2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DECLARE_RWSEM(css_set_rwsem);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_rwsem);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DECLARE_RWSEM(css_set_rwsem);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 rcu_lockdep_assert(rcu_read_lock_held() || \
111 lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct *cgroup_destroy_wq;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
143 * The default hierarchy, reserved for the subsystems that are otherwise
144 * unattached - it never has more than a single cgroup, and all tasks are
145 * part of that cgroup.
147 struct cgroup_root cgrp_dfl_root;
150 * The default hierarchy always exists but is hidden until mounted for the
151 * first time. This is for backward compatibility.
153 static bool cgrp_dfl_root_visible;
156 * Set by the boot param of the same name and makes subsystems with NULL
157 * ->dfl_files to use ->legacy_files on the default hierarchy.
159 static bool cgroup_legacy_files_on_dfl;
161 /* some controllers are not supported in the default hierarchy */
162 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
164 /* The list of hierarchy roots */
166 static LIST_HEAD(cgroup_roots);
167 static int cgroup_root_count;
169 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
170 static DEFINE_IDR(cgroup_hierarchy_idr);
173 * Assign a monotonically increasing serial number to csses. It guarantees
174 * cgroups with bigger numbers are newer than those with smaller numbers.
175 * Also, as csses are always appended to the parent's ->children list, it
176 * guarantees that sibling csses are always sorted in the ascending serial
177 * number order on the list. Protected by cgroup_mutex.
179 static u64 css_serial_nr_next = 1;
182 * These bitmask flags indicate whether tasks in the fork and exit paths have
183 * fork/exit handlers to call. This avoids us having to do extra work in the
184 * fork/exit path to check which subsystems have fork/exit callbacks.
186 static unsigned long have_fork_callback __read_mostly;
187 static unsigned long have_exit_callback __read_mostly;
189 /* Ditto for the can_fork callback. */
190 static unsigned long have_canfork_callback __read_mostly;
192 static struct cftype cgroup_dfl_base_files[];
193 static struct cftype cgroup_legacy_base_files[];
195 static int rebind_subsystems(struct cgroup_root *dst_root,
196 unsigned long ss_mask);
197 static int cgroup_destroy_locked(struct cgroup *cgrp);
198 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
200 static void css_release(struct percpu_ref *ref);
201 static void kill_css(struct cgroup_subsys_state *css);
202 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
205 /* IDR wrappers which synchronize using cgroup_idr_lock */
206 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
211 idr_preload(gfp_mask);
212 spin_lock_bh(&cgroup_idr_lock);
213 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
214 spin_unlock_bh(&cgroup_idr_lock);
219 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
223 spin_lock_bh(&cgroup_idr_lock);
224 ret = idr_replace(idr, ptr, id);
225 spin_unlock_bh(&cgroup_idr_lock);
229 static void cgroup_idr_remove(struct idr *idr, int id)
231 spin_lock_bh(&cgroup_idr_lock);
233 spin_unlock_bh(&cgroup_idr_lock);
236 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
238 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
241 return container_of(parent_css, struct cgroup, self);
246 * cgroup_css - obtain a cgroup's css for the specified subsystem
247 * @cgrp: the cgroup of interest
248 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
250 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
251 * function must be called either under cgroup_mutex or rcu_read_lock() and
252 * the caller is responsible for pinning the returned css if it wants to
253 * keep accessing it outside the said locks. This function may return
254 * %NULL if @cgrp doesn't have @subsys_id enabled.
256 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
257 struct cgroup_subsys *ss)
260 return rcu_dereference_check(cgrp->subsys[ss->id],
261 lockdep_is_held(&cgroup_mutex));
267 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
268 * @cgrp: the cgroup of interest
269 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
271 * Similar to cgroup_css() but returns the effective css, which is defined
272 * as the matching css of the nearest ancestor including self which has @ss
273 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
274 * function is guaranteed to return non-NULL css.
276 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
277 struct cgroup_subsys *ss)
279 lockdep_assert_held(&cgroup_mutex);
284 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
288 * This function is used while updating css associations and thus
289 * can't test the csses directly. Use ->child_subsys_mask.
291 while (cgroup_parent(cgrp) &&
292 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
293 cgrp = cgroup_parent(cgrp);
295 return cgroup_css(cgrp, ss);
299 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
300 * @cgrp: the cgroup of interest
301 * @ss: the subsystem of interest
303 * Find and get the effective css of @cgrp for @ss. The effective css is
304 * defined as the matching css of the nearest ancestor including self which
305 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
306 * the root css is returned, so this function always returns a valid css.
307 * The returned css must be put using css_put().
309 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
310 struct cgroup_subsys *ss)
312 struct cgroup_subsys_state *css;
317 css = cgroup_css(cgrp, ss);
319 if (css && css_tryget_online(css))
321 cgrp = cgroup_parent(cgrp);
324 css = init_css_set.subsys[ss->id];
331 /* convenient tests for these bits */
332 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
334 return !(cgrp->self.flags & CSS_ONLINE);
337 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
339 struct cgroup *cgrp = of->kn->parent->priv;
340 struct cftype *cft = of_cft(of);
343 * This is open and unprotected implementation of cgroup_css().
344 * seq_css() is only called from a kernfs file operation which has
345 * an active reference on the file. Because all the subsystem
346 * files are drained before a css is disassociated with a cgroup,
347 * the matching css from the cgroup's subsys table is guaranteed to
348 * be and stay valid until the enclosing operation is complete.
351 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
355 EXPORT_SYMBOL_GPL(of_css);
358 * cgroup_is_descendant - test ancestry
359 * @cgrp: the cgroup to be tested
360 * @ancestor: possible ancestor of @cgrp
362 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
363 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
364 * and @ancestor are accessible.
366 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
369 if (cgrp == ancestor)
371 cgrp = cgroup_parent(cgrp);
376 static int notify_on_release(const struct cgroup *cgrp)
378 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
382 * for_each_css - iterate all css's of a cgroup
383 * @css: the iteration cursor
384 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
385 * @cgrp: the target cgroup to iterate css's of
387 * Should be called under cgroup_[tree_]mutex.
389 #define for_each_css(css, ssid, cgrp) \
390 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
391 if (!((css) = rcu_dereference_check( \
392 (cgrp)->subsys[(ssid)], \
393 lockdep_is_held(&cgroup_mutex)))) { } \
397 * for_each_e_css - iterate all effective css's of a cgroup
398 * @css: the iteration cursor
399 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
400 * @cgrp: the target cgroup to iterate css's of
402 * Should be called under cgroup_[tree_]mutex.
404 #define for_each_e_css(css, ssid, cgrp) \
405 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
406 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
411 * for_each_subsys - iterate all enabled cgroup subsystems
412 * @ss: the iteration cursor
413 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
415 #define for_each_subsys(ss, ssid) \
416 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
417 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
420 * for_each_subsys_which - filter for_each_subsys with a bitmask
421 * @ss: the iteration cursor
422 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
423 * @ss_maskp: a pointer to the bitmask
425 * The block will only run for cases where the ssid-th bit (1 << ssid) of
428 #define for_each_subsys_which(ss, ssid, ss_maskp) \
429 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
432 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
433 if (((ss) = cgroup_subsys[ssid]) && false) \
437 /* iterate across the hierarchies */
438 #define for_each_root(root) \
439 list_for_each_entry((root), &cgroup_roots, root_list)
441 /* iterate over child cgrps, lock should be held throughout iteration */
442 #define cgroup_for_each_live_child(child, cgrp) \
443 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
444 if (({ lockdep_assert_held(&cgroup_mutex); \
445 cgroup_is_dead(child); })) \
449 static void cgroup_release_agent(struct work_struct *work);
450 static void check_for_release(struct cgroup *cgrp);
453 * A cgroup can be associated with multiple css_sets as different tasks may
454 * belong to different cgroups on different hierarchies. In the other
455 * direction, a css_set is naturally associated with multiple cgroups.
456 * This M:N relationship is represented by the following link structure
457 * which exists for each association and allows traversing the associations
460 struct cgrp_cset_link {
461 /* the cgroup and css_set this link associates */
463 struct css_set *cset;
465 /* list of cgrp_cset_links anchored at cgrp->cset_links */
466 struct list_head cset_link;
468 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
469 struct list_head cgrp_link;
473 * The default css_set - used by init and its children prior to any
474 * hierarchies being mounted. It contains a pointer to the root state
475 * for each subsystem. Also used to anchor the list of css_sets. Not
476 * reference-counted, to improve performance when child cgroups
477 * haven't been created.
479 struct css_set init_css_set = {
480 .refcount = ATOMIC_INIT(1),
481 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
482 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
483 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
484 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
485 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
488 static int css_set_count = 1; /* 1 for init_css_set */
491 * cgroup_update_populated - updated populated count of a cgroup
492 * @cgrp: the target cgroup
493 * @populated: inc or dec populated count
495 * @cgrp is either getting the first task (css_set) or losing the last.
496 * Update @cgrp->populated_cnt accordingly. The count is propagated
497 * towards root so that a given cgroup's populated_cnt is zero iff the
498 * cgroup and all its descendants are empty.
500 * @cgrp's interface file "cgroup.populated" is zero if
501 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
502 * changes from or to zero, userland is notified that the content of the
503 * interface file has changed. This can be used to detect when @cgrp and
504 * its descendants become populated or empty.
506 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
508 lockdep_assert_held(&css_set_rwsem);
514 trigger = !cgrp->populated_cnt++;
516 trigger = !--cgrp->populated_cnt;
521 if (cgrp->populated_kn)
522 kernfs_notify(cgrp->populated_kn);
523 cgrp = cgroup_parent(cgrp);
528 * hash table for cgroup groups. This improves the performance to find
529 * an existing css_set. This hash doesn't (currently) take into
530 * account cgroups in empty hierarchies.
532 #define CSS_SET_HASH_BITS 7
533 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
535 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
537 unsigned long key = 0UL;
538 struct cgroup_subsys *ss;
541 for_each_subsys(ss, i)
542 key += (unsigned long)css[i];
543 key = (key >> 16) ^ key;
548 static void put_css_set_locked(struct css_set *cset)
550 struct cgrp_cset_link *link, *tmp_link;
551 struct cgroup_subsys *ss;
554 lockdep_assert_held(&css_set_rwsem);
556 if (!atomic_dec_and_test(&cset->refcount))
559 /* This css_set is dead. unlink it and release cgroup refcounts */
560 for_each_subsys(ss, ssid)
561 list_del(&cset->e_cset_node[ssid]);
562 hash_del(&cset->hlist);
565 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
566 struct cgroup *cgrp = link->cgrp;
568 list_del(&link->cset_link);
569 list_del(&link->cgrp_link);
571 /* @cgrp can't go away while we're holding css_set_rwsem */
572 if (list_empty(&cgrp->cset_links)) {
573 cgroup_update_populated(cgrp, false);
574 check_for_release(cgrp);
580 kfree_rcu(cset, rcu_head);
583 static void put_css_set(struct css_set *cset)
586 * Ensure that the refcount doesn't hit zero while any readers
587 * can see it. Similar to atomic_dec_and_lock(), but for an
590 if (atomic_add_unless(&cset->refcount, -1, 1))
593 down_write(&css_set_rwsem);
594 put_css_set_locked(cset);
595 up_write(&css_set_rwsem);
599 * refcounted get/put for css_set objects
601 static inline void get_css_set(struct css_set *cset)
603 atomic_inc(&cset->refcount);
607 * compare_css_sets - helper function for find_existing_css_set().
608 * @cset: candidate css_set being tested
609 * @old_cset: existing css_set for a task
610 * @new_cgrp: cgroup that's being entered by the task
611 * @template: desired set of css pointers in css_set (pre-calculated)
613 * Returns true if "cset" matches "old_cset" except for the hierarchy
614 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
616 static bool compare_css_sets(struct css_set *cset,
617 struct css_set *old_cset,
618 struct cgroup *new_cgrp,
619 struct cgroup_subsys_state *template[])
621 struct list_head *l1, *l2;
624 * On the default hierarchy, there can be csets which are
625 * associated with the same set of cgroups but different csses.
626 * Let's first ensure that csses match.
628 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
632 * Compare cgroup pointers in order to distinguish between
633 * different cgroups in hierarchies. As different cgroups may
634 * share the same effective css, this comparison is always
637 l1 = &cset->cgrp_links;
638 l2 = &old_cset->cgrp_links;
640 struct cgrp_cset_link *link1, *link2;
641 struct cgroup *cgrp1, *cgrp2;
645 /* See if we reached the end - both lists are equal length. */
646 if (l1 == &cset->cgrp_links) {
647 BUG_ON(l2 != &old_cset->cgrp_links);
650 BUG_ON(l2 == &old_cset->cgrp_links);
652 /* Locate the cgroups associated with these links. */
653 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
654 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
657 /* Hierarchies should be linked in the same order. */
658 BUG_ON(cgrp1->root != cgrp2->root);
661 * If this hierarchy is the hierarchy of the cgroup
662 * that's changing, then we need to check that this
663 * css_set points to the new cgroup; if it's any other
664 * hierarchy, then this css_set should point to the
665 * same cgroup as the old css_set.
667 if (cgrp1->root == new_cgrp->root) {
668 if (cgrp1 != new_cgrp)
679 * find_existing_css_set - init css array and find the matching css_set
680 * @old_cset: the css_set that we're using before the cgroup transition
681 * @cgrp: the cgroup that we're moving into
682 * @template: out param for the new set of csses, should be clear on entry
684 static struct css_set *find_existing_css_set(struct css_set *old_cset,
686 struct cgroup_subsys_state *template[])
688 struct cgroup_root *root = cgrp->root;
689 struct cgroup_subsys *ss;
690 struct css_set *cset;
695 * Build the set of subsystem state objects that we want to see in the
696 * new css_set. while subsystems can change globally, the entries here
697 * won't change, so no need for locking.
699 for_each_subsys(ss, i) {
700 if (root->subsys_mask & (1UL << i)) {
702 * @ss is in this hierarchy, so we want the
703 * effective css from @cgrp.
705 template[i] = cgroup_e_css(cgrp, ss);
708 * @ss is not in this hierarchy, so we don't want
711 template[i] = old_cset->subsys[i];
715 key = css_set_hash(template);
716 hash_for_each_possible(css_set_table, cset, hlist, key) {
717 if (!compare_css_sets(cset, old_cset, cgrp, template))
720 /* This css_set matches what we need */
724 /* No existing cgroup group matched */
728 static void free_cgrp_cset_links(struct list_head *links_to_free)
730 struct cgrp_cset_link *link, *tmp_link;
732 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
733 list_del(&link->cset_link);
739 * allocate_cgrp_cset_links - allocate cgrp_cset_links
740 * @count: the number of links to allocate
741 * @tmp_links: list_head the allocated links are put on
743 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
744 * through ->cset_link. Returns 0 on success or -errno.
746 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
748 struct cgrp_cset_link *link;
751 INIT_LIST_HEAD(tmp_links);
753 for (i = 0; i < count; i++) {
754 link = kzalloc(sizeof(*link), GFP_KERNEL);
756 free_cgrp_cset_links(tmp_links);
759 list_add(&link->cset_link, tmp_links);
765 * link_css_set - a helper function to link a css_set to a cgroup
766 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
767 * @cset: the css_set to be linked
768 * @cgrp: the destination cgroup
770 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
773 struct cgrp_cset_link *link;
775 BUG_ON(list_empty(tmp_links));
777 if (cgroup_on_dfl(cgrp))
778 cset->dfl_cgrp = cgrp;
780 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
784 if (list_empty(&cgrp->cset_links))
785 cgroup_update_populated(cgrp, true);
786 list_move(&link->cset_link, &cgrp->cset_links);
789 * Always add links to the tail of the list so that the list
790 * is sorted by order of hierarchy creation
792 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
796 * find_css_set - return a new css_set with one cgroup updated
797 * @old_cset: the baseline css_set
798 * @cgrp: the cgroup to be updated
800 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
801 * substituted into the appropriate hierarchy.
