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 /* some controllers are not supported in the default hierarchy */
152 static const unsigned int cgrp_dfl_root_inhibit_ss_mask = 0
153 #ifdef CONFIG_CGROUP_DEBUG
154 | (1 << debug_cgrp_id)
158 /* The list of hierarchy roots */
160 static LIST_HEAD(cgroup_roots);
161 static int cgroup_root_count;
163 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
164 static DEFINE_IDR(cgroup_hierarchy_idr);
167 * Assign a monotonically increasing serial number to csses. It guarantees
168 * cgroups with bigger numbers are newer than those with smaller numbers.
169 * Also, as csses are always appended to the parent's ->children list, it
170 * guarantees that sibling csses are always sorted in the ascending serial
171 * number order on the list. Protected by cgroup_mutex.
173 static u64 css_serial_nr_next = 1;
175 /* This flag indicates whether tasks in the fork and exit paths should
176 * check for fork/exit handlers to call. This avoids us having to do
177 * extra work in the fork/exit path if none of the subsystems need to
180 static int need_forkexit_callback __read_mostly;
182 static struct cftype cgroup_base_files[];
184 static void cgroup_put(struct cgroup *cgrp);
185 static int rebind_subsystems(struct cgroup_root *dst_root,
186 unsigned int ss_mask);
187 static int cgroup_destroy_locked(struct cgroup *cgrp);
188 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss);
189 static void css_release(struct percpu_ref *ref);
190 static void kill_css(struct cgroup_subsys_state *css);
191 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
193 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
195 /* IDR wrappers which synchronize using cgroup_idr_lock */
196 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
201 idr_preload(gfp_mask);
202 spin_lock_bh(&cgroup_idr_lock);
203 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
204 spin_unlock_bh(&cgroup_idr_lock);
209 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
213 spin_lock_bh(&cgroup_idr_lock);
214 ret = idr_replace(idr, ptr, id);
215 spin_unlock_bh(&cgroup_idr_lock);
219 static void cgroup_idr_remove(struct idr *idr, int id)
221 spin_lock_bh(&cgroup_idr_lock);
223 spin_unlock_bh(&cgroup_idr_lock);
226 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
228 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
231 return container_of(parent_css, struct cgroup, self);
236 * cgroup_css - obtain a cgroup's css for the specified subsystem
237 * @cgrp: the cgroup of interest
238 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
240 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
241 * function must be called either under cgroup_mutex or rcu_read_lock() and
242 * the caller is responsible for pinning the returned css if it wants to
243 * keep accessing it outside the said locks. This function may return
244 * %NULL if @cgrp doesn't have @subsys_id enabled.
246 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
247 struct cgroup_subsys *ss)
250 return rcu_dereference_check(cgrp->subsys[ss->id],
251 lockdep_is_held(&cgroup_mutex));
257 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
258 * @cgrp: the cgroup of interest
259 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
261 * Similar to cgroup_css() but returns the effctive css, which is defined
262 * as the matching css of the nearest ancestor including self which has @ss
263 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
264 * function is guaranteed to return non-NULL css.
266 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
267 struct cgroup_subsys *ss)
269 lockdep_assert_held(&cgroup_mutex);
274 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
277 while (cgroup_parent(cgrp) &&
278 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
279 cgrp = cgroup_parent(cgrp);
281 return cgroup_css(cgrp, ss);
284 /* convenient tests for these bits */
285 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
287 return !(cgrp->self.flags & CSS_ONLINE);
290 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
292 struct cgroup *cgrp = of->kn->parent->priv;
293 struct cftype *cft = of_cft(of);
296 * This is open and unprotected implementation of cgroup_css().
297 * seq_css() is only called from a kernfs file operation which has
298 * an active reference on the file. Because all the subsystem
299 * files are drained before a css is disassociated with a cgroup,
300 * the matching css from the cgroup's subsys table is guaranteed to
301 * be and stay valid until the enclosing operation is complete.
304 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
308 EXPORT_SYMBOL_GPL(of_css);
311 * cgroup_is_descendant - test ancestry
312 * @cgrp: the cgroup to be tested
313 * @ancestor: possible ancestor of @cgrp
315 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
316 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
317 * and @ancestor are accessible.
319 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
322 if (cgrp == ancestor)
324 cgrp = cgroup_parent(cgrp);
329 static int cgroup_is_releasable(const struct cgroup *cgrp)
332 (1 << CGRP_RELEASABLE) |
333 (1 << CGRP_NOTIFY_ON_RELEASE);
334 return (cgrp->flags & bits) == bits;
337 static int notify_on_release(const struct cgroup *cgrp)
339 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
343 * for_each_css - iterate all 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_css(css, ssid, cgrp) \
351 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
352 if (!((css) = rcu_dereference_check( \
353 (cgrp)->subsys[(ssid)], \
354 lockdep_is_held(&cgroup_mutex)))) { } \
358 * for_each_e_css - iterate all effective css's of a cgroup
359 * @css: the iteration cursor
360 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
361 * @cgrp: the target cgroup to iterate css's of
363 * Should be called under cgroup_[tree_]mutex.
365 #define for_each_e_css(css, ssid, cgrp) \
366 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
367 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
372 * for_each_subsys - iterate all enabled cgroup subsystems
373 * @ss: the iteration cursor
374 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
376 #define for_each_subsys(ss, ssid) \
377 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
378 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
380 /* iterate across the hierarchies */
381 #define for_each_root(root) \
382 list_for_each_entry((root), &cgroup_roots, root_list)
384 /* iterate over child cgrps, lock should be held throughout iteration */
385 #define cgroup_for_each_live_child(child, cgrp) \
386 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
387 if (({ lockdep_assert_held(&cgroup_mutex); \
388 cgroup_is_dead(child); })) \
392 /* the list of cgroups eligible for automatic release. Protected by
393 * release_list_lock */
394 static LIST_HEAD(release_list);
395 static DEFINE_RAW_SPINLOCK(release_list_lock);
396 static void cgroup_release_agent(struct work_struct *work);
397 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
398 static void check_for_release(struct cgroup *cgrp);
401 * A cgroup can be associated with multiple css_sets as different tasks may
402 * belong to different cgroups on different hierarchies. In the other
403 * direction, a css_set is naturally associated with multiple cgroups.
404 * This M:N relationship is represented by the following link structure
405 * which exists for each association and allows traversing the associations
408 struct cgrp_cset_link {
409 /* the cgroup and css_set this link associates */
411 struct css_set *cset;
413 /* list of cgrp_cset_links anchored at cgrp->cset_links */
414 struct list_head cset_link;
416 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
417 struct list_head cgrp_link;
421 * The default css_set - used by init and its children prior to any
422 * hierarchies being mounted. It contains a pointer to the root state
423 * for each subsystem. Also used to anchor the list of css_sets. Not
424 * reference-counted, to improve performance when child cgroups
425 * haven't been created.
427 struct css_set init_css_set = {
428 .refcount = ATOMIC_INIT(1),
429 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
430 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
431 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
432 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
433 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
436 static int css_set_count = 1; /* 1 for init_css_set */
439 * cgroup_update_populated - updated populated count of a cgroup
440 * @cgrp: the target cgroup
441 * @populated: inc or dec populated count
443 * @cgrp is either getting the first task (css_set) or losing the last.
444 * Update @cgrp->populated_cnt accordingly. The count is propagated
445 * towards root so that a given cgroup's populated_cnt is zero iff the
446 * cgroup and all its descendants are empty.
448 * @cgrp's interface file "cgroup.populated" is zero if
449 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
450 * changes from or to zero, userland is notified that the content of the
451 * interface file has changed. This can be used to detect when @cgrp and
452 * its descendants become populated or empty.
454 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
456 lockdep_assert_held(&css_set_rwsem);
462 trigger = !cgrp->populated_cnt++;
464 trigger = !--cgrp->populated_cnt;
469 if (cgrp->populated_kn)
470 kernfs_notify(cgrp->populated_kn);
471 cgrp = cgroup_parent(cgrp);
476 * hash table for cgroup groups. This improves the performance to find
477 * an existing css_set. This hash doesn't (currently) take into
478 * account cgroups in empty hierarchies.
480 #define CSS_SET_HASH_BITS 7
481 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
483 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
485 unsigned long key = 0UL;
486 struct cgroup_subsys *ss;
489 for_each_subsys(ss, i)
490 key += (unsigned long)css[i];
491 key = (key >> 16) ^ key;
496 static void put_css_set_locked(struct css_set *cset, bool taskexit)
498 struct cgrp_cset_link *link, *tmp_link;
499 struct cgroup_subsys *ss;
502 lockdep_assert_held(&css_set_rwsem);
504 if (!atomic_dec_and_test(&cset->refcount))
507 /* This css_set is dead. unlink it and release cgroup refcounts */
508 for_each_subsys(ss, ssid)
509 list_del(&cset->e_cset_node[ssid]);
510 hash_del(&cset->hlist);
513 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
514 struct cgroup *cgrp = link->cgrp;
516 list_del(&link->cset_link);
517 list_del(&link->cgrp_link);
519 /* @cgrp can't go away while we're holding css_set_rwsem */
520 if (list_empty(&cgrp->cset_links)) {
521 cgroup_update_populated(cgrp, false);
522 if (notify_on_release(cgrp)) {
524 set_bit(CGRP_RELEASABLE, &cgrp->flags);
525 check_for_release(cgrp);
532 kfree_rcu(cset, rcu_head);
535 static void put_css_set(struct css_set *cset, bool taskexit)
538 * Ensure that the refcount doesn't hit zero while any readers
539 * can see it. Similar to atomic_dec_and_lock(), but for an
542 if (atomic_add_unless(&cset->refcount, -1, 1))
545 down_write(&css_set_rwsem);
546 put_css_set_locked(cset, taskexit);
547 up_write(&css_set_rwsem);
551 * refcounted get/put for css_set objects
553 static inline void get_css_set(struct css_set *cset)
555 atomic_inc(&cset->refcount);
559 * compare_css_sets - helper function for find_existing_css_set().
560 * @cset: candidate css_set being tested
561 * @old_cset: existing css_set for a task
562 * @new_cgrp: cgroup that's being entered by the task
563 * @template: desired set of css pointers in css_set (pre-calculated)
565 * Returns true if "cset" matches "old_cset" except for the hierarchy
566 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
568 static bool compare_css_sets(struct css_set *cset,
569 struct css_set *old_cset,
570 struct cgroup *new_cgrp,
571 struct cgroup_subsys_state *template[])
573 struct list_head *l1, *l2;
576 * On the default hierarchy, there can be csets which are
577 * associated with the same set of cgroups but different csses.
578 * Let's first ensure that csses match.
580 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
584 * Compare cgroup pointers in order to distinguish between
585 * different cgroups in hierarchies. As different cgroups may
586 * share the same effective css, this comparison is always
589 l1 = &cset->cgrp_links;
590 l2 = &old_cset->cgrp_links;
592 struct cgrp_cset_link *link1, *link2;
593 struct cgroup *cgrp1, *cgrp2;
597 /* See if we reached the end - both lists are equal length. */
598 if (l1 == &cset->cgrp_links) {
599 BUG_ON(l2 != &old_cset->cgrp_links);
602 BUG_ON(l2 == &old_cset->cgrp_links);
604 /* Locate the cgroups associated with these links. */
605 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
606 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
609 /* Hierarchies should be linked in the same order. */
610 BUG_ON(cgrp1->root != cgrp2->root);
613 * If this hierarchy is the hierarchy of the cgroup
614 * that's changing, then we need to check that this
615 * css_set points to the new cgroup; if it's any other
616 * hierarchy, then this css_set should point to the
617 * same cgroup as the old css_set.
619 if (cgrp1->root == new_cgrp->root) {
620 if (cgrp1 != new_cgrp)
631 * find_existing_css_set - init css array and find the matching css_set
632 * @old_cset: the css_set that we're using before the cgroup transition
633 * @cgrp: the cgroup that we're moving into
634 * @template: out param for the new set of csses, should be clear on entry
636 static struct css_set *find_existing_css_set(struct css_set *old_cset,
638 struct cgroup_subsys_state *template[])
640 struct cgroup_root *root = cgrp->root;
641 struct cgroup_subsys *ss;
642 struct css_set *cset;
647 * Build the set of subsystem state objects that we want to see in the
648 * new css_set. while subsystems can change globally, the entries here
649 * won't change, so no need for locking.
651 for_each_subsys(ss, i) {
652 if (root->subsys_mask & (1UL << i)) {
654 * @ss is in this hierarchy, so we want the
655 * effective css from @cgrp.
657 template[i] = cgroup_e_css(cgrp, ss);
660 * @ss is not in this hierarchy, so we don't want
663 template[i] = old_cset->subsys[i];
667 key = css_set_hash(template);
668 hash_for_each_possible(css_set_table, cset, hlist, key) {
669 if (!compare_css_sets(cset, old_cset, cgrp, template))
672 /* This css_set matches what we need */
676 /* No existing cgroup group matched */
680 static void free_cgrp_cset_links(struct list_head *links_to_free)
682 struct cgrp_cset_link *link, *tmp_link;
684 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
685 list_del(&link->cset_link);
691 * allocate_cgrp_cset_links - allocate cgrp_cset_links
692 * @count: the number of links to allocate
693 * @tmp_links: list_head the allocated links are put on
695 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
696 * through ->cset_link. Returns 0 on success or -errno.
698 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
700 struct cgrp_cset_link *link;
703 INIT_LIST_HEAD(tmp_links);
705 for (i = 0; i < count; i++) {
706 link = kzalloc(sizeof(*link), GFP_KERNEL);
708 free_cgrp_cset_links(tmp_links);
711 list_add(&link->cset_link, tmp_links);
717 * link_css_set - a helper function to link a css_set to a cgroup
718 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
719 * @cset: the css_set to be linked
720 * @cgrp: the destination cgroup
722 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
725 struct cgrp_cset_link *link;
727 BUG_ON(list_empty(tmp_links));
729 if (cgroup_on_dfl(cgrp))
730 cset->dfl_cgrp = cgrp;
732 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
736 if (list_empty(&cgrp->cset_links))
737 cgroup_update_populated(cgrp, true);
738 list_move(&link->cset_link, &cgrp->cset_links);
741 * Always add links to the tail of the list so that the list
742 * is sorted by order of hierarchy creation
744 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
748 * find_css_set - return a new css_set with one cgroup updated
749 * @old_cset: the baseline css_set
750 * @cgrp: the cgroup to be updated
752 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
753 * substituted into the appropriate hierarchy.
