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/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
60 #include <linux/cpuset.h>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_lock protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex);
86 DEFINE_SPINLOCK(css_set_lock);
87 EXPORT_SYMBOL_GPL(cgroup_mutex);
88 EXPORT_SYMBOL_GPL(css_set_lock);
90 static DEFINE_MUTEX(cgroup_mutex);
91 static DEFINE_SPINLOCK(css_set_lock);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock);
101 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
102 * against file removal/re-creation across css hiding.
104 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
107 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
108 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
110 static DEFINE_SPINLOCK(release_agent_path_lock);
112 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
114 #define cgroup_assert_mutex_or_rcu_locked() \
115 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
116 !lockdep_is_held(&cgroup_mutex), \
117 "cgroup_mutex or RCU read lock required");
120 * cgroup destruction makes heavy use of work items and there can be a lot
121 * of concurrent destructions. Use a separate workqueue so that cgroup
122 * destruction work items don't end up filling up max_active of system_wq
123 * which may lead to deadlock.
125 static struct workqueue_struct *cgroup_destroy_wq;
128 * pidlist destructions need to be flushed on cgroup destruction. Use a
129 * separate workqueue as flush domain.
131 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
133 /* generate an array of cgroup subsystem pointers */
134 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
135 static struct cgroup_subsys *cgroup_subsys[] = {
136 #include <linux/cgroup_subsys.h>
140 /* array of cgroup subsystem names */
141 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
142 static const char *cgroup_subsys_name[] = {
143 #include <linux/cgroup_subsys.h>
147 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
149 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
150 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
151 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
152 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
153 #include <linux/cgroup_subsys.h>
156 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
157 static struct static_key_true *cgroup_subsys_enabled_key[] = {
158 #include <linux/cgroup_subsys.h>
162 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
163 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
164 #include <linux/cgroup_subsys.h>
169 * The default hierarchy, reserved for the subsystems that are otherwise
170 * unattached - it never has more than a single cgroup, and all tasks are
171 * part of that cgroup.
173 struct cgroup_root cgrp_dfl_root;
174 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
177 * The default hierarchy always exists but is hidden until mounted for the
178 * first time. This is for backward compatibility.
180 static bool cgrp_dfl_root_visible;
182 /* some controllers are not supported in the default hierarchy */
183 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
185 /* The list of hierarchy roots */
187 static LIST_HEAD(cgroup_roots);
188 static int cgroup_root_count;
190 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
191 static DEFINE_IDR(cgroup_hierarchy_idr);
194 * Assign a monotonically increasing serial number to csses. It guarantees
195 * cgroups with bigger numbers are newer than those with smaller numbers.
196 * Also, as csses are always appended to the parent's ->children list, it
197 * guarantees that sibling csses are always sorted in the ascending serial
198 * number order on the list. Protected by cgroup_mutex.
200 static u64 css_serial_nr_next = 1;
203 * These bitmask flags indicate whether tasks in the fork and exit paths have
204 * fork/exit handlers to call. This avoids us having to do extra work in the
205 * fork/exit path to check which subsystems have fork/exit callbacks.
207 static unsigned long have_fork_callback __read_mostly;
208 static unsigned long have_exit_callback __read_mostly;
209 static unsigned long have_free_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 void css_task_iter_advance(struct css_task_iter *it);
220 static int cgroup_destroy_locked(struct cgroup *cgrp);
221 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
223 static void css_release(struct percpu_ref *ref);
224 static void kill_css(struct cgroup_subsys_state *css);
225 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
226 struct cgroup *cgrp, struct cftype cfts[],
230 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
231 * @ssid: subsys ID of interest
233 * cgroup_subsys_enabled() can only be used with literal subsys names which
234 * is fine for individual subsystems but unsuitable for cgroup core. This
235 * is slower static_key_enabled() based test indexed by @ssid.
237 static bool cgroup_ssid_enabled(int ssid)
239 if (CGROUP_SUBSYS_COUNT == 0)
242 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
246 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
247 * @cgrp: the cgroup of interest
249 * The default hierarchy is the v2 interface of cgroup and this function
250 * can be used to test whether a cgroup is on the default hierarchy for
251 * cases where a subsystem should behave differnetly depending on the
254 * The set of behaviors which change on the default hierarchy are still
255 * being determined and the mount option is prefixed with __DEVEL__.
257 * List of changed behaviors:
259 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
260 * and "name" are disallowed.
262 * - When mounting an existing superblock, mount options should match.
264 * - Remount is disallowed.
266 * - rename(2) is disallowed.
268 * - "tasks" is removed. Everything should be at process granularity. Use
269 * "cgroup.procs" instead.
271 * - "cgroup.procs" is not sorted. pids will be unique unless they got
272 * recycled inbetween reads.
274 * - "release_agent" and "notify_on_release" are removed. Replacement
275 * notification mechanism will be implemented.
277 * - "cgroup.clone_children" is removed.
279 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
280 * and its descendants contain no task; otherwise, 1. The file also
281 * generates kernfs notification which can be monitored through poll and
282 * [di]notify when the value of the file changes.
284 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
285 * take masks of ancestors with non-empty cpus/mems, instead of being
286 * moved to an ancestor.
288 * - cpuset: a task can be moved into an empty cpuset, and again it takes
289 * masks of ancestors.
291 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
294 * - blkcg: blk-throttle becomes properly hierarchical.
296 * - debug: disallowed on the default hierarchy.
298 static bool cgroup_on_dfl(const struct cgroup *cgrp)
300 return cgrp->root == &cgrp_dfl_root;
303 /* IDR wrappers which synchronize using cgroup_idr_lock */
304 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
309 idr_preload(gfp_mask);
310 spin_lock_bh(&cgroup_idr_lock);
311 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
312 spin_unlock_bh(&cgroup_idr_lock);
317 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
321 spin_lock_bh(&cgroup_idr_lock);
322 ret = idr_replace(idr, ptr, id);
323 spin_unlock_bh(&cgroup_idr_lock);
327 static void cgroup_idr_remove(struct idr *idr, int id)
329 spin_lock_bh(&cgroup_idr_lock);
331 spin_unlock_bh(&cgroup_idr_lock);
334 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
336 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
339 return container_of(parent_css, struct cgroup, self);
344 * cgroup_css - obtain a cgroup's css for the specified subsystem
345 * @cgrp: the cgroup of interest
346 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
348 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
349 * function must be called either under cgroup_mutex or rcu_read_lock() and
350 * the caller is responsible for pinning the returned css if it wants to
351 * keep accessing it outside the said locks. This function may return
352 * %NULL if @cgrp doesn't have @subsys_id enabled.
354 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
355 struct cgroup_subsys *ss)
358 return rcu_dereference_check(cgrp->subsys[ss->id],
359 lockdep_is_held(&cgroup_mutex));
365 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
366 * @cgrp: the cgroup of interest
367 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
369 * Similar to cgroup_css() but returns the effective css, which is defined
370 * as the matching css of the nearest ancestor including self which has @ss
371 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
372 * function is guaranteed to return non-NULL css.
374 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
375 struct cgroup_subsys *ss)
377 lockdep_assert_held(&cgroup_mutex);
382 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
386 * This function is used while updating css associations and thus
387 * can't test the csses directly. Use ->child_subsys_mask.
389 while (cgroup_parent(cgrp) &&
390 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
391 cgrp = cgroup_parent(cgrp);
393 return cgroup_css(cgrp, ss);
397 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest
401 * Find and get the effective css of @cgrp for @ss. The effective css is
402 * defined as the matching css of the nearest ancestor including self which
403 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
404 * the root css is returned, so this function always returns a valid css.
405 * The returned css must be put using css_put().
407 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
408 struct cgroup_subsys *ss)
410 struct cgroup_subsys_state *css;
415 css = cgroup_css(cgrp, ss);
417 if (css && css_tryget_online(css))
419 cgrp = cgroup_parent(cgrp);
422 css = init_css_set.subsys[ss->id];
429 /* convenient tests for these bits */
430 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
432 return !(cgrp->self.flags & CSS_ONLINE);
435 static void cgroup_get(struct cgroup *cgrp)
437 WARN_ON_ONCE(cgroup_is_dead(cgrp));
438 css_get(&cgrp->self);
441 static bool cgroup_tryget(struct cgroup *cgrp)
443 return css_tryget(&cgrp->self);
446 static void cgroup_put(struct cgroup *cgrp)
448 css_put(&cgrp->self);
451 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
453 struct cgroup *cgrp = of->kn->parent->priv;
454 struct cftype *cft = of_cft(of);
457 * This is open and unprotected implementation of cgroup_css().
458 * seq_css() is only called from a kernfs file operation which has
459 * an active reference on the file. Because all the subsystem
460 * files are drained before a css is disassociated with a cgroup,
461 * the matching css from the cgroup's subsys table is guaranteed to
462 * be and stay valid until the enclosing operation is complete.
465 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
469 EXPORT_SYMBOL_GPL(of_css);
472 * cgroup_is_descendant - test ancestry
473 * @cgrp: the cgroup to be tested
474 * @ancestor: possible ancestor of @cgrp
476 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
477 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
478 * and @ancestor are accessible.
480 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
483 if (cgrp == ancestor)
485 cgrp = cgroup_parent(cgrp);
490 static int notify_on_release(const struct cgroup *cgrp)
492 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
496 * for_each_css - iterate all css's of a cgroup
497 * @css: the iteration cursor
498 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
499 * @cgrp: the target cgroup to iterate css's of
501 * Should be called under cgroup_[tree_]mutex.
503 #define for_each_css(css, ssid, cgrp) \
504 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
505 if (!((css) = rcu_dereference_check( \
506 (cgrp)->subsys[(ssid)], \
507 lockdep_is_held(&cgroup_mutex)))) { } \
511 * for_each_e_css - iterate all effective css's of a cgroup
512 * @css: the iteration cursor
513 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
514 * @cgrp: the target cgroup to iterate css's of
516 * Should be called under cgroup_[tree_]mutex.
518 #define for_each_e_css(css, ssid, cgrp) \
519 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
520 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
525 * for_each_subsys - iterate all enabled cgroup subsystems
526 * @ss: the iteration cursor
527 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
529 #define for_each_subsys(ss, ssid) \
530 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
531 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
534 * for_each_subsys_which - filter for_each_subsys with a bitmask
535 * @ss: the iteration cursor
536 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
537 * @ss_maskp: a pointer to the bitmask
539 * The block will only run for cases where the ssid-th bit (1 << ssid) of
542 #define for_each_subsys_which(ss, ssid, ss_maskp) \
543 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
546 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
547 if (((ss) = cgroup_subsys[ssid]) && false) \
551 /* iterate across the hierarchies */
552 #define for_each_root(root) \
553 list_for_each_entry((root), &cgroup_roots, root_list)
555 /* iterate over child cgrps, lock should be held throughout iteration */
556 #define cgroup_for_each_live_child(child, cgrp) \
557 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
558 if (({ lockdep_assert_held(&cgroup_mutex); \
559 cgroup_is_dead(child); })) \
563 static void cgroup_release_agent(struct work_struct *work);
564 static void check_for_release(struct cgroup *cgrp);
567 * A cgroup can be associated with multiple css_sets as different tasks may
568 * belong to different cgroups on different hierarchies. In the other
569 * direction, a css_set is naturally associated with multiple cgroups.
570 * This M:N relationship is represented by the following link structure
571 * which exists for each association and allows traversing the associations
574 struct cgrp_cset_link {
575 /* the cgroup and css_set this link associates */
577 struct css_set *cset;
579 /* list of cgrp_cset_links anchored at cgrp->cset_links */
580 struct list_head cset_link;
582 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
583 struct list_head cgrp_link;
587 * The default css_set - used by init and its children prior to any
588 * hierarchies being mounted. It contains a pointer to the root state
589 * for each subsystem. Also used to anchor the list of css_sets. Not
590 * reference-counted, to improve performance when child cgroups
591 * haven't been created.
593 struct css_set init_css_set = {
594 .refcount = ATOMIC_INIT(1),
595 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
596 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
597 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
598 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
599 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
600 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
603 static int css_set_count = 1; /* 1 for init_css_set */
606 * css_set_populated - does a css_set contain any tasks?
607 * @cset: target css_set
609 static bool css_set_populated(struct css_set *cset)
611 lockdep_assert_held(&css_set_lock);
613 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
617 * cgroup_update_populated - updated populated count of a cgroup
618 * @cgrp: the target cgroup
619 * @populated: inc or dec populated count
621 * One of the css_sets associated with @cgrp is either getting its first
622 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
623 * count is propagated towards root so that a given cgroup's populated_cnt
624 * is zero iff the cgroup and all its descendants don't contain any tasks.
626 * @cgrp's interface file "cgroup.populated" is zero if
627 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
628 * changes from or to zero, userland is notified that the content of the
629 * interface file has changed. This can be used to detect when @cgrp and
630 * its descendants become populated or empty.
632 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
634 lockdep_assert_held(&css_set_lock);
640 trigger = !cgrp->populated_cnt++;
642 trigger = !--cgrp->populated_cnt;
647 check_for_release(cgrp);
648 cgroup_file_notify(&cgrp->events_file);
650 cgrp = cgroup_parent(cgrp);
655 * css_set_update_populated - update populated state of a css_set
656 * @cset: target css_set
657 * @populated: whether @cset is populated or depopulated
659 * @cset is either getting the first task or losing the last. Update the
660 * ->populated_cnt of all associated cgroups accordingly.
662 static void css_set_update_populated(struct css_set *cset, bool populated)
664 struct cgrp_cset_link *link;
666 lockdep_assert_held(&css_set_lock);
668 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
669 cgroup_update_populated(link->cgrp, populated);
673 * css_set_move_task - move a task from one css_set to another
674 * @task: task being moved
675 * @from_cset: css_set @task currently belongs to (may be NULL)
676 * @to_cset: new css_set @task is being moved to (may be NULL)
677 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
679 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
680 * css_set, @from_cset can be NULL. If @task is being disassociated
681 * instead of moved, @to_cset can be NULL.
683 * This function automatically handles populated_cnt updates and
684 * css_task_iter adjustments but the caller is responsible for managing
685 * @from_cset and @to_cset's reference counts.
687 static void css_set_move_task(struct task_struct *task,
688 struct css_set *from_cset, struct css_set *to_cset,
691 lockdep_assert_held(&css_set_lock);
694 struct css_task_iter *it, *pos;
696 WARN_ON_ONCE(list_empty(&task->cg_list));
699 * @task is leaving, advance task iterators which are
700 * pointing to it so that they can resume at the next
701 * position. Advancing an iterator might remove it from
702 * the list, use safe walk. See css_task_iter_advance*()
705 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
707 if (it->task_pos == &task->cg_list)
708 css_task_iter_advance(it);
710 list_del_init(&task->cg_list);
711 if (!css_set_populated(from_cset))
712 css_set_update_populated(from_cset, false);
714 WARN_ON_ONCE(!list_empty(&task->cg_list));
719 * We are synchronized through cgroup_threadgroup_rwsem
720 * against PF_EXITING setting such that we can't race
721 * against cgroup_exit() changing the css_set to
722 * init_css_set and dropping the old one.
724 WARN_ON_ONCE(task->flags & PF_EXITING);
726 if (!css_set_populated(to_cset))
727 css_set_update_populated(to_cset, true);
728 rcu_assign_pointer(task->cgroups, to_cset);
729 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
735 * hash table for cgroup groups. This improves the performance to find
736 * an existing css_set. This hash doesn't (currently) take into
737 * account cgroups in empty hierarchies.
739 #define CSS_SET_HASH_BITS 7
740 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
742 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
744 unsigned long key = 0UL;
745 struct cgroup_subsys *ss;
748 for_each_subsys(ss, i)
749 key += (unsigned long)css[i];
750 key = (key >> 16) ^ key;
755 static void put_css_set_locked(struct css_set *cset)
757 struct cgrp_cset_link *link, *tmp_link;
758 struct cgroup_subsys *ss;
761 lockdep_assert_held(&css_set_lock);
763 if (!atomic_dec_and_test(&cset->refcount))
766 /* This css_set is dead. unlink it and release cgroup and css refs */
767 for_each_subsys(ss, ssid) {
768 list_del(&cset->e_cset_node[ssid]);
769 css_put(cset->subsys[ssid]);
771 hash_del(&cset->hlist);
774 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
775 list_del(&link->cset_link);
776 list_del(&link->cgrp_link);
777 if (cgroup_parent(link->cgrp))
778 cgroup_put(link->cgrp);
782 kfree_rcu(cset, rcu_head);
785 static void put_css_set(struct css_set *cset)
788 * Ensure that the refcount doesn't hit zero while any readers
789 * can see it. Similar to atomic_dec_and_lock(), but for an
792 if (atomic_add_unless(&cset->refcount, -1, 1))
795 spin_lock_bh(&css_set_lock);
796 put_css_set_locked(cset);
797 spin_unlock_bh(&css_set_lock);
801 * refcounted get/put for css_set objects
803 static inline void get_css_set(struct css_set *cset)
805 atomic_inc(&cset->refcount);
809 * compare_css_sets - helper function for find_existing_css_set().
