2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DECLARE_RWSEM(css_set_rwsem);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_rwsem);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DECLARE_RWSEM(css_set_rwsem);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 rcu_lockdep_assert(rcu_read_lock_held() || \
111 lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct *cgroup_destroy_wq;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
143 * The default hierarchy, reserved for the subsystems that are otherwise
144 * unattached - it never has more than a single cgroup, and all tasks are
145 * part of that cgroup.
147 struct cgroup_root cgrp_dfl_root;
150 * The default hierarchy always exists but is hidden until mounted for the
151 * first time. This is for backward compatibility.
153 static bool cgrp_dfl_root_visible;
156 * Set by the boot param of the same name and makes subsystems with NULL
157 * ->dfl_files to use ->legacy_files on the default hierarchy.
159 static bool cgroup_legacy_files_on_dfl;
161 /* some controllers are not supported in the default hierarchy */
162 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
164 /* The list of hierarchy roots */
166 static LIST_HEAD(cgroup_roots);
167 static int cgroup_root_count;
169 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
170 static DEFINE_IDR(cgroup_hierarchy_idr);
173 * Assign a monotonically increasing serial number to csses. It guarantees
174 * cgroups with bigger numbers are newer than those with smaller numbers.
175 * Also, as csses are always appended to the parent's ->children list, it
176 * guarantees that sibling csses are always sorted in the ascending serial
177 * number order on the list. Protected by cgroup_mutex.
179 static u64 css_serial_nr_next = 1;
182 * These bitmask flags indicate whether tasks in the fork and exit paths have
183 * fork/exit handlers to call. This avoids us having to do extra work in the
184 * fork/exit path to check which subsystems have fork/exit callbacks.
186 static unsigned long have_fork_callback __read_mostly;
187 static unsigned long have_exit_callback __read_mostly;
189 static struct cftype cgroup_dfl_base_files[];
190 static struct cftype cgroup_legacy_base_files[];
192 static int rebind_subsystems(struct cgroup_root *dst_root,
193 unsigned long ss_mask);
194 static int cgroup_destroy_locked(struct cgroup *cgrp);
195 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
197 static void css_release(struct percpu_ref *ref);
198 static void kill_css(struct cgroup_subsys_state *css);
199 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
202 /* IDR wrappers which synchronize using cgroup_idr_lock */
203 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
208 idr_preload(gfp_mask);
209 spin_lock_bh(&cgroup_idr_lock);
210 ret = idr_alloc(idr, ptr, start, end, gfp_mask);
211 spin_unlock_bh(&cgroup_idr_lock);
216 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
220 spin_lock_bh(&cgroup_idr_lock);
221 ret = idr_replace(idr, ptr, id);
222 spin_unlock_bh(&cgroup_idr_lock);
226 static void cgroup_idr_remove(struct idr *idr, int id)
228 spin_lock_bh(&cgroup_idr_lock);
230 spin_unlock_bh(&cgroup_idr_lock);
233 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
235 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
238 return container_of(parent_css, struct cgroup, self);
243 * cgroup_css - obtain a cgroup's css for the specified subsystem
244 * @cgrp: the cgroup of interest
245 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
247 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
248 * function must be called either under cgroup_mutex or rcu_read_lock() and
249 * the caller is responsible for pinning the returned css if it wants to
250 * keep accessing it outside the said locks. This function may return
251 * %NULL if @cgrp doesn't have @subsys_id enabled.
253 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
254 struct cgroup_subsys *ss)
257 return rcu_dereference_check(cgrp->subsys[ss->id],
258 lockdep_is_held(&cgroup_mutex));
264 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
265 * @cgrp: the cgroup of interest
266 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
268 * Similar to cgroup_css() but returns the effective css, which is defined
269 * as the matching css of the nearest ancestor including self which has @ss
270 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
271 * function is guaranteed to return non-NULL css.
273 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
274 struct cgroup_subsys *ss)
276 lockdep_assert_held(&cgroup_mutex);
281 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
285 * This function is used while updating css associations and thus
286 * can't test the csses directly. Use ->child_subsys_mask.
288 while (cgroup_parent(cgrp) &&
289 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
290 cgrp = cgroup_parent(cgrp);
292 return cgroup_css(cgrp, ss);
296 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
297 * @cgrp: the cgroup of interest
298 * @ss: the subsystem of interest
300 * Find and get the effective css of @cgrp for @ss. The effective css is
301 * defined as the matching css of the nearest ancestor including self which
302 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
303 * the root css is returned, so this function always returns a valid css.
304 * The returned css must be put using css_put().
306 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
307 struct cgroup_subsys *ss)
309 struct cgroup_subsys_state *css;
314 css = cgroup_css(cgrp, ss);
316 if (css && css_tryget_online(css))
318 cgrp = cgroup_parent(cgrp);
321 css = init_css_set.subsys[ss->id];
328 /* convenient tests for these bits */
329 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
331 return !(cgrp->self.flags & CSS_ONLINE);
334 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
336 struct cgroup *cgrp = of->kn->parent->priv;
337 struct cftype *cft = of_cft(of);
340 * This is open and unprotected implementation of cgroup_css().
341 * seq_css() is only called from a kernfs file operation which has
342 * an active reference on the file. Because all the subsystem
343 * files are drained before a css is disassociated with a cgroup,
344 * the matching css from the cgroup's subsys table is guaranteed to
345 * be and stay valid until the enclosing operation is complete.
348 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
352 EXPORT_SYMBOL_GPL(of_css);
355 * cgroup_is_descendant - test ancestry
356 * @cgrp: the cgroup to be tested
357 * @ancestor: possible ancestor of @cgrp
359 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
360 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
361 * and @ancestor are accessible.
363 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
366 if (cgrp == ancestor)
368 cgrp = cgroup_parent(cgrp);
373 static int notify_on_release(const struct cgroup *cgrp)
375 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
379 * for_each_css - iterate all css's of a cgroup
380 * @css: the iteration cursor
381 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
382 * @cgrp: the target cgroup to iterate css's of
384 * Should be called under cgroup_[tree_]mutex.
386 #define for_each_css(css, ssid, cgrp) \
387 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
388 if (!((css) = rcu_dereference_check( \
389 (cgrp)->subsys[(ssid)], \
390 lockdep_is_held(&cgroup_mutex)))) { } \
394 * for_each_e_css - iterate all effective css's of a cgroup
395 * @css: the iteration cursor
396 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
397 * @cgrp: the target cgroup to iterate css's of
399 * Should be called under cgroup_[tree_]mutex.
401 #define for_each_e_css(css, ssid, cgrp) \
402 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
403 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
408 * for_each_subsys - iterate all enabled cgroup subsystems
409 * @ss: the iteration cursor
410 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
412 #define for_each_subsys(ss, ssid) \
413 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
414 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
417 * for_each_subsys_which - filter for_each_subsys with a bitmask
418 * @ss: the iteration cursor
419 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
420 * @ss_maskp: a pointer to the bitmask
422 * The block will only run for cases where the ssid-th bit (1 << ssid) of
425 #define for_each_subsys_which(ss, ssid, ss_maskp) \
426 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
429 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
430 if (((ss) = cgroup_subsys[ssid]) && false) \
434 /* iterate across the hierarchies */
435 #define for_each_root(root) \
436 list_for_each_entry((root), &cgroup_roots, root_list)
438 /* iterate over child cgrps, lock should be held throughout iteration */
439 #define cgroup_for_each_live_child(child, cgrp) \
440 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
441 if (({ lockdep_assert_held(&cgroup_mutex); \
442 cgroup_is_dead(child); })) \
446 static void cgroup_release_agent(struct work_struct *work);
447 static void check_for_release(struct cgroup *cgrp);
450 * A cgroup can be associated with multiple css_sets as different tasks may
451 * belong to different cgroups on different hierarchies. In the other
452 * direction, a css_set is naturally associated with multiple cgroups.
453 * This M:N relationship is represented by the following link structure
454 * which exists for each association and allows traversing the associations
457 struct cgrp_cset_link {
458 /* the cgroup and css_set this link associates */
460 struct css_set *cset;
462 /* list of cgrp_cset_links anchored at cgrp->cset_links */
463 struct list_head cset_link;
465 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
466 struct list_head cgrp_link;
470 * The default css_set - used by init and its children prior to any
471 * hierarchies being mounted. It contains a pointer to the root state
472 * for each subsystem. Also used to anchor the list of css_sets. Not
473 * reference-counted, to improve performance when child cgroups
474 * haven't been created.
476 struct css_set init_css_set = {
477 .refcount = ATOMIC_INIT(1),
478 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
479 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
480 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
481 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
482 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
485 static int css_set_count = 1; /* 1 for init_css_set */
488 * cgroup_update_populated - updated populated count of a cgroup
489 * @cgrp: the target cgroup
490 * @populated: inc or dec populated count
492 * @cgrp is either getting the first task (css_set) or losing the last.
493 * Update @cgrp->populated_cnt accordingly. The count is propagated
494 * towards root so that a given cgroup's populated_cnt is zero iff the
495 * cgroup and all its descendants are empty.
497 * @cgrp's interface file "cgroup.populated" is zero if
498 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
499 * changes from or to zero, userland is notified that the content of the
500 * interface file has changed. This can be used to detect when @cgrp and
501 * its descendants become populated or empty.
503 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
505 lockdep_assert_held(&css_set_rwsem);
511 trigger = !cgrp->populated_cnt++;
513 trigger = !--cgrp->populated_cnt;
518 if (cgrp->populated_kn)
519 kernfs_notify(cgrp->populated_kn);
520 cgrp = cgroup_parent(cgrp);
525 * hash table for cgroup groups. This improves the performance to find
526 * an existing css_set. This hash doesn't (currently) take into
527 * account cgroups in empty hierarchies.
529 #define CSS_SET_HASH_BITS 7
530 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
532 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
534 unsigned long key = 0UL;
535 struct cgroup_subsys *ss;
538 for_each_subsys(ss, i)
539 key += (unsigned long)css[i];
540 key = (key >> 16) ^ key;
545 static void put_css_set_locked(struct css_set *cset)
547 struct cgrp_cset_link *link, *tmp_link;
548 struct cgroup_subsys *ss;
551 lockdep_assert_held(&css_set_rwsem);
553 if (!atomic_dec_and_test(&cset->refcount))
556 /* This css_set is dead. unlink it and release cgroup refcounts */
557 for_each_subsys(ss, ssid)
558 list_del(&cset->e_cset_node[ssid]);
559 hash_del(&cset->hlist);
562 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
563 struct cgroup *cgrp = link->cgrp;
565 list_del(&link->cset_link);
566 list_del(&link->cgrp_link);
568 /* @cgrp can't go away while we're holding css_set_rwsem */
569 if (list_empty(&cgrp->cset_links)) {
570 cgroup_update_populated(cgrp, false);
571 check_for_release(cgrp);
577 kfree_rcu(cset, rcu_head);
580 static void put_css_set(struct css_set *cset)
583 * Ensure that the refcount doesn't hit zero while any readers
584 * can see it. Similar to atomic_dec_and_lock(), but for an
587 if (atomic_add_unless(&cset->refcount, -1, 1))
590 down_write(&css_set_rwsem);
591 put_css_set_locked(cset);
592 up_write(&css_set_rwsem);
596 * refcounted get/put for css_set objects
598 static inline void get_css_set(struct css_set *cset)
600 atomic_inc(&cset->refcount);
604 * compare_css_sets - helper function for find_existing_css_set().
605 * @cset: candidate css_set being tested
606 * @old_cset: existing css_set for a task
607 * @new_cgrp: cgroup that's being entered by the task
608 * @template: desired set of css pointers in css_set (pre-calculated)
610 * Returns true if "cset" matches "old_cset" except for the hierarchy
611 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
613 static bool compare_css_sets(struct css_set *cset,
614 struct css_set *old_cset,
615 struct cgroup *new_cgrp,
616 struct cgroup_subsys_state *template[])
618 struct list_head *l1, *l2;
621 * On the default hierarchy, there can be csets which are
622 * associated with the same set of cgroups but different csses.
623 * Let's first ensure that csses match.
625 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
629 * Compare cgroup pointers in order to distinguish between
630 * different cgroups in hierarchies. As different cgroups may
631 * share the same effective css, this comparison is always
634 l1 = &cset->cgrp_links;
635 l2 = &old_cset->cgrp_links;
637 struct cgrp_cset_link *link1, *link2;
638 struct cgroup *cgrp1, *cgrp2;
642 /* See if we reached the end - both lists are equal length. */
643 if (l1 == &cset->cgrp_links) {
644 BUG_ON(l2 != &old_cset->cgrp_links);
647 BUG_ON(l2 == &old_cset->cgrp_links);
649 /* Locate the cgroups associated with these links. */
650 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
651 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
654 /* Hierarchies should be linked in the same order. */
655 BUG_ON(cgrp1->root != cgrp2->root);
658 * If this hierarchy is the hierarchy of the cgroup
659 * that's changing, then we need to check that this
660 * css_set points to the new cgroup; if it's any other
661 * hierarchy, then this css_set should point to the
662 * same cgroup as the old css_set.
664 if (cgrp1->root == new_cgrp->root) {
665 if (cgrp1 != new_cgrp)
676 * find_existing_css_set - init css array and find the matching css_set
677 * @old_cset: the css_set that we're using before the cgroup transition
678 * @cgrp: the cgroup that we're moving into
679 * @template: out param for the new set of csses, should be clear on entry
681 static struct css_set *find_existing_css_set(struct css_set *old_cset,
683 struct cgroup_subsys_state *template[])
685 struct cgroup_root *root = cgrp->root;
686 struct cgroup_subsys *ss;
687 struct css_set *cset;
692 * Build the set of subsystem state objects that we want to see in the
693 * new css_set. while subsystems can change globally, the entries here
694 * won't change, so no need for locking.
696 for_each_subsys(ss, i) {
697 if (root->subsys_mask & (1UL << i)) {
699 * @ss is in this hierarchy, so we want the
700 * effective css from @cgrp.
702 template[i] = cgroup_e_css(cgrp, ss);
705 * @ss is not in this hierarchy, so we don't want
708 template[i] = old_cset->subsys[i];
712 key = css_set_hash(template);
713 hash_for_each_possible(css_set_table, cset, hlist, key) {
714 if (!compare_css_sets(cset, old_cset, cgrp, template))
717 /* This css_set matches what we need */
721 /* No existing cgroup group matched */
725 static void free_cgrp_cset_links(struct list_head *links_to_free)
727 struct cgrp_cset_link *link, *tmp_link;
729 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
730 list_del(&link->cset_link);
736 * allocate_cgrp_cset_links - allocate cgrp_cset_links
737 * @count: the number of links to allocate
738 * @tmp_links: list_head the allocated links are put on
740 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
741 * through ->cset_link. Returns 0 on success or -errno.
743 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
745 struct cgrp_cset_link *link;
748 INIT_LIST_HEAD(tmp_links);
750 for (i = 0; i < count; i++) {
751 link = kzalloc(sizeof(*link), GFP_KERNEL);
753 free_cgrp_cset_links(tmp_links);
756 list_add(&link->cset_link, tmp_links);
762 * link_css_set - a helper function to link a css_set to a cgroup
763 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
764 * @cset: the css_set to be linked
765 * @cgrp: the destination cgroup
767 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
770 struct cgrp_cset_link *link;
772 BUG_ON(list_empty(tmp_links));
774 if (cgroup_on_dfl(cgrp))
775 cset->dfl_cgrp = cgrp;
777 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
781 if (list_empty(&cgrp->cset_links))
782 cgroup_update_populated(cgrp, true);
783 list_move(&link->cset_link, &cgrp->cset_links);
786 * Always add links to the tail of the list so that the list
787 * is sorted by order of hierarchy creation
789 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
793 * find_css_set - return a new css_set with one cgroup updated
794 * @old_cset: the baseline css_set
795 * @cgrp: the cgroup to be updated
797 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
798 * substituted into the appropriate hierarchy.
