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>
39 #include <linux/mutex.h>
40 #include <linux/mount.h>
41 #include <linux/pagemap.h>
42 #include <linux/proc_fs.h>
43 #include <linux/rcupdate.h>
44 #include <linux/sched.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/rwsem.h>
48 #include <linux/string.h>
49 #include <linux/sort.h>
50 #include <linux/kmod.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hashtable.h>
54 #include <linux/pid_namespace.h>
55 #include <linux/idr.h>
56 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
57 #include <linux/kthread.h>
58 #include <linux/delay.h>
60 #include <linux/atomic.h>
63 * pidlists linger the following amount before being destroyed. The goal
64 * is avoiding frequent destruction in the middle of consecutive read calls
65 * Expiring in the middle is a performance problem not a correctness one.
66 * 1 sec should be enough.
68 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
70 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
74 * cgroup_mutex is the master lock. Any modification to cgroup or its
75 * hierarchy must be performed while holding it.
77 * css_set_rwsem protects task->cgroups pointer, the list of css_set
78 * objects, and the chain of tasks off each css_set.
80 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
81 * cgroup.h can use them for lockdep annotations.
83 #ifdef CONFIG_PROVE_RCU
84 DEFINE_MUTEX(cgroup_mutex);
85 DECLARE_RWSEM(css_set_rwsem);
86 EXPORT_SYMBOL_GPL(cgroup_mutex);
87 EXPORT_SYMBOL_GPL(css_set_rwsem);
89 static DEFINE_MUTEX(cgroup_mutex);
90 static DECLARE_RWSEM(css_set_rwsem);
94 * Protects cgroup_idr and css_idr so that IDs can be released without
95 * grabbing cgroup_mutex.
97 static DEFINE_SPINLOCK(cgroup_idr_lock);
100 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
101 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
103 static DEFINE_SPINLOCK(release_agent_path_lock);
105 #define cgroup_assert_mutex_or_rcu_locked() \
106 rcu_lockdep_assert(rcu_read_lock_held() || \
107 lockdep_is_held(&cgroup_mutex), \
108 "cgroup_mutex or RCU read lock required");
111 * cgroup destruction makes heavy use of work items and there can be a lot
112 * of concurrent destructions. Use a separate workqueue so that cgroup
113 * destruction work items don't end up filling up max_active of system_wq
114 * which may lead to deadlock.
116 static struct workqueue_struct *cgroup_destroy_wq;
119 * pidlist destructions need to be flushed on cgroup destruction. Use a
120 * separate workqueue as flush domain.
122 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
124 /* generate an array of cgroup subsystem pointers */
125 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
126 static struct cgroup_subsys *cgroup_subsys[] = {
127 #include <linux/cgroup_subsys.h>
131 /* array of cgroup subsystem names */
132 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
133 static const char *cgroup_subsys_name[] = {
134 #include <linux/cgroup_subsys.h>
139 * The default hierarchy, reserved for the subsystems that are otherwise
140 * unattached - it never has more than a single cgroup, and all tasks are
141 * part of that cgroup.
143 struct cgroup_root cgrp_dfl_root;
146 * The default hierarchy always exists but is hidden until mounted for the
147 * first time. This is for backward compatibility.
149 static bool cgrp_dfl_root_visible;
151 /* The list of hierarchy roots */
153 static LIST_HEAD(cgroup_roots);
154 static int cgroup_root_count;
156 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
157 static DEFINE_IDR(cgroup_hierarchy_idr);
160 * Assign a monotonically increasing serial number to cgroups. It
161 * guarantees cgroups with bigger numbers are newer than those with smaller
162 * numbers. Also, as cgroups are always appended to the parent's
163 * ->children list, it guarantees that sibling cgroups are always sorted in
164 * the ascending serial number order on the list. Protected by
167 static u64 cgroup_serial_nr_next = 1;
169 /* This flag indicates whether tasks in the fork and exit paths should
170 * check for fork/exit handlers to call. This avoids us having to do
171 * extra work in the fork/exit path if none of the subsystems need to
174 static int need_forkexit_callback __read_mostly;
176 static struct cftype cgroup_base_files[];
178 static void cgroup_put(struct cgroup *cgrp);
179 static int rebind_subsystems(struct cgroup_root *dst_root,
180 unsigned int ss_mask);
181 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
182 static int cgroup_destroy_locked(struct cgroup *cgrp);
183 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss);
184 static void kill_css(struct cgroup_subsys_state *css);
185 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
187 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
189 /* IDR wrappers which synchronize using cgroup_idr_lock */
190 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
195 idr_preload(gfp_mask);
196 spin_lock_bh(&cgroup_idr_lock);
197 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
198 spin_unlock_bh(&cgroup_idr_lock);
203 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
207 spin_lock_bh(&cgroup_idr_lock);
208 ret = idr_replace(idr, ptr, id);
209 spin_unlock_bh(&cgroup_idr_lock);
213 static void cgroup_idr_remove(struct idr *idr, int id)
215 spin_lock_bh(&cgroup_idr_lock);
217 spin_unlock_bh(&cgroup_idr_lock);
221 * cgroup_css - obtain a cgroup's css for the specified subsystem
222 * @cgrp: the cgroup of interest
223 * @ss: the subsystem of interest (%NULL returns the dummy_css)
225 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
226 * function must be called either under cgroup_mutex or rcu_read_lock() and
227 * the caller is responsible for pinning the returned css if it wants to
228 * keep accessing it outside the said locks. This function may return
229 * %NULL if @cgrp doesn't have @subsys_id enabled.
231 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
232 struct cgroup_subsys *ss)
235 return rcu_dereference_check(cgrp->subsys[ss->id],
236 lockdep_is_held(&cgroup_mutex));
238 return &cgrp->dummy_css;
242 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
243 * @cgrp: the cgroup of interest
244 * @ss: the subsystem of interest (%NULL returns the dummy_css)
246 * Similar to cgroup_css() but returns the effctive css, which is defined
247 * as the matching css of the nearest ancestor including self which has @ss
248 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
249 * function is guaranteed to return non-NULL css.
251 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
252 struct cgroup_subsys *ss)
254 lockdep_assert_held(&cgroup_mutex);
257 return &cgrp->dummy_css;
259 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
262 while (cgrp->parent &&
263 !(cgrp->parent->child_subsys_mask & (1 << ss->id)))
266 return cgroup_css(cgrp, ss);
269 /* convenient tests for these bits */
270 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
272 return test_bit(CGRP_DEAD, &cgrp->flags);
275 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
277 struct cgroup *cgrp = of->kn->parent->priv;
278 struct cftype *cft = of_cft(of);
281 * This is open and unprotected implementation of cgroup_css().
282 * seq_css() is only called from a kernfs file operation which has
283 * an active reference on the file. Because all the subsystem
284 * files are drained before a css is disassociated with a cgroup,
285 * the matching css from the cgroup's subsys table is guaranteed to
286 * be and stay valid until the enclosing operation is complete.
289 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
291 return &cgrp->dummy_css;
293 EXPORT_SYMBOL_GPL(of_css);
296 * cgroup_is_descendant - test ancestry
297 * @cgrp: the cgroup to be tested
298 * @ancestor: possible ancestor of @cgrp
300 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
301 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
302 * and @ancestor are accessible.
304 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
307 if (cgrp == ancestor)
314 static int cgroup_is_releasable(const struct cgroup *cgrp)
317 (1 << CGRP_RELEASABLE) |
318 (1 << CGRP_NOTIFY_ON_RELEASE);
319 return (cgrp->flags & bits) == bits;
322 static int notify_on_release(const struct cgroup *cgrp)
324 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
328 * for_each_css - iterate all css's of a cgroup
329 * @css: the iteration cursor
330 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
331 * @cgrp: the target cgroup to iterate css's of
333 * Should be called under cgroup_[tree_]mutex.
335 #define for_each_css(css, ssid, cgrp) \
336 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
337 if (!((css) = rcu_dereference_check( \
338 (cgrp)->subsys[(ssid)], \
339 lockdep_is_held(&cgroup_mutex)))) { } \
343 * for_each_e_css - iterate all effective css's of a cgroup
344 * @css: the iteration cursor
345 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
346 * @cgrp: the target cgroup to iterate css's of
348 * Should be called under cgroup_[tree_]mutex.
350 #define for_each_e_css(css, ssid, cgrp) \
351 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
352 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
357 * for_each_subsys - iterate all enabled cgroup subsystems
358 * @ss: the iteration cursor
359 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
361 #define for_each_subsys(ss, ssid) \
362 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
363 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
365 /* iterate across the hierarchies */
366 #define for_each_root(root) \
367 list_for_each_entry((root), &cgroup_roots, root_list)
369 /* iterate over child cgrps, lock should be held throughout iteration */
370 #define cgroup_for_each_live_child(child, cgrp) \
371 list_for_each_entry((child), &(cgrp)->children, sibling) \
372 if (({ lockdep_assert_held(&cgroup_mutex); \
373 cgroup_is_dead(child); })) \
377 /* the list of cgroups eligible for automatic release. Protected by
378 * release_list_lock */
379 static LIST_HEAD(release_list);
380 static DEFINE_RAW_SPINLOCK(release_list_lock);
381 static void cgroup_release_agent(struct work_struct *work);
382 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
383 static void check_for_release(struct cgroup *cgrp);
386 * A cgroup can be associated with multiple css_sets as different tasks may
387 * belong to different cgroups on different hierarchies. In the other
388 * direction, a css_set is naturally associated with multiple cgroups.
389 * This M:N relationship is represented by the following link structure
390 * which exists for each association and allows traversing the associations
393 struct cgrp_cset_link {
394 /* the cgroup and css_set this link associates */
396 struct css_set *cset;
398 /* list of cgrp_cset_links anchored at cgrp->cset_links */
399 struct list_head cset_link;
401 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
402 struct list_head cgrp_link;
406 * The default css_set - used by init and its children prior to any
407 * hierarchies being mounted. It contains a pointer to the root state
408 * for each subsystem. Also used to anchor the list of css_sets. Not
409 * reference-counted, to improve performance when child cgroups
410 * haven't been created.
412 struct css_set init_css_set = {
413 .refcount = ATOMIC_INIT(1),
414 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
415 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
416 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
417 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
418 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
421 static int css_set_count = 1; /* 1 for init_css_set */
424 * cgroup_update_populated - updated populated count of a cgroup
425 * @cgrp: the target cgroup
426 * @populated: inc or dec populated count
428 * @cgrp is either getting the first task (css_set) or losing the last.
429 * Update @cgrp->populated_cnt accordingly. The count is propagated
430 * towards root so that a given cgroup's populated_cnt is zero iff the
431 * cgroup and all its descendants are empty.
433 * @cgrp's interface file "cgroup.populated" is zero if
434 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
435 * changes from or to zero, userland is notified that the content of the
436 * interface file has changed. This can be used to detect when @cgrp and
437 * its descendants become populated or empty.
439 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
441 lockdep_assert_held(&css_set_rwsem);
447 trigger = !cgrp->populated_cnt++;
449 trigger = !--cgrp->populated_cnt;
454 if (cgrp->populated_kn)
455 kernfs_notify(cgrp->populated_kn);
461 * hash table for cgroup groups. This improves the performance to find
462 * an existing css_set. This hash doesn't (currently) take into
463 * account cgroups in empty hierarchies.
465 #define CSS_SET_HASH_BITS 7
466 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
468 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
470 unsigned long key = 0UL;
471 struct cgroup_subsys *ss;
474 for_each_subsys(ss, i)
475 key += (unsigned long)css[i];
476 key = (key >> 16) ^ key;
481 static void put_css_set_locked(struct css_set *cset, bool taskexit)
483 struct cgrp_cset_link *link, *tmp_link;
484 struct cgroup_subsys *ss;
487 lockdep_assert_held(&css_set_rwsem);
489 if (!atomic_dec_and_test(&cset->refcount))
492 /* This css_set is dead. unlink it and release cgroup refcounts */
493 for_each_subsys(ss, ssid)
494 list_del(&cset->e_cset_node[ssid]);
495 hash_del(&cset->hlist);
498 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
499 struct cgroup *cgrp = link->cgrp;
501 list_del(&link->cset_link);
502 list_del(&link->cgrp_link);
504 /* @cgrp can't go away while we're holding css_set_rwsem */
505 if (list_empty(&cgrp->cset_links)) {
506 cgroup_update_populated(cgrp, false);
507 if (notify_on_release(cgrp)) {
509 set_bit(CGRP_RELEASABLE, &cgrp->flags);
510 check_for_release(cgrp);
517 kfree_rcu(cset, rcu_head);
520 static void put_css_set(struct css_set *cset, bool taskexit)
523 * Ensure that the refcount doesn't hit zero while any readers
524 * can see it. Similar to atomic_dec_and_lock(), but for an
527 if (atomic_add_unless(&cset->refcount, -1, 1))
530 down_write(&css_set_rwsem);
531 put_css_set_locked(cset, taskexit);
532 up_write(&css_set_rwsem);
536 * refcounted get/put for css_set objects
538 static inline void get_css_set(struct css_set *cset)
540 atomic_inc(&cset->refcount);
544 * compare_css_sets - helper function for find_existing_css_set().
545 * @cset: candidate css_set being tested
546 * @old_cset: existing css_set for a task
547 * @new_cgrp: cgroup that's being entered by the task
548 * @template: desired set of css pointers in css_set (pre-calculated)
550 * Returns true if "cset" matches "old_cset" except for the hierarchy
551 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
553 static bool compare_css_sets(struct css_set *cset,
554 struct css_set *old_cset,
555 struct cgroup *new_cgrp,
556 struct cgroup_subsys_state *template[])
558 struct list_head *l1, *l2;
561 * On the default hierarchy, there can be csets which are
562 * associated with the same set of cgroups but different csses.
563 * Let's first ensure that csses match.
565 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
569 * Compare cgroup pointers in order to distinguish between
570 * different cgroups in hierarchies. As different cgroups may
571 * share the same effective css, this comparison is always
574 l1 = &cset->cgrp_links;
575 l2 = &old_cset->cgrp_links;
577 struct cgrp_cset_link *link1, *link2;
578 struct cgroup *cgrp1, *cgrp2;
582 /* See if we reached the end - both lists are equal length. */
583 if (l1 == &cset->cgrp_links) {
584 BUG_ON(l2 != &old_cset->cgrp_links);
587 BUG_ON(l2 == &old_cset->cgrp_links);
589 /* Locate the cgroups associated with these links. */
590 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
591 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
594 /* Hierarchies should be linked in the same order. */
595 BUG_ON(cgrp1->root != cgrp2->root);
598 * If this hierarchy is the hierarchy of the cgroup
599 * that's changing, then we need to check that this
600 * css_set points to the new cgroup; if it's any other
601 * hierarchy, then this css_set should point to the
602 * same cgroup as the old css_set.
604 if (cgrp1->root == new_cgrp->root) {
605 if (cgrp1 != new_cgrp)
616 * find_existing_css_set - init css array and find the matching css_set
617 * @old_cset: the css_set that we're using before the cgroup transition
618 * @cgrp: the cgroup that we're moving into
619 * @template: out param for the new set of csses, should be clear on entry
621 static struct css_set *find_existing_css_set(struct css_set *old_cset,
623 struct cgroup_subsys_state *template[])
625 struct cgroup_root *root = cgrp->root;
626 struct cgroup_subsys *ss;
627 struct css_set *cset;
632 * Build the set of subsystem state objects that we want to see in the
633 * new css_set. while subsystems can change globally, the entries here
634 * won't change, so no need for locking.
