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/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DECLARE_RWSEM(css_set_rwsem);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_rwsem);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DECLARE_RWSEM(css_set_rwsem);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 rcu_lockdep_assert(rcu_read_lock_held() || \
111 lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct *cgroup_destroy_wq;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
143 * The default hierarchy, reserved for the subsystems that are otherwise
144 * unattached - it never has more than a single cgroup, and all tasks are
145 * part of that cgroup.
147 struct cgroup_root cgrp_dfl_root;
150 * The default hierarchy always exists but is hidden until mounted for the
151 * first time. This is for backward compatibility.
153 static bool cgrp_dfl_root_visible;
156 * Set by the boot param of the same name and makes subsystems with NULL
157 * ->dfl_files to use ->legacy_files on the default hierarchy.
159 static bool cgroup_legacy_files_on_dfl;
161 /* some controllers are not supported in the default hierarchy */
162 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
164 /* The list of hierarchy roots */
166 static LIST_HEAD(cgroup_roots);
167 static int cgroup_root_count;
169 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
170 static DEFINE_IDR(cgroup_hierarchy_idr);
173 * Assign a monotonically increasing serial number to csses. It guarantees
174 * cgroups with bigger numbers are newer than those with smaller numbers.
175 * Also, as csses are always appended to the parent's ->children list, it
176 * guarantees that sibling csses are always sorted in the ascending serial
177 * number order on the list. Protected by cgroup_mutex.
179 static u64 css_serial_nr_next = 1;
181 /* This flag indicates whether tasks in the fork and exit paths should
182 * check for fork/exit handlers to call. This avoids us having to do
183 * extra work in the fork/exit path if none of the subsystems need to
186 static int need_forkexit_callback __read_mostly;
188 static struct cftype cgroup_dfl_base_files[];
189 static struct cftype cgroup_legacy_base_files[];
191 static int rebind_subsystems(struct cgroup_root *dst_root,
192 unsigned long ss_mask);
193 static int cgroup_destroy_locked(struct cgroup *cgrp);
194 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
196 static void css_release(struct percpu_ref *ref);
197 static void kill_css(struct cgroup_subsys_state *css);
198 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
201 /* IDR wrappers which synchronize using cgroup_idr_lock */
202 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
207 idr_preload(gfp_mask);
208 spin_lock_bh(&cgroup_idr_lock);
209 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
210 spin_unlock_bh(&cgroup_idr_lock);
215 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
219 spin_lock_bh(&cgroup_idr_lock);
220 ret = idr_replace(idr, ptr, id);
221 spin_unlock_bh(&cgroup_idr_lock);
225 static void cgroup_idr_remove(struct idr *idr, int id)
227 spin_lock_bh(&cgroup_idr_lock);
229 spin_unlock_bh(&cgroup_idr_lock);
232 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
234 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
237 return container_of(parent_css, struct cgroup, self);
242 * cgroup_css - obtain a cgroup's css for the specified subsystem
243 * @cgrp: the cgroup of interest
244 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
246 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
247 * function must be called either under cgroup_mutex or rcu_read_lock() and
248 * the caller is responsible for pinning the returned css if it wants to
249 * keep accessing it outside the said locks. This function may return
250 * %NULL if @cgrp doesn't have @subsys_id enabled.
252 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
253 struct cgroup_subsys *ss)
256 return rcu_dereference_check(cgrp->subsys[ss->id],
257 lockdep_is_held(&cgroup_mutex));
263 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
264 * @cgrp: the cgroup of interest
265 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
267 * Similar to cgroup_css() but returns the effective css, which is defined
268 * as the matching css of the nearest ancestor including self which has @ss
269 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
270 * function is guaranteed to return non-NULL css.
272 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
273 struct cgroup_subsys *ss)
275 lockdep_assert_held(&cgroup_mutex);
280 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
284 * This function is used while updating css associations and thus
285 * can't test the csses directly. Use ->child_subsys_mask.
287 while (cgroup_parent(cgrp) &&
288 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
289 cgrp = cgroup_parent(cgrp);
291 return cgroup_css(cgrp, ss);
295 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
296 * @cgrp: the cgroup of interest
297 * @ss: the subsystem of interest
299 * Find and get the effective css of @cgrp for @ss. The effective css is
300 * defined as the matching css of the nearest ancestor including self which
301 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
302 * the root css is returned, so this function always returns a valid css.
303 * The returned css must be put using css_put().
305 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
306 struct cgroup_subsys *ss)
308 struct cgroup_subsys_state *css;
313 css = cgroup_css(cgrp, ss);
315 if (css && css_tryget_online(css))
317 cgrp = cgroup_parent(cgrp);
320 css = init_css_set.subsys[ss->id];
327 /* convenient tests for these bits */
328 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
330 return !(cgrp->self.flags & CSS_ONLINE);
333 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
335 struct cgroup *cgrp = of->kn->parent->priv;
336 struct cftype *cft = of_cft(of);
339 * This is open and unprotected implementation of cgroup_css().
340 * seq_css() is only called from a kernfs file operation which has
341 * an active reference on the file. Because all the subsystem
342 * files are drained before a css is disassociated with a cgroup,
343 * the matching css from the cgroup's subsys table is guaranteed to
344 * be and stay valid until the enclosing operation is complete.
347 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
351 EXPORT_SYMBOL_GPL(of_css);
354 * cgroup_is_descendant - test ancestry
355 * @cgrp: the cgroup to be tested
356 * @ancestor: possible ancestor of @cgrp
358 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
359 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
360 * and @ancestor are accessible.
362 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
365 if (cgrp == ancestor)
367 cgrp = cgroup_parent(cgrp);
372 static int notify_on_release(const struct cgroup *cgrp)
374 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
378 * for_each_css - iterate all css's of a cgroup
379 * @css: the iteration cursor
380 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
381 * @cgrp: the target cgroup to iterate css's of
383 * Should be called under cgroup_[tree_]mutex.
385 #define for_each_css(css, ssid, cgrp) \
386 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
387 if (!((css) = rcu_dereference_check( \
388 (cgrp)->subsys[(ssid)], \
389 lockdep_is_held(&cgroup_mutex)))) { } \
393 * for_each_e_css - iterate all effective css's of a cgroup
394 * @css: the iteration cursor
395 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
396 * @cgrp: the target cgroup to iterate css's of
398 * Should be called under cgroup_[tree_]mutex.
400 #define for_each_e_css(css, ssid, cgrp) \
401 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
402 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
407 * for_each_subsys - iterate all enabled cgroup subsystems
408 * @ss: the iteration cursor
409 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
411 #define for_each_subsys(ss, ssid) \
412 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
413 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
415 /* iterate across the hierarchies */
416 #define for_each_root(root) \
417 list_for_each_entry((root), &cgroup_roots, root_list)
419 /* iterate over child cgrps, lock should be held throughout iteration */
420 #define cgroup_for_each_live_child(child, cgrp) \
421 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
422 if (({ lockdep_assert_held(&cgroup_mutex); \
423 cgroup_is_dead(child); })) \
427 static void cgroup_release_agent(struct work_struct *work);
428 static void check_for_release(struct cgroup *cgrp);
431 * A cgroup can be associated with multiple css_sets as different tasks may
432 * belong to different cgroups on different hierarchies. In the other
433 * direction, a css_set is naturally associated with multiple cgroups.
434 * This M:N relationship is represented by the following link structure
435 * which exists for each association and allows traversing the associations
438 struct cgrp_cset_link {
439 /* the cgroup and css_set this link associates */
441 struct css_set *cset;
443 /* list of cgrp_cset_links anchored at cgrp->cset_links */
444 struct list_head cset_link;
446 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
447 struct list_head cgrp_link;
451 * The default css_set - used by init and its children prior to any
452 * hierarchies being mounted. It contains a pointer to the root state
453 * for each subsystem. Also used to anchor the list of css_sets. Not
454 * reference-counted, to improve performance when child cgroups
455 * haven't been created.
457 struct css_set init_css_set = {
458 .refcount = ATOMIC_INIT(1),
459 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
460 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
461 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
462 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
463 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
466 static int css_set_count = 1; /* 1 for init_css_set */
469 * cgroup_update_populated - updated populated count of a cgroup
470 * @cgrp: the target cgroup
471 * @populated: inc or dec populated count
473 * @cgrp is either getting the first task (css_set) or losing the last.
474 * Update @cgrp->populated_cnt accordingly. The count is propagated
475 * towards root so that a given cgroup's populated_cnt is zero iff the
476 * cgroup and all its descendants are empty.
478 * @cgrp's interface file "cgroup.populated" is zero if
479 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
480 * changes from or to zero, userland is notified that the content of the
481 * interface file has changed. This can be used to detect when @cgrp and
482 * its descendants become populated or empty.
484 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
486 lockdep_assert_held(&css_set_rwsem);
492 trigger = !cgrp->populated_cnt++;
494 trigger = !--cgrp->populated_cnt;
499 if (cgrp->populated_kn)
500 kernfs_notify(cgrp->populated_kn);
501 cgrp = cgroup_parent(cgrp);
506 * hash table for cgroup groups. This improves the performance to find
507 * an existing css_set. This hash doesn't (currently) take into
508 * account cgroups in empty hierarchies.
510 #define CSS_SET_HASH_BITS 7
511 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
513 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
515 unsigned long key = 0UL;
516 struct cgroup_subsys *ss;
519 for_each_subsys(ss, i)
520 key += (unsigned long)css[i];
521 key = (key >> 16) ^ key;
526 static void put_css_set_locked(struct css_set *cset)
528 struct cgrp_cset_link *link, *tmp_link;
529 struct cgroup_subsys *ss;
532 lockdep_assert_held(&css_set_rwsem);
534 if (!atomic_dec_and_test(&cset->refcount))
537 /* This css_set is dead. unlink it and release cgroup refcounts */
538 for_each_subsys(ss, ssid)
539 list_del(&cset->e_cset_node[ssid]);
540 hash_del(&cset->hlist);
543 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
544 struct cgroup *cgrp = link->cgrp;
546 list_del(&link->cset_link);
547 list_del(&link->cgrp_link);
549 /* @cgrp can't go away while we're holding css_set_rwsem */
550 if (list_empty(&cgrp->cset_links)) {
551 cgroup_update_populated(cgrp, false);
552 check_for_release(cgrp);
558 kfree_rcu(cset, rcu_head);
561 static void put_css_set(struct css_set *cset)
564 * Ensure that the refcount doesn't hit zero while any readers
565 * can see it. Similar to atomic_dec_and_lock(), but for an
568 if (atomic_add_unless(&cset->refcount, -1, 1))
571 down_write(&css_set_rwsem);
572 put_css_set_locked(cset);
573 up_write(&css_set_rwsem);
577 * refcounted get/put for css_set objects
579 static inline void get_css_set(struct css_set *cset)
581 atomic_inc(&cset->refcount);
585 * compare_css_sets - helper function for find_existing_css_set().
586 * @cset: candidate css_set being tested
587 * @old_cset: existing css_set for a task
588 * @new_cgrp: cgroup that's being entered by the task
589 * @template: desired set of css pointers in css_set (pre-calculated)
591 * Returns true if "cset" matches "old_cset" except for the hierarchy
592 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
594 static bool compare_css_sets(struct css_set *cset,
595 struct css_set *old_cset,
596 struct cgroup *new_cgrp,
597 struct cgroup_subsys_state *template[])
599 struct list_head *l1, *l2;
602 * On the default hierarchy, there can be csets which are
603 * associated with the same set of cgroups but different csses.
604 * Let's first ensure that csses match.
606 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
610 * Compare cgroup pointers in order to distinguish between
611 * different cgroups in hierarchies. As different cgroups may
612 * share the same effective css, this comparison is always
615 l1 = &cset->cgrp_links;
616 l2 = &old_cset->cgrp_links;
618 struct cgrp_cset_link *link1, *link2;
619 struct cgroup *cgrp1, *cgrp2;
623 /* See if we reached the end - both lists are equal length. */
624 if (l1 == &cset->cgrp_links) {
625 BUG_ON(l2 != &old_cset->cgrp_links);
628 BUG_ON(l2 == &old_cset->cgrp_links);
630 /* Locate the cgroups associated with these links. */
631 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
632 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
635 /* Hierarchies should be linked in the same order. */
636 BUG_ON(cgrp1->root != cgrp2->root);
639 * If this hierarchy is the hierarchy of the cgroup
640 * that's changing, then we need to check that this
641 * css_set points to the new cgroup; if it's any other
642 * hierarchy, then this css_set should point to the
643 * same cgroup as the old css_set.
645 if (cgrp1->root == new_cgrp->root) {
646 if (cgrp1 != new_cgrp)
657 * find_existing_css_set - init css array and find the matching css_set
658 * @old_cset: the css_set that we're using before the cgroup transition
659 * @cgrp: the cgroup that we're moving into
660 * @template: out param for the new set of csses, should be clear on entry
662 static struct css_set *find_existing_css_set(struct css_set *old_cset,
664 struct cgroup_subsys_state *template[])
666 struct cgroup_root *root = cgrp->root;
667 struct cgroup_subsys *ss;
668 struct css_set *cset;
673 * Build the set of subsystem state objects that we want to see in the
674 * new css_set. while subsystems can change globally, the entries here
675 * won't change, so no need for locking.
677 for_each_subsys(ss, i) {
678 if (root->subsys_mask & (1UL << i)) {
680 * @ss is in this hierarchy, so we want the
681 * effective css from @cgrp.
683 template[i] = cgroup_e_css(cgrp, ss);
686 * @ss is not in this hierarchy, so we don't want
689 template[i] = old_cset->subsys[i];
693 key = css_set_hash(template);
694 hash_for_each_possible(css_set_table, cset, hlist, key) {
695 if (!compare_css_sets(cset, old_cset, cgrp, template))
698 /* This css_set matches what we need */
702 /* No existing cgroup group matched */
706 static void free_cgrp_cset_links(struct list_head *links_to_free)
708 struct cgrp_cset_link *link, *tmp_link;
710 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
711 list_del(&link->cset_link);
717 * allocate_cgrp_cset_links - allocate cgrp_cset_links
718 * @count: the number of links to allocate
719 * @tmp_links: list_head the allocated links are put on
721 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
722 * through ->cset_link. Returns 0 on success or -errno.
724 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
726 struct cgrp_cset_link *link;
729 INIT_LIST_HEAD(tmp_links);
731 for (i = 0; i < count; i++) {
732 link = kzalloc(sizeof(*link), GFP_KERNEL);
734 free_cgrp_cset_links(tmp_links);
737 list_add(&link->cset_link, tmp_links);
743 * link_css_set - a helper function to link a css_set to a cgroup
744 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
745 * @cset: the css_set to be linked
746 * @cgrp: the destination cgroup
748 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
751 struct cgrp_cset_link *link;
753 BUG_ON(list_empty(tmp_links));
755 if (cgroup_on_dfl(cgrp))
756 cset->dfl_cgrp = cgrp;
758 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
762 if (list_empty(&cgrp->cset_links))
763 cgroup_update_populated(cgrp, true);
764 list_move(&link->cset_link, &cgrp->cset_links);
767 * Always add links to the tail of the list so that the list
768 * is sorted by order of hierarchy creation
770 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
774 * find_css_set - return a new css_set with one cgroup updated
775 * @old_cset: the baseline css_set
776 * @cgrp: the cgroup to be updated
778 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
779 * substituted into the appropriate hierarchy.