810 * @cset: candidate css_set being tested
811 * @old_cset: existing css_set for a task
812 * @new_cgrp: cgroup that's being entered by the task
813 * @template: desired set of css pointers in css_set (pre-calculated)
815 * Returns true if "cset" matches "old_cset" except for the hierarchy
816 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
818 static bool compare_css_sets(struct css_set *cset,
819 struct css_set *old_cset,
820 struct cgroup *new_cgrp,
821 struct cgroup_subsys_state *template[])
823 struct list_head *l1, *l2;
826 * On the default hierarchy, there can be csets which are
827 * associated with the same set of cgroups but different csses.
828 * Let's first ensure that csses match.
830 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
834 * Compare cgroup pointers in order to distinguish between
835 * different cgroups in hierarchies. As different cgroups may
836 * share the same effective css, this comparison is always
839 l1 = &cset->cgrp_links;
840 l2 = &old_cset->cgrp_links;
842 struct cgrp_cset_link *link1, *link2;
843 struct cgroup *cgrp1, *cgrp2;
847 /* See if we reached the end - both lists are equal length. */
848 if (l1 == &cset->cgrp_links) {
849 BUG_ON(l2 != &old_cset->cgrp_links);
852 BUG_ON(l2 == &old_cset->cgrp_links);
854 /* Locate the cgroups associated with these links. */
855 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
856 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
859 /* Hierarchies should be linked in the same order. */
860 BUG_ON(cgrp1->root != cgrp2->root);
863 * If this hierarchy is the hierarchy of the cgroup
864 * that's changing, then we need to check that this
865 * css_set points to the new cgroup; if it's any other
866 * hierarchy, then this css_set should point to the
867 * same cgroup as the old css_set.
869 if (cgrp1->root == new_cgrp->root) {
870 if (cgrp1 != new_cgrp)
881 * find_existing_css_set - init css array and find the matching css_set
882 * @old_cset: the css_set that we're using before the cgroup transition
883 * @cgrp: the cgroup that we're moving into
884 * @template: out param for the new set of csses, should be clear on entry
886 static struct css_set *find_existing_css_set(struct css_set *old_cset,
888 struct cgroup_subsys_state *template[])
890 struct cgroup_root *root = cgrp->root;
891 struct cgroup_subsys *ss;
892 struct css_set *cset;
897 * Build the set of subsystem state objects that we want to see in the
898 * new css_set. while subsystems can change globally, the entries here
899 * won't change, so no need for locking.
901 for_each_subsys(ss, i) {
902 if (root->subsys_mask & (1UL << i)) {
904 * @ss is in this hierarchy, so we want the
905 * effective css from @cgrp.
907 template[i] = cgroup_e_css(cgrp, ss);
910 * @ss is not in this hierarchy, so we don't want
913 template[i] = old_cset->subsys[i];
917 key = css_set_hash(template);
918 hash_for_each_possible(css_set_table, cset, hlist, key) {
919 if (!compare_css_sets(cset, old_cset, cgrp, template))
922 /* This css_set matches what we need */
926 /* No existing cgroup group matched */
930 static void free_cgrp_cset_links(struct list_head *links_to_free)
932 struct cgrp_cset_link *link, *tmp_link;
934 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
935 list_del(&link->cset_link);
941 * allocate_cgrp_cset_links - allocate cgrp_cset_links
942 * @count: the number of links to allocate
943 * @tmp_links: list_head the allocated links are put on
945 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
946 * through ->cset_link. Returns 0 on success or -errno.
948 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
950 struct cgrp_cset_link *link;
953 INIT_LIST_HEAD(tmp_links);
955 for (i = 0; i < count; i++) {
956 link = kzalloc(sizeof(*link), GFP_KERNEL);
958 free_cgrp_cset_links(tmp_links);
961 list_add(&link->cset_link, tmp_links);
967 * link_css_set - a helper function to link a css_set to a cgroup
968 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
969 * @cset: the css_set to be linked
970 * @cgrp: the destination cgroup
972 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
975 struct cgrp_cset_link *link;
977 BUG_ON(list_empty(tmp_links));
979 if (cgroup_on_dfl(cgrp))
980 cset->dfl_cgrp = cgrp;
982 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
987 * Always add links to the tail of the lists so that the lists are
988 * in choronological order.
990 list_move_tail(&link->cset_link, &cgrp->cset_links);
991 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
993 if (cgroup_parent(cgrp))
998 * find_css_set - return a new css_set with one cgroup updated
999 * @old_cset: the baseline css_set
1000 * @cgrp: the cgroup to be updated
1002 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1003 * substituted into the appropriate hierarchy.
1005 static struct css_set *find_css_set(struct css_set *old_cset,
1006 struct cgroup *cgrp)
1008 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1009 struct css_set *cset;
1010 struct list_head tmp_links;
1011 struct cgrp_cset_link *link;
1012 struct cgroup_subsys *ss;
1016 lockdep_assert_held(&cgroup_mutex);
1018 /* First see if we already have a cgroup group that matches
1019 * the desired set */
1020 spin_lock_bh(&css_set_lock);
1021 cset = find_existing_css_set(old_cset, cgrp, template);
1024 spin_unlock_bh(&css_set_lock);
1029 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1033 /* Allocate all the cgrp_cset_link objects that we'll need */
1034 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1039 atomic_set(&cset->refcount, 1);
1040 INIT_LIST_HEAD(&cset->cgrp_links);
1041 INIT_LIST_HEAD(&cset->tasks);
1042 INIT_LIST_HEAD(&cset->mg_tasks);
1043 INIT_LIST_HEAD(&cset->mg_preload_node);
1044 INIT_LIST_HEAD(&cset->mg_node);
1045 INIT_LIST_HEAD(&cset->task_iters);
1046 INIT_HLIST_NODE(&cset->hlist);
1048 /* Copy the set of subsystem state objects generated in
1049 * find_existing_css_set() */
1050 memcpy(cset->subsys, template, sizeof(cset->subsys));
1052 spin_lock_bh(&css_set_lock);
1053 /* Add reference counts and links from the new css_set. */
1054 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1055 struct cgroup *c = link->cgrp;
1057 if (c->root == cgrp->root)
1059 link_css_set(&tmp_links, cset, c);
1062 BUG_ON(!list_empty(&tmp_links));
1066 /* Add @cset to the hash table */
1067 key = css_set_hash(cset->subsys);
1068 hash_add(css_set_table, &cset->hlist, key);
1070 for_each_subsys(ss, ssid) {
1071 struct cgroup_subsys_state *css = cset->subsys[ssid];
1073 list_add_tail(&cset->e_cset_node[ssid],
1074 &css->cgroup->e_csets[ssid]);
1078 spin_unlock_bh(&css_set_lock);
1083 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1085 struct cgroup *root_cgrp = kf_root->kn->priv;
1087 return root_cgrp->root;
1090 static int cgroup_init_root_id(struct cgroup_root *root)
1094 lockdep_assert_held(&cgroup_mutex);
1096 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1100 root->hierarchy_id = id;
1104 static void cgroup_exit_root_id(struct cgroup_root *root)
1106 lockdep_assert_held(&cgroup_mutex);
1108 if (root->hierarchy_id) {
1109 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1110 root->hierarchy_id = 0;
1114 static void cgroup_free_root(struct cgroup_root *root)
1117 /* hierarchy ID should already have been released */
1118 WARN_ON_ONCE(root->hierarchy_id);
1120 idr_destroy(&root->cgroup_idr);
1125 static void cgroup_destroy_root(struct cgroup_root *root)
1127 struct cgroup *cgrp = &root->cgrp;
1128 struct cgrp_cset_link *link, *tmp_link;
1130 mutex_lock(&cgroup_mutex);
1132 BUG_ON(atomic_read(&root->nr_cgrps));
1133 BUG_ON(!list_empty(&cgrp->self.children));
1135 /* Rebind all subsystems back to the default hierarchy */
1136 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1139 * Release all the links from cset_links to this hierarchy's
1142 spin_lock_bh(&css_set_lock);
1144 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1145 list_del(&link->cset_link);
1146 list_del(&link->cgrp_link);
1150 spin_unlock_bh(&css_set_lock);
1152 if (!list_empty(&root->root_list)) {
1153 list_del(&root->root_list);
1154 cgroup_root_count--;
1157 cgroup_exit_root_id(root);
1159 mutex_unlock(&cgroup_mutex);
1161 kernfs_destroy_root(root->kf_root);
1162 cgroup_free_root(root);
1165 /* look up cgroup associated with given css_set on the specified hierarchy */
1166 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1167 struct cgroup_root *root)
1169 struct cgroup *res = NULL;
1171 lockdep_assert_held(&cgroup_mutex);
1172 lockdep_assert_held(&css_set_lock);
1174 if (cset == &init_css_set) {
1177 struct cgrp_cset_link *link;
1179 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1180 struct cgroup *c = link->cgrp;
1182 if (c->root == root) {
1194 * Return the cgroup for "task" from the given hierarchy. Must be
1195 * called with cgroup_mutex and css_set_lock held.
1197 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1198 struct cgroup_root *root)
1201 * No need to lock the task - since we hold cgroup_mutex the
1202 * task can't change groups, so the only thing that can happen
1203 * is that it exits and its css is set back to init_css_set.
1205 return cset_cgroup_from_root(task_css_set(task), root);
1209 * A task must hold cgroup_mutex to modify cgroups.
1211 * Any task can increment and decrement the count field without lock.
1212 * So in general, code holding cgroup_mutex can't rely on the count
1213 * field not changing. However, if the count goes to zero, then only
1214 * cgroup_attach_task() can increment it again. Because a count of zero
1215 * means that no tasks are currently attached, therefore there is no
1216 * way a task attached to that cgroup can fork (the other way to
1217 * increment the count). So code holding cgroup_mutex can safely
1218 * assume that if the count is zero, it will stay zero. Similarly, if
1219 * a task holds cgroup_mutex on a cgroup with zero count, it
1220 * knows that the cgroup won't be removed, as cgroup_rmdir()
1223 * A cgroup can only be deleted if both its 'count' of using tasks
1224 * is zero, and its list of 'children' cgroups is empty. Since all
1225 * tasks in the system use _some_ cgroup, and since there is always at
1226 * least one task in the system (init, pid == 1), therefore, root cgroup
1227 * always has either children cgroups and/or using tasks. So we don't
1228 * need a special hack to ensure that root cgroup cannot be deleted.
1230 * P.S. One more locking exception. RCU is used to guard the
1231 * update of a tasks cgroup pointer by cgroup_attach_task()
1234 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1235 static const struct file_operations proc_cgroupstats_operations;
1237 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1240 struct cgroup_subsys *ss = cft->ss;
1242 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1243 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1244 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1245 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1248 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1253 * cgroup_file_mode - deduce file mode of a control file
1254 * @cft: the control file in question
1256 * S_IRUGO for read, S_IWUSR for write.
1258 static umode_t cgroup_file_mode(const struct cftype *cft)
1262 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1265 if (cft->write_u64 || cft->write_s64 || cft->write) {
1266 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1276 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1277 * @cgrp: the target cgroup
1278 * @subtree_control: the new subtree_control mask to consider
1280 * On the default hierarchy, a subsystem may request other subsystems to be
1281 * enabled together through its ->depends_on mask. In such cases, more
1282 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1284 * This function calculates which subsystems need to be enabled if
1285 * @subtree_control is to be applied to @cgrp. The returned mask is always
1286 * a superset of @subtree_control and follows the usual hierarchy rules.
1288 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1289 unsigned long subtree_control)
1291 struct cgroup *parent = cgroup_parent(cgrp);
1292 unsigned long cur_ss_mask = subtree_control;
1293 struct cgroup_subsys *ss;
1296 lockdep_assert_held(&cgroup_mutex);
1298 if (!cgroup_on_dfl(cgrp))
1302 unsigned long new_ss_mask = cur_ss_mask;
1304 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1305 new_ss_mask |= ss->depends_on;
1308 * Mask out subsystems which aren't available. This can
1309 * happen only if some depended-upon subsystems were bound
1310 * to non-default hierarchies.
1313 new_ss_mask &= parent->child_subsys_mask;
1315 new_ss_mask &= cgrp->root->subsys_mask;
1317 if (new_ss_mask == cur_ss_mask)
1319 cur_ss_mask = new_ss_mask;
1326 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1327 * @cgrp: the target cgroup
1329 * Update @cgrp->child_subsys_mask according to the current
1330 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1332 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1334 cgrp->child_subsys_mask =
1335 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1339 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1340 * @kn: the kernfs_node being serviced
1342 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1343 * the method finishes if locking succeeded. Note that once this function
1344 * returns the cgroup returned by cgroup_kn_lock_live() may become
1345 * inaccessible any time. If the caller intends to continue to access the
1346 * cgroup, it should pin it before invoking this function.
1348 static void cgroup_kn_unlock(struct kernfs_node *kn)
1350 struct cgroup *cgrp;
1352 if (kernfs_type(kn) == KERNFS_DIR)
1355 cgrp = kn->parent->priv;
1357 mutex_unlock(&cgroup_mutex);
1359 kernfs_unbreak_active_protection(kn);
1364 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1365 * @kn: the kernfs_node being serviced
1367 * This helper is to be used by a cgroup kernfs method currently servicing
1368 * @kn. It breaks the active protection, performs cgroup locking and
1369 * verifies that the associated cgroup is alive. Returns the cgroup if
1370 * alive; otherwise, %NULL. A successful return should be undone by a
1371 * matching cgroup_kn_unlock() invocation.
1373 * Any cgroup kernfs method implementation which requires locking the
1374 * associated cgroup should use this helper. It avoids nesting cgroup
1375 * locking under kernfs active protection and allows all kernfs operations
1376 * including self-removal.
1378 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1380 struct cgroup *cgrp;
1382 if (kernfs_type(kn) == KERNFS_DIR)
1385 cgrp = kn->parent->priv;
1388 * We're gonna grab cgroup_mutex which nests outside kernfs
1389 * active_ref. cgroup liveliness check alone provides enough
1390 * protection against removal. Ensure @cgrp stays accessible and
1391 * break the active_ref protection.
1393 if (!cgroup_tryget(cgrp))
1395 kernfs_break_active_protection(kn);
1397 mutex_lock(&cgroup_mutex);
1399 if (!cgroup_is_dead(cgrp))
1402 cgroup_kn_unlock(kn);
1406 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1408 char name[CGROUP_FILE_NAME_MAX];
1410 lockdep_assert_held(&cgroup_mutex);
1412 if (cft->file_offset) {
1413 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1414 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1416 spin_lock_irq(&cgroup_file_kn_lock);
1418 spin_unlock_irq(&cgroup_file_kn_lock);
1421 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1425 * css_clear_dir - remove subsys files in a cgroup directory
1427 * @cgrp_override: specify if target cgroup is different from css->cgroup
1429 static void css_clear_dir(struct cgroup_subsys_state *css,
1430 struct cgroup *cgrp_override)
1432 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1433 struct cftype *cfts;
1435 list_for_each_entry(cfts, &css->ss->cfts, node)
1436 cgroup_addrm_files(css, cgrp, cfts, false);
1440 * css_populate_dir - create subsys files in a cgroup directory
1442 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1444 * On failure, no file is added.
1446 static int css_populate_dir(struct cgroup_subsys_state *css,
1447 struct cgroup *cgrp_override)
1449 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1450 struct cftype *cfts, *failed_cfts;
1454 if (cgroup_on_dfl(cgrp))
1455 cfts = cgroup_dfl_base_files;
1457 cfts = cgroup_legacy_base_files;
1459 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1462 list_for_each_entry(cfts, &css->ss->cfts, node) {
1463 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1471 list_for_each_entry(cfts, &css->ss->cfts, node) {
1472 if (cfts == failed_cfts)
1474 cgroup_addrm_files(css, cgrp, cfts, false);
1479 static int rebind_subsystems(struct cgroup_root *dst_root,
1480 unsigned long ss_mask)
1482 struct cgroup *dcgrp = &dst_root->cgrp;
1483 struct cgroup_subsys *ss;
1484 unsigned long tmp_ss_mask;
1487 lockdep_assert_held(&cgroup_mutex);
1489 for_each_subsys_which(ss, ssid, &ss_mask) {
1490 /* if @ss has non-root csses attached to it, can't move */
1491 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1494 /* can't move between two non-dummy roots either */
1495 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1499 /* skip creating root files on dfl_root for inhibited subsystems */
1500 tmp_ss_mask = ss_mask;
1501 if (dst_root == &cgrp_dfl_root)
1502 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1504 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1505 struct cgroup *scgrp = &ss->root->cgrp;
1508 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1513 * Rebinding back to the default root is not allowed to
1514 * fail. Using both default and non-default roots should
1515 * be rare. Moving subsystems back and forth even more so.
1516 * Just warn about it and continue.
1518 if (dst_root == &cgrp_dfl_root) {
1519 if (cgrp_dfl_root_visible) {
1520 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1522 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1527 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1530 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1536 * Nothing can fail from this point on. Remove files for the
1537 * removed subsystems and rebind each subsystem.
