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_WARN(!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>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true *cgroup_subsys_enabled_key[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
159 #include <linux/cgroup_subsys.h>
164 * The default hierarchy, reserved for the subsystems that are otherwise
165 * unattached - it never has more than a single cgroup, and all tasks are
166 * part of that cgroup.
168 struct cgroup_root cgrp_dfl_root;
169 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
172 * The default hierarchy always exists but is hidden until mounted for the
173 * first time. This is for backward compatibility.
175 static bool cgrp_dfl_root_visible;
178 * Set by the boot param of the same name and makes subsystems with NULL
179 * ->dfl_files to use ->legacy_files on the default hierarchy.
181 static bool cgroup_legacy_files_on_dfl;
183 /* some controllers are not supported in the default hierarchy */
184 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
186 /* The list of hierarchy roots */
188 static LIST_HEAD(cgroup_roots);
189 static int cgroup_root_count;
191 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
192 static DEFINE_IDR(cgroup_hierarchy_idr);
195 * Assign a monotonically increasing serial number to csses. It guarantees
196 * cgroups with bigger numbers are newer than those with smaller numbers.
197 * Also, as csses are always appended to the parent's ->children list, it
198 * guarantees that sibling csses are always sorted in the ascending serial
199 * number order on the list. Protected by cgroup_mutex.
201 static u64 css_serial_nr_next = 1;
204 * These bitmask flags indicate whether tasks in the fork and exit paths have
205 * fork/exit handlers to call. This avoids us having to do extra work in the
206 * fork/exit path to check which subsystems have fork/exit callbacks.
208 static unsigned long have_fork_callback __read_mostly;
209 static unsigned long have_exit_callback __read_mostly;
211 /* Ditto for the can_fork callback. */
212 static unsigned long have_canfork_callback __read_mostly;
214 static struct cftype cgroup_dfl_base_files[];
215 static struct cftype cgroup_legacy_base_files[];
217 static int rebind_subsystems(struct cgroup_root *dst_root,
218 unsigned long ss_mask);
219 static int cgroup_destroy_locked(struct cgroup *cgrp);
220 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
222 static void css_release(struct percpu_ref *ref);
223 static void kill_css(struct cgroup_subsys_state *css);
224 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
225 struct cgroup *cgrp, struct cftype cfts[],
229 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
230 * @ssid: subsys ID of interest
232 * cgroup_subsys_enabled() can only be used with literal subsys names which
233 * is fine for individual subsystems but unsuitable for cgroup core. This
234 * is slower static_key_enabled() based test indexed by @ssid.
236 static bool cgroup_ssid_enabled(int ssid)
238 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
242 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
243 * @cgrp: the cgroup of interest
245 * The default hierarchy is the v2 interface of cgroup and this function
246 * can be used to test whether a cgroup is on the default hierarchy for
247 * cases where a subsystem should behave differnetly depending on the
250 * The set of behaviors which change on the default hierarchy are still
251 * being determined and the mount option is prefixed with __DEVEL__.
253 * List of changed behaviors:
255 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
256 * and "name" are disallowed.
258 * - When mounting an existing superblock, mount options should match.
260 * - Remount is disallowed.
262 * - rename(2) is disallowed.
264 * - "tasks" is removed. Everything should be at process granularity. Use
265 * "cgroup.procs" instead.
267 * - "cgroup.procs" is not sorted. pids will be unique unless they got
268 * recycled inbetween reads.
270 * - "release_agent" and "notify_on_release" are removed. Replacement
271 * notification mechanism will be implemented.
273 * - "cgroup.clone_children" is removed.
275 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
276 * and its descendants contain no task; otherwise, 1. The file also
277 * generates kernfs notification which can be monitored through poll and
278 * [di]notify when the value of the file changes.
280 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
281 * take masks of ancestors with non-empty cpus/mems, instead of being
282 * moved to an ancestor.
284 * - cpuset: a task can be moved into an empty cpuset, and again it takes
285 * masks of ancestors.
287 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
290 * - blkcg: blk-throttle becomes properly hierarchical.
292 * - debug: disallowed on the default hierarchy.
294 static bool cgroup_on_dfl(const struct cgroup *cgrp)
296 return cgrp->root == &cgrp_dfl_root;
299 /* IDR wrappers which synchronize using cgroup_idr_lock */
300 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
305 idr_preload(gfp_mask);
306 spin_lock_bh(&cgroup_idr_lock);
307 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_WAIT);
308 spin_unlock_bh(&cgroup_idr_lock);
313 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
317 spin_lock_bh(&cgroup_idr_lock);
318 ret = idr_replace(idr, ptr, id);
319 spin_unlock_bh(&cgroup_idr_lock);
323 static void cgroup_idr_remove(struct idr *idr, int id)
325 spin_lock_bh(&cgroup_idr_lock);
327 spin_unlock_bh(&cgroup_idr_lock);
330 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
332 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
335 return container_of(parent_css, struct cgroup, self);
340 * cgroup_css - obtain a cgroup's css for the specified subsystem
341 * @cgrp: the cgroup of interest
342 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
344 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
345 * function must be called either under cgroup_mutex or rcu_read_lock() and
346 * the caller is responsible for pinning the returned css if it wants to
347 * keep accessing it outside the said locks. This function may return
348 * %NULL if @cgrp doesn't have @subsys_id enabled.
350 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
351 struct cgroup_subsys *ss)
354 return rcu_dereference_check(cgrp->subsys[ss->id],
355 lockdep_is_held(&cgroup_mutex));
361 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
362 * @cgrp: the cgroup of interest
363 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
365 * Similar to cgroup_css() but returns the effective css, which is defined
366 * as the matching css of the nearest ancestor including self which has @ss
367 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
368 * function is guaranteed to return non-NULL css.
370 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
371 struct cgroup_subsys *ss)
373 lockdep_assert_held(&cgroup_mutex);
378 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
382 * This function is used while updating css associations and thus
383 * can't test the csses directly. Use ->child_subsys_mask.
385 while (cgroup_parent(cgrp) &&
386 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
387 cgrp = cgroup_parent(cgrp);
389 return cgroup_css(cgrp, ss);
393 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
394 * @cgrp: the cgroup of interest
395 * @ss: the subsystem of interest
397 * Find and get the effective css of @cgrp for @ss. The effective css is
398 * defined as the matching css of the nearest ancestor including self which
399 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
400 * the root css is returned, so this function always returns a valid css.
401 * The returned css must be put using css_put().
403 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
404 struct cgroup_subsys *ss)
406 struct cgroup_subsys_state *css;
411 css = cgroup_css(cgrp, ss);
413 if (css && css_tryget_online(css))
415 cgrp = cgroup_parent(cgrp);
418 css = init_css_set.subsys[ss->id];
425 /* convenient tests for these bits */
426 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
428 return !(cgrp->self.flags & CSS_ONLINE);
431 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
433 struct cgroup *cgrp = of->kn->parent->priv;
434 struct cftype *cft = of_cft(of);
437 * This is open and unprotected implementation of cgroup_css().
438 * seq_css() is only called from a kernfs file operation which has
439 * an active reference on the file. Because all the subsystem
440 * files are drained before a css is disassociated with a cgroup,
441 * the matching css from the cgroup's subsys table is guaranteed to
442 * be and stay valid until the enclosing operation is complete.
445 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
449 EXPORT_SYMBOL_GPL(of_css);
452 * cgroup_is_descendant - test ancestry
453 * @cgrp: the cgroup to be tested
454 * @ancestor: possible ancestor of @cgrp
456 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
457 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
458 * and @ancestor are accessible.
460 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
463 if (cgrp == ancestor)
465 cgrp = cgroup_parent(cgrp);
470 static int notify_on_release(const struct cgroup *cgrp)
472 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
476 * for_each_css - iterate all css's of a cgroup
477 * @css: the iteration cursor
478 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
479 * @cgrp: the target cgroup to iterate css's of
481 * Should be called under cgroup_[tree_]mutex.
483 #define for_each_css(css, ssid, cgrp) \
484 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
485 if (!((css) = rcu_dereference_check( \
486 (cgrp)->subsys[(ssid)], \
487 lockdep_is_held(&cgroup_mutex)))) { } \
491 * for_each_e_css - iterate all effective css's of a cgroup
492 * @css: the iteration cursor
493 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
494 * @cgrp: the target cgroup to iterate css's of
496 * Should be called under cgroup_[tree_]mutex.
498 #define for_each_e_css(css, ssid, cgrp) \
499 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
500 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
505 * for_each_subsys - iterate all enabled cgroup subsystems
506 * @ss: the iteration cursor
507 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
509 #define for_each_subsys(ss, ssid) \
510 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
511 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
514 * for_each_subsys_which - filter for_each_subsys with a bitmask
515 * @ss: the iteration cursor
516 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
517 * @ss_maskp: a pointer to the bitmask
519 * The block will only run for cases where the ssid-th bit (1 << ssid) of
522 #define for_each_subsys_which(ss, ssid, ss_maskp) \
523 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
526 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
527 if (((ss) = cgroup_subsys[ssid]) && false) \
531 /* iterate across the hierarchies */
532 #define for_each_root(root) \
533 list_for_each_entry((root), &cgroup_roots, root_list)
535 /* iterate over child cgrps, lock should be held throughout iteration */
536 #define cgroup_for_each_live_child(child, cgrp) \
537 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
538 if (({ lockdep_assert_held(&cgroup_mutex); \
539 cgroup_is_dead(child); })) \
543 static void cgroup_release_agent(struct work_struct *work);
544 static void check_for_release(struct cgroup *cgrp);
547 * A cgroup can be associated with multiple css_sets as different tasks may
548 * belong to different cgroups on different hierarchies. In the other
549 * direction, a css_set is naturally associated with multiple cgroups.
550 * This M:N relationship is represented by the following link structure
551 * which exists for each association and allows traversing the associations
554 struct cgrp_cset_link {
555 /* the cgroup and css_set this link associates */
557 struct css_set *cset;
559 /* list of cgrp_cset_links anchored at cgrp->cset_links */
560 struct list_head cset_link;
562 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
563 struct list_head cgrp_link;
567 * The default css_set - used by init and its children prior to any
568 * hierarchies being mounted. It contains a pointer to the root state
569 * for each subsystem. Also used to anchor the list of css_sets. Not
570 * reference-counted, to improve performance when child cgroups
571 * haven't been created.
573 struct css_set init_css_set = {
574 .refcount = ATOMIC_INIT(1),
575 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
576 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
577 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
578 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
579 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
582 static int css_set_count = 1; /* 1 for init_css_set */
585 * cgroup_update_populated - updated populated count of a cgroup
586 * @cgrp: the target cgroup
587 * @populated: inc or dec populated count
589 * @cgrp is either getting the first task (css_set) or losing the last.
590 * Update @cgrp->populated_cnt accordingly. The count is propagated
591 * towards root so that a given cgroup's populated_cnt is zero iff the
592 * cgroup and all its descendants are empty.
594 * @cgrp's interface file "cgroup.populated" is zero if
595 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
596 * changes from or to zero, userland is notified that the content of the
597 * interface file has changed. This can be used to detect when @cgrp and
598 * its descendants become populated or empty.
600 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
602 lockdep_assert_held(&css_set_rwsem);
608 trigger = !cgrp->populated_cnt++;
610 trigger = !--cgrp->populated_cnt;
615 cgroup_file_notify(&cgrp->events_file);
617 cgrp = cgroup_parent(cgrp);
622 * hash table for cgroup groups. This improves the performance to find
623 * an existing css_set. This hash doesn't (currently) take into
624 * account cgroups in empty hierarchies.
626 #define CSS_SET_HASH_BITS 7
627 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
629 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
631 unsigned long key = 0UL;
632 struct cgroup_subsys *ss;
635 for_each_subsys(ss, i)
636 key += (unsigned long)css[i];
637 key = (key >> 16) ^ key;
642 static void put_css_set_locked(struct css_set *cset)
644 struct cgrp_cset_link *link, *tmp_link;
645 struct cgroup_subsys *ss;
648 lockdep_assert_held(&css_set_rwsem);
650 if (!atomic_dec_and_test(&cset->refcount))
653 /* This css_set is dead. unlink it and release cgroup refcounts */
654 for_each_subsys(ss, ssid)
655 list_del(&cset->e_cset_node[ssid]);
656 hash_del(&cset->hlist);
659 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
660 struct cgroup *cgrp = link->cgrp;
662 list_del(&link->cset_link);
663 list_del(&link->cgrp_link);
665 /* @cgrp can't go away while we're holding css_set_rwsem */
666 if (list_empty(&cgrp->cset_links)) {
667 cgroup_update_populated(cgrp, false);
668 check_for_release(cgrp);
674 kfree_rcu(cset, rcu_head);
677 static void put_css_set(struct css_set *cset)
680 * Ensure that the refcount doesn't hit zero while any readers
681 * can see it. Similar to atomic_dec_and_lock(), but for an
684 if (atomic_add_unless(&cset->refcount, -1, 1))
687 down_write(&css_set_rwsem);
688 put_css_set_locked(cset);
689 up_write(&css_set_rwsem);
693 * refcounted get/put for css_set objects
695 static inline void get_css_set(struct css_set *cset)
697 atomic_inc(&cset->refcount);
701 * compare_css_sets - helper function for find_existing_css_set().
702 * @cset: candidate css_set being tested
703 * @old_cset: existing css_set for a task
704 * @new_cgrp: cgroup that's being entered by the task
705 * @template: desired set of css pointers in css_set (pre-calculated)
707 * Returns true if "cset" matches "old_cset" except for the hierarchy
708 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
710 static bool compare_css_sets(struct css_set *cset,
711 struct css_set *old_cset,
712 struct cgroup *new_cgrp,
713 struct cgroup_subsys_state *template[])
715 struct list_head *l1, *l2;
718 * On the default hierarchy, there can be csets which are
719 * associated with the same set of cgroups but different csses.
720 * Let's first ensure that csses match.
722 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
726 * Compare cgroup pointers in order to distinguish between
727 * different cgroups in hierarchies. As different cgroups may
728 * share the same effective css, this comparison is always
731 l1 = &cset->cgrp_links;
732 l2 = &old_cset->cgrp_links;
734 struct cgrp_cset_link *link1, *link2;
735 struct cgroup *cgrp1, *cgrp2;
739 /* See if we reached the end - both lists are equal length. */
740 if (l1 == &cset->cgrp_links) {
741 BUG_ON(l2 != &old_cset->cgrp_links);
744 BUG_ON(l2 == &old_cset->cgrp_links);
746 /* Locate the cgroups associated with these links. */
747 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
748 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
751 /* Hierarchies should be linked in the same order. */
752 BUG_ON(cgrp1->root != cgrp2->root);
755 * If this hierarchy is the hierarchy of the cgroup
756 * that's changing, then we need to check that this
757 * css_set points to the new cgroup; if it's any other
758 * hierarchy, then this css_set should point to the
759 * same cgroup as the old css_set.
761 if (cgrp1->root == new_cgrp->root) {
762 if (cgrp1 != new_cgrp)
773 * find_existing_css_set - init css array and find the matching css_set
774 * @old_cset: the css_set that we're using before the cgroup transition
775 * @cgrp: the cgroup that we're moving into
776 * @template: out param for the new set of csses, should be clear on entry
778 static struct css_set *find_existing_css_set(struct css_set *old_cset,
780 struct cgroup_subsys_state *template[])
782 struct cgroup_root *root = cgrp->root;
783 struct cgroup_subsys *ss;
784 struct css_set *cset;
789 * Build the set of subsystem state objects that we want to see in the
790 * new css_set. while subsystems can change globally, the entries here
791 * won't change, so no need for locking.
793 for_each_subsys(ss, i) {
794 if (root->subsys_mask & (1UL << i)) {
796 * @ss is in this hierarchy, so we want the
797 * effective css from @cgrp.
799 template[i] = cgroup_e_css(cgrp, ss);
802 * @ss is not in this hierarchy, so we don't want
805 template[i] = old_cset->subsys[i];
809 key = css_set_hash(template);
810 hash_for_each_possible(css_set_table, cset, hlist, key) {
811 if (!compare_css_sets(cset, old_cset, cgrp, template))
814 /* This css_set matches what we need */
818 /* No existing cgroup group matched */
822 static void free_cgrp_cset_links(struct list_head *links_to_free)
824 struct cgrp_cset_link *link, *tmp_link;
826 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
827 list_del(&link->cset_link);
833 * allocate_cgrp_cset_links - allocate cgrp_cset_links
834 * @count: the number of links to allocate
835 * @tmp_links: list_head the allocated links are put on
837 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
838 * through ->cset_link. Returns 0 on success or -errno.
