct_is_task_waiting_on_io(tsk))
3865 stats->nr_io_wait++;
3869 css_task_iter_end(&it);
3871 mutex_unlock(&cgroup_mutex);
3877 * seq_file methods for the tasks/procs files. The seq_file position is the
3878 * next pid to display; the seq_file iterator is a pointer to the pid
3879 * in the cgroup->l->list array.
3882 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3885 * Initially we receive a position value that corresponds to
3886 * one more than the last pid shown (or 0 on the first call or
3887 * after a seek to the start). Use a binary-search to find the
3888 * next pid to display, if any
3890 struct kernfs_open_file *of = s->private;
3891 struct cgroup *cgrp = seq_css(s)->cgroup;
3892 struct cgroup_pidlist *l;
3893 enum cgroup_filetype type = seq_cft(s)->private;
3894 int index = 0, pid = *pos;
3897 mutex_lock(&cgrp->pidlist_mutex);
3900 * !NULL @of->priv indicates that this isn't the first start()
3901 * after open. If the matching pidlist is around, we can use that.
3902 * Look for it. Note that @of->priv can't be used directly. It
3903 * could already have been destroyed.
3906 of->priv = cgroup_pidlist_find(cgrp, type);
3909 * Either this is the first start() after open or the matching
3910 * pidlist has been destroyed inbetween. Create a new one.
3913 ret = pidlist_array_load(cgrp, type,
3914 (struct cgroup_pidlist **)&of->priv);
3916 return ERR_PTR(ret);
3921 int end = l->length;
3923 while (index < end) {
3924 int mid = (index + end) / 2;
3925 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3928 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3934 /* If we're off the end of the array, we're done */
3935 if (index >= l->length)
3937 /* Update the abstract position to be the actual pid that we found */
3938 iter = l->list + index;
3939 *pos = cgroup_pid_fry(cgrp, *iter);
3943 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3945 struct kernfs_open_file *of = s->private;
3946 struct cgroup_pidlist *l = of->priv;
3949 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3950 CGROUP_PIDLIST_DESTROY_DELAY);
3951 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3954 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3956 struct kernfs_open_file *of = s->private;
3957 struct cgroup_pidlist *l = of->priv;
3959 pid_t *end = l->list + l->length;
3961 * Advance to the next pid in the array. If this goes off the
3968 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3973 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3975 return seq_printf(s, "%d\n", *(int *)v);
3978 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3981 return notify_on_release(css->cgroup);
3984 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3985 struct cftype *cft, u64 val)
3987 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3989 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3991 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3995 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3998 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4001 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4002 struct cftype *cft, u64 val)
4005 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4007 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4011 static struct cftype cgroup_base_files[] = {
4013 .name = "cgroup.procs",
4014 .seq_start = cgroup_pidlist_start,
4015 .seq_next = cgroup_pidlist_next,
4016 .seq_stop = cgroup_pidlist_stop,
4017 .seq_show = cgroup_pidlist_show,
4018 .private = CGROUP_FILE_PROCS,
4019 .write = cgroup_procs_write,
4020 .mode = S_IRUGO | S_IWUSR,
4023 .name = "cgroup.clone_children",
4024 .flags = CFTYPE_INSANE,
4025 .read_u64 = cgroup_clone_children_read,
4026 .write_u64 = cgroup_clone_children_write,
4029 .name = "cgroup.sane_behavior",
4030 .flags = CFTYPE_ONLY_ON_ROOT,
4031 .seq_show = cgroup_sane_behavior_show,
4034 .name = "cgroup.controllers",
4035 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_ONLY_ON_ROOT,
4036 .seq_show = cgroup_root_controllers_show,
4039 .name = "cgroup.controllers",
4040 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4041 .seq_show = cgroup_controllers_show,
4044 .name = "cgroup.subtree_control",
4045 .flags = CFTYPE_ONLY_ON_DFL,
4046 .seq_show = cgroup_subtree_control_show,
4047 .write = cgroup_subtree_control_write,
4050 .name = "cgroup.populated",
4051 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4052 .seq_show = cgroup_populated_show,
4056 * Historical crazy stuff. These don't have "cgroup." prefix and
4057 * don't exist if sane_behavior. If you're depending on these, be
4058 * prepared to be burned.
