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>
61 #include <linux/cpuset.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_lock protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DEFINE_SPINLOCK(css_set_lock);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_lock);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DEFINE_SPINLOCK(css_set_lock);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
103 * against file removal/re-creation across css hiding.
105 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
108 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
109 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
111 static DEFINE_SPINLOCK(release_agent_path_lock);
113 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct *cgroup_destroy_wq;
129 * pidlist destructions need to be flushed on cgroup destruction. Use a
130 * separate workqueue as flush domain.
132 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
134 /* generate an array of cgroup subsystem pointers */
135 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
136 static struct cgroup_subsys *cgroup_subsys[] = {
137 #include <linux/cgroup_subsys.h>
141 /* array of cgroup subsystem names */
142 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
143 static const char *cgroup_subsys_name[] = {
144 #include <linux/cgroup_subsys.h>
148 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
150 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
151 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
152 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
153 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
154 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
158 static struct static_key_true *cgroup_subsys_enabled_key[] = {
159 #include <linux/cgroup_subsys.h>
163 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
164 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
165 #include <linux/cgroup_subsys.h>
170 * The default hierarchy, reserved for the subsystems that are otherwise
171 * unattached - it never has more than a single cgroup, and all tasks are
172 * part of that cgroup.
174 struct cgroup_root cgrp_dfl_root;
175 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
178 * The default hierarchy always exists but is hidden until mounted for the
179 * first time. This is for backward compatibility.
181 static bool cgrp_dfl_root_visible;
183 /* Controllers blocked by the commandline in v1 */
184 static unsigned long cgroup_no_v1_mask;
186 /* some controllers are not supported in the default hierarchy */
187 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
189 /* The list of hierarchy roots */
191 static LIST_HEAD(cgroup_roots);
192 static int cgroup_root_count;
194 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
195 static DEFINE_IDR(cgroup_hierarchy_idr);
198 * Assign a monotonically increasing serial number to csses. It guarantees
199 * cgroups with bigger numbers are newer than those with smaller numbers.
200 * Also, as csses are always appended to the parent's ->children list, it
201 * guarantees that sibling csses are always sorted in the ascending serial
202 * number order on the list. Protected by cgroup_mutex.
204 static u64 css_serial_nr_next = 1;
207 * These bitmask flags indicate whether tasks in the fork and exit paths have
208 * fork/exit handlers to call. This avoids us having to do extra work in the
209 * fork/exit path to check which subsystems have fork/exit callbacks.
211 static unsigned long have_fork_callback __read_mostly;
212 static unsigned long have_exit_callback __read_mostly;
213 static unsigned long have_free_callback __read_mostly;
215 /* Ditto for the can_fork callback. */
216 static unsigned long have_canfork_callback __read_mostly;
218 static struct file_system_type cgroup2_fs_type;
219 static struct cftype cgroup_dfl_base_files[];
220 static struct cftype cgroup_legacy_base_files[];
222 static int rebind_subsystems(struct cgroup_root *dst_root,
223 unsigned long ss_mask);
224 static void css_task_iter_advance(struct css_task_iter *it);
225 static int cgroup_destroy_locked(struct cgroup *cgrp);
226 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
228 static void css_release(struct percpu_ref *ref);
229 static void kill_css(struct cgroup_subsys_state *css);
230 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
231 struct cgroup *cgrp, struct cftype cfts[],
235 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
236 * @ssid: subsys ID of interest
238 * cgroup_subsys_enabled() can only be used with literal subsys names which
239 * is fine for individual subsystems but unsuitable for cgroup core. This
240 * is slower static_key_enabled() based test indexed by @ssid.
242 static bool cgroup_ssid_enabled(int ssid)
244 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
247 static bool cgroup_ssid_no_v1(int ssid)
249 return cgroup_no_v1_mask & (1 << ssid);
253 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
254 * @cgrp: the cgroup of interest
256 * The default hierarchy is the v2 interface of cgroup and this function
257 * can be used to test whether a cgroup is on the default hierarchy for
258 * cases where a subsystem should behave differnetly depending on the
261 * The set of behaviors which change on the default hierarchy are still
262 * being determined and the mount option is prefixed with __DEVEL__.
