2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 DECLARE_RWSEM(css_set_rwsem);
87 EXPORT_SYMBOL_GPL(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(css_set_rwsem);
90 static DEFINE_MUTEX(cgroup_mutex);
91 static DECLARE_RWSEM(css_set_rwsem);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
123 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys *cgroup_subsys[] = {
128 #include <linux/cgroup_subsys.h>
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name[] = {
135 #include <linux/cgroup_subsys.h>
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
144 struct cgroup_root cgrp_dfl_root;
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
150 static bool cgrp_dfl_root_visible;
153 * Set by the boot param of the same name and makes subsystems with NULL
154 * ->dfl_files to use ->legacy_files on the default hierarchy.
156 static bool cgroup_legacy_files_on_dfl;
158 /* some controllers are not supported in the default hierarchy */
159 static unsigned int cgrp_dfl_root_inhibit_ss_mask;
161 /* The list of hierarchy roots */
163 static LIST_HEAD(cgroup_roots);
164 static int cgroup_root_count;
166 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
167 static DEFINE_IDR(cgroup_hierarchy_idr);
170 * Assign a monotonically increasing serial number to csses. It guarantees
171 * cgroups with bigger numbers are newer than those with smaller numbers.
172 * Also, as csses are always appended to the parent's ->children list, it
173 * guarantees that sibling csses are always sorted in the ascending serial
174 * number order on the list. Protected by cgroup_mutex.
176 static u64 css_serial_nr_next = 1;
178 /* This flag indicates whether tasks in the fork and exit paths should
179 * check for fork/exit handlers to call. This avoids us having to do
180 * extra work in the fork/exit path if none of the subsystems need to
183 static int need_forkexit_callback __read_mostly;
185 static struct cftype cgroup_dfl_base_files[];
186 static struct cftype cgroup_legacy_base_files[];
188 static int rebind_subsystems(struct cgroup_root *dst_root,
189 unsigned int ss_mask);
190 static int cgroup_destroy_locked(struct cgroup *cgrp);
191 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
193 static void css_release(struct percpu_ref *ref);
194 static void kill_css(struct cgroup_subsys_state *css);
195 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
198 /* IDR wrappers which synchronize using cgroup_idr_lock */
199 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
204 idr_preload(gfp_mask);
205 spin_lock_bh(&cgroup_idr_lock);
206 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
207 spin_unlock_bh(&cgroup_idr_lock);
212 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
216 spin_lock_bh(&cgroup_idr_lock);
217 ret = idr_replace(idr, ptr, id);
218 spin_unlock_bh(&cgroup_idr_lock);
222 static void cgroup_idr_remove(struct idr *idr, int id)
224 spin_lock_bh(&cgroup_idr_lock);
226 spin_unlock_bh(&cgroup_idr_lock);
229 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
231 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
234 return container_of(parent_css, struct cgroup, self);
239 * cgroup_css - obtain a cgroup's css for the specified subsystem
240 * @cgrp: the cgroup of interest
241 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
243 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
244 * function must be called either under cgroup_mutex or rcu_read_lock() and
245 * the caller is responsible for pinning the returned css if it wants to
246 * keep accessing it outside the said locks. This function may return
247 * %NULL if @cgrp doesn't have @subsys_id enabled.
249 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
250 struct cgroup_subsys *ss)
253 return rcu_dereference_check(cgrp->subsys[ss->id],
254 lockdep_is_held(&cgroup_mutex));
260 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
261 * @cgrp: the cgroup of interest
262 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
264 * Similar to cgroup_css() but returns the effctive css, which is defined
265 * as the matching css of the nearest ancestor including self which has @ss
266 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
267 * function is guaranteed to return non-NULL css.
269 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
270 struct cgroup_subsys *ss)
272 lockdep_assert_held(&cgroup_mutex);
277 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
280 while (cgroup_parent(cgrp) &&
281 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
282 cgrp = cgroup_parent(cgrp);
284 return cgroup_css(cgrp, ss);
287 /* convenient tests for these bits */
288 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
290 return !(cgrp->self.flags & CSS_ONLINE);
293 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
295 struct cgroup *cgrp = of->kn->parent->priv;
296 struct cftype *cft = of_cft(of);
299 * This is open and unprotected implementation of cgroup_css().
300 * seq_css() is only called from a kernfs file operation which has
301 * an active reference on the file. Because all the subsystem
302 * files are drained before a css is disassociated with a cgroup,
303 * the matching css from the cgroup's subsys table is guaranteed to
304 * be and stay valid until the enclosing operation is complete.
307 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
311 EXPORT_SYMBOL_GPL(of_css);
314 * cgroup_is_descendant - test ancestry
315 * @cgrp: the cgroup to be tested
316 * @ancestor: possible ancestor of @cgrp
318 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
319 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
320 * and @ancestor are accessible.
322 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
325 if (cgrp == ancestor)
327 cgrp = cgroup_parent(cgrp);
332 static int notify_on_release(const struct cgroup *cgrp)
334 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
338 * for_each_css - iterate all css's of a cgroup
339 * @css: the iteration cursor
340 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
341 * @cgrp: the target cgroup to iterate css's of
343 * Should be called under cgroup_[tree_]mutex.
345 #define for_each_css(css, ssid, cgrp) \
346 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
347 if (!((css) = rcu_dereference_check( \
348 (cgrp)->subsys[(ssid)], \
349 lockdep_is_held(&cgroup_mutex)))) { } \
353 * for_each_e_css - iterate all effective css's of a cgroup
354 * @css: the iteration cursor
355 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
356 * @cgrp: the target cgroup to iterate css's of
358 * Should be called under cgroup_[tree_]mutex.
360 #define for_each_e_css(css, ssid, cgrp) \
361 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
362 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
367 * for_each_subsys - iterate all enabled cgroup subsystems
368 * @ss: the iteration cursor
369 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
371 #define for_each_subsys(ss, ssid) \
372 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
373 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
375 /* iterate across the hierarchies */
376 #define for_each_root(root) \
377 list_for_each_entry((root), &cgroup_roots, root_list)
379 /* iterate over child cgrps, lock should be held throughout iteration */
380 #define cgroup_for_each_live_child(child, cgrp) \
381 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
382 if (({ lockdep_assert_held(&cgroup_mutex); \
383 cgroup_is_dead(child); })) \
387 static void cgroup_release_agent(struct work_struct *work);
388 static void check_for_release(struct cgroup *cgrp);
391 * A cgroup can be associated with multiple css_sets as different tasks may
392 * belong to different cgroups on different hierarchies. In the other
393 * direction, a css_set is naturally associated with multiple cgroups.
394 * This M:N relationship is represented by the following link structure
395 * which exists for each association and allows traversing the associations
398 struct cgrp_cset_link {
399 /* the cgroup and css_set this link associates */
401 struct css_set *cset;
403 /* list of cgrp_cset_links anchored at cgrp->cset_links */
404 struct list_head cset_link;
406 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
407 struct list_head cgrp_link;
411 * The default css_set - used by init and its children prior to any
412 * hierarchies being mounted. It contains a pointer to the root state
413 * for each subsystem. Also used to anchor the list of css_sets. Not
414 * reference-counted, to improve performance when child cgroups
415 * haven't been created.
417 struct css_set init_css_set = {
418 .refcount = ATOMIC_INIT(1),
419 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
420 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
421 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
422 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
423 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
426 static int css_set_count = 1; /* 1 for init_css_set */
429 * cgroup_update_populated - updated populated count of a cgroup
430 * @cgrp: the target cgroup
431 * @populated: inc or dec populated count
433 * @cgrp is either getting the first task (css_set) or losing the last.
434 * Update @cgrp->populated_cnt accordingly. The count is propagated
435 * towards root so that a given cgroup's populated_cnt is zero iff the
436 * cgroup and all its descendants are empty.
438 * @cgrp's interface file "cgroup.populated" is zero if
439 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
440 * changes from or to zero, userland is notified that the content of the
441 * interface file has changed. This can be used to detect when @cgrp and
442 * its descendants become populated or empty.
444 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
446 lockdep_assert_held(&css_set_rwsem);
452 trigger = !cgrp->populated_cnt++;
454 trigger = !--cgrp->populated_cnt;
459 if (cgrp->populated_kn)
460 kernfs_notify(cgrp->populated_kn);
461 cgrp = cgroup_parent(cgrp);
466 * hash table for cgroup groups. This improves the performance to find
467 * an existing css_set. This hash doesn't (currently) take into
468 * account cgroups in empty hierarchies.
470 #define CSS_SET_HASH_BITS 7
471 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
473 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
475 unsigned long key = 0UL;
476 struct cgroup_subsys *ss;
479 for_each_subsys(ss, i)
480 key += (unsigned long)css[i];
481 key = (key >> 16) ^ key;
486 static void put_css_set_locked(struct css_set *cset)
488 struct cgrp_cset_link *link, *tmp_link;
489 struct cgroup_subsys *ss;
492 lockdep_assert_held(&css_set_rwsem);
494 if (!atomic_dec_and_test(&cset->refcount))
497 /* This css_set is dead. unlink it and release cgroup refcounts */
498 for_each_subsys(ss, ssid)
499 list_del(&cset->e_cset_node[ssid]);
500 hash_del(&cset->hlist);
503 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
504 struct cgroup *cgrp = link->cgrp;
506 list_del(&link->cset_link);
507 list_del(&link->cgrp_link);
509 /* @cgrp can't go away while we're holding css_set_rwsem */
510 if (list_empty(&cgrp->cset_links)) {
511 cgroup_update_populated(cgrp, false);
512 check_for_release(cgrp);
518 kfree_rcu(cset, rcu_head);
521 static void put_css_set(struct css_set *cset)
524 * Ensure that the refcount doesn't hit zero while any readers
525 * can see it. Similar to atomic_dec_and_lock(), but for an
528 if (atomic_add_unless(&cset->refcount, -1, 1))
531 down_write(&css_set_rwsem);
532 put_css_set_locked(cset);
533 up_write(&css_set_rwsem);
537 * refcounted get/put for css_set objects
539 static inline void get_css_set(struct css_set *cset)
541 atomic_inc(&cset->refcount);
545 * compare_css_sets - helper function for find_existing_css_set().
546 * @cset: candidate css_set being tested
547 * @old_cset: existing css_set for a task
548 * @new_cgrp: cgroup that's being entered by the task
549 * @template: desired set of css pointers in css_set (pre-calculated)
551 * Returns true if "cset" matches "old_cset" except for the hierarchy
552 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
554 static bool compare_css_sets(struct css_set *cset,
555 struct css_set *old_cset,
556 struct cgroup *new_cgrp,
557 struct cgroup_subsys_state *template[])
559 struct list_head *l1, *l2;
562 * On the default hierarchy, there can be csets which are
563 * associated with the same set of cgroups but different csses.
564 * Let's first ensure that csses match.
566 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
570 * Compare cgroup pointers in order to distinguish between
571 * different cgroups in hierarchies. As different cgroups may
572 * share the same effective css, this comparison is always
575 l1 = &cset->cgrp_links;
576 l2 = &old_cset->cgrp_links;
578 struct cgrp_cset_link *link1, *link2;
579 struct cgroup *cgrp1, *cgrp2;
583 /* See if we reached the end - both lists are equal length. */
584 if (l1 == &cset->cgrp_links) {
585 BUG_ON(l2 != &old_cset->cgrp_links);
588 BUG_ON(l2 == &old_cset->cgrp_links);
590 /* Locate the cgroups associated with these links. */
591 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
592 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
595 /* Hierarchies should be linked in the same order. */
596 BUG_ON(cgrp1->root != cgrp2->root);
599 * If this hierarchy is the hierarchy of the cgroup
600 * that's changing, then we need to check that this
601 * css_set points to the new cgroup; if it's any other
602 * hierarchy, then this css_set should point to the
603 * same cgroup as the old css_set.
605 if (cgrp1->root == new_cgrp->root) {
606 if (cgrp1 != new_cgrp)
617 * find_existing_css_set - init css array and find the matching css_set
618 * @old_cset: the css_set that we're using before the cgroup transition
619 * @cgrp: the cgroup that we're moving into
620 * @template: out param for the new set of csses, should be clear on entry
622 static struct css_set *find_existing_css_set(struct css_set *old_cset,
624 struct cgroup_subsys_state *template[])
626 struct cgroup_root *root = cgrp->root;
627 struct cgroup_subsys *ss;
628 struct css_set *cset;
633 * Build the set of subsystem state objects that we want to see in the
634 * new css_set. while subsystems can change globally, the entries here
635 * won't change, so no need for locking.
637 for_each_subsys(ss, i) {
638 if (root->subsys_mask & (1UL << i)) {
640 * @ss is in this hierarchy, so we want the
641 * effective css from @cgrp.
643 template[i] = cgroup_e_css(cgrp, ss);
646 * @ss is not in this hierarchy, so we don't want
649 template[i] = old_cset->subsys[i];
653 key = css_set_hash(template);
654 hash_for_each_possible(css_set_table, cset, hlist, key) {
655 if (!compare_css_sets(cset, old_cset, cgrp, template))
658 /* This css_set matches what we need */
662 /* No existing cgroup group matched */
666 static void free_cgrp_cset_links(struct list_head *links_to_free)
668 struct cgrp_cset_link *link, *tmp_link;
670 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
671 list_del(&link->cset_link);
677 * allocate_cgrp_cset_links - allocate cgrp_cset_links
678 * @count: the number of links to allocate
679 * @tmp_links: list_head the allocated links are put on
681 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
682 * through ->cset_link. Returns 0 on success or -errno.
684 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
686 struct cgrp_cset_link *link;
689 INIT_LIST_HEAD(tmp_links);
691 for (i = 0; i < count; i++) {
692 link = kzalloc(sizeof(*link), GFP_KERNEL);
694 free_cgrp_cset_links(tmp_links);
697 list_add(&link->cset_link, tmp_links);
703 * link_css_set - a helper function to link a css_set to a cgroup
704 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
705 * @cset: the css_set to be linked
706 * @cgrp: the destination cgroup
708 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
711 struct cgrp_cset_link *link;
713 BUG_ON(list_empty(tmp_links));
715 if (cgroup_on_dfl(cgrp))
716 cset->dfl_cgrp = cgrp;
718 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
722 if (list_empty(&cgrp->cset_links))
723 cgroup_update_populated(cgrp, true);
724 list_move(&link->cset_link, &cgrp->cset_links);
727 * Always add links to the tail of the list so that the list
728 * is sorted by order of hierarchy creation
730 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
734 * find_css_set - return a new css_set with one cgroup updated
735 * @old_cset: the baseline css_set
736 * @cgrp: the cgroup to be updated
738 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
739 * substituted into the appropriate hierarchy.