781 static struct css_set *find_css_set(struct css_set *old_cset,
784 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
785 struct css_set *cset;
786 struct list_head tmp_links;
787 struct cgrp_cset_link *link;
788 struct cgroup_subsys *ss;
792 lockdep_assert_held(&cgroup_mutex);
794 /* First see if we already have a cgroup group that matches
796 down_read(&css_set_rwsem);
797 cset = find_existing_css_set(old_cset, cgrp, template);
800 up_read(&css_set_rwsem);
805 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
809 /* Allocate all the cgrp_cset_link objects that we'll need */
810 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
815 atomic_set(&cset->refcount, 1);
816 INIT_LIST_HEAD(&cset->cgrp_links);
817 INIT_LIST_HEAD(&cset->tasks);
818 INIT_LIST_HEAD(&cset->mg_tasks);
819 INIT_LIST_HEAD(&cset->mg_preload_node);
820 INIT_LIST_HEAD(&cset->mg_node);
821 INIT_HLIST_NODE(&cset->hlist);
823 /* Copy the set of subsystem state objects generated in
824 * find_existing_css_set() */
825 memcpy(cset->subsys, template, sizeof(cset->subsys));
827 down_write(&css_set_rwsem);
828 /* Add reference counts and links from the new css_set. */
829 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
830 struct cgroup *c = link->cgrp;
832 if (c->root == cgrp->root)
834 link_css_set(&tmp_links, cset, c);
837 BUG_ON(!list_empty(&tmp_links));
841 /* Add @cset to the hash table */
842 key = css_set_hash(cset->subsys);
843 hash_add(css_set_table, &cset->hlist, key);
845 for_each_subsys(ss, ssid)
846 list_add_tail(&cset->e_cset_node[ssid],
847 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
849 up_write(&css_set_rwsem);
854 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
856 struct cgroup *root_cgrp = kf_root->kn->priv;
858 return root_cgrp->root;
861 static int cgroup_init_root_id(struct cgroup_root *root)
865 lockdep_assert_held(&cgroup_mutex);
867 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
871 root->hierarchy_id = id;
875 static void cgroup_exit_root_id(struct cgroup_root *root)
877 lockdep_assert_held(&cgroup_mutex);
879 if (root->hierarchy_id) {
880 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
881 root->hierarchy_id = 0;
885 static void cgroup_free_root(struct cgroup_root *root)
888 /* hierarchy ID should already have been released */
889 WARN_ON_ONCE(root->hierarchy_id);
891 idr_destroy(&root->cgroup_idr);
896 static void cgroup_destroy_root(struct cgroup_root *root)
898 struct cgroup *cgrp = &root->cgrp;
899 struct cgrp_cset_link *link, *tmp_link;
901 mutex_lock(&cgroup_mutex);
903 BUG_ON(atomic_read(&root->nr_cgrps));
904 BUG_ON(!list_empty(&cgrp->self.children));
906 /* Rebind all subsystems back to the default hierarchy */
907 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
910 * Release all the links from cset_links to this hierarchy's
913 down_write(&css_set_rwsem);
915 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
916 list_del(&link->cset_link);
917 list_del(&link->cgrp_link);
920 up_write(&css_set_rwsem);
922 if (!list_empty(&root->root_list)) {
923 list_del(&root->root_list);
927 cgroup_exit_root_id(root);
929 mutex_unlock(&cgroup_mutex);
931 kernfs_destroy_root(root->kf_root);
932 cgroup_free_root(root);
935 /* look up cgroup associated with given css_set on the specified hierarchy */
936 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
937 struct cgroup_root *root)
939 struct cgroup *res = NULL;
941 lockdep_assert_held(&cgroup_mutex);
942 lockdep_assert_held(&css_set_rwsem);
944 if (cset == &init_css_set) {
947 struct cgrp_cset_link *link;
949 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
950 struct cgroup *c = link->cgrp;
952 if (c->root == root) {
964 * Return the cgroup for "task" from the given hierarchy. Must be
965 * called with cgroup_mutex and css_set_rwsem held.
967 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
968 struct cgroup_root *root)
971 * No need to lock the task - since we hold cgroup_mutex the
972 * task can't change groups, so the only thing that can happen
973 * is that it exits and its css is set back to init_css_set.
975 return cset_cgroup_from_root(task_css_set(task), root);
979 * A task must hold cgroup_mutex to modify cgroups.
981 * Any task can increment and decrement the count field without lock.
982 * So in general, code holding cgroup_mutex can't rely on the count
983 * field not changing. However, if the count goes to zero, then only
984 * cgroup_attach_task() can increment it again. Because a count of zero
985 * means that no tasks are currently attached, therefore there is no
986 * way a task attached to that cgroup can fork (the other way to
987 * increment the count). So code holding cgroup_mutex can safely
988 * assume that if the count is zero, it will stay zero. Similarly, if
989 * a task holds cgroup_mutex on a cgroup with zero count, it
990 * knows that the cgroup won't be removed, as cgroup_rmdir()
993 * A cgroup can only be deleted if both its 'count' of using tasks
994 * is zero, and its list of 'children' cgroups is empty. Since all
995 * tasks in the system use _some_ cgroup, and since there is always at
996 * least one task in the system (init, pid == 1), therefore, root cgroup
997 * always has either children cgroups and/or using tasks. So we don't
998 * need a special hack to ensure that root cgroup cannot be deleted.
1000 * P.S. One more locking exception. RCU is used to guard the
1001 * update of a tasks cgroup pointer by cgroup_attach_task()
1004 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
1005 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1006 static const struct file_operations proc_cgroupstats_operations;
1008 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1011 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1012 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1013 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1014 cft->ss->name, cft->name);
1016 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1021 * cgroup_file_mode - deduce file mode of a control file
1022 * @cft: the control file in question
1024 * returns cft->mode if ->mode is not 0
1025 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1026 * returns S_IRUGO if it has only a read handler
1027 * returns S_IWUSR if it has only a write hander
1029 static umode_t cgroup_file_mode(const struct cftype *cft)
1036 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1039 if (cft->write_u64 || cft->write_s64 || cft->write)
1045 static void cgroup_get(struct cgroup *cgrp)
1047 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1048 css_get(&cgrp->self);
1051 static bool cgroup_tryget(struct cgroup *cgrp)
1053 return css_tryget(&cgrp->self);
1056 static void cgroup_put(struct cgroup *cgrp)
1058 css_put(&cgrp->self);
1062 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1063 * @cgrp: the target cgroup
1064 * @subtree_control: the new subtree_control mask to consider
1066 * On the default hierarchy, a subsystem may request other subsystems to be
1067 * enabled together through its ->depends_on mask. In such cases, more
1068 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1070 * This function calculates which subsystems need to be enabled if
1071 * @subtree_control is to be applied to @cgrp. The returned mask is always
1072 * a superset of @subtree_control and follows the usual hierarchy rules.
1074 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1075 unsigned long subtree_control)
1077 struct cgroup *parent = cgroup_parent(cgrp);
1078 unsigned long cur_ss_mask = subtree_control;
1079 struct cgroup_subsys *ss;
1082 lockdep_assert_held(&cgroup_mutex);
1084 if (!cgroup_on_dfl(cgrp))
1088 unsigned long new_ss_mask = cur_ss_mask;
1090 for_each_subsys(ss, ssid)
1091 if (cur_ss_mask & (1 << ssid))
1092 new_ss_mask |= ss->depends_on;
1095 * Mask out subsystems which aren't available. This can
1096 * happen only if some depended-upon subsystems were bound
1097 * to non-default hierarchies.
1100 new_ss_mask &= parent->child_subsys_mask;
1102 new_ss_mask &= cgrp->root->subsys_mask;
1104 if (new_ss_mask == cur_ss_mask)
1106 cur_ss_mask = new_ss_mask;
1113 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1114 * @cgrp: the target cgroup
1116 * Update @cgrp->child_subsys_mask according to the current
1117 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1119 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1121 cgrp->child_subsys_mask =
1122 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1126 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1127 * @kn: the kernfs_node being serviced
1129 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1130 * the method finishes if locking succeeded. Note that once this function
1131 * returns the cgroup returned by cgroup_kn_lock_live() may become
1132 * inaccessible any time. If the caller intends to continue to access the
1133 * cgroup, it should pin it before invoking this function.
1135 static void cgroup_kn_unlock(struct kernfs_node *kn)
1137 struct cgroup *cgrp;
1139 if (kernfs_type(kn) == KERNFS_DIR)
1142 cgrp = kn->parent->priv;
1144 mutex_unlock(&cgroup_mutex);
1146 kernfs_unbreak_active_protection(kn);
1151 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1152 * @kn: the kernfs_node being serviced
1154 * This helper is to be used by a cgroup kernfs method currently servicing
1155 * @kn. It breaks the active protection, performs cgroup locking and
1156 * verifies that the associated cgroup is alive. Returns the cgroup if
1157 * alive; otherwise, %NULL. A successful return should be undone by a
1158 * matching cgroup_kn_unlock() invocation.
1160 * Any cgroup kernfs method implementation which requires locking the
1161 * associated cgroup should use this helper. It avoids nesting cgroup
1162 * locking under kernfs active protection and allows all kernfs operations
1163 * including self-removal.
1165 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1167 struct cgroup *cgrp;
1169 if (kernfs_type(kn) == KERNFS_DIR)
1172 cgrp = kn->parent->priv;
1175 * We're gonna grab cgroup_mutex which nests outside kernfs
1176 * active_ref. cgroup liveliness check alone provides enough
1177 * protection against removal. Ensure @cgrp stays accessible and
1178 * break the active_ref protection.
1180 if (!cgroup_tryget(cgrp))
1182 kernfs_break_active_protection(kn);
1184 mutex_lock(&cgroup_mutex);
1186 if (!cgroup_is_dead(cgrp))
1189 cgroup_kn_unlock(kn);
1193 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1195 char name[CGROUP_FILE_NAME_MAX];
1197 lockdep_assert_held(&cgroup_mutex);
1198 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1202 * cgroup_clear_dir - remove subsys files in a cgroup directory
1203 * @cgrp: target cgroup
1204 * @subsys_mask: mask of the subsystem ids whose files should be removed
1206 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1208 struct cgroup_subsys *ss;
1211 for_each_subsys(ss, i) {
1212 struct cftype *cfts;
1214 if (!(subsys_mask & (1 << i)))
1216 list_for_each_entry(cfts, &ss->cfts, node)
1217 cgroup_addrm_files(cgrp, cfts, false);
1221 static int rebind_subsystems(struct cgroup_root *dst_root,
1222 unsigned long ss_mask)
1224 struct cgroup_subsys *ss;
1225 unsigned long tmp_ss_mask;
1228 lockdep_assert_held(&cgroup_mutex);
1230 for_each_subsys(ss, ssid) {
1231 if (!(ss_mask & (1 << ssid)))
1234 /* if @ss has non-root csses attached to it, can't move */
1235 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1238 /* can't move between two non-dummy roots either */
1239 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1243 /* skip creating root files on dfl_root for inhibited subsystems */
1244 tmp_ss_mask = ss_mask;
1245 if (dst_root == &cgrp_dfl_root)
1246 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1248 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1250 if (dst_root != &cgrp_dfl_root)
1254 * Rebinding back to the default root is not allowed to
1255 * fail. Using both default and non-default roots should
1256 * be rare. Moving subsystems back and forth even more so.
1257 * Just warn about it and continue.
1259 if (cgrp_dfl_root_visible) {
1260 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1262 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1267 * Nothing can fail from this point on. Remove files for the
1268 * removed subsystems and rebind each subsystem.
1270 for_each_subsys(ss, ssid)
1271 if (ss_mask & (1 << ssid))
1272 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1274 for_each_subsys(ss, ssid) {
1275 struct cgroup_root *src_root;
1276 struct cgroup_subsys_state *css;
1277 struct css_set *cset;
1279 if (!(ss_mask & (1 << ssid)))
1282 src_root = ss->root;
1283 css = cgroup_css(&src_root->cgrp, ss);
1285 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1287 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1288 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1289 ss->root = dst_root;
1290 css->cgroup = &dst_root->cgrp;
1292 down_write(&css_set_rwsem);
1293 hash_for_each(css_set_table, i, cset, hlist)
1294 list_move_tail(&cset->e_cset_node[ss->id],
1295 &dst_root->cgrp.e_csets[ss->id]);
1296 up_write(&css_set_rwsem);
1298 src_root->subsys_mask &= ~(1 << ssid);
1299 src_root->cgrp.subtree_control &= ~(1 << ssid);
1300 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1302 /* default hierarchy doesn't enable controllers by default */
1303 dst_root->subsys_mask |= 1 << ssid;
1304 if (dst_root != &cgrp_dfl_root) {
1305 dst_root->cgrp.subtree_control |= 1 << ssid;
1306 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1313 kernfs_activate(dst_root->cgrp.kn);
1317 static int cgroup_show_options(struct seq_file *seq,
1318 struct kernfs_root *kf_root)
1320 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1321 struct cgroup_subsys *ss;
1324 for_each_subsys(ss, ssid)
1325 if (root->subsys_mask & (1 << ssid))
1326 seq_printf(seq, ",%s", ss->name);
1327 if (root->flags & CGRP_ROOT_NOPREFIX)
1328 seq_puts(seq, ",noprefix");
1329 if (root->flags & CGRP_ROOT_XATTR)
1330 seq_puts(seq, ",xattr");
1332 spin_lock(&release_agent_path_lock);
1333 if (strlen(root->release_agent_path))
1334 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1335 spin_unlock(&release_agent_path_lock);
1337 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1338 seq_puts(seq, ",clone_children");
1339 if (strlen(root->name))
1340 seq_printf(seq, ",name=%s", root->name);
1344 struct cgroup_sb_opts {
1345 unsigned long subsys_mask;
1347 char *release_agent;
1348 bool cpuset_clone_children;
1350 /* User explicitly requested empty subsystem */
1354 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1356 char *token, *o = data;
1357 bool all_ss = false, one_ss = false;
1358 unsigned long mask = -1UL;
1359 struct cgroup_subsys *ss;
1363 #ifdef CONFIG_CPUSETS
1364 mask = ~(1U << cpuset_cgrp_id);
1367 memset(opts, 0, sizeof(*opts));
1369 while ((token = strsep(&o, ",")) != NULL) {
1374 if (!strcmp(token, "none")) {
1375 /* Explicitly have no subsystems */
1379 if (!strcmp(token, "all")) {
1380 /* Mutually exclusive option 'all' + subsystem name */
1386 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1387 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1390 if (!strcmp(token, "noprefix")) {
1391 opts->flags |= CGRP_ROOT_NOPREFIX;
1394 if (!strcmp(token, "clone_children")) {
1395 opts->cpuset_clone_children = true;
1398 if (!strcmp(token, "xattr")) {
1399 opts->flags |= CGRP_ROOT_XATTR;
1402 if (!strncmp(token, "release_agent=", 14)) {
1403 /* Specifying two release agents is forbidden */
1404 if (opts->release_agent)
1406 opts->release_agent =
1407 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1408 if (!opts->release_agent)
1412 if (!strncmp(token, "name=", 5)) {
1413 const char *name = token + 5;
1414 /* Can't specify an empty name */
1417 /* Must match [\w.-]+ */
1418 for (i = 0; i < strlen(name); i++) {
1422 if ((c == '.') || (c == '-') || (c == '_'))
1426 /* Specifying two names is forbidden */
1429 opts->name = kstrndup(name,
1430 MAX_CGROUP_ROOT_NAMELEN - 1,
1438 for_each_subsys(ss, i) {
1439 if (strcmp(token, ss->name))
1444 /* Mutually exclusive option 'all' + subsystem name */
1447 opts->subsys_mask |= (1 << i);
1452 if (i == CGROUP_SUBSYS_COUNT)
1456 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1457 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1459 pr_err("sane_behavior: no other mount options allowed\n");
1466 * If the 'all' option was specified select all the subsystems,
1467 * otherwise if 'none', 'name=' and a subsystem name options were
1468 * not specified, let's default to 'all'
1470 if (all_ss || (!one_ss && !opts->none && !opts->name))
1471 for_each_subsys(ss, i)
1473 opts->subsys_mask |= (1 << i);
1476 * We either have to specify by name or by subsystems. (So all
1477 * empty hierarchies must have a name).
1479 if (!opts->subsys_mask && !opts->name)
1483 * Option noprefix was introduced just for backward compatibility
1484 * with the old cpuset, so we allow noprefix only if mounting just
1485 * the cpuset subsystem.
