2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 DECLARE_RWSEM(css_set_rwsem);
87 EXPORT_SYMBOL_GPL(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(css_set_rwsem);
90 static DEFINE_MUTEX(cgroup_mutex);
91 static DECLARE_RWSEM(css_set_rwsem);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
123 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys *cgroup_subsys[] = {
128 #include <linux/cgroup_subsys.h>
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name[] = {
135 #include <linux/cgroup_subsys.h>
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
144 struct cgroup_root cgrp_dfl_root;
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
150 static bool cgrp_dfl_root_visible;
152 /* some controllers are not supported in the default hierarchy */
153 static const unsigned int cgrp_dfl_root_inhibit_ss_mask = 0
154 #ifdef CONFIG_CGROUP_DEBUG
155 | (1 << debug_cgrp_id)
159 /* The list of hierarchy roots */
161 static LIST_HEAD(cgroup_roots);
162 static int cgroup_root_count;
164 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
165 static DEFINE_IDR(cgroup_hierarchy_idr);
168 * Assign a monotonically increasing serial number to csses. It guarantees
169 * cgroups with bigger numbers are newer than those with smaller numbers.
170 * Also, as csses are always appended to the parent's ->children list, it
171 * guarantees that sibling csses are always sorted in the ascending serial
172 * number order on the list. Protected by cgroup_mutex.
174 static u64 css_serial_nr_next = 1;
176 /* This flag indicates whether tasks in the fork and exit paths should
177 * check for fork/exit handlers to call. This avoids us having to do
178 * extra work in the fork/exit path if none of the subsystems need to
181 static int need_forkexit_callback __read_mostly;
183 static struct cftype cgroup_base_files[];
185 static void cgroup_put(struct cgroup *cgrp);
186 static int rebind_subsystems(struct cgroup_root *dst_root,
187 unsigned int ss_mask);
188 static int cgroup_destroy_locked(struct cgroup *cgrp);
189 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
191 static void css_release(struct percpu_ref *ref);
192 static void kill_css(struct cgroup_subsys_state *css);
193 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
195 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
197 /* IDR wrappers which synchronize using cgroup_idr_lock */
198 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
203 idr_preload(gfp_mask);
204 spin_lock_bh(&cgroup_idr_lock);
205 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
206 spin_unlock_bh(&cgroup_idr_lock);
211 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
215 spin_lock_bh(&cgroup_idr_lock);
216 ret = idr_replace(idr, ptr, id);
217 spin_unlock_bh(&cgroup_idr_lock);
221 static void cgroup_idr_remove(struct idr *idr, int id)
223 spin_lock_bh(&cgroup_idr_lock);
225 spin_unlock_bh(&cgroup_idr_lock);
228 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
230 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
233 return container_of(parent_css, struct cgroup, self);
238 * cgroup_css - obtain a cgroup's css for the specified subsystem
239 * @cgrp: the cgroup of interest
240 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
242 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
243 * function must be called either under cgroup_mutex or rcu_read_lock() and
244 * the caller is responsible for pinning the returned css if it wants to
245 * keep accessing it outside the said locks. This function may return
246 * %NULL if @cgrp doesn't have @subsys_id enabled.
248 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
249 struct cgroup_subsys *ss)
252 return rcu_dereference_check(cgrp->subsys[ss->id],
253 lockdep_is_held(&cgroup_mutex));
259 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
260 * @cgrp: the cgroup of interest
261 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
263 * Similar to cgroup_css() but returns the effctive css, which is defined
264 * as the matching css of the nearest ancestor including self which has @ss
265 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
266 * function is guaranteed to return non-NULL css.
268 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
269 struct cgroup_subsys *ss)
271 lockdep_assert_held(&cgroup_mutex);
276 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
279 while (cgroup_parent(cgrp) &&
280 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
281 cgrp = cgroup_parent(cgrp);
283 return cgroup_css(cgrp, ss);
286 /* convenient tests for these bits */
287 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
289 return !(cgrp->self.flags & CSS_ONLINE);
292 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
294 struct cgroup *cgrp = of->kn->parent->priv;
295 struct cftype *cft = of_cft(of);
298 * This is open and unprotected implementation of cgroup_css().
299 * seq_css() is only called from a kernfs file operation which has
300 * an active reference on the file. Because all the subsystem
301 * files are drained before a css is disassociated with a cgroup,
302 * the matching css from the cgroup's subsys table is guaranteed to
303 * be and stay valid until the enclosing operation is complete.
306 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
310 EXPORT_SYMBOL_GPL(of_css);
313 * cgroup_is_descendant - test ancestry
314 * @cgrp: the cgroup to be tested
315 * @ancestor: possible ancestor of @cgrp
317 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
318 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
319 * and @ancestor are accessible.
321 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
324 if (cgrp == ancestor)
326 cgrp = cgroup_parent(cgrp);
331 static int cgroup_is_releasable(const struct cgroup *cgrp)
334 (1 << CGRP_RELEASABLE) |
335 (1 << CGRP_NOTIFY_ON_RELEASE);
336 return (cgrp->flags & bits) == bits;
339 static int notify_on_release(const struct cgroup *cgrp)
341 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
345 * for_each_css - iterate all css's of a cgroup
346 * @css: the iteration cursor
347 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
348 * @cgrp: the target cgroup to iterate css's of
350 * Should be called under cgroup_[tree_]mutex.
352 #define for_each_css(css, ssid, cgrp) \
353 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
354 if (!((css) = rcu_dereference_check( \
355 (cgrp)->subsys[(ssid)], \
356 lockdep_is_held(&cgroup_mutex)))) { } \
360 * for_each_e_css - iterate all effective css's of a cgroup
361 * @css: the iteration cursor
362 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
363 * @cgrp: the target cgroup to iterate css's of
365 * Should be called under cgroup_[tree_]mutex.
367 #define for_each_e_css(css, ssid, cgrp) \
368 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
369 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
374 * for_each_subsys - iterate all enabled cgroup subsystems
375 * @ss: the iteration cursor
376 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
378 #define for_each_subsys(ss, ssid) \
379 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
380 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
382 /* iterate across the hierarchies */
383 #define for_each_root(root) \
384 list_for_each_entry((root), &cgroup_roots, root_list)
386 /* iterate over child cgrps, lock should be held throughout iteration */
387 #define cgroup_for_each_live_child(child, cgrp) \
388 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
389 if (({ lockdep_assert_held(&cgroup_mutex); \
390 cgroup_is_dead(child); })) \
394 /* the list of cgroups eligible for automatic release. Protected by
395 * release_list_lock */
396 static LIST_HEAD(release_list);
397 static DEFINE_RAW_SPINLOCK(release_list_lock);
398 static void cgroup_release_agent(struct work_struct *work);
399 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
400 static void check_for_release(struct cgroup *cgrp);
403 * A cgroup can be associated with multiple css_sets as different tasks may
404 * belong to different cgroups on different hierarchies. In the other
405 * direction, a css_set is naturally associated with multiple cgroups.
406 * This M:N relationship is represented by the following link structure
407 * which exists for each association and allows traversing the associations
410 struct cgrp_cset_link {
411 /* the cgroup and css_set this link associates */
413 struct css_set *cset;
415 /* list of cgrp_cset_links anchored at cgrp->cset_links */
416 struct list_head cset_link;
418 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
419 struct list_head cgrp_link;
423 * The default css_set - used by init and its children prior to any
424 * hierarchies being mounted. It contains a pointer to the root state
425 * for each subsystem. Also used to anchor the list of css_sets. Not
426 * reference-counted, to improve performance when child cgroups
427 * haven't been created.
429 struct css_set init_css_set = {
430 .refcount = ATOMIC_INIT(1),
431 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
432 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
433 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
434 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
435 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
438 static int css_set_count = 1; /* 1 for init_css_set */
441 * cgroup_update_populated - updated populated count of a cgroup
442 * @cgrp: the target cgroup
443 * @populated: inc or dec populated count
445 * @cgrp is either getting the first task (css_set) or losing the last.
446 * Update @cgrp->populated_cnt accordingly. The count is propagated
447 * towards root so that a given cgroup's populated_cnt is zero iff the
448 * cgroup and all its descendants are empty.
450 * @cgrp's interface file "cgroup.populated" is zero if
451 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
452 * changes from or to zero, userland is notified that the content of the
453 * interface file has changed. This can be used to detect when @cgrp and
454 * its descendants become populated or empty.
456 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
458 lockdep_assert_held(&css_set_rwsem);
464 trigger = !cgrp->populated_cnt++;
466 trigger = !--cgrp->populated_cnt;
471 if (cgrp->populated_kn)
472 kernfs_notify(cgrp->populated_kn);
473 cgrp = cgroup_parent(cgrp);
478 * hash table for cgroup groups. This improves the performance to find
479 * an existing css_set. This hash doesn't (currently) take into
480 * account cgroups in empty hierarchies.
482 #define CSS_SET_HASH_BITS 7
483 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
485 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
487 unsigned long key = 0UL;
488 struct cgroup_subsys *ss;
491 for_each_subsys(ss, i)
492 key += (unsigned long)css[i];
493 key = (key >> 16) ^ key;
498 static void put_css_set_locked(struct css_set *cset, bool taskexit)
500 struct cgrp_cset_link *link, *tmp_link;
501 struct cgroup_subsys *ss;
504 lockdep_assert_held(&css_set_rwsem);
506 if (!atomic_dec_and_test(&cset->refcount))
509 /* This css_set is dead. unlink it and release cgroup refcounts */
510 for_each_subsys(ss, ssid)
511 list_del(&cset->e_cset_node[ssid]);
512 hash_del(&cset->hlist);
515 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
516 struct cgroup *cgrp = link->cgrp;
518 list_del(&link->cset_link);
519 list_del(&link->cgrp_link);
521 /* @cgrp can't go away while we're holding css_set_rwsem */
522 if (list_empty(&cgrp->cset_links)) {
523 cgroup_update_populated(cgrp, false);
524 if (notify_on_release(cgrp)) {
526 set_bit(CGRP_RELEASABLE, &cgrp->flags);
527 check_for_release(cgrp);
534 kfree_rcu(cset, rcu_head);
537 static void put_css_set(struct css_set *cset, bool taskexit)
540 * Ensure that the refcount doesn't hit zero while any readers
541 * can see it. Similar to atomic_dec_and_lock(), but for an
544 if (atomic_add_unless(&cset->refcount, -1, 1))
547 down_write(&css_set_rwsem);
548 put_css_set_locked(cset, taskexit);
549 up_write(&css_set_rwsem);
553 * refcounted get/put for css_set objects
555 static inline void get_css_set(struct css_set *cset)
557 atomic_inc(&cset->refcount);
561 * compare_css_sets - helper function for find_existing_css_set().
562 * @cset: candidate css_set being tested
563 * @old_cset: existing css_set for a task
564 * @new_cgrp: cgroup that's being entered by the task
565 * @template: desired set of css pointers in css_set (pre-calculated)
567 * Returns true if "cset" matches "old_cset" except for the hierarchy
568 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
570 static bool compare_css_sets(struct css_set *cset,
571 struct css_set *old_cset,
572 struct cgroup *new_cgrp,
573 struct cgroup_subsys_state *template[])
575 struct list_head *l1, *l2;
578 * On the default hierarchy, there can be csets which are
579 * associated with the same set of cgroups but different csses.
580 * Let's first ensure that csses match.
582 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
586 * Compare cgroup pointers in order to distinguish between
587 * different cgroups in hierarchies. As different cgroups may
588 * share the same effective css, this comparison is always
591 l1 = &cset->cgrp_links;
592 l2 = &old_cset->cgrp_links;
594 struct cgrp_cset_link *link1, *link2;
595 struct cgroup *cgrp1, *cgrp2;
599 /* See if we reached the end - both lists are equal length. */
600 if (l1 == &cset->cgrp_links) {
601 BUG_ON(l2 != &old_cset->cgrp_links);
604 BUG_ON(l2 == &old_cset->cgrp_links);
606 /* Locate the cgroups associated with these links. */
607 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
608 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
611 /* Hierarchies should be linked in the same order. */
612 BUG_ON(cgrp1->root != cgrp2->root);
615 * If this hierarchy is the hierarchy of the cgroup
616 * that's changing, then we need to check that this
617 * css_set points to the new cgroup; if it's any other
618 * hierarchy, then this css_set should point to the
619 * same cgroup as the old css_set.
621 if (cgrp1->root == new_cgrp->root) {
622 if (cgrp1 != new_cgrp)
633 * find_existing_css_set - init css array and find the matching css_set
634 * @old_cset: the css_set that we're using before the cgroup transition
635 * @cgrp: the cgroup that we're moving into
636 * @template: out param for the new set of csses, should be clear on entry
638 static struct css_set *find_existing_css_set(struct css_set *old_cset,
640 struct cgroup_subsys_state *template[])
642 struct cgroup_root *root = cgrp->root;
643 struct cgroup_subsys *ss;
644 struct css_set *cset;
649 * Build the set of subsystem state objects that we want to see in the
650 * new css_set. while subsystems can change globally, the entries here
651 * won't change, so no need for locking.
653 for_each_subsys(ss, i) {
654 if (root->subsys_mask & (1UL << i)) {
656 * @ss is in this hierarchy, so we want the
657 * effective css from @cgrp.
659 template[i] = cgroup_e_css(cgrp, ss);
662 * @ss is not in this hierarchy, so we don't want
665 template[i] = old_cset->subsys[i];
669 key = css_set_hash(template);
670 hash_for_each_possible(css_set_table, cset, hlist, key) {
671 if (!compare_css_sets(cset, old_cset, cgrp, template))
674 /* This css_set matches what we need */
678 /* No existing cgroup group matched */
682 static void free_cgrp_cset_links(struct list_head *links_to_free)
684 struct cgrp_cset_link *link, *tmp_link;
686 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
687 list_del(&link->cset_link);
693 * allocate_cgrp_cset_links - allocate cgrp_cset_links
694 * @count: the number of links to allocate
695 * @tmp_links: list_head the allocated links are put on
697 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
698 * through ->cset_link. Returns 0 on success or -errno.
700 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
702 struct cgrp_cset_link *link;
705 INIT_LIST_HEAD(tmp_links);
707 for (i = 0; i < count; i++) {
708 link = kzalloc(sizeof(*link), GFP_KERNEL);
710 free_cgrp_cset_links(tmp_links);
713 list_add(&link->cset_link, tmp_links);
719 * link_css_set - a helper function to link a css_set to a cgroup
720 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
721 * @cset: the css_set to be linked
722 * @cgrp: the destination cgroup
724 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
727 struct cgrp_cset_link *link;
729 BUG_ON(list_empty(tmp_links));
731 if (cgroup_on_dfl(cgrp))
732 cset->dfl_cgrp = cgrp;
734 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
738 if (list_empty(&cgrp->cset_links))
739 cgroup_update_populated(cgrp, true);
740 list_move(&link->cset_link, &cgrp->cset_links);
743 * Always add links to the tail of the list so that the list
744 * is sorted by order of hierarchy creation
746 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
750 * find_css_set - return a new css_set with one cgroup updated
751 * @old_cset: the baseline css_set
752 * @cgrp: the cgroup to be updated
754 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
755 * substituted into the appropriate hierarchy.
