2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 DECLARE_RWSEM(css_set_rwsem);
87 EXPORT_SYMBOL_GPL(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(css_set_rwsem);
90 static DEFINE_MUTEX(cgroup_mutex);
91 static DECLARE_RWSEM(css_set_rwsem);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct *cgroup_destroy_wq;
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
123 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys *cgroup_subsys[] = {
128 #include <linux/cgroup_subsys.h>
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name[] = {
135 #include <linux/cgroup_subsys.h>
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
144 struct cgroup_root cgrp_dfl_root;
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
150 static bool cgrp_dfl_root_visible;
153 * Set by the boot param of the same name and makes subsystems with NULL
154 * ->dfl_files to use ->legacy_files on the default hierarchy.
156 static bool cgroup_legacy_files_on_dfl;
158 /* some controllers are not supported in the default hierarchy */
159 static unsigned int cgrp_dfl_root_inhibit_ss_mask;
161 /* The list of hierarchy roots */
163 static LIST_HEAD(cgroup_roots);
164 static int cgroup_root_count;
166 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
167 static DEFINE_IDR(cgroup_hierarchy_idr);
170 * Assign a monotonically increasing serial number to csses. It guarantees
171 * cgroups with bigger numbers are newer than those with smaller numbers.
172 * Also, as csses are always appended to the parent's ->children list, it
173 * guarantees that sibling csses are always sorted in the ascending serial
174 * number order on the list. Protected by cgroup_mutex.
176 static u64 css_serial_nr_next = 1;
178 /* This flag indicates whether tasks in the fork and exit paths should
179 * check for fork/exit handlers to call. This avoids us having to do
180 * extra work in the fork/exit path if none of the subsystems need to
183 static int need_forkexit_callback __read_mostly;
185 static struct cftype cgroup_dfl_base_files[];
186 static struct cftype cgroup_legacy_base_files[];
188 static void cgroup_put(struct cgroup *cgrp);
189 static int rebind_subsystems(struct cgroup_root *dst_root,
190 unsigned int ss_mask);
191 static int cgroup_destroy_locked(struct cgroup *cgrp);
192 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
194 static void css_release(struct percpu_ref *ref);
195 static void kill_css(struct cgroup_subsys_state *css);
196 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
198 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
200 /* IDR wrappers which synchronize using cgroup_idr_lock */
201 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
206 idr_preload(gfp_mask);
207 spin_lock_bh(&cgroup_idr_lock);
208 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
209 spin_unlock_bh(&cgroup_idr_lock);
214 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
218 spin_lock_bh(&cgroup_idr_lock);
219 ret = idr_replace(idr, ptr, id);
220 spin_unlock_bh(&cgroup_idr_lock);
224 static void cgroup_idr_remove(struct idr *idr, int id)
226 spin_lock_bh(&cgroup_idr_lock);
228 spin_unlock_bh(&cgroup_idr_lock);
231 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
233 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
236 return container_of(parent_css, struct cgroup, self);
241 * cgroup_css - obtain a cgroup's css for the specified subsystem
242 * @cgrp: the cgroup of interest
243 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
245 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
246 * function must be called either under cgroup_mutex or rcu_read_lock() and
247 * the caller is responsible for pinning the returned css if it wants to
248 * keep accessing it outside the said locks. This function may return
249 * %NULL if @cgrp doesn't have @subsys_id enabled.
251 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
252 struct cgroup_subsys *ss)
255 return rcu_dereference_check(cgrp->subsys[ss->id],
256 lockdep_is_held(&cgroup_mutex));
262 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
263 * @cgrp: the cgroup of interest
264 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
266 * Similar to cgroup_css() but returns the effctive css, which is defined
267 * as the matching css of the nearest ancestor including self which has @ss
268 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
269 * function is guaranteed to return non-NULL css.
271 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
272 struct cgroup_subsys *ss)
274 lockdep_assert_held(&cgroup_mutex);
279 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
282 while (cgroup_parent(cgrp) &&
283 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
284 cgrp = cgroup_parent(cgrp);
286 return cgroup_css(cgrp, ss);
289 /* convenient tests for these bits */
290 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
292 return !(cgrp->self.flags & CSS_ONLINE);
295 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
297 struct cgroup *cgrp = of->kn->parent->priv;
298 struct cftype *cft = of_cft(of);
301 * This is open and unprotected implementation of cgroup_css().
302 * seq_css() is only called from a kernfs file operation which has
303 * an active reference on the file. Because all the subsystem
304 * files are drained before a css is disassociated with a cgroup,
305 * the matching css from the cgroup's subsys table is guaranteed to
306 * be and stay valid until the enclosing operation is complete.
309 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
313 EXPORT_SYMBOL_GPL(of_css);
316 * cgroup_is_descendant - test ancestry
317 * @cgrp: the cgroup to be tested
318 * @ancestor: possible ancestor of @cgrp
320 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
321 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
322 * and @ancestor are accessible.
324 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
327 if (cgrp == ancestor)
329 cgrp = cgroup_parent(cgrp);
334 static int cgroup_is_releasable(const struct cgroup *cgrp)
337 (1 << CGRP_RELEASABLE) |
338 (1 << CGRP_NOTIFY_ON_RELEASE);
339 return (cgrp->flags & bits) == bits;
342 static int notify_on_release(const struct cgroup *cgrp)
344 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
348 * for_each_css - iterate all css's of a cgroup
349 * @css: the iteration cursor
350 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
351 * @cgrp: the target cgroup to iterate css's of
353 * Should be called under cgroup_[tree_]mutex.
355 #define for_each_css(css, ssid, cgrp) \
356 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
357 if (!((css) = rcu_dereference_check( \
358 (cgrp)->subsys[(ssid)], \
359 lockdep_is_held(&cgroup_mutex)))) { } \
363 * for_each_e_css - iterate all effective css's of a cgroup
364 * @css: the iteration cursor
365 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
366 * @cgrp: the target cgroup to iterate css's of
368 * Should be called under cgroup_[tree_]mutex.
370 #define for_each_e_css(css, ssid, cgrp) \
371 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
372 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
377 * for_each_subsys - iterate all enabled cgroup subsystems
378 * @ss: the iteration cursor
379 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
381 #define for_each_subsys(ss, ssid) \
382 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
383 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
385 /* iterate across the hierarchies */
386 #define for_each_root(root) \
387 list_for_each_entry((root), &cgroup_roots, root_list)
389 /* iterate over child cgrps, lock should be held throughout iteration */
390 #define cgroup_for_each_live_child(child, cgrp) \
391 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
392 if (({ lockdep_assert_held(&cgroup_mutex); \
393 cgroup_is_dead(child); })) \
397 /* the list of cgroups eligible for automatic release. Protected by
398 * release_list_lock */
399 static LIST_HEAD(release_list);
400 static DEFINE_RAW_SPINLOCK(release_list_lock);
401 static void cgroup_release_agent(struct work_struct *work);
402 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
403 static void check_for_release(struct cgroup *cgrp);
406 * A cgroup can be associated with multiple css_sets as different tasks may
407 * belong to different cgroups on different hierarchies. In the other
408 * direction, a css_set is naturally associated with multiple cgroups.
409 * This M:N relationship is represented by the following link structure
410 * which exists for each association and allows traversing the associations
413 struct cgrp_cset_link {
414 /* the cgroup and css_set this link associates */
416 struct css_set *cset;
418 /* list of cgrp_cset_links anchored at cgrp->cset_links */
419 struct list_head cset_link;
421 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
422 struct list_head cgrp_link;
426 * The default css_set - used by init and its children prior to any
427 * hierarchies being mounted. It contains a pointer to the root state
428 * for each subsystem. Also used to anchor the list of css_sets. Not
429 * reference-counted, to improve performance when child cgroups
430 * haven't been created.
432 struct css_set init_css_set = {
433 .refcount = ATOMIC_INIT(1),
434 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
435 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
436 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
437 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
438 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
441 static int css_set_count = 1; /* 1 for init_css_set */
444 * cgroup_update_populated - updated populated count of a cgroup
445 * @cgrp: the target cgroup
446 * @populated: inc or dec populated count
448 * @cgrp is either getting the first task (css_set) or losing the last.
449 * Update @cgrp->populated_cnt accordingly. The count is propagated
450 * towards root so that a given cgroup's populated_cnt is zero iff the
451 * cgroup and all its descendants are empty.
453 * @cgrp's interface file "cgroup.populated" is zero if
454 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
455 * changes from or to zero, userland is notified that the content of the
456 * interface file has changed. This can be used to detect when @cgrp and
457 * its descendants become populated or empty.
459 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
461 lockdep_assert_held(&css_set_rwsem);
467 trigger = !cgrp->populated_cnt++;
469 trigger = !--cgrp->populated_cnt;
474 if (cgrp->populated_kn)
475 kernfs_notify(cgrp->populated_kn);
476 cgrp = cgroup_parent(cgrp);
481 * hash table for cgroup groups. This improves the performance to find
482 * an existing css_set. This hash doesn't (currently) take into
483 * account cgroups in empty hierarchies.
485 #define CSS_SET_HASH_BITS 7
486 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
488 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
490 unsigned long key = 0UL;
491 struct cgroup_subsys *ss;
494 for_each_subsys(ss, i)
495 key += (unsigned long)css[i];
496 key = (key >> 16) ^ key;
501 static void put_css_set_locked(struct css_set *cset, bool taskexit)
503 struct cgrp_cset_link *link, *tmp_link;
504 struct cgroup_subsys *ss;
507 lockdep_assert_held(&css_set_rwsem);
509 if (!atomic_dec_and_test(&cset->refcount))
512 /* This css_set is dead. unlink it and release cgroup refcounts */
513 for_each_subsys(ss, ssid)
514 list_del(&cset->e_cset_node[ssid]);
515 hash_del(&cset->hlist);
518 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
519 struct cgroup *cgrp = link->cgrp;
521 list_del(&link->cset_link);
522 list_del(&link->cgrp_link);
524 /* @cgrp can't go away while we're holding css_set_rwsem */
525 if (list_empty(&cgrp->cset_links)) {
526 cgroup_update_populated(cgrp, false);
527 if (notify_on_release(cgrp)) {
529 set_bit(CGRP_RELEASABLE, &cgrp->flags);
530 check_for_release(cgrp);
537 kfree_rcu(cset, rcu_head);
540 static void put_css_set(struct css_set *cset, bool taskexit)
543 * Ensure that the refcount doesn't hit zero while any readers
544 * can see it. Similar to atomic_dec_and_lock(), but for an
547 if (atomic_add_unless(&cset->refcount, -1, 1))
550 down_write(&css_set_rwsem);
551 put_css_set_locked(cset, taskexit);
552 up_write(&css_set_rwsem);
556 * refcounted get/put for css_set objects
558 static inline void get_css_set(struct css_set *cset)
560 atomic_inc(&cset->refcount);
564 * compare_css_sets - helper function for find_existing_css_set().
565 * @cset: candidate css_set being tested
566 * @old_cset: existing css_set for a task
567 * @new_cgrp: cgroup that's being entered by the task
568 * @template: desired set of css pointers in css_set (pre-calculated)
570 * Returns true if "cset" matches "old_cset" except for the hierarchy
571 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
573 static bool compare_css_sets(struct css_set *cset,
574 struct css_set *old_cset,
575 struct cgroup *new_cgrp,
576 struct cgroup_subsys_state *template[])
578 struct list_head *l1, *l2;
581 * On the default hierarchy, there can be csets which are
582 * associated with the same set of cgroups but different csses.
583 * Let's first ensure that csses match.
585 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
589 * Compare cgroup pointers in order to distinguish between
590 * different cgroups in hierarchies. As different cgroups may
591 * share the same effective css, this comparison is always
594 l1 = &cset->cgrp_links;
595 l2 = &old_cset->cgrp_links;
597 struct cgrp_cset_link *link1, *link2;
598 struct cgroup *cgrp1, *cgrp2;
602 /* See if we reached the end - both lists are equal length. */
603 if (l1 == &cset->cgrp_links) {
604 BUG_ON(l2 != &old_cset->cgrp_links);
607 BUG_ON(l2 == &old_cset->cgrp_links);
609 /* Locate the cgroups associated with these links. */
610 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
611 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
614 /* Hierarchies should be linked in the same order. */
615 BUG_ON(cgrp1->root != cgrp2->root);
618 * If this hierarchy is the hierarchy of the cgroup
619 * that's changing, then we need to check that this
620 * css_set points to the new cgroup; if it's any other
621 * hierarchy, then this css_set should point to the
622 * same cgroup as the old css_set.
624 if (cgrp1->root == new_cgrp->root) {
625 if (cgrp1 != new_cgrp)
636 * find_existing_css_set - init css array and find the matching css_set
637 * @old_cset: the css_set that we're using before the cgroup transition
638 * @cgrp: the cgroup that we're moving into
639 * @template: out param for the new set of csses, should be clear on entry
641 static struct css_set *find_existing_css_set(struct css_set *old_cset,
643 struct cgroup_subsys_state *template[])
645 struct cgroup_root *root = cgrp->root;
646 struct cgroup_subsys *ss;
647 struct css_set *cset;
652 * Build the set of subsystem state objects that we want to see in the
653 * new css_set. while subsystems can change globally, the entries here
654 * won't change, so no need for locking.
656 for_each_subsys(ss, i) {
657 if (root->subsys_mask & (1UL << i)) {
659 * @ss is in this hierarchy, so we want the
660 * effective css from @cgrp.
662 template[i] = cgroup_e_css(cgrp, ss);
665 * @ss is not in this hierarchy, so we don't want
668 template[i] = old_cset->subsys[i];
672 key = css_set_hash(template);
673 hash_for_each_possible(css_set_table, cset, hlist, key) {
674 if (!compare_css_sets(cset, old_cset, cgrp, template))
677 /* This css_set matches what we need */
681 /* No existing cgroup group matched */
685 static void free_cgrp_cset_links(struct list_head *links_to_free)
687 struct cgrp_cset_link *link, *tmp_link;
689 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
690 list_del(&link->cset_link);
696 * allocate_cgrp_cset_links - allocate cgrp_cset_links
697 * @count: the number of links to allocate
698 * @tmp_links: list_head the allocated links are put on
700 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
701 * through ->cset_link. Returns 0 on success or -errno.
703 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
705 struct cgrp_cset_link *link;
708 INIT_LIST_HEAD(tmp_links);
710 for (i = 0; i < count; i++) {
711 link = kzalloc(sizeof(*link), GFP_KERNEL);
713 free_cgrp_cset_links(tmp_links);
716 list_add(&link->cset_link, tmp_links);
722 * link_css_set - a helper function to link a css_set to a cgroup
723 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
724 * @cset: the css_set to be linked
725 * @cgrp: the destination cgroup
727 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
730 struct cgrp_cset_link *link;
732 BUG_ON(list_empty(tmp_links));
734 if (cgroup_on_dfl(cgrp))
735 cset->dfl_cgrp = cgrp;
737 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
741 if (list_empty(&cgrp->cset_links))
742 cgroup_update_populated(cgrp, true);
743 list_move(&link->cset_link, &cgrp->cset_links);
746 * Always add links to the tail of the list so that the list
747 * is sorted by order of hierarchy creation
749 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
753 * find_css_set - return a new css_set with one cgroup updated
754 * @old_cset: the baseline css_set
755 * @cgrp: the cgroup to be updated
757 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
758 * substituted into the appropriate hierarchy.