4062 .flags = CFTYPE_INSANE, /* use "procs" instead */
4063 .seq_start = cgroup_pidlist_start,
4064 .seq_next = cgroup_pidlist_next,
4065 .seq_stop = cgroup_pidlist_stop,
4066 .seq_show = cgroup_pidlist_show,
4067 .private = CGROUP_FILE_TASKS,
4068 .write = cgroup_tasks_write,
4069 .mode = S_IRUGO | S_IWUSR,
4072 .name = "notify_on_release",
4073 .flags = CFTYPE_INSANE,
4074 .read_u64 = cgroup_read_notify_on_release,
4075 .write_u64 = cgroup_write_notify_on_release,
4078 .name = "release_agent",
4079 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4080 .seq_show = cgroup_release_agent_show,
4081 .write = cgroup_release_agent_write,
4082 .max_write_len = PATH_MAX - 1,
4088 * cgroup_populate_dir - create subsys files in a cgroup directory
4089 * @cgrp: target cgroup
4090 * @subsys_mask: mask of the subsystem ids whose files should be added
4092 * On failure, no file is added.
4094 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4096 struct cgroup_subsys *ss;
4099 /* process cftsets of each subsystem */
4100 for_each_subsys(ss, i) {
4101 struct cftype *cfts;
4103 if (!(subsys_mask & (1 << i)))
4106 list_for_each_entry(cfts, &ss->cfts, node) {
4107 ret = cgroup_addrm_files(cgrp, cfts, true);
4114 cgroup_clear_dir(cgrp, subsys_mask);
4119 * css destruction is four-stage process.
4121 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4122 * Implemented in kill_css().
4124 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4125 * and thus css_tryget_online() is guaranteed to fail, the css can be
4126 * offlined by invoking offline_css(). After offlining, the base ref is
4127 * put. Implemented in css_killed_work_fn().
4129 * 3. When the percpu_ref reaches zero, the only possible remaining
4130 * accessors are inside RCU read sections. css_release() schedules the
4133 * 4. After the grace period, the css can be freed. Implemented in
4134 * css_free_work_fn().
4136 * It is actually hairier because both step 2 and 4 require process context
4137 * and thus involve punting to css->destroy_work adding two additional
4138 * steps to the already complex sequence.
4140 static void css_free_work_fn(struct work_struct *work)
4142 struct cgroup_subsys_state *css =
4143 container_of(work, struct cgroup_subsys_state, destroy_work);
4144 struct cgroup *cgrp = css->cgroup;
4149 css_put(css->parent);
4151 css->ss->css_free(css);
4154 /* cgroup free path */
4155 atomic_dec(&cgrp->root->nr_cgrps);
4156 cgroup_pidlist_destroy_all(cgrp);
4158 if (cgroup_parent(cgrp)) {
4160 * We get a ref to the parent, and put the ref when
4161 * this cgroup is being freed, so it's guaranteed
4162 * that the parent won't be destroyed before its
4165 cgroup_put(cgroup_parent(cgrp));
4166 kernfs_put(cgrp->kn);
4170 * This is root cgroup's refcnt reaching zero,
4171 * which indicates that the root should be
4174 cgroup_destroy_root(cgrp->root);
4179 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4181 struct cgroup_subsys_state *css =
4182 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4184 INIT_WORK(&css->destroy_work, css_free_work_fn);
4185 queue_work(cgroup_destroy_wq, &css->destroy_work);
4188 static void css_release_work_fn(struct work_struct *work)
4190 struct cgroup_subsys_state *css =
4191 container_of(work, struct cgroup_subsys_state, destroy_work);
4192 struct cgroup_subsys *ss = css->ss;
4193 struct cgroup *cgrp = css->cgroup;
4195 mutex_lock(&cgroup_mutex);
4197 css->flags |= CSS_RELEASED;
4198 list_del_rcu(&css->sibling);
4201 /* css release path */
4202 cgroup_idr_remove(&ss->css_idr, css->id);
4204 /* cgroup release path */
4205 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4209 mutex_unlock(&cgroup_mutex);
4211 call_rcu(&css->rcu_head, css_free_rcu_fn);
4214 static void css_release(struct percpu_ref *ref)
4216 struct cgroup_subsys_state *css =
4217 container_of(ref, struct cgroup_subsys_state, refcnt);
4219 INIT_WORK(&css->destroy_work, css_release_work_fn);
4220 queue_work(cgroup_destroy_wq, &css->destroy_work);
4223 static void init_and_link_css(struct cgroup_subsys_state *css,
4224 struct cgroup_subsys *ss, struct cgroup *cgrp)
4226 lockdep_assert_held(&cgroup_mutex);
4230 memset(css, 0, sizeof(*css));
4233 INIT_LIST_HEAD(&css->sibling);
4234 INIT_LIST_HEAD(&css->children);
4235 css->serial_nr = css_serial_nr_next++;
4237 if (cgroup_parent(cgrp)) {
4238 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4239 css_get(css->parent);
4242 BUG_ON(cgroup_css(cgrp, ss));
4245 /* invoke ->css_online() on a new CSS and mark it online if successful */
4246 static int online_css(struct cgroup_subsys_state *css)