1539 for_each_subsys_which(ss, ssid, &ss_mask) {
1540 struct cgroup_root *src_root = ss->root;
1541 struct cgroup *scgrp = &src_root->cgrp;
1542 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1543 struct css_set *cset;
1545 WARN_ON(!css || cgroup_css(dcgrp, ss));
1547 css_clear_dir(css, NULL);
1549 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1550 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1551 ss->root = dst_root;
1552 css->cgroup = dcgrp;
1554 spin_lock_bh(&css_set_lock);
1555 hash_for_each(css_set_table, i, cset, hlist)
1556 list_move_tail(&cset->e_cset_node[ss->id],
1557 &dcgrp->e_csets[ss->id]);
1558 spin_unlock_bh(&css_set_lock);
1560 src_root->subsys_mask &= ~(1 << ssid);
1561 scgrp->subtree_control &= ~(1 << ssid);
1562 cgroup_refresh_child_subsys_mask(scgrp);
1564 /* default hierarchy doesn't enable controllers by default */
1565 dst_root->subsys_mask |= 1 << ssid;
1566 if (dst_root == &cgrp_dfl_root) {
1567 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1569 dcgrp->subtree_control |= 1 << ssid;
1570 cgroup_refresh_child_subsys_mask(dcgrp);
1571 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1578 kernfs_activate(dcgrp->kn);
1582 static int cgroup_show_options(struct seq_file *seq,
1583 struct kernfs_root *kf_root)
1585 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1586 struct cgroup_subsys *ss;
1589 if (root != &cgrp_dfl_root)
1590 for_each_subsys(ss, ssid)
1591 if (root->subsys_mask & (1 << ssid))
1592 seq_show_option(seq, ss->legacy_name, NULL);
1593 if (root->flags & CGRP_ROOT_NOPREFIX)
1594 seq_puts(seq, ",noprefix");
1595 if (root->flags & CGRP_ROOT_XATTR)
1596 seq_puts(seq, ",xattr");
1598 spin_lock(&release_agent_path_lock);
1599 if (strlen(root->release_agent_path))
1600 seq_show_option(seq, "release_agent",
1601 root->release_agent_path);
1602 spin_unlock(&release_agent_path_lock);
1604 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1605 seq_puts(seq, ",clone_children");
1606 if (strlen(root->name))
1607 seq_show_option(seq, "name", root->name);
1611 struct cgroup_sb_opts {
1612 unsigned long subsys_mask;
1614 char *release_agent;
1615 bool cpuset_clone_children;
1617 /* User explicitly requested empty subsystem */
1621 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1623 char *token, *o = data;
1624 bool all_ss = false, one_ss = false;
1625 unsigned long mask = -1UL;
1626 struct cgroup_subsys *ss;
1630 #ifdef CONFIG_CPUSETS
1631 mask = ~(1U << cpuset_cgrp_id);
1634 memset(opts, 0, sizeof(*opts));
1636 while ((token = strsep(&o, ",")) != NULL) {
1641 if (!strcmp(token, "none")) {
1642 /* Explicitly have no subsystems */
1646 if (!strcmp(token, "all")) {
1647 /* Mutually exclusive option 'all' + subsystem name */
1653 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1654 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1657 if (!strcmp(token, "noprefix")) {
1658 opts->flags |= CGRP_ROOT_NOPREFIX;
1661 if (!strcmp(token, "clone_children")) {
1662 opts->cpuset_clone_children = true;
1665 if (!strcmp(token, "xattr")) {
1666 opts->flags |= CGRP_ROOT_XATTR;
1669 if (!strncmp(token, "release_agent=", 14)) {
1670 /* Specifying two release agents is forbidden */
1671 if (opts->release_agent)
1673 opts->release_agent =
1674 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1675 if (!opts->release_agent)
1679 if (!strncmp(token, "name=", 5)) {
1680 const char *name = token + 5;
1681 /* Can't specify an empty name */
1684 /* Must match [\w.-]+ */
1685 for (i = 0; i < strlen(name); i++) {
1689 if ((c == '.') || (c == '-') || (c == '_'))
1693 /* Specifying two names is forbidden */
1696 opts->name = kstrndup(name,
1697 MAX_CGROUP_ROOT_NAMELEN - 1,
1705 for_each_subsys(ss, i) {
1706 if (strcmp(token, ss->legacy_name))
1708 if (!cgroup_ssid_enabled(i))
1711 /* Mutually exclusive option 'all' + subsystem name */
1714 opts->subsys_mask |= (1 << i);
1719 if (i == CGROUP_SUBSYS_COUNT)
1723 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1724 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1726 pr_err("sane_behavior: no other mount options allowed\n");
1733 * If the 'all' option was specified select all the subsystems,
1734 * otherwise if 'none', 'name=' and a subsystem name options were
1735 * not specified, let's default to 'all'
1737 if (all_ss || (!one_ss && !opts->none && !opts->name))
1738 for_each_subsys(ss, i)
1739 if (cgroup_ssid_enabled(i))
1740 opts->subsys_mask |= (1 << i);
1743 * We either have to specify by name or by subsystems. (So all
1744 * empty hierarchies must have a name).
1746 if (!opts->subsys_mask && !opts->name)
1750 * Option noprefix was introduced just for backward compatibility
1751 * with the old cpuset, so we allow noprefix only if mounting just
1752 * the cpuset subsystem.
1754 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1757 /* Can't specify "none" and some subsystems */
1758 if (opts->subsys_mask && opts->none)
1764 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1767 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1768 struct cgroup_sb_opts opts;
1769 unsigned long added_mask, removed_mask;
1771 if (root == &cgrp_dfl_root) {
1772 pr_err("remount is not allowed\n");
1776 mutex_lock(&cgroup_mutex);
1778 /* See what subsystems are wanted */
1779 ret = parse_cgroupfs_options(data, &opts);
1783 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1784 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1785 task_tgid_nr(current), current->comm);
1787 added_mask = opts.subsys_mask & ~root->subsys_mask;
1788 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1790 /* Don't allow flags or name to change at remount */
1791 if ((opts.flags ^ root->flags) ||
1792 (opts.name && strcmp(opts.name, root->name))) {
1793 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1794 opts.flags, opts.name ?: "", root->flags, root->name);
1799 /* remounting is not allowed for populated hierarchies */
1800 if (!list_empty(&root->cgrp.self.children)) {
1805 ret = rebind_subsystems(root, added_mask);
1809 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1811 if (opts.release_agent) {
1812 spin_lock(&release_agent_path_lock);
1813 strcpy(root->release_agent_path, opts.release_agent);
1814 spin_unlock(&release_agent_path_lock);
1817 kfree(opts.release_agent);
1819 mutex_unlock(&cgroup_mutex);
1824 * To reduce the fork() overhead for systems that are not actually using
1825 * their cgroups capability, we don't maintain the lists running through
1826 * each css_set to its tasks until we see the list actually used - in other
1827 * words after the first mount.
1829 static bool use_task_css_set_links __read_mostly;
1831 static void cgroup_enable_task_cg_lists(void)
1833 struct task_struct *p, *g;
1835 spin_lock_bh(&css_set_lock);
1837 if (use_task_css_set_links)
1840 use_task_css_set_links = true;
1843 * We need tasklist_lock because RCU is not safe against
1844 * while_each_thread(). Besides, a forking task that has passed
1845 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1846 * is not guaranteed to have its child immediately visible in the
1847 * tasklist if we walk through it with RCU.
1849 read_lock(&tasklist_lock);
1850 do_each_thread(g, p) {
1851 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1852 task_css_set(p) != &init_css_set);
1855 * We should check if the process is exiting, otherwise
1856 * it will race with cgroup_exit() in that the list
1857 * entry won't be deleted though the process has exited.
1858 * Do it while holding siglock so that we don't end up
1859 * racing against cgroup_exit().
1861 spin_lock_irq(&p->sighand->siglock);
1862 if (!(p->flags & PF_EXITING)) {
1863 struct css_set *cset = task_css_set(p);
1865 if (!css_set_populated(cset))
1866 css_set_update_populated(cset, true);
1867 list_add_tail(&p->cg_list, &cset->tasks);
1870 spin_unlock_irq(&p->sighand->siglock);
1871 } while_each_thread(g, p);
1872 read_unlock(&tasklist_lock);
1874 spin_unlock_bh(&css_set_lock);
1877 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1879 struct cgroup_subsys *ss;
1882 INIT_LIST_HEAD(&cgrp->self.sibling);
1883 INIT_LIST_HEAD(&cgrp->self.children);
1884 INIT_LIST_HEAD(&cgrp->cset_links);
1885 INIT_LIST_HEAD(&cgrp->pidlists);
1886 mutex_init(&cgrp->pidlist_mutex);
1887 cgrp->self.cgroup = cgrp;
1888 cgrp->self.flags |= CSS_ONLINE;
1890 for_each_subsys(ss, ssid)
1891 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1893 init_waitqueue_head(&cgrp->offline_waitq);
1894 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1897 static void init_cgroup_root(struct cgroup_root *root,
1898 struct cgroup_sb_opts *opts)
1900 struct cgroup *cgrp = &root->cgrp;
1902 INIT_LIST_HEAD(&root->root_list);
1903 atomic_set(&root->nr_cgrps, 1);
1905 init_cgroup_housekeeping(cgrp);
1906 idr_init(&root->cgroup_idr);
1908 root->flags = opts->flags;
1909 if (opts->release_agent)
1910 strcpy(root->release_agent_path, opts->release_agent);
1912 strcpy(root->name, opts->name);
1913 if (opts->cpuset_clone_children)
1914 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1917 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1919 LIST_HEAD(tmp_links);
1920 struct cgroup *root_cgrp = &root->cgrp;
1921 struct css_set *cset;
1924 lockdep_assert_held(&cgroup_mutex);
1926 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1929 root_cgrp->id = ret;
1931 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1937 * We're accessing css_set_count without locking css_set_lock here,
1938 * but that's OK - it can only be increased by someone holding
1939 * cgroup_lock, and that's us. The worst that can happen is that we
1940 * have some link structures left over
1942 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1946 ret = cgroup_init_root_id(root);
1950 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1951 KERNFS_ROOT_CREATE_DEACTIVATED,
1953 if (IS_ERR(root->kf_root)) {
1954 ret = PTR_ERR(root->kf_root);
1957 root_cgrp->kn = root->kf_root->kn;
1959 ret = css_populate_dir(&root_cgrp->self, NULL);
1963 ret = rebind_subsystems(root, ss_mask);
1968 * There must be no failure case after here, since rebinding takes
1969 * care of subsystems' refcounts, which are explicitly dropped in
1970 * the failure exit path.
1972 list_add(&root->root_list, &cgroup_roots);
1973 cgroup_root_count++;
1976 * Link the root cgroup in this hierarchy into all the css_set
1979 spin_lock_bh(&css_set_lock);
1980 hash_for_each(css_set_table, i, cset, hlist) {
1981 link_css_set(&tmp_links, cset, root_cgrp);
1982 if (css_set_populated(cset))
1983 cgroup_update_populated(root_cgrp, true);
1985 spin_unlock_bh(&css_set_lock);
1987 BUG_ON(!list_empty(&root_cgrp->self.children));
1988 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1990 kernfs_activate(root_cgrp->kn);
1995 kernfs_destroy_root(root->kf_root);
1996 root->kf_root = NULL;
1998 cgroup_exit_root_id(root);
2000 percpu_ref_exit(&root_cgrp->self.refcnt);
2002 free_cgrp_cset_links(&tmp_links);
2006 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2007 int flags, const char *unused_dev_name,
2010 struct super_block *pinned_sb = NULL;
2011 struct cgroup_subsys *ss;
2012 struct cgroup_root *root;
2013 struct cgroup_sb_opts opts;
2014 struct dentry *dentry;
2020 * The first time anyone tries to mount a cgroup, enable the list
2021 * linking each css_set to its tasks and fix up all existing tasks.
2023 if (!use_task_css_set_links)
2024 cgroup_enable_task_cg_lists();
2026 mutex_lock(&cgroup_mutex);
2028 /* First find the desired set of subsystems */
2029 ret = parse_cgroupfs_options(data, &opts);
2033 /* look for a matching existing root */
2034 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
2035 cgrp_dfl_root_visible = true;
2036 root = &cgrp_dfl_root;
2037 cgroup_get(&root->cgrp);
2043 * Destruction of cgroup root is asynchronous, so subsystems may
2044 * still be dying after the previous unmount. Let's drain the
2045 * dying subsystems. We just need to ensure that the ones
2046 * unmounted previously finish dying and don't care about new ones
2047 * starting. Testing ref liveliness is good enough.
2049 for_each_subsys(ss, i) {
2050 if (!(opts.subsys_mask & (1 << i)) ||
2051 ss->root == &cgrp_dfl_root)
2054 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2055 mutex_unlock(&cgroup_mutex);
2057 ret = restart_syscall();
2060 cgroup_put(&ss->root->cgrp);
2063 for_each_root(root) {
2064 bool name_match = false;
2066 if (root == &cgrp_dfl_root)
2070 * If we asked for a name then it must match. Also, if
2071 * name matches but sybsys_mask doesn't, we should fail.
2072 * Remember whether name matched.
2075 if (strcmp(opts.name, root->name))
2081 * If we asked for subsystems (or explicitly for no
2082 * subsystems) then they must match.
2084 if ((opts.subsys_mask || opts.none) &&
2085 (opts.subsys_mask != root->subsys_mask)) {
2092 if (root->flags ^ opts.flags)
2093 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2096 * We want to reuse @root whose lifetime is governed by its
2097 * ->cgrp. Let's check whether @root is alive and keep it
2098 * that way. As cgroup_kill_sb() can happen anytime, we
2099 * want to block it by pinning the sb so that @root doesn't
2100 * get killed before mount is complete.
2102 * With the sb pinned, tryget_live can reliably indicate
2103 * whether @root can be reused. If it's being killed,
2104 * drain it. We can use wait_queue for the wait but this
2105 * path is super cold. Let's just sleep a bit and retry.
2107 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2108 if (IS_ERR(pinned_sb) ||
2109 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2110 mutex_unlock(&cgroup_mutex);
2111 if (!IS_ERR_OR_NULL(pinned_sb))
2112 deactivate_super(pinned_sb);
2114 ret = restart_syscall();
2123 * No such thing, create a new one. name= matching without subsys
2124 * specification is allowed for already existing hierarchies but we
2125 * can't create new one without subsys specification.
2127 if (!opts.subsys_mask && !opts.none) {
2132 root = kzalloc(sizeof(*root), GFP_KERNEL);
2138 init_cgroup_root(root, &opts);
2140 ret = cgroup_setup_root(root, opts.subsys_mask);
2142 cgroup_free_root(root);
2145 mutex_unlock(&cgroup_mutex);
2147 kfree(opts.release_agent);
2151 return ERR_PTR(ret);
2153 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2154 CGROUP_SUPER_MAGIC, &new_sb);
2155 if (IS_ERR(dentry) || !new_sb)
2156 cgroup_put(&root->cgrp);
2159 * If @pinned_sb, we're reusing an existing root and holding an
2160 * extra ref on its sb. Mount is complete. Put the extra ref.
2164 deactivate_super(pinned_sb);
2170 static void cgroup_kill_sb(struct super_block *sb)
2172 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2173 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2176 * If @root doesn't have any mounts or children, start killing it.
2177 * This prevents new mounts by disabling percpu_ref_tryget_live().
2178 * cgroup_mount() may wait for @root's release.
2180 * And don't kill the default root.
2182 if (!list_empty(&root->cgrp.self.children) ||
2183 root == &cgrp_dfl_root)
2184 cgroup_put(&root->cgrp);
2186 percpu_ref_kill(&root->cgrp.self.refcnt);
2191 static struct file_system_type cgroup_fs_type = {
2193 .mount = cgroup_mount,
2194 .kill_sb = cgroup_kill_sb,
2198 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2199 * @task: target task
2200 * @buf: the buffer to write the path into
2201 * @buflen: the length of the buffer
2203 * Determine @task's cgroup on the first (the one with the lowest non-zero
2204 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2205 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2206 * cgroup controller callbacks.
2208 * Return value is the same as kernfs_path().
2210 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2212 struct cgroup_root *root;
2213 struct cgroup *cgrp;
2214 int hierarchy_id = 1;
2217 mutex_lock(&cgroup_mutex);
2218 spin_lock_bh(&css_set_lock);
2220 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2223 cgrp = task_cgroup_from_root(task, root);
2224 path = cgroup_path(cgrp, buf, buflen);
2226 /* if no hierarchy exists, everyone is in "/" */
2227 if (strlcpy(buf, "/", buflen) < buflen)
2231 spin_unlock_bh(&css_set_lock);
2232 mutex_unlock(&cgroup_mutex);
2235 EXPORT_SYMBOL_GPL(task_cgroup_path);
2237 /* used to track tasks and other necessary states during migration */
2238 struct cgroup_taskset {
2239 /* the src and dst cset list running through cset->mg_node */
2240 struct list_head src_csets;
2241 struct list_head dst_csets;
2243 /* the subsys currently being processed */
2247 * Fields for cgroup_taskset_*() iteration.
2249 * Before migration is committed, the target migration tasks are on
2250 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2251 * the csets on ->dst_csets. ->csets point to either ->src_csets
2252 * or ->dst_csets depending on whether migration is committed.
2254 * ->cur_csets and ->cur_task point to the current task position
2257 struct list_head *csets;
2258 struct css_set *cur_cset;
2259 struct task_struct *cur_task;
2262 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2263 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2264 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2265 .csets = &tset.src_csets, \
2269 * cgroup_taskset_add - try to add a migration target task to a taskset
2270 * @task: target task
2271 * @tset: target taskset
2273 * Add @task, which is a migration target, to @tset. This function becomes
2274 * noop if @task doesn't need to be migrated. @task's css_set should have
2275 * been added as a migration source and @task->cg_list will be moved from
2276 * the css_set's tasks list to mg_tasks one.
