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/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 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
243 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
244 * @cgrp: the cgroup of interest
246 * The default hierarchy is the v2 interface of cgroup and this function
247 * can be used to test whether a cgroup is on the default hierarchy for
248 * cases where a subsystem should behave differnetly depending on the
251 * The set of behaviors which change on the default hierarchy are still
252 * being determined and the mount option is prefixed with __DEVEL__.
254 * List of changed behaviors:
256 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
257 * and "name" are disallowed.
259 * - When mounting an existing superblock, mount options should match.
261 * - Remount is disallowed.
263 * - rename(2) is disallowed.
265 * - "tasks" is removed. Everything should be at process granularity. Use
266 * "cgroup.procs" instead.
268 * - "cgroup.procs" is not sorted. pids will be unique unless they got
269 * recycled inbetween reads.
271 * - "release_agent" and "notify_on_release" are removed. Replacement
272 * notification mechanism will be implemented.
274 * - "cgroup.clone_children" is removed.
276 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
277 * and its descendants contain no task; otherwise, 1. The file also
278 * generates kernfs notification which can be monitored through poll and
279 * [di]notify when the value of the file changes.
281 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
282 * take masks of ancestors with non-empty cpus/mems, instead of being
283 * moved to an ancestor.
285 * - cpuset: a task can be moved into an empty cpuset, and again it takes
286 * masks of ancestors.
288 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
291 * - blkcg: blk-throttle becomes properly hierarchical.
293 * - debug: disallowed on the default hierarchy.
295 static bool cgroup_on_dfl(const struct cgroup *cgrp)
297 return cgrp->root == &cgrp_dfl_root;
300 /* IDR wrappers which synchronize using cgroup_idr_lock */
301 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
306 idr_preload(gfp_mask);
307 spin_lock_bh(&cgroup_idr_lock);
308 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
309 spin_unlock_bh(&cgroup_idr_lock);
314 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
318 spin_lock_bh(&cgroup_idr_lock);
319 ret = idr_replace(idr, ptr, id);
320 spin_unlock_bh(&cgroup_idr_lock);
324 static void cgroup_idr_remove(struct idr *idr, int id)
326 spin_lock_bh(&cgroup_idr_lock);
328 spin_unlock_bh(&cgroup_idr_lock);
331 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
333 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
336 return container_of(parent_css, struct cgroup, self);
341 * cgroup_css - obtain a cgroup's css for the specified subsystem
342 * @cgrp: the cgroup of interest
343 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
345 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
346 * function must be called either under cgroup_mutex or rcu_read_lock() and
347 * the caller is responsible for pinning the returned css if it wants to
348 * keep accessing it outside the said locks. This function may return
349 * %NULL if @cgrp doesn't have @subsys_id enabled.
351 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
352 struct cgroup_subsys *ss)
355 return rcu_dereference_check(cgrp->subsys[ss->id],
356 lockdep_is_held(&cgroup_mutex));
362 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
363 * @cgrp: the cgroup of interest
364 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
366 * Similar to cgroup_css() but returns the effective css, which is defined
367 * as the matching css of the nearest ancestor including self which has @ss
368 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
369 * function is guaranteed to return non-NULL css.
371 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
372 struct cgroup_subsys *ss)
374 lockdep_assert_held(&cgroup_mutex);
379 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
383 * This function is used while updating css associations and thus
384 * can't test the csses directly. Use ->child_subsys_mask.
386 while (cgroup_parent(cgrp) &&
387 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
388 cgrp = cgroup_parent(cgrp);
390 return cgroup_css(cgrp, ss);
394 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
395 * @cgrp: the cgroup of interest
396 * @ss: the subsystem of interest
398 * Find and get the effective css of @cgrp for @ss. The effective css is
399 * defined as the matching css of the nearest ancestor including self which
400 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
401 * the root css is returned, so this function always returns a valid css.
402 * The returned css must be put using css_put().
404 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
405 struct cgroup_subsys *ss)
407 struct cgroup_subsys_state *css;
412 css = cgroup_css(cgrp, ss);
414 if (css && css_tryget_online(css))
416 cgrp = cgroup_parent(cgrp);
419 css = init_css_set.subsys[ss->id];
426 /* convenient tests for these bits */
427 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
429 return !(cgrp->self.flags & CSS_ONLINE);
432 static void cgroup_get(struct cgroup *cgrp)
434 WARN_ON_ONCE(cgroup_is_dead(cgrp));
435 css_get(&cgrp->self);
438 static bool cgroup_tryget(struct cgroup *cgrp)
440 return css_tryget(&cgrp->self);
443 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
445 struct cgroup *cgrp = of->kn->parent->priv;
446 struct cftype *cft = of_cft(of);
449 * This is open and unprotected implementation of cgroup_css().
450 * seq_css() is only called from a kernfs file operation which has
451 * an active reference on the file. Because all the subsystem
452 * files are drained before a css is disassociated with a cgroup,
453 * the matching css from the cgroup's subsys table is guaranteed to
454 * be and stay valid until the enclosing operation is complete.
457 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
461 EXPORT_SYMBOL_GPL(of_css);
463 static int notify_on_release(const struct cgroup *cgrp)
465 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
469 * for_each_css - iterate all css's of a cgroup
470 * @css: the iteration cursor
471 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
472 * @cgrp: the target cgroup to iterate css's of
474 * Should be called under cgroup_[tree_]mutex.
476 #define for_each_css(css, ssid, cgrp) \
477 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
478 if (!((css) = rcu_dereference_check( \
479 (cgrp)->subsys[(ssid)], \
480 lockdep_is_held(&cgroup_mutex)))) { } \
484 * for_each_e_css - iterate all effective css's of a cgroup
485 * @css: the iteration cursor
486 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
487 * @cgrp: the target cgroup to iterate css's of
489 * Should be called under cgroup_[tree_]mutex.
491 #define for_each_e_css(css, ssid, cgrp) \
492 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
493 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
498 * for_each_subsys - iterate all enabled cgroup subsystems
499 * @ss: the iteration cursor
500 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
502 #define for_each_subsys(ss, ssid) \
503 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
504 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
507 * for_each_subsys_which - filter for_each_subsys with a bitmask
508 * @ss: the iteration cursor
509 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
510 * @ss_maskp: a pointer to the bitmask
512 * The block will only run for cases where the ssid-th bit (1 << ssid) of
515 #define for_each_subsys_which(ss, ssid, ss_maskp) \
516 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
519 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
520 if (((ss) = cgroup_subsys[ssid]) && false) \
524 /* iterate across the hierarchies */
525 #define for_each_root(root) \
526 list_for_each_entry((root), &cgroup_roots, root_list)
528 /* iterate over child cgrps, lock should be held throughout iteration */
529 #define cgroup_for_each_live_child(child, cgrp) \
530 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
531 if (({ lockdep_assert_held(&cgroup_mutex); \
532 cgroup_is_dead(child); })) \
536 static void cgroup_release_agent(struct work_struct *work);
537 static void check_for_release(struct cgroup *cgrp);
540 * A cgroup can be associated with multiple css_sets as different tasks may
541 * belong to different cgroups on different hierarchies. In the other
542 * direction, a css_set is naturally associated with multiple cgroups.
543 * This M:N relationship is represented by the following link structure
544 * which exists for each association and allows traversing the associations
547 struct cgrp_cset_link {
548 /* the cgroup and css_set this link associates */
550 struct css_set *cset;
552 /* list of cgrp_cset_links anchored at cgrp->cset_links */
553 struct list_head cset_link;
555 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
556 struct list_head cgrp_link;
560 * The default css_set - used by init and its children prior to any
561 * hierarchies being mounted. It contains a pointer to the root state
562 * for each subsystem. Also used to anchor the list of css_sets. Not
563 * reference-counted, to improve performance when child cgroups
564 * haven't been created.
566 struct css_set init_css_set = {
567 .refcount = ATOMIC_INIT(1),
568 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
569 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
570 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
571 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
572 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
573 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
576 static int css_set_count = 1; /* 1 for init_css_set */
579 * css_set_populated - does a css_set contain any tasks?
580 * @cset: target css_set
582 static bool css_set_populated(struct css_set *cset)
584 lockdep_assert_held(&css_set_lock);
586 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
590 * cgroup_update_populated - updated populated count of a cgroup
591 * @cgrp: the target cgroup
592 * @populated: inc or dec populated count
594 * One of the css_sets associated with @cgrp is either getting its first
595 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
596 * count is propagated towards root so that a given cgroup's populated_cnt
597 * is zero iff the cgroup and all its descendants don't contain any tasks.
599 * @cgrp's interface file "cgroup.populated" is zero if
600 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
601 * changes from or to zero, userland is notified that the content of the
602 * interface file has changed. This can be used to detect when @cgrp and
603 * its descendants become populated or empty.
605 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
607 lockdep_assert_held(&css_set_lock);
613 trigger = !cgrp->populated_cnt++;
615 trigger = !--cgrp->populated_cnt;
620 check_for_release(cgrp);
621 cgroup_file_notify(&cgrp->events_file);
623 cgrp = cgroup_parent(cgrp);
628 * css_set_update_populated - update populated state of a css_set
629 * @cset: target css_set
630 * @populated: whether @cset is populated or depopulated
632 * @cset is either getting the first task or losing the last. Update the
633 * ->populated_cnt of all associated cgroups accordingly.
635 static void css_set_update_populated(struct css_set *cset, bool populated)
637 struct cgrp_cset_link *link;
639 lockdep_assert_held(&css_set_lock);
641 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
642 cgroup_update_populated(link->cgrp, populated);
646 * css_set_move_task - move a task from one css_set to another
647 * @task: task being moved
648 * @from_cset: css_set @task currently belongs to (may be NULL)
649 * @to_cset: new css_set @task is being moved to (may be NULL)
650 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
652 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
653 * css_set, @from_cset can be NULL. If @task is being disassociated
654 * instead of moved, @to_cset can be NULL.
656 * This function automatically handles populated_cnt updates and
657 * css_task_iter adjustments but the caller is responsible for managing
658 * @from_cset and @to_cset's reference counts.
660 static void css_set_move_task(struct task_struct *task,
661 struct css_set *from_cset, struct css_set *to_cset,
664 lockdep_assert_held(&css_set_lock);
667 struct css_task_iter *it, *pos;
669 WARN_ON_ONCE(list_empty(&task->cg_list));
672 * @task is leaving, advance task iterators which are
673 * pointing to it so that they can resume at the next
674 * position. Advancing an iterator might remove it from
675 * the list, use safe walk. See css_task_iter_advance*()
678 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
680 if (it->task_pos == &task->cg_list)
681 css_task_iter_advance(it);
683 list_del_init(&task->cg_list);
684 if (!css_set_populated(from_cset))
685 css_set_update_populated(from_cset, false);
687 WARN_ON_ONCE(!list_empty(&task->cg_list));
692 * We are synchronized through cgroup_threadgroup_rwsem
693 * against PF_EXITING setting such that we can't race
694 * against cgroup_exit() changing the css_set to
695 * init_css_set and dropping the old one.
697 WARN_ON_ONCE(task->flags & PF_EXITING);
699 if (!css_set_populated(to_cset))
700 css_set_update_populated(to_cset, true);
701 rcu_assign_pointer(task->cgroups, to_cset);
702 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
708 * hash table for cgroup groups. This improves the performance to find
709 * an existing css_set. This hash doesn't (currently) take into
710 * account cgroups in empty hierarchies.
712 #define CSS_SET_HASH_BITS 7
713 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
715 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
717 unsigned long key = 0UL;
718 struct cgroup_subsys *ss;
721 for_each_subsys(ss, i)
722 key += (unsigned long)css[i];
723 key = (key >> 16) ^ key;
728 static void put_css_set_locked(struct css_set *cset)
730 struct cgrp_cset_link *link, *tmp_link;
731 struct cgroup_subsys *ss;
734 lockdep_assert_held(&css_set_lock);
736 if (!atomic_dec_and_test(&cset->refcount))
739 /* This css_set is dead. unlink it and release cgroup and css refs */
740 for_each_subsys(ss, ssid) {
741 list_del(&cset->e_cset_node[ssid]);
742 css_put(cset->subsys[ssid]);
744 hash_del(&cset->hlist);
747 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
748 list_del(&link->cset_link);
749 list_del(&link->cgrp_link);
750 if (cgroup_parent(link->cgrp))
751 cgroup_put(link->cgrp);
755 kfree_rcu(cset, rcu_head);
758 static void put_css_set(struct css_set *cset)
761 * Ensure that the refcount doesn't hit zero while any readers
762 * can see it. Similar to atomic_dec_and_lock(), but for an
765 if (atomic_add_unless(&cset->refcount, -1, 1))
768 spin_lock_bh(&css_set_lock);
769 put_css_set_locked(cset);
770 spin_unlock_bh(&css_set_lock);
774 * refcounted get/put for css_set objects
776 static inline void get_css_set(struct css_set *cset)
778 atomic_inc(&cset->refcount);
782 * compare_css_sets - helper function for find_existing_css_set().
783 * @cset: candidate css_set being tested
784 * @old_cset: existing css_set for a task
785 * @new_cgrp: cgroup that's being entered by the task
786 * @template: desired set of css pointers in css_set (pre-calculated)
788 * Returns true if "cset" matches "old_cset" except for the hierarchy
789 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
791 static bool compare_css_sets(struct css_set *cset,
792 struct css_set *old_cset,
793 struct cgroup *new_cgrp,
794 struct cgroup_subsys_state *template[])
796 struct list_head *l1, *l2;
799 * On the default hierarchy, there can be csets which are
800 * associated with the same set of cgroups but different csses.
801 * Let's first ensure that csses match.
803 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
807 * Compare cgroup pointers in order to distinguish between
808 * different cgroups in hierarchies. As different cgroups may
809 * share the same effective css, this comparison is always
812 l1 = &cset->cgrp_links;
813 l2 = &old_cset->cgrp_links;
815 struct cgrp_cset_link *link1, *link2;
816 struct cgroup *cgrp1, *cgrp2;
820 /* See if we reached the end - both lists are equal length. */
821 if (l1 == &cset->cgrp_links) {
822 BUG_ON(l2 != &old_cset->cgrp_links);
825 BUG_ON(l2 == &old_cset->cgrp_links);
827 /* Locate the cgroups associated with these links. */
828 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
829 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
832 /* Hierarchies should be linked in the same order. */
833 BUG_ON(cgrp1->root != cgrp2->root);
836 * If this hierarchy is the hierarchy of the cgroup
837 * that's changing, then we need to check that this
838 * css_set points to the new cgroup; if it's any other
839 * hierarchy, then this css_set should point to the
840 * same cgroup as the old css_set.
842 if (cgrp1->root == new_cgrp->root) {
843 if (cgrp1 != new_cgrp)
854 * find_existing_css_set - init css array and find the matching css_set
855 * @old_cset: the css_set that we're using before the cgroup transition
856 * @cgrp: the cgroup that we're moving into
857 * @template: out param for the new set of csses, should be clear on entry
859 static struct css_set *find_existing_css_set(struct css_set *old_cset,
861 struct cgroup_subsys_state *template[])
863 struct cgroup_root *root = cgrp->root;
864 struct cgroup_subsys *ss;
865 struct css_set *cset;
870 * Build the set of subsystem state objects that we want to see in the
871 * new css_set. while subsystems can change globally, the entries here
872 * won't change, so no need for locking.
874 for_each_subsys(ss, i) {
875 if (root->subsys_mask & (1UL << i)) {
877 * @ss is in this hierarchy, so we want the
878 * effective css from @cgrp.
880 template[i] = cgroup_e_css(cgrp, ss);
883 * @ss is not in this hierarchy, so we don't want
886 template[i] = old_cset->subsys[i];
890 key = css_set_hash(template);
891 hash_for_each_possible(css_set_table, cset, hlist, key) {
892 if (!compare_css_sets(cset, old_cset, cgrp, template))
895 /* This css_set matches what we need */
899 /* No existing cgroup group matched */
903 static void free_cgrp_cset_links(struct list_head *links_to_free)
905 struct cgrp_cset_link *link, *tmp_link;
907 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
908 list_del(&link->cset_link);
914 * allocate_cgrp_cset_links - allocate cgrp_cset_links
915 * @count: the number of links to allocate
916 * @tmp_links: list_head the allocated links are put on
918 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
919 * through ->cset_link. Returns 0 on success or -errno.
921 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
923 struct cgrp_cset_link *link;
926 INIT_LIST_HEAD(tmp_links);
928 for (i = 0; i < count; i++) {
929 link = kzalloc(sizeof(*link), GFP_KERNEL);
931 free_cgrp_cset_links(tmp_links);
934 list_add(&link->cset_link, tmp_links);
940 * link_css_set - a helper function to link a css_set to a cgroup
941 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
942 * @cset: the css_set to be linked
943 * @cgrp: the destination cgroup
945 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
948 struct cgrp_cset_link *link;
950 BUG_ON(list_empty(tmp_links));
952 if (cgroup_on_dfl(cgrp))
953 cset->dfl_cgrp = cgrp;
955 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
960 * Always add links to the tail of the lists so that the lists are
961 * in choronological order.