800 static struct css_set *find_css_set(struct css_set *old_cset,
803 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
804 struct css_set *cset;
805 struct list_head tmp_links;
806 struct cgrp_cset_link *link;
807 struct cgroup_subsys *ss;
811 lockdep_assert_held(&cgroup_mutex);
813 /* First see if we already have a cgroup group that matches
815 down_read(&css_set_rwsem);
816 cset = find_existing_css_set(old_cset, cgrp, template);
819 up_read(&css_set_rwsem);
824 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
828 /* Allocate all the cgrp_cset_link objects that we'll need */
829 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
834 atomic_set(&cset->refcount, 1);
835 INIT_LIST_HEAD(&cset->cgrp_links);
836 INIT_LIST_HEAD(&cset->tasks);
837 INIT_LIST_HEAD(&cset->mg_tasks);
838 INIT_LIST_HEAD(&cset->mg_preload_node);
839 INIT_LIST_HEAD(&cset->mg_node);
840 INIT_HLIST_NODE(&cset->hlist);
842 /* Copy the set of subsystem state objects generated in
843 * find_existing_css_set() */
844 memcpy(cset->subsys, template, sizeof(cset->subsys));
846 down_write(&css_set_rwsem);
847 /* Add reference counts and links from the new css_set. */
848 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
849 struct cgroup *c = link->cgrp;
851 if (c->root == cgrp->root)
853 link_css_set(&tmp_links, cset, c);
856 BUG_ON(!list_empty(&tmp_links));
860 /* Add @cset to the hash table */
861 key = css_set_hash(cset->subsys);
862 hash_add(css_set_table, &cset->hlist, key);
864 for_each_subsys(ss, ssid)
865 list_add_tail(&cset->e_cset_node[ssid],
866 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
868 up_write(&css_set_rwsem);
873 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
875 struct cgroup *root_cgrp = kf_root->kn->priv;
877 return root_cgrp->root;
880 static int cgroup_init_root_id(struct cgroup_root *root)
884 lockdep_assert_held(&cgroup_mutex);
886 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
890 root->hierarchy_id = id;
894 static void cgroup_exit_root_id(struct cgroup_root *root)
896 lockdep_assert_held(&cgroup_mutex);
898 if (root->hierarchy_id) {
899 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
900 root->hierarchy_id = 0;
904 static void cgroup_free_root(struct cgroup_root *root)
907 /* hierarchy ID should already have been released */
908 WARN_ON_ONCE(root->hierarchy_id);
910 idr_destroy(&root->cgroup_idr);
915 static void cgroup_destroy_root(struct cgroup_root *root)
917 struct cgroup *cgrp = &root->cgrp;
918 struct cgrp_cset_link *link, *tmp_link;
920 mutex_lock(&cgroup_mutex);
922 BUG_ON(atomic_read(&root->nr_cgrps));
923 BUG_ON(!list_empty(&cgrp->self.children));
925 /* Rebind all subsystems back to the default hierarchy */
926 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
929 * Release all the links from cset_links to this hierarchy's
932 down_write(&css_set_rwsem);
934 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
935 list_del(&link->cset_link);
936 list_del(&link->cgrp_link);
939 up_write(&css_set_rwsem);
941 if (!list_empty(&root->root_list)) {
942 list_del(&root->root_list);
946 cgroup_exit_root_id(root);
948 mutex_unlock(&cgroup_mutex);
950 kernfs_destroy_root(root->kf_root);
951 cgroup_free_root(root);
954 /* look up cgroup associated with given css_set on the specified hierarchy */
955 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
956 struct cgroup_root *root)
958 struct cgroup *res = NULL;
960 lockdep_assert_held(&cgroup_mutex);
961 lockdep_assert_held(&css_set_rwsem);
963 if (cset == &init_css_set) {
966 struct cgrp_cset_link *link;
968 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
969 struct cgroup *c = link->cgrp;
971 if (c->root == root) {
983 * Return the cgroup for "task" from the given hierarchy. Must be
984 * called with cgroup_mutex and css_set_rwsem held.
986 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
987 struct cgroup_root *root)
990 * No need to lock the task - since we hold cgroup_mutex the
991 * task can't change groups, so the only thing that can happen
992 * is that it exits and its css is set back to init_css_set.
994 return cset_cgroup_from_root(task_css_set(task), root);
998 * A task must hold cgroup_mutex to modify cgroups.
1000 * Any task can increment and decrement the count field without lock.
1001 * So in general, code holding cgroup_mutex can't rely on the count
1002 * field not changing. However, if the count goes to zero, then only
1003 * cgroup_attach_task() can increment it again. Because a count of zero
1004 * means that no tasks are currently attached, therefore there is no
1005 * way a task attached to that cgroup can fork (the other way to
1006 * increment the count). So code holding cgroup_mutex can safely
1007 * assume that if the count is zero, it will stay zero. Similarly, if
1008 * a task holds cgroup_mutex on a cgroup with zero count, it
1009 * knows that the cgroup won't be removed, as cgroup_rmdir()
1012 * A cgroup can only be deleted if both its 'count' of using tasks
1013 * is zero, and its list of 'children' cgroups is empty. Since all
1014 * tasks in the system use _some_ cgroup, and since there is always at
1015 * least one task in the system (init, pid == 1), therefore, root cgroup
1016 * always has either children cgroups and/or using tasks. So we don't
1017 * need a special hack to ensure that root cgroup cannot be deleted.
1019 * P.S. One more locking exception. RCU is used to guard the
1020 * update of a tasks cgroup pointer by cgroup_attach_task()
1023 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
1024 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1025 static const struct file_operations proc_cgroupstats_operations;
1027 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1030 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1031 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1032 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1033 cft->ss->name, cft->name);
1035 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1040 * cgroup_file_mode - deduce file mode of a control file
1041 * @cft: the control file in question
1043 * returns cft->mode if ->mode is not 0
1044 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1045 * returns S_IRUGO if it has only a read handler
1046 * returns S_IWUSR if it has only a write hander
1048 static umode_t cgroup_file_mode(const struct cftype *cft)
1055 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1058 if (cft->write_u64 || cft->write_s64 || cft->write)
1064 static void cgroup_get(struct cgroup *cgrp)
1066 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1067 css_get(&cgrp->self);
1070 static bool cgroup_tryget(struct cgroup *cgrp)
1072 return css_tryget(&cgrp->self);
1075 static void cgroup_put(struct cgroup *cgrp)
1077 css_put(&cgrp->self);
1081 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1082 * @cgrp: the target cgroup
1083 * @subtree_control: the new subtree_control mask to consider
1085 * On the default hierarchy, a subsystem may request other subsystems to be
1086 * enabled together through its ->depends_on mask. In such cases, more
1087 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1089 * This function calculates which subsystems need to be enabled if
1090 * @subtree_control is to be applied to @cgrp. The returned mask is always
1091 * a superset of @subtree_control and follows the usual hierarchy rules.
1093 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1094 unsigned long subtree_control)
1096 struct cgroup *parent = cgroup_parent(cgrp);
1097 unsigned long cur_ss_mask = subtree_control;
1098 struct cgroup_subsys *ss;
1101 lockdep_assert_held(&cgroup_mutex);
1103 if (!cgroup_on_dfl(cgrp))
1107 unsigned long new_ss_mask = cur_ss_mask;
1109 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1110 new_ss_mask |= ss->depends_on;
1113 * Mask out subsystems which aren't available. This can
1114 * happen only if some depended-upon subsystems were bound
1115 * to non-default hierarchies.
1118 new_ss_mask &= parent->child_subsys_mask;
1120 new_ss_mask &= cgrp->root->subsys_mask;
1122 if (new_ss_mask == cur_ss_mask)
1124 cur_ss_mask = new_ss_mask;
1131 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1132 * @cgrp: the target cgroup
1134 * Update @cgrp->child_subsys_mask according to the current
1135 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1137 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1139 cgrp->child_subsys_mask =
1140 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1144 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1145 * @kn: the kernfs_node being serviced
1147 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1148 * the method finishes if locking succeeded. Note that once this function
1149 * returns the cgroup returned by cgroup_kn_lock_live() may become
1150 * inaccessible any time. If the caller intends to continue to access the
1151 * cgroup, it should pin it before invoking this function.
1153 static void cgroup_kn_unlock(struct kernfs_node *kn)
1155 struct cgroup *cgrp;
1157 if (kernfs_type(kn) == KERNFS_DIR)
1160 cgrp = kn->parent->priv;
1162 mutex_unlock(&cgroup_mutex);
1164 kernfs_unbreak_active_protection(kn);
1169 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1170 * @kn: the kernfs_node being serviced
1172 * This helper is to be used by a cgroup kernfs method currently servicing
1173 * @kn. It breaks the active protection, performs cgroup locking and
1174 * verifies that the associated cgroup is alive. Returns the cgroup if
1175 * alive; otherwise, %NULL. A successful return should be undone by a
1176 * matching cgroup_kn_unlock() invocation.
1178 * Any cgroup kernfs method implementation which requires locking the
1179 * associated cgroup should use this helper. It avoids nesting cgroup
1180 * locking under kernfs active protection and allows all kernfs operations
1181 * including self-removal.
1183 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1185 struct cgroup *cgrp;
1187 if (kernfs_type(kn) == KERNFS_DIR)
1190 cgrp = kn->parent->priv;
1193 * We're gonna grab cgroup_mutex which nests outside kernfs
1194 * active_ref. cgroup liveliness check alone provides enough
1195 * protection against removal. Ensure @cgrp stays accessible and
1196 * break the active_ref protection.
1198 if (!cgroup_tryget(cgrp))
1200 kernfs_break_active_protection(kn);
1202 mutex_lock(&cgroup_mutex);
1204 if (!cgroup_is_dead(cgrp))
1207 cgroup_kn_unlock(kn);
1211 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1213 char name[CGROUP_FILE_NAME_MAX];
1215 lockdep_assert_held(&cgroup_mutex);
1216 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1220 * cgroup_clear_dir - remove subsys files in a cgroup directory
1221 * @cgrp: target cgroup
1222 * @subsys_mask: mask of the subsystem ids whose files should be removed
1224 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1226 struct cgroup_subsys *ss;
1229 for_each_subsys(ss, i) {
1230 struct cftype *cfts;
1232 if (!(subsys_mask & (1 << i)))
1234 list_for_each_entry(cfts, &ss->cfts, node)
1235 cgroup_addrm_files(cgrp, cfts, false);
1239 static int rebind_subsystems(struct cgroup_root *dst_root,
1240 unsigned long ss_mask)
1242 struct cgroup_subsys *ss;
1243 unsigned long tmp_ss_mask;
1246 lockdep_assert_held(&cgroup_mutex);
1248 for_each_subsys_which(ss, ssid, &ss_mask) {
1249 /* if @ss has non-root csses attached to it, can't move */
1250 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1253 /* can't move between two non-dummy roots either */
1254 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1258 /* skip creating root files on dfl_root for inhibited subsystems */
1259 tmp_ss_mask = ss_mask;
1260 if (dst_root == &cgrp_dfl_root)
1261 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1263 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1265 if (dst_root != &cgrp_dfl_root)
1269 * Rebinding back to the default root is not allowed to
1270 * fail. Using both default and non-default roots should
1271 * be rare. Moving subsystems back and forth even more so.
1272 * Just warn about it and continue.
1274 if (cgrp_dfl_root_visible) {
1275 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1277 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1282 * Nothing can fail from this point on. Remove files for the
1283 * removed subsystems and rebind each subsystem.
1285 for_each_subsys_which(ss, ssid, &ss_mask)
1286 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1288 for_each_subsys_which(ss, ssid, &ss_mask) {
1289 struct cgroup_root *src_root;
1290 struct cgroup_subsys_state *css;
1291 struct css_set *cset;
1293 src_root = ss->root;
1294 css = cgroup_css(&src_root->cgrp, ss);
1296 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1298 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1299 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1300 ss->root = dst_root;
1301 css->cgroup = &dst_root->cgrp;
1303 down_write(&css_set_rwsem);
1304 hash_for_each(css_set_table, i, cset, hlist)
1305 list_move_tail(&cset->e_cset_node[ss->id],
1306 &dst_root->cgrp.e_csets[ss->id]);
1307 up_write(&css_set_rwsem);
1309 src_root->subsys_mask &= ~(1 << ssid);
1310 src_root->cgrp.subtree_control &= ~(1 << ssid);
1311 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1313 /* default hierarchy doesn't enable controllers by default */
1314 dst_root->subsys_mask |= 1 << ssid;
1315 if (dst_root != &cgrp_dfl_root) {
1316 dst_root->cgrp.subtree_control |= 1 << ssid;
1317 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1324 kernfs_activate(dst_root->cgrp.kn);
1328 static int cgroup_show_options(struct seq_file *seq,
1329 struct kernfs_root *kf_root)
1331 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1332 struct cgroup_subsys *ss;
1335 for_each_subsys(ss, ssid)
1336 if (root->subsys_mask & (1 << ssid))
1337 seq_printf(seq, ",%s", ss->name);
1338 if (root->flags & CGRP_ROOT_NOPREFIX)
1339 seq_puts(seq, ",noprefix");
1340 if (root->flags & CGRP_ROOT_XATTR)
1341 seq_puts(seq, ",xattr");
1343 spin_lock(&release_agent_path_lock);
1344 if (strlen(root->release_agent_path))
1345 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1346 spin_unlock(&release_agent_path_lock);
1348 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1349 seq_puts(seq, ",clone_children");
1350 if (strlen(root->name))
1351 seq_printf(seq, ",name=%s", root->name);
1355 struct cgroup_sb_opts {
1356 unsigned long subsys_mask;
1358 char *release_agent;
1359 bool cpuset_clone_children;
1361 /* User explicitly requested empty subsystem */
1365 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1367 char *token, *o = data;
1368 bool all_ss = false, one_ss = false;
1369 unsigned long mask = -1UL;
1370 struct cgroup_subsys *ss;
1374 #ifdef CONFIG_CPUSETS
1375 mask = ~(1U << cpuset_cgrp_id);
1378 memset(opts, 0, sizeof(*opts));
1380 while ((token = strsep(&o, ",")) != NULL) {
1385 if (!strcmp(token, "none")) {
1386 /* Explicitly have no subsystems */
1390 if (!strcmp(token, "all")) {
1391 /* Mutually exclusive option 'all' + subsystem name */
1397 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1398 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1401 if (!strcmp(token, "noprefix")) {
1402 opts->flags |= CGRP_ROOT_NOPREFIX;
1405 if (!strcmp(token, "clone_children")) {
1406 opts->cpuset_clone_children = true;
1409 if (!strcmp(token, "xattr")) {
1410 opts->flags |= CGRP_ROOT_XATTR;
1413 if (!strncmp(token, "release_agent=", 14)) {
1414 /* Specifying two release agents is forbidden */
1415 if (opts->release_agent)
1417 opts->release_agent =
1418 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1419 if (!opts->release_agent)
1423 if (!strncmp(token, "name=", 5)) {
1424 const char *name = token + 5;
1425 /* Can't specify an empty name */
1428 /* Must match [\w.-]+ */
1429 for (i = 0; i < strlen(name); i++) {
1433 if ((c == '.') || (c == '-') || (c == '_'))
1437 /* Specifying two names is forbidden */
1440 opts->name = kstrndup(name,
1441 MAX_CGROUP_ROOT_NAMELEN - 1,
1449 for_each_subsys(ss, i) {
1450 if (strcmp(token, ss->name))
1455 /* Mutually exclusive option 'all' + subsystem name */
1458 opts->subsys_mask |= (1 << i);
1463 if (i == CGROUP_SUBSYS_COUNT)
1467 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1468 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1470 pr_err("sane_behavior: no other mount options allowed\n");
1477 * If the 'all' option was specified select all the subsystems,
1478 * otherwise if 'none', 'name=' and a subsystem name options were
1479 * not specified, let's default to 'all'
1481 if (all_ss || (!one_ss && !opts->none && !opts->name))
1482 for_each_subsys(ss, i)
1484 opts->subsys_mask |= (1 << i);
1487 * We either have to specify by name or by subsystems. (So all
1488 * empty hierarchies must have a name).