636 for_each_subsys(ss, i) {
637 if (root->subsys_mask & (1UL << i)) {
639 * @ss is in this hierarchy, so we want the
640 * effective css from @cgrp.
642 template[i] = cgroup_e_css(cgrp, ss);
645 * @ss is not in this hierarchy, so we don't want
648 template[i] = old_cset->subsys[i];
652 key = css_set_hash(template);
653 hash_for_each_possible(css_set_table, cset, hlist, key) {
654 if (!compare_css_sets(cset, old_cset, cgrp, template))
657 /* This css_set matches what we need */
661 /* No existing cgroup group matched */
665 static void free_cgrp_cset_links(struct list_head *links_to_free)
667 struct cgrp_cset_link *link, *tmp_link;
669 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
670 list_del(&link->cset_link);
676 * allocate_cgrp_cset_links - allocate cgrp_cset_links
677 * @count: the number of links to allocate
678 * @tmp_links: list_head the allocated links are put on
680 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
681 * through ->cset_link. Returns 0 on success or -errno.
683 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
685 struct cgrp_cset_link *link;
688 INIT_LIST_HEAD(tmp_links);
690 for (i = 0; i < count; i++) {
691 link = kzalloc(sizeof(*link), GFP_KERNEL);
693 free_cgrp_cset_links(tmp_links);
696 list_add(&link->cset_link, tmp_links);
702 * link_css_set - a helper function to link a css_set to a cgroup
703 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
704 * @cset: the css_set to be linked
705 * @cgrp: the destination cgroup
707 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
710 struct cgrp_cset_link *link;
712 BUG_ON(list_empty(tmp_links));
714 if (cgroup_on_dfl(cgrp))
715 cset->dfl_cgrp = cgrp;
717 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
721 if (list_empty(&cgrp->cset_links))
722 cgroup_update_populated(cgrp, true);
723 list_move(&link->cset_link, &cgrp->cset_links);
726 * Always add links to the tail of the list so that the list
727 * is sorted by order of hierarchy creation
729 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
733 * find_css_set - return a new css_set with one cgroup updated
734 * @old_cset: the baseline css_set
735 * @cgrp: the cgroup to be updated
737 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
738 * substituted into the appropriate hierarchy.
740 static struct css_set *find_css_set(struct css_set *old_cset,
743 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
744 struct css_set *cset;
745 struct list_head tmp_links;
746 struct cgrp_cset_link *link;
747 struct cgroup_subsys *ss;
751 lockdep_assert_held(&cgroup_mutex);
753 /* First see if we already have a cgroup group that matches
755 down_read(&css_set_rwsem);
756 cset = find_existing_css_set(old_cset, cgrp, template);
759 up_read(&css_set_rwsem);
764 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
768 /* Allocate all the cgrp_cset_link objects that we'll need */
769 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
774 atomic_set(&cset->refcount, 1);
775 INIT_LIST_HEAD(&cset->cgrp_links);
776 INIT_LIST_HEAD(&cset->tasks);
777 INIT_LIST_HEAD(&cset->mg_tasks);
778 INIT_LIST_HEAD(&cset->mg_preload_node);
779 INIT_LIST_HEAD(&cset->mg_node);
780 INIT_HLIST_NODE(&cset->hlist);
782 /* Copy the set of subsystem state objects generated in
783 * find_existing_css_set() */
784 memcpy(cset->subsys, template, sizeof(cset->subsys));
786 down_write(&css_set_rwsem);
787 /* Add reference counts and links from the new css_set. */
788 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
789 struct cgroup *c = link->cgrp;
791 if (c->root == cgrp->root)
793 link_css_set(&tmp_links, cset, c);
796 BUG_ON(!list_empty(&tmp_links));
800 /* Add @cset to the hash table */
801 key = css_set_hash(cset->subsys);
802 hash_add(css_set_table, &cset->hlist, key);
804 for_each_subsys(ss, ssid)
805 list_add_tail(&cset->e_cset_node[ssid],
806 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
808 up_write(&css_set_rwsem);
813 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
815 struct cgroup *root_cgrp = kf_root->kn->priv;
817 return root_cgrp->root;
820 static int cgroup_init_root_id(struct cgroup_root *root)
824 lockdep_assert_held(&cgroup_mutex);
826 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
830 root->hierarchy_id = id;
834 static void cgroup_exit_root_id(struct cgroup_root *root)
836 lockdep_assert_held(&cgroup_mutex);
838 if (root->hierarchy_id) {
839 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
840 root->hierarchy_id = 0;
844 static void cgroup_free_root(struct cgroup_root *root)
847 /* hierarhcy ID shoulid already have been released */
848 WARN_ON_ONCE(root->hierarchy_id);
850 idr_destroy(&root->cgroup_idr);
855 static void cgroup_destroy_root(struct cgroup_root *root)
857 struct cgroup *cgrp = &root->cgrp;
858 struct cgrp_cset_link *link, *tmp_link;
860 mutex_lock(&cgroup_mutex);
862 BUG_ON(atomic_read(&root->nr_cgrps));
863 BUG_ON(!list_empty(&cgrp->children));
865 /* Rebind all subsystems back to the default hierarchy */
866 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
869 * Release all the links from cset_links to this hierarchy's
872 down_write(&css_set_rwsem);
874 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
875 list_del(&link->cset_link);
876 list_del(&link->cgrp_link);
879 up_write(&css_set_rwsem);
881 if (!list_empty(&root->root_list)) {
882 list_del(&root->root_list);
886 cgroup_exit_root_id(root);
888 mutex_unlock(&cgroup_mutex);
890 kernfs_destroy_root(root->kf_root);
891 cgroup_free_root(root);
894 /* look up cgroup associated with given css_set on the specified hierarchy */
895 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
896 struct cgroup_root *root)
898 struct cgroup *res = NULL;
900 lockdep_assert_held(&cgroup_mutex);
901 lockdep_assert_held(&css_set_rwsem);
903 if (cset == &init_css_set) {
906 struct cgrp_cset_link *link;
908 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
909 struct cgroup *c = link->cgrp;
911 if (c->root == root) {
923 * Return the cgroup for "task" from the given hierarchy. Must be
924 * called with cgroup_mutex and css_set_rwsem held.
926 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
927 struct cgroup_root *root)
930 * No need to lock the task - since we hold cgroup_mutex the
931 * task can't change groups, so the only thing that can happen
932 * is that it exits and its css is set back to init_css_set.
934 return cset_cgroup_from_root(task_css_set(task), root);
938 * A task must hold cgroup_mutex to modify cgroups.
940 * Any task can increment and decrement the count field without lock.
941 * So in general, code holding cgroup_mutex can't rely on the count
942 * field not changing. However, if the count goes to zero, then only
943 * cgroup_attach_task() can increment it again. Because a count of zero
944 * means that no tasks are currently attached, therefore there is no
945 * way a task attached to that cgroup can fork (the other way to
946 * increment the count). So code holding cgroup_mutex can safely
947 * assume that if the count is zero, it will stay zero. Similarly, if
948 * a task holds cgroup_mutex on a cgroup with zero count, it
949 * knows that the cgroup won't be removed, as cgroup_rmdir()
952 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
953 * (usually) take cgroup_mutex. These are the two most performance
954 * critical pieces of code here. The exception occurs on cgroup_exit(),
955 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
956 * is taken, and if the cgroup count is zero, a usermode call made
957 * to the release agent with the name of the cgroup (path relative to
958 * the root of cgroup file system) as the argument.
960 * A cgroup can only be deleted if both its 'count' of using tasks
961 * is zero, and its list of 'children' cgroups is empty. Since all
962 * tasks in the system use _some_ cgroup, and since there is always at
963 * least one task in the system (init, pid == 1), therefore, root cgroup
964 * always has either children cgroups and/or using tasks. So we don't
965 * need a special hack to ensure that root cgroup cannot be deleted.
967 * P.S. One more locking exception. RCU is used to guard the
968 * update of a tasks cgroup pointer by cgroup_attach_task()
971 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
972 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
973 static const struct file_operations proc_cgroupstats_operations;
975 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
978 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
979 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
980 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
981 cft->ss->name, cft->name);
983 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
988 * cgroup_file_mode - deduce file mode of a control file
989 * @cft: the control file in question
991 * returns cft->mode if ->mode is not 0
992 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
993 * returns S_IRUGO if it has only a read handler
994 * returns S_IWUSR if it has only a write hander
996 static umode_t cgroup_file_mode(const struct cftype *cft)
1003 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1006 if (cft->write_u64 || cft->write_s64 || cft->write)
1012 static void cgroup_free_fn(struct work_struct *work)
1014 struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
1016 atomic_dec(&cgrp->root->nr_cgrps);
1017 cgroup_pidlist_destroy_all(cgrp);
1021 * We get a ref to the parent, and put the ref when this
1022 * cgroup is being freed, so it's guaranteed that the
1023 * parent won't be destroyed before its children.
1025 cgroup_put(cgrp->parent);
1026 kernfs_put(cgrp->kn);
1030 * This is root cgroup's refcnt reaching zero, which
1031 * indicates that the root should be released.
1033 cgroup_destroy_root(cgrp->root);
1037 static void cgroup_free_rcu(struct rcu_head *head)
1039 struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
1041 INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
1042 queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
1045 static void cgroup_get(struct cgroup *cgrp)
1047 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1048 WARN_ON_ONCE(atomic_read(&cgrp->refcnt) <= 0);
1049 atomic_inc(&cgrp->refcnt);
1052 static void cgroup_put(struct cgroup *cgrp)
1054 if (!atomic_dec_and_test(&cgrp->refcnt))
1056 if (WARN_ON_ONCE(cgrp->parent && !cgroup_is_dead(cgrp)))
1059 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
1062 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
1066 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1067 * @kn: the kernfs_node being serviced
1069 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1070 * the method finishes if locking succeeded. Note that once this function
1071 * returns the cgroup returned by cgroup_kn_lock_live() may become
1072 * inaccessible any time. If the caller intends to continue to access the
1073 * cgroup, it should pin it before invoking this function.
1075 static void cgroup_kn_unlock(struct kernfs_node *kn)
1077 struct cgroup *cgrp;
1079 if (kernfs_type(kn) == KERNFS_DIR)
1082 cgrp = kn->parent->priv;
1084 mutex_unlock(&cgroup_mutex);
1086 kernfs_unbreak_active_protection(kn);
1091 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1092 * @kn: the kernfs_node being serviced
1094 * This helper is to be used by a cgroup kernfs method currently servicing
1095 * @kn. It breaks the active protection, performs cgroup locking and
1096 * verifies that the associated cgroup is alive. Returns the cgroup if
1097 * alive; otherwise, %NULL. A successful return should be undone by a
1098 * matching cgroup_kn_unlock() invocation.
1100 * Any cgroup kernfs method implementation which requires locking the
1101 * associated cgroup should use this helper. It avoids nesting cgroup
1102 * locking under kernfs active protection and allows all kernfs operations
1103 * including self-removal.
1105 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1107 struct cgroup *cgrp;
1109 if (kernfs_type(kn) == KERNFS_DIR)
1112 cgrp = kn->parent->priv;
1115 * We're gonna grab cgroup_mutex which nests outside kernfs
1116 * active_ref. cgroup liveliness check alone provides enough
1117 * protection against removal. Ensure @cgrp stays accessible and
1118 * break the active_ref protection.
1121 kernfs_break_active_protection(kn);
1123 mutex_lock(&cgroup_mutex);
1125 if (!cgroup_is_dead(cgrp))
1128 cgroup_kn_unlock(kn);
1132 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1134 char name[CGROUP_FILE_NAME_MAX];
1136 lockdep_assert_held(&cgroup_mutex);
1137 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1141 * cgroup_clear_dir - remove subsys files in a cgroup directory
1142 * @cgrp: target cgroup
1143 * @subsys_mask: mask of the subsystem ids whose files should be removed
1145 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1147 struct cgroup_subsys *ss;
1150 for_each_subsys(ss, i) {
1151 struct cftype *cfts;
1153 if (!(subsys_mask & (1 << i)))
1155 list_for_each_entry(cfts, &ss->cfts, node)
1156 cgroup_addrm_files(cgrp, cfts, false);
1160 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1162 struct cgroup_subsys *ss;
1165 lockdep_assert_held(&cgroup_mutex);
1167 for_each_subsys(ss, ssid) {
1168 if (!(ss_mask & (1 << ssid)))
1171 /* if @ss has non-root csses attached to it, can't move */
1172 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1175 /* can't move between two non-dummy roots either */
1176 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1180 ret = cgroup_populate_dir(&dst_root->cgrp, ss_mask);
1182 if (dst_root != &cgrp_dfl_root)
1186 * Rebinding back to the default root is not allowed to
1187 * fail. Using both default and non-default roots should
1188 * be rare. Moving subsystems back and forth even more so.
1189 * Just warn about it and continue.
1191 if (cgrp_dfl_root_visible) {
1192 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1194 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1199 * Nothing can fail from this point on. Remove files for the
1200 * removed subsystems and rebind each subsystem.
1202 for_each_subsys(ss, ssid)
1203 if (ss_mask & (1 << ssid))
1204 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1206 for_each_subsys(ss, ssid) {
1207 struct cgroup_root *src_root;
1208 struct cgroup_subsys_state *css;
1209 struct css_set *cset;
1211 if (!(ss_mask & (1 << ssid)))
1214 src_root = ss->root;
1215 css = cgroup_css(&src_root->cgrp, ss);
1217 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1219 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1220 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1221 ss->root = dst_root;
1222 css->cgroup = &dst_root->cgrp;
1224 down_write(&css_set_rwsem);
1225 hash_for_each(css_set_table, i, cset, hlist)
1226 list_move_tail(&cset->e_cset_node[ss->id],
1227 &dst_root->cgrp.e_csets[ss->id]);
1228 up_write(&css_set_rwsem);
1230 src_root->subsys_mask &= ~(1 << ssid);
1231 src_root->cgrp.child_subsys_mask &= ~(1 << ssid);
1233 /* default hierarchy doesn't enable controllers by default */
1234 dst_root->subsys_mask |= 1 << ssid;
1235 if (dst_root != &cgrp_dfl_root)
1236 dst_root->cgrp.child_subsys_mask |= 1 << ssid;
1242 kernfs_activate(dst_root->cgrp.kn);
1246 static int cgroup_show_options(struct seq_file *seq,
1247 struct kernfs_root *kf_root)
1249 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1250 struct cgroup_subsys *ss;
1253 for_each_subsys(ss, ssid)
1254 if (root->subsys_mask & (1 << ssid))
1255 seq_printf(seq, ",%s", ss->name);
1256 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1257 seq_puts(seq, ",sane_behavior");
1258 if (root->flags & CGRP_ROOT_NOPREFIX)
1259 seq_puts(seq, ",noprefix");
1260 if (root->flags & CGRP_ROOT_XATTR)
1261 seq_puts(seq, ",xattr");
1263 spin_lock(&release_agent_path_lock);
1264 if (strlen(root->release_agent_path))
1265 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1266 spin_unlock(&release_agent_path_lock);
1268 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1269 seq_puts(seq, ",clone_children");
1270 if (strlen(root->name))
1271 seq_printf(seq, ",name=%s", root->name);
1275 struct cgroup_sb_opts {
1276 unsigned int subsys_mask;
1278 char *release_agent;
1279 bool cpuset_clone_children;
1281 /* User explicitly requested empty subsystem */
1285 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1287 char *token, *o = data;
1288 bool all_ss = false, one_ss = false;
1289 unsigned int mask = -1U;
1290 struct cgroup_subsys *ss;
1293 #ifdef CONFIG_CPUSETS
1294 mask = ~(1U << cpuset_cgrp_id);
1297 memset(opts, 0, sizeof(*opts));
1299 while ((token = strsep(&o, ",")) != NULL) {
1302 if (!strcmp(token, "none")) {
1303 /* Explicitly have no subsystems */
1307 if (!strcmp(token, "all")) {
1308 /* Mutually exclusive option 'all' + subsystem name */
1314 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1315 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1318 if (!strcmp(token, "noprefix")) {
1319 opts->flags |= CGRP_ROOT_NOPREFIX;
1322 if (!strcmp(token, "clone_children")) {
1323 opts->cpuset_clone_children = true;
1326 if (!strcmp(token, "xattr")) {
1327 opts->flags |= CGRP_ROOT_XATTR;
1330 if (!strncmp(token, "release_agent=", 14)) {
1331 /* Specifying two release agents is forbidden */
1332 if (opts->release_agent)
1334 opts->release_agent =
1335 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1336 if (!opts->release_agent)
1340 if (!strncmp(token, "name=", 5)) {
1341 const char *name = token + 5;
1342 /* Can't specify an empty name */
1345 /* Must match [\w.-]+ */
1346 for (i = 0; i < strlen(name); i++) {
1350 if ((c == '.') || (c == '-') || (c == '_'))
1354 /* Specifying two names is forbidden */
1357 opts->name = kstrndup(name,
1358 MAX_CGROUP_ROOT_NAMELEN - 1,
1366 for_each_subsys(ss, i) {
1367 if (strcmp(token, ss->name))
1372 /* Mutually exclusive option 'all' + subsystem name */
1375 opts->subsys_mask |= (1 << i);
1380 if (i == CGROUP_SUBSYS_COUNT)
1384 /* Consistency checks */
1386 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1387 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1389 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1390 opts->cpuset_clone_children || opts->release_agent ||
1392 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1397 * If the 'all' option was specified select all the
1398 * subsystems, otherwise if 'none', 'name=' and a subsystem
1399 * name options were not specified, let's default to 'all'
1401 if (all_ss || (!one_ss && !opts->none && !opts->name))
1402 for_each_subsys(ss, i)
1404 opts->subsys_mask |= (1 << i);
1407 * We either have to specify by name or by subsystems. (So
1408 * all empty hierarchies must have a name).