963 list_move_tail(&link->cset_link, &cgrp->cset_links);
964 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
966 if (cgroup_parent(cgrp))
971 * find_css_set - return a new css_set with one cgroup updated
972 * @old_cset: the baseline css_set
973 * @cgrp: the cgroup to be updated
975 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
976 * substituted into the appropriate hierarchy.
978 static struct css_set *find_css_set(struct css_set *old_cset,
981 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
982 struct css_set *cset;
983 struct list_head tmp_links;
984 struct cgrp_cset_link *link;
985 struct cgroup_subsys *ss;
989 lockdep_assert_held(&cgroup_mutex);
991 /* First see if we already have a cgroup group that matches
993 spin_lock_bh(&css_set_lock);
994 cset = find_existing_css_set(old_cset, cgrp, template);
997 spin_unlock_bh(&css_set_lock);
1002 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1006 /* Allocate all the cgrp_cset_link objects that we'll need */
1007 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1012 atomic_set(&cset->refcount, 1);
1013 INIT_LIST_HEAD(&cset->cgrp_links);
1014 INIT_LIST_HEAD(&cset->tasks);
1015 INIT_LIST_HEAD(&cset->mg_tasks);
1016 INIT_LIST_HEAD(&cset->mg_preload_node);
1017 INIT_LIST_HEAD(&cset->mg_node);
1018 INIT_LIST_HEAD(&cset->task_iters);
1019 INIT_HLIST_NODE(&cset->hlist);
1021 /* Copy the set of subsystem state objects generated in
1022 * find_existing_css_set() */
1023 memcpy(cset->subsys, template, sizeof(cset->subsys));
1025 spin_lock_bh(&css_set_lock);
1026 /* Add reference counts and links from the new css_set. */
1027 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1028 struct cgroup *c = link->cgrp;
1030 if (c->root == cgrp->root)
1032 link_css_set(&tmp_links, cset, c);
1035 BUG_ON(!list_empty(&tmp_links));
1039 /* Add @cset to the hash table */
1040 key = css_set_hash(cset->subsys);
1041 hash_add(css_set_table, &cset->hlist, key);
1043 for_each_subsys(ss, ssid) {
1044 struct cgroup_subsys_state *css = cset->subsys[ssid];
1046 list_add_tail(&cset->e_cset_node[ssid],
1047 &css->cgroup->e_csets[ssid]);
1051 spin_unlock_bh(&css_set_lock);
1056 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1058 struct cgroup *root_cgrp = kf_root->kn->priv;
1060 return root_cgrp->root;
1063 static int cgroup_init_root_id(struct cgroup_root *root)
1067 lockdep_assert_held(&cgroup_mutex);
1069 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1073 root->hierarchy_id = id;
1077 static void cgroup_exit_root_id(struct cgroup_root *root)
1079 lockdep_assert_held(&cgroup_mutex);
1081 if (root->hierarchy_id) {
1082 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1083 root->hierarchy_id = 0;
1087 static void cgroup_free_root(struct cgroup_root *root)
1090 /* hierarchy ID should already have been released */
1091 WARN_ON_ONCE(root->hierarchy_id);
1093 idr_destroy(&root->cgroup_idr);
1098 static void cgroup_destroy_root(struct cgroup_root *root)
1100 struct cgroup *cgrp = &root->cgrp;
1101 struct cgrp_cset_link *link, *tmp_link;
1103 mutex_lock(&cgroup_mutex);
1105 BUG_ON(atomic_read(&root->nr_cgrps));
1106 BUG_ON(!list_empty(&cgrp->self.children));
1108 /* Rebind all subsystems back to the default hierarchy */
1109 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1112 * Release all the links from cset_links to this hierarchy's
1115 spin_lock_bh(&css_set_lock);
1117 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1118 list_del(&link->cset_link);
1119 list_del(&link->cgrp_link);
1123 spin_unlock_bh(&css_set_lock);
1125 if (!list_empty(&root->root_list)) {
1126 list_del(&root->root_list);
1127 cgroup_root_count--;
1130 cgroup_exit_root_id(root);
1132 mutex_unlock(&cgroup_mutex);
1134 kernfs_destroy_root(root->kf_root);
1135 cgroup_free_root(root);
1138 /* look up cgroup associated with given css_set on the specified hierarchy */
1139 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1140 struct cgroup_root *root)
1142 struct cgroup *res = NULL;
1144 lockdep_assert_held(&cgroup_mutex);
1145 lockdep_assert_held(&css_set_lock);
1147 if (cset == &init_css_set) {
1150 struct cgrp_cset_link *link;
1152 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1153 struct cgroup *c = link->cgrp;
1155 if (c->root == root) {
1167 * Return the cgroup for "task" from the given hierarchy. Must be
1168 * called with cgroup_mutex and css_set_lock held.
1170 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1171 struct cgroup_root *root)
1174 * No need to lock the task - since we hold cgroup_mutex the
1175 * task can't change groups, so the only thing that can happen
1176 * is that it exits and its css is set back to init_css_set.
1178 return cset_cgroup_from_root(task_css_set(task), root);
1182 * A task must hold cgroup_mutex to modify cgroups.
1184 * Any task can increment and decrement the count field without lock.
1185 * So in general, code holding cgroup_mutex can't rely on the count
1186 * field not changing. However, if the count goes to zero, then only
1187 * cgroup_attach_task() can increment it again. Because a count of zero
1188 * means that no tasks are currently attached, therefore there is no
1189 * way a task attached to that cgroup can fork (the other way to
1190 * increment the count). So code holding cgroup_mutex can safely
1191 * assume that if the count is zero, it will stay zero. Similarly, if
1192 * a task holds cgroup_mutex on a cgroup with zero count, it
1193 * knows that the cgroup won't be removed, as cgroup_rmdir()
1196 * A cgroup can only be deleted if both its 'count' of using tasks
1197 * is zero, and its list of 'children' cgroups is empty. Since all
1198 * tasks in the system use _some_ cgroup, and since there is always at
1199 * least one task in the system (init, pid == 1), therefore, root cgroup
1200 * always has either children cgroups and/or using tasks. So we don't
1201 * need a special hack to ensure that root cgroup cannot be deleted.
1203 * P.S. One more locking exception. RCU is used to guard the
1204 * update of a tasks cgroup pointer by cgroup_attach_task()
1207 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1208 static const struct file_operations proc_cgroupstats_operations;
1210 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1213 struct cgroup_subsys *ss = cft->ss;
1215 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1216 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1217 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1218 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1221 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1226 * cgroup_file_mode - deduce file mode of a control file
1227 * @cft: the control file in question
1229 * S_IRUGO for read, S_IWUSR for write.
1231 static umode_t cgroup_file_mode(const struct cftype *cft)
1235 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1238 if (cft->write_u64 || cft->write_s64 || cft->write) {
1239 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1249 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1250 * @cgrp: the target cgroup
1251 * @subtree_control: the new subtree_control mask to consider
1253 * On the default hierarchy, a subsystem may request other subsystems to be
1254 * enabled together through its ->depends_on mask. In such cases, more
1255 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1257 * This function calculates which subsystems need to be enabled if
1258 * @subtree_control is to be applied to @cgrp. The returned mask is always
1259 * a superset of @subtree_control and follows the usual hierarchy rules.
1261 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1262 unsigned long subtree_control)
1264 struct cgroup *parent = cgroup_parent(cgrp);
1265 unsigned long cur_ss_mask = subtree_control;
1266 struct cgroup_subsys *ss;
1269 lockdep_assert_held(&cgroup_mutex);
1271 if (!cgroup_on_dfl(cgrp))
1275 unsigned long new_ss_mask = cur_ss_mask;
1277 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1278 new_ss_mask |= ss->depends_on;
1281 * Mask out subsystems which aren't available. This can
1282 * happen only if some depended-upon subsystems were bound
1283 * to non-default hierarchies.
1286 new_ss_mask &= parent->child_subsys_mask;
1288 new_ss_mask &= cgrp->root->subsys_mask;
1290 if (new_ss_mask == cur_ss_mask)
1292 cur_ss_mask = new_ss_mask;
1299 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1300 * @cgrp: the target cgroup
1302 * Update @cgrp->child_subsys_mask according to the current
1303 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1305 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1307 cgrp->child_subsys_mask =
1308 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1312 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1313 * @kn: the kernfs_node being serviced
1315 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1316 * the method finishes if locking succeeded. Note that once this function
1317 * returns the cgroup returned by cgroup_kn_lock_live() may become
1318 * inaccessible any time. If the caller intends to continue to access the
1319 * cgroup, it should pin it before invoking this function.
1321 static void cgroup_kn_unlock(struct kernfs_node *kn)
1323 struct cgroup *cgrp;
1325 if (kernfs_type(kn) == KERNFS_DIR)
1328 cgrp = kn->parent->priv;
1330 mutex_unlock(&cgroup_mutex);
1332 kernfs_unbreak_active_protection(kn);
1337 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1338 * @kn: the kernfs_node being serviced
1340 * This helper is to be used by a cgroup kernfs method currently servicing
1341 * @kn. It breaks the active protection, performs cgroup locking and
1342 * verifies that the associated cgroup is alive. Returns the cgroup if
1343 * alive; otherwise, %NULL. A successful return should be undone by a
1344 * matching cgroup_kn_unlock() invocation.
1346 * Any cgroup kernfs method implementation which requires locking the
1347 * associated cgroup should use this helper. It avoids nesting cgroup
1348 * locking under kernfs active protection and allows all kernfs operations
1349 * including self-removal.
1351 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1353 struct cgroup *cgrp;
1355 if (kernfs_type(kn) == KERNFS_DIR)
1358 cgrp = kn->parent->priv;
1361 * We're gonna grab cgroup_mutex which nests outside kernfs
1362 * active_ref. cgroup liveliness check alone provides enough
1363 * protection against removal. Ensure @cgrp stays accessible and
1364 * break the active_ref protection.
1366 if (!cgroup_tryget(cgrp))
1368 kernfs_break_active_protection(kn);
1370 mutex_lock(&cgroup_mutex);
1372 if (!cgroup_is_dead(cgrp))
1375 cgroup_kn_unlock(kn);
1379 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1381 char name[CGROUP_FILE_NAME_MAX];
1383 lockdep_assert_held(&cgroup_mutex);
1385 if (cft->file_offset) {
1386 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1387 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1389 spin_lock_irq(&cgroup_file_kn_lock);
1391 spin_unlock_irq(&cgroup_file_kn_lock);
1394 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1398 * css_clear_dir - remove subsys files in a cgroup directory
1400 * @cgrp_override: specify if target cgroup is different from css->cgroup
1402 static void css_clear_dir(struct cgroup_subsys_state *css,
1403 struct cgroup *cgrp_override)
1405 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1406 struct cftype *cfts;
1408 list_for_each_entry(cfts, &css->ss->cfts, node)
1409 cgroup_addrm_files(css, cgrp, cfts, false);
1413 * css_populate_dir - create subsys files in a cgroup directory
1415 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1417 * On failure, no file is added.
1419 static int css_populate_dir(struct cgroup_subsys_state *css,
1420 struct cgroup *cgrp_override)
1422 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1423 struct cftype *cfts, *failed_cfts;
1427 if (cgroup_on_dfl(cgrp))
1428 cfts = cgroup_dfl_base_files;
1430 cfts = cgroup_legacy_base_files;
1432 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1435 list_for_each_entry(cfts, &css->ss->cfts, node) {
1436 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1444 list_for_each_entry(cfts, &css->ss->cfts, node) {
1445 if (cfts == failed_cfts)
1447 cgroup_addrm_files(css, cgrp, cfts, false);
1452 static int rebind_subsystems(struct cgroup_root *dst_root,
1453 unsigned long ss_mask)
1455 struct cgroup *dcgrp = &dst_root->cgrp;
1456 struct cgroup_subsys *ss;
1457 unsigned long tmp_ss_mask;
1460 lockdep_assert_held(&cgroup_mutex);
1462 for_each_subsys_which(ss, ssid, &ss_mask) {
1463 /* if @ss has non-root csses attached to it, can't move */
1464 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1467 /* can't move between two non-dummy roots either */
1468 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1472 /* skip creating root files on dfl_root for inhibited subsystems */
1473 tmp_ss_mask = ss_mask;
1474 if (dst_root == &cgrp_dfl_root)
1475 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1477 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1478 struct cgroup *scgrp = &ss->root->cgrp;
1481 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1486 * Rebinding back to the default root is not allowed to
1487 * fail. Using both default and non-default roots should
1488 * be rare. Moving subsystems back and forth even more so.
1489 * Just warn about it and continue.
1491 if (dst_root == &cgrp_dfl_root) {
1492 if (cgrp_dfl_root_visible) {
1493 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1495 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1500 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1503 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1509 * Nothing can fail from this point on. Remove files for the
1510 * removed subsystems and rebind each subsystem.
1512 for_each_subsys_which(ss, ssid, &ss_mask) {
1513 struct cgroup_root *src_root = ss->root;
1514 struct cgroup *scgrp = &src_root->cgrp;
1515 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1516 struct css_set *cset;
1518 WARN_ON(!css || cgroup_css(dcgrp, ss));
1520 css_clear_dir(css, NULL);
1522 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1523 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1524 ss->root = dst_root;
1525 css->cgroup = dcgrp;
1527 spin_lock_bh(&css_set_lock);
1528 hash_for_each(css_set_table, i, cset, hlist)
1529 list_move_tail(&cset->e_cset_node[ss->id],
1530 &dcgrp->e_csets[ss->id]);
1531 spin_unlock_bh(&css_set_lock);
1533 src_root->subsys_mask &= ~(1 << ssid);
1534 scgrp->subtree_control &= ~(1 << ssid);
1535 cgroup_refresh_child_subsys_mask(scgrp);
1537 /* default hierarchy doesn't enable controllers by default */
1538 dst_root->subsys_mask |= 1 << ssid;
1539 if (dst_root == &cgrp_dfl_root) {
1540 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1542 dcgrp->subtree_control |= 1 << ssid;
1543 cgroup_refresh_child_subsys_mask(dcgrp);
1544 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1551 kernfs_activate(dcgrp->kn);
1555 static int cgroup_show_options(struct seq_file *seq,
1556 struct kernfs_root *kf_root)
1558 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1559 struct cgroup_subsys *ss;
1562 if (root != &cgrp_dfl_root)
1563 for_each_subsys(ss, ssid)
1564 if (root->subsys_mask & (1 << ssid))
1565 seq_show_option(seq, ss->legacy_name, NULL);
1566 if (root->flags & CGRP_ROOT_NOPREFIX)
1567 seq_puts(seq, ",noprefix");
1568 if (root->flags & CGRP_ROOT_XATTR)
1569 seq_puts(seq, ",xattr");
1571 spin_lock(&release_agent_path_lock);
1572 if (strlen(root->release_agent_path))
1573 seq_show_option(seq, "release_agent",
1574 root->release_agent_path);
1575 spin_unlock(&release_agent_path_lock);
1577 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1578 seq_puts(seq, ",clone_children");
1579 if (strlen(root->name))
1580 seq_show_option(seq, "name", root->name);
1584 struct cgroup_sb_opts {
1585 unsigned long subsys_mask;
1587 char *release_agent;
1588 bool cpuset_clone_children;
1590 /* User explicitly requested empty subsystem */
1594 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1596 char *token, *o = data;
1597 bool all_ss = false, one_ss = false;
1598 unsigned long mask = -1UL;
1599 struct cgroup_subsys *ss;
1603 #ifdef CONFIG_CPUSETS
1604 mask = ~(1U << cpuset_cgrp_id);
1607 memset(opts, 0, sizeof(*opts));
1609 while ((token = strsep(&o, ",")) != NULL) {
1614 if (!strcmp(token, "none")) {
1615 /* Explicitly have no subsystems */
1619 if (!strcmp(token, "all")) {
1620 /* Mutually exclusive option 'all' + subsystem name */
1626 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1627 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1630 if (!strcmp(token, "noprefix")) {
1631 opts->flags |= CGRP_ROOT_NOPREFIX;
1634 if (!strcmp(token, "clone_children")) {
1635 opts->cpuset_clone_children = true;
1638 if (!strcmp(token, "xattr")) {
1639 opts->flags |= CGRP_ROOT_XATTR;
1642 if (!strncmp(token, "release_agent=", 14)) {
1643 /* Specifying two release agents is forbidden */
1644 if (opts->release_agent)
1646 opts->release_agent =
1647 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1648 if (!opts->release_agent)
1652 if (!strncmp(token, "name=", 5)) {
1653 const char *name = token + 5;
1654 /* Can't specify an empty name */
1657 /* Must match [\w.-]+ */
1658 for (i = 0; i < strlen(name); i++) {
1662 if ((c == '.') || (c == '-') || (c == '_'))
1666 /* Specifying two names is forbidden */
1669 opts->name = kstrndup(name,
1670 MAX_CGROUP_ROOT_NAMELEN - 1,
1678 for_each_subsys(ss, i) {
1679 if (strcmp(token, ss->legacy_name))
1681 if (!cgroup_ssid_enabled(i))
1684 /* Mutually exclusive option 'all' + subsystem name */
1687 opts->subsys_mask |= (1 << i);
1692 if (i == CGROUP_SUBSYS_COUNT)
1696 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1697 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1699 pr_err("sane_behavior: no other mount options allowed\n");
1706 * If the 'all' option was specified select all the subsystems,
1707 * otherwise if 'none', 'name=' and a subsystem name options were
1708 * not specified, let's default to 'all'
1710 if (all_ss || (!one_ss && !opts->none && !opts->name))
1711 for_each_subsys(ss, i)
1712 if (cgroup_ssid_enabled(i))
1713 opts->subsys_mask |= (1 << i);
1716 * We either have to specify by name or by subsystems. (So all
1717 * empty hierarchies must have a name).