757 static struct css_set *find_css_set(struct css_set *old_cset,
760 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
761 struct css_set *cset;
762 struct list_head tmp_links;
763 struct cgrp_cset_link *link;
764 struct cgroup_subsys *ss;
768 lockdep_assert_held(&cgroup_mutex);
770 /* First see if we already have a cgroup group that matches
772 down_read(&css_set_rwsem);
773 cset = find_existing_css_set(old_cset, cgrp, template);
776 up_read(&css_set_rwsem);
781 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
785 /* Allocate all the cgrp_cset_link objects that we'll need */
786 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
791 atomic_set(&cset->refcount, 1);
792 INIT_LIST_HEAD(&cset->cgrp_links);
793 INIT_LIST_HEAD(&cset->tasks);
794 INIT_LIST_HEAD(&cset->mg_tasks);
795 INIT_LIST_HEAD(&cset->mg_preload_node);
796 INIT_LIST_HEAD(&cset->mg_node);
797 INIT_HLIST_NODE(&cset->hlist);
799 /* Copy the set of subsystem state objects generated in
800 * find_existing_css_set() */
801 memcpy(cset->subsys, template, sizeof(cset->subsys));
803 down_write(&css_set_rwsem);
804 /* Add reference counts and links from the new css_set. */
805 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
806 struct cgroup *c = link->cgrp;
808 if (c->root == cgrp->root)
810 link_css_set(&tmp_links, cset, c);
813 BUG_ON(!list_empty(&tmp_links));
817 /* Add @cset to the hash table */
818 key = css_set_hash(cset->subsys);
819 hash_add(css_set_table, &cset->hlist, key);
821 for_each_subsys(ss, ssid)
822 list_add_tail(&cset->e_cset_node[ssid],
823 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
825 up_write(&css_set_rwsem);
830 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
832 struct cgroup *root_cgrp = kf_root->kn->priv;
834 return root_cgrp->root;
837 static int cgroup_init_root_id(struct cgroup_root *root)
841 lockdep_assert_held(&cgroup_mutex);
843 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
847 root->hierarchy_id = id;
851 static void cgroup_exit_root_id(struct cgroup_root *root)
853 lockdep_assert_held(&cgroup_mutex);
855 if (root->hierarchy_id) {
856 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
857 root->hierarchy_id = 0;
861 static void cgroup_free_root(struct cgroup_root *root)
864 /* hierarhcy ID shoulid already have been released */
865 WARN_ON_ONCE(root->hierarchy_id);
867 idr_destroy(&root->cgroup_idr);
872 static void cgroup_destroy_root(struct cgroup_root *root)
874 struct cgroup *cgrp = &root->cgrp;
875 struct cgrp_cset_link *link, *tmp_link;
877 mutex_lock(&cgroup_mutex);
879 BUG_ON(atomic_read(&root->nr_cgrps));
880 BUG_ON(!list_empty(&cgrp->self.children));
882 /* Rebind all subsystems back to the default hierarchy */
883 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
886 * Release all the links from cset_links to this hierarchy's
889 down_write(&css_set_rwsem);
891 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
892 list_del(&link->cset_link);
893 list_del(&link->cgrp_link);
896 up_write(&css_set_rwsem);
898 if (!list_empty(&root->root_list)) {
899 list_del(&root->root_list);
903 cgroup_exit_root_id(root);
905 mutex_unlock(&cgroup_mutex);
907 kernfs_destroy_root(root->kf_root);
908 cgroup_free_root(root);
911 /* look up cgroup associated with given css_set on the specified hierarchy */
912 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
913 struct cgroup_root *root)
915 struct cgroup *res = NULL;
917 lockdep_assert_held(&cgroup_mutex);
918 lockdep_assert_held(&css_set_rwsem);
920 if (cset == &init_css_set) {
923 struct cgrp_cset_link *link;
925 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
926 struct cgroup *c = link->cgrp;
928 if (c->root == root) {
940 * Return the cgroup for "task" from the given hierarchy. Must be
941 * called with cgroup_mutex and css_set_rwsem held.
943 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
944 struct cgroup_root *root)
947 * No need to lock the task - since we hold cgroup_mutex the
948 * task can't change groups, so the only thing that can happen
949 * is that it exits and its css is set back to init_css_set.
951 return cset_cgroup_from_root(task_css_set(task), root);
955 * A task must hold cgroup_mutex to modify cgroups.
957 * Any task can increment and decrement the count field without lock.
958 * So in general, code holding cgroup_mutex can't rely on the count
959 * field not changing. However, if the count goes to zero, then only
960 * cgroup_attach_task() can increment it again. Because a count of zero
961 * means that no tasks are currently attached, therefore there is no
962 * way a task attached to that cgroup can fork (the other way to
963 * increment the count). So code holding cgroup_mutex can safely
964 * assume that if the count is zero, it will stay zero. Similarly, if
965 * a task holds cgroup_mutex on a cgroup with zero count, it
966 * knows that the cgroup won't be removed, as cgroup_rmdir()
969 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
970 * (usually) take cgroup_mutex. These are the two most performance
971 * critical pieces of code here. The exception occurs on cgroup_exit(),
972 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
973 * is taken, and if the cgroup count is zero, a usermode call made
974 * to the release agent with the name of the cgroup (path relative to
975 * the root of cgroup file system) as the argument.
977 * A cgroup can only be deleted if both its 'count' of using tasks
978 * is zero, and its list of 'children' cgroups is empty. Since all
979 * tasks in the system use _some_ cgroup, and since there is always at
980 * least one task in the system (init, pid == 1), therefore, root cgroup
981 * always has either children cgroups and/or using tasks. So we don't
982 * need a special hack to ensure that root cgroup cannot be deleted.
984 * P.S. One more locking exception. RCU is used to guard the
985 * update of a tasks cgroup pointer by cgroup_attach_task()
988 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
989 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
990 static const struct file_operations proc_cgroupstats_operations;
992 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
995 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
996 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
997 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
998 cft->ss->name, cft->name);
1000 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1005 * cgroup_file_mode - deduce file mode of a control file
1006 * @cft: the control file in question
1008 * returns cft->mode if ->mode is not 0
1009 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1010 * returns S_IRUGO if it has only a read handler
1011 * returns S_IWUSR if it has only a write hander
1013 static umode_t cgroup_file_mode(const struct cftype *cft)
1020 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1023 if (cft->write_u64 || cft->write_s64 || cft->write)
1029 static void cgroup_get(struct cgroup *cgrp)
1031 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1032 css_get(&cgrp->self);
1035 static void cgroup_put(struct cgroup *cgrp)
1037 css_put(&cgrp->self);
1040 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1042 cgrp->child_subsys_mask = cgrp->subtree_control;
1046 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1047 * @kn: the kernfs_node being serviced
1049 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1050 * the method finishes if locking succeeded. Note that once this function
1051 * returns the cgroup returned by cgroup_kn_lock_live() may become
1052 * inaccessible any time. If the caller intends to continue to access the
1053 * cgroup, it should pin it before invoking this function.
1055 static void cgroup_kn_unlock(struct kernfs_node *kn)
1057 struct cgroup *cgrp;
1059 if (kernfs_type(kn) == KERNFS_DIR)
1062 cgrp = kn->parent->priv;
1064 mutex_unlock(&cgroup_mutex);
1066 kernfs_unbreak_active_protection(kn);
1071 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1072 * @kn: the kernfs_node being serviced
1074 * This helper is to be used by a cgroup kernfs method currently servicing
1075 * @kn. It breaks the active protection, performs cgroup locking and
1076 * verifies that the associated cgroup is alive. Returns the cgroup if
1077 * alive; otherwise, %NULL. A successful return should be undone by a
1078 * matching cgroup_kn_unlock() invocation.
1080 * Any cgroup kernfs method implementation which requires locking the
1081 * associated cgroup should use this helper. It avoids nesting cgroup
1082 * locking under kernfs active protection and allows all kernfs operations
1083 * including self-removal.
1085 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1087 struct cgroup *cgrp;
1089 if (kernfs_type(kn) == KERNFS_DIR)
1092 cgrp = kn->parent->priv;
1095 * We're gonna grab cgroup_mutex which nests outside kernfs
1096 * active_ref. cgroup liveliness check alone provides enough
1097 * protection against removal. Ensure @cgrp stays accessible and
1098 * break the active_ref protection.
1101 kernfs_break_active_protection(kn);
1103 mutex_lock(&cgroup_mutex);
1105 if (!cgroup_is_dead(cgrp))
1108 cgroup_kn_unlock(kn);
1112 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1114 char name[CGROUP_FILE_NAME_MAX];
1116 lockdep_assert_held(&cgroup_mutex);
1117 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1121 * cgroup_clear_dir - remove subsys files in a cgroup directory
1122 * @cgrp: target cgroup
1123 * @subsys_mask: mask of the subsystem ids whose files should be removed
1125 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1127 struct cgroup_subsys *ss;
1130 for_each_subsys(ss, i) {
1131 struct cftype *cfts;
1133 if (!(subsys_mask & (1 << i)))
1135 list_for_each_entry(cfts, &ss->cfts, node)
1136 cgroup_addrm_files(cgrp, cfts, false);
1140 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1142 struct cgroup_subsys *ss;
1143 unsigned int tmp_ss_mask;
1146 lockdep_assert_held(&cgroup_mutex);
1148 for_each_subsys(ss, ssid) {
1149 if (!(ss_mask & (1 << ssid)))
1152 /* if @ss has non-root csses attached to it, can't move */
1153 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1156 /* can't move between two non-dummy roots either */
1157 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1161 /* skip creating root files on dfl_root for inhibited subsystems */
1162 tmp_ss_mask = ss_mask;
1163 if (dst_root == &cgrp_dfl_root)
1164 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1166 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1168 if (dst_root != &cgrp_dfl_root)
1172 * Rebinding back to the default root is not allowed to
1173 * fail. Using both default and non-default roots should
1174 * be rare. Moving subsystems back and forth even more so.
1175 * Just warn about it and continue.
1177 if (cgrp_dfl_root_visible) {
1178 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1180 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1185 * Nothing can fail from this point on. Remove files for the
1186 * removed subsystems and rebind each subsystem.
1188 for_each_subsys(ss, ssid)
1189 if (ss_mask & (1 << ssid))
1190 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1192 for_each_subsys(ss, ssid) {
1193 struct cgroup_root *src_root;
1194 struct cgroup_subsys_state *css;
1195 struct css_set *cset;
1197 if (!(ss_mask & (1 << ssid)))
1200 src_root = ss->root;
1201 css = cgroup_css(&src_root->cgrp, ss);
1203 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1205 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1206 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1207 ss->root = dst_root;
1208 css->cgroup = &dst_root->cgrp;
1210 down_write(&css_set_rwsem);
1211 hash_for_each(css_set_table, i, cset, hlist)
1212 list_move_tail(&cset->e_cset_node[ss->id],
1213 &dst_root->cgrp.e_csets[ss->id]);
1214 up_write(&css_set_rwsem);
1216 src_root->subsys_mask &= ~(1 << ssid);
1217 src_root->cgrp.subtree_control &= ~(1 << ssid);
1218 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1220 /* default hierarchy doesn't enable controllers by default */
1221 dst_root->subsys_mask |= 1 << ssid;
1222 if (dst_root != &cgrp_dfl_root) {
1223 dst_root->cgrp.subtree_control |= 1 << ssid;
1224 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1231 kernfs_activate(dst_root->cgrp.kn);
1235 static int cgroup_show_options(struct seq_file *seq,
1236 struct kernfs_root *kf_root)
1238 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1239 struct cgroup_subsys *ss;
1242 for_each_subsys(ss, ssid)
1243 if (root->subsys_mask & (1 << ssid))
1244 seq_printf(seq, ",%s", ss->name);
1245 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1246 seq_puts(seq, ",sane_behavior");
1247 if (root->flags & CGRP_ROOT_NOPREFIX)
1248 seq_puts(seq, ",noprefix");
1249 if (root->flags & CGRP_ROOT_XATTR)
1250 seq_puts(seq, ",xattr");
1252 spin_lock(&release_agent_path_lock);
1253 if (strlen(root->release_agent_path))
1254 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1255 spin_unlock(&release_agent_path_lock);
1257 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1258 seq_puts(seq, ",clone_children");
1259 if (strlen(root->name))
1260 seq_printf(seq, ",name=%s", root->name);
1264 struct cgroup_sb_opts {
1265 unsigned int subsys_mask;
1267 char *release_agent;
1268 bool cpuset_clone_children;
1270 /* User explicitly requested empty subsystem */
1274 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1276 char *token, *o = data;
1277 bool all_ss = false, one_ss = false;
1278 unsigned int mask = -1U;
1279 struct cgroup_subsys *ss;
1282 #ifdef CONFIG_CPUSETS
1283 mask = ~(1U << cpuset_cgrp_id);
1286 memset(opts, 0, sizeof(*opts));
1288 while ((token = strsep(&o, ",")) != NULL) {
1291 if (!strcmp(token, "none")) {
1292 /* Explicitly have no subsystems */
1296 if (!strcmp(token, "all")) {
1297 /* Mutually exclusive option 'all' + subsystem name */
1303 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1304 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1307 if (!strcmp(token, "noprefix")) {
1308 opts->flags |= CGRP_ROOT_NOPREFIX;
1311 if (!strcmp(token, "clone_children")) {
1312 opts->cpuset_clone_children = true;
1315 if (!strcmp(token, "xattr")) {
1316 opts->flags |= CGRP_ROOT_XATTR;
1319 if (!strncmp(token, "release_agent=", 14)) {
1320 /* Specifying two release agents is forbidden */
1321 if (opts->release_agent)
1323 opts->release_agent =
1324 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1325 if (!opts->release_agent)
1329 if (!strncmp(token, "name=", 5)) {
1330 const char *name = token + 5;
1331 /* Can't specify an empty name */
1334 /* Must match [\w.-]+ */
1335 for (i = 0; i < strlen(name); i++) {
1339 if ((c == '.') || (c == '-') || (c == '_'))
1343 /* Specifying two names is forbidden */
1346 opts->name = kstrndup(name,
1347 MAX_CGROUP_ROOT_NAMELEN - 1,
1355 for_each_subsys(ss, i) {
1356 if (strcmp(token, ss->name))
1361 /* Mutually exclusive option 'all' + subsystem name */
1364 opts->subsys_mask |= (1 << i);
1369 if (i == CGROUP_SUBSYS_COUNT)
1373 /* Consistency checks */
1375 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1376 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1378 if ((opts->flags & (CGRP_ROOT_NOPREFIX | CGRP_ROOT_XATTR)) ||
1379 opts->cpuset_clone_children || opts->release_agent ||
1381 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1386 * If the 'all' option was specified select all the
1387 * subsystems, otherwise if 'none', 'name=' and a subsystem
1388 * name options were not specified, let's default to 'all'
1390 if (all_ss || (!one_ss && !opts->none && !opts->name))
1391 for_each_subsys(ss, i)
1393 opts->subsys_mask |= (1 << i);
1396 * We either have to specify by name or by subsystems. (So
1397 * all empty hierarchies must have a name).
1399 if (!opts->subsys_mask && !opts->name)
1404 * Option noprefix was introduced just for backward compatibility
1405 * with the old cpuset, so we allow noprefix only if mounting just
1406 * the cpuset subsystem.