4248 struct cgroup_subsys *ss = css->ss;
4251 lockdep_assert_held(&cgroup_mutex);
4254 ret = ss->css_online(css);
4256 css->flags |= CSS_ONLINE;
4257 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4262 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4263 static void offline_css(struct cgroup_subsys_state *css)
4265 struct cgroup_subsys *ss = css->ss;
4267 lockdep_assert_held(&cgroup_mutex);
4269 if (!(css->flags & CSS_ONLINE))
4272 if (ss->css_offline)
4273 ss->css_offline(css);
4275 css->flags &= ~CSS_ONLINE;
4276 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4278 wake_up_all(&css->cgroup->offline_waitq);
4282 * create_css - create a cgroup_subsys_state
4283 * @cgrp: the cgroup new css will be associated with
4284 * @ss: the subsys of new css
4286 * Create a new css associated with @cgrp - @ss pair. On success, the new
4287 * css is online and installed in @cgrp with all interface files created.
4288 * Returns 0 on success, -errno on failure.
4290 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4292 struct cgroup *parent = cgroup_parent(cgrp);
4293 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4294 struct cgroup_subsys_state *css;
4297 lockdep_assert_held(&cgroup_mutex);
4299 css = ss->css_alloc(parent_css);
4301 return PTR_ERR(css);
4303 init_and_link_css(css, ss, cgrp);
4305 err = percpu_ref_init(&css->refcnt, css_release);
4309 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4311 goto err_free_percpu_ref;
4314 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4318 /* @css is ready to be brought online now, make it visible */
4319 list_add_tail_rcu(&css->sibling, &parent_css->children);
4320 cgroup_idr_replace(&ss->css_idr, css, css->id);
4322 err = online_css(css);
4326 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4327 cgroup_parent(parent)) {
4328 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4329 current->comm, current->pid, ss->name);
4330 if (!strcmp(ss->name, "memory"))
4331 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4332 ss->warned_broken_hierarchy = true;
4338 list_del_rcu(&css->sibling);
4339 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4341 cgroup_idr_remove(&ss->css_idr, css->id);
4342 err_free_percpu_ref:
4343 percpu_ref_cancel_init(&css->refcnt);
4345 call_rcu(&css->rcu_head, css_free_rcu_fn);
4349 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4352 struct cgroup *parent, *cgrp;
4353 struct cgroup_root *root;
4354 struct cgroup_subsys *ss;
4355 struct kernfs_node *kn;
4358 parent = cgroup_kn_lock_live(parent_kn);
4361 root = parent->root;
4363 /* allocate the cgroup and its ID, 0 is reserved for the root */
4364 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4370 ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
4375 * Temporarily set the pointer to NULL, so idr_find() won't return
4376 * a half-baked cgroup.
4378 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4381 goto out_cancel_ref;
4384 init_cgroup_housekeeping(cgrp);
4386 cgrp->self.parent = &parent->self;
4389 if (notify_on_release(parent))
4390 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4392 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4393 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4395 /* create the directory */
4396 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4404 * This extra ref will be put in cgroup_free_fn() and guarantees
4405 * that @cgrp->kn is always accessible.
4409 cgrp->self.serial_nr = css_serial_nr_next++;
4411 /* allocation complete, commit to creation */
4412 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4413 atomic_inc(&root->nr_cgrps);
4417 * @cgrp is now fully operational. If something fails after this
4418 * point, it'll be released via the normal destruction path.
4420 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4422 ret = cgroup_kn_set_ugid(kn);
4426 ret = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4430 /* let's create and online css's */
4431 for_each_subsys(ss, ssid) {
4432 if (parent->child_subsys_mask & (1 << ssid)) {
4433 ret = create_css(cgrp, ss);
4440 * On the default hierarchy, a child doesn't automatically inherit
4441 * subtree_control from the parent. Each is configured manually.
4443 if (!cgroup_on_dfl(cgrp)) {
4444 cgrp->subtree_control = parent->subtree_control;
4445 cgroup_refresh_child_subsys_mask(cgrp);
4448 kernfs_activate(kn);
4454 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4456 percpu_ref_cancel_init(&cgrp->self.refcnt);
4460 cgroup_kn_unlock(parent_kn);
4464 cgroup_destroy_locked(cgrp);
4469 * This is called when the refcnt of a css is confirmed to be killed.