2278 static void cgroup_taskset_add(struct task_struct *task,
2279 struct cgroup_taskset *tset)
2281 struct css_set *cset;
2283 lockdep_assert_held(&css_set_lock);
2285 /* @task either already exited or can't exit until the end */
2286 if (task->flags & PF_EXITING)
2289 /* leave @task alone if post_fork() hasn't linked it yet */
2290 if (list_empty(&task->cg_list))
2293 cset = task_css_set(task);
2294 if (!cset->mg_src_cgrp)
2297 list_move_tail(&task->cg_list, &cset->mg_tasks);
2298 if (list_empty(&cset->mg_node))
2299 list_add_tail(&cset->mg_node, &tset->src_csets);
2300 if (list_empty(&cset->mg_dst_cset->mg_node))
2301 list_move_tail(&cset->mg_dst_cset->mg_node,
2306 * cgroup_taskset_first - reset taskset and return the first task
2307 * @tset: taskset of interest
2308 * @dst_cssp: output variable for the destination css
2310 * @tset iteration is initialized and the first task is returned.
2312 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2313 struct cgroup_subsys_state **dst_cssp)
2315 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2316 tset->cur_task = NULL;
2318 return cgroup_taskset_next(tset, dst_cssp);
2322 * cgroup_taskset_next - iterate to the next task in taskset
2323 * @tset: taskset of interest
2324 * @dst_cssp: output variable for the destination css
2326 * Return the next task in @tset. Iteration must have been initialized
2327 * with cgroup_taskset_first().
2329 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2330 struct cgroup_subsys_state **dst_cssp)
2332 struct css_set *cset = tset->cur_cset;
2333 struct task_struct *task = tset->cur_task;
2335 while (&cset->mg_node != tset->csets) {
2337 task = list_first_entry(&cset->mg_tasks,
2338 struct task_struct, cg_list);
2340 task = list_next_entry(task, cg_list);
2342 if (&task->cg_list != &cset->mg_tasks) {
2343 tset->cur_cset = cset;
2344 tset->cur_task = task;
2347 * This function may be called both before and
2348 * after cgroup_taskset_migrate(). The two cases
2349 * can be distinguished by looking at whether @cset
2350 * has its ->mg_dst_cset set.
2352 if (cset->mg_dst_cset)
2353 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2355 *dst_cssp = cset->subsys[tset->ssid];
2360 cset = list_next_entry(cset, mg_node);
2368 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2369 * @tset: taget taskset
2370 * @dst_cgrp: destination cgroup
2372 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2373 * ->can_attach callbacks fails and guarantees that either all or none of
2374 * the tasks in @tset are migrated. @tset is consumed regardless of
2377 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2378 struct cgroup *dst_cgrp)
2380 struct cgroup_subsys_state *css, *failed_css = NULL;
2381 struct task_struct *task, *tmp_task;
2382 struct css_set *cset, *tmp_cset;
2385 /* methods shouldn't be called if no task is actually migrating */
2386 if (list_empty(&tset->src_csets))
2389 /* check that we can legitimately attach to the cgroup */
2390 for_each_e_css(css, i, dst_cgrp) {
2391 if (css->ss->can_attach) {
2393 ret = css->ss->can_attach(tset);
2396 goto out_cancel_attach;
2402 * Now that we're guaranteed success, proceed to move all tasks to
2403 * the new cgroup. There are no failure cases after here, so this
2404 * is the commit point.
2406 spin_lock_bh(&css_set_lock);
2407 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2408 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2409 struct css_set *from_cset = task_css_set(task);
2410 struct css_set *to_cset = cset->mg_dst_cset;
2412 get_css_set(to_cset);
2413 css_set_move_task(task, from_cset, to_cset, true);
2414 put_css_set_locked(from_cset);
2417 spin_unlock_bh(&css_set_lock);
2420 * Migration is committed, all target tasks are now on dst_csets.
2421 * Nothing is sensitive to fork() after this point. Notify
2422 * controllers that migration is complete.
2424 tset->csets = &tset->dst_csets;
2426 for_each_e_css(css, i, dst_cgrp) {
2427 if (css->ss->attach) {
2429 css->ss->attach(tset);
2434 goto out_release_tset;
2437 for_each_e_css(css, i, dst_cgrp) {
2438 if (css == failed_css)
2440 if (css->ss->cancel_attach) {
2442 css->ss->cancel_attach(tset);
2446 spin_lock_bh(&css_set_lock);
2447 list_splice_init(&tset->dst_csets, &tset->src_csets);
2448 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2449 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2450 list_del_init(&cset->mg_node);
2452 spin_unlock_bh(&css_set_lock);
2457 * cgroup_migrate_finish - cleanup after attach
2458 * @preloaded_csets: list of preloaded css_sets
2460 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2461 * those functions for details.
2463 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2465 struct css_set *cset, *tmp_cset;
2467 lockdep_assert_held(&cgroup_mutex);
2469 spin_lock_bh(&css_set_lock);
2470 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2471 cset->mg_src_cgrp = NULL;
2472 cset->mg_dst_cset = NULL;
2473 list_del_init(&cset->mg_preload_node);
2474 put_css_set_locked(cset);
2476 spin_unlock_bh(&css_set_lock);
2480 * cgroup_migrate_add_src - add a migration source css_set
2481 * @src_cset: the source css_set to add
2482 * @dst_cgrp: the destination cgroup
2483 * @preloaded_csets: list of preloaded css_sets
2485 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2486 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2487 * up by cgroup_migrate_finish().
2489 * This function may be called without holding cgroup_threadgroup_rwsem
2490 * even if the target is a process. Threads may be created and destroyed
2491 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2492 * into play and the preloaded css_sets are guaranteed to cover all
2495 static void cgroup_migrate_add_src(struct css_set *src_cset,
2496 struct cgroup *dst_cgrp,
2497 struct list_head *preloaded_csets)
2499 struct cgroup *src_cgrp;
2501 lockdep_assert_held(&cgroup_mutex);
2502 lockdep_assert_held(&css_set_lock);
2505 * If ->dead, @src_set is associated with one or more dead cgroups
2506 * and doesn't contain any migratable tasks. Ignore it early so
2507 * that the rest of migration path doesn't get confused by it.
2512 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2514 if (!list_empty(&src_cset->mg_preload_node))
2517 WARN_ON(src_cset->mg_src_cgrp);
2518 WARN_ON(!list_empty(&src_cset->mg_tasks));
2519 WARN_ON(!list_empty(&src_cset->mg_node));
2521 src_cset->mg_src_cgrp = src_cgrp;
2522 get_css_set(src_cset);
2523 list_add(&src_cset->mg_preload_node, preloaded_csets);
2527 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2528 * @dst_cgrp: the destination cgroup (may be %NULL)
2529 * @preloaded_csets: list of preloaded source css_sets
2531 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2532 * have been preloaded to @preloaded_csets. This function looks up and
2533 * pins all destination css_sets, links each to its source, and append them
2534 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2535 * source css_set is assumed to be its cgroup on the default hierarchy.
2537 * This function must be called after cgroup_migrate_add_src() has been
2538 * called on each migration source css_set. After migration is performed
2539 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2542 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2543 struct list_head *preloaded_csets)
2546 struct css_set *src_cset, *tmp_cset;
2548 lockdep_assert_held(&cgroup_mutex);
2551 * Except for the root, child_subsys_mask must be zero for a cgroup
2552 * with tasks so that child cgroups don't compete against tasks.
2554 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2555 dst_cgrp->child_subsys_mask)
2558 /* look up the dst cset for each src cset and link it to src */
2559 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2560 struct css_set *dst_cset;
2562 dst_cset = find_css_set(src_cset,
2563 dst_cgrp ?: src_cset->dfl_cgrp);
2567 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2570 * If src cset equals dst, it's noop. Drop the src.
2571 * cgroup_migrate() will skip the cset too. Note that we
2572 * can't handle src == dst as some nodes are used by both.
2574 if (src_cset == dst_cset) {
2575 src_cset->mg_src_cgrp = NULL;
2576 list_del_init(&src_cset->mg_preload_node);
2577 put_css_set(src_cset);
2578 put_css_set(dst_cset);
2582 src_cset->mg_dst_cset = dst_cset;
2584 if (list_empty(&dst_cset->mg_preload_node))
2585 list_add(&dst_cset->mg_preload_node, &csets);
2587 put_css_set(dst_cset);
2590 list_splice_tail(&csets, preloaded_csets);
2593 cgroup_migrate_finish(&csets);
2598 * cgroup_migrate - migrate a process or task to a cgroup
2599 * @leader: the leader of the process or the task to migrate
2600 * @threadgroup: whether @leader points to the whole process or a single task
2601 * @cgrp: the destination cgroup
2603 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2604 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2605 * caller is also responsible for invoking cgroup_migrate_add_src() and
2606 * cgroup_migrate_prepare_dst() on the targets before invoking this
2607 * function and following up with cgroup_migrate_finish().
2609 * As long as a controller's ->can_attach() doesn't fail, this function is
2610 * guaranteed to succeed. This means that, excluding ->can_attach()
2611 * failure, when migrating multiple targets, the success or failure can be
2612 * decided for all targets by invoking group_migrate_prepare_dst() before
2613 * actually starting migrating.
2615 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2616 struct cgroup *cgrp)
2618 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2619 struct task_struct *task;
2622 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2623 * already PF_EXITING could be freed from underneath us unless we
2624 * take an rcu_read_lock.
2626 spin_lock_bh(&css_set_lock);
2630 cgroup_taskset_add(task, &tset);
2633 } while_each_thread(leader, task);
2635 spin_unlock_bh(&css_set_lock);
2637 return cgroup_taskset_migrate(&tset, cgrp);
2641 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2642 * @dst_cgrp: the cgroup to attach to
2643 * @leader: the task or the leader of the threadgroup to be attached
2644 * @threadgroup: attach the whole threadgroup?
2646 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2648 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2649 struct task_struct *leader, bool threadgroup)
2651 LIST_HEAD(preloaded_csets);
2652 struct task_struct *task;
2655 /* look up all src csets */
2656 spin_lock_bh(&css_set_lock);
2660 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2664 } while_each_thread(leader, task);
2666 spin_unlock_bh(&css_set_lock);
2668 /* prepare dst csets and commit */
2669 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2671 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2673 cgroup_migrate_finish(&preloaded_csets);
2677 static int cgroup_procs_write_permission(struct task_struct *task,
2678 struct cgroup *dst_cgrp,
2679 struct kernfs_open_file *of)
2681 const struct cred *cred = current_cred();
2682 const struct cred *tcred = get_task_cred(task);
2686 * even if we're attaching all tasks in the thread group, we only
2687 * need to check permissions on one of them.
2689 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2690 !uid_eq(cred->euid, tcred->uid) &&
2691 !uid_eq(cred->euid, tcred->suid))
2694 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2695 struct super_block *sb = of->file->f_path.dentry->d_sb;
2696 struct cgroup *cgrp;
2697 struct inode *inode;
2699 spin_lock_bh(&css_set_lock);
2700 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2701 spin_unlock_bh(&css_set_lock);
2703 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2704 cgrp = cgroup_parent(cgrp);
2707 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2709 ret = inode_permission(inode, MAY_WRITE);
2719 * Find the task_struct of the task to attach by vpid and pass it along to the
2720 * function to attach either it or all tasks in its threadgroup. Will lock
2721 * cgroup_mutex and threadgroup.
2723 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2724 size_t nbytes, loff_t off, bool threadgroup)
2726 struct task_struct *tsk;
2727 struct cgroup_subsys *ss;
2728 struct cgroup *cgrp;
2732 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2735 cgrp = cgroup_kn_lock_live(of->kn);
2739 percpu_down_write(&cgroup_threadgroup_rwsem);
2742 tsk = find_task_by_vpid(pid);
2745 goto out_unlock_rcu;
2752 tsk = tsk->group_leader;
2755 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2756 * trapped in a cpuset, or RT worker may be born in a cgroup
2757 * with no rt_runtime allocated. Just say no.
2759 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2761 goto out_unlock_rcu;
2764 get_task_struct(tsk);
2767 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2769 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2771 put_task_struct(tsk);
2772 goto out_unlock_threadgroup;
2776 out_unlock_threadgroup:
2777 percpu_up_write(&cgroup_threadgroup_rwsem);
2778 for_each_subsys(ss, ssid)
2779 if (ss->post_attach)
2781 cgroup_kn_unlock(of->kn);
2782 return ret ?: nbytes;
2786 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2787 * @from: attach to all cgroups of a given task
2788 * @tsk: the task to be attached
2790 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2792 struct cgroup_root *root;
2795 mutex_lock(&cgroup_mutex);
2796 for_each_root(root) {
2797 struct cgroup *from_cgrp;
2799 if (root == &cgrp_dfl_root)
2802 spin_lock_bh(&css_set_lock);
2803 from_cgrp = task_cgroup_from_root(from, root);
2804 spin_unlock_bh(&css_set_lock);
2806 retval = cgroup_attach_task(from_cgrp, tsk, false);
2810 mutex_unlock(&cgroup_mutex);
2814 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2816 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2817 char *buf, size_t nbytes, loff_t off)
2819 return __cgroup_procs_write(of, buf, nbytes, off, false);
2822 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2823 char *buf, size_t nbytes, loff_t off)
2825 return __cgroup_procs_write(of, buf, nbytes, off, true);
2828 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2829 char *buf, size_t nbytes, loff_t off)
2831 struct cgroup *cgrp;
2833 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2835 cgrp = cgroup_kn_lock_live(of->kn);
2838 spin_lock(&release_agent_path_lock);
2839 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2840 sizeof(cgrp->root->release_agent_path));
2841 spin_unlock(&release_agent_path_lock);
2842 cgroup_kn_unlock(of->kn);
2846 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2848 struct cgroup *cgrp = seq_css(seq)->cgroup;
2850 spin_lock(&release_agent_path_lock);
2851 seq_puts(seq, cgrp->root->release_agent_path);
2852 spin_unlock(&release_agent_path_lock);
2853 seq_putc(seq, '\n');
2857 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2859 seq_puts(seq, "0\n");
2863 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2865 struct cgroup_subsys *ss;
2866 bool printed = false;
2869 for_each_subsys_which(ss, ssid, &ss_mask) {
2872 seq_printf(seq, "%s", ss->name);
2876 seq_putc(seq, '\n');
2879 /* show controllers which are currently attached to the default hierarchy */
2880 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2882 struct cgroup *cgrp = seq_css(seq)->cgroup;
2884 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2885 ~cgrp_dfl_root_inhibit_ss_mask);
2889 /* show controllers which are enabled from the parent */
2890 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2892 struct cgroup *cgrp = seq_css(seq)->cgroup;
2894 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2898 /* show controllers which are enabled for a given cgroup's children */
2899 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2901 struct cgroup *cgrp = seq_css(seq)->cgroup;
2903 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2908 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2909 * @cgrp: root of the subtree to update csses for
2911 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2912 * css associations need to be updated accordingly. This function looks up
2913 * all css_sets which are attached to the subtree, creates the matching
2914 * updated css_sets and migrates the tasks to the new ones.
2916 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2918 LIST_HEAD(preloaded_csets);
2919 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2920 struct cgroup_subsys_state *css;
2921 struct css_set *src_cset;
2924 lockdep_assert_held(&cgroup_mutex);
2926 percpu_down_write(&cgroup_threadgroup_rwsem);
2928 /* look up all csses currently attached to @cgrp's subtree */
2929 spin_lock_bh(&css_set_lock);
2930 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2931 struct cgrp_cset_link *link;
2933 /* self is not affected by child_subsys_mask change */
2934 if (css->cgroup == cgrp)
2937 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2938 cgroup_migrate_add_src(link->cset, cgrp,
2941 spin_unlock_bh(&css_set_lock);
2943 /* NULL dst indicates self on default hierarchy */
2944 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2948 spin_lock_bh(&css_set_lock);
2949 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2950 struct task_struct *task, *ntask;
2952 /* src_csets precede dst_csets, break on the first dst_cset */
2953 if (!src_cset->mg_src_cgrp)
2956 /* all tasks in src_csets need to be migrated */
2957 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2958 cgroup_taskset_add(task, &tset);
2960 spin_unlock_bh(&css_set_lock);
2962 ret = cgroup_taskset_migrate(&tset, cgrp);
2964 cgroup_migrate_finish(&preloaded_csets);
2965 percpu_up_write(&cgroup_threadgroup_rwsem);
2969 /* change the enabled child controllers for a cgroup in the default hierarchy */
2970 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2971 char *buf, size_t nbytes,
2974 unsigned long enable = 0, disable = 0;
2975 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2976 struct cgroup *cgrp, *child;
2977 struct cgroup_subsys *ss;
2982 * Parse input - space separated list of subsystem names prefixed
2983 * with either + or -.
2985 buf = strstrip(buf);
2986 while ((tok = strsep(&buf, " "))) {
2987 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2991 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2992 if (!cgroup_ssid_enabled(ssid) ||
2993 strcmp(tok + 1, ss->name))
2997 enable |= 1 << ssid;
2998 disable &= ~(1 << ssid);
2999 } else if (*tok == '-') {
3000 disable |= 1 << ssid;
3001 enable &= ~(1 << ssid);
3007 if (ssid == CGROUP_SUBSYS_COUNT)
3011 cgrp = cgroup_kn_lock_live(of->kn);
3015 for_each_subsys(ss, ssid) {
3016 if (enable & (1 << ssid)) {
3017 if (cgrp->subtree_control & (1 << ssid)) {
3018 enable &= ~(1 << ssid);
3022 /* unavailable or not enabled on the parent? */
3023 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
3024 (cgroup_parent(cgrp) &&
3025 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
3029 } else if (disable & (1 << ssid)) {
3030 if (!(cgrp->subtree_control & (1 << ssid))) {
3031 disable &= ~(1 << ssid);
3035 /* a child has it enabled? */
3036 cgroup_for_each_live_child(child, cgrp) {
3037 if (child->subtree_control & (1 << ssid)) {
3045 if (!enable && !disable) {
3051 * Except for the root, subtree_control must be zero for a cgroup
3052 * with tasks so that child cgroups don't compete against tasks.