760 static struct css_set *find_css_set(struct css_set *old_cset,
763 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
764 struct css_set *cset;
765 struct list_head tmp_links;
766 struct cgrp_cset_link *link;
767 struct cgroup_subsys *ss;
771 lockdep_assert_held(&cgroup_mutex);
773 /* First see if we already have a cgroup group that matches
775 down_read(&css_set_rwsem);
776 cset = find_existing_css_set(old_cset, cgrp, template);
779 up_read(&css_set_rwsem);
784 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
788 /* Allocate all the cgrp_cset_link objects that we'll need */
789 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
794 atomic_set(&cset->refcount, 1);
795 INIT_LIST_HEAD(&cset->cgrp_links);
796 INIT_LIST_HEAD(&cset->tasks);
797 INIT_LIST_HEAD(&cset->mg_tasks);
798 INIT_LIST_HEAD(&cset->mg_preload_node);
799 INIT_LIST_HEAD(&cset->mg_node);
800 INIT_HLIST_NODE(&cset->hlist);
802 /* Copy the set of subsystem state objects generated in
803 * find_existing_css_set() */
804 memcpy(cset->subsys, template, sizeof(cset->subsys));
806 down_write(&css_set_rwsem);
807 /* Add reference counts and links from the new css_set. */
808 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
809 struct cgroup *c = link->cgrp;
811 if (c->root == cgrp->root)
813 link_css_set(&tmp_links, cset, c);
816 BUG_ON(!list_empty(&tmp_links));
820 /* Add @cset to the hash table */
821 key = css_set_hash(cset->subsys);
822 hash_add(css_set_table, &cset->hlist, key);
824 for_each_subsys(ss, ssid)
825 list_add_tail(&cset->e_cset_node[ssid],
826 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
828 up_write(&css_set_rwsem);
833 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
835 struct cgroup *root_cgrp = kf_root->kn->priv;
837 return root_cgrp->root;
840 static int cgroup_init_root_id(struct cgroup_root *root)
844 lockdep_assert_held(&cgroup_mutex);
846 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
850 root->hierarchy_id = id;
854 static void cgroup_exit_root_id(struct cgroup_root *root)
856 lockdep_assert_held(&cgroup_mutex);
858 if (root->hierarchy_id) {
859 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
860 root->hierarchy_id = 0;
864 static void cgroup_free_root(struct cgroup_root *root)
867 /* hierarhcy ID shoulid already have been released */
868 WARN_ON_ONCE(root->hierarchy_id);
870 idr_destroy(&root->cgroup_idr);
875 static void cgroup_destroy_root(struct cgroup_root *root)
877 struct cgroup *cgrp = &root->cgrp;
878 struct cgrp_cset_link *link, *tmp_link;
880 mutex_lock(&cgroup_mutex);
882 BUG_ON(atomic_read(&root->nr_cgrps));
883 BUG_ON(!list_empty(&cgrp->self.children));
885 /* Rebind all subsystems back to the default hierarchy */
886 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
889 * Release all the links from cset_links to this hierarchy's
892 down_write(&css_set_rwsem);
894 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
895 list_del(&link->cset_link);
896 list_del(&link->cgrp_link);
899 up_write(&css_set_rwsem);
901 if (!list_empty(&root->root_list)) {
902 list_del(&root->root_list);
906 cgroup_exit_root_id(root);
908 mutex_unlock(&cgroup_mutex);
910 kernfs_destroy_root(root->kf_root);
911 cgroup_free_root(root);
914 /* look up cgroup associated with given css_set on the specified hierarchy */
915 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
916 struct cgroup_root *root)
918 struct cgroup *res = NULL;
920 lockdep_assert_held(&cgroup_mutex);
921 lockdep_assert_held(&css_set_rwsem);
923 if (cset == &init_css_set) {
926 struct cgrp_cset_link *link;
928 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
929 struct cgroup *c = link->cgrp;
931 if (c->root == root) {
943 * Return the cgroup for "task" from the given hierarchy. Must be
944 * called with cgroup_mutex and css_set_rwsem held.
946 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
947 struct cgroup_root *root)
950 * No need to lock the task - since we hold cgroup_mutex the
951 * task can't change groups, so the only thing that can happen
952 * is that it exits and its css is set back to init_css_set.
954 return cset_cgroup_from_root(task_css_set(task), root);
958 * A task must hold cgroup_mutex to modify cgroups.
960 * Any task can increment and decrement the count field without lock.
961 * So in general, code holding cgroup_mutex can't rely on the count
962 * field not changing. However, if the count goes to zero, then only
963 * cgroup_attach_task() can increment it again. Because a count of zero
964 * means that no tasks are currently attached, therefore there is no
965 * way a task attached to that cgroup can fork (the other way to
966 * increment the count). So code holding cgroup_mutex can safely
967 * assume that if the count is zero, it will stay zero. Similarly, if
968 * a task holds cgroup_mutex on a cgroup with zero count, it
969 * knows that the cgroup won't be removed, as cgroup_rmdir()
972 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
973 * (usually) take cgroup_mutex. These are the two most performance
974 * critical pieces of code here. The exception occurs on cgroup_exit(),
975 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
976 * is taken, and if the cgroup count is zero, a usermode call made
977 * to the release agent with the name of the cgroup (path relative to
978 * the root of cgroup file system) as the argument.
980 * A cgroup can only be deleted if both its 'count' of using tasks
981 * is zero, and its list of 'children' cgroups is empty. Since all
982 * tasks in the system use _some_ cgroup, and since there is always at
983 * least one task in the system (init, pid == 1), therefore, root cgroup
984 * always has either children cgroups and/or using tasks. So we don't
985 * need a special hack to ensure that root cgroup cannot be deleted.
987 * P.S. One more locking exception. RCU is used to guard the
988 * update of a tasks cgroup pointer by cgroup_attach_task()
991 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
992 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
993 static const struct file_operations proc_cgroupstats_operations;
995 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
998 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
999 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1000 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1001 cft->ss->name, cft->name);
1003 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1008 * cgroup_file_mode - deduce file mode of a control file
1009 * @cft: the control file in question
1011 * returns cft->mode if ->mode is not 0
1012 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1013 * returns S_IRUGO if it has only a read handler
1014 * returns S_IWUSR if it has only a write hander
1016 static umode_t cgroup_file_mode(const struct cftype *cft)
1023 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1026 if (cft->write_u64 || cft->write_s64 || cft->write)
1032 static void cgroup_get(struct cgroup *cgrp)
1034 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1035 css_get(&cgrp->self);
1038 static void cgroup_put(struct cgroup *cgrp)
1040 css_put(&cgrp->self);
1044 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1045 * @cgrp: the target cgroup
1047 * On the default hierarchy, a subsystem may request other subsystems to be
1048 * enabled together through its ->depends_on mask. In such cases, more
1049 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1051 * This function determines which subsystems need to be enabled given the
1052 * current @cgrp->subtree_control and records it in
1053 * @cgrp->child_subsys_mask. The resulting mask is always a superset of
1054 * @cgrp->subtree_control and follows the usual hierarchy rules.
1056 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1058 struct cgroup *parent = cgroup_parent(cgrp);
1059 unsigned int cur_ss_mask = cgrp->subtree_control;
1060 struct cgroup_subsys *ss;
1063 lockdep_assert_held(&cgroup_mutex);
1065 if (!cgroup_on_dfl(cgrp)) {
1066 cgrp->child_subsys_mask = cur_ss_mask;
1071 unsigned int new_ss_mask = cur_ss_mask;
1073 for_each_subsys(ss, ssid)
1074 if (cur_ss_mask & (1 << ssid))
1075 new_ss_mask |= ss->depends_on;
1078 * Mask out subsystems which aren't available. This can
1079 * happen only if some depended-upon subsystems were bound
1080 * to non-default hierarchies.
1083 new_ss_mask &= parent->child_subsys_mask;
1085 new_ss_mask &= cgrp->root->subsys_mask;
1087 if (new_ss_mask == cur_ss_mask)
1089 cur_ss_mask = new_ss_mask;
1092 cgrp->child_subsys_mask = cur_ss_mask;
1096 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1097 * @kn: the kernfs_node being serviced
1099 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1100 * the method finishes if locking succeeded. Note that once this function
1101 * returns the cgroup returned by cgroup_kn_lock_live() may become
1102 * inaccessible any time. If the caller intends to continue to access the
1103 * cgroup, it should pin it before invoking this function.
1105 static void cgroup_kn_unlock(struct kernfs_node *kn)
1107 struct cgroup *cgrp;
1109 if (kernfs_type(kn) == KERNFS_DIR)
1112 cgrp = kn->parent->priv;
1114 mutex_unlock(&cgroup_mutex);
1116 kernfs_unbreak_active_protection(kn);
1121 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1122 * @kn: the kernfs_node being serviced
1124 * This helper is to be used by a cgroup kernfs method currently servicing
1125 * @kn. It breaks the active protection, performs cgroup locking and
1126 * verifies that the associated cgroup is alive. Returns the cgroup if
1127 * alive; otherwise, %NULL. A successful return should be undone by a
1128 * matching cgroup_kn_unlock() invocation.
1130 * Any cgroup kernfs method implementation which requires locking the
1131 * associated cgroup should use this helper. It avoids nesting cgroup
1132 * locking under kernfs active protection and allows all kernfs operations
1133 * including self-removal.
1135 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1137 struct cgroup *cgrp;
1139 if (kernfs_type(kn) == KERNFS_DIR)
1142 cgrp = kn->parent->priv;
1145 * We're gonna grab cgroup_mutex which nests outside kernfs
1146 * active_ref. cgroup liveliness check alone provides enough
1147 * protection against removal. Ensure @cgrp stays accessible and
1148 * break the active_ref protection.
1151 kernfs_break_active_protection(kn);
1153 mutex_lock(&cgroup_mutex);
1155 if (!cgroup_is_dead(cgrp))
1158 cgroup_kn_unlock(kn);
1162 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1164 char name[CGROUP_FILE_NAME_MAX];
1166 lockdep_assert_held(&cgroup_mutex);
1167 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1171 * cgroup_clear_dir - remove subsys files in a cgroup directory
1172 * @cgrp: target cgroup
1173 * @subsys_mask: mask of the subsystem ids whose files should be removed
1175 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1177 struct cgroup_subsys *ss;
1180 for_each_subsys(ss, i) {
1181 struct cftype *cfts;
1183 if (!(subsys_mask & (1 << i)))
1185 list_for_each_entry(cfts, &ss->cfts, node)
1186 cgroup_addrm_files(cgrp, cfts, false);
1190 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1192 struct cgroup_subsys *ss;
1193 unsigned int tmp_ss_mask;
1196 lockdep_assert_held(&cgroup_mutex);
1198 for_each_subsys(ss, ssid) {
1199 if (!(ss_mask & (1 << ssid)))
1202 /* if @ss has non-root csses attached to it, can't move */
1203 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1206 /* can't move between two non-dummy roots either */
1207 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1211 /* skip creating root files on dfl_root for inhibited subsystems */
1212 tmp_ss_mask = ss_mask;
1213 if (dst_root == &cgrp_dfl_root)
1214 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1216 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1218 if (dst_root != &cgrp_dfl_root)
1222 * Rebinding back to the default root is not allowed to
1223 * fail. Using both default and non-default roots should
1224 * be rare. Moving subsystems back and forth even more so.
1225 * Just warn about it and continue.
1227 if (cgrp_dfl_root_visible) {
1228 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1230 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1235 * Nothing can fail from this point on. Remove files for the
1236 * removed subsystems and rebind each subsystem.
1238 for_each_subsys(ss, ssid)
1239 if (ss_mask & (1 << ssid))
1240 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1242 for_each_subsys(ss, ssid) {
1243 struct cgroup_root *src_root;
1244 struct cgroup_subsys_state *css;
1245 struct css_set *cset;
1247 if (!(ss_mask & (1 << ssid)))
1250 src_root = ss->root;
1251 css = cgroup_css(&src_root->cgrp, ss);
1253 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1255 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1256 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1257 ss->root = dst_root;
1258 css->cgroup = &dst_root->cgrp;
1260 down_write(&css_set_rwsem);
1261 hash_for_each(css_set_table, i, cset, hlist)
1262 list_move_tail(&cset->e_cset_node[ss->id],
1263 &dst_root->cgrp.e_csets[ss->id]);
1264 up_write(&css_set_rwsem);
1266 src_root->subsys_mask &= ~(1 << ssid);
1267 src_root->cgrp.subtree_control &= ~(1 << ssid);
1268 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1270 /* default hierarchy doesn't enable controllers by default */
1271 dst_root->subsys_mask |= 1 << ssid;
1272 if (dst_root != &cgrp_dfl_root) {
1273 dst_root->cgrp.subtree_control |= 1 << ssid;
1274 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1281 kernfs_activate(dst_root->cgrp.kn);
1285 static int cgroup_show_options(struct seq_file *seq,
1286 struct kernfs_root *kf_root)
1288 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1289 struct cgroup_subsys *ss;
1292 for_each_subsys(ss, ssid)
1293 if (root->subsys_mask & (1 << ssid))
1294 seq_printf(seq, ",%s", ss->name);
1295 if (root->flags & CGRP_ROOT_NOPREFIX)
1296 seq_puts(seq, ",noprefix");
1297 if (root->flags & CGRP_ROOT_XATTR)
1298 seq_puts(seq, ",xattr");
1300 spin_lock(&release_agent_path_lock);
1301 if (strlen(root->release_agent_path))
1302 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1303 spin_unlock(&release_agent_path_lock);
1305 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1306 seq_puts(seq, ",clone_children");
1307 if (strlen(root->name))
1308 seq_printf(seq, ",name=%s", root->name);
1312 struct cgroup_sb_opts {
1313 unsigned int subsys_mask;
1315 char *release_agent;
1316 bool cpuset_clone_children;
1318 /* User explicitly requested empty subsystem */
1322 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1324 char *token, *o = data;
1325 bool all_ss = false, one_ss = false;
1326 unsigned int mask = -1U;
1327 struct cgroup_subsys *ss;
1331 #ifdef CONFIG_CPUSETS
1332 mask = ~(1U << cpuset_cgrp_id);
1335 memset(opts, 0, sizeof(*opts));
1337 while ((token = strsep(&o, ",")) != NULL) {
1342 if (!strcmp(token, "none")) {
1343 /* Explicitly have no subsystems */
1347 if (!strcmp(token, "all")) {
1348 /* Mutually exclusive option 'all' + subsystem name */
1354 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1355 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1358 if (!strcmp(token, "noprefix")) {
1359 opts->flags |= CGRP_ROOT_NOPREFIX;
1362 if (!strcmp(token, "clone_children")) {
1363 opts->cpuset_clone_children = true;
1366 if (!strcmp(token, "xattr")) {
1367 opts->flags |= CGRP_ROOT_XATTR;
1370 if (!strncmp(token, "release_agent=", 14)) {
1371 /* Specifying two release agents is forbidden */
1372 if (opts->release_agent)
1374 opts->release_agent =
1375 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1376 if (!opts->release_agent)
1380 if (!strncmp(token, "name=", 5)) {
1381 const char *name = token + 5;
1382 /* Can't specify an empty name */
1385 /* Must match [\w.-]+ */
1386 for (i = 0; i < strlen(name); i++) {
1390 if ((c == '.') || (c == '-') || (c == '_'))
1394 /* Specifying two names is forbidden */
1397 opts->name = kstrndup(name,
1398 MAX_CGROUP_ROOT_NAMELEN - 1,
1406 for_each_subsys(ss, i) {
1407 if (strcmp(token, ss->name))
1412 /* Mutually exclusive option 'all' + subsystem name */
1415 opts->subsys_mask |= (1 << i);
1420 if (i == CGROUP_SUBSYS_COUNT)
1424 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1425 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1427 pr_err("sane_behavior: no other mount options allowed\n");
1434 * If the 'all' option was specified select all the subsystems,
1435 * otherwise if 'none', 'name=' and a subsystem name options were
1436 * not specified, let's default to 'all'
1438 if (all_ss || (!one_ss && !opts->none && !opts->name))
1439 for_each_subsys(ss, i)
1441 opts->subsys_mask |= (1 << i);
1444 * We either have to specify by name or by subsystems. (So all
1445 * empty hierarchies must have a name).
1447 if (!opts->subsys_mask && !opts->name)
1451 * Option noprefix was introduced just for backward compatibility
1452 * with the old cpuset, so we allow noprefix only if mounting just
1453 * the cpuset subsystem.