1490 if (!opts->subsys_mask && !opts->name)
1494 * Option noprefix was introduced just for backward compatibility
1495 * with the old cpuset, so we allow noprefix only if mounting just
1496 * the cpuset subsystem.
1498 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1501 /* Can't specify "none" and some subsystems */
1502 if (opts->subsys_mask && opts->none)
1508 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1511 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1512 struct cgroup_sb_opts opts;
1513 unsigned long added_mask, removed_mask;
1515 if (root == &cgrp_dfl_root) {
1516 pr_err("remount is not allowed\n");
1520 mutex_lock(&cgroup_mutex);
1522 /* See what subsystems are wanted */
1523 ret = parse_cgroupfs_options(data, &opts);
1527 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1528 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1529 task_tgid_nr(current), current->comm);
1531 added_mask = opts.subsys_mask & ~root->subsys_mask;
1532 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1534 /* Don't allow flags or name to change at remount */
1535 if ((opts.flags ^ root->flags) ||
1536 (opts.name && strcmp(opts.name, root->name))) {
1537 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1538 opts.flags, opts.name ?: "", root->flags, root->name);
1543 /* remounting is not allowed for populated hierarchies */
1544 if (!list_empty(&root->cgrp.self.children)) {
1549 ret = rebind_subsystems(root, added_mask);
1553 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1555 if (opts.release_agent) {
1556 spin_lock(&release_agent_path_lock);
1557 strcpy(root->release_agent_path, opts.release_agent);
1558 spin_unlock(&release_agent_path_lock);
1561 kfree(opts.release_agent);
1563 mutex_unlock(&cgroup_mutex);
1568 * To reduce the fork() overhead for systems that are not actually using
1569 * their cgroups capability, we don't maintain the lists running through
1570 * each css_set to its tasks until we see the list actually used - in other
1571 * words after the first mount.
1573 static bool use_task_css_set_links __read_mostly;
1575 static void cgroup_enable_task_cg_lists(void)
1577 struct task_struct *p, *g;
1579 down_write(&css_set_rwsem);
1581 if (use_task_css_set_links)
1584 use_task_css_set_links = true;
1587 * We need tasklist_lock because RCU is not safe against
1588 * while_each_thread(). Besides, a forking task that has passed
1589 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1590 * is not guaranteed to have its child immediately visible in the
1591 * tasklist if we walk through it with RCU.
1593 read_lock(&tasklist_lock);
1594 do_each_thread(g, p) {
1595 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1596 task_css_set(p) != &init_css_set);
1599 * We should check if the process is exiting, otherwise
1600 * it will race with cgroup_exit() in that the list
1601 * entry won't be deleted though the process has exited.
1602 * Do it while holding siglock so that we don't end up
1603 * racing against cgroup_exit().
1605 spin_lock_irq(&p->sighand->siglock);
1606 if (!(p->flags & PF_EXITING)) {
1607 struct css_set *cset = task_css_set(p);
1609 list_add(&p->cg_list, &cset->tasks);
1612 spin_unlock_irq(&p->sighand->siglock);
1613 } while_each_thread(g, p);
1614 read_unlock(&tasklist_lock);
1616 up_write(&css_set_rwsem);
1619 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1621 struct cgroup_subsys *ss;
1624 INIT_LIST_HEAD(&cgrp->self.sibling);
1625 INIT_LIST_HEAD(&cgrp->self.children);
1626 INIT_LIST_HEAD(&cgrp->cset_links);
1627 INIT_LIST_HEAD(&cgrp->pidlists);
1628 mutex_init(&cgrp->pidlist_mutex);
1629 cgrp->self.cgroup = cgrp;
1630 cgrp->self.flags |= CSS_ONLINE;
1632 for_each_subsys(ss, ssid)
1633 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1635 init_waitqueue_head(&cgrp->offline_waitq);
1636 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1639 static void init_cgroup_root(struct cgroup_root *root,
1640 struct cgroup_sb_opts *opts)
1642 struct cgroup *cgrp = &root->cgrp;
1644 INIT_LIST_HEAD(&root->root_list);
1645 atomic_set(&root->nr_cgrps, 1);
1647 init_cgroup_housekeeping(cgrp);
1648 idr_init(&root->cgroup_idr);
1650 root->flags = opts->flags;
1651 if (opts->release_agent)
1652 strcpy(root->release_agent_path, opts->release_agent);
1654 strcpy(root->name, opts->name);
1655 if (opts->cpuset_clone_children)
1656 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1659 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1661 LIST_HEAD(tmp_links);
1662 struct cgroup *root_cgrp = &root->cgrp;
1663 struct cftype *base_files;
1664 struct css_set *cset;
1667 lockdep_assert_held(&cgroup_mutex);
1669 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1672 root_cgrp->id = ret;
1674 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1680 * We're accessing css_set_count without locking css_set_rwsem here,
1681 * but that's OK - it can only be increased by someone holding
1682 * cgroup_lock, and that's us. The worst that can happen is that we
1683 * have some link structures left over
1685 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1689 ret = cgroup_init_root_id(root);
1693 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1694 KERNFS_ROOT_CREATE_DEACTIVATED,
1696 if (IS_ERR(root->kf_root)) {
1697 ret = PTR_ERR(root->kf_root);
1700 root_cgrp->kn = root->kf_root->kn;
1702 if (root == &cgrp_dfl_root)
1703 base_files = cgroup_dfl_base_files;
1705 base_files = cgroup_legacy_base_files;
1707 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1711 ret = rebind_subsystems(root, ss_mask);
1716 * There must be no failure case after here, since rebinding takes
1717 * care of subsystems' refcounts, which are explicitly dropped in
1718 * the failure exit path.
1720 list_add(&root->root_list, &cgroup_roots);
1721 cgroup_root_count++;
1724 * Link the root cgroup in this hierarchy into all the css_set
1727 down_write(&css_set_rwsem);
1728 hash_for_each(css_set_table, i, cset, hlist)
1729 link_css_set(&tmp_links, cset, root_cgrp);
1730 up_write(&css_set_rwsem);
1732 BUG_ON(!list_empty(&root_cgrp->self.children));
1733 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1735 kernfs_activate(root_cgrp->kn);
1740 kernfs_destroy_root(root->kf_root);
1741 root->kf_root = NULL;
1743 cgroup_exit_root_id(root);
1745 percpu_ref_exit(&root_cgrp->self.refcnt);
1747 free_cgrp_cset_links(&tmp_links);
1751 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1752 int flags, const char *unused_dev_name,
1755 struct super_block *pinned_sb = NULL;
1756 struct cgroup_subsys *ss;
1757 struct cgroup_root *root;
1758 struct cgroup_sb_opts opts;
1759 struct dentry *dentry;
1765 * The first time anyone tries to mount a cgroup, enable the list
1766 * linking each css_set to its tasks and fix up all existing tasks.
1768 if (!use_task_css_set_links)
1769 cgroup_enable_task_cg_lists();
1771 mutex_lock(&cgroup_mutex);
1773 /* First find the desired set of subsystems */
1774 ret = parse_cgroupfs_options(data, &opts);
1778 /* look for a matching existing root */
1779 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1780 cgrp_dfl_root_visible = true;
1781 root = &cgrp_dfl_root;
1782 cgroup_get(&root->cgrp);
1788 * Destruction of cgroup root is asynchronous, so subsystems may
1789 * still be dying after the previous unmount. Let's drain the
1790 * dying subsystems. We just need to ensure that the ones
1791 * unmounted previously finish dying and don't care about new ones
1792 * starting. Testing ref liveliness is good enough.
1794 for_each_subsys(ss, i) {
1795 if (!(opts.subsys_mask & (1 << i)) ||
1796 ss->root == &cgrp_dfl_root)
1799 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1800 mutex_unlock(&cgroup_mutex);
1802 ret = restart_syscall();
1805 cgroup_put(&ss->root->cgrp);
1808 for_each_root(root) {
1809 bool name_match = false;
1811 if (root == &cgrp_dfl_root)
1815 * If we asked for a name then it must match. Also, if
1816 * name matches but sybsys_mask doesn't, we should fail.
1817 * Remember whether name matched.
1820 if (strcmp(opts.name, root->name))
1826 * If we asked for subsystems (or explicitly for no
1827 * subsystems) then they must match.
1829 if ((opts.subsys_mask || opts.none) &&
1830 (opts.subsys_mask != root->subsys_mask)) {
1837 if (root->flags ^ opts.flags)
1838 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1841 * We want to reuse @root whose lifetime is governed by its
1842 * ->cgrp. Let's check whether @root is alive and keep it
1843 * that way. As cgroup_kill_sb() can happen anytime, we
1844 * want to block it by pinning the sb so that @root doesn't
1845 * get killed before mount is complete.
1847 * With the sb pinned, tryget_live can reliably indicate
1848 * whether @root can be reused. If it's being killed,
1849 * drain it. We can use wait_queue for the wait but this
1850 * path is super cold. Let's just sleep a bit and retry.
1852 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1853 if (IS_ERR(pinned_sb) ||
1854 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1855 mutex_unlock(&cgroup_mutex);
1856 if (!IS_ERR_OR_NULL(pinned_sb))
1857 deactivate_super(pinned_sb);
1859 ret = restart_syscall();
1868 * No such thing, create a new one. name= matching without subsys
1869 * specification is allowed for already existing hierarchies but we
1870 * can't create new one without subsys specification.
1872 if (!opts.subsys_mask && !opts.none) {
1877 root = kzalloc(sizeof(*root), GFP_KERNEL);
1883 init_cgroup_root(root, &opts);
1885 ret = cgroup_setup_root(root, opts.subsys_mask);
1887 cgroup_free_root(root);
1890 mutex_unlock(&cgroup_mutex);
1892 kfree(opts.release_agent);
1896 return ERR_PTR(ret);
1898 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1899 CGROUP_SUPER_MAGIC, &new_sb);
1900 if (IS_ERR(dentry) || !new_sb)
1901 cgroup_put(&root->cgrp);
1904 * If @pinned_sb, we're reusing an existing root and holding an
1905 * extra ref on its sb. Mount is complete. Put the extra ref.
1909 deactivate_super(pinned_sb);
1915 static void cgroup_kill_sb(struct super_block *sb)
1917 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1918 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1921 * If @root doesn't have any mounts or children, start killing it.
1922 * This prevents new mounts by disabling percpu_ref_tryget_live().
1923 * cgroup_mount() may wait for @root's release.
1925 * And don't kill the default root.
1927 if (!list_empty(&root->cgrp.self.children) ||
1928 root == &cgrp_dfl_root)
1929 cgroup_put(&root->cgrp);
1931 percpu_ref_kill(&root->cgrp.self.refcnt);
1936 static struct file_system_type cgroup_fs_type = {
1938 .mount = cgroup_mount,
1939 .kill_sb = cgroup_kill_sb,
1942 static struct kobject *cgroup_kobj;
1945 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1946 * @task: target task
1947 * @buf: the buffer to write the path into
1948 * @buflen: the length of the buffer
1950 * Determine @task's cgroup on the first (the one with the lowest non-zero
1951 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1952 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1953 * cgroup controller callbacks.
1955 * Return value is the same as kernfs_path().
1957 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1959 struct cgroup_root *root;
1960 struct cgroup *cgrp;
1961 int hierarchy_id = 1;
1964 mutex_lock(&cgroup_mutex);
1965 down_read(&css_set_rwsem);
1967 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1970 cgrp = task_cgroup_from_root(task, root);
1971 path = cgroup_path(cgrp, buf, buflen);
1973 /* if no hierarchy exists, everyone is in "/" */
1974 if (strlcpy(buf, "/", buflen) < buflen)
1978 up_read(&css_set_rwsem);
1979 mutex_unlock(&cgroup_mutex);
1982 EXPORT_SYMBOL_GPL(task_cgroup_path);
1984 /* used to track tasks and other necessary states during migration */
1985 struct cgroup_taskset {
1986 /* the src and dst cset list running through cset->mg_node */
1987 struct list_head src_csets;
1988 struct list_head dst_csets;
1991 * Fields for cgroup_taskset_*() iteration.
1993 * Before migration is committed, the target migration tasks are on
1994 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1995 * the csets on ->dst_csets. ->csets point to either ->src_csets
1996 * or ->dst_csets depending on whether migration is committed.
1998 * ->cur_csets and ->cur_task point to the current task position
2001 struct list_head *csets;
2002 struct css_set *cur_cset;
2003 struct task_struct *cur_task;
2007 * cgroup_taskset_first - reset taskset and return the first task
2008 * @tset: taskset of interest
2010 * @tset iteration is initialized and the first task is returned.
2012 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2014 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2015 tset->cur_task = NULL;
2017 return cgroup_taskset_next(tset);
2021 * cgroup_taskset_next - iterate to the next task in taskset
2022 * @tset: taskset of interest
2024 * Return the next task in @tset. Iteration must have been initialized
2025 * with cgroup_taskset_first().
2027 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2029 struct css_set *cset = tset->cur_cset;
2030 struct task_struct *task = tset->cur_task;
2032 while (&cset->mg_node != tset->csets) {
2034 task = list_first_entry(&cset->mg_tasks,
2035 struct task_struct, cg_list);
2037 task = list_next_entry(task, cg_list);
2039 if (&task->cg_list != &cset->mg_tasks) {
2040 tset->cur_cset = cset;
2041 tset->cur_task = task;
2045 cset = list_next_entry(cset, mg_node);
2053 * cgroup_task_migrate - move a task from one cgroup to another.
2054 * @old_cgrp: the cgroup @tsk is being migrated from
2055 * @tsk: the task being migrated
2056 * @new_cset: the new css_set @tsk is being attached to
2058 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2060 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2061 struct task_struct *tsk,
2062 struct css_set *new_cset)
2064 struct css_set *old_cset;
2066 lockdep_assert_held(&cgroup_mutex);
2067 lockdep_assert_held(&css_set_rwsem);
2070 * We are synchronized through cgroup_threadgroup_rwsem against
2071 * PF_EXITING setting such that we can't race against cgroup_exit()
2072 * changing the css_set to init_css_set and dropping the old one.
2074 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2075 old_cset = task_css_set(tsk);
2077 get_css_set(new_cset);
2078 rcu_assign_pointer(tsk->cgroups, new_cset);
2081 * Use move_tail so that cgroup_taskset_first() still returns the
2082 * leader after migration. This works because cgroup_migrate()
2083 * ensures that the dst_cset of the leader is the first on the
2084 * tset's dst_csets list.
2086 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2089 * We just gained a reference on old_cset by taking it from the
2090 * task. As trading it for new_cset is protected by cgroup_mutex,
2091 * we're safe to drop it here; it will be freed under RCU.
2093 put_css_set_locked(old_cset);
2097 * cgroup_migrate_finish - cleanup after attach
2098 * @preloaded_csets: list of preloaded css_sets
2100 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2101 * those functions for details.
2103 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2105 struct css_set *cset, *tmp_cset;
2107 lockdep_assert_held(&cgroup_mutex);
2109 down_write(&css_set_rwsem);
2110 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2111 cset->mg_src_cgrp = NULL;
2112 cset->mg_dst_cset = NULL;
2113 list_del_init(&cset->mg_preload_node);
2114 put_css_set_locked(cset);
2116 up_write(&css_set_rwsem);
2120 * cgroup_migrate_add_src - add a migration source css_set
2121 * @src_cset: the source css_set to add
2122 * @dst_cgrp: the destination cgroup
2123 * @preloaded_csets: list of preloaded css_sets
2125 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2126 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2127 * up by cgroup_migrate_finish().