755 static struct css_set *find_css_set(struct css_set *old_cset,
758 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
759 struct css_set *cset;
760 struct list_head tmp_links;
761 struct cgrp_cset_link *link;
762 struct cgroup_subsys *ss;
766 lockdep_assert_held(&cgroup_mutex);
768 /* First see if we already have a cgroup group that matches
770 down_read(&css_set_rwsem);
771 cset = find_existing_css_set(old_cset, cgrp, template);
774 up_read(&css_set_rwsem);
779 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
783 /* Allocate all the cgrp_cset_link objects that we'll need */
784 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
789 atomic_set(&cset->refcount, 1);
790 INIT_LIST_HEAD(&cset->cgrp_links);
791 INIT_LIST_HEAD(&cset->tasks);
792 INIT_LIST_HEAD(&cset->mg_tasks);
793 INIT_LIST_HEAD(&cset->mg_preload_node);
794 INIT_LIST_HEAD(&cset->mg_node);
795 INIT_HLIST_NODE(&cset->hlist);
797 /* Copy the set of subsystem state objects generated in
798 * find_existing_css_set() */
799 memcpy(cset->subsys, template, sizeof(cset->subsys));
801 down_write(&css_set_rwsem);
802 /* Add reference counts and links from the new css_set. */
803 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
804 struct cgroup *c = link->cgrp;
806 if (c->root == cgrp->root)
808 link_css_set(&tmp_links, cset, c);
811 BUG_ON(!list_empty(&tmp_links));
815 /* Add @cset to the hash table */
816 key = css_set_hash(cset->subsys);
817 hash_add(css_set_table, &cset->hlist, key);
819 for_each_subsys(ss, ssid)
820 list_add_tail(&cset->e_cset_node[ssid],
821 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
823 up_write(&css_set_rwsem);
828 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
830 struct cgroup *root_cgrp = kf_root->kn->priv;
832 return root_cgrp->root;
835 static int cgroup_init_root_id(struct cgroup_root *root)
839 lockdep_assert_held(&cgroup_mutex);
841 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
845 root->hierarchy_id = id;
849 static void cgroup_exit_root_id(struct cgroup_root *root)
851 lockdep_assert_held(&cgroup_mutex);
853 if (root->hierarchy_id) {
854 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
855 root->hierarchy_id = 0;
859 static void cgroup_free_root(struct cgroup_root *root)
862 /* hierarhcy ID shoulid already have been released */
863 WARN_ON_ONCE(root->hierarchy_id);
865 idr_destroy(&root->cgroup_idr);
870 static void cgroup_destroy_root(struct cgroup_root *root)
872 struct cgroup *cgrp = &root->cgrp;
873 struct cgrp_cset_link *link, *tmp_link;
875 mutex_lock(&cgroup_mutex);
877 BUG_ON(atomic_read(&root->nr_cgrps));
878 BUG_ON(!list_empty(&cgrp->self.children));
880 /* Rebind all subsystems back to the default hierarchy */
881 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
884 * Release all the links from cset_links to this hierarchy's
887 down_write(&css_set_rwsem);
889 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
890 list_del(&link->cset_link);
891 list_del(&link->cgrp_link);
894 up_write(&css_set_rwsem);
896 if (!list_empty(&root->root_list)) {
897 list_del(&root->root_list);
901 cgroup_exit_root_id(root);
903 mutex_unlock(&cgroup_mutex);
905 kernfs_destroy_root(root->kf_root);
906 cgroup_free_root(root);
909 /* look up cgroup associated with given css_set on the specified hierarchy */
910 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
911 struct cgroup_root *root)
913 struct cgroup *res = NULL;
915 lockdep_assert_held(&cgroup_mutex);
916 lockdep_assert_held(&css_set_rwsem);
918 if (cset == &init_css_set) {
921 struct cgrp_cset_link *link;
923 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
924 struct cgroup *c = link->cgrp;
926 if (c->root == root) {
938 * Return the cgroup for "task" from the given hierarchy. Must be
939 * called with cgroup_mutex and css_set_rwsem held.
941 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
942 struct cgroup_root *root)
945 * No need to lock the task - since we hold cgroup_mutex the
946 * task can't change groups, so the only thing that can happen
947 * is that it exits and its css is set back to init_css_set.
949 return cset_cgroup_from_root(task_css_set(task), root);
953 * A task must hold cgroup_mutex to modify cgroups.
955 * Any task can increment and decrement the count field without lock.
956 * So in general, code holding cgroup_mutex can't rely on the count
957 * field not changing. However, if the count goes to zero, then only
958 * cgroup_attach_task() can increment it again. Because a count of zero
959 * means that no tasks are currently attached, therefore there is no
960 * way a task attached to that cgroup can fork (the other way to
961 * increment the count). So code holding cgroup_mutex can safely
962 * assume that if the count is zero, it will stay zero. Similarly, if
963 * a task holds cgroup_mutex on a cgroup with zero count, it
964 * knows that the cgroup won't be removed, as cgroup_rmdir()
967 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
968 * (usually) take cgroup_mutex. These are the two most performance
969 * critical pieces of code here. The exception occurs on cgroup_exit(),
970 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
971 * is taken, and if the cgroup count is zero, a usermode call made
972 * to the release agent with the name of the cgroup (path relative to
973 * the root of cgroup file system) as the argument.
975 * A cgroup can only be deleted if both its 'count' of using tasks
976 * is zero, and its list of 'children' cgroups is empty. Since all
977 * tasks in the system use _some_ cgroup, and since there is always at
978 * least one task in the system (init, pid == 1), therefore, root cgroup
979 * always has either children cgroups and/or using tasks. So we don't
980 * need a special hack to ensure that root cgroup cannot be deleted.
982 * P.S. One more locking exception. RCU is used to guard the
983 * update of a tasks cgroup pointer by cgroup_attach_task()
986 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
987 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
988 static const struct file_operations proc_cgroupstats_operations;
990 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
993 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
994 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
995 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
996 cft->ss->name, cft->name);
998 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1003 * cgroup_file_mode - deduce file mode of a control file
1004 * @cft: the control file in question
1006 * returns cft->mode if ->mode is not 0
1007 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1008 * returns S_IRUGO if it has only a read handler
1009 * returns S_IWUSR if it has only a write hander
1011 static umode_t cgroup_file_mode(const struct cftype *cft)
1018 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1021 if (cft->write_u64 || cft->write_s64 || cft->write)
1027 static void cgroup_get(struct cgroup *cgrp)
1029 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1030 css_get(&cgrp->self);
1033 static void cgroup_put(struct cgroup *cgrp)
1035 css_put(&cgrp->self);
1039 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1040 * @kn: the kernfs_node being serviced
1042 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1043 * the method finishes if locking succeeded. Note that once this function
1044 * returns the cgroup returned by cgroup_kn_lock_live() may become
1045 * inaccessible any time. If the caller intends to continue to access the
1046 * cgroup, it should pin it before invoking this function.
1048 static void cgroup_kn_unlock(struct kernfs_node *kn)
1050 struct cgroup *cgrp;
1052 if (kernfs_type(kn) == KERNFS_DIR)
1055 cgrp = kn->parent->priv;
1057 mutex_unlock(&cgroup_mutex);
1059 kernfs_unbreak_active_protection(kn);
1064 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1065 * @kn: the kernfs_node being serviced
1067 * This helper is to be used by a cgroup kernfs method currently servicing
1068 * @kn. It breaks the active protection, performs cgroup locking and
1069 * verifies that the associated cgroup is alive. Returns the cgroup if
1070 * alive; otherwise, %NULL. A successful return should be undone by a
1071 * matching cgroup_kn_unlock() invocation.
1073 * Any cgroup kernfs method implementation which requires locking the
1074 * associated cgroup should use this helper. It avoids nesting cgroup
1075 * locking under kernfs active protection and allows all kernfs operations
1076 * including self-removal.
1078 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1080 struct cgroup *cgrp;
1082 if (kernfs_type(kn) == KERNFS_DIR)
1085 cgrp = kn->parent->priv;
1088 * We're gonna grab cgroup_mutex which nests outside kernfs
1089 * active_ref. cgroup liveliness check alone provides enough
1090 * protection against removal. Ensure @cgrp stays accessible and
1091 * break the active_ref protection.
1094 kernfs_break_active_protection(kn);
1096 mutex_lock(&cgroup_mutex);
1098 if (!cgroup_is_dead(cgrp))
1101 cgroup_kn_unlock(kn);
1105 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1107 char name[CGROUP_FILE_NAME_MAX];
1109 lockdep_assert_held(&cgroup_mutex);
1110 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1114 * cgroup_clear_dir - remove subsys files in a cgroup directory
1115 * @cgrp: target cgroup
1116 * @subsys_mask: mask of the subsystem ids whose files should be removed
1118 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1120 struct cgroup_subsys *ss;
1123 for_each_subsys(ss, i) {
1124 struct cftype *cfts;
1126 if (!(subsys_mask & (1 << i)))
1128 list_for_each_entry(cfts, &ss->cfts, node)
1129 cgroup_addrm_files(cgrp, cfts, false);
1133 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1135 struct cgroup_subsys *ss;
1136 unsigned int tmp_ss_mask;
1139 lockdep_assert_held(&cgroup_mutex);
1141 for_each_subsys(ss, ssid) {
1142 if (!(ss_mask & (1 << ssid)))
1145 /* if @ss has non-root csses attached to it, can't move */
1146 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1149 /* can't move between two non-dummy roots either */
1150 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1154 /* skip creating root files on dfl_root for inhibited subsystems */
1155 tmp_ss_mask = ss_mask;
1156 if (dst_root == &cgrp_dfl_root)
1157 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1159 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1161 if (dst_root != &cgrp_dfl_root)
1165 * Rebinding back to the default root is not allowed to
1166 * fail. Using both default and non-default roots should
1167 * be rare. Moving subsystems back and forth even more so.
1168 * Just warn about it and continue.
1170 if (cgrp_dfl_root_visible) {
1171 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1173 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1178 * Nothing can fail from this point on. Remove files for the
1179 * removed subsystems and rebind each subsystem.
1181 for_each_subsys(ss, ssid)
1182 if (ss_mask & (1 << ssid))
1183 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1185 for_each_subsys(ss, ssid) {
1186 struct cgroup_root *src_root;
1187 struct cgroup_subsys_state *css;
1188 struct css_set *cset;
1190 if (!(ss_mask & (1 << ssid)))
1193 src_root = ss->root;
1194 css = cgroup_css(&src_root->cgrp, ss);
1196 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1198 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1199 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1200 ss->root = dst_root;
1201 css->cgroup = &dst_root->cgrp;
1203 down_write(&css_set_rwsem);
1204 hash_for_each(css_set_table, i, cset, hlist)
1205 list_move_tail(&cset->e_cset_node[ss->id],
1206 &dst_root->cgrp.e_csets[ss->id]);
1207 up_write(&css_set_rwsem);
1209 src_root->subsys_mask &= ~(1 << ssid);
1210 src_root->cgrp.child_subsys_mask &= ~(1 << ssid);
1212 /* default hierarchy doesn't enable controllers by default */
1213 dst_root->subsys_mask |= 1 << ssid;
1214 if (dst_root != &cgrp_dfl_root)
1215 dst_root->cgrp.child_subsys_mask |= 1 << ssid;
1221 kernfs_activate(dst_root->cgrp.kn);
1225 static int cgroup_show_options(struct seq_file *seq,
1226 struct kernfs_root *kf_root)
1228 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1229 struct cgroup_subsys *ss;
1232 for_each_subsys(ss, ssid)
1233 if (root->subsys_mask & (1 << ssid))
1234 seq_printf(seq, ",%s", ss->name);
1235 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1236 seq_puts(seq, ",sane_behavior");
1237 if (root->flags & CGRP_ROOT_NOPREFIX)
1238 seq_puts(seq, ",noprefix");
1239 if (root->flags & CGRP_ROOT_XATTR)
1240 seq_puts(seq, ",xattr");
1242 spin_lock(&release_agent_path_lock);
1243 if (strlen(root->release_agent_path))
1244 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1245 spin_unlock(&release_agent_path_lock);
1247 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1248 seq_puts(seq, ",clone_children");
1249 if (strlen(root->name))
1250 seq_printf(seq, ",name=%s", root->name);
1254 struct cgroup_sb_opts {
1255 unsigned int subsys_mask;
1257 char *release_agent;
1258 bool cpuset_clone_children;
1260 /* User explicitly requested empty subsystem */
1264 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1266 char *token, *o = data;
1267 bool all_ss = false, one_ss = false;
1268 unsigned int mask = -1U;
1269 struct cgroup_subsys *ss;
1272 #ifdef CONFIG_CPUSETS
1273 mask = ~(1U << cpuset_cgrp_id);
1276 memset(opts, 0, sizeof(*opts));
1278 while ((token = strsep(&o, ",")) != NULL) {
1281 if (!strcmp(token, "none")) {
1282 /* Explicitly have no subsystems */
1286 if (!strcmp(token, "all")) {
1287 /* Mutually exclusive option 'all' + subsystem name */
1293 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1294 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1297 if (!strcmp(token, "noprefix")) {
1298 opts->flags |= CGRP_ROOT_NOPREFIX;
1301 if (!strcmp(token, "clone_children")) {
1302 opts->cpuset_clone_children = true;
1305 if (!strcmp(token, "xattr")) {
1306 opts->flags |= CGRP_ROOT_XATTR;
1309 if (!strncmp(token, "release_agent=", 14)) {
1310 /* Specifying two release agents is forbidden */
1311 if (opts->release_agent)
1313 opts->release_agent =
1314 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1315 if (!opts->release_agent)
1319 if (!strncmp(token, "name=", 5)) {
1320 const char *name = token + 5;
1321 /* Can't specify an empty name */
1324 /* Must match [\w.-]+ */
1325 for (i = 0; i < strlen(name); i++) {
1329 if ((c == '.') || (c == '-') || (c == '_'))
1333 /* Specifying two names is forbidden */
1336 opts->name = kstrndup(name,
1337 MAX_CGROUP_ROOT_NAMELEN - 1,
1345 for_each_subsys(ss, i) {
1346 if (strcmp(token, ss->name))
1351 /* Mutually exclusive option 'all' + subsystem name */
1354 opts->subsys_mask |= (1 << i);
1359 if (i == CGROUP_SUBSYS_COUNT)
1363 /* Consistency checks */
1365 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1366 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1368 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1369 opts->cpuset_clone_children || opts->release_agent ||
1371 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1376 * If the 'all' option was specified select all the
1377 * subsystems, otherwise if 'none', 'name=' and a subsystem
1378 * name options were not specified, let's default to 'all'
1380 if (all_ss || (!one_ss && !opts->none && !opts->name))
1381 for_each_subsys(ss, i)
1383 opts->subsys_mask |= (1 << i);
1386 * We either have to specify by name or by subsystems. (So
1387 * all empty hierarchies must have a name).