840 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
842 struct cgrp_cset_link *link;
845 INIT_LIST_HEAD(tmp_links);
847 for (i = 0; i < count; i++) {
848 link = kzalloc(sizeof(*link), GFP_KERNEL);
850 free_cgrp_cset_links(tmp_links);
853 list_add(&link->cset_link, tmp_links);
859 * link_css_set - a helper function to link a css_set to a cgroup
860 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
861 * @cset: the css_set to be linked
862 * @cgrp: the destination cgroup
864 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
867 struct cgrp_cset_link *link;
869 BUG_ON(list_empty(tmp_links));
871 if (cgroup_on_dfl(cgrp))
872 cset->dfl_cgrp = cgrp;
874 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
878 if (list_empty(&cgrp->cset_links))
879 cgroup_update_populated(cgrp, true);
880 list_move(&link->cset_link, &cgrp->cset_links);
883 * Always add links to the tail of the list so that the list
884 * is sorted by order of hierarchy creation
886 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
890 * find_css_set - return a new css_set with one cgroup updated
891 * @old_cset: the baseline css_set
892 * @cgrp: the cgroup to be updated
894 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
895 * substituted into the appropriate hierarchy.
897 static struct css_set *find_css_set(struct css_set *old_cset,
900 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
901 struct css_set *cset;
902 struct list_head tmp_links;
903 struct cgrp_cset_link *link;
904 struct cgroup_subsys *ss;
908 lockdep_assert_held(&cgroup_mutex);
910 /* First see if we already have a cgroup group that matches
912 down_read(&css_set_rwsem);
913 cset = find_existing_css_set(old_cset, cgrp, template);
916 up_read(&css_set_rwsem);
921 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
925 /* Allocate all the cgrp_cset_link objects that we'll need */
926 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
931 atomic_set(&cset->refcount, 1);
932 INIT_LIST_HEAD(&cset->cgrp_links);
933 INIT_LIST_HEAD(&cset->tasks);
934 INIT_LIST_HEAD(&cset->mg_tasks);
935 INIT_LIST_HEAD(&cset->mg_preload_node);
936 INIT_LIST_HEAD(&cset->mg_node);
937 INIT_HLIST_NODE(&cset->hlist);
939 /* Copy the set of subsystem state objects generated in
940 * find_existing_css_set() */
941 memcpy(cset->subsys, template, sizeof(cset->subsys));
943 down_write(&css_set_rwsem);
944 /* Add reference counts and links from the new css_set. */
945 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
946 struct cgroup *c = link->cgrp;
948 if (c->root == cgrp->root)
950 link_css_set(&tmp_links, cset, c);
953 BUG_ON(!list_empty(&tmp_links));
957 /* Add @cset to the hash table */
958 key = css_set_hash(cset->subsys);
959 hash_add(css_set_table, &cset->hlist, key);
961 for_each_subsys(ss, ssid)
962 list_add_tail(&cset->e_cset_node[ssid],
963 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
965 up_write(&css_set_rwsem);
970 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
972 struct cgroup *root_cgrp = kf_root->kn->priv;
974 return root_cgrp->root;
977 static int cgroup_init_root_id(struct cgroup_root *root)
981 lockdep_assert_held(&cgroup_mutex);
983 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
987 root->hierarchy_id = id;
991 static void cgroup_exit_root_id(struct cgroup_root *root)
993 lockdep_assert_held(&cgroup_mutex);
995 if (root->hierarchy_id) {
996 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
997 root->hierarchy_id = 0;
1001 static void cgroup_free_root(struct cgroup_root *root)
1004 /* hierarchy ID should already have been released */
1005 WARN_ON_ONCE(root->hierarchy_id);
1007 idr_destroy(&root->cgroup_idr);
1012 static void cgroup_destroy_root(struct cgroup_root *root)
1014 struct cgroup *cgrp = &root->cgrp;
1015 struct cgrp_cset_link *link, *tmp_link;
1017 mutex_lock(&cgroup_mutex);
1019 BUG_ON(atomic_read(&root->nr_cgrps));
1020 BUG_ON(!list_empty(&cgrp->self.children));
1022 /* Rebind all subsystems back to the default hierarchy */
1023 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1026 * Release all the links from cset_links to this hierarchy's
1029 down_write(&css_set_rwsem);
1031 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1032 list_del(&link->cset_link);
1033 list_del(&link->cgrp_link);
1036 up_write(&css_set_rwsem);
1038 if (!list_empty(&root->root_list)) {
1039 list_del(&root->root_list);
1040 cgroup_root_count--;
1043 cgroup_exit_root_id(root);
1045 mutex_unlock(&cgroup_mutex);
1047 kernfs_destroy_root(root->kf_root);
1048 cgroup_free_root(root);
1051 /* look up cgroup associated with given css_set on the specified hierarchy */
1052 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1053 struct cgroup_root *root)
1055 struct cgroup *res = NULL;
1057 lockdep_assert_held(&cgroup_mutex);
1058 lockdep_assert_held(&css_set_rwsem);
1060 if (cset == &init_css_set) {
1063 struct cgrp_cset_link *link;
1065 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1066 struct cgroup *c = link->cgrp;
1068 if (c->root == root) {
1080 * Return the cgroup for "task" from the given hierarchy. Must be
1081 * called with cgroup_mutex and css_set_rwsem held.
1083 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1084 struct cgroup_root *root)
1087 * No need to lock the task - since we hold cgroup_mutex the
1088 * task can't change groups, so the only thing that can happen
1089 * is that it exits and its css is set back to init_css_set.
1091 return cset_cgroup_from_root(task_css_set(task), root);
1095 * A task must hold cgroup_mutex to modify cgroups.
1097 * Any task can increment and decrement the count field without lock.
1098 * So in general, code holding cgroup_mutex can't rely on the count
1099 * field not changing. However, if the count goes to zero, then only
1100 * cgroup_attach_task() can increment it again. Because a count of zero
1101 * means that no tasks are currently attached, therefore there is no
1102 * way a task attached to that cgroup can fork (the other way to
1103 * increment the count). So code holding cgroup_mutex can safely
1104 * assume that if the count is zero, it will stay zero. Similarly, if
1105 * a task holds cgroup_mutex on a cgroup with zero count, it
1106 * knows that the cgroup won't be removed, as cgroup_rmdir()
1109 * A cgroup can only be deleted if both its 'count' of using tasks
1110 * is zero, and its list of 'children' cgroups is empty. Since all
1111 * tasks in the system use _some_ cgroup, and since there is always at
1112 * least one task in the system (init, pid == 1), therefore, root cgroup
1113 * always has either children cgroups and/or using tasks. So we don't
1114 * need a special hack to ensure that root cgroup cannot be deleted.
1116 * P.S. One more locking exception. RCU is used to guard the
1117 * update of a tasks cgroup pointer by cgroup_attach_task()
1120 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1121 static const struct file_operations proc_cgroupstats_operations;
1123 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1126 struct cgroup_subsys *ss = cft->ss;
1128 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1129 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1130 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1131 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1134 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1139 * cgroup_file_mode - deduce file mode of a control file
1140 * @cft: the control file in question
1142 * S_IRUGO for read, S_IWUSR for write.
1144 static umode_t cgroup_file_mode(const struct cftype *cft)
1148 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1151 if (cft->write_u64 || cft->write_s64 || cft->write) {
1152 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1161 static void cgroup_get(struct cgroup *cgrp)
1163 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1164 css_get(&cgrp->self);
1167 static bool cgroup_tryget(struct cgroup *cgrp)
1169 return css_tryget(&cgrp->self);
1172 static void cgroup_put(struct cgroup *cgrp)
1174 css_put(&cgrp->self);
1178 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1179 * @cgrp: the target cgroup
1180 * @subtree_control: the new subtree_control mask to consider
1182 * On the default hierarchy, a subsystem may request other subsystems to be
1183 * enabled together through its ->depends_on mask. In such cases, more
1184 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1186 * This function calculates which subsystems need to be enabled if
1187 * @subtree_control is to be applied to @cgrp. The returned mask is always
1188 * a superset of @subtree_control and follows the usual hierarchy rules.
1190 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1191 unsigned long subtree_control)
1193 struct cgroup *parent = cgroup_parent(cgrp);
1194 unsigned long cur_ss_mask = subtree_control;
1195 struct cgroup_subsys *ss;
1198 lockdep_assert_held(&cgroup_mutex);
1200 if (!cgroup_on_dfl(cgrp))
1204 unsigned long new_ss_mask = cur_ss_mask;
1206 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1207 new_ss_mask |= ss->depends_on;
1210 * Mask out subsystems which aren't available. This can
1211 * happen only if some depended-upon subsystems were bound
1212 * to non-default hierarchies.
1215 new_ss_mask &= parent->child_subsys_mask;
1217 new_ss_mask &= cgrp->root->subsys_mask;
1219 if (new_ss_mask == cur_ss_mask)
1221 cur_ss_mask = new_ss_mask;
1228 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1229 * @cgrp: the target cgroup
1231 * Update @cgrp->child_subsys_mask according to the current
1232 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1234 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1236 cgrp->child_subsys_mask =
1237 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1241 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1242 * @kn: the kernfs_node being serviced
1244 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1245 * the method finishes if locking succeeded. Note that once this function
1246 * returns the cgroup returned by cgroup_kn_lock_live() may become
1247 * inaccessible any time. If the caller intends to continue to access the
1248 * cgroup, it should pin it before invoking this function.
1250 static void cgroup_kn_unlock(struct kernfs_node *kn)
1252 struct cgroup *cgrp;
1254 if (kernfs_type(kn) == KERNFS_DIR)
1257 cgrp = kn->parent->priv;
1259 mutex_unlock(&cgroup_mutex);
1261 kernfs_unbreak_active_protection(kn);
1266 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1267 * @kn: the kernfs_node being serviced
1269 * This helper is to be used by a cgroup kernfs method currently servicing
1270 * @kn. It breaks the active protection, performs cgroup locking and
1271 * verifies that the associated cgroup is alive. Returns the cgroup if
1272 * alive; otherwise, %NULL. A successful return should be undone by a
1273 * matching cgroup_kn_unlock() invocation.
1275 * Any cgroup kernfs method implementation which requires locking the
1276 * associated cgroup should use this helper. It avoids nesting cgroup
1277 * locking under kernfs active protection and allows all kernfs operations
1278 * including self-removal.
1280 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1282 struct cgroup *cgrp;
1284 if (kernfs_type(kn) == KERNFS_DIR)
1287 cgrp = kn->parent->priv;
1290 * We're gonna grab cgroup_mutex which nests outside kernfs
1291 * active_ref. cgroup liveliness check alone provides enough
1292 * protection against removal. Ensure @cgrp stays accessible and
1293 * break the active_ref protection.
1295 if (!cgroup_tryget(cgrp))
1297 kernfs_break_active_protection(kn);
1299 mutex_lock(&cgroup_mutex);
1301 if (!cgroup_is_dead(cgrp))
1304 cgroup_kn_unlock(kn);
1308 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1310 char name[CGROUP_FILE_NAME_MAX];
1312 lockdep_assert_held(&cgroup_mutex);
1313 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1317 * css_clear_dir - remove subsys files in a cgroup directory
1319 * @cgrp_override: specify if target cgroup is different from css->cgroup
1321 static void css_clear_dir(struct cgroup_subsys_state *css,
1322 struct cgroup *cgrp_override)
1324 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1325 struct cftype *cfts;
1327 list_for_each_entry(cfts, &css->ss->cfts, node)
1328 cgroup_addrm_files(css, cgrp, cfts, false);
1332 * css_populate_dir - create subsys files in a cgroup directory
1334 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1336 * On failure, no file is added.
1338 static int css_populate_dir(struct cgroup_subsys_state *css,
1339 struct cgroup *cgrp_override)
1341 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1342 struct cftype *cfts, *failed_cfts;
1346 if (cgroup_on_dfl(cgrp))
1347 cfts = cgroup_dfl_base_files;
1349 cfts = cgroup_legacy_base_files;
1351 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1354 list_for_each_entry(cfts, &css->ss->cfts, node) {
1355 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1363 list_for_each_entry(cfts, &css->ss->cfts, node) {
1364 if (cfts == failed_cfts)
1366 cgroup_addrm_files(css, cgrp, cfts, false);
1371 static int rebind_subsystems(struct cgroup_root *dst_root,
1372 unsigned long ss_mask)
1374 struct cgroup *dcgrp = &dst_root->cgrp;
1375 struct cgroup_subsys *ss;
1376 unsigned long tmp_ss_mask;
1379 lockdep_assert_held(&cgroup_mutex);
1381 for_each_subsys_which(ss, ssid, &ss_mask) {
1382 /* if @ss has non-root csses attached to it, can't move */
1383 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1386 /* can't move between two non-dummy roots either */
1387 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1391 /* skip creating root files on dfl_root for inhibited subsystems */
1392 tmp_ss_mask = ss_mask;
1393 if (dst_root == &cgrp_dfl_root)
1394 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1396 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1397 struct cgroup *scgrp = &ss->root->cgrp;
1400 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1405 * Rebinding back to the default root is not allowed to
1406 * fail. Using both default and non-default roots should
1407 * be rare. Moving subsystems back and forth even more so.
1408 * Just warn about it and continue.
1410 if (dst_root == &cgrp_dfl_root) {
1411 if (cgrp_dfl_root_visible) {
1412 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1414 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1419 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1422 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1428 * Nothing can fail from this point on. Remove files for the
1429 * removed subsystems and rebind each subsystem.
1431 for_each_subsys_which(ss, ssid, &ss_mask) {
1432 struct cgroup_root *src_root = ss->root;
1433 struct cgroup *scgrp = &src_root->cgrp;
1434 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1435 struct css_set *cset;
1437 WARN_ON(!css || cgroup_css(dcgrp, ss));
1439 css_clear_dir(css, NULL);
1441 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1442 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1443 ss->root = dst_root;
1444 css->cgroup = dcgrp;
1446 down_write(&css_set_rwsem);
1447 hash_for_each(css_set_table, i, cset, hlist)
1448 list_move_tail(&cset->e_cset_node[ss->id],
1449 &dcgrp->e_csets[ss->id]);
1450 up_write(&css_set_rwsem);
1452 src_root->subsys_mask &= ~(1 << ssid);
1453 scgrp->subtree_control &= ~(1 << ssid);
1454 cgroup_refresh_child_subsys_mask(scgrp);
1456 /* default hierarchy doesn't enable controllers by default */
1457 dst_root->subsys_mask |= 1 << ssid;
1458 if (dst_root == &cgrp_dfl_root) {
1459 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1461 dcgrp->subtree_control |= 1 << ssid;
1462 cgroup_refresh_child_subsys_mask(dcgrp);
1463 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1470 kernfs_activate(dcgrp->kn);
1474 static int cgroup_show_options(struct seq_file *seq,
1475 struct kernfs_root *kf_root)
1477 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1478 struct cgroup_subsys *ss;
1481 if (root != &cgrp_dfl_root)
1482 for_each_subsys(ss, ssid)
1483 if (root->subsys_mask & (1 << ssid))
1484 seq_show_option(seq, ss->legacy_name, NULL);
1485 if (root->flags & CGRP_ROOT_NOPREFIX)
1486 seq_puts(seq, ",noprefix");
1487 if (root->flags & CGRP_ROOT_XATTR)
1488 seq_puts(seq, ",xattr");
1490 spin_lock(&release_agent_path_lock);
1491 if (strlen(root->release_agent_path))
1492 seq_show_option(seq, "release_agent",
1493 root->release_agent_path);
1494 spin_unlock(&release_agent_path_lock);
1496 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1497 seq_puts(seq, ",clone_children");
1498 if (strlen(root->name))
1499 seq_show_option(seq, "name", root->name);
1503 struct cgroup_sb_opts {
1504 unsigned long subsys_mask;
1506 char *release_agent;
1507 bool cpuset_clone_children;
1509 /* User explicitly requested empty subsystem */
1513 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1515 char *token, *o = data;
1516 bool all_ss = false, one_ss = false;
1517 unsigned long mask = -1UL;
1518 struct cgroup_subsys *ss;
1522 #ifdef CONFIG_CPUSETS
1523 mask = ~(1U << cpuset_cgrp_id);
1526 memset(opts, 0, sizeof(*opts));
1528 while ((token = strsep(&o, ",")) != NULL) {
1533 if (!strcmp(token, "none")) {
1534 /* Explicitly have no subsystems */
1538 if (!strcmp(token, "all")) {
1539 /* Mutually exclusive option 'all' + subsystem name */
1545 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1546 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1549 if (!strcmp(token, "noprefix")) {
1550 opts->flags |= CGRP_ROOT_NOPREFIX;
1553 if (!strcmp(token, "clone_children")) {
1554 opts->cpuset_clone_children = true;
1557 if (!strcmp(token, "xattr")) {
1558 opts->flags |= CGRP_ROOT_XATTR;
1561 if (!strncmp(token, "release_agent=", 14)) {
1562 /* Specifying two release agents is forbidden */
1563 if (opts->release_agent)
1565 opts->release_agent =
1566 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1567 if (!opts->release_agent)
1571 if (!strncmp(token, "name=", 5)) {
1572 const char *name = token + 5;
1573 /* Can't specify an empty name */
1576 /* Must match [\w.-]+ */
1577 for (i = 0; i < strlen(name); i++) {
1581 if ((c == '.') || (c == '-') || (c == '_'))
1585 /* Specifying two names is forbidden */
1588 opts->name = kstrndup(name,
1589 MAX_CGROUP_ROOT_NAMELEN - 1,
1597 for_each_subsys(ss, i) {
1598 if (strcmp(token, ss->legacy_name))
1600 if (!cgroup_ssid_enabled(i))
1603 /* Mutually exclusive option 'all' + subsystem name */
1606 opts->subsys_mask |= (1 << i);
1611 if (i == CGROUP_SUBSYS_COUNT)
1615 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1616 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1618 pr_err("sane_behavior: no other mount options allowed\n");
1625 * If the 'all' option was specified select all the subsystems,
1626 * otherwise if 'none', 'name=' and a subsystem name options were
1627 * not specified, let's default to 'all'
1629 if (all_ss || (!one_ss && !opts->none && !opts->name))
1630 for_each_subsys(ss, i)
1631 if (cgroup_ssid_enabled(i))
1632 opts->subsys_mask |= (1 << i);
1635 * We either have to specify by name or by subsystems. (So all
1636 * empty hierarchies must have a name).