3054 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3060 * Update subsys masks and calculate what needs to be done. More
3061 * subsystems than specified may need to be enabled or disabled
3062 * depending on subsystem dependencies.
3064 old_sc = cgrp->subtree_control;
3065 old_ss = cgrp->child_subsys_mask;
3066 new_sc = (old_sc | enable) & ~disable;
3067 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
3069 css_enable = ~old_ss & new_ss;
3070 css_disable = old_ss & ~new_ss;
3071 enable |= css_enable;
3072 disable |= css_disable;
3075 * Because css offlining is asynchronous, userland might try to
3076 * re-enable the same controller while the previous instance is
3077 * still around. In such cases, wait till it's gone using
3080 for_each_subsys_which(ss, ssid, &css_enable) {
3081 cgroup_for_each_live_child(child, cgrp) {
3084 if (!cgroup_css(child, ss))
3088 prepare_to_wait(&child->offline_waitq, &wait,
3089 TASK_UNINTERRUPTIBLE);
3090 cgroup_kn_unlock(of->kn);
3092 finish_wait(&child->offline_waitq, &wait);
3095 return restart_syscall();
3099 cgrp->subtree_control = new_sc;
3100 cgrp->child_subsys_mask = new_ss;
3103 * Create new csses or make the existing ones visible. A css is
3104 * created invisible if it's being implicitly enabled through
3105 * dependency. An invisible css is made visible when the userland
3106 * explicitly enables it.
3108 for_each_subsys(ss, ssid) {
3109 if (!(enable & (1 << ssid)))
3112 cgroup_for_each_live_child(child, cgrp) {
3113 if (css_enable & (1 << ssid))
3114 ret = create_css(child, ss,
3115 cgrp->subtree_control & (1 << ssid));
3117 ret = css_populate_dir(cgroup_css(child, ss),
3125 * At this point, cgroup_e_css() results reflect the new csses
3126 * making the following cgroup_update_dfl_csses() properly update
3127 * css associations of all tasks in the subtree.
3129 ret = cgroup_update_dfl_csses(cgrp);
3134 * All tasks are migrated out of disabled csses. Kill or hide
3135 * them. A css is hidden when the userland requests it to be
3136 * disabled while other subsystems are still depending on it. The
3137 * css must not actively control resources and be in the vanilla
3138 * state if it's made visible again later. Controllers which may
3139 * be depended upon should provide ->css_reset() for this purpose.
3141 for_each_subsys(ss, ssid) {
3142 if (!(disable & (1 << ssid)))
3145 cgroup_for_each_live_child(child, cgrp) {
3146 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3148 if (css_disable & (1 << ssid)) {
3151 css_clear_dir(css, NULL);
3159 * The effective csses of all the descendants (excluding @cgrp) may
3160 * have changed. Subsystems can optionally subscribe to this event
3161 * by implementing ->css_e_css_changed() which is invoked if any of
3162 * the effective csses seen from the css's cgroup may have changed.
3164 for_each_subsys(ss, ssid) {
3165 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3166 struct cgroup_subsys_state *css;
3168 if (!ss->css_e_css_changed || !this_css)
3171 css_for_each_descendant_pre(css, this_css)
3172 if (css != this_css)
3173 ss->css_e_css_changed(css);
3176 kernfs_activate(cgrp->kn);
3179 cgroup_kn_unlock(of->kn);
3180 return ret ?: nbytes;
3183 cgrp->subtree_control = old_sc;
3184 cgrp->child_subsys_mask = old_ss;
3186 for_each_subsys(ss, ssid) {
3187 if (!(enable & (1 << ssid)))
3190 cgroup_for_each_live_child(child, cgrp) {
3191 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3196 if (css_enable & (1 << ssid))
3199 css_clear_dir(css, NULL);
3205 static int cgroup_events_show(struct seq_file *seq, void *v)
3207 seq_printf(seq, "populated %d\n",
3208 cgroup_is_populated(seq_css(seq)->cgroup));
3212 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3213 size_t nbytes, loff_t off)
3215 struct cgroup *cgrp = of->kn->parent->priv;
3216 struct cftype *cft = of->kn->priv;
3217 struct cgroup_subsys_state *css;
3221 return cft->write(of, buf, nbytes, off);
3224 * kernfs guarantees that a file isn't deleted with operations in
3225 * flight, which means that the matching css is and stays alive and
3226 * doesn't need to be pinned. The RCU locking is not necessary
3227 * either. It's just for the convenience of using cgroup_css().
3230 css = cgroup_css(cgrp, cft->ss);
3233 if (cft->write_u64) {
3234 unsigned long long v;
3235 ret = kstrtoull(buf, 0, &v);
3237 ret = cft->write_u64(css, cft, v);
3238 } else if (cft->write_s64) {
3240 ret = kstrtoll(buf, 0, &v);
3242 ret = cft->write_s64(css, cft, v);
3247 return ret ?: nbytes;
3250 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3252 return seq_cft(seq)->seq_start(seq, ppos);
3255 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3257 return seq_cft(seq)->seq_next(seq, v, ppos);
3260 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3262 seq_cft(seq)->seq_stop(seq, v);
3265 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3267 struct cftype *cft = seq_cft(m);
3268 struct cgroup_subsys_state *css = seq_css(m);
3271 return cft->seq_show(m, arg);
3274 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3275 else if (cft->read_s64)
3276 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3282 static struct kernfs_ops cgroup_kf_single_ops = {
3283 .atomic_write_len = PAGE_SIZE,
3284 .write = cgroup_file_write,
3285 .seq_show = cgroup_seqfile_show,
3288 static struct kernfs_ops cgroup_kf_ops = {
3289 .atomic_write_len = PAGE_SIZE,
3290 .write = cgroup_file_write,
3291 .seq_start = cgroup_seqfile_start,
3292 .seq_next = cgroup_seqfile_next,
3293 .seq_stop = cgroup_seqfile_stop,
3294 .seq_show = cgroup_seqfile_show,
3298 * cgroup_rename - Only allow simple rename of directories in place.
3300 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3301 const char *new_name_str)
3303 struct cgroup *cgrp = kn->priv;
3306 if (kernfs_type(kn) != KERNFS_DIR)
3308 if (kn->parent != new_parent)
3312 * This isn't a proper migration and its usefulness is very
3313 * limited. Disallow on the default hierarchy.
3315 if (cgroup_on_dfl(cgrp))
3319 * We're gonna grab cgroup_mutex which nests outside kernfs
3320 * active_ref. kernfs_rename() doesn't require active_ref
3321 * protection. Break them before grabbing cgroup_mutex.
3323 kernfs_break_active_protection(new_parent);
3324 kernfs_break_active_protection(kn);
3326 mutex_lock(&cgroup_mutex);
3328 ret = kernfs_rename(kn, new_parent, new_name_str);
3330 mutex_unlock(&cgroup_mutex);
3332 kernfs_unbreak_active_protection(kn);
3333 kernfs_unbreak_active_protection(new_parent);
3337 /* set uid and gid of cgroup dirs and files to that of the creator */
3338 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3340 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3341 .ia_uid = current_fsuid(),
3342 .ia_gid = current_fsgid(), };
3344 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3345 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3348 return kernfs_setattr(kn, &iattr);
3351 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3354 char name[CGROUP_FILE_NAME_MAX];
3355 struct kernfs_node *kn;
3356 struct lock_class_key *key = NULL;
3359 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3360 key = &cft->lockdep_key;
3362 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3363 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3368 ret = cgroup_kn_set_ugid(kn);
3374 if (cft->file_offset) {
3375 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3377 spin_lock_irq(&cgroup_file_kn_lock);
3379 spin_unlock_irq(&cgroup_file_kn_lock);
3386 * cgroup_addrm_files - add or remove files to a cgroup directory
3387 * @css: the target css
3388 * @cgrp: the target cgroup (usually css->cgroup)
3389 * @cfts: array of cftypes to be added
3390 * @is_add: whether to add or remove
3392 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3393 * For removals, this function never fails.
3395 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3396 struct cgroup *cgrp, struct cftype cfts[],
3399 struct cftype *cft, *cft_end = NULL;
3402 lockdep_assert_held(&cgroup_mutex);
3405 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3406 /* does cft->flags tell us to skip this file on @cgrp? */
3407 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3409 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3411 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3413 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3417 ret = cgroup_add_file(css, cgrp, cft);
3419 pr_warn("%s: failed to add %s, err=%d\n",
3420 __func__, cft->name, ret);
3426 cgroup_rm_file(cgrp, cft);
3432 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3435 struct cgroup_subsys *ss = cfts[0].ss;
3436 struct cgroup *root = &ss->root->cgrp;
3437 struct cgroup_subsys_state *css;
3440 lockdep_assert_held(&cgroup_mutex);
3442 /* add/rm files for all cgroups created before */
3443 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3444 struct cgroup *cgrp = css->cgroup;
3446 if (cgroup_is_dead(cgrp))
3449 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3455 kernfs_activate(root->kn);
3459 static void cgroup_exit_cftypes(struct cftype *cfts)
3463 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3464 /* free copy for custom atomic_write_len, see init_cftypes() */
3465 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3470 /* revert flags set by cgroup core while adding @cfts */
3471 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3475 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3479 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3480 struct kernfs_ops *kf_ops;
3482 WARN_ON(cft->ss || cft->kf_ops);
3485 kf_ops = &cgroup_kf_ops;
3487 kf_ops = &cgroup_kf_single_ops;
3490 * Ugh... if @cft wants a custom max_write_len, we need to
3491 * make a copy of kf_ops to set its atomic_write_len.
3493 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3494 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3496 cgroup_exit_cftypes(cfts);
3499 kf_ops->atomic_write_len = cft->max_write_len;
3502 cft->kf_ops = kf_ops;
3509 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3511 lockdep_assert_held(&cgroup_mutex);
3513 if (!cfts || !cfts[0].ss)
3516 list_del(&cfts->node);
3517 cgroup_apply_cftypes(cfts, false);
3518 cgroup_exit_cftypes(cfts);
3523 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3524 * @cfts: zero-length name terminated array of cftypes
3526 * Unregister @cfts. Files described by @cfts are removed from all
3527 * existing cgroups and all future cgroups won't have them either. This
3528 * function can be called anytime whether @cfts' subsys is attached or not.
3530 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3533 int cgroup_rm_cftypes(struct cftype *cfts)
3537 mutex_lock(&cgroup_mutex);
3538 ret = cgroup_rm_cftypes_locked(cfts);
3539 mutex_unlock(&cgroup_mutex);
3544 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3545 * @ss: target cgroup subsystem
3546 * @cfts: zero-length name terminated array of cftypes
3548 * Register @cfts to @ss. Files described by @cfts are created for all
3549 * existing cgroups to which @ss is attached and all future cgroups will
3550 * have them too. This function can be called anytime whether @ss is
3553 * Returns 0 on successful registration, -errno on failure. Note that this
3554 * function currently returns 0 as long as @cfts registration is successful
3555 * even if some file creation attempts on existing cgroups fail.
3557 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3561 if (!cgroup_ssid_enabled(ss->id))
3564 if (!cfts || cfts[0].name[0] == '\0')
3567 ret = cgroup_init_cftypes(ss, cfts);
3571 mutex_lock(&cgroup_mutex);
3573 list_add_tail(&cfts->node, &ss->cfts);
3574 ret = cgroup_apply_cftypes(cfts, true);
3576 cgroup_rm_cftypes_locked(cfts);
3578 mutex_unlock(&cgroup_mutex);
3583 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3584 * @ss: target cgroup subsystem
3585 * @cfts: zero-length name terminated array of cftypes
3587 * Similar to cgroup_add_cftypes() but the added files are only used for
3588 * the default hierarchy.
3590 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3594 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3595 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3596 return cgroup_add_cftypes(ss, cfts);
3600 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3601 * @ss: target cgroup subsystem
3602 * @cfts: zero-length name terminated array of cftypes
3604 * Similar to cgroup_add_cftypes() but the added files are only used for
3605 * the legacy hierarchies.
3607 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3611 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3612 cft->flags |= __CFTYPE_NOT_ON_DFL;
3613 return cgroup_add_cftypes(ss, cfts);
3617 * cgroup_file_notify - generate a file modified event for a cgroup_file
3618 * @cfile: target cgroup_file
3620 * @cfile must have been obtained by setting cftype->file_offset.
3622 void cgroup_file_notify(struct cgroup_file *cfile)
3624 unsigned long flags;
3626 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3628 kernfs_notify(cfile->kn);
3629 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3633 * cgroup_task_count - count the number of tasks in a cgroup.
3634 * @cgrp: the cgroup in question
3636 * Return the number of tasks in the cgroup.
3638 static int cgroup_task_count(const struct cgroup *cgrp)
3641 struct cgrp_cset_link *link;
3643 spin_lock_bh(&css_set_lock);
3644 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3645 count += atomic_read(&link->cset->refcount);
3646 spin_unlock_bh(&css_set_lock);
3651 * css_next_child - find the next child of a given css
3652 * @pos: the current position (%NULL to initiate traversal)
3653 * @parent: css whose children to walk
3655 * This function returns the next child of @parent and should be called
3656 * under either cgroup_mutex or RCU read lock. The only requirement is
3657 * that @parent and @pos are accessible. The next sibling is guaranteed to
3658 * be returned regardless of their states.
3660 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3661 * css which finished ->css_online() is guaranteed to be visible in the
3662 * future iterations and will stay visible until the last reference is put.
3663 * A css which hasn't finished ->css_online() or already finished
3664 * ->css_offline() may show up during traversal. It's each subsystem's
3665 * responsibility to synchronize against on/offlining.
3667 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3668 struct cgroup_subsys_state *parent)
3670 struct cgroup_subsys_state *next;
3672 cgroup_assert_mutex_or_rcu_locked();
3675 * @pos could already have been unlinked from the sibling list.
3676 * Once a cgroup is removed, its ->sibling.next is no longer
3677 * updated when its next sibling changes. CSS_RELEASED is set when
3678 * @pos is taken off list, at which time its next pointer is valid,
3679 * and, as releases are serialized, the one pointed to by the next
3680 * pointer is guaranteed to not have started release yet. This
3681 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3682 * critical section, the one pointed to by its next pointer is
3683 * guaranteed to not have finished its RCU grace period even if we
3684 * have dropped rcu_read_lock() inbetween iterations.
3686 * If @pos has CSS_RELEASED set, its next pointer can't be
3687 * dereferenced; however, as each css is given a monotonically
3688 * increasing unique serial number and always appended to the
3689 * sibling list, the next one can be found by walking the parent's
3690 * children until the first css with higher serial number than
3691 * @pos's. While this path can be slower, it happens iff iteration
3692 * races against release and the race window is very small.
3695 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3696 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3697 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3699 list_for_each_entry_rcu(next, &parent->children, sibling)
3700 if (next->serial_nr > pos->serial_nr)
3705 * @next, if not pointing to the head, can be dereferenced and is
3708 if (&next->sibling != &parent->children)
3714 * css_next_descendant_pre - find the next descendant for pre-order walk
3715 * @pos: the current position (%NULL to initiate traversal)
3716 * @root: css whose descendants to walk
3718 * To be used by css_for_each_descendant_pre(). Find the next descendant
3719 * to visit for pre-order traversal of @root's descendants. @root is
3720 * included in the iteration and the first node to be visited.
3722 * While this function requires cgroup_mutex or RCU read locking, it
3723 * doesn't require the whole traversal to be contained in a single critical
3724 * section. This function will return the correct next descendant as long
3725 * as both @pos and @root are accessible and @pos is a descendant of @root.
3727 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3728 * css which finished ->css_online() is guaranteed to be visible in the
3729 * future iterations and will stay visible until the last reference is put.
3730 * A css which hasn't finished ->css_online() or already finished
3731 * ->css_offline() may show up during traversal. It's each subsystem's
3732 * responsibility to synchronize against on/offlining.
3734 struct cgroup_subsys_state *
3735 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3736 struct cgroup_subsys_state *root)
3738 struct cgroup_subsys_state *next;
3740 cgroup_assert_mutex_or_rcu_locked();
3742 /* if first iteration, visit @root */
3746 /* visit the first child if exists */
3747 next = css_next_child(NULL, pos);
3751 /* no child, visit my or the closest ancestor's next sibling */
3752 while (pos != root) {
3753 next = css_next_child(pos, pos->parent);
3763 * css_rightmost_descendant - return the rightmost descendant of a css
3764 * @pos: css of interest
3766 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3767 * is returned. This can be used during pre-order traversal to skip
3770 * While this function requires cgroup_mutex or RCU read locking, it
3771 * doesn't require the whole traversal to be contained in a single critical
3772 * section. This function will return the correct rightmost descendant as
3773 * long as @pos is accessible.