803 static struct css_set *find_css_set(struct css_set *old_cset,
806 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
807 struct css_set *cset;
808 struct list_head tmp_links;
809 struct cgrp_cset_link *link;
810 struct cgroup_subsys *ss;
814 lockdep_assert_held(&cgroup_mutex);
816 /* First see if we already have a cgroup group that matches
818 down_read(&css_set_rwsem);
819 cset = find_existing_css_set(old_cset, cgrp, template);
822 up_read(&css_set_rwsem);
827 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
831 /* Allocate all the cgrp_cset_link objects that we'll need */
832 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
837 atomic_set(&cset->refcount, 1);
838 INIT_LIST_HEAD(&cset->cgrp_links);
839 INIT_LIST_HEAD(&cset->tasks);
840 INIT_LIST_HEAD(&cset->mg_tasks);
841 INIT_LIST_HEAD(&cset->mg_preload_node);
842 INIT_LIST_HEAD(&cset->mg_node);
843 INIT_HLIST_NODE(&cset->hlist);
845 /* Copy the set of subsystem state objects generated in
846 * find_existing_css_set() */
847 memcpy(cset->subsys, template, sizeof(cset->subsys));
849 down_write(&css_set_rwsem);
850 /* Add reference counts and links from the new css_set. */
851 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
852 struct cgroup *c = link->cgrp;
854 if (c->root == cgrp->root)
856 link_css_set(&tmp_links, cset, c);
859 BUG_ON(!list_empty(&tmp_links));
863 /* Add @cset to the hash table */
864 key = css_set_hash(cset->subsys);
865 hash_add(css_set_table, &cset->hlist, key);
867 for_each_subsys(ss, ssid)
868 list_add_tail(&cset->e_cset_node[ssid],
869 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
871 up_write(&css_set_rwsem);
876 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
878 struct cgroup *root_cgrp = kf_root->kn->priv;
880 return root_cgrp->root;
883 static int cgroup_init_root_id(struct cgroup_root *root)
887 lockdep_assert_held(&cgroup_mutex);
889 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
893 root->hierarchy_id = id;
897 static void cgroup_exit_root_id(struct cgroup_root *root)
899 lockdep_assert_held(&cgroup_mutex);
901 if (root->hierarchy_id) {
902 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
903 root->hierarchy_id = 0;
907 static void cgroup_free_root(struct cgroup_root *root)
910 /* hierarchy ID should already have been released */
911 WARN_ON_ONCE(root->hierarchy_id);
913 idr_destroy(&root->cgroup_idr);
918 static void cgroup_destroy_root(struct cgroup_root *root)
920 struct cgroup *cgrp = &root->cgrp;
921 struct cgrp_cset_link *link, *tmp_link;
923 mutex_lock(&cgroup_mutex);
925 BUG_ON(atomic_read(&root->nr_cgrps));
926 BUG_ON(!list_empty(&cgrp->self.children));
928 /* Rebind all subsystems back to the default hierarchy */
929 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
932 * Release all the links from cset_links to this hierarchy's
935 down_write(&css_set_rwsem);
937 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
938 list_del(&link->cset_link);
939 list_del(&link->cgrp_link);
942 up_write(&css_set_rwsem);
944 if (!list_empty(&root->root_list)) {
945 list_del(&root->root_list);
949 cgroup_exit_root_id(root);
951 mutex_unlock(&cgroup_mutex);
953 kernfs_destroy_root(root->kf_root);
954 cgroup_free_root(root);
957 /* look up cgroup associated with given css_set on the specified hierarchy */
958 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
959 struct cgroup_root *root)
961 struct cgroup *res = NULL;
963 lockdep_assert_held(&cgroup_mutex);
964 lockdep_assert_held(&css_set_rwsem);
966 if (cset == &init_css_set) {
969 struct cgrp_cset_link *link;
971 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
972 struct cgroup *c = link->cgrp;
974 if (c->root == root) {
986 * Return the cgroup for "task" from the given hierarchy. Must be
987 * called with cgroup_mutex and css_set_rwsem held.
989 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
990 struct cgroup_root *root)
993 * No need to lock the task - since we hold cgroup_mutex the
994 * task can't change groups, so the only thing that can happen
995 * is that it exits and its css is set back to init_css_set.
997 return cset_cgroup_from_root(task_css_set(task), root);
1001 * A task must hold cgroup_mutex to modify cgroups.
1003 * Any task can increment and decrement the count field without lock.
1004 * So in general, code holding cgroup_mutex can't rely on the count
1005 * field not changing. However, if the count goes to zero, then only
1006 * cgroup_attach_task() can increment it again. Because a count of zero
1007 * means that no tasks are currently attached, therefore there is no
1008 * way a task attached to that cgroup can fork (the other way to
1009 * increment the count). So code holding cgroup_mutex can safely
1010 * assume that if the count is zero, it will stay zero. Similarly, if
1011 * a task holds cgroup_mutex on a cgroup with zero count, it
1012 * knows that the cgroup won't be removed, as cgroup_rmdir()
1015 * A cgroup can only be deleted if both its 'count' of using tasks
1016 * is zero, and its list of 'children' cgroups is empty. Since all
1017 * tasks in the system use _some_ cgroup, and since there is always at
1018 * least one task in the system (init, pid == 1), therefore, root cgroup
1019 * always has either children cgroups and/or using tasks. So we don't
1020 * need a special hack to ensure that root cgroup cannot be deleted.
1022 * P.S. One more locking exception. RCU is used to guard the
1023 * update of a tasks cgroup pointer by cgroup_attach_task()
1026 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
1027 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1028 static const struct file_operations proc_cgroupstats_operations;
1030 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1033 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1034 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1035 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1036 cft->ss->name, cft->name);
1038 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1043 * cgroup_file_mode - deduce file mode of a control file
1044 * @cft: the control file in question
1046 * returns cft->mode if ->mode is not 0
1047 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1048 * returns S_IRUGO if it has only a read handler
1049 * returns S_IWUSR if it has only a write hander
1051 static umode_t cgroup_file_mode(const struct cftype *cft)
1058 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1061 if (cft->write_u64 || cft->write_s64 || cft->write)
1067 static void cgroup_get(struct cgroup *cgrp)
1069 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1070 css_get(&cgrp->self);
1073 static bool cgroup_tryget(struct cgroup *cgrp)
1075 return css_tryget(&cgrp->self);
1078 static void cgroup_put(struct cgroup *cgrp)
1080 css_put(&cgrp->self);
1084 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1085 * @cgrp: the target cgroup
1086 * @subtree_control: the new subtree_control mask to consider
1088 * On the default hierarchy, a subsystem may request other subsystems to be
1089 * enabled together through its ->depends_on mask. In such cases, more
1090 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1092 * This function calculates which subsystems need to be enabled if
1093 * @subtree_control is to be applied to @cgrp. The returned mask is always
1094 * a superset of @subtree_control and follows the usual hierarchy rules.
1096 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1097 unsigned long subtree_control)
1099 struct cgroup *parent = cgroup_parent(cgrp);
1100 unsigned long cur_ss_mask = subtree_control;
1101 struct cgroup_subsys *ss;
1104 lockdep_assert_held(&cgroup_mutex);
1106 if (!cgroup_on_dfl(cgrp))
1110 unsigned long new_ss_mask = cur_ss_mask;
1112 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1113 new_ss_mask |= ss->depends_on;
1116 * Mask out subsystems which aren't available. This can
1117 * happen only if some depended-upon subsystems were bound
1118 * to non-default hierarchies.
1121 new_ss_mask &= parent->child_subsys_mask;
1123 new_ss_mask &= cgrp->root->subsys_mask;
1125 if (new_ss_mask == cur_ss_mask)
1127 cur_ss_mask = new_ss_mask;
1134 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1135 * @cgrp: the target cgroup
1137 * Update @cgrp->child_subsys_mask according to the current
1138 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1140 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1142 cgrp->child_subsys_mask =
1143 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1147 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1148 * @kn: the kernfs_node being serviced
1150 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1151 * the method finishes if locking succeeded. Note that once this function
1152 * returns the cgroup returned by cgroup_kn_lock_live() may become
1153 * inaccessible any time. If the caller intends to continue to access the
1154 * cgroup, it should pin it before invoking this function.
1156 static void cgroup_kn_unlock(struct kernfs_node *kn)
1158 struct cgroup *cgrp;
1160 if (kernfs_type(kn) == KERNFS_DIR)
1163 cgrp = kn->parent->priv;
1165 mutex_unlock(&cgroup_mutex);
1167 kernfs_unbreak_active_protection(kn);
1172 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1173 * @kn: the kernfs_node being serviced
1175 * This helper is to be used by a cgroup kernfs method currently servicing
1176 * @kn. It breaks the active protection, performs cgroup locking and
1177 * verifies that the associated cgroup is alive. Returns the cgroup if
1178 * alive; otherwise, %NULL. A successful return should be undone by a
1179 * matching cgroup_kn_unlock() invocation.
1181 * Any cgroup kernfs method implementation which requires locking the
1182 * associated cgroup should use this helper. It avoids nesting cgroup
1183 * locking under kernfs active protection and allows all kernfs operations
1184 * including self-removal.
1186 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1188 struct cgroup *cgrp;
1190 if (kernfs_type(kn) == KERNFS_DIR)
1193 cgrp = kn->parent->priv;
1196 * We're gonna grab cgroup_mutex which nests outside kernfs
1197 * active_ref. cgroup liveliness check alone provides enough
1198 * protection against removal. Ensure @cgrp stays accessible and
1199 * break the active_ref protection.
1201 if (!cgroup_tryget(cgrp))
1203 kernfs_break_active_protection(kn);
1205 mutex_lock(&cgroup_mutex);
1207 if (!cgroup_is_dead(cgrp))
1210 cgroup_kn_unlock(kn);
1214 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1216 char name[CGROUP_FILE_NAME_MAX];
1218 lockdep_assert_held(&cgroup_mutex);
1219 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1223 * cgroup_clear_dir - remove subsys files in a cgroup directory
1224 * @cgrp: target cgroup
1225 * @subsys_mask: mask of the subsystem ids whose files should be removed
1227 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1229 struct cgroup_subsys *ss;
1232 for_each_subsys(ss, i) {
1233 struct cftype *cfts;
1235 if (!(subsys_mask & (1 << i)))
1237 list_for_each_entry(cfts, &ss->cfts, node)
1238 cgroup_addrm_files(cgrp, cfts, false);
1242 static int rebind_subsystems(struct cgroup_root *dst_root,
1243 unsigned long ss_mask)
1245 struct cgroup_subsys *ss;
1246 unsigned long tmp_ss_mask;
1249 lockdep_assert_held(&cgroup_mutex);
1251 for_each_subsys_which(ss, ssid, &ss_mask) {
1252 /* if @ss has non-root csses attached to it, can't move */
1253 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1256 /* can't move between two non-dummy roots either */
1257 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1261 /* skip creating root files on dfl_root for inhibited subsystems */
1262 tmp_ss_mask = ss_mask;
1263 if (dst_root == &cgrp_dfl_root)
1264 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1266 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1268 if (dst_root != &cgrp_dfl_root)
1272 * Rebinding back to the default root is not allowed to
1273 * fail. Using both default and non-default roots should
1274 * be rare. Moving subsystems back and forth even more so.
1275 * Just warn about it and continue.
1277 if (cgrp_dfl_root_visible) {
1278 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1280 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1285 * Nothing can fail from this point on. Remove files for the
1286 * removed subsystems and rebind each subsystem.
1288 for_each_subsys_which(ss, ssid, &ss_mask)
1289 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1291 for_each_subsys_which(ss, ssid, &ss_mask) {
1292 struct cgroup_root *src_root;
1293 struct cgroup_subsys_state *css;
1294 struct css_set *cset;
1296 src_root = ss->root;
1297 css = cgroup_css(&src_root->cgrp, ss);
1299 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1301 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1302 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1303 ss->root = dst_root;
1304 css->cgroup = &dst_root->cgrp;
1306 down_write(&css_set_rwsem);
1307 hash_for_each(css_set_table, i, cset, hlist)
1308 list_move_tail(&cset->e_cset_node[ss->id],
1309 &dst_root->cgrp.e_csets[ss->id]);
1310 up_write(&css_set_rwsem);
1312 src_root->subsys_mask &= ~(1 << ssid);
1313 src_root->cgrp.subtree_control &= ~(1 << ssid);
1314 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1316 /* default hierarchy doesn't enable controllers by default */
1317 dst_root->subsys_mask |= 1 << ssid;
1318 if (dst_root != &cgrp_dfl_root) {
1319 dst_root->cgrp.subtree_control |= 1 << ssid;
1320 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1327 kernfs_activate(dst_root->cgrp.kn);
1331 static int cgroup_show_options(struct seq_file *seq,
1332 struct kernfs_root *kf_root)
1334 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1335 struct cgroup_subsys *ss;
1338 for_each_subsys(ss, ssid)
1339 if (root->subsys_mask & (1 << ssid))
1340 seq_printf(seq, ",%s", ss->name);
1341 if (root->flags & CGRP_ROOT_NOPREFIX)
1342 seq_puts(seq, ",noprefix");
1343 if (root->flags & CGRP_ROOT_XATTR)
1344 seq_puts(seq, ",xattr");
1346 spin_lock(&release_agent_path_lock);
1347 if (strlen(root->release_agent_path))
1348 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1349 spin_unlock(&release_agent_path_lock);
1351 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1352 seq_puts(seq, ",clone_children");
1353 if (strlen(root->name))
1354 seq_printf(seq, ",name=%s", root->name);
1358 struct cgroup_sb_opts {
1359 unsigned long subsys_mask;
1361 char *release_agent;
1362 bool cpuset_clone_children;
1364 /* User explicitly requested empty subsystem */
1368 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1370 char *token, *o = data;
1371 bool all_ss = false, one_ss = false;
1372 unsigned long mask = -1UL;
1373 struct cgroup_subsys *ss;
1377 #ifdef CONFIG_CPUSETS
1378 mask = ~(1U << cpuset_cgrp_id);
1381 memset(opts, 0, sizeof(*opts));
1383 while ((token = strsep(&o, ",")) != NULL) {
1388 if (!strcmp(token, "none")) {
1389 /* Explicitly have no subsystems */
1393 if (!strcmp(token, "all")) {
1394 /* Mutually exclusive option 'all' + subsystem name */
1400 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1401 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1404 if (!strcmp(token, "noprefix")) {
1405 opts->flags |= CGRP_ROOT_NOPREFIX;
1408 if (!strcmp(token, "clone_children")) {
1409 opts->cpuset_clone_children = true;
1412 if (!strcmp(token, "xattr")) {
1413 opts->flags |= CGRP_ROOT_XATTR;
1416 if (!strncmp(token, "release_agent=", 14)) {
1417 /* Specifying two release agents is forbidden */
1418 if (opts->release_agent)
1420 opts->release_agent =
1421 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1422 if (!opts->release_agent)
1426 if (!strncmp(token, "name=", 5)) {
1427 const char *name = token + 5;
1428 /* Can't specify an empty name */
1431 /* Must match [\w.-]+ */
1432 for (i = 0; i < strlen(name); i++) {
1436 if ((c == '.') || (c == '-') || (c == '_'))
1440 /* Specifying two names is forbidden */
1443 opts->name = kstrndup(name,
1444 MAX_CGROUP_ROOT_NAMELEN - 1,
1452 for_each_subsys(ss, i) {
1453 if (strcmp(token, ss->name))
1458 /* Mutually exclusive option 'all' + subsystem name */
1461 opts->subsys_mask |= (1 << i);
1466 if (i == CGROUP_SUBSYS_COUNT)
1470 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1471 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1473 pr_err("sane_behavior: no other mount options allowed\n");
1480 * If the 'all' option was specified select all the subsystems,
1481 * otherwise if 'none', 'name=' and a subsystem name options were
1482 * not specified, let's default to 'all'
1484 if (all_ss || (!one_ss && !opts->none && !opts->name))
1485 for_each_subsys(ss, i)
1487 opts->subsys_mask |= (1 << i);
1490 * We either have to specify by name or by subsystems. (So all
1491 * empty hierarchies must have a name).
1493 if (!opts->subsys_mask && !opts->name)
1497 * Option noprefix was introduced just for backward compatibility
1498 * with the old cpuset, so we allow noprefix only if mounting just
1499 * the cpuset subsystem.