1487 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1490 /* Can't specify "none" and some subsystems */
1491 if (opts->subsys_mask && opts->none)
1497 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1500 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1501 struct cgroup_sb_opts opts;
1502 unsigned long added_mask, removed_mask;
1504 if (root == &cgrp_dfl_root) {
1505 pr_err("remount is not allowed\n");
1509 mutex_lock(&cgroup_mutex);
1511 /* See what subsystems are wanted */
1512 ret = parse_cgroupfs_options(data, &opts);
1516 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1517 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1518 task_tgid_nr(current), current->comm);
1520 added_mask = opts.subsys_mask & ~root->subsys_mask;
1521 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1523 /* Don't allow flags or name to change at remount */
1524 if ((opts.flags ^ root->flags) ||
1525 (opts.name && strcmp(opts.name, root->name))) {
1526 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1527 opts.flags, opts.name ?: "", root->flags, root->name);
1532 /* remounting is not allowed for populated hierarchies */
1533 if (!list_empty(&root->cgrp.self.children)) {
1538 ret = rebind_subsystems(root, added_mask);
1542 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1544 if (opts.release_agent) {
1545 spin_lock(&release_agent_path_lock);
1546 strcpy(root->release_agent_path, opts.release_agent);
1547 spin_unlock(&release_agent_path_lock);
1550 kfree(opts.release_agent);
1552 mutex_unlock(&cgroup_mutex);
1557 * To reduce the fork() overhead for systems that are not actually using
1558 * their cgroups capability, we don't maintain the lists running through
1559 * each css_set to its tasks until we see the list actually used - in other
1560 * words after the first mount.
1562 static bool use_task_css_set_links __read_mostly;
1564 static void cgroup_enable_task_cg_lists(void)
1566 struct task_struct *p, *g;
1568 down_write(&css_set_rwsem);
1570 if (use_task_css_set_links)
1573 use_task_css_set_links = true;
1576 * We need tasklist_lock because RCU is not safe against
1577 * while_each_thread(). Besides, a forking task that has passed
1578 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1579 * is not guaranteed to have its child immediately visible in the
1580 * tasklist if we walk through it with RCU.
1582 read_lock(&tasklist_lock);
1583 do_each_thread(g, p) {
1584 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1585 task_css_set(p) != &init_css_set);
1588 * We should check if the process is exiting, otherwise
1589 * it will race with cgroup_exit() in that the list
1590 * entry won't be deleted though the process has exited.
1591 * Do it while holding siglock so that we don't end up
1592 * racing against cgroup_exit().
1594 spin_lock_irq(&p->sighand->siglock);
1595 if (!(p->flags & PF_EXITING)) {
1596 struct css_set *cset = task_css_set(p);
1598 list_add(&p->cg_list, &cset->tasks);
1601 spin_unlock_irq(&p->sighand->siglock);
1602 } while_each_thread(g, p);
1603 read_unlock(&tasklist_lock);
1605 up_write(&css_set_rwsem);
1608 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1610 struct cgroup_subsys *ss;
1613 INIT_LIST_HEAD(&cgrp->self.sibling);
1614 INIT_LIST_HEAD(&cgrp->self.children);
1615 INIT_LIST_HEAD(&cgrp->cset_links);
1616 INIT_LIST_HEAD(&cgrp->pidlists);
1617 mutex_init(&cgrp->pidlist_mutex);
1618 cgrp->self.cgroup = cgrp;
1619 cgrp->self.flags |= CSS_ONLINE;
1621 for_each_subsys(ss, ssid)
1622 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1624 init_waitqueue_head(&cgrp->offline_waitq);
1625 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1628 static void init_cgroup_root(struct cgroup_root *root,
1629 struct cgroup_sb_opts *opts)
1631 struct cgroup *cgrp = &root->cgrp;
1633 INIT_LIST_HEAD(&root->root_list);
1634 atomic_set(&root->nr_cgrps, 1);
1636 init_cgroup_housekeeping(cgrp);
1637 idr_init(&root->cgroup_idr);
1639 root->flags = opts->flags;
1640 if (opts->release_agent)
1641 strcpy(root->release_agent_path, opts->release_agent);
1643 strcpy(root->name, opts->name);
1644 if (opts->cpuset_clone_children)
1645 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1648 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1650 LIST_HEAD(tmp_links);
1651 struct cgroup *root_cgrp = &root->cgrp;
1652 struct cftype *base_files;
1653 struct css_set *cset;
1656 lockdep_assert_held(&cgroup_mutex);
1658 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1661 root_cgrp->id = ret;
1663 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1669 * We're accessing css_set_count without locking css_set_rwsem here,
1670 * but that's OK - it can only be increased by someone holding
1671 * cgroup_lock, and that's us. The worst that can happen is that we
1672 * have some link structures left over
1674 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1678 ret = cgroup_init_root_id(root);
1682 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1683 KERNFS_ROOT_CREATE_DEACTIVATED,
1685 if (IS_ERR(root->kf_root)) {
1686 ret = PTR_ERR(root->kf_root);
1689 root_cgrp->kn = root->kf_root->kn;
1691 if (root == &cgrp_dfl_root)
1692 base_files = cgroup_dfl_base_files;
1694 base_files = cgroup_legacy_base_files;
1696 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1700 ret = rebind_subsystems(root, ss_mask);
1705 * There must be no failure case after here, since rebinding takes
1706 * care of subsystems' refcounts, which are explicitly dropped in
1707 * the failure exit path.
1709 list_add(&root->root_list, &cgroup_roots);
1710 cgroup_root_count++;
1713 * Link the root cgroup in this hierarchy into all the css_set
1716 down_write(&css_set_rwsem);
1717 hash_for_each(css_set_table, i, cset, hlist)
1718 link_css_set(&tmp_links, cset, root_cgrp);
1719 up_write(&css_set_rwsem);
1721 BUG_ON(!list_empty(&root_cgrp->self.children));
1722 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1724 kernfs_activate(root_cgrp->kn);
1729 kernfs_destroy_root(root->kf_root);
1730 root->kf_root = NULL;
1732 cgroup_exit_root_id(root);
1734 percpu_ref_exit(&root_cgrp->self.refcnt);
1736 free_cgrp_cset_links(&tmp_links);
1740 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1741 int flags, const char *unused_dev_name,
1744 struct super_block *pinned_sb = NULL;
1745 struct cgroup_subsys *ss;
1746 struct cgroup_root *root;
1747 struct cgroup_sb_opts opts;
1748 struct dentry *dentry;
1754 * The first time anyone tries to mount a cgroup, enable the list
1755 * linking each css_set to its tasks and fix up all existing tasks.
1757 if (!use_task_css_set_links)
1758 cgroup_enable_task_cg_lists();
1760 mutex_lock(&cgroup_mutex);
1762 /* First find the desired set of subsystems */
1763 ret = parse_cgroupfs_options(data, &opts);
1767 /* look for a matching existing root */
1768 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1769 cgrp_dfl_root_visible = true;
1770 root = &cgrp_dfl_root;
1771 cgroup_get(&root->cgrp);
1777 * Destruction of cgroup root is asynchronous, so subsystems may
1778 * still be dying after the previous unmount. Let's drain the
1779 * dying subsystems. We just need to ensure that the ones
1780 * unmounted previously finish dying and don't care about new ones
1781 * starting. Testing ref liveliness is good enough.
1783 for_each_subsys(ss, i) {
1784 if (!(opts.subsys_mask & (1 << i)) ||
1785 ss->root == &cgrp_dfl_root)
1788 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1789 mutex_unlock(&cgroup_mutex);
1791 ret = restart_syscall();
1794 cgroup_put(&ss->root->cgrp);
1797 for_each_root(root) {
1798 bool name_match = false;
1800 if (root == &cgrp_dfl_root)
1804 * If we asked for a name then it must match. Also, if
1805 * name matches but sybsys_mask doesn't, we should fail.
1806 * Remember whether name matched.
1809 if (strcmp(opts.name, root->name))
1815 * If we asked for subsystems (or explicitly for no
1816 * subsystems) then they must match.
1818 if ((opts.subsys_mask || opts.none) &&
1819 (opts.subsys_mask != root->subsys_mask)) {
1826 if (root->flags ^ opts.flags)
1827 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1830 * We want to reuse @root whose lifetime is governed by its
1831 * ->cgrp. Let's check whether @root is alive and keep it
1832 * that way. As cgroup_kill_sb() can happen anytime, we
1833 * want to block it by pinning the sb so that @root doesn't
1834 * get killed before mount is complete.
1836 * With the sb pinned, tryget_live can reliably indicate
1837 * whether @root can be reused. If it's being killed,
1838 * drain it. We can use wait_queue for the wait but this
1839 * path is super cold. Let's just sleep a bit and retry.
1841 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1842 if (IS_ERR(pinned_sb) ||
1843 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1844 mutex_unlock(&cgroup_mutex);
1845 if (!IS_ERR_OR_NULL(pinned_sb))
1846 deactivate_super(pinned_sb);
1848 ret = restart_syscall();
1857 * No such thing, create a new one. name= matching without subsys
1858 * specification is allowed for already existing hierarchies but we
1859 * can't create new one without subsys specification.
1861 if (!opts.subsys_mask && !opts.none) {
1866 root = kzalloc(sizeof(*root), GFP_KERNEL);
1872 init_cgroup_root(root, &opts);
1874 ret = cgroup_setup_root(root, opts.subsys_mask);
1876 cgroup_free_root(root);
1879 mutex_unlock(&cgroup_mutex);
1881 kfree(opts.release_agent);
1885 return ERR_PTR(ret);
1887 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1888 CGROUP_SUPER_MAGIC, &new_sb);
1889 if (IS_ERR(dentry) || !new_sb)
1890 cgroup_put(&root->cgrp);
1893 * If @pinned_sb, we're reusing an existing root and holding an
1894 * extra ref on its sb. Mount is complete. Put the extra ref.
1898 deactivate_super(pinned_sb);
1904 static void cgroup_kill_sb(struct super_block *sb)
1906 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1907 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1910 * If @root doesn't have any mounts or children, start killing it.
1911 * This prevents new mounts by disabling percpu_ref_tryget_live().
1912 * cgroup_mount() may wait for @root's release.
1914 * And don't kill the default root.
1916 if (!list_empty(&root->cgrp.self.children) ||
1917 root == &cgrp_dfl_root)
1918 cgroup_put(&root->cgrp);
1920 percpu_ref_kill(&root->cgrp.self.refcnt);
1925 static struct file_system_type cgroup_fs_type = {
1927 .mount = cgroup_mount,
1928 .kill_sb = cgroup_kill_sb,
1931 static struct kobject *cgroup_kobj;
1934 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1935 * @task: target task
1936 * @buf: the buffer to write the path into
1937 * @buflen: the length of the buffer
1939 * Determine @task's cgroup on the first (the one with the lowest non-zero
1940 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1941 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1942 * cgroup controller callbacks.
1944 * Return value is the same as kernfs_path().
1946 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1948 struct cgroup_root *root;
1949 struct cgroup *cgrp;
1950 int hierarchy_id = 1;
1953 mutex_lock(&cgroup_mutex);
1954 down_read(&css_set_rwsem);
1956 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1959 cgrp = task_cgroup_from_root(task, root);
1960 path = cgroup_path(cgrp, buf, buflen);
1962 /* if no hierarchy exists, everyone is in "/" */
1963 if (strlcpy(buf, "/", buflen) < buflen)
1967 up_read(&css_set_rwsem);
1968 mutex_unlock(&cgroup_mutex);
1971 EXPORT_SYMBOL_GPL(task_cgroup_path);
1973 /* used to track tasks and other necessary states during migration */
1974 struct cgroup_taskset {
1975 /* the src and dst cset list running through cset->mg_node */
1976 struct list_head src_csets;
1977 struct list_head dst_csets;
1980 * Fields for cgroup_taskset_*() iteration.
1982 * Before migration is committed, the target migration tasks are on
1983 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1984 * the csets on ->dst_csets. ->csets point to either ->src_csets
1985 * or ->dst_csets depending on whether migration is committed.
1987 * ->cur_csets and ->cur_task point to the current task position
1990 struct list_head *csets;
1991 struct css_set *cur_cset;
1992 struct task_struct *cur_task;
1996 * cgroup_taskset_first - reset taskset and return the first task
1997 * @tset: taskset of interest
1999 * @tset iteration is initialized and the first task is returned.
2001 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2003 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2004 tset->cur_task = NULL;
2006 return cgroup_taskset_next(tset);
2010 * cgroup_taskset_next - iterate to the next task in taskset
2011 * @tset: taskset of interest
2013 * Return the next task in @tset. Iteration must have been initialized
2014 * with cgroup_taskset_first().
2016 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2018 struct css_set *cset = tset->cur_cset;
2019 struct task_struct *task = tset->cur_task;
2021 while (&cset->mg_node != tset->csets) {
2023 task = list_first_entry(&cset->mg_tasks,
2024 struct task_struct, cg_list);
2026 task = list_next_entry(task, cg_list);
2028 if (&task->cg_list != &cset->mg_tasks) {
2029 tset->cur_cset = cset;
2030 tset->cur_task = task;
2034 cset = list_next_entry(cset, mg_node);
2042 * cgroup_task_migrate - move a task from one cgroup to another.
2043 * @old_cgrp: the cgroup @tsk is being migrated from
2044 * @tsk: the task being migrated
2045 * @new_cset: the new css_set @tsk is being attached to
2047 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2049 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2050 struct task_struct *tsk,
2051 struct css_set *new_cset)
2053 struct css_set *old_cset;
2055 lockdep_assert_held(&cgroup_mutex);
2056 lockdep_assert_held(&css_set_rwsem);
2059 * We are synchronized through cgroup_threadgroup_rwsem against
2060 * PF_EXITING setting such that we can't race against cgroup_exit()
2061 * changing the css_set to init_css_set and dropping the old one.
2063 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2064 old_cset = task_css_set(tsk);
2066 get_css_set(new_cset);
2067 rcu_assign_pointer(tsk->cgroups, new_cset);
2070 * Use move_tail so that cgroup_taskset_first() still returns the
2071 * leader after migration. This works because cgroup_migrate()
2072 * ensures that the dst_cset of the leader is the first on the
2073 * tset's dst_csets list.
2075 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2078 * We just gained a reference on old_cset by taking it from the
2079 * task. As trading it for new_cset is protected by cgroup_mutex,
2080 * we're safe to drop it here; it will be freed under RCU.
2082 put_css_set_locked(old_cset);
2086 * cgroup_migrate_finish - cleanup after attach
2087 * @preloaded_csets: list of preloaded css_sets
2089 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2090 * those functions for details.
2092 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2094 struct css_set *cset, *tmp_cset;
2096 lockdep_assert_held(&cgroup_mutex);
2098 down_write(&css_set_rwsem);
2099 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2100 cset->mg_src_cgrp = NULL;
2101 cset->mg_dst_cset = NULL;
2102 list_del_init(&cset->mg_preload_node);
2103 put_css_set_locked(cset);
2105 up_write(&css_set_rwsem);
2109 * cgroup_migrate_add_src - add a migration source css_set
2110 * @src_cset: the source css_set to add
2111 * @dst_cgrp: the destination cgroup
2112 * @preloaded_csets: list of preloaded css_sets
2114 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2115 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2116 * up by cgroup_migrate_finish().