1719 if (!opts->subsys_mask && !opts->name)
1723 * Option noprefix was introduced just for backward compatibility
1724 * with the old cpuset, so we allow noprefix only if mounting just
1725 * the cpuset subsystem.
1727 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1730 /* Can't specify "none" and some subsystems */
1731 if (opts->subsys_mask && opts->none)
1737 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1740 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1741 struct cgroup_sb_opts opts;
1742 unsigned long added_mask, removed_mask;
1744 if (root == &cgrp_dfl_root) {
1745 pr_err("remount is not allowed\n");
1749 mutex_lock(&cgroup_mutex);
1751 /* See what subsystems are wanted */
1752 ret = parse_cgroupfs_options(data, &opts);
1756 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1757 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1758 task_tgid_nr(current), current->comm);
1760 added_mask = opts.subsys_mask & ~root->subsys_mask;
1761 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1763 /* Don't allow flags or name to change at remount */
1764 if ((opts.flags ^ root->flags) ||
1765 (opts.name && strcmp(opts.name, root->name))) {
1766 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1767 opts.flags, opts.name ?: "", root->flags, root->name);
1772 /* remounting is not allowed for populated hierarchies */
1773 if (!list_empty(&root->cgrp.self.children)) {
1778 ret = rebind_subsystems(root, added_mask);
1782 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1784 if (opts.release_agent) {
1785 spin_lock(&release_agent_path_lock);
1786 strcpy(root->release_agent_path, opts.release_agent);
1787 spin_unlock(&release_agent_path_lock);
1790 kfree(opts.release_agent);
1792 mutex_unlock(&cgroup_mutex);
1797 * To reduce the fork() overhead for systems that are not actually using
1798 * their cgroups capability, we don't maintain the lists running through
1799 * each css_set to its tasks until we see the list actually used - in other
1800 * words after the first mount.
1802 static bool use_task_css_set_links __read_mostly;
1804 static void cgroup_enable_task_cg_lists(void)
1806 struct task_struct *p, *g;
1808 spin_lock_bh(&css_set_lock);
1810 if (use_task_css_set_links)
1813 use_task_css_set_links = true;
1816 * We need tasklist_lock because RCU is not safe against
1817 * while_each_thread(). Besides, a forking task that has passed
1818 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1819 * is not guaranteed to have its child immediately visible in the
1820 * tasklist if we walk through it with RCU.
1822 read_lock(&tasklist_lock);
1823 do_each_thread(g, p) {
1824 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1825 task_css_set(p) != &init_css_set);
1828 * We should check if the process is exiting, otherwise
1829 * it will race with cgroup_exit() in that the list
1830 * entry won't be deleted though the process has exited.
1831 * Do it while holding siglock so that we don't end up
1832 * racing against cgroup_exit().
1834 spin_lock_irq(&p->sighand->siglock);
1835 if (!(p->flags & PF_EXITING)) {
1836 struct css_set *cset = task_css_set(p);
1838 if (!css_set_populated(cset))
1839 css_set_update_populated(cset, true);
1840 list_add_tail(&p->cg_list, &cset->tasks);
1843 spin_unlock_irq(&p->sighand->siglock);
1844 } while_each_thread(g, p);
1845 read_unlock(&tasklist_lock);
1847 spin_unlock_bh(&css_set_lock);
1850 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1852 struct cgroup_subsys *ss;
1855 INIT_LIST_HEAD(&cgrp->self.sibling);
1856 INIT_LIST_HEAD(&cgrp->self.children);
1857 INIT_LIST_HEAD(&cgrp->cset_links);
1858 INIT_LIST_HEAD(&cgrp->pidlists);
1859 mutex_init(&cgrp->pidlist_mutex);
1860 cgrp->self.cgroup = cgrp;
1861 cgrp->self.flags |= CSS_ONLINE;
1863 for_each_subsys(ss, ssid)
1864 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1866 init_waitqueue_head(&cgrp->offline_waitq);
1867 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1870 static void init_cgroup_root(struct cgroup_root *root,
1871 struct cgroup_sb_opts *opts)
1873 struct cgroup *cgrp = &root->cgrp;
1875 INIT_LIST_HEAD(&root->root_list);
1876 atomic_set(&root->nr_cgrps, 1);
1878 init_cgroup_housekeeping(cgrp);
1879 idr_init(&root->cgroup_idr);
1881 root->flags = opts->flags;
1882 if (opts->release_agent)
1883 strcpy(root->release_agent_path, opts->release_agent);
1885 strcpy(root->name, opts->name);
1886 if (opts->cpuset_clone_children)
1887 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1890 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1892 LIST_HEAD(tmp_links);
1893 struct cgroup *root_cgrp = &root->cgrp;
1894 struct css_set *cset;
1897 lockdep_assert_held(&cgroup_mutex);
1899 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1902 root_cgrp->id = ret;
1903 root_cgrp->ancestor_ids[0] = ret;
1905 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1911 * We're accessing css_set_count without locking css_set_lock here,
1912 * but that's OK - it can only be increased by someone holding
1913 * cgroup_lock, and that's us. The worst that can happen is that we
1914 * have some link structures left over
1916 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1920 ret = cgroup_init_root_id(root);
1924 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1925 KERNFS_ROOT_CREATE_DEACTIVATED,
1927 if (IS_ERR(root->kf_root)) {
1928 ret = PTR_ERR(root->kf_root);
1931 root_cgrp->kn = root->kf_root->kn;
1933 ret = css_populate_dir(&root_cgrp->self, NULL);
1937 ret = rebind_subsystems(root, ss_mask);
1942 * There must be no failure case after here, since rebinding takes
1943 * care of subsystems' refcounts, which are explicitly dropped in
1944 * the failure exit path.
1946 list_add(&root->root_list, &cgroup_roots);
1947 cgroup_root_count++;
1950 * Link the root cgroup in this hierarchy into all the css_set
1953 spin_lock_bh(&css_set_lock);
1954 hash_for_each(css_set_table, i, cset, hlist) {
1955 link_css_set(&tmp_links, cset, root_cgrp);
1956 if (css_set_populated(cset))
1957 cgroup_update_populated(root_cgrp, true);
1959 spin_unlock_bh(&css_set_lock);
1961 BUG_ON(!list_empty(&root_cgrp->self.children));
1962 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1964 kernfs_activate(root_cgrp->kn);
1969 kernfs_destroy_root(root->kf_root);
1970 root->kf_root = NULL;
1972 cgroup_exit_root_id(root);
1974 percpu_ref_exit(&root_cgrp->self.refcnt);
1976 free_cgrp_cset_links(&tmp_links);
1980 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1981 int flags, const char *unused_dev_name,
1984 struct super_block *pinned_sb = NULL;
1985 struct cgroup_subsys *ss;
1986 struct cgroup_root *root;
1987 struct cgroup_sb_opts opts;
1988 struct dentry *dentry;
1994 * The first time anyone tries to mount a cgroup, enable the list
1995 * linking each css_set to its tasks and fix up all existing tasks.
1997 if (!use_task_css_set_links)
1998 cgroup_enable_task_cg_lists();
2000 mutex_lock(&cgroup_mutex);
2002 /* First find the desired set of subsystems */
2003 ret = parse_cgroupfs_options(data, &opts);
2007 /* look for a matching existing root */
2008 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
2009 cgrp_dfl_root_visible = true;
2010 root = &cgrp_dfl_root;
2011 cgroup_get(&root->cgrp);
2017 * Destruction of cgroup root is asynchronous, so subsystems may
2018 * still be dying after the previous unmount. Let's drain the
2019 * dying subsystems. We just need to ensure that the ones
2020 * unmounted previously finish dying and don't care about new ones
2021 * starting. Testing ref liveliness is good enough.
2023 for_each_subsys(ss, i) {
2024 if (!(opts.subsys_mask & (1 << i)) ||
2025 ss->root == &cgrp_dfl_root)
2028 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2029 mutex_unlock(&cgroup_mutex);
2031 ret = restart_syscall();
2034 cgroup_put(&ss->root->cgrp);
2037 for_each_root(root) {
2038 bool name_match = false;
2040 if (root == &cgrp_dfl_root)
2044 * If we asked for a name then it must match. Also, if
2045 * name matches but sybsys_mask doesn't, we should fail.
2046 * Remember whether name matched.
2049 if (strcmp(opts.name, root->name))
2055 * If we asked for subsystems (or explicitly for no
2056 * subsystems) then they must match.
2058 if ((opts.subsys_mask || opts.none) &&
2059 (opts.subsys_mask != root->subsys_mask)) {
2066 if (root->flags ^ opts.flags)
2067 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2070 * We want to reuse @root whose lifetime is governed by its
2071 * ->cgrp. Let's check whether @root is alive and keep it
2072 * that way. As cgroup_kill_sb() can happen anytime, we
2073 * want to block it by pinning the sb so that @root doesn't
2074 * get killed before mount is complete.
2076 * With the sb pinned, tryget_live can reliably indicate
2077 * whether @root can be reused. If it's being killed,
2078 * drain it. We can use wait_queue for the wait but this
2079 * path is super cold. Let's just sleep a bit and retry.
2081 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2082 if (IS_ERR(pinned_sb) ||
2083 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2084 mutex_unlock(&cgroup_mutex);
2085 if (!IS_ERR_OR_NULL(pinned_sb))
2086 deactivate_super(pinned_sb);
2088 ret = restart_syscall();
2097 * No such thing, create a new one. name= matching without subsys
2098 * specification is allowed for already existing hierarchies but we
2099 * can't create new one without subsys specification.
2101 if (!opts.subsys_mask && !opts.none) {
2106 root = kzalloc(sizeof(*root), GFP_KERNEL);
2112 init_cgroup_root(root, &opts);
2114 ret = cgroup_setup_root(root, opts.subsys_mask);
2116 cgroup_free_root(root);
2119 mutex_unlock(&cgroup_mutex);
2121 kfree(opts.release_agent);
2125 return ERR_PTR(ret);
2127 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2128 CGROUP_SUPER_MAGIC, &new_sb);
2129 if (IS_ERR(dentry) || !new_sb)
2130 cgroup_put(&root->cgrp);
2133 * If @pinned_sb, we're reusing an existing root and holding an
2134 * extra ref on its sb. Mount is complete. Put the extra ref.
2138 deactivate_super(pinned_sb);
2144 static void cgroup_kill_sb(struct super_block *sb)
2146 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2147 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2150 * If @root doesn't have any mounts or children, start killing it.
2151 * This prevents new mounts by disabling percpu_ref_tryget_live().
2152 * cgroup_mount() may wait for @root's release.
2154 * And don't kill the default root.
2156 if (!list_empty(&root->cgrp.self.children) ||
2157 root == &cgrp_dfl_root)
2158 cgroup_put(&root->cgrp);
2160 percpu_ref_kill(&root->cgrp.self.refcnt);
2165 static struct file_system_type cgroup_fs_type = {
2167 .mount = cgroup_mount,
2168 .kill_sb = cgroup_kill_sb,
2172 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2173 * @task: target task
2174 * @buf: the buffer to write the path into
2175 * @buflen: the length of the buffer
2177 * Determine @task's cgroup on the first (the one with the lowest non-zero
2178 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2179 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2180 * cgroup controller callbacks.
2182 * Return value is the same as kernfs_path().
2184 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2186 struct cgroup_root *root;
2187 struct cgroup *cgrp;
2188 int hierarchy_id = 1;
2191 mutex_lock(&cgroup_mutex);
2192 spin_lock_bh(&css_set_lock);
2194 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2197 cgrp = task_cgroup_from_root(task, root);
2198 path = cgroup_path(cgrp, buf, buflen);
2200 /* if no hierarchy exists, everyone is in "/" */
2201 if (strlcpy(buf, "/", buflen) < buflen)
2205 spin_unlock_bh(&css_set_lock);
2206 mutex_unlock(&cgroup_mutex);
2209 EXPORT_SYMBOL_GPL(task_cgroup_path);
2211 /* used to track tasks and other necessary states during migration */
2212 struct cgroup_taskset {
2213 /* the src and dst cset list running through cset->mg_node */
2214 struct list_head src_csets;
2215 struct list_head dst_csets;
2217 /* the subsys currently being processed */
2221 * Fields for cgroup_taskset_*() iteration.
2223 * Before migration is committed, the target migration tasks are on
2224 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2225 * the csets on ->dst_csets. ->csets point to either ->src_csets
2226 * or ->dst_csets depending on whether migration is committed.
2228 * ->cur_csets and ->cur_task point to the current task position
2231 struct list_head *csets;
2232 struct css_set *cur_cset;
2233 struct task_struct *cur_task;
2236 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2237 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2238 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2239 .csets = &tset.src_csets, \
2243 * cgroup_taskset_add - try to add a migration target task to a taskset
2244 * @task: target task
2245 * @tset: target taskset
2247 * Add @task, which is a migration target, to @tset. This function becomes
2248 * noop if @task doesn't need to be migrated. @task's css_set should have
2249 * been added as a migration source and @task->cg_list will be moved from
2250 * the css_set's tasks list to mg_tasks one.
2252 static void cgroup_taskset_add(struct task_struct *task,
2253 struct cgroup_taskset *tset)
2255 struct css_set *cset;
2257 lockdep_assert_held(&css_set_lock);
2259 /* @task either already exited or can't exit until the end */
2260 if (task->flags & PF_EXITING)
2263 /* leave @task alone if post_fork() hasn't linked it yet */
2264 if (list_empty(&task->cg_list))
2267 cset = task_css_set(task);
2268 if (!cset->mg_src_cgrp)
2271 list_move_tail(&task->cg_list, &cset->mg_tasks);
2272 if (list_empty(&cset->mg_node))
2273 list_add_tail(&cset->mg_node, &tset->src_csets);
2274 if (list_empty(&cset->mg_dst_cset->mg_node))
2275 list_move_tail(&cset->mg_dst_cset->mg_node,
2280 * cgroup_taskset_first - reset taskset and return the first task
2281 * @tset: taskset of interest
2282 * @dst_cssp: output variable for the destination css
2284 * @tset iteration is initialized and the first task is returned.
2286 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2287 struct cgroup_subsys_state **dst_cssp)
2289 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2290 tset->cur_task = NULL;
2292 return cgroup_taskset_next(tset, dst_cssp);
2296 * cgroup_taskset_next - iterate to the next task in taskset
2297 * @tset: taskset of interest
2298 * @dst_cssp: output variable for the destination css
2300 * Return the next task in @tset. Iteration must have been initialized
2301 * with cgroup_taskset_first().
2303 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2304 struct cgroup_subsys_state **dst_cssp)
2306 struct css_set *cset = tset->cur_cset;
2307 struct task_struct *task = tset->cur_task;
2309 while (&cset->mg_node != tset->csets) {
2311 task = list_first_entry(&cset->mg_tasks,
2312 struct task_struct, cg_list);
2314 task = list_next_entry(task, cg_list);
2316 if (&task->cg_list != &cset->mg_tasks) {
2317 tset->cur_cset = cset;
2318 tset->cur_task = task;
2321 * This function may be called both before and
2322 * after cgroup_taskset_migrate(). The two cases
2323 * can be distinguished by looking at whether @cset
2324 * has its ->mg_dst_cset set.
2326 if (cset->mg_dst_cset)
2327 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2329 *dst_cssp = cset->subsys[tset->ssid];
2334 cset = list_next_entry(cset, mg_node);
2342 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2343 * @tset: taget taskset
2344 * @dst_cgrp: destination cgroup
2346 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2347 * ->can_attach callbacks fails and guarantees that either all or none of
2348 * the tasks in @tset are migrated. @tset is consumed regardless of
2351 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2352 struct cgroup *dst_cgrp)
2354 struct cgroup_subsys_state *css, *failed_css = NULL;
2355 struct task_struct *task, *tmp_task;
2356 struct css_set *cset, *tmp_cset;
2359 /* methods shouldn't be called if no task is actually migrating */
2360 if (list_empty(&tset->src_csets))
2363 /* check that we can legitimately attach to the cgroup */
2364 for_each_e_css(css, i, dst_cgrp) {
2365 if (css->ss->can_attach) {
2367 ret = css->ss->can_attach(tset);
2370 goto out_cancel_attach;
2376 * Now that we're guaranteed success, proceed to move all tasks to
2377 * the new cgroup. There are no failure cases after here, so this
2378 * is the commit point.