1410 if (!opts->subsys_mask && !opts->name)
1415 * Option noprefix was introduced just for backward compatibility
1416 * with the old cpuset, so we allow noprefix only if mounting just
1417 * the cpuset subsystem.
1419 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1423 /* Can't specify "none" and some subsystems */
1424 if (opts->subsys_mask && opts->none)
1430 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1433 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1434 struct cgroup_sb_opts opts;
1435 unsigned int added_mask, removed_mask;
1437 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1438 pr_err("sane_behavior: remount is not allowed\n");
1442 mutex_lock(&cgroup_mutex);
1444 /* See what subsystems are wanted */
1445 ret = parse_cgroupfs_options(data, &opts);
1449 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1450 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1451 task_tgid_nr(current), current->comm);
1453 added_mask = opts.subsys_mask & ~root->subsys_mask;
1454 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1456 /* Don't allow flags or name to change at remount */
1457 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1458 (opts.name && strcmp(opts.name, root->name))) {
1459 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1460 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1461 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1466 /* remounting is not allowed for populated hierarchies */
1467 if (!list_empty(&root->cgrp.children)) {
1472 ret = rebind_subsystems(root, added_mask);
1476 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1478 if (opts.release_agent) {
1479 spin_lock(&release_agent_path_lock);
1480 strcpy(root->release_agent_path, opts.release_agent);
1481 spin_unlock(&release_agent_path_lock);
1484 kfree(opts.release_agent);
1486 mutex_unlock(&cgroup_mutex);
1491 * To reduce the fork() overhead for systems that are not actually using
1492 * their cgroups capability, we don't maintain the lists running through
1493 * each css_set to its tasks until we see the list actually used - in other
1494 * words after the first mount.
1496 static bool use_task_css_set_links __read_mostly;
1498 static void cgroup_enable_task_cg_lists(void)
1500 struct task_struct *p, *g;
1502 down_write(&css_set_rwsem);
1504 if (use_task_css_set_links)
1507 use_task_css_set_links = true;
1510 * We need tasklist_lock because RCU is not safe against
1511 * while_each_thread(). Besides, a forking task that has passed
1512 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1513 * is not guaranteed to have its child immediately visible in the
1514 * tasklist if we walk through it with RCU.
1516 read_lock(&tasklist_lock);
1517 do_each_thread(g, p) {
1518 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1519 task_css_set(p) != &init_css_set);
1522 * We should check if the process is exiting, otherwise
1523 * it will race with cgroup_exit() in that the list
1524 * entry won't be deleted though the process has exited.
1525 * Do it while holding siglock so that we don't end up
1526 * racing against cgroup_exit().
1528 spin_lock_irq(&p->sighand->siglock);
1529 if (!(p->flags & PF_EXITING)) {
1530 struct css_set *cset = task_css_set(p);
1532 list_add(&p->cg_list, &cset->tasks);
1535 spin_unlock_irq(&p->sighand->siglock);
1536 } while_each_thread(g, p);
1537 read_unlock(&tasklist_lock);
1539 up_write(&css_set_rwsem);
1542 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1544 struct cgroup_subsys *ss;
1547 atomic_set(&cgrp->refcnt, 1);
1548 INIT_LIST_HEAD(&cgrp->sibling);
1549 INIT_LIST_HEAD(&cgrp->children);
1550 INIT_LIST_HEAD(&cgrp->cset_links);
1551 INIT_LIST_HEAD(&cgrp->release_list);
1552 INIT_LIST_HEAD(&cgrp->pidlists);
1553 mutex_init(&cgrp->pidlist_mutex);
1554 cgrp->dummy_css.cgroup = cgrp;
1556 for_each_subsys(ss, ssid)
1557 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1559 init_waitqueue_head(&cgrp->offline_waitq);
1562 static void init_cgroup_root(struct cgroup_root *root,
1563 struct cgroup_sb_opts *opts)
1565 struct cgroup *cgrp = &root->cgrp;
1567 INIT_LIST_HEAD(&root->root_list);
1568 atomic_set(&root->nr_cgrps, 1);
1570 init_cgroup_housekeeping(cgrp);
1571 idr_init(&root->cgroup_idr);
1573 root->flags = opts->flags;
1574 if (opts->release_agent)
1575 strcpy(root->release_agent_path, opts->release_agent);
1577 strcpy(root->name, opts->name);
1578 if (opts->cpuset_clone_children)
1579 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1582 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1584 LIST_HEAD(tmp_links);
1585 struct cgroup *root_cgrp = &root->cgrp;
1586 struct css_set *cset;
1589 lockdep_assert_held(&cgroup_mutex);
1591 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1594 root_cgrp->id = ret;
1597 * We're accessing css_set_count without locking css_set_rwsem here,
1598 * but that's OK - it can only be increased by someone holding
1599 * cgroup_lock, and that's us. The worst that can happen is that we
1600 * have some link structures left over
1602 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1606 ret = cgroup_init_root_id(root);
1610 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1611 KERNFS_ROOT_CREATE_DEACTIVATED,
1613 if (IS_ERR(root->kf_root)) {
1614 ret = PTR_ERR(root->kf_root);
1617 root_cgrp->kn = root->kf_root->kn;
1619 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1623 ret = rebind_subsystems(root, ss_mask);
1628 * There must be no failure case after here, since rebinding takes
1629 * care of subsystems' refcounts, which are explicitly dropped in
1630 * the failure exit path.
1632 list_add(&root->root_list, &cgroup_roots);
1633 cgroup_root_count++;
1636 * Link the root cgroup in this hierarchy into all the css_set
1639 down_write(&css_set_rwsem);
1640 hash_for_each(css_set_table, i, cset, hlist)
1641 link_css_set(&tmp_links, cset, root_cgrp);
1642 up_write(&css_set_rwsem);
1644 BUG_ON(!list_empty(&root_cgrp->children));
1645 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1647 kernfs_activate(root_cgrp->kn);
1652 kernfs_destroy_root(root->kf_root);
1653 root->kf_root = NULL;
1655 cgroup_exit_root_id(root);
1657 free_cgrp_cset_links(&tmp_links);
1661 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1662 int flags, const char *unused_dev_name,
1665 struct cgroup_root *root;
1666 struct cgroup_sb_opts opts;
1667 struct dentry *dentry;
1672 * The first time anyone tries to mount a cgroup, enable the list
1673 * linking each css_set to its tasks and fix up all existing tasks.
1675 if (!use_task_css_set_links)
1676 cgroup_enable_task_cg_lists();
1678 mutex_lock(&cgroup_mutex);
1680 /* First find the desired set of subsystems */
1681 ret = parse_cgroupfs_options(data, &opts);
1685 /* look for a matching existing root */
1686 if (!opts.subsys_mask && !opts.none && !opts.name) {
1687 cgrp_dfl_root_visible = true;
1688 root = &cgrp_dfl_root;
1689 cgroup_get(&root->cgrp);
1694 for_each_root(root) {
1695 bool name_match = false;
1697 if (root == &cgrp_dfl_root)
1701 * If we asked for a name then it must match. Also, if
1702 * name matches but sybsys_mask doesn't, we should fail.
1703 * Remember whether name matched.
1706 if (strcmp(opts.name, root->name))
1712 * If we asked for subsystems (or explicitly for no
1713 * subsystems) then they must match.
1715 if ((opts.subsys_mask || opts.none) &&
1716 (opts.subsys_mask != root->subsys_mask)) {
1723 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1724 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1725 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1729 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1734 * A root's lifetime is governed by its root cgroup. Zero
1735 * ref indicate that the root is being destroyed. Wait for
1736 * destruction to complete so that the subsystems are free.
1737 * We can use wait_queue for the wait but this path is
1738 * super cold. Let's just sleep for a bit and retry.
1740 if (!atomic_inc_not_zero(&root->cgrp.refcnt)) {
1741 mutex_unlock(&cgroup_mutex);
1743 ret = restart_syscall();
1752 * No such thing, create a new one. name= matching without subsys
1753 * specification is allowed for already existing hierarchies but we
1754 * can't create new one without subsys specification.
1756 if (!opts.subsys_mask && !opts.none) {
1761 root = kzalloc(sizeof(*root), GFP_KERNEL);
1767 init_cgroup_root(root, &opts);
1769 ret = cgroup_setup_root(root, opts.subsys_mask);
1771 cgroup_free_root(root);
1774 mutex_unlock(&cgroup_mutex);
1776 kfree(opts.release_agent);
1780 return ERR_PTR(ret);
1782 dentry = kernfs_mount(fs_type, flags, root->kf_root, &new_sb);
1783 if (IS_ERR(dentry) || !new_sb)
1784 cgroup_put(&root->cgrp);
1788 static void cgroup_kill_sb(struct super_block *sb)
1790 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1791 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1793 cgroup_put(&root->cgrp);
1797 static struct file_system_type cgroup_fs_type = {
1799 .mount = cgroup_mount,
1800 .kill_sb = cgroup_kill_sb,
1803 static struct kobject *cgroup_kobj;
1806 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1807 * @task: target task
1808 * @buf: the buffer to write the path into
1809 * @buflen: the length of the buffer
1811 * Determine @task's cgroup on the first (the one with the lowest non-zero
1812 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1813 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1814 * cgroup controller callbacks.
1816 * Return value is the same as kernfs_path().
1818 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1820 struct cgroup_root *root;
1821 struct cgroup *cgrp;
1822 int hierarchy_id = 1;
1825 mutex_lock(&cgroup_mutex);
1826 down_read(&css_set_rwsem);
1828 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1831 cgrp = task_cgroup_from_root(task, root);
1832 path = cgroup_path(cgrp, buf, buflen);
1834 /* if no hierarchy exists, everyone is in "/" */
1835 if (strlcpy(buf, "/", buflen) < buflen)
1839 up_read(&css_set_rwsem);
1840 mutex_unlock(&cgroup_mutex);
1843 EXPORT_SYMBOL_GPL(task_cgroup_path);
1845 /* used to track tasks and other necessary states during migration */
1846 struct cgroup_taskset {
1847 /* the src and dst cset list running through cset->mg_node */
1848 struct list_head src_csets;
1849 struct list_head dst_csets;
1852 * Fields for cgroup_taskset_*() iteration.
1854 * Before migration is committed, the target migration tasks are on
1855 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1856 * the csets on ->dst_csets. ->csets point to either ->src_csets
1857 * or ->dst_csets depending on whether migration is committed.
1859 * ->cur_csets and ->cur_task point to the current task position
1862 struct list_head *csets;
1863 struct css_set *cur_cset;
1864 struct task_struct *cur_task;
1868 * cgroup_taskset_first - reset taskset and return the first task
1869 * @tset: taskset of interest
1871 * @tset iteration is initialized and the first task is returned.
1873 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1875 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1876 tset->cur_task = NULL;
1878 return cgroup_taskset_next(tset);
1882 * cgroup_taskset_next - iterate to the next task in taskset
1883 * @tset: taskset of interest
1885 * Return the next task in @tset. Iteration must have been initialized
1886 * with cgroup_taskset_first().
1888 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1890 struct css_set *cset = tset->cur_cset;
1891 struct task_struct *task = tset->cur_task;
1893 while (&cset->mg_node != tset->csets) {
1895 task = list_first_entry(&cset->mg_tasks,
1896 struct task_struct, cg_list);
1898 task = list_next_entry(task, cg_list);
1900 if (&task->cg_list != &cset->mg_tasks) {
1901 tset->cur_cset = cset;
1902 tset->cur_task = task;
1906 cset = list_next_entry(cset, mg_node);
1914 * cgroup_task_migrate - move a task from one cgroup to another.
1915 * @old_cgrp: the cgroup @tsk is being migrated from
1916 * @tsk: the task being migrated
1917 * @new_cset: the new css_set @tsk is being attached to
1919 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1921 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1922 struct task_struct *tsk,
1923 struct css_set *new_cset)
1925 struct css_set *old_cset;
1927 lockdep_assert_held(&cgroup_mutex);
1928 lockdep_assert_held(&css_set_rwsem);
1931 * We are synchronized through threadgroup_lock() against PF_EXITING
1932 * setting such that we can't race against cgroup_exit() changing the
1933 * css_set to init_css_set and dropping the old one.
1935 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1936 old_cset = task_css_set(tsk);
1938 get_css_set(new_cset);
1939 rcu_assign_pointer(tsk->cgroups, new_cset);
1942 * Use move_tail so that cgroup_taskset_first() still returns the
1943 * leader after migration. This works because cgroup_migrate()
1944 * ensures that the dst_cset of the leader is the first on the
1945 * tset's dst_csets list.
1947 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1950 * We just gained a reference on old_cset by taking it from the
1951 * task. As trading it for new_cset is protected by cgroup_mutex,
1952 * we're safe to drop it here; it will be freed under RCU.
1954 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1955 put_css_set_locked(old_cset, false);
1959 * cgroup_migrate_finish - cleanup after attach
1960 * @preloaded_csets: list of preloaded css_sets
1962 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1963 * those functions for details.
1965 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1967 struct css_set *cset, *tmp_cset;
1969 lockdep_assert_held(&cgroup_mutex);
1971 down_write(&css_set_rwsem);
1972 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1973 cset->mg_src_cgrp = NULL;
1974 cset->mg_dst_cset = NULL;
1975 list_del_init(&cset->mg_preload_node);
1976 put_css_set_locked(cset, false);
1978 up_write(&css_set_rwsem);
1982 * cgroup_migrate_add_src - add a migration source css_set
1983 * @src_cset: the source css_set to add
1984 * @dst_cgrp: the destination cgroup
1985 * @preloaded_csets: list of preloaded css_sets
1987 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1988 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1989 * up by cgroup_migrate_finish().