2118 * This function may be called without holding cgroup_threadgroup_rwsem
2119 * even if the target is a process. Threads may be created and destroyed
2120 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2121 * into play and the preloaded css_sets are guaranteed to cover all
2124 static void cgroup_migrate_add_src(struct css_set *src_cset,
2125 struct cgroup *dst_cgrp,
2126 struct list_head *preloaded_csets)
2128 struct cgroup *src_cgrp;
2130 lockdep_assert_held(&cgroup_mutex);
2131 lockdep_assert_held(&css_set_rwsem);
2133 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2135 if (!list_empty(&src_cset->mg_preload_node))
2138 WARN_ON(src_cset->mg_src_cgrp);
2139 WARN_ON(!list_empty(&src_cset->mg_tasks));
2140 WARN_ON(!list_empty(&src_cset->mg_node));
2142 src_cset->mg_src_cgrp = src_cgrp;
2143 get_css_set(src_cset);
2144 list_add(&src_cset->mg_preload_node, preloaded_csets);
2148 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2149 * @dst_cgrp: the destination cgroup (may be %NULL)
2150 * @preloaded_csets: list of preloaded source css_sets
2152 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2153 * have been preloaded to @preloaded_csets. This function looks up and
2154 * pins all destination css_sets, links each to its source, and append them
2155 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2156 * source css_set is assumed to be its cgroup on the default hierarchy.
2158 * This function must be called after cgroup_migrate_add_src() has been
2159 * called on each migration source css_set. After migration is performed
2160 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2163 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2164 struct list_head *preloaded_csets)
2167 struct css_set *src_cset, *tmp_cset;
2169 lockdep_assert_held(&cgroup_mutex);
2172 * Except for the root, child_subsys_mask must be zero for a cgroup
2173 * with tasks so that child cgroups don't compete against tasks.
2175 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2176 dst_cgrp->child_subsys_mask)
2179 /* look up the dst cset for each src cset and link it to src */
2180 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2181 struct css_set *dst_cset;
2183 dst_cset = find_css_set(src_cset,
2184 dst_cgrp ?: src_cset->dfl_cgrp);
2188 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2191 * If src cset equals dst, it's noop. Drop the src.
2192 * cgroup_migrate() will skip the cset too. Note that we
2193 * can't handle src == dst as some nodes are used by both.
2195 if (src_cset == dst_cset) {
2196 src_cset->mg_src_cgrp = NULL;
2197 list_del_init(&src_cset->mg_preload_node);
2198 put_css_set(src_cset);
2199 put_css_set(dst_cset);
2203 src_cset->mg_dst_cset = dst_cset;
2205 if (list_empty(&dst_cset->mg_preload_node))
2206 list_add(&dst_cset->mg_preload_node, &csets);
2208 put_css_set(dst_cset);
2211 list_splice_tail(&csets, preloaded_csets);
2214 cgroup_migrate_finish(&csets);
2219 * cgroup_migrate - migrate a process or task to a cgroup
2220 * @cgrp: the destination cgroup
2221 * @leader: the leader of the process or the task to migrate
2222 * @threadgroup: whether @leader points to the whole process or a single task
2224 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2225 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2226 * caller is also responsible for invoking cgroup_migrate_add_src() and
2227 * cgroup_migrate_prepare_dst() on the targets before invoking this
2228 * function and following up with cgroup_migrate_finish().
2230 * As long as a controller's ->can_attach() doesn't fail, this function is
2231 * guaranteed to succeed. This means that, excluding ->can_attach()
2232 * failure, when migrating multiple targets, the success or failure can be
2233 * decided for all targets by invoking group_migrate_prepare_dst() before
2234 * actually starting migrating.
2236 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2239 struct cgroup_taskset tset = {
2240 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2241 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2242 .csets = &tset.src_csets,
2244 struct cgroup_subsys_state *css, *failed_css = NULL;
2245 struct css_set *cset, *tmp_cset;
2246 struct task_struct *task, *tmp_task;
2250 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2251 * already PF_EXITING could be freed from underneath us unless we
2252 * take an rcu_read_lock.
2254 down_write(&css_set_rwsem);
2258 /* @task either already exited or can't exit until the end */
2259 if (task->flags & PF_EXITING)
2262 /* leave @task alone if post_fork() hasn't linked it yet */
2263 if (list_empty(&task->cg_list))
2266 cset = task_css_set(task);
2267 if (!cset->mg_src_cgrp)
2271 * cgroup_taskset_first() must always return the leader.
2272 * Take care to avoid disturbing the ordering.
2274 list_move_tail(&task->cg_list, &cset->mg_tasks);
2275 if (list_empty(&cset->mg_node))
2276 list_add_tail(&cset->mg_node, &tset.src_csets);
2277 if (list_empty(&cset->mg_dst_cset->mg_node))
2278 list_move_tail(&cset->mg_dst_cset->mg_node,
2283 } while_each_thread(leader, task);
2285 up_write(&css_set_rwsem);
2287 /* methods shouldn't be called if no task is actually migrating */
2288 if (list_empty(&tset.src_csets))
2291 /* check that we can legitimately attach to the cgroup */
2292 for_each_e_css(css, i, cgrp) {
2293 if (css->ss->can_attach) {
2294 ret = css->ss->can_attach(css, &tset);
2297 goto out_cancel_attach;
2303 * Now that we're guaranteed success, proceed to move all tasks to
2304 * the new cgroup. There are no failure cases after here, so this
2305 * is the commit point.
2307 down_write(&css_set_rwsem);
2308 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2309 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2310 cgroup_task_migrate(cset->mg_src_cgrp, task,
2313 up_write(&css_set_rwsem);
2316 * Migration is committed, all target tasks are now on dst_csets.
2317 * Nothing is sensitive to fork() after this point. Notify
2318 * controllers that migration is complete.
2320 tset.csets = &tset.dst_csets;
2322 for_each_e_css(css, i, cgrp)
2323 if (css->ss->attach)
2324 css->ss->attach(css, &tset);
2327 goto out_release_tset;
2330 for_each_e_css(css, i, cgrp) {
2331 if (css == failed_css)
2333 if (css->ss->cancel_attach)
2334 css->ss->cancel_attach(css, &tset);
2337 down_write(&css_set_rwsem);
2338 list_splice_init(&tset.dst_csets, &tset.src_csets);
2339 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2340 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2341 list_del_init(&cset->mg_node);
2343 up_write(&css_set_rwsem);
2348 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2349 * @dst_cgrp: the cgroup to attach to
2350 * @leader: the task or the leader of the threadgroup to be attached
2351 * @threadgroup: attach the whole threadgroup?
2353 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2355 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2356 struct task_struct *leader, bool threadgroup)
2358 LIST_HEAD(preloaded_csets);
2359 struct task_struct *task;
2362 /* look up all src csets */
2363 down_read(&css_set_rwsem);
2367 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2371 } while_each_thread(leader, task);
2373 up_read(&css_set_rwsem);
2375 /* prepare dst csets and commit */
2376 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2378 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2380 cgroup_migrate_finish(&preloaded_csets);
2385 * Find the task_struct of the task to attach by vpid and pass it along to the
2386 * function to attach either it or all tasks in its threadgroup. Will lock
2387 * cgroup_mutex and threadgroup.
2389 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2390 size_t nbytes, loff_t off, bool threadgroup)
2392 struct task_struct *tsk;
2393 const struct cred *cred = current_cred(), *tcred;
2394 struct cgroup *cgrp;
2398 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2401 cgrp = cgroup_kn_lock_live(of->kn);
2405 percpu_down_write(&cgroup_threadgroup_rwsem);
2408 tsk = find_task_by_vpid(pid);
2411 goto out_unlock_rcu;
2414 * even if we're attaching all tasks in the thread group, we
2415 * only need to check permissions on one of them.
2417 tcred = __task_cred(tsk);
2418 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2419 !uid_eq(cred->euid, tcred->uid) &&
2420 !uid_eq(cred->euid, tcred->suid)) {
2422 goto out_unlock_rcu;
2428 tsk = tsk->group_leader;
2431 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2432 * trapped in a cpuset, or RT worker may be born in a cgroup
2433 * with no rt_runtime allocated. Just say no.
2435 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2437 goto out_unlock_rcu;
2440 get_task_struct(tsk);
2443 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2445 put_task_struct(tsk);
2446 goto out_unlock_threadgroup;
2450 out_unlock_threadgroup:
2451 percpu_up_write(&cgroup_threadgroup_rwsem);
2452 cgroup_kn_unlock(of->kn);
2453 return ret ?: nbytes;
2457 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2458 * @from: attach to all cgroups of a given task
2459 * @tsk: the task to be attached
2461 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2463 struct cgroup_root *root;
2466 mutex_lock(&cgroup_mutex);
2467 for_each_root(root) {
2468 struct cgroup *from_cgrp;
2470 if (root == &cgrp_dfl_root)
2473 down_read(&css_set_rwsem);
2474 from_cgrp = task_cgroup_from_root(from, root);
2475 up_read(&css_set_rwsem);
2477 retval = cgroup_attach_task(from_cgrp, tsk, false);
2481 mutex_unlock(&cgroup_mutex);
2485 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2487 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2488 char *buf, size_t nbytes, loff_t off)
2490 return __cgroup_procs_write(of, buf, nbytes, off, false);
2493 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2494 char *buf, size_t nbytes, loff_t off)
2496 return __cgroup_procs_write(of, buf, nbytes, off, true);
2499 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2500 char *buf, size_t nbytes, loff_t off)
2502 struct cgroup *cgrp;
2504 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2506 cgrp = cgroup_kn_lock_live(of->kn);
2509 spin_lock(&release_agent_path_lock);
2510 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2511 sizeof(cgrp->root->release_agent_path));
2512 spin_unlock(&release_agent_path_lock);
2513 cgroup_kn_unlock(of->kn);
2517 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2519 struct cgroup *cgrp = seq_css(seq)->cgroup;
2521 spin_lock(&release_agent_path_lock);
2522 seq_puts(seq, cgrp->root->release_agent_path);
2523 spin_unlock(&release_agent_path_lock);
2524 seq_putc(seq, '\n');
2528 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2530 seq_puts(seq, "0\n");
2534 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2536 struct cgroup_subsys *ss;
2537 bool printed = false;
2540 for_each_subsys(ss, ssid) {
2541 if (ss_mask & (1 << ssid)) {
2544 seq_printf(seq, "%s", ss->name);
2549 seq_putc(seq, '\n');
2552 /* show controllers which are currently attached to the default hierarchy */
2553 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2555 struct cgroup *cgrp = seq_css(seq)->cgroup;
2557 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2558 ~cgrp_dfl_root_inhibit_ss_mask);
2562 /* show controllers which are enabled from the parent */
2563 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2565 struct cgroup *cgrp = seq_css(seq)->cgroup;
2567 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2571 /* show controllers which are enabled for a given cgroup's children */
2572 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2574 struct cgroup *cgrp = seq_css(seq)->cgroup;
2576 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2581 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2582 * @cgrp: root of the subtree to update csses for
2584 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2585 * css associations need to be updated accordingly. This function looks up
2586 * all css_sets which are attached to the subtree, creates the matching
2587 * updated css_sets and migrates the tasks to the new ones.
2589 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2591 LIST_HEAD(preloaded_csets);
2592 struct cgroup_subsys_state *css;
2593 struct css_set *src_cset;
2596 lockdep_assert_held(&cgroup_mutex);
2598 percpu_down_write(&cgroup_threadgroup_rwsem);
2600 /* look up all csses currently attached to @cgrp's subtree */
2601 down_read(&css_set_rwsem);
2602 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2603 struct cgrp_cset_link *link;
2605 /* self is not affected by child_subsys_mask change */
2606 if (css->cgroup == cgrp)
2609 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2610 cgroup_migrate_add_src(link->cset, cgrp,
2613 up_read(&css_set_rwsem);
2615 /* NULL dst indicates self on default hierarchy */
2616 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2620 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2621 struct task_struct *last_task = NULL, *task;
2623 /* src_csets precede dst_csets, break on the first dst_cset */
2624 if (!src_cset->mg_src_cgrp)
2628 * All tasks in src_cset need to be migrated to the
2629 * matching dst_cset. Empty it process by process. We
2630 * walk tasks but migrate processes. The leader might even
2631 * belong to a different cset but such src_cset would also
2632 * be among the target src_csets because the default
2633 * hierarchy enforces per-process membership.
2636 down_read(&css_set_rwsem);
2637 task = list_first_entry_or_null(&src_cset->tasks,
2638 struct task_struct, cg_list);
2640 task = task->group_leader;
2641 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2642 get_task_struct(task);
2644 up_read(&css_set_rwsem);
2649 /* guard against possible infinite loop */
2650 if (WARN(last_task == task,
2651 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2655 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2657 put_task_struct(task);
2659 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2665 cgroup_migrate_finish(&preloaded_csets);
2666 percpu_up_write(&cgroup_threadgroup_rwsem);
2670 /* change the enabled child controllers for a cgroup in the default hierarchy */
2671 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2672 char *buf, size_t nbytes,
2675 unsigned long enable = 0, disable = 0;
2676 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2677 struct cgroup *cgrp, *child;
2678 struct cgroup_subsys *ss;
2683 * Parse input - space separated list of subsystem names prefixed
2684 * with either + or -.
2686 buf = strstrip(buf);
2687 while ((tok = strsep(&buf, " "))) {
2690 for_each_subsys(ss, ssid) {
2691 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2692 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2696 enable |= 1 << ssid;
2697 disable &= ~(1 << ssid);
2698 } else if (*tok == '-') {
2699 disable |= 1 << ssid;
2700 enable &= ~(1 << ssid);
2706 if (ssid == CGROUP_SUBSYS_COUNT)
2710 cgrp = cgroup_kn_lock_live(of->kn);
2714 for_each_subsys(ss, ssid) {
2715 if (enable & (1 << ssid)) {
2716 if (cgrp->subtree_control & (1 << ssid)) {
2717 enable &= ~(1 << ssid);
2721 /* unavailable or not enabled on the parent? */
2722 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2723 (cgroup_parent(cgrp) &&
2724 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2728 } else if (disable & (1 << ssid)) {
2729 if (!(cgrp->subtree_control & (1 << ssid))) {
2730 disable &= ~(1 << ssid);
2734 /* a child has it enabled? */
2735 cgroup_for_each_live_child(child, cgrp) {
2736 if (child->subtree_control & (1 << ssid)) {
2744 if (!enable && !disable) {
2750 * Except for the root, subtree_control must be zero for a cgroup
2751 * with tasks so that child cgroups don't compete against tasks.
2753 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2759 * Update subsys masks and calculate what needs to be done. More
2760 * subsystems than specified may need to be enabled or disabled
2761 * depending on subsystem dependencies.
2763 old_sc = cgrp->subtree_control;
2764 old_ss = cgrp->child_subsys_mask;
2765 new_sc = (old_sc | enable) & ~disable;
2766 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2768 css_enable = ~old_ss & new_ss;
2769 css_disable = old_ss & ~new_ss;
2770 enable |= css_enable;
2771 disable |= css_disable;
2774 * Because css offlining is asynchronous, userland might try to
2775 * re-enable the same controller while the previous instance is
2776 * still around. In such cases, wait till it's gone using
2779 for_each_subsys(ss, ssid) {
2780 if (!(css_enable & (1 << ssid)))
2783 cgroup_for_each_live_child(child, cgrp) {
2786 if (!cgroup_css(child, ss))
2790 prepare_to_wait(&child->offline_waitq, &wait,
2791 TASK_UNINTERRUPTIBLE);
2792 cgroup_kn_unlock(of->kn);
2794 finish_wait(&child->offline_waitq, &wait);
2797 return restart_syscall();
2801 cgrp->subtree_control = new_sc;
2802 cgrp->child_subsys_mask = new_ss;
2805 * Create new csses or make the existing ones visible. A css is
2806 * created invisible if it's being implicitly enabled through
2807 * dependency. An invisible css is made visible when the userland
2808 * explicitly enables it.
2810 for_each_subsys(ss, ssid) {
2811 if (!(enable & (1 << ssid)))
2814 cgroup_for_each_live_child(child, cgrp) {
2815 if (css_enable & (1 << ssid))
2816 ret = create_css(child, ss,
2817 cgrp->subtree_control & (1 << ssid));
2819 ret = cgroup_populate_dir(child, 1 << ssid);
2826 * At this point, cgroup_e_css() results reflect the new csses
2827 * making the following cgroup_update_dfl_csses() properly update
2828 * css associations of all tasks in the subtree.