264 * List of changed behaviors:
266 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
267 * and "name" are disallowed.
269 * - When mounting an existing superblock, mount options should match.
271 * - Remount is disallowed.
273 * - rename(2) is disallowed.
275 * - "tasks" is removed. Everything should be at process granularity. Use
276 * "cgroup.procs" instead.
278 * - "cgroup.procs" is not sorted. pids will be unique unless they got
279 * recycled inbetween reads.
281 * - "release_agent" and "notify_on_release" are removed. Replacement
282 * notification mechanism will be implemented.
284 * - "cgroup.clone_children" is removed.
286 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
287 * and its descendants contain no task; otherwise, 1. The file also
288 * generates kernfs notification which can be monitored through poll and
289 * [di]notify when the value of the file changes.
291 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
292 * take masks of ancestors with non-empty cpus/mems, instead of being
293 * moved to an ancestor.
295 * - cpuset: a task can be moved into an empty cpuset, and again it takes
296 * masks of ancestors.
298 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
301 * - blkcg: blk-throttle becomes properly hierarchical.
303 * - debug: disallowed on the default hierarchy.
305 static bool cgroup_on_dfl(const struct cgroup *cgrp)
307 return cgrp->root == &cgrp_dfl_root;
310 /* IDR wrappers which synchronize using cgroup_idr_lock */
311 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
316 idr_preload(gfp_mask);
317 spin_lock_bh(&cgroup_idr_lock);
318 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
319 spin_unlock_bh(&cgroup_idr_lock);
324 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
328 spin_lock_bh(&cgroup_idr_lock);
329 ret = idr_replace(idr, ptr, id);
330 spin_unlock_bh(&cgroup_idr_lock);
334 static void cgroup_idr_remove(struct idr *idr, int id)
336 spin_lock_bh(&cgroup_idr_lock);
338 spin_unlock_bh(&cgroup_idr_lock);
341 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
343 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
346 return container_of(parent_css, struct cgroup, self);
351 * cgroup_css - obtain a cgroup's css for the specified subsystem
352 * @cgrp: the cgroup of interest
353 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
355 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
356 * function must be called either under cgroup_mutex or rcu_read_lock() and
357 * the caller is responsible for pinning the returned css if it wants to
358 * keep accessing it outside the said locks. This function may return
359 * %NULL if @cgrp doesn't have @subsys_id enabled.
361 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
362 struct cgroup_subsys *ss)
365 return rcu_dereference_check(cgrp->subsys[ss->id],
366 lockdep_is_held(&cgroup_mutex));
372 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
373 * @cgrp: the cgroup of interest
374 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
376 * Similar to cgroup_css() but returns the effective css, which is defined
377 * as the matching css of the nearest ancestor including self which has @ss
378 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
379 * function is guaranteed to return non-NULL css.
381 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
382 struct cgroup_subsys *ss)
384 lockdep_assert_held(&cgroup_mutex);
389 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
393 * This function is used while updating css associations and thus
394 * can't test the csses directly. Use ->subtree_ss_mask.
396 while (cgroup_parent(cgrp) &&
397 !(cgroup_parent(cgrp)->subtree_ss_mask & (1 << ss->id)))
398 cgrp = cgroup_parent(cgrp);
400 return cgroup_css(cgrp, ss);
404 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
405 * @cgrp: the cgroup of interest
406 * @ss: the subsystem of interest
408 * Find and get the effective css of @cgrp for @ss. The effective css is
409 * defined as the matching css of the nearest ancestor including self which
410 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
411 * the root css is returned, so this function always returns a valid css.
412 * The returned css must be put using css_put().