1638 if (!opts->subsys_mask && !opts->name)
1642 * Option noprefix was introduced just for backward compatibility
1643 * with the old cpuset, so we allow noprefix only if mounting just
1644 * the cpuset subsystem.
1646 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1649 /* Can't specify "none" and some subsystems */
1650 if (opts->subsys_mask && opts->none)
1656 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1659 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1660 struct cgroup_sb_opts opts;
1661 unsigned long added_mask, removed_mask;
1663 if (root == &cgrp_dfl_root) {
1664 pr_err("remount is not allowed\n");
1668 mutex_lock(&cgroup_mutex);
1670 /* See what subsystems are wanted */
1671 ret = parse_cgroupfs_options(data, &opts);
1675 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1676 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1677 task_tgid_nr(current), current->comm);
1679 added_mask = opts.subsys_mask & ~root->subsys_mask;
1680 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1682 /* Don't allow flags or name to change at remount */
1683 if ((opts.flags ^ root->flags) ||
1684 (opts.name && strcmp(opts.name, root->name))) {
1685 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1686 opts.flags, opts.name ?: "", root->flags, root->name);
1691 /* remounting is not allowed for populated hierarchies */
1692 if (!list_empty(&root->cgrp.self.children)) {
1697 ret = rebind_subsystems(root, added_mask);
1701 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1703 if (opts.release_agent) {
1704 spin_lock(&release_agent_path_lock);
1705 strcpy(root->release_agent_path, opts.release_agent);
1706 spin_unlock(&release_agent_path_lock);
1709 kfree(opts.release_agent);
1711 mutex_unlock(&cgroup_mutex);
1716 * To reduce the fork() overhead for systems that are not actually using
1717 * their cgroups capability, we don't maintain the lists running through
1718 * each css_set to its tasks until we see the list actually used - in other
1719 * words after the first mount.
1721 static bool use_task_css_set_links __read_mostly;
1723 static void cgroup_enable_task_cg_lists(void)
1725 struct task_struct *p, *g;
1727 down_write(&css_set_rwsem);
1729 if (use_task_css_set_links)
1732 use_task_css_set_links = true;
1735 * We need tasklist_lock because RCU is not safe against
1736 * while_each_thread(). Besides, a forking task that has passed
1737 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1738 * is not guaranteed to have its child immediately visible in the
1739 * tasklist if we walk through it with RCU.
1741 read_lock(&tasklist_lock);
1742 do_each_thread(g, p) {
1743 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1744 task_css_set(p) != &init_css_set);
1747 * We should check if the process is exiting, otherwise
1748 * it will race with cgroup_exit() in that the list
1749 * entry won't be deleted though the process has exited.
1750 * Do it while holding siglock so that we don't end up
1751 * racing against cgroup_exit().
1753 spin_lock_irq(&p->sighand->siglock);
1754 if (!(p->flags & PF_EXITING)) {
1755 struct css_set *cset = task_css_set(p);
1757 list_add(&p->cg_list, &cset->tasks);
1760 spin_unlock_irq(&p->sighand->siglock);
1761 } while_each_thread(g, p);
1762 read_unlock(&tasklist_lock);
1764 up_write(&css_set_rwsem);
1767 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1769 struct cgroup_subsys *ss;
1772 INIT_LIST_HEAD(&cgrp->self.sibling);
1773 INIT_LIST_HEAD(&cgrp->self.children);
1774 INIT_LIST_HEAD(&cgrp->self.files);
1775 INIT_LIST_HEAD(&cgrp->cset_links);
1776 INIT_LIST_HEAD(&cgrp->pidlists);
1777 mutex_init(&cgrp->pidlist_mutex);
1778 cgrp->self.cgroup = cgrp;
1779 cgrp->self.flags |= CSS_ONLINE;
1781 for_each_subsys(ss, ssid)
1782 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1784 init_waitqueue_head(&cgrp->offline_waitq);
1785 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1788 static void init_cgroup_root(struct cgroup_root *root,
1789 struct cgroup_sb_opts *opts)
1791 struct cgroup *cgrp = &root->cgrp;
1793 INIT_LIST_HEAD(&root->root_list);
1794 atomic_set(&root->nr_cgrps, 1);
1796 init_cgroup_housekeeping(cgrp);
1797 idr_init(&root->cgroup_idr);
1799 root->flags = opts->flags;
1800 if (opts->release_agent)
1801 strcpy(root->release_agent_path, opts->release_agent);
1803 strcpy(root->name, opts->name);
1804 if (opts->cpuset_clone_children)
1805 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1808 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1810 LIST_HEAD(tmp_links);
1811 struct cgroup *root_cgrp = &root->cgrp;
1812 struct css_set *cset;
1815 lockdep_assert_held(&cgroup_mutex);
1817 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1820 root_cgrp->id = ret;
1822 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1828 * We're accessing css_set_count without locking css_set_rwsem here,
1829 * but that's OK - it can only be increased by someone holding
1830 * cgroup_lock, and that's us. The worst that can happen is that we
1831 * have some link structures left over
1833 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1837 ret = cgroup_init_root_id(root);
1841 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1842 KERNFS_ROOT_CREATE_DEACTIVATED,
1844 if (IS_ERR(root->kf_root)) {
1845 ret = PTR_ERR(root->kf_root);
1848 root_cgrp->kn = root->kf_root->kn;
1850 ret = css_populate_dir(&root_cgrp->self, NULL);
1854 ret = rebind_subsystems(root, ss_mask);
1859 * There must be no failure case after here, since rebinding takes
1860 * care of subsystems' refcounts, which are explicitly dropped in
1861 * the failure exit path.
1863 list_add(&root->root_list, &cgroup_roots);
1864 cgroup_root_count++;
1867 * Link the root cgroup in this hierarchy into all the css_set
1870 down_write(&css_set_rwsem);
1871 hash_for_each(css_set_table, i, cset, hlist)
1872 link_css_set(&tmp_links, cset, root_cgrp);
1873 up_write(&css_set_rwsem);
1875 BUG_ON(!list_empty(&root_cgrp->self.children));
1876 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1878 kernfs_activate(root_cgrp->kn);
1883 kernfs_destroy_root(root->kf_root);
1884 root->kf_root = NULL;
1886 cgroup_exit_root_id(root);
1888 percpu_ref_exit(&root_cgrp->self.refcnt);
1890 free_cgrp_cset_links(&tmp_links);
1894 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1895 int flags, const char *unused_dev_name,
1898 struct super_block *pinned_sb = NULL;
1899 struct cgroup_subsys *ss;
1900 struct cgroup_root *root;
1901 struct cgroup_sb_opts opts;
1902 struct dentry *dentry;
1908 * The first time anyone tries to mount a cgroup, enable the list
1909 * linking each css_set to its tasks and fix up all existing tasks.
1911 if (!use_task_css_set_links)
1912 cgroup_enable_task_cg_lists();
1914 mutex_lock(&cgroup_mutex);
1916 /* First find the desired set of subsystems */
1917 ret = parse_cgroupfs_options(data, &opts);
1921 /* look for a matching existing root */
1922 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1923 cgrp_dfl_root_visible = true;
1924 root = &cgrp_dfl_root;
1925 cgroup_get(&root->cgrp);
1931 * Destruction of cgroup root is asynchronous, so subsystems may
1932 * still be dying after the previous unmount. Let's drain the
1933 * dying subsystems. We just need to ensure that the ones
1934 * unmounted previously finish dying and don't care about new ones
1935 * starting. Testing ref liveliness is good enough.
1937 for_each_subsys(ss, i) {
1938 if (!(opts.subsys_mask & (1 << i)) ||
1939 ss->root == &cgrp_dfl_root)
1942 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1943 mutex_unlock(&cgroup_mutex);
1945 ret = restart_syscall();
1948 cgroup_put(&ss->root->cgrp);
1951 for_each_root(root) {
1952 bool name_match = false;
1954 if (root == &cgrp_dfl_root)
1958 * If we asked for a name then it must match. Also, if
1959 * name matches but sybsys_mask doesn't, we should fail.
1960 * Remember whether name matched.
1963 if (strcmp(opts.name, root->name))
1969 * If we asked for subsystems (or explicitly for no
1970 * subsystems) then they must match.
1972 if ((opts.subsys_mask || opts.none) &&
1973 (opts.subsys_mask != root->subsys_mask)) {
1980 if (root->flags ^ opts.flags)
1981 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1984 * We want to reuse @root whose lifetime is governed by its
1985 * ->cgrp. Let's check whether @root is alive and keep it
1986 * that way. As cgroup_kill_sb() can happen anytime, we
1987 * want to block it by pinning the sb so that @root doesn't
1988 * get killed before mount is complete.
1990 * With the sb pinned, tryget_live can reliably indicate
1991 * whether @root can be reused. If it's being killed,
1992 * drain it. We can use wait_queue for the wait but this
1993 * path is super cold. Let's just sleep a bit and retry.
1995 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1996 if (IS_ERR(pinned_sb) ||
1997 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1998 mutex_unlock(&cgroup_mutex);
1999 if (!IS_ERR_OR_NULL(pinned_sb))
2000 deactivate_super(pinned_sb);
2002 ret = restart_syscall();
2011 * No such thing, create a new one. name= matching without subsys
2012 * specification is allowed for already existing hierarchies but we
2013 * can't create new one without subsys specification.
2015 if (!opts.subsys_mask && !opts.none) {
2020 root = kzalloc(sizeof(*root), GFP_KERNEL);
2026 init_cgroup_root(root, &opts);
2028 ret = cgroup_setup_root(root, opts.subsys_mask);
2030 cgroup_free_root(root);
2033 mutex_unlock(&cgroup_mutex);
2035 kfree(opts.release_agent);
2039 return ERR_PTR(ret);
2041 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2042 CGROUP_SUPER_MAGIC, &new_sb);
2043 if (IS_ERR(dentry) || !new_sb)
2044 cgroup_put(&root->cgrp);
2047 * If @pinned_sb, we're reusing an existing root and holding an
2048 * extra ref on its sb. Mount is complete. Put the extra ref.
2052 deactivate_super(pinned_sb);
2058 static void cgroup_kill_sb(struct super_block *sb)
2060 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2061 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2064 * If @root doesn't have any mounts or children, start killing it.
2065 * This prevents new mounts by disabling percpu_ref_tryget_live().
2066 * cgroup_mount() may wait for @root's release.
2068 * And don't kill the default root.
2070 if (!list_empty(&root->cgrp.self.children) ||
2071 root == &cgrp_dfl_root)
2072 cgroup_put(&root->cgrp);
2074 percpu_ref_kill(&root->cgrp.self.refcnt);
2079 static struct file_system_type cgroup_fs_type = {
2081 .mount = cgroup_mount,
2082 .kill_sb = cgroup_kill_sb,
2086 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2087 * @task: target task
2088 * @buf: the buffer to write the path into
2089 * @buflen: the length of the buffer
2091 * Determine @task's cgroup on the first (the one with the lowest non-zero
2092 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2093 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2094 * cgroup controller callbacks.
2096 * Return value is the same as kernfs_path().
2098 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2100 struct cgroup_root *root;
2101 struct cgroup *cgrp;
2102 int hierarchy_id = 1;
2105 mutex_lock(&cgroup_mutex);
2106 down_read(&css_set_rwsem);
2108 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2111 cgrp = task_cgroup_from_root(task, root);
2112 path = cgroup_path(cgrp, buf, buflen);
2114 /* if no hierarchy exists, everyone is in "/" */
2115 if (strlcpy(buf, "/", buflen) < buflen)
2119 up_read(&css_set_rwsem);
2120 mutex_unlock(&cgroup_mutex);
2123 EXPORT_SYMBOL_GPL(task_cgroup_path);
2125 /* used to track tasks and other necessary states during migration */
2126 struct cgroup_taskset {
2127 /* the src and dst cset list running through cset->mg_node */
2128 struct list_head src_csets;
2129 struct list_head dst_csets;
2132 * Fields for cgroup_taskset_*() iteration.
2134 * Before migration is committed, the target migration tasks are on
2135 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2136 * the csets on ->dst_csets. ->csets point to either ->src_csets
2137 * or ->dst_csets depending on whether migration is committed.
2139 * ->cur_csets and ->cur_task point to the current task position
2142 struct list_head *csets;
2143 struct css_set *cur_cset;
2144 struct task_struct *cur_task;
2148 * cgroup_taskset_first - reset taskset and return the first task
2149 * @tset: taskset of interest
2151 * @tset iteration is initialized and the first task is returned.
2153 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2155 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2156 tset->cur_task = NULL;
2158 return cgroup_taskset_next(tset);
2162 * cgroup_taskset_next - iterate to the next task in taskset
2163 * @tset: taskset of interest
2165 * Return the next task in @tset. Iteration must have been initialized
2166 * with cgroup_taskset_first().
2168 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2170 struct css_set *cset = tset->cur_cset;
2171 struct task_struct *task = tset->cur_task;
2173 while (&cset->mg_node != tset->csets) {
2175 task = list_first_entry(&cset->mg_tasks,
2176 struct task_struct, cg_list);
2178 task = list_next_entry(task, cg_list);
2180 if (&task->cg_list != &cset->mg_tasks) {
2181 tset->cur_cset = cset;
2182 tset->cur_task = task;
2186 cset = list_next_entry(cset, mg_node);
2194 * cgroup_task_migrate - move a task from one cgroup to another.
2195 * @old_cgrp: the cgroup @tsk is being migrated from
2196 * @tsk: the task being migrated
2197 * @new_cset: the new css_set @tsk is being attached to
2199 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2201 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2202 struct task_struct *tsk,
2203 struct css_set *new_cset)
2205 struct css_set *old_cset;
2207 lockdep_assert_held(&cgroup_mutex);
2208 lockdep_assert_held(&css_set_rwsem);
2211 * We are synchronized through cgroup_threadgroup_rwsem against
2212 * PF_EXITING setting such that we can't race against cgroup_exit()
2213 * changing the css_set to init_css_set and dropping the old one.
2215 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2216 old_cset = task_css_set(tsk);
2218 get_css_set(new_cset);
2219 rcu_assign_pointer(tsk->cgroups, new_cset);
2220 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2223 * We just gained a reference on old_cset by taking it from the
2224 * task. As trading it for new_cset is protected by cgroup_mutex,
2225 * we're safe to drop it here; it will be freed under RCU.
2227 put_css_set_locked(old_cset);
2231 * cgroup_migrate_finish - cleanup after attach
2232 * @preloaded_csets: list of preloaded css_sets
2234 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2235 * those functions for details.