1408 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1412 /* Can't specify "none" and some subsystems */
1413 if (opts->subsys_mask && opts->none)
1419 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1422 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1423 struct cgroup_sb_opts opts;
1424 unsigned int added_mask, removed_mask;
1426 if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1427 pr_err("sane_behavior: remount is not allowed\n");
1431 mutex_lock(&cgroup_mutex);
1433 /* See what subsystems are wanted */
1434 ret = parse_cgroupfs_options(data, &opts);
1438 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1439 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1440 task_tgid_nr(current), current->comm);
1442 added_mask = opts.subsys_mask & ~root->subsys_mask;
1443 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1445 /* Don't allow flags or name to change at remount */
1446 if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1447 (opts.name && strcmp(opts.name, root->name))) {
1448 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1449 opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1450 root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1455 /* remounting is not allowed for populated hierarchies */
1456 if (!list_empty(&root->cgrp.self.children)) {
1461 ret = rebind_subsystems(root, added_mask);
1465 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1467 if (opts.release_agent) {
1468 spin_lock(&release_agent_path_lock);
1469 strcpy(root->release_agent_path, opts.release_agent);
1470 spin_unlock(&release_agent_path_lock);
1473 kfree(opts.release_agent);
1475 mutex_unlock(&cgroup_mutex);
1480 * To reduce the fork() overhead for systems that are not actually using
1481 * their cgroups capability, we don't maintain the lists running through
1482 * each css_set to its tasks until we see the list actually used - in other
1483 * words after the first mount.
1485 static bool use_task_css_set_links __read_mostly;
1487 static void cgroup_enable_task_cg_lists(void)
1489 struct task_struct *p, *g;
1491 down_write(&css_set_rwsem);
1493 if (use_task_css_set_links)
1496 use_task_css_set_links = true;
1499 * We need tasklist_lock because RCU is not safe against
1500 * while_each_thread(). Besides, a forking task that has passed
1501 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1502 * is not guaranteed to have its child immediately visible in the
1503 * tasklist if we walk through it with RCU.
1505 read_lock(&tasklist_lock);
1506 do_each_thread(g, p) {
1507 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1508 task_css_set(p) != &init_css_set);
1511 * We should check if the process is exiting, otherwise
1512 * it will race with cgroup_exit() in that the list
1513 * entry won't be deleted though the process has exited.
1514 * Do it while holding siglock so that we don't end up
1515 * racing against cgroup_exit().
1517 spin_lock_irq(&p->sighand->siglock);
1518 if (!(p->flags & PF_EXITING)) {
1519 struct css_set *cset = task_css_set(p);
1521 list_add(&p->cg_list, &cset->tasks);
1524 spin_unlock_irq(&p->sighand->siglock);
1525 } while_each_thread(g, p);
1526 read_unlock(&tasklist_lock);
1528 up_write(&css_set_rwsem);
1531 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1533 struct cgroup_subsys *ss;
1536 INIT_LIST_HEAD(&cgrp->self.sibling);
1537 INIT_LIST_HEAD(&cgrp->self.children);
1538 INIT_LIST_HEAD(&cgrp->cset_links);
1539 INIT_LIST_HEAD(&cgrp->release_list);
1540 INIT_LIST_HEAD(&cgrp->pidlists);
1541 mutex_init(&cgrp->pidlist_mutex);
1542 cgrp->self.cgroup = cgrp;
1543 cgrp->self.flags |= CSS_ONLINE;
1545 for_each_subsys(ss, ssid)
1546 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1548 init_waitqueue_head(&cgrp->offline_waitq);
1551 static void init_cgroup_root(struct cgroup_root *root,
1552 struct cgroup_sb_opts *opts)
1554 struct cgroup *cgrp = &root->cgrp;
1556 INIT_LIST_HEAD(&root->root_list);
1557 atomic_set(&root->nr_cgrps, 1);
1559 init_cgroup_housekeeping(cgrp);
1560 idr_init(&root->cgroup_idr);
1562 root->flags = opts->flags;
1563 if (opts->release_agent)
1564 strcpy(root->release_agent_path, opts->release_agent);
1566 strcpy(root->name, opts->name);
1567 if (opts->cpuset_clone_children)
1568 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1571 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1573 LIST_HEAD(tmp_links);
1574 struct cgroup *root_cgrp = &root->cgrp;
1575 struct css_set *cset;
1578 lockdep_assert_held(&cgroup_mutex);
1580 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1583 root_cgrp->id = ret;
1585 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
1590 * We're accessing css_set_count without locking css_set_rwsem here,
1591 * but that's OK - it can only be increased by someone holding
1592 * cgroup_lock, and that's us. The worst that can happen is that we
1593 * have some link structures left over
1595 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1599 ret = cgroup_init_root_id(root);
1603 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1604 KERNFS_ROOT_CREATE_DEACTIVATED,
1606 if (IS_ERR(root->kf_root)) {
1607 ret = PTR_ERR(root->kf_root);
1610 root_cgrp->kn = root->kf_root->kn;
1612 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1616 ret = rebind_subsystems(root, ss_mask);
1621 * There must be no failure case after here, since rebinding takes
1622 * care of subsystems' refcounts, which are explicitly dropped in
1623 * the failure exit path.
1625 list_add(&root->root_list, &cgroup_roots);
1626 cgroup_root_count++;
1629 * Link the root cgroup in this hierarchy into all the css_set
1632 down_write(&css_set_rwsem);
1633 hash_for_each(css_set_table, i, cset, hlist)
1634 link_css_set(&tmp_links, cset, root_cgrp);
1635 up_write(&css_set_rwsem);
1637 BUG_ON(!list_empty(&root_cgrp->self.children));
1638 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1640 kernfs_activate(root_cgrp->kn);
1645 kernfs_destroy_root(root->kf_root);
1646 root->kf_root = NULL;
1648 cgroup_exit_root_id(root);
1650 percpu_ref_cancel_init(&root_cgrp->self.refcnt);
1652 free_cgrp_cset_links(&tmp_links);
1656 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1657 int flags, const char *unused_dev_name,
1660 struct cgroup_root *root;
1661 struct cgroup_sb_opts opts;
1662 struct dentry *dentry;
1667 * The first time anyone tries to mount a cgroup, enable the list
1668 * linking each css_set to its tasks and fix up all existing tasks.
1670 if (!use_task_css_set_links)
1671 cgroup_enable_task_cg_lists();
1673 mutex_lock(&cgroup_mutex);
1675 /* First find the desired set of subsystems */
1676 ret = parse_cgroupfs_options(data, &opts);
1680 /* look for a matching existing root */
1681 if (!opts.subsys_mask && !opts.none && !opts.name) {
1682 cgrp_dfl_root_visible = true;
1683 root = &cgrp_dfl_root;
1684 cgroup_get(&root->cgrp);
1689 for_each_root(root) {
1690 bool name_match = false;
1692 if (root == &cgrp_dfl_root)
1696 * If we asked for a name then it must match. Also, if
1697 * name matches but sybsys_mask doesn't, we should fail.
1698 * Remember whether name matched.
1701 if (strcmp(opts.name, root->name))
1707 * If we asked for subsystems (or explicitly for no
1708 * subsystems) then they must match.
1710 if ((opts.subsys_mask || opts.none) &&
1711 (opts.subsys_mask != root->subsys_mask)) {
1718 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1719 if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1720 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1724 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1729 * A root's lifetime is governed by its root cgroup.
1730 * tryget_live failure indicate that the root is being
1731 * destroyed. Wait for destruction to complete so that the
1732 * subsystems are free. We can use wait_queue for the wait
1733 * but this path is super cold. Let's just sleep for a bit
1736 if (!percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1737 mutex_unlock(&cgroup_mutex);
1739 ret = restart_syscall();
1748 * No such thing, create a new one. name= matching without subsys
1749 * specification is allowed for already existing hierarchies but we
1750 * can't create new one without subsys specification.
1752 if (!opts.subsys_mask && !opts.none) {
1757 root = kzalloc(sizeof(*root), GFP_KERNEL);
1763 init_cgroup_root(root, &opts);
1765 ret = cgroup_setup_root(root, opts.subsys_mask);
1767 cgroup_free_root(root);
1770 mutex_unlock(&cgroup_mutex);
1772 kfree(opts.release_agent);
1776 return ERR_PTR(ret);
1778 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1779 CGROUP_SUPER_MAGIC, &new_sb);
1780 if (IS_ERR(dentry) || !new_sb)
1781 cgroup_put(&root->cgrp);
1785 static void cgroup_kill_sb(struct super_block *sb)
1787 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1788 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1791 * If @root doesn't have any mounts or children, start killing it.
1792 * This prevents new mounts by disabling percpu_ref_tryget_live().
1793 * cgroup_mount() may wait for @root's release.
1795 * And don't kill the default root.
1797 if (css_has_online_children(&root->cgrp.self) ||
1798 root == &cgrp_dfl_root)
1799 cgroup_put(&root->cgrp);
1801 percpu_ref_kill(&root->cgrp.self.refcnt);
1806 static struct file_system_type cgroup_fs_type = {
1808 .mount = cgroup_mount,
1809 .kill_sb = cgroup_kill_sb,
1812 static struct kobject *cgroup_kobj;
1815 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1816 * @task: target task
1817 * @buf: the buffer to write the path into
1818 * @buflen: the length of the buffer
1820 * Determine @task's cgroup on the first (the one with the lowest non-zero
1821 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1822 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1823 * cgroup controller callbacks.
1825 * Return value is the same as kernfs_path().
1827 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1829 struct cgroup_root *root;
1830 struct cgroup *cgrp;
1831 int hierarchy_id = 1;
1834 mutex_lock(&cgroup_mutex);
1835 down_read(&css_set_rwsem);
1837 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1840 cgrp = task_cgroup_from_root(task, root);
1841 path = cgroup_path(cgrp, buf, buflen);
1843 /* if no hierarchy exists, everyone is in "/" */
1844 if (strlcpy(buf, "/", buflen) < buflen)
1848 up_read(&css_set_rwsem);
1849 mutex_unlock(&cgroup_mutex);
1852 EXPORT_SYMBOL_GPL(task_cgroup_path);
1854 /* used to track tasks and other necessary states during migration */
1855 struct cgroup_taskset {
1856 /* the src and dst cset list running through cset->mg_node */
1857 struct list_head src_csets;
1858 struct list_head dst_csets;
1861 * Fields for cgroup_taskset_*() iteration.
1863 * Before migration is committed, the target migration tasks are on
1864 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1865 * the csets on ->dst_csets. ->csets point to either ->src_csets
1866 * or ->dst_csets depending on whether migration is committed.
1868 * ->cur_csets and ->cur_task point to the current task position
1871 struct list_head *csets;
1872 struct css_set *cur_cset;
1873 struct task_struct *cur_task;
1877 * cgroup_taskset_first - reset taskset and return the first task
1878 * @tset: taskset of interest
1880 * @tset iteration is initialized and the first task is returned.
1882 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1884 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1885 tset->cur_task = NULL;
1887 return cgroup_taskset_next(tset);
1891 * cgroup_taskset_next - iterate to the next task in taskset
1892 * @tset: taskset of interest
1894 * Return the next task in @tset. Iteration must have been initialized
1895 * with cgroup_taskset_first().
1897 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1899 struct css_set *cset = tset->cur_cset;
1900 struct task_struct *task = tset->cur_task;
1902 while (&cset->mg_node != tset->csets) {
1904 task = list_first_entry(&cset->mg_tasks,
1905 struct task_struct, cg_list);
1907 task = list_next_entry(task, cg_list);
1909 if (&task->cg_list != &cset->mg_tasks) {
1910 tset->cur_cset = cset;
1911 tset->cur_task = task;
1915 cset = list_next_entry(cset, mg_node);
1923 * cgroup_task_migrate - move a task from one cgroup to another.
1924 * @old_cgrp: the cgroup @tsk is being migrated from
1925 * @tsk: the task being migrated
1926 * @new_cset: the new css_set @tsk is being attached to
1928 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1930 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1931 struct task_struct *tsk,
1932 struct css_set *new_cset)
1934 struct css_set *old_cset;
1936 lockdep_assert_held(&cgroup_mutex);
1937 lockdep_assert_held(&css_set_rwsem);
1940 * We are synchronized through threadgroup_lock() against PF_EXITING
1941 * setting such that we can't race against cgroup_exit() changing the
1942 * css_set to init_css_set and dropping the old one.
1944 WARN_ON_ONCE(tsk->flags & PF_EXITING);
1945 old_cset = task_css_set(tsk);
1947 get_css_set(new_cset);
1948 rcu_assign_pointer(tsk->cgroups, new_cset);
1951 * Use move_tail so that cgroup_taskset_first() still returns the
1952 * leader after migration. This works because cgroup_migrate()
1953 * ensures that the dst_cset of the leader is the first on the
1954 * tset's dst_csets list.
1956 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
1959 * We just gained a reference on old_cset by taking it from the
1960 * task. As trading it for new_cset is protected by cgroup_mutex,
1961 * we're safe to drop it here; it will be freed under RCU.
1963 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1964 put_css_set_locked(old_cset, false);
1968 * cgroup_migrate_finish - cleanup after attach
1969 * @preloaded_csets: list of preloaded css_sets
1971 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1972 * those functions for details.
1974 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
1976 struct css_set *cset, *tmp_cset;
1978 lockdep_assert_held(&cgroup_mutex);
1980 down_write(&css_set_rwsem);
1981 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
1982 cset->mg_src_cgrp = NULL;
1983 cset->mg_dst_cset = NULL;
1984 list_del_init(&cset->mg_preload_node);
1985 put_css_set_locked(cset, false);
1987 up_write(&css_set_rwsem);
1991 * cgroup_migrate_add_src - add a migration source css_set
1992 * @src_cset: the source css_set to add
1993 * @dst_cgrp: the destination cgroup
1994 * @preloaded_csets: list of preloaded css_sets
1996 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1997 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1998 * up by cgroup_migrate_finish().
2000 * This function may be called without holding threadgroup_lock even if the
2001 * target is a process. Threads may be created and destroyed but as long
2002 * as cgroup_mutex is not dropped, no new css_set can be put into play and
2003 * the preloaded css_sets are guaranteed to cover all migrations.
2005 static void cgroup_migrate_add_src(struct css_set *src_cset,
2006 struct cgroup *dst_cgrp,
2007 struct list_head *preloaded_csets)
2009 struct cgroup *src_cgrp;
2011 lockdep_assert_held(&cgroup_mutex);
2012 lockdep_assert_held(&css_set_rwsem);
2014 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2016 if (!list_empty(&src_cset->mg_preload_node))
2019 WARN_ON(src_cset->mg_src_cgrp);
2020 WARN_ON(!list_empty(&src_cset->mg_tasks));
2021 WARN_ON(!list_empty(&src_cset->mg_node));
2023 src_cset->mg_src_cgrp = src_cgrp;
2024 get_css_set(src_cset);
2025 list_add(&src_cset->mg_preload_node, preloaded_csets);
2029 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2030 * @dst_cgrp: the destination cgroup (may be %NULL)
2031 * @preloaded_csets: list of preloaded source css_sets
2033 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2034 * have been preloaded to @preloaded_csets. This function looks up and
2035 * pins all destination css_sets, links each to its source, and append them
2036 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2037 * source css_set is assumed to be its cgroup on the default hierarchy.
2039 * This function must be called after cgroup_migrate_add_src() has been
2040 * called on each migration source css_set. After migration is performed
2041 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2044 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2045 struct list_head *preloaded_csets)
2048 struct css_set *src_cset, *tmp_cset;
2050 lockdep_assert_held(&cgroup_mutex);
2053 * Except for the root, child_subsys_mask must be zero for a cgroup
2054 * with tasks so that child cgroups don't compete against tasks.
2056 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2057 dst_cgrp->child_subsys_mask)
2060 /* look up the dst cset for each src cset and link it to src */
2061 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2062 struct css_set *dst_cset;
2064 dst_cset = find_css_set(src_cset,
2065 dst_cgrp ?: src_cset->dfl_cgrp);
2069 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2072 * If src cset equals dst, it's noop. Drop the src.
2073 * cgroup_migrate() will skip the cset too. Note that we
2074 * can't handle src == dst as some nodes are used by both.