414 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
415 struct cgroup_subsys *ss)
417 struct cgroup_subsys_state *css;
422 css = cgroup_css(cgrp, ss);
424 if (css && css_tryget_online(css))
426 cgrp = cgroup_parent(cgrp);
429 css = init_css_set.subsys[ss->id];
436 /* convenient tests for these bits */
437 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
439 return !(cgrp->self.flags & CSS_ONLINE);
442 static void cgroup_get(struct cgroup *cgrp)
444 WARN_ON_ONCE(cgroup_is_dead(cgrp));
445 css_get(&cgrp->self);
448 static bool cgroup_tryget(struct cgroup *cgrp)
450 return css_tryget(&cgrp->self);
453 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
455 struct cgroup *cgrp = of->kn->parent->priv;
456 struct cftype *cft = of_cft(of);
459 * This is open and unprotected implementation of cgroup_css().
460 * seq_css() is only called from a kernfs file operation which has
461 * an active reference on the file. Because all the subsystem
462 * files are drained before a css is disassociated with a cgroup,
463 * the matching css from the cgroup's subsys table is guaranteed to
464 * be and stay valid until the enclosing operation is complete.
467 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
471 EXPORT_SYMBOL_GPL(of_css);
473 static int notify_on_release(const struct cgroup *cgrp)
475 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
479 * for_each_css - iterate all css's of a cgroup
480 * @css: the iteration cursor
481 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
482 * @cgrp: the target cgroup to iterate css's of
484 * Should be called under cgroup_[tree_]mutex.
486 #define for_each_css(css, ssid, cgrp) \
487 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
488 if (!((css) = rcu_dereference_check( \
489 (cgrp)->subsys[(ssid)], \
490 lockdep_is_held(&cgroup_mutex)))) { } \
494 * for_each_e_css - iterate all effective css's of a cgroup
495 * @css: the iteration cursor
496 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
497 * @cgrp: the target cgroup to iterate css's of
499 * Should be called under cgroup_[tree_]mutex.
501 #define for_each_e_css(css, ssid, cgrp) \
502 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
503 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
508 * for_each_subsys - iterate all enabled cgroup subsystems
509 * @ss: the iteration cursor
510 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
512 #define for_each_subsys(ss, ssid) \
513 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
514 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
517 * for_each_subsys_which - filter for_each_subsys with a bitmask
518 * @ss: the iteration cursor
519 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
520 * @ss_maskp: a pointer to the bitmask
522 * The block will only run for cases where the ssid-th bit (1 << ssid) of
525 #define for_each_subsys_which(ss, ssid, ss_maskp) \
526 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
529 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
530 if (((ss) = cgroup_subsys[ssid]) && false) \
534 /* iterate across the hierarchies */
535 #define for_each_root(root) \
536 list_for_each_entry((root), &cgroup_roots, root_list)
538 /* iterate over child cgrps, lock should be held throughout iteration */
539 #define cgroup_for_each_live_child(child, cgrp) \
540 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
541 if (({ lockdep_assert_held(&cgroup_mutex); \
542 cgroup_is_dead(child); })) \
546 static void cgroup_release_agent(struct work_struct *work);
547 static void check_for_release(struct cgroup *cgrp);
550 * A cgroup can be associated with multiple css_sets as different tasks may
551 * belong to different cgroups on different hierarchies. In the other
552 * direction, a css_set is naturally associated with multiple cgroups.
553 * This M:N relationship is represented by the following link structure
554 * which exists for each association and allows traversing the associations
557 struct cgrp_cset_link {
558 /* the cgroup and css_set this link associates */
560 struct css_set *cset;
562 /* list of cgrp_cset_links anchored at cgrp->cset_links */
563 struct list_head cset_link;
565 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
566 struct list_head cgrp_link;
570 * The default css_set - used by init and its children prior to any
571 * hierarchies being mounted. It contains a pointer to the root state
572 * for each subsystem. Also used to anchor the list of css_sets. Not
573 * reference-counted, to improve performance when child cgroups
574 * haven't been created.