1991 * This function may be called without holding threadgroup_lock even if the
1992 * target is a process. Threads may be created and destroyed but as long
1993 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1994 * the preloaded css_sets are guaranteed to cover all migrations.
1996 static void cgroup_migrate_add_src(struct css_set *src_cset,
1997 struct cgroup *dst_cgrp,
1998 struct list_head *preloaded_csets)
2000 struct cgroup *src_cgrp;
2002 lockdep_assert_held(&cgroup_mutex);
2003 lockdep_assert_held(&css_set_rwsem);
2005 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2007 if (!list_empty(&src_cset->mg_preload_node))
2010 WARN_ON(src_cset->mg_src_cgrp);
2011 WARN_ON(!list_empty(&src_cset->mg_tasks));
2012 WARN_ON(!list_empty(&src_cset->mg_node));
2014 src_cset->mg_src_cgrp = src_cgrp;
2015 get_css_set(src_cset);
2016 list_add(&src_cset->mg_preload_node, preloaded_csets);
2020 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2021 * @dst_cgrp: the destination cgroup (may be %NULL)
2022 * @preloaded_csets: list of preloaded source css_sets
2024 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2025 * have been preloaded to @preloaded_csets. This function looks up and
2026 * pins all destination css_sets, links each to its source, and append them
2027 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2028 * source css_set is assumed to be its cgroup on the default hierarchy.
2030 * This function must be called after cgroup_migrate_add_src() has been
2031 * called on each migration source css_set. After migration is performed
2032 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2035 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2036 struct list_head *preloaded_csets)
2039 struct css_set *src_cset, *tmp_cset;
2041 lockdep_assert_held(&cgroup_mutex);
2044 * Except for the root, child_subsys_mask must be zero for a cgroup
2045 * with tasks so that child cgroups don't compete against tasks.
2047 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && dst_cgrp->parent &&
2048 dst_cgrp->child_subsys_mask)
2051 /* look up the dst cset for each src cset and link it to src */
2052 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2053 struct css_set *dst_cset;
2055 dst_cset = find_css_set(src_cset,
2056 dst_cgrp ?: src_cset->dfl_cgrp);
2060 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2063 * If src cset equals dst, it's noop. Drop the src.
2064 * cgroup_migrate() will skip the cset too. Note that we
2065 * can't handle src == dst as some nodes are used by both.
2067 if (src_cset == dst_cset) {
2068 src_cset->mg_src_cgrp = NULL;
2069 list_del_init(&src_cset->mg_preload_node);
2070 put_css_set(src_cset, false);
2071 put_css_set(dst_cset, false);
2075 src_cset->mg_dst_cset = dst_cset;
2077 if (list_empty(&dst_cset->mg_preload_node))
2078 list_add(&dst_cset->mg_preload_node, &csets);
2080 put_css_set(dst_cset, false);
2083 list_splice_tail(&csets, preloaded_csets);
2086 cgroup_migrate_finish(&csets);
2091 * cgroup_migrate - migrate a process or task to a cgroup
2092 * @cgrp: the destination cgroup
2093 * @leader: the leader of the process or the task to migrate
2094 * @threadgroup: whether @leader points to the whole process or a single task
2096 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2097 * process, the caller must be holding threadgroup_lock of @leader. The
2098 * caller is also responsible for invoking cgroup_migrate_add_src() and
2099 * cgroup_migrate_prepare_dst() on the targets before invoking this
2100 * function and following up with cgroup_migrate_finish().
2102 * As long as a controller's ->can_attach() doesn't fail, this function is
2103 * guaranteed to succeed. This means that, excluding ->can_attach()
2104 * failure, when migrating multiple targets, the success or failure can be
2105 * decided for all targets by invoking group_migrate_prepare_dst() before
2106 * actually starting migrating.
2108 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2111 struct cgroup_taskset tset = {
2112 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2113 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2114 .csets = &tset.src_csets,
2116 struct cgroup_subsys_state *css, *failed_css = NULL;
2117 struct css_set *cset, *tmp_cset;
2118 struct task_struct *task, *tmp_task;
2122 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2123 * already PF_EXITING could be freed from underneath us unless we
2124 * take an rcu_read_lock.
2126 down_write(&css_set_rwsem);
2130 /* @task either already exited or can't exit until the end */
2131 if (task->flags & PF_EXITING)
2134 /* leave @task alone if post_fork() hasn't linked it yet */
2135 if (list_empty(&task->cg_list))
2138 cset = task_css_set(task);
2139 if (!cset->mg_src_cgrp)
2143 * cgroup_taskset_first() must always return the leader.
2144 * Take care to avoid disturbing the ordering.
2146 list_move_tail(&task->cg_list, &cset->mg_tasks);
2147 if (list_empty(&cset->mg_node))
2148 list_add_tail(&cset->mg_node, &tset.src_csets);
2149 if (list_empty(&cset->mg_dst_cset->mg_node))
2150 list_move_tail(&cset->mg_dst_cset->mg_node,
2155 } while_each_thread(leader, task);
2157 up_write(&css_set_rwsem);
2159 /* methods shouldn't be called if no task is actually migrating */
2160 if (list_empty(&tset.src_csets))
2163 /* check that we can legitimately attach to the cgroup */
2164 for_each_e_css(css, i, cgrp) {
2165 if (css->ss->can_attach) {
2166 ret = css->ss->can_attach(css, &tset);
2169 goto out_cancel_attach;
2175 * Now that we're guaranteed success, proceed to move all tasks to
2176 * the new cgroup. There are no failure cases after here, so this
2177 * is the commit point.
2179 down_write(&css_set_rwsem);
2180 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2181 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2182 cgroup_task_migrate(cset->mg_src_cgrp, task,
2185 up_write(&css_set_rwsem);
2188 * Migration is committed, all target tasks are now on dst_csets.
2189 * Nothing is sensitive to fork() after this point. Notify
2190 * controllers that migration is complete.
2192 tset.csets = &tset.dst_csets;
2194 for_each_e_css(css, i, cgrp)
2195 if (css->ss->attach)
2196 css->ss->attach(css, &tset);
2199 goto out_release_tset;
2202 for_each_e_css(css, i, cgrp) {
2203 if (css == failed_css)
2205 if (css->ss->cancel_attach)
2206 css->ss->cancel_attach(css, &tset);
2209 down_write(&css_set_rwsem);
2210 list_splice_init(&tset.dst_csets, &tset.src_csets);
2211 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2212 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2213 list_del_init(&cset->mg_node);
2215 up_write(&css_set_rwsem);
2220 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2221 * @dst_cgrp: the cgroup to attach to
2222 * @leader: the task or the leader of the threadgroup to be attached
2223 * @threadgroup: attach the whole threadgroup?
2225 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2227 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2228 struct task_struct *leader, bool threadgroup)
2230 LIST_HEAD(preloaded_csets);
2231 struct task_struct *task;
2234 /* look up all src csets */
2235 down_read(&css_set_rwsem);
2239 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2243 } while_each_thread(leader, task);
2245 up_read(&css_set_rwsem);
2247 /* prepare dst csets and commit */
2248 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2250 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2252 cgroup_migrate_finish(&preloaded_csets);
2257 * Find the task_struct of the task to attach by vpid and pass it along to the
2258 * function to attach either it or all tasks in its threadgroup. Will lock
2259 * cgroup_mutex and threadgroup.
2261 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2262 size_t nbytes, loff_t off, bool threadgroup)
2264 struct task_struct *tsk;
2265 const struct cred *cred = current_cred(), *tcred;
2266 struct cgroup *cgrp;
2270 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2273 cgrp = cgroup_kn_lock_live(of->kn);
2280 tsk = find_task_by_vpid(pid);
2284 goto out_unlock_cgroup;
2287 * even if we're attaching all tasks in the thread group, we
2288 * only need to check permissions on one of them.
2290 tcred = __task_cred(tsk);
2291 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2292 !uid_eq(cred->euid, tcred->uid) &&
2293 !uid_eq(cred->euid, tcred->suid)) {
2296 goto out_unlock_cgroup;
2302 tsk = tsk->group_leader;
2305 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2306 * trapped in a cpuset, or RT worker may be born in a cgroup
2307 * with no rt_runtime allocated. Just say no.
2309 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2312 goto out_unlock_cgroup;
2315 get_task_struct(tsk);
2318 threadgroup_lock(tsk);
2320 if (!thread_group_leader(tsk)) {
2322 * a race with de_thread from another thread's exec()
2323 * may strip us of our leadership, if this happens,
2324 * there is no choice but to throw this task away and
2325 * try again; this is
2326 * "double-double-toil-and-trouble-check locking".
2328 threadgroup_unlock(tsk);
2329 put_task_struct(tsk);
2330 goto retry_find_task;
2334 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2336 threadgroup_unlock(tsk);
2338 put_task_struct(tsk);
2340 cgroup_kn_unlock(of->kn);
2341 return ret ?: nbytes;
2345 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2346 * @from: attach to all cgroups of a given task
2347 * @tsk: the task to be attached
2349 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2351 struct cgroup_root *root;
2354 mutex_lock(&cgroup_mutex);
2355 for_each_root(root) {
2356 struct cgroup *from_cgrp;
2358 if (root == &cgrp_dfl_root)
2361 down_read(&css_set_rwsem);
2362 from_cgrp = task_cgroup_from_root(from, root);
2363 up_read(&css_set_rwsem);
2365 retval = cgroup_attach_task(from_cgrp, tsk, false);
2369 mutex_unlock(&cgroup_mutex);
2373 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2375 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2376 char *buf, size_t nbytes, loff_t off)
2378 return __cgroup_procs_write(of, buf, nbytes, off, false);
2381 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2382 char *buf, size_t nbytes, loff_t off)
2384 return __cgroup_procs_write(of, buf, nbytes, off, true);
2387 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2388 char *buf, size_t nbytes, loff_t off)
2390 struct cgroup *cgrp;
2392 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2394 cgrp = cgroup_kn_lock_live(of->kn);
2397 spin_lock(&release_agent_path_lock);
2398 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2399 sizeof(cgrp->root->release_agent_path));
2400 spin_unlock(&release_agent_path_lock);
2401 cgroup_kn_unlock(of->kn);
2405 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2407 struct cgroup *cgrp = seq_css(seq)->cgroup;
2409 spin_lock(&release_agent_path_lock);
2410 seq_puts(seq, cgrp->root->release_agent_path);
2411 spin_unlock(&release_agent_path_lock);
2412 seq_putc(seq, '\n');
2416 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2418 struct cgroup *cgrp = seq_css(seq)->cgroup;
2420 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2424 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2426 struct cgroup_subsys *ss;
2427 bool printed = false;
2430 for_each_subsys(ss, ssid) {
2431 if (ss_mask & (1 << ssid)) {
2434 seq_printf(seq, "%s", ss->name);
2439 seq_putc(seq, '\n');
2442 /* show controllers which are currently attached to the default hierarchy */
2443 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2445 struct cgroup *cgrp = seq_css(seq)->cgroup;
2447 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask);
2451 /* show controllers which are enabled from the parent */
2452 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2454 struct cgroup *cgrp = seq_css(seq)->cgroup;
2456 cgroup_print_ss_mask(seq, cgrp->parent->child_subsys_mask);
2460 /* show controllers which are enabled for a given cgroup's children */
2461 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2463 struct cgroup *cgrp = seq_css(seq)->cgroup;
2465 cgroup_print_ss_mask(seq, cgrp->child_subsys_mask);
2470 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2471 * @cgrp: root of the subtree to update csses for
2473 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2474 * css associations need to be updated accordingly. This function looks up
2475 * all css_sets which are attached to the subtree, creates the matching
2476 * updated css_sets and migrates the tasks to the new ones.
2478 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2480 LIST_HEAD(preloaded_csets);
2481 struct cgroup_subsys_state *css;
2482 struct css_set *src_cset;
2485 lockdep_assert_held(&cgroup_mutex);
2487 /* look up all csses currently attached to @cgrp's subtree */
2488 down_read(&css_set_rwsem);
2489 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2490 struct cgrp_cset_link *link;
2492 /* self is not affected by child_subsys_mask change */
2493 if (css->cgroup == cgrp)
2496 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2497 cgroup_migrate_add_src(link->cset, cgrp,
2500 up_read(&css_set_rwsem);
2502 /* NULL dst indicates self on default hierarchy */
2503 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2507 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2508 struct task_struct *last_task = NULL, *task;
2510 /* src_csets precede dst_csets, break on the first dst_cset */
2511 if (!src_cset->mg_src_cgrp)
2515 * All tasks in src_cset need to be migrated to the
2516 * matching dst_cset. Empty it process by process. We
2517 * walk tasks but migrate processes. The leader might even
2518 * belong to a different cset but such src_cset would also
2519 * be among the target src_csets because the default
2520 * hierarchy enforces per-process membership.
2523 down_read(&css_set_rwsem);
2524 task = list_first_entry_or_null(&src_cset->tasks,
2525 struct task_struct, cg_list);
2527 task = task->group_leader;
2528 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2529 get_task_struct(task);
2531 up_read(&css_set_rwsem);
2536 /* guard against possible infinite loop */
2537 if (WARN(last_task == task,
2538 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2542 threadgroup_lock(task);
2543 /* raced against de_thread() from another thread? */
2544 if (!thread_group_leader(task)) {
2545 threadgroup_unlock(task);
2546 put_task_struct(task);
2550 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2552 threadgroup_unlock(task);
2553 put_task_struct(task);
2555 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2561 cgroup_migrate_finish(&preloaded_csets);
2565 /* change the enabled child controllers for a cgroup in the default hierarchy */
2566 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2567 char *buf, size_t nbytes,
2570 unsigned int enable = 0, disable = 0;
2571 struct cgroup *cgrp, *child;
2572 struct cgroup_subsys *ss;
2577 * Parse input - space separated list of subsystem names prefixed
2578 * with either + or -.
2580 buf = strstrip(buf);
2581 while ((tok = strsep(&buf, " "))) {
2584 for_each_subsys(ss, ssid) {
2585 if (ss->disabled || strcmp(tok + 1, ss->name))
2589 enable |= 1 << ssid;
2590 disable &= ~(1 << ssid);
2591 } else if (*tok == '-') {
2592 disable |= 1 << ssid;
2593 enable &= ~(1 << ssid);
2599 if (ssid == CGROUP_SUBSYS_COUNT)
2603 cgrp = cgroup_kn_lock_live(of->kn);
2607 for_each_subsys(ss, ssid) {
2608 if (enable & (1 << ssid)) {
2609 if (cgrp->child_subsys_mask & (1 << ssid)) {
2610 enable &= ~(1 << ssid);
2615 * Because css offlining is asynchronous, userland
2616 * might try to re-enable the same controller while
2617 * the previous instance is still around. In such
2618 * cases, wait till it's gone using offline_waitq.