2076 if (src_cset == dst_cset) {
2077 src_cset->mg_src_cgrp = NULL;
2078 list_del_init(&src_cset->mg_preload_node);
2079 put_css_set(src_cset, false);
2080 put_css_set(dst_cset, false);
2084 src_cset->mg_dst_cset = dst_cset;
2086 if (list_empty(&dst_cset->mg_preload_node))
2087 list_add(&dst_cset->mg_preload_node, &csets);
2089 put_css_set(dst_cset, false);
2092 list_splice_tail(&csets, preloaded_csets);
2095 cgroup_migrate_finish(&csets);
2100 * cgroup_migrate - migrate a process or task to a cgroup
2101 * @cgrp: the destination cgroup
2102 * @leader: the leader of the process or the task to migrate
2103 * @threadgroup: whether @leader points to the whole process or a single task
2105 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2106 * process, the caller must be holding threadgroup_lock of @leader. The
2107 * caller is also responsible for invoking cgroup_migrate_add_src() and
2108 * cgroup_migrate_prepare_dst() on the targets before invoking this
2109 * function and following up with cgroup_migrate_finish().
2111 * As long as a controller's ->can_attach() doesn't fail, this function is
2112 * guaranteed to succeed. This means that, excluding ->can_attach()
2113 * failure, when migrating multiple targets, the success or failure can be
2114 * decided for all targets by invoking group_migrate_prepare_dst() before
2115 * actually starting migrating.
2117 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2120 struct cgroup_taskset tset = {
2121 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2122 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2123 .csets = &tset.src_csets,
2125 struct cgroup_subsys_state *css, *failed_css = NULL;
2126 struct css_set *cset, *tmp_cset;
2127 struct task_struct *task, *tmp_task;
2131 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2132 * already PF_EXITING could be freed from underneath us unless we
2133 * take an rcu_read_lock.
2135 down_write(&css_set_rwsem);
2139 /* @task either already exited or can't exit until the end */
2140 if (task->flags & PF_EXITING)
2143 /* leave @task alone if post_fork() hasn't linked it yet */
2144 if (list_empty(&task->cg_list))
2147 cset = task_css_set(task);
2148 if (!cset->mg_src_cgrp)
2152 * cgroup_taskset_first() must always return the leader.
2153 * Take care to avoid disturbing the ordering.
2155 list_move_tail(&task->cg_list, &cset->mg_tasks);
2156 if (list_empty(&cset->mg_node))
2157 list_add_tail(&cset->mg_node, &tset.src_csets);
2158 if (list_empty(&cset->mg_dst_cset->mg_node))
2159 list_move_tail(&cset->mg_dst_cset->mg_node,
2164 } while_each_thread(leader, task);
2166 up_write(&css_set_rwsem);
2168 /* methods shouldn't be called if no task is actually migrating */
2169 if (list_empty(&tset.src_csets))
2172 /* check that we can legitimately attach to the cgroup */
2173 for_each_e_css(css, i, cgrp) {
2174 if (css->ss->can_attach) {
2175 ret = css->ss->can_attach(css, &tset);
2178 goto out_cancel_attach;
2184 * Now that we're guaranteed success, proceed to move all tasks to
2185 * the new cgroup. There are no failure cases after here, so this
2186 * is the commit point.
2188 down_write(&css_set_rwsem);
2189 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2190 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2191 cgroup_task_migrate(cset->mg_src_cgrp, task,
2194 up_write(&css_set_rwsem);
2197 * Migration is committed, all target tasks are now on dst_csets.
2198 * Nothing is sensitive to fork() after this point. Notify
2199 * controllers that migration is complete.
2201 tset.csets = &tset.dst_csets;
2203 for_each_e_css(css, i, cgrp)
2204 if (css->ss->attach)
2205 css->ss->attach(css, &tset);
2208 goto out_release_tset;
2211 for_each_e_css(css, i, cgrp) {
2212 if (css == failed_css)
2214 if (css->ss->cancel_attach)
2215 css->ss->cancel_attach(css, &tset);
2218 down_write(&css_set_rwsem);
2219 list_splice_init(&tset.dst_csets, &tset.src_csets);
2220 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2221 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2222 list_del_init(&cset->mg_node);
2224 up_write(&css_set_rwsem);
2229 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2230 * @dst_cgrp: the cgroup to attach to
2231 * @leader: the task or the leader of the threadgroup to be attached
2232 * @threadgroup: attach the whole threadgroup?
2234 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2236 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2237 struct task_struct *leader, bool threadgroup)
2239 LIST_HEAD(preloaded_csets);
2240 struct task_struct *task;
2243 /* look up all src csets */
2244 down_read(&css_set_rwsem);
2248 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2252 } while_each_thread(leader, task);
2254 up_read(&css_set_rwsem);
2256 /* prepare dst csets and commit */
2257 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2259 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2261 cgroup_migrate_finish(&preloaded_csets);
2266 * Find the task_struct of the task to attach by vpid and pass it along to the
2267 * function to attach either it or all tasks in its threadgroup. Will lock
2268 * cgroup_mutex and threadgroup.
2270 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2271 size_t nbytes, loff_t off, bool threadgroup)
2273 struct task_struct *tsk;
2274 const struct cred *cred = current_cred(), *tcred;
2275 struct cgroup *cgrp;
2279 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2282 cgrp = cgroup_kn_lock_live(of->kn);
2289 tsk = find_task_by_vpid(pid);
2293 goto out_unlock_cgroup;
2296 * even if we're attaching all tasks in the thread group, we
2297 * only need to check permissions on one of them.
2299 tcred = __task_cred(tsk);
2300 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2301 !uid_eq(cred->euid, tcred->uid) &&
2302 !uid_eq(cred->euid, tcred->suid)) {
2305 goto out_unlock_cgroup;
2311 tsk = tsk->group_leader;
2314 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2315 * trapped in a cpuset, or RT worker may be born in a cgroup
2316 * with no rt_runtime allocated. Just say no.
2318 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2321 goto out_unlock_cgroup;
2324 get_task_struct(tsk);
2327 threadgroup_lock(tsk);
2329 if (!thread_group_leader(tsk)) {
2331 * a race with de_thread from another thread's exec()
2332 * may strip us of our leadership, if this happens,
2333 * there is no choice but to throw this task away and
2334 * try again; this is
2335 * "double-double-toil-and-trouble-check locking".
2337 threadgroup_unlock(tsk);
2338 put_task_struct(tsk);
2339 goto retry_find_task;
2343 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2345 threadgroup_unlock(tsk);
2347 put_task_struct(tsk);
2349 cgroup_kn_unlock(of->kn);
2350 return ret ?: nbytes;
2354 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2355 * @from: attach to all cgroups of a given task
2356 * @tsk: the task to be attached
2358 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2360 struct cgroup_root *root;
2363 mutex_lock(&cgroup_mutex);
2364 for_each_root(root) {
2365 struct cgroup *from_cgrp;
2367 if (root == &cgrp_dfl_root)
2370 down_read(&css_set_rwsem);
2371 from_cgrp = task_cgroup_from_root(from, root);
2372 up_read(&css_set_rwsem);
2374 retval = cgroup_attach_task(from_cgrp, tsk, false);
2378 mutex_unlock(&cgroup_mutex);
2382 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2384 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2385 char *buf, size_t nbytes, loff_t off)
2387 return __cgroup_procs_write(of, buf, nbytes, off, false);
2390 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2391 char *buf, size_t nbytes, loff_t off)
2393 return __cgroup_procs_write(of, buf, nbytes, off, true);
2396 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2397 char *buf, size_t nbytes, loff_t off)
2399 struct cgroup *cgrp;
2401 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2403 cgrp = cgroup_kn_lock_live(of->kn);
2406 spin_lock(&release_agent_path_lock);
2407 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2408 sizeof(cgrp->root->release_agent_path));
2409 spin_unlock(&release_agent_path_lock);
2410 cgroup_kn_unlock(of->kn);
2414 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2416 struct cgroup *cgrp = seq_css(seq)->cgroup;
2418 spin_lock(&release_agent_path_lock);
2419 seq_puts(seq, cgrp->root->release_agent_path);
2420 spin_unlock(&release_agent_path_lock);
2421 seq_putc(seq, '\n');
2425 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2427 struct cgroup *cgrp = seq_css(seq)->cgroup;
2429 seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2433 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2435 struct cgroup_subsys *ss;
2436 bool printed = false;
2439 for_each_subsys(ss, ssid) {
2440 if (ss_mask & (1 << ssid)) {
2443 seq_printf(seq, "%s", ss->name);
2448 seq_putc(seq, '\n');
2451 /* show controllers which are currently attached to the default hierarchy */
2452 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2454 struct cgroup *cgrp = seq_css(seq)->cgroup;
2456 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2457 ~cgrp_dfl_root_inhibit_ss_mask);
2461 /* show controllers which are enabled from the parent */
2462 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2464 struct cgroup *cgrp = seq_css(seq)->cgroup;
2466 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2470 /* show controllers which are enabled for a given cgroup's children */
2471 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2473 struct cgroup *cgrp = seq_css(seq)->cgroup;
2475 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2480 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2481 * @cgrp: root of the subtree to update csses for
2483 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2484 * css associations need to be updated accordingly. This function looks up
2485 * all css_sets which are attached to the subtree, creates the matching
2486 * updated css_sets and migrates the tasks to the new ones.
2488 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2490 LIST_HEAD(preloaded_csets);
2491 struct cgroup_subsys_state *css;
2492 struct css_set *src_cset;
2495 lockdep_assert_held(&cgroup_mutex);
2497 /* look up all csses currently attached to @cgrp's subtree */
2498 down_read(&css_set_rwsem);
2499 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2500 struct cgrp_cset_link *link;
2502 /* self is not affected by child_subsys_mask change */
2503 if (css->cgroup == cgrp)
2506 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2507 cgroup_migrate_add_src(link->cset, cgrp,
2510 up_read(&css_set_rwsem);
2512 /* NULL dst indicates self on default hierarchy */
2513 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2517 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2518 struct task_struct *last_task = NULL, *task;
2520 /* src_csets precede dst_csets, break on the first dst_cset */
2521 if (!src_cset->mg_src_cgrp)
2525 * All tasks in src_cset need to be migrated to the
2526 * matching dst_cset. Empty it process by process. We
2527 * walk tasks but migrate processes. The leader might even
2528 * belong to a different cset but such src_cset would also
2529 * be among the target src_csets because the default
2530 * hierarchy enforces per-process membership.
2533 down_read(&css_set_rwsem);
2534 task = list_first_entry_or_null(&src_cset->tasks,
2535 struct task_struct, cg_list);
2537 task = task->group_leader;
2538 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2539 get_task_struct(task);
2541 up_read(&css_set_rwsem);
2546 /* guard against possible infinite loop */
2547 if (WARN(last_task == task,
2548 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2552 threadgroup_lock(task);
2553 /* raced against de_thread() from another thread? */
2554 if (!thread_group_leader(task)) {
2555 threadgroup_unlock(task);
2556 put_task_struct(task);
2560 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2562 threadgroup_unlock(task);
2563 put_task_struct(task);
2565 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2571 cgroup_migrate_finish(&preloaded_csets);
2575 /* change the enabled child controllers for a cgroup in the default hierarchy */
2576 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2577 char *buf, size_t nbytes,
2580 unsigned int enable = 0, disable = 0;
2581 unsigned int css_enable, css_disable, old_ctrl, new_ctrl;
2582 struct cgroup *cgrp, *child;
2583 struct cgroup_subsys *ss;
2588 * Parse input - space separated list of subsystem names prefixed
2589 * with either + or -.
2591 buf = strstrip(buf);
2592 while ((tok = strsep(&buf, " "))) {
2595 for_each_subsys(ss, ssid) {
2596 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2597 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2601 enable |= 1 << ssid;
2602 disable &= ~(1 << ssid);
2603 } else if (*tok == '-') {
2604 disable |= 1 << ssid;
2605 enable &= ~(1 << ssid);
2611 if (ssid == CGROUP_SUBSYS_COUNT)
2615 cgrp = cgroup_kn_lock_live(of->kn);
2619 for_each_subsys(ss, ssid) {
2620 if (enable & (1 << ssid)) {
2621 if (cgrp->subtree_control & (1 << ssid)) {
2622 enable &= ~(1 << ssid);
2626 /* unavailable or not enabled on the parent? */
2627 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2628 (cgroup_parent(cgrp) &&
2629 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2635 * @ss is already enabled through dependency and
2636 * we'll just make it visible. Skip draining.
2638 if (cgrp->child_subsys_mask & (1 << ssid))
2642 * Because css offlining is asynchronous, userland
2643 * might try to re-enable the same controller while
2644 * the previous instance is still around. In such
2645 * cases, wait till it's gone using offline_waitq.
2647 cgroup_for_each_live_child(child, cgrp) {
2650 if (!cgroup_css(child, ss))
2654 prepare_to_wait(&child->offline_waitq, &wait,
2655 TASK_UNINTERRUPTIBLE);
2656 cgroup_kn_unlock(of->kn);
2658 finish_wait(&child->offline_waitq, &wait);
2661 return restart_syscall();
2663 } else if (disable & (1 << ssid)) {
2664 if (!(cgrp->subtree_control & (1 << ssid))) {
2665 disable &= ~(1 << ssid);
2669 /* a child has it enabled? */
2670 cgroup_for_each_live_child(child, cgrp) {
2671 if (child->subtree_control & (1 << ssid)) {
2679 if (!enable && !disable) {
2685 * Except for the root, subtree_control must be zero for a cgroup
2686 * with tasks so that child cgroups don't compete against tasks.
2688 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2694 * Update subsys masks and calculate what needs to be done. More
2695 * subsystems than specified may need to be enabled or disabled
2696 * depending on subsystem dependencies.
2698 cgrp->subtree_control |= enable;
2699 cgrp->subtree_control &= ~disable;
2701 old_ctrl = cgrp->child_subsys_mask;
2702 cgroup_refresh_child_subsys_mask(cgrp);
2703 new_ctrl = cgrp->child_subsys_mask;
2705 css_enable = ~old_ctrl & new_ctrl;
2706 css_disable = old_ctrl & ~new_ctrl;
2707 enable |= css_enable;
2708 disable |= css_disable;
2711 * Create new csses or make the existing ones visible. A css is
2712 * created invisible if it's being implicitly enabled through
2713 * dependency. An invisible css is made visible when the userland
2714 * explicitly enables it.
2716 for_each_subsys(ss, ssid) {
2717 if (!(enable & (1 << ssid)))
2720 cgroup_for_each_live_child(child, cgrp) {
2721 if (css_enable & (1 << ssid))
2722 ret = create_css(child, ss,
2723 cgrp->subtree_control & (1 << ssid));
2725 ret = cgroup_populate_dir(child, 1 << ssid);
2732 * At this point, cgroup_e_css() results reflect the new csses
2733 * making the following cgroup_update_dfl_csses() properly update
2734 * css associations of all tasks in the subtree.
2736 ret = cgroup_update_dfl_csses(cgrp);
2741 * All tasks are migrated out of disabled csses. Kill or hide
2742 * them. A css is hidden when the userland requests it to be
2743 * disabled while other subsystems are still depending on it.