2830 ret = cgroup_update_dfl_csses(cgrp);
2835 * All tasks are migrated out of disabled csses. Kill or hide
2836 * them. A css is hidden when the userland requests it to be
2837 * disabled while other subsystems are still depending on it. The
2838 * css must not actively control resources and be in the vanilla
2839 * state if it's made visible again later. Controllers which may
2840 * be depended upon should provide ->css_reset() for this purpose.
2842 for_each_subsys(ss, ssid) {
2843 if (!(disable & (1 << ssid)))
2846 cgroup_for_each_live_child(child, cgrp) {
2847 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2849 if (css_disable & (1 << ssid)) {
2852 cgroup_clear_dir(child, 1 << ssid);
2860 * The effective csses of all the descendants (excluding @cgrp) may
2861 * have changed. Subsystems can optionally subscribe to this event
2862 * by implementing ->css_e_css_changed() which is invoked if any of
2863 * the effective csses seen from the css's cgroup may have changed.
2865 for_each_subsys(ss, ssid) {
2866 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2867 struct cgroup_subsys_state *css;
2869 if (!ss->css_e_css_changed || !this_css)
2872 css_for_each_descendant_pre(css, this_css)
2873 if (css != this_css)
2874 ss->css_e_css_changed(css);
2877 kernfs_activate(cgrp->kn);
2880 cgroup_kn_unlock(of->kn);
2881 return ret ?: nbytes;
2884 cgrp->subtree_control = old_sc;
2885 cgrp->child_subsys_mask = old_ss;
2887 for_each_subsys(ss, ssid) {
2888 if (!(enable & (1 << ssid)))
2891 cgroup_for_each_live_child(child, cgrp) {
2892 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2897 if (css_enable & (1 << ssid))
2900 cgroup_clear_dir(child, 1 << ssid);
2906 static int cgroup_populated_show(struct seq_file *seq, void *v)
2908 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2912 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2913 size_t nbytes, loff_t off)
2915 struct cgroup *cgrp = of->kn->parent->priv;
2916 struct cftype *cft = of->kn->priv;
2917 struct cgroup_subsys_state *css;
2921 return cft->write(of, buf, nbytes, off);
2924 * kernfs guarantees that a file isn't deleted with operations in
2925 * flight, which means that the matching css is and stays alive and
2926 * doesn't need to be pinned. The RCU locking is not necessary
2927 * either. It's just for the convenience of using cgroup_css().
2930 css = cgroup_css(cgrp, cft->ss);
2933 if (cft->write_u64) {
2934 unsigned long long v;
2935 ret = kstrtoull(buf, 0, &v);
2937 ret = cft->write_u64(css, cft, v);
2938 } else if (cft->write_s64) {
2940 ret = kstrtoll(buf, 0, &v);
2942 ret = cft->write_s64(css, cft, v);
2947 return ret ?: nbytes;
2950 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2952 return seq_cft(seq)->seq_start(seq, ppos);
2955 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2957 return seq_cft(seq)->seq_next(seq, v, ppos);
2960 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2962 seq_cft(seq)->seq_stop(seq, v);
2965 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2967 struct cftype *cft = seq_cft(m);
2968 struct cgroup_subsys_state *css = seq_css(m);
2971 return cft->seq_show(m, arg);
2974 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2975 else if (cft->read_s64)
2976 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2982 static struct kernfs_ops cgroup_kf_single_ops = {
2983 .atomic_write_len = PAGE_SIZE,
2984 .write = cgroup_file_write,
2985 .seq_show = cgroup_seqfile_show,
2988 static struct kernfs_ops cgroup_kf_ops = {
2989 .atomic_write_len = PAGE_SIZE,
2990 .write = cgroup_file_write,
2991 .seq_start = cgroup_seqfile_start,
2992 .seq_next = cgroup_seqfile_next,
2993 .seq_stop = cgroup_seqfile_stop,
2994 .seq_show = cgroup_seqfile_show,
2998 * cgroup_rename - Only allow simple rename of directories in place.
3000 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3001 const char *new_name_str)
3003 struct cgroup *cgrp = kn->priv;
3006 if (kernfs_type(kn) != KERNFS_DIR)
3008 if (kn->parent != new_parent)
3012 * This isn't a proper migration and its usefulness is very
3013 * limited. Disallow on the default hierarchy.
3015 if (cgroup_on_dfl(cgrp))
3019 * We're gonna grab cgroup_mutex which nests outside kernfs
3020 * active_ref. kernfs_rename() doesn't require active_ref
3021 * protection. Break them before grabbing cgroup_mutex.
3023 kernfs_break_active_protection(new_parent);
3024 kernfs_break_active_protection(kn);
3026 mutex_lock(&cgroup_mutex);
3028 ret = kernfs_rename(kn, new_parent, new_name_str);
3030 mutex_unlock(&cgroup_mutex);
3032 kernfs_unbreak_active_protection(kn);
3033 kernfs_unbreak_active_protection(new_parent);
3037 /* set uid and gid of cgroup dirs and files to that of the creator */
3038 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3040 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3041 .ia_uid = current_fsuid(),
3042 .ia_gid = current_fsgid(), };
3044 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3045 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3048 return kernfs_setattr(kn, &iattr);
3051 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3053 char name[CGROUP_FILE_NAME_MAX];
3054 struct kernfs_node *kn;
3055 struct lock_class_key *key = NULL;
3058 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3059 key = &cft->lockdep_key;
3061 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3062 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3067 ret = cgroup_kn_set_ugid(kn);
3073 if (cft->seq_show == cgroup_populated_show)
3074 cgrp->populated_kn = kn;
3079 * cgroup_addrm_files - add or remove files to a cgroup directory
3080 * @cgrp: the target cgroup
3081 * @cfts: array of cftypes to be added
3082 * @is_add: whether to add or remove
3084 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3085 * For removals, this function never fails. If addition fails, this
3086 * function doesn't remove files already added. The caller is responsible
3089 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3095 lockdep_assert_held(&cgroup_mutex);
3097 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3098 /* does cft->flags tell us to skip this file on @cgrp? */
3099 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3101 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3103 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3105 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3109 ret = cgroup_add_file(cgrp, cft);
3111 pr_warn("%s: failed to add %s, err=%d\n",
3112 __func__, cft->name, ret);
3116 cgroup_rm_file(cgrp, cft);
3122 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3125 struct cgroup_subsys *ss = cfts[0].ss;
3126 struct cgroup *root = &ss->root->cgrp;
3127 struct cgroup_subsys_state *css;
3130 lockdep_assert_held(&cgroup_mutex);
3132 /* add/rm files for all cgroups created before */
3133 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3134 struct cgroup *cgrp = css->cgroup;
3136 if (cgroup_is_dead(cgrp))
3139 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3145 kernfs_activate(root->kn);
3149 static void cgroup_exit_cftypes(struct cftype *cfts)
3153 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3154 /* free copy for custom atomic_write_len, see init_cftypes() */
3155 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3160 /* revert flags set by cgroup core while adding @cfts */
3161 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3165 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3169 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3170 struct kernfs_ops *kf_ops;
3172 WARN_ON(cft->ss || cft->kf_ops);
3175 kf_ops = &cgroup_kf_ops;
3177 kf_ops = &cgroup_kf_single_ops;
3180 * Ugh... if @cft wants a custom max_write_len, we need to
3181 * make a copy of kf_ops to set its atomic_write_len.
3183 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3184 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3186 cgroup_exit_cftypes(cfts);
3189 kf_ops->atomic_write_len = cft->max_write_len;
3192 cft->kf_ops = kf_ops;
3199 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3201 lockdep_assert_held(&cgroup_mutex);
3203 if (!cfts || !cfts[0].ss)
3206 list_del(&cfts->node);
3207 cgroup_apply_cftypes(cfts, false);
3208 cgroup_exit_cftypes(cfts);
3213 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3214 * @cfts: zero-length name terminated array of cftypes
3216 * Unregister @cfts. Files described by @cfts are removed from all
3217 * existing cgroups and all future cgroups won't have them either. This
3218 * function can be called anytime whether @cfts' subsys is attached or not.
3220 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3223 int cgroup_rm_cftypes(struct cftype *cfts)
3227 mutex_lock(&cgroup_mutex);
3228 ret = cgroup_rm_cftypes_locked(cfts);
3229 mutex_unlock(&cgroup_mutex);
3234 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3235 * @ss: target cgroup subsystem
3236 * @cfts: zero-length name terminated array of cftypes
3238 * Register @cfts to @ss. Files described by @cfts are created for all
3239 * existing cgroups to which @ss is attached and all future cgroups will
3240 * have them too. This function can be called anytime whether @ss is
3243 * Returns 0 on successful registration, -errno on failure. Note that this
3244 * function currently returns 0 as long as @cfts registration is successful
3245 * even if some file creation attempts on existing cgroups fail.
3247 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3254 if (!cfts || cfts[0].name[0] == '\0')
3257 ret = cgroup_init_cftypes(ss, cfts);
3261 mutex_lock(&cgroup_mutex);
3263 list_add_tail(&cfts->node, &ss->cfts);
3264 ret = cgroup_apply_cftypes(cfts, true);
3266 cgroup_rm_cftypes_locked(cfts);
3268 mutex_unlock(&cgroup_mutex);
3273 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3274 * @ss: target cgroup subsystem
3275 * @cfts: zero-length name terminated array of cftypes
3277 * Similar to cgroup_add_cftypes() but the added files are only used for
3278 * the default hierarchy.
3280 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3284 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3285 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3286 return cgroup_add_cftypes(ss, cfts);
3290 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3291 * @ss: target cgroup subsystem
3292 * @cfts: zero-length name terminated array of cftypes
3294 * Similar to cgroup_add_cftypes() but the added files are only used for
3295 * the legacy hierarchies.
3297 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3302 * If legacy_flies_on_dfl, we want to show the legacy files on the
3303 * dfl hierarchy but iff the target subsystem hasn't been updated
3304 * for the dfl hierarchy yet.
3306 if (!cgroup_legacy_files_on_dfl ||
3307 ss->dfl_cftypes != ss->legacy_cftypes) {
3308 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3309 cft->flags |= __CFTYPE_NOT_ON_DFL;
3312 return cgroup_add_cftypes(ss, cfts);
3316 * cgroup_task_count - count the number of tasks in a cgroup.
3317 * @cgrp: the cgroup in question
3319 * Return the number of tasks in the cgroup.
3321 static int cgroup_task_count(const struct cgroup *cgrp)
3324 struct cgrp_cset_link *link;
3326 down_read(&css_set_rwsem);
3327 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3328 count += atomic_read(&link->cset->refcount);
3329 up_read(&css_set_rwsem);
3334 * css_next_child - find the next child of a given css
3335 * @pos: the current position (%NULL to initiate traversal)
3336 * @parent: css whose children to walk
3338 * This function returns the next child of @parent and should be called
3339 * under either cgroup_mutex or RCU read lock. The only requirement is
3340 * that @parent and @pos are accessible. The next sibling is guaranteed to
3341 * be returned regardless of their states.
3343 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3344 * css which finished ->css_online() is guaranteed to be visible in the
3345 * future iterations and will stay visible until the last reference is put.
3346 * A css which hasn't finished ->css_online() or already finished
3347 * ->css_offline() may show up during traversal. It's each subsystem's
3348 * responsibility to synchronize against on/offlining.
3350 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3351 struct cgroup_subsys_state *parent)
3353 struct cgroup_subsys_state *next;
3355 cgroup_assert_mutex_or_rcu_locked();
3358 * @pos could already have been unlinked from the sibling list.
3359 * Once a cgroup is removed, its ->sibling.next is no longer
3360 * updated when its next sibling changes. CSS_RELEASED is set when
3361 * @pos is taken off list, at which time its next pointer is valid,
3362 * and, as releases are serialized, the one pointed to by the next
3363 * pointer is guaranteed to not have started release yet. This
3364 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3365 * critical section, the one pointed to by its next pointer is
3366 * guaranteed to not have finished its RCU grace period even if we
3367 * have dropped rcu_read_lock() inbetween iterations.
3369 * If @pos has CSS_RELEASED set, its next pointer can't be
3370 * dereferenced; however, as each css is given a monotonically
3371 * increasing unique serial number and always appended to the
3372 * sibling list, the next one can be found by walking the parent's
3373 * children until the first css with higher serial number than
3374 * @pos's. While this path can be slower, it happens iff iteration
3375 * races against release and the race window is very small.
3378 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3379 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3380 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3382 list_for_each_entry_rcu(next, &parent->children, sibling)
3383 if (next->serial_nr > pos->serial_nr)
3388 * @next, if not pointing to the head, can be dereferenced and is
3391 if (&next->sibling != &parent->children)
3397 * css_next_descendant_pre - find the next descendant for pre-order walk
3398 * @pos: the current position (%NULL to initiate traversal)
3399 * @root: css whose descendants to walk
3401 * To be used by css_for_each_descendant_pre(). Find the next descendant
3402 * to visit for pre-order traversal of @root's descendants. @root is
3403 * included in the iteration and the first node to be visited.
3405 * While this function requires cgroup_mutex or RCU read locking, it
3406 * doesn't require the whole traversal to be contained in a single critical
3407 * section. This function will return the correct next descendant as long
3408 * as both @pos and @root are accessible and @pos is a descendant of @root.
3410 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3411 * css which finished ->css_online() is guaranteed to be visible in the
3412 * future iterations and will stay visible until the last reference is put.
3413 * A css which hasn't finished ->css_online() or already finished
3414 * ->css_offline() may show up during traversal. It's each subsystem's
3415 * responsibility to synchronize against on/offlining.
3417 struct cgroup_subsys_state *
3418 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3419 struct cgroup_subsys_state *root)
3421 struct cgroup_subsys_state *next;
3423 cgroup_assert_mutex_or_rcu_locked();
3425 /* if first iteration, visit @root */
3429 /* visit the first child if exists */
3430 next = css_next_child(NULL, pos);
3434 /* no child, visit my or the closest ancestor's next sibling */
3435 while (pos != root) {
3436 next = css_next_child(pos, pos->parent);
3446 * css_rightmost_descendant - return the rightmost descendant of a css
3447 * @pos: css of interest
3449 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3450 * is returned. This can be used during pre-order traversal to skip
3453 * While this function requires cgroup_mutex or RCU read locking, it
3454 * doesn't require the whole traversal to be contained in a single critical
3455 * section. This function will return the correct rightmost descendant as
3456 * long as @pos is accessible.
3458 struct cgroup_subsys_state *
3459 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3461 struct cgroup_subsys_state *last, *tmp;
3463 cgroup_assert_mutex_or_rcu_locked();
3467 /* ->prev isn't RCU safe, walk ->next till the end */
3469 css_for_each_child(tmp, last)
3476 static struct cgroup_subsys_state *
3477 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3479 struct cgroup_subsys_state *last;
3483 pos = css_next_child(NULL, pos);
3490 * css_next_descendant_post - find the next descendant for post-order walk
3491 * @pos: the current position (%NULL to initiate traversal)
3492 * @root: css whose descendants to walk
3494 * To be used by css_for_each_descendant_post(). Find the next descendant
3495 * to visit for post-order traversal of @root's descendants. @root is
3496 * included in the iteration and the last node to be visited.
3498 * While this function requires cgroup_mutex or RCU read locking, it
3499 * doesn't require the whole traversal to be contained in a single critical
3500 * section. This function will return the correct next descendant as long
3501 * as both @pos and @cgroup are accessible and @pos is a descendant of
3504 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3505 * css which finished ->css_online() is guaranteed to be visible in the
3506 * future iterations and will stay visible until the last reference is put.
3507 * A css which hasn't finished ->css_online() or already finished
3508 * ->css_offline() may show up during traversal. It's each subsystem's
3509 * responsibility to synchronize against on/offlining.