576 struct css_set init_css_set = {
577 .refcount = ATOMIC_INIT(1),
578 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
579 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
580 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
581 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
582 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
583 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
586 static int css_set_count = 1; /* 1 for init_css_set */
589 * css_set_populated - does a css_set contain any tasks?
590 * @cset: target css_set
592 static bool css_set_populated(struct css_set *cset)
594 lockdep_assert_held(&css_set_lock);
596 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
600 * cgroup_update_populated - updated populated count of a cgroup
601 * @cgrp: the target cgroup
602 * @populated: inc or dec populated count
604 * One of the css_sets associated with @cgrp is either getting its first
605 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
606 * count is propagated towards root so that a given cgroup's populated_cnt
607 * is zero iff the cgroup and all its descendants don't contain any tasks.
609 * @cgrp's interface file "cgroup.populated" is zero if
610 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
611 * changes from or to zero, userland is notified that the content of the
612 * interface file has changed. This can be used to detect when @cgrp and
613 * its descendants become populated or empty.
615 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
617 lockdep_assert_held(&css_set_lock);
623 trigger = !cgrp->populated_cnt++;
625 trigger = !--cgrp->populated_cnt;
630 check_for_release(cgrp);
631 cgroup_file_notify(&cgrp->events_file);
633 cgrp = cgroup_parent(cgrp);
638 * css_set_update_populated - update populated state of a css_set
639 * @cset: target css_set
640 * @populated: whether @cset is populated or depopulated
642 * @cset is either getting the first task or losing the last. Update the
643 * ->populated_cnt of all associated cgroups accordingly.
645 static void css_set_update_populated(struct css_set *cset, bool populated)
647 struct cgrp_cset_link *link;
649 lockdep_assert_held(&css_set_lock);
651 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
652 cgroup_update_populated(link->cgrp, populated);
656 * css_set_move_task - move a task from one css_set to another
657 * @task: task being moved
658 * @from_cset: css_set @task currently belongs to (may be NULL)
659 * @to_cset: new css_set @task is being moved to (may be NULL)
660 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
662 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
663 * css_set, @from_cset can be NULL. If @task is being disassociated
664 * instead of moved, @to_cset can be NULL.
666 * This function automatically handles populated_cnt updates and
667 * css_task_iter adjustments but the caller is responsible for managing
668 * @from_cset and @to_cset's reference counts.
670 static void css_set_move_task(struct task_struct *task,
671 struct css_set *from_cset, struct css_set *to_cset,
674 lockdep_assert_held(&css_set_lock);
677 struct css_task_iter *it, *pos;
679 WARN_ON_ONCE(list_empty(&task->cg_list));
682 * @task is leaving, advance task iterators which are
683 * pointing to it so that they can resume at the next
684 * position. Advancing an iterator might remove it from
685 * the list, use safe walk. See css_task_iter_advance*()
688 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
690 if (it->task_pos == &task->cg_list)
691 css_task_iter_advance(it);
693 list_del_init(&task->cg_list);
694 if (!css_set_populated(from_cset))
695 css_set_update_populated(from_cset, false);
697 WARN_ON_ONCE(!list_empty(&task->cg_list));
702 * We are synchronized through cgroup_threadgroup_rwsem
703 * against PF_EXITING setting such that we can't race
704 * against cgroup_exit() changing the css_set to
705 * init_css_set and dropping the old one.
707 WARN_ON_ONCE(task->flags & PF_EXITING);
709 if (!css_set_populated(to_cset))
710 css_set_update_populated(to_cset, true);
711 rcu_assign_pointer(task->cgroups, to_cset);
712 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
718 * hash table for cgroup groups. This improves the performance to find
719 * an existing css_set. This hash doesn't (currently) take into
720 * account cgroups in empty hierarchies.