2380 spin_lock_bh(&css_set_lock);
2381 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2382 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2383 struct css_set *from_cset = task_css_set(task);
2384 struct css_set *to_cset = cset->mg_dst_cset;
2386 get_css_set(to_cset);
2387 css_set_move_task(task, from_cset, to_cset, true);
2388 put_css_set_locked(from_cset);
2391 spin_unlock_bh(&css_set_lock);
2394 * Migration is committed, all target tasks are now on dst_csets.
2395 * Nothing is sensitive to fork() after this point. Notify
2396 * controllers that migration is complete.
2398 tset->csets = &tset->dst_csets;
2400 for_each_e_css(css, i, dst_cgrp) {
2401 if (css->ss->attach) {
2403 css->ss->attach(tset);
2408 goto out_release_tset;
2411 for_each_e_css(css, i, dst_cgrp) {
2412 if (css == failed_css)
2414 if (css->ss->cancel_attach) {
2416 css->ss->cancel_attach(tset);
2420 spin_lock_bh(&css_set_lock);
2421 list_splice_init(&tset->dst_csets, &tset->src_csets);
2422 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2423 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2424 list_del_init(&cset->mg_node);
2426 spin_unlock_bh(&css_set_lock);
2431 * cgroup_migrate_finish - cleanup after attach
2432 * @preloaded_csets: list of preloaded css_sets
2434 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2435 * those functions for details.
2437 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2439 struct css_set *cset, *tmp_cset;
2441 lockdep_assert_held(&cgroup_mutex);
2443 spin_lock_bh(&css_set_lock);
2444 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2445 cset->mg_src_cgrp = NULL;
2446 cset->mg_dst_cset = NULL;
2447 list_del_init(&cset->mg_preload_node);
2448 put_css_set_locked(cset);
2450 spin_unlock_bh(&css_set_lock);
2454 * cgroup_migrate_add_src - add a migration source css_set
2455 * @src_cset: the source css_set to add
2456 * @dst_cgrp: the destination cgroup
2457 * @preloaded_csets: list of preloaded css_sets
2459 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2460 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2461 * up by cgroup_migrate_finish().
2463 * This function may be called without holding cgroup_threadgroup_rwsem
2464 * even if the target is a process. Threads may be created and destroyed
2465 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2466 * into play and the preloaded css_sets are guaranteed to cover all
2469 static void cgroup_migrate_add_src(struct css_set *src_cset,
2470 struct cgroup *dst_cgrp,
2471 struct list_head *preloaded_csets)
2473 struct cgroup *src_cgrp;
2475 lockdep_assert_held(&cgroup_mutex);
2476 lockdep_assert_held(&css_set_lock);
2478 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2480 if (!list_empty(&src_cset->mg_preload_node))
2483 WARN_ON(src_cset->mg_src_cgrp);
2484 WARN_ON(!list_empty(&src_cset->mg_tasks));
2485 WARN_ON(!list_empty(&src_cset->mg_node));
2487 src_cset->mg_src_cgrp = src_cgrp;
2488 get_css_set(src_cset);
2489 list_add(&src_cset->mg_preload_node, preloaded_csets);
2493 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2494 * @dst_cgrp: the destination cgroup (may be %NULL)
2495 * @preloaded_csets: list of preloaded source css_sets
2497 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2498 * have been preloaded to @preloaded_csets. This function looks up and
2499 * pins all destination css_sets, links each to its source, and append them
2500 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2501 * source css_set is assumed to be its cgroup on the default hierarchy.
2503 * This function must be called after cgroup_migrate_add_src() has been
2504 * called on each migration source css_set. After migration is performed
2505 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2508 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2509 struct list_head *preloaded_csets)
2512 struct css_set *src_cset, *tmp_cset;
2514 lockdep_assert_held(&cgroup_mutex);
2517 * Except for the root, child_subsys_mask must be zero for a cgroup
2518 * with tasks so that child cgroups don't compete against tasks.
2520 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2521 dst_cgrp->child_subsys_mask)
2524 /* look up the dst cset for each src cset and link it to src */
2525 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2526 struct css_set *dst_cset;
2528 dst_cset = find_css_set(src_cset,
2529 dst_cgrp ?: src_cset->dfl_cgrp);
2533 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2536 * If src cset equals dst, it's noop. Drop the src.
2537 * cgroup_migrate() will skip the cset too. Note that we
2538 * can't handle src == dst as some nodes are used by both.
2540 if (src_cset == dst_cset) {
2541 src_cset->mg_src_cgrp = NULL;
2542 list_del_init(&src_cset->mg_preload_node);
2543 put_css_set(src_cset);
2544 put_css_set(dst_cset);
2548 src_cset->mg_dst_cset = dst_cset;
2550 if (list_empty(&dst_cset->mg_preload_node))
2551 list_add(&dst_cset->mg_preload_node, &csets);
2553 put_css_set(dst_cset);
2556 list_splice_tail(&csets, preloaded_csets);
2559 cgroup_migrate_finish(&csets);
2564 * cgroup_migrate - migrate a process or task to a cgroup
2565 * @leader: the leader of the process or the task to migrate
2566 * @threadgroup: whether @leader points to the whole process or a single task
2567 * @cgrp: the destination cgroup
2569 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2570 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2571 * caller is also responsible for invoking cgroup_migrate_add_src() and
2572 * cgroup_migrate_prepare_dst() on the targets before invoking this
2573 * function and following up with cgroup_migrate_finish().
2575 * As long as a controller's ->can_attach() doesn't fail, this function is
2576 * guaranteed to succeed. This means that, excluding ->can_attach()
2577 * failure, when migrating multiple targets, the success or failure can be
2578 * decided for all targets by invoking group_migrate_prepare_dst() before
2579 * actually starting migrating.
2581 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2582 struct cgroup *cgrp)
2584 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2585 struct task_struct *task;
2588 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2589 * already PF_EXITING could be freed from underneath us unless we
2590 * take an rcu_read_lock.
2592 spin_lock_bh(&css_set_lock);
2596 cgroup_taskset_add(task, &tset);
2599 } while_each_thread(leader, task);
2601 spin_unlock_bh(&css_set_lock);
2603 return cgroup_taskset_migrate(&tset, cgrp);
2607 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2608 * @dst_cgrp: the cgroup to attach to
2609 * @leader: the task or the leader of the threadgroup to be attached
2610 * @threadgroup: attach the whole threadgroup?
2612 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2614 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2615 struct task_struct *leader, bool threadgroup)
2617 LIST_HEAD(preloaded_csets);
2618 struct task_struct *task;
2621 /* look up all src csets */
2622 spin_lock_bh(&css_set_lock);
2626 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2630 } while_each_thread(leader, task);
2632 spin_unlock_bh(&css_set_lock);
2634 /* prepare dst csets and commit */
2635 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2637 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2639 cgroup_migrate_finish(&preloaded_csets);
2643 static int cgroup_procs_write_permission(struct task_struct *task,
2644 struct cgroup *dst_cgrp,
2645 struct kernfs_open_file *of)
2647 const struct cred *cred = current_cred();
2648 const struct cred *tcred = get_task_cred(task);
2652 * even if we're attaching all tasks in the thread group, we only
2653 * need to check permissions on one of them.
2655 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2656 !uid_eq(cred->euid, tcred->uid) &&
2657 !uid_eq(cred->euid, tcred->suid))
2660 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2661 struct super_block *sb = of->file->f_path.dentry->d_sb;
2662 struct cgroup *cgrp;
2663 struct inode *inode;
2665 spin_lock_bh(&css_set_lock);
2666 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2667 spin_unlock_bh(&css_set_lock);
2669 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2670 cgrp = cgroup_parent(cgrp);
2673 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2675 ret = inode_permission(inode, MAY_WRITE);
2685 * Find the task_struct of the task to attach by vpid and pass it along to the
2686 * function to attach either it or all tasks in its threadgroup. Will lock
2687 * cgroup_mutex and threadgroup.
2689 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2690 size_t nbytes, loff_t off, bool threadgroup)
2692 struct task_struct *tsk;
2693 struct cgroup *cgrp;
2697 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2700 cgrp = cgroup_kn_lock_live(of->kn);
2704 percpu_down_write(&cgroup_threadgroup_rwsem);
2707 tsk = find_task_by_vpid(pid);
2710 goto out_unlock_rcu;
2717 tsk = tsk->group_leader;
2720 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2721 * trapped in a cpuset, or RT worker may be born in a cgroup
2722 * with no rt_runtime allocated. Just say no.
2724 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2726 goto out_unlock_rcu;
2729 get_task_struct(tsk);
2732 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2734 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2736 put_task_struct(tsk);
2737 goto out_unlock_threadgroup;
2741 out_unlock_threadgroup:
2742 percpu_up_write(&cgroup_threadgroup_rwsem);
2743 cgroup_kn_unlock(of->kn);
2744 return ret ?: nbytes;
2748 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2749 * @from: attach to all cgroups of a given task
2750 * @tsk: the task to be attached
2752 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2754 struct cgroup_root *root;
2757 mutex_lock(&cgroup_mutex);
2758 for_each_root(root) {
2759 struct cgroup *from_cgrp;
2761 if (root == &cgrp_dfl_root)
2764 spin_lock_bh(&css_set_lock);
2765 from_cgrp = task_cgroup_from_root(from, root);
2766 spin_unlock_bh(&css_set_lock);
2768 retval = cgroup_attach_task(from_cgrp, tsk, false);
2772 mutex_unlock(&cgroup_mutex);
2776 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2778 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2779 char *buf, size_t nbytes, loff_t off)
2781 return __cgroup_procs_write(of, buf, nbytes, off, false);
2784 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2785 char *buf, size_t nbytes, loff_t off)
2787 return __cgroup_procs_write(of, buf, nbytes, off, true);
2790 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2791 char *buf, size_t nbytes, loff_t off)
2793 struct cgroup *cgrp;
2795 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2797 cgrp = cgroup_kn_lock_live(of->kn);
2800 spin_lock(&release_agent_path_lock);
2801 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2802 sizeof(cgrp->root->release_agent_path));
2803 spin_unlock(&release_agent_path_lock);
2804 cgroup_kn_unlock(of->kn);
2808 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2810 struct cgroup *cgrp = seq_css(seq)->cgroup;
2812 spin_lock(&release_agent_path_lock);
2813 seq_puts(seq, cgrp->root->release_agent_path);
2814 spin_unlock(&release_agent_path_lock);
2815 seq_putc(seq, '\n');
2819 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2821 seq_puts(seq, "0\n");
2825 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2827 struct cgroup_subsys *ss;
2828 bool printed = false;
2831 for_each_subsys_which(ss, ssid, &ss_mask) {
2834 seq_printf(seq, "%s", ss->name);
2838 seq_putc(seq, '\n');
2841 /* show controllers which are currently attached to the default hierarchy */
2842 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2844 struct cgroup *cgrp = seq_css(seq)->cgroup;
2846 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2847 ~cgrp_dfl_root_inhibit_ss_mask);
2851 /* show controllers which are enabled from the parent */
2852 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2854 struct cgroup *cgrp = seq_css(seq)->cgroup;
2856 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2860 /* show controllers which are enabled for a given cgroup's children */
2861 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2863 struct cgroup *cgrp = seq_css(seq)->cgroup;
2865 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2870 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2871 * @cgrp: root of the subtree to update csses for
2873 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2874 * css associations need to be updated accordingly. This function looks up
2875 * all css_sets which are attached to the subtree, creates the matching
2876 * updated css_sets and migrates the tasks to the new ones.
2878 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2880 LIST_HEAD(preloaded_csets);
2881 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2882 struct cgroup_subsys_state *css;
2883 struct css_set *src_cset;
2886 lockdep_assert_held(&cgroup_mutex);
2888 percpu_down_write(&cgroup_threadgroup_rwsem);
2890 /* look up all csses currently attached to @cgrp's subtree */
2891 spin_lock_bh(&css_set_lock);
2892 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2893 struct cgrp_cset_link *link;
2895 /* self is not affected by child_subsys_mask change */
2896 if (css->cgroup == cgrp)
2899 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2900 cgroup_migrate_add_src(link->cset, cgrp,
2903 spin_unlock_bh(&css_set_lock);
2905 /* NULL dst indicates self on default hierarchy */
2906 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2910 spin_lock_bh(&css_set_lock);
2911 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2912 struct task_struct *task, *ntask;
2914 /* src_csets precede dst_csets, break on the first dst_cset */
2915 if (!src_cset->mg_src_cgrp)
2918 /* all tasks in src_csets need to be migrated */
2919 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2920 cgroup_taskset_add(task, &tset);
2922 spin_unlock_bh(&css_set_lock);
2924 ret = cgroup_taskset_migrate(&tset, cgrp);
2926 cgroup_migrate_finish(&preloaded_csets);
2927 percpu_up_write(&cgroup_threadgroup_rwsem);
2931 /* change the enabled child controllers for a cgroup in the default hierarchy */
2932 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2933 char *buf, size_t nbytes,
2936 unsigned long enable = 0, disable = 0;
2937 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2938 struct cgroup *cgrp, *child;
2939 struct cgroup_subsys *ss;
2944 * Parse input - space separated list of subsystem names prefixed
2945 * with either + or -.
2947 buf = strstrip(buf);
2948 while ((tok = strsep(&buf, " "))) {
2949 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2953 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2954 if (!cgroup_ssid_enabled(ssid) ||
2955 strcmp(tok + 1, ss->name))
2959 enable |= 1 << ssid;
2960 disable &= ~(1 << ssid);
2961 } else if (*tok == '-') {
2962 disable |= 1 << ssid;
2963 enable &= ~(1 << ssid);
2969 if (ssid == CGROUP_SUBSYS_COUNT)
2973 cgrp = cgroup_kn_lock_live(of->kn);
2977 for_each_subsys(ss, ssid) {
2978 if (enable & (1 << ssid)) {
2979 if (cgrp->subtree_control & (1 << ssid)) {
2980 enable &= ~(1 << ssid);
2984 /* unavailable or not enabled on the parent? */
2985 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2986 (cgroup_parent(cgrp) &&
2987 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2991 } else if (disable & (1 << ssid)) {
2992 if (!(cgrp->subtree_control & (1 << ssid))) {
2993 disable &= ~(1 << ssid);
2997 /* a child has it enabled? */
2998 cgroup_for_each_live_child(child, cgrp) {
2999 if (child->subtree_control & (1 << ssid)) {
3007 if (!enable && !disable) {
3013 * Except for the root, subtree_control must be zero for a cgroup
3014 * with tasks so that child cgroups don't compete against tasks.
3016 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3022 * Update subsys masks and calculate what needs to be done. More
3023 * subsystems than specified may need to be enabled or disabled
3024 * depending on subsystem dependencies.
3026 old_sc = cgrp->subtree_control;
3027 old_ss = cgrp->child_subsys_mask;
3028 new_sc = (old_sc | enable) & ~disable;
3029 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
3031 css_enable = ~old_ss & new_ss;
3032 css_disable = old_ss & ~new_ss;
3033 enable |= css_enable;
3034 disable |= css_disable;
3037 * Because css offlining is asynchronous, userland might try to
3038 * re-enable the same controller while the previous instance is
3039 * still around. In such cases, wait till it's gone using
3042 for_each_subsys_which(ss, ssid, &css_enable) {
3043 cgroup_for_each_live_child(child, cgrp) {
3046 if (!cgroup_css(child, ss))
3050 prepare_to_wait(&child->offline_waitq, &wait,
3051 TASK_UNINTERRUPTIBLE);
3052 cgroup_kn_unlock(of->kn);
3054 finish_wait(&child->offline_waitq, &wait);
3057 return restart_syscall();
3061 cgrp->subtree_control = new_sc;
3062 cgrp->child_subsys_mask = new_ss;
3065 * Create new csses or make the existing ones visible. A css is
3066 * created invisible if it's being implicitly enabled through
3067 * dependency. An invisible css is made visible when the userland
3068 * explicitly enables it.
3070 for_each_subsys(ss, ssid) {
3071 if (!(enable & (1 << ssid)))
3074 cgroup_for_each_live_child(child, cgrp) {
3075 if (css_enable & (1 << ssid))
3076 ret = create_css(child, ss,
3077 cgrp->subtree_control & (1 << ssid));
3079 ret = css_populate_dir(cgroup_css(child, ss),
3087 * At this point, cgroup_e_css() results reflect the new csses
3088 * making the following cgroup_update_dfl_csses() properly update
3089 * css associations of all tasks in the subtree.
3091 ret = cgroup_update_dfl_csses(cgrp);
3096 * All tasks are migrated out of disabled csses. Kill or hide
3097 * them. A css is hidden when the userland requests it to be
3098 * disabled while other subsystems are still depending on it. The
3099 * css must not actively control resources and be in the vanilla
3100 * state if it's made visible again later. Controllers which may
3101 * be depended upon should provide ->css_reset() for this purpose.