1389 if (!opts->subsys_mask && !opts->name)
1394 * Option noprefix was introduced just for backward compatibility
1395 * with the old cpuset, so we allow noprefix only if mounting just
1396 * the cpuset subsystem.
1398 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1402 /* Can't specify "none" and some subsystems */
1403 if (opts->subsys_mask && opts->none)
1409 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1412 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1413 struct cgroup_sb_opts opts;
1414 unsigned int added_mask, removed_mask;
1416 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1417 pr_err("sane_behavior: remount is not allowed\n");
1421 mutex_lock(&cgroup_mutex);
1423 /* See what subsystems are wanted */
1424 ret = parse_cgroupfs_options(data, &opts);
1428 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1429 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1430 task_tgid_nr(current), current->comm);
1432 added_mask = opts.subsys_mask & ~root->subsys_mask;
1433 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1435 /* Don't allow flags or name to change at remount */
1436 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1437 (opts.name && strcmp(opts.name, root->name))) {
1438 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1439 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1440 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1445 /* remounting is not allowed for populated hierarchies */
1446 if (!list_empty(&root->cgrp.self.children)) {
1451 ret = rebind_subsystems(root, added_mask);
1455 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1457 if (opts.release_agent) {
1458 spin_lock(&release_agent_path_lock);
1459 strcpy(root->release_agent_path, opts.release_agent);
1460 spin_unlock(&release_agent_path_lock);
1463 kfree(opts.release_agent);
1465 mutex_unlock(&cgroup_mutex);
1470 * To reduce the fork() overhead for systems that are not actually using
1471 * their cgroups capability, we don't maintain the lists running through
1472 * each css_set to its tasks until we see the list actually used - in other
1473 * words after the first mount.
1475 static bool use_task_css_set_links __read_mostly;
1477 static void cgroup_enable_task_cg_lists(void)
1479 struct task_struct *p, *g;
1481 down_write(&css_set_rwsem);
1483 if (use_task_css_set_links)
1486 use_task_css_set_links = true;
1489 * We need tasklist_lock because RCU is not safe against
1490 * while_each_thread(). Besides, a forking task that has passed
1491 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1492 * is not guaranteed to have its child immediately visible in the
1493 * tasklist if we walk through it with RCU.
1495 read_lock(&tasklist_lock);
1496 do_each_thread(g, p) {
1497 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1498 task_css_set(p) != &init_css_set);
1501 * We should check if the process is exiting, otherwise
1502 * it will race with cgroup_exit() in that the list
1503 * entry won't be deleted though the process has exited.
1504 * Do it while holding siglock so that we don't end up
1505 * racing against cgroup_exit().
1507 spin_lock_irq(&p->sighand->siglock);
1508 if (!(p->flags & PF_EXITING)) {
1509 struct css_set *cset = task_css_set(p);
1511 list_add(&p->cg_list, &cset->tasks);
1514 spin_unlock_irq(&p->sighand->siglock);
1515 } while_each_thread(g, p);
1516 read_unlock(&tasklist_lock);
1518 up_write(&css_set_rwsem);
1521 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1523 struct cgroup_subsys *ss;
1526 INIT_LIST_HEAD(&cgrp->self.sibling);
1527 INIT_LIST_HEAD(&cgrp->self.children);
1528 INIT_LIST_HEAD(&cgrp->cset_links);
1529 INIT_LIST_HEAD(&cgrp->release_list);
1530 INIT_LIST_HEAD(&cgrp->pidlists);
1531 mutex_init(&cgrp->pidlist_mutex);
1532 cgrp->self.cgroup = cgrp;
1533 cgrp->self.flags |= CSS_ONLINE;
1535 for_each_subsys(ss, ssid)
1536 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1538 init_waitqueue_head(&cgrp->offline_waitq);
1541 static void init_cgroup_root(struct cgroup_root *root,
1542 struct cgroup_sb_opts *opts)
1544 struct cgroup *cgrp = &root->cgrp;
1546 INIT_LIST_HEAD(&root->root_list);
1547 atomic_set(&root->nr_cgrps, 1);
1549 init_cgroup_housekeeping(cgrp);
1550 idr_init(&root->cgroup_idr);
1552 root->flags = opts->flags;
1553 if (opts->release_agent)
1554 strcpy(root->release_agent_path, opts->release_agent);
1556 strcpy(root->name, opts->name);
1557 if (opts->cpuset_clone_children)
1558 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1561 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1563 LIST_HEAD(tmp_links);
1564 struct cgroup *root_cgrp = &root->cgrp;
1565 struct css_set *cset;
1568 lockdep_assert_held(&cgroup_mutex);
1570 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1573 root_cgrp->id = ret;
1575 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
1580 * We're accessing css_set_count without locking css_set_rwsem here,
1581 * but that's OK - it can only be increased by someone holding
1582 * cgroup_lock, and that's us. The worst that can happen is that we
1583 * have some link structures left over
1585 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1589 ret = cgroup_init_root_id(root);
1593 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1594 KERNFS_ROOT_CREATE_DEACTIVATED,
1596 if (IS_ERR(root->kf_root)) {
1597 ret = PTR_ERR(root->kf_root);
1600 root_cgrp->kn = root->kf_root->kn;
1602 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1606 ret = rebind_subsystems(root, ss_mask);
1611 * There must be no failure case after here, since rebinding takes
1612 * care of subsystems' refcounts, which are explicitly dropped in
1613 * the failure exit path.
1615 list_add(&root->root_list, &cgroup_roots);
1616 cgroup_root_count++;
1619 * Link the root cgroup in this hierarchy into all the css_set
1622 down_write(&css_set_rwsem);
1623 hash_for_each(css_set_table, i, cset, hlist)
1624 link_css_set(&tmp_links, cset, root_cgrp);
1625 up_write(&css_set_rwsem);
1627 BUG_ON(!list_empty(&root_cgrp->self.children));
1628 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1630 kernfs_activate(root_cgrp->kn);
1635 kernfs_destroy_root(root->kf_root);
1636 root->kf_root = NULL;
1638 cgroup_exit_root_id(root);
1640 percpu_ref_cancel_init(&root_cgrp->self.refcnt);
1642 free_cgrp_cset_links(&tmp_links);
1646 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1647 int flags, const char *unused_dev_name,
1650 struct cgroup_root *root;
1651 struct cgroup_sb_opts opts;
1652 struct dentry *dentry;
1657 * The first time anyone tries to mount a cgroup, enable the list
1658 * linking each css_set to its tasks and fix up all existing tasks.
1660 if (!use_task_css_set_links)
1661 cgroup_enable_task_cg_lists();
1663 mutex_lock(&cgroup_mutex);
1665 /* First find the desired set of subsystems */
1666 ret = parse_cgroupfs_options(data, &opts);
1670 /* look for a matching existing root */
1671 if (!opts.subsys_mask && !opts.none && !opts.name) {
1672 cgrp_dfl_root_visible = true;
1673 root = &cgrp_dfl_root;
1674 cgroup_get(&root->cgrp);
1679 for_each_root(root) {
1680 bool name_match = false;
1682 if (root == &cgrp_dfl_root)
1686 * If we asked for a name then it must match. Also, if
1687 * name matches but sybsys_mask doesn't, we should fail.
1688 * Remember whether name matched.
1691 if (strcmp(opts.name, root->name))
1697 * If we asked for subsystems (or explicitly for no
1698 * subsystems) then they must match.
1700 if ((opts.subsys_mask || opts.none) &&
1701 (opts.subsys_mask != root->subsys_mask)) {
1708 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1709 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1710 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1714 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1719 * A root's lifetime is governed by its root cgroup.
1720 * tryget_live failure indicate that the root is being
1721 * destroyed. Wait for destruction to complete so that the
1722 * subsystems are free. We can use wait_queue for the wait
1723 * but this path is super cold. Let's just sleep for a bit
1726 if (!percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1727 mutex_unlock(&cgroup_mutex);
1729 ret = restart_syscall();
1738 * No such thing, create a new one. name= matching without subsys
1739 * specification is allowed for already existing hierarchies but we
1740 * can't create new one without subsys specification.
1742 if (!opts.subsys_mask && !opts.none) {
1747 root = kzalloc(sizeof(*root), GFP_KERNEL);
1753 init_cgroup_root(root, &opts);
1755 ret = cgroup_setup_root(root, opts.subsys_mask);
1757 cgroup_free_root(root);
1760 mutex_unlock(&cgroup_mutex);
1762 kfree(opts.release_agent);
1766 return ERR_PTR(ret);
1768 dentry = kernfs_mount(fs_type, flags, root->kf_root, &new_sb);
1769 if (IS_ERR(dentry) || !new_sb)
1770 cgroup_put(&root->cgrp);
1774 static void cgroup_kill_sb(struct super_block *sb)
1776 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1777 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1780 * If @root doesn't have any mounts or children, start killing it.
1781 * This prevents new mounts by disabling percpu_ref_tryget_live().
1782 * cgroup_mount() may wait for @root's release.
1784 * And don't kill the default root.
1786 if (css_has_online_children(&root->cgrp.self) ||
1787 root == &cgrp_dfl_root)
1788 cgroup_put(&root->cgrp);
1790 percpu_ref_kill(&root->cgrp.self.refcnt);
1795 static struct file_system_type cgroup_fs_type = {
1797 .mount = cgroup_mount,
1798 .kill_sb = cgroup_kill_sb,
1801 static struct kobject *cgroup_kobj;
1804 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1805 * @task: target task
1806 * @buf: the buffer to write the path into
1807 * @buflen: the length of the buffer
1809 * Determine @task's cgroup on the first (the one with the lowest non-zero
1810 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1811 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1812 * cgroup controller callbacks.
1814 * Return value is the same as kernfs_path().
1816 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1818 struct cgroup_root *root;
1819 struct cgroup *cgrp;
1820 int hierarchy_id = 1;
1823 mutex_lock(&cgroup_mutex);
1824 down_read(&css_set_rwsem);
1826 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1829 cgrp = task_cgroup_from_root(task, root);
1830 path = cgroup_path(cgrp, buf, buflen);
1832 /* if no hierarchy exists, everyone is in "/" */
1833 if (strlcpy(buf, "/", buflen) < buflen)
1837 up_read(&css_set_rwsem);
1838 mutex_unlock(&cgroup_mutex);
1841 EXPORT_SYMBOL_GPL(task_cgroup_path);
1843 /* used to track tasks and other necessary states during migration */
1844 struct cgroup_taskset {
1845 /* the src and dst cset list running through cset->mg_node */
1846 struct list_head src_csets;
1847 struct list_head dst_csets;
1850 * Fields for cgroup_taskset_*() iteration.
1852 * Before migration is committed, the target migration tasks are on
1853 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1854 * the csets on ->dst_csets. ->csets point to either ->src_csets
1855 * or ->dst_csets depending on whether migration is committed.
1857 * ->cur_csets and ->cur_task point to the current task position
1860 struct list_head *csets;
1861 struct css_set *cur_cset;
1862 struct task_struct *cur_task;
1866 * cgroup_taskset_first - reset taskset and return the first task
1867 * @tset: taskset of interest
1869 * @tset iteration is initialized and the first task is returned.
1871 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1873 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1874 tset->cur_task = NULL;
1876 return cgroup_taskset_next(tset);
1880 * cgroup_taskset_next - iterate to the next task in taskset
1881 * @tset: taskset of interest
1883 * Return the next task in @tset. Iteration must have been initialized
1884 * with cgroup_taskset_first().
1886 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1888 struct css_set *cset = tset->cur_cset;
1889 struct task_struct *task = tset->cur_task;
1891 while (&cset->mg_node != tset->csets) {
1893 task = list_first_entry(&cset->mg_tasks,
1894 struct task_struct, cg_list);
1896 task = list_next_entry(task, cg_list);
1898 if (&task->cg_list != &cset->mg_tasks) {
1899 tset->cur_cset = cset;
1900 tset->cur_task = task;
1904 cset = list_next_entry(cset, mg_node);
1912 * cgroup_task_migrate - move a task from one cgroup to another.
1913 * @old_cgrp: the cgroup @tsk is being migrated from
1914 * @tsk: the task being migrated
1915 * @new_cset: the new css_set @tsk is being attached to
1917 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1919 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1920 struct task_struct *tsk,
1921 struct css_set *new_cset)
1923 struct css_set *old_cset;
1925 lockdep_assert_held(&cgroup_mutex);
1926 lockdep_assert_held(&css_set_rwsem);
1929 * We are synchronized through threadgroup_lock() against PF_EXITING
1930 * setting such that we can't race against cgroup_exit() changing the
1931 * css_set to init_css_set and dropping the old one.