2237 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2239 struct css_set *cset, *tmp_cset;
2241 lockdep_assert_held(&cgroup_mutex);
2243 down_write(&css_set_rwsem);
2244 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2245 cset->mg_src_cgrp = NULL;
2246 cset->mg_dst_cset = NULL;
2247 list_del_init(&cset->mg_preload_node);
2248 put_css_set_locked(cset);
2250 up_write(&css_set_rwsem);
2254 * cgroup_migrate_add_src - add a migration source css_set
2255 * @src_cset: the source css_set to add
2256 * @dst_cgrp: the destination cgroup
2257 * @preloaded_csets: list of preloaded css_sets
2259 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2260 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2261 * up by cgroup_migrate_finish().
2263 * This function may be called without holding cgroup_threadgroup_rwsem
2264 * even if the target is a process. Threads may be created and destroyed
2265 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2266 * into play and the preloaded css_sets are guaranteed to cover all
2269 static void cgroup_migrate_add_src(struct css_set *src_cset,
2270 struct cgroup *dst_cgrp,
2271 struct list_head *preloaded_csets)
2273 struct cgroup *src_cgrp;
2275 lockdep_assert_held(&cgroup_mutex);
2276 lockdep_assert_held(&css_set_rwsem);
2278 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2280 if (!list_empty(&src_cset->mg_preload_node))
2283 WARN_ON(src_cset->mg_src_cgrp);
2284 WARN_ON(!list_empty(&src_cset->mg_tasks));
2285 WARN_ON(!list_empty(&src_cset->mg_node));
2287 src_cset->mg_src_cgrp = src_cgrp;
2288 get_css_set(src_cset);
2289 list_add(&src_cset->mg_preload_node, preloaded_csets);
2293 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2294 * @dst_cgrp: the destination cgroup (may be %NULL)
2295 * @preloaded_csets: list of preloaded source css_sets
2297 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2298 * have been preloaded to @preloaded_csets. This function looks up and
2299 * pins all destination css_sets, links each to its source, and append them
2300 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2301 * source css_set is assumed to be its cgroup on the default hierarchy.
2303 * This function must be called after cgroup_migrate_add_src() has been
2304 * called on each migration source css_set. After migration is performed
2305 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2308 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2309 struct list_head *preloaded_csets)
2312 struct css_set *src_cset, *tmp_cset;
2314 lockdep_assert_held(&cgroup_mutex);
2317 * Except for the root, child_subsys_mask must be zero for a cgroup
2318 * with tasks so that child cgroups don't compete against tasks.
2320 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2321 dst_cgrp->child_subsys_mask)
2324 /* look up the dst cset for each src cset and link it to src */
2325 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2326 struct css_set *dst_cset;
2328 dst_cset = find_css_set(src_cset,
2329 dst_cgrp ?: src_cset->dfl_cgrp);
2333 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2336 * If src cset equals dst, it's noop. Drop the src.
2337 * cgroup_migrate() will skip the cset too. Note that we
2338 * can't handle src == dst as some nodes are used by both.
2340 if (src_cset == dst_cset) {
2341 src_cset->mg_src_cgrp = NULL;
2342 list_del_init(&src_cset->mg_preload_node);
2343 put_css_set(src_cset);
2344 put_css_set(dst_cset);
2348 src_cset->mg_dst_cset = dst_cset;
2350 if (list_empty(&dst_cset->mg_preload_node))
2351 list_add(&dst_cset->mg_preload_node, &csets);
2353 put_css_set(dst_cset);
2356 list_splice_tail(&csets, preloaded_csets);
2359 cgroup_migrate_finish(&csets);
2364 * cgroup_migrate - migrate a process or task to a cgroup
2365 * @cgrp: the destination cgroup
2366 * @leader: the leader of the process or the task to migrate
2367 * @threadgroup: whether @leader points to the whole process or a single task
2369 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2370 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2371 * caller is also responsible for invoking cgroup_migrate_add_src() and
2372 * cgroup_migrate_prepare_dst() on the targets before invoking this
2373 * function and following up with cgroup_migrate_finish().
2375 * As long as a controller's ->can_attach() doesn't fail, this function is
2376 * guaranteed to succeed. This means that, excluding ->can_attach()
2377 * failure, when migrating multiple targets, the success or failure can be
2378 * decided for all targets by invoking group_migrate_prepare_dst() before
2379 * actually starting migrating.
2381 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2384 struct cgroup_taskset tset = {
2385 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2386 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2387 .csets = &tset.src_csets,
2389 struct cgroup_subsys_state *css, *failed_css = NULL;
2390 struct css_set *cset, *tmp_cset;
2391 struct task_struct *task, *tmp_task;
2395 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2396 * already PF_EXITING could be freed from underneath us unless we
2397 * take an rcu_read_lock.
2399 down_write(&css_set_rwsem);
2403 /* @task either already exited or can't exit until the end */
2404 if (task->flags & PF_EXITING)
2407 /* leave @task alone if post_fork() hasn't linked it yet */
2408 if (list_empty(&task->cg_list))
2411 cset = task_css_set(task);
2412 if (!cset->mg_src_cgrp)
2415 list_move_tail(&task->cg_list, &cset->mg_tasks);
2416 if (list_empty(&cset->mg_node))
2417 list_add_tail(&cset->mg_node, &tset.src_csets);
2418 if (list_empty(&cset->mg_dst_cset->mg_node))
2419 list_move_tail(&cset->mg_dst_cset->mg_node,
2424 } while_each_thread(leader, task);
2426 up_write(&css_set_rwsem);
2428 /* methods shouldn't be called if no task is actually migrating */
2429 if (list_empty(&tset.src_csets))
2432 /* check that we can legitimately attach to the cgroup */
2433 for_each_e_css(css, i, cgrp) {
2434 if (css->ss->can_attach) {
2435 ret = css->ss->can_attach(css, &tset);
2438 goto out_cancel_attach;
2444 * Now that we're guaranteed success, proceed to move all tasks to
2445 * the new cgroup. There are no failure cases after here, so this
2446 * is the commit point.
2448 down_write(&css_set_rwsem);
2449 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2450 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2451 cgroup_task_migrate(cset->mg_src_cgrp, task,
2454 up_write(&css_set_rwsem);
2457 * Migration is committed, all target tasks are now on dst_csets.
2458 * Nothing is sensitive to fork() after this point. Notify
2459 * controllers that migration is complete.
2461 tset.csets = &tset.dst_csets;
2463 for_each_e_css(css, i, cgrp)
2464 if (css->ss->attach)
2465 css->ss->attach(css, &tset);
2468 goto out_release_tset;
2471 for_each_e_css(css, i, cgrp) {
2472 if (css == failed_css)
2474 if (css->ss->cancel_attach)
2475 css->ss->cancel_attach(css, &tset);
2478 down_write(&css_set_rwsem);
2479 list_splice_init(&tset.dst_csets, &tset.src_csets);
2480 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2481 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2482 list_del_init(&cset->mg_node);
2484 up_write(&css_set_rwsem);
2489 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2490 * @dst_cgrp: the cgroup to attach to
2491 * @leader: the task or the leader of the threadgroup to be attached
2492 * @threadgroup: attach the whole threadgroup?
2494 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2496 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2497 struct task_struct *leader, bool threadgroup)
2499 LIST_HEAD(preloaded_csets);
2500 struct task_struct *task;
2503 /* look up all src csets */
2504 down_read(&css_set_rwsem);
2508 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2512 } while_each_thread(leader, task);
2514 up_read(&css_set_rwsem);
2516 /* prepare dst csets and commit */
2517 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2519 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2521 cgroup_migrate_finish(&preloaded_csets);
2525 static int cgroup_procs_write_permission(struct task_struct *task,
2526 struct cgroup *dst_cgrp,
2527 struct kernfs_open_file *of)
2529 const struct cred *cred = current_cred();
2530 const struct cred *tcred = get_task_cred(task);
2534 * even if we're attaching all tasks in the thread group, we only
2535 * need to check permissions on one of them.
2537 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2538 !uid_eq(cred->euid, tcred->uid) &&
2539 !uid_eq(cred->euid, tcred->suid))
2542 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2543 struct super_block *sb = of->file->f_path.dentry->d_sb;
2544 struct cgroup *cgrp;
2545 struct inode *inode;
2547 down_read(&css_set_rwsem);
2548 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2549 up_read(&css_set_rwsem);
2551 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2552 cgrp = cgroup_parent(cgrp);
2555 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2557 ret = inode_permission(inode, MAY_WRITE);
2567 * Find the task_struct of the task to attach by vpid and pass it along to the
2568 * function to attach either it or all tasks in its threadgroup. Will lock
2569 * cgroup_mutex and threadgroup.
2571 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2572 size_t nbytes, loff_t off, bool threadgroup)
2574 struct task_struct *tsk;
2575 struct cgroup *cgrp;
2579 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2582 cgrp = cgroup_kn_lock_live(of->kn);
2586 percpu_down_write(&cgroup_threadgroup_rwsem);
2589 tsk = find_task_by_vpid(pid);
2592 goto out_unlock_rcu;
2599 tsk = tsk->group_leader;
2602 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2603 * trapped in a cpuset, or RT worker may be born in a cgroup
2604 * with no rt_runtime allocated. Just say no.
2606 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2608 goto out_unlock_rcu;
2611 get_task_struct(tsk);
2614 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2616 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2618 put_task_struct(tsk);
2619 goto out_unlock_threadgroup;
2623 out_unlock_threadgroup:
2624 percpu_up_write(&cgroup_threadgroup_rwsem);
2625 cgroup_kn_unlock(of->kn);
2626 return ret ?: nbytes;
2630 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2631 * @from: attach to all cgroups of a given task
2632 * @tsk: the task to be attached
2634 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2636 struct cgroup_root *root;
2639 mutex_lock(&cgroup_mutex);
2640 for_each_root(root) {
2641 struct cgroup *from_cgrp;
2643 if (root == &cgrp_dfl_root)
2646 down_read(&css_set_rwsem);
2647 from_cgrp = task_cgroup_from_root(from, root);
2648 up_read(&css_set_rwsem);
2650 retval = cgroup_attach_task(from_cgrp, tsk, false);
2654 mutex_unlock(&cgroup_mutex);
2658 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2660 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2661 char *buf, size_t nbytes, loff_t off)
2663 return __cgroup_procs_write(of, buf, nbytes, off, false);
2666 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2667 char *buf, size_t nbytes, loff_t off)
2669 return __cgroup_procs_write(of, buf, nbytes, off, true);
2672 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2673 char *buf, size_t nbytes, loff_t off)
2675 struct cgroup *cgrp;
2677 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2679 cgrp = cgroup_kn_lock_live(of->kn);
2682 spin_lock(&release_agent_path_lock);
2683 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2684 sizeof(cgrp->root->release_agent_path));
2685 spin_unlock(&release_agent_path_lock);
2686 cgroup_kn_unlock(of->kn);
2690 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2692 struct cgroup *cgrp = seq_css(seq)->cgroup;
2694 spin_lock(&release_agent_path_lock);
2695 seq_puts(seq, cgrp->root->release_agent_path);
2696 spin_unlock(&release_agent_path_lock);
2697 seq_putc(seq, '\n');
2701 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2703 seq_puts(seq, "0\n");
2707 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2709 struct cgroup_subsys *ss;
2710 bool printed = false;
2713 for_each_subsys_which(ss, ssid, &ss_mask) {
2716 seq_printf(seq, "%s", ss->name);
2720 seq_putc(seq, '\n');
2723 /* show controllers which are currently attached to the default hierarchy */
2724 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2726 struct cgroup *cgrp = seq_css(seq)->cgroup;
2728 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2729 ~cgrp_dfl_root_inhibit_ss_mask);
2733 /* show controllers which are enabled from the parent */
2734 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2736 struct cgroup *cgrp = seq_css(seq)->cgroup;
2738 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2742 /* show controllers which are enabled for a given cgroup's children */
2743 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2745 struct cgroup *cgrp = seq_css(seq)->cgroup;
2747 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2752 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2753 * @cgrp: root of the subtree to update csses for
2755 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2756 * css associations need to be updated accordingly. This function looks up
2757 * all css_sets which are attached to the subtree, creates the matching
2758 * updated css_sets and migrates the tasks to the new ones.
2760 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2762 LIST_HEAD(preloaded_csets);
2763 struct cgroup_subsys_state *css;
2764 struct css_set *src_cset;
2767 lockdep_assert_held(&cgroup_mutex);
2769 percpu_down_write(&cgroup_threadgroup_rwsem);
2771 /* look up all csses currently attached to @cgrp's subtree */
2772 down_read(&css_set_rwsem);
2773 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2774 struct cgrp_cset_link *link;
2776 /* self is not affected by child_subsys_mask change */
2777 if (css->cgroup == cgrp)
2780 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2781 cgroup_migrate_add_src(link->cset, cgrp,
2784 up_read(&css_set_rwsem);
2786 /* NULL dst indicates self on default hierarchy */
2787 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2791 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2792 struct task_struct *last_task = NULL, *task;
2794 /* src_csets precede dst_csets, break on the first dst_cset */
2795 if (!src_cset->mg_src_cgrp)
2799 * All tasks in src_cset need to be migrated to the
2800 * matching dst_cset. Empty it process by process. We
2801 * walk tasks but migrate processes. The leader might even
2802 * belong to a different cset but such src_cset would also
2803 * be among the target src_csets because the default
2804 * hierarchy enforces per-process membership.
2807 down_read(&css_set_rwsem);
2808 task = list_first_entry_or_null(&src_cset->tasks,
2809 struct task_struct, cg_list);
2811 task = task->group_leader;
2812 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2813 get_task_struct(task);
2815 up_read(&css_set_rwsem);
2820 /* guard against possible infinite loop */
2821 if (WARN(last_task == task,
2822 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2826 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2828 put_task_struct(task);
2830 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2836 cgroup_migrate_finish(&preloaded_csets);
2837 percpu_up_write(&cgroup_threadgroup_rwsem);
2841 /* change the enabled child controllers for a cgroup in the default hierarchy */
2842 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2843 char *buf, size_t nbytes,
2846 unsigned long enable = 0, disable = 0;
2847 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2848 struct cgroup *cgrp, *child;
2849 struct cgroup_subsys *ss;
2854 * Parse input - space separated list of subsystem names prefixed
2855 * with either + or -.
2857 buf = strstrip(buf);
2858 while ((tok = strsep(&buf, " "))) {
2859 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2863 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2864 if (!cgroup_ssid_enabled(ssid) ||
2865 strcmp(tok + 1, ss->name))
2869 enable |= 1 << ssid;
2870 disable &= ~(1 << ssid);
2871 } else if (*tok == '-') {
2872 disable |= 1 << ssid;
2873 enable &= ~(1 << ssid);
2879 if (ssid == CGROUP_SUBSYS_COUNT)
2883 cgrp = cgroup_kn_lock_live(of->kn);
2887 for_each_subsys(ss, ssid) {
2888 if (enable & (1 << ssid)) {
2889 if (cgrp->subtree_control & (1 << ssid)) {
2890 enable &= ~(1 << ssid);
2894 /* unavailable or not enabled on the parent? */
2895 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2896 (cgroup_parent(cgrp) &&
2897 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2901 } else if (disable & (1 << ssid)) {
2902 if (!(cgrp->subtree_control & (1 << ssid))) {
2903 disable &= ~(1 << ssid);
2907 /* a child has it enabled? */
2908 cgroup_for_each_live_child(child, cgrp) {
2909 if (child->subtree_control & (1 << ssid)) {
2917 if (!enable && !disable) {
2923 * Except for the root, subtree_control must be zero for a cgroup
2924 * with tasks so that child cgroups don't compete against tasks.
2926 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2932 * Update subsys masks and calculate what needs to be done. More
2933 * subsystems than specified may need to be enabled or disabled
2934 * depending on subsystem dependencies.
2936 old_sc = cgrp->subtree_control;
2937 old_ss = cgrp->child_subsys_mask;
2938 new_sc = (old_sc | enable) & ~disable;
2939 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2941 css_enable = ~old_ss & new_ss;
2942 css_disable = old_ss & ~new_ss;
2943 enable |= css_enable;
2944 disable |= css_disable;
2947 * Because css offlining is asynchronous, userland might try to
2948 * re-enable the same controller while the previous instance is
2949 * still around. In such cases, wait till it's gone using
2952 for_each_subsys_which(ss, ssid, &css_enable) {
2953 cgroup_for_each_live_child(child, cgrp) {
2956 if (!cgroup_css(child, ss))
2960 prepare_to_wait(&child->offline_waitq, &wait,
2961 TASK_UNINTERRUPTIBLE);
2962 cgroup_kn_unlock(of->kn);
2964 finish_wait(&child->offline_waitq, &wait);
2967 return restart_syscall();
2971 cgrp->subtree_control = new_sc;
2972 cgrp->child_subsys_mask = new_ss;
2975 * Create new csses or make the existing ones visible. A css is
2976 * created invisible if it's being implicitly enabled through
2977 * dependency. An invisible css is made visible when the userland
2978 * explicitly enables it.