2620 cgroup_for_each_live_child(child, cgrp) {
2623 if (!cgroup_css(child, ss))
2627 prepare_to_wait(&child->offline_waitq, &wait,
2628 TASK_UNINTERRUPTIBLE);
2629 cgroup_kn_unlock(of->kn);
2631 finish_wait(&child->offline_waitq, &wait);
2634 return restart_syscall();
2637 /* unavailable or not enabled on the parent? */
2638 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2640 !(cgrp->parent->child_subsys_mask & (1 << ssid)))) {
2644 } else if (disable & (1 << ssid)) {
2645 if (!(cgrp->child_subsys_mask & (1 << ssid))) {
2646 disable &= ~(1 << ssid);
2650 /* a child has it enabled? */
2651 cgroup_for_each_live_child(child, cgrp) {
2652 if (child->child_subsys_mask & (1 << ssid)) {
2660 if (!enable && !disable) {
2666 * Except for the root, child_subsys_mask must be zero for a cgroup
2667 * with tasks so that child cgroups don't compete against tasks.
2669 if (enable && cgrp->parent && !list_empty(&cgrp->cset_links)) {
2675 * Create csses for enables and update child_subsys_mask. This
2676 * changes cgroup_e_css() results which in turn makes the
2677 * subsequent cgroup_update_dfl_csses() associate all tasks in the
2678 * subtree to the updated csses.
2680 for_each_subsys(ss, ssid) {
2681 if (!(enable & (1 << ssid)))
2684 cgroup_for_each_live_child(child, cgrp) {
2685 ret = create_css(child, ss);
2691 cgrp->child_subsys_mask |= enable;
2692 cgrp->child_subsys_mask &= ~disable;
2694 ret = cgroup_update_dfl_csses(cgrp);
2698 /* all tasks are now migrated away from the old csses, kill them */
2699 for_each_subsys(ss, ssid) {
2700 if (!(disable & (1 << ssid)))
2703 cgroup_for_each_live_child(child, cgrp)
2704 kill_css(cgroup_css(child, ss));
2707 kernfs_activate(cgrp->kn);
2710 cgroup_kn_unlock(of->kn);
2711 return ret ?: nbytes;
2714 cgrp->child_subsys_mask &= ~enable;
2715 cgrp->child_subsys_mask |= disable;
2717 for_each_subsys(ss, ssid) {
2718 if (!(enable & (1 << ssid)))
2721 cgroup_for_each_live_child(child, cgrp) {
2722 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2730 static int cgroup_populated_show(struct seq_file *seq, void *v)
2732 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2736 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2737 size_t nbytes, loff_t off)
2739 struct cgroup *cgrp = of->kn->parent->priv;
2740 struct cftype *cft = of->kn->priv;
2741 struct cgroup_subsys_state *css;
2745 return cft->write(of, buf, nbytes, off);
2748 * kernfs guarantees that a file isn't deleted with operations in
2749 * flight, which means that the matching css is and stays alive and
2750 * doesn't need to be pinned. The RCU locking is not necessary
2751 * either. It's just for the convenience of using cgroup_css().
2754 css = cgroup_css(cgrp, cft->ss);
2757 if (cft->write_u64) {
2758 unsigned long long v;
2759 ret = kstrtoull(buf, 0, &v);
2761 ret = cft->write_u64(css, cft, v);
2762 } else if (cft->write_s64) {
2764 ret = kstrtoll(buf, 0, &v);
2766 ret = cft->write_s64(css, cft, v);
2771 return ret ?: nbytes;
2774 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2776 return seq_cft(seq)->seq_start(seq, ppos);
2779 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2781 return seq_cft(seq)->seq_next(seq, v, ppos);
2784 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2786 seq_cft(seq)->seq_stop(seq, v);
2789 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2791 struct cftype *cft = seq_cft(m);
2792 struct cgroup_subsys_state *css = seq_css(m);
2795 return cft->seq_show(m, arg);
2798 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2799 else if (cft->read_s64)
2800 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2806 static struct kernfs_ops cgroup_kf_single_ops = {
2807 .atomic_write_len = PAGE_SIZE,
2808 .write = cgroup_file_write,
2809 .seq_show = cgroup_seqfile_show,
2812 static struct kernfs_ops cgroup_kf_ops = {
2813 .atomic_write_len = PAGE_SIZE,
2814 .write = cgroup_file_write,
2815 .seq_start = cgroup_seqfile_start,
2816 .seq_next = cgroup_seqfile_next,
2817 .seq_stop = cgroup_seqfile_stop,
2818 .seq_show = cgroup_seqfile_show,
2822 * cgroup_rename - Only allow simple rename of directories in place.
2824 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2825 const char *new_name_str)
2827 struct cgroup *cgrp = kn->priv;
2830 if (kernfs_type(kn) != KERNFS_DIR)
2832 if (kn->parent != new_parent)
2836 * This isn't a proper migration and its usefulness is very
2837 * limited. Disallow if sane_behavior.
2839 if (cgroup_sane_behavior(cgrp))
2843 * We're gonna grab cgroup_mutex which nests outside kernfs
2844 * active_ref. kernfs_rename() doesn't require active_ref
2845 * protection. Break them before grabbing cgroup_mutex.
2847 kernfs_break_active_protection(new_parent);
2848 kernfs_break_active_protection(kn);
2850 mutex_lock(&cgroup_mutex);
2852 ret = kernfs_rename(kn, new_parent, new_name_str);
2854 mutex_unlock(&cgroup_mutex);
2856 kernfs_unbreak_active_protection(kn);
2857 kernfs_unbreak_active_protection(new_parent);
2861 /* set uid and gid of cgroup dirs and files to that of the creator */
2862 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2864 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2865 .ia_uid = current_fsuid(),
2866 .ia_gid = current_fsgid(), };
2868 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2869 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2872 return kernfs_setattr(kn, &iattr);
2875 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2877 char name[CGROUP_FILE_NAME_MAX];
2878 struct kernfs_node *kn;
2879 struct lock_class_key *key = NULL;
2882 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2883 key = &cft->lockdep_key;
2885 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2886 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2891 ret = cgroup_kn_set_ugid(kn);
2897 if (cft->seq_show == cgroup_populated_show)
2898 cgrp->populated_kn = kn;
2903 * cgroup_addrm_files - add or remove files to a cgroup directory
2904 * @cgrp: the target cgroup
2905 * @cfts: array of cftypes to be added
2906 * @is_add: whether to add or remove
2908 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2909 * For removals, this function never fails. If addition fails, this
2910 * function doesn't remove files already added. The caller is responsible
2913 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2919 lockdep_assert_held(&cgroup_mutex);
2921 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2922 /* does cft->flags tell us to skip this file on @cgrp? */
2923 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2925 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2927 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2929 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2933 ret = cgroup_add_file(cgrp, cft);
2935 pr_warn("%s: failed to add %s, err=%d\n",
2936 __func__, cft->name, ret);
2940 cgroup_rm_file(cgrp, cft);
2946 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
2949 struct cgroup_subsys *ss = cfts[0].ss;
2950 struct cgroup *root = &ss->root->cgrp;
2951 struct cgroup_subsys_state *css;
2954 lockdep_assert_held(&cgroup_mutex);
2956 /* add/rm files for all cgroups created before */
2957 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2958 struct cgroup *cgrp = css->cgroup;
2960 if (cgroup_is_dead(cgrp))
2963 ret = cgroup_addrm_files(cgrp, cfts, is_add);
2969 kernfs_activate(root->kn);
2973 static void cgroup_exit_cftypes(struct cftype *cfts)
2977 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2978 /* free copy for custom atomic_write_len, see init_cftypes() */
2979 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
2986 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2990 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2991 struct kernfs_ops *kf_ops;
2993 WARN_ON(cft->ss || cft->kf_ops);
2996 kf_ops = &cgroup_kf_ops;
2998 kf_ops = &cgroup_kf_single_ops;
3001 * Ugh... if @cft wants a custom max_write_len, we need to
3002 * make a copy of kf_ops to set its atomic_write_len.
3004 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3005 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3007 cgroup_exit_cftypes(cfts);
3010 kf_ops->atomic_write_len = cft->max_write_len;
3013 cft->kf_ops = kf_ops;
3020 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3022 lockdep_assert_held(&cgroup_mutex);
3024 if (!cfts || !cfts[0].ss)
3027 list_del(&cfts->node);
3028 cgroup_apply_cftypes(cfts, false);
3029 cgroup_exit_cftypes(cfts);
3034 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3035 * @cfts: zero-length name terminated array of cftypes
3037 * Unregister @cfts. Files described by @cfts are removed from all
3038 * existing cgroups and all future cgroups won't have them either. This
3039 * function can be called anytime whether @cfts' subsys is attached or not.
3041 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3044 int cgroup_rm_cftypes(struct cftype *cfts)
3048 mutex_lock(&cgroup_mutex);
3049 ret = cgroup_rm_cftypes_locked(cfts);
3050 mutex_unlock(&cgroup_mutex);
3055 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3056 * @ss: target cgroup subsystem
3057 * @cfts: zero-length name terminated array of cftypes
3059 * Register @cfts to @ss. Files described by @cfts are created for all
3060 * existing cgroups to which @ss is attached and all future cgroups will
3061 * have them too. This function can be called anytime whether @ss is
3064 * Returns 0 on successful registration, -errno on failure. Note that this
3065 * function currently returns 0 as long as @cfts registration is successful
3066 * even if some file creation attempts on existing cgroups fail.
3068 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3072 if (!cfts || cfts[0].name[0] == '\0')
3075 ret = cgroup_init_cftypes(ss, cfts);
3079 mutex_lock(&cgroup_mutex);
3081 list_add_tail(&cfts->node, &ss->cfts);
3082 ret = cgroup_apply_cftypes(cfts, true);
3084 cgroup_rm_cftypes_locked(cfts);
3086 mutex_unlock(&cgroup_mutex);
3091 * cgroup_task_count - count the number of tasks in a cgroup.
3092 * @cgrp: the cgroup in question
3094 * Return the number of tasks in the cgroup.
3096 static int cgroup_task_count(const struct cgroup *cgrp)
3099 struct cgrp_cset_link *link;
3101 down_read(&css_set_rwsem);
3102 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3103 count += atomic_read(&link->cset->refcount);
3104 up_read(&css_set_rwsem);
3109 * css_next_child - find the next child of a given css
3110 * @pos_css: the current position (%NULL to initiate traversal)
3111 * @parent_css: css whose children to walk
3113 * This function returns the next child of @parent_css and should be called
3114 * under either cgroup_mutex or RCU read lock. The only requirement is
3115 * that @parent_css and @pos_css are accessible. The next sibling is
3116 * guaranteed to be returned regardless of their states.
3118 struct cgroup_subsys_state *
3119 css_next_child(struct cgroup_subsys_state *pos_css,
3120 struct cgroup_subsys_state *parent_css)
3122 struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
3123 struct cgroup *cgrp = parent_css->cgroup;
3124 struct cgroup *next;
3126 cgroup_assert_mutex_or_rcu_locked();
3129 * @pos could already have been removed. Once a cgroup is removed,
3130 * its ->sibling.next is no longer updated when its next sibling
3131 * changes. As CGRP_DEAD assertion is serialized and happens
3132 * before the cgroup is taken off the ->sibling list, if we see it
3133 * unasserted, it's guaranteed that the next sibling hasn't
3134 * finished its grace period even if it's already removed, and thus
3135 * safe to dereference from this RCU critical section. If
3136 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
3137 * to be visible as %true here.
3139 * If @pos is dead, its next pointer can't be dereferenced;
3140 * however, as each cgroup is given a monotonically increasing
3141 * unique serial number and always appended to the sibling list,
3142 * the next one can be found by walking the parent's children until
3143 * we see a cgroup with higher serial number than @pos's. While
3144 * this path can be slower, it's taken only when either the current
3145 * cgroup is removed or iteration and removal race.
3148 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
3149 } else if (likely(!cgroup_is_dead(pos))) {
3150 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
3152 list_for_each_entry_rcu(next, &cgrp->children, sibling)
3153 if (next->serial_nr > pos->serial_nr)
3158 * @next, if not pointing to the head, can be dereferenced and is
3159 * the next sibling; however, it might have @ss disabled. If so,
3160 * fast-forward to the next enabled one.
3162 while (&next->sibling != &cgrp->children) {
3163 struct cgroup_subsys_state *next_css = cgroup_css(next, parent_css->ss);
3167 next = list_entry_rcu(next->sibling.next, struct cgroup, sibling);
3173 * css_next_descendant_pre - find the next descendant for pre-order walk
3174 * @pos: the current position (%NULL to initiate traversal)
3175 * @root: css whose descendants to walk
3177 * To be used by css_for_each_descendant_pre(). Find the next descendant
3178 * to visit for pre-order traversal of @root's descendants. @root is
3179 * included in the iteration and the first node to be visited.
3181 * While this function requires cgroup_mutex or RCU read locking, it
3182 * doesn't require the whole traversal to be contained in a single critical
3183 * section. This function will return the correct next descendant as long
3184 * as both @pos and @root are accessible and @pos is a descendant of @root.
3186 struct cgroup_subsys_state *
3187 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3188 struct cgroup_subsys_state *root)
3190 struct cgroup_subsys_state *next;
3192 cgroup_assert_mutex_or_rcu_locked();
3194 /* if first iteration, visit @root */
3198 /* visit the first child if exists */
3199 next = css_next_child(NULL, pos);
3203 /* no child, visit my or the closest ancestor's next sibling */
3204 while (pos != root) {
3205 next = css_next_child(pos, css_parent(pos));
3208 pos = css_parent(pos);
3215 * css_rightmost_descendant - return the rightmost descendant of a css
3216 * @pos: css of interest
3218 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3219 * is returned. This can be used during pre-order traversal to skip
3222 * While this function requires cgroup_mutex or RCU read locking, it
3223 * doesn't require the whole traversal to be contained in a single critical
3224 * section. This function will return the correct rightmost descendant as
3225 * long as @pos is accessible.
3227 struct cgroup_subsys_state *
3228 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3230 struct cgroup_subsys_state *last, *tmp;
3232 cgroup_assert_mutex_or_rcu_locked();
3236 /* ->prev isn't RCU safe, walk ->next till the end */
3238 css_for_each_child(tmp, last)
3245 static struct cgroup_subsys_state *
3246 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3248 struct cgroup_subsys_state *last;
3252 pos = css_next_child(NULL, pos);
3259 * css_next_descendant_post - find the next descendant for post-order walk
3260 * @pos: the current position (%NULL to initiate traversal)
3261 * @root: css whose descendants to walk
3263 * To be used by css_for_each_descendant_post(). Find the next descendant
3264 * to visit for post-order traversal of @root's descendants. @root is
3265 * included in the iteration and the last node to be visited.
3267 * While this function requires cgroup_mutex or RCU read locking, it
3268 * doesn't require the whole traversal to be contained in a single critical
3269 * section. This function will return the correct next descendant as long
3270 * as both @pos and @cgroup are accessible and @pos is a descendant of
3273 struct cgroup_subsys_state *
3274 css_next_descendant_post(struct cgroup_subsys_state *pos,
3275 struct cgroup_subsys_state *root)
3277 struct cgroup_subsys_state *next;
3279 cgroup_assert_mutex_or_rcu_locked();
3281 /* if first iteration, visit leftmost descendant which may be @root */
3283 return css_leftmost_descendant(root);
3285 /* if we visited @root, we're done */
3289 /* if there's an unvisited sibling, visit its leftmost descendant */
3290 next = css_next_child(pos, css_parent(pos));
3292 return css_leftmost_descendant(next);
3294 /* no sibling left, visit parent */
3295 return css_parent(pos);
3299 * css_advance_task_iter - advance a task itererator to the next css_set
3300 * @it: the iterator to advance
3302 * Advance @it to the next css_set to walk.