3775 struct cgroup_subsys_state *
3776 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3778 struct cgroup_subsys_state *last, *tmp;
3780 cgroup_assert_mutex_or_rcu_locked();
3784 /* ->prev isn't RCU safe, walk ->next till the end */
3786 css_for_each_child(tmp, last)
3793 static struct cgroup_subsys_state *
3794 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3796 struct cgroup_subsys_state *last;
3800 pos = css_next_child(NULL, pos);
3807 * css_next_descendant_post - find the next descendant for post-order walk
3808 * @pos: the current position (%NULL to initiate traversal)
3809 * @root: css whose descendants to walk
3811 * To be used by css_for_each_descendant_post(). Find the next descendant
3812 * to visit for post-order traversal of @root's descendants. @root is
3813 * included in the iteration and the last node to be visited.
3815 * While this function requires cgroup_mutex or RCU read locking, it
3816 * doesn't require the whole traversal to be contained in a single critical
3817 * section. This function will return the correct next descendant as long
3818 * as both @pos and @cgroup are accessible and @pos is a descendant of
3821 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3822 * css which finished ->css_online() is guaranteed to be visible in the
3823 * future iterations and will stay visible until the last reference is put.
3824 * A css which hasn't finished ->css_online() or already finished
3825 * ->css_offline() may show up during traversal. It's each subsystem's
3826 * responsibility to synchronize against on/offlining.
3828 struct cgroup_subsys_state *
3829 css_next_descendant_post(struct cgroup_subsys_state *pos,
3830 struct cgroup_subsys_state *root)
3832 struct cgroup_subsys_state *next;
3834 cgroup_assert_mutex_or_rcu_locked();
3836 /* if first iteration, visit leftmost descendant which may be @root */
3838 return css_leftmost_descendant(root);
3840 /* if we visited @root, we're done */
3844 /* if there's an unvisited sibling, visit its leftmost descendant */
3845 next = css_next_child(pos, pos->parent);
3847 return css_leftmost_descendant(next);
3849 /* no sibling left, visit parent */
3854 * css_has_online_children - does a css have online children
3855 * @css: the target css
3857 * Returns %true if @css has any online children; otherwise, %false. This
3858 * function can be called from any context but the caller is responsible
3859 * for synchronizing against on/offlining as necessary.
3861 bool css_has_online_children(struct cgroup_subsys_state *css)
3863 struct cgroup_subsys_state *child;
3867 css_for_each_child(child, css) {
3868 if (child->flags & CSS_ONLINE) {
3878 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3879 * @it: the iterator to advance
3881 * Advance @it to the next css_set to walk.
3883 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3885 struct list_head *l = it->cset_pos;
3886 struct cgrp_cset_link *link;
3887 struct css_set *cset;
3889 lockdep_assert_held(&css_set_lock);
3891 /* Advance to the next non-empty css_set */
3894 if (l == it->cset_head) {
3895 it->cset_pos = NULL;
3896 it->task_pos = NULL;
3901 cset = container_of(l, struct css_set,
3902 e_cset_node[it->ss->id]);
3904 link = list_entry(l, struct cgrp_cset_link, cset_link);
3907 } while (!css_set_populated(cset));
3911 if (!list_empty(&cset->tasks))
3912 it->task_pos = cset->tasks.next;
3914 it->task_pos = cset->mg_tasks.next;
3916 it->tasks_head = &cset->tasks;
3917 it->mg_tasks_head = &cset->mg_tasks;
3920 * We don't keep css_sets locked across iteration steps and thus
3921 * need to take steps to ensure that iteration can be resumed after
3922 * the lock is re-acquired. Iteration is performed at two levels -
3923 * css_sets and tasks in them.
3925 * Once created, a css_set never leaves its cgroup lists, so a
3926 * pinned css_set is guaranteed to stay put and we can resume
3927 * iteration afterwards.
3929 * Tasks may leave @cset across iteration steps. This is resolved
3930 * by registering each iterator with the css_set currently being
3931 * walked and making css_set_move_task() advance iterators whose
3932 * next task is leaving.
3935 list_del(&it->iters_node);
3936 put_css_set_locked(it->cur_cset);
3939 it->cur_cset = cset;
3940 list_add(&it->iters_node, &cset->task_iters);
3943 static void css_task_iter_advance(struct css_task_iter *it)
3945 struct list_head *l = it->task_pos;
3947 lockdep_assert_held(&css_set_lock);
3951 * Advance iterator to find next entry. cset->tasks is consumed
3952 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3957 if (l == it->tasks_head)
3958 l = it->mg_tasks_head->next;
3960 if (l == it->mg_tasks_head)
3961 css_task_iter_advance_css_set(it);
3967 * css_task_iter_start - initiate task iteration
3968 * @css: the css to walk tasks of
3969 * @it: the task iterator to use
3971 * Initiate iteration through the tasks of @css. The caller can call
3972 * css_task_iter_next() to walk through the tasks until the function
3973 * returns NULL. On completion of iteration, css_task_iter_end() must be
3976 void css_task_iter_start(struct cgroup_subsys_state *css,
3977 struct css_task_iter *it)
3979 /* no one should try to iterate before mounting cgroups */
3980 WARN_ON_ONCE(!use_task_css_set_links);
3982 memset(it, 0, sizeof(*it));
3984 spin_lock_bh(&css_set_lock);
3989 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3991 it->cset_pos = &css->cgroup->cset_links;
3993 it->cset_head = it->cset_pos;
3995 css_task_iter_advance_css_set(it);
3997 spin_unlock_bh(&css_set_lock);
4001 * css_task_iter_next - return the next task for the iterator
4002 * @it: the task iterator being iterated
4004 * The "next" function for task iteration. @it should have been
4005 * initialized via css_task_iter_start(). Returns NULL when the iteration
4008 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4011 put_task_struct(it->cur_task);
4012 it->cur_task = NULL;
4015 spin_lock_bh(&css_set_lock);
4018 it->cur_task = list_entry(it->task_pos, struct task_struct,
4020 get_task_struct(it->cur_task);
4021 css_task_iter_advance(it);
4024 spin_unlock_bh(&css_set_lock);
4026 return it->cur_task;
4030 * css_task_iter_end - finish task iteration
4031 * @it: the task iterator to finish
4033 * Finish task iteration started by css_task_iter_start().
4035 void css_task_iter_end(struct css_task_iter *it)
4038 spin_lock_bh(&css_set_lock);
4039 list_del(&it->iters_node);
4040 put_css_set_locked(it->cur_cset);
4041 spin_unlock_bh(&css_set_lock);
4045 put_task_struct(it->cur_task);
4049 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4050 * @to: cgroup to which the tasks will be moved
4051 * @from: cgroup in which the tasks currently reside
4053 * Locking rules between cgroup_post_fork() and the migration path
4054 * guarantee that, if a task is forking while being migrated, the new child
4055 * is guaranteed to be either visible in the source cgroup after the
4056 * parent's migration is complete or put into the target cgroup. No task
4057 * can slip out of migration through forking.
4059 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4061 LIST_HEAD(preloaded_csets);
4062 struct cgrp_cset_link *link;
4063 struct css_task_iter it;
4064 struct task_struct *task;
4067 mutex_lock(&cgroup_mutex);
4069 /* all tasks in @from are being moved, all csets are source */
4070 spin_lock_bh(&css_set_lock);
4071 list_for_each_entry(link, &from->cset_links, cset_link)
4072 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4073 spin_unlock_bh(&css_set_lock);
4075 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4080 * Migrate tasks one-by-one until @form is empty. This fails iff
4081 * ->can_attach() fails.
4084 css_task_iter_start(&from->self, &it);
4085 task = css_task_iter_next(&it);
4087 get_task_struct(task);
4088 css_task_iter_end(&it);
4091 ret = cgroup_migrate(task, false, to);
4092 put_task_struct(task);
4094 } while (task && !ret);
4096 cgroup_migrate_finish(&preloaded_csets);
4097 mutex_unlock(&cgroup_mutex);
4102 * Stuff for reading the 'tasks'/'procs' files.
4104 * Reading this file can return large amounts of data if a cgroup has
4105 * *lots* of attached tasks. So it may need several calls to read(),
4106 * but we cannot guarantee that the information we produce is correct
4107 * unless we produce it entirely atomically.
4111 /* which pidlist file are we talking about? */
4112 enum cgroup_filetype {
4118 * A pidlist is a list of pids that virtually represents the contents of one
4119 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4120 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4123 struct cgroup_pidlist {
4125 * used to find which pidlist is wanted. doesn't change as long as
4126 * this particular list stays in the list.
4128 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4131 /* how many elements the above list has */
4133 /* each of these stored in a list by its cgroup */
4134 struct list_head links;
4135 /* pointer to the cgroup we belong to, for list removal purposes */
4136 struct cgroup *owner;
4137 /* for delayed destruction */
4138 struct delayed_work destroy_dwork;
4142 * The following two functions "fix" the issue where there are more pids
4143 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4144 * TODO: replace with a kernel-wide solution to this problem
4146 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4147 static void *pidlist_allocate(int count)
4149 if (PIDLIST_TOO_LARGE(count))
4150 return vmalloc(count * sizeof(pid_t));
4152 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4155 static void pidlist_free(void *p)
4161 * Used to destroy all pidlists lingering waiting for destroy timer. None
4162 * should be left afterwards.
4164 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4166 struct cgroup_pidlist *l, *tmp_l;
4168 mutex_lock(&cgrp->pidlist_mutex);
4169 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4170 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4171 mutex_unlock(&cgrp->pidlist_mutex);
4173 flush_workqueue(cgroup_pidlist_destroy_wq);
4174 BUG_ON(!list_empty(&cgrp->pidlists));
4177 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4179 struct delayed_work *dwork = to_delayed_work(work);
4180 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4182 struct cgroup_pidlist *tofree = NULL;
4184 mutex_lock(&l->owner->pidlist_mutex);
4187 * Destroy iff we didn't get queued again. The state won't change
4188 * as destroy_dwork can only be queued while locked.
4190 if (!delayed_work_pending(dwork)) {
4191 list_del(&l->links);
4192 pidlist_free(l->list);
4193 put_pid_ns(l->key.ns);
4197 mutex_unlock(&l->owner->pidlist_mutex);
4202 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4203 * Returns the number of unique elements.
4205 static int pidlist_uniq(pid_t *list, int length)
4210 * we presume the 0th element is unique, so i starts at 1. trivial
4211 * edge cases first; no work needs to be done for either
4213 if (length == 0 || length == 1)
4215 /* src and dest walk down the list; dest counts unique elements */
4216 for (src = 1; src < length; src++) {
4217 /* find next unique element */
4218 while (list[src] == list[src-1]) {
4223 /* dest always points to where the next unique element goes */
4224 list[dest] = list[src];
4232 * The two pid files - task and cgroup.procs - guaranteed that the result
4233 * is sorted, which forced this whole pidlist fiasco. As pid order is
4234 * different per namespace, each namespace needs differently sorted list,
4235 * making it impossible to use, for example, single rbtree of member tasks
4236 * sorted by task pointer. As pidlists can be fairly large, allocating one
4237 * per open file is dangerous, so cgroup had to implement shared pool of
4238 * pidlists keyed by cgroup and namespace.
4240 * All this extra complexity was caused by the original implementation
4241 * committing to an entirely unnecessary property. In the long term, we
4242 * want to do away with it. Explicitly scramble sort order if on the
4243 * default hierarchy so that no such expectation exists in the new
4246 * Scrambling is done by swapping every two consecutive bits, which is
4247 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4249 static pid_t pid_fry(pid_t pid)
4251 unsigned a = pid & 0x55555555;
4252 unsigned b = pid & 0xAAAAAAAA;
4254 return (a << 1) | (b >> 1);
4257 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4259 if (cgroup_on_dfl(cgrp))
4260 return pid_fry(pid);
4265 static int cmppid(const void *a, const void *b)
4267 return *(pid_t *)a - *(pid_t *)b;
4270 static int fried_cmppid(const void *a, const void *b)
4272 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4275 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4276 enum cgroup_filetype type)
4278 struct cgroup_pidlist *l;
4279 /* don't need task_nsproxy() if we're looking at ourself */
4280 struct pid_namespace *ns = task_active_pid_ns(current);
4282 lockdep_assert_held(&cgrp->pidlist_mutex);
4284 list_for_each_entry(l, &cgrp->pidlists, links)
4285 if (l->key.type == type && l->key.ns == ns)
4291 * find the appropriate pidlist for our purpose (given procs vs tasks)
4292 * returns with the lock on that pidlist already held, and takes care
4293 * of the use count, or returns NULL with no locks held if we're out of
4296 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4297 enum cgroup_filetype type)
4299 struct cgroup_pidlist *l;
4301 lockdep_assert_held(&cgrp->pidlist_mutex);
4303 l = cgroup_pidlist_find(cgrp, type);
4307 /* entry not found; create a new one */
4308 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4312 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4314 /* don't need task_nsproxy() if we're looking at ourself */
4315 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4317 list_add(&l->links, &cgrp->pidlists);
4322 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4324 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4325 struct cgroup_pidlist **lp)
4329 int pid, n = 0; /* used for populating the array */
4330 struct css_task_iter it;
4331 struct task_struct *tsk;
4332 struct cgroup_pidlist *l;
4334 lockdep_assert_held(&cgrp->pidlist_mutex);
4337 * If cgroup gets more users after we read count, we won't have
4338 * enough space - tough. This race is indistinguishable to the
4339 * caller from the case that the additional cgroup users didn't
4340 * show up until sometime later on.
4342 length = cgroup_task_count(cgrp);
4343 array = pidlist_allocate(length);
4346 /* now, populate the array */
4347 css_task_iter_start(&cgrp->self, &it);
4348 while ((tsk = css_task_iter_next(&it))) {
4349 if (unlikely(n == length))
4351 /* get tgid or pid for procs or tasks file respectively */
4352 if (type == CGROUP_FILE_PROCS)
4353 pid = task_tgid_vnr(tsk);
4355 pid = task_pid_vnr(tsk);
4356 if (pid > 0) /* make sure to only use valid results */
4359 css_task_iter_end(&it);
4361 /* now sort & (if procs) strip out duplicates */
4362 if (cgroup_on_dfl(cgrp))
4363 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4365 sort(array, length, sizeof(pid_t), cmppid, NULL);
4366 if (type == CGROUP_FILE_PROCS)
4367 length = pidlist_uniq(array, length);
4369 l = cgroup_pidlist_find_create(cgrp, type);
4371 pidlist_free(array);
4375 /* store array, freeing old if necessary */
4376 pidlist_free(l->list);
4384 * cgroupstats_build - build and fill cgroupstats
4385 * @stats: cgroupstats to fill information into
4386 * @dentry: A dentry entry belonging to the cgroup for which stats have
4389 * Build and fill cgroupstats so that taskstats can export it to user
4392 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4394 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4395 struct cgroup *cgrp;
4396 struct css_task_iter it;
4397 struct task_struct *tsk;
4399 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4400 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4401 kernfs_type(kn) != KERNFS_DIR)
4404 mutex_lock(&cgroup_mutex);
4407 * We aren't being called from kernfs and there's no guarantee on
4408 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4409 * @kn->priv is RCU safe. Let's do the RCU dancing.
4412 cgrp = rcu_dereference(kn->priv);
4413 if (!cgrp || cgroup_is_dead(cgrp)) {
4415 mutex_unlock(&cgroup_mutex);
4420 css_task_iter_start(&cgrp->self, &it);
4421 while ((tsk = css_task_iter_next(&it))) {
4422 switch (tsk->state) {
4424 stats->nr_running++;
4426 case TASK_INTERRUPTIBLE:
4427 stats->nr_sleeping++;
4429 case TASK_UNINTERRUPTIBLE:
4430 stats->nr_uninterruptible++;
4433 stats->nr_stopped++;
4436 if (delayacct_is_task_waiting_on_io(tsk))
4437 stats->nr_io_wait++;
4441 css_task_iter_end(&it);
4443 mutex_unlock(&cgroup_mutex);
4449 * seq_file methods for the tasks/procs files. The seq_file position is the
4450 * next pid to display; the seq_file iterator is a pointer to the pid
4451 * in the cgroup->l->list array.
4454 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4457 * Initially we receive a position value that corresponds to
4458 * one more than the last pid shown (or 0 on the first call or
4459 * after a seek to the start). Use a binary-search to find the
4460 * next pid to display, if any
4462 struct kernfs_open_file *of = s->private;
4463 struct cgroup *cgrp = seq_css(s)->cgroup;
4464 struct cgroup_pidlist *l;
4465 enum cgroup_filetype type = seq_cft(s)->private;
4466 int index = 0, pid = *pos;
4469 mutex_lock(&cgrp->pidlist_mutex);
4472 * !NULL @of->priv indicates that this isn't the first start()
4473 * after open. If the matching pidlist is around, we can use that.
4474 * Look for it. Note that @of->priv can't be used directly. It
4475 * could already have been destroyed.
4478 of->priv = cgroup_pidlist_find(cgrp, type);
4481 * Either this is the first start() after open or the matching
4482 * pidlist has been destroyed inbetween. Create a new one.