3511 struct cgroup_subsys_state *
3512 css_next_descendant_post(struct cgroup_subsys_state *pos,
3513 struct cgroup_subsys_state *root)
3515 struct cgroup_subsys_state *next;
3517 cgroup_assert_mutex_or_rcu_locked();
3519 /* if first iteration, visit leftmost descendant which may be @root */
3521 return css_leftmost_descendant(root);
3523 /* if we visited @root, we're done */
3527 /* if there's an unvisited sibling, visit its leftmost descendant */
3528 next = css_next_child(pos, pos->parent);
3530 return css_leftmost_descendant(next);
3532 /* no sibling left, visit parent */
3537 * css_has_online_children - does a css have online children
3538 * @css: the target css
3540 * Returns %true if @css has any online children; otherwise, %false. This
3541 * function can be called from any context but the caller is responsible
3542 * for synchronizing against on/offlining as necessary.
3544 bool css_has_online_children(struct cgroup_subsys_state *css)
3546 struct cgroup_subsys_state *child;
3550 css_for_each_child(child, css) {
3551 if (child->flags & CSS_ONLINE) {
3561 * css_advance_task_iter - advance a task itererator to the next css_set
3562 * @it: the iterator to advance
3564 * Advance @it to the next css_set to walk.
3566 static void css_advance_task_iter(struct css_task_iter *it)
3568 struct list_head *l = it->cset_pos;
3569 struct cgrp_cset_link *link;
3570 struct css_set *cset;
3572 /* Advance to the next non-empty css_set */
3575 if (l == it->cset_head) {
3576 it->cset_pos = NULL;
3581 cset = container_of(l, struct css_set,
3582 e_cset_node[it->ss->id]);
3584 link = list_entry(l, struct cgrp_cset_link, cset_link);
3587 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3591 if (!list_empty(&cset->tasks))
3592 it->task_pos = cset->tasks.next;
3594 it->task_pos = cset->mg_tasks.next;
3596 it->tasks_head = &cset->tasks;
3597 it->mg_tasks_head = &cset->mg_tasks;
3601 * css_task_iter_start - initiate task iteration
3602 * @css: the css to walk tasks of
3603 * @it: the task iterator to use
3605 * Initiate iteration through the tasks of @css. The caller can call
3606 * css_task_iter_next() to walk through the tasks until the function
3607 * returns NULL. On completion of iteration, css_task_iter_end() must be
3610 * Note that this function acquires a lock which is released when the
3611 * iteration finishes. The caller can't sleep while iteration is in
3614 void css_task_iter_start(struct cgroup_subsys_state *css,
3615 struct css_task_iter *it)
3616 __acquires(css_set_rwsem)
3618 /* no one should try to iterate before mounting cgroups */
3619 WARN_ON_ONCE(!use_task_css_set_links);
3621 down_read(&css_set_rwsem);
3626 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3628 it->cset_pos = &css->cgroup->cset_links;
3630 it->cset_head = it->cset_pos;
3632 css_advance_task_iter(it);
3636 * css_task_iter_next - return the next task for the iterator
3637 * @it: the task iterator being iterated
3639 * The "next" function for task iteration. @it should have been
3640 * initialized via css_task_iter_start(). Returns NULL when the iteration
3643 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3645 struct task_struct *res;
3646 struct list_head *l = it->task_pos;
3648 /* If the iterator cg is NULL, we have no tasks */
3651 res = list_entry(l, struct task_struct, cg_list);
3654 * Advance iterator to find next entry. cset->tasks is consumed
3655 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3660 if (l == it->tasks_head)
3661 l = it->mg_tasks_head->next;
3663 if (l == it->mg_tasks_head)
3664 css_advance_task_iter(it);
3672 * css_task_iter_end - finish task iteration
3673 * @it: the task iterator to finish
3675 * Finish task iteration started by css_task_iter_start().
3677 void css_task_iter_end(struct css_task_iter *it)
3678 __releases(css_set_rwsem)
3680 up_read(&css_set_rwsem);
3684 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3685 * @to: cgroup to which the tasks will be moved
3686 * @from: cgroup in which the tasks currently reside
3688 * Locking rules between cgroup_post_fork() and the migration path
3689 * guarantee that, if a task is forking while being migrated, the new child
3690 * is guaranteed to be either visible in the source cgroup after the
3691 * parent's migration is complete or put into the target cgroup. No task
3692 * can slip out of migration through forking.
3694 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3696 LIST_HEAD(preloaded_csets);
3697 struct cgrp_cset_link *link;
3698 struct css_task_iter it;
3699 struct task_struct *task;
3702 mutex_lock(&cgroup_mutex);
3704 /* all tasks in @from are being moved, all csets are source */
3705 down_read(&css_set_rwsem);
3706 list_for_each_entry(link, &from->cset_links, cset_link)
3707 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3708 up_read(&css_set_rwsem);
3710 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3715 * Migrate tasks one-by-one until @form is empty. This fails iff
3716 * ->can_attach() fails.
3719 css_task_iter_start(&from->self, &it);
3720 task = css_task_iter_next(&it);
3722 get_task_struct(task);
3723 css_task_iter_end(&it);
3726 ret = cgroup_migrate(to, task, false);
3727 put_task_struct(task);
3729 } while (task && !ret);
3731 cgroup_migrate_finish(&preloaded_csets);
3732 mutex_unlock(&cgroup_mutex);
3737 * Stuff for reading the 'tasks'/'procs' files.
3739 * Reading this file can return large amounts of data if a cgroup has
3740 * *lots* of attached tasks. So it may need several calls to read(),
3741 * but we cannot guarantee that the information we produce is correct
3742 * unless we produce it entirely atomically.
3746 /* which pidlist file are we talking about? */
3747 enum cgroup_filetype {
3753 * A pidlist is a list of pids that virtually represents the contents of one
3754 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3755 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3758 struct cgroup_pidlist {
3760 * used to find which pidlist is wanted. doesn't change as long as
3761 * this particular list stays in the list.
3763 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3766 /* how many elements the above list has */
3768 /* each of these stored in a list by its cgroup */
3769 struct list_head links;
3770 /* pointer to the cgroup we belong to, for list removal purposes */
3771 struct cgroup *owner;
3772 /* for delayed destruction */
3773 struct delayed_work destroy_dwork;
3777 * The following two functions "fix" the issue where there are more pids
3778 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3779 * TODO: replace with a kernel-wide solution to this problem
3781 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3782 static void *pidlist_allocate(int count)
3784 if (PIDLIST_TOO_LARGE(count))
3785 return vmalloc(count * sizeof(pid_t));
3787 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3790 static void pidlist_free(void *p)
3796 * Used to destroy all pidlists lingering waiting for destroy timer. None
3797 * should be left afterwards.
3799 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3801 struct cgroup_pidlist *l, *tmp_l;
3803 mutex_lock(&cgrp->pidlist_mutex);
3804 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3805 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3806 mutex_unlock(&cgrp->pidlist_mutex);
3808 flush_workqueue(cgroup_pidlist_destroy_wq);
3809 BUG_ON(!list_empty(&cgrp->pidlists));
3812 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3814 struct delayed_work *dwork = to_delayed_work(work);
3815 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3817 struct cgroup_pidlist *tofree = NULL;
3819 mutex_lock(&l->owner->pidlist_mutex);
3822 * Destroy iff we didn't get queued again. The state won't change
3823 * as destroy_dwork can only be queued while locked.
3825 if (!delayed_work_pending(dwork)) {
3826 list_del(&l->links);
3827 pidlist_free(l->list);
3828 put_pid_ns(l->key.ns);
3832 mutex_unlock(&l->owner->pidlist_mutex);
3837 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3838 * Returns the number of unique elements.
3840 static int pidlist_uniq(pid_t *list, int length)
3845 * we presume the 0th element is unique, so i starts at 1. trivial
3846 * edge cases first; no work needs to be done for either
3848 if (length == 0 || length == 1)
3850 /* src and dest walk down the list; dest counts unique elements */
3851 for (src = 1; src < length; src++) {
3852 /* find next unique element */
3853 while (list[src] == list[src-1]) {
3858 /* dest always points to where the next unique element goes */
3859 list[dest] = list[src];
3867 * The two pid files - task and cgroup.procs - guaranteed that the result
3868 * is sorted, which forced this whole pidlist fiasco. As pid order is
3869 * different per namespace, each namespace needs differently sorted list,
3870 * making it impossible to use, for example, single rbtree of member tasks
3871 * sorted by task pointer. As pidlists can be fairly large, allocating one
3872 * per open file is dangerous, so cgroup had to implement shared pool of
3873 * pidlists keyed by cgroup and namespace.
3875 * All this extra complexity was caused by the original implementation
3876 * committing to an entirely unnecessary property. In the long term, we
3877 * want to do away with it. Explicitly scramble sort order if on the
3878 * default hierarchy so that no such expectation exists in the new
3881 * Scrambling is done by swapping every two consecutive bits, which is
3882 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3884 static pid_t pid_fry(pid_t pid)
3886 unsigned a = pid & 0x55555555;
3887 unsigned b = pid & 0xAAAAAAAA;
3889 return (a << 1) | (b >> 1);
3892 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3894 if (cgroup_on_dfl(cgrp))
3895 return pid_fry(pid);
3900 static int cmppid(const void *a, const void *b)
3902 return *(pid_t *)a - *(pid_t *)b;
3905 static int fried_cmppid(const void *a, const void *b)
3907 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3910 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3911 enum cgroup_filetype type)
3913 struct cgroup_pidlist *l;
3914 /* don't need task_nsproxy() if we're looking at ourself */
3915 struct pid_namespace *ns = task_active_pid_ns(current);
3917 lockdep_assert_held(&cgrp->pidlist_mutex);
3919 list_for_each_entry(l, &cgrp->pidlists, links)
3920 if (l->key.type == type && l->key.ns == ns)
3926 * find the appropriate pidlist for our purpose (given procs vs tasks)
3927 * returns with the lock on that pidlist already held, and takes care
3928 * of the use count, or returns NULL with no locks held if we're out of
3931 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3932 enum cgroup_filetype type)
3934 struct cgroup_pidlist *l;
3936 lockdep_assert_held(&cgrp->pidlist_mutex);
3938 l = cgroup_pidlist_find(cgrp, type);
3942 /* entry not found; create a new one */
3943 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3947 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3949 /* don't need task_nsproxy() if we're looking at ourself */
3950 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3952 list_add(&l->links, &cgrp->pidlists);
3957 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3959 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3960 struct cgroup_pidlist **lp)
3964 int pid, n = 0; /* used for populating the array */
3965 struct css_task_iter it;
3966 struct task_struct *tsk;
3967 struct cgroup_pidlist *l;
3969 lockdep_assert_held(&cgrp->pidlist_mutex);
3972 * If cgroup gets more users after we read count, we won't have
3973 * enough space - tough. This race is indistinguishable to the
3974 * caller from the case that the additional cgroup users didn't
3975 * show up until sometime later on.
3977 length = cgroup_task_count(cgrp);
3978 array = pidlist_allocate(length);
3981 /* now, populate the array */
3982 css_task_iter_start(&cgrp->self, &it);
3983 while ((tsk = css_task_iter_next(&it))) {
3984 if (unlikely(n == length))
3986 /* get tgid or pid for procs or tasks file respectively */
3987 if (type == CGROUP_FILE_PROCS)
3988 pid = task_tgid_vnr(tsk);
3990 pid = task_pid_vnr(tsk);
3991 if (pid > 0) /* make sure to only use valid results */
3994 css_task_iter_end(&it);
3996 /* now sort & (if procs) strip out duplicates */
3997 if (cgroup_on_dfl(cgrp))
3998 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4000 sort(array, length, sizeof(pid_t), cmppid, NULL);
4001 if (type == CGROUP_FILE_PROCS)
4002 length = pidlist_uniq(array, length);
4004 l = cgroup_pidlist_find_create(cgrp, type);
4006 pidlist_free(array);
4010 /* store array, freeing old if necessary */
4011 pidlist_free(l->list);
4019 * cgroupstats_build - build and fill cgroupstats
4020 * @stats: cgroupstats to fill information into
4021 * @dentry: A dentry entry belonging to the cgroup for which stats have
4024 * Build and fill cgroupstats so that taskstats can export it to user
4027 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4029 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4030 struct cgroup *cgrp;
4031 struct css_task_iter it;
4032 struct task_struct *tsk;
4034 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4035 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4036 kernfs_type(kn) != KERNFS_DIR)
4039 mutex_lock(&cgroup_mutex);
4042 * We aren't being called from kernfs and there's no guarantee on
4043 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4044 * @kn->priv is RCU safe. Let's do the RCU dancing.
4047 cgrp = rcu_dereference(kn->priv);
4048 if (!cgrp || cgroup_is_dead(cgrp)) {
4050 mutex_unlock(&cgroup_mutex);
4055 css_task_iter_start(&cgrp->self, &it);
4056 while ((tsk = css_task_iter_next(&it))) {
4057 switch (tsk->state) {
4059 stats->nr_running++;
4061 case TASK_INTERRUPTIBLE:
4062 stats->nr_sleeping++;
4064 case TASK_UNINTERRUPTIBLE:
4065 stats->nr_uninterruptible++;
4068 stats->nr_stopped++;
4071 if (delayacct_is_task_waiting_on_io(tsk))
4072 stats->nr_io_wait++;
4076 css_task_iter_end(&it);
4078 mutex_unlock(&cgroup_mutex);
4084 * seq_file methods for the tasks/procs files. The seq_file position is the
4085 * next pid to display; the seq_file iterator is a pointer to the pid
4086 * in the cgroup->l->list array.
4089 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4092 * Initially we receive a position value that corresponds to
4093 * one more than the last pid shown (or 0 on the first call or
4094 * after a seek to the start). Use a binary-search to find the
4095 * next pid to display, if any
4097 struct kernfs_open_file *of = s->private;
4098 struct cgroup *cgrp = seq_css(s)->cgroup;
4099 struct cgroup_pidlist *l;
4100 enum cgroup_filetype type = seq_cft(s)->private;
4101 int index = 0, pid = *pos;
4104 mutex_lock(&cgrp->pidlist_mutex);
4107 * !NULL @of->priv indicates that this isn't the first start()
4108 * after open. If the matching pidlist is around, we can use that.
4109 * Look for it. Note that @of->priv can't be used directly. It
4110 * could already have been destroyed.
4113 of->priv = cgroup_pidlist_find(cgrp, type);
4116 * Either this is the first start() after open or the matching
4117 * pidlist has been destroyed inbetween. Create a new one.