1933 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1934 old_cset = task_css_set(tsk);
1936 get_css_set(new_cset);
1937 rcu_assign_pointer(tsk->cgroups, new_cset);
1940 * Use move_tail so that cgroup_taskset_first() still returns the
1941 * leader after migration. This works because cgroup_migrate()
1942 * ensures that the dst_cset of the leader is the first on the
1943 * tset's dst_csets list.
1945 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1948 * We just gained a reference on old_cset by taking it from the
1949 * task. As trading it for new_cset is protected by cgroup_mutex,
1950 * we're safe to drop it here; it will be freed under RCU.
1952 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1953 put_css_set_locked(old_cset, false);
1957 * cgroup_migrate_finish - cleanup after attach
1958 * @preloaded_csets: list of preloaded css_sets
1960 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1961 * those functions for details.
1963 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1965 struct css_set *cset, *tmp_cset;
1967 lockdep_assert_held(&cgroup_mutex);
1969 down_write(&css_set_rwsem);
1970 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1971 cset->mg_src_cgrp = NULL;
1972 cset->mg_dst_cset = NULL;
1973 list_del_init(&cset->mg_preload_node);
1974 put_css_set_locked(cset, false);
1976 up_write(&css_set_rwsem);
1980 * cgroup_migrate_add_src - add a migration source css_set
1981 * @src_cset: the source css_set to add
1982 * @dst_cgrp: the destination cgroup
1983 * @preloaded_csets: list of preloaded css_sets
1985 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1986 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1987 * up by cgroup_migrate_finish().
1989 * This function may be called without holding threadgroup_lock even if the
1990 * target is a process. Threads may be created and destroyed but as long
1991 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1992 * the preloaded css_sets are guaranteed to cover all migrations.
1994 static void cgroup_migrate_add_src(struct css_set *src_cset,
1995 struct cgroup *dst_cgrp,
1996 struct list_head *preloaded_csets)
1998 struct cgroup *src_cgrp;
2000 lockdep_assert_held(&cgroup_mutex);
2001 lockdep_assert_held(&css_set_rwsem);
2003 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2005 if (!list_empty(&src_cset->mg_preload_node))
2008 WARN_ON(src_cset->mg_src_cgrp);
2009 WARN_ON(!list_empty(&src_cset->mg_tasks));
2010 WARN_ON(!list_empty(&src_cset->mg_node));
2012 src_cset->mg_src_cgrp = src_cgrp;
2013 get_css_set(src_cset);
2014 list_add(&src_cset->mg_preload_node, preloaded_csets);
2018 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2019 * @dst_cgrp: the destination cgroup (may be %NULL)
2020 * @preloaded_csets: list of preloaded source css_sets
2022 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2023 * have been preloaded to @preloaded_csets. This function looks up and
2024 * pins all destination css_sets, links each to its source, and append them
2025 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2026 * source css_set is assumed to be its cgroup on the default hierarchy.
2028 * This function must be called after cgroup_migrate_add_src() has been
2029 * called on each migration source css_set. After migration is performed
2030 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2033 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2034 struct list_head *preloaded_csets)
2037 struct css_set *src_cset, *tmp_cset;
2039 lockdep_assert_held(&cgroup_mutex);
2042 * Except for the root, child_subsys_mask must be zero for a cgroup
2043 * with tasks so that child cgroups don't compete against tasks.
2045 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2046 dst_cgrp->child_subsys_mask)
2049 /* look up the dst cset for each src cset and link it to src */
2050 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2051 struct css_set *dst_cset;
2053 dst_cset = find_css_set(src_cset,
2054 dst_cgrp ?: src_cset->dfl_cgrp);
2058 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2061 * If src cset equals dst, it's noop. Drop the src.
2062 * cgroup_migrate() will skip the cset too. Note that we
2063 * can't handle src == dst as some nodes are used by both.
2065 if (src_cset == dst_cset) {
2066 src_cset->mg_src_cgrp = NULL;
2067 list_del_init(&src_cset->mg_preload_node);
2068 put_css_set(src_cset, false);
2069 put_css_set(dst_cset, false);
2073 src_cset->mg_dst_cset = dst_cset;
2075 if (list_empty(&dst_cset->mg_preload_node))
2076 list_add(&dst_cset->mg_preload_node, &csets);
2078 put_css_set(dst_cset, false);
2081 list_splice_tail(&csets, preloaded_csets);
2084 cgroup_migrate_finish(&csets);
2089 * cgroup_migrate - migrate a process or task to a cgroup
2090 * @cgrp: the destination cgroup
2091 * @leader: the leader of the process or the task to migrate
2092 * @threadgroup: whether @leader points to the whole process or a single task
2094 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2095 * process, the caller must be holding threadgroup_lock of @leader. The
2096 * caller is also responsible for invoking cgroup_migrate_add_src() and
2097 * cgroup_migrate_prepare_dst() on the targets before invoking this
2098 * function and following up with cgroup_migrate_finish().
2100 * As long as a controller's ->can_attach() doesn't fail, this function is
2101 * guaranteed to succeed. This means that, excluding ->can_attach()
2102 * failure, when migrating multiple targets, the success or failure can be
2103 * decided for all targets by invoking group_migrate_prepare_dst() before
2104 * actually starting migrating.
2106 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2109 struct cgroup_taskset tset = {
2110 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2111 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2112 .csets = &tset.src_csets,
2114 struct cgroup_subsys_state *css, *failed_css = NULL;
2115 struct css_set *cset, *tmp_cset;
2116 struct task_struct *task, *tmp_task;
2120 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2121 * already PF_EXITING could be freed from underneath us unless we
2122 * take an rcu_read_lock.
2124 down_write(&css_set_rwsem);
2128 /* @task either already exited or can't exit until the end */
2129 if (task->flags & PF_EXITING)
2132 /* leave @task alone if post_fork() hasn't linked it yet */
2133 if (list_empty(&task->cg_list))
2136 cset = task_css_set(task);
2137 if (!cset->mg_src_cgrp)
2141 * cgroup_taskset_first() must always return the leader.
2142 * Take care to avoid disturbing the ordering.
2144 list_move_tail(&task->cg_list, &cset->mg_tasks);
2145 if (list_empty(&cset->mg_node))
2146 list_add_tail(&cset->mg_node, &tset.src_csets);
2147 if (list_empty(&cset->mg_dst_cset->mg_node))
2148 list_move_tail(&cset->mg_dst_cset->mg_node,
2153 } while_each_thread(leader, task);
2155 up_write(&css_set_rwsem);
2157 /* methods shouldn't be called if no task is actually migrating */
2158 if (list_empty(&tset.src_csets))
2161 /* check that we can legitimately attach to the cgroup */
2162 for_each_e_css(css, i, cgrp) {
2163 if (css->ss->can_attach) {
2164 ret = css->ss->can_attach(css, &tset);
2167 goto out_cancel_attach;
2173 * Now that we're guaranteed success, proceed to move all tasks to
2174 * the new cgroup. There are no failure cases after here, so this
2175 * is the commit point.
2177 down_write(&css_set_rwsem);
2178 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2179 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2180 cgroup_task_migrate(cset->mg_src_cgrp, task,
2183 up_write(&css_set_rwsem);
2186 * Migration is committed, all target tasks are now on dst_csets.
2187 * Nothing is sensitive to fork() after this point. Notify
2188 * controllers that migration is complete.
2190 tset.csets = &tset.dst_csets;
2192 for_each_e_css(css, i, cgrp)
2193 if (css->ss->attach)
2194 css->ss->attach(css, &tset);
2197 goto out_release_tset;
2200 for_each_e_css(css, i, cgrp) {
2201 if (css == failed_css)
2203 if (css->ss->cancel_attach)
2204 css->ss->cancel_attach(css, &tset);
2207 down_write(&css_set_rwsem);
2208 list_splice_init(&tset.dst_csets, &tset.src_csets);
2209 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2210 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2211 list_del_init(&cset->mg_node);
2213 up_write(&css_set_rwsem);
2218 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2219 * @dst_cgrp: the cgroup to attach to
2220 * @leader: the task or the leader of the threadgroup to be attached
2221 * @threadgroup: attach the whole threadgroup?
2223 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2225 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2226 struct task_struct *leader, bool threadgroup)
2228 LIST_HEAD(preloaded_csets);
2229 struct task_struct *task;
2232 /* look up all src csets */
2233 down_read(&css_set_rwsem);
2237 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2241 } while_each_thread(leader, task);
2243 up_read(&css_set_rwsem);
2245 /* prepare dst csets and commit */
2246 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2248 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2250 cgroup_migrate_finish(&preloaded_csets);
2255 * Find the task_struct of the task to attach by vpid and pass it along to the
2256 * function to attach either it or all tasks in its threadgroup. Will lock
2257 * cgroup_mutex and threadgroup.
2259 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2260 size_t nbytes, loff_t off, bool threadgroup)
2262 struct task_struct *tsk;
2263 const struct cred *cred = current_cred(), *tcred;
2264 struct cgroup *cgrp;
2268 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2271 cgrp = cgroup_kn_lock_live(of->kn);
2278 tsk = find_task_by_vpid(pid);
2282 goto out_unlock_cgroup;
2285 * even if we're attaching all tasks in the thread group, we
2286 * only need to check permissions on one of them.
2288 tcred = __task_cred(tsk);
2289 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2290 !uid_eq(cred->euid, tcred->uid) &&
2291 !uid_eq(cred->euid, tcred->suid)) {
2294 goto out_unlock_cgroup;
2300 tsk = tsk->group_leader;
2303 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2304 * trapped in a cpuset, or RT worker may be born in a cgroup
2305 * with no rt_runtime allocated. Just say no.
2307 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2310 goto out_unlock_cgroup;
2313 get_task_struct(tsk);
2316 threadgroup_lock(tsk);
2318 if (!thread_group_leader(tsk)) {
2320 * a race with de_thread from another thread's exec()
2321 * may strip us of our leadership, if this happens,
2322 * there is no choice but to throw this task away and
2323 * try again; this is
2324 * "double-double-toil-and-trouble-check locking".
2326 threadgroup_unlock(tsk);
2327 put_task_struct(tsk);
2328 goto retry_find_task;
2332 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2334 threadgroup_unlock(tsk);
2336 put_task_struct(tsk);
2338 cgroup_kn_unlock(of->kn);
2339 return ret ?: nbytes;
2343 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2344 * @from: attach to all cgroups of a given task
2345 * @tsk: the task to be attached
2347 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2349 struct cgroup_root *root;
2352 mutex_lock(&cgroup_mutex);
2353 for_each_root(root) {
2354 struct cgroup *from_cgrp;
2356 if (root == &cgrp_dfl_root)
2359 down_read(&css_set_rwsem);
2360 from_cgrp = task_cgroup_from_root(from, root);
2361 up_read(&css_set_rwsem);
2363 retval = cgroup_attach_task(from_cgrp, tsk, false);
2367 mutex_unlock(&cgroup_mutex);
2371 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2373 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2374 char *buf, size_t nbytes, loff_t off)
2376 return __cgroup_procs_write(of, buf, nbytes, off, false);
2379 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2380 char *buf, size_t nbytes, loff_t off)
2382 return __cgroup_procs_write(of, buf, nbytes, off, true);
2385 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2386 char *buf, size_t nbytes, loff_t off)
2388 struct cgroup *cgrp;
2390 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2392 cgrp = cgroup_kn_lock_live(of->kn);
2395 spin_lock(&release_agent_path_lock);
2396 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2397 sizeof(cgrp->root->release_agent_path));
2398 spin_unlock(&release_agent_path_lock);
2399 cgroup_kn_unlock(of->kn);
2403 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2405 struct cgroup *cgrp = seq_css(seq)->cgroup;
2407 spin_lock(&release_agent_path_lock);
2408 seq_puts(seq, cgrp->root->release_agent_path);
2409 spin_unlock(&release_agent_path_lock);
2410 seq_putc(seq, '\n');
2414 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2416 struct cgroup *cgrp = seq_css(seq)->cgroup;
2418 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2422 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2424 struct cgroup_subsys *ss;
2425 bool printed = false;
2428 for_each_subsys(ss, ssid) {
2429 if (ss_mask & (1 << ssid)) {
2432 seq_printf(seq, "%s", ss->name);
2437 seq_putc(seq, '\n');
2440 /* show controllers which are currently attached to the default hierarchy */
2441 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2443 struct cgroup *cgrp = seq_css(seq)->cgroup;
2445 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2446 ~cgrp_dfl_root_inhibit_ss_mask);
2450 /* show controllers which are enabled from the parent */
2451 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2453 struct cgroup *cgrp = seq_css(seq)->cgroup;
2455 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->child_subsys_mask);
2459 /* show controllers which are enabled for a given cgroup's children */
2460 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2462 struct cgroup *cgrp = seq_css(seq)->cgroup;
2464 cgroup_print_ss_mask(seq, cgrp->child_subsys_mask);
2469 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2470 * @cgrp: root of the subtree to update csses for
2472 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2473 * css associations need to be updated accordingly. This function looks up
2474 * all css_sets which are attached to the subtree, creates the matching
2475 * updated css_sets and migrates the tasks to the new ones.
2477 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2479 LIST_HEAD(preloaded_csets);
2480 struct cgroup_subsys_state *css;
2481 struct css_set *src_cset;
2484 lockdep_assert_held(&cgroup_mutex);
2486 /* look up all csses currently attached to @cgrp's subtree */
2487 down_read(&css_set_rwsem);
2488 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2489 struct cgrp_cset_link *link;
2491 /* self is not affected by child_subsys_mask change */
2492 if (css->cgroup == cgrp)
2495 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2496 cgroup_migrate_add_src(link->cset, cgrp,
2499 up_read(&css_set_rwsem);
2501 /* NULL dst indicates self on default hierarchy */
2502 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2506 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2507 struct task_struct *last_task = NULL, *task;
2509 /* src_csets precede dst_csets, break on the first dst_cset */
2510 if (!src_cset->mg_src_cgrp)
2514 * All tasks in src_cset need to be migrated to the
2515 * matching dst_cset. Empty it process by process. We
2516 * walk tasks but migrate processes. The leader might even
2517 * belong to a different cset but such src_cset would also
2518 * be among the target src_csets because the default
2519 * hierarchy enforces per-process membership.