3304 static void css_advance_task_iter(struct css_task_iter *it)
3306 struct list_head *l = it->cset_pos;
3307 struct cgrp_cset_link *link;
3308 struct css_set *cset;
3310 /* Advance to the next non-empty css_set */
3313 if (l == it->cset_head) {
3314 it->cset_pos = NULL;
3319 cset = container_of(l, struct css_set,
3320 e_cset_node[it->ss->id]);
3322 link = list_entry(l, struct cgrp_cset_link, cset_link);
3325 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3329 if (!list_empty(&cset->tasks))
3330 it->task_pos = cset->tasks.next;
3332 it->task_pos = cset->mg_tasks.next;
3334 it->tasks_head = &cset->tasks;
3335 it->mg_tasks_head = &cset->mg_tasks;
3339 * css_task_iter_start - initiate task iteration
3340 * @css: the css to walk tasks of
3341 * @it: the task iterator to use
3343 * Initiate iteration through the tasks of @css. The caller can call
3344 * css_task_iter_next() to walk through the tasks until the function
3345 * returns NULL. On completion of iteration, css_task_iter_end() must be
3348 * Note that this function acquires a lock which is released when the
3349 * iteration finishes. The caller can't sleep while iteration is in
3352 void css_task_iter_start(struct cgroup_subsys_state *css,
3353 struct css_task_iter *it)
3354 __acquires(css_set_rwsem)
3356 /* no one should try to iterate before mounting cgroups */
3357 WARN_ON_ONCE(!use_task_css_set_links);
3359 down_read(&css_set_rwsem);
3364 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3366 it->cset_pos = &css->cgroup->cset_links;
3368 it->cset_head = it->cset_pos;
3370 css_advance_task_iter(it);
3374 * css_task_iter_next - return the next task for the iterator
3375 * @it: the task iterator being iterated
3377 * The "next" function for task iteration. @it should have been
3378 * initialized via css_task_iter_start(). Returns NULL when the iteration
3381 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3383 struct task_struct *res;
3384 struct list_head *l = it->task_pos;
3386 /* If the iterator cg is NULL, we have no tasks */
3389 res = list_entry(l, struct task_struct, cg_list);
3392 * Advance iterator to find next entry. cset->tasks is consumed
3393 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3398 if (l == it->tasks_head)
3399 l = it->mg_tasks_head->next;
3401 if (l == it->mg_tasks_head)
3402 css_advance_task_iter(it);
3410 * css_task_iter_end - finish task iteration
3411 * @it: the task iterator to finish
3413 * Finish task iteration started by css_task_iter_start().
3415 void css_task_iter_end(struct css_task_iter *it)
3416 __releases(css_set_rwsem)
3418 up_read(&css_set_rwsem);
3422 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3423 * @to: cgroup to which the tasks will be moved
3424 * @from: cgroup in which the tasks currently reside
3426 * Locking rules between cgroup_post_fork() and the migration path
3427 * guarantee that, if a task is forking while being migrated, the new child
3428 * is guaranteed to be either visible in the source cgroup after the
3429 * parent's migration is complete or put into the target cgroup. No task
3430 * can slip out of migration through forking.
3432 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3434 LIST_HEAD(preloaded_csets);
3435 struct cgrp_cset_link *link;
3436 struct css_task_iter it;
3437 struct task_struct *task;
3440 mutex_lock(&cgroup_mutex);
3442 /* all tasks in @from are being moved, all csets are source */
3443 down_read(&css_set_rwsem);
3444 list_for_each_entry(link, &from->cset_links, cset_link)
3445 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3446 up_read(&css_set_rwsem);
3448 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3453 * Migrate tasks one-by-one until @form is empty. This fails iff
3454 * ->can_attach() fails.
3457 css_task_iter_start(&from->dummy_css, &it);
3458 task = css_task_iter_next(&it);
3460 get_task_struct(task);
3461 css_task_iter_end(&it);
3464 ret = cgroup_migrate(to, task, false);
3465 put_task_struct(task);
3467 } while (task && !ret);
3469 cgroup_migrate_finish(&preloaded_csets);
3470 mutex_unlock(&cgroup_mutex);
3475 * Stuff for reading the 'tasks'/'procs' files.
3477 * Reading this file can return large amounts of data if a cgroup has
3478 * *lots* of attached tasks. So it may need several calls to read(),
3479 * but we cannot guarantee that the information we produce is correct
3480 * unless we produce it entirely atomically.
3484 /* which pidlist file are we talking about? */
3485 enum cgroup_filetype {
3491 * A pidlist is a list of pids that virtually represents the contents of one
3492 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3493 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3496 struct cgroup_pidlist {
3498 * used to find which pidlist is wanted. doesn't change as long as
3499 * this particular list stays in the list.
3501 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3504 /* how many elements the above list has */
3506 /* each of these stored in a list by its cgroup */
3507 struct list_head links;
3508 /* pointer to the cgroup we belong to, for list removal purposes */
3509 struct cgroup *owner;
3510 /* for delayed destruction */
3511 struct delayed_work destroy_dwork;
3515 * The following two functions "fix" the issue where there are more pids
3516 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3517 * TODO: replace with a kernel-wide solution to this problem
3519 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3520 static void *pidlist_allocate(int count)
3522 if (PIDLIST_TOO_LARGE(count))
3523 return vmalloc(count * sizeof(pid_t));
3525 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3528 static void pidlist_free(void *p)
3530 if (is_vmalloc_addr(p))
3537 * Used to destroy all pidlists lingering waiting for destroy timer. None
3538 * should be left afterwards.
3540 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3542 struct cgroup_pidlist *l, *tmp_l;
3544 mutex_lock(&cgrp->pidlist_mutex);
3545 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3546 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3547 mutex_unlock(&cgrp->pidlist_mutex);
3549 flush_workqueue(cgroup_pidlist_destroy_wq);
3550 BUG_ON(!list_empty(&cgrp->pidlists));
3553 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3555 struct delayed_work *dwork = to_delayed_work(work);
3556 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3558 struct cgroup_pidlist *tofree = NULL;
3560 mutex_lock(&l->owner->pidlist_mutex);
3563 * Destroy iff we didn't get queued again. The state won't change
3564 * as destroy_dwork can only be queued while locked.
3566 if (!delayed_work_pending(dwork)) {
3567 list_del(&l->links);
3568 pidlist_free(l->list);
3569 put_pid_ns(l->key.ns);
3573 mutex_unlock(&l->owner->pidlist_mutex);
3578 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3579 * Returns the number of unique elements.
3581 static int pidlist_uniq(pid_t *list, int length)
3586 * we presume the 0th element is unique, so i starts at 1. trivial
3587 * edge cases first; no work needs to be done for either
3589 if (length == 0 || length == 1)
3591 /* src and dest walk down the list; dest counts unique elements */
3592 for (src = 1; src < length; src++) {
3593 /* find next unique element */
3594 while (list[src] == list[src-1]) {
3599 /* dest always points to where the next unique element goes */
3600 list[dest] = list[src];
3608 * The two pid files - task and cgroup.procs - guaranteed that the result
3609 * is sorted, which forced this whole pidlist fiasco. As pid order is
3610 * different per namespace, each namespace needs differently sorted list,
3611 * making it impossible to use, for example, single rbtree of member tasks
3612 * sorted by task pointer. As pidlists can be fairly large, allocating one
3613 * per open file is dangerous, so cgroup had to implement shared pool of
3614 * pidlists keyed by cgroup and namespace.
3616 * All this extra complexity was caused by the original implementation
3617 * committing to an entirely unnecessary property. In the long term, we
3618 * want to do away with it. Explicitly scramble sort order if
3619 * sane_behavior so that no such expectation exists in the new interface.
3621 * Scrambling is done by swapping every two consecutive bits, which is
3622 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3624 static pid_t pid_fry(pid_t pid)
3626 unsigned a = pid & 0x55555555;
3627 unsigned b = pid & 0xAAAAAAAA;
3629 return (a << 1) | (b >> 1);
3632 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3634 if (cgroup_sane_behavior(cgrp))
3635 return pid_fry(pid);
3640 static int cmppid(const void *a, const void *b)
3642 return *(pid_t *)a - *(pid_t *)b;
3645 static int fried_cmppid(const void *a, const void *b)
3647 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3650 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3651 enum cgroup_filetype type)
3653 struct cgroup_pidlist *l;
3654 /* don't need task_nsproxy() if we're looking at ourself */
3655 struct pid_namespace *ns = task_active_pid_ns(current);
3657 lockdep_assert_held(&cgrp->pidlist_mutex);
3659 list_for_each_entry(l, &cgrp->pidlists, links)
3660 if (l->key.type == type && l->key.ns == ns)
3666 * find the appropriate pidlist for our purpose (given procs vs tasks)
3667 * returns with the lock on that pidlist already held, and takes care
3668 * of the use count, or returns NULL with no locks held if we're out of
3671 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3672 enum cgroup_filetype type)
3674 struct cgroup_pidlist *l;
3676 lockdep_assert_held(&cgrp->pidlist_mutex);
3678 l = cgroup_pidlist_find(cgrp, type);
3682 /* entry not found; create a new one */
3683 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3687 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3689 /* don't need task_nsproxy() if we're looking at ourself */
3690 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3692 list_add(&l->links, &cgrp->pidlists);
3697 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3699 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3700 struct cgroup_pidlist **lp)
3704 int pid, n = 0; /* used for populating the array */
3705 struct css_task_iter it;
3706 struct task_struct *tsk;
3707 struct cgroup_pidlist *l;
3709 lockdep_assert_held(&cgrp->pidlist_mutex);
3712 * If cgroup gets more users after we read count, we won't have
3713 * enough space - tough. This race is indistinguishable to the
3714 * caller from the case that the additional cgroup users didn't
3715 * show up until sometime later on.
3717 length = cgroup_task_count(cgrp);
3718 array = pidlist_allocate(length);
3721 /* now, populate the array */
3722 css_task_iter_start(&cgrp->dummy_css, &it);
3723 while ((tsk = css_task_iter_next(&it))) {
3724 if (unlikely(n == length))
3726 /* get tgid or pid for procs or tasks file respectively */
3727 if (type == CGROUP_FILE_PROCS)
3728 pid = task_tgid_vnr(tsk);
3730 pid = task_pid_vnr(tsk);
3731 if (pid > 0) /* make sure to only use valid results */
3734 css_task_iter_end(&it);
3736 /* now sort & (if procs) strip out duplicates */
3737 if (cgroup_sane_behavior(cgrp))
3738 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3740 sort(array, length, sizeof(pid_t), cmppid, NULL);
3741 if (type == CGROUP_FILE_PROCS)
3742 length = pidlist_uniq(array, length);
3744 l = cgroup_pidlist_find_create(cgrp, type);
3746 mutex_unlock(&cgrp->pidlist_mutex);
3747 pidlist_free(array);
3751 /* store array, freeing old if necessary */
3752 pidlist_free(l->list);
3760 * cgroupstats_build - build and fill cgroupstats
3761 * @stats: cgroupstats to fill information into
3762 * @dentry: A dentry entry belonging to the cgroup for which stats have
3765 * Build and fill cgroupstats so that taskstats can export it to user
3768 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3770 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3771 struct cgroup *cgrp;
3772 struct css_task_iter it;
3773 struct task_struct *tsk;
3775 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3776 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3777 kernfs_type(kn) != KERNFS_DIR)
3780 mutex_lock(&cgroup_mutex);
3783 * We aren't being called from kernfs and there's no guarantee on
3784 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3785 * @kn->priv is RCU safe. Let's do the RCU dancing.
3788 cgrp = rcu_dereference(kn->priv);
3789 if (!cgrp || cgroup_is_dead(cgrp)) {
3791 mutex_unlock(&cgroup_mutex);
3796 css_task_iter_start(&cgrp->dummy_css, &it);
3797 while ((tsk = css_task_iter_next(&it))) {
3798 switch (tsk->state) {
3800 stats->nr_running++;
3802 case TASK_INTERRUPTIBLE:
3803 stats->nr_sleeping++;
3805 case TASK_UNINTERRUPTIBLE:
3806 stats->nr_uninterruptible++;
3809 stats->nr_stopped++;
3812 if (delayacct_is_task_waiting_on_io(tsk))
3813 stats->nr_io_wait++;
3817 css_task_iter_end(&it);
3819 mutex_unlock(&cgroup_mutex);
3825 * seq_file methods for the tasks/procs files. The seq_file position is the
3826 * next pid to display; the seq_file iterator is a pointer to the pid
3827 * in the cgroup->l->list array.
3830 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3833 * Initially we receive a position value that corresponds to
3834 * one more than the last pid shown (or 0 on the first call or
3835 * after a seek to the start). Use a binary-search to find the
3836 * next pid to display, if any
3838 struct kernfs_open_file *of = s->private;
3839 struct cgroup *cgrp = seq_css(s)->cgroup;
3840 struct cgroup_pidlist *l;
3841 enum cgroup_filetype type = seq_cft(s)->private;
3842 int index = 0, pid = *pos;
3845 mutex_lock(&cgrp->pidlist_mutex);
3848 * !NULL @of->priv indicates that this isn't the first start()
3849 * after open. If the matching pidlist is around, we can use that.
3850 * Look for it. Note that @of->priv can't be used directly. It
3851 * could already have been destroyed.
3854 of->priv = cgroup_pidlist_find(cgrp, type);
3857 * Either this is the first start() after open or the matching
3858 * pidlist has been destroyed inbetween. Create a new one.
3861 ret = pidlist_array_load(cgrp, type,
3862 (struct cgroup_pidlist **)&of->priv);
3864 return ERR_PTR(ret);
3869 int end = l->length;
3871 while (index < end) {
3872 int mid = (index + end) / 2;
3873 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3876 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3882 /* If we're off the end of the array, we're done */
3883 if (index >= l->length)
3885 /* Update the abstract position to be the actual pid that we found */
3886 iter = l->list + index;
3887 *pos = cgroup_pid_fry(cgrp, *iter);
3891 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3893 struct kernfs_open_file *of = s->private;
3894 struct cgroup_pidlist *l = of->priv;
3897 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3898 CGROUP_PIDLIST_DESTROY_DELAY);
3899 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3902 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3904 struct kernfs_open_file *of = s->private;
3905 struct cgroup_pidlist *l = of->priv;
3907 pid_t *end = l->list + l->length;
3909 * Advance to the next pid in the array. If this goes off the
3916 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3921 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3923 return seq_printf(s, "%d\n", *(int *)v);
3926 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3929 return notify_on_release(css->cgroup);
3932 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3933 struct cftype *cft, u64 val)
3935 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3937 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3939 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3943 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3946 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3949 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3950 struct cftype *cft, u64 val)
3953 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3955 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3959 static struct cftype cgroup_base_files[] = {
3961 .name = "cgroup.procs",
3962 .seq_start = cgroup_pidlist_start,
3963 .seq_next = cgroup_pidlist_next,
3964 .seq_stop = cgroup_pidlist_stop,
3965 .seq_show = cgroup_pidlist_show,
3966 .private = CGROUP_FILE_PROCS,
3967 .write = cgroup_procs_write,
3968 .mode = S_IRUGO | S_IWUSR,
3971 .name = "cgroup.clone_children",
3972 .flags = CFTYPE_INSANE,
3973 .read_u64 = cgroup_clone_children_read,
3974 .write_u64 = cgroup_clone_children_write,
3977 .name = "cgroup.sane_behavior",
3978 .flags = CFTYPE_ONLY_ON_ROOT,
3979 .seq_show = cgroup_sane_behavior_show,
3982 .name = "cgroup.controllers",
3983 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_ONLY_ON_ROOT,
3984 .seq_show = cgroup_root_controllers_show,
3987 .name = "cgroup.controllers",
3988 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
3989 .seq_show = cgroup_controllers_show,
3992 .name = "cgroup.subtree_control",
3993 .flags = CFTYPE_ONLY_ON_DFL,
3994 .seq_show = cgroup_subtree_control_show,
3995 .write = cgroup_subtree_control_write,
3998 .name = "cgroup.populated",
3999 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4000 .seq_show = cgroup_populated_show,
4004 * Historical crazy stuff. These don't have "cgroup." prefix and
4005 * don't exist if sane_behavior. If you're depending on these, be
4006 * prepared to be burned.