4485 ret = pidlist_array_load(cgrp, type,
4486 (struct cgroup_pidlist **)&of->priv);
4488 return ERR_PTR(ret);
4493 int end = l->length;
4495 while (index < end) {
4496 int mid = (index + end) / 2;
4497 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4500 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4506 /* If we're off the end of the array, we're done */
4507 if (index >= l->length)
4509 /* Update the abstract position to be the actual pid that we found */
4510 iter = l->list + index;
4511 *pos = cgroup_pid_fry(cgrp, *iter);
4515 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4517 struct kernfs_open_file *of = s->private;
4518 struct cgroup_pidlist *l = of->priv;
4521 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4522 CGROUP_PIDLIST_DESTROY_DELAY);
4523 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4526 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4528 struct kernfs_open_file *of = s->private;
4529 struct cgroup_pidlist *l = of->priv;
4531 pid_t *end = l->list + l->length;
4533 * Advance to the next pid in the array. If this goes off the
4540 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4545 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4547 seq_printf(s, "%d\n", *(int *)v);
4552 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4555 return notify_on_release(css->cgroup);
4558 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4559 struct cftype *cft, u64 val)
4562 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4564 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4568 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4571 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4574 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4575 struct cftype *cft, u64 val)
4578 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4580 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4584 /* cgroup core interface files for the default hierarchy */
4585 static struct cftype cgroup_dfl_base_files[] = {
4587 .name = "cgroup.procs",
4588 .file_offset = offsetof(struct cgroup, procs_file),
4589 .seq_start = cgroup_pidlist_start,
4590 .seq_next = cgroup_pidlist_next,
4591 .seq_stop = cgroup_pidlist_stop,
4592 .seq_show = cgroup_pidlist_show,
4593 .private = CGROUP_FILE_PROCS,
4594 .write = cgroup_procs_write,
4597 .name = "cgroup.controllers",
4598 .flags = CFTYPE_ONLY_ON_ROOT,
4599 .seq_show = cgroup_root_controllers_show,
4602 .name = "cgroup.controllers",
4603 .flags = CFTYPE_NOT_ON_ROOT,
4604 .seq_show = cgroup_controllers_show,
4607 .name = "cgroup.subtree_control",
4608 .seq_show = cgroup_subtree_control_show,
4609 .write = cgroup_subtree_control_write,
4612 .name = "cgroup.events",
4613 .flags = CFTYPE_NOT_ON_ROOT,
4614 .file_offset = offsetof(struct cgroup, events_file),
4615 .seq_show = cgroup_events_show,
4620 /* cgroup core interface files for the legacy hierarchies */
4621 static struct cftype cgroup_legacy_base_files[] = {
4623 .name = "cgroup.procs",
4624 .seq_start = cgroup_pidlist_start,
4625 .seq_next = cgroup_pidlist_next,
4626 .seq_stop = cgroup_pidlist_stop,
4627 .seq_show = cgroup_pidlist_show,
4628 .private = CGROUP_FILE_PROCS,
4629 .write = cgroup_procs_write,
4632 .name = "cgroup.clone_children",
4633 .read_u64 = cgroup_clone_children_read,
4634 .write_u64 = cgroup_clone_children_write,
4637 .name = "cgroup.sane_behavior",
4638 .flags = CFTYPE_ONLY_ON_ROOT,
4639 .seq_show = cgroup_sane_behavior_show,
4643 .seq_start = cgroup_pidlist_start,
4644 .seq_next = cgroup_pidlist_next,
4645 .seq_stop = cgroup_pidlist_stop,
4646 .seq_show = cgroup_pidlist_show,
4647 .private = CGROUP_FILE_TASKS,
4648 .write = cgroup_tasks_write,
4651 .name = "notify_on_release",
4652 .read_u64 = cgroup_read_notify_on_release,
4653 .write_u64 = cgroup_write_notify_on_release,
4656 .name = "release_agent",
4657 .flags = CFTYPE_ONLY_ON_ROOT,
4658 .seq_show = cgroup_release_agent_show,
4659 .write = cgroup_release_agent_write,
4660 .max_write_len = PATH_MAX - 1,
4666 * css destruction is four-stage process.
4668 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4669 * Implemented in kill_css().
4671 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4672 * and thus css_tryget_online() is guaranteed to fail, the css can be
4673 * offlined by invoking offline_css(). After offlining, the base ref is
4674 * put. Implemented in css_killed_work_fn().
4676 * 3. When the percpu_ref reaches zero, the only possible remaining
4677 * accessors are inside RCU read sections. css_release() schedules the
4680 * 4. After the grace period, the css can be freed. Implemented in
4681 * css_free_work_fn().
4683 * It is actually hairier because both step 2 and 4 require process context
4684 * and thus involve punting to css->destroy_work adding two additional
4685 * steps to the already complex sequence.
4687 static void css_free_work_fn(struct work_struct *work)
4689 struct cgroup_subsys_state *css =
4690 container_of(work, struct cgroup_subsys_state, destroy_work);
4691 struct cgroup_subsys *ss = css->ss;
4692 struct cgroup *cgrp = css->cgroup;
4694 percpu_ref_exit(&css->refcnt);
4698 struct cgroup_subsys_state *parent = css->parent;
4702 cgroup_idr_remove(&ss->css_idr, id);
4708 /* cgroup free path */
4709 atomic_dec(&cgrp->root->nr_cgrps);
4710 cgroup_pidlist_destroy_all(cgrp);
4711 cancel_work_sync(&cgrp->release_agent_work);
4713 if (cgroup_parent(cgrp)) {
4715 * We get a ref to the parent, and put the ref when
4716 * this cgroup is being freed, so it's guaranteed
4717 * that the parent won't be destroyed before its
4720 cgroup_put(cgroup_parent(cgrp));
4721 kernfs_put(cgrp->kn);
4725 * This is root cgroup's refcnt reaching zero,
4726 * which indicates that the root should be
4729 cgroup_destroy_root(cgrp->root);
4734 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4736 struct cgroup_subsys_state *css =
4737 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4739 INIT_WORK(&css->destroy_work, css_free_work_fn);
4740 queue_work(cgroup_destroy_wq, &css->destroy_work);
4743 static void css_release_work_fn(struct work_struct *work)
4745 struct cgroup_subsys_state *css =
4746 container_of(work, struct cgroup_subsys_state, destroy_work);
4747 struct cgroup_subsys *ss = css->ss;
4748 struct cgroup *cgrp = css->cgroup;
4750 mutex_lock(&cgroup_mutex);
4752 css->flags |= CSS_RELEASED;
4753 list_del_rcu(&css->sibling);
4756 /* css release path */
4757 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4758 if (ss->css_released)
4759 ss->css_released(css);
4761 /* cgroup release path */
4762 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4766 * There are two control paths which try to determine
4767 * cgroup from dentry without going through kernfs -
4768 * cgroupstats_build() and css_tryget_online_from_dir().
4769 * Those are supported by RCU protecting clearing of
4770 * cgrp->kn->priv backpointer.
4772 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4775 mutex_unlock(&cgroup_mutex);
4777 call_rcu(&css->rcu_head, css_free_rcu_fn);
4780 static void css_release(struct percpu_ref *ref)
4782 struct cgroup_subsys_state *css =
4783 container_of(ref, struct cgroup_subsys_state, refcnt);
4785 INIT_WORK(&css->destroy_work, css_release_work_fn);
4786 queue_work(cgroup_destroy_wq, &css->destroy_work);
4789 static void init_and_link_css(struct cgroup_subsys_state *css,
4790 struct cgroup_subsys *ss, struct cgroup *cgrp)
4792 lockdep_assert_held(&cgroup_mutex);
4796 memset(css, 0, sizeof(*css));
4800 INIT_LIST_HEAD(&css->sibling);
4801 INIT_LIST_HEAD(&css->children);
4802 css->serial_nr = css_serial_nr_next++;
4803 atomic_set(&css->online_cnt, 0);
4805 if (cgroup_parent(cgrp)) {
4806 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4807 css_get(css->parent);
4810 BUG_ON(cgroup_css(cgrp, ss));
4813 /* invoke ->css_online() on a new CSS and mark it online if successful */
4814 static int online_css(struct cgroup_subsys_state *css)
4816 struct cgroup_subsys *ss = css->ss;
4819 lockdep_assert_held(&cgroup_mutex);
4822 ret = ss->css_online(css);
4824 css->flags |= CSS_ONLINE;
4825 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4827 atomic_inc(&css->online_cnt);
4829 atomic_inc(&css->parent->online_cnt);
4834 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4835 static void offline_css(struct cgroup_subsys_state *css)
4837 struct cgroup_subsys *ss = css->ss;
4839 lockdep_assert_held(&cgroup_mutex);
4841 if (!(css->flags & CSS_ONLINE))
4844 if (ss->css_offline)
4845 ss->css_offline(css);
4847 css->flags &= ~CSS_ONLINE;
4848 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4850 wake_up_all(&css->cgroup->offline_waitq);
4854 * create_css - create a cgroup_subsys_state
4855 * @cgrp: the cgroup new css will be associated with
4856 * @ss: the subsys of new css
4857 * @visible: whether to create control knobs for the new css or not
4859 * Create a new css associated with @cgrp - @ss pair. On success, the new
4860 * css is online and installed in @cgrp with all interface files created if
4861 * @visible. Returns 0 on success, -errno on failure.
4863 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4866 struct cgroup *parent = cgroup_parent(cgrp);
4867 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4868 struct cgroup_subsys_state *css;
4871 lockdep_assert_held(&cgroup_mutex);
4873 css = ss->css_alloc(parent_css);
4875 return PTR_ERR(css);
4877 init_and_link_css(css, ss, cgrp);
4879 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4883 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4885 goto err_free_percpu_ref;
4889 err = css_populate_dir(css, NULL);
4894 /* @css is ready to be brought online now, make it visible */
4895 list_add_tail_rcu(&css->sibling, &parent_css->children);
4896 cgroup_idr_replace(&ss->css_idr, css, css->id);
4898 err = online_css(css);
4902 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4903 cgroup_parent(parent)) {
4904 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4905 current->comm, current->pid, ss->name);
4906 if (!strcmp(ss->name, "memory"))
4907 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4908 ss->warned_broken_hierarchy = true;
4914 list_del_rcu(&css->sibling);
4915 css_clear_dir(css, NULL);
4917 cgroup_idr_remove(&ss->css_idr, css->id);
4918 err_free_percpu_ref:
4919 percpu_ref_exit(&css->refcnt);
4921 call_rcu(&css->rcu_head, css_free_rcu_fn);
4925 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4928 struct cgroup *parent, *cgrp;
4929 struct cgroup_root *root;
4930 struct cgroup_subsys *ss;
4931 struct kernfs_node *kn;
4934 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4936 if (strchr(name, '\n'))
4939 parent = cgroup_kn_lock_live(parent_kn);
4942 root = parent->root;
4944 /* allocate the cgroup and its ID, 0 is reserved for the root */
4945 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4951 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4956 * Temporarily set the pointer to NULL, so idr_find() won't return
4957 * a half-baked cgroup.
4959 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4962 goto out_cancel_ref;
4965 init_cgroup_housekeeping(cgrp);
4967 cgrp->self.parent = &parent->self;
4970 if (notify_on_release(parent))
4971 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4973 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4974 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4976 /* create the directory */
4977 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4985 * This extra ref will be put in cgroup_free_fn() and guarantees
4986 * that @cgrp->kn is always accessible.
4990 cgrp->self.serial_nr = css_serial_nr_next++;
4992 /* allocation complete, commit to creation */
4993 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4994 atomic_inc(&root->nr_cgrps);
4998 * @cgrp is now fully operational. If something fails after this
4999 * point, it'll be released via the normal destruction path.
5001 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5003 ret = cgroup_kn_set_ugid(kn);
5007 ret = css_populate_dir(&cgrp->self, NULL);
5011 /* let's create and online css's */
5012 for_each_subsys(ss, ssid) {
5013 if (parent->child_subsys_mask & (1 << ssid)) {
5014 ret = create_css(cgrp, ss,
5015 parent->subtree_control & (1 << ssid));
5022 * On the default hierarchy, a child doesn't automatically inherit
5023 * subtree_control from the parent. Each is configured manually.
5025 if (!cgroup_on_dfl(cgrp)) {
5026 cgrp->subtree_control = parent->subtree_control;
5027 cgroup_refresh_child_subsys_mask(cgrp);
5030 kernfs_activate(kn);
5036 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5038 percpu_ref_exit(&cgrp->self.refcnt);
5042 cgroup_kn_unlock(parent_kn);
5046 cgroup_destroy_locked(cgrp);
5051 * This is called when the refcnt of a css is confirmed to be killed.
5052 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5053 * initate destruction and put the css ref from kill_css().
5055 static void css_killed_work_fn(struct work_struct *work)
5057 struct cgroup_subsys_state *css =
5058 container_of(work, struct cgroup_subsys_state, destroy_work);
5060 mutex_lock(&cgroup_mutex);
5065 /* @css can't go away while we're holding cgroup_mutex */
5067 } while (css && atomic_dec_and_test(&css->online_cnt));
5069 mutex_unlock(&cgroup_mutex);
5072 /* css kill confirmation processing requires process context, bounce */
5073 static void css_killed_ref_fn(struct percpu_ref *ref)
5075 struct cgroup_subsys_state *css =
5076 container_of(ref, struct cgroup_subsys_state, refcnt);
5078 if (atomic_dec_and_test(&css->online_cnt)) {
5079 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5080 queue_work(cgroup_destroy_wq, &css->destroy_work);
5085 * kill_css - destroy a css
5086 * @css: css to destroy
5088 * This function initiates destruction of @css by removing cgroup interface
5089 * files and putting its base reference. ->css_offline() will be invoked
5090 * asynchronously once css_tryget_online() is guaranteed to fail and when
5091 * the reference count reaches zero, @css will be released.
5093 static void kill_css(struct cgroup_subsys_state *css)
5095 lockdep_assert_held(&cgroup_mutex);
5098 * This must happen before css is disassociated with its cgroup.
5099 * See seq_css() for details.
5101 css_clear_dir(css, NULL);
5104 * Killing would put the base ref, but we need to keep it alive
5105 * until after ->css_offline().
5110 * cgroup core guarantees that, by the time ->css_offline() is
5111 * invoked, no new css reference will be given out via
5112 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5113 * proceed to offlining css's because percpu_ref_kill() doesn't
5114 * guarantee that the ref is seen as killed on all CPUs on return.
5116 * Use percpu_ref_kill_and_confirm() to get notifications as each
5117 * css is confirmed to be seen as killed on all CPUs.
5119 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5123 * cgroup_destroy_locked - the first stage of cgroup destruction
5124 * @cgrp: cgroup to be destroyed
5126 * css's make use of percpu refcnts whose killing latency shouldn't be
5127 * exposed to userland and are RCU protected. Also, cgroup core needs to
5128 * guarantee that css_tryget_online() won't succeed by the time
5129 * ->css_offline() is invoked. To satisfy all the requirements,
5130 * destruction is implemented in the following two steps.
5132 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5133 * userland visible parts and start killing the percpu refcnts of
5134 * css's. Set up so that the next stage will be kicked off once all
5135 * the percpu refcnts are confirmed to be killed.
5137 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5138 * rest of destruction. Once all cgroup references are gone, the
5139 * cgroup is RCU-freed.
5141 * This function implements s1. After this step, @cgrp is gone as far as
5142 * the userland is concerned and a new cgroup with the same name may be
5143 * created. As cgroup doesn't care about the names internally, this
5144 * doesn't cause any problem.
5146 static int cgroup_destroy_locked(struct cgroup *cgrp)
5147 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5149 struct cgroup_subsys_state *css;
5150 struct cgrp_cset_link *link;
5153 lockdep_assert_held(&cgroup_mutex);
5156 * Only migration can raise populated from zero and we're already
5157 * holding cgroup_mutex.
5159 if (cgroup_is_populated(cgrp))
5163 * Make sure there's no live children. We can't test emptiness of
5164 * ->self.children as dead children linger on it while being
5165 * drained; otherwise, "rmdir parent/child parent" may fail.
5167 if (css_has_online_children(&cgrp->self))
5171 * Mark @cgrp and the associated csets dead. The former prevents
5172 * further task migration and child creation by disabling
5173 * cgroup_lock_live_group(). The latter makes the csets ignored by
5174 * the migration path.
5176 cgrp->self.flags &= ~CSS_ONLINE;
5178 spin_lock_bh(&css_set_lock);
5179 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5180 link->cset->dead = true;
5181 spin_unlock_bh(&css_set_lock);
5183 /* initiate massacre of all css's */
5184 for_each_css(css, ssid, cgrp)
5188 * Remove @cgrp directory along with the base files. @cgrp has an
5189 * extra ref on its kn.
5191 kernfs_remove(cgrp->kn);
5193 check_for_release(cgroup_parent(cgrp));
5195 /* put the base reference */
5196 percpu_ref_kill(&cgrp->self.refcnt);
5201 static int cgroup_rmdir(struct kernfs_node *kn)
5203 struct cgroup *cgrp;
5206 cgrp = cgroup_kn_lock_live(kn);
5210 ret = cgroup_destroy_locked(cgrp);
5212 cgroup_kn_unlock(kn);
5216 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5217 .remount_fs = cgroup_remount,
5218 .show_options = cgroup_show_options,
5219 .mkdir = cgroup_mkdir,
5220 .rmdir = cgroup_rmdir,
5221 .rename = cgroup_rename,
5224 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5226 struct cgroup_subsys_state *css;
5228 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
5230 mutex_lock(&cgroup_mutex);
5232 idr_init(&ss->css_idr);
5233 INIT_LIST_HEAD(&ss->cfts);
5235 /* Create the root cgroup state for this subsystem */
5236 ss->root = &cgrp_dfl_root;
5237 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5238 /* We don't handle early failures gracefully */
5239 BUG_ON(IS_ERR(css));
5240 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5243 * Root csses are never destroyed and we can't initialize
5244 * percpu_ref during early init. Disable refcnting.