2522 down_read(&css_set_rwsem);
2523 task = list_first_entry_or_null(&src_cset->tasks,
2524 struct task_struct, cg_list);
2526 task = task->group_leader;
2527 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2528 get_task_struct(task);
2530 up_read(&css_set_rwsem);
2535 /* guard against possible infinite loop */
2536 if (WARN(last_task == task,
2537 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2541 threadgroup_lock(task);
2542 /* raced against de_thread() from another thread? */
2543 if (!thread_group_leader(task)) {
2544 threadgroup_unlock(task);
2545 put_task_struct(task);
2549 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2551 threadgroup_unlock(task);
2552 put_task_struct(task);
2554 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2560 cgroup_migrate_finish(&preloaded_csets);
2564 /* change the enabled child controllers for a cgroup in the default hierarchy */
2565 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2566 char *buf, size_t nbytes,
2569 unsigned int enable = 0, disable = 0;
2570 struct cgroup *cgrp, *child;
2571 struct cgroup_subsys *ss;
2576 * Parse input - space separated list of subsystem names prefixed
2577 * with either + or -.
2579 buf = strstrip(buf);
2580 while ((tok = strsep(&buf, " "))) {
2583 for_each_subsys(ss, ssid) {
2584 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2585 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
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)) ||
2639 (cgroup_parent(cgrp) &&
2640 !(cgroup_parent(cgrp)->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 && cgroup_parent(cgrp) && !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) && !cgroup_parent(cgrp))
2929 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
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)
3075 if (!cfts || cfts[0].name[0] == '\0')
3078 ret = cgroup_init_cftypes(ss, cfts);
3082 mutex_lock(&cgroup_mutex);
3084 list_add_tail(&cfts->node, &ss->cfts);
3085 ret = cgroup_apply_cftypes(cfts, true);
3087 cgroup_rm_cftypes_locked(cfts);
3089 mutex_unlock(&cgroup_mutex);
3094 * cgroup_task_count - count the number of tasks in a cgroup.
3095 * @cgrp: the cgroup in question
3097 * Return the number of tasks in the cgroup.
3099 static int cgroup_task_count(const struct cgroup *cgrp)
3102 struct cgrp_cset_link *link;
3104 down_read(&css_set_rwsem);
3105 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3106 count += atomic_read(&link->cset->refcount);
3107 up_read(&css_set_rwsem);
3112 * css_next_child - find the next child of a given css
3113 * @pos: the current position (%NULL to initiate traversal)
3114 * @parent: css whose children to walk
3116 * This function returns the next child of @parent and should be called
3117 * under either cgroup_mutex or RCU read lock. The only requirement is
3118 * that @parent and @pos are accessible. The next sibling is guaranteed to
3119 * be returned regardless of their states.
3121 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3122 * css which finished ->css_online() is guaranteed to be visible in the
3123 * future iterations and will stay visible until the last reference is put.
3124 * A css which hasn't finished ->css_online() or already finished
3125 * ->css_offline() may show up during traversal. It's each subsystem's
3126 * responsibility to synchronize against on/offlining.
3128 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3129 struct cgroup_subsys_state *parent)
3131 struct cgroup_subsys_state *next;
3133 cgroup_assert_mutex_or_rcu_locked();
3136 * @pos could already have been unlinked from the sibling list.
3137 * Once a cgroup is removed, its ->sibling.next is no longer
3138 * updated when its next sibling changes. CSS_RELEASED is set when
3139 * @pos is taken off list, at which time its next pointer is valid,
3140 * and, as releases are serialized, the one pointed to by the next
3141 * pointer is guaranteed to not have started release yet. This
3142 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3143 * critical section, the one pointed to by its next pointer is
3144 * guaranteed to not have finished its RCU grace period even if we
3145 * have dropped rcu_read_lock() inbetween iterations.
3147 * If @pos has CSS_RELEASED set, its next pointer can't be
3148 * dereferenced; however, as each css is given a monotonically
3149 * increasing unique serial number and always appended to the
3150 * sibling list, the next one can be found by walking the parent's
3151 * children until the first css with higher serial number than
3152 * @pos's. While this path can be slower, it happens iff iteration
3153 * races against release and the race window is very small.
3156 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3157 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3158 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3160 list_for_each_entry_rcu(next, &parent->children, sibling)
3161 if (next->serial_nr > pos->serial_nr)
3166 * @next, if not pointing to the head, can be dereferenced and is
3169 if (&next->sibling != &parent->children)
3175 * css_next_descendant_pre - find the next descendant for pre-order walk
3176 * @pos: the current position (%NULL to initiate traversal)
3177 * @root: css whose descendants to walk
3179 * To be used by css_for_each_descendant_pre(). Find the next descendant
3180 * to visit for pre-order traversal of @root's descendants. @root is
3181 * included in the iteration and the first node to be visited.
3183 * While this function requires cgroup_mutex or RCU read locking, it
3184 * doesn't require the whole traversal to be contained in a single critical
3185 * section. This function will return the correct next descendant as long
3186 * as both @pos and @root are accessible and @pos is a descendant of @root.
3188 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3189 * css which finished ->css_online() is guaranteed to be visible in the
3190 * future iterations and will stay visible until the last reference is put.
3191 * A css which hasn't finished ->css_online() or already finished
3192 * ->css_offline() may show up during traversal. It's each subsystem's
3193 * responsibility to synchronize against on/offlining.
3195 struct cgroup_subsys_state *
3196 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3197 struct cgroup_subsys_state *root)
3199 struct cgroup_subsys_state *next;
3201 cgroup_assert_mutex_or_rcu_locked();
3203 /* if first iteration, visit @root */
3207 /* visit the first child if exists */
3208 next = css_next_child(NULL, pos);
3212 /* no child, visit my or the closest ancestor's next sibling */
3213 while (pos != root) {
3214 next = css_next_child(pos, pos->parent);
3224 * css_rightmost_descendant - return the rightmost descendant of a css
3225 * @pos: css of interest
3227 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3228 * is returned. This can be used during pre-order traversal to skip
3231 * While this function requires cgroup_mutex or RCU read locking, it
3232 * doesn't require the whole traversal to be contained in a single critical
3233 * section. This function will return the correct rightmost descendant as
3234 * long as @pos is accessible.
3236 struct cgroup_subsys_state *
3237 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3239 struct cgroup_subsys_state *last, *tmp;
3241 cgroup_assert_mutex_or_rcu_locked();
3245 /* ->prev isn't RCU safe, walk ->next till the end */
3247 css_for_each_child(tmp, last)
3254 static struct cgroup_subsys_state *
3255 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3257 struct cgroup_subsys_state *last;
3261 pos = css_next_child(NULL, pos);
3268 * css_next_descendant_post - find the next descendant for post-order walk
3269 * @pos: the current position (%NULL to initiate traversal)
3270 * @root: css whose descendants to walk
3272 * To be used by css_for_each_descendant_post(). Find the next descendant
3273 * to visit for post-order traversal of @root's descendants. @root is
3274 * included in the iteration and the last node to be visited.
3276 * While this function requires cgroup_mutex or RCU read locking, it
3277 * doesn't require the whole traversal to be contained in a single critical
3278 * section. This function will return the correct next descendant as long
3279 * as both @pos and @cgroup are accessible and @pos is a descendant of
3282 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3283 * css which finished ->css_online() is guaranteed to be visible in the
3284 * future iterations and will stay visible until the last reference is put.
3285 * A css which hasn't finished ->css_online() or already finished
3286 * ->css_offline() may show up during traversal. It's each subsystem's
3287 * responsibility to synchronize against on/offlining.
3289 struct cgroup_subsys_state *
3290 css_next_descendant_post(struct cgroup_subsys_state *pos,
3291 struct cgroup_subsys_state *root)
3293 struct cgroup_subsys_state *next;
3295 cgroup_assert_mutex_or_rcu_locked();
3297 /* if first iteration, visit leftmost descendant which may be @root */
3299 return css_leftmost_descendant(root);
3301 /* if we visited @root, we're done */
3305 /* if there's an unvisited sibling, visit its leftmost descendant */
3306 next = css_next_child(pos, pos->parent);
3308 return css_leftmost_descendant(next);
3310 /* no sibling left, visit parent */
3315 * css_has_online_children - does a css have online children
3316 * @css: the target css
3318 * Returns %true if @css has any online children; otherwise, %false. This
3319 * function can be called from any context but the caller is responsible
3320 * for synchronizing against on/offlining as necessary.
3322 bool css_has_online_children(struct cgroup_subsys_state *css)
3324 struct cgroup_subsys_state *child;
3328 css_for_each_child(child, css) {
3329 if (css->flags & CSS_ONLINE) {
3339 * css_advance_task_iter - advance a task itererator to the next css_set
3340 * @it: the iterator to advance
3342 * Advance @it to the next css_set to walk.
3344 static void css_advance_task_iter(struct css_task_iter *it)
3346 struct list_head *l = it->cset_pos;
3347 struct cgrp_cset_link *link;
3348 struct css_set *cset;
3350 /* Advance to the next non-empty css_set */
3353 if (l == it->cset_head) {
3354 it->cset_pos = NULL;
3359 cset = container_of(l, struct css_set,
3360 e_cset_node[it->ss->id]);
3362 link = list_entry(l, struct cgrp_cset_link, cset_link);
3365 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3369 if (!list_empty(&cset->tasks))
3370 it->task_pos = cset->tasks.next;
3372 it->task_pos = cset->mg_tasks.next;
3374 it->tasks_head = &cset->tasks;
3375 it->mg_tasks_head = &cset->mg_tasks;
3379 * css_task_iter_start - initiate task iteration
3380 * @css: the css to walk tasks of
3381 * @it: the task iterator to use
3383 * Initiate iteration through the tasks of @css. The caller can call
3384 * css_task_iter_next() to walk through the tasks until the function
3385 * returns NULL. On completion of iteration, css_task_iter_end() must be
3388 * Note that this function acquires a lock which is released when the
3389 * iteration finishes. The caller can't sleep while iteration is in
3392 void css_task_iter_start(struct cgroup_subsys_state *css,
3393 struct css_task_iter *it)
3394 __acquires(css_set_rwsem)
3396 /* no one should try to iterate before mounting cgroups */
3397 WARN_ON_ONCE(!use_task_css_set_links);
3399 down_read(&css_set_rwsem);
3404 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3406 it->cset_pos = &css->cgroup->cset_links;
3408 it->cset_head = it->cset_pos;
3410 css_advance_task_iter(it);
3414 * css_task_iter_next - return the next task for the iterator
3415 * @it: the task iterator being iterated
3417 * The "next" function for task iteration. @it should have been
3418 * initialized via css_task_iter_start(). Returns NULL when the iteration
3421 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3423 struct task_struct *res;
3424 struct list_head *l = it->task_pos;
3426 /* If the iterator cg is NULL, we have no tasks */
3429 res = list_entry(l, struct task_struct, cg_list);
3432 * Advance iterator to find next entry. cset->tasks is consumed
3433 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3438 if (l == it->tasks_head)
3439 l = it->mg_tasks_head->next;
3441 if (l == it->mg_tasks_head)
3442 css_advance_task_iter(it);
3450 * css_task_iter_end - finish task iteration
3451 * @it: the task iterator to finish
3453 * Finish task iteration started by css_task_iter_start().
3455 void css_task_iter_end(struct css_task_iter *it)
3456 __releases(css_set_rwsem)
3458 up_read(&css_set_rwsem);
3462 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3463 * @to: cgroup to which the tasks will be moved
3464 * @from: cgroup in which the tasks currently reside
3466 * Locking rules between cgroup_post_fork() and the migration path
3467 * guarantee that, if a task is forking while being migrated, the new child
3468 * is guaranteed to be either visible in the source cgroup after the
3469 * parent's migration is complete or put into the target cgroup. No task
3470 * can slip out of migration through forking.
3472 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3474 LIST_HEAD(preloaded_csets);
3475 struct cgrp_cset_link *link;
3476 struct css_task_iter it;
3477 struct task_struct *task;
3480 mutex_lock(&cgroup_mutex);
3482 /* all tasks in @from are being moved, all csets are source */
3483 down_read(&css_set_rwsem);
3484 list_for_each_entry(link, &from->cset_links, cset_link)
3485 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3486 up_read(&css_set_rwsem);
3488 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3493 * Migrate tasks one-by-one until @form is empty. This fails iff
3494 * ->can_attach() fails.
3497 css_task_iter_start(&from->self, &it);
3498 task = css_task_iter_next(&it);
3500 get_task_struct(task);
3501 css_task_iter_end(&it);
3504 ret = cgroup_migrate(to, task, false);
3505 put_task_struct(task);
3507 } while (task && !ret);
3509 cgroup_migrate_finish(&preloaded_csets);
3510 mutex_unlock(&cgroup_mutex);
3515 * Stuff for reading the 'tasks'/'procs' files.
3517 * Reading this file can return large amounts of data if a cgroup has
3518 * *lots* of attached tasks. So it may need several calls to read(),
3519 * but we cannot guarantee that the information we produce is correct
3520 * unless we produce it entirely atomically.
3524 /* which pidlist file are we talking about? */
3525 enum cgroup_filetype {
3531 * A pidlist is a list of pids that virtually represents the contents of one
3532 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3533 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3536 struct cgroup_pidlist {
3538 * used to find which pidlist is wanted. doesn't change as long as
3539 * this particular list stays in the list.
3541 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3544 /* how many elements the above list has */
3546 /* each of these stored in a list by its cgroup */
3547 struct list_head links;
3548 /* pointer to the cgroup we belong to, for list removal purposes */
3549 struct cgroup *owner;
3550 /* for delayed destruction */
3551 struct delayed_work destroy_dwork;
3555 * The following two functions "fix" the issue where there are more pids
3556 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3557 * TODO: replace with a kernel-wide solution to this problem
3559 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3560 static void *pidlist_allocate(int count)
3562 if (PIDLIST_TOO_LARGE(count))
3563 return vmalloc(count * sizeof(pid_t));
3565 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3568 static void pidlist_free(void *p)
3570 if (is_vmalloc_addr(p))
3577 * Used to destroy all pidlists lingering waiting for destroy timer. None
3578 * should be left afterwards.