741 static struct css_set *find_css_set(struct css_set *old_cset,
744 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
745 struct css_set *cset;
746 struct list_head tmp_links;
747 struct cgrp_cset_link *link;
748 struct cgroup_subsys *ss;
752 lockdep_assert_held(&cgroup_mutex);
754 /* First see if we already have a cgroup group that matches
756 down_read(&css_set_rwsem);
757 cset = find_existing_css_set(old_cset, cgrp, template);
760 up_read(&css_set_rwsem);
765 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
769 /* Allocate all the cgrp_cset_link objects that we'll need */
770 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
775 atomic_set(&cset->refcount, 1);
776 INIT_LIST_HEAD(&cset->cgrp_links);
777 INIT_LIST_HEAD(&cset->tasks);
778 INIT_LIST_HEAD(&cset->mg_tasks);
779 INIT_LIST_HEAD(&cset->mg_preload_node);
780 INIT_LIST_HEAD(&cset->mg_node);
781 INIT_HLIST_NODE(&cset->hlist);
783 /* Copy the set of subsystem state objects generated in
784 * find_existing_css_set() */
785 memcpy(cset->subsys, template, sizeof(cset->subsys));
787 down_write(&css_set_rwsem);
788 /* Add reference counts and links from the new css_set. */
789 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
790 struct cgroup *c = link->cgrp;
792 if (c->root == cgrp->root)
794 link_css_set(&tmp_links, cset, c);
797 BUG_ON(!list_empty(&tmp_links));
801 /* Add @cset to the hash table */
802 key = css_set_hash(cset->subsys);
803 hash_add(css_set_table, &cset->hlist, key);
805 for_each_subsys(ss, ssid)
806 list_add_tail(&cset->e_cset_node[ssid],
807 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
809 up_write(&css_set_rwsem);
814 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
816 struct cgroup *root_cgrp = kf_root->kn->priv;
818 return root_cgrp->root;
821 static int cgroup_init_root_id(struct cgroup_root *root)
825 lockdep_assert_held(&cgroup_mutex);
827 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
831 root->hierarchy_id = id;
835 static void cgroup_exit_root_id(struct cgroup_root *root)
837 lockdep_assert_held(&cgroup_mutex);
839 if (root->hierarchy_id) {
840 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
841 root->hierarchy_id = 0;
845 static void cgroup_free_root(struct cgroup_root *root)
848 /* hierarhcy ID shoulid already have been released */
849 WARN_ON_ONCE(root->hierarchy_id);
851 idr_destroy(&root->cgroup_idr);
856 static void cgroup_destroy_root(struct cgroup_root *root)
858 struct cgroup *cgrp = &root->cgrp;
859 struct cgrp_cset_link *link, *tmp_link;
861 mutex_lock(&cgroup_mutex);
863 BUG_ON(atomic_read(&root->nr_cgrps));
864 BUG_ON(!list_empty(&cgrp->self.children));
866 /* Rebind all subsystems back to the default hierarchy */
867 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
870 * Release all the links from cset_links to this hierarchy's
873 down_write(&css_set_rwsem);
875 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
876 list_del(&link->cset_link);
877 list_del(&link->cgrp_link);
880 up_write(&css_set_rwsem);
882 if (!list_empty(&root->root_list)) {
883 list_del(&root->root_list);
887 cgroup_exit_root_id(root);
889 mutex_unlock(&cgroup_mutex);
891 kernfs_destroy_root(root->kf_root);
892 cgroup_free_root(root);
895 /* look up cgroup associated with given css_set on the specified hierarchy */
896 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
897 struct cgroup_root *root)
899 struct cgroup *res = NULL;
901 lockdep_assert_held(&cgroup_mutex);
902 lockdep_assert_held(&css_set_rwsem);
904 if (cset == &init_css_set) {
907 struct cgrp_cset_link *link;
909 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
910 struct cgroup *c = link->cgrp;
912 if (c->root == root) {
924 * Return the cgroup for "task" from the given hierarchy. Must be
925 * called with cgroup_mutex and css_set_rwsem held.
927 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
928 struct cgroup_root *root)
931 * No need to lock the task - since we hold cgroup_mutex the
932 * task can't change groups, so the only thing that can happen
933 * is that it exits and its css is set back to init_css_set.
935 return cset_cgroup_from_root(task_css_set(task), root);
939 * A task must hold cgroup_mutex to modify cgroups.
941 * Any task can increment and decrement the count field without lock.
942 * So in general, code holding cgroup_mutex can't rely on the count
943 * field not changing. However, if the count goes to zero, then only
944 * cgroup_attach_task() can increment it again. Because a count of zero
945 * means that no tasks are currently attached, therefore there is no
946 * way a task attached to that cgroup can fork (the other way to
947 * increment the count). So code holding cgroup_mutex can safely
948 * assume that if the count is zero, it will stay zero. Similarly, if
949 * a task holds cgroup_mutex on a cgroup with zero count, it
950 * knows that the cgroup won't be removed, as cgroup_rmdir()
953 * A cgroup can only be deleted if both its 'count' of using tasks
954 * is zero, and its list of 'children' cgroups is empty. Since all
955 * tasks in the system use _some_ cgroup, and since there is always at
956 * least one task in the system (init, pid == 1), therefore, root cgroup
957 * always has either children cgroups and/or using tasks. So we don't
958 * need a special hack to ensure that root cgroup cannot be deleted.
960 * P.S. One more locking exception. RCU is used to guard the
961 * update of a tasks cgroup pointer by cgroup_attach_task()
964 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
965 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
966 static const struct file_operations proc_cgroupstats_operations;
968 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
971 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
972 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
973 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
974 cft->ss->name, cft->name);
976 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
981 * cgroup_file_mode - deduce file mode of a control file
982 * @cft: the control file in question
984 * returns cft->mode if ->mode is not 0
985 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
986 * returns S_IRUGO if it has only a read handler
987 * returns S_IWUSR if it has only a write hander
989 static umode_t cgroup_file_mode(const struct cftype *cft)
996 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
999 if (cft->write_u64 || cft->write_s64 || cft->write)
1005 static void cgroup_get(struct cgroup *cgrp)
1007 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1008 css_get(&cgrp->self);
1011 static bool cgroup_tryget(struct cgroup *cgrp)
1013 return css_tryget(&cgrp->self);
1016 static void cgroup_put(struct cgroup *cgrp)
1018 css_put(&cgrp->self);
1022 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1023 * @cgrp: the target cgroup
1024 * @subtree_control: the new subtree_control mask to consider
1026 * On the default hierarchy, a subsystem may request other subsystems to be
1027 * enabled together through its ->depends_on mask. In such cases, more
1028 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1030 * This function calculates which subsystems need to be enabled if
1031 * @subtree_control is to be applied to @cgrp. The returned mask is always
1032 * a superset of @subtree_control and follows the usual hierarchy rules.
1034 static unsigned int cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1035 unsigned int subtree_control)
1037 struct cgroup *parent = cgroup_parent(cgrp);
1038 unsigned int cur_ss_mask = subtree_control;
1039 struct cgroup_subsys *ss;
1042 lockdep_assert_held(&cgroup_mutex);
1044 if (!cgroup_on_dfl(cgrp))
1048 unsigned int new_ss_mask = cur_ss_mask;
1050 for_each_subsys(ss, ssid)
1051 if (cur_ss_mask & (1 << ssid))
1052 new_ss_mask |= ss->depends_on;
1055 * Mask out subsystems which aren't available. This can
1056 * happen only if some depended-upon subsystems were bound
1057 * to non-default hierarchies.
1060 new_ss_mask &= parent->child_subsys_mask;
1062 new_ss_mask &= cgrp->root->subsys_mask;
1064 if (new_ss_mask == cur_ss_mask)
1066 cur_ss_mask = new_ss_mask;
1073 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1074 * @cgrp: the target cgroup
1076 * Update @cgrp->child_subsys_mask according to the current
1077 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1079 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1081 cgrp->child_subsys_mask =
1082 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1086 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1087 * @kn: the kernfs_node being serviced
1089 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1090 * the method finishes if locking succeeded. Note that once this function
1091 * returns the cgroup returned by cgroup_kn_lock_live() may become
1092 * inaccessible any time. If the caller intends to continue to access the
1093 * cgroup, it should pin it before invoking this function.
1095 static void cgroup_kn_unlock(struct kernfs_node *kn)
1097 struct cgroup *cgrp;
1099 if (kernfs_type(kn) == KERNFS_DIR)
1102 cgrp = kn->parent->priv;
1104 mutex_unlock(&cgroup_mutex);
1106 kernfs_unbreak_active_protection(kn);
1111 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1112 * @kn: the kernfs_node being serviced
1114 * This helper is to be used by a cgroup kernfs method currently servicing
1115 * @kn. It breaks the active protection, performs cgroup locking and
1116 * verifies that the associated cgroup is alive. Returns the cgroup if
1117 * alive; otherwise, %NULL. A successful return should be undone by a
1118 * matching cgroup_kn_unlock() invocation.
1120 * Any cgroup kernfs method implementation which requires locking the
1121 * associated cgroup should use this helper. It avoids nesting cgroup
1122 * locking under kernfs active protection and allows all kernfs operations
1123 * including self-removal.
1125 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1127 struct cgroup *cgrp;
1129 if (kernfs_type(kn) == KERNFS_DIR)
1132 cgrp = kn->parent->priv;
1135 * We're gonna grab cgroup_mutex which nests outside kernfs
1136 * active_ref. cgroup liveliness check alone provides enough
1137 * protection against removal. Ensure @cgrp stays accessible and
1138 * break the active_ref protection.
1140 if (!cgroup_tryget(cgrp))
1142 kernfs_break_active_protection(kn);
1144 mutex_lock(&cgroup_mutex);
1146 if (!cgroup_is_dead(cgrp))
1149 cgroup_kn_unlock(kn);
1153 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1155 char name[CGROUP_FILE_NAME_MAX];
1157 lockdep_assert_held(&cgroup_mutex);
1158 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1162 * cgroup_clear_dir - remove subsys files in a cgroup directory
1163 * @cgrp: target cgroup
1164 * @subsys_mask: mask of the subsystem ids whose files should be removed
1166 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1168 struct cgroup_subsys *ss;
1171 for_each_subsys(ss, i) {
1172 struct cftype *cfts;
1174 if (!(subsys_mask & (1 << i)))
1176 list_for_each_entry(cfts, &ss->cfts, node)
1177 cgroup_addrm_files(cgrp, cfts, false);
1181 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1183 struct cgroup_subsys *ss;
1184 unsigned int tmp_ss_mask;
1187 lockdep_assert_held(&cgroup_mutex);
1189 for_each_subsys(ss, ssid) {
1190 if (!(ss_mask & (1 << ssid)))
1193 /* if @ss has non-root csses attached to it, can't move */
1194 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1197 /* can't move between two non-dummy roots either */
1198 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1202 /* skip creating root files on dfl_root for inhibited subsystems */
1203 tmp_ss_mask = ss_mask;
1204 if (dst_root == &cgrp_dfl_root)
1205 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1207 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1209 if (dst_root != &cgrp_dfl_root)
1213 * Rebinding back to the default root is not allowed to
1214 * fail. Using both default and non-default roots should
1215 * be rare. Moving subsystems back and forth even more so.
1216 * Just warn about it and continue.
1218 if (cgrp_dfl_root_visible) {
1219 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1221 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1226 * Nothing can fail from this point on. Remove files for the
1227 * removed subsystems and rebind each subsystem.
1229 for_each_subsys(ss, ssid)
1230 if (ss_mask & (1 << ssid))
1231 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1233 for_each_subsys(ss, ssid) {
1234 struct cgroup_root *src_root;
1235 struct cgroup_subsys_state *css;
1236 struct css_set *cset;
1238 if (!(ss_mask & (1 << ssid)))
1241 src_root = ss->root;
1242 css = cgroup_css(&src_root->cgrp, ss);
1244 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1246 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1247 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1248 ss->root = dst_root;
1249 css->cgroup = &dst_root->cgrp;
1251 down_write(&css_set_rwsem);
1252 hash_for_each(css_set_table, i, cset, hlist)
1253 list_move_tail(&cset->e_cset_node[ss->id],
1254 &dst_root->cgrp.e_csets[ss->id]);
1255 up_write(&css_set_rwsem);
1257 src_root->subsys_mask &= ~(1 << ssid);
1258 src_root->cgrp.subtree_control &= ~(1 << ssid);
1259 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1261 /* default hierarchy doesn't enable controllers by default */
1262 dst_root->subsys_mask |= 1 << ssid;
1263 if (dst_root != &cgrp_dfl_root) {
1264 dst_root->cgrp.subtree_control |= 1 << ssid;
1265 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1272 kernfs_activate(dst_root->cgrp.kn);
1276 static int cgroup_show_options(struct seq_file *seq,
1277 struct kernfs_root *kf_root)
1279 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1280 struct cgroup_subsys *ss;
1283 for_each_subsys(ss, ssid)
1284 if (root->subsys_mask & (1 << ssid))
1285 seq_printf(seq, ",%s", ss->name);
1286 if (root->flags & CGRP_ROOT_NOPREFIX)
1287 seq_puts(seq, ",noprefix");
1288 if (root->flags & CGRP_ROOT_XATTR)
1289 seq_puts(seq, ",xattr");
1291 spin_lock(&release_agent_path_lock);
1292 if (strlen(root->release_agent_path))
1293 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1294 spin_unlock(&release_agent_path_lock);
1296 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1297 seq_puts(seq, ",clone_children");
1298 if (strlen(root->name))
1299 seq_printf(seq, ",name=%s", root->name);
1303 struct cgroup_sb_opts {
1304 unsigned int subsys_mask;
1306 char *release_agent;
1307 bool cpuset_clone_children;
1309 /* User explicitly requested empty subsystem */
1313 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1315 char *token, *o = data;
1316 bool all_ss = false, one_ss = false;
1317 unsigned int mask = -1U;
1318 struct cgroup_subsys *ss;
1322 #ifdef CONFIG_CPUSETS
1323 mask = ~(1U << cpuset_cgrp_id);
1326 memset(opts, 0, sizeof(*opts));
1328 while ((token = strsep(&o, ",")) != NULL) {
1333 if (!strcmp(token, "none")) {
1334 /* Explicitly have no subsystems */
1338 if (!strcmp(token, "all")) {
1339 /* Mutually exclusive option 'all' + subsystem name */
1345 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1346 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1349 if (!strcmp(token, "noprefix")) {
1350 opts->flags |= CGRP_ROOT_NOPREFIX;
1353 if (!strcmp(token, "clone_children")) {
1354 opts->cpuset_clone_children = true;
1357 if (!strcmp(token, "xattr")) {
1358 opts->flags |= CGRP_ROOT_XATTR;
1361 if (!strncmp(token, "release_agent=", 14)) {
1362 /* Specifying two release agents is forbidden */
1363 if (opts->release_agent)
1365 opts->release_agent =
1366 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1367 if (!opts->release_agent)
1371 if (!strncmp(token, "name=", 5)) {
1372 const char *name = token + 5;
1373 /* Can't specify an empty name */
1376 /* Must match [\w.-]+ */
1377 for (i = 0; i < strlen(name); i++) {
1381 if ((c == '.') || (c == '-') || (c == '_'))
1385 /* Specifying two names is forbidden */
1388 opts->name = kstrndup(name,
1389 MAX_CGROUP_ROOT_NAMELEN - 1,
1397 for_each_subsys(ss, i) {
1398 if (strcmp(token, ss->name))
1403 /* Mutually exclusive option 'all' + subsystem name */
1406 opts->subsys_mask |= (1 << i);
1411 if (i == CGROUP_SUBSYS_COUNT)
1415 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1416 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1418 pr_err("sane_behavior: no other mount options allowed\n");
1425 * If the 'all' option was specified select all the subsystems,
1426 * otherwise if 'none', 'name=' and a subsystem name options were
1427 * not specified, let's default to 'all'
1429 if (all_ss || (!one_ss && !opts->none && !opts->name))
1430 for_each_subsys(ss, i)
1432 opts->subsys_mask |= (1 << i);
1435 * We either have to specify by name or by subsystems. (So all
1436 * empty hierarchies must have a name).
1438 if (!opts->subsys_mask && !opts->name)
1442 * Option noprefix was introduced just for backward compatibility
1443 * with the old cpuset, so we allow noprefix only if mounting just
1444 * the cpuset subsystem.