2980 for_each_subsys(ss, ssid) {
2981 if (!(enable & (1 << ssid)))
2984 cgroup_for_each_live_child(child, cgrp) {
2985 if (css_enable & (1 << ssid))
2986 ret = create_css(child, ss,
2987 cgrp->subtree_control & (1 << ssid));
2989 ret = css_populate_dir(cgroup_css(child, ss),
2997 * At this point, cgroup_e_css() results reflect the new csses
2998 * making the following cgroup_update_dfl_csses() properly update
2999 * css associations of all tasks in the subtree.
3001 ret = cgroup_update_dfl_csses(cgrp);
3006 * All tasks are migrated out of disabled csses. Kill or hide
3007 * them. A css is hidden when the userland requests it to be
3008 * disabled while other subsystems are still depending on it. The
3009 * css must not actively control resources and be in the vanilla
3010 * state if it's made visible again later. Controllers which may
3011 * be depended upon should provide ->css_reset() for this purpose.
3013 for_each_subsys(ss, ssid) {
3014 if (!(disable & (1 << ssid)))
3017 cgroup_for_each_live_child(child, cgrp) {
3018 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3020 if (css_disable & (1 << ssid)) {
3023 css_clear_dir(css, NULL);
3031 * The effective csses of all the descendants (excluding @cgrp) may
3032 * have changed. Subsystems can optionally subscribe to this event
3033 * by implementing ->css_e_css_changed() which is invoked if any of
3034 * the effective csses seen from the css's cgroup may have changed.
3036 for_each_subsys(ss, ssid) {
3037 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3038 struct cgroup_subsys_state *css;
3040 if (!ss->css_e_css_changed || !this_css)
3043 css_for_each_descendant_pre(css, this_css)
3044 if (css != this_css)
3045 ss->css_e_css_changed(css);
3048 kernfs_activate(cgrp->kn);
3051 cgroup_kn_unlock(of->kn);
3052 return ret ?: nbytes;
3055 cgrp->subtree_control = old_sc;
3056 cgrp->child_subsys_mask = old_ss;
3058 for_each_subsys(ss, ssid) {
3059 if (!(enable & (1 << ssid)))
3062 cgroup_for_each_live_child(child, cgrp) {
3063 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3068 if (css_enable & (1 << ssid))
3071 css_clear_dir(css, NULL);
3077 static int cgroup_events_show(struct seq_file *seq, void *v)
3079 seq_printf(seq, "populated %d\n",
3080 (bool)seq_css(seq)->cgroup->populated_cnt);
3084 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3085 size_t nbytes, loff_t off)
3087 struct cgroup *cgrp = of->kn->parent->priv;
3088 struct cftype *cft = of->kn->priv;
3089 struct cgroup_subsys_state *css;
3093 return cft->write(of, buf, nbytes, off);
3096 * kernfs guarantees that a file isn't deleted with operations in
3097 * flight, which means that the matching css is and stays alive and
3098 * doesn't need to be pinned. The RCU locking is not necessary
3099 * either. It's just for the convenience of using cgroup_css().
3102 css = cgroup_css(cgrp, cft->ss);
3105 if (cft->write_u64) {
3106 unsigned long long v;
3107 ret = kstrtoull(buf, 0, &v);
3109 ret = cft->write_u64(css, cft, v);
3110 } else if (cft->write_s64) {
3112 ret = kstrtoll(buf, 0, &v);
3114 ret = cft->write_s64(css, cft, v);
3119 return ret ?: nbytes;
3122 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3124 return seq_cft(seq)->seq_start(seq, ppos);
3127 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3129 return seq_cft(seq)->seq_next(seq, v, ppos);
3132 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3134 seq_cft(seq)->seq_stop(seq, v);
3137 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3139 struct cftype *cft = seq_cft(m);
3140 struct cgroup_subsys_state *css = seq_css(m);
3143 return cft->seq_show(m, arg);
3146 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3147 else if (cft->read_s64)
3148 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3154 static struct kernfs_ops cgroup_kf_single_ops = {
3155 .atomic_write_len = PAGE_SIZE,
3156 .write = cgroup_file_write,
3157 .seq_show = cgroup_seqfile_show,
3160 static struct kernfs_ops cgroup_kf_ops = {
3161 .atomic_write_len = PAGE_SIZE,
3162 .write = cgroup_file_write,
3163 .seq_start = cgroup_seqfile_start,
3164 .seq_next = cgroup_seqfile_next,
3165 .seq_stop = cgroup_seqfile_stop,
3166 .seq_show = cgroup_seqfile_show,
3170 * cgroup_rename - Only allow simple rename of directories in place.
3172 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3173 const char *new_name_str)
3175 struct cgroup *cgrp = kn->priv;
3178 if (kernfs_type(kn) != KERNFS_DIR)
3180 if (kn->parent != new_parent)
3184 * This isn't a proper migration and its usefulness is very
3185 * limited. Disallow on the default hierarchy.
3187 if (cgroup_on_dfl(cgrp))
3191 * We're gonna grab cgroup_mutex which nests outside kernfs
3192 * active_ref. kernfs_rename() doesn't require active_ref
3193 * protection. Break them before grabbing cgroup_mutex.
3195 kernfs_break_active_protection(new_parent);
3196 kernfs_break_active_protection(kn);
3198 mutex_lock(&cgroup_mutex);
3200 ret = kernfs_rename(kn, new_parent, new_name_str);
3202 mutex_unlock(&cgroup_mutex);
3204 kernfs_unbreak_active_protection(kn);
3205 kernfs_unbreak_active_protection(new_parent);
3209 /* set uid and gid of cgroup dirs and files to that of the creator */
3210 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3212 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3213 .ia_uid = current_fsuid(),
3214 .ia_gid = current_fsgid(), };
3216 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3217 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3220 return kernfs_setattr(kn, &iattr);
3223 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3226 char name[CGROUP_FILE_NAME_MAX];
3227 struct kernfs_node *kn;
3228 struct lock_class_key *key = NULL;
3231 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3232 key = &cft->lockdep_key;
3234 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3235 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3240 ret = cgroup_kn_set_ugid(kn);
3246 if (cft->file_offset) {
3247 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3251 list_add(&cfile->node, &css->files);
3258 * cgroup_addrm_files - add or remove files to a cgroup directory
3259 * @css: the target css
3260 * @cgrp: the target cgroup (usually css->cgroup)
3261 * @cfts: array of cftypes to be added
3262 * @is_add: whether to add or remove
3264 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3265 * For removals, this function never fails.
3267 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3268 struct cgroup *cgrp, struct cftype cfts[],
3271 struct cftype *cft, *cft_end = NULL;
3274 lockdep_assert_held(&cgroup_mutex);
3277 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3278 /* does cft->flags tell us to skip this file on @cgrp? */
3279 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3281 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3283 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3285 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3289 ret = cgroup_add_file(css, cgrp, cft);
3291 pr_warn("%s: failed to add %s, err=%d\n",
3292 __func__, cft->name, ret);
3298 cgroup_rm_file(cgrp, cft);
3304 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3307 struct cgroup_subsys *ss = cfts[0].ss;
3308 struct cgroup *root = &ss->root->cgrp;
3309 struct cgroup_subsys_state *css;
3312 lockdep_assert_held(&cgroup_mutex);
3314 /* add/rm files for all cgroups created before */
3315 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3316 struct cgroup *cgrp = css->cgroup;
3318 if (cgroup_is_dead(cgrp))
3321 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3327 kernfs_activate(root->kn);
3331 static void cgroup_exit_cftypes(struct cftype *cfts)
3335 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3336 /* free copy for custom atomic_write_len, see init_cftypes() */
3337 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3342 /* revert flags set by cgroup core while adding @cfts */
3343 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3347 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3351 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3352 struct kernfs_ops *kf_ops;
3354 WARN_ON(cft->ss || cft->kf_ops);
3357 kf_ops = &cgroup_kf_ops;
3359 kf_ops = &cgroup_kf_single_ops;
3362 * Ugh... if @cft wants a custom max_write_len, we need to
3363 * make a copy of kf_ops to set its atomic_write_len.
3365 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3366 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3368 cgroup_exit_cftypes(cfts);
3371 kf_ops->atomic_write_len = cft->max_write_len;
3374 cft->kf_ops = kf_ops;
3381 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3383 lockdep_assert_held(&cgroup_mutex);
3385 if (!cfts || !cfts[0].ss)
3388 list_del(&cfts->node);
3389 cgroup_apply_cftypes(cfts, false);
3390 cgroup_exit_cftypes(cfts);
3395 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3396 * @cfts: zero-length name terminated array of cftypes
3398 * Unregister @cfts. Files described by @cfts are removed from all
3399 * existing cgroups and all future cgroups won't have them either. This
3400 * function can be called anytime whether @cfts' subsys is attached or not.
3402 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3405 int cgroup_rm_cftypes(struct cftype *cfts)
3409 mutex_lock(&cgroup_mutex);
3410 ret = cgroup_rm_cftypes_locked(cfts);
3411 mutex_unlock(&cgroup_mutex);
3416 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3417 * @ss: target cgroup subsystem
3418 * @cfts: zero-length name terminated array of cftypes
3420 * Register @cfts to @ss. Files described by @cfts are created for all
3421 * existing cgroups to which @ss is attached and all future cgroups will
3422 * have them too. This function can be called anytime whether @ss is
3425 * Returns 0 on successful registration, -errno on failure. Note that this
3426 * function currently returns 0 as long as @cfts registration is successful
3427 * even if some file creation attempts on existing cgroups fail.
3429 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3433 if (!cgroup_ssid_enabled(ss->id))
3436 if (!cfts || cfts[0].name[0] == '\0')
3439 ret = cgroup_init_cftypes(ss, cfts);
3443 mutex_lock(&cgroup_mutex);
3445 list_add_tail(&cfts->node, &ss->cfts);
3446 ret = cgroup_apply_cftypes(cfts, true);
3448 cgroup_rm_cftypes_locked(cfts);
3450 mutex_unlock(&cgroup_mutex);
3455 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3456 * @ss: target cgroup subsystem
3457 * @cfts: zero-length name terminated array of cftypes
3459 * Similar to cgroup_add_cftypes() but the added files are only used for
3460 * the default hierarchy.
3462 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3466 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3467 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3468 return cgroup_add_cftypes(ss, cfts);
3472 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3473 * @ss: target cgroup subsystem
3474 * @cfts: zero-length name terminated array of cftypes
3476 * Similar to cgroup_add_cftypes() but the added files are only used for
3477 * the legacy hierarchies.
3479 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3484 * If legacy_flies_on_dfl, we want to show the legacy files on the
3485 * dfl hierarchy but iff the target subsystem hasn't been updated
3486 * for the dfl hierarchy yet.
3488 if (!cgroup_legacy_files_on_dfl ||
3489 ss->dfl_cftypes != ss->legacy_cftypes) {
3490 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3491 cft->flags |= __CFTYPE_NOT_ON_DFL;
3494 return cgroup_add_cftypes(ss, cfts);
3498 * cgroup_task_count - count the number of tasks in a cgroup.
3499 * @cgrp: the cgroup in question
3501 * Return the number of tasks in the cgroup.
3503 static int cgroup_task_count(const struct cgroup *cgrp)
3506 struct cgrp_cset_link *link;
3508 down_read(&css_set_rwsem);
3509 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3510 count += atomic_read(&link->cset->refcount);
3511 up_read(&css_set_rwsem);
3516 * css_next_child - find the next child of a given css
3517 * @pos: the current position (%NULL to initiate traversal)
3518 * @parent: css whose children to walk
3520 * This function returns the next child of @parent and should be called
3521 * under either cgroup_mutex or RCU read lock. The only requirement is
3522 * that @parent and @pos are accessible. The next sibling is guaranteed to
3523 * be returned regardless of their states.
3525 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3526 * css which finished ->css_online() is guaranteed to be visible in the
3527 * future iterations and will stay visible until the last reference is put.
3528 * A css which hasn't finished ->css_online() or already finished
3529 * ->css_offline() may show up during traversal. It's each subsystem's
3530 * responsibility to synchronize against on/offlining.
3532 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3533 struct cgroup_subsys_state *parent)
3535 struct cgroup_subsys_state *next;
3537 cgroup_assert_mutex_or_rcu_locked();
3540 * @pos could already have been unlinked from the sibling list.
3541 * Once a cgroup is removed, its ->sibling.next is no longer
3542 * updated when its next sibling changes. CSS_RELEASED is set when
3543 * @pos is taken off list, at which time its next pointer is valid,
3544 * and, as releases are serialized, the one pointed to by the next
3545 * pointer is guaranteed to not have started release yet. This
3546 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3547 * critical section, the one pointed to by its next pointer is
3548 * guaranteed to not have finished its RCU grace period even if we
3549 * have dropped rcu_read_lock() inbetween iterations.
3551 * If @pos has CSS_RELEASED set, its next pointer can't be
3552 * dereferenced; however, as each css is given a monotonically
3553 * increasing unique serial number and always appended to the
3554 * sibling list, the next one can be found by walking the parent's
3555 * children until the first css with higher serial number than
3556 * @pos's. While this path can be slower, it happens iff iteration
3557 * races against release and the race window is very small.
3560 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3561 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3562 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3564 list_for_each_entry_rcu(next, &parent->children, sibling)
3565 if (next->serial_nr > pos->serial_nr)
3570 * @next, if not pointing to the head, can be dereferenced and is
3573 if (&next->sibling != &parent->children)
3579 * css_next_descendant_pre - find the next descendant for pre-order walk
3580 * @pos: the current position (%NULL to initiate traversal)
3581 * @root: css whose descendants to walk
3583 * To be used by css_for_each_descendant_pre(). Find the next descendant
3584 * to visit for pre-order traversal of @root's descendants. @root is
3585 * included in the iteration and the first node to be visited.
3587 * While this function requires cgroup_mutex or RCU read locking, it
3588 * doesn't require the whole traversal to be contained in a single critical
3589 * section. This function will return the correct next descendant as long
3590 * as both @pos and @root are accessible and @pos is a descendant of @root.
3592 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3593 * css which finished ->css_online() is guaranteed to be visible in the
3594 * future iterations and will stay visible until the last reference is put.
3595 * A css which hasn't finished ->css_online() or already finished
3596 * ->css_offline() may show up during traversal. It's each subsystem's
3597 * responsibility to synchronize against on/offlining.
3599 struct cgroup_subsys_state *
3600 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3601 struct cgroup_subsys_state *root)
3603 struct cgroup_subsys_state *next;
3605 cgroup_assert_mutex_or_rcu_locked();
3607 /* if first iteration, visit @root */
3611 /* visit the first child if exists */
3612 next = css_next_child(NULL, pos);
3616 /* no child, visit my or the closest ancestor's next sibling */
3617 while (pos != root) {
3618 next = css_next_child(pos, pos->parent);
3628 * css_rightmost_descendant - return the rightmost descendant of a css
3629 * @pos: css of interest
3631 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3632 * is returned. This can be used during pre-order traversal to skip
3635 * While this function requires cgroup_mutex or RCU read locking, it
3636 * doesn't require the whole traversal to be contained in a single critical
3637 * section. This function will return the correct rightmost descendant as
3638 * long as @pos is accessible.
3640 struct cgroup_subsys_state *
3641 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3643 struct cgroup_subsys_state *last, *tmp;
3645 cgroup_assert_mutex_or_rcu_locked();
3649 /* ->prev isn't RCU safe, walk ->next till the end */
3651 css_for_each_child(tmp, last)
3658 static struct cgroup_subsys_state *
3659 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3661 struct cgroup_subsys_state *last;
3665 pos = css_next_child(NULL, pos);
3672 * css_next_descendant_post - find the next descendant for post-order walk
3673 * @pos: the current position (%NULL to initiate traversal)
3674 * @root: css whose descendants to walk
3676 * To be used by css_for_each_descendant_post(). Find the next descendant
3677 * to visit for post-order traversal of @root's descendants. @root is
3678 * included in the iteration and the last node to be visited.