3580 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3582 struct cgroup_pidlist *l, *tmp_l;
3584 mutex_lock(&cgrp->pidlist_mutex);
3585 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3586 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3587 mutex_unlock(&cgrp->pidlist_mutex);
3589 flush_workqueue(cgroup_pidlist_destroy_wq);
3590 BUG_ON(!list_empty(&cgrp->pidlists));
3593 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3595 struct delayed_work *dwork = to_delayed_work(work);
3596 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3598 struct cgroup_pidlist *tofree = NULL;
3600 mutex_lock(&l->owner->pidlist_mutex);
3603 * Destroy iff we didn't get queued again. The state won't change
3604 * as destroy_dwork can only be queued while locked.
3606 if (!delayed_work_pending(dwork)) {
3607 list_del(&l->links);
3608 pidlist_free(l->list);
3609 put_pid_ns(l->key.ns);
3613 mutex_unlock(&l->owner->pidlist_mutex);
3618 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3619 * Returns the number of unique elements.
3621 static int pidlist_uniq(pid_t *list, int length)
3626 * we presume the 0th element is unique, so i starts at 1. trivial
3627 * edge cases first; no work needs to be done for either
3629 if (length == 0 || length == 1)
3631 /* src and dest walk down the list; dest counts unique elements */
3632 for (src = 1; src < length; src++) {
3633 /* find next unique element */
3634 while (list[src] == list[src-1]) {
3639 /* dest always points to where the next unique element goes */
3640 list[dest] = list[src];
3648 * The two pid files - task and cgroup.procs - guaranteed that the result
3649 * is sorted, which forced this whole pidlist fiasco. As pid order is
3650 * different per namespace, each namespace needs differently sorted list,
3651 * making it impossible to use, for example, single rbtree of member tasks
3652 * sorted by task pointer. As pidlists can be fairly large, allocating one
3653 * per open file is dangerous, so cgroup had to implement shared pool of
3654 * pidlists keyed by cgroup and namespace.
3656 * All this extra complexity was caused by the original implementation
3657 * committing to an entirely unnecessary property. In the long term, we
3658 * want to do away with it. Explicitly scramble sort order if
3659 * sane_behavior so that no such expectation exists in the new interface.
3661 * Scrambling is done by swapping every two consecutive bits, which is
3662 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3664 static pid_t pid_fry(pid_t pid)
3666 unsigned a = pid & 0x55555555;
3667 unsigned b = pid & 0xAAAAAAAA;
3669 return (a << 1) | (b >> 1);
3672 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3674 if (cgroup_sane_behavior(cgrp))
3675 return pid_fry(pid);
3680 static int cmppid(const void *a, const void *b)
3682 return *(pid_t *)a - *(pid_t *)b;
3685 static int fried_cmppid(const void *a, const void *b)
3687 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3690 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3691 enum cgroup_filetype type)
3693 struct cgroup_pidlist *l;
3694 /* don't need task_nsproxy() if we're looking at ourself */
3695 struct pid_namespace *ns = task_active_pid_ns(current);
3697 lockdep_assert_held(&cgrp->pidlist_mutex);
3699 list_for_each_entry(l, &cgrp->pidlists, links)
3700 if (l->key.type == type && l->key.ns == ns)
3706 * find the appropriate pidlist for our purpose (given procs vs tasks)
3707 * returns with the lock on that pidlist already held, and takes care
3708 * of the use count, or returns NULL with no locks held if we're out of
3711 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3712 enum cgroup_filetype type)
3714 struct cgroup_pidlist *l;
3716 lockdep_assert_held(&cgrp->pidlist_mutex);
3718 l = cgroup_pidlist_find(cgrp, type);
3722 /* entry not found; create a new one */
3723 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3727 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3729 /* don't need task_nsproxy() if we're looking at ourself */
3730 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3732 list_add(&l->links, &cgrp->pidlists);
3737 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3739 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3740 struct cgroup_pidlist **lp)
3744 int pid, n = 0; /* used for populating the array */
3745 struct css_task_iter it;
3746 struct task_struct *tsk;
3747 struct cgroup_pidlist *l;
3749 lockdep_assert_held(&cgrp->pidlist_mutex);
3752 * If cgroup gets more users after we read count, we won't have
3753 * enough space - tough. This race is indistinguishable to the
3754 * caller from the case that the additional cgroup users didn't
3755 * show up until sometime later on.
3757 length = cgroup_task_count(cgrp);
3758 array = pidlist_allocate(length);
3761 /* now, populate the array */
3762 css_task_iter_start(&cgrp->self, &it);
3763 while ((tsk = css_task_iter_next(&it))) {
3764 if (unlikely(n == length))
3766 /* get tgid or pid for procs or tasks file respectively */
3767 if (type == CGROUP_FILE_PROCS)
3768 pid = task_tgid_vnr(tsk);
3770 pid = task_pid_vnr(tsk);
3771 if (pid > 0) /* make sure to only use valid results */
3774 css_task_iter_end(&it);
3776 /* now sort & (if procs) strip out duplicates */
3777 if (cgroup_sane_behavior(cgrp))
3778 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3780 sort(array, length, sizeof(pid_t), cmppid, NULL);
3781 if (type == CGROUP_FILE_PROCS)
3782 length = pidlist_uniq(array, length);
3784 l = cgroup_pidlist_find_create(cgrp, type);
3786 mutex_unlock(&cgrp->pidlist_mutex);
3787 pidlist_free(array);
3791 /* store array, freeing old if necessary */
3792 pidlist_free(l->list);
3800 * cgroupstats_build - build and fill cgroupstats
3801 * @stats: cgroupstats to fill information into
3802 * @dentry: A dentry entry belonging to the cgroup for which stats have
3805 * Build and fill cgroupstats so that taskstats can export it to user
3808 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3810 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3811 struct cgroup *cgrp;
3812 struct css_task_iter it;
3813 struct task_struct *tsk;
3815 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3816 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3817 kernfs_type(kn) != KERNFS_DIR)
3820 mutex_lock(&cgroup_mutex);
3823 * We aren't being called from kernfs and there's no guarantee on
3824 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3825 * @kn->priv is RCU safe. Let's do the RCU dancing.
3828 cgrp = rcu_dereference(kn->priv);
3829 if (!cgrp || cgroup_is_dead(cgrp)) {
3831 mutex_unlock(&cgroup_mutex);
3836 css_task_iter_start(&cgrp->self, &it);
3837 while ((tsk = css_task_iter_next(&it))) {
3838 switch (tsk->state) {
3840 stats->nr_running++;
3842 case TASK_INTERRUPTIBLE:
3843 stats->nr_sleeping++;
3845 case TASK_UNINTERRUPTIBLE:
3846 stats->nr_uninterruptible++;
3849 stats->nr_stopped++;
3852 if (delayacct_is_task_waiting_on_io(tsk))
3853 stats->nr_io_wait++;
3857 css_task_iter_end(&it);
3859 mutex_unlock(&cgroup_mutex);
3865 * seq_file methods for the tasks/procs files. The seq_file position is the
3866 * next pid to display; the seq_file iterator is a pointer to the pid
3867 * in the cgroup->l->list array.
3870 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3873 * Initially we receive a position value that corresponds to
3874 * one more than the last pid shown (or 0 on the first call or
3875 * after a seek to the start). Use a binary-search to find the
3876 * next pid to display, if any
3878 struct kernfs_open_file *of = s->private;
3879 struct cgroup *cgrp = seq_css(s)->cgroup;
3880 struct cgroup_pidlist *l;
3881 enum cgroup_filetype type = seq_cft(s)->private;
3882 int index = 0, pid = *pos;
3885 mutex_lock(&cgrp->pidlist_mutex);
3888 * !NULL @of->priv indicates that this isn't the first start()
3889 * after open. If the matching pidlist is around, we can use that.
3890 * Look for it. Note that @of->priv can't be used directly. It
3891 * could already have been destroyed.
3894 of->priv = cgroup_pidlist_find(cgrp, type);
3897 * Either this is the first start() after open or the matching
3898 * pidlist has been destroyed inbetween. Create a new one.
3901 ret = pidlist_array_load(cgrp, type,
3902 (struct cgroup_pidlist **)&of->priv);
3904 return ERR_PTR(ret);
3909 int end = l->length;
3911 while (index < end) {
3912 int mid = (index + end) / 2;
3913 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3916 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3922 /* If we're off the end of the array, we're done */
3923 if (index >= l->length)
3925 /* Update the abstract position to be the actual pid that we found */
3926 iter = l->list + index;
3927 *pos = cgroup_pid_fry(cgrp, *iter);
3931 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3933 struct kernfs_open_file *of = s->private;
3934 struct cgroup_pidlist *l = of->priv;
3937 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3938 CGROUP_PIDLIST_DESTROY_DELAY);
3939 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3942 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3944 struct kernfs_open_file *of = s->private;
3945 struct cgroup_pidlist *l = of->priv;
3947 pid_t *end = l->list + l->length;
3949 * Advance to the next pid in the array. If this goes off the
3956 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3961 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3963 return seq_printf(s, "%d\n", *(int *)v);
3966 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3969 return notify_on_release(css->cgroup);
3972 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3973 struct cftype *cft, u64 val)
3975 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3977 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3979 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3983 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3986 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3989 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3990 struct cftype *cft, u64 val)
3993 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3995 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3999 static struct cftype cgroup_base_files[] = {
4001 .name = "cgroup.procs",
4002 .seq_start = cgroup_pidlist_start,
4003 .seq_next = cgroup_pidlist_next,
4004 .seq_stop = cgroup_pidlist_stop,
4005 .seq_show = cgroup_pidlist_show,
4006 .private = CGROUP_FILE_PROCS,
4007 .write = cgroup_procs_write,
4008 .mode = S_IRUGO | S_IWUSR,
4011 .name = "cgroup.clone_children",
4012 .flags = CFTYPE_INSANE,
4013 .read_u64 = cgroup_clone_children_read,
4014 .write_u64 = cgroup_clone_children_write,
4017 .name = "cgroup.sane_behavior",
4018 .flags = CFTYPE_ONLY_ON_ROOT,
4019 .seq_show = cgroup_sane_behavior_show,
4022 .name = "cgroup.controllers",
4023 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_ONLY_ON_ROOT,
4024 .seq_show = cgroup_root_controllers_show,
4027 .name = "cgroup.controllers",
4028 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4029 .seq_show = cgroup_controllers_show,
4032 .name = "cgroup.subtree_control",
4033 .flags = CFTYPE_ONLY_ON_DFL,
4034 .seq_show = cgroup_subtree_control_show,
4035 .write = cgroup_subtree_control_write,
4038 .name = "cgroup.populated",
4039 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4040 .seq_show = cgroup_populated_show,
4044 * Historical crazy stuff. These don't have "cgroup." prefix and
4045 * don't exist if sane_behavior. If you're depending on these, be
4046 * prepared to be burned.
4050 .flags = CFTYPE_INSANE, /* use "procs" instead */
4051 .seq_start = cgroup_pidlist_start,
4052 .seq_next = cgroup_pidlist_next,
4053 .seq_stop = cgroup_pidlist_stop,
4054 .seq_show = cgroup_pidlist_show,
4055 .private = CGROUP_FILE_TASKS,
4056 .write = cgroup_tasks_write,
4057 .mode = S_IRUGO | S_IWUSR,
4060 .name = "notify_on_release",
4061 .flags = CFTYPE_INSANE,
4062 .read_u64 = cgroup_read_notify_on_release,
4063 .write_u64 = cgroup_write_notify_on_release,
4066 .name = "release_agent",
4067 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4068 .seq_show = cgroup_release_agent_show,
4069 .write = cgroup_release_agent_write,
4070 .max_write_len = PATH_MAX - 1,
4076 * cgroup_populate_dir - create subsys files in a cgroup directory
4077 * @cgrp: target cgroup
4078 * @subsys_mask: mask of the subsystem ids whose files should be added
4080 * On failure, no file is added.
4082 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4084 struct cgroup_subsys *ss;
4087 /* process cftsets of each subsystem */
4088 for_each_subsys(ss, i) {
4089 struct cftype *cfts;
4091 if (!(subsys_mask & (1 << i)))
4094 list_for_each_entry(cfts, &ss->cfts, node) {
4095 ret = cgroup_addrm_files(cgrp, cfts, true);
4102 cgroup_clear_dir(cgrp, subsys_mask);
4107 * css destruction is four-stage process.
4109 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4110 * Implemented in kill_css().
4112 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4113 * and thus css_tryget_online() is guaranteed to fail, the css can be
4114 * offlined by invoking offline_css(). After offlining, the base ref is
4115 * put. Implemented in css_killed_work_fn().
4117 * 3. When the percpu_ref reaches zero, the only possible remaining
4118 * accessors are inside RCU read sections. css_release() schedules the
4121 * 4. After the grace period, the css can be freed. Implemented in
4122 * css_free_work_fn().
4124 * It is actually hairier because both step 2 and 4 require process context
4125 * and thus involve punting to css->destroy_work adding two additional
4126 * steps to the already complex sequence.