1501 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1504 /* Can't specify "none" and some subsystems */
1505 if (opts->subsys_mask && opts->none)
1511 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1514 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1515 struct cgroup_sb_opts opts;
1516 unsigned long added_mask, removed_mask;
1518 if (root == &cgrp_dfl_root) {
1519 pr_err("remount is not allowed\n");
1523 mutex_lock(&cgroup_mutex);
1525 /* See what subsystems are wanted */
1526 ret = parse_cgroupfs_options(data, &opts);
1530 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1531 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1532 task_tgid_nr(current), current->comm);
1534 added_mask = opts.subsys_mask & ~root->subsys_mask;
1535 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1537 /* Don't allow flags or name to change at remount */
1538 if ((opts.flags ^ root->flags) ||
1539 (opts.name && strcmp(opts.name, root->name))) {
1540 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1541 opts.flags, opts.name ?: "", root->flags, root->name);
1546 /* remounting is not allowed for populated hierarchies */
1547 if (!list_empty(&root->cgrp.self.children)) {
1552 ret = rebind_subsystems(root, added_mask);
1556 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1558 if (opts.release_agent) {
1559 spin_lock(&release_agent_path_lock);
1560 strcpy(root->release_agent_path, opts.release_agent);
1561 spin_unlock(&release_agent_path_lock);
1564 kfree(opts.release_agent);
1566 mutex_unlock(&cgroup_mutex);
1571 * To reduce the fork() overhead for systems that are not actually using
1572 * their cgroups capability, we don't maintain the lists running through
1573 * each css_set to its tasks until we see the list actually used - in other
1574 * words after the first mount.
1576 static bool use_task_css_set_links __read_mostly;
1578 static void cgroup_enable_task_cg_lists(void)
1580 struct task_struct *p, *g;
1582 down_write(&css_set_rwsem);
1584 if (use_task_css_set_links)
1587 use_task_css_set_links = true;
1590 * We need tasklist_lock because RCU is not safe against
1591 * while_each_thread(). Besides, a forking task that has passed
1592 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1593 * is not guaranteed to have its child immediately visible in the
1594 * tasklist if we walk through it with RCU.
1596 read_lock(&tasklist_lock);
1597 do_each_thread(g, p) {
1598 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1599 task_css_set(p) != &init_css_set);
1602 * We should check if the process is exiting, otherwise
1603 * it will race with cgroup_exit() in that the list
1604 * entry won't be deleted though the process has exited.
1605 * Do it while holding siglock so that we don't end up
1606 * racing against cgroup_exit().
1608 spin_lock_irq(&p->sighand->siglock);
1609 if (!(p->flags & PF_EXITING)) {
1610 struct css_set *cset = task_css_set(p);
1612 list_add(&p->cg_list, &cset->tasks);
1615 spin_unlock_irq(&p->sighand->siglock);
1616 } while_each_thread(g, p);
1617 read_unlock(&tasklist_lock);
1619 up_write(&css_set_rwsem);
1622 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1624 struct cgroup_subsys *ss;
1627 INIT_LIST_HEAD(&cgrp->self.sibling);
1628 INIT_LIST_HEAD(&cgrp->self.children);
1629 INIT_LIST_HEAD(&cgrp->cset_links);
1630 INIT_LIST_HEAD(&cgrp->pidlists);
1631 mutex_init(&cgrp->pidlist_mutex);
1632 cgrp->self.cgroup = cgrp;
1633 cgrp->self.flags |= CSS_ONLINE;
1635 for_each_subsys(ss, ssid)
1636 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1638 init_waitqueue_head(&cgrp->offline_waitq);
1639 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1642 static void init_cgroup_root(struct cgroup_root *root,
1643 struct cgroup_sb_opts *opts)
1645 struct cgroup *cgrp = &root->cgrp;
1647 INIT_LIST_HEAD(&root->root_list);
1648 atomic_set(&root->nr_cgrps, 1);
1650 init_cgroup_housekeeping(cgrp);
1651 idr_init(&root->cgroup_idr);
1653 root->flags = opts->flags;
1654 if (opts->release_agent)
1655 strcpy(root->release_agent_path, opts->release_agent);
1657 strcpy(root->name, opts->name);
1658 if (opts->cpuset_clone_children)
1659 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1662 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1664 LIST_HEAD(tmp_links);
1665 struct cgroup *root_cgrp = &root->cgrp;
1666 struct cftype *base_files;
1667 struct css_set *cset;
1670 lockdep_assert_held(&cgroup_mutex);
1672 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1675 root_cgrp->id = ret;
1677 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1683 * We're accessing css_set_count without locking css_set_rwsem here,
1684 * but that's OK - it can only be increased by someone holding
1685 * cgroup_lock, and that's us. The worst that can happen is that we
1686 * have some link structures left over
1688 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1692 ret = cgroup_init_root_id(root);
1696 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1697 KERNFS_ROOT_CREATE_DEACTIVATED,
1699 if (IS_ERR(root->kf_root)) {
1700 ret = PTR_ERR(root->kf_root);
1703 root_cgrp->kn = root->kf_root->kn;
1705 if (root == &cgrp_dfl_root)
1706 base_files = cgroup_dfl_base_files;
1708 base_files = cgroup_legacy_base_files;
1710 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1714 ret = rebind_subsystems(root, ss_mask);
1719 * There must be no failure case after here, since rebinding takes
1720 * care of subsystems' refcounts, which are explicitly dropped in
1721 * the failure exit path.
1723 list_add(&root->root_list, &cgroup_roots);
1724 cgroup_root_count++;
1727 * Link the root cgroup in this hierarchy into all the css_set
1730 down_write(&css_set_rwsem);
1731 hash_for_each(css_set_table, i, cset, hlist)
1732 link_css_set(&tmp_links, cset, root_cgrp);
1733 up_write(&css_set_rwsem);
1735 BUG_ON(!list_empty(&root_cgrp->self.children));
1736 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1738 kernfs_activate(root_cgrp->kn);
1743 kernfs_destroy_root(root->kf_root);
1744 root->kf_root = NULL;
1746 cgroup_exit_root_id(root);
1748 percpu_ref_exit(&root_cgrp->self.refcnt);
1750 free_cgrp_cset_links(&tmp_links);
1754 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1755 int flags, const char *unused_dev_name,
1758 struct super_block *pinned_sb = NULL;
1759 struct cgroup_subsys *ss;
1760 struct cgroup_root *root;
1761 struct cgroup_sb_opts opts;
1762 struct dentry *dentry;
1768 * The first time anyone tries to mount a cgroup, enable the list
1769 * linking each css_set to its tasks and fix up all existing tasks.
1771 if (!use_task_css_set_links)
1772 cgroup_enable_task_cg_lists();
1774 mutex_lock(&cgroup_mutex);
1776 /* First find the desired set of subsystems */
1777 ret = parse_cgroupfs_options(data, &opts);
1781 /* look for a matching existing root */
1782 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1783 cgrp_dfl_root_visible = true;
1784 root = &cgrp_dfl_root;
1785 cgroup_get(&root->cgrp);
1791 * Destruction of cgroup root is asynchronous, so subsystems may
1792 * still be dying after the previous unmount. Let's drain the
1793 * dying subsystems. We just need to ensure that the ones
1794 * unmounted previously finish dying and don't care about new ones
1795 * starting. Testing ref liveliness is good enough.
1797 for_each_subsys(ss, i) {
1798 if (!(opts.subsys_mask & (1 << i)) ||
1799 ss->root == &cgrp_dfl_root)
1802 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1803 mutex_unlock(&cgroup_mutex);
1805 ret = restart_syscall();
1808 cgroup_put(&ss->root->cgrp);
1811 for_each_root(root) {
1812 bool name_match = false;
1814 if (root == &cgrp_dfl_root)
1818 * If we asked for a name then it must match. Also, if
1819 * name matches but sybsys_mask doesn't, we should fail.
1820 * Remember whether name matched.
1823 if (strcmp(opts.name, root->name))
1829 * If we asked for subsystems (or explicitly for no
1830 * subsystems) then they must match.
1832 if ((opts.subsys_mask || opts.none) &&
1833 (opts.subsys_mask != root->subsys_mask)) {
1840 if (root->flags ^ opts.flags)
1841 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1844 * We want to reuse @root whose lifetime is governed by its
1845 * ->cgrp. Let's check whether @root is alive and keep it
1846 * that way. As cgroup_kill_sb() can happen anytime, we
1847 * want to block it by pinning the sb so that @root doesn't
1848 * get killed before mount is complete.
1850 * With the sb pinned, tryget_live can reliably indicate
1851 * whether @root can be reused. If it's being killed,
1852 * drain it. We can use wait_queue for the wait but this
1853 * path is super cold. Let's just sleep a bit and retry.
1855 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1856 if (IS_ERR(pinned_sb) ||
1857 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1858 mutex_unlock(&cgroup_mutex);
1859 if (!IS_ERR_OR_NULL(pinned_sb))
1860 deactivate_super(pinned_sb);
1862 ret = restart_syscall();
1871 * No such thing, create a new one. name= matching without subsys
1872 * specification is allowed for already existing hierarchies but we
1873 * can't create new one without subsys specification.
1875 if (!opts.subsys_mask && !opts.none) {
1880 root = kzalloc(sizeof(*root), GFP_KERNEL);
1886 init_cgroup_root(root, &opts);
1888 ret = cgroup_setup_root(root, opts.subsys_mask);
1890 cgroup_free_root(root);
1893 mutex_unlock(&cgroup_mutex);
1895 kfree(opts.release_agent);
1899 return ERR_PTR(ret);
1901 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1902 CGROUP_SUPER_MAGIC, &new_sb);
1903 if (IS_ERR(dentry) || !new_sb)
1904 cgroup_put(&root->cgrp);
1907 * If @pinned_sb, we're reusing an existing root and holding an
1908 * extra ref on its sb. Mount is complete. Put the extra ref.
1912 deactivate_super(pinned_sb);
1918 static void cgroup_kill_sb(struct super_block *sb)
1920 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1921 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1924 * If @root doesn't have any mounts or children, start killing it.
1925 * This prevents new mounts by disabling percpu_ref_tryget_live().
1926 * cgroup_mount() may wait for @root's release.
1928 * And don't kill the default root.
1930 if (!list_empty(&root->cgrp.self.children) ||
1931 root == &cgrp_dfl_root)
1932 cgroup_put(&root->cgrp);
1934 percpu_ref_kill(&root->cgrp.self.refcnt);
1939 static struct file_system_type cgroup_fs_type = {
1941 .mount = cgroup_mount,
1942 .kill_sb = cgroup_kill_sb,
1946 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1947 * @task: target task
1948 * @buf: the buffer to write the path into
1949 * @buflen: the length of the buffer
1951 * Determine @task's cgroup on the first (the one with the lowest non-zero
1952 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1953 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1954 * cgroup controller callbacks.
1956 * Return value is the same as kernfs_path().
1958 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1960 struct cgroup_root *root;
1961 struct cgroup *cgrp;
1962 int hierarchy_id = 1;
1965 mutex_lock(&cgroup_mutex);
1966 down_read(&css_set_rwsem);
1968 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1971 cgrp = task_cgroup_from_root(task, root);
1972 path = cgroup_path(cgrp, buf, buflen);
1974 /* if no hierarchy exists, everyone is in "/" */
1975 if (strlcpy(buf, "/", buflen) < buflen)
1979 up_read(&css_set_rwsem);
1980 mutex_unlock(&cgroup_mutex);
1983 EXPORT_SYMBOL_GPL(task_cgroup_path);
1985 /* used to track tasks and other necessary states during migration */
1986 struct cgroup_taskset {
1987 /* the src and dst cset list running through cset->mg_node */
1988 struct list_head src_csets;
1989 struct list_head dst_csets;
1992 * Fields for cgroup_taskset_*() iteration.
1994 * Before migration is committed, the target migration tasks are on
1995 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1996 * the csets on ->dst_csets. ->csets point to either ->src_csets
1997 * or ->dst_csets depending on whether migration is committed.
1999 * ->cur_csets and ->cur_task point to the current task position
2002 struct list_head *csets;
2003 struct css_set *cur_cset;
2004 struct task_struct *cur_task;
2008 * cgroup_taskset_first - reset taskset and return the first task
2009 * @tset: taskset of interest
2011 * @tset iteration is initialized and the first task is returned.
2013 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2015 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2016 tset->cur_task = NULL;
2018 return cgroup_taskset_next(tset);
2022 * cgroup_taskset_next - iterate to the next task in taskset
2023 * @tset: taskset of interest
2025 * Return the next task in @tset. Iteration must have been initialized
2026 * with cgroup_taskset_first().
2028 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2030 struct css_set *cset = tset->cur_cset;
2031 struct task_struct *task = tset->cur_task;
2033 while (&cset->mg_node != tset->csets) {
2035 task = list_first_entry(&cset->mg_tasks,
2036 struct task_struct, cg_list);
2038 task = list_next_entry(task, cg_list);
2040 if (&task->cg_list != &cset->mg_tasks) {
2041 tset->cur_cset = cset;
2042 tset->cur_task = task;
2046 cset = list_next_entry(cset, mg_node);
2054 * cgroup_task_migrate - move a task from one cgroup to another.
2055 * @old_cgrp: the cgroup @tsk is being migrated from
2056 * @tsk: the task being migrated
2057 * @new_cset: the new css_set @tsk is being attached to
2059 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2061 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2062 struct task_struct *tsk,
2063 struct css_set *new_cset)
2065 struct css_set *old_cset;
2067 lockdep_assert_held(&cgroup_mutex);
2068 lockdep_assert_held(&css_set_rwsem);
2071 * We are synchronized through cgroup_threadgroup_rwsem against
2072 * PF_EXITING setting such that we can't race against cgroup_exit()
2073 * changing the css_set to init_css_set and dropping the old one.
2075 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2076 old_cset = task_css_set(tsk);
2078 get_css_set(new_cset);
2079 rcu_assign_pointer(tsk->cgroups, new_cset);
2082 * Use move_tail so that cgroup_taskset_first() still returns the
2083 * leader after migration. This works because cgroup_migrate()
2084 * ensures that the dst_cset of the leader is the first on the
2085 * tset's dst_csets list.
2087 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2090 * We just gained a reference on old_cset by taking it from the
2091 * task. As trading it for new_cset is protected by cgroup_mutex,
2092 * we're safe to drop it here; it will be freed under RCU.
2094 put_css_set_locked(old_cset);
2098 * cgroup_migrate_finish - cleanup after attach
2099 * @preloaded_csets: list of preloaded css_sets
2101 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2102 * those functions for details.
2104 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2106 struct css_set *cset, *tmp_cset;
2108 lockdep_assert_held(&cgroup_mutex);
2110 down_write(&css_set_rwsem);
2111 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2112 cset->mg_src_cgrp = NULL;
2113 cset->mg_dst_cset = NULL;
2114 list_del_init(&cset->mg_preload_node);
2115 put_css_set_locked(cset);
2117 up_write(&css_set_rwsem);
2121 * cgroup_migrate_add_src - add a migration source css_set
2122 * @src_cset: the source css_set to add
2123 * @dst_cgrp: the destination cgroup
2124 * @preloaded_csets: list of preloaded css_sets
2126 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2127 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2128 * up by cgroup_migrate_finish().
2130 * This function may be called without holding cgroup_threadgroup_rwsem
2131 * even if the target is a process. Threads may be created and destroyed
2132 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2133 * into play and the preloaded css_sets are guaranteed to cover all
2136 static void cgroup_migrate_add_src(struct css_set *src_cset,
2137 struct cgroup *dst_cgrp,
2138 struct list_head *preloaded_csets)
2140 struct cgroup *src_cgrp;
2142 lockdep_assert_held(&cgroup_mutex);
2143 lockdep_assert_held(&css_set_rwsem);
2145 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2147 if (!list_empty(&src_cset->mg_preload_node))
2150 WARN_ON(src_cset->mg_src_cgrp);
2151 WARN_ON(!list_empty(&src_cset->mg_tasks));
2152 WARN_ON(!list_empty(&src_cset->mg_node));
2154 src_cset->mg_src_cgrp = src_cgrp;
2155 get_css_set(src_cset);
2156 list_add(&src_cset->mg_preload_node, preloaded_csets);
2160 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2161 * @dst_cgrp: the destination cgroup (may be %NULL)
2162 * @preloaded_csets: list of preloaded source css_sets
2164 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2165 * have been preloaded to @preloaded_csets. This function looks up and
2166 * pins all destination css_sets, links each to its source, and append them
2167 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2168 * source css_set is assumed to be its cgroup on the default hierarchy.