4010 .flags = CFTYPE_INSANE, /* use "procs" instead */
4011 .seq_start = cgroup_pidlist_start,
4012 .seq_next = cgroup_pidlist_next,
4013 .seq_stop = cgroup_pidlist_stop,
4014 .seq_show = cgroup_pidlist_show,
4015 .private = CGROUP_FILE_TASKS,
4016 .write = cgroup_tasks_write,
4017 .mode = S_IRUGO | S_IWUSR,
4020 .name = "notify_on_release",
4021 .flags = CFTYPE_INSANE,
4022 .read_u64 = cgroup_read_notify_on_release,
4023 .write_u64 = cgroup_write_notify_on_release,
4026 .name = "release_agent",
4027 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4028 .seq_show = cgroup_release_agent_show,
4029 .write = cgroup_release_agent_write,
4030 .max_write_len = PATH_MAX - 1,
4036 * cgroup_populate_dir - create subsys files in a cgroup directory
4037 * @cgrp: target cgroup
4038 * @subsys_mask: mask of the subsystem ids whose files should be added
4040 * On failure, no file is added.
4042 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4044 struct cgroup_subsys *ss;
4047 /* process cftsets of each subsystem */
4048 for_each_subsys(ss, i) {
4049 struct cftype *cfts;
4051 if (!(subsys_mask & (1 << i)))
4054 list_for_each_entry(cfts, &ss->cfts, node) {
4055 ret = cgroup_addrm_files(cgrp, cfts, true);
4062 cgroup_clear_dir(cgrp, subsys_mask);
4067 * css destruction is four-stage process.
4069 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4070 * Implemented in kill_css().
4072 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4073 * and thus css_tryget_online() is guaranteed to fail, the css can be
4074 * offlined by invoking offline_css(). After offlining, the base ref is
4075 * put. Implemented in css_killed_work_fn().
4077 * 3. When the percpu_ref reaches zero, the only possible remaining
4078 * accessors are inside RCU read sections. css_release() schedules the
4081 * 4. After the grace period, the css can be freed. Implemented in
4082 * css_free_work_fn().
4084 * It is actually hairier because both step 2 and 4 require process context
4085 * and thus involve punting to css->destroy_work adding two additional
4086 * steps to the already complex sequence.
4088 static void css_free_work_fn(struct work_struct *work)
4090 struct cgroup_subsys_state *css =
4091 container_of(work, struct cgroup_subsys_state, destroy_work);
4092 struct cgroup *cgrp = css->cgroup;
4095 css_put(css->parent);
4097 css->ss->css_free(css);
4101 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4103 struct cgroup_subsys_state *css =
4104 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4106 INIT_WORK(&css->destroy_work, css_free_work_fn);
4107 queue_work(cgroup_destroy_wq, &css->destroy_work);
4110 static void css_release(struct percpu_ref *ref)
4112 struct cgroup_subsys_state *css =
4113 container_of(ref, struct cgroup_subsys_state, refcnt);
4114 struct cgroup_subsys *ss = css->ss;
4116 cgroup_idr_remove(&ss->css_idr, css->id);
4118 call_rcu(&css->rcu_head, css_free_rcu_fn);
4121 static void init_and_link_css(struct cgroup_subsys_state *css,
4122 struct cgroup_subsys *ss, struct cgroup *cgrp)
4131 css->parent = cgroup_css(cgrp->parent, ss);
4132 css_get(css->parent);
4134 css->flags |= CSS_ROOT;
4137 BUG_ON(cgroup_css(cgrp, ss));
4140 /* invoke ->css_online() on a new CSS and mark it online if successful */
4141 static int online_css(struct cgroup_subsys_state *css)
4143 struct cgroup_subsys *ss = css->ss;
4146 lockdep_assert_held(&cgroup_mutex);
4149 ret = ss->css_online(css);
4151 css->flags |= CSS_ONLINE;
4152 css->cgroup->nr_css++;
4153 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4158 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4159 static void offline_css(struct cgroup_subsys_state *css)
4161 struct cgroup_subsys *ss = css->ss;
4163 lockdep_assert_held(&cgroup_mutex);
4165 if (!(css->flags & CSS_ONLINE))
4168 if (ss->css_offline)
4169 ss->css_offline(css);
4171 css->flags &= ~CSS_ONLINE;
4172 css->cgroup->nr_css--;
4173 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4175 wake_up_all(&css->cgroup->offline_waitq);
4179 * create_css - create a cgroup_subsys_state
4180 * @cgrp: the cgroup new css will be associated with
4181 * @ss: the subsys of new css
4183 * Create a new css associated with @cgrp - @ss pair. On success, the new
4184 * css is online and installed in @cgrp with all interface files created.
4185 * Returns 0 on success, -errno on failure.
4187 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4189 struct cgroup *parent = cgrp->parent;
4190 struct cgroup_subsys_state *css;
4193 lockdep_assert_held(&cgroup_mutex);
4195 css = ss->css_alloc(cgroup_css(parent, ss));
4197 return PTR_ERR(css);
4199 init_and_link_css(css, ss, cgrp);
4201 err = percpu_ref_init(&css->refcnt, css_release);
4205 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4207 goto err_free_percpu_ref;
4210 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4214 /* @css is ready to be brought online now, make it visible */
4215 cgroup_idr_replace(&ss->css_idr, css, css->id);
4217 err = online_css(css);
4221 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4223 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4224 current->comm, current->pid, ss->name);
4225 if (!strcmp(ss->name, "memory"))
4226 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4227 ss->warned_broken_hierarchy = true;
4233 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4235 cgroup_idr_remove(&ss->css_idr, css->id);
4236 err_free_percpu_ref:
4237 percpu_ref_cancel_init(&css->refcnt);
4239 call_rcu(&css->rcu_head, css_free_rcu_fn);
4243 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4246 struct cgroup *parent, *cgrp;
4247 struct cgroup_root *root;
4248 struct cgroup_subsys *ss;
4249 struct kernfs_node *kn;
4252 parent = cgroup_kn_lock_live(parent_kn);
4255 root = parent->root;
4257 /* allocate the cgroup and its ID, 0 is reserved for the root */
4258 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4265 * Temporarily set the pointer to NULL, so idr_find() won't return
4266 * a half-baked cgroup.
4268 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4274 init_cgroup_housekeeping(cgrp);
4276 cgrp->parent = parent;
4277 cgrp->dummy_css.parent = &parent->dummy_css;
4280 if (notify_on_release(parent))
4281 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4283 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4284 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4286 /* create the directory */
4287 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4295 * This extra ref will be put in cgroup_free_fn() and guarantees
4296 * that @cgrp->kn is always accessible.
4300 cgrp->serial_nr = cgroup_serial_nr_next++;
4302 /* allocation complete, commit to creation */
4303 list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4304 atomic_inc(&root->nr_cgrps);
4308 * @cgrp is now fully operational. If something fails after this
4309 * point, it'll be released via the normal destruction path.
4311 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4313 ret = cgroup_kn_set_ugid(kn);
4317 ret = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4321 /* let's create and online css's */
4322 for_each_subsys(ss, ssid) {
4323 if (parent->child_subsys_mask & (1 << ssid)) {
4324 ret = create_css(cgrp, ss);
4331 * On the default hierarchy, a child doesn't automatically inherit
4332 * child_subsys_mask from the parent. Each is configured manually.
4334 if (!cgroup_on_dfl(cgrp))
4335 cgrp->child_subsys_mask = parent->child_subsys_mask;
4337 kernfs_activate(kn);
4343 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4347 cgroup_kn_unlock(parent_kn);
4351 cgroup_destroy_locked(cgrp);
4356 * This is called when the refcnt of a css is confirmed to be killed.
4357 * css_tryget_online() is now guaranteed to fail.
4359 static void css_killed_work_fn(struct work_struct *work)
4361 struct cgroup_subsys_state *css =
4362 container_of(work, struct cgroup_subsys_state, destroy_work);
4363 struct cgroup *cgrp = css->cgroup;
4365 mutex_lock(&cgroup_mutex);
4368 * css_tryget_online() is guaranteed to fail now. Tell subsystems
4369 * to initate destruction.
4374 * If @cgrp is marked dead, it's waiting for refs of all css's to
4375 * be disabled before proceeding to the second phase of cgroup
4376 * destruction. If we are the last one, kick it off.
4378 if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4379 cgroup_destroy_css_killed(cgrp);
4381 mutex_unlock(&cgroup_mutex);
4384 * Put the css refs from kill_css(). Each css holds an extra
4385 * reference to the cgroup's dentry and cgroup removal proceeds
4386 * regardless of css refs. On the last put of each css, whenever
4387 * that may be, the extra dentry ref is put so that dentry
4388 * destruction happens only after all css's are released.
4393 /* css kill confirmation processing requires process context, bounce */
4394 static void css_killed_ref_fn(struct percpu_ref *ref)
4396 struct cgroup_subsys_state *css =
4397 container_of(ref, struct cgroup_subsys_state, refcnt);
4399 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4400 queue_work(cgroup_destroy_wq, &css->destroy_work);
4404 * kill_css - destroy a css
4405 * @css: css to destroy
4407 * This function initiates destruction of @css by removing cgroup interface
4408 * files and putting its base reference. ->css_offline() will be invoked
4409 * asynchronously once css_tryget_online() is guaranteed to fail and when
4410 * the reference count reaches zero, @css will be released.
4412 static void kill_css(struct cgroup_subsys_state *css)
4414 lockdep_assert_held(&cgroup_mutex);
4417 * This must happen before css is disassociated with its cgroup.
4418 * See seq_css() for details.
4420 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4423 * Killing would put the base ref, but we need to keep it alive
4424 * until after ->css_offline().
4429 * cgroup core guarantees that, by the time ->css_offline() is
4430 * invoked, no new css reference will be given out via
4431 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4432 * proceed to offlining css's because percpu_ref_kill() doesn't
4433 * guarantee that the ref is seen as killed on all CPUs on return.
4435 * Use percpu_ref_kill_and_confirm() to get notifications as each
4436 * css is confirmed to be seen as killed on all CPUs.
4438 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4442 * cgroup_destroy_locked - the first stage of cgroup destruction
4443 * @cgrp: cgroup to be destroyed
4445 * css's make use of percpu refcnts whose killing latency shouldn't be
4446 * exposed to userland and are RCU protected. Also, cgroup core needs to
4447 * guarantee that css_tryget_online() won't succeed by the time
4448 * ->css_offline() is invoked. To satisfy all the requirements,
4449 * destruction is implemented in the following two steps.
4451 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4452 * userland visible parts and start killing the percpu refcnts of
4453 * css's. Set up so that the next stage will be kicked off once all
4454 * the percpu refcnts are confirmed to be killed.
4456 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4457 * rest of destruction. Once all cgroup references are gone, the
4458 * cgroup is RCU-freed.
4460 * This function implements s1. After this step, @cgrp is gone as far as
4461 * the userland is concerned and a new cgroup with the same name may be
4462 * created. As cgroup doesn't care about the names internally, this
4463 * doesn't cause any problem.
4465 static int cgroup_destroy_locked(struct cgroup *cgrp)
4466 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4468 struct cgroup *child;
4469 struct cgroup_subsys_state *css;
4473 lockdep_assert_held(&cgroup_mutex);
4476 * css_set_rwsem synchronizes access to ->cset_links and prevents
4477 * @cgrp from being removed while put_css_set() is in progress.
4479 down_read(&css_set_rwsem);
4480 empty = list_empty(&cgrp->cset_links);
4481 up_read(&css_set_rwsem);
4486 * Make sure there's no live children. We can't test ->children
4487 * emptiness as dead children linger on it while being destroyed;
4488 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4492 list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4493 empty = cgroup_is_dead(child);
4502 * Mark @cgrp dead. This prevents further task migration and child
4503 * creation by disabling cgroup_lock_live_group(). Note that
4504 * CGRP_DEAD assertion is depended upon by css_next_child() to
4505 * resume iteration after dropping RCU read lock. See
4506 * css_next_child() for details.
4508 set_bit(CGRP_DEAD, &cgrp->flags);
4511 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4512 * will be invoked to perform the rest of destruction once the
4513 * percpu refs of all css's are confirmed to be killed.
4515 for_each_css(css, ssid, cgrp)
4518 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4519 raw_spin_lock(&release_list_lock);
4520 if (!list_empty(&cgrp->release_list))
4521 list_del_init(&cgrp->release_list);
4522 raw_spin_unlock(&release_list_lock);
4525 * If @cgrp has css's attached, the second stage of cgroup
4526 * destruction is kicked off from css_killed_work_fn() after the
4527 * refs of all attached css's are killed. If @cgrp doesn't have
4528 * any css, we kick it off here.
4531 cgroup_destroy_css_killed(cgrp);
4534 * Remove @cgrp directory along with the base files. @cgrp has an
4535 * extra ref on its kn.
4537 kernfs_remove(cgrp->kn);
4543 * cgroup_destroy_css_killed - the second step of cgroup destruction
4544 * @cgrp: the cgroup whose csses have just finished offlining
4546 * This function is invoked from a work item for a cgroup which is being
4547 * destroyed after all css's are offlined and performs the rest of
4548 * destruction. This is the second step of destruction described in the
4549 * comment above cgroup_destroy_locked().
4551 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4553 struct cgroup *parent = cgrp->parent;
4555 lockdep_assert_held(&cgroup_mutex);
4557 /* delete this cgroup from parent->children */
4558 list_del_rcu(&cgrp->sibling);
4562 set_bit(CGRP_RELEASABLE, &parent->flags);
4563 check_for_release(parent);
4566 static int cgroup_rmdir(struct kernfs_node *kn)
4568 struct cgroup *cgrp;
4571 cgrp = cgroup_kn_lock_live(kn);
4574 cgroup_get(cgrp); /* for @kn->priv clearing */
4576 ret = cgroup_destroy_locked(cgrp);
4578 cgroup_kn_unlock(kn);
4581 * There are two control paths which try to determine cgroup from
4582 * dentry without going through kernfs - cgroupstats_build() and
4583 * css_tryget_online_from_dir(). Those are supported by RCU
4584 * protecting clearing of cgrp->kn->priv backpointer, which should
4585 * happen after all files under it have been removed.
4588 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4594 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4595 .remount_fs = cgroup_remount,
4596 .show_options = cgroup_show_options,
4597 .mkdir = cgroup_mkdir,
4598 .rmdir = cgroup_rmdir,
4599 .rename = cgroup_rename,
4602 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4604 struct cgroup_subsys_state *css;
4606 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4608 mutex_lock(&cgroup_mutex);
4610 idr_init(&ss->css_idr);
4611 INIT_LIST_HEAD(&ss->cfts);
4613 /* Create the root cgroup state for this subsystem */
4614 ss->root = &cgrp_dfl_root;
4615 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4616 /* We don't handle early failures gracefully */
4617 BUG_ON(IS_ERR(css));
4618 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4620 /* idr_alloc() can't be called safely during early init */
4623 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4624 BUG_ON(css->id < 0);
4627 /* Update the init_css_set to contain a subsys
4628 * pointer to this state - since the subsystem is
4629 * newly registered, all tasks and hence the
4630 * init_css_set is in the subsystem's root cgroup. */
4631 init_css_set.subsys[ss->id] = css;
4633 need_forkexit_callback |= ss->fork || ss->exit;
4635 /* At system boot, before all subsystems have been
4636 * registered, no tasks have been forked, so we don't
4637 * need to invoke fork callbacks here. */
4638 BUG_ON(!list_empty(&init_task.tasks));
4640 BUG_ON(online_css(css));
4642 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4644 mutex_unlock(&cgroup_mutex);
4648 * cgroup_init_early - cgroup initialization at system boot
4650 * Initialize cgroups at system boot, and initialize any
4651 * subsystems that request early init.