1446 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1449 /* Can't specify "none" and some subsystems */
1450 if (opts->subsys_mask && opts->none)
1456 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1459 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1460 struct cgroup_sb_opts opts;
1461 unsigned int added_mask, removed_mask;
1463 if (root == &cgrp_dfl_root) {
1464 pr_err("remount is not allowed\n");
1468 mutex_lock(&cgroup_mutex);
1470 /* See what subsystems are wanted */
1471 ret = parse_cgroupfs_options(data, &opts);
1475 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1476 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1477 task_tgid_nr(current), current->comm);
1479 added_mask = opts.subsys_mask & ~root->subsys_mask;
1480 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1482 /* Don't allow flags or name to change at remount */
1483 if ((opts.flags ^ root->flags) ||
1484 (opts.name && strcmp(opts.name, root->name))) {
1485 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1486 opts.flags, opts.name ?: "", root->flags, root->name);
1491 /* remounting is not allowed for populated hierarchies */
1492 if (!list_empty(&root->cgrp.self.children)) {
1497 ret = rebind_subsystems(root, added_mask);
1501 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1503 if (opts.release_agent) {
1504 spin_lock(&release_agent_path_lock);
1505 strcpy(root->release_agent_path, opts.release_agent);
1506 spin_unlock(&release_agent_path_lock);
1509 kfree(opts.release_agent);
1511 mutex_unlock(&cgroup_mutex);
1516 * To reduce the fork() overhead for systems that are not actually using
1517 * their cgroups capability, we don't maintain the lists running through
1518 * each css_set to its tasks until we see the list actually used - in other
1519 * words after the first mount.
1521 static bool use_task_css_set_links __read_mostly;
1523 static void cgroup_enable_task_cg_lists(void)
1525 struct task_struct *p, *g;
1527 down_write(&css_set_rwsem);
1529 if (use_task_css_set_links)
1532 use_task_css_set_links = true;
1535 * We need tasklist_lock because RCU is not safe against
1536 * while_each_thread(). Besides, a forking task that has passed
1537 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1538 * is not guaranteed to have its child immediately visible in the
1539 * tasklist if we walk through it with RCU.
1541 read_lock(&tasklist_lock);
1542 do_each_thread(g, p) {
1543 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1544 task_css_set(p) != &init_css_set);
1547 * We should check if the process is exiting, otherwise
1548 * it will race with cgroup_exit() in that the list
1549 * entry won't be deleted though the process has exited.
1550 * Do it while holding siglock so that we don't end up
1551 * racing against cgroup_exit().
1553 spin_lock_irq(&p->sighand->siglock);
1554 if (!(p->flags & PF_EXITING)) {
1555 struct css_set *cset = task_css_set(p);
1557 list_add(&p->cg_list, &cset->tasks);
1560 spin_unlock_irq(&p->sighand->siglock);
1561 } while_each_thread(g, p);
1562 read_unlock(&tasklist_lock);
1564 up_write(&css_set_rwsem);
1567 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1569 struct cgroup_subsys *ss;
1572 INIT_LIST_HEAD(&cgrp->self.sibling);
1573 INIT_LIST_HEAD(&cgrp->self.children);
1574 INIT_LIST_HEAD(&cgrp->cset_links);
1575 INIT_LIST_HEAD(&cgrp->pidlists);
1576 mutex_init(&cgrp->pidlist_mutex);
1577 cgrp->self.cgroup = cgrp;
1578 cgrp->self.flags |= CSS_ONLINE;
1580 for_each_subsys(ss, ssid)
1581 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1583 init_waitqueue_head(&cgrp->offline_waitq);
1584 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1587 static void init_cgroup_root(struct cgroup_root *root,
1588 struct cgroup_sb_opts *opts)
1590 struct cgroup *cgrp = &root->cgrp;
1592 INIT_LIST_HEAD(&root->root_list);
1593 atomic_set(&root->nr_cgrps, 1);
1595 init_cgroup_housekeeping(cgrp);
1596 idr_init(&root->cgroup_idr);
1598 root->flags = opts->flags;
1599 if (opts->release_agent)
1600 strcpy(root->release_agent_path, opts->release_agent);
1602 strcpy(root->name, opts->name);
1603 if (opts->cpuset_clone_children)
1604 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1607 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1609 LIST_HEAD(tmp_links);
1610 struct cgroup *root_cgrp = &root->cgrp;
1611 struct cftype *base_files;
1612 struct css_set *cset;
1615 lockdep_assert_held(&cgroup_mutex);
1617 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1620 root_cgrp->id = ret;
1622 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1628 * We're accessing css_set_count without locking css_set_rwsem here,
1629 * but that's OK - it can only be increased by someone holding
1630 * cgroup_lock, and that's us. The worst that can happen is that we
1631 * have some link structures left over
1633 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1637 ret = cgroup_init_root_id(root);
1641 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1642 KERNFS_ROOT_CREATE_DEACTIVATED,
1644 if (IS_ERR(root->kf_root)) {
1645 ret = PTR_ERR(root->kf_root);
1648 root_cgrp->kn = root->kf_root->kn;
1650 if (root == &cgrp_dfl_root)
1651 base_files = cgroup_dfl_base_files;
1653 base_files = cgroup_legacy_base_files;
1655 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1659 ret = rebind_subsystems(root, ss_mask);
1664 * There must be no failure case after here, since rebinding takes
1665 * care of subsystems' refcounts, which are explicitly dropped in
1666 * the failure exit path.
1668 list_add(&root->root_list, &cgroup_roots);
1669 cgroup_root_count++;
1672 * Link the root cgroup in this hierarchy into all the css_set
1675 down_write(&css_set_rwsem);
1676 hash_for_each(css_set_table, i, cset, hlist)
1677 link_css_set(&tmp_links, cset, root_cgrp);
1678 up_write(&css_set_rwsem);
1680 BUG_ON(!list_empty(&root_cgrp->self.children));
1681 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1683 kernfs_activate(root_cgrp->kn);
1688 kernfs_destroy_root(root->kf_root);
1689 root->kf_root = NULL;
1691 cgroup_exit_root_id(root);
1693 percpu_ref_exit(&root_cgrp->self.refcnt);
1695 free_cgrp_cset_links(&tmp_links);
1699 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1700 int flags, const char *unused_dev_name,
1703 struct super_block *pinned_sb = NULL;
1704 struct cgroup_subsys *ss;
1705 struct cgroup_root *root;
1706 struct cgroup_sb_opts opts;
1707 struct dentry *dentry;
1713 * The first time anyone tries to mount a cgroup, enable the list
1714 * linking each css_set to its tasks and fix up all existing tasks.
1716 if (!use_task_css_set_links)
1717 cgroup_enable_task_cg_lists();
1719 mutex_lock(&cgroup_mutex);
1721 /* First find the desired set of subsystems */
1722 ret = parse_cgroupfs_options(data, &opts);
1726 /* look for a matching existing root */
1727 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1728 cgrp_dfl_root_visible = true;
1729 root = &cgrp_dfl_root;
1730 cgroup_get(&root->cgrp);
1736 * Destruction of cgroup root is asynchronous, so subsystems may
1737 * still be dying after the previous unmount. Let's drain the
1738 * dying subsystems. We just need to ensure that the ones
1739 * unmounted previously finish dying and don't care about new ones
1740 * starting. Testing ref liveliness is good enough.
1742 for_each_subsys(ss, i) {
1743 if (!(opts.subsys_mask & (1 << i)) ||
1744 ss->root == &cgrp_dfl_root)
1747 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1748 mutex_unlock(&cgroup_mutex);
1750 ret = restart_syscall();
1753 cgroup_put(&ss->root->cgrp);
1756 for_each_root(root) {
1757 bool name_match = false;
1759 if (root == &cgrp_dfl_root)
1763 * If we asked for a name then it must match. Also, if
1764 * name matches but sybsys_mask doesn't, we should fail.
1765 * Remember whether name matched.
1768 if (strcmp(opts.name, root->name))
1774 * If we asked for subsystems (or explicitly for no
1775 * subsystems) then they must match.
1777 if ((opts.subsys_mask || opts.none) &&
1778 (opts.subsys_mask != root->subsys_mask)) {
1785 if (root->flags ^ opts.flags)
1786 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1789 * We want to reuse @root whose lifetime is governed by its
1790 * ->cgrp. Let's check whether @root is alive and keep it
1791 * that way. As cgroup_kill_sb() can happen anytime, we
1792 * want to block it by pinning the sb so that @root doesn't
1793 * get killed before mount is complete.
1795 * With the sb pinned, tryget_live can reliably indicate
1796 * whether @root can be reused. If it's being killed,
1797 * drain it. We can use wait_queue for the wait but this
1798 * path is super cold. Let's just sleep a bit and retry.
1800 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1801 if (IS_ERR(pinned_sb) ||
1802 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1803 mutex_unlock(&cgroup_mutex);
1804 if (!IS_ERR_OR_NULL(pinned_sb))
1805 deactivate_super(pinned_sb);
1807 ret = restart_syscall();
1816 * No such thing, create a new one. name= matching without subsys
1817 * specification is allowed for already existing hierarchies but we
1818 * can't create new one without subsys specification.
1820 if (!opts.subsys_mask && !opts.none) {
1825 root = kzalloc(sizeof(*root), GFP_KERNEL);
1831 init_cgroup_root(root, &opts);
1833 ret = cgroup_setup_root(root, opts.subsys_mask);
1835 cgroup_free_root(root);
1838 mutex_unlock(&cgroup_mutex);
1840 kfree(opts.release_agent);
1844 return ERR_PTR(ret);
1846 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1847 CGROUP_SUPER_MAGIC, &new_sb);
1848 if (IS_ERR(dentry) || !new_sb)
1849 cgroup_put(&root->cgrp);
1852 * If @pinned_sb, we're reusing an existing root and holding an
1853 * extra ref on its sb. Mount is complete. Put the extra ref.
1857 deactivate_super(pinned_sb);
1863 static void cgroup_kill_sb(struct super_block *sb)
1865 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1866 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1869 * If @root doesn't have any mounts or children, start killing it.
1870 * This prevents new mounts by disabling percpu_ref_tryget_live().
1871 * cgroup_mount() may wait for @root's release.
1873 * And don't kill the default root.
1875 if (css_has_online_children(&root->cgrp.self) ||
1876 root == &cgrp_dfl_root)
1877 cgroup_put(&root->cgrp);
1879 percpu_ref_kill(&root->cgrp.self.refcnt);
1884 static struct file_system_type cgroup_fs_type = {
1886 .mount = cgroup_mount,
1887 .kill_sb = cgroup_kill_sb,
1890 static struct kobject *cgroup_kobj;
1893 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1894 * @task: target task
1895 * @buf: the buffer to write the path into
1896 * @buflen: the length of the buffer
1898 * Determine @task's cgroup on the first (the one with the lowest non-zero
1899 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1900 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1901 * cgroup controller callbacks.
1903 * Return value is the same as kernfs_path().
1905 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1907 struct cgroup_root *root;
1908 struct cgroup *cgrp;
1909 int hierarchy_id = 1;
1912 mutex_lock(&cgroup_mutex);
1913 down_read(&css_set_rwsem);
1915 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1918 cgrp = task_cgroup_from_root(task, root);
1919 path = cgroup_path(cgrp, buf, buflen);
1921 /* if no hierarchy exists, everyone is in "/" */
1922 if (strlcpy(buf, "/", buflen) < buflen)
1926 up_read(&css_set_rwsem);
1927 mutex_unlock(&cgroup_mutex);
1930 EXPORT_SYMBOL_GPL(task_cgroup_path);
1932 /* used to track tasks and other necessary states during migration */
1933 struct cgroup_taskset {
1934 /* the src and dst cset list running through cset->mg_node */
1935 struct list_head src_csets;
1936 struct list_head dst_csets;
1939 * Fields for cgroup_taskset_*() iteration.
1941 * Before migration is committed, the target migration tasks are on
1942 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1943 * the csets on ->dst_csets. ->csets point to either ->src_csets
1944 * or ->dst_csets depending on whether migration is committed.
1946 * ->cur_csets and ->cur_task point to the current task position
1949 struct list_head *csets;
1950 struct css_set *cur_cset;
1951 struct task_struct *cur_task;
1955 * cgroup_taskset_first - reset taskset and return the first task
1956 * @tset: taskset of interest
1958 * @tset iteration is initialized and the first task is returned.
1960 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1962 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1963 tset->cur_task = NULL;
1965 return cgroup_taskset_next(tset);
1969 * cgroup_taskset_next - iterate to the next task in taskset
1970 * @tset: taskset of interest
1972 * Return the next task in @tset. Iteration must have been initialized
1973 * with cgroup_taskset_first().
1975 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1977 struct css_set *cset = tset->cur_cset;
1978 struct task_struct *task = tset->cur_task;
1980 while (&cset->mg_node != tset->csets) {
1982 task = list_first_entry(&cset->mg_tasks,
1983 struct task_struct, cg_list);
1985 task = list_next_entry(task, cg_list);
1987 if (&task->cg_list != &cset->mg_tasks) {
1988 tset->cur_cset = cset;
1989 tset->cur_task = task;
1993 cset = list_next_entry(cset, mg_node);
2001 * cgroup_task_migrate - move a task from one cgroup to another.
2002 * @old_cgrp: the cgroup @tsk is being migrated from
2003 * @tsk: the task being migrated
2004 * @new_cset: the new css_set @tsk is being attached to
2006 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2008 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2009 struct task_struct *tsk,
2010 struct css_set *new_cset)
2012 struct css_set *old_cset;
2014 lockdep_assert_held(&cgroup_mutex);
2015 lockdep_assert_held(&css_set_rwsem);
2018 * We are synchronized through threadgroup_lock() against PF_EXITING
2019 * setting such that we can't race against cgroup_exit() changing the
2020 * css_set to init_css_set and dropping the old one.
2022 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2023 old_cset = task_css_set(tsk);
2025 get_css_set(new_cset);
2026 rcu_assign_pointer(tsk->cgroups, new_cset);
2029 * Use move_tail so that cgroup_taskset_first() still returns the
2030 * leader after migration. This works because cgroup_migrate()
2031 * ensures that the dst_cset of the leader is the first on the
2032 * tset's dst_csets list.
2034 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2037 * We just gained a reference on old_cset by taking it from the
2038 * task. As trading it for new_cset is protected by cgroup_mutex,
2039 * we're safe to drop it here; it will be freed under RCU.
2041 put_css_set_locked(old_cset);
2045 * cgroup_migrate_finish - cleanup after attach
2046 * @preloaded_csets: list of preloaded css_sets
2048 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2049 * those functions for details.
2051 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2053 struct css_set *cset, *tmp_cset;
2055 lockdep_assert_held(&cgroup_mutex);
2057 down_write(&css_set_rwsem);
2058 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2059 cset->mg_src_cgrp = NULL;
2060 cset->mg_dst_cset = NULL;
2061 list_del_init(&cset->mg_preload_node);
2062 put_css_set_locked(cset);
2064 up_write(&css_set_rwsem);
2068 * cgroup_migrate_add_src - add a migration source css_set
2069 * @src_cset: the source css_set to add
2070 * @dst_cgrp: the destination cgroup
2071 * @preloaded_csets: list of preloaded css_sets
2073 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2074 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2075 * up by cgroup_migrate_finish().
2077 * This function may be called without holding threadgroup_lock even if the
2078 * target is a process. Threads may be created and destroyed but as long
2079 * as cgroup_mutex is not dropped, no new css_set can be put into play and
2080 * the preloaded css_sets are guaranteed to cover all migrations.
2082 static void cgroup_migrate_add_src(struct css_set *src_cset,
2083 struct cgroup *dst_cgrp,
2084 struct list_head *preloaded_csets)
2086 struct cgroup *src_cgrp;
2088 lockdep_assert_held(&cgroup_mutex);
2089 lockdep_assert_held(&css_set_rwsem);
2091 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2093 if (!list_empty(&src_cset->mg_preload_node))
2096 WARN_ON(src_cset->mg_src_cgrp);
2097 WARN_ON(!list_empty(&src_cset->mg_tasks));
2098 WARN_ON(!list_empty(&src_cset->mg_node));
2100 src_cset->mg_src_cgrp = src_cgrp;
2101 get_css_set(src_cset);
2102 list_add(&src_cset->mg_preload_node, preloaded_csets);
2106 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2107 * @dst_cgrp: the destination cgroup (may be %NULL)
2108 * @preloaded_csets: list of preloaded source css_sets
2110 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2111 * have been preloaded to @preloaded_csets. This function looks up and
2112 * pins all destination css_sets, links each to its source, and append them
2113 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2114 * source css_set is assumed to be its cgroup on the default hierarchy.