5246 css->flags |= CSS_NO_REF;
5249 /* allocation can't be done safely during early init */
5252 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5253 BUG_ON(css->id < 0);
5256 /* Update the init_css_set to contain a subsys
5257 * pointer to this state - since the subsystem is
5258 * newly registered, all tasks and hence the
5259 * init_css_set is in the subsystem's root cgroup. */
5260 init_css_set.subsys[ss->id] = css;
5262 have_fork_callback |= (bool)ss->fork << ss->id;
5263 have_exit_callback |= (bool)ss->exit << ss->id;
5264 have_free_callback |= (bool)ss->free << ss->id;
5265 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5267 /* At system boot, before all subsystems have been
5268 * registered, no tasks have been forked, so we don't
5269 * need to invoke fork callbacks here. */
5270 BUG_ON(!list_empty(&init_task.tasks));
5272 BUG_ON(online_css(css));
5274 mutex_unlock(&cgroup_mutex);
5278 * cgroup_init_early - cgroup initialization at system boot
5280 * Initialize cgroups at system boot, and initialize any
5281 * subsystems that request early init.
5283 int __init cgroup_init_early(void)
5285 static struct cgroup_sb_opts __initdata opts;
5286 struct cgroup_subsys *ss;
5289 init_cgroup_root(&cgrp_dfl_root, &opts);
5290 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5292 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5294 for_each_subsys(ss, i) {
5295 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5296 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5297 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5299 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5300 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5303 ss->name = cgroup_subsys_name[i];
5304 if (!ss->legacy_name)
5305 ss->legacy_name = cgroup_subsys_name[i];
5308 cgroup_init_subsys(ss, true);
5313 static unsigned long cgroup_disable_mask __initdata;
5316 * cgroup_init - cgroup initialization
5318 * Register cgroup filesystem and /proc file, and initialize
5319 * any subsystems that didn't request early init.
5321 int __init cgroup_init(void)
5323 struct cgroup_subsys *ss;
5327 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5328 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5329 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5331 mutex_lock(&cgroup_mutex);
5333 /* Add init_css_set to the hash table */
5334 key = css_set_hash(init_css_set.subsys);
5335 hash_add(css_set_table, &init_css_set.hlist, key);
5337 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5339 mutex_unlock(&cgroup_mutex);
5341 for_each_subsys(ss, ssid) {
5342 if (ss->early_init) {
5343 struct cgroup_subsys_state *css =
5344 init_css_set.subsys[ss->id];
5346 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5348 BUG_ON(css->id < 0);
5350 cgroup_init_subsys(ss, false);
5353 list_add_tail(&init_css_set.e_cset_node[ssid],
5354 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5357 * Setting dfl_root subsys_mask needs to consider the
5358 * disabled flag and cftype registration needs kmalloc,
5359 * both of which aren't available during early_init.
5361 if (cgroup_disable_mask & (1 << ssid)) {
5362 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5363 printk(KERN_INFO "Disabling %s control group subsystem\n",
5368 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5370 if (!ss->dfl_cftypes)
5371 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5373 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5374 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5376 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5377 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5381 ss->bind(init_css_set.subsys[ssid]);
5384 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5385 WARN_ON(register_filesystem(&cgroup_fs_type));
5386 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5391 static int __init cgroup_wq_init(void)
5394 * There isn't much point in executing destruction path in
5395 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5396 * Use 1 for @max_active.
5398 * We would prefer to do this in cgroup_init() above, but that
5399 * is called before init_workqueues(): so leave this until after.
5401 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5402 BUG_ON(!cgroup_destroy_wq);
5405 * Used to destroy pidlists and separate to serve as flush domain.
5406 * Cap @max_active to 1 too.
5408 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5410 BUG_ON(!cgroup_pidlist_destroy_wq);
5414 core_initcall(cgroup_wq_init);
5417 * proc_cgroup_show()
5418 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5419 * - Used for /proc/<pid>/cgroup.
5421 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5422 struct pid *pid, struct task_struct *tsk)
5426 struct cgroup_root *root;
5429 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5433 mutex_lock(&cgroup_mutex);
5434 spin_lock_bh(&css_set_lock);
5436 for_each_root(root) {
5437 struct cgroup_subsys *ss;
5438 struct cgroup *cgrp;
5439 int ssid, count = 0;
5441 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5444 seq_printf(m, "%d:", root->hierarchy_id);
5445 if (root != &cgrp_dfl_root)
5446 for_each_subsys(ss, ssid)
5447 if (root->subsys_mask & (1 << ssid))
5448 seq_printf(m, "%s%s", count++ ? "," : "",
5450 if (strlen(root->name))
5451 seq_printf(m, "%sname=%s", count ? "," : "",
5455 cgrp = task_cgroup_from_root(tsk, root);
5458 * On traditional hierarchies, all zombie tasks show up as
5459 * belonging to the root cgroup. On the default hierarchy,
5460 * while a zombie doesn't show up in "cgroup.procs" and
5461 * thus can't be migrated, its /proc/PID/cgroup keeps
5462 * reporting the cgroup it belonged to before exiting. If
5463 * the cgroup is removed before the zombie is reaped,
5464 * " (deleted)" is appended to the cgroup path.
5466 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5467 path = cgroup_path(cgrp, buf, PATH_MAX);
5469 retval = -ENAMETOOLONG;
5478 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5479 seq_puts(m, " (deleted)\n");
5486 spin_unlock_bh(&css_set_lock);
5487 mutex_unlock(&cgroup_mutex);
5493 /* Display information about each subsystem and each hierarchy */
5494 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5496 struct cgroup_subsys *ss;
5499 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5501 * ideally we don't want subsystems moving around while we do this.
5502 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5503 * subsys/hierarchy state.
5505 mutex_lock(&cgroup_mutex);
5507 for_each_subsys(ss, i)
5508 seq_printf(m, "%s\t%d\t%d\t%d\n",
5509 ss->legacy_name, ss->root->hierarchy_id,
5510 atomic_read(&ss->root->nr_cgrps),
5511 cgroup_ssid_enabled(i));
5513 mutex_unlock(&cgroup_mutex);
5517 static int cgroupstats_open(struct inode *inode, struct file *file)
5519 return single_open(file, proc_cgroupstats_show, NULL);
5522 static const struct file_operations proc_cgroupstats_operations = {
5523 .open = cgroupstats_open,
5525 .llseek = seq_lseek,
5526 .release = single_release,
5529 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5531 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5532 return &ss_priv[i - CGROUP_CANFORK_START];
5536 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5538 void **private = subsys_canfork_priv_p(ss_priv, i);
5539 return private ? *private : NULL;
5543 * cgroup_fork - initialize cgroup related fields during copy_process()
5544 * @child: pointer to task_struct of forking parent process.
5546 * A task is associated with the init_css_set until cgroup_post_fork()
5547 * attaches it to the parent's css_set. Empty cg_list indicates that
5548 * @child isn't holding reference to its css_set.
5550 void cgroup_fork(struct task_struct *child)
5552 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5553 INIT_LIST_HEAD(&child->cg_list);
5557 * cgroup_can_fork - called on a new task before the process is exposed
5558 * @child: the task in question.
5560 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5561 * returns an error, the fork aborts with that error code. This allows for
5562 * a cgroup subsystem to conditionally allow or deny new forks.
5564 int cgroup_can_fork(struct task_struct *child,
5565 void *ss_priv[CGROUP_CANFORK_COUNT])
5567 struct cgroup_subsys *ss;
5570 for_each_subsys_which(ss, i, &have_canfork_callback) {
5571 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5579 for_each_subsys(ss, j) {
5582 if (ss->cancel_fork)
5583 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5590 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5591 * @child: the task in question
5593 * This calls the cancel_fork() callbacks if a fork failed *after*
5594 * cgroup_can_fork() succeded.
5596 void cgroup_cancel_fork(struct task_struct *child,
5597 void *ss_priv[CGROUP_CANFORK_COUNT])
5599 struct cgroup_subsys *ss;
5602 for_each_subsys(ss, i)
5603 if (ss->cancel_fork)
5604 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5608 * cgroup_post_fork - called on a new task after adding it to the task list
5609 * @child: the task in question
5611 * Adds the task to the list running through its css_set if necessary and
5612 * call the subsystem fork() callbacks. Has to be after the task is
5613 * visible on the task list in case we race with the first call to
5614 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5617 void cgroup_post_fork(struct task_struct *child,
5618 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5620 struct cgroup_subsys *ss;
5624 * This may race against cgroup_enable_task_cg_lists(). As that
5625 * function sets use_task_css_set_links before grabbing
5626 * tasklist_lock and we just went through tasklist_lock to add
5627 * @child, it's guaranteed that either we see the set
5628 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5629 * @child during its iteration.
5631 * If we won the race, @child is associated with %current's
5632 * css_set. Grabbing css_set_lock guarantees both that the
5633 * association is stable, and, on completion of the parent's
5634 * migration, @child is visible in the source of migration or
5635 * already in the destination cgroup. This guarantee is necessary
5636 * when implementing operations which need to migrate all tasks of
5637 * a cgroup to another.
5639 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5640 * will remain in init_css_set. This is safe because all tasks are
5641 * in the init_css_set before cg_links is enabled and there's no
5642 * operation which transfers all tasks out of init_css_set.
5644 if (use_task_css_set_links) {
5645 struct css_set *cset;
5647 spin_lock_bh(&css_set_lock);
5648 cset = task_css_set(current);
5649 if (list_empty(&child->cg_list)) {
5651 css_set_move_task(child, NULL, cset, false);
5653 spin_unlock_bh(&css_set_lock);
5657 * Call ss->fork(). This must happen after @child is linked on
5658 * css_set; otherwise, @child might change state between ->fork()
5659 * and addition to css_set.
5661 for_each_subsys_which(ss, i, &have_fork_callback)
5662 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5666 * cgroup_exit - detach cgroup from exiting task
5667 * @tsk: pointer to task_struct of exiting process
5669 * Description: Detach cgroup from @tsk and release it.
5671 * Note that cgroups marked notify_on_release force every task in
5672 * them to take the global cgroup_mutex mutex when exiting.
5673 * This could impact scaling on very large systems. Be reluctant to
5674 * use notify_on_release cgroups where very high task exit scaling
5675 * is required on large systems.
5677 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5678 * call cgroup_exit() while the task is still competent to handle
5679 * notify_on_release(), then leave the task attached to the root cgroup in
5680 * each hierarchy for the remainder of its exit. No need to bother with
5681 * init_css_set refcnting. init_css_set never goes away and we can't race
5682 * with migration path - PF_EXITING is visible to migration path.
5684 void cgroup_exit(struct task_struct *tsk)
5686 struct cgroup_subsys *ss;
5687 struct css_set *cset;
5691 * Unlink from @tsk from its css_set. As migration path can't race
5692 * with us, we can check css_set and cg_list without synchronization.
5694 cset = task_css_set(tsk);
5696 if (!list_empty(&tsk->cg_list)) {
5697 spin_lock_bh(&css_set_lock);
5698 css_set_move_task(tsk, cset, NULL, false);
5699 spin_unlock_bh(&css_set_lock);
5704 /* see cgroup_post_fork() for details */
5705 for_each_subsys_which(ss, i, &have_exit_callback)
5709 void cgroup_free(struct task_struct *task)
5711 struct css_set *cset = task_css_set(task);
5712 struct cgroup_subsys *ss;
5715 for_each_subsys_which(ss, ssid, &have_free_callback)
5721 static void check_for_release(struct cgroup *cgrp)
5723 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5724 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5725 schedule_work(&cgrp->release_agent_work);
5729 * Notify userspace when a cgroup is released, by running the
5730 * configured release agent with the name of the cgroup (path
5731 * relative to the root of cgroup file system) as the argument.
5733 * Most likely, this user command will try to rmdir this cgroup.
5735 * This races with the possibility that some other task will be
5736 * attached to this cgroup before it is removed, or that some other
5737 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5738 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5739 * unused, and this cgroup will be reprieved from its death sentence,
5740 * to continue to serve a useful existence. Next time it's released,
5741 * we will get notified again, if it still has 'notify_on_release' set.
5743 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5744 * means only wait until the task is successfully execve()'d. The
5745 * separate release agent task is forked by call_usermodehelper(),
5746 * then control in this thread returns here, without waiting for the
5747 * release agent task. We don't bother to wait because the caller of
5748 * this routine has no use for the exit status of the release agent
5749 * task, so no sense holding our caller up for that.
5751 static void cgroup_release_agent(struct work_struct *work)
5753 struct cgroup *cgrp =
5754 container_of(work, struct cgroup, release_agent_work);
5755 char *pathbuf = NULL, *agentbuf = NULL, *path;
5756 char *argv[3], *envp[3];
5758 mutex_lock(&cgroup_mutex);
5760 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5761 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5762 if (!pathbuf || !agentbuf)
5765 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5773 /* minimal command environment */
5775 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5778 mutex_unlock(&cgroup_mutex);
5779 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5782 mutex_unlock(&cgroup_mutex);
5788 static int __init cgroup_disable(char *str)
5790 struct cgroup_subsys *ss;
5794 while ((token = strsep(&str, ",")) != NULL) {
5798 for_each_subsys(ss, i) {
5799 if (strcmp(token, ss->name) &&
5800 strcmp(token, ss->legacy_name))
5802 cgroup_disable_mask |= 1 << i;
5807 __setup("cgroup_disable=", cgroup_disable);
5810 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5811 * @dentry: directory dentry of interest
5812 * @ss: subsystem of interest
5814 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5815 * to get the corresponding css and return it. If such css doesn't exist
5816 * or can't be pinned, an ERR_PTR value is returned.
5818 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5819 struct cgroup_subsys *ss)
5821 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5822 struct cgroup_subsys_state *css = NULL;
5823 struct cgroup *cgrp;
5825 /* is @dentry a cgroup dir? */
5826 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5827 kernfs_type(kn) != KERNFS_DIR)
5828 return ERR_PTR(-EBADF);
5833 * This path doesn't originate from kernfs and @kn could already
5834 * have been or be removed at any point. @kn->priv is RCU
5835 * protected for this access. See css_release_work_fn() for details.
5837 cgrp = rcu_dereference(kn->priv);
5839 css = cgroup_css(cgrp, ss);
5841 if (!css || !css_tryget_online(css))
5842 css = ERR_PTR(-ENOENT);
5849 * css_from_id - lookup css by id
5850 * @id: the cgroup id
5851 * @ss: cgroup subsys to be looked into
5853 * Returns the css if there's valid one with @id, otherwise returns NULL.
5854 * Should be called under rcu_read_lock().
5856 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5858 WARN_ON_ONCE(!rcu_read_lock_held());
5859 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5862 #ifdef CONFIG_CGROUP_DEBUG
5863 static struct cgroup_subsys_state *
5864 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5866 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5869 return ERR_PTR(-ENOMEM);
5874 static void debug_css_free(struct cgroup_subsys_state *css)
5879 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5882 return cgroup_task_count(css->cgroup);
5885 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5888 return (u64)(unsigned long)current->cgroups;
5891 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5897 count = atomic_read(&task_css_set(current)->refcount);
5902 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5904 struct cgrp_cset_link *link;
5905 struct css_set *cset;
5908 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5912 spin_lock_bh(&css_set_lock);
5914 cset = rcu_dereference(current->cgroups);
5915 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5916 struct cgroup *c = link->cgrp;
5918 cgroup_name(c, name_buf, NAME_MAX + 1);
5919 seq_printf(seq, "Root %d group %s\n",
5920 c->root->hierarchy_id, name_buf);
5923 spin_unlock_bh(&css_set_lock);
5928 #define MAX_TASKS_SHOWN_PER_CSS 25
5929 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5931 struct cgroup_subsys_state *css = seq_css(seq);
5932 struct cgrp_cset_link *link;
5934 spin_lock_bh(&css_set_lock);
5935 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5936 struct css_set *cset = link->cset;
5937 struct task_struct *task;
5940 seq_printf(seq, "css_set %p\n", cset);
5942 list_for_each_entry(task, &cset->tasks, cg_list) {
5943 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5945 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5948 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5949 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5951 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5955 seq_puts(seq, " ...\n");
5957 spin_unlock_bh(&css_set_lock);
5961 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5963 return (!cgroup_is_populated(css->cgroup) &&
5964 !css_has_online_children(&css->cgroup->self));
5967 static struct cftype debug_files[] = {
5969 .name = "taskcount",
5970 .read_u64 = debug_taskcount_read,
5974 .name = "current_css_set",
5975 .read_u64 = current_css_set_read,
5979 .name = "current_css_set_refcount",
5980 .read_u64 = current_css_set_refcount_read,
5984 .name = "current_css_set_cg_links",
5985 .seq_show = current_css_set_cg_links_read,
5989 .name = "cgroup_css_links",
5990 .seq_show = cgroup_css_links_read,
5994 .name = "releasable",
5995 .read_u64 = releasable_read,
6001 struct cgroup_subsys debug_cgrp_subsys = {
6002 .css_alloc = debug_css_alloc,
6003 .css_free = debug_css_free,
6004 .legacy_cftypes = debug_files,
6006 #endif /* CONFIG_CGROUP_DEBUG */