2170 * This function must be called after cgroup_migrate_add_src() has been
2171 * called on each migration source css_set. After migration is performed
2172 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2175 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2176 struct list_head *preloaded_csets)
2179 struct css_set *src_cset, *tmp_cset;
2181 lockdep_assert_held(&cgroup_mutex);
2184 * Except for the root, child_subsys_mask must be zero for a cgroup
2185 * with tasks so that child cgroups don't compete against tasks.
2187 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2188 dst_cgrp->child_subsys_mask)
2191 /* look up the dst cset for each src cset and link it to src */
2192 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2193 struct css_set *dst_cset;
2195 dst_cset = find_css_set(src_cset,
2196 dst_cgrp ?: src_cset->dfl_cgrp);
2200 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2203 * If src cset equals dst, it's noop. Drop the src.
2204 * cgroup_migrate() will skip the cset too. Note that we
2205 * can't handle src == dst as some nodes are used by both.
2207 if (src_cset == dst_cset) {
2208 src_cset->mg_src_cgrp = NULL;
2209 list_del_init(&src_cset->mg_preload_node);
2210 put_css_set(src_cset);
2211 put_css_set(dst_cset);
2215 src_cset->mg_dst_cset = dst_cset;
2217 if (list_empty(&dst_cset->mg_preload_node))
2218 list_add(&dst_cset->mg_preload_node, &csets);
2220 put_css_set(dst_cset);
2223 list_splice_tail(&csets, preloaded_csets);
2226 cgroup_migrate_finish(&csets);
2231 * cgroup_migrate - migrate a process or task to a cgroup
2232 * @cgrp: the destination cgroup
2233 * @leader: the leader of the process or the task to migrate
2234 * @threadgroup: whether @leader points to the whole process or a single task
2236 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2237 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2238 * caller is also responsible for invoking cgroup_migrate_add_src() and
2239 * cgroup_migrate_prepare_dst() on the targets before invoking this
2240 * function and following up with cgroup_migrate_finish().
2242 * As long as a controller's ->can_attach() doesn't fail, this function is
2243 * guaranteed to succeed. This means that, excluding ->can_attach()
2244 * failure, when migrating multiple targets, the success or failure can be
2245 * decided for all targets by invoking group_migrate_prepare_dst() before
2246 * actually starting migrating.
2248 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2251 struct cgroup_taskset tset = {
2252 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2253 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2254 .csets = &tset.src_csets,
2256 struct cgroup_subsys_state *css, *failed_css = NULL;
2257 struct css_set *cset, *tmp_cset;
2258 struct task_struct *task, *tmp_task;
2262 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2263 * already PF_EXITING could be freed from underneath us unless we
2264 * take an rcu_read_lock.
2266 down_write(&css_set_rwsem);
2270 /* @task either already exited or can't exit until the end */
2271 if (task->flags & PF_EXITING)
2274 /* leave @task alone if post_fork() hasn't linked it yet */
2275 if (list_empty(&task->cg_list))
2278 cset = task_css_set(task);
2279 if (!cset->mg_src_cgrp)
2283 * cgroup_taskset_first() must always return the leader.
2284 * Take care to avoid disturbing the ordering.
2286 list_move_tail(&task->cg_list, &cset->mg_tasks);
2287 if (list_empty(&cset->mg_node))
2288 list_add_tail(&cset->mg_node, &tset.src_csets);
2289 if (list_empty(&cset->mg_dst_cset->mg_node))
2290 list_move_tail(&cset->mg_dst_cset->mg_node,
2295 } while_each_thread(leader, task);
2297 up_write(&css_set_rwsem);
2299 /* methods shouldn't be called if no task is actually migrating */
2300 if (list_empty(&tset.src_csets))
2303 /* check that we can legitimately attach to the cgroup */
2304 for_each_e_css(css, i, cgrp) {
2305 if (css->ss->can_attach) {
2306 ret = css->ss->can_attach(css, &tset);
2309 goto out_cancel_attach;
2315 * Now that we're guaranteed success, proceed to move all tasks to
2316 * the new cgroup. There are no failure cases after here, so this
2317 * is the commit point.
2319 down_write(&css_set_rwsem);
2320 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2321 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2322 cgroup_task_migrate(cset->mg_src_cgrp, task,
2325 up_write(&css_set_rwsem);
2328 * Migration is committed, all target tasks are now on dst_csets.
2329 * Nothing is sensitive to fork() after this point. Notify
2330 * controllers that migration is complete.
2332 tset.csets = &tset.dst_csets;
2334 for_each_e_css(css, i, cgrp)
2335 if (css->ss->attach)
2336 css->ss->attach(css, &tset);
2339 goto out_release_tset;
2342 for_each_e_css(css, i, cgrp) {
2343 if (css == failed_css)
2345 if (css->ss->cancel_attach)
2346 css->ss->cancel_attach(css, &tset);
2349 down_write(&css_set_rwsem);
2350 list_splice_init(&tset.dst_csets, &tset.src_csets);
2351 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2352 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2353 list_del_init(&cset->mg_node);
2355 up_write(&css_set_rwsem);
2360 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2361 * @dst_cgrp: the cgroup to attach to
2362 * @leader: the task or the leader of the threadgroup to be attached
2363 * @threadgroup: attach the whole threadgroup?
2365 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2367 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2368 struct task_struct *leader, bool threadgroup)
2370 LIST_HEAD(preloaded_csets);
2371 struct task_struct *task;
2374 /* look up all src csets */
2375 down_read(&css_set_rwsem);
2379 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2383 } while_each_thread(leader, task);
2385 up_read(&css_set_rwsem);
2387 /* prepare dst csets and commit */
2388 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2390 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2392 cgroup_migrate_finish(&preloaded_csets);
2396 static int cgroup_procs_write_permission(struct task_struct *task,
2397 struct cgroup *dst_cgrp,
2398 struct kernfs_open_file *of)
2400 const struct cred *cred = current_cred();
2401 const struct cred *tcred = get_task_cred(task);
2405 * even if we're attaching all tasks in the thread group, we only
2406 * need to check permissions on one of them.
2408 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2409 !uid_eq(cred->euid, tcred->uid) &&
2410 !uid_eq(cred->euid, tcred->suid))
2413 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2414 struct super_block *sb = of->file->f_path.dentry->d_sb;
2415 struct cgroup *cgrp;
2416 struct inode *inode;
2418 down_read(&css_set_rwsem);
2419 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2420 up_read(&css_set_rwsem);
2422 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2423 cgrp = cgroup_parent(cgrp);
2426 inode = kernfs_get_inode(sb, cgrp->procs_kn);
2428 ret = inode_permission(inode, MAY_WRITE);
2438 * Find the task_struct of the task to attach by vpid and pass it along to the
2439 * function to attach either it or all tasks in its threadgroup. Will lock
2440 * cgroup_mutex and threadgroup.
2442 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2443 size_t nbytes, loff_t off, bool threadgroup)
2445 struct task_struct *tsk;
2446 struct cgroup *cgrp;
2450 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2453 cgrp = cgroup_kn_lock_live(of->kn);
2457 percpu_down_write(&cgroup_threadgroup_rwsem);
2460 tsk = find_task_by_vpid(pid);
2463 goto out_unlock_rcu;
2470 tsk = tsk->group_leader;
2473 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2474 * trapped in a cpuset, or RT worker may be born in a cgroup
2475 * with no rt_runtime allocated. Just say no.
2477 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2479 goto out_unlock_rcu;
2482 get_task_struct(tsk);
2485 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2487 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2489 put_task_struct(tsk);
2490 goto out_unlock_threadgroup;
2494 out_unlock_threadgroup:
2495 percpu_up_write(&cgroup_threadgroup_rwsem);
2496 cgroup_kn_unlock(of->kn);
2497 return ret ?: nbytes;
2501 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2502 * @from: attach to all cgroups of a given task
2503 * @tsk: the task to be attached
2505 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2507 struct cgroup_root *root;
2510 mutex_lock(&cgroup_mutex);
2511 for_each_root(root) {
2512 struct cgroup *from_cgrp;
2514 if (root == &cgrp_dfl_root)
2517 down_read(&css_set_rwsem);
2518 from_cgrp = task_cgroup_from_root(from, root);
2519 up_read(&css_set_rwsem);
2521 retval = cgroup_attach_task(from_cgrp, tsk, false);
2525 mutex_unlock(&cgroup_mutex);
2529 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2531 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2532 char *buf, size_t nbytes, loff_t off)
2534 return __cgroup_procs_write(of, buf, nbytes, off, false);
2537 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2538 char *buf, size_t nbytes, loff_t off)
2540 return __cgroup_procs_write(of, buf, nbytes, off, true);
2543 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2544 char *buf, size_t nbytes, loff_t off)
2546 struct cgroup *cgrp;
2548 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2550 cgrp = cgroup_kn_lock_live(of->kn);
2553 spin_lock(&release_agent_path_lock);
2554 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2555 sizeof(cgrp->root->release_agent_path));
2556 spin_unlock(&release_agent_path_lock);
2557 cgroup_kn_unlock(of->kn);
2561 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2563 struct cgroup *cgrp = seq_css(seq)->cgroup;
2565 spin_lock(&release_agent_path_lock);
2566 seq_puts(seq, cgrp->root->release_agent_path);
2567 spin_unlock(&release_agent_path_lock);
2568 seq_putc(seq, '\n');
2572 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2574 seq_puts(seq, "0\n");
2578 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2580 struct cgroup_subsys *ss;
2581 bool printed = false;
2584 for_each_subsys_which(ss, ssid, &ss_mask) {
2587 seq_printf(seq, "%s", ss->name);
2591 seq_putc(seq, '\n');
2594 /* show controllers which are currently attached to the default hierarchy */
2595 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2597 struct cgroup *cgrp = seq_css(seq)->cgroup;
2599 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2600 ~cgrp_dfl_root_inhibit_ss_mask);
2604 /* show controllers which are enabled from the parent */
2605 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2607 struct cgroup *cgrp = seq_css(seq)->cgroup;
2609 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2613 /* show controllers which are enabled for a given cgroup's children */
2614 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2616 struct cgroup *cgrp = seq_css(seq)->cgroup;
2618 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2623 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2624 * @cgrp: root of the subtree to update csses for
2626 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2627 * css associations need to be updated accordingly. This function looks up
2628 * all css_sets which are attached to the subtree, creates the matching
2629 * updated css_sets and migrates the tasks to the new ones.
2631 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2633 LIST_HEAD(preloaded_csets);
2634 struct cgroup_subsys_state *css;
2635 struct css_set *src_cset;
2638 lockdep_assert_held(&cgroup_mutex);
2640 percpu_down_write(&cgroup_threadgroup_rwsem);
2642 /* look up all csses currently attached to @cgrp's subtree */
2643 down_read(&css_set_rwsem);
2644 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2645 struct cgrp_cset_link *link;
2647 /* self is not affected by child_subsys_mask change */
2648 if (css->cgroup == cgrp)
2651 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2652 cgroup_migrate_add_src(link->cset, cgrp,
2655 up_read(&css_set_rwsem);
2657 /* NULL dst indicates self on default hierarchy */
2658 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2662 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2663 struct task_struct *last_task = NULL, *task;
2665 /* src_csets precede dst_csets, break on the first dst_cset */
2666 if (!src_cset->mg_src_cgrp)
2670 * All tasks in src_cset need to be migrated to the
2671 * matching dst_cset. Empty it process by process. We
2672 * walk tasks but migrate processes. The leader might even
2673 * belong to a different cset but such src_cset would also
2674 * be among the target src_csets because the default
2675 * hierarchy enforces per-process membership.
2678 down_read(&css_set_rwsem);
2679 task = list_first_entry_or_null(&src_cset->tasks,
2680 struct task_struct, cg_list);
2682 task = task->group_leader;
2683 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2684 get_task_struct(task);
2686 up_read(&css_set_rwsem);
2691 /* guard against possible infinite loop */
2692 if (WARN(last_task == task,
2693 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2697 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2699 put_task_struct(task);
2701 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2707 cgroup_migrate_finish(&preloaded_csets);
2708 percpu_up_write(&cgroup_threadgroup_rwsem);
2712 /* change the enabled child controllers for a cgroup in the default hierarchy */
2713 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2714 char *buf, size_t nbytes,
2717 unsigned long enable = 0, disable = 0;
2718 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2719 struct cgroup *cgrp, *child;
2720 struct cgroup_subsys *ss;
2725 * Parse input - space separated list of subsystem names prefixed
2726 * with either + or -.
2728 buf = strstrip(buf);
2729 while ((tok = strsep(&buf, " "))) {
2730 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2734 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2735 if (ss->disabled || strcmp(tok + 1, ss->name))
2739 enable |= 1 << ssid;
2740 disable &= ~(1 << ssid);
2741 } else if (*tok == '-') {
2742 disable |= 1 << ssid;
2743 enable &= ~(1 << ssid);
2749 if (ssid == CGROUP_SUBSYS_COUNT)
2753 cgrp = cgroup_kn_lock_live(of->kn);
2757 for_each_subsys(ss, ssid) {
2758 if (enable & (1 << ssid)) {
2759 if (cgrp->subtree_control & (1 << ssid)) {
2760 enable &= ~(1 << ssid);
2764 /* unavailable or not enabled on the parent? */
2765 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2766 (cgroup_parent(cgrp) &&
2767 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2771 } else if (disable & (1 << ssid)) {
2772 if (!(cgrp->subtree_control & (1 << ssid))) {
2773 disable &= ~(1 << ssid);
2777 /* a child has it enabled? */
2778 cgroup_for_each_live_child(child, cgrp) {
2779 if (child->subtree_control & (1 << ssid)) {
2787 if (!enable && !disable) {
2793 * Except for the root, subtree_control must be zero for a cgroup
2794 * with tasks so that child cgroups don't compete against tasks.
2796 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2802 * Update subsys masks and calculate what needs to be done. More
2803 * subsystems than specified may need to be enabled or disabled
2804 * depending on subsystem dependencies.
2806 old_sc = cgrp->subtree_control;
2807 old_ss = cgrp->child_subsys_mask;
2808 new_sc = (old_sc | enable) & ~disable;
2809 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2811 css_enable = ~old_ss & new_ss;
2812 css_disable = old_ss & ~new_ss;
2813 enable |= css_enable;
2814 disable |= css_disable;
2817 * Because css offlining is asynchronous, userland might try to
2818 * re-enable the same controller while the previous instance is
2819 * still around. In such cases, wait till it's gone using
2822 for_each_subsys_which(ss, ssid, &css_enable) {
2823 cgroup_for_each_live_child(child, cgrp) {
2826 if (!cgroup_css(child, ss))
2830 prepare_to_wait(&child->offline_waitq, &wait,
2831 TASK_UNINTERRUPTIBLE);
2832 cgroup_kn_unlock(of->kn);
2834 finish_wait(&child->offline_waitq, &wait);
2837 return restart_syscall();
2841 cgrp->subtree_control = new_sc;
2842 cgrp->child_subsys_mask = new_ss;
2845 * Create new csses or make the existing ones visible. A css is
2846 * created invisible if it's being implicitly enabled through
2847 * dependency. An invisible css is made visible when the userland
2848 * explicitly enables it.
2850 for_each_subsys(ss, ssid) {
2851 if (!(enable & (1 << ssid)))
2854 cgroup_for_each_live_child(child, cgrp) {
2855 if (css_enable & (1 << ssid))
2856 ret = create_css(child, ss,
2857 cgrp->subtree_control & (1 << ssid));
2859 ret = cgroup_populate_dir(child, 1 << ssid);
2866 * At this point, cgroup_e_css() results reflect the new csses
2867 * making the following cgroup_update_dfl_csses() properly update
2868 * css associations of all tasks in the subtree.
2870 ret = cgroup_update_dfl_csses(cgrp);
2875 * All tasks are migrated out of disabled csses. Kill or hide
2876 * them. A css is hidden when the userland requests it to be
2877 * disabled while other subsystems are still depending on it. The
2878 * css must not actively control resources and be in the vanilla
2879 * state if it's made visible again later. Controllers which may
2880 * be depended upon should provide ->css_reset() for this purpose.