1455 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1458 /* Can't specify "none" and some subsystems */
1459 if (opts->subsys_mask && opts->none)
1465 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1468 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1469 struct cgroup_sb_opts opts;
1470 unsigned int added_mask, removed_mask;
1472 if (root == &cgrp_dfl_root) {
1473 pr_err("remount is not allowed\n");
1477 mutex_lock(&cgroup_mutex);
1479 /* See what subsystems are wanted */
1480 ret = parse_cgroupfs_options(data, &opts);
1484 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1485 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1486 task_tgid_nr(current), current->comm);
1488 added_mask = opts.subsys_mask & ~root->subsys_mask;
1489 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1491 /* Don't allow flags or name to change at remount */
1492 if ((opts.flags ^ root->flags) ||
1493 (opts.name && strcmp(opts.name, root->name))) {
1494 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1495 opts.flags, opts.name ?: "", root->flags, root->name);
1500 /* remounting is not allowed for populated hierarchies */
1501 if (!list_empty(&root->cgrp.self.children)) {
1506 ret = rebind_subsystems(root, added_mask);
1510 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1512 if (opts.release_agent) {
1513 spin_lock(&release_agent_path_lock);
1514 strcpy(root->release_agent_path, opts.release_agent);
1515 spin_unlock(&release_agent_path_lock);
1518 kfree(opts.release_agent);
1520 mutex_unlock(&cgroup_mutex);
1525 * To reduce the fork() overhead for systems that are not actually using
1526 * their cgroups capability, we don't maintain the lists running through
1527 * each css_set to its tasks until we see the list actually used - in other
1528 * words after the first mount.
1530 static bool use_task_css_set_links __read_mostly;
1532 static void cgroup_enable_task_cg_lists(void)
1534 struct task_struct *p, *g;
1536 down_write(&css_set_rwsem);
1538 if (use_task_css_set_links)
1541 use_task_css_set_links = true;
1544 * We need tasklist_lock because RCU is not safe against
1545 * while_each_thread(). Besides, a forking task that has passed
1546 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1547 * is not guaranteed to have its child immediately visible in the
1548 * tasklist if we walk through it with RCU.
1550 read_lock(&tasklist_lock);
1551 do_each_thread(g, p) {
1552 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1553 task_css_set(p) != &init_css_set);
1556 * We should check if the process is exiting, otherwise
1557 * it will race with cgroup_exit() in that the list
1558 * entry won't be deleted though the process has exited.
1559 * Do it while holding siglock so that we don't end up
1560 * racing against cgroup_exit().
1562 spin_lock_irq(&p->sighand->siglock);
1563 if (!(p->flags & PF_EXITING)) {
1564 struct css_set *cset = task_css_set(p);
1566 list_add(&p->cg_list, &cset->tasks);
1569 spin_unlock_irq(&p->sighand->siglock);
1570 } while_each_thread(g, p);
1571 read_unlock(&tasklist_lock);
1573 up_write(&css_set_rwsem);
1576 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1578 struct cgroup_subsys *ss;
1581 INIT_LIST_HEAD(&cgrp->self.sibling);
1582 INIT_LIST_HEAD(&cgrp->self.children);
1583 INIT_LIST_HEAD(&cgrp->cset_links);
1584 INIT_LIST_HEAD(&cgrp->release_list);
1585 INIT_LIST_HEAD(&cgrp->pidlists);
1586 mutex_init(&cgrp->pidlist_mutex);
1587 cgrp->self.cgroup = cgrp;
1588 cgrp->self.flags |= CSS_ONLINE;
1590 for_each_subsys(ss, ssid)
1591 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1593 init_waitqueue_head(&cgrp->offline_waitq);
1596 static void init_cgroup_root(struct cgroup_root *root,
1597 struct cgroup_sb_opts *opts)
1599 struct cgroup *cgrp = &root->cgrp;
1601 INIT_LIST_HEAD(&root->root_list);
1602 atomic_set(&root->nr_cgrps, 1);
1604 init_cgroup_housekeeping(cgrp);
1605 idr_init(&root->cgroup_idr);
1607 root->flags = opts->flags;
1608 if (opts->release_agent)
1609 strcpy(root->release_agent_path, opts->release_agent);
1611 strcpy(root->name, opts->name);
1612 if (opts->cpuset_clone_children)
1613 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1616 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1618 LIST_HEAD(tmp_links);
1619 struct cgroup *root_cgrp = &root->cgrp;
1620 struct cftype *base_files;
1621 struct css_set *cset;
1624 lockdep_assert_held(&cgroup_mutex);
1626 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1629 root_cgrp->id = ret;
1631 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
1636 * We're accessing css_set_count without locking css_set_rwsem here,
1637 * but that's OK - it can only be increased by someone holding
1638 * cgroup_lock, and that's us. The worst that can happen is that we
1639 * have some link structures left over
1641 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1645 ret = cgroup_init_root_id(root);
1649 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1650 KERNFS_ROOT_CREATE_DEACTIVATED,
1652 if (IS_ERR(root->kf_root)) {
1653 ret = PTR_ERR(root->kf_root);
1656 root_cgrp->kn = root->kf_root->kn;
1658 if (root == &cgrp_dfl_root)
1659 base_files = cgroup_dfl_base_files;
1661 base_files = cgroup_legacy_base_files;
1663 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1667 ret = rebind_subsystems(root, ss_mask);
1672 * There must be no failure case after here, since rebinding takes
1673 * care of subsystems' refcounts, which are explicitly dropped in
1674 * the failure exit path.
1676 list_add(&root->root_list, &cgroup_roots);
1677 cgroup_root_count++;
1680 * Link the root cgroup in this hierarchy into all the css_set
1683 down_write(&css_set_rwsem);
1684 hash_for_each(css_set_table, i, cset, hlist)
1685 link_css_set(&tmp_links, cset, root_cgrp);
1686 up_write(&css_set_rwsem);
1688 BUG_ON(!list_empty(&root_cgrp->self.children));
1689 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1691 kernfs_activate(root_cgrp->kn);
1696 kernfs_destroy_root(root->kf_root);
1697 root->kf_root = NULL;
1699 cgroup_exit_root_id(root);
1701 percpu_ref_exit(&root_cgrp->self.refcnt);
1703 free_cgrp_cset_links(&tmp_links);
1707 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1708 int flags, const char *unused_dev_name,
1711 struct super_block *pinned_sb = NULL;
1712 struct cgroup_subsys *ss;
1713 struct cgroup_root *root;
1714 struct cgroup_sb_opts opts;
1715 struct dentry *dentry;
1721 * The first time anyone tries to mount a cgroup, enable the list
1722 * linking each css_set to its tasks and fix up all existing tasks.
1724 if (!use_task_css_set_links)
1725 cgroup_enable_task_cg_lists();
1727 mutex_lock(&cgroup_mutex);
1729 /* First find the desired set of subsystems */
1730 ret = parse_cgroupfs_options(data, &opts);
1734 /* look for a matching existing root */
1735 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1736 cgrp_dfl_root_visible = true;
1737 root = &cgrp_dfl_root;
1738 cgroup_get(&root->cgrp);
1744 * Destruction of cgroup root is asynchronous, so subsystems may
1745 * still be dying after the previous unmount. Let's drain the
1746 * dying subsystems. We just need to ensure that the ones
1747 * unmounted previously finish dying and don't care about new ones
1748 * starting. Testing ref liveliness is good enough.
1750 for_each_subsys(ss, i) {
1751 if (!(opts.subsys_mask & (1 << i)) ||
1752 ss->root == &cgrp_dfl_root)
1755 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1756 mutex_unlock(&cgroup_mutex);
1758 ret = restart_syscall();
1761 cgroup_put(&ss->root->cgrp);
1764 for_each_root(root) {
1765 bool name_match = false;
1767 if (root == &cgrp_dfl_root)
1771 * If we asked for a name then it must match. Also, if
1772 * name matches but sybsys_mask doesn't, we should fail.
1773 * Remember whether name matched.
1776 if (strcmp(opts.name, root->name))
1782 * If we asked for subsystems (or explicitly for no
1783 * subsystems) then they must match.
1785 if ((opts.subsys_mask || opts.none) &&
1786 (opts.subsys_mask != root->subsys_mask)) {
1793 if (root->flags ^ opts.flags)
1794 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1797 * We want to reuse @root whose lifetime is governed by its
1798 * ->cgrp. Let's check whether @root is alive and keep it
1799 * that way. As cgroup_kill_sb() can happen anytime, we
1800 * want to block it by pinning the sb so that @root doesn't
1801 * get killed before mount is complete.
1803 * With the sb pinned, tryget_live can reliably indicate
1804 * whether @root can be reused. If it's being killed,
1805 * drain it. We can use wait_queue for the wait but this
1806 * path is super cold. Let's just sleep a bit and retry.
1808 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1809 if (IS_ERR(pinned_sb) ||
1810 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1811 mutex_unlock(&cgroup_mutex);
1812 if (!IS_ERR_OR_NULL(pinned_sb))
1813 deactivate_super(pinned_sb);
1815 ret = restart_syscall();
1824 * No such thing, create a new one. name= matching without subsys
1825 * specification is allowed for already existing hierarchies but we
1826 * can't create new one without subsys specification.
1828 if (!opts.subsys_mask && !opts.none) {
1833 root = kzalloc(sizeof(*root), GFP_KERNEL);
1839 init_cgroup_root(root, &opts);
1841 ret = cgroup_setup_root(root, opts.subsys_mask);
1843 cgroup_free_root(root);
1846 mutex_unlock(&cgroup_mutex);
1848 kfree(opts.release_agent);
1852 return ERR_PTR(ret);
1854 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1855 CGROUP_SUPER_MAGIC, &new_sb);
1856 if (IS_ERR(dentry) || !new_sb)
1857 cgroup_put(&root->cgrp);
1860 * If @pinned_sb, we're reusing an existing root and holding an
1861 * extra ref on its sb. Mount is complete. Put the extra ref.
1865 deactivate_super(pinned_sb);
1871 static void cgroup_kill_sb(struct super_block *sb)
1873 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1874 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1877 * If @root doesn't have any mounts or children, start killing it.
1878 * This prevents new mounts by disabling percpu_ref_tryget_live().
1879 * cgroup_mount() may wait for @root's release.
1881 * And don't kill the default root.
1883 if (css_has_online_children(&root->cgrp.self) ||
1884 root == &cgrp_dfl_root)
1885 cgroup_put(&root->cgrp);
1887 percpu_ref_kill(&root->cgrp.self.refcnt);
1892 static struct file_system_type cgroup_fs_type = {
1894 .mount = cgroup_mount,
1895 .kill_sb = cgroup_kill_sb,
1898 static struct kobject *cgroup_kobj;
1901 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1902 * @task: target task
1903 * @buf: the buffer to write the path into
1904 * @buflen: the length of the buffer
1906 * Determine @task's cgroup on the first (the one with the lowest non-zero
1907 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1908 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1909 * cgroup controller callbacks.
1911 * Return value is the same as kernfs_path().
1913 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1915 struct cgroup_root *root;
1916 struct cgroup *cgrp;
1917 int hierarchy_id = 1;
1920 mutex_lock(&cgroup_mutex);
1921 down_read(&css_set_rwsem);
1923 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1926 cgrp = task_cgroup_from_root(task, root);
1927 path = cgroup_path(cgrp, buf, buflen);
1929 /* if no hierarchy exists, everyone is in "/" */
1930 if (strlcpy(buf, "/", buflen) < buflen)
1934 up_read(&css_set_rwsem);
1935 mutex_unlock(&cgroup_mutex);
1938 EXPORT_SYMBOL_GPL(task_cgroup_path);
1940 /* used to track tasks and other necessary states during migration */
1941 struct cgroup_taskset {
1942 /* the src and dst cset list running through cset->mg_node */
1943 struct list_head src_csets;
1944 struct list_head dst_csets;
1947 * Fields for cgroup_taskset_*() iteration.
1949 * Before migration is committed, the target migration tasks are on
1950 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1951 * the csets on ->dst_csets. ->csets point to either ->src_csets
1952 * or ->dst_csets depending on whether migration is committed.
1954 * ->cur_csets and ->cur_task point to the current task position
1957 struct list_head *csets;
1958 struct css_set *cur_cset;
1959 struct task_struct *cur_task;
1963 * cgroup_taskset_first - reset taskset and return the first task
1964 * @tset: taskset of interest
1966 * @tset iteration is initialized and the first task is returned.
1968 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1970 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
1971 tset->cur_task = NULL;
1973 return cgroup_taskset_next(tset);
1977 * cgroup_taskset_next - iterate to the next task in taskset
1978 * @tset: taskset of interest
1980 * Return the next task in @tset. Iteration must have been initialized
1981 * with cgroup_taskset_first().
1983 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1985 struct css_set *cset = tset->cur_cset;
1986 struct task_struct *task = tset->cur_task;
1988 while (&cset->mg_node != tset->csets) {
1990 task = list_first_entry(&cset->mg_tasks,
1991 struct task_struct, cg_list);
1993 task = list_next_entry(task, cg_list);
1995 if (&task->cg_list != &cset->mg_tasks) {
1996 tset->cur_cset = cset;
1997 tset->cur_task = task;
2001 cset = list_next_entry(cset, mg_node);
2009 * cgroup_task_migrate - move a task from one cgroup to another.
2010 * @old_cgrp: the cgroup @tsk is being migrated from
2011 * @tsk: the task being migrated
2012 * @new_cset: the new css_set @tsk is being attached to
2014 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2016 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2017 struct task_struct *tsk,
2018 struct css_set *new_cset)
2020 struct css_set *old_cset;
2022 lockdep_assert_held(&cgroup_mutex);
2023 lockdep_assert_held(&css_set_rwsem);
2026 * We are synchronized through threadgroup_lock() against PF_EXITING
2027 * setting such that we can't race against cgroup_exit() changing the
2028 * css_set to init_css_set and dropping the old one.
2030 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2031 old_cset = task_css_set(tsk);
2033 get_css_set(new_cset);
2034 rcu_assign_pointer(tsk->cgroups, new_cset);
2037 * Use move_tail so that cgroup_taskset_first() still returns the
2038 * leader after migration. This works because cgroup_migrate()
2039 * ensures that the dst_cset of the leader is the first on the
2040 * tset's dst_csets list.
2042 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2045 * We just gained a reference on old_cset by taking it from the
2046 * task. As trading it for new_cset is protected by cgroup_mutex,
2047 * we're safe to drop it here; it will be freed under RCU.
2049 set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
2050 put_css_set_locked(old_cset, false);
2054 * cgroup_migrate_finish - cleanup after attach
2055 * @preloaded_csets: list of preloaded css_sets
2057 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2058 * those functions for details.
2060 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2062 struct css_set *cset, *tmp_cset;
2064 lockdep_assert_held(&cgroup_mutex);
2066 down_write(&css_set_rwsem);
2067 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2068 cset->mg_src_cgrp = NULL;
2069 cset->mg_dst_cset = NULL;
2070 list_del_init(&cset->mg_preload_node);
2071 put_css_set_locked(cset, false);
2073 up_write(&css_set_rwsem);
2077 * cgroup_migrate_add_src - add a migration source css_set
2078 * @src_cset: the source css_set to add
2079 * @dst_cgrp: the destination cgroup
2080 * @preloaded_csets: list of preloaded css_sets
2082 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2083 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2084 * up by cgroup_migrate_finish().
2086 * This function may be called without holding threadgroup_lock even if the
2087 * target is a process. Threads may be created and destroyed but as long
2088 * as cgroup_mutex is not dropped, no new css_set can be put into play and
2089 * the preloaded css_sets are guaranteed to cover all migrations.
2091 static void cgroup_migrate_add_src(struct css_set *src_cset,
2092 struct cgroup *dst_cgrp,
2093 struct list_head *preloaded_csets)
2095 struct cgroup *src_cgrp;
2097 lockdep_assert_held(&cgroup_mutex);
2098 lockdep_assert_held(&css_set_rwsem);
2100 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2102 if (!list_empty(&src_cset->mg_preload_node))
2105 WARN_ON(src_cset->mg_src_cgrp);
2106 WARN_ON(!list_empty(&src_cset->mg_tasks));
2107 WARN_ON(!list_empty(&src_cset->mg_node));
2109 src_cset->mg_src_cgrp = src_cgrp;
2110 get_css_set(src_cset);
2111 list_add(&src_cset->mg_preload_node, preloaded_csets);
2115 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2116 * @dst_cgrp: the destination cgroup (may be %NULL)
2117 * @preloaded_csets: list of preloaded source css_sets
2119 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2120 * have been preloaded to @preloaded_csets. This function looks up and
2121 * pins all destination css_sets, links each to its source, and append them
2122 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2123 * source css_set is assumed to be its cgroup on the default hierarchy.