3680 * While this function requires cgroup_mutex or RCU read locking, it
3681 * doesn't require the whole traversal to be contained in a single critical
3682 * section. This function will return the correct next descendant as long
3683 * as both @pos and @cgroup are accessible and @pos is a descendant of
3686 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3687 * css which finished ->css_online() is guaranteed to be visible in the
3688 * future iterations and will stay visible until the last reference is put.
3689 * A css which hasn't finished ->css_online() or already finished
3690 * ->css_offline() may show up during traversal. It's each subsystem's
3691 * responsibility to synchronize against on/offlining.
3693 struct cgroup_subsys_state *
3694 css_next_descendant_post(struct cgroup_subsys_state *pos,
3695 struct cgroup_subsys_state *root)
3697 struct cgroup_subsys_state *next;
3699 cgroup_assert_mutex_or_rcu_locked();
3701 /* if first iteration, visit leftmost descendant which may be @root */
3703 return css_leftmost_descendant(root);
3705 /* if we visited @root, we're done */
3709 /* if there's an unvisited sibling, visit its leftmost descendant */
3710 next = css_next_child(pos, pos->parent);
3712 return css_leftmost_descendant(next);
3714 /* no sibling left, visit parent */
3719 * css_has_online_children - does a css have online children
3720 * @css: the target css
3722 * Returns %true if @css has any online children; otherwise, %false. This
3723 * function can be called from any context but the caller is responsible
3724 * for synchronizing against on/offlining as necessary.
3726 bool css_has_online_children(struct cgroup_subsys_state *css)
3728 struct cgroup_subsys_state *child;
3732 css_for_each_child(child, css) {
3733 if (child->flags & CSS_ONLINE) {
3743 * css_advance_task_iter - advance a task itererator to the next css_set
3744 * @it: the iterator to advance
3746 * Advance @it to the next css_set to walk.
3748 static void css_advance_task_iter(struct css_task_iter *it)
3750 struct list_head *l = it->cset_pos;
3751 struct cgrp_cset_link *link;
3752 struct css_set *cset;
3754 /* Advance to the next non-empty css_set */
3757 if (l == it->cset_head) {
3758 it->cset_pos = NULL;
3763 cset = container_of(l, struct css_set,
3764 e_cset_node[it->ss->id]);
3766 link = list_entry(l, struct cgrp_cset_link, cset_link);
3769 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3773 if (!list_empty(&cset->tasks))
3774 it->task_pos = cset->tasks.next;
3776 it->task_pos = cset->mg_tasks.next;
3778 it->tasks_head = &cset->tasks;
3779 it->mg_tasks_head = &cset->mg_tasks;
3783 * css_task_iter_start - initiate task iteration
3784 * @css: the css to walk tasks of
3785 * @it: the task iterator to use
3787 * Initiate iteration through the tasks of @css. The caller can call
3788 * css_task_iter_next() to walk through the tasks until the function
3789 * returns NULL. On completion of iteration, css_task_iter_end() must be
3792 * Note that this function acquires a lock which is released when the
3793 * iteration finishes. The caller can't sleep while iteration is in
3796 void css_task_iter_start(struct cgroup_subsys_state *css,
3797 struct css_task_iter *it)
3798 __acquires(css_set_rwsem)
3800 /* no one should try to iterate before mounting cgroups */
3801 WARN_ON_ONCE(!use_task_css_set_links);
3803 down_read(&css_set_rwsem);
3808 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3810 it->cset_pos = &css->cgroup->cset_links;
3812 it->cset_head = it->cset_pos;
3814 css_advance_task_iter(it);
3818 * css_task_iter_next - return the next task for the iterator
3819 * @it: the task iterator being iterated
3821 * The "next" function for task iteration. @it should have been
3822 * initialized via css_task_iter_start(). Returns NULL when the iteration
3825 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3827 struct task_struct *res;
3828 struct list_head *l = it->task_pos;
3830 /* If the iterator cg is NULL, we have no tasks */
3833 res = list_entry(l, struct task_struct, cg_list);
3836 * Advance iterator to find next entry. cset->tasks is consumed
3837 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3842 if (l == it->tasks_head)
3843 l = it->mg_tasks_head->next;
3845 if (l == it->mg_tasks_head)
3846 css_advance_task_iter(it);
3854 * css_task_iter_end - finish task iteration
3855 * @it: the task iterator to finish
3857 * Finish task iteration started by css_task_iter_start().
3859 void css_task_iter_end(struct css_task_iter *it)
3860 __releases(css_set_rwsem)
3862 up_read(&css_set_rwsem);
3866 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3867 * @to: cgroup to which the tasks will be moved
3868 * @from: cgroup in which the tasks currently reside
3870 * Locking rules between cgroup_post_fork() and the migration path
3871 * guarantee that, if a task is forking while being migrated, the new child
3872 * is guaranteed to be either visible in the source cgroup after the
3873 * parent's migration is complete or put into the target cgroup. No task
3874 * can slip out of migration through forking.
3876 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3878 LIST_HEAD(preloaded_csets);
3879 struct cgrp_cset_link *link;
3880 struct css_task_iter it;
3881 struct task_struct *task;
3884 mutex_lock(&cgroup_mutex);
3886 /* all tasks in @from are being moved, all csets are source */
3887 down_read(&css_set_rwsem);
3888 list_for_each_entry(link, &from->cset_links, cset_link)
3889 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3890 up_read(&css_set_rwsem);
3892 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3897 * Migrate tasks one-by-one until @form is empty. This fails iff
3898 * ->can_attach() fails.
3901 css_task_iter_start(&from->self, &it);
3902 task = css_task_iter_next(&it);
3904 get_task_struct(task);
3905 css_task_iter_end(&it);
3908 ret = cgroup_migrate(to, task, false);
3909 put_task_struct(task);
3911 } while (task && !ret);
3913 cgroup_migrate_finish(&preloaded_csets);
3914 mutex_unlock(&cgroup_mutex);
3919 * Stuff for reading the 'tasks'/'procs' files.
3921 * Reading this file can return large amounts of data if a cgroup has
3922 * *lots* of attached tasks. So it may need several calls to read(),
3923 * but we cannot guarantee that the information we produce is correct
3924 * unless we produce it entirely atomically.
3928 /* which pidlist file are we talking about? */
3929 enum cgroup_filetype {
3935 * A pidlist is a list of pids that virtually represents the contents of one
3936 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3937 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3940 struct cgroup_pidlist {
3942 * used to find which pidlist is wanted. doesn't change as long as
3943 * this particular list stays in the list.
3945 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3948 /* how many elements the above list has */
3950 /* each of these stored in a list by its cgroup */
3951 struct list_head links;
3952 /* pointer to the cgroup we belong to, for list removal purposes */
3953 struct cgroup *owner;
3954 /* for delayed destruction */
3955 struct delayed_work destroy_dwork;
3959 * The following two functions "fix" the issue where there are more pids
3960 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3961 * TODO: replace with a kernel-wide solution to this problem
3963 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3964 static void *pidlist_allocate(int count)
3966 if (PIDLIST_TOO_LARGE(count))
3967 return vmalloc(count * sizeof(pid_t));
3969 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3972 static void pidlist_free(void *p)
3978 * Used to destroy all pidlists lingering waiting for destroy timer. None
3979 * should be left afterwards.
3981 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3983 struct cgroup_pidlist *l, *tmp_l;
3985 mutex_lock(&cgrp->pidlist_mutex);
3986 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3987 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3988 mutex_unlock(&cgrp->pidlist_mutex);
3990 flush_workqueue(cgroup_pidlist_destroy_wq);
3991 BUG_ON(!list_empty(&cgrp->pidlists));
3994 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3996 struct delayed_work *dwork = to_delayed_work(work);
3997 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3999 struct cgroup_pidlist *tofree = NULL;
4001 mutex_lock(&l->owner->pidlist_mutex);
4004 * Destroy iff we didn't get queued again. The state won't change
4005 * as destroy_dwork can only be queued while locked.
4007 if (!delayed_work_pending(dwork)) {
4008 list_del(&l->links);
4009 pidlist_free(l->list);
4010 put_pid_ns(l->key.ns);
4014 mutex_unlock(&l->owner->pidlist_mutex);
4019 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4020 * Returns the number of unique elements.
4022 static int pidlist_uniq(pid_t *list, int length)
4027 * we presume the 0th element is unique, so i starts at 1. trivial
4028 * edge cases first; no work needs to be done for either
4030 if (length == 0 || length == 1)
4032 /* src and dest walk down the list; dest counts unique elements */
4033 for (src = 1; src < length; src++) {
4034 /* find next unique element */
4035 while (list[src] == list[src-1]) {
4040 /* dest always points to where the next unique element goes */
4041 list[dest] = list[src];
4049 * The two pid files - task and cgroup.procs - guaranteed that the result
4050 * is sorted, which forced this whole pidlist fiasco. As pid order is
4051 * different per namespace, each namespace needs differently sorted list,
4052 * making it impossible to use, for example, single rbtree of member tasks
4053 * sorted by task pointer. As pidlists can be fairly large, allocating one
4054 * per open file is dangerous, so cgroup had to implement shared pool of
4055 * pidlists keyed by cgroup and namespace.
4057 * All this extra complexity was caused by the original implementation
4058 * committing to an entirely unnecessary property. In the long term, we
4059 * want to do away with it. Explicitly scramble sort order if on the
4060 * default hierarchy so that no such expectation exists in the new
4063 * Scrambling is done by swapping every two consecutive bits, which is
4064 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4066 static pid_t pid_fry(pid_t pid)
4068 unsigned a = pid & 0x55555555;
4069 unsigned b = pid & 0xAAAAAAAA;
4071 return (a << 1) | (b >> 1);
4074 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4076 if (cgroup_on_dfl(cgrp))
4077 return pid_fry(pid);
4082 static int cmppid(const void *a, const void *b)
4084 return *(pid_t *)a - *(pid_t *)b;
4087 static int fried_cmppid(const void *a, const void *b)
4089 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4092 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4093 enum cgroup_filetype type)
4095 struct cgroup_pidlist *l;
4096 /* don't need task_nsproxy() if we're looking at ourself */
4097 struct pid_namespace *ns = task_active_pid_ns(current);
4099 lockdep_assert_held(&cgrp->pidlist_mutex);
4101 list_for_each_entry(l, &cgrp->pidlists, links)
4102 if (l->key.type == type && l->key.ns == ns)
4108 * find the appropriate pidlist for our purpose (given procs vs tasks)
4109 * returns with the lock on that pidlist already held, and takes care
4110 * of the use count, or returns NULL with no locks held if we're out of
4113 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4114 enum cgroup_filetype type)
4116 struct cgroup_pidlist *l;
4118 lockdep_assert_held(&cgrp->pidlist_mutex);
4120 l = cgroup_pidlist_find(cgrp, type);
4124 /* entry not found; create a new one */
4125 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4129 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4131 /* don't need task_nsproxy() if we're looking at ourself */
4132 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4134 list_add(&l->links, &cgrp->pidlists);
4139 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4141 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4142 struct cgroup_pidlist **lp)
4146 int pid, n = 0; /* used for populating the array */
4147 struct css_task_iter it;
4148 struct task_struct *tsk;
4149 struct cgroup_pidlist *l;
4151 lockdep_assert_held(&cgrp->pidlist_mutex);
4154 * If cgroup gets more users after we read count, we won't have
4155 * enough space - tough. This race is indistinguishable to the
4156 * caller from the case that the additional cgroup users didn't
4157 * show up until sometime later on.
4159 length = cgroup_task_count(cgrp);
4160 array = pidlist_allocate(length);
4163 /* now, populate the array */
4164 css_task_iter_start(&cgrp->self, &it);
4165 while ((tsk = css_task_iter_next(&it))) {
4166 if (unlikely(n == length))
4168 /* get tgid or pid for procs or tasks file respectively */
4169 if (type == CGROUP_FILE_PROCS)
4170 pid = task_tgid_vnr(tsk);
4172 pid = task_pid_vnr(tsk);
4173 if (pid > 0) /* make sure to only use valid results */
4176 css_task_iter_end(&it);
4178 /* now sort & (if procs) strip out duplicates */
4179 if (cgroup_on_dfl(cgrp))
4180 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4182 sort(array, length, sizeof(pid_t), cmppid, NULL);
4183 if (type == CGROUP_FILE_PROCS)
4184 length = pidlist_uniq(array, length);
4186 l = cgroup_pidlist_find_create(cgrp, type);
4188 pidlist_free(array);
4192 /* store array, freeing old if necessary */
4193 pidlist_free(l->list);
4201 * cgroupstats_build - build and fill cgroupstats
4202 * @stats: cgroupstats to fill information into
4203 * @dentry: A dentry entry belonging to the cgroup for which stats have
4206 * Build and fill cgroupstats so that taskstats can export it to user
4209 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4211 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4212 struct cgroup *cgrp;
4213 struct css_task_iter it;
4214 struct task_struct *tsk;
4216 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4217 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4218 kernfs_type(kn) != KERNFS_DIR)
4221 mutex_lock(&cgroup_mutex);
4224 * We aren't being called from kernfs and there's no guarantee on
4225 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4226 * @kn->priv is RCU safe. Let's do the RCU dancing.
4229 cgrp = rcu_dereference(kn->priv);
4230 if (!cgrp || cgroup_is_dead(cgrp)) {
4232 mutex_unlock(&cgroup_mutex);
4237 css_task_iter_start(&cgrp->self, &it);
4238 while ((tsk = css_task_iter_next(&it))) {
4239 switch (tsk->state) {
4241 stats->nr_running++;
4243 case TASK_INTERRUPTIBLE:
4244 stats->nr_sleeping++;
4246 case TASK_UNINTERRUPTIBLE:
4247 stats->nr_uninterruptible++;
4250 stats->nr_stopped++;
4253 if (delayacct_is_task_waiting_on_io(tsk))
4254 stats->nr_io_wait++;
4258 css_task_iter_end(&it);
4260 mutex_unlock(&cgroup_mutex);
4266 * seq_file methods for the tasks/procs files. The seq_file position is the
4267 * next pid to display; the seq_file iterator is a pointer to the pid
4268 * in the cgroup->l->list array.
4271 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4274 * Initially we receive a position value that corresponds to
4275 * one more than the last pid shown (or 0 on the first call or
4276 * after a seek to the start). Use a binary-search to find the
4277 * next pid to display, if any
4279 struct kernfs_open_file *of = s->private;
4280 struct cgroup *cgrp = seq_css(s)->cgroup;
4281 struct cgroup_pidlist *l;
4282 enum cgroup_filetype type = seq_cft(s)->private;
4283 int index = 0, pid = *pos;
4286 mutex_lock(&cgrp->pidlist_mutex);
4289 * !NULL @of->priv indicates that this isn't the first start()
4290 * after open. If the matching pidlist is around, we can use that.
4291 * Look for it. Note that @of->priv can't be used directly. It
4292 * could already have been destroyed.
4295 of->priv = cgroup_pidlist_find(cgrp, type);
4298 * Either this is the first start() after open or the matching
4299 * pidlist has been destroyed inbetween. Create a new one.