4470 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4471 * initate destruction and put the css ref from kill_css().
4473 static void css_killed_work_fn(struct work_struct *work)
4475 struct cgroup_subsys_state *css =
4476 container_of(work, struct cgroup_subsys_state, destroy_work);
4478 mutex_lock(&cgroup_mutex);
4480 mutex_unlock(&cgroup_mutex);
4485 /* css kill confirmation processing requires process context, bounce */
4486 static void css_killed_ref_fn(struct percpu_ref *ref)
4488 struct cgroup_subsys_state *css =
4489 container_of(ref, struct cgroup_subsys_state, refcnt);
4491 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4492 queue_work(cgroup_destroy_wq, &css->destroy_work);
4496 * kill_css - destroy a css
4497 * @css: css to destroy
4499 * This function initiates destruction of @css by removing cgroup interface
4500 * files and putting its base reference. ->css_offline() will be invoked
4501 * asynchronously once css_tryget_online() is guaranteed to fail and when
4502 * the reference count reaches zero, @css will be released.
4504 static void kill_css(struct cgroup_subsys_state *css)
4506 lockdep_assert_held(&cgroup_mutex);
4509 * This must happen before css is disassociated with its cgroup.
4510 * See seq_css() for details.
4512 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4515 * Killing would put the base ref, but we need to keep it alive
4516 * until after ->css_offline().
4521 * cgroup core guarantees that, by the time ->css_offline() is
4522 * invoked, no new css reference will be given out via
4523 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4524 * proceed to offlining css's because percpu_ref_kill() doesn't
4525 * guarantee that the ref is seen as killed on all CPUs on return.
4527 * Use percpu_ref_kill_and_confirm() to get notifications as each
4528 * css is confirmed to be seen as killed on all CPUs.
4530 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4534 * cgroup_destroy_locked - the first stage of cgroup destruction
4535 * @cgrp: cgroup to be destroyed
4537 * css's make use of percpu refcnts whose killing latency shouldn't be
4538 * exposed to userland and are RCU protected. Also, cgroup core needs to
4539 * guarantee that css_tryget_online() won't succeed by the time
4540 * ->css_offline() is invoked. To satisfy all the requirements,
4541 * destruction is implemented in the following two steps.
4543 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4544 * userland visible parts and start killing the percpu refcnts of
4545 * css's. Set up so that the next stage will be kicked off once all
4546 * the percpu refcnts are confirmed to be killed.
4548 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4549 * rest of destruction. Once all cgroup references are gone, the
4550 * cgroup is RCU-freed.
4552 * This function implements s1. After this step, @cgrp is gone as far as
4553 * the userland is concerned and a new cgroup with the same name may be
4554 * created. As cgroup doesn't care about the names internally, this
4555 * doesn't cause any problem.
4557 static int cgroup_destroy_locked(struct cgroup *cgrp)
4558 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4560 struct cgroup_subsys_state *css;
4564 lockdep_assert_held(&cgroup_mutex);
4567 * css_set_rwsem synchronizes access to ->cset_links and prevents
4568 * @cgrp from being removed while put_css_set() is in progress.
4570 down_read(&css_set_rwsem);
4571 empty = list_empty(&cgrp->cset_links);
4572 up_read(&css_set_rwsem);
4577 * Make sure there's no live children. We can't test emptiness of
4578 * ->self.children as dead children linger on it while being
4579 * drained; otherwise, "rmdir parent/child parent" may fail.
4581 if (css_has_online_children(&cgrp->self))
4585 * Mark @cgrp dead. This prevents further task migration and child
4586 * creation by disabling cgroup_lock_live_group().
4588 cgrp->self.flags &= ~CSS_ONLINE;
4590 /* initiate massacre of all css's */
4591 for_each_css(css, ssid, cgrp)
4594 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4595 raw_spin_lock(&release_list_lock);
4596 if (!list_empty(&cgrp->release_list))
4597 list_del_init(&cgrp->release_list);
4598 raw_spin_unlock(&release_list_lock);
4601 * Remove @cgrp directory along with the base files. @cgrp has an
4602 * extra ref on its kn.