2125 * This function must be called after cgroup_migrate_add_src() has been
2126 * called on each migration source css_set. After migration is performed
2127 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2130 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2131 struct list_head *preloaded_csets)
2134 struct css_set *src_cset, *tmp_cset;
2136 lockdep_assert_held(&cgroup_mutex);
2139 * Except for the root, child_subsys_mask must be zero for a cgroup
2140 * with tasks so that child cgroups don't compete against tasks.
2142 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2143 dst_cgrp->child_subsys_mask)
2146 /* look up the dst cset for each src cset and link it to src */
2147 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2148 struct css_set *dst_cset;
2150 dst_cset = find_css_set(src_cset,
2151 dst_cgrp ?: src_cset->dfl_cgrp);
2155 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2158 * If src cset equals dst, it's noop. Drop the src.
2159 * cgroup_migrate() will skip the cset too. Note that we
2160 * can't handle src == dst as some nodes are used by both.
2162 if (src_cset == dst_cset) {
2163 src_cset->mg_src_cgrp = NULL;
2164 list_del_init(&src_cset->mg_preload_node);
2165 put_css_set(src_cset, false);
2166 put_css_set(dst_cset, false);
2170 src_cset->mg_dst_cset = dst_cset;
2172 if (list_empty(&dst_cset->mg_preload_node))
2173 list_add(&dst_cset->mg_preload_node, &csets);
2175 put_css_set(dst_cset, false);
2178 list_splice_tail(&csets, preloaded_csets);
2181 cgroup_migrate_finish(&csets);
2186 * cgroup_migrate - migrate a process or task to a cgroup
2187 * @cgrp: the destination cgroup
2188 * @leader: the leader of the process or the task to migrate
2189 * @threadgroup: whether @leader points to the whole process or a single task
2191 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2192 * process, the caller must be holding threadgroup_lock of @leader. The
2193 * caller is also responsible for invoking cgroup_migrate_add_src() and
2194 * cgroup_migrate_prepare_dst() on the targets before invoking this
2195 * function and following up with cgroup_migrate_finish().
2197 * As long as a controller's ->can_attach() doesn't fail, this function is
2198 * guaranteed to succeed. This means that, excluding ->can_attach()
2199 * failure, when migrating multiple targets, the success or failure can be
2200 * decided for all targets by invoking group_migrate_prepare_dst() before
2201 * actually starting migrating.
2203 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2206 struct cgroup_taskset tset = {
2207 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2208 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2209 .csets = &tset.src_csets,
2211 struct cgroup_subsys_state *css, *failed_css = NULL;
2212 struct css_set *cset, *tmp_cset;
2213 struct task_struct *task, *tmp_task;
2217 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2218 * already PF_EXITING could be freed from underneath us unless we
2219 * take an rcu_read_lock.
2221 down_write(&css_set_rwsem);
2225 /* @task either already exited or can't exit until the end */
2226 if (task->flags & PF_EXITING)
2229 /* leave @task alone if post_fork() hasn't linked it yet */
2230 if (list_empty(&task->cg_list))
2233 cset = task_css_set(task);
2234 if (!cset->mg_src_cgrp)
2238 * cgroup_taskset_first() must always return the leader.
2239 * Take care to avoid disturbing the ordering.
2241 list_move_tail(&task->cg_list, &cset->mg_tasks);
2242 if (list_empty(&cset->mg_node))
2243 list_add_tail(&cset->mg_node, &tset.src_csets);
2244 if (list_empty(&cset->mg_dst_cset->mg_node))
2245 list_move_tail(&cset->mg_dst_cset->mg_node,
2250 } while_each_thread(leader, task);
2252 up_write(&css_set_rwsem);
2254 /* methods shouldn't be called if no task is actually migrating */
2255 if (list_empty(&tset.src_csets))
2258 /* check that we can legitimately attach to the cgroup */
2259 for_each_e_css(css, i, cgrp) {
2260 if (css->ss->can_attach) {
2261 ret = css->ss->can_attach(css, &tset);
2264 goto out_cancel_attach;
2270 * Now that we're guaranteed success, proceed to move all tasks to
2271 * the new cgroup. There are no failure cases after here, so this
2272 * is the commit point.
2274 down_write(&css_set_rwsem);
2275 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2276 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2277 cgroup_task_migrate(cset->mg_src_cgrp, task,
2280 up_write(&css_set_rwsem);
2283 * Migration is committed, all target tasks are now on dst_csets.
2284 * Nothing is sensitive to fork() after this point. Notify
2285 * controllers that migration is complete.
2287 tset.csets = &tset.dst_csets;
2289 for_each_e_css(css, i, cgrp)
2290 if (css->ss->attach)
2291 css->ss->attach(css, &tset);
2294 goto out_release_tset;
2297 for_each_e_css(css, i, cgrp) {
2298 if (css == failed_css)
2300 if (css->ss->cancel_attach)
2301 css->ss->cancel_attach(css, &tset);
2304 down_write(&css_set_rwsem);
2305 list_splice_init(&tset.dst_csets, &tset.src_csets);
2306 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2307 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2308 list_del_init(&cset->mg_node);
2310 up_write(&css_set_rwsem);
2315 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2316 * @dst_cgrp: the cgroup to attach to
2317 * @leader: the task or the leader of the threadgroup to be attached
2318 * @threadgroup: attach the whole threadgroup?
2320 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2322 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2323 struct task_struct *leader, bool threadgroup)
2325 LIST_HEAD(preloaded_csets);
2326 struct task_struct *task;
2329 /* look up all src csets */
2330 down_read(&css_set_rwsem);
2334 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2338 } while_each_thread(leader, task);
2340 up_read(&css_set_rwsem);
2342 /* prepare dst csets and commit */
2343 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2345 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2347 cgroup_migrate_finish(&preloaded_csets);
2352 * Find the task_struct of the task to attach by vpid and pass it along to the
2353 * function to attach either it or all tasks in its threadgroup. Will lock
2354 * cgroup_mutex and threadgroup.
2356 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2357 size_t nbytes, loff_t off, bool threadgroup)
2359 struct task_struct *tsk;
2360 const struct cred *cred = current_cred(), *tcred;
2361 struct cgroup *cgrp;
2365 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2368 cgrp = cgroup_kn_lock_live(of->kn);
2375 tsk = find_task_by_vpid(pid);
2379 goto out_unlock_cgroup;
2382 * even if we're attaching all tasks in the thread group, we
2383 * only need to check permissions on one of them.
2385 tcred = __task_cred(tsk);
2386 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2387 !uid_eq(cred->euid, tcred->uid) &&
2388 !uid_eq(cred->euid, tcred->suid)) {
2391 goto out_unlock_cgroup;
2397 tsk = tsk->group_leader;
2400 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2401 * trapped in a cpuset, or RT worker may be born in a cgroup
2402 * with no rt_runtime allocated. Just say no.
2404 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2407 goto out_unlock_cgroup;
2410 get_task_struct(tsk);
2413 threadgroup_lock(tsk);
2415 if (!thread_group_leader(tsk)) {
2417 * a race with de_thread from another thread's exec()
2418 * may strip us of our leadership, if this happens,
2419 * there is no choice but to throw this task away and
2420 * try again; this is
2421 * "double-double-toil-and-trouble-check locking".
2423 threadgroup_unlock(tsk);
2424 put_task_struct(tsk);
2425 goto retry_find_task;
2429 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2431 threadgroup_unlock(tsk);
2433 put_task_struct(tsk);
2435 cgroup_kn_unlock(of->kn);
2436 return ret ?: nbytes;
2440 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2441 * @from: attach to all cgroups of a given task
2442 * @tsk: the task to be attached
2444 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2446 struct cgroup_root *root;
2449 mutex_lock(&cgroup_mutex);
2450 for_each_root(root) {
2451 struct cgroup *from_cgrp;
2453 if (root == &cgrp_dfl_root)
2456 down_read(&css_set_rwsem);
2457 from_cgrp = task_cgroup_from_root(from, root);
2458 up_read(&css_set_rwsem);
2460 retval = cgroup_attach_task(from_cgrp, tsk, false);
2464 mutex_unlock(&cgroup_mutex);
2468 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2470 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2471 char *buf, size_t nbytes, loff_t off)
2473 return __cgroup_procs_write(of, buf, nbytes, off, false);
2476 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2477 char *buf, size_t nbytes, loff_t off)
2479 return __cgroup_procs_write(of, buf, nbytes, off, true);
2482 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2483 char *buf, size_t nbytes, loff_t off)
2485 struct cgroup *cgrp;
2487 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2489 cgrp = cgroup_kn_lock_live(of->kn);
2492 spin_lock(&release_agent_path_lock);
2493 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2494 sizeof(cgrp->root->release_agent_path));
2495 spin_unlock(&release_agent_path_lock);
2496 cgroup_kn_unlock(of->kn);
2500 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2502 struct cgroup *cgrp = seq_css(seq)->cgroup;
2504 spin_lock(&release_agent_path_lock);
2505 seq_puts(seq, cgrp->root->release_agent_path);
2506 spin_unlock(&release_agent_path_lock);
2507 seq_putc(seq, '\n');
2511 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2513 seq_puts(seq, "0\n");
2517 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2519 struct cgroup_subsys *ss;
2520 bool printed = false;
2523 for_each_subsys(ss, ssid) {
2524 if (ss_mask & (1 << ssid)) {
2527 seq_printf(seq, "%s", ss->name);
2532 seq_putc(seq, '\n');
2535 /* show controllers which are currently attached to the default hierarchy */
2536 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2538 struct cgroup *cgrp = seq_css(seq)->cgroup;
2540 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2541 ~cgrp_dfl_root_inhibit_ss_mask);
2545 /* show controllers which are enabled from the parent */
2546 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2548 struct cgroup *cgrp = seq_css(seq)->cgroup;
2550 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2554 /* show controllers which are enabled for a given cgroup's children */
2555 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2557 struct cgroup *cgrp = seq_css(seq)->cgroup;
2559 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2564 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2565 * @cgrp: root of the subtree to update csses for
2567 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2568 * css associations need to be updated accordingly. This function looks up
2569 * all css_sets which are attached to the subtree, creates the matching
2570 * updated css_sets and migrates the tasks to the new ones.
2572 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2574 LIST_HEAD(preloaded_csets);
2575 struct cgroup_subsys_state *css;
2576 struct css_set *src_cset;
2579 lockdep_assert_held(&cgroup_mutex);
2581 /* look up all csses currently attached to @cgrp's subtree */
2582 down_read(&css_set_rwsem);
2583 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2584 struct cgrp_cset_link *link;
2586 /* self is not affected by child_subsys_mask change */
2587 if (css->cgroup == cgrp)
2590 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2591 cgroup_migrate_add_src(link->cset, cgrp,
2594 up_read(&css_set_rwsem);
2596 /* NULL dst indicates self on default hierarchy */
2597 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2601 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2602 struct task_struct *last_task = NULL, *task;
2604 /* src_csets precede dst_csets, break on the first dst_cset */
2605 if (!src_cset->mg_src_cgrp)
2609 * All tasks in src_cset need to be migrated to the
2610 * matching dst_cset. Empty it process by process. We
2611 * walk tasks but migrate processes. The leader might even
2612 * belong to a different cset but such src_cset would also
2613 * be among the target src_csets because the default
2614 * hierarchy enforces per-process membership.
2617 down_read(&css_set_rwsem);
2618 task = list_first_entry_or_null(&src_cset->tasks,
2619 struct task_struct, cg_list);
2621 task = task->group_leader;
2622 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2623 get_task_struct(task);
2625 up_read(&css_set_rwsem);
2630 /* guard against possible infinite loop */
2631 if (WARN(last_task == task,
2632 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2636 threadgroup_lock(task);
2637 /* raced against de_thread() from another thread? */
2638 if (!thread_group_leader(task)) {
2639 threadgroup_unlock(task);
2640 put_task_struct(task);
2644 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2646 threadgroup_unlock(task);
2647 put_task_struct(task);
2649 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2655 cgroup_migrate_finish(&preloaded_csets);
2659 /* change the enabled child controllers for a cgroup in the default hierarchy */
2660 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2661 char *buf, size_t nbytes,
2664 unsigned int enable = 0, disable = 0;
2665 unsigned int css_enable, css_disable, old_ctrl, new_ctrl;
2666 struct cgroup *cgrp, *child;
2667 struct cgroup_subsys *ss;
2672 * Parse input - space separated list of subsystem names prefixed
2673 * with either + or -.
2675 buf = strstrip(buf);
2676 while ((tok = strsep(&buf, " "))) {
2679 for_each_subsys(ss, ssid) {
2680 if (ss->disabled || strcmp(tok + 1, ss->name) ||
2681 ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2685 enable |= 1 << ssid;
2686 disable &= ~(1 << ssid);
2687 } else if (*tok == '-') {
2688 disable |= 1 << ssid;
2689 enable &= ~(1 << ssid);
2695 if (ssid == CGROUP_SUBSYS_COUNT)
2699 cgrp = cgroup_kn_lock_live(of->kn);
2703 for_each_subsys(ss, ssid) {
2704 if (enable & (1 << ssid)) {
2705 if (cgrp->subtree_control & (1 << ssid)) {
2706 enable &= ~(1 << ssid);
2710 /* unavailable or not enabled on the parent? */
2711 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2712 (cgroup_parent(cgrp) &&
2713 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2719 * @ss is already enabled through dependency and
2720 * we'll just make it visible. Skip draining.
2722 if (cgrp->child_subsys_mask & (1 << ssid))
2726 * Because css offlining is asynchronous, userland
2727 * might try to re-enable the same controller while
2728 * the previous instance is still around. In such
2729 * cases, wait till it's gone using offline_waitq.
2731 cgroup_for_each_live_child(child, cgrp) {
2734 if (!cgroup_css(child, ss))
2738 prepare_to_wait(&child->offline_waitq, &wait,
2739 TASK_UNINTERRUPTIBLE);
2740 cgroup_kn_unlock(of->kn);
2742 finish_wait(&child->offline_waitq, &wait);
2745 return restart_syscall();
2747 } else if (disable & (1 << ssid)) {
2748 if (!(cgrp->subtree_control & (1 << ssid))) {
2749 disable &= ~(1 << ssid);
2753 /* a child has it enabled? */
2754 cgroup_for_each_live_child(child, cgrp) {
2755 if (child->subtree_control & (1 << ssid)) {
2763 if (!enable && !disable) {
2769 * Except for the root, subtree_control must be zero for a cgroup
2770 * with tasks so that child cgroups don't compete against tasks.
2772 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2778 * Update subsys masks and calculate what needs to be done. More
2779 * subsystems than specified may need to be enabled or disabled
2780 * depending on subsystem dependencies.
2782 cgrp->subtree_control |= enable;
2783 cgrp->subtree_control &= ~disable;
2785 old_ctrl = cgrp->child_subsys_mask;
2786 cgroup_refresh_child_subsys_mask(cgrp);
2787 new_ctrl = cgrp->child_subsys_mask;
2789 css_enable = ~old_ctrl & new_ctrl;
2790 css_disable = old_ctrl & ~new_ctrl;
2791 enable |= css_enable;
2792 disable |= css_disable;
2795 * Create new csses or make the existing ones visible. A css is
2796 * created invisible if it's being implicitly enabled through
2797 * dependency. An invisible css is made visible when the userland
2798 * explicitly enables it.
2800 for_each_subsys(ss, ssid) {
2801 if (!(enable & (1 << ssid)))
2804 cgroup_for_each_live_child(child, cgrp) {
2805 if (css_enable & (1 << ssid))
2806 ret = create_css(child, ss,
2807 cgrp->subtree_control & (1 << ssid));
2809 ret = cgroup_populate_dir(child, 1 << ssid);
2816 * At this point, cgroup_e_css() results reflect the new csses
2817 * making the following cgroup_update_dfl_csses() properly update
2818 * css associations of all tasks in the subtree.