4302 ret = pidlist_array_load(cgrp, type,
4303 (struct cgroup_pidlist **)&of->priv);
4305 return ERR_PTR(ret);
4310 int end = l->length;
4312 while (index < end) {
4313 int mid = (index + end) / 2;
4314 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4317 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4323 /* If we're off the end of the array, we're done */
4324 if (index >= l->length)
4326 /* Update the abstract position to be the actual pid that we found */
4327 iter = l->list + index;
4328 *pos = cgroup_pid_fry(cgrp, *iter);
4332 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4334 struct kernfs_open_file *of = s->private;
4335 struct cgroup_pidlist *l = of->priv;
4338 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4339 CGROUP_PIDLIST_DESTROY_DELAY);
4340 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4343 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4345 struct kernfs_open_file *of = s->private;
4346 struct cgroup_pidlist *l = of->priv;
4348 pid_t *end = l->list + l->length;
4350 * Advance to the next pid in the array. If this goes off the
4357 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4362 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4364 seq_printf(s, "%d\n", *(int *)v);
4369 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4372 return notify_on_release(css->cgroup);
4375 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4376 struct cftype *cft, u64 val)
4379 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4381 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4385 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4388 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4391 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4392 struct cftype *cft, u64 val)
4395 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4397 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4401 /* cgroup core interface files for the default hierarchy */
4402 static struct cftype cgroup_dfl_base_files[] = {
4404 .name = "cgroup.procs",
4405 .file_offset = offsetof(struct cgroup, procs_file),
4406 .seq_start = cgroup_pidlist_start,
4407 .seq_next = cgroup_pidlist_next,
4408 .seq_stop = cgroup_pidlist_stop,
4409 .seq_show = cgroup_pidlist_show,
4410 .private = CGROUP_FILE_PROCS,
4411 .write = cgroup_procs_write,
4414 .name = "cgroup.controllers",
4415 .flags = CFTYPE_ONLY_ON_ROOT,
4416 .seq_show = cgroup_root_controllers_show,
4419 .name = "cgroup.controllers",
4420 .flags = CFTYPE_NOT_ON_ROOT,
4421 .seq_show = cgroup_controllers_show,
4424 .name = "cgroup.subtree_control",
4425 .seq_show = cgroup_subtree_control_show,
4426 .write = cgroup_subtree_control_write,
4429 .name = "cgroup.events",
4430 .flags = CFTYPE_NOT_ON_ROOT,
4431 .file_offset = offsetof(struct cgroup, events_file),
4432 .seq_show = cgroup_events_show,
4437 /* cgroup core interface files for the legacy hierarchies */
4438 static struct cftype cgroup_legacy_base_files[] = {
4440 .name = "cgroup.procs",
4441 .seq_start = cgroup_pidlist_start,
4442 .seq_next = cgroup_pidlist_next,
4443 .seq_stop = cgroup_pidlist_stop,
4444 .seq_show = cgroup_pidlist_show,
4445 .private = CGROUP_FILE_PROCS,
4446 .write = cgroup_procs_write,
4449 .name = "cgroup.clone_children",
4450 .read_u64 = cgroup_clone_children_read,
4451 .write_u64 = cgroup_clone_children_write,
4454 .name = "cgroup.sane_behavior",
4455 .flags = CFTYPE_ONLY_ON_ROOT,
4456 .seq_show = cgroup_sane_behavior_show,
4460 .seq_start = cgroup_pidlist_start,
4461 .seq_next = cgroup_pidlist_next,
4462 .seq_stop = cgroup_pidlist_stop,
4463 .seq_show = cgroup_pidlist_show,
4464 .private = CGROUP_FILE_TASKS,
4465 .write = cgroup_tasks_write,
4468 .name = "notify_on_release",
4469 .read_u64 = cgroup_read_notify_on_release,
4470 .write_u64 = cgroup_write_notify_on_release,
4473 .name = "release_agent",
4474 .flags = CFTYPE_ONLY_ON_ROOT,
4475 .seq_show = cgroup_release_agent_show,
4476 .write = cgroup_release_agent_write,
4477 .max_write_len = PATH_MAX - 1,
4483 * css destruction is four-stage process.
4485 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4486 * Implemented in kill_css().
4488 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4489 * and thus css_tryget_online() is guaranteed to fail, the css can be
4490 * offlined by invoking offline_css(). After offlining, the base ref is
4491 * put. Implemented in css_killed_work_fn().
4493 * 3. When the percpu_ref reaches zero, the only possible remaining
4494 * accessors are inside RCU read sections. css_release() schedules the
4497 * 4. After the grace period, the css can be freed. Implemented in
4498 * css_free_work_fn().
4500 * It is actually hairier because both step 2 and 4 require process context
4501 * and thus involve punting to css->destroy_work adding two additional
4502 * steps to the already complex sequence.
4504 static void css_free_work_fn(struct work_struct *work)
4506 struct cgroup_subsys_state *css =
4507 container_of(work, struct cgroup_subsys_state, destroy_work);
4508 struct cgroup_subsys *ss = css->ss;
4509 struct cgroup *cgrp = css->cgroup;
4510 struct cgroup_file *cfile;
4512 percpu_ref_exit(&css->refcnt);
4514 list_for_each_entry(cfile, &css->files, node)
4515 kernfs_put(cfile->kn);
4522 css_put(css->parent);
4525 cgroup_idr_remove(&ss->css_idr, id);
4528 /* cgroup free path */
4529 atomic_dec(&cgrp->root->nr_cgrps);
4530 cgroup_pidlist_destroy_all(cgrp);
4531 cancel_work_sync(&cgrp->release_agent_work);
4533 if (cgroup_parent(cgrp)) {
4535 * We get a ref to the parent, and put the ref when
4536 * this cgroup is being freed, so it's guaranteed
4537 * that the parent won't be destroyed before its
4540 cgroup_put(cgroup_parent(cgrp));
4541 kernfs_put(cgrp->kn);
4545 * This is root cgroup's refcnt reaching zero,
4546 * which indicates that the root should be
4549 cgroup_destroy_root(cgrp->root);
4554 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4556 struct cgroup_subsys_state *css =
4557 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4559 INIT_WORK(&css->destroy_work, css_free_work_fn);
4560 queue_work(cgroup_destroy_wq, &css->destroy_work);
4563 static void css_release_work_fn(struct work_struct *work)
4565 struct cgroup_subsys_state *css =
4566 container_of(work, struct cgroup_subsys_state, destroy_work);
4567 struct cgroup_subsys *ss = css->ss;
4568 struct cgroup *cgrp = css->cgroup;
4570 mutex_lock(&cgroup_mutex);
4572 css->flags |= CSS_RELEASED;
4573 list_del_rcu(&css->sibling);
4576 /* css release path */
4577 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4578 if (ss->css_released)
4579 ss->css_released(css);
4581 /* cgroup release path */
4582 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4586 * There are two control paths which try to determine
4587 * cgroup from dentry without going through kernfs -
4588 * cgroupstats_build() and css_tryget_online_from_dir().
4589 * Those are supported by RCU protecting clearing of
4590 * cgrp->kn->priv backpointer.
4592 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4595 mutex_unlock(&cgroup_mutex);
4597 call_rcu(&css->rcu_head, css_free_rcu_fn);
4600 static void css_release(struct percpu_ref *ref)
4602 struct cgroup_subsys_state *css =
4603 container_of(ref, struct cgroup_subsys_state, refcnt);
4605 INIT_WORK(&css->destroy_work, css_release_work_fn);
4606 queue_work(cgroup_destroy_wq, &css->destroy_work);
4609 static void init_and_link_css(struct cgroup_subsys_state *css,
4610 struct cgroup_subsys *ss, struct cgroup *cgrp)
4612 lockdep_assert_held(&cgroup_mutex);
4616 memset(css, 0, sizeof(*css));
4619 INIT_LIST_HEAD(&css->sibling);
4620 INIT_LIST_HEAD(&css->children);
4621 INIT_LIST_HEAD(&css->files);
4622 css->serial_nr = css_serial_nr_next++;
4624 if (cgroup_parent(cgrp)) {
4625 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4626 css_get(css->parent);
4629 BUG_ON(cgroup_css(cgrp, ss));
4632 /* invoke ->css_online() on a new CSS and mark it online if successful */
4633 static int online_css(struct cgroup_subsys_state *css)
4635 struct cgroup_subsys *ss = css->ss;
4638 lockdep_assert_held(&cgroup_mutex);
4641 ret = ss->css_online(css);
4643 css->flags |= CSS_ONLINE;
4644 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4649 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4650 static void offline_css(struct cgroup_subsys_state *css)
4652 struct cgroup_subsys *ss = css->ss;
4654 lockdep_assert_held(&cgroup_mutex);
4656 if (!(css->flags & CSS_ONLINE))
4659 if (ss->css_offline)
4660 ss->css_offline(css);
4662 css->flags &= ~CSS_ONLINE;
4663 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4665 wake_up_all(&css->cgroup->offline_waitq);
4669 * create_css - create a cgroup_subsys_state
4670 * @cgrp: the cgroup new css will be associated with
4671 * @ss: the subsys of new css
4672 * @visible: whether to create control knobs for the new css or not
4674 * Create a new css associated with @cgrp - @ss pair. On success, the new
4675 * css is online and installed in @cgrp with all interface files created if
4676 * @visible. Returns 0 on success, -errno on failure.
4678 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4681 struct cgroup *parent = cgroup_parent(cgrp);
4682 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4683 struct cgroup_subsys_state *css;
4686 lockdep_assert_held(&cgroup_mutex);
4688 css = ss->css_alloc(parent_css);
4690 return PTR_ERR(css);
4692 init_and_link_css(css, ss, cgrp);
4694 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4698 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4700 goto err_free_percpu_ref;
4704 err = css_populate_dir(css, NULL);
4709 /* @css is ready to be brought online now, make it visible */
4710 list_add_tail_rcu(&css->sibling, &parent_css->children);
4711 cgroup_idr_replace(&ss->css_idr, css, css->id);
4713 err = online_css(css);
4717 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4718 cgroup_parent(parent)) {
4719 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4720 current->comm, current->pid, ss->name);
4721 if (!strcmp(ss->name, "memory"))
4722 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4723 ss->warned_broken_hierarchy = true;
4729 list_del_rcu(&css->sibling);
4730 css_clear_dir(css, NULL);
4732 cgroup_idr_remove(&ss->css_idr, css->id);
4733 err_free_percpu_ref:
4734 percpu_ref_exit(&css->refcnt);
4736 call_rcu(&css->rcu_head, css_free_rcu_fn);
4740 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4743 struct cgroup *parent, *cgrp;
4744 struct cgroup_root *root;
4745 struct cgroup_subsys *ss;
4746 struct kernfs_node *kn;
4749 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4751 if (strchr(name, '\n'))
4754 parent = cgroup_kn_lock_live(parent_kn);
4757 root = parent->root;
4759 /* allocate the cgroup and its ID, 0 is reserved for the root */
4760 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4766 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4771 * Temporarily set the pointer to NULL, so idr_find() won't return
4772 * a half-baked cgroup.
4774 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4777 goto out_cancel_ref;
4780 init_cgroup_housekeeping(cgrp);
4782 cgrp->self.parent = &parent->self;
4785 if (notify_on_release(parent))
4786 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4788 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4789 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4791 /* create the directory */
4792 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4800 * This extra ref will be put in cgroup_free_fn() and guarantees
4801 * that @cgrp->kn is always accessible.
4805 cgrp->self.serial_nr = css_serial_nr_next++;
4807 /* allocation complete, commit to creation */
4808 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4809 atomic_inc(&root->nr_cgrps);
4813 * @cgrp is now fully operational. If something fails after this
4814 * point, it'll be released via the normal destruction path.
4816 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4818 ret = cgroup_kn_set_ugid(kn);
4822 ret = css_populate_dir(&cgrp->self, NULL);
4826 /* let's create and online css's */
4827 for_each_subsys(ss, ssid) {
4828 if (parent->child_subsys_mask & (1 << ssid)) {
4829 ret = create_css(cgrp, ss,
4830 parent->subtree_control & (1 << ssid));
4837 * On the default hierarchy, a child doesn't automatically inherit
4838 * subtree_control from the parent. Each is configured manually.
4840 if (!cgroup_on_dfl(cgrp)) {
4841 cgrp->subtree_control = parent->subtree_control;
4842 cgroup_refresh_child_subsys_mask(cgrp);
4845 kernfs_activate(kn);
4851 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4853 percpu_ref_exit(&cgrp->self.refcnt);
4857 cgroup_kn_unlock(parent_kn);
4861 cgroup_destroy_locked(cgrp);
4866 * This is called when the refcnt of a css is confirmed to be killed.
4867 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4868 * initate destruction and put the css ref from kill_css().
4870 static void css_killed_work_fn(struct work_struct *work)
4872 struct cgroup_subsys_state *css =
4873 container_of(work, struct cgroup_subsys_state, destroy_work);
4875 mutex_lock(&cgroup_mutex);
4877 mutex_unlock(&cgroup_mutex);
4882 /* css kill confirmation processing requires process context, bounce */
4883 static void css_killed_ref_fn(struct percpu_ref *ref)
4885 struct cgroup_subsys_state *css =
4886 container_of(ref, struct cgroup_subsys_state, refcnt);
4888 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4889 queue_work(cgroup_destroy_wq, &css->destroy_work);
4893 * kill_css - destroy a css
4894 * @css: css to destroy
4896 * This function initiates destruction of @css by removing cgroup interface
4897 * files and putting its base reference. ->css_offline() will be invoked
4898 * asynchronously once css_tryget_online() is guaranteed to fail and when
4899 * the reference count reaches zero, @css will be released.
4901 static void kill_css(struct cgroup_subsys_state *css)
4903 lockdep_assert_held(&cgroup_mutex);
4906 * This must happen before css is disassociated with its cgroup.
4907 * See seq_css() for details.
4909 css_clear_dir(css, NULL);
4912 * Killing would put the base ref, but we need to keep it alive
4913 * until after ->css_offline().
4918 * cgroup core guarantees that, by the time ->css_offline() is
4919 * invoked, no new css reference will be given out via
4920 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4921 * proceed to offlining css's because percpu_ref_kill() doesn't
4922 * guarantee that the ref is seen as killed on all CPUs on return.
4924 * Use percpu_ref_kill_and_confirm() to get notifications as each
4925 * css is confirmed to be seen as killed on all CPUs.
4927 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4931 * cgroup_destroy_locked - the first stage of cgroup destruction
4932 * @cgrp: cgroup to be destroyed
4934 * css's make use of percpu refcnts whose killing latency shouldn't be
4935 * exposed to userland and are RCU protected. Also, cgroup core needs to
4936 * guarantee that css_tryget_online() won't succeed by the time
4937 * ->css_offline() is invoked. To satisfy all the requirements,
4938 * destruction is implemented in the following two steps.
4940 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4941 * userland visible parts and start killing the percpu refcnts of
4942 * css's. Set up so that the next stage will be kicked off once all
4943 * the percpu refcnts are confirmed to be killed.
4945 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4946 * rest of destruction. Once all cgroup references are gone, the
4947 * cgroup is RCU-freed.
4949 * This function implements s1. After this step, @cgrp is gone as far as
4950 * the userland is concerned and a new cgroup with the same name may be
4951 * created. As cgroup doesn't care about the names internally, this
4952 * doesn't cause any problem.
4954 static int cgroup_destroy_locked(struct cgroup *cgrp)
4955 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4957 struct cgroup_subsys_state *css;
4961 lockdep_assert_held(&cgroup_mutex);
4964 * css_set_rwsem synchronizes access to ->cset_links and prevents
4965 * @cgrp from being removed while put_css_set() is in progress.
4967 down_read(&css_set_rwsem);
4968 empty = list_empty(&cgrp->cset_links);
4969 up_read(&css_set_rwsem);
4974 * Make sure there's no live children. We can't test emptiness of
4975 * ->self.children as dead children linger on it while being
4976 * drained; otherwise, "rmdir parent/child parent" may fail.
4978 if (css_has_online_children(&cgrp->self))
4982 * Mark @cgrp dead. This prevents further task migration and child
4983 * creation by disabling cgroup_lock_live_group().
4985 cgrp->self.flags &= ~CSS_ONLINE;
4987 /* initiate massacre of all css's */
4988 for_each_css(css, ssid, cgrp)
4992 * Remove @cgrp directory along with the base files. @cgrp has an
4993 * extra ref on its kn.
4995 kernfs_remove(cgrp->kn);
4997 check_for_release(cgroup_parent(cgrp));
4999 /* put the base reference */
5000 percpu_ref_kill(&cgrp->self.refcnt);
5005 static int cgroup_rmdir(struct kernfs_node *kn)
5007 struct cgroup *cgrp;
5010 cgrp = cgroup_kn_lock_live(kn);
5014 ret = cgroup_destroy_locked(cgrp);
5016 cgroup_kn_unlock(kn);
5020 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5021 .remount_fs = cgroup_remount,
5022 .show_options = cgroup_show_options,
5023 .mkdir = cgroup_mkdir,
5024 .rmdir = cgroup_rmdir,
5025 .rename = cgroup_rename,
5028 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5030 struct cgroup_subsys_state *css;
5032 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
5034 mutex_lock(&cgroup_mutex);
5036 idr_init(&ss->css_idr);
5037 INIT_LIST_HEAD(&ss->cfts);
5039 /* Create the root cgroup state for this subsystem */
5040 ss->root = &cgrp_dfl_root;
5041 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5042 /* We don't handle early failures gracefully */
5043 BUG_ON(IS_ERR(css));
5044 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5047 * Root csses are never destroyed and we can't initialize
5048 * percpu_ref during early init. Disable refcnting.