722 #define CSS_SET_HASH_BITS 7
723 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
725 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
727 unsigned long key = 0UL;
728 struct cgroup_subsys *ss;
731 for_each_subsys(ss, i)
732 key += (unsigned long)css[i];
733 key = (key >> 16) ^ key;
738 static void put_css_set_locked(struct css_set *cset)
740 struct cgrp_cset_link *link, *tmp_link;
741 struct cgroup_subsys *ss;
744 lockdep_assert_held(&css_set_lock);
746 if (!atomic_dec_and_test(&cset->refcount))
749 /* This css_set is dead. unlink it and release cgroup and css refs */
750 for_each_subsys(ss, ssid) {
751 list_del(&cset->e_cset_node[ssid]);
752 css_put(cset->subsys[ssid]);
754 hash_del(&cset->hlist);
757 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
758 list_del(&link->cset_link);
759 list_del(&link->cgrp_link);
760 if (cgroup_parent(link->cgrp))
761 cgroup_put(link->cgrp);
765 kfree_rcu(cset, rcu_head);
768 static void put_css_set(struct css_set *cset)
771 * Ensure that the refcount doesn't hit zero while any readers
772 * can see it. Similar to atomic_dec_and_lock(), but for an
775 if (atomic_add_unless(&cset->refcount, -1, 1))
778 spin_lock_bh(&css_set_lock);
779 put_css_set_locked(cset);
780 spin_unlock_bh(&css_set_lock);
784 * refcounted get/put for css_set objects
786 static inline void get_css_set(struct css_set *cset)
788 atomic_inc(&cset->refcount);
792 * compare_css_sets - helper function for find_existing_css_set().
793 * @cset: candidate css_set being tested
794 * @old_cset: existing css_set for a task
795 * @new_cgrp: cgroup that's being entered by the task
796 * @template: desired set of css pointers in css_set (pre-calculated)
798 * Returns true if "cset" matches "old_cset" except for the hierarchy
799 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
801 static bool compare_css_sets(struct css_set *cset,
802 struct css_set *old_cset,
803 struct cgroup *new_cgrp,
804 struct cgroup_subsys_state *template[])
806 struct list_head *l1, *l2;
809 * On the default hierarchy, there can be csets which are
810 * associated with the same set of cgroups but different csses.
811 * Let's first ensure that csses match.
813 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
817 * Compare cgroup pointers in order to distinguish between
818 * different cgroups in hierarchies. As different cgroups may
819 * share the same effective css, this comparison is always
822 l1 = &cset->cgrp_links;
823 l2 = &old_cset->cgrp_links;
825 struct cgrp_cset_link *link1, *link2;
826 struct cgroup *cgrp1, *cgrp2;
830 /* See if we reached the end - both lists are equal length. */
831 if (l1 == &cset->cgrp_links) {
832 BUG_ON(l2 != &old_cset->cgrp_links);
835 BUG_ON(l2 == &old_cset->cgrp_links);
837 /* Locate the cgroups associated with these links. */
838 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
839 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
842 /* Hierarchies should be linked in the same order. */
843 BUG_ON(cgrp1->root != cgrp2->root);
846 * If this hierarchy is the hierarchy of the cgroup
847 * that's changing, then we need to check that this
848 * css_set points to the new cgroup; if it's any other
849 * hierarchy, then this css_set should point to the
850 * same cgroup as the old css_set.
852 if (cgrp1->root == new_cgrp->root) {
853 if (cgrp1 != new_cgrp)
864 * find_existing_css_set - init css array and find the matching css_set
865 * @old_cset: the css_set that we're using before the cgroup transition
866 * @cgrp: the cgroup that we're moving into
867 * @template: out param for the new set of csses, should be clear on entry
869 static struct css_set *find_existing_css_set(struct css_set *old_cset,
871 struct cgroup_subsys_state *template[])
873 struct cgroup_root *root = cgrp->root;
874 struct cgroup_subsys *ss;
875 struct css_set *cset;
880 * Build the set of subsystem state objects that we want to see in the
881 * new css_set. while subsystems can change globally, the entries here
882 * won't change, so no need for locking.