2820 ret = cgroup_update_dfl_csses(cgrp);
2825 * All tasks are migrated out of disabled csses. Kill or hide
2826 * them. A css is hidden when the userland requests it to be
2827 * disabled while other subsystems are still depending on it. The
2828 * css must not actively control resources and be in the vanilla
2829 * state if it's made visible again later. Controllers which may
2830 * be depended upon should provide ->css_reset() for this purpose.
2832 for_each_subsys(ss, ssid) {
2833 if (!(disable & (1 << ssid)))
2836 cgroup_for_each_live_child(child, cgrp) {
2837 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2839 if (css_disable & (1 << ssid)) {
2842 cgroup_clear_dir(child, 1 << ssid);
2849 kernfs_activate(cgrp->kn);
2852 cgroup_kn_unlock(of->kn);
2853 return ret ?: nbytes;
2856 cgrp->subtree_control &= ~enable;
2857 cgrp->subtree_control |= disable;
2858 cgroup_refresh_child_subsys_mask(cgrp);
2860 for_each_subsys(ss, ssid) {
2861 if (!(enable & (1 << ssid)))
2864 cgroup_for_each_live_child(child, cgrp) {
2865 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2870 if (css_enable & (1 << ssid))
2873 cgroup_clear_dir(child, 1 << ssid);
2879 static int cgroup_populated_show(struct seq_file *seq, void *v)
2881 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2885 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2886 size_t nbytes, loff_t off)
2888 struct cgroup *cgrp = of->kn->parent->priv;
2889 struct cftype *cft = of->kn->priv;
2890 struct cgroup_subsys_state *css;
2894 return cft->write(of, buf, nbytes, off);
2897 * kernfs guarantees that a file isn't deleted with operations in
2898 * flight, which means that the matching css is and stays alive and
2899 * doesn't need to be pinned. The RCU locking is not necessary
2900 * either. It's just for the convenience of using cgroup_css().
2903 css = cgroup_css(cgrp, cft->ss);
2906 if (cft->write_u64) {
2907 unsigned long long v;
2908 ret = kstrtoull(buf, 0, &v);
2910 ret = cft->write_u64(css, cft, v);
2911 } else if (cft->write_s64) {
2913 ret = kstrtoll(buf, 0, &v);
2915 ret = cft->write_s64(css, cft, v);
2920 return ret ?: nbytes;
2923 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2925 return seq_cft(seq)->seq_start(seq, ppos);
2928 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2930 return seq_cft(seq)->seq_next(seq, v, ppos);
2933 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2935 seq_cft(seq)->seq_stop(seq, v);
2938 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2940 struct cftype *cft = seq_cft(m);
2941 struct cgroup_subsys_state *css = seq_css(m);
2944 return cft->seq_show(m, arg);
2947 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2948 else if (cft->read_s64)
2949 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2955 static struct kernfs_ops cgroup_kf_single_ops = {
2956 .atomic_write_len = PAGE_SIZE,
2957 .write = cgroup_file_write,
2958 .seq_show = cgroup_seqfile_show,
2961 static struct kernfs_ops cgroup_kf_ops = {
2962 .atomic_write_len = PAGE_SIZE,
2963 .write = cgroup_file_write,
2964 .seq_start = cgroup_seqfile_start,
2965 .seq_next = cgroup_seqfile_next,
2966 .seq_stop = cgroup_seqfile_stop,
2967 .seq_show = cgroup_seqfile_show,
2971 * cgroup_rename - Only allow simple rename of directories in place.
2973 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
2974 const char *new_name_str)
2976 struct cgroup *cgrp = kn->priv;
2979 if (kernfs_type(kn) != KERNFS_DIR)
2981 if (kn->parent != new_parent)
2985 * This isn't a proper migration and its usefulness is very
2986 * limited. Disallow on the default hierarchy.
2988 if (cgroup_on_dfl(cgrp))
2992 * We're gonna grab cgroup_mutex which nests outside kernfs
2993 * active_ref. kernfs_rename() doesn't require active_ref
2994 * protection. Break them before grabbing cgroup_mutex.
2996 kernfs_break_active_protection(new_parent);
2997 kernfs_break_active_protection(kn);
2999 mutex_lock(&cgroup_mutex);
3001 ret = kernfs_rename(kn, new_parent, new_name_str);
3003 mutex_unlock(&cgroup_mutex);
3005 kernfs_unbreak_active_protection(kn);
3006 kernfs_unbreak_active_protection(new_parent);
3010 /* set uid and gid of cgroup dirs and files to that of the creator */
3011 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3013 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3014 .ia_uid = current_fsuid(),
3015 .ia_gid = current_fsgid(), };
3017 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3018 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3021 return kernfs_setattr(kn, &iattr);
3024 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3026 char name[CGROUP_FILE_NAME_MAX];
3027 struct kernfs_node *kn;
3028 struct lock_class_key *key = NULL;
3031 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3032 key = &cft->lockdep_key;
3034 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3035 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3040 ret = cgroup_kn_set_ugid(kn);
3046 if (cft->seq_show == cgroup_populated_show)
3047 cgrp->populated_kn = kn;
3052 * cgroup_addrm_files - add or remove files to a cgroup directory
3053 * @cgrp: the target cgroup
3054 * @cfts: array of cftypes to be added
3055 * @is_add: whether to add or remove
3057 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3058 * For removals, this function never fails. If addition fails, this
3059 * function doesn't remove files already added. The caller is responsible
3062 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3068 lockdep_assert_held(&cgroup_mutex);
3070 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3071 /* does cft->flags tell us to skip this file on @cgrp? */
3072 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3074 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3076 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3078 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3082 ret = cgroup_add_file(cgrp, cft);
3084 pr_warn("%s: failed to add %s, err=%d\n",
3085 __func__, cft->name, ret);
3089 cgroup_rm_file(cgrp, cft);
3095 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3098 struct cgroup_subsys *ss = cfts[0].ss;
3099 struct cgroup *root = &ss->root->cgrp;
3100 struct cgroup_subsys_state *css;
3103 lockdep_assert_held(&cgroup_mutex);
3105 /* add/rm files for all cgroups created before */
3106 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3107 struct cgroup *cgrp = css->cgroup;
3109 if (cgroup_is_dead(cgrp))
3112 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3118 kernfs_activate(root->kn);
3122 static void cgroup_exit_cftypes(struct cftype *cfts)
3126 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3127 /* free copy for custom atomic_write_len, see init_cftypes() */
3128 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3133 /* revert flags set by cgroup core while adding @cfts */
3134 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3138 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3142 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3143 struct kernfs_ops *kf_ops;
3145 WARN_ON(cft->ss || cft->kf_ops);
3148 kf_ops = &cgroup_kf_ops;
3150 kf_ops = &cgroup_kf_single_ops;
3153 * Ugh... if @cft wants a custom max_write_len, we need to
3154 * make a copy of kf_ops to set its atomic_write_len.
3156 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3157 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3159 cgroup_exit_cftypes(cfts);
3162 kf_ops->atomic_write_len = cft->max_write_len;
3165 cft->kf_ops = kf_ops;
3172 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3174 lockdep_assert_held(&cgroup_mutex);
3176 if (!cfts || !cfts[0].ss)
3179 list_del(&cfts->node);
3180 cgroup_apply_cftypes(cfts, false);
3181 cgroup_exit_cftypes(cfts);
3186 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3187 * @cfts: zero-length name terminated array of cftypes
3189 * Unregister @cfts. Files described by @cfts are removed from all
3190 * existing cgroups and all future cgroups won't have them either. This
3191 * function can be called anytime whether @cfts' subsys is attached or not.
3193 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3196 int cgroup_rm_cftypes(struct cftype *cfts)
3200 mutex_lock(&cgroup_mutex);
3201 ret = cgroup_rm_cftypes_locked(cfts);
3202 mutex_unlock(&cgroup_mutex);
3207 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3208 * @ss: target cgroup subsystem
3209 * @cfts: zero-length name terminated array of cftypes
3211 * Register @cfts to @ss. Files described by @cfts are created for all
3212 * existing cgroups to which @ss is attached and all future cgroups will
3213 * have them too. This function can be called anytime whether @ss is
3216 * Returns 0 on successful registration, -errno on failure. Note that this
3217 * function currently returns 0 as long as @cfts registration is successful
3218 * even if some file creation attempts on existing cgroups fail.
3220 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3227 if (!cfts || cfts[0].name[0] == '\0')
3230 ret = cgroup_init_cftypes(ss, cfts);
3234 mutex_lock(&cgroup_mutex);
3236 list_add_tail(&cfts->node, &ss->cfts);
3237 ret = cgroup_apply_cftypes(cfts, true);
3239 cgroup_rm_cftypes_locked(cfts);
3241 mutex_unlock(&cgroup_mutex);
3246 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3247 * @ss: target cgroup subsystem
3248 * @cfts: zero-length name terminated array of cftypes
3250 * Similar to cgroup_add_cftypes() but the added files are only used for
3251 * the default hierarchy.
3253 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3257 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3258 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3259 return cgroup_add_cftypes(ss, cfts);
3263 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3264 * @ss: target cgroup subsystem
3265 * @cfts: zero-length name terminated array of cftypes
3267 * Similar to cgroup_add_cftypes() but the added files are only used for
3268 * the legacy hierarchies.
3270 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3274 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3275 cft->flags |= __CFTYPE_NOT_ON_DFL;
3276 return cgroup_add_cftypes(ss, cfts);
3280 * cgroup_task_count - count the number of tasks in a cgroup.
3281 * @cgrp: the cgroup in question
3283 * Return the number of tasks in the cgroup.
3285 static int cgroup_task_count(const struct cgroup *cgrp)
3288 struct cgrp_cset_link *link;
3290 down_read(&css_set_rwsem);
3291 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3292 count += atomic_read(&link->cset->refcount);
3293 up_read(&css_set_rwsem);
3298 * css_next_child - find the next child of a given css
3299 * @pos: the current position (%NULL to initiate traversal)
3300 * @parent: css whose children to walk
3302 * This function returns the next child of @parent and should be called
3303 * under either cgroup_mutex or RCU read lock. The only requirement is
3304 * that @parent and @pos are accessible. The next sibling is guaranteed to
3305 * be returned regardless of their states.
3307 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3308 * css which finished ->css_online() is guaranteed to be visible in the
3309 * future iterations and will stay visible until the last reference is put.
3310 * A css which hasn't finished ->css_online() or already finished
3311 * ->css_offline() may show up during traversal. It's each subsystem's
3312 * responsibility to synchronize against on/offlining.
3314 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3315 struct cgroup_subsys_state *parent)
3317 struct cgroup_subsys_state *next;
3319 cgroup_assert_mutex_or_rcu_locked();
3322 * @pos could already have been unlinked from the sibling list.
3323 * Once a cgroup is removed, its ->sibling.next is no longer
3324 * updated when its next sibling changes. CSS_RELEASED is set when
3325 * @pos is taken off list, at which time its next pointer is valid,
3326 * and, as releases are serialized, the one pointed to by the next
3327 * pointer is guaranteed to not have started release yet. This
3328 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3329 * critical section, the one pointed to by its next pointer is
3330 * guaranteed to not have finished its RCU grace period even if we
3331 * have dropped rcu_read_lock() inbetween iterations.
3333 * If @pos has CSS_RELEASED set, its next pointer can't be
3334 * dereferenced; however, as each css is given a monotonically
3335 * increasing unique serial number and always appended to the
3336 * sibling list, the next one can be found by walking the parent's
3337 * children until the first css with higher serial number than
3338 * @pos's. While this path can be slower, it happens iff iteration
3339 * races against release and the race window is very small.
3342 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3343 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3344 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3346 list_for_each_entry_rcu(next, &parent->children, sibling)
3347 if (next->serial_nr > pos->serial_nr)
3352 * @next, if not pointing to the head, can be dereferenced and is
3355 if (&next->sibling != &parent->children)
3361 * css_next_descendant_pre - find the next descendant for pre-order walk
3362 * @pos: the current position (%NULL to initiate traversal)
3363 * @root: css whose descendants to walk
3365 * To be used by css_for_each_descendant_pre(). Find the next descendant
3366 * to visit for pre-order traversal of @root's descendants. @root is
3367 * included in the iteration and the first node to be visited.
3369 * While this function requires cgroup_mutex or RCU read locking, it
3370 * doesn't require the whole traversal to be contained in a single critical
3371 * section. This function will return the correct next descendant as long
3372 * as both @pos and @root are accessible and @pos is a descendant of @root.
3374 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3375 * css which finished ->css_online() is guaranteed to be visible in the
3376 * future iterations and will stay visible until the last reference is put.
3377 * A css which hasn't finished ->css_online() or already finished
3378 * ->css_offline() may show up during traversal. It's each subsystem's
3379 * responsibility to synchronize against on/offlining.
3381 struct cgroup_subsys_state *
3382 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3383 struct cgroup_subsys_state *root)
3385 struct cgroup_subsys_state *next;
3387 cgroup_assert_mutex_or_rcu_locked();
3389 /* if first iteration, visit @root */
3393 /* visit the first child if exists */
3394 next = css_next_child(NULL, pos);
3398 /* no child, visit my or the closest ancestor's next sibling */
3399 while (pos != root) {
3400 next = css_next_child(pos, pos->parent);
3410 * css_rightmost_descendant - return the rightmost descendant of a css
3411 * @pos: css of interest
3413 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3414 * is returned. This can be used during pre-order traversal to skip
3417 * While this function requires cgroup_mutex or RCU read locking, it
3418 * doesn't require the whole traversal to be contained in a single critical
3419 * section. This function will return the correct rightmost descendant as
3420 * long as @pos is accessible.
3422 struct cgroup_subsys_state *
3423 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3425 struct cgroup_subsys_state *last, *tmp;
3427 cgroup_assert_mutex_or_rcu_locked();
3431 /* ->prev isn't RCU safe, walk ->next till the end */
3433 css_for_each_child(tmp, last)
3440 static struct cgroup_subsys_state *
3441 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3443 struct cgroup_subsys_state *last;
3447 pos = css_next_child(NULL, pos);
3454 * css_next_descendant_post - find the next descendant for post-order walk
3455 * @pos: the current position (%NULL to initiate traversal)
3456 * @root: css whose descendants to walk
3458 * To be used by css_for_each_descendant_post(). Find the next descendant
3459 * to visit for post-order traversal of @root's descendants. @root is
3460 * included in the iteration and the last node to be visited.
3462 * While this function requires cgroup_mutex or RCU read locking, it
3463 * doesn't require the whole traversal to be contained in a single critical
3464 * section. This function will return the correct next descendant as long
3465 * as both @pos and @cgroup are accessible and @pos is a descendant of
3468 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3469 * css which finished ->css_online() is guaranteed to be visible in the
3470 * future iterations and will stay visible until the last reference is put.
3471 * A css which hasn't finished ->css_online() or already finished
3472 * ->css_offline() may show up during traversal. It's each subsystem's
3473 * responsibility to synchronize against on/offlining.
3475 struct cgroup_subsys_state *
3476 css_next_descendant_post(struct cgroup_subsys_state *pos,
3477 struct cgroup_subsys_state *root)
3479 struct cgroup_subsys_state *next;
3481 cgroup_assert_mutex_or_rcu_locked();
3483 /* if first iteration, visit leftmost descendant which may be @root */
3485 return css_leftmost_descendant(root);
3487 /* if we visited @root, we're done */
3491 /* if there's an unvisited sibling, visit its leftmost descendant */
3492 next = css_next_child(pos, pos->parent);
3494 return css_leftmost_descendant(next);
3496 /* no sibling left, visit parent */
3501 * css_has_online_children - does a css have online children
3502 * @css: the target css
3504 * Returns %true if @css has any online children; otherwise, %false. This
3505 * function can be called from any context but the caller is responsible
3506 * for synchronizing against on/offlining as necessary.
3508 bool css_has_online_children(struct cgroup_subsys_state *css)
3510 struct cgroup_subsys_state *child;
3514 css_for_each_child(child, css) {
3515 if (child->flags & CSS_ONLINE) {
3525 * css_advance_task_iter - advance a task itererator to the next css_set
3526 * @it: the iterator to advance
3528 * Advance @it to the next css_set to walk.