4128 static void css_free_work_fn(struct work_struct *work)
4130 struct cgroup_subsys_state *css =
4131 container_of(work, struct cgroup_subsys_state, destroy_work);
4132 struct cgroup *cgrp = css->cgroup;
4137 css_put(css->parent);
4139 css->ss->css_free(css);
4142 /* cgroup free path */
4143 atomic_dec(&cgrp->root->nr_cgrps);
4144 cgroup_pidlist_destroy_all(cgrp);
4146 if (cgroup_parent(cgrp)) {
4148 * We get a ref to the parent, and put the ref when
4149 * this cgroup is being freed, so it's guaranteed
4150 * that the parent won't be destroyed before its
4153 cgroup_put(cgroup_parent(cgrp));
4154 kernfs_put(cgrp->kn);
4158 * This is root cgroup's refcnt reaching zero,
4159 * which indicates that the root should be
4162 cgroup_destroy_root(cgrp->root);
4167 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4169 struct cgroup_subsys_state *css =
4170 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4172 INIT_WORK(&css->destroy_work, css_free_work_fn);
4173 queue_work(cgroup_destroy_wq, &css->destroy_work);
4176 static void css_release_work_fn(struct work_struct *work)
4178 struct cgroup_subsys_state *css =
4179 container_of(work, struct cgroup_subsys_state, destroy_work);
4180 struct cgroup_subsys *ss = css->ss;
4181 struct cgroup *cgrp = css->cgroup;
4183 mutex_lock(&cgroup_mutex);
4185 css->flags |= CSS_RELEASED;
4186 list_del_rcu(&css->sibling);
4189 /* css release path */
4190 cgroup_idr_remove(&ss->css_idr, css->id);
4192 /* cgroup release path */
4193 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4197 mutex_unlock(&cgroup_mutex);
4199 call_rcu(&css->rcu_head, css_free_rcu_fn);
4202 static void css_release(struct percpu_ref *ref)
4204 struct cgroup_subsys_state *css =
4205 container_of(ref, struct cgroup_subsys_state, refcnt);
4207 INIT_WORK(&css->destroy_work, css_release_work_fn);
4208 queue_work(cgroup_destroy_wq, &css->destroy_work);
4211 static void init_and_link_css(struct cgroup_subsys_state *css,
4212 struct cgroup_subsys *ss, struct cgroup *cgrp)
4214 lockdep_assert_held(&cgroup_mutex);
4218 memset(css, 0, sizeof(*css));
4221 INIT_LIST_HEAD(&css->sibling);
4222 INIT_LIST_HEAD(&css->children);
4223 css->serial_nr = css_serial_nr_next++;
4225 if (cgroup_parent(cgrp)) {
4226 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4227 css_get(css->parent);
4230 BUG_ON(cgroup_css(cgrp, ss));
4233 /* invoke ->css_online() on a new CSS and mark it online if successful */
4234 static int online_css(struct cgroup_subsys_state *css)
4236 struct cgroup_subsys *ss = css->ss;
4239 lockdep_assert_held(&cgroup_mutex);
4242 ret = ss->css_online(css);
4244 css->flags |= CSS_ONLINE;
4245 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4250 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4251 static void offline_css(struct cgroup_subsys_state *css)
4253 struct cgroup_subsys *ss = css->ss;
4255 lockdep_assert_held(&cgroup_mutex);
4257 if (!(css->flags & CSS_ONLINE))
4260 if (ss->css_offline)
4261 ss->css_offline(css);
4263 css->flags &= ~CSS_ONLINE;
4264 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4266 wake_up_all(&css->cgroup->offline_waitq);
4270 * create_css - create a cgroup_subsys_state
4271 * @cgrp: the cgroup new css will be associated with
4272 * @ss: the subsys of new css
4274 * Create a new css associated with @cgrp - @ss pair. On success, the new
4275 * css is online and installed in @cgrp with all interface files created.
4276 * Returns 0 on success, -errno on failure.
4278 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4280 struct cgroup *parent = cgroup_parent(cgrp);
4281 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4282 struct cgroup_subsys_state *css;
4285 lockdep_assert_held(&cgroup_mutex);
4287 css = ss->css_alloc(parent_css);
4289 return PTR_ERR(css);
4291 init_and_link_css(css, ss, cgrp);
4293 err = percpu_ref_init(&css->refcnt, css_release);
4297 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4299 goto err_free_percpu_ref;
4302 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4306 /* @css is ready to be brought online now, make it visible */
4307 list_add_tail_rcu(&css->sibling, &parent_css->children);
4308 cgroup_idr_replace(&ss->css_idr, css, css->id);
4310 err = online_css(css);
4314 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4315 cgroup_parent(parent)) {
4316 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4317 current->comm, current->pid, ss->name);
4318 if (!strcmp(ss->name, "memory"))
4319 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4320 ss->warned_broken_hierarchy = true;
4326 list_del_rcu(&css->sibling);
4327 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4329 cgroup_idr_remove(&ss->css_idr, css->id);
4330 err_free_percpu_ref:
4331 percpu_ref_cancel_init(&css->refcnt);
4333 call_rcu(&css->rcu_head, css_free_rcu_fn);
4337 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4340 struct cgroup *parent, *cgrp;
4341 struct cgroup_root *root;
4342 struct cgroup_subsys *ss;
4343 struct kernfs_node *kn;
4346 parent = cgroup_kn_lock_live(parent_kn);
4349 root = parent->root;
4351 /* allocate the cgroup and its ID, 0 is reserved for the root */
4352 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4358 ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
4363 * Temporarily set the pointer to NULL, so idr_find() won't return
4364 * a half-baked cgroup.
4366 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4369 goto out_cancel_ref;
4372 init_cgroup_housekeeping(cgrp);
4374 cgrp->self.parent = &parent->self;
4377 if (notify_on_release(parent))
4378 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4380 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4381 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4383 /* create the directory */
4384 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4392 * This extra ref will be put in cgroup_free_fn() and guarantees
4393 * that @cgrp->kn is always accessible.
4397 cgrp->self.serial_nr = css_serial_nr_next++;
4399 /* allocation complete, commit to creation */
4400 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4401 atomic_inc(&root->nr_cgrps);
4405 * @cgrp is now fully operational. If something fails after this
4406 * point, it'll be released via the normal destruction path.
4408 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4410 ret = cgroup_kn_set_ugid(kn);
4414 ret = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4418 /* let's create and online css's */
4419 for_each_subsys(ss, ssid) {
4420 if (parent->child_subsys_mask & (1 << ssid)) {
4421 ret = create_css(cgrp, ss);
4428 * On the default hierarchy, a child doesn't automatically inherit
4429 * child_subsys_mask from the parent. Each is configured manually.
4431 if (!cgroup_on_dfl(cgrp))
4432 cgrp->child_subsys_mask = parent->child_subsys_mask;
4434 kernfs_activate(kn);
4440 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4442 percpu_ref_cancel_init(&cgrp->self.refcnt);
4446 cgroup_kn_unlock(parent_kn);
4450 cgroup_destroy_locked(cgrp);
4455 * This is called when the refcnt of a css is confirmed to be killed.
4456 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4457 * initate destruction and put the css ref from kill_css().
4459 static void css_killed_work_fn(struct work_struct *work)
4461 struct cgroup_subsys_state *css =
4462 container_of(work, struct cgroup_subsys_state, destroy_work);
4464 mutex_lock(&cgroup_mutex);
4466 mutex_unlock(&cgroup_mutex);
4471 /* css kill confirmation processing requires process context, bounce */
4472 static void css_killed_ref_fn(struct percpu_ref *ref)
4474 struct cgroup_subsys_state *css =
4475 container_of(ref, struct cgroup_subsys_state, refcnt);
4477 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4478 queue_work(cgroup_destroy_wq, &css->destroy_work);
4482 * kill_css - destroy a css
4483 * @css: css to destroy
4485 * This function initiates destruction of @css by removing cgroup interface
4486 * files and putting its base reference. ->css_offline() will be invoked
4487 * asynchronously once css_tryget_online() is guaranteed to fail and when
4488 * the reference count reaches zero, @css will be released.
4490 static void kill_css(struct cgroup_subsys_state *css)
4492 lockdep_assert_held(&cgroup_mutex);
4495 * This must happen before css is disassociated with its cgroup.
4496 * See seq_css() for details.
4498 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4501 * Killing would put the base ref, but we need to keep it alive
4502 * until after ->css_offline().
4507 * cgroup core guarantees that, by the time ->css_offline() is
4508 * invoked, no new css reference will be given out via
4509 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4510 * proceed to offlining css's because percpu_ref_kill() doesn't
4511 * guarantee that the ref is seen as killed on all CPUs on return.
4513 * Use percpu_ref_kill_and_confirm() to get notifications as each
4514 * css is confirmed to be seen as killed on all CPUs.
4516 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4520 * cgroup_destroy_locked - the first stage of cgroup destruction
4521 * @cgrp: cgroup to be destroyed
4523 * css's make use of percpu refcnts whose killing latency shouldn't be
4524 * exposed to userland and are RCU protected. Also, cgroup core needs to
4525 * guarantee that css_tryget_online() won't succeed by the time
4526 * ->css_offline() is invoked. To satisfy all the requirements,
4527 * destruction is implemented in the following two steps.
4529 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4530 * userland visible parts and start killing the percpu refcnts of
4531 * css's. Set up so that the next stage will be kicked off once all
4532 * the percpu refcnts are confirmed to be killed.
4534 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4535 * rest of destruction. Once all cgroup references are gone, the
4536 * cgroup is RCU-freed.
4538 * This function implements s1. After this step, @cgrp is gone as far as
4539 * the userland is concerned and a new cgroup with the same name may be
4540 * created. As cgroup doesn't care about the names internally, this
4541 * doesn't cause any problem.
4543 static int cgroup_destroy_locked(struct cgroup *cgrp)
4544 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4546 struct cgroup_subsys_state *css;
4550 lockdep_assert_held(&cgroup_mutex);
4553 * css_set_rwsem synchronizes access to ->cset_links and prevents
4554 * @cgrp from being removed while put_css_set() is in progress.
4556 down_read(&css_set_rwsem);
4557 empty = list_empty(&cgrp->cset_links);
4558 up_read(&css_set_rwsem);
4563 * Make sure there's no live children. We can't test emptiness of
4564 * ->self.children as dead children linger on it while being
4565 * drained; otherwise, "rmdir parent/child parent" may fail.
4567 if (css_has_online_children(&cgrp->self))
4571 * Mark @cgrp dead. This prevents further task migration and child
4572 * creation by disabling cgroup_lock_live_group().
4574 cgrp->self.flags &= ~CSS_ONLINE;
4576 /* initiate massacre of all css's */
4577 for_each_css(css, ssid, cgrp)
4580 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4581 raw_spin_lock(&release_list_lock);
4582 if (!list_empty(&cgrp->release_list))
4583 list_del_init(&cgrp->release_list);
4584 raw_spin_unlock(&release_list_lock);
4587 * Remove @cgrp directory along with the base files. @cgrp has an
4588 * extra ref on its kn.
4590 kernfs_remove(cgrp->kn);
4592 set_bit(CGRP_RELEASABLE, &cgroup_parent(cgrp)->flags);
4593 check_for_release(cgroup_parent(cgrp));
4595 /* put the base reference */
4596 percpu_ref_kill(&cgrp->self.refcnt);
4601 static int cgroup_rmdir(struct kernfs_node *kn)
4603 struct cgroup *cgrp;
4606 cgrp = cgroup_kn_lock_live(kn);
4609 cgroup_get(cgrp); /* for @kn->priv clearing */
4611 ret = cgroup_destroy_locked(cgrp);
4613 cgroup_kn_unlock(kn);
4616 * There are two control paths which try to determine cgroup from
4617 * dentry without going through kernfs - cgroupstats_build() and
4618 * css_tryget_online_from_dir(). Those are supported by RCU
4619 * protecting clearing of cgrp->kn->priv backpointer, which should
4620 * happen after all files under it have been removed.
4623 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4629 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4630 .remount_fs = cgroup_remount,
4631 .show_options = cgroup_show_options,
4632 .mkdir = cgroup_mkdir,
4633 .rmdir = cgroup_rmdir,
4634 .rename = cgroup_rename,
4637 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4639 struct cgroup_subsys_state *css;
4641 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4643 mutex_lock(&cgroup_mutex);
4645 idr_init(&ss->css_idr);
4646 INIT_LIST_HEAD(&ss->cfts);
4648 /* Create the root cgroup state for this subsystem */
4649 ss->root = &cgrp_dfl_root;
4650 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4651 /* We don't handle early failures gracefully */
4652 BUG_ON(IS_ERR(css));
4653 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4656 * Root csses are never destroyed and we can't initialize
4657 * percpu_ref during early init. Disable refcnting.
4659 css->flags |= CSS_NO_REF;
4662 /* allocation can't be done safely during early init */
4665 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4666 BUG_ON(css->id < 0);
4669 /* Update the init_css_set to contain a subsys
4670 * pointer to this state - since the subsystem is
4671 * newly registered, all tasks and hence the
4672 * init_css_set is in the subsystem's root cgroup. */
4673 init_css_set.subsys[ss->id] = css;
4675 need_forkexit_callback |= ss->fork || ss->exit;
4677 /* At system boot, before all subsystems have been
4678 * registered, no tasks have been forked, so we don't
4679 * need to invoke fork callbacks here. */
4680 BUG_ON(!list_empty(&init_task.tasks));
4682 BUG_ON(online_css(css));
4684 mutex_unlock(&cgroup_mutex);
4688 * cgroup_init_early - cgroup initialization at system boot
4690 * Initialize cgroups at system boot, and initialize any
4691 * subsystems that request early init.
4693 int __init cgroup_init_early(void)
4695 static struct cgroup_sb_opts __initdata opts =
4696 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4697 struct cgroup_subsys *ss;
4700 init_cgroup_root(&cgrp_dfl_root, &opts);
4701 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4703 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4705 for_each_subsys(ss, i) {
4706 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4707 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4708 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4710 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4711 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4714 ss->name = cgroup_subsys_name[i];
4717 cgroup_init_subsys(ss, true);
4723 * cgroup_init - cgroup initialization
4725 * Register cgroup filesystem and /proc file, and initialize
4726 * any subsystems that didn't request early init.