2129 * This function may be called without holding cgroup_threadgroup_rwsem
2130 * even if the target is a process. Threads may be created and destroyed
2131 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2132 * into play and the preloaded css_sets are guaranteed to cover all
2135 static void cgroup_migrate_add_src(struct css_set *src_cset,
2136 struct cgroup *dst_cgrp,
2137 struct list_head *preloaded_csets)
2139 struct cgroup *src_cgrp;
2141 lockdep_assert_held(&cgroup_mutex);
2142 lockdep_assert_held(&css_set_rwsem);
2144 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2146 if (!list_empty(&src_cset->mg_preload_node))
2149 WARN_ON(src_cset->mg_src_cgrp);
2150 WARN_ON(!list_empty(&src_cset->mg_tasks));
2151 WARN_ON(!list_empty(&src_cset->mg_node));
2153 src_cset->mg_src_cgrp = src_cgrp;
2154 get_css_set(src_cset);
2155 list_add(&src_cset->mg_preload_node, preloaded_csets);
2159 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2160 * @dst_cgrp: the destination cgroup (may be %NULL)
2161 * @preloaded_csets: list of preloaded source css_sets
2163 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2164 * have been preloaded to @preloaded_csets. This function looks up and
2165 * pins all destination css_sets, links each to its source, and append them
2166 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2167 * source css_set is assumed to be its cgroup on the default hierarchy.
2169 * This function must be called after cgroup_migrate_add_src() has been
2170 * called on each migration source css_set. After migration is performed
2171 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2174 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2175 struct list_head *preloaded_csets)
2178 struct css_set *src_cset, *tmp_cset;
2180 lockdep_assert_held(&cgroup_mutex);
2183 * Except for the root, child_subsys_mask must be zero for a cgroup
2184 * with tasks so that child cgroups don't compete against tasks.
2186 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2187 dst_cgrp->child_subsys_mask)
2190 /* look up the dst cset for each src cset and link it to src */
2191 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2192 struct css_set *dst_cset;
2194 dst_cset = find_css_set(src_cset,
2195 dst_cgrp ?: src_cset->dfl_cgrp);
2199 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2202 * If src cset equals dst, it's noop. Drop the src.
2203 * cgroup_migrate() will skip the cset too. Note that we
2204 * can't handle src == dst as some nodes are used by both.
2206 if (src_cset == dst_cset) {
2207 src_cset->mg_src_cgrp = NULL;
2208 list_del_init(&src_cset->mg_preload_node);
2209 put_css_set(src_cset);
2210 put_css_set(dst_cset);
2214 src_cset->mg_dst_cset = dst_cset;
2216 if (list_empty(&dst_cset->mg_preload_node))
2217 list_add(&dst_cset->mg_preload_node, &csets);
2219 put_css_set(dst_cset);
2222 list_splice_tail(&csets, preloaded_csets);
2225 cgroup_migrate_finish(&csets);
2230 * cgroup_migrate - migrate a process or task to a cgroup
2231 * @cgrp: the destination cgroup
2232 * @leader: the leader of the process or the task to migrate
2233 * @threadgroup: whether @leader points to the whole process or a single task
2235 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2236 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2237 * caller is also responsible for invoking cgroup_migrate_add_src() and
2238 * cgroup_migrate_prepare_dst() on the targets before invoking this
2239 * function and following up with cgroup_migrate_finish().
2241 * As long as a controller's ->can_attach() doesn't fail, this function is
2242 * guaranteed to succeed. This means that, excluding ->can_attach()
2243 * failure, when migrating multiple targets, the success or failure can be
2244 * decided for all targets by invoking group_migrate_prepare_dst() before
2245 * actually starting migrating.
2247 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2250 struct cgroup_taskset tset = {
2251 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2252 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2253 .csets = &tset.src_csets,
2255 struct cgroup_subsys_state *css, *failed_css = NULL;
2256 struct css_set *cset, *tmp_cset;
2257 struct task_struct *task, *tmp_task;
2261 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2262 * already PF_EXITING could be freed from underneath us unless we
2263 * take an rcu_read_lock.
2265 down_write(&css_set_rwsem);
2269 /* @task either already exited or can't exit until the end */
2270 if (task->flags & PF_EXITING)
2273 /* leave @task alone if post_fork() hasn't linked it yet */
2274 if (list_empty(&task->cg_list))
2277 cset = task_css_set(task);
2278 if (!cset->mg_src_cgrp)
2282 * cgroup_taskset_first() must always return the leader.
2283 * Take care to avoid disturbing the ordering.
2285 list_move_tail(&task->cg_list, &cset->mg_tasks);
2286 if (list_empty(&cset->mg_node))
2287 list_add_tail(&cset->mg_node, &tset.src_csets);
2288 if (list_empty(&cset->mg_dst_cset->mg_node))
2289 list_move_tail(&cset->mg_dst_cset->mg_node,
2294 } while_each_thread(leader, task);
2296 up_write(&css_set_rwsem);
2298 /* methods shouldn't be called if no task is actually migrating */
2299 if (list_empty(&tset.src_csets))
2302 /* check that we can legitimately attach to the cgroup */
2303 for_each_e_css(css, i, cgrp) {
2304 if (css->ss->can_attach) {
2305 ret = css->ss->can_attach(css, &tset);
2308 goto out_cancel_attach;
2314 * Now that we're guaranteed success, proceed to move all tasks to
2315 * the new cgroup. There are no failure cases after here, so this
2316 * is the commit point.
2318 down_write(&css_set_rwsem);
2319 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2320 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2321 cgroup_task_migrate(cset->mg_src_cgrp, task,
2324 up_write(&css_set_rwsem);
2327 * Migration is committed, all target tasks are now on dst_csets.
2328 * Nothing is sensitive to fork() after this point. Notify
2329 * controllers that migration is complete.
2331 tset.csets = &tset.dst_csets;
2333 for_each_e_css(css, i, cgrp)
2334 if (css->ss->attach)
2335 css->ss->attach(css, &tset);
2338 goto out_release_tset;
2341 for_each_e_css(css, i, cgrp) {
2342 if (css == failed_css)
2344 if (css->ss->cancel_attach)
2345 css->ss->cancel_attach(css, &tset);
2348 down_write(&css_set_rwsem);
2349 list_splice_init(&tset.dst_csets, &tset.src_csets);
2350 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2351 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2352 list_del_init(&cset->mg_node);
2354 up_write(&css_set_rwsem);
2359 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2360 * @dst_cgrp: the cgroup to attach to
2361 * @leader: the task or the leader of the threadgroup to be attached
2362 * @threadgroup: attach the whole threadgroup?
2364 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2366 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2367 struct task_struct *leader, bool threadgroup)
2369 LIST_HEAD(preloaded_csets);
2370 struct task_struct *task;
2373 /* look up all src csets */
2374 down_read(&css_set_rwsem);
2378 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2382 } while_each_thread(leader, task);
2384 up_read(&css_set_rwsem);
2386 /* prepare dst csets and commit */
2387 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2389 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2391 cgroup_migrate_finish(&preloaded_csets);
2395 static int cgroup_procs_write_permission(struct task_struct *task,
2396 struct cgroup *dst_cgrp,
2397 struct kernfs_open_file *of)
2399 const struct cred *cred = current_cred();
2400 const struct cred *tcred = get_task_cred(task);
2404 * even if we're attaching all tasks in the thread group, we only
2405 * need to check permissions on one of them.
2407 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2408 !uid_eq(cred->euid, tcred->uid) &&
2409 !uid_eq(cred->euid, tcred->suid))
2412 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2413 struct super_block *sb = of->file->f_path.dentry->d_sb;
2414 struct cgroup *cgrp;
2415 struct inode *inode;
2417 down_read(&css_set_rwsem);
2418 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2419 up_read(&css_set_rwsem);
2421 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2422 cgrp = cgroup_parent(cgrp);
2425 inode = kernfs_get_inode(sb, cgrp->procs_kn);
2427 ret = inode_permission(inode, MAY_WRITE);
2437 * Find the task_struct of the task to attach by vpid and pass it along to the
2438 * function to attach either it or all tasks in its threadgroup. Will lock
2439 * cgroup_mutex and threadgroup.
2441 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2442 size_t nbytes, loff_t off, bool threadgroup)
2444 struct task_struct *tsk;
2445 struct cgroup *cgrp;
2449 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2452 cgrp = cgroup_kn_lock_live(of->kn);
2456 percpu_down_write(&cgroup_threadgroup_rwsem);
2459 tsk = find_task_by_vpid(pid);
2462 goto out_unlock_rcu;
2469 tsk = tsk->group_leader;
2472 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2473 * trapped in a cpuset, or RT worker may be born in a cgroup
2474 * with no rt_runtime allocated. Just say no.
2476 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2478 goto out_unlock_rcu;
2481 get_task_struct(tsk);
2484 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2486 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2488 put_task_struct(tsk);
2489 goto out_unlock_threadgroup;
2493 out_unlock_threadgroup:
2494 percpu_up_write(&cgroup_threadgroup_rwsem);
2495 cgroup_kn_unlock(of->kn);
2496 return ret ?: nbytes;
2500 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2501 * @from: attach to all cgroups of a given task
2502 * @tsk: the task to be attached
2504 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2506 struct cgroup_root *root;
2509 mutex_lock(&cgroup_mutex);
2510 for_each_root(root) {
2511 struct cgroup *from_cgrp;
2513 if (root == &cgrp_dfl_root)
2516 down_read(&css_set_rwsem);
2517 from_cgrp = task_cgroup_from_root(from, root);
2518 up_read(&css_set_rwsem);
2520 retval = cgroup_attach_task(from_cgrp, tsk, false);
2524 mutex_unlock(&cgroup_mutex);
2528 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2530 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2531 char *buf, size_t nbytes, loff_t off)
2533 return __cgroup_procs_write(of, buf, nbytes, off, false);
2536 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2537 char *buf, size_t nbytes, loff_t off)
2539 return __cgroup_procs_write(of, buf, nbytes, off, true);
2542 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2543 char *buf, size_t nbytes, loff_t off)
2545 struct cgroup *cgrp;
2547 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2549 cgrp = cgroup_kn_lock_live(of->kn);
2552 spin_lock(&release_agent_path_lock);
2553 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2554 sizeof(cgrp->root->release_agent_path));
2555 spin_unlock(&release_agent_path_lock);
2556 cgroup_kn_unlock(of->kn);
2560 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2562 struct cgroup *cgrp = seq_css(seq)->cgroup;
2564 spin_lock(&release_agent_path_lock);
2565 seq_puts(seq, cgrp->root->release_agent_path);
2566 spin_unlock(&release_agent_path_lock);
2567 seq_putc(seq, '\n');
2571 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2573 seq_puts(seq, "0\n");
2577 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2579 struct cgroup_subsys *ss;
2580 bool printed = false;
2583 for_each_subsys_which(ss, ssid, &ss_mask) {
2586 seq_printf(seq, "%s", ss->name);
2590 seq_putc(seq, '\n');
2593 /* show controllers which are currently attached to the default hierarchy */
2594 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2596 struct cgroup *cgrp = seq_css(seq)->cgroup;
2598 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2599 ~cgrp_dfl_root_inhibit_ss_mask);
2603 /* show controllers which are enabled from the parent */
2604 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2606 struct cgroup *cgrp = seq_css(seq)->cgroup;
2608 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2612 /* show controllers which are enabled for a given cgroup's children */
2613 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2615 struct cgroup *cgrp = seq_css(seq)->cgroup;
2617 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2622 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2623 * @cgrp: root of the subtree to update csses for
2625 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2626 * css associations need to be updated accordingly. This function looks up
2627 * all css_sets which are attached to the subtree, creates the matching
2628 * updated css_sets and migrates the tasks to the new ones.
2630 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2632 LIST_HEAD(preloaded_csets);
2633 struct cgroup_subsys_state *css;
2634 struct css_set *src_cset;
2637 lockdep_assert_held(&cgroup_mutex);
2639 percpu_down_write(&cgroup_threadgroup_rwsem);
2641 /* look up all csses currently attached to @cgrp's subtree */
2642 down_read(&css_set_rwsem);
2643 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2644 struct cgrp_cset_link *link;
2646 /* self is not affected by child_subsys_mask change */
2647 if (css->cgroup == cgrp)
2650 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2651 cgroup_migrate_add_src(link->cset, cgrp,
2654 up_read(&css_set_rwsem);
2656 /* NULL dst indicates self on default hierarchy */
2657 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2661 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2662 struct task_struct *last_task = NULL, *task;
2664 /* src_csets precede dst_csets, break on the first dst_cset */
2665 if (!src_cset->mg_src_cgrp)
2669 * All tasks in src_cset need to be migrated to the
2670 * matching dst_cset. Empty it process by process. We
2671 * walk tasks but migrate processes. The leader might even
2672 * belong to a different cset but such src_cset would also
2673 * be among the target src_csets because the default
2674 * hierarchy enforces per-process membership.
2677 down_read(&css_set_rwsem);
2678 task = list_first_entry_or_null(&src_cset->tasks,
2679 struct task_struct, cg_list);
2681 task = task->group_leader;
2682 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2683 get_task_struct(task);
2685 up_read(&css_set_rwsem);
2690 /* guard against possible infinite loop */
2691 if (WARN(last_task == task,
2692 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2696 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2698 put_task_struct(task);
2700 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2706 cgroup_migrate_finish(&preloaded_csets);
2707 percpu_up_write(&cgroup_threadgroup_rwsem);
2711 /* change the enabled child controllers for a cgroup in the default hierarchy */
2712 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2713 char *buf, size_t nbytes,
2716 unsigned long enable = 0, disable = 0;
2717 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2718 struct cgroup *cgrp, *child;
2719 struct cgroup_subsys *ss;
2724 * Parse input - space separated list of subsystem names prefixed
2725 * with either + or -.
2727 buf = strstrip(buf);
2728 while ((tok = strsep(&buf, " "))) {
2729 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2733 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2734 if (ss->disabled || strcmp(tok + 1, ss->name))
2738 enable |= 1 << ssid;
2739 disable &= ~(1 << ssid);
2740 } else if (*tok == '-') {
2741 disable |= 1 << ssid;
2742 enable &= ~(1 << ssid);
2748 if (ssid == CGROUP_SUBSYS_COUNT)
2752 cgrp = cgroup_kn_lock_live(of->kn);
2756 for_each_subsys(ss, ssid) {
2757 if (enable & (1 << ssid)) {
2758 if (cgrp->subtree_control & (1 << ssid)) {
2759 enable &= ~(1 << ssid);
2763 /* unavailable or not enabled on the parent? */
2764 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2765 (cgroup_parent(cgrp) &&
2766 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2770 } else if (disable & (1 << ssid)) {
2771 if (!(cgrp->subtree_control & (1 << ssid))) {
2772 disable &= ~(1 << ssid);
2776 /* a child has it enabled? */
2777 cgroup_for_each_live_child(child, cgrp) {
2778 if (child->subtree_control & (1 << ssid)) {
2786 if (!enable && !disable) {
2792 * Except for the root, subtree_control must be zero for a cgroup
2793 * with tasks so that child cgroups don't compete against tasks.
2795 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2801 * Update subsys masks and calculate what needs to be done. More
2802 * subsystems than specified may need to be enabled or disabled
2803 * depending on subsystem dependencies.
2805 old_sc = cgrp->subtree_control;
2806 old_ss = cgrp->child_subsys_mask;
2807 new_sc = (old_sc | enable) & ~disable;
2808 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2810 css_enable = ~old_ss & new_ss;
2811 css_disable = old_ss & ~new_ss;
2812 enable |= css_enable;
2813 disable |= css_disable;
2816 * Because css offlining is asynchronous, userland might try to
2817 * re-enable the same controller while the previous instance is
2818 * still around. In such cases, wait till it's gone using
2821 for_each_subsys_which(ss, ssid, &css_enable) {
2822 cgroup_for_each_live_child(child, cgrp) {
2825 if (!cgroup_css(child, ss))
2829 prepare_to_wait(&child->offline_waitq, &wait,
2830 TASK_UNINTERRUPTIBLE);
2831 cgroup_kn_unlock(of->kn);
2833 finish_wait(&child->offline_waitq, &wait);
2836 return restart_syscall();
2840 cgrp->subtree_control = new_sc;
2841 cgrp->child_subsys_mask = new_ss;
2844 * Create new csses or make the existing ones visible. A css is
2845 * created invisible if it's being implicitly enabled through
2846 * dependency. An invisible css is made visible when the userland
2847 * explicitly enables it.