2116 * This function must be called after cgroup_migrate_add_src() has been
2117 * called on each migration source css_set. After migration is performed
2118 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2121 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2122 struct list_head *preloaded_csets)
2125 struct css_set *src_cset, *tmp_cset;
2127 lockdep_assert_held(&cgroup_mutex);
2130 * Except for the root, child_subsys_mask must be zero for a cgroup
2131 * with tasks so that child cgroups don't compete against tasks.
2133 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2134 dst_cgrp->child_subsys_mask)
2137 /* look up the dst cset for each src cset and link it to src */
2138 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2139 struct css_set *dst_cset;
2141 dst_cset = find_css_set(src_cset,
2142 dst_cgrp ?: src_cset->dfl_cgrp);
2146 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2149 * If src cset equals dst, it's noop. Drop the src.
2150 * cgroup_migrate() will skip the cset too. Note that we
2151 * can't handle src == dst as some nodes are used by both.
2153 if (src_cset == dst_cset) {
2154 src_cset->mg_src_cgrp = NULL;
2155 list_del_init(&src_cset->mg_preload_node);
2156 put_css_set(src_cset);
2157 put_css_set(dst_cset);
2161 src_cset->mg_dst_cset = dst_cset;
2163 if (list_empty(&dst_cset->mg_preload_node))
2164 list_add(&dst_cset->mg_preload_node, &csets);
2166 put_css_set(dst_cset);
2169 list_splice_tail(&csets, preloaded_csets);
2172 cgroup_migrate_finish(&csets);
2177 * cgroup_migrate - migrate a process or task to a cgroup
2178 * @cgrp: the destination cgroup
2179 * @leader: the leader of the process or the task to migrate
2180 * @threadgroup: whether @leader points to the whole process or a single task
2182 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2183 * process, the caller must be holding threadgroup_lock of @leader. The
2184 * caller is also responsible for invoking cgroup_migrate_add_src() and
2185 * cgroup_migrate_prepare_dst() on the targets before invoking this
2186 * function and following up with cgroup_migrate_finish().
2188 * As long as a controller's ->can_attach() doesn't fail, this function is
2189 * guaranteed to succeed. This means that, excluding ->can_attach()
2190 * failure, when migrating multiple targets, the success or failure can be
2191 * decided for all targets by invoking group_migrate_prepare_dst() before
2192 * actually starting migrating.
2194 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2197 struct cgroup_taskset tset = {
2198 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2199 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2200 .csets = &tset.src_csets,
2202 struct cgroup_subsys_state *css, *failed_css = NULL;
2203 struct css_set *cset, *tmp_cset;
2204 struct task_struct *task, *tmp_task;
2208 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2209 * already PF_EXITING could be freed from underneath us unless we
2210 * take an rcu_read_lock.
2212 down_write(&css_set_rwsem);
2216 /* @task either already exited or can't exit until the end */
2217 if (task->flags & PF_EXITING)
2220 /* leave @task alone if post_fork() hasn't linked it yet */
2221 if (list_empty(&task->cg_list))
2224 cset = task_css_set(task);
2225 if (!cset->mg_src_cgrp)
2229 * cgroup_taskset_first() must always return the leader.
2230 * Take care to avoid disturbing the ordering.
2232 list_move_tail(&task->cg_list, &cset->mg_tasks);
2233 if (list_empty(&cset->mg_node))
2234 list_add_tail(&cset->mg_node, &tset.src_csets);
2235 if (list_empty(&cset->mg_dst_cset->mg_node))
2236 list_move_tail(&cset->mg_dst_cset->mg_node,
2241 } while_each_thread(leader, task);
2243 up_write(&css_set_rwsem);
2245 /* methods shouldn't be called if no task is actually migrating */
2246 if (list_empty(&tset.src_csets))
2249 /* check that we can legitimately attach to the cgroup */
2250 for_each_e_css(css, i, cgrp) {
2251 if (css->ss->can_attach) {
2252 ret = css->ss->can_attach(css, &tset);
2255 goto out_cancel_attach;
2261 * Now that we're guaranteed success, proceed to move all tasks to
2262 * the new cgroup. There are no failure cases after here, so this
2263 * is the commit point.
2265 down_write(&css_set_rwsem);
2266 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2267 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2268 cgroup_task_migrate(cset->mg_src_cgrp, task,
2271 up_write(&css_set_rwsem);
2274 * Migration is committed, all target tasks are now on dst_csets.
2275 * Nothing is sensitive to fork() after this point. Notify
2276 * controllers that migration is complete.
2278 tset.csets = &tset.dst_csets;
2280 for_each_e_css(css, i, cgrp)
2281 if (css->ss->attach)
2282 css->ss->attach(css, &tset);
2285 goto out_release_tset;
2288 for_each_e_css(css, i, cgrp) {
2289 if (css == failed_css)
2291 if (css->ss->cancel_attach)
2292 css->ss->cancel_attach(css, &tset);
2295 down_write(&css_set_rwsem);
2296 list_splice_init(&tset.dst_csets, &tset.src_csets);
2297 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2298 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2299 list_del_init(&cset->mg_node);
2301 up_write(&css_set_rwsem);
2306 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2307 * @dst_cgrp: the cgroup to attach to
2308 * @leader: the task or the leader of the threadgroup to be attached
2309 * @threadgroup: attach the whole threadgroup?
2311 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2313 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2314 struct task_struct *leader, bool threadgroup)
2316 LIST_HEAD(preloaded_csets);
2317 struct task_struct *task;
2320 /* look up all src csets */
2321 down_read(&css_set_rwsem);
2325 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2329 } while_each_thread(leader, task);
2331 up_read(&css_set_rwsem);
2333 /* prepare dst csets and commit */
2334 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2336 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2338 cgroup_migrate_finish(&preloaded_csets);
2343 * Find the task_struct of the task to attach by vpid and pass it along to the
2344 * function to attach either it or all tasks in its threadgroup. Will lock
2345 * cgroup_mutex and threadgroup.
2347 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2348 size_t nbytes, loff_t off, bool threadgroup)
2350 struct task_struct *tsk;
2351 const struct cred *cred = current_cred(), *tcred;
2352 struct cgroup *cgrp;
2356 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2359 cgrp = cgroup_kn_lock_live(of->kn);
2366 tsk = find_task_by_vpid(pid);
2370 goto out_unlock_cgroup;
2373 * even if we're attaching all tasks in the thread group, we
2374 * only need to check permissions on one of them.
2376 tcred = __task_cred(tsk);
2377 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2378 !uid_eq(cred->euid, tcred->uid) &&
2379 !uid_eq(cred->euid, tcred->suid)) {
2382 goto out_unlock_cgroup;
2388 tsk = tsk->group_leader;
2391 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2392 * trapped in a cpuset, or RT worker may be born in a cgroup
2393 * with no rt_runtime allocated. Just say no.
2395 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2398 goto out_unlock_cgroup;
2401 get_task_struct(tsk);
2404 threadgroup_lock(tsk);
2406 if (!thread_group_leader(tsk)) {
2408 * a race with de_thread from another thread's exec()
2409 * may strip us of our leadership, if this happens,
2410 * there is no choice but to throw this task away and
2411 * try again; this is
2412 * "double-double-toil-and-trouble-check locking".
2414 threadgroup_unlock(tsk);
2415 put_task_struct(tsk);
2416 goto retry_find_task;
2420 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2422 threadgroup_unlock(tsk);
2424 put_task_struct(tsk);
2426 cgroup_kn_unlock(of->kn);
2427 return ret ?: nbytes;
2431 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2432 * @from: attach to all cgroups of a given task
2433 * @tsk: the task to be attached
2435 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2437 struct cgroup_root *root;
2440 mutex_lock(&cgroup_mutex);
2441 for_each_root(root) {
2442 struct cgroup *from_cgrp;
2444 if (root == &cgrp_dfl_root)
2447 down_read(&css_set_rwsem);
2448 from_cgrp = task_cgroup_from_root(from, root);
2449 up_read(&css_set_rwsem);
2451 retval = cgroup_attach_task(from_cgrp, tsk, false);
2455 mutex_unlock(&cgroup_mutex);
2459 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2461 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2462 char *buf, size_t nbytes, loff_t off)
2464 return __cgroup_procs_write(of, buf, nbytes, off, false);
2467 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2468 char *buf, size_t nbytes, loff_t off)
2470 return __cgroup_procs_write(of, buf, nbytes, off, true);
2473 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2474 char *buf, size_t nbytes, loff_t off)
2476 struct cgroup *cgrp;
2478 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2480 cgrp = cgroup_kn_lock_live(of->kn);
2483 spin_lock(&release_agent_path_lock);
2484 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2485 sizeof(cgrp->root->release_agent_path));
2486 spin_unlock(&release_agent_path_lock);
2487 cgroup_kn_unlock(of->kn);
2491 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2493 struct cgroup *cgrp = seq_css(seq)->cgroup;
2495 spin_lock(&release_agent_path_lock);
2496 seq_puts(seq, cgrp->root->release_agent_path);
2497 spin_unlock(&release_agent_path_lock);
2498 seq_putc(seq, '\n');
2502 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2504 seq_puts(seq, "0\n");
2508 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2510 struct cgroup_subsys *ss;
2511 bool printed = false;
2514 for_each_subsys(ss, ssid) {
2515 if (ss_mask & (1 << ssid)) {
2518 seq_printf(seq, "%s", ss->name);
2523 seq_putc(seq, '\n');
2526 /* show controllers which are currently attached to the default hierarchy */
2527 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2529 struct cgroup *cgrp = seq_css(seq)->cgroup;
2531 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2532 ~cgrp_dfl_root_inhibit_ss_mask);
2536 /* show controllers which are enabled from the parent */
2537 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2539 struct cgroup *cgrp = seq_css(seq)->cgroup;
2541 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2545 /* show controllers which are enabled for a given cgroup's children */
2546 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2548 struct cgroup *cgrp = seq_css(seq)->cgroup;
2550 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2555 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2556 * @cgrp: root of the subtree to update csses for
2558 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2559 * css associations need to be updated accordingly. This function looks up
2560 * all css_sets which are attached to the subtree, creates the matching
2561 * updated css_sets and migrates the tasks to the new ones.
2563 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2565 LIST_HEAD(preloaded_csets);
2566 struct cgroup_subsys_state *css;
2567 struct css_set *src_cset;
2570 lockdep_assert_held(&cgroup_mutex);
2572 /* look up all csses currently attached to @cgrp's subtree */
2573 down_read(&css_set_rwsem);
2574 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2575 struct cgrp_cset_link *link;
2577 /* self is not affected by child_subsys_mask change */
2578 if (css->cgroup == cgrp)
2581 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2582 cgroup_migrate_add_src(link->cset, cgrp,
2585 up_read(&css_set_rwsem);
2587 /* NULL dst indicates self on default hierarchy */
2588 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2592 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2593 struct task_struct *last_task = NULL, *task;
2595 /* src_csets precede dst_csets, break on the first dst_cset */
2596 if (!src_cset->mg_src_cgrp)
2600 * All tasks in src_cset need to be migrated to the
2601 * matching dst_cset. Empty it process by process. We
2602 * walk tasks but migrate processes. The leader might even
2603 * belong to a different cset but such src_cset would also
2604 * be among the target src_csets because the default
2605 * hierarchy enforces per-process membership.
2608 down_read(&css_set_rwsem);
2609 task = list_first_entry_or_null(&src_cset->tasks,
2610 struct task_struct, cg_list);
2612 task = task->group_leader;
2613 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2614 get_task_struct(task);
2616 up_read(&css_set_rwsem);
2621 /* guard against possible infinite loop */
2622 if (WARN(last_task == task,
2623 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2627 threadgroup_lock(task);
2628 /* raced against de_thread() from another thread? */
2629 if (!thread_group_leader(task)) {
2630 threadgroup_unlock(task);
2631 put_task_struct(task);
2635 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2637 threadgroup_unlock(task);
2638 put_task_struct(task);
2640 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2646 cgroup_migrate_finish(&preloaded_csets);
2650 /* change the enabled child controllers for a cgroup in the default hierarchy */
2651 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2652 char *buf, size_t nbytes,
2655 unsigned int enable = 0, disable = 0;
2656 unsigned int css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2657 struct cgroup *cgrp, *child;
2658 struct cgroup_subsys *ss;
2663 * Parse input - space separated list of subsystem names prefixed
2664 * with either + or -.
2666 buf = strstrip(buf);
2667 while ((tok = strsep(&buf, " "))) {
2670 for_each_subsys(ss, ssid) {
2671 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2672 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2676 enable |= 1 << ssid;
2677 disable &= ~(1 << ssid);
2678 } else if (*tok == '-') {
2679 disable |= 1 << ssid;
2680 enable &= ~(1 << ssid);
2686 if (ssid == CGROUP_SUBSYS_COUNT)
2690 cgrp = cgroup_kn_lock_live(of->kn);
2694 for_each_subsys(ss, ssid) {
2695 if (enable & (1 << ssid)) {
2696 if (cgrp->subtree_control & (1 << ssid)) {
2697 enable &= ~(1 << ssid);
2701 /* unavailable or not enabled on the parent? */
2702 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2703 (cgroup_parent(cgrp) &&
2704 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2708 } else if (disable & (1 << ssid)) {
2709 if (!(cgrp->subtree_control & (1 << ssid))) {
2710 disable &= ~(1 << ssid);
2714 /* a child has it enabled? */
2715 cgroup_for_each_live_child(child, cgrp) {
2716 if (child->subtree_control & (1 << ssid)) {
2724 if (!enable && !disable) {
2730 * Except for the root, subtree_control must be zero for a cgroup
2731 * with tasks so that child cgroups don't compete against tasks.
2733 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2739 * Update subsys masks and calculate what needs to be done. More
2740 * subsystems than specified may need to be enabled or disabled
2741 * depending on subsystem dependencies.
2743 old_sc = cgrp->subtree_control;
2744 old_ss = cgrp->child_subsys_mask;
2745 new_sc = (old_sc | enable) & ~disable;
2746 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2748 css_enable = ~old_ss & new_ss;
2749 css_disable = old_ss & ~new_ss;
2750 enable |= css_enable;
2751 disable |= css_disable;
2754 * Because css offlining is asynchronous, userland might try to
2755 * re-enable the same controller while the previous instance is
2756 * still around. In such cases, wait till it's gone using
2759 for_each_subsys(ss, ssid) {
2760 if (!(css_enable & (1 << ssid)))
2763 cgroup_for_each_live_child(child, cgrp) {
2766 if (!cgroup_css(child, ss))
2770 prepare_to_wait(&child->offline_waitq, &wait,
2771 TASK_UNINTERRUPTIBLE);
2772 cgroup_kn_unlock(of->kn);
2774 finish_wait(&child->offline_waitq, &wait);
2777 return restart_syscall();
2781 cgrp->subtree_control = new_sc;
2782 cgrp->child_subsys_mask = new_ss;
2785 * Create new csses or make the existing ones visible. A css is
2786 * created invisible if it's being implicitly enabled through
2787 * dependency. An invisible css is made visible when the userland
2788 * explicitly enables it.
2790 for_each_subsys(ss, ssid) {
2791 if (!(enable & (1 << ssid)))
2794 cgroup_for_each_live_child(child, cgrp) {
2795 if (css_enable & (1 << ssid))
2796 ret = create_css(child, ss,
2797 cgrp->subtree_control & (1 << ssid));
2799 ret = cgroup_populate_dir(child, 1 << ssid);
2806 * At this point, cgroup_e_css() results reflect the new csses
2807 * making the following cgroup_update_dfl_csses() properly update
2808 * css associations of all tasks in the subtree.