4653 int __init cgroup_init_early(void)
4655 static struct cgroup_sb_opts __initdata opts =
4656 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4657 struct cgroup_subsys *ss;
4660 init_cgroup_root(&cgrp_dfl_root, &opts);
4661 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4663 for_each_subsys(ss, i) {
4664 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4665 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4666 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4668 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4669 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4672 ss->name = cgroup_subsys_name[i];
4675 cgroup_init_subsys(ss, true);
4681 * cgroup_init - cgroup initialization
4683 * Register cgroup filesystem and /proc file, and initialize
4684 * any subsystems that didn't request early init.
4686 int __init cgroup_init(void)
4688 struct cgroup_subsys *ss;
4692 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4694 mutex_lock(&cgroup_mutex);
4696 /* Add init_css_set to the hash table */
4697 key = css_set_hash(init_css_set.subsys);
4698 hash_add(css_set_table, &init_css_set.hlist, key);
4700 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4702 mutex_unlock(&cgroup_mutex);
4704 for_each_subsys(ss, ssid) {
4705 if (ss->early_init) {
4706 struct cgroup_subsys_state *css =
4707 init_css_set.subsys[ss->id];
4709 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4711 BUG_ON(css->id < 0);
4713 cgroup_init_subsys(ss, false);
4716 list_add_tail(&init_css_set.e_cset_node[ssid],
4717 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4720 * cftype registration needs kmalloc and can't be done
4721 * during early_init. Register base cftypes separately.
4723 if (ss->base_cftypes)
4724 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4727 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4731 err = register_filesystem(&cgroup_fs_type);
4733 kobject_put(cgroup_kobj);
4737 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4741 static int __init cgroup_wq_init(void)
4744 * There isn't much point in executing destruction path in
4745 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4746 * Use 1 for @max_active.
4748 * We would prefer to do this in cgroup_init() above, but that
4749 * is called before init_workqueues(): so leave this until after.
4751 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4752 BUG_ON(!cgroup_destroy_wq);
4755 * Used to destroy pidlists and separate to serve as flush domain.
4756 * Cap @max_active to 1 too.
4758 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4760 BUG_ON(!cgroup_pidlist_destroy_wq);
4764 core_initcall(cgroup_wq_init);
4767 * proc_cgroup_show()
4768 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4769 * - Used for /proc/<pid>/cgroup.
4772 /* TODO: Use a proper seq_file iterator */
4773 int proc_cgroup_show(struct seq_file *m, void *v)
4776 struct task_struct *tsk;
4779 struct cgroup_root *root;
4782 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4788 tsk = get_pid_task(pid, PIDTYPE_PID);
4794 mutex_lock(&cgroup_mutex);
4795 down_read(&css_set_rwsem);
4797 for_each_root(root) {
4798 struct cgroup_subsys *ss;
4799 struct cgroup *cgrp;
4800 int ssid, count = 0;
4802 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4805 seq_printf(m, "%d:", root->hierarchy_id);
4806 for_each_subsys(ss, ssid)
4807 if (root->subsys_mask & (1 << ssid))
4808 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4809 if (strlen(root->name))
4810 seq_printf(m, "%sname=%s", count ? "," : "",
4813 cgrp = task_cgroup_from_root(tsk, root);
4814 path = cgroup_path(cgrp, buf, PATH_MAX);
4816 retval = -ENAMETOOLONG;
4824 up_read(&css_set_rwsem);
4825 mutex_unlock(&cgroup_mutex);
4826 put_task_struct(tsk);
4833 /* Display information about each subsystem and each hierarchy */
4834 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4836 struct cgroup_subsys *ss;
4839 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4841 * ideally we don't want subsystems moving around while we do this.
4842 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4843 * subsys/hierarchy state.
4845 mutex_lock(&cgroup_mutex);
4847 for_each_subsys(ss, i)
4848 seq_printf(m, "%s\t%d\t%d\t%d\n",
4849 ss->name, ss->root->hierarchy_id,
4850 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4852 mutex_unlock(&cgroup_mutex);
4856 static int cgroupstats_open(struct inode *inode, struct file *file)
4858 return single_open(file, proc_cgroupstats_show, NULL);
4861 static const struct file_operations proc_cgroupstats_operations = {
4862 .open = cgroupstats_open,
4864 .llseek = seq_lseek,
4865 .release = single_release,
4869 * cgroup_fork - initialize cgroup related fields during copy_process()
4870 * @child: pointer to task_struct of forking parent process.
4872 * A task is associated with the init_css_set until cgroup_post_fork()
4873 * attaches it to the parent's css_set. Empty cg_list indicates that
4874 * @child isn't holding reference to its css_set.
4876 void cgroup_fork(struct task_struct *child)
4878 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4879 INIT_LIST_HEAD(&child->cg_list);
4883 * cgroup_post_fork - called on a new task after adding it to the task list
4884 * @child: the task in question
4886 * Adds the task to the list running through its css_set if necessary and
4887 * call the subsystem fork() callbacks. Has to be after the task is
4888 * visible on the task list in case we race with the first call to
4889 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4892 void cgroup_post_fork(struct task_struct *child)
4894 struct cgroup_subsys *ss;
4898 * This may race against cgroup_enable_task_cg_links(). As that
4899 * function sets use_task_css_set_links before grabbing
4900 * tasklist_lock and we just went through tasklist_lock to add
4901 * @child, it's guaranteed that either we see the set
4902 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4903 * @child during its iteration.
4905 * If we won the race, @child is associated with %current's
4906 * css_set. Grabbing css_set_rwsem guarantees both that the
4907 * association is stable, and, on completion of the parent's
4908 * migration, @child is visible in the source of migration or
4909 * already in the destination cgroup. This guarantee is necessary
4910 * when implementing operations which need to migrate all tasks of
4911 * a cgroup to another.
4913 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4914 * will remain in init_css_set. This is safe because all tasks are
4915 * in the init_css_set before cg_links is enabled and there's no
4916 * operation which transfers all tasks out of init_css_set.
4918 if (use_task_css_set_links) {
4919 struct css_set *cset;
4921 down_write(&css_set_rwsem);
4922 cset = task_css_set(current);
4923 if (list_empty(&child->cg_list)) {
4924 rcu_assign_pointer(child->cgroups, cset);
4925 list_add(&child->cg_list, &cset->tasks);
4928 up_write(&css_set_rwsem);
4932 * Call ss->fork(). This must happen after @child is linked on
4933 * css_set; otherwise, @child might change state between ->fork()
4934 * and addition to css_set.
4936 if (need_forkexit_callback) {
4937 for_each_subsys(ss, i)
4944 * cgroup_exit - detach cgroup from exiting task
4945 * @tsk: pointer to task_struct of exiting process
4947 * Description: Detach cgroup from @tsk and release it.
4949 * Note that cgroups marked notify_on_release force every task in
4950 * them to take the global cgroup_mutex mutex when exiting.
4951 * This could impact scaling on very large systems. Be reluctant to
4952 * use notify_on_release cgroups where very high task exit scaling
4953 * is required on large systems.
4955 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
4956 * call cgroup_exit() while the task is still competent to handle
4957 * notify_on_release(), then leave the task attached to the root cgroup in
4958 * each hierarchy for the remainder of its exit. No need to bother with
4959 * init_css_set refcnting. init_css_set never goes away and we can't race
4960 * with migration path - PF_EXITING is visible to migration path.
4962 void cgroup_exit(struct task_struct *tsk)
4964 struct cgroup_subsys *ss;
4965 struct css_set *cset;
4966 bool put_cset = false;
4970 * Unlink from @tsk from its css_set. As migration path can't race
4971 * with us, we can check cg_list without grabbing css_set_rwsem.
4973 if (!list_empty(&tsk->cg_list)) {
4974 down_write(&css_set_rwsem);
4975 list_del_init(&tsk->cg_list);
4976 up_write(&css_set_rwsem);
4980 /* Reassign the task to the init_css_set. */
4981 cset = task_css_set(tsk);
4982 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
4984 if (need_forkexit_callback) {
4985 /* see cgroup_post_fork() for details */
4986 for_each_subsys(ss, i) {
4988 struct cgroup_subsys_state *old_css = cset->subsys[i];
4989 struct cgroup_subsys_state *css = task_css(tsk, i);
4991 ss->exit(css, old_css, tsk);
4997 put_css_set(cset, true);
5000 static void check_for_release(struct cgroup *cgrp)
5002 if (cgroup_is_releasable(cgrp) &&
5003 list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5005 * Control Group is currently removeable. If it's not
5006 * already queued for a userspace notification, queue
5009 int need_schedule_work = 0;
5011 raw_spin_lock(&release_list_lock);
5012 if (!cgroup_is_dead(cgrp) &&
5013 list_empty(&cgrp->release_list)) {
5014 list_add(&cgrp->release_list, &release_list);
5015 need_schedule_work = 1;
5017 raw_spin_unlock(&release_list_lock);
5018 if (need_schedule_work)
5019 schedule_work(&release_agent_work);
5024 * Notify userspace when a cgroup is released, by running the
5025 * configured release agent with the name of the cgroup (path
5026 * relative to the root of cgroup file system) as the argument.
5028 * Most likely, this user command will try to rmdir this cgroup.
5030 * This races with the possibility that some other task will be
5031 * attached to this cgroup before it is removed, or that some other
5032 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5033 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5034 * unused, and this cgroup will be reprieved from its death sentence,
5035 * to continue to serve a useful existence. Next time it's released,
5036 * we will get notified again, if it still has 'notify_on_release' set.
5038 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5039 * means only wait until the task is successfully execve()'d. The
5040 * separate release agent task is forked by call_usermodehelper(),
5041 * then control in this thread returns here, without waiting for the
5042 * release agent task. We don't bother to wait because the caller of
5043 * this routine has no use for the exit status of the release agent
5044 * task, so no sense holding our caller up for that.
5046 static void cgroup_release_agent(struct work_struct *work)
5048 BUG_ON(work != &release_agent_work);
5049 mutex_lock(&cgroup_mutex);
5050 raw_spin_lock(&release_list_lock);
5051 while (!list_empty(&release_list)) {
5052 char *argv[3], *envp[3];
5054 char *pathbuf = NULL, *agentbuf = NULL, *path;
5055 struct cgroup *cgrp = list_entry(release_list.next,
5058 list_del_init(&cgrp->release_list);
5059 raw_spin_unlock(&release_list_lock);
5060 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5063 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5066 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5071 argv[i++] = agentbuf;
5076 /* minimal command environment */
5077 envp[i++] = "HOME=/";
5078 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5081 /* Drop the lock while we invoke the usermode helper,
5082 * since the exec could involve hitting disk and hence
5083 * be a slow process */
5084 mutex_unlock(&cgroup_mutex);
5085 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5086 mutex_lock(&cgroup_mutex);
5090 raw_spin_lock(&release_list_lock);
5092 raw_spin_unlock(&release_list_lock);
5093 mutex_unlock(&cgroup_mutex);
5096 static int __init cgroup_disable(char *str)
5098 struct cgroup_subsys *ss;
5102 while ((token = strsep(&str, ",")) != NULL) {
5106 for_each_subsys(ss, i) {
5107 if (!strcmp(token, ss->name)) {
5109 printk(KERN_INFO "Disabling %s control group"
5110 " subsystem\n", ss->name);
5117 __setup("cgroup_disable=", cgroup_disable);
5120 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5121 * @dentry: directory dentry of interest
5122 * @ss: subsystem of interest
5124 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5125 * to get the corresponding css and return it. If such css doesn't exist
5126 * or can't be pinned, an ERR_PTR value is returned.
5128 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5129 struct cgroup_subsys *ss)
5131 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5132 struct cgroup_subsys_state *css = NULL;
5133 struct cgroup *cgrp;
5135 /* is @dentry a cgroup dir? */
5136 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5137 kernfs_type(kn) != KERNFS_DIR)
5138 return ERR_PTR(-EBADF);
5143 * This path doesn't originate from kernfs and @kn could already
5144 * have been or be removed at any point. @kn->priv is RCU
5145 * protected for this access. See cgroup_rmdir() for details.
5147 cgrp = rcu_dereference(kn->priv);
5149 css = cgroup_css(cgrp, ss);
5151 if (!css || !css_tryget_online(css))
5152 css = ERR_PTR(-ENOENT);
5159 * css_from_id - lookup css by id
5160 * @id: the cgroup id
5161 * @ss: cgroup subsys to be looked into
5163 * Returns the css if there's valid one with @id, otherwise returns NULL.
5164 * Should be called under rcu_read_lock().
5166 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5168 WARN_ON_ONCE(!rcu_read_lock_held());
5169 return idr_find(&ss->css_idr, id);
5172 #ifdef CONFIG_CGROUP_DEBUG
5173 static struct cgroup_subsys_state *
5174 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5176 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5179 return ERR_PTR(-ENOMEM);
5184 static void debug_css_free(struct cgroup_subsys_state *css)
5189 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5192 return cgroup_task_count(css->cgroup);
5195 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5198 return (u64)(unsigned long)current->cgroups;
5201 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5207 count = atomic_read(&task_css_set(current)->refcount);
5212 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5214 struct cgrp_cset_link *link;
5215 struct css_set *cset;
5218 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5222 down_read(&css_set_rwsem);
5224 cset = rcu_dereference(current->cgroups);
5225 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5226 struct cgroup *c = link->cgrp;
5228 cgroup_name(c, name_buf, NAME_MAX + 1);
5229 seq_printf(seq, "Root %d group %s\n",
5230 c->root->hierarchy_id, name_buf);
5233 up_read(&css_set_rwsem);
5238 #define MAX_TASKS_SHOWN_PER_CSS 25
5239 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5241 struct cgroup_subsys_state *css = seq_css(seq);
5242 struct cgrp_cset_link *link;
5244 down_read(&css_set_rwsem);
5245 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5246 struct css_set *cset = link->cset;
5247 struct task_struct *task;
5250 seq_printf(seq, "css_set %p\n", cset);
5252 list_for_each_entry(task, &cset->tasks, cg_list) {
5253 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5255 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5258 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5259 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5261 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5265 seq_puts(seq, " ...\n");
5267 up_read(&css_set_rwsem);
5271 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5273 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5276 static struct cftype debug_files[] = {
5278 .name = "taskcount",
5279 .read_u64 = debug_taskcount_read,
5283 .name = "current_css_set",
5284 .read_u64 = current_css_set_read,
5288 .name = "current_css_set_refcount",
5289 .read_u64 = current_css_set_refcount_read,
5293 .name = "current_css_set_cg_links",
5294 .seq_show = current_css_set_cg_links_read,
5298 .name = "cgroup_css_links",
5299 .seq_show = cgroup_css_links_read,
5303 .name = "releasable",
5304 .read_u64 = releasable_read,
5310 struct cgroup_subsys debug_cgrp_subsys = {
5311 .css_alloc = debug_css_alloc,
5312 .css_free = debug_css_free,
5313 .base_cftypes = debug_files,
5315 #endif /* CONFIG_CGROUP_DEBUG */