884 for_each_subsys(ss, i) {
885 if (root->subsys_mask & (1UL << i)) {
887 * @ss is in this hierarchy, so we want the
888 * effective css from @cgrp.
890 template[i] = cgroup_e_css(cgrp, ss);
893 * @ss is not in this hierarchy, so we don't want
896 template[i] = old_cset->subsys[i];
900 key = css_set_hash(template);
901 hash_for_each_possible(css_set_table, cset, hlist, key) {
902 if (!compare_css_sets(cset, old_cset, cgrp, template))
905 /* This css_set matches what we need */
909 /* No existing cgroup group matched */
913 static void free_cgrp_cset_links(struct list_head *links_to_free)
915 struct cgrp_cset_link *link, *tmp_link;
917 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
918 list_del(&link->cset_link);
924 * allocate_cgrp_cset_links - allocate cgrp_cset_links
925 * @count: the number of links to allocate
926 * @tmp_links: list_head the allocated links are put on
928 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
929 * through ->cset_link. Returns 0 on success or -errno.
931 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
933 struct cgrp_cset_link *link;
936 INIT_LIST_HEAD(tmp_links);
938 for (i = 0; i < count; i++) {
939 link = kzalloc(sizeof(*link), GFP_KERNEL);
941 free_cgrp_cset_links(tmp_links);
944 list_add(&link->cset_link, tmp_links);
950 * link_css_set - a helper function to link a css_set to a cgroup
951 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
952 * @cset: the css_set to be linked
953 * @cgrp: the destination cgroup
955 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
958 struct cgrp_cset_link *link;
960 BUG_ON(list_empty(tmp_links));
962 if (cgroup_on_dfl(cgrp))
963 cset->dfl_cgrp = cgrp;
965 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
970 * Always add links to the tail of the lists so that the lists are
971 * in choronological order.
973 list_move_tail(&link->cset_link, &cgrp->cset_links);
974 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
976 if (cgroup_parent(cgrp))
981 * find_css_set - return a new css_set with one cgroup updated
982 * @old_cset: the baseline css_set
983 * @cgrp: the cgroup to be updated
985 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
986 * substituted into the appropriate hierarchy.
988 static struct css_set *find_css_set(struct css_set *old_cset,
991 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
992 struct css_set *cset;
993 struct list_head tmp_links;
994 struct cgrp_cset_link *link;
995 struct cgroup_subsys *ss;
999 lockdep_assert_held(&cgroup_mutex);
1001 /* First see if we already have a cgroup group that matches
1002 * the desired set */
1003 spin_lock_bh(&css_set_lock);
1004 cset = find_existing_css_set(old_cset, cgrp, template);
1007 spin_unlock_bh(&css_set_lock);
1012 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1016 /* Allocate all the cgrp_cset_link objects that we'll need */
1017 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1022 atomic_set(&cset->refcount, 1);
1023 INIT_LIST_HEAD(&cset->cgrp_links);
1024 INIT_LIST_HEAD(&cset->tasks);
1025 INIT_LIST_HEAD(&cset->mg_tasks);
1026 INIT_LIST_HEAD(&cset->mg_preload_node);
1027 INIT_LIST_HEAD(&cset->mg_node);
1028 INIT_LIST_HEAD(&cset->task_iters);
1029 INIT_HLIST_NODE(&cset->hlist);
1031 /* Copy the set of subsystem state objects generated in
1032 * find_existing_css_set() */
1033 memcpy(cset->subsys, template, sizeof(cset->subsys));
1035 spin_lock_bh(&css_set_lock);
1036 /* Add reference counts and links from the new css_set. */
1037 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1038 struct cgroup *c = link->cgrp;
1040 if (c->root == cgrp->root)
1042 link_css_set(&tmp_links, cset, c);
1045 BUG_ON(!list_empty(&tmp_links));
1049 /* Add @cset to the hash table */
1050 key = css_set_hash(cset->subsys);
1051 hash_add(css_set_table, &cset->hlist, key);
1053 for_each_subsys(ss, ssid) {
1054 struct cgroup_subsys_state *css = cset->subsys[ssid];