2810 ret = cgroup_update_dfl_csses(cgrp);
2815 * All tasks are migrated out of disabled csses. Kill or hide
2816 * them. A css is hidden when the userland requests it to be
2817 * disabled while other subsystems are still depending on it. The
2818 * css must not actively control resources and be in the vanilla
2819 * state if it's made visible again later. Controllers which may
2820 * be depended upon should provide ->css_reset() for this purpose.
2822 for_each_subsys(ss, ssid) {
2823 if (!(disable & (1 << ssid)))
2826 cgroup_for_each_live_child(child, cgrp) {
2827 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2829 if (css_disable & (1 << ssid)) {
2832 cgroup_clear_dir(child, 1 << ssid);
2839 kernfs_activate(cgrp->kn);
2842 cgroup_kn_unlock(of->kn);
2843 return ret ?: nbytes;
2846 cgrp->subtree_control = old_sc;
2847 cgrp->child_subsys_mask = old_ss;
2849 for_each_subsys(ss, ssid) {
2850 if (!(enable & (1 << ssid)))
2853 cgroup_for_each_live_child(child, cgrp) {
2854 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2859 if (css_enable & (1 << ssid))
2862 cgroup_clear_dir(child, 1 << ssid);
2868 static int cgroup_populated_show(struct seq_file *seq, void *v)
2870 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2874 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2875 size_t nbytes, loff_t off)
2877 struct cgroup *cgrp = of->kn->parent->priv;
2878 struct cftype *cft = of->kn->priv;
2879 struct cgroup_subsys_state *css;
2883 return cft->write(of, buf, nbytes, off);
2886 * kernfs guarantees that a file isn't deleted with operations in
2887 * flight, which means that the matching css is and stays alive and
2888 * doesn't need to be pinned. The RCU locking is not necessary
2889 * either. It's just for the convenience of using cgroup_css().
2892 css = cgroup_css(cgrp, cft->ss);
2895 if (cft->write_u64) {
2896 unsigned long long v;
2897 ret = kstrtoull(buf, 0, &v);
2899 ret = cft->write_u64(css, cft, v);
2900 } else if (cft->write_s64) {
2902 ret = kstrtoll(buf, 0, &v);
2904 ret = cft->write_s64(css, cft, v);
2909 return ret ?: nbytes;
2912 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2914 return seq_cft(seq)->seq_start(seq, ppos);
2917 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2919 return seq_cft(seq)->seq_next(seq, v, ppos);
2922 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2924 seq_cft(seq)->seq_stop(seq, v);
2927 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2929 struct cftype *cft = seq_cft(m);
2930 struct cgroup_subsys_state *css = seq_css(m);
2933 return cft->seq_show(m, arg);
2936 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2937 else if (cft->read_s64)
2938 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2944 static struct kernfs_ops cgroup_kf_single_ops = {
2945 .atomic_write_len = PAGE_SIZE,
2946 .write = cgroup_file_write,
2947 .seq_show = cgroup_seqfile_show,
2950 static struct kernfs_ops cgroup_kf_ops = {
2951 .atomic_write_len = PAGE_SIZE,
2952 .write = cgroup_file_write,
2953 .seq_start = cgroup_seqfile_start,
2954 .seq_next = cgroup_seqfile_next,
2955 .seq_stop = cgroup_seqfile_stop,
2956 .seq_show = cgroup_seqfile_show,
2960 * cgroup_rename - Only allow simple rename of directories in place.
2962 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2963 const char *new_name_str)
2965 struct cgroup *cgrp = kn->priv;
2968 if (kernfs_type(kn) != KERNFS_DIR)
2970 if (kn->parent != new_parent)
2974 * This isn't a proper migration and its usefulness is very
2975 * limited. Disallow on the default hierarchy.
2977 if (cgroup_on_dfl(cgrp))
2981 * We're gonna grab cgroup_mutex which nests outside kernfs
2982 * active_ref. kernfs_rename() doesn't require active_ref
2983 * protection. Break them before grabbing cgroup_mutex.
2985 kernfs_break_active_protection(new_parent);
2986 kernfs_break_active_protection(kn);
2988 mutex_lock(&cgroup_mutex);
2990 ret = kernfs_rename(kn, new_parent, new_name_str);
2992 mutex_unlock(&cgroup_mutex);
2994 kernfs_unbreak_active_protection(kn);
2995 kernfs_unbreak_active_protection(new_parent);
2999 /* set uid and gid of cgroup dirs and files to that of the creator */
3000 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3002 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3003 .ia_uid = current_fsuid(),
3004 .ia_gid = current_fsgid(), };
3006 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3007 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3010 return kernfs_setattr(kn, &iattr);
3013 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3015 char name[CGROUP_FILE_NAME_MAX];
3016 struct kernfs_node *kn;
3017 struct lock_class_key *key = NULL;
3020 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3021 key = &cft->lockdep_key;
3023 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3024 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3029 ret = cgroup_kn_set_ugid(kn);
3035 if (cft->seq_show == cgroup_populated_show)
3036 cgrp->populated_kn = kn;
3041 * cgroup_addrm_files - add or remove files to a cgroup directory
3042 * @cgrp: the target cgroup
3043 * @cfts: array of cftypes to be added
3044 * @is_add: whether to add or remove
3046 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3047 * For removals, this function never fails. If addition fails, this
3048 * function doesn't remove files already added. The caller is responsible
3051 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3057 lockdep_assert_held(&cgroup_mutex);
3059 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3060 /* does cft->flags tell us to skip this file on @cgrp? */
3061 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3063 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3065 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3067 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3071 ret = cgroup_add_file(cgrp, cft);
3073 pr_warn("%s: failed to add %s, err=%d\n",
3074 __func__, cft->name, ret);
3078 cgroup_rm_file(cgrp, cft);
3084 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3087 struct cgroup_subsys *ss = cfts[0].ss;
3088 struct cgroup *root = &ss->root->cgrp;
3089 struct cgroup_subsys_state *css;
3092 lockdep_assert_held(&cgroup_mutex);
3094 /* add/rm files for all cgroups created before */
3095 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3096 struct cgroup *cgrp = css->cgroup;
3098 if (cgroup_is_dead(cgrp))
3101 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3107 kernfs_activate(root->kn);
3111 static void cgroup_exit_cftypes(struct cftype *cfts)
3115 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3116 /* free copy for custom atomic_write_len, see init_cftypes() */
3117 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3122 /* revert flags set by cgroup core while adding @cfts */
3123 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3127 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3131 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3132 struct kernfs_ops *kf_ops;
3134 WARN_ON(cft->ss || cft->kf_ops);
3137 kf_ops = &cgroup_kf_ops;
3139 kf_ops = &cgroup_kf_single_ops;
3142 * Ugh... if @cft wants a custom max_write_len, we need to
3143 * make a copy of kf_ops to set its atomic_write_len.
3145 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3146 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3148 cgroup_exit_cftypes(cfts);
3151 kf_ops->atomic_write_len = cft->max_write_len;
3154 cft->kf_ops = kf_ops;
3161 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3163 lockdep_assert_held(&cgroup_mutex);
3165 if (!cfts || !cfts[0].ss)
3168 list_del(&cfts->node);
3169 cgroup_apply_cftypes(cfts, false);
3170 cgroup_exit_cftypes(cfts);
3175 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3176 * @cfts: zero-length name terminated array of cftypes
3178 * Unregister @cfts. Files described by @cfts are removed from all
3179 * existing cgroups and all future cgroups won't have them either. This
3180 * function can be called anytime whether @cfts' subsys is attached or not.
3182 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3185 int cgroup_rm_cftypes(struct cftype *cfts)
3189 mutex_lock(&cgroup_mutex);
3190 ret = cgroup_rm_cftypes_locked(cfts);
3191 mutex_unlock(&cgroup_mutex);
3196 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3197 * @ss: target cgroup subsystem
3198 * @cfts: zero-length name terminated array of cftypes
3200 * Register @cfts to @ss. Files described by @cfts are created for all
3201 * existing cgroups to which @ss is attached and all future cgroups will
3202 * have them too. This function can be called anytime whether @ss is
3205 * Returns 0 on successful registration, -errno on failure. Note that this
3206 * function currently returns 0 as long as @cfts registration is successful
3207 * even if some file creation attempts on existing cgroups fail.
3209 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3216 if (!cfts || cfts[0].name[0] == '\0')
3219 ret = cgroup_init_cftypes(ss, cfts);
3223 mutex_lock(&cgroup_mutex);
3225 list_add_tail(&cfts->node, &ss->cfts);
3226 ret = cgroup_apply_cftypes(cfts, true);
3228 cgroup_rm_cftypes_locked(cfts);
3230 mutex_unlock(&cgroup_mutex);
3235 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3236 * @ss: target cgroup subsystem
3237 * @cfts: zero-length name terminated array of cftypes
3239 * Similar to cgroup_add_cftypes() but the added files are only used for
3240 * the default hierarchy.
3242 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3246 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3247 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3248 return cgroup_add_cftypes(ss, cfts);
3252 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3253 * @ss: target cgroup subsystem
3254 * @cfts: zero-length name terminated array of cftypes
3256 * Similar to cgroup_add_cftypes() but the added files are only used for
3257 * the legacy hierarchies.
3259 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3264 * If legacy_flies_on_dfl, we want to show the legacy files on the
3265 * dfl hierarchy but iff the target subsystem hasn't been updated
3266 * for the dfl hierarchy yet.
3268 if (!cgroup_legacy_files_on_dfl ||
3269 ss->dfl_cftypes != ss->legacy_cftypes) {
3270 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3271 cft->flags |= __CFTYPE_NOT_ON_DFL;
3274 return cgroup_add_cftypes(ss, cfts);
3278 * cgroup_task_count - count the number of tasks in a cgroup.
3279 * @cgrp: the cgroup in question
3281 * Return the number of tasks in the cgroup.
3283 static int cgroup_task_count(const struct cgroup *cgrp)
3286 struct cgrp_cset_link *link;
3288 down_read(&css_set_rwsem);
3289 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3290 count += atomic_read(&link->cset->refcount);
3291 up_read(&css_set_rwsem);
3296 * css_next_child - find the next child of a given css
3297 * @pos: the current position (%NULL to initiate traversal)
3298 * @parent: css whose children to walk
3300 * This function returns the next child of @parent and should be called
3301 * under either cgroup_mutex or RCU read lock. The only requirement is
3302 * that @parent and @pos are accessible. The next sibling is guaranteed to
3303 * be returned regardless of their states.
3305 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3306 * css which finished ->css_online() is guaranteed to be visible in the
3307 * future iterations and will stay visible until the last reference is put.
3308 * A css which hasn't finished ->css_online() or already finished
3309 * ->css_offline() may show up during traversal. It's each subsystem's
3310 * responsibility to synchronize against on/offlining.
3312 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3313 struct cgroup_subsys_state *parent)
3315 struct cgroup_subsys_state *next;
3317 cgroup_assert_mutex_or_rcu_locked();
3320 * @pos could already have been unlinked from the sibling list.
3321 * Once a cgroup is removed, its ->sibling.next is no longer
3322 * updated when its next sibling changes. CSS_RELEASED is set when
3323 * @pos is taken off list, at which time its next pointer is valid,
3324 * and, as releases are serialized, the one pointed to by the next
3325 * pointer is guaranteed to not have started release yet. This
3326 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3327 * critical section, the one pointed to by its next pointer is
3328 * guaranteed to not have finished its RCU grace period even if we
3329 * have dropped rcu_read_lock() inbetween iterations.
3331 * If @pos has CSS_RELEASED set, its next pointer can't be
3332 * dereferenced; however, as each css is given a monotonically
3333 * increasing unique serial number and always appended to the
3334 * sibling list, the next one can be found by walking the parent's
3335 * children until the first css with higher serial number than
3336 * @pos's. While this path can be slower, it happens iff iteration
3337 * races against release and the race window is very small.
3340 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3341 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3342 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3344 list_for_each_entry_rcu(next, &parent->children, sibling)
3345 if (next->serial_nr > pos->serial_nr)
3350 * @next, if not pointing to the head, can be dereferenced and is
3353 if (&next->sibling != &parent->children)
3359 * css_next_descendant_pre - find the next descendant for pre-order walk
3360 * @pos: the current position (%NULL to initiate traversal)
3361 * @root: css whose descendants to walk
3363 * To be used by css_for_each_descendant_pre(). Find the next descendant
3364 * to visit for pre-order traversal of @root's descendants. @root is
3365 * included in the iteration and the first node to be visited.
3367 * While this function requires cgroup_mutex or RCU read locking, it
3368 * doesn't require the whole traversal to be contained in a single critical
3369 * section. This function will return the correct next descendant as long
3370 * as both @pos and @root are accessible and @pos is a descendant of @root.
3372 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3373 * css which finished ->css_online() is guaranteed to be visible in the
3374 * future iterations and will stay visible until the last reference is put.
3375 * A css which hasn't finished ->css_online() or already finished
3376 * ->css_offline() may show up during traversal. It's each subsystem's
3377 * responsibility to synchronize against on/offlining.
3379 struct cgroup_subsys_state *
3380 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3381 struct cgroup_subsys_state *root)
3383 struct cgroup_subsys_state *next;
3385 cgroup_assert_mutex_or_rcu_locked();
3387 /* if first iteration, visit @root */
3391 /* visit the first child if exists */
3392 next = css_next_child(NULL, pos);
3396 /* no child, visit my or the closest ancestor's next sibling */
3397 while (pos != root) {
3398 next = css_next_child(pos, pos->parent);
3408 * css_rightmost_descendant - return the rightmost descendant of a css
3409 * @pos: css of interest
3411 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3412 * is returned. This can be used during pre-order traversal to skip
3415 * While this function requires cgroup_mutex or RCU read locking, it
3416 * doesn't require the whole traversal to be contained in a single critical
3417 * section. This function will return the correct rightmost descendant as
3418 * long as @pos is accessible.
3420 struct cgroup_subsys_state *
3421 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3423 struct cgroup_subsys_state *last, *tmp;
3425 cgroup_assert_mutex_or_rcu_locked();
3429 /* ->prev isn't RCU safe, walk ->next till the end */
3431 css_for_each_child(tmp, last)
3438 static struct cgroup_subsys_state *
3439 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3441 struct cgroup_subsys_state *last;
3445 pos = css_next_child(NULL, pos);
3452 * css_next_descendant_post - find the next descendant for post-order walk
3453 * @pos: the current position (%NULL to initiate traversal)
3454 * @root: css whose descendants to walk
3456 * To be used by css_for_each_descendant_post(). Find the next descendant
3457 * to visit for post-order traversal of @root's descendants. @root is
3458 * included in the iteration and the last node to be visited.
3460 * While this function requires cgroup_mutex or RCU read locking, it
3461 * doesn't require the whole traversal to be contained in a single critical
3462 * section. This function will return the correct next descendant as long
3463 * as both @pos and @cgroup are accessible and @pos is a descendant of
3466 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3467 * css which finished ->css_online() is guaranteed to be visible in the
3468 * future iterations and will stay visible until the last reference is put.
3469 * A css which hasn't finished ->css_online() or already finished
3470 * ->css_offline() may show up during traversal. It's each subsystem's
3471 * responsibility to synchronize against on/offlining.
3473 struct cgroup_subsys_state *
3474 css_next_descendant_post(struct cgroup_subsys_state *pos,
3475 struct cgroup_subsys_state *root)
3477 struct cgroup_subsys_state *next;
3479 cgroup_assert_mutex_or_rcu_locked();
3481 /* if first iteration, visit leftmost descendant which may be @root */
3483 return css_leftmost_descendant(root);
3485 /* if we visited @root, we're done */
3489 /* if there's an unvisited sibling, visit its leftmost descendant */
3490 next = css_next_child(pos, pos->parent);
3492 return css_leftmost_descendant(next);
3494 /* no sibling left, visit parent */
3499 * css_has_online_children - does a css have online children
3500 * @css: the target css
3502 * Returns %true if @css has any online children; otherwise, %false. This
3503 * function can be called from any context but the caller is responsible
3504 * for synchronizing against on/offlining as necessary.