4604 kernfs_remove(cgrp->kn);
4606 set_bit(CGRP_RELEASABLE, &cgroup_parent(cgrp)->flags);
4607 check_for_release(cgroup_parent(cgrp));
4609 /* put the base reference */
4610 percpu_ref_kill(&cgrp->self.refcnt);
4615 static int cgroup_rmdir(struct kernfs_node *kn)
4617 struct cgroup *cgrp;
4620 cgrp = cgroup_kn_lock_live(kn);
4623 cgroup_get(cgrp); /* for @kn->priv clearing */
4625 ret = cgroup_destroy_locked(cgrp);
4627 cgroup_kn_unlock(kn);
4630 * There are two control paths which try to determine cgroup from
4631 * dentry without going through kernfs - cgroupstats_build() and
4632 * css_tryget_online_from_dir(). Those are supported by RCU
4633 * protecting clearing of cgrp->kn->priv backpointer, which should
4634 * happen after all files under it have been removed.
4637 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4643 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4644 .remount_fs = cgroup_remount,
4645 .show_options = cgroup_show_options,
4646 .mkdir = cgroup_mkdir,
4647 .rmdir = cgroup_rmdir,
4648 .rename = cgroup_rename,
4651 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4653 struct cgroup_subsys_state *css;
4655 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4657 mutex_lock(&cgroup_mutex);
4659 idr_init(&ss->css_idr);
4660 INIT_LIST_HEAD(&ss->cfts);
4662 /* Create the root cgroup state for this subsystem */
4663 ss->root = &cgrp_dfl_root;
4664 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4665 /* We don't handle early failures gracefully */
4666 BUG_ON(IS_ERR(css));
4667 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4670 * Root csses are never destroyed and we can't initialize
4671 * percpu_ref during early init. Disable refcnting.
4673 css->flags |= CSS_NO_REF;
4676 /* allocation can't be done safely during early init */
4679 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4680 BUG_ON(css->id < 0);
4683 /* Update the init_css_set to contain a subsys
4684 * pointer to this state - since the subsystem is
4685 * newly registered, all tasks and hence the
4686 * init_css_set is in the subsystem's root cgroup. */
4687 init_css_set.subsys[ss->id] = css;
4689 need_forkexit_callback |= ss->fork || ss->exit;
4691 /* At system boot, before all subsystems have been
4692 * registered, no tasks have been forked, so we don't
4693 * need to invoke fork callbacks here. */
4694 BUG_ON(!list_empty(&init_task.tasks));
4696 BUG_ON(online_css(css));
4698 mutex_unlock(&cgroup_mutex);
4702 * cgroup_init_early - cgroup initialization at system boot
4704 * Initialize cgroups at system boot, and initialize any
4705 * subsystems that request early init.
4707 int __init cgroup_init_early(void)
4709 static struct cgroup_sb_opts __initdata opts =
4710 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4711 struct cgroup_subsys *ss;
4714 init_cgroup_root(&cgrp_dfl_root, &opts);
4715 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4717 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4719 for_each_subsys(ss, i) {
4720 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4721 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4722 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4724 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4725 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4728 ss->name = cgroup_subsys_name[i];
4731 cgroup_init_subsys(ss, true);
4737 * cgroup_init - cgroup initialization
4739 * Register cgroup filesystem and /proc file, and initialize
4740 * any subsystems that didn't request early init.
4742 int __init cgroup_init(void)
4744 struct cgroup_subsys *ss;
4748 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4750 mutex_lock(&cgroup_mutex);
4752 /* Add init_css_set to the hash table */
4753 key = css_set_hash(init_css_set.subsys);
4754 hash_add(css_set_table, &init_css_set.hlist, key);
4756 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4758 mutex_unlock(&cgroup_mutex);
4760 for_each_subsys(ss, ssid) {
4761 if (ss->early_init) {
4762 struct cgroup_subsys_state *css =
4763 init_css_set.subsys[ss->id];
4765 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4767 BUG_ON(css->id < 0);
4769 cgroup_init_subsys(ss, false);
4772 list_add_tail(&init_css_set.e_cset_node[ssid],
4773 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4776 * Setting dfl_root subsys_mask needs to consider the
4777 * disabled flag and cftype registration needs kmalloc,
4778 * both of which aren't available during early_init.
4780 if (!ss->disabled) {
4781 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4782 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4786 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4790 err = register_filesystem(&cgroup_fs_type);
4792 kobject_put(cgroup_kobj);
4796 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4800 static int __init cgroup_wq_init(void)
4803 * There isn't much point in executing destruction path in
4804 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4805 * Use 1 for @max_active.
4807 * We would prefer to do this in cgroup_init() above, but that
4808 * is called before init_workqueues(): so leave this until after.