2849 for_each_subsys(ss, ssid) {
2850 if (!(enable & (1 << ssid)))
2853 cgroup_for_each_live_child(child, cgrp) {
2854 if (css_enable & (1 << ssid))
2855 ret = create_css(child, ss,
2856 cgrp->subtree_control & (1 << ssid));
2858 ret = cgroup_populate_dir(child, 1 << ssid);
2865 * At this point, cgroup_e_css() results reflect the new csses
2866 * making the following cgroup_update_dfl_csses() properly update
2867 * css associations of all tasks in the subtree.
2869 ret = cgroup_update_dfl_csses(cgrp);
2874 * All tasks are migrated out of disabled csses. Kill or hide
2875 * them. A css is hidden when the userland requests it to be
2876 * disabled while other subsystems are still depending on it. The
2877 * css must not actively control resources and be in the vanilla
2878 * state if it's made visible again later. Controllers which may
2879 * be depended upon should provide ->css_reset() for this purpose.
2881 for_each_subsys(ss, ssid) {
2882 if (!(disable & (1 << ssid)))
2885 cgroup_for_each_live_child(child, cgrp) {
2886 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2888 if (css_disable & (1 << ssid)) {
2891 cgroup_clear_dir(child, 1 << ssid);
2899 * The effective csses of all the descendants (excluding @cgrp) may
2900 * have changed. Subsystems can optionally subscribe to this event
2901 * by implementing ->css_e_css_changed() which is invoked if any of
2902 * the effective csses seen from the css's cgroup may have changed.
2904 for_each_subsys(ss, ssid) {
2905 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2906 struct cgroup_subsys_state *css;
2908 if (!ss->css_e_css_changed || !this_css)
2911 css_for_each_descendant_pre(css, this_css)
2912 if (css != this_css)
2913 ss->css_e_css_changed(css);
2916 kernfs_activate(cgrp->kn);
2919 cgroup_kn_unlock(of->kn);
2920 return ret ?: nbytes;
2923 cgrp->subtree_control = old_sc;
2924 cgrp->child_subsys_mask = old_ss;
2926 for_each_subsys(ss, ssid) {
2927 if (!(enable & (1 << ssid)))
2930 cgroup_for_each_live_child(child, cgrp) {
2931 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2936 if (css_enable & (1 << ssid))
2939 cgroup_clear_dir(child, 1 << ssid);
2945 static int cgroup_populated_show(struct seq_file *seq, void *v)
2947 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2951 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2952 size_t nbytes, loff_t off)
2954 struct cgroup *cgrp = of->kn->parent->priv;
2955 struct cftype *cft = of->kn->priv;
2956 struct cgroup_subsys_state *css;
2960 return cft->write(of, buf, nbytes, off);
2963 * kernfs guarantees that a file isn't deleted with operations in
2964 * flight, which means that the matching css is and stays alive and
2965 * doesn't need to be pinned. The RCU locking is not necessary
2966 * either. It's just for the convenience of using cgroup_css().
2969 css = cgroup_css(cgrp, cft->ss);
2972 if (cft->write_u64) {
2973 unsigned long long v;
2974 ret = kstrtoull(buf, 0, &v);
2976 ret = cft->write_u64(css, cft, v);
2977 } else if (cft->write_s64) {
2979 ret = kstrtoll(buf, 0, &v);
2981 ret = cft->write_s64(css, cft, v);
2986 return ret ?: nbytes;
2989 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2991 return seq_cft(seq)->seq_start(seq, ppos);
2994 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2996 return seq_cft(seq)->seq_next(seq, v, ppos);
2999 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3001 seq_cft(seq)->seq_stop(seq, v);
3004 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3006 struct cftype *cft = seq_cft(m);
3007 struct cgroup_subsys_state *css = seq_css(m);
3010 return cft->seq_show(m, arg);
3013 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3014 else if (cft->read_s64)
3015 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3021 static struct kernfs_ops cgroup_kf_single_ops = {
3022 .atomic_write_len = PAGE_SIZE,
3023 .write = cgroup_file_write,
3024 .seq_show = cgroup_seqfile_show,
3027 static struct kernfs_ops cgroup_kf_ops = {
3028 .atomic_write_len = PAGE_SIZE,
3029 .write = cgroup_file_write,
3030 .seq_start = cgroup_seqfile_start,
3031 .seq_next = cgroup_seqfile_next,
3032 .seq_stop = cgroup_seqfile_stop,
3033 .seq_show = cgroup_seqfile_show,
3037 * cgroup_rename - Only allow simple rename of directories in place.
3039 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3040 const char *new_name_str)
3042 struct cgroup *cgrp = kn->priv;
3045 if (kernfs_type(kn) != KERNFS_DIR)
3047 if (kn->parent != new_parent)
3051 * This isn't a proper migration and its usefulness is very
3052 * limited. Disallow on the default hierarchy.
3054 if (cgroup_on_dfl(cgrp))
3058 * We're gonna grab cgroup_mutex which nests outside kernfs
3059 * active_ref. kernfs_rename() doesn't require active_ref
3060 * protection. Break them before grabbing cgroup_mutex.
3062 kernfs_break_active_protection(new_parent);
3063 kernfs_break_active_protection(kn);
3065 mutex_lock(&cgroup_mutex);
3067 ret = kernfs_rename(kn, new_parent, new_name_str);
3069 mutex_unlock(&cgroup_mutex);
3071 kernfs_unbreak_active_protection(kn);
3072 kernfs_unbreak_active_protection(new_parent);
3076 /* set uid and gid of cgroup dirs and files to that of the creator */
3077 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3079 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3080 .ia_uid = current_fsuid(),
3081 .ia_gid = current_fsgid(), };
3083 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3084 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3087 return kernfs_setattr(kn, &iattr);
3090 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3092 char name[CGROUP_FILE_NAME_MAX];
3093 struct kernfs_node *kn;
3094 struct lock_class_key *key = NULL;
3097 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3098 key = &cft->lockdep_key;
3100 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3101 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3106 ret = cgroup_kn_set_ugid(kn);
3112 if (cft->write == cgroup_procs_write)
3113 cgrp->procs_kn = kn;
3114 else if (cft->seq_show == cgroup_populated_show)
3115 cgrp->populated_kn = kn;
3120 * cgroup_addrm_files - add or remove files to a cgroup directory
3121 * @cgrp: the target cgroup
3122 * @cfts: array of cftypes to be added
3123 * @is_add: whether to add or remove
3125 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3126 * For removals, this function never fails. If addition fails, this
3127 * function doesn't remove files already added. The caller is responsible
3130 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3136 lockdep_assert_held(&cgroup_mutex);
3138 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3139 /* does cft->flags tell us to skip this file on @cgrp? */
3140 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3142 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3144 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3146 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3150 ret = cgroup_add_file(cgrp, cft);
3152 pr_warn("%s: failed to add %s, err=%d\n",
3153 __func__, cft->name, ret);
3157 cgroup_rm_file(cgrp, cft);
3163 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3166 struct cgroup_subsys *ss = cfts[0].ss;
3167 struct cgroup *root = &ss->root->cgrp;
3168 struct cgroup_subsys_state *css;
3171 lockdep_assert_held(&cgroup_mutex);
3173 /* add/rm files for all cgroups created before */
3174 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3175 struct cgroup *cgrp = css->cgroup;
3177 if (cgroup_is_dead(cgrp))
3180 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3186 kernfs_activate(root->kn);
3190 static void cgroup_exit_cftypes(struct cftype *cfts)
3194 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3195 /* free copy for custom atomic_write_len, see init_cftypes() */
3196 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3201 /* revert flags set by cgroup core while adding @cfts */
3202 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3206 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3210 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3211 struct kernfs_ops *kf_ops;
3213 WARN_ON(cft->ss || cft->kf_ops);
3216 kf_ops = &cgroup_kf_ops;
3218 kf_ops = &cgroup_kf_single_ops;
3221 * Ugh... if @cft wants a custom max_write_len, we need to
3222 * make a copy of kf_ops to set its atomic_write_len.
3224 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3225 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3227 cgroup_exit_cftypes(cfts);
3230 kf_ops->atomic_write_len = cft->max_write_len;
3233 cft->kf_ops = kf_ops;
3240 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3242 lockdep_assert_held(&cgroup_mutex);
3244 if (!cfts || !cfts[0].ss)
3247 list_del(&cfts->node);
3248 cgroup_apply_cftypes(cfts, false);
3249 cgroup_exit_cftypes(cfts);
3254 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3255 * @cfts: zero-length name terminated array of cftypes
3257 * Unregister @cfts. Files described by @cfts are removed from all
3258 * existing cgroups and all future cgroups won't have them either. This
3259 * function can be called anytime whether @cfts' subsys is attached or not.
3261 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3264 int cgroup_rm_cftypes(struct cftype *cfts)
3268 mutex_lock(&cgroup_mutex);
3269 ret = cgroup_rm_cftypes_locked(cfts);
3270 mutex_unlock(&cgroup_mutex);
3275 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3276 * @ss: target cgroup subsystem
3277 * @cfts: zero-length name terminated array of cftypes
3279 * Register @cfts to @ss. Files described by @cfts are created for all
3280 * existing cgroups to which @ss is attached and all future cgroups will
3281 * have them too. This function can be called anytime whether @ss is
3284 * Returns 0 on successful registration, -errno on failure. Note that this
3285 * function currently returns 0 as long as @cfts registration is successful
3286 * even if some file creation attempts on existing cgroups fail.
3288 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3295 if (!cfts || cfts[0].name[0] == '\0')
3298 ret = cgroup_init_cftypes(ss, cfts);
3302 mutex_lock(&cgroup_mutex);
3304 list_add_tail(&cfts->node, &ss->cfts);
3305 ret = cgroup_apply_cftypes(cfts, true);
3307 cgroup_rm_cftypes_locked(cfts);
3309 mutex_unlock(&cgroup_mutex);
3314 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3315 * @ss: target cgroup subsystem
3316 * @cfts: zero-length name terminated array of cftypes
3318 * Similar to cgroup_add_cftypes() but the added files are only used for
3319 * the default hierarchy.
3321 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3325 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3326 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3327 return cgroup_add_cftypes(ss, cfts);
3331 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3332 * @ss: target cgroup subsystem
3333 * @cfts: zero-length name terminated array of cftypes
3335 * Similar to cgroup_add_cftypes() but the added files are only used for
3336 * the legacy hierarchies.
3338 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3343 * If legacy_flies_on_dfl, we want to show the legacy files on the
3344 * dfl hierarchy but iff the target subsystem hasn't been updated
3345 * for the dfl hierarchy yet.
3347 if (!cgroup_legacy_files_on_dfl ||
3348 ss->dfl_cftypes != ss->legacy_cftypes) {
3349 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3350 cft->flags |= __CFTYPE_NOT_ON_DFL;
3353 return cgroup_add_cftypes(ss, cfts);
3357 * cgroup_task_count - count the number of tasks in a cgroup.
3358 * @cgrp: the cgroup in question
3360 * Return the number of tasks in the cgroup.
3362 static int cgroup_task_count(const struct cgroup *cgrp)
3365 struct cgrp_cset_link *link;
3367 down_read(&css_set_rwsem);
3368 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3369 count += atomic_read(&link->cset->refcount);
3370 up_read(&css_set_rwsem);
3375 * css_next_child - find the next child of a given css
3376 * @pos: the current position (%NULL to initiate traversal)
3377 * @parent: css whose children to walk
3379 * This function returns the next child of @parent and should be called
3380 * under either cgroup_mutex or RCU read lock. The only requirement is
3381 * that @parent and @pos are accessible. The next sibling is guaranteed to
3382 * be returned regardless of their states.
3384 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3385 * css which finished ->css_online() is guaranteed to be visible in the
3386 * future iterations and will stay visible until the last reference is put.
3387 * A css which hasn't finished ->css_online() or already finished
3388 * ->css_offline() may show up during traversal. It's each subsystem's
3389 * responsibility to synchronize against on/offlining.
3391 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3392 struct cgroup_subsys_state *parent)
3394 struct cgroup_subsys_state *next;
3396 cgroup_assert_mutex_or_rcu_locked();
3399 * @pos could already have been unlinked from the sibling list.
3400 * Once a cgroup is removed, its ->sibling.next is no longer
3401 * updated when its next sibling changes. CSS_RELEASED is set when
3402 * @pos is taken off list, at which time its next pointer is valid,
3403 * and, as releases are serialized, the one pointed to by the next
3404 * pointer is guaranteed to not have started release yet. This
3405 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3406 * critical section, the one pointed to by its next pointer is
3407 * guaranteed to not have finished its RCU grace period even if we
3408 * have dropped rcu_read_lock() inbetween iterations.
3410 * If @pos has CSS_RELEASED set, its next pointer can't be
3411 * dereferenced; however, as each css is given a monotonically
3412 * increasing unique serial number and always appended to the
3413 * sibling list, the next one can be found by walking the parent's
3414 * children until the first css with higher serial number than
3415 * @pos's. While this path can be slower, it happens iff iteration
3416 * races against release and the race window is very small.
3419 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3420 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3421 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3423 list_for_each_entry_rcu(next, &parent->children, sibling)
3424 if (next->serial_nr > pos->serial_nr)
3429 * @next, if not pointing to the head, can be dereferenced and is
3432 if (&next->sibling != &parent->children)
3438 * css_next_descendant_pre - find the next descendant for pre-order walk
3439 * @pos: the current position (%NULL to initiate traversal)
3440 * @root: css whose descendants to walk
3442 * To be used by css_for_each_descendant_pre(). Find the next descendant
3443 * to visit for pre-order traversal of @root's descendants. @root is
3444 * included in the iteration and the first node to be visited.
3446 * While this function requires cgroup_mutex or RCU read locking, it
3447 * doesn't require the whole traversal to be contained in a single critical
3448 * section. This function will return the correct next descendant as long
3449 * as both @pos and @root are accessible and @pos is a descendant of @root.
3451 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3452 * css which finished ->css_online() is guaranteed to be visible in the
3453 * future iterations and will stay visible until the last reference is put.
3454 * A css which hasn't finished ->css_online() or already finished
3455 * ->css_offline() may show up during traversal. It's each subsystem's
3456 * responsibility to synchronize against on/offlining.
3458 struct cgroup_subsys_state *
3459 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3460 struct cgroup_subsys_state *root)
3462 struct cgroup_subsys_state *next;
3464 cgroup_assert_mutex_or_rcu_locked();
3466 /* if first iteration, visit @root */
3470 /* visit the first child if exists */
3471 next = css_next_child(NULL, pos);
3475 /* no child, visit my or the closest ancestor's next sibling */
3476 while (pos != root) {
3477 next = css_next_child(pos, pos->parent);
3487 * css_rightmost_descendant - return the rightmost descendant of a css
3488 * @pos: css of interest
3490 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3491 * is returned. This can be used during pre-order traversal to skip
3494 * While this function requires cgroup_mutex or RCU read locking, it
3495 * doesn't require the whole traversal to be contained in a single critical
3496 * section. This function will return the correct rightmost descendant as
3497 * long as @pos is accessible.
3499 struct cgroup_subsys_state *
3500 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3502 struct cgroup_subsys_state *last, *tmp;
3504 cgroup_assert_mutex_or_rcu_locked();
3508 /* ->prev isn't RCU safe, walk ->next till the end */
3510 css_for_each_child(tmp, last)
3517 static struct cgroup_subsys_state *
3518 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3520 struct cgroup_subsys_state *last;
3524 pos = css_next_child(NULL, pos);
3531 * css_next_descendant_post - find the next descendant for post-order walk
3532 * @pos: the current position (%NULL to initiate traversal)
3533 * @root: css whose descendants to walk
3535 * To be used by css_for_each_descendant_post(). Find the next descendant
3536 * to visit for post-order traversal of @root's descendants. @root is
3537 * included in the iteration and the last node to be visited.
3539 * While this function requires cgroup_mutex or RCU read locking, it
3540 * doesn't require the whole traversal to be contained in a single critical
3541 * section. This function will return the correct next descendant as long
3542 * as both @pos and @cgroup are accessible and @pos is a descendant of
3545 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3546 * css which finished ->css_online() is guaranteed to be visible in the
3547 * future iterations and will stay visible until the last reference is put.