2882 for_each_subsys(ss, ssid) {
2883 if (!(disable & (1 << ssid)))
2886 cgroup_for_each_live_child(child, cgrp) {
2887 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2889 if (css_disable & (1 << ssid)) {
2892 cgroup_clear_dir(child, 1 << ssid);
2900 * The effective csses of all the descendants (excluding @cgrp) may
2901 * have changed. Subsystems can optionally subscribe to this event
2902 * by implementing ->css_e_css_changed() which is invoked if any of
2903 * the effective csses seen from the css's cgroup may have changed.
2905 for_each_subsys(ss, ssid) {
2906 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2907 struct cgroup_subsys_state *css;
2909 if (!ss->css_e_css_changed || !this_css)
2912 css_for_each_descendant_pre(css, this_css)
2913 if (css != this_css)
2914 ss->css_e_css_changed(css);
2917 kernfs_activate(cgrp->kn);
2920 cgroup_kn_unlock(of->kn);
2921 return ret ?: nbytes;
2924 cgrp->subtree_control = old_sc;
2925 cgrp->child_subsys_mask = old_ss;
2927 for_each_subsys(ss, ssid) {
2928 if (!(enable & (1 << ssid)))
2931 cgroup_for_each_live_child(child, cgrp) {
2932 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2937 if (css_enable & (1 << ssid))
2940 cgroup_clear_dir(child, 1 << ssid);
2946 static int cgroup_populated_show(struct seq_file *seq, void *v)
2948 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2952 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2953 size_t nbytes, loff_t off)
2955 struct cgroup *cgrp = of->kn->parent->priv;
2956 struct cftype *cft = of->kn->priv;
2957 struct cgroup_subsys_state *css;
2961 return cft->write(of, buf, nbytes, off);
2964 * kernfs guarantees that a file isn't deleted with operations in
2965 * flight, which means that the matching css is and stays alive and
2966 * doesn't need to be pinned. The RCU locking is not necessary
2967 * either. It's just for the convenience of using cgroup_css().
2970 css = cgroup_css(cgrp, cft->ss);
2973 if (cft->write_u64) {
2974 unsigned long long v;
2975 ret = kstrtoull(buf, 0, &v);
2977 ret = cft->write_u64(css, cft, v);
2978 } else if (cft->write_s64) {
2980 ret = kstrtoll(buf, 0, &v);
2982 ret = cft->write_s64(css, cft, v);
2987 return ret ?: nbytes;
2990 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2992 return seq_cft(seq)->seq_start(seq, ppos);
2995 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2997 return seq_cft(seq)->seq_next(seq, v, ppos);
3000 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3002 seq_cft(seq)->seq_stop(seq, v);
3005 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3007 struct cftype *cft = seq_cft(m);
3008 struct cgroup_subsys_state *css = seq_css(m);
3011 return cft->seq_show(m, arg);
3014 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3015 else if (cft->read_s64)
3016 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3022 static struct kernfs_ops cgroup_kf_single_ops = {
3023 .atomic_write_len = PAGE_SIZE,
3024 .write = cgroup_file_write,
3025 .seq_show = cgroup_seqfile_show,
3028 static struct kernfs_ops cgroup_kf_ops = {
3029 .atomic_write_len = PAGE_SIZE,
3030 .write = cgroup_file_write,
3031 .seq_start = cgroup_seqfile_start,
3032 .seq_next = cgroup_seqfile_next,
3033 .seq_stop = cgroup_seqfile_stop,
3034 .seq_show = cgroup_seqfile_show,
3038 * cgroup_rename - Only allow simple rename of directories in place.
3040 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3041 const char *new_name_str)
3043 struct cgroup *cgrp = kn->priv;
3046 if (kernfs_type(kn) != KERNFS_DIR)
3048 if (kn->parent != new_parent)
3052 * This isn't a proper migration and its usefulness is very
3053 * limited. Disallow on the default hierarchy.
3055 if (cgroup_on_dfl(cgrp))
3059 * We're gonna grab cgroup_mutex which nests outside kernfs
3060 * active_ref. kernfs_rename() doesn't require active_ref
3061 * protection. Break them before grabbing cgroup_mutex.
3063 kernfs_break_active_protection(new_parent);
3064 kernfs_break_active_protection(kn);
3066 mutex_lock(&cgroup_mutex);
3068 ret = kernfs_rename(kn, new_parent, new_name_str);
3070 mutex_unlock(&cgroup_mutex);
3072 kernfs_unbreak_active_protection(kn);
3073 kernfs_unbreak_active_protection(new_parent);
3077 /* set uid and gid of cgroup dirs and files to that of the creator */
3078 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3080 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3081 .ia_uid = current_fsuid(),
3082 .ia_gid = current_fsgid(), };
3084 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3085 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3088 return kernfs_setattr(kn, &iattr);
3091 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3093 char name[CGROUP_FILE_NAME_MAX];
3094 struct kernfs_node *kn;
3095 struct lock_class_key *key = NULL;
3098 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3099 key = &cft->lockdep_key;
3101 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3102 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3107 ret = cgroup_kn_set_ugid(kn);
3113 if (cft->write == cgroup_procs_write)
3114 cgrp->procs_kn = kn;
3115 else if (cft->seq_show == cgroup_populated_show)
3116 cgrp->populated_kn = kn;
3121 * cgroup_addrm_files - add or remove files to a cgroup directory
3122 * @cgrp: the target cgroup
3123 * @cfts: array of cftypes to be added
3124 * @is_add: whether to add or remove
3126 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3127 * For removals, this function never fails. If addition fails, this
3128 * function doesn't remove files already added. The caller is responsible
3131 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3137 lockdep_assert_held(&cgroup_mutex);
3139 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3140 /* does cft->flags tell us to skip this file on @cgrp? */
3141 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3143 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3145 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3147 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3151 ret = cgroup_add_file(cgrp, cft);
3153 pr_warn("%s: failed to add %s, err=%d\n",
3154 __func__, cft->name, ret);
3158 cgroup_rm_file(cgrp, cft);
3164 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3167 struct cgroup_subsys *ss = cfts[0].ss;
3168 struct cgroup *root = &ss->root->cgrp;
3169 struct cgroup_subsys_state *css;
3172 lockdep_assert_held(&cgroup_mutex);
3174 /* add/rm files for all cgroups created before */
3175 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3176 struct cgroup *cgrp = css->cgroup;
3178 if (cgroup_is_dead(cgrp))
3181 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3187 kernfs_activate(root->kn);
3191 static void cgroup_exit_cftypes(struct cftype *cfts)
3195 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3196 /* free copy for custom atomic_write_len, see init_cftypes() */
3197 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3202 /* revert flags set by cgroup core while adding @cfts */
3203 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3207 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3211 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3212 struct kernfs_ops *kf_ops;
3214 WARN_ON(cft->ss || cft->kf_ops);
3217 kf_ops = &cgroup_kf_ops;
3219 kf_ops = &cgroup_kf_single_ops;
3222 * Ugh... if @cft wants a custom max_write_len, we need to
3223 * make a copy of kf_ops to set its atomic_write_len.
3225 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3226 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3228 cgroup_exit_cftypes(cfts);
3231 kf_ops->atomic_write_len = cft->max_write_len;
3234 cft->kf_ops = kf_ops;
3241 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3243 lockdep_assert_held(&cgroup_mutex);
3245 if (!cfts || !cfts[0].ss)
3248 list_del(&cfts->node);
3249 cgroup_apply_cftypes(cfts, false);
3250 cgroup_exit_cftypes(cfts);
3255 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3256 * @cfts: zero-length name terminated array of cftypes
3258 * Unregister @cfts. Files described by @cfts are removed from all
3259 * existing cgroups and all future cgroups won't have them either. This
3260 * function can be called anytime whether @cfts' subsys is attached or not.
3262 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3265 int cgroup_rm_cftypes(struct cftype *cfts)
3269 mutex_lock(&cgroup_mutex);
3270 ret = cgroup_rm_cftypes_locked(cfts);
3271 mutex_unlock(&cgroup_mutex);
3276 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3277 * @ss: target cgroup subsystem
3278 * @cfts: zero-length name terminated array of cftypes
3280 * Register @cfts to @ss. Files described by @cfts are created for all
3281 * existing cgroups to which @ss is attached and all future cgroups will
3282 * have them too. This function can be called anytime whether @ss is
3285 * Returns 0 on successful registration, -errno on failure. Note that this
3286 * function currently returns 0 as long as @cfts registration is successful
3287 * even if some file creation attempts on existing cgroups fail.
3289 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3296 if (!cfts || cfts[0].name[0] == '\0')
3299 ret = cgroup_init_cftypes(ss, cfts);
3303 mutex_lock(&cgroup_mutex);
3305 list_add_tail(&cfts->node, &ss->cfts);
3306 ret = cgroup_apply_cftypes(cfts, true);
3308 cgroup_rm_cftypes_locked(cfts);
3310 mutex_unlock(&cgroup_mutex);
3315 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3316 * @ss: target cgroup subsystem
3317 * @cfts: zero-length name terminated array of cftypes
3319 * Similar to cgroup_add_cftypes() but the added files are only used for
3320 * the default hierarchy.
3322 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3326 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3327 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3328 return cgroup_add_cftypes(ss, cfts);
3332 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3333 * @ss: target cgroup subsystem
3334 * @cfts: zero-length name terminated array of cftypes
3336 * Similar to cgroup_add_cftypes() but the added files are only used for
3337 * the legacy hierarchies.
3339 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3344 * If legacy_flies_on_dfl, we want to show the legacy files on the
3345 * dfl hierarchy but iff the target subsystem hasn't been updated
3346 * for the dfl hierarchy yet.
3348 if (!cgroup_legacy_files_on_dfl ||
3349 ss->dfl_cftypes != ss->legacy_cftypes) {
3350 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3351 cft->flags |= __CFTYPE_NOT_ON_DFL;
3354 return cgroup_add_cftypes(ss, cfts);
3358 * cgroup_task_count - count the number of tasks in a cgroup.
3359 * @cgrp: the cgroup in question
3361 * Return the number of tasks in the cgroup.
3363 static int cgroup_task_count(const struct cgroup *cgrp)
3366 struct cgrp_cset_link *link;
3368 down_read(&css_set_rwsem);
3369 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3370 count += atomic_read(&link->cset->refcount);
3371 up_read(&css_set_rwsem);
3376 * css_next_child - find the next child of a given css
3377 * @pos: the current position (%NULL to initiate traversal)
3378 * @parent: css whose children to walk
3380 * This function returns the next child of @parent and should be called
3381 * under either cgroup_mutex or RCU read lock. The only requirement is
3382 * that @parent and @pos are accessible. The next sibling is guaranteed to
3383 * be returned regardless of their states.
3385 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3386 * css which finished ->css_online() is guaranteed to be visible in the
3387 * future iterations and will stay visible until the last reference is put.
3388 * A css which hasn't finished ->css_online() or already finished
3389 * ->css_offline() may show up during traversal. It's each subsystem's
3390 * responsibility to synchronize against on/offlining.
3392 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3393 struct cgroup_subsys_state *parent)
3395 struct cgroup_subsys_state *next;
3397 cgroup_assert_mutex_or_rcu_locked();
3400 * @pos could already have been unlinked from the sibling list.
3401 * Once a cgroup is removed, its ->sibling.next is no longer
3402 * updated when its next sibling changes. CSS_RELEASED is set when
3403 * @pos is taken off list, at which time its next pointer is valid,
3404 * and, as releases are serialized, the one pointed to by the next
3405 * pointer is guaranteed to not have started release yet. This
3406 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3407 * critical section, the one pointed to by its next pointer is
3408 * guaranteed to not have finished its RCU grace period even if we
3409 * have dropped rcu_read_lock() inbetween iterations.
3411 * If @pos has CSS_RELEASED set, its next pointer can't be
3412 * dereferenced; however, as each css is given a monotonically
3413 * increasing unique serial number and always appended to the
3414 * sibling list, the next one can be found by walking the parent's
3415 * children until the first css with higher serial number than
3416 * @pos's. While this path can be slower, it happens iff iteration
3417 * races against release and the race window is very small.
3420 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3421 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3422 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3424 list_for_each_entry_rcu(next, &parent->children, sibling)
3425 if (next->serial_nr > pos->serial_nr)
3430 * @next, if not pointing to the head, can be dereferenced and is
3433 if (&next->sibling != &parent->children)
3439 * css_next_descendant_pre - find the next descendant for pre-order walk
3440 * @pos: the current position (%NULL to initiate traversal)
3441 * @root: css whose descendants to walk
3443 * To be used by css_for_each_descendant_pre(). Find the next descendant
3444 * to visit for pre-order traversal of @root's descendants. @root is
3445 * included in the iteration and the first node to be visited.
3447 * While this function requires cgroup_mutex or RCU read locking, it
3448 * doesn't require the whole traversal to be contained in a single critical
3449 * section. This function will return the correct next descendant as long
3450 * as both @pos and @root are accessible and @pos is a descendant of @root.
3452 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3453 * css which finished ->css_online() is guaranteed to be visible in the
3454 * future iterations and will stay visible until the last reference is put.
3455 * A css which hasn't finished ->css_online() or already finished
3456 * ->css_offline() may show up during traversal. It's each subsystem's
3457 * responsibility to synchronize against on/offlining.
3459 struct cgroup_subsys_state *
3460 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3461 struct cgroup_subsys_state *root)
3463 struct cgroup_subsys_state *next;
3465 cgroup_assert_mutex_or_rcu_locked();
3467 /* if first iteration, visit @root */
3471 /* visit the first child if exists */
3472 next = css_next_child(NULL, pos);
3476 /* no child, visit my or the closest ancestor's next sibling */
3477 while (pos != root) {
3478 next = css_next_child(pos, pos->parent);
3488 * css_rightmost_descendant - return the rightmost descendant of a css
3489 * @pos: css of interest
3491 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3492 * is returned. This can be used during pre-order traversal to skip
3495 * While this function requires cgroup_mutex or RCU read locking, it
3496 * doesn't require the whole traversal to be contained in a single critical
3497 * section. This function will return the correct rightmost descendant as
3498 * long as @pos is accessible.
3500 struct cgroup_subsys_state *
3501 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3503 struct cgroup_subsys_state *last, *tmp;
3505 cgroup_assert_mutex_or_rcu_locked();
3509 /* ->prev isn't RCU safe, walk ->next till the end */
3511 css_for_each_child(tmp, last)
3518 static struct cgroup_subsys_state *
3519 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3521 struct cgroup_subsys_state *last;
3525 pos = css_next_child(NULL, pos);
3532 * css_next_descendant_post - find the next descendant for post-order walk
3533 * @pos: the current position (%NULL to initiate traversal)
3534 * @root: css whose descendants to walk
3536 * To be used by css_for_each_descendant_post(). Find the next descendant
3537 * to visit for post-order traversal of @root's descendants. @root is
3538 * included in the iteration and the last node to be visited.
3540 * While this function requires cgroup_mutex or RCU read locking, it
3541 * doesn't require the whole traversal to be contained in a single critical
3542 * section. This function will return the correct next descendant as long
3543 * as both @pos and @cgroup are accessible and @pos is a descendant of
3546 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3547 * css which finished ->css_online() is guaranteed to be visible in the
3548 * future iterations and will stay visible until the last reference is put.
3549 * A css which hasn't finished ->css_online() or already finished
3550 * ->css_offline() may show up during traversal. It's each subsystem's
3551 * responsibility to synchronize against on/offlining.
3553 struct cgroup_subsys_state *
3554 css_next_descendant_post(struct cgroup_subsys_state *pos,
3555 struct cgroup_subsys_state *root)
3557 struct cgroup_subsys_state *next;
3559 cgroup_assert_mutex_or_rcu_locked();
3561 /* if first iteration, visit leftmost descendant which may be @root */
3563 return css_leftmost_descendant(root);
3565 /* if we visited @root, we're done */
3569 /* if there's an unvisited sibling, visit its leftmost descendant */
3570 next = css_next_child(pos, pos->parent);
3572 return css_leftmost_descendant(next);
3574 /* no sibling left, visit parent */
3579 * css_has_online_children - does a css have online children
3580 * @css: the target css
3582 * Returns %true if @css has any online children; otherwise, %false. This
3583 * function can be called from any context but the caller is responsible
3584 * for synchronizing against on/offlining as necessary.