2745 for_each_subsys(ss, ssid) {
2746 if (!(disable & (1 << ssid)))
2749 cgroup_for_each_live_child(child, cgrp) {
2750 if (css_disable & (1 << ssid))
2751 kill_css(cgroup_css(child, ss));
2753 cgroup_clear_dir(child, 1 << ssid);
2757 kernfs_activate(cgrp->kn);
2760 cgroup_kn_unlock(of->kn);
2761 return ret ?: nbytes;
2764 cgrp->subtree_control &= ~enable;
2765 cgrp->subtree_control |= disable;
2766 cgroup_refresh_child_subsys_mask(cgrp);
2768 for_each_subsys(ss, ssid) {
2769 if (!(enable & (1 << ssid)))
2772 cgroup_for_each_live_child(child, cgrp) {
2773 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2778 if (css_enable & (1 << ssid))
2781 cgroup_clear_dir(child, 1 << ssid);
2787 static int cgroup_populated_show(struct seq_file *seq, void *v)
2789 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2793 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2794 size_t nbytes, loff_t off)
2796 struct cgroup *cgrp = of->kn->parent->priv;
2797 struct cftype *cft = of->kn->priv;
2798 struct cgroup_subsys_state *css;
2802 return cft->write(of, buf, nbytes, off);
2805 * kernfs guarantees that a file isn't deleted with operations in
2806 * flight, which means that the matching css is and stays alive and
2807 * doesn't need to be pinned. The RCU locking is not necessary
2808 * either. It's just for the convenience of using cgroup_css().
2811 css = cgroup_css(cgrp, cft->ss);
2814 if (cft->write_u64) {
2815 unsigned long long v;
2816 ret = kstrtoull(buf, 0, &v);
2818 ret = cft->write_u64(css, cft, v);
2819 } else if (cft->write_s64) {
2821 ret = kstrtoll(buf, 0, &v);
2823 ret = cft->write_s64(css, cft, v);
2828 return ret ?: nbytes;
2831 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2833 return seq_cft(seq)->seq_start(seq, ppos);
2836 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2838 return seq_cft(seq)->seq_next(seq, v, ppos);
2841 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2843 seq_cft(seq)->seq_stop(seq, v);
2846 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2848 struct cftype *cft = seq_cft(m);
2849 struct cgroup_subsys_state *css = seq_css(m);
2852 return cft->seq_show(m, arg);
2855 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2856 else if (cft->read_s64)
2857 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2863 static struct kernfs_ops cgroup_kf_single_ops = {
2864 .atomic_write_len = PAGE_SIZE,
2865 .write = cgroup_file_write,
2866 .seq_show = cgroup_seqfile_show,
2869 static struct kernfs_ops cgroup_kf_ops = {
2870 .atomic_write_len = PAGE_SIZE,
2871 .write = cgroup_file_write,
2872 .seq_start = cgroup_seqfile_start,
2873 .seq_next = cgroup_seqfile_next,
2874 .seq_stop = cgroup_seqfile_stop,
2875 .seq_show = cgroup_seqfile_show,
2879 * cgroup_rename - Only allow simple rename of directories in place.
2881 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2882 const char *new_name_str)
2884 struct cgroup *cgrp = kn->priv;
2887 if (kernfs_type(kn) != KERNFS_DIR)
2889 if (kn->parent != new_parent)
2893 * This isn't a proper migration and its usefulness is very
2894 * limited. Disallow if sane_behavior.
2896 if (cgroup_sane_behavior(cgrp))
2900 * We're gonna grab cgroup_mutex which nests outside kernfs
2901 * active_ref. kernfs_rename() doesn't require active_ref
2902 * protection. Break them before grabbing cgroup_mutex.
2904 kernfs_break_active_protection(new_parent);
2905 kernfs_break_active_protection(kn);
2907 mutex_lock(&cgroup_mutex);
2909 ret = kernfs_rename(kn, new_parent, new_name_str);
2911 mutex_unlock(&cgroup_mutex);
2913 kernfs_unbreak_active_protection(kn);
2914 kernfs_unbreak_active_protection(new_parent);
2918 /* set uid and gid of cgroup dirs and files to that of the creator */
2919 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
2921 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
2922 .ia_uid = current_fsuid(),
2923 .ia_gid = current_fsgid(), };
2925 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
2926 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
2929 return kernfs_setattr(kn, &iattr);
2932 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2934 char name[CGROUP_FILE_NAME_MAX];
2935 struct kernfs_node *kn;
2936 struct lock_class_key *key = NULL;
2939 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2940 key = &cft->lockdep_key;
2942 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
2943 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
2948 ret = cgroup_kn_set_ugid(kn);
2954 if (cft->seq_show == cgroup_populated_show)
2955 cgrp->populated_kn = kn;
2960 * cgroup_addrm_files - add or remove files to a cgroup directory
2961 * @cgrp: the target cgroup
2962 * @cfts: array of cftypes to be added
2963 * @is_add: whether to add or remove
2965 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2966 * For removals, this function never fails. If addition fails, this
2967 * function doesn't remove files already added. The caller is responsible
2970 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2976 lockdep_assert_held(&cgroup_mutex);
2978 for (cft = cfts; cft->name[0] != '\0'; cft++) {
2979 /* does cft->flags tell us to skip this file on @cgrp? */
2980 if ((cft->flags & CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
2982 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2984 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
2986 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
2990 ret = cgroup_add_file(cgrp, cft);
2992 pr_warn("%s: failed to add %s, err=%d\n",
2993 __func__, cft->name, ret);
2997 cgroup_rm_file(cgrp, cft);
3003 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3006 struct cgroup_subsys *ss = cfts[0].ss;
3007 struct cgroup *root = &ss->root->cgrp;
3008 struct cgroup_subsys_state *css;
3011 lockdep_assert_held(&cgroup_mutex);
3013 /* add/rm files for all cgroups created before */
3014 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3015 struct cgroup *cgrp = css->cgroup;
3017 if (cgroup_is_dead(cgrp))
3020 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3026 kernfs_activate(root->kn);
3030 static void cgroup_exit_cftypes(struct cftype *cfts)
3034 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3035 /* free copy for custom atomic_write_len, see init_cftypes() */
3036 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3043 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3047 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3048 struct kernfs_ops *kf_ops;
3050 WARN_ON(cft->ss || cft->kf_ops);
3053 kf_ops = &cgroup_kf_ops;
3055 kf_ops = &cgroup_kf_single_ops;
3058 * Ugh... if @cft wants a custom max_write_len, we need to
3059 * make a copy of kf_ops to set its atomic_write_len.
3061 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3062 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3064 cgroup_exit_cftypes(cfts);
3067 kf_ops->atomic_write_len = cft->max_write_len;
3070 cft->kf_ops = kf_ops;
3077 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3079 lockdep_assert_held(&cgroup_mutex);
3081 if (!cfts || !cfts[0].ss)
3084 list_del(&cfts->node);
3085 cgroup_apply_cftypes(cfts, false);
3086 cgroup_exit_cftypes(cfts);
3091 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3092 * @cfts: zero-length name terminated array of cftypes
3094 * Unregister @cfts. Files described by @cfts are removed from all
3095 * existing cgroups and all future cgroups won't have them either. This
3096 * function can be called anytime whether @cfts' subsys is attached or not.
3098 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3101 int cgroup_rm_cftypes(struct cftype *cfts)
3105 mutex_lock(&cgroup_mutex);
3106 ret = cgroup_rm_cftypes_locked(cfts);
3107 mutex_unlock(&cgroup_mutex);
3112 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3113 * @ss: target cgroup subsystem
3114 * @cfts: zero-length name terminated array of cftypes
3116 * Register @cfts to @ss. Files described by @cfts are created for all
3117 * existing cgroups to which @ss is attached and all future cgroups will
3118 * have them too. This function can be called anytime whether @ss is
3121 * Returns 0 on successful registration, -errno on failure. Note that this
3122 * function currently returns 0 as long as @cfts registration is successful
3123 * even if some file creation attempts on existing cgroups fail.
3125 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3132 if (!cfts || cfts[0].name[0] == '\0')
3135 ret = cgroup_init_cftypes(ss, cfts);
3139 mutex_lock(&cgroup_mutex);
3141 list_add_tail(&cfts->node, &ss->cfts);
3142 ret = cgroup_apply_cftypes(cfts, true);
3144 cgroup_rm_cftypes_locked(cfts);
3146 mutex_unlock(&cgroup_mutex);
3151 * cgroup_task_count - count the number of tasks in a cgroup.
3152 * @cgrp: the cgroup in question
3154 * Return the number of tasks in the cgroup.
3156 static int cgroup_task_count(const struct cgroup *cgrp)
3159 struct cgrp_cset_link *link;
3161 down_read(&css_set_rwsem);
3162 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3163 count += atomic_read(&link->cset->refcount);
3164 up_read(&css_set_rwsem);
3169 * css_next_child - find the next child of a given css
3170 * @pos: the current position (%NULL to initiate traversal)
3171 * @parent: css whose children to walk
3173 * This function returns the next child of @parent and should be called
3174 * under either cgroup_mutex or RCU read lock. The only requirement is
3175 * that @parent and @pos are accessible. The next sibling is guaranteed to
3176 * be returned regardless of their states.
3178 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3179 * css which finished ->css_online() is guaranteed to be visible in the
3180 * future iterations and will stay visible until the last reference is put.
3181 * A css which hasn't finished ->css_online() or already finished
3182 * ->css_offline() may show up during traversal. It's each subsystem's
3183 * responsibility to synchronize against on/offlining.
3185 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3186 struct cgroup_subsys_state *parent)
3188 struct cgroup_subsys_state *next;
3190 cgroup_assert_mutex_or_rcu_locked();
3193 * @pos could already have been unlinked from the sibling list.
3194 * Once a cgroup is removed, its ->sibling.next is no longer
3195 * updated when its next sibling changes. CSS_RELEASED is set when
3196 * @pos is taken off list, at which time its next pointer is valid,
3197 * and, as releases are serialized, the one pointed to by the next
3198 * pointer is guaranteed to not have started release yet. This
3199 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3200 * critical section, the one pointed to by its next pointer is
3201 * guaranteed to not have finished its RCU grace period even if we
3202 * have dropped rcu_read_lock() inbetween iterations.
3204 * If @pos has CSS_RELEASED set, its next pointer can't be
3205 * dereferenced; however, as each css is given a monotonically
3206 * increasing unique serial number and always appended to the
3207 * sibling list, the next one can be found by walking the parent's
3208 * children until the first css with higher serial number than
3209 * @pos's. While this path can be slower, it happens iff iteration
3210 * races against release and the race window is very small.
3213 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3214 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3215 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3217 list_for_each_entry_rcu(next, &parent->children, sibling)
3218 if (next->serial_nr > pos->serial_nr)
3223 * @next, if not pointing to the head, can be dereferenced and is
3226 if (&next->sibling != &parent->children)
3232 * css_next_descendant_pre - find the next descendant for pre-order walk
3233 * @pos: the current position (%NULL to initiate traversal)
3234 * @root: css whose descendants to walk
3236 * To be used by css_for_each_descendant_pre(). Find the next descendant
3237 * to visit for pre-order traversal of @root's descendants. @root is
3238 * included in the iteration and the first node to be visited.
3240 * While this function requires cgroup_mutex or RCU read locking, it
3241 * doesn't require the whole traversal to be contained in a single critical
3242 * section. This function will return the correct next descendant as long
3243 * as both @pos and @root are accessible and @pos is a descendant of @root.
3245 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3246 * css which finished ->css_online() is guaranteed to be visible in the
3247 * future iterations and will stay visible until the last reference is put.
3248 * A css which hasn't finished ->css_online() or already finished
3249 * ->css_offline() may show up during traversal. It's each subsystem's
3250 * responsibility to synchronize against on/offlining.
3252 struct cgroup_subsys_state *
3253 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3254 struct cgroup_subsys_state *root)
3256 struct cgroup_subsys_state *next;
3258 cgroup_assert_mutex_or_rcu_locked();
3260 /* if first iteration, visit @root */
3264 /* visit the first child if exists */
3265 next = css_next_child(NULL, pos);
3269 /* no child, visit my or the closest ancestor's next sibling */
3270 while (pos != root) {
3271 next = css_next_child(pos, pos->parent);
3281 * css_rightmost_descendant - return the rightmost descendant of a css
3282 * @pos: css of interest
3284 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3285 * is returned. This can be used during pre-order traversal to skip
3288 * While this function requires cgroup_mutex or RCU read locking, it
3289 * doesn't require the whole traversal to be contained in a single critical
3290 * section. This function will return the correct rightmost descendant as
3291 * long as @pos is accessible.
3293 struct cgroup_subsys_state *
3294 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3296 struct cgroup_subsys_state *last, *tmp;
3298 cgroup_assert_mutex_or_rcu_locked();
3302 /* ->prev isn't RCU safe, walk ->next till the end */
3304 css_for_each_child(tmp, last)
3311 static struct cgroup_subsys_state *
3312 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3314 struct cgroup_subsys_state *last;
3318 pos = css_next_child(NULL, pos);
3325 * css_next_descendant_post - find the next descendant for post-order walk
3326 * @pos: the current position (%NULL to initiate traversal)
3327 * @root: css whose descendants to walk
3329 * To be used by css_for_each_descendant_post(). Find the next descendant
3330 * to visit for post-order traversal of @root's descendants. @root is
3331 * included in the iteration and the last node to be visited.
3333 * While this function requires cgroup_mutex or RCU read locking, it
3334 * doesn't require the whole traversal to be contained in a single critical
3335 * section. This function will return the correct next descendant as long
3336 * as both @pos and @cgroup are accessible and @pos is a descendant of
3339 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3340 * css which finished ->css_online() is guaranteed to be visible in the
3341 * future iterations and will stay visible until the last reference is put.
3342 * A css which hasn't finished ->css_online() or already finished
3343 * ->css_offline() may show up during traversal. It's each subsystem's
3344 * responsibility to synchronize against on/offlining.
3346 struct cgroup_subsys_state *
3347 css_next_descendant_post(struct cgroup_subsys_state *pos,
3348 struct cgroup_subsys_state *root)
3350 struct cgroup_subsys_state *next;
3352 cgroup_assert_mutex_or_rcu_locked();
3354 /* if first iteration, visit leftmost descendant which may be @root */
3356 return css_leftmost_descendant(root);
3358 /* if we visited @root, we're done */
3362 /* if there's an unvisited sibling, visit its leftmost descendant */
3363 next = css_next_child(pos, pos->parent);
3365 return css_leftmost_descendant(next);
3367 /* no sibling left, visit parent */
3372 * css_has_online_children - does a css have online children
3373 * @css: the target css
3375 * Returns %true if @css has any online children; otherwise, %false. This
3376 * function can be called from any context but the caller is responsible
3377 * for synchronizing against on/offlining as necessary.
3379 bool css_has_online_children(struct cgroup_subsys_state *css)
3381 struct cgroup_subsys_state *child;
3385 css_for_each_child(child, css) {
3386 if (css->flags & CSS_ONLINE) {
3396 * css_advance_task_iter - advance a task itererator to the next css_set
3397 * @it: the iterator to advance
3399 * Advance @it to the next css_set to walk.