3548 * A css which hasn't finished ->css_online() or already finished
3549 * ->css_offline() may show up during traversal. It's each subsystem's
3550 * responsibility to synchronize against on/offlining.
3552 struct cgroup_subsys_state *
3553 css_next_descendant_post(struct cgroup_subsys_state *pos,
3554 struct cgroup_subsys_state *root)
3556 struct cgroup_subsys_state *next;
3558 cgroup_assert_mutex_or_rcu_locked();
3560 /* if first iteration, visit leftmost descendant which may be @root */
3562 return css_leftmost_descendant(root);
3564 /* if we visited @root, we're done */
3568 /* if there's an unvisited sibling, visit its leftmost descendant */
3569 next = css_next_child(pos, pos->parent);
3571 return css_leftmost_descendant(next);
3573 /* no sibling left, visit parent */
3578 * css_has_online_children - does a css have online children
3579 * @css: the target css
3581 * Returns %true if @css has any online children; otherwise, %false. This
3582 * function can be called from any context but the caller is responsible
3583 * for synchronizing against on/offlining as necessary.
3585 bool css_has_online_children(struct cgroup_subsys_state *css)
3587 struct cgroup_subsys_state *child;
3591 css_for_each_child(child, css) {
3592 if (child->flags & CSS_ONLINE) {
3602 * css_advance_task_iter - advance a task itererator to the next css_set
3603 * @it: the iterator to advance
3605 * Advance @it to the next css_set to walk.
3607 static void css_advance_task_iter(struct css_task_iter *it)
3609 struct list_head *l = it->cset_pos;
3610 struct cgrp_cset_link *link;
3611 struct css_set *cset;
3613 /* Advance to the next non-empty css_set */
3616 if (l == it->cset_head) {
3617 it->cset_pos = NULL;
3622 cset = container_of(l, struct css_set,
3623 e_cset_node[it->ss->id]);
3625 link = list_entry(l, struct cgrp_cset_link, cset_link);
3628 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3632 if (!list_empty(&cset->tasks))
3633 it->task_pos = cset->tasks.next;
3635 it->task_pos = cset->mg_tasks.next;
3637 it->tasks_head = &cset->tasks;
3638 it->mg_tasks_head = &cset->mg_tasks;
3642 * css_task_iter_start - initiate task iteration
3643 * @css: the css to walk tasks of
3644 * @it: the task iterator to use
3646 * Initiate iteration through the tasks of @css. The caller can call
3647 * css_task_iter_next() to walk through the tasks until the function
3648 * returns NULL. On completion of iteration, css_task_iter_end() must be
3651 * Note that this function acquires a lock which is released when the
3652 * iteration finishes. The caller can't sleep while iteration is in
3655 void css_task_iter_start(struct cgroup_subsys_state *css,
3656 struct css_task_iter *it)
3657 __acquires(css_set_rwsem)
3659 /* no one should try to iterate before mounting cgroups */
3660 WARN_ON_ONCE(!use_task_css_set_links);
3662 down_read(&css_set_rwsem);
3667 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3669 it->cset_pos = &css->cgroup->cset_links;
3671 it->cset_head = it->cset_pos;
3673 css_advance_task_iter(it);
3677 * css_task_iter_next - return the next task for the iterator
3678 * @it: the task iterator being iterated
3680 * The "next" function for task iteration. @it should have been
3681 * initialized via css_task_iter_start(). Returns NULL when the iteration
3684 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3686 struct task_struct *res;
3687 struct list_head *l = it->task_pos;
3689 /* If the iterator cg is NULL, we have no tasks */
3692 res = list_entry(l, struct task_struct, cg_list);
3695 * Advance iterator to find next entry. cset->tasks is consumed
3696 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3701 if (l == it->tasks_head)
3702 l = it->mg_tasks_head->next;
3704 if (l == it->mg_tasks_head)
3705 css_advance_task_iter(it);
3713 * css_task_iter_end - finish task iteration
3714 * @it: the task iterator to finish
3716 * Finish task iteration started by css_task_iter_start().
3718 void css_task_iter_end(struct css_task_iter *it)
3719 __releases(css_set_rwsem)
3721 up_read(&css_set_rwsem);
3725 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3726 * @to: cgroup to which the tasks will be moved
3727 * @from: cgroup in which the tasks currently reside
3729 * Locking rules between cgroup_post_fork() and the migration path
3730 * guarantee that, if a task is forking while being migrated, the new child
3731 * is guaranteed to be either visible in the source cgroup after the
3732 * parent's migration is complete or put into the target cgroup. No task
3733 * can slip out of migration through forking.
3735 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3737 LIST_HEAD(preloaded_csets);
3738 struct cgrp_cset_link *link;
3739 struct css_task_iter it;
3740 struct task_struct *task;
3743 mutex_lock(&cgroup_mutex);
3745 /* all tasks in @from are being moved, all csets are source */
3746 down_read(&css_set_rwsem);
3747 list_for_each_entry(link, &from->cset_links, cset_link)
3748 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3749 up_read(&css_set_rwsem);
3751 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3756 * Migrate tasks one-by-one until @form is empty. This fails iff
3757 * ->can_attach() fails.
3760 css_task_iter_start(&from->self, &it);
3761 task = css_task_iter_next(&it);
3763 get_task_struct(task);
3764 css_task_iter_end(&it);
3767 ret = cgroup_migrate(to, task, false);
3768 put_task_struct(task);
3770 } while (task && !ret);
3772 cgroup_migrate_finish(&preloaded_csets);
3773 mutex_unlock(&cgroup_mutex);
3778 * Stuff for reading the 'tasks'/'procs' files.
3780 * Reading this file can return large amounts of data if a cgroup has
3781 * *lots* of attached tasks. So it may need several calls to read(),
3782 * but we cannot guarantee that the information we produce is correct
3783 * unless we produce it entirely atomically.
3787 /* which pidlist file are we talking about? */
3788 enum cgroup_filetype {
3794 * A pidlist is a list of pids that virtually represents the contents of one
3795 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3796 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3799 struct cgroup_pidlist {
3801 * used to find which pidlist is wanted. doesn't change as long as
3802 * this particular list stays in the list.
3804 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3807 /* how many elements the above list has */
3809 /* each of these stored in a list by its cgroup */
3810 struct list_head links;
3811 /* pointer to the cgroup we belong to, for list removal purposes */
3812 struct cgroup *owner;
3813 /* for delayed destruction */
3814 struct delayed_work destroy_dwork;
3818 * The following two functions "fix" the issue where there are more pids
3819 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3820 * TODO: replace with a kernel-wide solution to this problem
3822 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3823 static void *pidlist_allocate(int count)
3825 if (PIDLIST_TOO_LARGE(count))
3826 return vmalloc(count * sizeof(pid_t));
3828 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3831 static void pidlist_free(void *p)
3837 * Used to destroy all pidlists lingering waiting for destroy timer. None
3838 * should be left afterwards.
3840 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3842 struct cgroup_pidlist *l, *tmp_l;
3844 mutex_lock(&cgrp->pidlist_mutex);
3845 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3846 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3847 mutex_unlock(&cgrp->pidlist_mutex);
3849 flush_workqueue(cgroup_pidlist_destroy_wq);
3850 BUG_ON(!list_empty(&cgrp->pidlists));
3853 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3855 struct delayed_work *dwork = to_delayed_work(work);
3856 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3858 struct cgroup_pidlist *tofree = NULL;
3860 mutex_lock(&l->owner->pidlist_mutex);
3863 * Destroy iff we didn't get queued again. The state won't change
3864 * as destroy_dwork can only be queued while locked.
3866 if (!delayed_work_pending(dwork)) {
3867 list_del(&l->links);
3868 pidlist_free(l->list);
3869 put_pid_ns(l->key.ns);
3873 mutex_unlock(&l->owner->pidlist_mutex);
3878 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3879 * Returns the number of unique elements.
3881 static int pidlist_uniq(pid_t *list, int length)
3886 * we presume the 0th element is unique, so i starts at 1. trivial
3887 * edge cases first; no work needs to be done for either
3889 if (length == 0 || length == 1)
3891 /* src and dest walk down the list; dest counts unique elements */
3892 for (src = 1; src < length; src++) {
3893 /* find next unique element */
3894 while (list[src] == list[src-1]) {
3899 /* dest always points to where the next unique element goes */
3900 list[dest] = list[src];
3908 * The two pid files - task and cgroup.procs - guaranteed that the result
3909 * is sorted, which forced this whole pidlist fiasco. As pid order is
3910 * different per namespace, each namespace needs differently sorted list,
3911 * making it impossible to use, for example, single rbtree of member tasks
3912 * sorted by task pointer. As pidlists can be fairly large, allocating one
3913 * per open file is dangerous, so cgroup had to implement shared pool of
3914 * pidlists keyed by cgroup and namespace.
3916 * All this extra complexity was caused by the original implementation
3917 * committing to an entirely unnecessary property. In the long term, we
3918 * want to do away with it. Explicitly scramble sort order if on the
3919 * default hierarchy so that no such expectation exists in the new
3922 * Scrambling is done by swapping every two consecutive bits, which is
3923 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3925 static pid_t pid_fry(pid_t pid)
3927 unsigned a = pid & 0x55555555;
3928 unsigned b = pid & 0xAAAAAAAA;
3930 return (a << 1) | (b >> 1);
3933 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3935 if (cgroup_on_dfl(cgrp))
3936 return pid_fry(pid);
3941 static int cmppid(const void *a, const void *b)
3943 return *(pid_t *)a - *(pid_t *)b;
3946 static int fried_cmppid(const void *a, const void *b)
3948 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3951 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3952 enum cgroup_filetype type)
3954 struct cgroup_pidlist *l;
3955 /* don't need task_nsproxy() if we're looking at ourself */
3956 struct pid_namespace *ns = task_active_pid_ns(current);
3958 lockdep_assert_held(&cgrp->pidlist_mutex);
3960 list_for_each_entry(l, &cgrp->pidlists, links)
3961 if (l->key.type == type && l->key.ns == ns)
3967 * find the appropriate pidlist for our purpose (given procs vs tasks)
3968 * returns with the lock on that pidlist already held, and takes care
3969 * of the use count, or returns NULL with no locks held if we're out of
3972 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3973 enum cgroup_filetype type)
3975 struct cgroup_pidlist *l;
3977 lockdep_assert_held(&cgrp->pidlist_mutex);
3979 l = cgroup_pidlist_find(cgrp, type);
3983 /* entry not found; create a new one */
3984 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3988 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3990 /* don't need task_nsproxy() if we're looking at ourself */
3991 l->key.ns = get_pid_ns(task_active_pid_ns(current));
3993 list_add(&l->links, &cgrp->pidlists);
3998 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4000 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4001 struct cgroup_pidlist **lp)
4005 int pid, n = 0; /* used for populating the array */
4006 struct css_task_iter it;
4007 struct task_struct *tsk;
4008 struct cgroup_pidlist *l;
4010 lockdep_assert_held(&cgrp->pidlist_mutex);
4013 * If cgroup gets more users after we read count, we won't have
4014 * enough space - tough. This race is indistinguishable to the
4015 * caller from the case that the additional cgroup users didn't
4016 * show up until sometime later on.
4018 length = cgroup_task_count(cgrp);
4019 array = pidlist_allocate(length);
4022 /* now, populate the array */
4023 css_task_iter_start(&cgrp->self, &it);
4024 while ((tsk = css_task_iter_next(&it))) {
4025 if (unlikely(n == length))
4027 /* get tgid or pid for procs or tasks file respectively */
4028 if (type == CGROUP_FILE_PROCS)
4029 pid = task_tgid_vnr(tsk);
4031 pid = task_pid_vnr(tsk);
4032 if (pid > 0) /* make sure to only use valid results */
4035 css_task_iter_end(&it);
4037 /* now sort & (if procs) strip out duplicates */
4038 if (cgroup_on_dfl(cgrp))
4039 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4041 sort(array, length, sizeof(pid_t), cmppid, NULL);
4042 if (type == CGROUP_FILE_PROCS)
4043 length = pidlist_uniq(array, length);
4045 l = cgroup_pidlist_find_create(cgrp, type);
4047 pidlist_free(array);
4051 /* store array, freeing old if necessary */
4052 pidlist_free(l->list);
4060 * cgroupstats_build - build and fill cgroupstats
4061 * @stats: cgroupstats to fill information into
4062 * @dentry: A dentry entry belonging to the cgroup for which stats have
4065 * Build and fill cgroupstats so that taskstats can export it to user
4068 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4070 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4071 struct cgroup *cgrp;
4072 struct css_task_iter it;
4073 struct task_struct *tsk;
4075 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4076 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4077 kernfs_type(kn) != KERNFS_DIR)
4080 mutex_lock(&cgroup_mutex);
4083 * We aren't being called from kernfs and there's no guarantee on
4084 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4085 * @kn->priv is RCU safe. Let's do the RCU dancing.
4088 cgrp = rcu_dereference(kn->priv);
4089 if (!cgrp || cgroup_is_dead(cgrp)) {
4091 mutex_unlock(&cgroup_mutex);
4096 css_task_iter_start(&cgrp->self, &it);
4097 while ((tsk = css_task_iter_next(&it))) {
4098 switch (tsk->state) {
4100 stats->nr_running++;
4102 case TASK_INTERRUPTIBLE:
4103 stats->nr_sleeping++;
4105 case TASK_UNINTERRUPTIBLE:
4106 stats->nr_uninterruptible++;
4109 stats->nr_stopped++;
4112 if (delayacct_is_task_waiting_on_io(tsk))
4113 stats->nr_io_wait++;
4117 css_task_iter_end(&it);
4119 mutex_unlock(&cgroup_mutex);
4125 * seq_file methods for the tasks/procs files. The seq_file position is the
4126 * next pid to display; the seq_file iterator is a pointer to the pid
4127 * in the cgroup->l->list array.
4130 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4133 * Initially we receive a position value that corresponds to
4134 * one more than the last pid shown (or 0 on the first call or
4135 * after a seek to the start). Use a binary-search to find the
4136 * next pid to display, if any
4138 struct kernfs_open_file *of = s->private;
4139 struct cgroup *cgrp = seq_css(s)->cgroup;
4140 struct cgroup_pidlist *l;
4141 enum cgroup_filetype type = seq_cft(s)->private;
4142 int index = 0, pid = *pos;
4145 mutex_lock(&cgrp->pidlist_mutex);
4148 * !NULL @of->priv indicates that this isn't the first start()
4149 * after open. If the matching pidlist is around, we can use that.
4150 * Look for it. Note that @of->priv can't be used directly. It
4151 * could already have been destroyed.
4154 of->priv = cgroup_pidlist_find(cgrp, type);
4157 * Either this is the first start() after open or the matching
4158 * pidlist has been destroyed inbetween. Create a new one.