4120 ret = pidlist_array_load(cgrp, type,
4121 (struct cgroup_pidlist **)&of->priv);
4123 return ERR_PTR(ret);
4128 int end = l->length;
4130 while (index < end) {
4131 int mid = (index + end) / 2;
4132 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4135 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4141 /* If we're off the end of the array, we're done */
4142 if (index >= l->length)
4144 /* Update the abstract position to be the actual pid that we found */
4145 iter = l->list + index;
4146 *pos = cgroup_pid_fry(cgrp, *iter);
4150 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4152 struct kernfs_open_file *of = s->private;
4153 struct cgroup_pidlist *l = of->priv;
4156 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4157 CGROUP_PIDLIST_DESTROY_DELAY);
4158 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4161 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4163 struct kernfs_open_file *of = s->private;
4164 struct cgroup_pidlist *l = of->priv;
4166 pid_t *end = l->list + l->length;
4168 * Advance to the next pid in the array. If this goes off the
4175 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4180 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4182 seq_printf(s, "%d\n", *(int *)v);
4187 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4190 return notify_on_release(css->cgroup);
4193 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4194 struct cftype *cft, u64 val)
4197 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4199 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4203 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4206 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4209 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4210 struct cftype *cft, u64 val)
4213 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4215 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4219 /* cgroup core interface files for the default hierarchy */
4220 static struct cftype cgroup_dfl_base_files[] = {
4222 .name = "cgroup.procs",
4223 .seq_start = cgroup_pidlist_start,
4224 .seq_next = cgroup_pidlist_next,
4225 .seq_stop = cgroup_pidlist_stop,
4226 .seq_show = cgroup_pidlist_show,
4227 .private = CGROUP_FILE_PROCS,
4228 .write = cgroup_procs_write,
4229 .mode = S_IRUGO | S_IWUSR,
4232 .name = "cgroup.controllers",
4233 .flags = CFTYPE_ONLY_ON_ROOT,
4234 .seq_show = cgroup_root_controllers_show,
4237 .name = "cgroup.controllers",
4238 .flags = CFTYPE_NOT_ON_ROOT,
4239 .seq_show = cgroup_controllers_show,
4242 .name = "cgroup.subtree_control",
4243 .seq_show = cgroup_subtree_control_show,
4244 .write = cgroup_subtree_control_write,
4247 .name = "cgroup.populated",
4248 .flags = CFTYPE_NOT_ON_ROOT,
4249 .seq_show = cgroup_populated_show,
4254 /* cgroup core interface files for the legacy hierarchies */
4255 static struct cftype cgroup_legacy_base_files[] = {
4257 .name = "cgroup.procs",
4258 .seq_start = cgroup_pidlist_start,
4259 .seq_next = cgroup_pidlist_next,
4260 .seq_stop = cgroup_pidlist_stop,
4261 .seq_show = cgroup_pidlist_show,
4262 .private = CGROUP_FILE_PROCS,
4263 .write = cgroup_procs_write,
4264 .mode = S_IRUGO | S_IWUSR,
4267 .name = "cgroup.clone_children",
4268 .read_u64 = cgroup_clone_children_read,
4269 .write_u64 = cgroup_clone_children_write,
4272 .name = "cgroup.sane_behavior",
4273 .flags = CFTYPE_ONLY_ON_ROOT,
4274 .seq_show = cgroup_sane_behavior_show,
4278 .seq_start = cgroup_pidlist_start,
4279 .seq_next = cgroup_pidlist_next,
4280 .seq_stop = cgroup_pidlist_stop,
4281 .seq_show = cgroup_pidlist_show,
4282 .private = CGROUP_FILE_TASKS,
4283 .write = cgroup_tasks_write,
4284 .mode = S_IRUGO | S_IWUSR,
4287 .name = "notify_on_release",
4288 .read_u64 = cgroup_read_notify_on_release,
4289 .write_u64 = cgroup_write_notify_on_release,
4292 .name = "release_agent",
4293 .flags = CFTYPE_ONLY_ON_ROOT,
4294 .seq_show = cgroup_release_agent_show,
4295 .write = cgroup_release_agent_write,
4296 .max_write_len = PATH_MAX - 1,
4302 * cgroup_populate_dir - create subsys files in a cgroup directory
4303 * @cgrp: target cgroup
4304 * @subsys_mask: mask of the subsystem ids whose files should be added
4306 * On failure, no file is added.
4308 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4310 struct cgroup_subsys *ss;
4313 /* process cftsets of each subsystem */
4314 for_each_subsys(ss, i) {
4315 struct cftype *cfts;
4317 if (!(subsys_mask & (1 << i)))
4320 list_for_each_entry(cfts, &ss->cfts, node) {
4321 ret = cgroup_addrm_files(cgrp, cfts, true);
4328 cgroup_clear_dir(cgrp, subsys_mask);
4333 * css destruction is four-stage process.
4335 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4336 * Implemented in kill_css().
4338 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4339 * and thus css_tryget_online() is guaranteed to fail, the css can be
4340 * offlined by invoking offline_css(). After offlining, the base ref is
4341 * put. Implemented in css_killed_work_fn().
4343 * 3. When the percpu_ref reaches zero, the only possible remaining
4344 * accessors are inside RCU read sections. css_release() schedules the
4347 * 4. After the grace period, the css can be freed. Implemented in
4348 * css_free_work_fn().
4350 * It is actually hairier because both step 2 and 4 require process context
4351 * and thus involve punting to css->destroy_work adding two additional
4352 * steps to the already complex sequence.
4354 static void css_free_work_fn(struct work_struct *work)
4356 struct cgroup_subsys_state *css =
4357 container_of(work, struct cgroup_subsys_state, destroy_work);
4358 struct cgroup_subsys *ss = css->ss;
4359 struct cgroup *cgrp = css->cgroup;
4361 percpu_ref_exit(&css->refcnt);
4368 css_put(css->parent);
4371 cgroup_idr_remove(&ss->css_idr, id);
4374 /* cgroup free path */
4375 atomic_dec(&cgrp->root->nr_cgrps);
4376 cgroup_pidlist_destroy_all(cgrp);
4377 cancel_work_sync(&cgrp->release_agent_work);
4379 if (cgroup_parent(cgrp)) {
4381 * We get a ref to the parent, and put the ref when
4382 * this cgroup is being freed, so it's guaranteed
4383 * that the parent won't be destroyed before its
4386 cgroup_put(cgroup_parent(cgrp));
4387 kernfs_put(cgrp->kn);
4391 * This is root cgroup's refcnt reaching zero,
4392 * which indicates that the root should be
4395 cgroup_destroy_root(cgrp->root);
4400 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4402 struct cgroup_subsys_state *css =
4403 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4405 INIT_WORK(&css->destroy_work, css_free_work_fn);
4406 queue_work(cgroup_destroy_wq, &css->destroy_work);
4409 static void css_release_work_fn(struct work_struct *work)
4411 struct cgroup_subsys_state *css =
4412 container_of(work, struct cgroup_subsys_state, destroy_work);
4413 struct cgroup_subsys *ss = css->ss;
4414 struct cgroup *cgrp = css->cgroup;
4416 mutex_lock(&cgroup_mutex);
4418 css->flags |= CSS_RELEASED;
4419 list_del_rcu(&css->sibling);
4422 /* css release path */
4423 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4424 if (ss->css_released)
4425 ss->css_released(css);
4427 /* cgroup release path */
4428 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4432 * There are two control paths which try to determine
4433 * cgroup from dentry without going through kernfs -
4434 * cgroupstats_build() and css_tryget_online_from_dir().
4435 * Those are supported by RCU protecting clearing of
4436 * cgrp->kn->priv backpointer.
4438 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4441 mutex_unlock(&cgroup_mutex);
4443 call_rcu(&css->rcu_head, css_free_rcu_fn);
4446 static void css_release(struct percpu_ref *ref)
4448 struct cgroup_subsys_state *css =
4449 container_of(ref, struct cgroup_subsys_state, refcnt);
4451 INIT_WORK(&css->destroy_work, css_release_work_fn);
4452 queue_work(cgroup_destroy_wq, &css->destroy_work);
4455 static void init_and_link_css(struct cgroup_subsys_state *css,
4456 struct cgroup_subsys *ss, struct cgroup *cgrp)
4458 lockdep_assert_held(&cgroup_mutex);
4462 memset(css, 0, sizeof(*css));
4465 INIT_LIST_HEAD(&css->sibling);
4466 INIT_LIST_HEAD(&css->children);
4467 css->serial_nr = css_serial_nr_next++;
4469 if (cgroup_parent(cgrp)) {
4470 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4471 css_get(css->parent);
4474 BUG_ON(cgroup_css(cgrp, ss));
4477 /* invoke ->css_online() on a new CSS and mark it online if successful */
4478 static int online_css(struct cgroup_subsys_state *css)
4480 struct cgroup_subsys *ss = css->ss;
4483 lockdep_assert_held(&cgroup_mutex);
4486 ret = ss->css_online(css);
4488 css->flags |= CSS_ONLINE;
4489 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4494 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4495 static void offline_css(struct cgroup_subsys_state *css)
4497 struct cgroup_subsys *ss = css->ss;
4499 lockdep_assert_held(&cgroup_mutex);
4501 if (!(css->flags & CSS_ONLINE))
4504 if (ss->css_offline)
4505 ss->css_offline(css);
4507 css->flags &= ~CSS_ONLINE;
4508 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4510 wake_up_all(&css->cgroup->offline_waitq);
4514 * create_css - create a cgroup_subsys_state
4515 * @cgrp: the cgroup new css will be associated with
4516 * @ss: the subsys of new css
4517 * @visible: whether to create control knobs for the new css or not
4519 * Create a new css associated with @cgrp - @ss pair. On success, the new
4520 * css is online and installed in @cgrp with all interface files created if
4521 * @visible. Returns 0 on success, -errno on failure.
4523 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4526 struct cgroup *parent = cgroup_parent(cgrp);
4527 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4528 struct cgroup_subsys_state *css;
4531 lockdep_assert_held(&cgroup_mutex);
4533 css = ss->css_alloc(parent_css);
4535 return PTR_ERR(css);
4537 init_and_link_css(css, ss, cgrp);
4539 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4543 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4545 goto err_free_percpu_ref;
4549 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4554 /* @css is ready to be brought online now, make it visible */
4555 list_add_tail_rcu(&css->sibling, &parent_css->children);
4556 cgroup_idr_replace(&ss->css_idr, css, css->id);
4558 err = online_css(css);
4562 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4563 cgroup_parent(parent)) {
4564 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4565 current->comm, current->pid, ss->name);
4566 if (!strcmp(ss->name, "memory"))
4567 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4568 ss->warned_broken_hierarchy = true;
4574 list_del_rcu(&css->sibling);
4575 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4577 cgroup_idr_remove(&ss->css_idr, css->id);
4578 err_free_percpu_ref:
4579 percpu_ref_exit(&css->refcnt);
4581 call_rcu(&css->rcu_head, css_free_rcu_fn);
4585 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4588 struct cgroup *parent, *cgrp;
4589 struct cgroup_root *root;
4590 struct cgroup_subsys *ss;
4591 struct kernfs_node *kn;
4592 struct cftype *base_files;
4595 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4597 if (strchr(name, '\n'))
4600 parent = cgroup_kn_lock_live(parent_kn);
4603 root = parent->root;
4605 /* allocate the cgroup and its ID, 0 is reserved for the root */
4606 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4612 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4617 * Temporarily set the pointer to NULL, so idr_find() won't return
4618 * a half-baked cgroup.
4620 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4623 goto out_cancel_ref;
4626 init_cgroup_housekeeping(cgrp);
4628 cgrp->self.parent = &parent->self;
4631 if (notify_on_release(parent))
4632 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4634 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4635 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4637 /* create the directory */
4638 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4646 * This extra ref will be put in cgroup_free_fn() and guarantees
4647 * that @cgrp->kn is always accessible.
4651 cgrp->self.serial_nr = css_serial_nr_next++;
4653 /* allocation complete, commit to creation */
4654 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4655 atomic_inc(&root->nr_cgrps);
4659 * @cgrp is now fully operational. If something fails after this
4660 * point, it'll be released via the normal destruction path.
4662 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4664 ret = cgroup_kn_set_ugid(kn);
4668 if (cgroup_on_dfl(cgrp))
4669 base_files = cgroup_dfl_base_files;
4671 base_files = cgroup_legacy_base_files;
4673 ret = cgroup_addrm_files(cgrp, base_files, true);
4677 /* let's create and online css's */
4678 for_each_subsys(ss, ssid) {
4679 if (parent->child_subsys_mask & (1 << ssid)) {
4680 ret = create_css(cgrp, ss,
4681 parent->subtree_control & (1 << ssid));
4688 * On the default hierarchy, a child doesn't automatically inherit
4689 * subtree_control from the parent. Each is configured manually.
4691 if (!cgroup_on_dfl(cgrp)) {
4692 cgrp->subtree_control = parent->subtree_control;
4693 cgroup_refresh_child_subsys_mask(cgrp);
4696 kernfs_activate(kn);
4702 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4704 percpu_ref_exit(&cgrp->self.refcnt);
4708 cgroup_kn_unlock(parent_kn);
4712 cgroup_destroy_locked(cgrp);
4717 * This is called when the refcnt of a css is confirmed to be killed.
4718 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4719 * initate destruction and put the css ref from kill_css().
4721 static void css_killed_work_fn(struct work_struct *work)
4723 struct cgroup_subsys_state *css =
4724 container_of(work, struct cgroup_subsys_state, destroy_work);
4726 mutex_lock(&cgroup_mutex);
4728 mutex_unlock(&cgroup_mutex);
4733 /* css kill confirmation processing requires process context, bounce */
4734 static void css_killed_ref_fn(struct percpu_ref *ref)
4736 struct cgroup_subsys_state *css =
4737 container_of(ref, struct cgroup_subsys_state, refcnt);
4739 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4740 queue_work(cgroup_destroy_wq, &css->destroy_work);
4744 * kill_css - destroy a css
4745 * @css: css to destroy
4747 * This function initiates destruction of @css by removing cgroup interface
4748 * files and putting its base reference. ->css_offline() will be invoked
4749 * asynchronously once css_tryget_online() is guaranteed to fail and when
4750 * the reference count reaches zero, @css will be released.
4752 static void kill_css(struct cgroup_subsys_state *css)
4754 lockdep_assert_held(&cgroup_mutex);
4757 * This must happen before css is disassociated with its cgroup.
4758 * See seq_css() for details.
4760 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4763 * Killing would put the base ref, but we need to keep it alive
4764 * until after ->css_offline().
4769 * cgroup core guarantees that, by the time ->css_offline() is
4770 * invoked, no new css reference will be given out via
4771 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4772 * proceed to offlining css's because percpu_ref_kill() doesn't
4773 * guarantee that the ref is seen as killed on all CPUs on return.
4775 * Use percpu_ref_kill_and_confirm() to get notifications as each
4776 * css is confirmed to be seen as killed on all CPUs.
4778 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4782 * cgroup_destroy_locked - the first stage of cgroup destruction
4783 * @cgrp: cgroup to be destroyed
4785 * css's make use of percpu refcnts whose killing latency shouldn't be
4786 * exposed to userland and are RCU protected. Also, cgroup core needs to
4787 * guarantee that css_tryget_online() won't succeed by the time
4788 * ->css_offline() is invoked. To satisfy all the requirements,
4789 * destruction is implemented in the following two steps.
4791 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4792 * userland visible parts and start killing the percpu refcnts of
4793 * css's. Set up so that the next stage will be kicked off once all
4794 * the percpu refcnts are confirmed to be killed.
4796 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4797 * rest of destruction. Once all cgroup references are gone, the
4798 * cgroup is RCU-freed.
4800 * This function implements s1. After this step, @cgrp is gone as far as
4801 * the userland is concerned and a new cgroup with the same name may be
4802 * created. As cgroup doesn't care about the names internally, this
4803 * doesn't cause any problem.
4805 static int cgroup_destroy_locked(struct cgroup *cgrp)
4806 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4808 struct cgroup_subsys_state *css;
4812 lockdep_assert_held(&cgroup_mutex);
4815 * css_set_rwsem synchronizes access to ->cset_links and prevents
4816 * @cgrp from being removed while put_css_set() is in progress.
4818 down_read(&css_set_rwsem);
4819 empty = list_empty(&cgrp->cset_links);
4820 up_read(&css_set_rwsem);
4825 * Make sure there's no live children. We can't test emptiness of
4826 * ->self.children as dead children linger on it while being
4827 * drained; otherwise, "rmdir parent/child parent" may fail.
4829 if (css_has_online_children(&cgrp->self))
4833 * Mark @cgrp dead. This prevents further task migration and child
4834 * creation by disabling cgroup_lock_live_group().
4836 cgrp->self.flags &= ~CSS_ONLINE;
4838 /* initiate massacre of all css's */
4839 for_each_css(css, ssid, cgrp)
4843 * Remove @cgrp directory along with the base files. @cgrp has an
4844 * extra ref on its kn.