3103 for_each_subsys(ss, ssid) {
3104 if (!(disable & (1 << ssid)))
3107 cgroup_for_each_live_child(child, cgrp) {
3108 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3110 if (css_disable & (1 << ssid)) {
3113 css_clear_dir(css, NULL);
3121 * The effective csses of all the descendants (excluding @cgrp) may
3122 * have changed. Subsystems can optionally subscribe to this event
3123 * by implementing ->css_e_css_changed() which is invoked if any of
3124 * the effective csses seen from the css's cgroup may have changed.
3126 for_each_subsys(ss, ssid) {
3127 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3128 struct cgroup_subsys_state *css;
3130 if (!ss->css_e_css_changed || !this_css)
3133 css_for_each_descendant_pre(css, this_css)
3134 if (css != this_css)
3135 ss->css_e_css_changed(css);
3138 kernfs_activate(cgrp->kn);
3141 cgroup_kn_unlock(of->kn);
3142 return ret ?: nbytes;
3145 cgrp->subtree_control = old_sc;
3146 cgrp->child_subsys_mask = old_ss;
3148 for_each_subsys(ss, ssid) {
3149 if (!(enable & (1 << ssid)))
3152 cgroup_for_each_live_child(child, cgrp) {
3153 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3158 if (css_enable & (1 << ssid))
3161 css_clear_dir(css, NULL);
3167 static int cgroup_events_show(struct seq_file *seq, void *v)
3169 seq_printf(seq, "populated %d\n",
3170 cgroup_is_populated(seq_css(seq)->cgroup));
3174 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3175 size_t nbytes, loff_t off)
3177 struct cgroup *cgrp = of->kn->parent->priv;
3178 struct cftype *cft = of->kn->priv;
3179 struct cgroup_subsys_state *css;
3183 return cft->write(of, buf, nbytes, off);
3186 * kernfs guarantees that a file isn't deleted with operations in
3187 * flight, which means that the matching css is and stays alive and
3188 * doesn't need to be pinned. The RCU locking is not necessary
3189 * either. It's just for the convenience of using cgroup_css().
3192 css = cgroup_css(cgrp, cft->ss);
3195 if (cft->write_u64) {
3196 unsigned long long v;
3197 ret = kstrtoull(buf, 0, &v);
3199 ret = cft->write_u64(css, cft, v);
3200 } else if (cft->write_s64) {
3202 ret = kstrtoll(buf, 0, &v);
3204 ret = cft->write_s64(css, cft, v);
3209 return ret ?: nbytes;
3212 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3214 return seq_cft(seq)->seq_start(seq, ppos);
3217 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3219 return seq_cft(seq)->seq_next(seq, v, ppos);
3222 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3224 seq_cft(seq)->seq_stop(seq, v);
3227 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3229 struct cftype *cft = seq_cft(m);
3230 struct cgroup_subsys_state *css = seq_css(m);
3233 return cft->seq_show(m, arg);
3236 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3237 else if (cft->read_s64)
3238 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3244 static struct kernfs_ops cgroup_kf_single_ops = {
3245 .atomic_write_len = PAGE_SIZE,
3246 .write = cgroup_file_write,
3247 .seq_show = cgroup_seqfile_show,
3250 static struct kernfs_ops cgroup_kf_ops = {
3251 .atomic_write_len = PAGE_SIZE,
3252 .write = cgroup_file_write,
3253 .seq_start = cgroup_seqfile_start,
3254 .seq_next = cgroup_seqfile_next,
3255 .seq_stop = cgroup_seqfile_stop,
3256 .seq_show = cgroup_seqfile_show,
3260 * cgroup_rename - Only allow simple rename of directories in place.
3262 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3263 const char *new_name_str)
3265 struct cgroup *cgrp = kn->priv;
3268 if (kernfs_type(kn) != KERNFS_DIR)
3270 if (kn->parent != new_parent)
3274 * This isn't a proper migration and its usefulness is very
3275 * limited. Disallow on the default hierarchy.
3277 if (cgroup_on_dfl(cgrp))
3281 * We're gonna grab cgroup_mutex which nests outside kernfs
3282 * active_ref. kernfs_rename() doesn't require active_ref
3283 * protection. Break them before grabbing cgroup_mutex.
3285 kernfs_break_active_protection(new_parent);
3286 kernfs_break_active_protection(kn);
3288 mutex_lock(&cgroup_mutex);
3290 ret = kernfs_rename(kn, new_parent, new_name_str);
3292 mutex_unlock(&cgroup_mutex);
3294 kernfs_unbreak_active_protection(kn);
3295 kernfs_unbreak_active_protection(new_parent);
3299 /* set uid and gid of cgroup dirs and files to that of the creator */
3300 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3302 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3303 .ia_uid = current_fsuid(),
3304 .ia_gid = current_fsgid(), };
3306 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3307 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3310 return kernfs_setattr(kn, &iattr);
3313 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3316 char name[CGROUP_FILE_NAME_MAX];
3317 struct kernfs_node *kn;
3318 struct lock_class_key *key = NULL;
3321 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3322 key = &cft->lockdep_key;
3324 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3325 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3330 ret = cgroup_kn_set_ugid(kn);
3336 if (cft->file_offset) {
3337 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3339 spin_lock_irq(&cgroup_file_kn_lock);
3341 spin_unlock_irq(&cgroup_file_kn_lock);
3348 * cgroup_addrm_files - add or remove files to a cgroup directory
3349 * @css: the target css
3350 * @cgrp: the target cgroup (usually css->cgroup)
3351 * @cfts: array of cftypes to be added
3352 * @is_add: whether to add or remove
3354 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3355 * For removals, this function never fails.
3357 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3358 struct cgroup *cgrp, struct cftype cfts[],
3361 struct cftype *cft, *cft_end = NULL;
3364 lockdep_assert_held(&cgroup_mutex);
3367 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3368 /* does cft->flags tell us to skip this file on @cgrp? */
3369 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3371 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3373 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3375 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3379 ret = cgroup_add_file(css, cgrp, cft);
3381 pr_warn("%s: failed to add %s, err=%d\n",
3382 __func__, cft->name, ret);
3388 cgroup_rm_file(cgrp, cft);
3394 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3397 struct cgroup_subsys *ss = cfts[0].ss;
3398 struct cgroup *root = &ss->root->cgrp;
3399 struct cgroup_subsys_state *css;
3402 lockdep_assert_held(&cgroup_mutex);
3404 /* add/rm files for all cgroups created before */
3405 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3406 struct cgroup *cgrp = css->cgroup;
3408 if (cgroup_is_dead(cgrp))
3411 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3417 kernfs_activate(root->kn);
3421 static void cgroup_exit_cftypes(struct cftype *cfts)
3425 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3426 /* free copy for custom atomic_write_len, see init_cftypes() */
3427 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3432 /* revert flags set by cgroup core while adding @cfts */
3433 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3437 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3441 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3442 struct kernfs_ops *kf_ops;
3444 WARN_ON(cft->ss || cft->kf_ops);
3447 kf_ops = &cgroup_kf_ops;
3449 kf_ops = &cgroup_kf_single_ops;
3452 * Ugh... if @cft wants a custom max_write_len, we need to
3453 * make a copy of kf_ops to set its atomic_write_len.
3455 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3456 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3458 cgroup_exit_cftypes(cfts);
3461 kf_ops->atomic_write_len = cft->max_write_len;
3464 cft->kf_ops = kf_ops;
3471 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3473 lockdep_assert_held(&cgroup_mutex);
3475 if (!cfts || !cfts[0].ss)
3478 list_del(&cfts->node);
3479 cgroup_apply_cftypes(cfts, false);
3480 cgroup_exit_cftypes(cfts);
3485 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3486 * @cfts: zero-length name terminated array of cftypes
3488 * Unregister @cfts. Files described by @cfts are removed from all
3489 * existing cgroups and all future cgroups won't have them either. This
3490 * function can be called anytime whether @cfts' subsys is attached or not.
3492 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3495 int cgroup_rm_cftypes(struct cftype *cfts)
3499 mutex_lock(&cgroup_mutex);
3500 ret = cgroup_rm_cftypes_locked(cfts);
3501 mutex_unlock(&cgroup_mutex);
3506 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3507 * @ss: target cgroup subsystem
3508 * @cfts: zero-length name terminated array of cftypes
3510 * Register @cfts to @ss. Files described by @cfts are created for all
3511 * existing cgroups to which @ss is attached and all future cgroups will
3512 * have them too. This function can be called anytime whether @ss is
3515 * Returns 0 on successful registration, -errno on failure. Note that this
3516 * function currently returns 0 as long as @cfts registration is successful
3517 * even if some file creation attempts on existing cgroups fail.
3519 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3523 if (!cgroup_ssid_enabled(ss->id))
3526 if (!cfts || cfts[0].name[0] == '\0')
3529 ret = cgroup_init_cftypes(ss, cfts);
3533 mutex_lock(&cgroup_mutex);
3535 list_add_tail(&cfts->node, &ss->cfts);
3536 ret = cgroup_apply_cftypes(cfts, true);
3538 cgroup_rm_cftypes_locked(cfts);
3540 mutex_unlock(&cgroup_mutex);
3545 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3546 * @ss: target cgroup subsystem
3547 * @cfts: zero-length name terminated array of cftypes
3549 * Similar to cgroup_add_cftypes() but the added files are only used for
3550 * the default hierarchy.
3552 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3556 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3557 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3558 return cgroup_add_cftypes(ss, cfts);
3562 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3563 * @ss: target cgroup subsystem
3564 * @cfts: zero-length name terminated array of cftypes
3566 * Similar to cgroup_add_cftypes() but the added files are only used for
3567 * the legacy hierarchies.
3569 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3573 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3574 cft->flags |= __CFTYPE_NOT_ON_DFL;
3575 return cgroup_add_cftypes(ss, cfts);
3579 * cgroup_file_notify - generate a file modified event for a cgroup_file
3580 * @cfile: target cgroup_file
3582 * @cfile must have been obtained by setting cftype->file_offset.
3584 void cgroup_file_notify(struct cgroup_file *cfile)
3586 unsigned long flags;
3588 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3590 kernfs_notify(cfile->kn);
3591 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3595 * cgroup_task_count - count the number of tasks in a cgroup.
3596 * @cgrp: the cgroup in question
3598 * Return the number of tasks in the cgroup.
3600 static int cgroup_task_count(const struct cgroup *cgrp)
3603 struct cgrp_cset_link *link;
3605 spin_lock_bh(&css_set_lock);
3606 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3607 count += atomic_read(&link->cset->refcount);
3608 spin_unlock_bh(&css_set_lock);
3613 * css_next_child - find the next child of a given css
3614 * @pos: the current position (%NULL to initiate traversal)
3615 * @parent: css whose children to walk
3617 * This function returns the next child of @parent and should be called
3618 * under either cgroup_mutex or RCU read lock. The only requirement is
3619 * that @parent and @pos are accessible. The next sibling is guaranteed to
3620 * be returned regardless of their states.
3622 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3623 * css which finished ->css_online() is guaranteed to be visible in the
3624 * future iterations and will stay visible until the last reference is put.
3625 * A css which hasn't finished ->css_online() or already finished
3626 * ->css_offline() may show up during traversal. It's each subsystem's
3627 * responsibility to synchronize against on/offlining.
3629 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3630 struct cgroup_subsys_state *parent)
3632 struct cgroup_subsys_state *next;
3634 cgroup_assert_mutex_or_rcu_locked();
3637 * @pos could already have been unlinked from the sibling list.
3638 * Once a cgroup is removed, its ->sibling.next is no longer
3639 * updated when its next sibling changes. CSS_RELEASED is set when
3640 * @pos is taken off list, at which time its next pointer is valid,
3641 * and, as releases are serialized, the one pointed to by the next
3642 * pointer is guaranteed to not have started release yet. This
3643 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3644 * critical section, the one pointed to by its next pointer is
3645 * guaranteed to not have finished its RCU grace period even if we
3646 * have dropped rcu_read_lock() inbetween iterations.
3648 * If @pos has CSS_RELEASED set, its next pointer can't be
3649 * dereferenced; however, as each css is given a monotonically
3650 * increasing unique serial number and always appended to the
3651 * sibling list, the next one can be found by walking the parent's
3652 * children until the first css with higher serial number than
3653 * @pos's. While this path can be slower, it happens iff iteration
3654 * races against release and the race window is very small.
3657 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3658 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3659 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3661 list_for_each_entry_rcu(next, &parent->children, sibling)
3662 if (next->serial_nr > pos->serial_nr)
3667 * @next, if not pointing to the head, can be dereferenced and is
3670 if (&next->sibling != &parent->children)
3676 * css_next_descendant_pre - find the next descendant for pre-order walk
3677 * @pos: the current position (%NULL to initiate traversal)
3678 * @root: css whose descendants to walk
3680 * To be used by css_for_each_descendant_pre(). Find the next descendant
3681 * to visit for pre-order traversal of @root's descendants. @root is
3682 * included in the iteration and the first node to be visited.
3684 * While this function requires cgroup_mutex or RCU read locking, it
3685 * doesn't require the whole traversal to be contained in a single critical
3686 * section. This function will return the correct next descendant as long
3687 * as both @pos and @root are accessible and @pos is a descendant of @root.
3689 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3690 * css which finished ->css_online() is guaranteed to be visible in the
3691 * future iterations and will stay visible until the last reference is put.
3692 * A css which hasn't finished ->css_online() or already finished
3693 * ->css_offline() may show up during traversal. It's each subsystem's
3694 * responsibility to synchronize against on/offlining.
3696 struct cgroup_subsys_state *
3697 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3698 struct cgroup_subsys_state *root)
3700 struct cgroup_subsys_state *next;
3702 cgroup_assert_mutex_or_rcu_locked();
3704 /* if first iteration, visit @root */
3708 /* visit the first child if exists */
3709 next = css_next_child(NULL, pos);
3713 /* no child, visit my or the closest ancestor's next sibling */
3714 while (pos != root) {
3715 next = css_next_child(pos, pos->parent);
3725 * css_rightmost_descendant - return the rightmost descendant of a css
3726 * @pos: css of interest
3728 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3729 * is returned. This can be used during pre-order traversal to skip
3732 * While this function requires cgroup_mutex or RCU read locking, it
3733 * doesn't require the whole traversal to be contained in a single critical
3734 * section. This function will return the correct rightmost descendant as
3735 * long as @pos is accessible.
3737 struct cgroup_subsys_state *
3738 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3740 struct cgroup_subsys_state *last, *tmp;
3742 cgroup_assert_mutex_or_rcu_locked();
3746 /* ->prev isn't RCU safe, walk ->next till the end */
3748 css_for_each_child(tmp, last)
3755 static struct cgroup_subsys_state *
3756 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3758 struct cgroup_subsys_state *last;
3762 pos = css_next_child(NULL, pos);
3769 * css_next_descendant_post - find the next descendant for post-order walk
3770 * @pos: the current position (%NULL to initiate traversal)
3771 * @root: css whose descendants to walk
3773 * To be used by css_for_each_descendant_post(). Find the next descendant
3774 * to visit for post-order traversal of @root's descendants. @root is
3775 * included in the iteration and the last node to be visited.
3777 * While this function requires cgroup_mutex or RCU read locking, it
3778 * doesn't require the whole traversal to be contained in a single critical
3779 * section. This function will return the correct next descendant as long
3780 * as both @pos and @cgroup are accessible and @pos is a descendant of
3783 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3784 * css which finished ->css_online() is guaranteed to be visible in the
3785 * future iterations and will stay visible until the last reference is put.
3786 * A css which hasn't finished ->css_online() or already finished
3787 * ->css_offline() may show up during traversal. It's each subsystem's
3788 * responsibility to synchronize against on/offlining.
3790 struct cgroup_subsys_state *
3791 css_next_descendant_post(struct cgroup_subsys_state *pos,
3792 struct cgroup_subsys_state *root)
3794 struct cgroup_subsys_state *next;
3796 cgroup_assert_mutex_or_rcu_locked();
3798 /* if first iteration, visit leftmost descendant which may be @root */
3800 return css_leftmost_descendant(root);
3802 /* if we visited @root, we're done */
3806 /* if there's an unvisited sibling, visit its leftmost descendant */
3807 next = css_next_child(pos, pos->parent);
3809 return css_leftmost_descendant(next);
3811 /* no sibling left, visit parent */
3816 * css_has_online_children - does a css have online children
3817 * @css: the target css
3819 * Returns %true if @css has any online children; otherwise, %false. This
3820 * function can be called from any context but the caller is responsible
3821 * for synchronizing against on/offlining as necessary.
3823 bool css_has_online_children(struct cgroup_subsys_state *css)
3825 struct cgroup_subsys_state *child;
3829 css_for_each_child(child, css) {
3830 if (child->flags & CSS_ONLINE) {
3840 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3841 * @it: the iterator to advance
3843 * Advance @it to the next css_set to walk.