3586 bool css_has_online_children(struct cgroup_subsys_state *css)
3588 struct cgroup_subsys_state *child;
3592 css_for_each_child(child, css) {
3593 if (child->flags & CSS_ONLINE) {
3603 * css_advance_task_iter - advance a task itererator to the next css_set
3604 * @it: the iterator to advance
3606 * Advance @it to the next css_set to walk.
3608 static void css_advance_task_iter(struct css_task_iter *it)
3610 struct list_head *l = it->cset_pos;
3611 struct cgrp_cset_link *link;
3612 struct css_set *cset;
3614 /* Advance to the next non-empty css_set */
3617 if (l == it->cset_head) {
3618 it->cset_pos = NULL;
3623 cset = container_of(l, struct css_set,
3624 e_cset_node[it->ss->id]);
3626 link = list_entry(l, struct cgrp_cset_link, cset_link);
3629 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3633 if (!list_empty(&cset->tasks))
3634 it->task_pos = cset->tasks.next;
3636 it->task_pos = cset->mg_tasks.next;
3638 it->tasks_head = &cset->tasks;
3639 it->mg_tasks_head = &cset->mg_tasks;
3643 * css_task_iter_start - initiate task iteration
3644 * @css: the css to walk tasks of
3645 * @it: the task iterator to use
3647 * Initiate iteration through the tasks of @css. The caller can call
3648 * css_task_iter_next() to walk through the tasks until the function
3649 * returns NULL. On completion of iteration, css_task_iter_end() must be
3652 * Note that this function acquires a lock which is released when the
3653 * iteration finishes. The caller can't sleep while iteration is in
3656 void css_task_iter_start(struct cgroup_subsys_state *css,
3657 struct css_task_iter *it)
3658 __acquires(css_set_rwsem)
3660 /* no one should try to iterate before mounting cgroups */
3661 WARN_ON_ONCE(!use_task_css_set_links);
3663 down_read(&css_set_rwsem);
3668 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3670 it->cset_pos = &css->cgroup->cset_links;
3672 it->cset_head = it->cset_pos;
3674 css_advance_task_iter(it);
3678 * css_task_iter_next - return the next task for the iterator
3679 * @it: the task iterator being iterated
3681 * The "next" function for task iteration. @it should have been
3682 * initialized via css_task_iter_start(). Returns NULL when the iteration
3685 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3687 struct task_struct *res;
3688 struct list_head *l = it->task_pos;
3690 /* If the iterator cg is NULL, we have no tasks */
3693 res = list_entry(l, struct task_struct, cg_list);
3696 * Advance iterator to find next entry. cset->tasks is consumed
3697 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3702 if (l == it->tasks_head)
3703 l = it->mg_tasks_head->next;
3705 if (l == it->mg_tasks_head)
3706 css_advance_task_iter(it);
3714 * css_task_iter_end - finish task iteration
3715 * @it: the task iterator to finish
3717 * Finish task iteration started by css_task_iter_start().
3719 void css_task_iter_end(struct css_task_iter *it)
3720 __releases(css_set_rwsem)
3722 up_read(&css_set_rwsem);
3726 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3727 * @to: cgroup to which the tasks will be moved
3728 * @from: cgroup in which the tasks currently reside
3730 * Locking rules between cgroup_post_fork() and the migration path
3731 * guarantee that, if a task is forking while being migrated, the new child
3732 * is guaranteed to be either visible in the source cgroup after the
3733 * parent's migration is complete or put into the target cgroup. No task
3734 * can slip out of migration through forking.
3736 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3738 LIST_HEAD(preloaded_csets);
3739 struct cgrp_cset_link *link;
3740 struct css_task_iter it;
3741 struct task_struct *task;
3744 mutex_lock(&cgroup_mutex);
3746 /* all tasks in @from are being moved, all csets are source */
3747 down_read(&css_set_rwsem);
3748 list_for_each_entry(link, &from->cset_links, cset_link)
3749 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3750 up_read(&css_set_rwsem);
3752 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3757 * Migrate tasks one-by-one until @form is empty. This fails iff
3758 * ->can_attach() fails.
3761 css_task_iter_start(&from->self, &it);
3762 task = css_task_iter_next(&it);
3764 get_task_struct(task);
3765 css_task_iter_end(&it);
3768 ret = cgroup_migrate(to, task, false);
3769 put_task_struct(task);
3771 } while (task && !ret);
3773 cgroup_migrate_finish(&preloaded_csets);
3774 mutex_unlock(&cgroup_mutex);
3779 * Stuff for reading the 'tasks'/'procs' files.
3781 * Reading this file can return large amounts of data if a cgroup has
3782 * *lots* of attached tasks. So it may need several calls to read(),
3783 * but we cannot guarantee that the information we produce is correct
3784 * unless we produce it entirely atomically.
3788 /* which pidlist file are we talking about? */
3789 enum cgroup_filetype {
3795 * A pidlist is a list of pids that virtually represents the contents of one
3796 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3797 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3800 struct cgroup_pidlist {
3802 * used to find which pidlist is wanted. doesn't change as long as
3803 * this particular list stays in the list.
3805 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3808 /* how many elements the above list has */
3810 /* each of these stored in a list by its cgroup */
3811 struct list_head links;
3812 /* pointer to the cgroup we belong to, for list removal purposes */
3813 struct cgroup *owner;
3814 /* for delayed destruction */
3815 struct delayed_work destroy_dwork;
3819 * The following two functions "fix" the issue where there are more pids
3820 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3821 * TODO: replace with a kernel-wide solution to this problem
3823 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3824 static void *pidlist_allocate(int count)
3826 if (PIDLIST_TOO_LARGE(count))
3827 return vmalloc(count * sizeof(pid_t));
3829 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3832 static void pidlist_free(void *p)
3838 * Used to destroy all pidlists lingering waiting for destroy timer. None
3839 * should be left afterwards.
3841 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3843 struct cgroup_pidlist *l, *tmp_l;
3845 mutex_lock(&cgrp->pidlist_mutex);
3846 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3847 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3848 mutex_unlock(&cgrp->pidlist_mutex);
3850 flush_workqueue(cgroup_pidlist_destroy_wq);
3851 BUG_ON(!list_empty(&cgrp->pidlists));
3854 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3856 struct delayed_work *dwork = to_delayed_work(work);
3857 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3859 struct cgroup_pidlist *tofree = NULL;
3861 mutex_lock(&l->owner->pidlist_mutex);
3864 * Destroy iff we didn't get queued again. The state won't change
3865 * as destroy_dwork can only be queued while locked.
3867 if (!delayed_work_pending(dwork)) {
3868 list_del(&l->links);
3869 pidlist_free(l->list);
3870 put_pid_ns(l->key.ns);
3874 mutex_unlock(&l->owner->pidlist_mutex);
3879 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3880 * Returns the number of unique elements.
3882 static int pidlist_uniq(pid_t *list, int length)
3887 * we presume the 0th element is unique, so i starts at 1. trivial
3888 * edge cases first; no work needs to be done for either
3890 if (length == 0 || length == 1)
3892 /* src and dest walk down the list; dest counts unique elements */
3893 for (src = 1; src < length; src++) {
3894 /* find next unique element */
3895 while (list[src] == list[src-1]) {
3900 /* dest always points to where the next unique element goes */
3901 list[dest] = list[src];
3909 * The two pid files - task and cgroup.procs - guaranteed that the result
3910 * is sorted, which forced this whole pidlist fiasco. As pid order is
3911 * different per namespace, each namespace needs differently sorted list,
3912 * making it impossible to use, for example, single rbtree of member tasks
3913 * sorted by task pointer. As pidlists can be fairly large, allocating one
3914 * per open file is dangerous, so cgroup had to implement shared pool of
3915 * pidlists keyed by cgroup and namespace.
3917 * All this extra complexity was caused by the original implementation
3918 * committing to an entirely unnecessary property. In the long term, we
3919 * want to do away with it. Explicitly scramble sort order if on the
3920 * default hierarchy so that no such expectation exists in the new
3923 * Scrambling is done by swapping every two consecutive bits, which is
3924 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3926 static pid_t pid_fry(pid_t pid)
3928 unsigned a = pid & 0x55555555;
3929 unsigned b = pid & 0xAAAAAAAA;
3931 return (a << 1) | (b >> 1);
3934 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3936 if (cgroup_on_dfl(cgrp))
3937 return pid_fry(pid);
3942 static int cmppid(const void *a, const void *b)
3944 return *(pid_t *)a - *(pid_t *)b;
3947 static int fried_cmppid(const void *a, const void *b)
3949 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3952 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3953 enum cgroup_filetype type)
3955 struct cgroup_pidlist *l;
3956 /* don't need task_nsproxy() if we're looking at ourself */
3957 struct pid_namespace *ns = task_active_pid_ns(current);
3959 lockdep_assert_held(&cgrp->pidlist_mutex);
3961 list_for_each_entry(l, &cgrp->pidlists, links)
3962 if (l->key.type == type && l->key.ns == ns)
3968 * find the appropriate pidlist for our purpose (given procs vs tasks)
3969 * returns with the lock on that pidlist already held, and takes care
3970 * of the use count, or returns NULL with no locks held if we're out of
3973 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3974 enum cgroup_filetype type)
3976 struct cgroup_pidlist *l;
3978 lockdep_assert_held(&cgrp->pidlist_mutex);
3980 l = cgroup_pidlist_find(cgrp, type);
3984 /* entry not found; create a new one */
3985 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3989 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3991 /* don't need task_nsproxy() if we're looking at ourself */
3992 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3994 list_add(&l->links, &cgrp->pidlists);
3999 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4001 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4002 struct cgroup_pidlist **lp)
4006 int pid, n = 0; /* used for populating the array */
4007 struct css_task_iter it;
4008 struct task_struct *tsk;
4009 struct cgroup_pidlist *l;
4011 lockdep_assert_held(&cgrp->pidlist_mutex);
4014 * If cgroup gets more users after we read count, we won't have
4015 * enough space - tough. This race is indistinguishable to the
4016 * caller from the case that the additional cgroup users didn't
4017 * show up until sometime later on.
4019 length = cgroup_task_count(cgrp);
4020 array = pidlist_allocate(length);
4023 /* now, populate the array */
4024 css_task_iter_start(&cgrp->self, &it);
4025 while ((tsk = css_task_iter_next(&it))) {
4026 if (unlikely(n == length))
4028 /* get tgid or pid for procs or tasks file respectively */
4029 if (type == CGROUP_FILE_PROCS)
4030 pid = task_tgid_vnr(tsk);
4032 pid = task_pid_vnr(tsk);
4033 if (pid > 0) /* make sure to only use valid results */
4036 css_task_iter_end(&it);
4038 /* now sort & (if procs) strip out duplicates */
4039 if (cgroup_on_dfl(cgrp))
4040 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4042 sort(array, length, sizeof(pid_t), cmppid, NULL);
4043 if (type == CGROUP_FILE_PROCS)
4044 length = pidlist_uniq(array, length);
4046 l = cgroup_pidlist_find_create(cgrp, type);
4048 pidlist_free(array);
4052 /* store array, freeing old if necessary */
4053 pidlist_free(l->list);
4061 * cgroupstats_build - build and fill cgroupstats
4062 * @stats: cgroupstats to fill information into
4063 * @dentry: A dentry entry belonging to the cgroup for which stats have
4066 * Build and fill cgroupstats so that taskstats can export it to user
4069 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4071 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4072 struct cgroup *cgrp;
4073 struct css_task_iter it;
4074 struct task_struct *tsk;
4076 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4077 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4078 kernfs_type(kn) != KERNFS_DIR)
4081 mutex_lock(&cgroup_mutex);
4084 * We aren't being called from kernfs and there's no guarantee on
4085 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4086 * @kn->priv is RCU safe. Let's do the RCU dancing.
4089 cgrp = rcu_dereference(kn->priv);
4090 if (!cgrp || cgroup_is_dead(cgrp)) {
4092 mutex_unlock(&cgroup_mutex);
4097 css_task_iter_start(&cgrp->self, &it);
4098 while ((tsk = css_task_iter_next(&it))) {
4099 switch (tsk->state) {
4101 stats->nr_running++;
4103 case TASK_INTERRUPTIBLE:
4104 stats->nr_sleeping++;
4106 case TASK_UNINTERRUPTIBLE:
4107 stats->nr_uninterruptible++;
4110 stats->nr_stopped++;
4113 if (delayacct_is_task_waiting_on_io(tsk))
4114 stats->nr_io_wait++;
4118 css_task_iter_end(&it);
4120 mutex_unlock(&cgroup_mutex);
4126 * seq_file methods for the tasks/procs files. The seq_file position is the
4127 * next pid to display; the seq_file iterator is a pointer to the pid
4128 * in the cgroup->l->list array.
4131 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4134 * Initially we receive a position value that corresponds to
4135 * one more than the last pid shown (or 0 on the first call or
4136 * after a seek to the start). Use a binary-search to find the
4137 * next pid to display, if any
4139 struct kernfs_open_file *of = s->private;
4140 struct cgroup *cgrp = seq_css(s)->cgroup;
4141 struct cgroup_pidlist *l;
4142 enum cgroup_filetype type = seq_cft(s)->private;
4143 int index = 0, pid = *pos;
4146 mutex_lock(&cgrp->pidlist_mutex);
4149 * !NULL @of->priv indicates that this isn't the first start()
4150 * after open. If the matching pidlist is around, we can use that.
4151 * Look for it. Note that @of->priv can't be used directly. It
4152 * could already have been destroyed.
4155 of->priv = cgroup_pidlist_find(cgrp, type);
4158 * Either this is the first start() after open or the matching
4159 * pidlist has been destroyed inbetween. Create a new one.
4162 ret = pidlist_array_load(cgrp, type,
4163 (struct cgroup_pidlist **)&of->priv);
4165 return ERR_PTR(ret);
4170 int end = l->length;
4172 while (index < end) {
4173 int mid = (index + end) / 2;
4174 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4177 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4183 /* If we're off the end of the array, we're done */
4184 if (index >= l->length)
4186 /* Update the abstract position to be the actual pid that we found */
4187 iter = l->list + index;
4188 *pos = cgroup_pid_fry(cgrp, *iter);
4192 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4194 struct kernfs_open_file *of = s->private;
4195 struct cgroup_pidlist *l = of->priv;
4198 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4199 CGROUP_PIDLIST_DESTROY_DELAY);
4200 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4203 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4205 struct kernfs_open_file *of = s->private;
4206 struct cgroup_pidlist *l = of->priv;
4208 pid_t *end = l->list + l->length;
4210 * Advance to the next pid in the array. If this goes off the
4217 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4222 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4224 seq_printf(s, "%d\n", *(int *)v);
4229 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4232 return notify_on_release(css->cgroup);
4235 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4236 struct cftype *cft, u64 val)
4239 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4241 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4245 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4248 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4251 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4252 struct cftype *cft, u64 val)
4255 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4257 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4261 /* cgroup core interface files for the default hierarchy */
4262 static struct cftype cgroup_dfl_base_files[] = {
4264 .name = "cgroup.procs",
4265 .seq_start = cgroup_pidlist_start,
4266 .seq_next = cgroup_pidlist_next,
4267 .seq_stop = cgroup_pidlist_stop,
4268 .seq_show = cgroup_pidlist_show,
4269 .private = CGROUP_FILE_PROCS,
4270 .write = cgroup_procs_write,
4271 .mode = S_IRUGO | S_IWUSR,
4274 .name = "cgroup.controllers",
4275 .flags = CFTYPE_ONLY_ON_ROOT,
4276 .seq_show = cgroup_root_controllers_show,
4279 .name = "cgroup.controllers",
4280 .flags = CFTYPE_NOT_ON_ROOT,
4281 .seq_show = cgroup_controllers_show,
4284 .name = "cgroup.subtree_control",
4285 .seq_show = cgroup_subtree_control_show,
4286 .write = cgroup_subtree_control_write,
4289 .name = "cgroup.populated",
4290 .flags = CFTYPE_NOT_ON_ROOT,
4291 .seq_show = cgroup_populated_show,
4296 /* cgroup core interface files for the legacy hierarchies */
4297 static struct cftype cgroup_legacy_base_files[] = {
4299 .name = "cgroup.procs",
4300 .seq_start = cgroup_pidlist_start,
4301 .seq_next = cgroup_pidlist_next,
4302 .seq_stop = cgroup_pidlist_stop,
4303 .seq_show = cgroup_pidlist_show,
4304 .private = CGROUP_FILE_PROCS,
4305 .write = cgroup_procs_write,
4306 .mode = S_IRUGO | S_IWUSR,
4309 .name = "cgroup.clone_children",
4310 .read_u64 = cgroup_clone_children_read,
4311 .write_u64 = cgroup_clone_children_write,
4314 .name = "cgroup.sane_behavior",
4315 .flags = CFTYPE_ONLY_ON_ROOT,
4316 .seq_show = cgroup_sane_behavior_show,
4320 .seq_start = cgroup_pidlist_start,
4321 .seq_next = cgroup_pidlist_next,
4322 .seq_stop = cgroup_pidlist_stop,
4323 .seq_show = cgroup_pidlist_show,
4324 .private = CGROUP_FILE_TASKS,
4325 .write = cgroup_tasks_write,
4326 .mode = S_IRUGO | S_IWUSR,
4329 .name = "notify_on_release",
4330 .read_u64 = cgroup_read_notify_on_release,
4331 .write_u64 = cgroup_write_notify_on_release,
4334 .name = "release_agent",
4335 .flags = CFTYPE_ONLY_ON_ROOT,
4336 .seq_show = cgroup_release_agent_show,
4337 .write = cgroup_release_agent_write,
4338 .max_write_len = PATH_MAX - 1,
4344 * cgroup_populate_dir - create subsys files in a cgroup directory
4345 * @cgrp: target cgroup
4346 * @subsys_mask: mask of the subsystem ids whose files should be added
4348 * On failure, no file is added.