1056 list_add_tail(&cset->e_cset_node[ssid],
1057 &css->cgroup->e_csets[ssid]);
1061 spin_unlock_bh(&css_set_lock);
1066 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1068 struct cgroup *root_cgrp = kf_root->kn->priv;
1070 return root_cgrp->root;
1073 static int cgroup_init_root_id(struct cgroup_root *root)
1077 lockdep_assert_held(&cgroup_mutex);
1079 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1083 root->hierarchy_id = id;
1087 static void cgroup_exit_root_id(struct cgroup_root *root)
1089 lockdep_assert_held(&cgroup_mutex);
1091 if (root->hierarchy_id) {
1092 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1093 root->hierarchy_id = 0;
1097 static void cgroup_free_root(struct cgroup_root *root)
1100 /* hierarchy ID should already have been released */
1101 WARN_ON_ONCE(root->hierarchy_id);
1103 idr_destroy(&root->cgroup_idr);
1108 static void cgroup_destroy_root(struct cgroup_root *root)
1110 struct cgroup *cgrp = &root->cgrp;
1111 struct cgrp_cset_link *link, *tmp_link;
1113 mutex_lock(&cgroup_mutex);
1115 BUG_ON(atomic_read(&root->nr_cgrps));
1116 BUG_ON(!list_empty(&cgrp->self.children));
1118 /* Rebind all subsystems back to the default hierarchy */
1119 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1122 * Release all the links from cset_links to this hierarchy's
1125 spin_lock_bh(&css_set_lock);
1127 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1128 list_del(&link->cset_link);
1129 list_del(&link->cgrp_link);
1133 spin_unlock_bh(&css_set_lock);
1135 if (!list_empty(&root->root_list)) {
1136 list_del(&root->root_list);
1137 cgroup_root_count--;
1140 cgroup_exit_root_id(root);
1142 mutex_unlock(&cgroup_mutex);
1144 kernfs_destroy_root(root->kf_root);
1145 cgroup_free_root(root);
1148 /* look up cgroup associated with given css_set on the specified hierarchy */
1149 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1150 struct cgroup_root *root)
1152 struct cgroup *res = NULL;
1154 lockdep_assert_held(&cgroup_mutex);
1155 lockdep_assert_held(&css_set_lock);
1157 if (cset == &init_css_set) {
1160 struct cgrp_cset_link *link;
1162 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1163 struct cgroup *c = link->cgrp;
1165 if (c->root == root) {
1177 * Return the cgroup for "task" from the given hierarchy. Must be
1178 * called with cgroup_mutex and css_set_lock held.
1180 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1181 struct cgroup_root *root)
1184 * No need to lock the task - since we hold cgroup_mutex the
1185 * task can't change groups, so the only thing that can happen
1186 * is that it exits and its css is set back to init_css_set.
1188 return cset_cgroup_from_root(task_css_set(task), root);
1192 * A task must hold cgroup_mutex to modify cgroups.
1194 * Any task can increment and decrement the count field without lock.
1195 * So in general, code holding cgroup_mutex can't rely on the count
1196 * field not changing. However, if the count goes to zero, then only
1197 * cgroup_attach_task() can increment it again. Because a count of zero
1198 * means that no tasks are currently attached, therefore there is no
1199 * way a task attached to that cgroup can fork (the other way to
1200 * increment the count). So code holding cgroup_mutex can safely
1201 * assume that if the count is zero, it will stay zero. Similarly, if
1202 * a task holds cgroup_mutex on a cgroup with zero count, it
1203 * knows that the cgroup won't be removed, as cgroup_rmdir()
1206 * A cgroup can only be deleted if both its 'count' of using tasks
1207 * is zero, and its list of 'children' cgroups is empty. Since all
1208 * tasks in the system use _some_ cgroup, and since there is always at
1209 * least one task in the system (init, pid == 1), therefore, root cgroup
1210 * always has either children cgroups and/or using tasks. So we don't
1211 * need a special hack to ensure that root cgroup cannot be deleted.