3845 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3847 struct list_head *l = it->cset_pos;
3848 struct cgrp_cset_link *link;
3849 struct css_set *cset;
3851 lockdep_assert_held(&css_set_lock);
3853 /* Advance to the next non-empty css_set */
3856 if (l == it->cset_head) {
3857 it->cset_pos = NULL;
3858 it->task_pos = NULL;
3863 cset = container_of(l, struct css_set,
3864 e_cset_node[it->ss->id]);
3866 link = list_entry(l, struct cgrp_cset_link, cset_link);
3869 } while (!css_set_populated(cset));
3873 if (!list_empty(&cset->tasks))
3874 it->task_pos = cset->tasks.next;
3876 it->task_pos = cset->mg_tasks.next;
3878 it->tasks_head = &cset->tasks;
3879 it->mg_tasks_head = &cset->mg_tasks;
3882 * We don't keep css_sets locked across iteration steps and thus
3883 * need to take steps to ensure that iteration can be resumed after
3884 * the lock is re-acquired. Iteration is performed at two levels -
3885 * css_sets and tasks in them.
3887 * Once created, a css_set never leaves its cgroup lists, so a
3888 * pinned css_set is guaranteed to stay put and we can resume
3889 * iteration afterwards.
3891 * Tasks may leave @cset across iteration steps. This is resolved
3892 * by registering each iterator with the css_set currently being
3893 * walked and making css_set_move_task() advance iterators whose
3894 * next task is leaving.
3897 list_del(&it->iters_node);
3898 put_css_set_locked(it->cur_cset);
3901 it->cur_cset = cset;
3902 list_add(&it->iters_node, &cset->task_iters);
3905 static void css_task_iter_advance(struct css_task_iter *it)
3907 struct list_head *l = it->task_pos;
3909 lockdep_assert_held(&css_set_lock);
3913 * Advance iterator to find next entry. cset->tasks is consumed
3914 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3919 if (l == it->tasks_head)
3920 l = it->mg_tasks_head->next;
3922 if (l == it->mg_tasks_head)
3923 css_task_iter_advance_css_set(it);
3929 * css_task_iter_start - initiate task iteration
3930 * @css: the css to walk tasks of
3931 * @it: the task iterator to use
3933 * Initiate iteration through the tasks of @css. The caller can call
3934 * css_task_iter_next() to walk through the tasks until the function
3935 * returns NULL. On completion of iteration, css_task_iter_end() must be
3938 void css_task_iter_start(struct cgroup_subsys_state *css,
3939 struct css_task_iter *it)
3941 /* no one should try to iterate before mounting cgroups */
3942 WARN_ON_ONCE(!use_task_css_set_links);
3944 memset(it, 0, sizeof(*it));
3946 spin_lock_bh(&css_set_lock);
3951 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3953 it->cset_pos = &css->cgroup->cset_links;
3955 it->cset_head = it->cset_pos;
3957 css_task_iter_advance_css_set(it);
3959 spin_unlock_bh(&css_set_lock);
3963 * css_task_iter_next - return the next task for the iterator
3964 * @it: the task iterator being iterated
3966 * The "next" function for task iteration. @it should have been
3967 * initialized via css_task_iter_start(). Returns NULL when the iteration
3970 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3973 put_task_struct(it->cur_task);
3974 it->cur_task = NULL;
3977 spin_lock_bh(&css_set_lock);
3980 it->cur_task = list_entry(it->task_pos, struct task_struct,
3982 get_task_struct(it->cur_task);
3983 css_task_iter_advance(it);
3986 spin_unlock_bh(&css_set_lock);
3988 return it->cur_task;
3992 * css_task_iter_end - finish task iteration
3993 * @it: the task iterator to finish
3995 * Finish task iteration started by css_task_iter_start().
3997 void css_task_iter_end(struct css_task_iter *it)
4000 spin_lock_bh(&css_set_lock);
4001 list_del(&it->iters_node);
4002 put_css_set_locked(it->cur_cset);
4003 spin_unlock_bh(&css_set_lock);
4007 put_task_struct(it->cur_task);
4011 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4012 * @to: cgroup to which the tasks will be moved
4013 * @from: cgroup in which the tasks currently reside
4015 * Locking rules between cgroup_post_fork() and the migration path
4016 * guarantee that, if a task is forking while being migrated, the new child
4017 * is guaranteed to be either visible in the source cgroup after the
4018 * parent's migration is complete or put into the target cgroup. No task
4019 * can slip out of migration through forking.
4021 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4023 LIST_HEAD(preloaded_csets);
4024 struct cgrp_cset_link *link;
4025 struct css_task_iter it;
4026 struct task_struct *task;
4029 mutex_lock(&cgroup_mutex);
4031 /* all tasks in @from are being moved, all csets are source */
4032 spin_lock_bh(&css_set_lock);
4033 list_for_each_entry(link, &from->cset_links, cset_link)
4034 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4035 spin_unlock_bh(&css_set_lock);
4037 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4042 * Migrate tasks one-by-one until @form is empty. This fails iff
4043 * ->can_attach() fails.
4046 css_task_iter_start(&from->self, &it);
4047 task = css_task_iter_next(&it);
4049 get_task_struct(task);
4050 css_task_iter_end(&it);
4053 ret = cgroup_migrate(task, false, to);
4054 put_task_struct(task);
4056 } while (task && !ret);
4058 cgroup_migrate_finish(&preloaded_csets);
4059 mutex_unlock(&cgroup_mutex);
4064 * Stuff for reading the 'tasks'/'procs' files.
4066 * Reading this file can return large amounts of data if a cgroup has
4067 * *lots* of attached tasks. So it may need several calls to read(),
4068 * but we cannot guarantee that the information we produce is correct
4069 * unless we produce it entirely atomically.
4073 /* which pidlist file are we talking about? */
4074 enum cgroup_filetype {
4080 * A pidlist is a list of pids that virtually represents the contents of one
4081 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4082 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4085 struct cgroup_pidlist {
4087 * used to find which pidlist is wanted. doesn't change as long as
4088 * this particular list stays in the list.
4090 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4093 /* how many elements the above list has */
4095 /* each of these stored in a list by its cgroup */
4096 struct list_head links;
4097 /* pointer to the cgroup we belong to, for list removal purposes */
4098 struct cgroup *owner;
4099 /* for delayed destruction */
4100 struct delayed_work destroy_dwork;
4104 * The following two functions "fix" the issue where there are more pids
4105 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4106 * TODO: replace with a kernel-wide solution to this problem
4108 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4109 static void *pidlist_allocate(int count)
4111 if (PIDLIST_TOO_LARGE(count))
4112 return vmalloc(count * sizeof(pid_t));
4114 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4117 static void pidlist_free(void *p)
4123 * Used to destroy all pidlists lingering waiting for destroy timer. None
4124 * should be left afterwards.
4126 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4128 struct cgroup_pidlist *l, *tmp_l;
4130 mutex_lock(&cgrp->pidlist_mutex);
4131 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4132 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4133 mutex_unlock(&cgrp->pidlist_mutex);
4135 flush_workqueue(cgroup_pidlist_destroy_wq);
4136 BUG_ON(!list_empty(&cgrp->pidlists));
4139 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4141 struct delayed_work *dwork = to_delayed_work(work);
4142 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4144 struct cgroup_pidlist *tofree = NULL;
4146 mutex_lock(&l->owner->pidlist_mutex);
4149 * Destroy iff we didn't get queued again. The state won't change
4150 * as destroy_dwork can only be queued while locked.
4152 if (!delayed_work_pending(dwork)) {
4153 list_del(&l->links);
4154 pidlist_free(l->list);
4155 put_pid_ns(l->key.ns);
4159 mutex_unlock(&l->owner->pidlist_mutex);
4164 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4165 * Returns the number of unique elements.
4167 static int pidlist_uniq(pid_t *list, int length)
4172 * we presume the 0th element is unique, so i starts at 1. trivial
4173 * edge cases first; no work needs to be done for either
4175 if (length == 0 || length == 1)
4177 /* src and dest walk down the list; dest counts unique elements */
4178 for (src = 1; src < length; src++) {
4179 /* find next unique element */
4180 while (list[src] == list[src-1]) {
4185 /* dest always points to where the next unique element goes */
4186 list[dest] = list[src];
4194 * The two pid files - task and cgroup.procs - guaranteed that the result
4195 * is sorted, which forced this whole pidlist fiasco. As pid order is
4196 * different per namespace, each namespace needs differently sorted list,
4197 * making it impossible to use, for example, single rbtree of member tasks
4198 * sorted by task pointer. As pidlists can be fairly large, allocating one
4199 * per open file is dangerous, so cgroup had to implement shared pool of
4200 * pidlists keyed by cgroup and namespace.
4202 * All this extra complexity was caused by the original implementation
4203 * committing to an entirely unnecessary property. In the long term, we
4204 * want to do away with it. Explicitly scramble sort order if on the
4205 * default hierarchy so that no such expectation exists in the new
4208 * Scrambling is done by swapping every two consecutive bits, which is
4209 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4211 static pid_t pid_fry(pid_t pid)
4213 unsigned a = pid & 0x55555555;
4214 unsigned b = pid & 0xAAAAAAAA;
4216 return (a << 1) | (b >> 1);
4219 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4221 if (cgroup_on_dfl(cgrp))
4222 return pid_fry(pid);
4227 static int cmppid(const void *a, const void *b)
4229 return *(pid_t *)a - *(pid_t *)b;
4232 static int fried_cmppid(const void *a, const void *b)
4234 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4237 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4238 enum cgroup_filetype type)
4240 struct cgroup_pidlist *l;
4241 /* don't need task_nsproxy() if we're looking at ourself */
4242 struct pid_namespace *ns = task_active_pid_ns(current);
4244 lockdep_assert_held(&cgrp->pidlist_mutex);
4246 list_for_each_entry(l, &cgrp->pidlists, links)
4247 if (l->key.type == type && l->key.ns == ns)
4253 * find the appropriate pidlist for our purpose (given procs vs tasks)
4254 * returns with the lock on that pidlist already held, and takes care
4255 * of the use count, or returns NULL with no locks held if we're out of
4258 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4259 enum cgroup_filetype type)
4261 struct cgroup_pidlist *l;
4263 lockdep_assert_held(&cgrp->pidlist_mutex);
4265 l = cgroup_pidlist_find(cgrp, type);
4269 /* entry not found; create a new one */
4270 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4274 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4276 /* don't need task_nsproxy() if we're looking at ourself */
4277 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4279 list_add(&l->links, &cgrp->pidlists);
4284 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4286 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4287 struct cgroup_pidlist **lp)
4291 int pid, n = 0; /* used for populating the array */
4292 struct css_task_iter it;
4293 struct task_struct *tsk;
4294 struct cgroup_pidlist *l;
4296 lockdep_assert_held(&cgrp->pidlist_mutex);
4299 * If cgroup gets more users after we read count, we won't have
4300 * enough space - tough. This race is indistinguishable to the
4301 * caller from the case that the additional cgroup users didn't
4302 * show up until sometime later on.
4304 length = cgroup_task_count(cgrp);
4305 array = pidlist_allocate(length);
4308 /* now, populate the array */
4309 css_task_iter_start(&cgrp->self, &it);
4310 while ((tsk = css_task_iter_next(&it))) {
4311 if (unlikely(n == length))
4313 /* get tgid or pid for procs or tasks file respectively */
4314 if (type == CGROUP_FILE_PROCS)
4315 pid = task_tgid_vnr(tsk);
4317 pid = task_pid_vnr(tsk);
4318 if (pid > 0) /* make sure to only use valid results */
4321 css_task_iter_end(&it);
4323 /* now sort & (if procs) strip out duplicates */
4324 if (cgroup_on_dfl(cgrp))
4325 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4327 sort(array, length, sizeof(pid_t), cmppid, NULL);
4328 if (type == CGROUP_FILE_PROCS)
4329 length = pidlist_uniq(array, length);
4331 l = cgroup_pidlist_find_create(cgrp, type);
4333 pidlist_free(array);
4337 /* store array, freeing old if necessary */
4338 pidlist_free(l->list);
4346 * cgroupstats_build - build and fill cgroupstats
4347 * @stats: cgroupstats to fill information into
4348 * @dentry: A dentry entry belonging to the cgroup for which stats have
4351 * Build and fill cgroupstats so that taskstats can export it to user
4354 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4356 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4357 struct cgroup *cgrp;
4358 struct css_task_iter it;
4359 struct task_struct *tsk;
4361 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4362 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4363 kernfs_type(kn) != KERNFS_DIR)
4366 mutex_lock(&cgroup_mutex);
4369 * We aren't being called from kernfs and there's no guarantee on
4370 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4371 * @kn->priv is RCU safe. Let's do the RCU dancing.
4374 cgrp = rcu_dereference(kn->priv);
4375 if (!cgrp || cgroup_is_dead(cgrp)) {
4377 mutex_unlock(&cgroup_mutex);
4382 css_task_iter_start(&cgrp->self, &it);
4383 while ((tsk = css_task_iter_next(&it))) {
4384 switch (tsk->state) {
4386 stats->nr_running++;
4388 case TASK_INTERRUPTIBLE:
4389 stats->nr_sleeping++;
4391 case TASK_UNINTERRUPTIBLE:
4392 stats->nr_uninterruptible++;
4395 stats->nr_stopped++;
4398 if (delayacct_is_task_waiting_on_io(tsk))
4399 stats->nr_io_wait++;
4403 css_task_iter_end(&it);
4405 mutex_unlock(&cgroup_mutex);
4411 * seq_file methods for the tasks/procs files. The seq_file position is the
4412 * next pid to display; the seq_file iterator is a pointer to the pid
4413 * in the cgroup->l->list array.
4416 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4419 * Initially we receive a position value that corresponds to
4420 * one more than the last pid shown (or 0 on the first call or
4421 * after a seek to the start). Use a binary-search to find the
4422 * next pid to display, if any
4424 struct kernfs_open_file *of = s->private;
4425 struct cgroup *cgrp = seq_css(s)->cgroup;
4426 struct cgroup_pidlist *l;
4427 enum cgroup_filetype type = seq_cft(s)->private;
4428 int index = 0, pid = *pos;
4431 mutex_lock(&cgrp->pidlist_mutex);
4434 * !NULL @of->priv indicates that this isn't the first start()
4435 * after open. If the matching pidlist is around, we can use that.
4436 * Look for it. Note that @of->priv can't be used directly. It
4437 * could already have been destroyed.
4440 of->priv = cgroup_pidlist_find(cgrp, type);
4443 * Either this is the first start() after open or the matching
4444 * pidlist has been destroyed inbetween. Create a new one.
4447 ret = pidlist_array_load(cgrp, type,
4448 (struct cgroup_pidlist **)&of->priv);
4450 return ERR_PTR(ret);
4455 int end = l->length;
4457 while (index < end) {
4458 int mid = (index + end) / 2;
4459 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4462 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4468 /* If we're off the end of the array, we're done */
4469 if (index >= l->length)
4471 /* Update the abstract position to be the actual pid that we found */
4472 iter = l->list + index;
4473 *pos = cgroup_pid_fry(cgrp, *iter);
4477 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4479 struct kernfs_open_file *of = s->private;
4480 struct cgroup_pidlist *l = of->priv;
4483 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4484 CGROUP_PIDLIST_DESTROY_DELAY);
4485 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4488 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4490 struct kernfs_open_file *of = s->private;
4491 struct cgroup_pidlist *l = of->priv;
4493 pid_t *end = l->list + l->length;
4495 * Advance to the next pid in the array. If this goes off the
4502 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4507 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4509 seq_printf(s, "%d\n", *(int *)v);
4514 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4517 return notify_on_release(css->cgroup);
4520 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4521 struct cftype *cft, u64 val)
4524 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4526 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4530 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4533 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4536 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4537 struct cftype *cft, u64 val)
4540 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4542 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4546 /* cgroup core interface files for the default hierarchy */
4547 static struct cftype cgroup_dfl_base_files[] = {
4549 .name = "cgroup.procs",
4550 .file_offset = offsetof(struct cgroup, procs_file),
4551 .seq_start = cgroup_pidlist_start,
4552 .seq_next = cgroup_pidlist_next,
4553 .seq_stop = cgroup_pidlist_stop,
4554 .seq_show = cgroup_pidlist_show,
4555 .private = CGROUP_FILE_PROCS,
4556 .write = cgroup_procs_write,
4559 .name = "cgroup.controllers",
4560 .flags = CFTYPE_ONLY_ON_ROOT,
4561 .seq_show = cgroup_root_controllers_show,
4564 .name = "cgroup.controllers",
4565 .flags = CFTYPE_NOT_ON_ROOT,
4566 .seq_show = cgroup_controllers_show,
4569 .name = "cgroup.subtree_control",
4570 .seq_show = cgroup_subtree_control_show,
4571 .write = cgroup_subtree_control_write,
4574 .name = "cgroup.events",
4575 .flags = CFTYPE_NOT_ON_ROOT,
4576 .file_offset = offsetof(struct cgroup, events_file),
4577 .seq_show = cgroup_events_show,
4582 /* cgroup core interface files for the legacy hierarchies */
4583 static struct cftype cgroup_legacy_base_files[] = {
4585 .name = "cgroup.procs",
4586 .seq_start = cgroup_pidlist_start,
4587 .seq_next = cgroup_pidlist_next,
4588 .seq_stop = cgroup_pidlist_stop,
4589 .seq_show = cgroup_pidlist_show,
4590 .private = CGROUP_FILE_PROCS,
4591 .write = cgroup_procs_write,
4594 .name = "cgroup.clone_children",
4595 .read_u64 = cgroup_clone_children_read,
4596 .write_u64 = cgroup_clone_children_write,
4599 .name = "cgroup.sane_behavior",
4600 .flags = CFTYPE_ONLY_ON_ROOT,
4601 .seq_show = cgroup_sane_behavior_show,
4605 .seq_start = cgroup_pidlist_start,
4606 .seq_next = cgroup_pidlist_next,
4607 .seq_stop = cgroup_pidlist_stop,
4608 .seq_show = cgroup_pidlist_show,
4609 .private = CGROUP_FILE_TASKS,
4610 .write = cgroup_tasks_write,
4613 .name = "notify_on_release",
4614 .read_u64 = cgroup_read_notify_on_release,
4615 .write_u64 = cgroup_write_notify_on_release,
4618 .name = "release_agent",
4619 .flags = CFTYPE_ONLY_ON_ROOT,
4620 .seq_show = cgroup_release_agent_show,
4621 .write = cgroup_release_agent_write,
4622 .max_write_len = PATH_MAX - 1,
4628 * css destruction is four-stage process.