4350 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4352 struct cgroup_subsys *ss;
4355 /* process cftsets of each subsystem */
4356 for_each_subsys(ss, i) {
4357 struct cftype *cfts;
4359 if (!(subsys_mask & (1 << i)))
4362 list_for_each_entry(cfts, &ss->cfts, node) {
4363 ret = cgroup_addrm_files(cgrp, cfts, true);
4370 cgroup_clear_dir(cgrp, subsys_mask);
4375 * css destruction is four-stage process.
4377 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4378 * Implemented in kill_css().
4380 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4381 * and thus css_tryget_online() is guaranteed to fail, the css can be
4382 * offlined by invoking offline_css(). After offlining, the base ref is
4383 * put. Implemented in css_killed_work_fn().
4385 * 3. When the percpu_ref reaches zero, the only possible remaining
4386 * accessors are inside RCU read sections. css_release() schedules the
4389 * 4. After the grace period, the css can be freed. Implemented in
4390 * css_free_work_fn().
4392 * It is actually hairier because both step 2 and 4 require process context
4393 * and thus involve punting to css->destroy_work adding two additional
4394 * steps to the already complex sequence.
4396 static void css_free_work_fn(struct work_struct *work)
4398 struct cgroup_subsys_state *css =
4399 container_of(work, struct cgroup_subsys_state, destroy_work);
4400 struct cgroup_subsys *ss = css->ss;
4401 struct cgroup *cgrp = css->cgroup;
4403 percpu_ref_exit(&css->refcnt);
4410 css_put(css->parent);
4413 cgroup_idr_remove(&ss->css_idr, id);
4416 /* cgroup free path */
4417 atomic_dec(&cgrp->root->nr_cgrps);
4418 cgroup_pidlist_destroy_all(cgrp);
4419 cancel_work_sync(&cgrp->release_agent_work);
4421 if (cgroup_parent(cgrp)) {
4423 * We get a ref to the parent, and put the ref when
4424 * this cgroup is being freed, so it's guaranteed
4425 * that the parent won't be destroyed before its
4428 cgroup_put(cgroup_parent(cgrp));
4429 kernfs_put(cgrp->kn);
4433 * This is root cgroup's refcnt reaching zero,
4434 * which indicates that the root should be
4437 cgroup_destroy_root(cgrp->root);
4442 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4444 struct cgroup_subsys_state *css =
4445 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4447 INIT_WORK(&css->destroy_work, css_free_work_fn);
4448 queue_work(cgroup_destroy_wq, &css->destroy_work);
4451 static void css_release_work_fn(struct work_struct *work)
4453 struct cgroup_subsys_state *css =
4454 container_of(work, struct cgroup_subsys_state, destroy_work);
4455 struct cgroup_subsys *ss = css->ss;
4456 struct cgroup *cgrp = css->cgroup;
4458 mutex_lock(&cgroup_mutex);
4460 css->flags |= CSS_RELEASED;
4461 list_del_rcu(&css->sibling);
4464 /* css release path */
4465 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4466 if (ss->css_released)
4467 ss->css_released(css);
4469 /* cgroup release path */
4470 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4474 * There are two control paths which try to determine
4475 * cgroup from dentry without going through kernfs -
4476 * cgroupstats_build() and css_tryget_online_from_dir().
4477 * Those are supported by RCU protecting clearing of
4478 * cgrp->kn->priv backpointer.
4480 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4483 mutex_unlock(&cgroup_mutex);
4485 call_rcu(&css->rcu_head, css_free_rcu_fn);
4488 static void css_release(struct percpu_ref *ref)
4490 struct cgroup_subsys_state *css =
4491 container_of(ref, struct cgroup_subsys_state, refcnt);
4493 INIT_WORK(&css->destroy_work, css_release_work_fn);
4494 queue_work(cgroup_destroy_wq, &css->destroy_work);
4497 static void init_and_link_css(struct cgroup_subsys_state *css,
4498 struct cgroup_subsys *ss, struct cgroup *cgrp)
4500 lockdep_assert_held(&cgroup_mutex);
4504 memset(css, 0, sizeof(*css));
4507 INIT_LIST_HEAD(&css->sibling);
4508 INIT_LIST_HEAD(&css->children);
4509 css->serial_nr = css_serial_nr_next++;
4511 if (cgroup_parent(cgrp)) {
4512 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4513 css_get(css->parent);
4516 BUG_ON(cgroup_css(cgrp, ss));
4519 /* invoke ->css_online() on a new CSS and mark it online if successful */
4520 static int online_css(struct cgroup_subsys_state *css)
4522 struct cgroup_subsys *ss = css->ss;
4525 lockdep_assert_held(&cgroup_mutex);
4528 ret = ss->css_online(css);
4530 css->flags |= CSS_ONLINE;
4531 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4536 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4537 static void offline_css(struct cgroup_subsys_state *css)
4539 struct cgroup_subsys *ss = css->ss;
4541 lockdep_assert_held(&cgroup_mutex);
4543 if (!(css->flags & CSS_ONLINE))
4546 if (ss->css_offline)
4547 ss->css_offline(css);
4549 css->flags &= ~CSS_ONLINE;
4550 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4552 wake_up_all(&css->cgroup->offline_waitq);
4556 * create_css - create a cgroup_subsys_state
4557 * @cgrp: the cgroup new css will be associated with
4558 * @ss: the subsys of new css
4559 * @visible: whether to create control knobs for the new css or not
4561 * Create a new css associated with @cgrp - @ss pair. On success, the new
4562 * css is online and installed in @cgrp with all interface files created if
4563 * @visible. Returns 0 on success, -errno on failure.
4565 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4568 struct cgroup *parent = cgroup_parent(cgrp);
4569 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4570 struct cgroup_subsys_state *css;
4573 lockdep_assert_held(&cgroup_mutex);
4575 css = ss->css_alloc(parent_css);
4577 return PTR_ERR(css);
4579 init_and_link_css(css, ss, cgrp);
4581 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4585 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4587 goto err_free_percpu_ref;
4591 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4596 /* @css is ready to be brought online now, make it visible */
4597 list_add_tail_rcu(&css->sibling, &parent_css->children);
4598 cgroup_idr_replace(&ss->css_idr, css, css->id);
4600 err = online_css(css);
4604 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4605 cgroup_parent(parent)) {
4606 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4607 current->comm, current->pid, ss->name);
4608 if (!strcmp(ss->name, "memory"))
4609 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4610 ss->warned_broken_hierarchy = true;
4616 list_del_rcu(&css->sibling);
4617 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4619 cgroup_idr_remove(&ss->css_idr, css->id);
4620 err_free_percpu_ref:
4621 percpu_ref_exit(&css->refcnt);
4623 call_rcu(&css->rcu_head, css_free_rcu_fn);
4627 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4630 struct cgroup *parent, *cgrp;
4631 struct cgroup_root *root;
4632 struct cgroup_subsys *ss;
4633 struct kernfs_node *kn;
4634 struct cftype *base_files;
4637 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4639 if (strchr(name, '\n'))
4642 parent = cgroup_kn_lock_live(parent_kn);
4645 root = parent->root;
4647 /* allocate the cgroup and its ID, 0 is reserved for the root */
4648 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4654 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4659 * Temporarily set the pointer to NULL, so idr_find() won't return
4660 * a half-baked cgroup.
4662 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4665 goto out_cancel_ref;
4668 init_cgroup_housekeeping(cgrp);
4670 cgrp->self.parent = &parent->self;
4673 if (notify_on_release(parent))
4674 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4676 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4677 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4679 /* create the directory */
4680 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4688 * This extra ref will be put in cgroup_free_fn() and guarantees
4689 * that @cgrp->kn is always accessible.
4693 cgrp->self.serial_nr = css_serial_nr_next++;
4695 /* allocation complete, commit to creation */
4696 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4697 atomic_inc(&root->nr_cgrps);
4701 * @cgrp is now fully operational. If something fails after this
4702 * point, it'll be released via the normal destruction path.
4704 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4706 ret = cgroup_kn_set_ugid(kn);
4710 if (cgroup_on_dfl(cgrp))
4711 base_files = cgroup_dfl_base_files;
4713 base_files = cgroup_legacy_base_files;
4715 ret = cgroup_addrm_files(cgrp, base_files, true);
4719 /* let's create and online css's */
4720 for_each_subsys(ss, ssid) {
4721 if (parent->child_subsys_mask & (1 << ssid)) {
4722 ret = create_css(cgrp, ss,
4723 parent->subtree_control & (1 << ssid));
4730 * On the default hierarchy, a child doesn't automatically inherit
4731 * subtree_control from the parent. Each is configured manually.
4733 if (!cgroup_on_dfl(cgrp)) {
4734 cgrp->subtree_control = parent->subtree_control;
4735 cgroup_refresh_child_subsys_mask(cgrp);
4738 kernfs_activate(kn);
4744 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4746 percpu_ref_exit(&cgrp->self.refcnt);
4750 cgroup_kn_unlock(parent_kn);
4754 cgroup_destroy_locked(cgrp);
4759 * This is called when the refcnt of a css is confirmed to be killed.
4760 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4761 * initate destruction and put the css ref from kill_css().
4763 static void css_killed_work_fn(struct work_struct *work)
4765 struct cgroup_subsys_state *css =
4766 container_of(work, struct cgroup_subsys_state, destroy_work);
4768 mutex_lock(&cgroup_mutex);
4770 mutex_unlock(&cgroup_mutex);
4775 /* css kill confirmation processing requires process context, bounce */
4776 static void css_killed_ref_fn(struct percpu_ref *ref)
4778 struct cgroup_subsys_state *css =
4779 container_of(ref, struct cgroup_subsys_state, refcnt);
4781 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4782 queue_work(cgroup_destroy_wq, &css->destroy_work);
4786 * kill_css - destroy a css
4787 * @css: css to destroy
4789 * This function initiates destruction of @css by removing cgroup interface
4790 * files and putting its base reference. ->css_offline() will be invoked
4791 * asynchronously once css_tryget_online() is guaranteed to fail and when
4792 * the reference count reaches zero, @css will be released.
4794 static void kill_css(struct cgroup_subsys_state *css)
4796 lockdep_assert_held(&cgroup_mutex);
4799 * This must happen before css is disassociated with its cgroup.
4800 * See seq_css() for details.
4802 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4805 * Killing would put the base ref, but we need to keep it alive
4806 * until after ->css_offline().
4811 * cgroup core guarantees that, by the time ->css_offline() is
4812 * invoked, no new css reference will be given out via
4813 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4814 * proceed to offlining css's because percpu_ref_kill() doesn't
4815 * guarantee that the ref is seen as killed on all CPUs on return.
4817 * Use percpu_ref_kill_and_confirm() to get notifications as each
4818 * css is confirmed to be seen as killed on all CPUs.
4820 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4824 * cgroup_destroy_locked - the first stage of cgroup destruction
4825 * @cgrp: cgroup to be destroyed
4827 * css's make use of percpu refcnts whose killing latency shouldn't be
4828 * exposed to userland and are RCU protected. Also, cgroup core needs to
4829 * guarantee that css_tryget_online() won't succeed by the time
4830 * ->css_offline() is invoked. To satisfy all the requirements,
4831 * destruction is implemented in the following two steps.
4833 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4834 * userland visible parts and start killing the percpu refcnts of
4835 * css's. Set up so that the next stage will be kicked off once all
4836 * the percpu refcnts are confirmed to be killed.
4838 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4839 * rest of destruction. Once all cgroup references are gone, the
4840 * cgroup is RCU-freed.
4842 * This function implements s1. After this step, @cgrp is gone as far as
4843 * the userland is concerned and a new cgroup with the same name may be
4844 * created. As cgroup doesn't care about the names internally, this
4845 * doesn't cause any problem.
4847 static int cgroup_destroy_locked(struct cgroup *cgrp)
4848 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4850 struct cgroup_subsys_state *css;
4854 lockdep_assert_held(&cgroup_mutex);
4857 * css_set_rwsem synchronizes access to ->cset_links and prevents
4858 * @cgrp from being removed while put_css_set() is in progress.
4860 down_read(&css_set_rwsem);
4861 empty = list_empty(&cgrp->cset_links);
4862 up_read(&css_set_rwsem);
4867 * Make sure there's no live children. We can't test emptiness of
4868 * ->self.children as dead children linger on it while being
4869 * drained; otherwise, "rmdir parent/child parent" may fail.
4871 if (css_has_online_children(&cgrp->self))
4875 * Mark @cgrp dead. This prevents further task migration and child
4876 * creation by disabling cgroup_lock_live_group().
4878 cgrp->self.flags &= ~CSS_ONLINE;
4880 /* initiate massacre of all css's */
4881 for_each_css(css, ssid, cgrp)
4885 * Remove @cgrp directory along with the base files. @cgrp has an
4886 * extra ref on its kn.
4888 kernfs_remove(cgrp->kn);
4890 check_for_release(cgroup_parent(cgrp));
4892 /* put the base reference */
4893 percpu_ref_kill(&cgrp->self.refcnt);
4898 static int cgroup_rmdir(struct kernfs_node *kn)
4900 struct cgroup *cgrp;
4903 cgrp = cgroup_kn_lock_live(kn);
4907 ret = cgroup_destroy_locked(cgrp);
4909 cgroup_kn_unlock(kn);
4913 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4914 .remount_fs = cgroup_remount,
4915 .show_options = cgroup_show_options,
4916 .mkdir = cgroup_mkdir,
4917 .rmdir = cgroup_rmdir,
4918 .rename = cgroup_rename,
4921 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4923 struct cgroup_subsys_state *css;
4925 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4927 mutex_lock(&cgroup_mutex);
4929 idr_init(&ss->css_idr);
4930 INIT_LIST_HEAD(&ss->cfts);
4932 /* Create the root cgroup state for this subsystem */
4933 ss->root = &cgrp_dfl_root;
4934 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4935 /* We don't handle early failures gracefully */
4936 BUG_ON(IS_ERR(css));
4937 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4940 * Root csses are never destroyed and we can't initialize
4941 * percpu_ref during early init. Disable refcnting.