4810 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4811 BUG_ON(!cgroup_destroy_wq);
4814 * Used to destroy pidlists and separate to serve as flush domain.
4815 * Cap @max_active to 1 too.
4817 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4819 BUG_ON(!cgroup_pidlist_destroy_wq);
4823 core_initcall(cgroup_wq_init);
4826 * proc_cgroup_show()
4827 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4828 * - Used for /proc/<pid>/cgroup.
4831 /* TODO: Use a proper seq_file iterator */
4832 int proc_cgroup_show(struct seq_file *m, void *v)
4835 struct task_struct *tsk;
4838 struct cgroup_root *root;
4841 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4847 tsk = get_pid_task(pid, PIDTYPE_PID);
4853 mutex_lock(&cgroup_mutex);
4854 down_read(&css_set_rwsem);
4856 for_each_root(root) {
4857 struct cgroup_subsys *ss;
4858 struct cgroup *cgrp;
4859 int ssid, count = 0;
4861 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4864 seq_printf(m, "%d:", root->hierarchy_id);
4865 for_each_subsys(ss, ssid)
4866 if (root->subsys_mask & (1 << ssid))
4867 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4868 if (strlen(root->name))
4869 seq_printf(m, "%sname=%s", count ? "," : "",
4872 cgrp = task_cgroup_from_root(tsk, root);
4873 path = cgroup_path(cgrp, buf, PATH_MAX);
4875 retval = -ENAMETOOLONG;
4883 up_read(&css_set_rwsem);
4884 mutex_unlock(&cgroup_mutex);
4885 put_task_struct(tsk);
4892 /* Display information about each subsystem and each hierarchy */
4893 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4895 struct cgroup_subsys *ss;
4898 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4900 * ideally we don't want subsystems moving around while we do this.
4901 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4902 * subsys/hierarchy state.
4904 mutex_lock(&cgroup_mutex);
4906 for_each_subsys(ss, i)
4907 seq_printf(m, "%s\t%d\t%d\t%d\n",
4908 ss->name, ss->root->hierarchy_id,
4909 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4911 mutex_unlock(&cgroup_mutex);
4915 static int cgroupstats_open(struct inode *inode, struct file *file)
4917 return single_open(file, proc_cgroupstats_show, NULL);
4920 static const struct file_operations proc_cgroupstats_operations = {
4921 .open = cgroupstats_open,
4923 .llseek = seq_lseek,
4924 .release = single_release,
4928 * cgroup_fork - initialize cgroup related fields during copy_process()
4929 * @child: pointer to task_struct of forking parent process.
4931 * A task is associated with the init_css_set until cgroup_post_fork()
4932 * attaches it to the parent's css_set. Empty cg_list indicates that
4933 * @child isn't holding reference to its css_set.
4935 void cgroup_fork(struct task_struct *child)
4937 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4938 INIT_LIST_HEAD(&child->cg_list);
4942 * cgroup_post_fork - called on a new task after adding it to the task list
4943 * @child: the task in question
4945 * Adds the task to the list running through its css_set if necessary and
4946 * call the subsystem fork() callbacks. Has to be after the task is
4947 * visible on the task list in case we race with the first call to
4948 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4951 void cgroup_post_fork(struct task_struct *child)
4953 struct cgroup_subsys *ss;
4957 * This may race against cgroup_enable_task_cg_links(). As that
4958 * function sets use_task_css_set_links before grabbing
4959 * tasklist_lock and we just went through tasklist_lock to add
4960 * @child, it's guaranteed that either we see the set
4961 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4962 * @child during its iteration.
4964 * If we won the race, @child is associated with %current's
4965 * css_set. Grabbing css_set_rwsem guarantees both that the
4966 * association is stable, and, on completion of the parent's
4967 * migration, @child is visible in the source of migration or
4968 * already in the destination cgroup. This guarantee is necessary
4969 * when implementing operations which need to migrate all tasks of
4970 * a cgroup to another.
4972 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4973 * will remain in init_css_set. This is safe because all tasks are
4974 * in the init_css_set before cg_links is enabled and there's no
4975 * operation which transfers all tasks out of init_css_set.
4977 if (use_task_css_set_links) {
4978 struct css_set *cset;
4980 down_write(&css_set_rwsem);
4981 cset = task_css_set(current);
4982 if (list_empty(&child->cg_list)) {
4983 rcu_assign_pointer(child->cgroups, cset);
4984 list_add(&child->cg_list, &cset->tasks);
4987 up_write(&css_set_rwsem);
4991 * Call ss->fork(). This must happen after @child is linked on
4992 * css_set; otherwise, @child might change state between ->fork()
4993 * and addition to css_set.