3530 static void css_advance_task_iter(struct css_task_iter *it)
3532 struct list_head *l = it->cset_pos;
3533 struct cgrp_cset_link *link;
3534 struct css_set *cset;
3536 /* Advance to the next non-empty css_set */
3539 if (l == it->cset_head) {
3540 it->cset_pos = NULL;
3545 cset = container_of(l, struct css_set,
3546 e_cset_node[it->ss->id]);
3548 link = list_entry(l, struct cgrp_cset_link, cset_link);
3551 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3555 if (!list_empty(&cset->tasks))
3556 it->task_pos = cset->tasks.next;
3558 it->task_pos = cset->mg_tasks.next;
3560 it->tasks_head = &cset->tasks;
3561 it->mg_tasks_head = &cset->mg_tasks;
3565 * css_task_iter_start - initiate task iteration
3566 * @css: the css to walk tasks of
3567 * @it: the task iterator to use
3569 * Initiate iteration through the tasks of @css. The caller can call
3570 * css_task_iter_next() to walk through the tasks until the function
3571 * returns NULL. On completion of iteration, css_task_iter_end() must be
3574 * Note that this function acquires a lock which is released when the
3575 * iteration finishes. The caller can't sleep while iteration is in
3578 void css_task_iter_start(struct cgroup_subsys_state *css,
3579 struct css_task_iter *it)
3580 __acquires(css_set_rwsem)
3582 /* no one should try to iterate before mounting cgroups */
3583 WARN_ON_ONCE(!use_task_css_set_links);
3585 down_read(&css_set_rwsem);
3590 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3592 it->cset_pos = &css->cgroup->cset_links;
3594 it->cset_head = it->cset_pos;
3596 css_advance_task_iter(it);
3600 * css_task_iter_next - return the next task for the iterator
3601 * @it: the task iterator being iterated
3603 * The "next" function for task iteration. @it should have been
3604 * initialized via css_task_iter_start(). Returns NULL when the iteration
3607 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3609 struct task_struct *res;
3610 struct list_head *l = it->task_pos;
3612 /* If the iterator cg is NULL, we have no tasks */
3615 res = list_entry(l, struct task_struct, cg_list);
3618 * Advance iterator to find next entry. cset->tasks is consumed
3619 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3624 if (l == it->tasks_head)
3625 l = it->mg_tasks_head->next;
3627 if (l == it->mg_tasks_head)
3628 css_advance_task_iter(it);
3636 * css_task_iter_end - finish task iteration
3637 * @it: the task iterator to finish
3639 * Finish task iteration started by css_task_iter_start().
3641 void css_task_iter_end(struct css_task_iter *it)
3642 __releases(css_set_rwsem)
3644 up_read(&css_set_rwsem);
3648 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3649 * @to: cgroup to which the tasks will be moved
3650 * @from: cgroup in which the tasks currently reside
3652 * Locking rules between cgroup_post_fork() and the migration path
3653 * guarantee that, if a task is forking while being migrated, the new child
3654 * is guaranteed to be either visible in the source cgroup after the
3655 * parent's migration is complete or put into the target cgroup. No task
3656 * can slip out of migration through forking.
3658 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3660 LIST_HEAD(preloaded_csets);
3661 struct cgrp_cset_link *link;
3662 struct css_task_iter it;
3663 struct task_struct *task;
3666 mutex_lock(&cgroup_mutex);
3668 /* all tasks in @from are being moved, all csets are source */
3669 down_read(&css_set_rwsem);
3670 list_for_each_entry(link, &from->cset_links, cset_link)
3671 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3672 up_read(&css_set_rwsem);
3674 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3679 * Migrate tasks one-by-one until @form is empty. This fails iff
3680 * ->can_attach() fails.
3683 css_task_iter_start(&from->self, &it);
3684 task = css_task_iter_next(&it);
3686 get_task_struct(task);
3687 css_task_iter_end(&it);
3690 ret = cgroup_migrate(to, task, false);
3691 put_task_struct(task);
3693 } while (task && !ret);
3695 cgroup_migrate_finish(&preloaded_csets);
3696 mutex_unlock(&cgroup_mutex);
3701 * Stuff for reading the 'tasks'/'procs' files.
3703 * Reading this file can return large amounts of data if a cgroup has
3704 * *lots* of attached tasks. So it may need several calls to read(),
3705 * but we cannot guarantee that the information we produce is correct
3706 * unless we produce it entirely atomically.
3710 /* which pidlist file are we talking about? */
3711 enum cgroup_filetype {
3717 * A pidlist is a list of pids that virtually represents the contents of one
3718 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3719 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3722 struct cgroup_pidlist {
3724 * used to find which pidlist is wanted. doesn't change as long as
3725 * this particular list stays in the list.
3727 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3730 /* how many elements the above list has */
3732 /* each of these stored in a list by its cgroup */
3733 struct list_head links;
3734 /* pointer to the cgroup we belong to, for list removal purposes */
3735 struct cgroup *owner;
3736 /* for delayed destruction */
3737 struct delayed_work destroy_dwork;
3741 * The following two functions "fix" the issue where there are more pids
3742 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3743 * TODO: replace with a kernel-wide solution to this problem
3745 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3746 static void *pidlist_allocate(int count)
3748 if (PIDLIST_TOO_LARGE(count))
3749 return vmalloc(count * sizeof(pid_t));
3751 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3754 static void pidlist_free(void *p)
3756 if (is_vmalloc_addr(p))
3763 * Used to destroy all pidlists lingering waiting for destroy timer. None
3764 * should be left afterwards.
3766 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3768 struct cgroup_pidlist *l, *tmp_l;
3770 mutex_lock(&cgrp->pidlist_mutex);
3771 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3772 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3773 mutex_unlock(&cgrp->pidlist_mutex);
3775 flush_workqueue(cgroup_pidlist_destroy_wq);
3776 BUG_ON(!list_empty(&cgrp->pidlists));
3779 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3781 struct delayed_work *dwork = to_delayed_work(work);
3782 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3784 struct cgroup_pidlist *tofree = NULL;
3786 mutex_lock(&l->owner->pidlist_mutex);
3789 * Destroy iff we didn't get queued again. The state won't change
3790 * as destroy_dwork can only be queued while locked.
3792 if (!delayed_work_pending(dwork)) {
3793 list_del(&l->links);
3794 pidlist_free(l->list);
3795 put_pid_ns(l->key.ns);
3799 mutex_unlock(&l->owner->pidlist_mutex);
3804 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3805 * Returns the number of unique elements.
3807 static int pidlist_uniq(pid_t *list, int length)
3812 * we presume the 0th element is unique, so i starts at 1. trivial
3813 * edge cases first; no work needs to be done for either
3815 if (length == 0 || length == 1)
3817 /* src and dest walk down the list; dest counts unique elements */
3818 for (src = 1; src < length; src++) {
3819 /* find next unique element */
3820 while (list[src] == list[src-1]) {
3825 /* dest always points to where the next unique element goes */
3826 list[dest] = list[src];
3834 * The two pid files - task and cgroup.procs - guaranteed that the result
3835 * is sorted, which forced this whole pidlist fiasco. As pid order is
3836 * different per namespace, each namespace needs differently sorted list,
3837 * making it impossible to use, for example, single rbtree of member tasks
3838 * sorted by task pointer. As pidlists can be fairly large, allocating one
3839 * per open file is dangerous, so cgroup had to implement shared pool of
3840 * pidlists keyed by cgroup and namespace.
3842 * All this extra complexity was caused by the original implementation
3843 * committing to an entirely unnecessary property. In the long term, we
3844 * want to do away with it. Explicitly scramble sort order if on the
3845 * default hierarchy so that no such expectation exists in the new
3848 * Scrambling is done by swapping every two consecutive bits, which is
3849 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3851 static pid_t pid_fry(pid_t pid)
3853 unsigned a = pid & 0x55555555;
3854 unsigned b = pid & 0xAAAAAAAA;
3856 return (a << 1) | (b >> 1);
3859 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3861 if (cgroup_on_dfl(cgrp))
3862 return pid_fry(pid);
3867 static int cmppid(const void *a, const void *b)
3869 return *(pid_t *)a - *(pid_t *)b;
3872 static int fried_cmppid(const void *a, const void *b)
3874 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3877 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3878 enum cgroup_filetype type)
3880 struct cgroup_pidlist *l;
3881 /* don't need task_nsproxy() if we're looking at ourself */
3882 struct pid_namespace *ns = task_active_pid_ns(current);
3884 lockdep_assert_held(&cgrp->pidlist_mutex);
3886 list_for_each_entry(l, &cgrp->pidlists, links)
3887 if (l->key.type == type && l->key.ns == ns)
3893 * find the appropriate pidlist for our purpose (given procs vs tasks)
3894 * returns with the lock on that pidlist already held, and takes care
3895 * of the use count, or returns NULL with no locks held if we're out of
3898 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3899 enum cgroup_filetype type)
3901 struct cgroup_pidlist *l;
3903 lockdep_assert_held(&cgrp->pidlist_mutex);
3905 l = cgroup_pidlist_find(cgrp, type);
3909 /* entry not found; create a new one */
3910 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3914 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3916 /* don't need task_nsproxy() if we're looking at ourself */
3917 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3919 list_add(&l->links, &cgrp->pidlists);
3924 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3926 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3927 struct cgroup_pidlist **lp)
3931 int pid, n = 0; /* used for populating the array */
3932 struct css_task_iter it;
3933 struct task_struct *tsk;
3934 struct cgroup_pidlist *l;
3936 lockdep_assert_held(&cgrp->pidlist_mutex);
3939 * If cgroup gets more users after we read count, we won't have
3940 * enough space - tough. This race is indistinguishable to the
3941 * caller from the case that the additional cgroup users didn't
3942 * show up until sometime later on.
3944 length = cgroup_task_count(cgrp);
3945 array = pidlist_allocate(length);
3948 /* now, populate the array */
3949 css_task_iter_start(&cgrp->self, &it);
3950 while ((tsk = css_task_iter_next(&it))) {
3951 if (unlikely(n == length))
3953 /* get tgid or pid for procs or tasks file respectively */
3954 if (type == CGROUP_FILE_PROCS)
3955 pid = task_tgid_vnr(tsk);
3957 pid = task_pid_vnr(tsk);
3958 if (pid > 0) /* make sure to only use valid results */
3961 css_task_iter_end(&it);
3963 /* now sort & (if procs) strip out duplicates */
3964 if (cgroup_on_dfl(cgrp))
3965 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3967 sort(array, length, sizeof(pid_t), cmppid, NULL);
3968 if (type == CGROUP_FILE_PROCS)
3969 length = pidlist_uniq(array, length);
3971 l = cgroup_pidlist_find_create(cgrp, type);
3973 mutex_unlock(&cgrp->pidlist_mutex);
3974 pidlist_free(array);
3978 /* store array, freeing old if necessary */
3979 pidlist_free(l->list);
3987 * cgroupstats_build - build and fill cgroupstats
3988 * @stats: cgroupstats to fill information into
3989 * @dentry: A dentry entry belonging to the cgroup for which stats have
3992 * Build and fill cgroupstats so that taskstats can export it to user
3995 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3997 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
3998 struct cgroup *cgrp;
3999 struct css_task_iter it;
4000 struct task_struct *tsk;
4002 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4003 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4004 kernfs_type(kn) != KERNFS_DIR)
4007 mutex_lock(&cgroup_mutex);
4010 * We aren't being called from kernfs and there's no guarantee on
4011 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4012 * @kn->priv is RCU safe. Let's do the RCU dancing.
4015 cgrp = rcu_dereference(kn->priv);
4016 if (!cgrp || cgroup_is_dead(cgrp)) {
4018 mutex_unlock(&cgroup_mutex);
4023 css_task_iter_start(&cgrp->self, &it);
4024 while ((tsk = css_task_iter_next(&it))) {
4025 switch (tsk->state) {
4027 stats->nr_running++;
4029 case TASK_INTERRUPTIBLE:
4030 stats->nr_sleeping++;
4032 case TASK_UNINTERRUPTIBLE:
4033 stats->nr_uninterruptible++;
4036 stats->nr_stopped++;
4039 if (delayacct_is_task_waiting_on_io(tsk))
4040 stats->nr_io_wait++;
4044 css_task_iter_end(&it);
4046 mutex_unlock(&cgroup_mutex);
4052 * seq_file methods for the tasks/procs files. The seq_file position is the
4053 * next pid to display; the seq_file iterator is a pointer to the pid
4054 * in the cgroup->l->list array.
4057 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4060 * Initially we receive a position value that corresponds to
4061 * one more than the last pid shown (or 0 on the first call or
4062 * after a seek to the start). Use a binary-search to find the
4063 * next pid to display, if any
4065 struct kernfs_open_file *of = s->private;
4066 struct cgroup *cgrp = seq_css(s)->cgroup;
4067 struct cgroup_pidlist *l;
4068 enum cgroup_filetype type = seq_cft(s)->private;
4069 int index = 0, pid = *pos;
4072 mutex_lock(&cgrp->pidlist_mutex);
4075 * !NULL @of->priv indicates that this isn't the first start()
4076 * after open. If the matching pidlist is around, we can use that.
4077 * Look for it. Note that @of->priv can't be used directly. It
4078 * could already have been destroyed.
4081 of->priv = cgroup_pidlist_find(cgrp, type);
4084 * Either this is the first start() after open or the matching
4085 * pidlist has been destroyed inbetween. Create a new one.
4088 ret = pidlist_array_load(cgrp, type,
4089 (struct cgroup_pidlist **)&of->priv);
4091 return ERR_PTR(ret);
4096 int end = l->length;
4098 while (index < end) {
4099 int mid = (index + end) / 2;
4100 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4103 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4109 /* If we're off the end of the array, we're done */
4110 if (index >= l->length)
4112 /* Update the abstract position to be the actual pid that we found */
4113 iter = l->list + index;
4114 *pos = cgroup_pid_fry(cgrp, *iter);
4118 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4120 struct kernfs_open_file *of = s->private;
4121 struct cgroup_pidlist *l = of->priv;
4124 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4125 CGROUP_PIDLIST_DESTROY_DELAY);
4126 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4129 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4131 struct kernfs_open_file *of = s->private;
4132 struct cgroup_pidlist *l = of->priv;
4134 pid_t *end = l->list + l->length;
4136 * Advance to the next pid in the array. If this goes off the
4143 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4148 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4150 return seq_printf(s, "%d\n", *(int *)v);
4153 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4156 return notify_on_release(css->cgroup);
4159 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4160 struct cftype *cft, u64 val)
4162 clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
4164 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4166 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4170 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4173 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4176 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4177 struct cftype *cft, u64 val)
4180 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4182 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4186 /* cgroup core interface files for the default hierarchy */
4187 static struct cftype cgroup_dfl_base_files[] = {
4189 .name = "cgroup.procs",
4190 .seq_start = cgroup_pidlist_start,
4191 .seq_next = cgroup_pidlist_next,
4192 .seq_stop = cgroup_pidlist_stop,
4193 .seq_show = cgroup_pidlist_show,
4194 .private = CGROUP_FILE_PROCS,
4195 .write = cgroup_procs_write,
4196 .mode = S_IRUGO | S_IWUSR,
4199 .name = "cgroup.controllers",
4200 .flags = CFTYPE_ONLY_ON_ROOT,
4201 .seq_show = cgroup_root_controllers_show,
4204 .name = "cgroup.controllers",
4205 .flags = CFTYPE_NOT_ON_ROOT,
4206 .seq_show = cgroup_controllers_show,
4209 .name = "cgroup.subtree_control",
4210 .seq_show = cgroup_subtree_control_show,
4211 .write = cgroup_subtree_control_write,
4214 .name = "cgroup.populated",
4215 .flags = CFTYPE_NOT_ON_ROOT,
4216 .seq_show = cgroup_populated_show,
4221 /* cgroup core interface files for the legacy hierarchies */
4222 static struct cftype cgroup_legacy_base_files[] = {
4224 .name = "cgroup.procs",
4225 .seq_start = cgroup_pidlist_start,
4226 .seq_next = cgroup_pidlist_next,
4227 .seq_stop = cgroup_pidlist_stop,
4228 .seq_show = cgroup_pidlist_show,
4229 .private = CGROUP_FILE_PROCS,
4230 .write = cgroup_procs_write,
4231 .mode = S_IRUGO | S_IWUSR,
4234 .name = "cgroup.clone_children",
4235 .read_u64 = cgroup_clone_children_read,
4236 .write_u64 = cgroup_clone_children_write,
4239 .name = "cgroup.sane_behavior",
4240 .flags = CFTYPE_ONLY_ON_ROOT,
4241 .seq_show = cgroup_sane_behavior_show,
4245 .seq_start = cgroup_pidlist_start,
4246 .seq_next = cgroup_pidlist_next,
4247 .seq_stop = cgroup_pidlist_stop,
4248 .seq_show = cgroup_pidlist_show,
4249 .private = CGROUP_FILE_TASKS,
4250 .write = cgroup_tasks_write,
4251 .mode = S_IRUGO | S_IWUSR,
4254 .name = "notify_on_release",
4255 .read_u64 = cgroup_read_notify_on_release,
4256 .write_u64 = cgroup_write_notify_on_release,
4259 .name = "release_agent",
4260 .flags = CFTYPE_ONLY_ON_ROOT,
4261 .seq_show = cgroup_release_agent_show,
4262 .write = cgroup_release_agent_write,
4263 .max_write_len = PATH_MAX - 1,
4269 * cgroup_populate_dir - create subsys files in a cgroup directory
4270 * @cgrp: target cgroup
4271 * @subsys_mask: mask of the subsystem ids whose files should be added
4273 * On failure, no file is added.