3506 bool css_has_online_children(struct cgroup_subsys_state *css)
3508 struct cgroup_subsys_state *child;
3512 css_for_each_child(child, css) {
3513 if (child->flags & CSS_ONLINE) {
3523 * css_advance_task_iter - advance a task itererator to the next css_set
3524 * @it: the iterator to advance
3526 * Advance @it to the next css_set to walk.
3528 static void css_advance_task_iter(struct css_task_iter *it)
3530 struct list_head *l = it->cset_pos;
3531 struct cgrp_cset_link *link;
3532 struct css_set *cset;
3534 /* Advance to the next non-empty css_set */
3537 if (l == it->cset_head) {
3538 it->cset_pos = NULL;
3543 cset = container_of(l, struct css_set,
3544 e_cset_node[it->ss->id]);
3546 link = list_entry(l, struct cgrp_cset_link, cset_link);
3549 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3553 if (!list_empty(&cset->tasks))
3554 it->task_pos = cset->tasks.next;
3556 it->task_pos = cset->mg_tasks.next;
3558 it->tasks_head = &cset->tasks;
3559 it->mg_tasks_head = &cset->mg_tasks;
3563 * css_task_iter_start - initiate task iteration
3564 * @css: the css to walk tasks of
3565 * @it: the task iterator to use
3567 * Initiate iteration through the tasks of @css. The caller can call
3568 * css_task_iter_next() to walk through the tasks until the function
3569 * returns NULL. On completion of iteration, css_task_iter_end() must be
3572 * Note that this function acquires a lock which is released when the
3573 * iteration finishes. The caller can't sleep while iteration is in
3576 void css_task_iter_start(struct cgroup_subsys_state *css,
3577 struct css_task_iter *it)
3578 __acquires(css_set_rwsem)
3580 /* no one should try to iterate before mounting cgroups */
3581 WARN_ON_ONCE(!use_task_css_set_links);
3583 down_read(&css_set_rwsem);
3588 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3590 it->cset_pos = &css->cgroup->cset_links;
3592 it->cset_head = it->cset_pos;
3594 css_advance_task_iter(it);
3598 * css_task_iter_next - return the next task for the iterator
3599 * @it: the task iterator being iterated
3601 * The "next" function for task iteration. @it should have been
3602 * initialized via css_task_iter_start(). Returns NULL when the iteration
3605 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3607 struct task_struct *res;
3608 struct list_head *l = it->task_pos;
3610 /* If the iterator cg is NULL, we have no tasks */
3613 res = list_entry(l, struct task_struct, cg_list);
3616 * Advance iterator to find next entry. cset->tasks is consumed
3617 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3622 if (l == it->tasks_head)
3623 l = it->mg_tasks_head->next;
3625 if (l == it->mg_tasks_head)
3626 css_advance_task_iter(it);
3634 * css_task_iter_end - finish task iteration
3635 * @it: the task iterator to finish
3637 * Finish task iteration started by css_task_iter_start().
3639 void css_task_iter_end(struct css_task_iter *it)
3640 __releases(css_set_rwsem)
3642 up_read(&css_set_rwsem);
3646 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3647 * @to: cgroup to which the tasks will be moved
3648 * @from: cgroup in which the tasks currently reside
3650 * Locking rules between cgroup_post_fork() and the migration path
3651 * guarantee that, if a task is forking while being migrated, the new child
3652 * is guaranteed to be either visible in the source cgroup after the
3653 * parent's migration is complete or put into the target cgroup. No task
3654 * can slip out of migration through forking.
3656 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3658 LIST_HEAD(preloaded_csets);
3659 struct cgrp_cset_link *link;
3660 struct css_task_iter it;
3661 struct task_struct *task;
3664 mutex_lock(&cgroup_mutex);
3666 /* all tasks in @from are being moved, all csets are source */
3667 down_read(&css_set_rwsem);
3668 list_for_each_entry(link, &from->cset_links, cset_link)
3669 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3670 up_read(&css_set_rwsem);
3672 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3677 * Migrate tasks one-by-one until @form is empty. This fails iff
3678 * ->can_attach() fails.
3681 css_task_iter_start(&from->self, &it);
3682 task = css_task_iter_next(&it);
3684 get_task_struct(task);
3685 css_task_iter_end(&it);
3688 ret = cgroup_migrate(to, task, false);
3689 put_task_struct(task);
3691 } while (task && !ret);
3693 cgroup_migrate_finish(&preloaded_csets);
3694 mutex_unlock(&cgroup_mutex);
3699 * Stuff for reading the 'tasks'/'procs' files.
3701 * Reading this file can return large amounts of data if a cgroup has
3702 * *lots* of attached tasks. So it may need several calls to read(),
3703 * but we cannot guarantee that the information we produce is correct
3704 * unless we produce it entirely atomically.
3708 /* which pidlist file are we talking about? */
3709 enum cgroup_filetype {
3715 * A pidlist is a list of pids that virtually represents the contents of one
3716 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3717 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3720 struct cgroup_pidlist {
3722 * used to find which pidlist is wanted. doesn't change as long as
3723 * this particular list stays in the list.
3725 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3728 /* how many elements the above list has */
3730 /* each of these stored in a list by its cgroup */
3731 struct list_head links;
3732 /* pointer to the cgroup we belong to, for list removal purposes */
3733 struct cgroup *owner;
3734 /* for delayed destruction */
3735 struct delayed_work destroy_dwork;
3739 * The following two functions "fix" the issue where there are more pids
3740 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3741 * TODO: replace with a kernel-wide solution to this problem
3743 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3744 static void *pidlist_allocate(int count)
3746 if (PIDLIST_TOO_LARGE(count))
3747 return vmalloc(count * sizeof(pid_t));
3749 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3752 static void pidlist_free(void *p)
3754 if (is_vmalloc_addr(p))
3761 * Used to destroy all pidlists lingering waiting for destroy timer. None
3762 * should be left afterwards.
3764 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3766 struct cgroup_pidlist *l, *tmp_l;
3768 mutex_lock(&cgrp->pidlist_mutex);
3769 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3770 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3771 mutex_unlock(&cgrp->pidlist_mutex);
3773 flush_workqueue(cgroup_pidlist_destroy_wq);
3774 BUG_ON(!list_empty(&cgrp->pidlists));
3777 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3779 struct delayed_work *dwork = to_delayed_work(work);
3780 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3782 struct cgroup_pidlist *tofree = NULL;
3784 mutex_lock(&l->owner->pidlist_mutex);
3787 * Destroy iff we didn't get queued again. The state won't change
3788 * as destroy_dwork can only be queued while locked.
3790 if (!delayed_work_pending(dwork)) {
3791 list_del(&l->links);
3792 pidlist_free(l->list);
3793 put_pid_ns(l->key.ns);
3797 mutex_unlock(&l->owner->pidlist_mutex);
3802 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3803 * Returns the number of unique elements.
3805 static int pidlist_uniq(pid_t *list, int length)
3810 * we presume the 0th element is unique, so i starts at 1. trivial
3811 * edge cases first; no work needs to be done for either
3813 if (length == 0 || length == 1)
3815 /* src and dest walk down the list; dest counts unique elements */
3816 for (src = 1; src < length; src++) {
3817 /* find next unique element */
3818 while (list[src] == list[src-1]) {
3823 /* dest always points to where the next unique element goes */
3824 list[dest] = list[src];
3832 * The two pid files - task and cgroup.procs - guaranteed that the result
3833 * is sorted, which forced this whole pidlist fiasco. As pid order is
3834 * different per namespace, each namespace needs differently sorted list,
3835 * making it impossible to use, for example, single rbtree of member tasks
3836 * sorted by task pointer. As pidlists can be fairly large, allocating one
3837 * per open file is dangerous, so cgroup had to implement shared pool of
3838 * pidlists keyed by cgroup and namespace.
3840 * All this extra complexity was caused by the original implementation
3841 * committing to an entirely unnecessary property. In the long term, we
3842 * want to do away with it. Explicitly scramble sort order if on the
3843 * default hierarchy so that no such expectation exists in the new
3846 * Scrambling is done by swapping every two consecutive bits, which is
3847 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3849 static pid_t pid_fry(pid_t pid)
3851 unsigned a = pid & 0x55555555;
3852 unsigned b = pid & 0xAAAAAAAA;
3854 return (a << 1) | (b >> 1);
3857 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3859 if (cgroup_on_dfl(cgrp))
3860 return pid_fry(pid);
3865 static int cmppid(const void *a, const void *b)
3867 return *(pid_t *)a - *(pid_t *)b;
3870 static int fried_cmppid(const void *a, const void *b)
3872 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3875 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3876 enum cgroup_filetype type)
3878 struct cgroup_pidlist *l;
3879 /* don't need task_nsproxy() if we're looking at ourself */
3880 struct pid_namespace *ns = task_active_pid_ns(current);
3882 lockdep_assert_held(&cgrp->pidlist_mutex);
3884 list_for_each_entry(l, &cgrp->pidlists, links)
3885 if (l->key.type == type && l->key.ns == ns)
3891 * find the appropriate pidlist for our purpose (given procs vs tasks)
3892 * returns with the lock on that pidlist already held, and takes care
3893 * of the use count, or returns NULL with no locks held if we're out of
3896 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3897 enum cgroup_filetype type)
3899 struct cgroup_pidlist *l;
3901 lockdep_assert_held(&cgrp->pidlist_mutex);
3903 l = cgroup_pidlist_find(cgrp, type);
3907 /* entry not found; create a new one */
3908 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3912 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3914 /* don't need task_nsproxy() if we're looking at ourself */
3915 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3917 list_add(&l->links, &cgrp->pidlists);
3922 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3924 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3925 struct cgroup_pidlist **lp)
3929 int pid, n = 0; /* used for populating the array */
3930 struct css_task_iter it;
3931 struct task_struct *tsk;
3932 struct cgroup_pidlist *l;
3934 lockdep_assert_held(&cgrp->pidlist_mutex);
3937 * If cgroup gets more users after we read count, we won't have
3938 * enough space - tough. This race is indistinguishable to the
3939 * caller from the case that the additional cgroup users didn't
3940 * show up until sometime later on.
3942 length = cgroup_task_count(cgrp);
3943 array = pidlist_allocate(length);
3946 /* now, populate the array */
3947 css_task_iter_start(&cgrp->self, &it);
3948 while ((tsk = css_task_iter_next(&it))) {
3949 if (unlikely(n == length))
3951 /* get tgid or pid for procs or tasks file respectively */
3952 if (type == CGROUP_FILE_PROCS)
3953 pid = task_tgid_vnr(tsk);
3955 pid = task_pid_vnr(tsk);
3956 if (pid > 0) /* make sure to only use valid results */
3959 css_task_iter_end(&it);
3961 /* now sort & (if procs) strip out duplicates */
3962 if (cgroup_on_dfl(cgrp))
3963 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3965 sort(array, length, sizeof(pid_t), cmppid, NULL);
3966 if (type == CGROUP_FILE_PROCS)
3967 length = pidlist_uniq(array, length);
3969 l = cgroup_pidlist_find_create(cgrp, type);
3971 pidlist_free(array);
3975 /* store array, freeing old if necessary */
3976 pidlist_free(l->list);
3984 * cgroupstats_build - build and fill cgroupstats
3985 * @stats: cgroupstats to fill information into
3986 * @dentry: A dentry entry belonging to the cgroup for which stats have
3989 * Build and fill cgroupstats so that taskstats can export it to user
3992 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3994 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3995 struct cgroup *cgrp;
3996 struct css_task_iter it;
3997 struct task_struct *tsk;
3999 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4000 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4001 kernfs_type(kn) != KERNFS_DIR)
4004 mutex_lock(&cgroup_mutex);
4007 * We aren't being called from kernfs and there's no guarantee on
4008 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4009 * @kn->priv is RCU safe. Let's do the RCU dancing.
4012 cgrp = rcu_dereference(kn->priv);
4013 if (!cgrp || cgroup_is_dead(cgrp)) {
4015 mutex_unlock(&cgroup_mutex);
4020 css_task_iter_start(&cgrp->self, &it);
4021 while ((tsk = css_task_iter_next(&it))) {
4022 switch (tsk->state) {
4024 stats->nr_running++;
4026 case TASK_INTERRUPTIBLE:
4027 stats->nr_sleeping++;
4029 case TASK_UNINTERRUPTIBLE:
4030 stats->nr_uninterruptible++;
4033 stats->nr_stopped++;
4036 if (delayacct_is_task_waiting_on_io(tsk))
4037 stats->nr_io_wait++;
4041 css_task_iter_end(&it);
4043 mutex_unlock(&cgroup_mutex);
4049 * seq_file methods for the tasks/procs files. The seq_file position is the
4050 * next pid to display; the seq_file iterator is a pointer to the pid
4051 * in the cgroup->l->list array.
4054 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4057 * Initially we receive a position value that corresponds to
4058 * one more than the last pid shown (or 0 on the first call or
4059 * after a seek to the start). Use a binary-search to find the
4060 * next pid to display, if any
4062 struct kernfs_open_file *of = s->private;
4063 struct cgroup *cgrp = seq_css(s)->cgroup;
4064 struct cgroup_pidlist *l;
4065 enum cgroup_filetype type = seq_cft(s)->private;
4066 int index = 0, pid = *pos;
4069 mutex_lock(&cgrp->pidlist_mutex);
4072 * !NULL @of->priv indicates that this isn't the first start()
4073 * after open. If the matching pidlist is around, we can use that.
4074 * Look for it. Note that @of->priv can't be used directly. It
4075 * could already have been destroyed.
4078 of->priv = cgroup_pidlist_find(cgrp, type);
4081 * Either this is the first start() after open or the matching
4082 * pidlist has been destroyed inbetween. Create a new one.