4995 if (need_forkexit_callback) {
4996 for_each_subsys(ss, i)
5003 * cgroup_exit - detach cgroup from exiting task
5004 * @tsk: pointer to task_struct of exiting process
5006 * Description: Detach cgroup from @tsk and release it.
5008 * Note that cgroups marked notify_on_release force every task in
5009 * them to take the global cgroup_mutex mutex when exiting.
5010 * This could impact scaling on very large systems. Be reluctant to
5011 * use notify_on_release cgroups where very high task exit scaling
5012 * is required on large systems.
5014 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5015 * call cgroup_exit() while the task is still competent to handle
5016 * notify_on_release(), then leave the task attached to the root cgroup in
5017 * each hierarchy for the remainder of its exit. No need to bother with
5018 * init_css_set refcnting. init_css_set never goes away and we can't race
5019 * with migration path - PF_EXITING is visible to migration path.
5021 void cgroup_exit(struct task_struct *tsk)
5023 struct cgroup_subsys *ss;
5024 struct css_set *cset;
5025 bool put_cset = false;
5029 * Unlink from @tsk from its css_set. As migration path can't race
5030 * with us, we can check cg_list without grabbing css_set_rwsem.
5032 if (!list_empty(&tsk->cg_list)) {
5033 down_write(&css_set_rwsem);
5034 list_del_init(&tsk->cg_list);
5035 up_write(&css_set_rwsem);
5039 /* Reassign the task to the init_css_set. */
5040 cset = task_css_set(tsk);
5041 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5043 if (need_forkexit_callback) {
5044 /* see cgroup_post_fork() for details */
5045 for_each_subsys(ss, i) {
5047 struct cgroup_subsys_state *old_css = cset->subsys[i];
5048 struct cgroup_subsys_state *css = task_css(tsk, i);
5050 ss->exit(css, old_css, tsk);
5056 put_css_set(cset, true);
5059 static void check_for_release(struct cgroup *cgrp)
5061 if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
5062 !css_has_online_children(&cgrp->self)) {
5064 * Control Group is currently removeable. If it's not
5065 * already queued for a userspace notification, queue
5068 int need_schedule_work = 0;
5070 raw_spin_lock(&release_list_lock);
5071 if (!cgroup_is_dead(cgrp) &&
5072 list_empty(&cgrp->release_list)) {
5073 list_add(&cgrp->release_list, &release_list);
5074 need_schedule_work = 1;
5076 raw_spin_unlock(&release_list_lock);
5077 if (need_schedule_work)
5078 schedule_work(&release_agent_work);
5083 * Notify userspace when a cgroup is released, by running the
5084 * configured release agent with the name of the cgroup (path
5085 * relative to the root of cgroup file system) as the argument.
5087 * Most likely, this user command will try to rmdir this cgroup.
5089 * This races with the possibility that some other task will be
5090 * attached to this cgroup before it is removed, or that some other
5091 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5092 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5093 * unused, and this cgroup will be reprieved from its death sentence,
5094 * to continue to serve a useful existence. Next time it's released,
5095 * we will get notified again, if it still has 'notify_on_release' set.
5097 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5098 * means only wait until the task is successfully execve()'d. The
5099 * separate release agent task is forked by call_usermodehelper(),
5100 * then control in this thread returns here, without waiting for the
5101 * release agent task. We don't bother to wait because the caller of
5102 * this routine has no use for the exit status of the release agent
5103 * task, so no sense holding our caller up for that.