5050 css->flags |= CSS_NO_REF;
5053 /* allocation can't be done safely during early init */
5056 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5057 BUG_ON(css->id < 0);
5060 /* Update the init_css_set to contain a subsys
5061 * pointer to this state - since the subsystem is
5062 * newly registered, all tasks and hence the
5063 * init_css_set is in the subsystem's root cgroup. */
5064 init_css_set.subsys[ss->id] = css;
5066 have_fork_callback |= (bool)ss->fork << ss->id;
5067 have_exit_callback |= (bool)ss->exit << ss->id;
5068 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5070 /* At system boot, before all subsystems have been
5071 * registered, no tasks have been forked, so we don't
5072 * need to invoke fork callbacks here. */
5073 BUG_ON(!list_empty(&init_task.tasks));
5075 BUG_ON(online_css(css));
5077 mutex_unlock(&cgroup_mutex);
5081 * cgroup_init_early - cgroup initialization at system boot
5083 * Initialize cgroups at system boot, and initialize any
5084 * subsystems that request early init.
5086 int __init cgroup_init_early(void)
5088 static struct cgroup_sb_opts __initdata opts;
5089 struct cgroup_subsys *ss;
5092 init_cgroup_root(&cgrp_dfl_root, &opts);
5093 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5095 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5097 for_each_subsys(ss, i) {
5098 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5099 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5100 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5102 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5103 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5106 ss->name = cgroup_subsys_name[i];
5107 if (!ss->legacy_name)
5108 ss->legacy_name = cgroup_subsys_name[i];
5111 cgroup_init_subsys(ss, true);
5117 * cgroup_init - cgroup initialization
5119 * Register cgroup filesystem and /proc file, and initialize
5120 * any subsystems that didn't request early init.
5122 int __init cgroup_init(void)
5124 struct cgroup_subsys *ss;
5128 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5129 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5130 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5132 mutex_lock(&cgroup_mutex);
5134 /* Add init_css_set to the hash table */
5135 key = css_set_hash(init_css_set.subsys);
5136 hash_add(css_set_table, &init_css_set.hlist, key);
5138 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5140 mutex_unlock(&cgroup_mutex);
5142 for_each_subsys(ss, ssid) {
5143 if (ss->early_init) {
5144 struct cgroup_subsys_state *css =
5145 init_css_set.subsys[ss->id];
5147 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5149 BUG_ON(css->id < 0);
5151 cgroup_init_subsys(ss, false);
5154 list_add_tail(&init_css_set.e_cset_node[ssid],
5155 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5158 * Setting dfl_root subsys_mask needs to consider the
5159 * disabled flag and cftype registration needs kmalloc,
5160 * both of which aren't available during early_init.
5162 if (!cgroup_ssid_enabled(ssid))
5165 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5167 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5168 ss->dfl_cftypes = ss->legacy_cftypes;
5170 if (!ss->dfl_cftypes)
5171 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5173 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5174 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5176 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5177 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5181 ss->bind(init_css_set.subsys[ssid]);
5184 err = sysfs_create_mount_point(fs_kobj, "cgroup");
5188 err = register_filesystem(&cgroup_fs_type);
5190 sysfs_remove_mount_point(fs_kobj, "cgroup");
5194 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5198 static int __init cgroup_wq_init(void)
5201 * There isn't much point in executing destruction path in
5202 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5203 * Use 1 for @max_active.
5205 * We would prefer to do this in cgroup_init() above, but that
5206 * is called before init_workqueues(): so leave this until after.
5208 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5209 BUG_ON(!cgroup_destroy_wq);
5212 * Used to destroy pidlists and separate to serve as flush domain.
5213 * Cap @max_active to 1 too.
5215 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5217 BUG_ON(!cgroup_pidlist_destroy_wq);
5221 core_initcall(cgroup_wq_init);
5224 * proc_cgroup_show()
5225 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5226 * - Used for /proc/<pid>/cgroup.
5228 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5229 struct pid *pid, struct task_struct *tsk)
5233 struct cgroup_root *root;
5236 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5240 mutex_lock(&cgroup_mutex);
5241 down_read(&css_set_rwsem);
5243 for_each_root(root) {
5244 struct cgroup_subsys *ss;
5245 struct cgroup *cgrp;
5246 int ssid, count = 0;
5248 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5251 seq_printf(m, "%d:", root->hierarchy_id);
5252 if (root != &cgrp_dfl_root)
5253 for_each_subsys(ss, ssid)
5254 if (root->subsys_mask & (1 << ssid))
5255 seq_printf(m, "%s%s", count++ ? "," : "",
5257 if (strlen(root->name))
5258 seq_printf(m, "%sname=%s", count ? "," : "",
5261 cgrp = task_cgroup_from_root(tsk, root);
5262 path = cgroup_path(cgrp, buf, PATH_MAX);
5264 retval = -ENAMETOOLONG;
5273 up_read(&css_set_rwsem);
5274 mutex_unlock(&cgroup_mutex);
5280 /* Display information about each subsystem and each hierarchy */
5281 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5283 struct cgroup_subsys *ss;
5286 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5288 * ideally we don't want subsystems moving around while we do this.
5289 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5290 * subsys/hierarchy state.
5292 mutex_lock(&cgroup_mutex);
5294 for_each_subsys(ss, i)
5295 seq_printf(m, "%s\t%d\t%d\t%d\n",
5296 ss->legacy_name, ss->root->hierarchy_id,
5297 atomic_read(&ss->root->nr_cgrps),
5298 cgroup_ssid_enabled(i));
5300 mutex_unlock(&cgroup_mutex);
5304 static int cgroupstats_open(struct inode *inode, struct file *file)
5306 return single_open(file, proc_cgroupstats_show, NULL);
5309 static const struct file_operations proc_cgroupstats_operations = {
5310 .open = cgroupstats_open,
5312 .llseek = seq_lseek,
5313 .release = single_release,
5316 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5318 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5319 return &ss_priv[i - CGROUP_CANFORK_START];
5323 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5325 void **private = subsys_canfork_priv_p(ss_priv, i);
5326 return private ? *private : NULL;
5330 * cgroup_fork - initialize cgroup related fields during copy_process()
5331 * @child: pointer to task_struct of forking parent process.
5333 * A task is associated with the init_css_set until cgroup_post_fork()
5334 * attaches it to the parent's css_set. Empty cg_list indicates that
5335 * @child isn't holding reference to its css_set.
5337 void cgroup_fork(struct task_struct *child)
5339 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5340 INIT_LIST_HEAD(&child->cg_list);
5344 * cgroup_can_fork - called on a new task before the process is exposed
5345 * @child: the task in question.
5347 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5348 * returns an error, the fork aborts with that error code. This allows for
5349 * a cgroup subsystem to conditionally allow or deny new forks.
5351 int cgroup_can_fork(struct task_struct *child,
5352 void *ss_priv[CGROUP_CANFORK_COUNT])
5354 struct cgroup_subsys *ss;
5357 for_each_subsys_which(ss, i, &have_canfork_callback) {
5358 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5366 for_each_subsys(ss, j) {
5369 if (ss->cancel_fork)
5370 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5377 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5378 * @child: the task in question
5380 * This calls the cancel_fork() callbacks if a fork failed *after*
5381 * cgroup_can_fork() succeded.
5383 void cgroup_cancel_fork(struct task_struct *child,
5384 void *ss_priv[CGROUP_CANFORK_COUNT])
5386 struct cgroup_subsys *ss;
5389 for_each_subsys(ss, i)
5390 if (ss->cancel_fork)
5391 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5395 * cgroup_post_fork - called on a new task after adding it to the task list
5396 * @child: the task in question
5398 * Adds the task to the list running through its css_set if necessary and
5399 * call the subsystem fork() callbacks. Has to be after the task is
5400 * visible on the task list in case we race with the first call to
5401 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5404 void cgroup_post_fork(struct task_struct *child,
5405 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5407 struct cgroup_subsys *ss;
5411 * This may race against cgroup_enable_task_cg_lists(). As that
5412 * function sets use_task_css_set_links before grabbing
5413 * tasklist_lock and we just went through tasklist_lock to add
5414 * @child, it's guaranteed that either we see the set
5415 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5416 * @child during its iteration.
5418 * If we won the race, @child is associated with %current's
5419 * css_set. Grabbing css_set_rwsem guarantees both that the
5420 * association is stable, and, on completion of the parent's
5421 * migration, @child is visible in the source of migration or
5422 * already in the destination cgroup. This guarantee is necessary
5423 * when implementing operations which need to migrate all tasks of
5424 * a cgroup to another.
5426 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5427 * will remain in init_css_set. This is safe because all tasks are
5428 * in the init_css_set before cg_links is enabled and there's no
5429 * operation which transfers all tasks out of init_css_set.
5431 if (use_task_css_set_links) {
5432 struct css_set *cset;
5434 down_write(&css_set_rwsem);
5435 cset = task_css_set(current);
5436 if (list_empty(&child->cg_list)) {
5437 rcu_assign_pointer(child->cgroups, cset);
5438 list_add(&child->cg_list, &cset->tasks);
5441 up_write(&css_set_rwsem);
5445 * Call ss->fork(). This must happen after @child is linked on
5446 * css_set; otherwise, @child might change state between ->fork()
5447 * and addition to css_set.
5449 for_each_subsys_which(ss, i, &have_fork_callback)
5450 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5454 * cgroup_exit - detach cgroup from exiting task
5455 * @tsk: pointer to task_struct of exiting process
5457 * Description: Detach cgroup from @tsk and release it.
5459 * Note that cgroups marked notify_on_release force every task in
5460 * them to take the global cgroup_mutex mutex when exiting.
5461 * This could impact scaling on very large systems. Be reluctant to
5462 * use notify_on_release cgroups where very high task exit scaling
5463 * is required on large systems.
5465 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5466 * call cgroup_exit() while the task is still competent to handle
5467 * notify_on_release(), then leave the task attached to the root cgroup in
5468 * each hierarchy for the remainder of its exit. No need to bother with
5469 * init_css_set refcnting. init_css_set never goes away and we can't race
5470 * with migration path - PF_EXITING is visible to migration path.
5472 void cgroup_exit(struct task_struct *tsk)
5474 struct cgroup_subsys *ss;
5475 struct css_set *cset;
5476 bool put_cset = false;
5480 * Unlink from @tsk from its css_set. As migration path can't race
5481 * with us, we can check cg_list without grabbing css_set_rwsem.
5483 if (!list_empty(&tsk->cg_list)) {
5484 down_write(&css_set_rwsem);
5485 list_del_init(&tsk->cg_list);
5486 up_write(&css_set_rwsem);
5490 /* Reassign the task to the init_css_set. */
5491 cset = task_css_set(tsk);
5492 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5494 /* see cgroup_post_fork() for details */
5495 for_each_subsys_which(ss, i, &have_exit_callback) {
5496 struct cgroup_subsys_state *old_css = cset->subsys[i];
5497 struct cgroup_subsys_state *css = task_css(tsk, i);
5499 ss->exit(css, old_css, tsk);
5506 static void check_for_release(struct cgroup *cgrp)
5508 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5509 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5510 schedule_work(&cgrp->release_agent_work);
5514 * Notify userspace when a cgroup is released, by running the
5515 * configured release agent with the name of the cgroup (path
5516 * relative to the root of cgroup file system) as the argument.
5518 * Most likely, this user command will try to rmdir this cgroup.
5520 * This races with the possibility that some other task will be
5521 * attached to this cgroup before it is removed, or that some other
5522 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5523 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5524 * unused, and this cgroup will be reprieved from its death sentence,
5525 * to continue to serve a useful existence. Next time it's released,
5526 * we will get notified again, if it still has 'notify_on_release' set.
5528 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5529 * means only wait until the task is successfully execve()'d. The
5530 * separate release agent task is forked by call_usermodehelper(),
5531 * then control in this thread returns here, without waiting for the
5532 * release agent task. We don't bother to wait because the caller of
5533 * this routine has no use for the exit status of the release agent
5534 * task, so no sense holding our caller up for that.
5536 static void cgroup_release_agent(struct work_struct *work)
5538 struct cgroup *cgrp =
5539 container_of(work, struct cgroup, release_agent_work);
5540 char *pathbuf = NULL, *agentbuf = NULL, *path;
5541 char *argv[3], *envp[3];
5543 mutex_lock(&cgroup_mutex);
5545 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5546 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5547 if (!pathbuf || !agentbuf)
5550 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5558 /* minimal command environment */
5560 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5563 mutex_unlock(&cgroup_mutex);
5564 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5567 mutex_unlock(&cgroup_mutex);
5573 static int __init cgroup_disable(char *str)
5575 struct cgroup_subsys *ss;
5579 while ((token = strsep(&str, ",")) != NULL) {
5583 for_each_subsys(ss, i) {
5584 if (strcmp(token, ss->name) &&
5585 strcmp(token, ss->legacy_name))
5588 static_branch_disable(cgroup_subsys_enabled_key[i]);
5589 printk(KERN_INFO "Disabling %s control group subsystem\n",
5596 __setup("cgroup_disable=", cgroup_disable);
5598 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5600 printk("cgroup: using legacy files on the default hierarchy\n");
5601 cgroup_legacy_files_on_dfl = true;
5604 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5607 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5608 * @dentry: directory dentry of interest
5609 * @ss: subsystem of interest
5611 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5612 * to get the corresponding css and return it. If such css doesn't exist
5613 * or can't be pinned, an ERR_PTR value is returned.
5615 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5616 struct cgroup_subsys *ss)
5618 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5619 struct cgroup_subsys_state *css = NULL;
5620 struct cgroup *cgrp;
5622 /* is @dentry a cgroup dir? */
5623 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5624 kernfs_type(kn) != KERNFS_DIR)
5625 return ERR_PTR(-EBADF);
5630 * This path doesn't originate from kernfs and @kn could already
5631 * have been or be removed at any point. @kn->priv is RCU
5632 * protected for this access. See css_release_work_fn() for details.
5634 cgrp = rcu_dereference(kn->priv);
5636 css = cgroup_css(cgrp, ss);
5638 if (!css || !css_tryget_online(css))
5639 css = ERR_PTR(-ENOENT);
5646 * css_from_id - lookup css by id
5647 * @id: the cgroup id
5648 * @ss: cgroup subsys to be looked into
5650 * Returns the css if there's valid one with @id, otherwise returns NULL.
5651 * Should be called under rcu_read_lock().
5653 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5655 WARN_ON_ONCE(!rcu_read_lock_held());
5656 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5659 #ifdef CONFIG_CGROUP_DEBUG
5660 static struct cgroup_subsys_state *
5661 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5663 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5666 return ERR_PTR(-ENOMEM);
5671 static void debug_css_free(struct cgroup_subsys_state *css)
5676 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5679 return cgroup_task_count(css->cgroup);
5682 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5685 return (u64)(unsigned long)current->cgroups;
5688 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5694 count = atomic_read(&task_css_set(current)->refcount);
5699 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5701 struct cgrp_cset_link *link;
5702 struct css_set *cset;
5705 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5709 down_read(&css_set_rwsem);
5711 cset = rcu_dereference(current->cgroups);
5712 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5713 struct cgroup *c = link->cgrp;
5715 cgroup_name(c, name_buf, NAME_MAX + 1);
5716 seq_printf(seq, "Root %d group %s\n",
5717 c->root->hierarchy_id, name_buf);
5720 up_read(&css_set_rwsem);
5725 #define MAX_TASKS_SHOWN_PER_CSS 25
5726 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5728 struct cgroup_subsys_state *css = seq_css(seq);
5729 struct cgrp_cset_link *link;
5731 down_read(&css_set_rwsem);
5732 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5733 struct css_set *cset = link->cset;
5734 struct task_struct *task;
5737 seq_printf(seq, "css_set %p\n", cset);
5739 list_for_each_entry(task, &cset->tasks, cg_list) {
5740 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5742 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5745 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5746 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5748 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5752 seq_puts(seq, " ...\n");
5754 up_read(&css_set_rwsem);
5758 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5760 return (!cgroup_has_tasks(css->cgroup) &&
5761 !css_has_online_children(&css->cgroup->self));
5764 static struct cftype debug_files[] = {
5766 .name = "taskcount",
5767 .read_u64 = debug_taskcount_read,
5771 .name = "current_css_set",
5772 .read_u64 = current_css_set_read,
5776 .name = "current_css_set_refcount",
5777 .read_u64 = current_css_set_refcount_read,
5781 .name = "current_css_set_cg_links",
5782 .seq_show = current_css_set_cg_links_read,
5786 .name = "cgroup_css_links",
5787 .seq_show = cgroup_css_links_read,
5791 .name = "releasable",
5792 .read_u64 = releasable_read,
5798 struct cgroup_subsys debug_cgrp_subsys = {
5799 .css_alloc = debug_css_alloc,
5800 .css_free = debug_css_free,
5801 .legacy_cftypes = debug_files,
5803 #endif /* CONFIG_CGROUP_DEBUG */