4085 ret = pidlist_array_load(cgrp, type,
4086 (struct cgroup_pidlist **)&of->priv);
4088 return ERR_PTR(ret);
4093 int end = l->length;
4095 while (index < end) {
4096 int mid = (index + end) / 2;
4097 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4100 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4106 /* If we're off the end of the array, we're done */
4107 if (index >= l->length)
4109 /* Update the abstract position to be the actual pid that we found */
4110 iter = l->list + index;
4111 *pos = cgroup_pid_fry(cgrp, *iter);
4115 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4117 struct kernfs_open_file *of = s->private;
4118 struct cgroup_pidlist *l = of->priv;
4121 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4122 CGROUP_PIDLIST_DESTROY_DELAY);
4123 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4126 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4128 struct kernfs_open_file *of = s->private;
4129 struct cgroup_pidlist *l = of->priv;
4131 pid_t *end = l->list + l->length;
4133 * Advance to the next pid in the array. If this goes off the
4140 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4145 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4147 return seq_printf(s, "%d\n", *(int *)v);
4150 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4153 return notify_on_release(css->cgroup);
4156 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4157 struct cftype *cft, u64 val)
4160 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4162 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4166 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4169 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4172 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4173 struct cftype *cft, u64 val)
4176 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4178 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4182 /* cgroup core interface files for the default hierarchy */
4183 static struct cftype cgroup_dfl_base_files[] = {
4185 .name = "cgroup.procs",
4186 .seq_start = cgroup_pidlist_start,
4187 .seq_next = cgroup_pidlist_next,
4188 .seq_stop = cgroup_pidlist_stop,
4189 .seq_show = cgroup_pidlist_show,
4190 .private = CGROUP_FILE_PROCS,
4191 .write = cgroup_procs_write,
4192 .mode = S_IRUGO | S_IWUSR,
4195 .name = "cgroup.controllers",
4196 .flags = CFTYPE_ONLY_ON_ROOT,
4197 .seq_show = cgroup_root_controllers_show,
4200 .name = "cgroup.controllers",
4201 .flags = CFTYPE_NOT_ON_ROOT,
4202 .seq_show = cgroup_controllers_show,
4205 .name = "cgroup.subtree_control",
4206 .seq_show = cgroup_subtree_control_show,
4207 .write = cgroup_subtree_control_write,
4210 .name = "cgroup.populated",
4211 .flags = CFTYPE_NOT_ON_ROOT,
4212 .seq_show = cgroup_populated_show,
4217 /* cgroup core interface files for the legacy hierarchies */
4218 static struct cftype cgroup_legacy_base_files[] = {
4220 .name = "cgroup.procs",
4221 .seq_start = cgroup_pidlist_start,
4222 .seq_next = cgroup_pidlist_next,
4223 .seq_stop = cgroup_pidlist_stop,
4224 .seq_show = cgroup_pidlist_show,
4225 .private = CGROUP_FILE_PROCS,
4226 .write = cgroup_procs_write,
4227 .mode = S_IRUGO | S_IWUSR,
4230 .name = "cgroup.clone_children",
4231 .read_u64 = cgroup_clone_children_read,
4232 .write_u64 = cgroup_clone_children_write,
4235 .name = "cgroup.sane_behavior",
4236 .flags = CFTYPE_ONLY_ON_ROOT,
4237 .seq_show = cgroup_sane_behavior_show,
4241 .seq_start = cgroup_pidlist_start,
4242 .seq_next = cgroup_pidlist_next,
4243 .seq_stop = cgroup_pidlist_stop,
4244 .seq_show = cgroup_pidlist_show,
4245 .private = CGROUP_FILE_TASKS,
4246 .write = cgroup_tasks_write,
4247 .mode = S_IRUGO | S_IWUSR,
4250 .name = "notify_on_release",
4251 .read_u64 = cgroup_read_notify_on_release,
4252 .write_u64 = cgroup_write_notify_on_release,
4255 .name = "release_agent",
4256 .flags = CFTYPE_ONLY_ON_ROOT,
4257 .seq_show = cgroup_release_agent_show,
4258 .write = cgroup_release_agent_write,
4259 .max_write_len = PATH_MAX - 1,
4265 * cgroup_populate_dir - create subsys files in a cgroup directory
4266 * @cgrp: target cgroup
4267 * @subsys_mask: mask of the subsystem ids whose files should be added
4269 * On failure, no file is added.
4271 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4273 struct cgroup_subsys *ss;
4276 /* process cftsets of each subsystem */
4277 for_each_subsys(ss, i) {
4278 struct cftype *cfts;
4280 if (!(subsys_mask & (1 << i)))
4283 list_for_each_entry(cfts, &ss->cfts, node) {
4284 ret = cgroup_addrm_files(cgrp, cfts, true);
4291 cgroup_clear_dir(cgrp, subsys_mask);
4296 * css destruction is four-stage process.
4298 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4299 * Implemented in kill_css().
4301 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4302 * and thus css_tryget_online() is guaranteed to fail, the css can be
4303 * offlined by invoking offline_css(). After offlining, the base ref is
4304 * put. Implemented in css_killed_work_fn().
4306 * 3. When the percpu_ref reaches zero, the only possible remaining
4307 * accessors are inside RCU read sections. css_release() schedules the
4310 * 4. After the grace period, the css can be freed. Implemented in
4311 * css_free_work_fn().
4313 * It is actually hairier because both step 2 and 4 require process context
4314 * and thus involve punting to css->destroy_work adding two additional
4315 * steps to the already complex sequence.
4317 static void css_free_work_fn(struct work_struct *work)
4319 struct cgroup_subsys_state *css =
4320 container_of(work, struct cgroup_subsys_state, destroy_work);
4321 struct cgroup *cgrp = css->cgroup;
4323 percpu_ref_exit(&css->refcnt);
4328 css_put(css->parent);
4330 css->ss->css_free(css);
4333 /* cgroup free path */
4334 atomic_dec(&cgrp->root->nr_cgrps);
4335 cgroup_pidlist_destroy_all(cgrp);
4336 cancel_work_sync(&cgrp->release_agent_work);
4338 if (cgroup_parent(cgrp)) {
4340 * We get a ref to the parent, and put the ref when
4341 * this cgroup is being freed, so it's guaranteed
4342 * that the parent won't be destroyed before its
4345 cgroup_put(cgroup_parent(cgrp));
4346 kernfs_put(cgrp->kn);
4350 * This is root cgroup's refcnt reaching zero,
4351 * which indicates that the root should be
4354 cgroup_destroy_root(cgrp->root);
4359 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4361 struct cgroup_subsys_state *css =
4362 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4364 INIT_WORK(&css->destroy_work, css_free_work_fn);
4365 queue_work(cgroup_destroy_wq, &css->destroy_work);
4368 static void css_release_work_fn(struct work_struct *work)
4370 struct cgroup_subsys_state *css =
4371 container_of(work, struct cgroup_subsys_state, destroy_work);
4372 struct cgroup_subsys *ss = css->ss;
4373 struct cgroup *cgrp = css->cgroup;
4375 mutex_lock(&cgroup_mutex);
4377 css->flags |= CSS_RELEASED;
4378 list_del_rcu(&css->sibling);
4381 /* css release path */
4382 cgroup_idr_remove(&ss->css_idr, css->id);
4384 /* cgroup release path */
4385 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4389 * There are two control paths which try to determine
4390 * cgroup from dentry without going through kernfs -
4391 * cgroupstats_build() and css_tryget_online_from_dir().
4392 * Those are supported by RCU protecting clearing of
4393 * cgrp->kn->priv backpointer.
4395 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4398 mutex_unlock(&cgroup_mutex);
4400 call_rcu(&css->rcu_head, css_free_rcu_fn);
4403 static void css_release(struct percpu_ref *ref)
4405 struct cgroup_subsys_state *css =
4406 container_of(ref, struct cgroup_subsys_state, refcnt);
4408 INIT_WORK(&css->destroy_work, css_release_work_fn);
4409 queue_work(cgroup_destroy_wq, &css->destroy_work);
4412 static void init_and_link_css(struct cgroup_subsys_state *css,
4413 struct cgroup_subsys *ss, struct cgroup *cgrp)
4415 lockdep_assert_held(&cgroup_mutex);
4419 memset(css, 0, sizeof(*css));
4422 INIT_LIST_HEAD(&css->sibling);
4423 INIT_LIST_HEAD(&css->children);
4424 css->serial_nr = css_serial_nr_next++;
4426 if (cgroup_parent(cgrp)) {
4427 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4428 css_get(css->parent);
4431 BUG_ON(cgroup_css(cgrp, ss));
4434 /* invoke ->css_online() on a new CSS and mark it online if successful */
4435 static int online_css(struct cgroup_subsys_state *css)
4437 struct cgroup_subsys *ss = css->ss;
4440 lockdep_assert_held(&cgroup_mutex);
4443 ret = ss->css_online(css);
4445 css->flags |= CSS_ONLINE;
4446 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4451 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4452 static void offline_css(struct cgroup_subsys_state *css)
4454 struct cgroup_subsys *ss = css->ss;
4456 lockdep_assert_held(&cgroup_mutex);
4458 if (!(css->flags & CSS_ONLINE))
4461 if (ss->css_offline)
4462 ss->css_offline(css);
4464 css->flags &= ~CSS_ONLINE;
4465 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4467 wake_up_all(&css->cgroup->offline_waitq);
4471 * create_css - create a cgroup_subsys_state
4472 * @cgrp: the cgroup new css will be associated with
4473 * @ss: the subsys of new css
4474 * @visible: whether to create control knobs for the new css or not
4476 * Create a new css associated with @cgrp - @ss pair. On success, the new
4477 * css is online and installed in @cgrp with all interface files created if
4478 * @visible. Returns 0 on success, -errno on failure.
4480 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4483 struct cgroup *parent = cgroup_parent(cgrp);
4484 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4485 struct cgroup_subsys_state *css;
4488 lockdep_assert_held(&cgroup_mutex);
4490 css = ss->css_alloc(parent_css);
4492 return PTR_ERR(css);
4494 init_and_link_css(css, ss, cgrp);
4496 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4500 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4502 goto err_free_percpu_ref;
4506 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4511 /* @css is ready to be brought online now, make it visible */
4512 list_add_tail_rcu(&css->sibling, &parent_css->children);
4513 cgroup_idr_replace(&ss->css_idr, css, css->id);
4515 err = online_css(css);
4519 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4520 cgroup_parent(parent)) {
4521 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4522 current->comm, current->pid, ss->name);
4523 if (!strcmp(ss->name, "memory"))
4524 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4525 ss->warned_broken_hierarchy = true;
4531 list_del_rcu(&css->sibling);
4532 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4534 cgroup_idr_remove(&ss->css_idr, css->id);
4535 err_free_percpu_ref:
4536 percpu_ref_exit(&css->refcnt);
4538 call_rcu(&css->rcu_head, css_free_rcu_fn);
4542 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4545 struct cgroup *parent, *cgrp;
4546 struct cgroup_root *root;
4547 struct cgroup_subsys *ss;
4548 struct kernfs_node *kn;
4549 struct cftype *base_files;
4552 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4554 if (strchr(name, '\n'))
4557 parent = cgroup_kn_lock_live(parent_kn);
4560 root = parent->root;
4562 /* allocate the cgroup and its ID, 0 is reserved for the root */
4563 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4569 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4574 * Temporarily set the pointer to NULL, so idr_find() won't return
4575 * a half-baked cgroup.
4577 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4580 goto out_cancel_ref;
4583 init_cgroup_housekeeping(cgrp);
4585 cgrp->self.parent = &parent->self;
4588 if (notify_on_release(parent))
4589 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4591 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4592 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4594 /* create the directory */
4595 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4603 * This extra ref will be put in cgroup_free_fn() and guarantees
4604 * that @cgrp->kn is always accessible.
4608 cgrp->self.serial_nr = css_serial_nr_next++;
4610 /* allocation complete, commit to creation */
4611 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4612 atomic_inc(&root->nr_cgrps);
4616 * @cgrp is now fully operational. If something fails after this
4617 * point, it'll be released via the normal destruction path.
4619 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4621 ret = cgroup_kn_set_ugid(kn);
4625 if (cgroup_on_dfl(cgrp))
4626 base_files = cgroup_dfl_base_files;
4628 base_files = cgroup_legacy_base_files;
4630 ret = cgroup_addrm_files(cgrp, base_files, true);
4634 /* let's create and online css's */
4635 for_each_subsys(ss, ssid) {
4636 if (parent->child_subsys_mask & (1 << ssid)) {
4637 ret = create_css(cgrp, ss,
4638 parent->subtree_control & (1 << ssid));
4645 * On the default hierarchy, a child doesn't automatically inherit
4646 * subtree_control from the parent. Each is configured manually.
4648 if (!cgroup_on_dfl(cgrp)) {
4649 cgrp->subtree_control = parent->subtree_control;
4650 cgroup_refresh_child_subsys_mask(cgrp);
4653 kernfs_activate(kn);
4659 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4661 percpu_ref_exit(&cgrp->self.refcnt);
4665 cgroup_kn_unlock(parent_kn);
4669 cgroup_destroy_locked(cgrp);
4674 * This is called when the refcnt of a css is confirmed to be killed.
4675 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4676 * initate destruction and put the css ref from kill_css().
4678 static void css_killed_work_fn(struct work_struct *work)
4680 struct cgroup_subsys_state *css =
4681 container_of(work, struct cgroup_subsys_state, destroy_work);
4683 mutex_lock(&cgroup_mutex);
4685 mutex_unlock(&cgroup_mutex);
4690 /* css kill confirmation processing requires process context, bounce */
4691 static void css_killed_ref_fn(struct percpu_ref *ref)
4693 struct cgroup_subsys_state *css =
4694 container_of(ref, struct cgroup_subsys_state, refcnt);
4696 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4697 queue_work(cgroup_destroy_wq, &css->destroy_work);
4701 * kill_css - destroy a css
4702 * @css: css to destroy
4704 * This function initiates destruction of @css by removing cgroup interface
4705 * files and putting its base reference. ->css_offline() will be invoked
4706 * asynchronously once css_tryget_online() is guaranteed to fail and when
4707 * the reference count reaches zero, @css will be released.
4709 static void kill_css(struct cgroup_subsys_state *css)
4711 lockdep_assert_held(&cgroup_mutex);
4714 * This must happen before css is disassociated with its cgroup.
4715 * See seq_css() for details.
4717 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4720 * Killing would put the base ref, but we need to keep it alive
4721 * until after ->css_offline().
4726 * cgroup core guarantees that, by the time ->css_offline() is
4727 * invoked, no new css reference will be given out via
4728 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4729 * proceed to offlining css's because percpu_ref_kill() doesn't
4730 * guarantee that the ref is seen as killed on all CPUs on return.
4732 * Use percpu_ref_kill_and_confirm() to get notifications as each
4733 * css is confirmed to be seen as killed on all CPUs.
4735 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4739 * cgroup_destroy_locked - the first stage of cgroup destruction
4740 * @cgrp: cgroup to be destroyed
4742 * css's make use of percpu refcnts whose killing latency shouldn't be
4743 * exposed to userland and are RCU protected. Also, cgroup core needs to
4744 * guarantee that css_tryget_online() won't succeed by the time
4745 * ->css_offline() is invoked. To satisfy all the requirements,
4746 * destruction is implemented in the following two steps.
4748 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4749 * userland visible parts and start killing the percpu refcnts of
4750 * css's. Set up so that the next stage will be kicked off once all
4751 * the percpu refcnts are confirmed to be killed.
4753 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4754 * rest of destruction. Once all cgroup references are gone, the
4755 * cgroup is RCU-freed.
4757 * This function implements s1. After this step, @cgrp is gone as far as
4758 * the userland is concerned and a new cgroup with the same name may be
4759 * created. As cgroup doesn't care about the names internally, this
4760 * doesn't cause any problem.
4762 static int cgroup_destroy_locked(struct cgroup *cgrp)
4763 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4765 struct cgroup_subsys_state *css;
4769 lockdep_assert_held(&cgroup_mutex);
4772 * css_set_rwsem synchronizes access to ->cset_links and prevents
4773 * @cgrp from being removed while put_css_set() is in progress.
4775 down_read(&css_set_rwsem);
4776 empty = list_empty(&cgrp->cset_links);
4777 up_read(&css_set_rwsem);
4782 * Make sure there's no live children. We can't test emptiness of
4783 * ->self.children as dead children linger on it while being
4784 * drained; otherwise, "rmdir parent/child parent" may fail.
4786 if (css_has_online_children(&cgrp->self))
4790 * Mark @cgrp dead. This prevents further task migration and child
4791 * creation by disabling cgroup_lock_live_group().
4793 cgrp->self.flags &= ~CSS_ONLINE;
4795 /* initiate massacre of all css's */
4796 for_each_css(css, ssid, cgrp)
4800 * Remove @cgrp directory along with the base files. @cgrp has an
4801 * extra ref on its kn.