4943 css->flags |= CSS_NO_REF;
4946 /* allocation can't be done safely during early init */
4949 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4950 BUG_ON(css->id < 0);
4953 /* Update the init_css_set to contain a subsys
4954 * pointer to this state - since the subsystem is
4955 * newly registered, all tasks and hence the
4956 * init_css_set is in the subsystem's root cgroup. */
4957 init_css_set.subsys[ss->id] = css;
4959 have_fork_callback |= (bool)ss->fork << ss->id;
4960 have_exit_callback |= (bool)ss->exit << ss->id;
4961 have_canfork_callback |= (bool)ss->can_fork << ss->id;
4963 /* At system boot, before all subsystems have been
4964 * registered, no tasks have been forked, so we don't
4965 * need to invoke fork callbacks here. */
4966 BUG_ON(!list_empty(&init_task.tasks));
4968 BUG_ON(online_css(css));
4970 mutex_unlock(&cgroup_mutex);
4974 * cgroup_init_early - cgroup initialization at system boot
4976 * Initialize cgroups at system boot, and initialize any
4977 * subsystems that request early init.
4979 int __init cgroup_init_early(void)
4981 static struct cgroup_sb_opts __initdata opts;
4982 struct cgroup_subsys *ss;
4985 init_cgroup_root(&cgrp_dfl_root, &opts);
4986 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4988 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4990 for_each_subsys(ss, i) {
4991 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4992 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4993 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4995 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4996 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4999 ss->name = cgroup_subsys_name[i];
5002 cgroup_init_subsys(ss, true);
5008 * cgroup_init - cgroup initialization
5010 * Register cgroup filesystem and /proc file, and initialize
5011 * any subsystems that didn't request early init.
5013 int __init cgroup_init(void)
5015 struct cgroup_subsys *ss;
5019 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5020 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5021 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5023 mutex_lock(&cgroup_mutex);
5025 /* Add init_css_set to the hash table */
5026 key = css_set_hash(init_css_set.subsys);
5027 hash_add(css_set_table, &init_css_set.hlist, key);
5029 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5031 mutex_unlock(&cgroup_mutex);
5033 for_each_subsys(ss, ssid) {
5034 if (ss->early_init) {
5035 struct cgroup_subsys_state *css =
5036 init_css_set.subsys[ss->id];
5038 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5040 BUG_ON(css->id < 0);
5042 cgroup_init_subsys(ss, false);
5045 list_add_tail(&init_css_set.e_cset_node[ssid],
5046 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5049 * Setting dfl_root subsys_mask needs to consider the
5050 * disabled flag and cftype registration needs kmalloc,
5051 * both of which aren't available during early_init.
5056 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5058 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5059 ss->dfl_cftypes = ss->legacy_cftypes;
5061 if (!ss->dfl_cftypes)
5062 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5064 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5065 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5067 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5068 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5072 ss->bind(init_css_set.subsys[ssid]);
5075 err = sysfs_create_mount_point(fs_kobj, "cgroup");
5079 err = register_filesystem(&cgroup_fs_type);
5081 sysfs_remove_mount_point(fs_kobj, "cgroup");
5085 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5089 static int __init cgroup_wq_init(void)
5092 * There isn't much point in executing destruction path in
5093 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5094 * Use 1 for @max_active.
5096 * We would prefer to do this in cgroup_init() above, but that
5097 * is called before init_workqueues(): so leave this until after.
5099 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5100 BUG_ON(!cgroup_destroy_wq);
5103 * Used to destroy pidlists and separate to serve as flush domain.
5104 * Cap @max_active to 1 too.
5106 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5108 BUG_ON(!cgroup_pidlist_destroy_wq);
5112 core_initcall(cgroup_wq_init);
5115 * proc_cgroup_show()
5116 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5117 * - Used for /proc/<pid>/cgroup.
5119 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5120 struct pid *pid, struct task_struct *tsk)
5124 struct cgroup_root *root;
5127 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5131 mutex_lock(&cgroup_mutex);
5132 down_read(&css_set_rwsem);
5134 for_each_root(root) {
5135 struct cgroup_subsys *ss;
5136 struct cgroup *cgrp;
5137 int ssid, count = 0;
5139 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5142 seq_printf(m, "%d:", root->hierarchy_id);
5143 for_each_subsys(ss, ssid)
5144 if (root->subsys_mask & (1 << ssid))
5145 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5146 if (strlen(root->name))
5147 seq_printf(m, "%sname=%s", count ? "," : "",
5150 cgrp = task_cgroup_from_root(tsk, root);
5151 path = cgroup_path(cgrp, buf, PATH_MAX);
5153 retval = -ENAMETOOLONG;
5162 up_read(&css_set_rwsem);
5163 mutex_unlock(&cgroup_mutex);
5169 /* Display information about each subsystem and each hierarchy */
5170 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5172 struct cgroup_subsys *ss;
5175 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5177 * ideally we don't want subsystems moving around while we do this.
5178 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5179 * subsys/hierarchy state.
5181 mutex_lock(&cgroup_mutex);
5183 for_each_subsys(ss, i)
5184 seq_printf(m, "%s\t%d\t%d\t%d\n",
5185 ss->name, ss->root->hierarchy_id,
5186 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5188 mutex_unlock(&cgroup_mutex);
5192 static int cgroupstats_open(struct inode *inode, struct file *file)
5194 return single_open(file, proc_cgroupstats_show, NULL);
5197 static const struct file_operations proc_cgroupstats_operations = {
5198 .open = cgroupstats_open,
5200 .llseek = seq_lseek,
5201 .release = single_release,
5204 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5206 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5207 return &ss_priv[i - CGROUP_CANFORK_START];
5211 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5213 void **private = subsys_canfork_priv_p(ss_priv, i);
5214 return private ? *private : NULL;
5218 * cgroup_fork - initialize cgroup related fields during copy_process()
5219 * @child: pointer to task_struct of forking parent process.
5221 * A task is associated with the init_css_set until cgroup_post_fork()
5222 * attaches it to the parent's css_set. Empty cg_list indicates that
5223 * @child isn't holding reference to its css_set.
5225 void cgroup_fork(struct task_struct *child)
5227 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5228 INIT_LIST_HEAD(&child->cg_list);
5232 * cgroup_can_fork - called on a new task before the process is exposed
5233 * @child: the task in question.
5235 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5236 * returns an error, the fork aborts with that error code. This allows for
5237 * a cgroup subsystem to conditionally allow or deny new forks.
5239 int cgroup_can_fork(struct task_struct *child,
5240 void *ss_priv[CGROUP_CANFORK_COUNT])
5242 struct cgroup_subsys *ss;
5245 for_each_subsys_which(ss, i, &have_canfork_callback) {
5246 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5254 for_each_subsys(ss, j) {
5257 if (ss->cancel_fork)
5258 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5265 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5266 * @child: the task in question
5268 * This calls the cancel_fork() callbacks if a fork failed *after*
5269 * cgroup_can_fork() succeded.
5271 void cgroup_cancel_fork(struct task_struct *child,
5272 void *ss_priv[CGROUP_CANFORK_COUNT])
5274 struct cgroup_subsys *ss;
5277 for_each_subsys(ss, i)
5278 if (ss->cancel_fork)
5279 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5283 * cgroup_post_fork - called on a new task after adding it to the task list
5284 * @child: the task in question
5286 * Adds the task to the list running through its css_set if necessary and
5287 * call the subsystem fork() callbacks. Has to be after the task is
5288 * visible on the task list in case we race with the first call to
5289 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5292 void cgroup_post_fork(struct task_struct *child,
5293 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5295 struct cgroup_subsys *ss;
5299 * This may race against cgroup_enable_task_cg_lists(). As that
5300 * function sets use_task_css_set_links before grabbing
5301 * tasklist_lock and we just went through tasklist_lock to add
5302 * @child, it's guaranteed that either we see the set
5303 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5304 * @child during its iteration.
5306 * If we won the race, @child is associated with %current's
5307 * css_set. Grabbing css_set_rwsem guarantees both that the
5308 * association is stable, and, on completion of the parent's
5309 * migration, @child is visible in the source of migration or
5310 * already in the destination cgroup. This guarantee is necessary
5311 * when implementing operations which need to migrate all tasks of
5312 * a cgroup to another.
5314 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5315 * will remain in init_css_set. This is safe because all tasks are
5316 * in the init_css_set before cg_links is enabled and there's no
5317 * operation which transfers all tasks out of init_css_set.
5319 if (use_task_css_set_links) {
5320 struct css_set *cset;
5322 down_write(&css_set_rwsem);
5323 cset = task_css_set(current);
5324 if (list_empty(&child->cg_list)) {
5325 rcu_assign_pointer(child->cgroups, cset);
5326 list_add(&child->cg_list, &cset->tasks);
5329 up_write(&css_set_rwsem);
5333 * Call ss->fork(). This must happen after @child is linked on
5334 * css_set; otherwise, @child might change state between ->fork()
5335 * and addition to css_set.
5337 for_each_subsys_which(ss, i, &have_fork_callback)
5338 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5342 * cgroup_exit - detach cgroup from exiting task
5343 * @tsk: pointer to task_struct of exiting process
5345 * Description: Detach cgroup from @tsk and release it.
5347 * Note that cgroups marked notify_on_release force every task in
5348 * them to take the global cgroup_mutex mutex when exiting.
5349 * This could impact scaling on very large systems. Be reluctant to
5350 * use notify_on_release cgroups where very high task exit scaling
5351 * is required on large systems.
5353 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5354 * call cgroup_exit() while the task is still competent to handle
5355 * notify_on_release(), then leave the task attached to the root cgroup in
5356 * each hierarchy for the remainder of its exit. No need to bother with
5357 * init_css_set refcnting. init_css_set never goes away and we can't race
5358 * with migration path - PF_EXITING is visible to migration path.
5360 void cgroup_exit(struct task_struct *tsk)
5362 struct cgroup_subsys *ss;
5363 struct css_set *cset;
5364 bool put_cset = false;
5368 * Unlink from @tsk from its css_set. As migration path can't race
5369 * with us, we can check cg_list without grabbing css_set_rwsem.
5371 if (!list_empty(&tsk->cg_list)) {
5372 down_write(&css_set_rwsem);
5373 list_del_init(&tsk->cg_list);
5374 up_write(&css_set_rwsem);
5378 /* Reassign the task to the init_css_set. */
5379 cset = task_css_set(tsk);
5380 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5382 /* see cgroup_post_fork() for details */
5383 for_each_subsys_which(ss, i, &have_exit_callback) {
5384 struct cgroup_subsys_state *old_css = cset->subsys[i];
5385 struct cgroup_subsys_state *css = task_css(tsk, i);
5387 ss->exit(css, old_css, tsk);
5394 static void check_for_release(struct cgroup *cgrp)
5396 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5397 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5398 schedule_work(&cgrp->release_agent_work);
5402 * Notify userspace when a cgroup is released, by running the
5403 * configured release agent with the name of the cgroup (path
5404 * relative to the root of cgroup file system) as the argument.
5406 * Most likely, this user command will try to rmdir this cgroup.
5408 * This races with the possibility that some other task will be
5409 * attached to this cgroup before it is removed, or that some other
5410 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5411 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5412 * unused, and this cgroup will be reprieved from its death sentence,
5413 * to continue to serve a useful existence. Next time it's released,
5414 * we will get notified again, if it still has 'notify_on_release' set.
5416 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5417 * means only wait until the task is successfully execve()'d. The
5418 * separate release agent task is forked by call_usermodehelper(),
5419 * then control in this thread returns here, without waiting for the
5420 * release agent task. We don't bother to wait because the caller of
5421 * this routine has no use for the exit status of the release agent
5422 * task, so no sense holding our caller up for that.
5424 static void cgroup_release_agent(struct work_struct *work)
5426 struct cgroup *cgrp =
5427 container_of(work, struct cgroup, release_agent_work);
5428 char *pathbuf = NULL, *agentbuf = NULL, *path;
5429 char *argv[3], *envp[3];
5431 mutex_lock(&cgroup_mutex);
5433 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5434 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5435 if (!pathbuf || !agentbuf)
5438 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5446 /* minimal command environment */
5448 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5451 mutex_unlock(&cgroup_mutex);
5452 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5455 mutex_unlock(&cgroup_mutex);
5461 static int __init cgroup_disable(char *str)
5463 struct cgroup_subsys *ss;
5467 while ((token = strsep(&str, ",")) != NULL) {
5471 for_each_subsys(ss, i) {
5472 if (!strcmp(token, ss->name)) {
5474 printk(KERN_INFO "Disabling %s control group"
5475 " subsystem\n", ss->name);
5482 __setup("cgroup_disable=", cgroup_disable);
5484 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5486 printk("cgroup: using legacy files on the default hierarchy\n");
5487 cgroup_legacy_files_on_dfl = true;
5490 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5493 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5494 * @dentry: directory dentry of interest
5495 * @ss: subsystem of interest
5497 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5498 * to get the corresponding css and return it. If such css doesn't exist
5499 * or can't be pinned, an ERR_PTR value is returned.
5501 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5502 struct cgroup_subsys *ss)
5504 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5505 struct cgroup_subsys_state *css = NULL;
5506 struct cgroup *cgrp;
5508 /* is @dentry a cgroup dir? */
5509 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5510 kernfs_type(kn) != KERNFS_DIR)
5511 return ERR_PTR(-EBADF);
5516 * This path doesn't originate from kernfs and @kn could already
5517 * have been or be removed at any point. @kn->priv is RCU
5518 * protected for this access. See css_release_work_fn() for details.
5520 cgrp = rcu_dereference(kn->priv);
5522 css = cgroup_css(cgrp, ss);
5524 if (!css || !css_tryget_online(css))
5525 css = ERR_PTR(-ENOENT);
5532 * css_from_id - lookup css by id
5533 * @id: the cgroup id
5534 * @ss: cgroup subsys to be looked into
5536 * Returns the css if there's valid one with @id, otherwise returns NULL.
5537 * Should be called under rcu_read_lock().
5539 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5541 WARN_ON_ONCE(!rcu_read_lock_held());
5542 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5545 #ifdef CONFIG_CGROUP_DEBUG
5546 static struct cgroup_subsys_state *
5547 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5549 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5552 return ERR_PTR(-ENOMEM);
5557 static void debug_css_free(struct cgroup_subsys_state *css)
5562 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5565 return cgroup_task_count(css->cgroup);
5568 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5571 return (u64)(unsigned long)current->cgroups;
5574 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5580 count = atomic_read(&task_css_set(current)->refcount);
5585 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5587 struct cgrp_cset_link *link;
5588 struct css_set *cset;
5591 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5595 down_read(&css_set_rwsem);
5597 cset = rcu_dereference(current->cgroups);
5598 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5599 struct cgroup *c = link->cgrp;
5601 cgroup_name(c, name_buf, NAME_MAX + 1);
5602 seq_printf(seq, "Root %d group %s\n",
5603 c->root->hierarchy_id, name_buf);
5606 up_read(&css_set_rwsem);
5611 #define MAX_TASKS_SHOWN_PER_CSS 25
5612 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5614 struct cgroup_subsys_state *css = seq_css(seq);
5615 struct cgrp_cset_link *link;
5617 down_read(&css_set_rwsem);
5618 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5619 struct css_set *cset = link->cset;
5620 struct task_struct *task;
5623 seq_printf(seq, "css_set %p\n", cset);
5625 list_for_each_entry(task, &cset->tasks, cg_list) {
5626 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5628 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5631 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5632 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5634 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5638 seq_puts(seq, " ...\n");
5640 up_read(&css_set_rwsem);
5644 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5646 return (!cgroup_has_tasks(css->cgroup) &&
5647 !css_has_online_children(&css->cgroup->self));
5650 static struct cftype debug_files[] = {
5652 .name = "taskcount",
5653 .read_u64 = debug_taskcount_read,
5657 .name = "current_css_set",
5658 .read_u64 = current_css_set_read,
5662 .name = "current_css_set_refcount",
5663 .read_u64 = current_css_set_refcount_read,
5667 .name = "current_css_set_cg_links",
5668 .seq_show = current_css_set_cg_links_read,
5672 .name = "cgroup_css_links",
5673 .seq_show = cgroup_css_links_read,
5677 .name = "releasable",
5678 .read_u64 = releasable_read,
5684 struct cgroup_subsys debug_cgrp_subsys = {
5685 .css_alloc = debug_css_alloc,
5686 .css_free = debug_css_free,
5687 .legacy_cftypes = debug_files,
5689 #endif /* CONFIG_CGROUP_DEBUG */