4846 kernfs_remove(cgrp->kn);
4848 check_for_release(cgroup_parent(cgrp));
4850 /* put the base reference */
4851 percpu_ref_kill(&cgrp->self.refcnt);
4856 static int cgroup_rmdir(struct kernfs_node *kn)
4858 struct cgroup *cgrp;
4861 cgrp = cgroup_kn_lock_live(kn);
4865 ret = cgroup_destroy_locked(cgrp);
4867 cgroup_kn_unlock(kn);
4871 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4872 .remount_fs = cgroup_remount,
4873 .show_options = cgroup_show_options,
4874 .mkdir = cgroup_mkdir,
4875 .rmdir = cgroup_rmdir,
4876 .rename = cgroup_rename,
4879 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4881 struct cgroup_subsys_state *css;
4883 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4885 mutex_lock(&cgroup_mutex);
4887 idr_init(&ss->css_idr);
4888 INIT_LIST_HEAD(&ss->cfts);
4890 /* Create the root cgroup state for this subsystem */
4891 ss->root = &cgrp_dfl_root;
4892 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4893 /* We don't handle early failures gracefully */
4894 BUG_ON(IS_ERR(css));
4895 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4898 * Root csses are never destroyed and we can't initialize
4899 * percpu_ref during early init. Disable refcnting.
4901 css->flags |= CSS_NO_REF;
4904 /* allocation can't be done safely during early init */
4907 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4908 BUG_ON(css->id < 0);
4911 /* Update the init_css_set to contain a subsys
4912 * pointer to this state - since the subsystem is
4913 * newly registered, all tasks and hence the
4914 * init_css_set is in the subsystem's root cgroup. */
4915 init_css_set.subsys[ss->id] = css;
4917 need_forkexit_callback |= ss->fork || ss->exit;
4919 /* At system boot, before all subsystems have been
4920 * registered, no tasks have been forked, so we don't
4921 * need to invoke fork callbacks here. */
4922 BUG_ON(!list_empty(&init_task.tasks));
4924 BUG_ON(online_css(css));
4926 mutex_unlock(&cgroup_mutex);
4930 * cgroup_init_early - cgroup initialization at system boot
4932 * Initialize cgroups at system boot, and initialize any
4933 * subsystems that request early init.
4935 int __init cgroup_init_early(void)
4937 static struct cgroup_sb_opts __initdata opts;
4938 struct cgroup_subsys *ss;
4941 init_cgroup_root(&cgrp_dfl_root, &opts);
4942 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4944 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4946 for_each_subsys(ss, i) {
4947 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4948 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4949 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4951 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4952 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4955 ss->name = cgroup_subsys_name[i];
4958 cgroup_init_subsys(ss, true);
4964 * cgroup_init - cgroup initialization
4966 * Register cgroup filesystem and /proc file, and initialize
4967 * any subsystems that didn't request early init.
4969 int __init cgroup_init(void)
4971 struct cgroup_subsys *ss;
4975 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
4976 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
4977 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
4979 mutex_lock(&cgroup_mutex);
4981 /* Add init_css_set to the hash table */
4982 key = css_set_hash(init_css_set.subsys);
4983 hash_add(css_set_table, &init_css_set.hlist, key);
4985 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4987 mutex_unlock(&cgroup_mutex);
4989 for_each_subsys(ss, ssid) {
4990 if (ss->early_init) {
4991 struct cgroup_subsys_state *css =
4992 init_css_set.subsys[ss->id];
4994 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4996 BUG_ON(css->id < 0);
4998 cgroup_init_subsys(ss, false);
5001 list_add_tail(&init_css_set.e_cset_node[ssid],
5002 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5005 * Setting dfl_root subsys_mask needs to consider the
5006 * disabled flag and cftype registration needs kmalloc,
5007 * both of which aren't available during early_init.
5012 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5014 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5015 ss->dfl_cftypes = ss->legacy_cftypes;
5017 if (!ss->dfl_cftypes)
5018 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5020 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5021 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5023 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5024 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5028 ss->bind(init_css_set.subsys[ssid]);
5031 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
5035 err = register_filesystem(&cgroup_fs_type);
5037 kobject_put(cgroup_kobj);
5041 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5045 static int __init cgroup_wq_init(void)
5048 * There isn't much point in executing destruction path in
5049 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5050 * Use 1 for @max_active.
5052 * We would prefer to do this in cgroup_init() above, but that
5053 * is called before init_workqueues(): so leave this until after.
5055 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5056 BUG_ON(!cgroup_destroy_wq);
5059 * Used to destroy pidlists and separate to serve as flush domain.
5060 * Cap @max_active to 1 too.
5062 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5064 BUG_ON(!cgroup_pidlist_destroy_wq);
5068 core_initcall(cgroup_wq_init);
5071 * proc_cgroup_show()
5072 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5073 * - Used for /proc/<pid>/cgroup.
5075 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5076 struct pid *pid, struct task_struct *tsk)
5080 struct cgroup_root *root;
5083 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5087 mutex_lock(&cgroup_mutex);
5088 down_read(&css_set_rwsem);
5090 for_each_root(root) {
5091 struct cgroup_subsys *ss;
5092 struct cgroup *cgrp;
5093 int ssid, count = 0;
5095 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5098 seq_printf(m, "%d:", root->hierarchy_id);
5099 for_each_subsys(ss, ssid)
5100 if (root->subsys_mask & (1 << ssid))
5101 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5102 if (strlen(root->name))
5103 seq_printf(m, "%sname=%s", count ? "," : "",
5106 cgrp = task_cgroup_from_root(tsk, root);
5107 path = cgroup_path(cgrp, buf, PATH_MAX);
5109 retval = -ENAMETOOLONG;
5118 up_read(&css_set_rwsem);
5119 mutex_unlock(&cgroup_mutex);
5125 /* Display information about each subsystem and each hierarchy */
5126 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5128 struct cgroup_subsys *ss;
5131 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5133 * ideally we don't want subsystems moving around while we do this.
5134 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5135 * subsys/hierarchy state.
5137 mutex_lock(&cgroup_mutex);
5139 for_each_subsys(ss, i)
5140 seq_printf(m, "%s\t%d\t%d\t%d\n",
5141 ss->name, ss->root->hierarchy_id,
5142 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5144 mutex_unlock(&cgroup_mutex);
5148 static int cgroupstats_open(struct inode *inode, struct file *file)
5150 return single_open(file, proc_cgroupstats_show, NULL);
5153 static const struct file_operations proc_cgroupstats_operations = {
5154 .open = cgroupstats_open,
5156 .llseek = seq_lseek,
5157 .release = single_release,
5161 * cgroup_fork - initialize cgroup related fields during copy_process()
5162 * @child: pointer to task_struct of forking parent process.
5164 * A task is associated with the init_css_set until cgroup_post_fork()
5165 * attaches it to the parent's css_set. Empty cg_list indicates that
5166 * @child isn't holding reference to its css_set.
5168 void cgroup_fork(struct task_struct *child)
5170 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5171 INIT_LIST_HEAD(&child->cg_list);
5175 * cgroup_post_fork - called on a new task after adding it to the task list
5176 * @child: the task in question
5178 * Adds the task to the list running through its css_set if necessary and
5179 * call the subsystem fork() callbacks. Has to be after the task is
5180 * visible on the task list in case we race with the first call to
5181 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5184 void cgroup_post_fork(struct task_struct *child)
5186 struct cgroup_subsys *ss;
5190 * This may race against cgroup_enable_task_cg_lists(). As that
5191 * function sets use_task_css_set_links before grabbing
5192 * tasklist_lock and we just went through tasklist_lock to add
5193 * @child, it's guaranteed that either we see the set
5194 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5195 * @child during its iteration.
5197 * If we won the race, @child is associated with %current's
5198 * css_set. Grabbing css_set_rwsem guarantees both that the
5199 * association is stable, and, on completion of the parent's
5200 * migration, @child is visible in the source of migration or
5201 * already in the destination cgroup. This guarantee is necessary
5202 * when implementing operations which need to migrate all tasks of
5203 * a cgroup to another.
5205 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5206 * will remain in init_css_set. This is safe because all tasks are
5207 * in the init_css_set before cg_links is enabled and there's no
5208 * operation which transfers all tasks out of init_css_set.
5210 if (use_task_css_set_links) {
5211 struct css_set *cset;
5213 down_write(&css_set_rwsem);
5214 cset = task_css_set(current);
5215 if (list_empty(&child->cg_list)) {
5216 rcu_assign_pointer(child->cgroups, cset);
5217 list_add(&child->cg_list, &cset->tasks);
5220 up_write(&css_set_rwsem);
5224 * Call ss->fork(). This must happen after @child is linked on
5225 * css_set; otherwise, @child might change state between ->fork()
5226 * and addition to css_set.
5228 if (need_forkexit_callback) {
5229 for_each_subsys(ss, i)
5236 * cgroup_exit - detach cgroup from exiting task
5237 * @tsk: pointer to task_struct of exiting process
5239 * Description: Detach cgroup from @tsk and release it.
5241 * Note that cgroups marked notify_on_release force every task in
5242 * them to take the global cgroup_mutex mutex when exiting.
5243 * This could impact scaling on very large systems. Be reluctant to
5244 * use notify_on_release cgroups where very high task exit scaling
5245 * is required on large systems.
5247 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5248 * call cgroup_exit() while the task is still competent to handle
5249 * notify_on_release(), then leave the task attached to the root cgroup in
5250 * each hierarchy for the remainder of its exit. No need to bother with
5251 * init_css_set refcnting. init_css_set never goes away and we can't race
5252 * with migration path - PF_EXITING is visible to migration path.
5254 void cgroup_exit(struct task_struct *tsk)
5256 struct cgroup_subsys *ss;
5257 struct css_set *cset;
5258 bool put_cset = false;
5262 * Unlink from @tsk from its css_set. As migration path can't race
5263 * with us, we can check cg_list without grabbing css_set_rwsem.
5265 if (!list_empty(&tsk->cg_list)) {
5266 down_write(&css_set_rwsem);
5267 list_del_init(&tsk->cg_list);
5268 up_write(&css_set_rwsem);
5272 /* Reassign the task to the init_css_set. */
5273 cset = task_css_set(tsk);
5274 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5276 if (need_forkexit_callback) {
5277 /* see cgroup_post_fork() for details */
5278 for_each_subsys(ss, i) {
5280 struct cgroup_subsys_state *old_css = cset->subsys[i];
5281 struct cgroup_subsys_state *css = task_css(tsk, i);
5283 ss->exit(css, old_css, tsk);
5292 static void check_for_release(struct cgroup *cgrp)
5294 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5295 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5296 schedule_work(&cgrp->release_agent_work);
5300 * Notify userspace when a cgroup is released, by running the
5301 * configured release agent with the name of the cgroup (path
5302 * relative to the root of cgroup file system) as the argument.
5304 * Most likely, this user command will try to rmdir this cgroup.
5306 * This races with the possibility that some other task will be
5307 * attached to this cgroup before it is removed, or that some other
5308 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5309 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5310 * unused, and this cgroup will be reprieved from its death sentence,
5311 * to continue to serve a useful existence. Next time it's released,
5312 * we will get notified again, if it still has 'notify_on_release' set.
5314 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5315 * means only wait until the task is successfully execve()'d. The
5316 * separate release agent task is forked by call_usermodehelper(),
5317 * then control in this thread returns here, without waiting for the
5318 * release agent task. We don't bother to wait because the caller of
5319 * this routine has no use for the exit status of the release agent
5320 * task, so no sense holding our caller up for that.
5322 static void cgroup_release_agent(struct work_struct *work)
5324 struct cgroup *cgrp =
5325 container_of(work, struct cgroup, release_agent_work);
5326 char *pathbuf = NULL, *agentbuf = NULL, *path;
5327 char *argv[3], *envp[3];
5329 mutex_lock(&cgroup_mutex);
5331 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5332 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5333 if (!pathbuf || !agentbuf)
5336 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5344 /* minimal command environment */
5346 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5349 mutex_unlock(&cgroup_mutex);
5350 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5353 mutex_unlock(&cgroup_mutex);
5359 static int __init cgroup_disable(char *str)
5361 struct cgroup_subsys *ss;
5365 while ((token = strsep(&str, ",")) != NULL) {
5369 for_each_subsys(ss, i) {
5370 if (!strcmp(token, ss->name)) {
5372 printk(KERN_INFO "Disabling %s control group"
5373 " subsystem\n", ss->name);
5380 __setup("cgroup_disable=", cgroup_disable);
5382 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5384 printk("cgroup: using legacy files on the default hierarchy\n");
5385 cgroup_legacy_files_on_dfl = true;
5388 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5391 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5392 * @dentry: directory dentry of interest
5393 * @ss: subsystem of interest
5395 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5396 * to get the corresponding css and return it. If such css doesn't exist
5397 * or can't be pinned, an ERR_PTR value is returned.
5399 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5400 struct cgroup_subsys *ss)
5402 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5403 struct cgroup_subsys_state *css = NULL;
5404 struct cgroup *cgrp;
5406 /* is @dentry a cgroup dir? */
5407 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5408 kernfs_type(kn) != KERNFS_DIR)
5409 return ERR_PTR(-EBADF);
5414 * This path doesn't originate from kernfs and @kn could already
5415 * have been or be removed at any point. @kn->priv is RCU
5416 * protected for this access. See css_release_work_fn() for details.
5418 cgrp = rcu_dereference(kn->priv);
5420 css = cgroup_css(cgrp, ss);
5422 if (!css || !css_tryget_online(css))
5423 css = ERR_PTR(-ENOENT);
5430 * css_from_id - lookup css by id
5431 * @id: the cgroup id
5432 * @ss: cgroup subsys to be looked into
5434 * Returns the css if there's valid one with @id, otherwise returns NULL.
5435 * Should be called under rcu_read_lock().
5437 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5439 WARN_ON_ONCE(!rcu_read_lock_held());
5440 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5443 #ifdef CONFIG_CGROUP_DEBUG
5444 static struct cgroup_subsys_state *
5445 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5447 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5450 return ERR_PTR(-ENOMEM);
5455 static void debug_css_free(struct cgroup_subsys_state *css)
5460 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5463 return cgroup_task_count(css->cgroup);
5466 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5469 return (u64)(unsigned long)current->cgroups;
5472 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5478 count = atomic_read(&task_css_set(current)->refcount);
5483 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5485 struct cgrp_cset_link *link;
5486 struct css_set *cset;
5489 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5493 down_read(&css_set_rwsem);
5495 cset = rcu_dereference(current->cgroups);
5496 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5497 struct cgroup *c = link->cgrp;
5499 cgroup_name(c, name_buf, NAME_MAX + 1);
5500 seq_printf(seq, "Root %d group %s\n",
5501 c->root->hierarchy_id, name_buf);
5504 up_read(&css_set_rwsem);
5509 #define MAX_TASKS_SHOWN_PER_CSS 25
5510 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5512 struct cgroup_subsys_state *css = seq_css(seq);
5513 struct cgrp_cset_link *link;
5515 down_read(&css_set_rwsem);
5516 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5517 struct css_set *cset = link->cset;
5518 struct task_struct *task;
5521 seq_printf(seq, "css_set %p\n", cset);
5523 list_for_each_entry(task, &cset->tasks, cg_list) {
5524 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5526 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5529 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5530 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5532 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5536 seq_puts(seq, " ...\n");
5538 up_read(&css_set_rwsem);
5542 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5544 return (!cgroup_has_tasks(css->cgroup) &&
5545 !css_has_online_children(&css->cgroup->self));
5548 static struct cftype debug_files[] = {
5550 .name = "taskcount",
5551 .read_u64 = debug_taskcount_read,
5555 .name = "current_css_set",
5556 .read_u64 = current_css_set_read,
5560 .name = "current_css_set_refcount",
5561 .read_u64 = current_css_set_refcount_read,
5565 .name = "current_css_set_cg_links",
5566 .seq_show = current_css_set_cg_links_read,
5570 .name = "cgroup_css_links",
5571 .seq_show = cgroup_css_links_read,
5575 .name = "releasable",
5576 .read_u64 = releasable_read,
5582 struct cgroup_subsys debug_cgrp_subsys = {
5583 .css_alloc = debug_css_alloc,
5584 .css_free = debug_css_free,
5585 .legacy_cftypes = debug_files,
5587 #endif /* CONFIG_CGROUP_DEBUG */