4161 ret = pidlist_array_load(cgrp, type,
4162 (struct cgroup_pidlist **)&of->priv);
4164 return ERR_PTR(ret);
4169 int end = l->length;
4171 while (index < end) {
4172 int mid = (index + end) / 2;
4173 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4176 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4182 /* If we're off the end of the array, we're done */
4183 if (index >= l->length)
4185 /* Update the abstract position to be the actual pid that we found */
4186 iter = l->list + index;
4187 *pos = cgroup_pid_fry(cgrp, *iter);
4191 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4193 struct kernfs_open_file *of = s->private;
4194 struct cgroup_pidlist *l = of->priv;
4197 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4198 CGROUP_PIDLIST_DESTROY_DELAY);
4199 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4202 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4204 struct kernfs_open_file *of = s->private;
4205 struct cgroup_pidlist *l = of->priv;
4207 pid_t *end = l->list + l->length;
4209 * Advance to the next pid in the array. If this goes off the
4216 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4221 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4223 seq_printf(s, "%d\n", *(int *)v);
4228 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4231 return notify_on_release(css->cgroup);
4234 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4235 struct cftype *cft, u64 val)
4238 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4240 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4244 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4247 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4250 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4251 struct cftype *cft, u64 val)
4254 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4256 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4260 /* cgroup core interface files for the default hierarchy */
4261 static struct cftype cgroup_dfl_base_files[] = {
4263 .name = "cgroup.procs",
4264 .seq_start = cgroup_pidlist_start,
4265 .seq_next = cgroup_pidlist_next,
4266 .seq_stop = cgroup_pidlist_stop,
4267 .seq_show = cgroup_pidlist_show,
4268 .private = CGROUP_FILE_PROCS,
4269 .write = cgroup_procs_write,
4270 .mode = S_IRUGO | S_IWUSR,
4273 .name = "cgroup.controllers",
4274 .flags = CFTYPE_ONLY_ON_ROOT,
4275 .seq_show = cgroup_root_controllers_show,
4278 .name = "cgroup.controllers",
4279 .flags = CFTYPE_NOT_ON_ROOT,
4280 .seq_show = cgroup_controllers_show,
4283 .name = "cgroup.subtree_control",
4284 .seq_show = cgroup_subtree_control_show,
4285 .write = cgroup_subtree_control_write,
4288 .name = "cgroup.populated",
4289 .flags = CFTYPE_NOT_ON_ROOT,
4290 .seq_show = cgroup_populated_show,
4295 /* cgroup core interface files for the legacy hierarchies */
4296 static struct cftype cgroup_legacy_base_files[] = {
4298 .name = "cgroup.procs",
4299 .seq_start = cgroup_pidlist_start,
4300 .seq_next = cgroup_pidlist_next,
4301 .seq_stop = cgroup_pidlist_stop,
4302 .seq_show = cgroup_pidlist_show,
4303 .private = CGROUP_FILE_PROCS,
4304 .write = cgroup_procs_write,
4305 .mode = S_IRUGO | S_IWUSR,
4308 .name = "cgroup.clone_children",
4309 .read_u64 = cgroup_clone_children_read,
4310 .write_u64 = cgroup_clone_children_write,
4313 .name = "cgroup.sane_behavior",
4314 .flags = CFTYPE_ONLY_ON_ROOT,
4315 .seq_show = cgroup_sane_behavior_show,
4319 .seq_start = cgroup_pidlist_start,
4320 .seq_next = cgroup_pidlist_next,
4321 .seq_stop = cgroup_pidlist_stop,
4322 .seq_show = cgroup_pidlist_show,
4323 .private = CGROUP_FILE_TASKS,
4324 .write = cgroup_tasks_write,
4325 .mode = S_IRUGO | S_IWUSR,
4328 .name = "notify_on_release",
4329 .read_u64 = cgroup_read_notify_on_release,
4330 .write_u64 = cgroup_write_notify_on_release,
4333 .name = "release_agent",
4334 .flags = CFTYPE_ONLY_ON_ROOT,
4335 .seq_show = cgroup_release_agent_show,
4336 .write = cgroup_release_agent_write,
4337 .max_write_len = PATH_MAX - 1,
4343 * cgroup_populate_dir - create subsys files in a cgroup directory
4344 * @cgrp: target cgroup
4345 * @subsys_mask: mask of the subsystem ids whose files should be added
4347 * On failure, no file is added.
4349 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4351 struct cgroup_subsys *ss;
4354 /* process cftsets of each subsystem */
4355 for_each_subsys(ss, i) {
4356 struct cftype *cfts;
4358 if (!(subsys_mask & (1 << i)))
4361 list_for_each_entry(cfts, &ss->cfts, node) {
4362 ret = cgroup_addrm_files(cgrp, cfts, true);
4369 cgroup_clear_dir(cgrp, subsys_mask);
4374 * css destruction is four-stage process.
4376 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4377 * Implemented in kill_css().
4379 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4380 * and thus css_tryget_online() is guaranteed to fail, the css can be
4381 * offlined by invoking offline_css(). After offlining, the base ref is
4382 * put. Implemented in css_killed_work_fn().
4384 * 3. When the percpu_ref reaches zero, the only possible remaining
4385 * accessors are inside RCU read sections. css_release() schedules the
4388 * 4. After the grace period, the css can be freed. Implemented in
4389 * css_free_work_fn().
4391 * It is actually hairier because both step 2 and 4 require process context
4392 * and thus involve punting to css->destroy_work adding two additional
4393 * steps to the already complex sequence.
4395 static void css_free_work_fn(struct work_struct *work)
4397 struct cgroup_subsys_state *css =
4398 container_of(work, struct cgroup_subsys_state, destroy_work);
4399 struct cgroup_subsys *ss = css->ss;
4400 struct cgroup *cgrp = css->cgroup;
4402 percpu_ref_exit(&css->refcnt);
4409 css_put(css->parent);
4412 cgroup_idr_remove(&ss->css_idr, id);
4415 /* cgroup free path */
4416 atomic_dec(&cgrp->root->nr_cgrps);
4417 cgroup_pidlist_destroy_all(cgrp);
4418 cancel_work_sync(&cgrp->release_agent_work);
4420 if (cgroup_parent(cgrp)) {
4422 * We get a ref to the parent, and put the ref when
4423 * this cgroup is being freed, so it's guaranteed
4424 * that the parent won't be destroyed before its
4427 cgroup_put(cgroup_parent(cgrp));
4428 kernfs_put(cgrp->kn);
4432 * This is root cgroup's refcnt reaching zero,
4433 * which indicates that the root should be
4436 cgroup_destroy_root(cgrp->root);
4441 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4443 struct cgroup_subsys_state *css =
4444 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4446 INIT_WORK(&css->destroy_work, css_free_work_fn);
4447 queue_work(cgroup_destroy_wq, &css->destroy_work);
4450 static void css_release_work_fn(struct work_struct *work)
4452 struct cgroup_subsys_state *css =
4453 container_of(work, struct cgroup_subsys_state, destroy_work);
4454 struct cgroup_subsys *ss = css->ss;
4455 struct cgroup *cgrp = css->cgroup;
4457 mutex_lock(&cgroup_mutex);
4459 css->flags |= CSS_RELEASED;
4460 list_del_rcu(&css->sibling);
4463 /* css release path */
4464 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4465 if (ss->css_released)
4466 ss->css_released(css);
4468 /* cgroup release path */
4469 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4473 * There are two control paths which try to determine
4474 * cgroup from dentry without going through kernfs -
4475 * cgroupstats_build() and css_tryget_online_from_dir().
4476 * Those are supported by RCU protecting clearing of
4477 * cgrp->kn->priv backpointer.
4479 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4482 mutex_unlock(&cgroup_mutex);
4484 call_rcu(&css->rcu_head, css_free_rcu_fn);
4487 static void css_release(struct percpu_ref *ref)
4489 struct cgroup_subsys_state *css =
4490 container_of(ref, struct cgroup_subsys_state, refcnt);
4492 INIT_WORK(&css->destroy_work, css_release_work_fn);
4493 queue_work(cgroup_destroy_wq, &css->destroy_work);
4496 static void init_and_link_css(struct cgroup_subsys_state *css,
4497 struct cgroup_subsys *ss, struct cgroup *cgrp)
4499 lockdep_assert_held(&cgroup_mutex);
4503 memset(css, 0, sizeof(*css));
4506 INIT_LIST_HEAD(&css->sibling);
4507 INIT_LIST_HEAD(&css->children);
4508 css->serial_nr = css_serial_nr_next++;
4510 if (cgroup_parent(cgrp)) {
4511 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4512 css_get(css->parent);
4515 BUG_ON(cgroup_css(cgrp, ss));
4518 /* invoke ->css_online() on a new CSS and mark it online if successful */
4519 static int online_css(struct cgroup_subsys_state *css)
4521 struct cgroup_subsys *ss = css->ss;
4524 lockdep_assert_held(&cgroup_mutex);
4527 ret = ss->css_online(css);
4529 css->flags |= CSS_ONLINE;
4530 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4535 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4536 static void offline_css(struct cgroup_subsys_state *css)
4538 struct cgroup_subsys *ss = css->ss;
4540 lockdep_assert_held(&cgroup_mutex);
4542 if (!(css->flags & CSS_ONLINE))
4545 if (ss->css_offline)
4546 ss->css_offline(css);
4548 css->flags &= ~CSS_ONLINE;
4549 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4551 wake_up_all(&css->cgroup->offline_waitq);
4555 * create_css - create a cgroup_subsys_state
4556 * @cgrp: the cgroup new css will be associated with
4557 * @ss: the subsys of new css
4558 * @visible: whether to create control knobs for the new css or not
4560 * Create a new css associated with @cgrp - @ss pair. On success, the new
4561 * css is online and installed in @cgrp with all interface files created if
4562 * @visible. Returns 0 on success, -errno on failure.
4564 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4567 struct cgroup *parent = cgroup_parent(cgrp);
4568 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4569 struct cgroup_subsys_state *css;
4572 lockdep_assert_held(&cgroup_mutex);
4574 css = ss->css_alloc(parent_css);
4576 return PTR_ERR(css);
4578 init_and_link_css(css, ss, cgrp);
4580 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4584 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4586 goto err_free_percpu_ref;
4590 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4595 /* @css is ready to be brought online now, make it visible */
4596 list_add_tail_rcu(&css->sibling, &parent_css->children);
4597 cgroup_idr_replace(&ss->css_idr, css, css->id);
4599 err = online_css(css);
4603 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4604 cgroup_parent(parent)) {
4605 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4606 current->comm, current->pid, ss->name);
4607 if (!strcmp(ss->name, "memory"))
4608 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4609 ss->warned_broken_hierarchy = true;
4615 list_del_rcu(&css->sibling);
4616 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4618 cgroup_idr_remove(&ss->css_idr, css->id);
4619 err_free_percpu_ref:
4620 percpu_ref_exit(&css->refcnt);
4622 call_rcu(&css->rcu_head, css_free_rcu_fn);
4626 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4629 struct cgroup *parent, *cgrp;
4630 struct cgroup_root *root;
4631 struct cgroup_subsys *ss;
4632 struct kernfs_node *kn;
4633 struct cftype *base_files;
4636 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4638 if (strchr(name, '\n'))
4641 parent = cgroup_kn_lock_live(parent_kn);
4644 root = parent->root;
4646 /* allocate the cgroup and its ID, 0 is reserved for the root */
4647 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4653 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4658 * Temporarily set the pointer to NULL, so idr_find() won't return
4659 * a half-baked cgroup.
4661 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4664 goto out_cancel_ref;
4667 init_cgroup_housekeeping(cgrp);
4669 cgrp->self.parent = &parent->self;
4672 if (notify_on_release(parent))
4673 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4675 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4676 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4678 /* create the directory */
4679 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4687 * This extra ref will be put in cgroup_free_fn() and guarantees
4688 * that @cgrp->kn is always accessible.
4692 cgrp->self.serial_nr = css_serial_nr_next++;
4694 /* allocation complete, commit to creation */
4695 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4696 atomic_inc(&root->nr_cgrps);
4700 * @cgrp is now fully operational. If something fails after this
4701 * point, it'll be released via the normal destruction path.
4703 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4705 ret = cgroup_kn_set_ugid(kn);
4709 if (cgroup_on_dfl(cgrp))
4710 base_files = cgroup_dfl_base_files;
4712 base_files = cgroup_legacy_base_files;
4714 ret = cgroup_addrm_files(cgrp, base_files, true);
4718 /* let's create and online css's */
4719 for_each_subsys(ss, ssid) {
4720 if (parent->child_subsys_mask & (1 << ssid)) {
4721 ret = create_css(cgrp, ss,
4722 parent->subtree_control & (1 << ssid));
4729 * On the default hierarchy, a child doesn't automatically inherit
4730 * subtree_control from the parent. Each is configured manually.
4732 if (!cgroup_on_dfl(cgrp)) {
4733 cgrp->subtree_control = parent->subtree_control;
4734 cgroup_refresh_child_subsys_mask(cgrp);
4737 kernfs_activate(kn);
4743 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4745 percpu_ref_exit(&cgrp->self.refcnt);
4749 cgroup_kn_unlock(parent_kn);
4753 cgroup_destroy_locked(cgrp);
4758 * This is called when the refcnt of a css is confirmed to be killed.
4759 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4760 * initate destruction and put the css ref from kill_css().
4762 static void css_killed_work_fn(struct work_struct *work)
4764 struct cgroup_subsys_state *css =
4765 container_of(work, struct cgroup_subsys_state, destroy_work);
4767 mutex_lock(&cgroup_mutex);
4769 mutex_unlock(&cgroup_mutex);
4774 /* css kill confirmation processing requires process context, bounce */
4775 static void css_killed_ref_fn(struct percpu_ref *ref)
4777 struct cgroup_subsys_state *css =
4778 container_of(ref, struct cgroup_subsys_state, refcnt);
4780 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4781 queue_work(cgroup_destroy_wq, &css->destroy_work);
4785 * kill_css - destroy a css
4786 * @css: css to destroy
4788 * This function initiates destruction of @css by removing cgroup interface
4789 * files and putting its base reference. ->css_offline() will be invoked
4790 * asynchronously once css_tryget_online() is guaranteed to fail and when
4791 * the reference count reaches zero, @css will be released.
4793 static void kill_css(struct cgroup_subsys_state *css)
4795 lockdep_assert_held(&cgroup_mutex);
4798 * This must happen before css is disassociated with its cgroup.
4799 * See seq_css() for details.
4801 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4804 * Killing would put the base ref, but we need to keep it alive
4805 * until after ->css_offline().
4810 * cgroup core guarantees that, by the time ->css_offline() is
4811 * invoked, no new css reference will be given out via
4812 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4813 * proceed to offlining css's because percpu_ref_kill() doesn't
4814 * guarantee that the ref is seen as killed on all CPUs on return.
4816 * Use percpu_ref_kill_and_confirm() to get notifications as each
4817 * css is confirmed to be seen as killed on all CPUs.
4819 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4823 * cgroup_destroy_locked - the first stage of cgroup destruction
4824 * @cgrp: cgroup to be destroyed
4826 * css's make use of percpu refcnts whose killing latency shouldn't be
4827 * exposed to userland and are RCU protected. Also, cgroup core needs to
4828 * guarantee that css_tryget_online() won't succeed by the time
4829 * ->css_offline() is invoked. To satisfy all the requirements,
4830 * destruction is implemented in the following two steps.
4832 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4833 * userland visible parts and start killing the percpu refcnts of
4834 * css's. Set up so that the next stage will be kicked off once all
4835 * the percpu refcnts are confirmed to be killed.
4837 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4838 * rest of destruction. Once all cgroup references are gone, the
4839 * cgroup is RCU-freed.
4841 * This function implements s1. After this step, @cgrp is gone as far as
4842 * the userland is concerned and a new cgroup with the same name may be
4843 * created. As cgroup doesn't care about the names internally, this
4844 * doesn't cause any problem.
4846 static int cgroup_destroy_locked(struct cgroup *cgrp)
4847 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4849 struct cgroup_subsys_state *css;
4853 lockdep_assert_held(&cgroup_mutex);
4856 * css_set_rwsem synchronizes access to ->cset_links and prevents
4857 * @cgrp from being removed while put_css_set() is in progress.
4859 down_read(&css_set_rwsem);
4860 empty = list_empty(&cgrp->cset_links);
4861 up_read(&css_set_rwsem);
4866 * Make sure there's no live children. We can't test emptiness of
4867 * ->self.children as dead children linger on it while being
4868 * drained; otherwise, "rmdir parent/child parent" may fail.
4870 if (css_has_online_children(&cgrp->self))
4874 * Mark @cgrp dead. This prevents further task migration and child
4875 * creation by disabling cgroup_lock_live_group().
4877 cgrp->self.flags &= ~CSS_ONLINE;
4879 /* initiate massacre of all css's */
4880 for_each_css(css, ssid, cgrp)
4884 * Remove @cgrp directory along with the base files. @cgrp has an
4885 * extra ref on its kn.