4275 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4277 struct cgroup_subsys *ss;
4280 /* process cftsets of each subsystem */
4281 for_each_subsys(ss, i) {
4282 struct cftype *cfts;
4284 if (!(subsys_mask & (1 << i)))
4287 list_for_each_entry(cfts, &ss->cfts, node) {
4288 ret = cgroup_addrm_files(cgrp, cfts, true);
4295 cgroup_clear_dir(cgrp, subsys_mask);
4300 * css destruction is four-stage process.
4302 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4303 * Implemented in kill_css().
4305 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4306 * and thus css_tryget_online() is guaranteed to fail, the css can be
4307 * offlined by invoking offline_css(). After offlining, the base ref is
4308 * put. Implemented in css_killed_work_fn().
4310 * 3. When the percpu_ref reaches zero, the only possible remaining
4311 * accessors are inside RCU read sections. css_release() schedules the
4314 * 4. After the grace period, the css can be freed. Implemented in
4315 * css_free_work_fn().
4317 * It is actually hairier because both step 2 and 4 require process context
4318 * and thus involve punting to css->destroy_work adding two additional
4319 * steps to the already complex sequence.
4321 static void css_free_work_fn(struct work_struct *work)
4323 struct cgroup_subsys_state *css =
4324 container_of(work, struct cgroup_subsys_state, destroy_work);
4325 struct cgroup *cgrp = css->cgroup;
4327 percpu_ref_exit(&css->refcnt);
4332 css_put(css->parent);
4334 css->ss->css_free(css);
4337 /* cgroup free path */
4338 atomic_dec(&cgrp->root->nr_cgrps);
4339 cgroup_pidlist_destroy_all(cgrp);
4341 if (cgroup_parent(cgrp)) {
4343 * We get a ref to the parent, and put the ref when
4344 * this cgroup is being freed, so it's guaranteed
4345 * that the parent won't be destroyed before its
4348 cgroup_put(cgroup_parent(cgrp));
4349 kernfs_put(cgrp->kn);
4353 * This is root cgroup's refcnt reaching zero,
4354 * which indicates that the root should be
4357 cgroup_destroy_root(cgrp->root);
4362 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4364 struct cgroup_subsys_state *css =
4365 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4367 INIT_WORK(&css->destroy_work, css_free_work_fn);
4368 queue_work(cgroup_destroy_wq, &css->destroy_work);
4371 static void css_release_work_fn(struct work_struct *work)
4373 struct cgroup_subsys_state *css =
4374 container_of(work, struct cgroup_subsys_state, destroy_work);
4375 struct cgroup_subsys *ss = css->ss;
4376 struct cgroup *cgrp = css->cgroup;
4378 mutex_lock(&cgroup_mutex);
4380 css->flags |= CSS_RELEASED;
4381 list_del_rcu(&css->sibling);
4384 /* css release path */
4385 cgroup_idr_remove(&ss->css_idr, css->id);
4387 /* cgroup release path */
4388 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4392 mutex_unlock(&cgroup_mutex);
4394 call_rcu(&css->rcu_head, css_free_rcu_fn);
4397 static void css_release(struct percpu_ref *ref)
4399 struct cgroup_subsys_state *css =
4400 container_of(ref, struct cgroup_subsys_state, refcnt);
4402 INIT_WORK(&css->destroy_work, css_release_work_fn);
4403 queue_work(cgroup_destroy_wq, &css->destroy_work);
4406 static void init_and_link_css(struct cgroup_subsys_state *css,
4407 struct cgroup_subsys *ss, struct cgroup *cgrp)
4409 lockdep_assert_held(&cgroup_mutex);
4413 memset(css, 0, sizeof(*css));
4416 INIT_LIST_HEAD(&css->sibling);
4417 INIT_LIST_HEAD(&css->children);
4418 css->serial_nr = css_serial_nr_next++;
4420 if (cgroup_parent(cgrp)) {
4421 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4422 css_get(css->parent);
4425 BUG_ON(cgroup_css(cgrp, ss));
4428 /* invoke ->css_online() on a new CSS and mark it online if successful */
4429 static int online_css(struct cgroup_subsys_state *css)
4431 struct cgroup_subsys *ss = css->ss;
4434 lockdep_assert_held(&cgroup_mutex);
4437 ret = ss->css_online(css);
4439 css->flags |= CSS_ONLINE;
4440 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4445 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4446 static void offline_css(struct cgroup_subsys_state *css)
4448 struct cgroup_subsys *ss = css->ss;
4450 lockdep_assert_held(&cgroup_mutex);
4452 if (!(css->flags & CSS_ONLINE))
4455 if (ss->css_offline)
4456 ss->css_offline(css);
4458 css->flags &= ~CSS_ONLINE;
4459 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4461 wake_up_all(&css->cgroup->offline_waitq);
4465 * create_css - create a cgroup_subsys_state
4466 * @cgrp: the cgroup new css will be associated with
4467 * @ss: the subsys of new css
4468 * @visible: whether to create control knobs for the new css or not
4470 * Create a new css associated with @cgrp - @ss pair. On success, the new
4471 * css is online and installed in @cgrp with all interface files created if
4472 * @visible. Returns 0 on success, -errno on failure.
4474 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4477 struct cgroup *parent = cgroup_parent(cgrp);
4478 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4479 struct cgroup_subsys_state *css;
4482 lockdep_assert_held(&cgroup_mutex);
4484 css = ss->css_alloc(parent_css);
4486 return PTR_ERR(css);
4488 init_and_link_css(css, ss, cgrp);
4490 err = percpu_ref_init(&css->refcnt, css_release);
4494 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4496 goto err_free_percpu_ref;
4500 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4505 /* @css is ready to be brought online now, make it visible */
4506 list_add_tail_rcu(&css->sibling, &parent_css->children);
4507 cgroup_idr_replace(&ss->css_idr, css, css->id);
4509 err = online_css(css);
4513 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4514 cgroup_parent(parent)) {
4515 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4516 current->comm, current->pid, ss->name);
4517 if (!strcmp(ss->name, "memory"))
4518 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4519 ss->warned_broken_hierarchy = true;
4525 list_del_rcu(&css->sibling);
4526 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4528 cgroup_idr_remove(&ss->css_idr, css->id);
4529 err_free_percpu_ref:
4530 percpu_ref_exit(&css->refcnt);
4532 call_rcu(&css->rcu_head, css_free_rcu_fn);
4536 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4539 struct cgroup *parent, *cgrp;
4540 struct cgroup_root *root;
4541 struct cgroup_subsys *ss;
4542 struct kernfs_node *kn;
4543 struct cftype *base_files;
4546 parent = cgroup_kn_lock_live(parent_kn);
4549 root = parent->root;
4551 /* allocate the cgroup and its ID, 0 is reserved for the root */
4552 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4558 ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
4563 * Temporarily set the pointer to NULL, so idr_find() won't return
4564 * a half-baked cgroup.
4566 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4569 goto out_cancel_ref;
4572 init_cgroup_housekeeping(cgrp);
4574 cgrp->self.parent = &parent->self;
4577 if (notify_on_release(parent))
4578 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4580 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4581 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4583 /* create the directory */
4584 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4592 * This extra ref will be put in cgroup_free_fn() and guarantees
4593 * that @cgrp->kn is always accessible.
4597 cgrp->self.serial_nr = css_serial_nr_next++;
4599 /* allocation complete, commit to creation */
4600 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4601 atomic_inc(&root->nr_cgrps);
4605 * @cgrp is now fully operational. If something fails after this
4606 * point, it'll be released via the normal destruction path.
4608 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4610 ret = cgroup_kn_set_ugid(kn);
4614 if (cgroup_on_dfl(cgrp))
4615 base_files = cgroup_dfl_base_files;
4617 base_files = cgroup_legacy_base_files;
4619 ret = cgroup_addrm_files(cgrp, base_files, true);
4623 /* let's create and online css's */
4624 for_each_subsys(ss, ssid) {
4625 if (parent->child_subsys_mask & (1 << ssid)) {
4626 ret = create_css(cgrp, ss,
4627 parent->subtree_control & (1 << ssid));
4634 * On the default hierarchy, a child doesn't automatically inherit
4635 * subtree_control from the parent. Each is configured manually.
4637 if (!cgroup_on_dfl(cgrp)) {
4638 cgrp->subtree_control = parent->subtree_control;
4639 cgroup_refresh_child_subsys_mask(cgrp);
4642 kernfs_activate(kn);
4648 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4650 percpu_ref_exit(&cgrp->self.refcnt);
4654 cgroup_kn_unlock(parent_kn);
4658 cgroup_destroy_locked(cgrp);
4663 * This is called when the refcnt of a css is confirmed to be killed.
4664 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4665 * initate destruction and put the css ref from kill_css().
4667 static void css_killed_work_fn(struct work_struct *work)
4669 struct cgroup_subsys_state *css =
4670 container_of(work, struct cgroup_subsys_state, destroy_work);
4672 mutex_lock(&cgroup_mutex);
4674 mutex_unlock(&cgroup_mutex);
4679 /* css kill confirmation processing requires process context, bounce */
4680 static void css_killed_ref_fn(struct percpu_ref *ref)
4682 struct cgroup_subsys_state *css =
4683 container_of(ref, struct cgroup_subsys_state, refcnt);
4685 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4686 queue_work(cgroup_destroy_wq, &css->destroy_work);
4690 * kill_css - destroy a css
4691 * @css: css to destroy
4693 * This function initiates destruction of @css by removing cgroup interface
4694 * files and putting its base reference. ->css_offline() will be invoked
4695 * asynchronously once css_tryget_online() is guaranteed to fail and when
4696 * the reference count reaches zero, @css will be released.
4698 static void kill_css(struct cgroup_subsys_state *css)
4700 lockdep_assert_held(&cgroup_mutex);
4703 * This must happen before css is disassociated with its cgroup.
4704 * See seq_css() for details.
4706 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4709 * Killing would put the base ref, but we need to keep it alive
4710 * until after ->css_offline().
4715 * cgroup core guarantees that, by the time ->css_offline() is
4716 * invoked, no new css reference will be given out via
4717 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4718 * proceed to offlining css's because percpu_ref_kill() doesn't
4719 * guarantee that the ref is seen as killed on all CPUs on return.
4721 * Use percpu_ref_kill_and_confirm() to get notifications as each
4722 * css is confirmed to be seen as killed on all CPUs.
4724 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4728 * cgroup_destroy_locked - the first stage of cgroup destruction
4729 * @cgrp: cgroup to be destroyed
4731 * css's make use of percpu refcnts whose killing latency shouldn't be
4732 * exposed to userland and are RCU protected. Also, cgroup core needs to
4733 * guarantee that css_tryget_online() won't succeed by the time
4734 * ->css_offline() is invoked. To satisfy all the requirements,
4735 * destruction is implemented in the following two steps.
4737 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4738 * userland visible parts and start killing the percpu refcnts of
4739 * css's. Set up so that the next stage will be kicked off once all
4740 * the percpu refcnts are confirmed to be killed.
4742 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4743 * rest of destruction. Once all cgroup references are gone, the
4744 * cgroup is RCU-freed.
4746 * This function implements s1. After this step, @cgrp is gone as far as
4747 * the userland is concerned and a new cgroup with the same name may be
4748 * created. As cgroup doesn't care about the names internally, this
4749 * doesn't cause any problem.
4751 static int cgroup_destroy_locked(struct cgroup *cgrp)
4752 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4754 struct cgroup_subsys_state *css;
4758 lockdep_assert_held(&cgroup_mutex);
4761 * css_set_rwsem synchronizes access to ->cset_links and prevents
4762 * @cgrp from being removed while put_css_set() is in progress.
4764 down_read(&css_set_rwsem);
4765 empty = list_empty(&cgrp->cset_links);
4766 up_read(&css_set_rwsem);
4771 * Make sure there's no live children. We can't test emptiness of
4772 * ->self.children as dead children linger on it while being
4773 * drained; otherwise, "rmdir parent/child parent" may fail.
4775 if (css_has_online_children(&cgrp->self))
4779 * Mark @cgrp dead. This prevents further task migration and child
4780 * creation by disabling cgroup_lock_live_group().
4782 cgrp->self.flags &= ~CSS_ONLINE;
4784 /* initiate massacre of all css's */
4785 for_each_css(css, ssid, cgrp)
4788 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4789 raw_spin_lock(&release_list_lock);
4790 if (!list_empty(&cgrp->release_list))
4791 list_del_init(&cgrp->release_list);
4792 raw_spin_unlock(&release_list_lock);
4795 * Remove @cgrp directory along with the base files. @cgrp has an
4796 * extra ref on its kn.
4798 kernfs_remove(cgrp->kn);
4800 set_bit(CGRP_RELEASABLE, &cgroup_parent(cgrp)->flags);
4801 check_for_release(cgroup_parent(cgrp));
4803 /* put the base reference */
4804 percpu_ref_kill(&cgrp->self.refcnt);
4809 static int cgroup_rmdir(struct kernfs_node *kn)
4811 struct cgroup *cgrp;
4814 cgrp = cgroup_kn_lock_live(kn);
4817 cgroup_get(cgrp); /* for @kn->priv clearing */
4819 ret = cgroup_destroy_locked(cgrp);
4821 cgroup_kn_unlock(kn);
4824 * There are two control paths which try to determine cgroup from
4825 * dentry without going through kernfs - cgroupstats_build() and
4826 * css_tryget_online_from_dir(). Those are supported by RCU
4827 * protecting clearing of cgrp->kn->priv backpointer, which should
4828 * happen after all files under it have been removed.
4831 RCU_INIT_POINTER(*(void __rcu __force **)&kn->priv, NULL);
4837 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4838 .remount_fs = cgroup_remount,
4839 .show_options = cgroup_show_options,
4840 .mkdir = cgroup_mkdir,
4841 .rmdir = cgroup_rmdir,
4842 .rename = cgroup_rename,
4845 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4847 struct cgroup_subsys_state *css;
4849 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4851 mutex_lock(&cgroup_mutex);
4853 idr_init(&ss->css_idr);
4854 INIT_LIST_HEAD(&ss->cfts);
4856 /* Create the root cgroup state for this subsystem */
4857 ss->root = &cgrp_dfl_root;
4858 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4859 /* We don't handle early failures gracefully */
4860 BUG_ON(IS_ERR(css));
4861 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4864 * Root csses are never destroyed and we can't initialize
4865 * percpu_ref during early init. Disable refcnting.