3401 static void css_advance_task_iter(struct css_task_iter *it)
3403 struct list_head *l = it->cset_pos;
3404 struct cgrp_cset_link *link;
3405 struct css_set *cset;
3407 /* Advance to the next non-empty css_set */
3410 if (l == it->cset_head) {
3411 it->cset_pos = NULL;
3416 cset = container_of(l, struct css_set,
3417 e_cset_node[it->ss->id]);
3419 link = list_entry(l, struct cgrp_cset_link, cset_link);
3422 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3426 if (!list_empty(&cset->tasks))
3427 it->task_pos = cset->tasks.next;
3429 it->task_pos = cset->mg_tasks.next;
3431 it->tasks_head = &cset->tasks;
3432 it->mg_tasks_head = &cset->mg_tasks;
3436 * css_task_iter_start - initiate task iteration
3437 * @css: the css to walk tasks of
3438 * @it: the task iterator to use
3440 * Initiate iteration through the tasks of @css. The caller can call
3441 * css_task_iter_next() to walk through the tasks until the function
3442 * returns NULL. On completion of iteration, css_task_iter_end() must be
3445 * Note that this function acquires a lock which is released when the
3446 * iteration finishes. The caller can't sleep while iteration is in
3449 void css_task_iter_start(struct cgroup_subsys_state *css,
3450 struct css_task_iter *it)
3451 __acquires(css_set_rwsem)
3453 /* no one should try to iterate before mounting cgroups */
3454 WARN_ON_ONCE(!use_task_css_set_links);
3456 down_read(&css_set_rwsem);
3461 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3463 it->cset_pos = &css->cgroup->cset_links;
3465 it->cset_head = it->cset_pos;
3467 css_advance_task_iter(it);
3471 * css_task_iter_next - return the next task for the iterator
3472 * @it: the task iterator being iterated
3474 * The "next" function for task iteration. @it should have been
3475 * initialized via css_task_iter_start(). Returns NULL when the iteration
3478 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3480 struct task_struct *res;
3481 struct list_head *l = it->task_pos;
3483 /* If the iterator cg is NULL, we have no tasks */
3486 res = list_entry(l, struct task_struct, cg_list);
3489 * Advance iterator to find next entry. cset->tasks is consumed
3490 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3495 if (l == it->tasks_head)
3496 l = it->mg_tasks_head->next;
3498 if (l == it->mg_tasks_head)
3499 css_advance_task_iter(it);
3507 * css_task_iter_end - finish task iteration
3508 * @it: the task iterator to finish
3510 * Finish task iteration started by css_task_iter_start().
3512 void css_task_iter_end(struct css_task_iter *it)
3513 __releases(css_set_rwsem)
3515 up_read(&css_set_rwsem);
3519 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3520 * @to: cgroup to which the tasks will be moved
3521 * @from: cgroup in which the tasks currently reside
3523 * Locking rules between cgroup_post_fork() and the migration path
3524 * guarantee that, if a task is forking while being migrated, the new child
3525 * is guaranteed to be either visible in the source cgroup after the
3526 * parent's migration is complete or put into the target cgroup. No task
3527 * can slip out of migration through forking.
3529 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3531 LIST_HEAD(preloaded_csets);
3532 struct cgrp_cset_link *link;
3533 struct css_task_iter it;
3534 struct task_struct *task;
3537 mutex_lock(&cgroup_mutex);
3539 /* all tasks in @from are being moved, all csets are source */
3540 down_read(&css_set_rwsem);
3541 list_for_each_entry(link, &from->cset_links, cset_link)
3542 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3543 up_read(&css_set_rwsem);
3545 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3550 * Migrate tasks one-by-one until @form is empty. This fails iff
3551 * ->can_attach() fails.
3554 css_task_iter_start(&from->self, &it);
3555 task = css_task_iter_next(&it);
3557 get_task_struct(task);
3558 css_task_iter_end(&it);
3561 ret = cgroup_migrate(to, task, false);
3562 put_task_struct(task);
3564 } while (task && !ret);
3566 cgroup_migrate_finish(&preloaded_csets);
3567 mutex_unlock(&cgroup_mutex);
3572 * Stuff for reading the 'tasks'/'procs' files.
3574 * Reading this file can return large amounts of data if a cgroup has
3575 * *lots* of attached tasks. So it may need several calls to read(),
3576 * but we cannot guarantee that the information we produce is correct
3577 * unless we produce it entirely atomically.
3581 /* which pidlist file are we talking about? */
3582 enum cgroup_filetype {
3588 * A pidlist is a list of pids that virtually represents the contents of one
3589 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3590 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3593 struct cgroup_pidlist {
3595 * used to find which pidlist is wanted. doesn't change as long as
3596 * this particular list stays in the list.
3598 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3601 /* how many elements the above list has */
3603 /* each of these stored in a list by its cgroup */
3604 struct list_head links;
3605 /* pointer to the cgroup we belong to, for list removal purposes */
3606 struct cgroup *owner;
3607 /* for delayed destruction */
3608 struct delayed_work destroy_dwork;
3612 * The following two functions "fix" the issue where there are more pids
3613 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3614 * TODO: replace with a kernel-wide solution to this problem
3616 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3617 static void *pidlist_allocate(int count)
3619 if (PIDLIST_TOO_LARGE(count))
3620 return vmalloc(count * sizeof(pid_t));
3622 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3625 static void pidlist_free(void *p)
3627 if (is_vmalloc_addr(p))
3634 * Used to destroy all pidlists lingering waiting for destroy timer. None
3635 * should be left afterwards.
3637 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3639 struct cgroup_pidlist *l, *tmp_l;
3641 mutex_lock(&cgrp->pidlist_mutex);
3642 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3643 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3644 mutex_unlock(&cgrp->pidlist_mutex);
3646 flush_workqueue(cgroup_pidlist_destroy_wq);
3647 BUG_ON(!list_empty(&cgrp->pidlists));
3650 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3652 struct delayed_work *dwork = to_delayed_work(work);
3653 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3655 struct cgroup_pidlist *tofree = NULL;
3657 mutex_lock(&l->owner->pidlist_mutex);
3660 * Destroy iff we didn't get queued again. The state won't change
3661 * as destroy_dwork can only be queued while locked.
3663 if (!delayed_work_pending(dwork)) {
3664 list_del(&l->links);
3665 pidlist_free(l->list);
3666 put_pid_ns(l->key.ns);
3670 mutex_unlock(&l->owner->pidlist_mutex);
3675 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3676 * Returns the number of unique elements.
3678 static int pidlist_uniq(pid_t *list, int length)
3683 * we presume the 0th element is unique, so i starts at 1. trivial
3684 * edge cases first; no work needs to be done for either
3686 if (length == 0 || length == 1)
3688 /* src and dest walk down the list; dest counts unique elements */
3689 for (src = 1; src < length; src++) {
3690 /* find next unique element */
3691 while (list[src] == list[src-1]) {
3696 /* dest always points to where the next unique element goes */
3697 list[dest] = list[src];
3705 * The two pid files - task and cgroup.procs - guaranteed that the result
3706 * is sorted, which forced this whole pidlist fiasco. As pid order is
3707 * different per namespace, each namespace needs differently sorted list,
3708 * making it impossible to use, for example, single rbtree of member tasks
3709 * sorted by task pointer. As pidlists can be fairly large, allocating one
3710 * per open file is dangerous, so cgroup had to implement shared pool of
3711 * pidlists keyed by cgroup and namespace.
3713 * All this extra complexity was caused by the original implementation
3714 * committing to an entirely unnecessary property. In the long term, we
3715 * want to do away with it. Explicitly scramble sort order if
3716 * sane_behavior so that no such expectation exists in the new interface.
3718 * Scrambling is done by swapping every two consecutive bits, which is
3719 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3721 static pid_t pid_fry(pid_t pid)
3723 unsigned a = pid & 0x55555555;
3724 unsigned b = pid & 0xAAAAAAAA;
3726 return (a << 1) | (b >> 1);
3729 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3731 if (cgroup_sane_behavior(cgrp))
3732 return pid_fry(pid);
3737 static int cmppid(const void *a, const void *b)
3739 return *(pid_t *)a - *(pid_t *)b;
3742 static int fried_cmppid(const void *a, const void *b)
3744 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3747 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3748 enum cgroup_filetype type)
3750 struct cgroup_pidlist *l;
3751 /* don't need task_nsproxy() if we're looking at ourself */
3752 struct pid_namespace *ns = task_active_pid_ns(current);
3754 lockdep_assert_held(&cgrp->pidlist_mutex);
3756 list_for_each_entry(l, &cgrp->pidlists, links)
3757 if (l->key.type == type && l->key.ns == ns)
3763 * find the appropriate pidlist for our purpose (given procs vs tasks)
3764 * returns with the lock on that pidlist already held, and takes care
3765 * of the use count, or returns NULL with no locks held if we're out of
3768 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3769 enum cgroup_filetype type)
3771 struct cgroup_pidlist *l;
3773 lockdep_assert_held(&cgrp->pidlist_mutex);
3775 l = cgroup_pidlist_find(cgrp, type);
3779 /* entry not found; create a new one */
3780 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3784 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3786 /* don't need task_nsproxy() if we're looking at ourself */
3787 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3789 list_add(&l->links, &cgrp->pidlists);
3794 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3796 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3797 struct cgroup_pidlist **lp)
3801 int pid, n = 0; /* used for populating the array */
3802 struct css_task_iter it;
3803 struct task_struct *tsk;
3804 struct cgroup_pidlist *l;
3806 lockdep_assert_held(&cgrp->pidlist_mutex);
3809 * If cgroup gets more users after we read count, we won't have
3810 * enough space - tough. This race is indistinguishable to the
3811 * caller from the case that the additional cgroup users didn't
3812 * show up until sometime later on.
3814 length = cgroup_task_count(cgrp);
3815 array = pidlist_allocate(length);
3818 /* now, populate the array */
3819 css_task_iter_start(&cgrp->self, &it);
3820 while ((tsk = css_task_iter_next(&it))) {
3821 if (unlikely(n == length))
3823 /* get tgid or pid for procs or tasks file respectively */
3824 if (type == CGROUP_FILE_PROCS)
3825 pid = task_tgid_vnr(tsk);
3827 pid = task_pid_vnr(tsk);
3828 if (pid > 0) /* make sure to only use valid results */
3831 css_task_iter_end(&it);
3833 /* now sort & (if procs) strip out duplicates */
3834 if (cgroup_sane_behavior(cgrp))
3835 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3837 sort(array, length, sizeof(pid_t), cmppid, NULL);
3838 if (type == CGROUP_FILE_PROCS)
3839 length = pidlist_uniq(array, length);
3841 l = cgroup_pidlist_find_create(cgrp, type);
3843 mutex_unlock(&cgrp->pidlist_mutex);
3844 pidlist_free(array);
3848 /* store array, freeing old if necessary */
3849 pidlist_free(l->list);
3857 * cgroupstats_build - build and fill cgroupstats
3858 * @stats: cgroupstats to fill information into
3859 * @dentry: A dentry entry belonging to the cgroup for which stats have
3862 * Build and fill cgroupstats so that taskstats can export it to user
3865 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3867 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3868 struct cgroup *cgrp;
3869 struct css_task_iter it;
3870 struct task_struct *tsk;
3872 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3873 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
3874 kernfs_type(kn) != KERNFS_DIR)
3877 mutex_lock(&cgroup_mutex);
3880 * We aren't being called from kernfs and there's no guarantee on
3881 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3882 * @kn->priv is RCU safe. Let's do the RCU dancing.
3885 cgrp = rcu_dereference(kn->priv);
3886 if (!cgrp || cgroup_is_dead(cgrp)) {
3888 mutex_unlock(&cgroup_mutex);
3893 css_task_iter_start(&cgrp->self, &it);
3894 while ((tsk = css_task_iter_next(&it))) {
3895 switch (tsk->state) {
3897 stats->nr_running++;
3899 case TASK_INTERRUPTIBLE:
3900 stats->nr_sleeping++;
3902 case TASK_UNINTERRUPTIBLE:
3903 stats->nr_uninterruptible++;
3906 stats->nr_stopped++;
3909 if (delayacct_is_task_waiting_on_io(tsk))
3910 stats->nr_io_wait++;
3914 css_task_iter_end(&it);
3916 mutex_unlock(&cgroup_mutex);
3922 * seq_file methods for the tasks/procs files. The seq_file position is the
3923 * next pid to display; the seq_file iterator is a pointer to the pid
3924 * in the cgroup->l->list array.
3927 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3930 * Initially we receive a position value that corresponds to
3931 * one more than the last pid shown (or 0 on the first call or
3932 * after a seek to the start). Use a binary-search to find the
3933 * next pid to display, if any
3935 struct kernfs_open_file *of = s->private;
3936 struct cgroup *cgrp = seq_css(s)->cgroup;
3937 struct cgroup_pidlist *l;
3938 enum cgroup_filetype type = seq_cft(s)->private;
3939 int index = 0, pid = *pos;
3942 mutex_lock(&cgrp->pidlist_mutex);
3945 * !NULL @of->priv indicates that this isn't the first start()
3946 * after open. If the matching pidlist is around, we can use that.
3947 * Look for it. Note that @of->priv can't be used directly. It
3948 * could already have been destroyed.
3951 of->priv = cgroup_pidlist_find(cgrp, type);
3954 * Either this is the first start() after open or the matching
3955 * pidlist has been destroyed inbetween. Create a new one.
3958 ret = pidlist_array_load(cgrp, type,
3959 (struct cgroup_pidlist **)&of->priv);
3961 return ERR_PTR(ret);
3966 int end = l->length;
3968 while (index < end) {
3969 int mid = (index + end) / 2;
3970 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3973 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3979 /* If we're off the end of the array, we're done */
3980 if (index >= l->length)
3982 /* Update the abstract position to be the actual pid that we found */
3983 iter = l->list + index;
3984 *pos = cgroup_pid_fry(cgrp, *iter);
3988 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3990 struct kernfs_open_file *of = s->private;
3991 struct cgroup_pidlist *l = of->priv;
3994 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3995 CGROUP_PIDLIST_DESTROY_DELAY);
3996 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3999 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4001 struct kernfs_open_file *of = s->private;
4002 struct cgroup_pidlist *l = of->priv;
4004 pid_t *end = l->list + l->length;
4006 * Advance to the next pid in the array. If this goes off the
4013 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4018 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4020 return seq_printf(s, "%d\n", *(int *)v);
4023 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4026 return notify_on_release(css->cgroup);
4029 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4030 struct cftype *cft, u64 val)
4032 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4034 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4036 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4040 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4043 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4046 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4047 struct cftype *cft, u64 val)
4050 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4052 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4056 static struct cftype cgroup_base_files[] = {
4058 .name = "cgroup.procs",
4059 .seq_start = cgroup_pidlist_start,
4060 .seq_next = cgroup_pidlist_next,
4061 .seq_stop = cgroup_pidlist_stop,
4062 .seq_show = cgroup_pidlist_show,
4063 .private = CGROUP_FILE_PROCS,
4064 .write = cgroup_procs_write,
4065 .mode = S_IRUGO | S_IWUSR,
4068 .name = "cgroup.clone_children",
4069 .flags = CFTYPE_INSANE,
4070 .read_u64 = cgroup_clone_children_read,
4071 .write_u64 = cgroup_clone_children_write,
4074 .name = "cgroup.sane_behavior",
4075 .flags = CFTYPE_ONLY_ON_ROOT,
4076 .seq_show = cgroup_sane_behavior_show,
4079 .name = "cgroup.controllers",
4080 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_ONLY_ON_ROOT,
4081 .seq_show = cgroup_root_controllers_show,
4084 .name = "cgroup.controllers",
4085 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4086 .seq_show = cgroup_controllers_show,
4089 .name = "cgroup.subtree_control",
4090 .flags = CFTYPE_ONLY_ON_DFL,
4091 .seq_show = cgroup_subtree_control_show,
4092 .write = cgroup_subtree_control_write,
4095 .name = "cgroup.populated",
4096 .flags = CFTYPE_ONLY_ON_DFL | CFTYPE_NOT_ON_ROOT,
4097 .seq_show = cgroup_populated_show,
4101 * Historical crazy stuff. These don't have "cgroup." prefix and
4102 * don't exist if sane_behavior. If you're depending on these, be
4103 * prepared to be burned.