4887 kernfs_remove(cgrp->kn);
4889 check_for_release(cgroup_parent(cgrp));
4891 /* put the base reference */
4892 percpu_ref_kill(&cgrp->self.refcnt);
4897 static int cgroup_rmdir(struct kernfs_node *kn)
4899 struct cgroup *cgrp;
4902 cgrp = cgroup_kn_lock_live(kn);
4906 ret = cgroup_destroy_locked(cgrp);
4908 cgroup_kn_unlock(kn);
4912 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4913 .remount_fs = cgroup_remount,
4914 .show_options = cgroup_show_options,
4915 .mkdir = cgroup_mkdir,
4916 .rmdir = cgroup_rmdir,
4917 .rename = cgroup_rename,
4920 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4922 struct cgroup_subsys_state *css;
4924 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4926 mutex_lock(&cgroup_mutex);
4928 idr_init(&ss->css_idr);
4929 INIT_LIST_HEAD(&ss->cfts);
4931 /* Create the root cgroup state for this subsystem */
4932 ss->root = &cgrp_dfl_root;
4933 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4934 /* We don't handle early failures gracefully */
4935 BUG_ON(IS_ERR(css));
4936 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4939 * Root csses are never destroyed and we can't initialize
4940 * percpu_ref during early init. Disable refcnting.
4942 css->flags |= CSS_NO_REF;
4945 /* allocation can't be done safely during early init */
4948 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4949 BUG_ON(css->id < 0);
4952 /* Update the init_css_set to contain a subsys
4953 * pointer to this state - since the subsystem is
4954 * newly registered, all tasks and hence the
4955 * init_css_set is in the subsystem's root cgroup. */
4956 init_css_set.subsys[ss->id] = css;
4958 have_fork_callback |= (bool)ss->fork << ss->id;
4959 have_exit_callback |= (bool)ss->exit << ss->id;
4961 /* At system boot, before all subsystems have been
4962 * registered, no tasks have been forked, so we don't
4963 * need to invoke fork callbacks here. */
4964 BUG_ON(!list_empty(&init_task.tasks));
4966 BUG_ON(online_css(css));
4968 mutex_unlock(&cgroup_mutex);
4972 * cgroup_init_early - cgroup initialization at system boot
4974 * Initialize cgroups at system boot, and initialize any
4975 * subsystems that request early init.
4977 int __init cgroup_init_early(void)
4979 static struct cgroup_sb_opts __initdata opts;
4980 struct cgroup_subsys *ss;
4983 init_cgroup_root(&cgrp_dfl_root, &opts);
4984 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4986 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4988 for_each_subsys(ss, i) {
4989 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4990 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4991 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4993 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4994 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4997 ss->name = cgroup_subsys_name[i];
5000 cgroup_init_subsys(ss, true);
5006 * cgroup_init - cgroup initialization
5008 * Register cgroup filesystem and /proc file, and initialize
5009 * any subsystems that didn't request early init.
5011 int __init cgroup_init(void)
5013 struct cgroup_subsys *ss;
5017 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5018 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5019 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5021 mutex_lock(&cgroup_mutex);
5023 /* Add init_css_set to the hash table */
5024 key = css_set_hash(init_css_set.subsys);
5025 hash_add(css_set_table, &init_css_set.hlist, key);
5027 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5029 mutex_unlock(&cgroup_mutex);
5031 for_each_subsys(ss, ssid) {
5032 if (ss->early_init) {
5033 struct cgroup_subsys_state *css =
5034 init_css_set.subsys[ss->id];
5036 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5038 BUG_ON(css->id < 0);
5040 cgroup_init_subsys(ss, false);
5043 list_add_tail(&init_css_set.e_cset_node[ssid],
5044 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5047 * Setting dfl_root subsys_mask needs to consider the
5048 * disabled flag and cftype registration needs kmalloc,
5049 * both of which aren't available during early_init.
5054 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5056 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5057 ss->dfl_cftypes = ss->legacy_cftypes;
5059 if (!ss->dfl_cftypes)
5060 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5062 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5063 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5065 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5066 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5070 ss->bind(init_css_set.subsys[ssid]);
5073 cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
5077 err = register_filesystem(&cgroup_fs_type);
5079 kobject_put(cgroup_kobj);
5083 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5087 static int __init cgroup_wq_init(void)
5090 * There isn't much point in executing destruction path in
5091 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5092 * Use 1 for @max_active.
5094 * We would prefer to do this in cgroup_init() above, but that
5095 * is called before init_workqueues(): so leave this until after.
5097 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5098 BUG_ON(!cgroup_destroy_wq);
5101 * Used to destroy pidlists and separate to serve as flush domain.
5102 * Cap @max_active to 1 too.
5104 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5106 BUG_ON(!cgroup_pidlist_destroy_wq);
5110 core_initcall(cgroup_wq_init);
5113 * proc_cgroup_show()
5114 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5115 * - Used for /proc/<pid>/cgroup.
5117 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5118 struct pid *pid, struct task_struct *tsk)
5122 struct cgroup_root *root;
5125 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5129 mutex_lock(&cgroup_mutex);
5130 down_read(&css_set_rwsem);
5132 for_each_root(root) {
5133 struct cgroup_subsys *ss;
5134 struct cgroup *cgrp;
5135 int ssid, count = 0;
5137 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5140 seq_printf(m, "%d:", root->hierarchy_id);
5141 for_each_subsys(ss, ssid)
5142 if (root->subsys_mask & (1 << ssid))
5143 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5144 if (strlen(root->name))
5145 seq_printf(m, "%sname=%s", count ? "," : "",
5148 cgrp = task_cgroup_from_root(tsk, root);
5149 path = cgroup_path(cgrp, buf, PATH_MAX);
5151 retval = -ENAMETOOLONG;
5160 up_read(&css_set_rwsem);
5161 mutex_unlock(&cgroup_mutex);
5167 /* Display information about each subsystem and each hierarchy */
5168 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5170 struct cgroup_subsys *ss;
5173 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5175 * ideally we don't want subsystems moving around while we do this.
5176 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5177 * subsys/hierarchy state.
5179 mutex_lock(&cgroup_mutex);
5181 for_each_subsys(ss, i)
5182 seq_printf(m, "%s\t%d\t%d\t%d\n",
5183 ss->name, ss->root->hierarchy_id,
5184 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5186 mutex_unlock(&cgroup_mutex);
5190 static int cgroupstats_open(struct inode *inode, struct file *file)
5192 return single_open(file, proc_cgroupstats_show, NULL);
5195 static const struct file_operations proc_cgroupstats_operations = {
5196 .open = cgroupstats_open,
5198 .llseek = seq_lseek,
5199 .release = single_release,
5203 * cgroup_fork - initialize cgroup related fields during copy_process()
5204 * @child: pointer to task_struct of forking parent process.
5206 * A task is associated with the init_css_set until cgroup_post_fork()
5207 * attaches it to the parent's css_set. Empty cg_list indicates that
5208 * @child isn't holding reference to its css_set.
5210 void cgroup_fork(struct task_struct *child)
5212 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5213 INIT_LIST_HEAD(&child->cg_list);
5217 * cgroup_post_fork - called on a new task after adding it to the task list
5218 * @child: the task in question
5220 * Adds the task to the list running through its css_set if necessary and
5221 * call the subsystem fork() callbacks. Has to be after the task is
5222 * visible on the task list in case we race with the first call to
5223 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5226 void cgroup_post_fork(struct task_struct *child)
5228 struct cgroup_subsys *ss;
5232 * This may race against cgroup_enable_task_cg_lists(). As that
5233 * function sets use_task_css_set_links before grabbing
5234 * tasklist_lock and we just went through tasklist_lock to add
5235 * @child, it's guaranteed that either we see the set
5236 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5237 * @child during its iteration.
5239 * If we won the race, @child is associated with %current's
5240 * css_set. Grabbing css_set_rwsem guarantees both that the
5241 * association is stable, and, on completion of the parent's
5242 * migration, @child is visible in the source of migration or
5243 * already in the destination cgroup. This guarantee is necessary
5244 * when implementing operations which need to migrate all tasks of
5245 * a cgroup to another.
5247 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5248 * will remain in init_css_set. This is safe because all tasks are
5249 * in the init_css_set before cg_links is enabled and there's no
5250 * operation which transfers all tasks out of init_css_set.
5252 if (use_task_css_set_links) {
5253 struct css_set *cset;
5255 down_write(&css_set_rwsem);
5256 cset = task_css_set(current);
5257 if (list_empty(&child->cg_list)) {
5258 rcu_assign_pointer(child->cgroups, cset);
5259 list_add(&child->cg_list, &cset->tasks);
5262 up_write(&css_set_rwsem);
5266 * Call ss->fork(). This must happen after @child is linked on
5267 * css_set; otherwise, @child might change state between ->fork()
5268 * and addition to css_set.
5270 for_each_subsys_which(ss, i, &have_fork_callback)
5275 * cgroup_exit - detach cgroup from exiting task
5276 * @tsk: pointer to task_struct of exiting process
5278 * Description: Detach cgroup from @tsk and release it.
5280 * Note that cgroups marked notify_on_release force every task in
5281 * them to take the global cgroup_mutex mutex when exiting.
5282 * This could impact scaling on very large systems. Be reluctant to
5283 * use notify_on_release cgroups where very high task exit scaling
5284 * is required on large systems.
5286 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5287 * call cgroup_exit() while the task is still competent to handle
5288 * notify_on_release(), then leave the task attached to the root cgroup in
5289 * each hierarchy for the remainder of its exit. No need to bother with
5290 * init_css_set refcnting. init_css_set never goes away and we can't race
5291 * with migration path - PF_EXITING is visible to migration path.
5293 void cgroup_exit(struct task_struct *tsk)
5295 struct cgroup_subsys *ss;
5296 struct css_set *cset;
5297 bool put_cset = false;
5301 * Unlink from @tsk from its css_set. As migration path can't race
5302 * with us, we can check cg_list without grabbing css_set_rwsem.
5304 if (!list_empty(&tsk->cg_list)) {
5305 down_write(&css_set_rwsem);
5306 list_del_init(&tsk->cg_list);
5307 up_write(&css_set_rwsem);
5311 /* Reassign the task to the init_css_set. */
5312 cset = task_css_set(tsk);
5313 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5315 /* see cgroup_post_fork() for details */
5316 for_each_subsys_which(ss, i, &have_exit_callback) {
5317 struct cgroup_subsys_state *old_css = cset->subsys[i];
5318 struct cgroup_subsys_state *css = task_css(tsk, i);
5320 ss->exit(css, old_css, tsk);
5327 static void check_for_release(struct cgroup *cgrp)
5329 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5330 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5331 schedule_work(&cgrp->release_agent_work);
5335 * Notify userspace when a cgroup is released, by running the
5336 * configured release agent with the name of the cgroup (path
5337 * relative to the root of cgroup file system) as the argument.
5339 * Most likely, this user command will try to rmdir this cgroup.
5341 * This races with the possibility that some other task will be
5342 * attached to this cgroup before it is removed, or that some other
5343 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5344 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5345 * unused, and this cgroup will be reprieved from its death sentence,
5346 * to continue to serve a useful existence. Next time it's released,
5347 * we will get notified again, if it still has 'notify_on_release' set.
5349 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5350 * means only wait until the task is successfully execve()'d. The
5351 * separate release agent task is forked by call_usermodehelper(),
5352 * then control in this thread returns here, without waiting for the
5353 * release agent task. We don't bother to wait because the caller of
5354 * this routine has no use for the exit status of the release agent
5355 * task, so no sense holding our caller up for that.
5357 static void cgroup_release_agent(struct work_struct *work)
5359 struct cgroup *cgrp =
5360 container_of(work, struct cgroup, release_agent_work);
5361 char *pathbuf = NULL, *agentbuf = NULL, *path;
5362 char *argv[3], *envp[3];
5364 mutex_lock(&cgroup_mutex);
5366 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5367 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5368 if (!pathbuf || !agentbuf)
5371 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5379 /* minimal command environment */
5381 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5384 mutex_unlock(&cgroup_mutex);
5385 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5388 mutex_unlock(&cgroup_mutex);
5394 static int __init cgroup_disable(char *str)
5396 struct cgroup_subsys *ss;
5400 while ((token = strsep(&str, ",")) != NULL) {
5404 for_each_subsys(ss, i) {
5405 if (!strcmp(token, ss->name)) {
5407 printk(KERN_INFO "Disabling %s control group"
5408 " subsystem\n", ss->name);
5415 __setup("cgroup_disable=", cgroup_disable);
5417 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5419 printk("cgroup: using legacy files on the default hierarchy\n");
5420 cgroup_legacy_files_on_dfl = true;
5423 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5426 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5427 * @dentry: directory dentry of interest
5428 * @ss: subsystem of interest
5430 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5431 * to get the corresponding css and return it. If such css doesn't exist
5432 * or can't be pinned, an ERR_PTR value is returned.
5434 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5435 struct cgroup_subsys *ss)
5437 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5438 struct cgroup_subsys_state *css = NULL;
5439 struct cgroup *cgrp;
5441 /* is @dentry a cgroup dir? */
5442 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5443 kernfs_type(kn) != KERNFS_DIR)
5444 return ERR_PTR(-EBADF);
5449 * This path doesn't originate from kernfs and @kn could already
5450 * have been or be removed at any point. @kn->priv is RCU
5451 * protected for this access. See css_release_work_fn() for details.
5453 cgrp = rcu_dereference(kn->priv);
5455 css = cgroup_css(cgrp, ss);
5457 if (!css || !css_tryget_online(css))
5458 css = ERR_PTR(-ENOENT);
5465 * css_from_id - lookup css by id
5466 * @id: the cgroup id
5467 * @ss: cgroup subsys to be looked into
5469 * Returns the css if there's valid one with @id, otherwise returns NULL.
5470 * Should be called under rcu_read_lock().
5472 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5474 WARN_ON_ONCE(!rcu_read_lock_held());
5475 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5478 #ifdef CONFIG_CGROUP_DEBUG
5479 static struct cgroup_subsys_state *
5480 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5482 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5485 return ERR_PTR(-ENOMEM);
5490 static void debug_css_free(struct cgroup_subsys_state *css)
5495 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5498 return cgroup_task_count(css->cgroup);
5501 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5504 return (u64)(unsigned long)current->cgroups;
5507 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5513 count = atomic_read(&task_css_set(current)->refcount);
5518 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5520 struct cgrp_cset_link *link;
5521 struct css_set *cset;
5524 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5528 down_read(&css_set_rwsem);
5530 cset = rcu_dereference(current->cgroups);
5531 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5532 struct cgroup *c = link->cgrp;
5534 cgroup_name(c, name_buf, NAME_MAX + 1);
5535 seq_printf(seq, "Root %d group %s\n",
5536 c->root->hierarchy_id, name_buf);
5539 up_read(&css_set_rwsem);
5544 #define MAX_TASKS_SHOWN_PER_CSS 25
5545 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5547 struct cgroup_subsys_state *css = seq_css(seq);
5548 struct cgrp_cset_link *link;
5550 down_read(&css_set_rwsem);
5551 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5552 struct css_set *cset = link->cset;
5553 struct task_struct *task;
5556 seq_printf(seq, "css_set %p\n", cset);
5558 list_for_each_entry(task, &cset->tasks, cg_list) {
5559 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5561 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5564 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5565 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5567 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5571 seq_puts(seq, " ...\n");
5573 up_read(&css_set_rwsem);
5577 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5579 return (!cgroup_has_tasks(css->cgroup) &&
5580 !css_has_online_children(&css->cgroup->self));
5583 static struct cftype debug_files[] = {
5585 .name = "taskcount",
5586 .read_u64 = debug_taskcount_read,
5590 .name = "current_css_set",
5591 .read_u64 = current_css_set_read,
5595 .name = "current_css_set_refcount",
5596 .read_u64 = current_css_set_refcount_read,
5600 .name = "current_css_set_cg_links",
5601 .seq_show = current_css_set_cg_links_read,
5605 .name = "cgroup_css_links",
5606 .seq_show = cgroup_css_links_read,
5610 .name = "releasable",
5611 .read_u64 = releasable_read,
5617 struct cgroup_subsys debug_cgrp_subsys = {
5618 .css_alloc = debug_css_alloc,
5619 .css_free = debug_css_free,
5620 .legacy_cftypes = debug_files,
5622 #endif /* CONFIG_CGROUP_DEBUG */