4630 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4631 * Implemented in kill_css().
4633 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4634 * and thus css_tryget_online() is guaranteed to fail, the css can be
4635 * offlined by invoking offline_css(). After offlining, the base ref is
4636 * put. Implemented in css_killed_work_fn().
4638 * 3. When the percpu_ref reaches zero, the only possible remaining
4639 * accessors are inside RCU read sections. css_release() schedules the
4642 * 4. After the grace period, the css can be freed. Implemented in
4643 * css_free_work_fn().
4645 * It is actually hairier because both step 2 and 4 require process context
4646 * and thus involve punting to css->destroy_work adding two additional
4647 * steps to the already complex sequence.
4649 static void css_free_work_fn(struct work_struct *work)
4651 struct cgroup_subsys_state *css =
4652 container_of(work, struct cgroup_subsys_state, destroy_work);
4653 struct cgroup_subsys *ss = css->ss;
4654 struct cgroup *cgrp = css->cgroup;
4656 percpu_ref_exit(&css->refcnt);
4663 css_put(css->parent);
4666 cgroup_idr_remove(&ss->css_idr, id);
4669 /* cgroup free path */
4670 atomic_dec(&cgrp->root->nr_cgrps);
4671 cgroup_pidlist_destroy_all(cgrp);
4672 cancel_work_sync(&cgrp->release_agent_work);
4674 if (cgroup_parent(cgrp)) {
4676 * We get a ref to the parent, and put the ref when
4677 * this cgroup is being freed, so it's guaranteed
4678 * that the parent won't be destroyed before its
4681 cgroup_put(cgroup_parent(cgrp));
4682 kernfs_put(cgrp->kn);
4686 * This is root cgroup's refcnt reaching zero,
4687 * which indicates that the root should be
4690 cgroup_destroy_root(cgrp->root);
4695 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4697 struct cgroup_subsys_state *css =
4698 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4700 INIT_WORK(&css->destroy_work, css_free_work_fn);
4701 queue_work(cgroup_destroy_wq, &css->destroy_work);
4704 static void css_release_work_fn(struct work_struct *work)
4706 struct cgroup_subsys_state *css =
4707 container_of(work, struct cgroup_subsys_state, destroy_work);
4708 struct cgroup_subsys *ss = css->ss;
4709 struct cgroup *cgrp = css->cgroup;
4711 mutex_lock(&cgroup_mutex);
4713 css->flags |= CSS_RELEASED;
4714 list_del_rcu(&css->sibling);
4717 /* css release path */
4718 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4719 if (ss->css_released)
4720 ss->css_released(css);
4722 /* cgroup release path */
4723 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4727 * There are two control paths which try to determine
4728 * cgroup from dentry without going through kernfs -
4729 * cgroupstats_build() and css_tryget_online_from_dir().
4730 * Those are supported by RCU protecting clearing of
4731 * cgrp->kn->priv backpointer.
4733 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4736 mutex_unlock(&cgroup_mutex);
4738 call_rcu(&css->rcu_head, css_free_rcu_fn);
4741 static void css_release(struct percpu_ref *ref)
4743 struct cgroup_subsys_state *css =
4744 container_of(ref, struct cgroup_subsys_state, refcnt);
4746 INIT_WORK(&css->destroy_work, css_release_work_fn);
4747 queue_work(cgroup_destroy_wq, &css->destroy_work);
4750 static void init_and_link_css(struct cgroup_subsys_state *css,
4751 struct cgroup_subsys *ss, struct cgroup *cgrp)
4753 lockdep_assert_held(&cgroup_mutex);
4757 memset(css, 0, sizeof(*css));
4760 INIT_LIST_HEAD(&css->sibling);
4761 INIT_LIST_HEAD(&css->children);
4762 css->serial_nr = css_serial_nr_next++;
4764 if (cgroup_parent(cgrp)) {
4765 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4766 css_get(css->parent);
4769 BUG_ON(cgroup_css(cgrp, ss));
4772 /* invoke ->css_online() on a new CSS and mark it online if successful */
4773 static int online_css(struct cgroup_subsys_state *css)
4775 struct cgroup_subsys *ss = css->ss;
4778 lockdep_assert_held(&cgroup_mutex);
4781 ret = ss->css_online(css);
4783 css->flags |= CSS_ONLINE;
4784 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4789 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4790 static void offline_css(struct cgroup_subsys_state *css)
4792 struct cgroup_subsys *ss = css->ss;
4794 lockdep_assert_held(&cgroup_mutex);
4796 if (!(css->flags & CSS_ONLINE))
4799 if (ss->css_offline)
4800 ss->css_offline(css);
4802 css->flags &= ~CSS_ONLINE;
4803 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4805 wake_up_all(&css->cgroup->offline_waitq);
4809 * create_css - create a cgroup_subsys_state
4810 * @cgrp: the cgroup new css will be associated with
4811 * @ss: the subsys of new css
4812 * @visible: whether to create control knobs for the new css or not
4814 * Create a new css associated with @cgrp - @ss pair. On success, the new
4815 * css is online and installed in @cgrp with all interface files created if
4816 * @visible. Returns 0 on success, -errno on failure.
4818 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4821 struct cgroup *parent = cgroup_parent(cgrp);
4822 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4823 struct cgroup_subsys_state *css;
4826 lockdep_assert_held(&cgroup_mutex);
4828 css = ss->css_alloc(parent_css);
4830 return PTR_ERR(css);
4832 init_and_link_css(css, ss, cgrp);
4834 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4838 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4840 goto err_free_percpu_ref;
4844 err = css_populate_dir(css, NULL);
4849 /* @css is ready to be brought online now, make it visible */
4850 list_add_tail_rcu(&css->sibling, &parent_css->children);
4851 cgroup_idr_replace(&ss->css_idr, css, css->id);
4853 err = online_css(css);
4857 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4858 cgroup_parent(parent)) {
4859 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4860 current->comm, current->pid, ss->name);
4861 if (!strcmp(ss->name, "memory"))
4862 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4863 ss->warned_broken_hierarchy = true;
4869 list_del_rcu(&css->sibling);
4870 css_clear_dir(css, NULL);
4872 cgroup_idr_remove(&ss->css_idr, css->id);
4873 err_free_percpu_ref:
4874 percpu_ref_exit(&css->refcnt);
4876 call_rcu(&css->rcu_head, css_free_rcu_fn);
4880 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4883 struct cgroup *parent, *cgrp, *tcgrp;
4884 struct cgroup_root *root;
4885 struct cgroup_subsys *ss;
4886 struct kernfs_node *kn;
4887 int level, ssid, ret;
4889 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4891 if (strchr(name, '\n'))
4894 parent = cgroup_kn_lock_live(parent_kn);
4897 root = parent->root;
4898 level = parent->level + 1;
4900 /* allocate the cgroup and its ID, 0 is reserved for the root */
4901 cgrp = kzalloc(sizeof(*cgrp) +
4902 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4908 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4913 * Temporarily set the pointer to NULL, so idr_find() won't return
4914 * a half-baked cgroup.
4916 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4919 goto out_cancel_ref;
4922 init_cgroup_housekeeping(cgrp);
4924 cgrp->self.parent = &parent->self;
4926 cgrp->level = level;
4928 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
4929 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4931 if (notify_on_release(parent))
4932 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4934 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4935 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4937 /* create the directory */
4938 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4946 * This extra ref will be put in cgroup_free_fn() and guarantees
4947 * that @cgrp->kn is always accessible.
4951 cgrp->self.serial_nr = css_serial_nr_next++;
4953 /* allocation complete, commit to creation */
4954 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4955 atomic_inc(&root->nr_cgrps);
4959 * @cgrp is now fully operational. If something fails after this
4960 * point, it'll be released via the normal destruction path.
4962 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4964 ret = cgroup_kn_set_ugid(kn);
4968 ret = css_populate_dir(&cgrp->self, NULL);
4972 /* let's create and online css's */
4973 for_each_subsys(ss, ssid) {
4974 if (parent->child_subsys_mask & (1 << ssid)) {
4975 ret = create_css(cgrp, ss,
4976 parent->subtree_control & (1 << ssid));
4983 * On the default hierarchy, a child doesn't automatically inherit
4984 * subtree_control from the parent. Each is configured manually.
4986 if (!cgroup_on_dfl(cgrp)) {
4987 cgrp->subtree_control = parent->subtree_control;
4988 cgroup_refresh_child_subsys_mask(cgrp);
4991 kernfs_activate(kn);
4997 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4999 percpu_ref_exit(&cgrp->self.refcnt);
5003 cgroup_kn_unlock(parent_kn);
5007 cgroup_destroy_locked(cgrp);
5012 * This is called when the refcnt of a css is confirmed to be killed.
5013 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5014 * initate destruction and put the css ref from kill_css().
5016 static void css_killed_work_fn(struct work_struct *work)
5018 struct cgroup_subsys_state *css =
5019 container_of(work, struct cgroup_subsys_state, destroy_work);
5021 mutex_lock(&cgroup_mutex);
5023 mutex_unlock(&cgroup_mutex);
5028 /* css kill confirmation processing requires process context, bounce */
5029 static void css_killed_ref_fn(struct percpu_ref *ref)
5031 struct cgroup_subsys_state *css =
5032 container_of(ref, struct cgroup_subsys_state, refcnt);
5034 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5035 queue_work(cgroup_destroy_wq, &css->destroy_work);
5039 * kill_css - destroy a css
5040 * @css: css to destroy
5042 * This function initiates destruction of @css by removing cgroup interface
5043 * files and putting its base reference. ->css_offline() will be invoked
5044 * asynchronously once css_tryget_online() is guaranteed to fail and when
5045 * the reference count reaches zero, @css will be released.
5047 static void kill_css(struct cgroup_subsys_state *css)
5049 lockdep_assert_held(&cgroup_mutex);
5052 * This must happen before css is disassociated with its cgroup.
5053 * See seq_css() for details.
5055 css_clear_dir(css, NULL);
5058 * Killing would put the base ref, but we need to keep it alive
5059 * until after ->css_offline().
5064 * cgroup core guarantees that, by the time ->css_offline() is
5065 * invoked, no new css reference will be given out via
5066 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5067 * proceed to offlining css's because percpu_ref_kill() doesn't
5068 * guarantee that the ref is seen as killed on all CPUs on return.
5070 * Use percpu_ref_kill_and_confirm() to get notifications as each
5071 * css is confirmed to be seen as killed on all CPUs.
5073 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5077 * cgroup_destroy_locked - the first stage of cgroup destruction
5078 * @cgrp: cgroup to be destroyed
5080 * css's make use of percpu refcnts whose killing latency shouldn't be
5081 * exposed to userland and are RCU protected. Also, cgroup core needs to
5082 * guarantee that css_tryget_online() won't succeed by the time
5083 * ->css_offline() is invoked. To satisfy all the requirements,
5084 * destruction is implemented in the following two steps.
5086 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5087 * userland visible parts and start killing the percpu refcnts of
5088 * css's. Set up so that the next stage will be kicked off once all
5089 * the percpu refcnts are confirmed to be killed.
5091 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5092 * rest of destruction. Once all cgroup references are gone, the
5093 * cgroup is RCU-freed.
5095 * This function implements s1. After this step, @cgrp is gone as far as
5096 * the userland is concerned and a new cgroup with the same name may be
5097 * created. As cgroup doesn't care about the names internally, this
5098 * doesn't cause any problem.
5100 static int cgroup_destroy_locked(struct cgroup *cgrp)
5101 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5103 struct cgroup_subsys_state *css;
5106 lockdep_assert_held(&cgroup_mutex);
5109 * Only migration can raise populated from zero and we're already
5110 * holding cgroup_mutex.
5112 if (cgroup_is_populated(cgrp))
5116 * Make sure there's no live children. We can't test emptiness of
5117 * ->self.children as dead children linger on it while being
5118 * drained; otherwise, "rmdir parent/child parent" may fail.
5120 if (css_has_online_children(&cgrp->self))
5124 * Mark @cgrp dead. This prevents further task migration and child
5125 * creation by disabling cgroup_lock_live_group().
5127 cgrp->self.flags &= ~CSS_ONLINE;
5129 /* initiate massacre of all css's */
5130 for_each_css(css, ssid, cgrp)
5134 * Remove @cgrp directory along with the base files. @cgrp has an
5135 * extra ref on its kn.
5137 kernfs_remove(cgrp->kn);
5139 check_for_release(cgroup_parent(cgrp));
5141 /* put the base reference */
5142 percpu_ref_kill(&cgrp->self.refcnt);
5147 static int cgroup_rmdir(struct kernfs_node *kn)
5149 struct cgroup *cgrp;
5152 cgrp = cgroup_kn_lock_live(kn);
5156 ret = cgroup_destroy_locked(cgrp);
5158 cgroup_kn_unlock(kn);
5162 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5163 .remount_fs = cgroup_remount,
5164 .show_options = cgroup_show_options,
5165 .mkdir = cgroup_mkdir,
5166 .rmdir = cgroup_rmdir,
5167 .rename = cgroup_rename,
5170 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5172 struct cgroup_subsys_state *css;
5174 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
5176 mutex_lock(&cgroup_mutex);
5178 idr_init(&ss->css_idr);
5179 INIT_LIST_HEAD(&ss->cfts);
5181 /* Create the root cgroup state for this subsystem */
5182 ss->root = &cgrp_dfl_root;
5183 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5184 /* We don't handle early failures gracefully */
5185 BUG_ON(IS_ERR(css));
5186 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5189 * Root csses are never destroyed and we can't initialize
5190 * percpu_ref during early init. Disable refcnting.
5192 css->flags |= CSS_NO_REF;
5195 /* allocation can't be done safely during early init */
5198 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5199 BUG_ON(css->id < 0);
5202 /* Update the init_css_set to contain a subsys
5203 * pointer to this state - since the subsystem is
5204 * newly registered, all tasks and hence the
5205 * init_css_set is in the subsystem's root cgroup. */
5206 init_css_set.subsys[ss->id] = css;
5208 have_fork_callback |= (bool)ss->fork << ss->id;
5209 have_exit_callback |= (bool)ss->exit << ss->id;
5210 have_free_callback |= (bool)ss->free << ss->id;
5211 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5213 /* At system boot, before all subsystems have been
5214 * registered, no tasks have been forked, so we don't
5215 * need to invoke fork callbacks here. */
5216 BUG_ON(!list_empty(&init_task.tasks));
5218 BUG_ON(online_css(css));
5220 mutex_unlock(&cgroup_mutex);
5224 * cgroup_init_early - cgroup initialization at system boot
5226 * Initialize cgroups at system boot, and initialize any
5227 * subsystems that request early init.
5229 int __init cgroup_init_early(void)
5231 static struct cgroup_sb_opts __initdata opts;
5232 struct cgroup_subsys *ss;
5235 init_cgroup_root(&cgrp_dfl_root, &opts);
5236 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5238 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5240 for_each_subsys(ss, i) {
5241 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5242 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5243 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5245 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5246 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5249 ss->name = cgroup_subsys_name[i];
5250 if (!ss->legacy_name)
5251 ss->legacy_name = cgroup_subsys_name[i];
5254 cgroup_init_subsys(ss, true);
5259 static unsigned long cgroup_disable_mask __initdata;
5262 * cgroup_init - cgroup initialization
5264 * Register cgroup filesystem and /proc file, and initialize
5265 * any subsystems that didn't request early init.