4867 css->flags |= CSS_NO_REF;
4870 /* allocation can't be done safely during early init */
4873 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4874 BUG_ON(css->id < 0);
4877 /* Update the init_css_set to contain a subsys
4878 * pointer to this state - since the subsystem is
4879 * newly registered, all tasks and hence the
4880 * init_css_set is in the subsystem's root cgroup. */
4881 init_css_set.subsys[ss->id] = css;
4883 need_forkexit_callback |= ss->fork || ss->exit;
4885 /* At system boot, before all subsystems have been
4886 * registered, no tasks have been forked, so we don't
4887 * need to invoke fork callbacks here. */
4888 BUG_ON(!list_empty(&init_task.tasks));
4890 BUG_ON(online_css(css));
4892 mutex_unlock(&cgroup_mutex);
4896 * cgroup_init_early - cgroup initialization at system boot
4898 * Initialize cgroups at system boot, and initialize any
4899 * subsystems that request early init.
4901 int __init cgroup_init_early(void)
4903 static struct cgroup_sb_opts __initdata opts;
4904 struct cgroup_subsys *ss;
4907 init_cgroup_root(&cgrp_dfl_root, &opts);
4908 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4910 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4912 for_each_subsys(ss, i) {
4913 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4914 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4915 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4917 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4918 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4921 ss->name = cgroup_subsys_name[i];
4924 cgroup_init_subsys(ss, true);
4930 * cgroup_init - cgroup initialization
4932 * Register cgroup filesystem and /proc file, and initialize
4933 * any subsystems that didn't request early init.
4935 int __init cgroup_init(void)
4937 struct cgroup_subsys *ss;
4941 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
4942 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
4944 mutex_lock(&cgroup_mutex);
4946 /* Add init_css_set to the hash table */
4947 key = css_set_hash(init_css_set.subsys);
4948 hash_add(css_set_table, &init_css_set.hlist, key);
4950 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
4952 mutex_unlock(&cgroup_mutex);
4954 for_each_subsys(ss, ssid) {
4955 if (ss->early_init) {
4956 struct cgroup_subsys_state *css =
4957 init_css_set.subsys[ss->id];
4959 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
4961 BUG_ON(css->id < 0);
4963 cgroup_init_subsys(ss, false);
4966 list_add_tail(&init_css_set.e_cset_node[ssid],
4967 &cgrp_dfl_root.cgrp.e_csets[ssid]);
4970 * Setting dfl_root subsys_mask needs to consider the
4971 * disabled flag and cftype registration needs kmalloc,
4972 * both of which aren't available during early_init.
4977 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
4979 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
4980 ss->dfl_cftypes = ss->legacy_cftypes;
4982 if (!ss->dfl_cftypes)
4983 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
4985 if (ss->dfl_cftypes == ss->legacy_cftypes) {
4986 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
4988 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
4989 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
4993 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4997 err = register_filesystem(&cgroup_fs_type);
4999 kobject_put(cgroup_kobj);
5003 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5007 static int __init cgroup_wq_init(void)
5010 * There isn't much point in executing destruction path in
5011 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5012 * Use 1 for @max_active.
5014 * We would prefer to do this in cgroup_init() above, but that
5015 * is called before init_workqueues(): so leave this until after.
5017 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5018 BUG_ON(!cgroup_destroy_wq);
5021 * Used to destroy pidlists and separate to serve as flush domain.
5022 * Cap @max_active to 1 too.
5024 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5026 BUG_ON(!cgroup_pidlist_destroy_wq);
5030 core_initcall(cgroup_wq_init);
5033 * proc_cgroup_show()
5034 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5035 * - Used for /proc/<pid>/cgroup.
5038 /* TODO: Use a proper seq_file iterator */
5039 int proc_cgroup_show(struct seq_file *m, void *v)
5042 struct task_struct *tsk;
5045 struct cgroup_root *root;
5048 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5054 tsk = get_pid_task(pid, PIDTYPE_PID);
5060 mutex_lock(&cgroup_mutex);
5061 down_read(&css_set_rwsem);
5063 for_each_root(root) {
5064 struct cgroup_subsys *ss;
5065 struct cgroup *cgrp;
5066 int ssid, count = 0;
5068 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5071 seq_printf(m, "%d:", root->hierarchy_id);
5072 for_each_subsys(ss, ssid)
5073 if (root->subsys_mask & (1 << ssid))
5074 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5075 if (strlen(root->name))
5076 seq_printf(m, "%sname=%s", count ? "," : "",
5079 cgrp = task_cgroup_from_root(tsk, root);
5080 path = cgroup_path(cgrp, buf, PATH_MAX);
5082 retval = -ENAMETOOLONG;
5090 up_read(&css_set_rwsem);
5091 mutex_unlock(&cgroup_mutex);
5092 put_task_struct(tsk);
5099 /* Display information about each subsystem and each hierarchy */
5100 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5102 struct cgroup_subsys *ss;
5105 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5107 * ideally we don't want subsystems moving around while we do this.
5108 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5109 * subsys/hierarchy state.
5111 mutex_lock(&cgroup_mutex);
5113 for_each_subsys(ss, i)
5114 seq_printf(m, "%s\t%d\t%d\t%d\n",
5115 ss->name, ss->root->hierarchy_id,
5116 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5118 mutex_unlock(&cgroup_mutex);
5122 static int cgroupstats_open(struct inode *inode, struct file *file)
5124 return single_open(file, proc_cgroupstats_show, NULL);
5127 static const struct file_operations proc_cgroupstats_operations = {
5128 .open = cgroupstats_open,
5130 .llseek = seq_lseek,
5131 .release = single_release,
5135 * cgroup_fork - initialize cgroup related fields during copy_process()
5136 * @child: pointer to task_struct of forking parent process.
5138 * A task is associated with the init_css_set until cgroup_post_fork()
5139 * attaches it to the parent's css_set. Empty cg_list indicates that
5140 * @child isn't holding reference to its css_set.
5142 void cgroup_fork(struct task_struct *child)
5144 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5145 INIT_LIST_HEAD(&child->cg_list);
5149 * cgroup_post_fork - called on a new task after adding it to the task list
5150 * @child: the task in question
5152 * Adds the task to the list running through its css_set if necessary and
5153 * call the subsystem fork() callbacks. Has to be after the task is
5154 * visible on the task list in case we race with the first call to
5155 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5158 void cgroup_post_fork(struct task_struct *child)
5160 struct cgroup_subsys *ss;
5164 * This may race against cgroup_enable_task_cg_links(). As that
5165 * function sets use_task_css_set_links before grabbing
5166 * tasklist_lock and we just went through tasklist_lock to add
5167 * @child, it's guaranteed that either we see the set
5168 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5169 * @child during its iteration.
5171 * If we won the race, @child is associated with %current's
5172 * css_set. Grabbing css_set_rwsem guarantees both that the
5173 * association is stable, and, on completion of the parent's
5174 * migration, @child is visible in the source of migration or
5175 * already in the destination cgroup. This guarantee is necessary
5176 * when implementing operations which need to migrate all tasks of
5177 * a cgroup to another.
5179 * Note that if we lose to cgroup_enable_task_cg_links(), @child
5180 * will remain in init_css_set. This is safe because all tasks are
5181 * in the init_css_set before cg_links is enabled and there's no
5182 * operation which transfers all tasks out of init_css_set.
5184 if (use_task_css_set_links) {
5185 struct css_set *cset;
5187 down_write(&css_set_rwsem);
5188 cset = task_css_set(current);
5189 if (list_empty(&child->cg_list)) {
5190 rcu_assign_pointer(child->cgroups, cset);
5191 list_add(&child->cg_list, &cset->tasks);
5194 up_write(&css_set_rwsem);
5198 * Call ss->fork(). This must happen after @child is linked on
5199 * css_set; otherwise, @child might change state between ->fork()
5200 * and addition to css_set.
5202 if (need_forkexit_callback) {
5203 for_each_subsys(ss, i)
5210 * cgroup_exit - detach cgroup from exiting task
5211 * @tsk: pointer to task_struct of exiting process
5213 * Description: Detach cgroup from @tsk and release it.
5215 * Note that cgroups marked notify_on_release force every task in
5216 * them to take the global cgroup_mutex mutex when exiting.
5217 * This could impact scaling on very large systems. Be reluctant to
5218 * use notify_on_release cgroups where very high task exit scaling
5219 * is required on large systems.
5221 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5222 * call cgroup_exit() while the task is still competent to handle
5223 * notify_on_release(), then leave the task attached to the root cgroup in
5224 * each hierarchy for the remainder of its exit. No need to bother with
5225 * init_css_set refcnting. init_css_set never goes away and we can't race
5226 * with migration path - PF_EXITING is visible to migration path.
5228 void cgroup_exit(struct task_struct *tsk)
5230 struct cgroup_subsys *ss;
5231 struct css_set *cset;
5232 bool put_cset = false;
5236 * Unlink from @tsk from its css_set. As migration path can't race
5237 * with us, we can check cg_list without grabbing css_set_rwsem.
5239 if (!list_empty(&tsk->cg_list)) {
5240 down_write(&css_set_rwsem);
5241 list_del_init(&tsk->cg_list);
5242 up_write(&css_set_rwsem);
5246 /* Reassign the task to the init_css_set. */
5247 cset = task_css_set(tsk);
5248 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5250 if (need_forkexit_callback) {
5251 /* see cgroup_post_fork() for details */
5252 for_each_subsys(ss, i) {
5254 struct cgroup_subsys_state *old_css = cset->subsys[i];
5255 struct cgroup_subsys_state *css = task_css(tsk, i);
5257 ss->exit(css, old_css, tsk);
5263 put_css_set(cset, true);
5266 static void check_for_release(struct cgroup *cgrp)
5268 if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
5269 !css_has_online_children(&cgrp->self)) {
5271 * Control Group is currently removeable. If it's not
5272 * already queued for a userspace notification, queue
5275 int need_schedule_work = 0;
5277 raw_spin_lock(&release_list_lock);
5278 if (!cgroup_is_dead(cgrp) &&
5279 list_empty(&cgrp->release_list)) {
5280 list_add(&cgrp->release_list, &release_list);
5281 need_schedule_work = 1;
5283 raw_spin_unlock(&release_list_lock);
5284 if (need_schedule_work)
5285 schedule_work(&release_agent_work);
5290 * Notify userspace when a cgroup is released, by running the
5291 * configured release agent with the name of the cgroup (path
5292 * relative to the root of cgroup file system) as the argument.
5294 * Most likely, this user command will try to rmdir this cgroup.
5296 * This races with the possibility that some other task will be
5297 * attached to this cgroup before it is removed, or that some other
5298 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5299 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5300 * unused, and this cgroup will be reprieved from its death sentence,
5301 * to continue to serve a useful existence. Next time it's released,
5302 * we will get notified again, if it still has 'notify_on_release' set.
5304 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5305 * means only wait until the task is successfully execve()'d. The
5306 * separate release agent task is forked by call_usermodehelper(),
5307 * then control in this thread returns here, without waiting for the
5308 * release agent task. We don't bother to wait because the caller of
5309 * this routine has no use for the exit status of the release agent
5310 * task, so no sense holding our caller up for that.
5312 static void cgroup_release_agent(struct work_struct *work)
5314 BUG_ON(work != &release_agent_work);
5315 mutex_lock(&cgroup_mutex);
5316 raw_spin_lock(&release_list_lock);
5317 while (!list_empty(&release_list)) {
5318 char *argv[3], *envp[3];
5320 char *pathbuf = NULL, *agentbuf = NULL, *path;
5321 struct cgroup *cgrp = list_entry(release_list.next,
5324 list_del_init(&cgrp->release_list);
5325 raw_spin_unlock(&release_list_lock);
5326 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5329 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5332 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5337 argv[i++] = agentbuf;
5342 /* minimal command environment */
5343 envp[i++] = "HOME=/";
5344 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5347 /* Drop the lock while we invoke the usermode helper,
5348 * since the exec could involve hitting disk and hence
5349 * be a slow process */
5350 mutex_unlock(&cgroup_mutex);
5351 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5352 mutex_lock(&cgroup_mutex);
5356 raw_spin_lock(&release_list_lock);
5358 raw_spin_unlock(&release_list_lock);
5359 mutex_unlock(&cgroup_mutex);
5362 static int __init cgroup_disable(char *str)
5364 struct cgroup_subsys *ss;
5368 while ((token = strsep(&str, ",")) != NULL) {
5372 for_each_subsys(ss, i) {
5373 if (!strcmp(token, ss->name)) {
5375 printk(KERN_INFO "Disabling %s control group"
5376 " subsystem\n", ss->name);
5383 __setup("cgroup_disable=", cgroup_disable);
5385 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5387 printk("cgroup: using legacy files on the default hierarchy\n");
5388 cgroup_legacy_files_on_dfl = true;
5391 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5394 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5395 * @dentry: directory dentry of interest
5396 * @ss: subsystem of interest
5398 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5399 * to get the corresponding css and return it. If such css doesn't exist
5400 * or can't be pinned, an ERR_PTR value is returned.
5402 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5403 struct cgroup_subsys *ss)
5405 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5406 struct cgroup_subsys_state *css = NULL;
5407 struct cgroup *cgrp;
5409 /* is @dentry a cgroup dir? */
5410 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5411 kernfs_type(kn) != KERNFS_DIR)
5412 return ERR_PTR(-EBADF);
5417 * This path doesn't originate from kernfs and @kn could already
5418 * have been or be removed at any point. @kn->priv is RCU
5419 * protected for this access. See cgroup_rmdir() for details.
5421 cgrp = rcu_dereference(kn->priv);
5423 css = cgroup_css(cgrp, ss);
5425 if (!css || !css_tryget_online(css))
5426 css = ERR_PTR(-ENOENT);
5433 * css_from_id - lookup css by id
5434 * @id: the cgroup id
5435 * @ss: cgroup subsys to be looked into
5437 * Returns the css if there's valid one with @id, otherwise returns NULL.
5438 * Should be called under rcu_read_lock().
5440 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5442 WARN_ON_ONCE(!rcu_read_lock_held());
5443 return idr_find(&ss->css_idr, id);
5446 #ifdef CONFIG_CGROUP_DEBUG
5447 static struct cgroup_subsys_state *
5448 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5450 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5453 return ERR_PTR(-ENOMEM);
5458 static void debug_css_free(struct cgroup_subsys_state *css)
5463 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5466 return cgroup_task_count(css->cgroup);
5469 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5472 return (u64)(unsigned long)current->cgroups;
5475 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5481 count = atomic_read(&task_css_set(current)->refcount);
5486 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5488 struct cgrp_cset_link *link;
5489 struct css_set *cset;
5492 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5496 down_read(&css_set_rwsem);
5498 cset = rcu_dereference(current->cgroups);
5499 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5500 struct cgroup *c = link->cgrp;
5502 cgroup_name(c, name_buf, NAME_MAX + 1);
5503 seq_printf(seq, "Root %d group %s\n",
5504 c->root->hierarchy_id, name_buf);
5507 up_read(&css_set_rwsem);
5512 #define MAX_TASKS_SHOWN_PER_CSS 25
5513 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5515 struct cgroup_subsys_state *css = seq_css(seq);
5516 struct cgrp_cset_link *link;
5518 down_read(&css_set_rwsem);
5519 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5520 struct css_set *cset = link->cset;
5521 struct task_struct *task;
5524 seq_printf(seq, "css_set %p\n", cset);
5526 list_for_each_entry(task, &cset->tasks, cg_list) {
5527 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5529 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5532 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5533 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5535 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5539 seq_puts(seq, " ...\n");
5541 up_read(&css_set_rwsem);
5545 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5547 return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5550 static struct cftype debug_files[] = {
5552 .name = "taskcount",
5553 .read_u64 = debug_taskcount_read,
5557 .name = "current_css_set",
5558 .read_u64 = current_css_set_read,
5562 .name = "current_css_set_refcount",
5563 .read_u64 = current_css_set_refcount_read,
5567 .name = "current_css_set_cg_links",
5568 .seq_show = current_css_set_cg_links_read,
5572 .name = "cgroup_css_links",
5573 .seq_show = cgroup_css_links_read,
5577 .name = "releasable",
5578 .read_u64 = releasable_read,
5584 struct cgroup_subsys debug_cgrp_subsys = {
5585 .css_alloc = debug_css_alloc,
5586 .css_free = debug_css_free,
5587 .legacy_cftypes = debug_files,
5589 #endif /* CONFIG_CGROUP_DEBUG */