4803 kernfs_remove(cgrp->kn);
4805 check_for_release(cgroup_parent(cgrp));
4807 /* put the base reference */
4808 percpu_ref_kill(&cgrp->self.refcnt);
4813 static int cgroup_rmdir(struct kernfs_node *kn)
4815 struct cgroup *cgrp;
4818 cgrp = cgroup_kn_lock_live(kn);
4822 ret = cgroup_destroy_locked(cgrp);
4824 cgroup_kn_unlock(kn);
4828 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4829 .remount_fs = cgroup_remount,
4830 .show_options = cgroup_show_options,
4831 .mkdir = cgroup_mkdir,
4832 .rmdir = cgroup_rmdir,
4833 .rename = cgroup_rename,
4836 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4838 struct cgroup_subsys_state *css;
4840 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4842 mutex_lock(&cgroup_mutex);
4844 idr_init(&ss->css_idr);
4845 INIT_LIST_HEAD(&ss->cfts);
4847 /* Create the root cgroup state for this subsystem */
4848 ss->root = &cgrp_dfl_root;
4849 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4850 /* We don't handle early failures gracefully */
4851 BUG_ON(IS_ERR(css));
4852 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4855 * Root csses are never destroyed and we can't initialize
4856 * percpu_ref during early init. Disable refcnting.
4858 css->flags |= CSS_NO_REF;
4861 /* allocation can't be done safely during early init */
4864 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4865 BUG_ON(css->id < 0);
4868 /* Update the init_css_set to contain a subsys
4869 * pointer to this state - since the subsystem is
4870 * newly registered, all tasks and hence the
4871 * init_css_set is in the subsystem's root cgroup. */
4872 init_css_set.subsys[ss->id] = css;
4874 need_forkexit_callback |= ss->fork || ss->exit;
4876 /* At system boot, before all subsystems have been
4877 * registered, no tasks have been forked, so we don't
4878 * need to invoke fork callbacks here. */
4879 BUG_ON(!list_empty(&init_task.tasks));
4881 BUG_ON(online_css(css));
4883 mutex_unlock(&cgroup_mutex);
4887 * cgroup_init_early - cgroup initialization at system boot
4889 * Initialize cgroups at system boot, and initialize any
4890 * subsystems that request early init.
4892 int __init cgroup_init_early(void)
4894 static struct cgroup_sb_opts __initdata opts;
4895 struct cgroup_subsys *ss;
4898 init_cgroup_root(&cgrp_dfl_root, &opts);
4899 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4901 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4903 for_each_subsys(ss, i) {
4904 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4905 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4906 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4908 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4909 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4912 ss->name = cgroup_subsys_name[i];
4915 cgroup_init_subsys(ss, true);
4921 * cgroup_init - cgroup initialization
4923 * Register cgroup filesystem and /proc file, and initialize
4924 * any subsystems that didn't request early init.
4926 int __init cgroup_init(void)
4928 struct cgroup_subsys *ss;
4932 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
4933 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
4935 mutex_lock(&cgroup_mutex);
4937 /* Add init_css_set to the hash table */
4938 key = css_set_hash(init_css_set.subsys);
4939 hash_add(css_set_table, &init_css_set.hlist, key);
4941 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4943 mutex_unlock(&cgroup_mutex);
4945 for_each_subsys(ss, ssid) {
4946 if (ss->early_init) {
4947 struct cgroup_subsys_state *css =
4948 init_css_set.subsys[ss->id];
4950 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4952 BUG_ON(css->id < 0);
4954 cgroup_init_subsys(ss, false);
4957 list_add_tail(&init_css_set.e_cset_node[ssid],
4958 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4961 * Setting dfl_root subsys_mask needs to consider the
4962 * disabled flag and cftype registration needs kmalloc,
4963 * both of which aren't available during early_init.
4968 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4970 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
4971 ss->dfl_cftypes = ss->legacy_cftypes;
4973 if (!ss->dfl_cftypes)
4974 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
4976 if (ss->dfl_cftypes == ss->legacy_cftypes) {
4977 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
4979 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
4980 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
4984 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4988 err = register_filesystem(&cgroup_fs_type);
4990 kobject_put(cgroup_kobj);
4994 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4998 static int __init cgroup_wq_init(void)
5001 * There isn't much point in executing destruction path in
5002 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5003 * Use 1 for @max_active.
5005 * We would prefer to do this in cgroup_init() above, but that
5006 * is called before init_workqueues(): so leave this until after.
5008 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5009 BUG_ON(!cgroup_destroy_wq);
5012 * Used to destroy pidlists and separate to serve as flush domain.
5013 * Cap @max_active to 1 too.
5015 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5017 BUG_ON(!cgroup_pidlist_destroy_wq);
5021 core_initcall(cgroup_wq_init);
5024 * proc_cgroup_show()
5025 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5026 * - Used for /proc/<pid>/cgroup.
5028 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5029 struct pid *pid, struct task_struct *tsk)
5033 struct cgroup_root *root;
5036 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5040 mutex_lock(&cgroup_mutex);
5041 down_read(&css_set_rwsem);
5043 for_each_root(root) {
5044 struct cgroup_subsys *ss;
5045 struct cgroup *cgrp;
5046 int ssid, count = 0;
5048 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5051 seq_printf(m, "%d:", root->hierarchy_id);
5052 for_each_subsys(ss, ssid)
5053 if (root->subsys_mask & (1 << ssid))
5054 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5055 if (strlen(root->name))
5056 seq_printf(m, "%sname=%s", count ? "," : "",
5059 cgrp = task_cgroup_from_root(tsk, root);
5060 path = cgroup_path(cgrp, buf, PATH_MAX);
5062 retval = -ENAMETOOLONG;
5071 up_read(&css_set_rwsem);
5072 mutex_unlock(&cgroup_mutex);
5078 /* Display information about each subsystem and each hierarchy */
5079 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5081 struct cgroup_subsys *ss;
5084 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5086 * ideally we don't want subsystems moving around while we do this.
5087 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5088 * subsys/hierarchy state.
5090 mutex_lock(&cgroup_mutex);
5092 for_each_subsys(ss, i)
5093 seq_printf(m, "%s\t%d\t%d\t%d\n",
5094 ss->name, ss->root->hierarchy_id,
5095 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5097 mutex_unlock(&cgroup_mutex);
5101 static int cgroupstats_open(struct inode *inode, struct file *file)
5103 return single_open(file, proc_cgroupstats_show, NULL);
5106 static const struct file_operations proc_cgroupstats_operations = {
5107 .open = cgroupstats_open,
5109 .llseek = seq_lseek,
5110 .release = single_release,
5114 * cgroup_fork - initialize cgroup related fields during copy_process()
5115 * @child: pointer to task_struct of forking parent process.
5117 * A task is associated with the init_css_set until cgroup_post_fork()
5118 * attaches it to the parent's css_set. Empty cg_list indicates that
5119 * @child isn't holding reference to its css_set.
5121 void cgroup_fork(struct task_struct *child)
5123 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5124 INIT_LIST_HEAD(&child->cg_list);
5128 * cgroup_post_fork - called on a new task after adding it to the task list
5129 * @child: the task in question
5131 * Adds the task to the list running through its css_set if necessary and
5132 * call the subsystem fork() callbacks. Has to be after the task is
5133 * visible on the task list in case we race with the first call to
5134 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5137 void cgroup_post_fork(struct task_struct *child)
5139 struct cgroup_subsys *ss;
5143 * This may race against cgroup_enable_task_cg_lists(). As that
5144 * function sets use_task_css_set_links before grabbing
5145 * tasklist_lock and we just went through tasklist_lock to add
5146 * @child, it's guaranteed that either we see the set
5147 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5148 * @child during its iteration.
5150 * If we won the race, @child is associated with %current's
5151 * css_set. Grabbing css_set_rwsem guarantees both that the
5152 * association is stable, and, on completion of the parent's
5153 * migration, @child is visible in the source of migration or
5154 * already in the destination cgroup. This guarantee is necessary
5155 * when implementing operations which need to migrate all tasks of
5156 * a cgroup to another.
5158 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5159 * will remain in init_css_set. This is safe because all tasks are
5160 * in the init_css_set before cg_links is enabled and there's no
5161 * operation which transfers all tasks out of init_css_set.
5163 if (use_task_css_set_links) {
5164 struct css_set *cset;
5166 down_write(&css_set_rwsem);
5167 cset = task_css_set(current);
5168 if (list_empty(&child->cg_list)) {
5169 rcu_assign_pointer(child->cgroups, cset);
5170 list_add(&child->cg_list, &cset->tasks);
5173 up_write(&css_set_rwsem);
5177 * Call ss->fork(). This must happen after @child is linked on
5178 * css_set; otherwise, @child might change state between ->fork()
5179 * and addition to css_set.
5181 if (need_forkexit_callback) {
5182 for_each_subsys(ss, i)
5189 * cgroup_exit - detach cgroup from exiting task
5190 * @tsk: pointer to task_struct of exiting process
5192 * Description: Detach cgroup from @tsk and release it.
5194 * Note that cgroups marked notify_on_release force every task in
5195 * them to take the global cgroup_mutex mutex when exiting.
5196 * This could impact scaling on very large systems. Be reluctant to
5197 * use notify_on_release cgroups where very high task exit scaling
5198 * is required on large systems.
5200 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5201 * call cgroup_exit() while the task is still competent to handle
5202 * notify_on_release(), then leave the task attached to the root cgroup in
5203 * each hierarchy for the remainder of its exit. No need to bother with
5204 * init_css_set refcnting. init_css_set never goes away and we can't race
5205 * with migration path - PF_EXITING is visible to migration path.
5207 void cgroup_exit(struct task_struct *tsk)
5209 struct cgroup_subsys *ss;
5210 struct css_set *cset;
5211 bool put_cset = false;
5215 * Unlink from @tsk from its css_set. As migration path can't race
5216 * with us, we can check cg_list without grabbing css_set_rwsem.
5218 if (!list_empty(&tsk->cg_list)) {
5219 down_write(&css_set_rwsem);
5220 list_del_init(&tsk->cg_list);
5221 up_write(&css_set_rwsem);
5225 /* Reassign the task to the init_css_set. */
5226 cset = task_css_set(tsk);
5227 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5229 if (need_forkexit_callback) {
5230 /* see cgroup_post_fork() for details */
5231 for_each_subsys(ss, i) {
5233 struct cgroup_subsys_state *old_css = cset->subsys[i];
5234 struct cgroup_subsys_state *css = task_css(tsk, i);
5236 ss->exit(css, old_css, tsk);
5245 static void check_for_release(struct cgroup *cgrp)
5247 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5248 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5249 schedule_work(&cgrp->release_agent_work);
5253 * Notify userspace when a cgroup is released, by running the
5254 * configured release agent with the name of the cgroup (path
5255 * relative to the root of cgroup file system) as the argument.
5257 * Most likely, this user command will try to rmdir this cgroup.
5259 * This races with the possibility that some other task will be
5260 * attached to this cgroup before it is removed, or that some other
5261 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5262 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5263 * unused, and this cgroup will be reprieved from its death sentence,
5264 * to continue to serve a useful existence. Next time it's released,
5265 * we will get notified again, if it still has 'notify_on_release' set.
5267 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5268 * means only wait until the task is successfully execve()'d. The
5269 * separate release agent task is forked by call_usermodehelper(),
5270 * then control in this thread returns here, without waiting for the
5271 * release agent task. We don't bother to wait because the caller of
5272 * this routine has no use for the exit status of the release agent
5273 * task, so no sense holding our caller up for that.
5275 static void cgroup_release_agent(struct work_struct *work)
5277 struct cgroup *cgrp =
5278 container_of(work, struct cgroup, release_agent_work);
5279 char *pathbuf = NULL, *agentbuf = NULL, *path;
5280 char *argv[3], *envp[3];
5282 mutex_lock(&cgroup_mutex);
5284 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5285 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5286 if (!pathbuf || !agentbuf)
5289 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5297 /* minimal command environment */
5299 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5302 mutex_unlock(&cgroup_mutex);
5303 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5306 mutex_unlock(&cgroup_mutex);
5312 static int __init cgroup_disable(char *str)
5314 struct cgroup_subsys *ss;
5318 while ((token = strsep(&str, ",")) != NULL) {
5322 for_each_subsys(ss, i) {
5323 if (!strcmp(token, ss->name)) {
5325 printk(KERN_INFO "Disabling %s control group"
5326 " subsystem\n", ss->name);
5333 __setup("cgroup_disable=", cgroup_disable);
5335 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5337 printk("cgroup: using legacy files on the default hierarchy\n");
5338 cgroup_legacy_files_on_dfl = true;
5341 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5344 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5345 * @dentry: directory dentry of interest
5346 * @ss: subsystem of interest
5348 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5349 * to get the corresponding css and return it. If such css doesn't exist
5350 * or can't be pinned, an ERR_PTR value is returned.
5352 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5353 struct cgroup_subsys *ss)
5355 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5356 struct cgroup_subsys_state *css = NULL;
5357 struct cgroup *cgrp;
5359 /* is @dentry a cgroup dir? */
5360 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5361 kernfs_type(kn) != KERNFS_DIR)
5362 return ERR_PTR(-EBADF);
5367 * This path doesn't originate from kernfs and @kn could already
5368 * have been or be removed at any point. @kn->priv is RCU
5369 * protected for this access. See css_release_work_fn() for details.
5371 cgrp = rcu_dereference(kn->priv);
5373 css = cgroup_css(cgrp, ss);
5375 if (!css || !css_tryget_online(css))
5376 css = ERR_PTR(-ENOENT);
5383 * css_from_id - lookup css by id
5384 * @id: the cgroup id
5385 * @ss: cgroup subsys to be looked into
5387 * Returns the css if there's valid one with @id, otherwise returns NULL.
5388 * Should be called under rcu_read_lock().
5390 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5392 WARN_ON_ONCE(!rcu_read_lock_held());
5393 return idr_find(&ss->css_idr, id);
5396 #ifdef CONFIG_CGROUP_DEBUG
5397 static struct cgroup_subsys_state *
5398 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5400 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5403 return ERR_PTR(-ENOMEM);
5408 static void debug_css_free(struct cgroup_subsys_state *css)
5413 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5416 return cgroup_task_count(css->cgroup);
5419 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5422 return (u64)(unsigned long)current->cgroups;
5425 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5431 count = atomic_read(&task_css_set(current)->refcount);
5436 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5438 struct cgrp_cset_link *link;
5439 struct css_set *cset;
5442 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5446 down_read(&css_set_rwsem);
5448 cset = rcu_dereference(current->cgroups);
5449 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5450 struct cgroup *c = link->cgrp;
5452 cgroup_name(c, name_buf, NAME_MAX + 1);
5453 seq_printf(seq, "Root %d group %s\n",
5454 c->root->hierarchy_id, name_buf);
5457 up_read(&css_set_rwsem);
5462 #define MAX_TASKS_SHOWN_PER_CSS 25
5463 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5465 struct cgroup_subsys_state *css = seq_css(seq);
5466 struct cgrp_cset_link *link;
5468 down_read(&css_set_rwsem);
5469 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5470 struct css_set *cset = link->cset;
5471 struct task_struct *task;
5474 seq_printf(seq, "css_set %p\n", cset);
5476 list_for_each_entry(task, &cset->tasks, cg_list) {
5477 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5479 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5482 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5483 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5485 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5489 seq_puts(seq, " ...\n");
5491 up_read(&css_set_rwsem);
5495 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5497 return (!cgroup_has_tasks(css->cgroup) &&
5498 !css_has_online_children(&css->cgroup->self));
5501 static struct cftype debug_files[] = {
5503 .name = "taskcount",
5504 .read_u64 = debug_taskcount_read,
5508 .name = "current_css_set",
5509 .read_u64 = current_css_set_read,
5513 .name = "current_css_set_refcount",
5514 .read_u64 = current_css_set_refcount_read,
5518 .name = "current_css_set_cg_links",
5519 .seq_show = current_css_set_cg_links_read,
5523 .name = "cgroup_css_links",
5524 .seq_show = cgroup_css_links_read,
5528 .name = "releasable",
5529 .read_u64 = releasable_read,
5535 struct cgroup_subsys debug_cgrp_subsys = {
5536 .css_alloc = debug_css_alloc,
5537 .css_free = debug_css_free,
5538 .legacy_cftypes = debug_files,
5540 #endif /* CONFIG_CGROUP_DEBUG */