4107 .flags = CFTYPE_INSANE, /* use "procs" instead */
4108 .seq_start = cgroup_pidlist_start,
4109 .seq_next = cgroup_pidlist_next,
4110 .seq_stop = cgroup_pidlist_stop,
4111 .seq_show = cgroup_pidlist_show,
4112 .private = CGROUP_FILE_TASKS,
4113 .write = cgroup_tasks_write,
4114 .mode = S_IRUGO | S_IWUSR,
4117 .name = "notify_on_release",
4118 .flags = CFTYPE_INSANE,
4119 .read_u64 = cgroup_read_notify_on_release,
4120 .write_u64 = cgroup_write_notify_on_release,
4123 .name = "release_agent",
4124 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
4125 .seq_show = cgroup_release_agent_show,
4126 .write = cgroup_release_agent_write,
4127 .max_write_len = PATH_MAX - 1,
4133 * cgroup_populate_dir - create subsys files in a cgroup directory
4134 * @cgrp: target cgroup
4135 * @subsys_mask: mask of the subsystem ids whose files should be added
4137 * On failure, no file is added.
4139 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4141 struct cgroup_subsys *ss;
4144 /* process cftsets of each subsystem */
4145 for_each_subsys(ss, i) {
4146 struct cftype *cfts;
4148 if (!(subsys_mask & (1 << i)))
4151 list_for_each_entry(cfts, &ss->cfts, node) {
4152 ret = cgroup_addrm_files(cgrp, cfts, true);
4159 cgroup_clear_dir(cgrp, subsys_mask);
4164 * css destruction is four-stage process.
4166 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4167 * Implemented in kill_css().
4169 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4170 * and thus css_tryget_online() is guaranteed to fail, the css can be
4171 * offlined by invoking offline_css(). After offlining, the base ref is
4172 * put. Implemented in css_killed_work_fn().
4174 * 3. When the percpu_ref reaches zero, the only possible remaining
4175 * accessors are inside RCU read sections. css_release() schedules the
4178 * 4. After the grace period, the css can be freed. Implemented in
4179 * css_free_work_fn().
4181 * It is actually hairier because both step 2 and 4 require process context
4182 * and thus involve punting to css->destroy_work adding two additional
4183 * steps to the already complex sequence.
4185 static void css_free_work_fn(struct work_struct *work)
4187 struct cgroup_subsys_state *css =
4188 container_of(work, struct cgroup_subsys_state, destroy_work);
4189 struct cgroup *cgrp = css->cgroup;
4194 css_put(css->parent);
4196 css->ss->css_free(css);
4199 /* cgroup free path */
4200 atomic_dec(&cgrp->root->nr_cgrps);
4201 cgroup_pidlist_destroy_all(cgrp);
4203 if (cgroup_parent(cgrp)) {
4205 * We get a ref to the parent, and put the ref when
4206 * this cgroup is being freed, so it's guaranteed
4207 * that the parent won't be destroyed before its
4210 cgroup_put(cgroup_parent(cgrp));
4211 kernfs_put(cgrp->kn);
4215 * This is root cgroup's refcnt reaching zero,
4216 * which indicates that the root should be
4219 cgroup_destroy_root(cgrp->root);
4224 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4226 struct cgroup_subsys_state *css =
4227 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4229 INIT_WORK(&css->destroy_work, css_free_work_fn);
4230 queue_work(cgroup_destroy_wq, &css->destroy_work);
4233 static void css_release_work_fn(struct work_struct *work)
4235 struct cgroup_subsys_state *css =
4236 container_of(work, struct cgroup_subsys_state, destroy_work);
4237 struct cgroup_subsys *ss = css->ss;
4238 struct cgroup *cgrp = css->cgroup;
4240 mutex_lock(&cgroup_mutex);
4242 css->flags |= CSS_RELEASED;
4243 list_del_rcu(&css->sibling);
4246 /* css release path */
4247 cgroup_idr_remove(&ss->css_idr, css->id);
4249 /* cgroup release path */
4250 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4254 mutex_unlock(&cgroup_mutex);
4256 call_rcu(&css->rcu_head, css_free_rcu_fn);
4259 static void css_release(struct percpu_ref *ref)
4261 struct cgroup_subsys_state *css =
4262 container_of(ref, struct cgroup_subsys_state, refcnt);
4264 INIT_WORK(&css->destroy_work, css_release_work_fn);
4265 queue_work(cgroup_destroy_wq, &css->destroy_work);
4268 static void init_and_link_css(struct cgroup_subsys_state *css,
4269 struct cgroup_subsys *ss, struct cgroup *cgrp)
4271 lockdep_assert_held(&cgroup_mutex);
4275 memset(css, 0, sizeof(*css));
4278 INIT_LIST_HEAD(&css->sibling);
4279 INIT_LIST_HEAD(&css->children);
4280 css->serial_nr = css_serial_nr_next++;
4282 if (cgroup_parent(cgrp)) {
4283 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4284 css_get(css->parent);
4287 BUG_ON(cgroup_css(cgrp, ss));
4290 /* invoke ->css_online() on a new CSS and mark it online if successful */
4291 static int online_css(struct cgroup_subsys_state *css)
4293 struct cgroup_subsys *ss = css->ss;
4296 lockdep_assert_held(&cgroup_mutex);
4299 ret = ss->css_online(css);
4301 css->flags |= CSS_ONLINE;
4302 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4307 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4308 static void offline_css(struct cgroup_subsys_state *css)
4310 struct cgroup_subsys *ss = css->ss;
4312 lockdep_assert_held(&cgroup_mutex);
4314 if (!(css->flags & CSS_ONLINE))
4317 if (ss->css_offline)
4318 ss->css_offline(css);
4320 css->flags &= ~CSS_ONLINE;
4321 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4323 wake_up_all(&css->cgroup->offline_waitq);
4327 * create_css - create a cgroup_subsys_state
4328 * @cgrp: the cgroup new css will be associated with
4329 * @ss: the subsys of new css
4330 * @visible: whether to create control knobs for the new css or not
4332 * Create a new css associated with @cgrp - @ss pair. On success, the new
4333 * css is online and installed in @cgrp with all interface files created if
4334 * @visible. Returns 0 on success, -errno on failure.
4336 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4339 struct cgroup *parent = cgroup_parent(cgrp);
4340 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4341 struct cgroup_subsys_state *css;
4344 lockdep_assert_held(&cgroup_mutex);
4346 css = ss->css_alloc(parent_css);
4348 return PTR_ERR(css);
4350 init_and_link_css(css, ss, cgrp);
4352 err = percpu_ref_init(&css->refcnt, css_release);
4356 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4358 goto err_free_percpu_ref;
4362 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4367 /* @css is ready to be brought online now, make it visible */
4368 list_add_tail_rcu(&css->sibling, &parent_css->children);
4369 cgroup_idr_replace(&ss->css_idr, css, css->id);
4371 err = online_css(css);
4375 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4376 cgroup_parent(parent)) {
4377 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4378 current->comm, current->pid, ss->name);
4379 if (!strcmp(ss->name, "memory"))
4380 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4381 ss->warned_broken_hierarchy = true;
4387 list_del_rcu(&css->sibling);
4388 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4390 cgroup_idr_remove(&ss->css_idr, css->id);
4391 err_free_percpu_ref:
4392 percpu_ref_cancel_init(&css->refcnt);
4394 call_rcu(&css->rcu_head, css_free_rcu_fn);
4398 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4401 struct cgroup *parent, *cgrp;
4402 struct cgroup_root *root;
4403 struct cgroup_subsys *ss;
4404 struct kernfs_node *kn;
4407 parent = cgroup_kn_lock_live(parent_kn);
4410 root = parent->root;
4412 /* allocate the cgroup and its ID, 0 is reserved for the root */
4413 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4419 ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
4424 * Temporarily set the pointer to NULL, so idr_find() won't return
4425 * a half-baked cgroup.
4427 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4430 goto out_cancel_ref;
4433 init_cgroup_housekeeping(cgrp);
4435 cgrp->self.parent = &parent->self;
4438 if (notify_on_release(parent))
4439 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4441 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4442 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4444 /* create the directory */
4445 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4453 * This extra ref will be put in cgroup_free_fn() and guarantees
4454 * that @cgrp->kn is always accessible.
4458 cgrp->self.serial_nr = css_serial_nr_next++;
4460 /* allocation complete, commit to creation */
4461 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4462 atomic_inc(&root->nr_cgrps);
4466 * @cgrp is now fully operational. If something fails after this
4467 * point, it'll be released via the normal destruction path.
4469 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4471 ret = cgroup_kn_set_ugid(kn);
4475 ret = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4479 /* let's create and online css's */
4480 for_each_subsys(ss, ssid) {
4481 if (parent->child_subsys_mask & (1 << ssid)) {
4482 ret = create_css(cgrp, ss,
4483 parent->subtree_control & (1 << ssid));
4490 * On the default hierarchy, a child doesn't automatically inherit
4491 * subtree_control from the parent. Each is configured manually.
4493 if (!cgroup_on_dfl(cgrp)) {
4494 cgrp->subtree_control = parent->subtree_control;
4495 cgroup_refresh_child_subsys_mask(cgrp);
4498 kernfs_activate(kn);
4504 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4506 percpu_ref_cancel_init(&cgrp->self.refcnt);
4510 cgroup_kn_unlock(parent_kn);
4514 cgroup_destroy_locked(cgrp);
4519 * This is called when the refcnt of a css is confirmed to be killed.
4520 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4521 * initate destruction and put the css ref from kill_css().
4523 static void css_killed_work_fn(struct work_struct *work)
4525 struct cgroup_subsys_state *css =
4526 container_of(work, struct cgroup_subsys_state, destroy_work);
4528 mutex_lock(&cgroup_mutex);
4530 mutex_unlock(&cgroup_mutex);
4535 /* css kill confirmation processing requires process context, bounce */
4536 static void css_killed_ref_fn(struct percpu_ref *ref)
4538 struct cgroup_subsys_state *css =
4539 container_of(ref, struct cgroup_subsys_state, refcnt);
4541 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4542 queue_work(cgroup_destroy_wq, &css->destroy_work);
4546 * kill_css - destroy a css
4547 * @css: css to destroy
4549 * This function initiates destruction of @css by removing cgroup interface
4550 * files and putting its base reference. ->css_offline() will be invoked
4551 * asynchronously once css_tryget_online() is guaranteed to fail and when
4552 * the reference count reaches zero, @css will be released.
4554 static void kill_css(struct cgroup_subsys_state *css)
4556 lockdep_assert_held(&cgroup_mutex);
4559 * This must happen before css is disassociated with its cgroup.
4560 * See seq_css() for details.
4562 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4565 * Killing would put the base ref, but we need to keep it alive
4566 * until after ->css_offline().
4571 * cgroup core guarantees that, by the time ->css_offline() is
4572 * invoked, no new css reference will be given out via
4573 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4574 * proceed to offlining css's because percpu_ref_kill() doesn't
4575 * guarantee that the ref is seen as killed on all CPUs on return.
4577 * Use percpu_ref_kill_and_confirm() to get notifications as each
4578 * css is confirmed to be seen as killed on all CPUs.
4580 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4584 * cgroup_destroy_locked - the first stage of cgroup destruction
4585 * @cgrp: cgroup to be destroyed
4587 * css's make use of percpu refcnts whose killing latency shouldn't be
4588 * exposed to userland and are RCU protected. Also, cgroup core needs to
4589 * guarantee that css_tryget_online() won't succeed by the time
4590 * ->css_offline() is invoked. To satisfy all the requirements,
4591 * destruction is implemented in the following two steps.
4593 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4594 * userland visible parts and start killing the percpu refcnts of
4595 * css's. Set up so that the next stage will be kicked off once all
4596 * the percpu refcnts are confirmed to be killed.
4598 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4599 * rest of destruction. Once all cgroup references are gone, the
4600 * cgroup is RCU-freed.
4602 * This function implements s1. After this step, @cgrp is gone as far as
4603 * the userland is concerned and a new cgroup with the same name may be
4604 * created. As cgroup doesn't care about the names internally, this
4605 * doesn't cause any problem.
4607 static int cgroup_destroy_locked(struct cgroup *cgrp)
4608 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4610 struct cgroup_subsys_state *css;
4614 lockdep_assert_held(&cgroup_mutex);
4617 * css_set_rwsem synchronizes access to ->cset_links and prevents
4618 * @cgrp from being removed while put_css_set() is in progress.
4620 down_read(&css_set_rwsem);
4621 empty = list_empty(&cgrp->cset_links);
4622 up_read(&css_set_rwsem);
4627 * Make sure there's no live children. We can't test emptiness of
4628 * ->self.children as dead children linger on it while being
4629 * drained; otherwise, "rmdir parent/child parent" may fail.
4631 if (css_has_online_children(&cgrp->self))
4635 * Mark @cgrp dead. This prevents further task migration and child
4636 * creation by disabling cgroup_lock_live_group().
4638 cgrp->self.flags &= ~CSS_ONLINE;
4640 /* initiate massacre of all css's */
4641 for_each_css(css, ssid, cgrp)
4644 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4645 raw_spin_lock(&release_list_lock);
4646 if (!list_empty(&cgrp->release_list))
4647 list_del_init(&cgrp->release_list);
4648 raw_spin_unlock(&release_list_lock);
4651 * Remove @cgrp directory along with the base files. @cgrp has an
4652 * extra ref on its kn.
4654 kernfs_remove(cgrp->kn);
4656 set_bit(CGRP_RELEASABLE, &cgroup_parent(cgrp)->flags);
4657 check_for_release(cgroup_parent(cgrp));
4659 /* put the base reference */
4660 percpu_ref_kill(&cgrp->self.refcnt);
4665 static int cgroup_rmdir(struct kernfs_node *kn)
4667 struct cgroup *cgrp;
4670 cgrp = cgroup_kn_lock_live(kn);
4673 cgroup_get(cgrp); /* for @kn->priv clearing */
4675 ret = cgroup_destroy_locked(cgrp);
4677 cgroup_kn_unlock(kn);
4680 * There are two control paths which try to determine cgroup from
4681 * dentry without going through kernfs - cgroupstats_build() and
4682 * css_tryget_online_from_dir(). Those are supported by RCU
4683 * protecting clearing of cgrp->kn->priv backpointer, which should
4684 * happen after all files under it have been removed.
4687 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4693 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4694 .remount_fs = cgroup_remount,
4695 .show_options = cgroup_show_options,
4696 .mkdir = cgroup_mkdir,
4697 .rmdir = cgroup_rmdir,
4698 .rename = cgroup_rename,
4701 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4703 struct cgroup_subsys_state *css;
4705 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4707 mutex_lock(&cgroup_mutex);
4709 idr_init(&ss->css_idr);
4710 INIT_LIST_HEAD(&ss->cfts);
4712 /* Create the root cgroup state for this subsystem */
4713 ss->root = &cgrp_dfl_root;
4714 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4715 /* We don't handle early failures gracefully */
4716 BUG_ON(IS_ERR(css));
4717 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4720 * Root csses are never destroyed and we can't initialize
4721 * percpu_ref during early init. Disable refcnting.
4723 css->flags |= CSS_NO_REF;
4726 /* allocation can't be done safely during early init */
4729 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4730 BUG_ON(css->id < 0);
4733 /* Update the init_css_set to contain a subsys
4734 * pointer to this state - since the subsystem is
4735 * newly registered, all tasks and hence the
4736 * init_css_set is in the subsystem's root cgroup. */
4737 init_css_set.subsys[ss->id] = css;
4739 need_forkexit_callback |= ss->fork || ss->exit;
4741 /* At system boot, before all subsystems have been
4742 * registered, no tasks have been forked, so we don't
4743 * need to invoke fork callbacks here. */
4744 BUG_ON(!list_empty(&init_task.tasks));
4746 BUG_ON(online_css(css));
4748 mutex_unlock(&cgroup_mutex);
4752 * cgroup_init_early - cgroup initialization at system boot
4754 * Initialize cgroups at system boot, and initialize any
4755 * subsystems that request early init.