4728 int __init cgroup_init(void)
4730 struct cgroup_subsys *ss;
4734 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4736 mutex_lock(&cgroup_mutex);
4738 /* Add init_css_set to the hash table */
4739 key = css_set_hash(init_css_set.subsys);
4740 hash_add(css_set_table, &init_css_set.hlist, key);
4742 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4744 mutex_unlock(&cgroup_mutex);
4746 for_each_subsys(ss, ssid) {
4747 if (ss->early_init) {
4748 struct cgroup_subsys_state *css =
4749 init_css_set.subsys[ss->id];
4751 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4753 BUG_ON(css->id < 0);
4755 cgroup_init_subsys(ss, false);
4758 list_add_tail(&init_css_set.e_cset_node[ssid],
4759 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4762 * Setting dfl_root subsys_mask needs to consider the
4763 * disabled flag and cftype registration needs kmalloc,
4764 * both of which aren't available during early_init.
4766 if (!ss->disabled) {
4767 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4768 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4772 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4776 err = register_filesystem(&cgroup_fs_type);
4778 kobject_put(cgroup_kobj);
4782 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4786 static int __init cgroup_wq_init(void)
4789 * There isn't much point in executing destruction path in
4790 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4791 * Use 1 for @max_active.
4793 * We would prefer to do this in cgroup_init() above, but that
4794 * is called before init_workqueues(): so leave this until after.
4796 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4797 BUG_ON(!cgroup_destroy_wq);
4800 * Used to destroy pidlists and separate to serve as flush domain.
4801 * Cap @max_active to 1 too.
4803 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4805 BUG_ON(!cgroup_pidlist_destroy_wq);
4809 core_initcall(cgroup_wq_init);
4812 * proc_cgroup_show()
4813 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4814 * - Used for /proc/<pid>/cgroup.
4817 /* TODO: Use a proper seq_file iterator */
4818 int proc_cgroup_show(struct seq_file *m, void *v)
4821 struct task_struct *tsk;
4824 struct cgroup_root *root;
4827 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4833 tsk = get_pid_task(pid, PIDTYPE_PID);
4839 mutex_lock(&cgroup_mutex);
4840 down_read(&css_set_rwsem);
4842 for_each_root(root) {
4843 struct cgroup_subsys *ss;
4844 struct cgroup *cgrp;
4845 int ssid, count = 0;
4847 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4850 seq_printf(m, "%d:", root->hierarchy_id);
4851 for_each_subsys(ss, ssid)
4852 if (root->subsys_mask & (1 << ssid))
4853 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4854 if (strlen(root->name))
4855 seq_printf(m, "%sname=%s", count ? "," : "",
4858 cgrp = task_cgroup_from_root(tsk, root);
4859 path = cgroup_path(cgrp, buf, PATH_MAX);
4861 retval = -ENAMETOOLONG;
4869 up_read(&css_set_rwsem);
4870 mutex_unlock(&cgroup_mutex);
4871 put_task_struct(tsk);
4878 /* Display information about each subsystem and each hierarchy */
4879 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4881 struct cgroup_subsys *ss;
4884 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4886 * ideally we don't want subsystems moving around while we do this.
4887 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4888 * subsys/hierarchy state.
4890 mutex_lock(&cgroup_mutex);
4892 for_each_subsys(ss, i)
4893 seq_printf(m, "%s\t%d\t%d\t%d\n",
4894 ss->name, ss->root->hierarchy_id,
4895 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4897 mutex_unlock(&cgroup_mutex);
4901 static int cgroupstats_open(struct inode *inode, struct file *file)
4903 return single_open(file, proc_cgroupstats_show, NULL);
4906 static const struct file_operations proc_cgroupstats_operations = {
4907 .open = cgroupstats_open,
4909 .llseek = seq_lseek,
4910 .release = single_release,
4914 * cgroup_fork - initialize cgroup related fields during copy_process()
4915 * @child: pointer to task_struct of forking parent process.
4917 * A task is associated with the init_css_set until cgroup_post_fork()
4918 * attaches it to the parent's css_set. Empty cg_list indicates that
4919 * @child isn't holding reference to its css_set.
4921 void cgroup_fork(struct task_struct *child)
4923 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4924 INIT_LIST_HEAD(&child->cg_list);
4928 * cgroup_post_fork - called on a new task after adding it to the task list
4929 * @child: the task in question
4931 * Adds the task to the list running through its css_set if necessary and
4932 * call the subsystem fork() callbacks. Has to be after the task is
4933 * visible on the task list in case we race with the first call to
4934 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4937 void cgroup_post_fork(struct task_struct *child)
4939 struct cgroup_subsys *ss;
4943 * This may race against cgroup_enable_task_cg_links(). As that
4944 * function sets use_task_css_set_links before grabbing
4945 * tasklist_lock and we just went through tasklist_lock to add
4946 * @child, it's guaranteed that either we see the set
4947 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4948 * @child during its iteration.
4950 * If we won the race, @child is associated with %current's
4951 * css_set. Grabbing css_set_rwsem guarantees both that the
4952 * association is stable, and, on completion of the parent's
4953 * migration, @child is visible in the source of migration or
4954 * already in the destination cgroup. This guarantee is necessary
4955 * when implementing operations which need to migrate all tasks of
4956 * a cgroup to another.
4958 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4959 * will remain in init_css_set. This is safe because all tasks are
4960 * in the init_css_set before cg_links is enabled and there's no
4961 * operation which transfers all tasks out of init_css_set.
4963 if (use_task_css_set_links) {
4964 struct css_set *cset;
4966 down_write(&css_set_rwsem);
4967 cset = task_css_set(current);
4968 if (list_empty(&child->cg_list)) {
4969 rcu_assign_pointer(child->cgroups, cset);
4970 list_add(&child->cg_list, &cset->tasks);
4973 up_write(&css_set_rwsem);
4977 * Call ss->fork(). This must happen after @child is linked on
4978 * css_set; otherwise, @child might change state between ->fork()
4979 * and addition to css_set.
4981 if (need_forkexit_callback) {
4982 for_each_subsys(ss, i)
4989 * cgroup_exit - detach cgroup from exiting task
4990 * @tsk: pointer to task_struct of exiting process
4992 * Description: Detach cgroup from @tsk and release it.
4994 * Note that cgroups marked notify_on_release force every task in
4995 * them to take the global cgroup_mutex mutex when exiting.
4996 * This could impact scaling on very large systems. Be reluctant to
4997 * use notify_on_release cgroups where very high task exit scaling
4998 * is required on large systems.
5000 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5001 * call cgroup_exit() while the task is still competent to handle
5002 * notify_on_release(), then leave the task attached to the root cgroup in
5003 * each hierarchy for the remainder of its exit. No need to bother with
5004 * init_css_set refcnting. init_css_set never goes away and we can't race
5005 * with migration path - PF_EXITING is visible to migration path.
5007 void cgroup_exit(struct task_struct *tsk)
5009 struct cgroup_subsys *ss;
5010 struct css_set *cset;
5011 bool put_cset = false;
5015 * Unlink from @tsk from its css_set. As migration path can't race
5016 * with us, we can check cg_list without grabbing css_set_rwsem.
5018 if (!list_empty(&tsk->cg_list)) {
5019 down_write(&css_set_rwsem);
5020 list_del_init(&tsk->cg_list);
5021 up_write(&css_set_rwsem);
5025 /* Reassign the task to the init_css_set. */
5026 cset = task_css_set(tsk);
5027 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5029 if (need_forkexit_callback) {
5030 /* see cgroup_post_fork() for details */
5031 for_each_subsys(ss, i) {
5033 struct cgroup_subsys_state *old_css = cset->subsys[i];
5034 struct cgroup_subsys_state *css = task_css(tsk, i);
5036 ss->exit(css, old_css, tsk);
5042 put_css_set(cset, true);
5045 static void check_for_release(struct cgroup *cgrp)
5047 if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
5048 !css_has_online_children(&cgrp->self)) {
5050 * Control Group is currently removeable. If it's not
5051 * already queued for a userspace notification, queue
5054 int need_schedule_work = 0;
5056 raw_spin_lock(&release_list_lock);
5057 if (!cgroup_is_dead(cgrp) &&
5058 list_empty(&cgrp->release_list)) {
5059 list_add(&cgrp->release_list, &release_list);
5060 need_schedule_work = 1;
5062 raw_spin_unlock(&release_list_lock);
5063 if (need_schedule_work)
5064 schedule_work(&release_agent_work);
5069 * Notify userspace when a cgroup is released, by running the
5070 * configured release agent with the name of the cgroup (path
5071 * relative to the root of cgroup file system) as the argument.
5073 * Most likely, this user command will try to rmdir this cgroup.
5075 * This races with the possibility that some other task will be
5076 * attached to this cgroup before it is removed, or that some other
5077 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5078 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5079 * unused, and this cgroup will be reprieved from its death sentence,
5080 * to continue to serve a useful existence. Next time it's released,
5081 * we will get notified again, if it still has 'notify_on_release' set.
5083 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5084 * means only wait until the task is successfully execve()'d. The
5085 * separate release agent task is forked by call_usermodehelper(),
5086 * then control in this thread returns here, without waiting for the
5087 * release agent task. We don't bother to wait because the caller of
5088 * this routine has no use for the exit status of the release agent
5089 * task, so no sense holding our caller up for that.
5091 static void cgroup_release_agent(struct work_struct *work)
5093 BUG_ON(work != &release_agent_work);
5094 mutex_lock(&cgroup_mutex);
5095 raw_spin_lock(&release_list_lock);
5096 while (!list_empty(&release_list)) {
5097 char *argv[3], *envp[3];
5099 char *pathbuf = NULL, *agentbuf = NULL, *path;
5100 struct cgroup *cgrp = list_entry(release_list.next,
5103 list_del_init(&cgrp->release_list);
5104 raw_spin_unlock(&release_list_lock);
5105 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5108 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5111 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5116 argv[i++] = agentbuf;
5121 /* minimal command environment */
5122 envp[i++] = "HOME=/";
5123 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5126 /* Drop the lock while we invoke the usermode helper,
5127 * since the exec could involve hitting disk and hence
5128 * be a slow process */
5129 mutex_unlock(&cgroup_mutex);
5130 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5131 mutex_lock(&cgroup_mutex);
5135 raw_spin_lock(&release_list_lock);
5137 raw_spin_unlock(&release_list_lock);
5138 mutex_unlock(&cgroup_mutex);
5141 static int __init cgroup_disable(char *str)
5143 struct cgroup_subsys *ss;
5147 while ((token = strsep(&str, ",")) != NULL) {
5151 for_each_subsys(ss, i) {
5152 if (!strcmp(token, ss->name)) {
5154 printk(KERN_INFO "Disabling %s control group"
5155 " subsystem\n", ss->name);
5162 __setup("cgroup_disable=", cgroup_disable);
5165 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5166 * @dentry: directory dentry of interest
5167 * @ss: subsystem of interest
5169 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5170 * to get the corresponding css and return it. If such css doesn't exist
5171 * or can't be pinned, an ERR_PTR value is returned.
5173 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5174 struct cgroup_subsys *ss)
5176 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5177 struct cgroup_subsys_state *css = NULL;
5178 struct cgroup *cgrp;
5180 /* is @dentry a cgroup dir? */
5181 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5182 kernfs_type(kn) != KERNFS_DIR)
5183 return ERR_PTR(-EBADF);
5188 * This path doesn't originate from kernfs and @kn could already
5189 * have been or be removed at any point. @kn->priv is RCU
5190 * protected for this access. See cgroup_rmdir() for details.
5192 cgrp = rcu_dereference(kn->priv);
5194 css = cgroup_css(cgrp, ss);
5196 if (!css || !css_tryget_online(css))
5197 css = ERR_PTR(-ENOENT);
5204 * css_from_id - lookup css by id
5205 * @id: the cgroup id
5206 * @ss: cgroup subsys to be looked into
5208 * Returns the css if there's valid one with @id, otherwise returns NULL.
5209 * Should be called under rcu_read_lock().
5211 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5213 WARN_ON_ONCE(!rcu_read_lock_held());
5214 return idr_find(&ss->css_idr, id);
5217 #ifdef CONFIG_CGROUP_DEBUG
5218 static struct cgroup_subsys_state *
5219 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5221 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5224 return ERR_PTR(-ENOMEM);
5229 static void debug_css_free(struct cgroup_subsys_state *css)
5234 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5237 return cgroup_task_count(css->cgroup);
5240 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5243 return (u64)(unsigned long)current->cgroups;
5246 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5252 count = atomic_read(&task_css_set(current)->refcount);
5257 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5259 struct cgrp_cset_link *link;
5260 struct css_set *cset;
5263 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5267 down_read(&css_set_rwsem);
5269 cset = rcu_dereference(current->cgroups);
5270 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5271 struct cgroup *c = link->cgrp;
5273 cgroup_name(c, name_buf, NAME_MAX + 1);
5274 seq_printf(seq, "Root %d group %s\n",
5275 c->root->hierarchy_id, name_buf);
5278 up_read(&css_set_rwsem);
5283 #define MAX_TASKS_SHOWN_PER_CSS 25
5284 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5286 struct cgroup_subsys_state *css = seq_css(seq);
5287 struct cgrp_cset_link *link;
5289 down_read(&css_set_rwsem);
5290 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5291 struct css_set *cset = link->cset;
5292 struct task_struct *task;
5295 seq_printf(seq, "css_set %p\n", cset);
5297 list_for_each_entry(task, &cset->tasks, cg_list) {
5298 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5300 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5303 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5304 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5306 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5310 seq_puts(seq, " ...\n");
5312 up_read(&css_set_rwsem);
5316 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5318 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5321 static struct cftype debug_files[] = {
5323 .name = "taskcount",
5324 .read_u64 = debug_taskcount_read,
5328 .name = "current_css_set",
5329 .read_u64 = current_css_set_read,
5333 .name = "current_css_set_refcount",
5334 .read_u64 = current_css_set_refcount_read,
5338 .name = "current_css_set_cg_links",
5339 .seq_show = current_css_set_cg_links_read,
5343 .name = "cgroup_css_links",
5344 .seq_show = cgroup_css_links_read,
5348 .name = "releasable",
5349 .read_u64 = releasable_read,
5355 struct cgroup_subsys debug_cgrp_subsys = {
5356 .css_alloc = debug_css_alloc,
5357 .css_free = debug_css_free,
5358 .base_cftypes = debug_files,
5360 #endif /* CONFIG_CGROUP_DEBUG */