5105 static void cgroup_release_agent(struct work_struct *work)
5107 BUG_ON(work != &release_agent_work);
5108 mutex_lock(&cgroup_mutex);
5109 raw_spin_lock(&release_list_lock);
5110 while (!list_empty(&release_list)) {
5111 char *argv[3], *envp[3];
5113 char *pathbuf = NULL, *agentbuf = NULL, *path;
5114 struct cgroup *cgrp = list_entry(release_list.next,
5117 list_del_init(&cgrp->release_list);
5118 raw_spin_unlock(&release_list_lock);
5119 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5122 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5125 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5130 argv[i++] = agentbuf;
5135 /* minimal command environment */
5136 envp[i++] = "HOME=/";
5137 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5140 /* Drop the lock while we invoke the usermode helper,
5141 * since the exec could involve hitting disk and hence
5142 * be a slow process */
5143 mutex_unlock(&cgroup_mutex);
5144 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5145 mutex_lock(&cgroup_mutex);
5149 raw_spin_lock(&release_list_lock);
5151 raw_spin_unlock(&release_list_lock);
5152 mutex_unlock(&cgroup_mutex);
5155 static int __init cgroup_disable(char *str)
5157 struct cgroup_subsys *ss;
5161 while ((token = strsep(&str, ",")) != NULL) {
5165 for_each_subsys(ss, i) {
5166 if (!strcmp(token, ss->name)) {
5168 printk(KERN_INFO "Disabling %s control group"
5169 " subsystem\n", ss->name);
5176 __setup("cgroup_disable=", cgroup_disable);
5179 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5180 * @dentry: directory dentry of interest
5181 * @ss: subsystem of interest
5183 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5184 * to get the corresponding css and return it. If such css doesn't exist
5185 * or can't be pinned, an ERR_PTR value is returned.
5187 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5188 struct cgroup_subsys *ss)
5190 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5191 struct cgroup_subsys_state *css = NULL;
5192 struct cgroup *cgrp;
5194 /* is @dentry a cgroup dir? */
5195 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5196 kernfs_type(kn) != KERNFS_DIR)
5197 return ERR_PTR(-EBADF);
5202 * This path doesn't originate from kernfs and @kn could already
5203 * have been or be removed at any point. @kn->priv is RCU
5204 * protected for this access. See cgroup_rmdir() for details.
5206 cgrp = rcu_dereference(kn->priv);
5208 css = cgroup_css(cgrp, ss);
5210 if (!css || !css_tryget_online(css))
5211 css = ERR_PTR(-ENOENT);
5218 * css_from_id - lookup css by id
5219 * @id: the cgroup id
5220 * @ss: cgroup subsys to be looked into
5222 * Returns the css if there's valid one with @id, otherwise returns NULL.
5223 * Should be called under rcu_read_lock().
5225 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5227 WARN_ON_ONCE(!rcu_read_lock_held());
5228 return idr_find(&ss->css_idr, id);
5231 #ifdef CONFIG_CGROUP_DEBUG
5232 static struct cgroup_subsys_state *
5233 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5235 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5238 return ERR_PTR(-ENOMEM);
5243 static void debug_css_free(struct cgroup_subsys_state *css)
5248 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5251 return cgroup_task_count(css->cgroup);
5254 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5257 return (u64)(unsigned long)current->cgroups;
5260 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5266 count = atomic_read(&task_css_set(current)->refcount);
5271 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5273 struct cgrp_cset_link *link;
5274 struct css_set *cset;
5277 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5281 down_read(&css_set_rwsem);
5283 cset = rcu_dereference(current->cgroups);
5284 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5285 struct cgroup *c = link->cgrp;
5287 cgroup_name(c, name_buf, NAME_MAX + 1);
5288 seq_printf(seq, "Root %d group %s\n",
5289 c->root->hierarchy_id, name_buf);
5292 up_read(&css_set_rwsem);
5297 #define MAX_TASKS_SHOWN_PER_CSS 25
5298 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5300 struct cgroup_subsys_state *css = seq_css(seq);
5301 struct cgrp_cset_link *link;
5303 down_read(&css_set_rwsem);
5304 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5305 struct css_set *cset = link->cset;
5306 struct task_struct *task;
5309 seq_printf(seq, "css_set %p\n", cset);
5311 list_for_each_entry(task, &cset->tasks, cg_list) {
5312 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5314 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5317 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5318 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5320 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5324 seq_puts(seq, " ...\n");
5326 up_read(&css_set_rwsem);
5330 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5332 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5335 static struct cftype debug_files[] = {
5337 .name = "taskcount",
5338 .read_u64 = debug_taskcount_read,
5342 .name = "current_css_set",
5343 .read_u64 = current_css_set_read,
5347 .name = "current_css_set_refcount",
5348 .read_u64 = current_css_set_refcount_read,
5352 .name = "current_css_set_cg_links",
5353 .seq_show = current_css_set_cg_links_read,
5357 .name = "cgroup_css_links",
5358 .seq_show = cgroup_css_links_read,
5362 .name = "releasable",
5363 .read_u64 = releasable_read,
5369 struct cgroup_subsys debug_cgrp_subsys = {
5370 .css_alloc = debug_css_alloc,
5371 .css_free = debug_css_free,
5372 .base_cftypes = debug_files,
5374 #endif /* CONFIG_CGROUP_DEBUG */