4757 int __init cgroup_init_early(void)
4759 static struct cgroup_sb_opts __initdata opts =
4760 { .flags = CGRP_ROOT_SANE_BEHAVIOR };
4761 struct cgroup_subsys *ss;
4764 init_cgroup_root(&cgrp_dfl_root, &opts);
4765 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4767 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4769 for_each_subsys(ss, i) {
4770 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4771 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4772 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4774 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4775 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4778 ss->name = cgroup_subsys_name[i];
4781 cgroup_init_subsys(ss, true);
4787 * cgroup_init - cgroup initialization
4789 * Register cgroup filesystem and /proc file, and initialize
4790 * any subsystems that didn't request early init.
4792 int __init cgroup_init(void)
4794 struct cgroup_subsys *ss;
4798 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
4800 mutex_lock(&cgroup_mutex);
4802 /* Add init_css_set to the hash table */
4803 key = css_set_hash(init_css_set.subsys);
4804 hash_add(css_set_table, &init_css_set.hlist, key);
4806 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4808 mutex_unlock(&cgroup_mutex);
4810 for_each_subsys(ss, ssid) {
4811 if (ss->early_init) {
4812 struct cgroup_subsys_state *css =
4813 init_css_set.subsys[ss->id];
4815 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4817 BUG_ON(css->id < 0);
4819 cgroup_init_subsys(ss, false);
4822 list_add_tail(&init_css_set.e_cset_node[ssid],
4823 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4826 * Setting dfl_root subsys_mask needs to consider the
4827 * disabled flag and cftype registration needs kmalloc,
4828 * both of which aren't available during early_init.
4830 if (!ss->disabled) {
4831 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4832 WARN_ON(cgroup_add_cftypes(ss, ss->base_cftypes));
4836 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4840 err = register_filesystem(&cgroup_fs_type);
4842 kobject_put(cgroup_kobj);
4846 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4850 static int __init cgroup_wq_init(void)
4853 * There isn't much point in executing destruction path in
4854 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4855 * Use 1 for @max_active.
4857 * We would prefer to do this in cgroup_init() above, but that
4858 * is called before init_workqueues(): so leave this until after.
4860 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4861 BUG_ON(!cgroup_destroy_wq);
4864 * Used to destroy pidlists and separate to serve as flush domain.
4865 * Cap @max_active to 1 too.
4867 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4869 BUG_ON(!cgroup_pidlist_destroy_wq);
4873 core_initcall(cgroup_wq_init);
4876 * proc_cgroup_show()
4877 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4878 * - Used for /proc/<pid>/cgroup.
4881 /* TODO: Use a proper seq_file iterator */
4882 int proc_cgroup_show(struct seq_file *m, void *v)
4885 struct task_struct *tsk;
4888 struct cgroup_root *root;
4891 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4897 tsk = get_pid_task(pid, PIDTYPE_PID);
4903 mutex_lock(&cgroup_mutex);
4904 down_read(&css_set_rwsem);
4906 for_each_root(root) {
4907 struct cgroup_subsys *ss;
4908 struct cgroup *cgrp;
4909 int ssid, count = 0;
4911 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
4914 seq_printf(m, "%d:", root->hierarchy_id);
4915 for_each_subsys(ss, ssid)
4916 if (root->subsys_mask & (1 << ssid))
4917 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4918 if (strlen(root->name))
4919 seq_printf(m, "%sname=%s", count ? "," : "",
4922 cgrp = task_cgroup_from_root(tsk, root);
4923 path = cgroup_path(cgrp, buf, PATH_MAX);
4925 retval = -ENAMETOOLONG;
4933 up_read(&css_set_rwsem);
4934 mutex_unlock(&cgroup_mutex);
4935 put_task_struct(tsk);
4942 /* Display information about each subsystem and each hierarchy */
4943 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4945 struct cgroup_subsys *ss;
4948 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4950 * ideally we don't want subsystems moving around while we do this.
4951 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4952 * subsys/hierarchy state.
4954 mutex_lock(&cgroup_mutex);
4956 for_each_subsys(ss, i)
4957 seq_printf(m, "%s\t%d\t%d\t%d\n",
4958 ss->name, ss->root->hierarchy_id,
4959 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
4961 mutex_unlock(&cgroup_mutex);
4965 static int cgroupstats_open(struct inode *inode, struct file *file)
4967 return single_open(file, proc_cgroupstats_show, NULL);
4970 static const struct file_operations proc_cgroupstats_operations = {
4971 .open = cgroupstats_open,
4973 .llseek = seq_lseek,
4974 .release = single_release,
4978 * cgroup_fork - initialize cgroup related fields during copy_process()
4979 * @child: pointer to task_struct of forking parent process.
4981 * A task is associated with the init_css_set until cgroup_post_fork()
4982 * attaches it to the parent's css_set. Empty cg_list indicates that
4983 * @child isn't holding reference to its css_set.
4985 void cgroup_fork(struct task_struct *child)
4987 RCU_INIT_POINTER(child->cgroups, &init_css_set);
4988 INIT_LIST_HEAD(&child->cg_list);
4992 * cgroup_post_fork - called on a new task after adding it to the task list
4993 * @child: the task in question
4995 * Adds the task to the list running through its css_set if necessary and
4996 * call the subsystem fork() callbacks. Has to be after the task is
4997 * visible on the task list in case we race with the first call to
4998 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5001 void cgroup_post_fork(struct task_struct *child)
5003 struct cgroup_subsys *ss;
5007 * This may race against cgroup_enable_task_cg_links(). As that
5008 * function sets use_task_css_set_links before grabbing
5009 * tasklist_lock and we just went through tasklist_lock to add
5010 * @child, it's guaranteed that either we see the set
5011 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5012 * @child during its iteration.
5014 * If we won the race, @child is associated with %current's
5015 * css_set. Grabbing css_set_rwsem guarantees both that the
5016 * association is stable, and, on completion of the parent's
5017 * migration, @child is visible in the source of migration or
5018 * already in the destination cgroup. This guarantee is necessary
5019 * when implementing operations which need to migrate all tasks of
5020 * a cgroup to another.
5022 * Note that if we lose to cgroup_enable_task_cg_links(), @child
5023 * will remain in init_css_set. This is safe because all tasks are
5024 * in the init_css_set before cg_links is enabled and there's no
5025 * operation which transfers all tasks out of init_css_set.
5027 if (use_task_css_set_links) {
5028 struct css_set *cset;
5030 down_write(&css_set_rwsem);
5031 cset = task_css_set(current);
5032 if (list_empty(&child->cg_list)) {
5033 rcu_assign_pointer(child->cgroups, cset);
5034 list_add(&child->cg_list, &cset->tasks);
5037 up_write(&css_set_rwsem);
5041 * Call ss->fork(). This must happen after @child is linked on
5042 * css_set; otherwise, @child might change state between ->fork()
5043 * and addition to css_set.
5045 if (need_forkexit_callback) {
5046 for_each_subsys(ss, i)
5053 * cgroup_exit - detach cgroup from exiting task
5054 * @tsk: pointer to task_struct of exiting process
5056 * Description: Detach cgroup from @tsk and release it.
5058 * Note that cgroups marked notify_on_release force every task in
5059 * them to take the global cgroup_mutex mutex when exiting.
5060 * This could impact scaling on very large systems. Be reluctant to
5061 * use notify_on_release cgroups where very high task exit scaling
5062 * is required on large systems.
5064 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5065 * call cgroup_exit() while the task is still competent to handle
5066 * notify_on_release(), then leave the task attached to the root cgroup in
5067 * each hierarchy for the remainder of its exit. No need to bother with
5068 * init_css_set refcnting. init_css_set never goes away and we can't race
5069 * with migration path - PF_EXITING is visible to migration path.
5071 void cgroup_exit(struct task_struct *tsk)
5073 struct cgroup_subsys *ss;
5074 struct css_set *cset;
5075 bool put_cset = false;
5079 * Unlink from @tsk from its css_set. As migration path can't race
5080 * with us, we can check cg_list without grabbing css_set_rwsem.
5082 if (!list_empty(&tsk->cg_list)) {
5083 down_write(&css_set_rwsem);
5084 list_del_init(&tsk->cg_list);
5085 up_write(&css_set_rwsem);
5089 /* Reassign the task to the init_css_set. */
5090 cset = task_css_set(tsk);
5091 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5093 if (need_forkexit_callback) {
5094 /* see cgroup_post_fork() for details */
5095 for_each_subsys(ss, i) {
5097 struct cgroup_subsys_state *old_css = cset->subsys[i];
5098 struct cgroup_subsys_state *css = task_css(tsk, i);
5100 ss->exit(css, old_css, tsk);
5106 put_css_set(cset, true);
5109 static void check_for_release(struct cgroup *cgrp)
5111 if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
5112 !css_has_online_children(&cgrp->self)) {
5114 * Control Group is currently removeable. If it's not
5115 * already queued for a userspace notification, queue
5118 int need_schedule_work = 0;
5120 raw_spin_lock(&release_list_lock);
5121 if (!cgroup_is_dead(cgrp) &&
5122 list_empty(&cgrp->release_list)) {
5123 list_add(&cgrp->release_list, &release_list);
5124 need_schedule_work = 1;
5126 raw_spin_unlock(&release_list_lock);
5127 if (need_schedule_work)
5128 schedule_work(&release_agent_work);
5133 * Notify userspace when a cgroup is released, by running the
5134 * configured release agent with the name of the cgroup (path
5135 * relative to the root of cgroup file system) as the argument.
5137 * Most likely, this user command will try to rmdir this cgroup.
5139 * This races with the possibility that some other task will be
5140 * attached to this cgroup before it is removed, or that some other
5141 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5142 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5143 * unused, and this cgroup will be reprieved from its death sentence,
5144 * to continue to serve a useful existence. Next time it's released,
5145 * we will get notified again, if it still has 'notify_on_release' set.
5147 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5148 * means only wait until the task is successfully execve()'d. The
5149 * separate release agent task is forked by call_usermodehelper(),
5150 * then control in this thread returns here, without waiting for the
5151 * release agent task. We don't bother to wait because the caller of
5152 * this routine has no use for the exit status of the release agent
5153 * task, so no sense holding our caller up for that.
5155 static void cgroup_release_agent(struct work_struct *work)
5157 BUG_ON(work != &release_agent_work);
5158 mutex_lock(&cgroup_mutex);
5159 raw_spin_lock(&release_list_lock);
5160 while (!list_empty(&release_list)) {
5161 char *argv[3], *envp[3];
5163 char *pathbuf = NULL, *agentbuf = NULL, *path;
5164 struct cgroup *cgrp = list_entry(release_list.next,
5167 list_del_init(&cgrp->release_list);
5168 raw_spin_unlock(&release_list_lock);
5169 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5172 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5175 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5180 argv[i++] = agentbuf;
5185 /* minimal command environment */
5186 envp[i++] = "HOME=/";
5187 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5190 /* Drop the lock while we invoke the usermode helper,
5191 * since the exec could involve hitting disk and hence
5192 * be a slow process */
5193 mutex_unlock(&cgroup_mutex);
5194 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5195 mutex_lock(&cgroup_mutex);
5199 raw_spin_lock(&release_list_lock);
5201 raw_spin_unlock(&release_list_lock);
5202 mutex_unlock(&cgroup_mutex);
5205 static int __init cgroup_disable(char *str)
5207 struct cgroup_subsys *ss;
5211 while ((token = strsep(&str, ",")) != NULL) {
5215 for_each_subsys(ss, i) {
5216 if (!strcmp(token, ss->name)) {
5218 printk(KERN_INFO "Disabling %s control group"
5219 " subsystem\n", ss->name);
5226 __setup("cgroup_disable=", cgroup_disable);
5229 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5230 * @dentry: directory dentry of interest
5231 * @ss: subsystem of interest
5233 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5234 * to get the corresponding css and return it. If such css doesn't exist
5235 * or can't be pinned, an ERR_PTR value is returned.
5237 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5238 struct cgroup_subsys *ss)
5240 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5241 struct cgroup_subsys_state *css = NULL;
5242 struct cgroup *cgrp;
5244 /* is @dentry a cgroup dir? */
5245 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5246 kernfs_type(kn) != KERNFS_DIR)
5247 return ERR_PTR(-EBADF);
5252 * This path doesn't originate from kernfs and @kn could already
5253 * have been or be removed at any point. @kn->priv is RCU
5254 * protected for this access. See cgroup_rmdir() for details.
5256 cgrp = rcu_dereference(kn->priv);
5258 css = cgroup_css(cgrp, ss);
5260 if (!css || !css_tryget_online(css))
5261 css = ERR_PTR(-ENOENT);
5268 * css_from_id - lookup css by id
5269 * @id: the cgroup id
5270 * @ss: cgroup subsys to be looked into
5272 * Returns the css if there's valid one with @id, otherwise returns NULL.
5273 * Should be called under rcu_read_lock().
5275 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5277 WARN_ON_ONCE(!rcu_read_lock_held());
5278 return idr_find(&ss->css_idr, id);
5281 #ifdef CONFIG_CGROUP_DEBUG
5282 static struct cgroup_subsys_state *
5283 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5285 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5288 return ERR_PTR(-ENOMEM);
5293 static void debug_css_free(struct cgroup_subsys_state *css)
5298 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5301 return cgroup_task_count(css->cgroup);
5304 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5307 return (u64)(unsigned long)current->cgroups;
5310 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5316 count = atomic_read(&task_css_set(current)->refcount);
5321 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5323 struct cgrp_cset_link *link;
5324 struct css_set *cset;
5327 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5331 down_read(&css_set_rwsem);
5333 cset = rcu_dereference(current->cgroups);
5334 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5335 struct cgroup *c = link->cgrp;
5337 cgroup_name(c, name_buf, NAME_MAX + 1);
5338 seq_printf(seq, "Root %d group %s\n",
5339 c->root->hierarchy_id, name_buf);
5342 up_read(&css_set_rwsem);
5347 #define MAX_TASKS_SHOWN_PER_CSS 25
5348 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5350 struct cgroup_subsys_state *css = seq_css(seq);
5351 struct cgrp_cset_link *link;
5353 down_read(&css_set_rwsem);
5354 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5355 struct css_set *cset = link->cset;
5356 struct task_struct *task;
5359 seq_printf(seq, "css_set %p\n", cset);
5361 list_for_each_entry(task, &cset->tasks, cg_list) {
5362 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5364 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5367 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5368 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5370 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5374 seq_puts(seq, " ...\n");
5376 up_read(&css_set_rwsem);
5380 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5382 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5385 static struct cftype debug_files[] = {
5387 .name = "taskcount",
5388 .read_u64 = debug_taskcount_read,
5392 .name = "current_css_set",
5393 .read_u64 = current_css_set_read,
5397 .name = "current_css_set_refcount",
5398 .read_u64 = current_css_set_refcount_read,
5402 .name = "current_css_set_cg_links",
5403 .seq_show = current_css_set_cg_links_read,
5407 .name = "cgroup_css_links",
5408 .seq_show = cgroup_css_links_read,
5412 .name = "releasable",
5413 .read_u64 = releasable_read,
5419 struct cgroup_subsys debug_cgrp_subsys = {
5420 .css_alloc = debug_css_alloc,
5421 .css_free = debug_css_free,
5422 .base_cftypes = debug_files,
5424 #endif /* CONFIG_CGROUP_DEBUG */