5267 int __init cgroup_init(void)
5269 struct cgroup_subsys *ss;
5273 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5274 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5275 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5277 mutex_lock(&cgroup_mutex);
5279 /* Add init_css_set to the hash table */
5280 key = css_set_hash(init_css_set.subsys);
5281 hash_add(css_set_table, &init_css_set.hlist, key);
5283 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5285 mutex_unlock(&cgroup_mutex);
5287 for_each_subsys(ss, ssid) {
5288 if (ss->early_init) {
5289 struct cgroup_subsys_state *css =
5290 init_css_set.subsys[ss->id];
5292 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5294 BUG_ON(css->id < 0);
5296 cgroup_init_subsys(ss, false);
5299 list_add_tail(&init_css_set.e_cset_node[ssid],
5300 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5303 * Setting dfl_root subsys_mask needs to consider the
5304 * disabled flag and cftype registration needs kmalloc,
5305 * both of which aren't available during early_init.
5307 if (cgroup_disable_mask & (1 << ssid)) {
5308 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5309 printk(KERN_INFO "Disabling %s control group subsystem\n",
5314 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5316 if (!ss->dfl_cftypes)
5317 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5319 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5320 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5322 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5323 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5327 ss->bind(init_css_set.subsys[ssid]);
5330 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5331 WARN_ON(register_filesystem(&cgroup_fs_type));
5332 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5337 static int __init cgroup_wq_init(void)
5340 * There isn't much point in executing destruction path in
5341 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5342 * Use 1 for @max_active.
5344 * We would prefer to do this in cgroup_init() above, but that
5345 * is called before init_workqueues(): so leave this until after.
5347 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5348 BUG_ON(!cgroup_destroy_wq);
5351 * Used to destroy pidlists and separate to serve as flush domain.
5352 * Cap @max_active to 1 too.
5354 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5356 BUG_ON(!cgroup_pidlist_destroy_wq);
5360 core_initcall(cgroup_wq_init);
5363 * proc_cgroup_show()
5364 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5365 * - Used for /proc/<pid>/cgroup.
5367 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5368 struct pid *pid, struct task_struct *tsk)
5372 struct cgroup_root *root;
5375 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5379 mutex_lock(&cgroup_mutex);
5380 spin_lock_bh(&css_set_lock);
5382 for_each_root(root) {
5383 struct cgroup_subsys *ss;
5384 struct cgroup *cgrp;
5385 int ssid, count = 0;
5387 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5390 seq_printf(m, "%d:", root->hierarchy_id);
5391 if (root != &cgrp_dfl_root)
5392 for_each_subsys(ss, ssid)
5393 if (root->subsys_mask & (1 << ssid))
5394 seq_printf(m, "%s%s", count++ ? "," : "",
5396 if (strlen(root->name))
5397 seq_printf(m, "%sname=%s", count ? "," : "",
5401 cgrp = task_cgroup_from_root(tsk, root);
5404 * On traditional hierarchies, all zombie tasks show up as
5405 * belonging to the root cgroup. On the default hierarchy,
5406 * while a zombie doesn't show up in "cgroup.procs" and
5407 * thus can't be migrated, its /proc/PID/cgroup keeps
5408 * reporting the cgroup it belonged to before exiting. If
5409 * the cgroup is removed before the zombie is reaped,
5410 * " (deleted)" is appended to the cgroup path.
5412 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5413 path = cgroup_path(cgrp, buf, PATH_MAX);
5415 retval = -ENAMETOOLONG;
5424 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5425 seq_puts(m, " (deleted)\n");
5432 spin_unlock_bh(&css_set_lock);
5433 mutex_unlock(&cgroup_mutex);
5439 /* Display information about each subsystem and each hierarchy */
5440 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5442 struct cgroup_subsys *ss;
5445 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5447 * ideally we don't want subsystems moving around while we do this.
5448 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5449 * subsys/hierarchy state.
5451 mutex_lock(&cgroup_mutex);
5453 for_each_subsys(ss, i)
5454 seq_printf(m, "%s\t%d\t%d\t%d\n",
5455 ss->legacy_name, ss->root->hierarchy_id,
5456 atomic_read(&ss->root->nr_cgrps),
5457 cgroup_ssid_enabled(i));
5459 mutex_unlock(&cgroup_mutex);
5463 static int cgroupstats_open(struct inode *inode, struct file *file)
5465 return single_open(file, proc_cgroupstats_show, NULL);
5468 static const struct file_operations proc_cgroupstats_operations = {
5469 .open = cgroupstats_open,
5471 .llseek = seq_lseek,
5472 .release = single_release,
5475 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5477 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5478 return &ss_priv[i - CGROUP_CANFORK_START];
5482 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5484 void **private = subsys_canfork_priv_p(ss_priv, i);
5485 return private ? *private : NULL;
5489 * cgroup_fork - initialize cgroup related fields during copy_process()
5490 * @child: pointer to task_struct of forking parent process.
5492 * A task is associated with the init_css_set until cgroup_post_fork()
5493 * attaches it to the parent's css_set. Empty cg_list indicates that
5494 * @child isn't holding reference to its css_set.
5496 void cgroup_fork(struct task_struct *child)
5498 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5499 INIT_LIST_HEAD(&child->cg_list);
5503 * cgroup_can_fork - called on a new task before the process is exposed
5504 * @child: the task in question.
5506 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5507 * returns an error, the fork aborts with that error code. This allows for
5508 * a cgroup subsystem to conditionally allow or deny new forks.
5510 int cgroup_can_fork(struct task_struct *child,
5511 void *ss_priv[CGROUP_CANFORK_COUNT])
5513 struct cgroup_subsys *ss;
5516 for_each_subsys_which(ss, i, &have_canfork_callback) {
5517 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5525 for_each_subsys(ss, j) {
5528 if (ss->cancel_fork)
5529 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5536 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5537 * @child: the task in question
5539 * This calls the cancel_fork() callbacks if a fork failed *after*
5540 * cgroup_can_fork() succeded.
5542 void cgroup_cancel_fork(struct task_struct *child,
5543 void *ss_priv[CGROUP_CANFORK_COUNT])
5545 struct cgroup_subsys *ss;
5548 for_each_subsys(ss, i)
5549 if (ss->cancel_fork)
5550 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5554 * cgroup_post_fork - called on a new task after adding it to the task list
5555 * @child: the task in question
5557 * Adds the task to the list running through its css_set if necessary and
5558 * call the subsystem fork() callbacks. Has to be after the task is
5559 * visible on the task list in case we race with the first call to
5560 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5563 void cgroup_post_fork(struct task_struct *child,
5564 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5566 struct cgroup_subsys *ss;
5570 * This may race against cgroup_enable_task_cg_lists(). As that
5571 * function sets use_task_css_set_links before grabbing
5572 * tasklist_lock and we just went through tasklist_lock to add
5573 * @child, it's guaranteed that either we see the set
5574 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5575 * @child during its iteration.
5577 * If we won the race, @child is associated with %current's
5578 * css_set. Grabbing css_set_lock guarantees both that the
5579 * association is stable, and, on completion of the parent's
5580 * migration, @child is visible in the source of migration or
5581 * already in the destination cgroup. This guarantee is necessary
5582 * when implementing operations which need to migrate all tasks of
5583 * a cgroup to another.
5585 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5586 * will remain in init_css_set. This is safe because all tasks are
5587 * in the init_css_set before cg_links is enabled and there's no
5588 * operation which transfers all tasks out of init_css_set.
5590 if (use_task_css_set_links) {
5591 struct css_set *cset;
5593 spin_lock_bh(&css_set_lock);
5594 cset = task_css_set(current);
5595 if (list_empty(&child->cg_list)) {
5597 css_set_move_task(child, NULL, cset, false);
5599 spin_unlock_bh(&css_set_lock);
5603 * Call ss->fork(). This must happen after @child is linked on
5604 * css_set; otherwise, @child might change state between ->fork()
5605 * and addition to css_set.
5607 for_each_subsys_which(ss, i, &have_fork_callback)
5608 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5612 * cgroup_exit - detach cgroup from exiting task
5613 * @tsk: pointer to task_struct of exiting process
5615 * Description: Detach cgroup from @tsk and release it.
5617 * Note that cgroups marked notify_on_release force every task in
5618 * them to take the global cgroup_mutex mutex when exiting.
5619 * This could impact scaling on very large systems. Be reluctant to
5620 * use notify_on_release cgroups where very high task exit scaling
5621 * is required on large systems.
5623 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5624 * call cgroup_exit() while the task is still competent to handle
5625 * notify_on_release(), then leave the task attached to the root cgroup in
5626 * each hierarchy for the remainder of its exit. No need to bother with
5627 * init_css_set refcnting. init_css_set never goes away and we can't race
5628 * with migration path - PF_EXITING is visible to migration path.
5630 void cgroup_exit(struct task_struct *tsk)
5632 struct cgroup_subsys *ss;
5633 struct css_set *cset;
5637 * Unlink from @tsk from its css_set. As migration path can't race
5638 * with us, we can check css_set and cg_list without synchronization.
5640 cset = task_css_set(tsk);
5642 if (!list_empty(&tsk->cg_list)) {
5643 spin_lock_bh(&css_set_lock);
5644 css_set_move_task(tsk, cset, NULL, false);
5645 spin_unlock_bh(&css_set_lock);
5650 /* see cgroup_post_fork() for details */
5651 for_each_subsys_which(ss, i, &have_exit_callback)
5655 void cgroup_free(struct task_struct *task)
5657 struct css_set *cset = task_css_set(task);
5658 struct cgroup_subsys *ss;
5661 for_each_subsys_which(ss, ssid, &have_free_callback)
5667 static void check_for_release(struct cgroup *cgrp)
5669 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5670 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5671 schedule_work(&cgrp->release_agent_work);
5675 * Notify userspace when a cgroup is released, by running the
5676 * configured release agent with the name of the cgroup (path
5677 * relative to the root of cgroup file system) as the argument.
5679 * Most likely, this user command will try to rmdir this cgroup.
5681 * This races with the possibility that some other task will be
5682 * attached to this cgroup before it is removed, or that some other
5683 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5684 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5685 * unused, and this cgroup will be reprieved from its death sentence,
5686 * to continue to serve a useful existence. Next time it's released,
5687 * we will get notified again, if it still has 'notify_on_release' set.
5689 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5690 * means only wait until the task is successfully execve()'d. The
5691 * separate release agent task is forked by call_usermodehelper(),
5692 * then control in this thread returns here, without waiting for the
5693 * release agent task. We don't bother to wait because the caller of
5694 * this routine has no use for the exit status of the release agent
5695 * task, so no sense holding our caller up for that.
5697 static void cgroup_release_agent(struct work_struct *work)
5699 struct cgroup *cgrp =
5700 container_of(work, struct cgroup, release_agent_work);
5701 char *pathbuf = NULL, *agentbuf = NULL, *path;
5702 char *argv[3], *envp[3];
5704 mutex_lock(&cgroup_mutex);
5706 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5707 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5708 if (!pathbuf || !agentbuf)
5711 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5719 /* minimal command environment */
5721 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5724 mutex_unlock(&cgroup_mutex);
5725 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5728 mutex_unlock(&cgroup_mutex);
5734 static int __init cgroup_disable(char *str)
5736 struct cgroup_subsys *ss;
5740 while ((token = strsep(&str, ",")) != NULL) {
5744 for_each_subsys(ss, i) {
5745 if (strcmp(token, ss->name) &&
5746 strcmp(token, ss->legacy_name))
5748 cgroup_disable_mask |= 1 << i;
5753 __setup("cgroup_disable=", cgroup_disable);
5756 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5757 * @dentry: directory dentry of interest
5758 * @ss: subsystem of interest
5760 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5761 * to get the corresponding css and return it. If such css doesn't exist
5762 * or can't be pinned, an ERR_PTR value is returned.
5764 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5765 struct cgroup_subsys *ss)
5767 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5768 struct cgroup_subsys_state *css = NULL;
5769 struct cgroup *cgrp;
5771 /* is @dentry a cgroup dir? */
5772 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5773 kernfs_type(kn) != KERNFS_DIR)
5774 return ERR_PTR(-EBADF);
5779 * This path doesn't originate from kernfs and @kn could already
5780 * have been or be removed at any point. @kn->priv is RCU
5781 * protected for this access. See css_release_work_fn() for details.
5783 cgrp = rcu_dereference(kn->priv);
5785 css = cgroup_css(cgrp, ss);
5787 if (!css || !css_tryget_online(css))
5788 css = ERR_PTR(-ENOENT);
5795 * css_from_id - lookup css by id
5796 * @id: the cgroup id
5797 * @ss: cgroup subsys to be looked into
5799 * Returns the css if there's valid one with @id, otherwise returns NULL.
5800 * Should be called under rcu_read_lock().
5802 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5804 WARN_ON_ONCE(!rcu_read_lock_held());
5805 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5809 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5810 * @path: path on the default hierarchy
5812 * Find the cgroup at @path on the default hierarchy, increment its
5813 * reference count and return it. Returns pointer to the found cgroup on
5814 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5815 * if @path points to a non-directory.
5817 struct cgroup *cgroup_get_from_path(const char *path)
5819 struct kernfs_node *kn;
5820 struct cgroup *cgrp;
5822 mutex_lock(&cgroup_mutex);
5824 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5826 if (kernfs_type(kn) == KERNFS_DIR) {
5830 cgrp = ERR_PTR(-ENOTDIR);
5834 cgrp = ERR_PTR(-ENOENT);
5837 mutex_unlock(&cgroup_mutex);
5840 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5843 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5844 * definition in cgroup-defs.h.
5846 #ifdef CONFIG_SOCK_CGROUP_DATA
5848 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5850 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5851 static bool cgroup_sk_alloc_disabled __read_mostly;
5853 void cgroup_sk_alloc_disable(void)
5855 if (cgroup_sk_alloc_disabled)
5857 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5858 cgroup_sk_alloc_disabled = true;
5863 #define cgroup_sk_alloc_disabled false
5867 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5869 if (cgroup_sk_alloc_disabled)
5875 struct css_set *cset;
5877 cset = task_css_set(current);
5878 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5879 skcd->val = (unsigned long)cset->dfl_cgrp;
5888 void cgroup_sk_free(struct sock_cgroup_data *skcd)
5890 cgroup_put(sock_cgroup_ptr(skcd));
5893 #endif /* CONFIG_SOCK_CGROUP_DATA */
5895 #ifdef CONFIG_CGROUP_DEBUG
5896 static struct cgroup_subsys_state *
5897 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5899 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5902 return ERR_PTR(-ENOMEM);
5907 static void debug_css_free(struct cgroup_subsys_state *css)
5912 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5915 return cgroup_task_count(css->cgroup);
5918 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5921 return (u64)(unsigned long)current->cgroups;
5924 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5930 count = atomic_read(&task_css_set(current)->refcount);
5935 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5937 struct cgrp_cset_link *link;
5938 struct css_set *cset;
5941 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5945 spin_lock_bh(&css_set_lock);
5947 cset = rcu_dereference(current->cgroups);
5948 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5949 struct cgroup *c = link->cgrp;
5951 cgroup_name(c, name_buf, NAME_MAX + 1);
5952 seq_printf(seq, "Root %d group %s\n",
5953 c->root->hierarchy_id, name_buf);
5956 spin_unlock_bh(&css_set_lock);
5961 #define MAX_TASKS_SHOWN_PER_CSS 25
5962 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5964 struct cgroup_subsys_state *css = seq_css(seq);
5965 struct cgrp_cset_link *link;
5967 spin_lock_bh(&css_set_lock);
5968 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5969 struct css_set *cset = link->cset;
5970 struct task_struct *task;
5973 seq_printf(seq, "css_set %p\n", cset);
5975 list_for_each_entry(task, &cset->tasks, cg_list) {
5976 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5978 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5981 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5982 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5984 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5988 seq_puts(seq, " ...\n");
5990 spin_unlock_bh(&css_set_lock);
5994 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5996 return (!cgroup_is_populated(css->cgroup) &&
5997 !css_has_online_children(&css->cgroup->self));
6000 static struct cftype debug_files[] = {
6002 .name = "taskcount",
6003 .read_u64 = debug_taskcount_read,
6007 .name = "current_css_set",
6008 .read_u64 = current_css_set_read,
6012 .name = "current_css_set_refcount",
6013 .read_u64 = current_css_set_refcount_read,
6017 .name = "current_css_set_cg_links",
6018 .seq_show = current_css_set_cg_links_read,
6022 .name = "cgroup_css_links",
6023 .seq_show = cgroup_css_links_read,
6027 .name = "releasable",
6028 .read_u64 = releasable_read,
6034 struct cgroup_subsys debug_cgrp_subsys = {
6035 .css_alloc = debug_css_alloc,
6036 .css_free = debug_css_free,
6037 .legacy_cftypes = debug_files,
6039 #endif /* CONFIG_CGROUP_DEBUG */