1213 * P.S. One more locking exception. RCU is used to guard the
1214 * update of a tasks cgroup pointer by cgroup_attach_task()
1217 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1218 static const struct file_operations proc_cgroupstats_operations;
1220 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1223 struct cgroup_subsys *ss = cft->ss;
1225 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1226 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1227 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1228 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1231 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1236 * cgroup_file_mode - deduce file mode of a control file
1237 * @cft: the control file in question
1239 * S_IRUGO for read, S_IWUSR for write.
1241 static umode_t cgroup_file_mode(const struct cftype *cft)
1245 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1248 if (cft->write_u64 || cft->write_s64 || cft->write) {
1249 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1259 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1260 * @cgrp: the target cgroup
1261 * @subtree_control: the new subtree_control mask to consider
1263 * On the default hierarchy, a subsystem may request other subsystems to be
1264 * enabled together through its ->depends_on mask. In such cases, more
1265 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1267 * This function calculates which subsystems need to be enabled if
1268 * @subtree_control is to be applied to @cgrp. The returned mask is always
1269 * a superset of @subtree_control and follows the usual hierarchy rules.
1271 static unsigned long cgroup_calc_subtree_ss_mask(struct cgroup *cgrp,
1272 unsigned long subtree_control)
1274 struct cgroup *parent = cgroup_parent(cgrp);
1275 unsigned long cur_ss_mask = subtree_control;
1276 struct cgroup_subsys *ss;
1279 lockdep_assert_held(&cgroup_mutex);
1281 if (!cgroup_on_dfl(cgrp))
1285 unsigned long new_ss_mask = cur_ss_mask;
1287 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1288 new_ss_mask |= ss->depends_on;
1291 * Mask out subsystems which aren't available. This can
1292 * happen only if some depended-upon subsystems were bound
1293 * to non-default hierarchies.
1296 new_ss_mask &= parent->subtree_ss_mask;
1298 new_ss_mask &= cgrp->root->subsys_mask;
1300 if (new_ss_mask == cur_ss_mask)
1302 cur_ss_mask = new_ss_mask;
1309 * cgroup_refresh_subtree_ss_mask - update subtree_ss_mask
1310 * @cgrp: the target cgroup
1312 * Update @cgrp->subtree_ss_mask according to the current
1313 * @cgrp->subtree_control using cgroup_calc_subtree_ss_mask().
1315 static void cgroup_refresh_subtree_ss_mask(struct cgroup *cgrp)
1317 cgrp->subtree_ss_mask =
1318 cgroup_calc_subtree_ss_mask(cgrp, cgrp->subtree_control);
1322 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1323 * @kn: the kernfs_node being serviced
1325 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1326 * the method finishes if locking succeeded. Note that once this function
1327 * returns the cgroup returned by cgroup_kn_lock_live() may become
1328 * inaccessible any time. If the caller intends to continue to access the
1329 * cgroup, it should pin it before invoking this function.
1331 static void cgroup_kn_unlock(struct kernfs_node *kn)
1333 struct cgroup *cgrp;
1335 if (kernfs_type(kn) == KERNFS_DIR)
1338 cgrp = kn->parent->priv;
1340 mutex_unlock(&cgroup_mutex);
1342 kernfs_unbreak_active_protection(kn);
1347 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1348 * @kn: the kernfs_node being serviced
1350 * This helper is to be used by a cgroup kernfs method currently servicing
1351 * @kn. It breaks the active protection, performs cgroup locking and
1352 * verifies that the associated cgroup is alive. Returns the cgroup if
1353 * alive; otherwise, %NULL. A successful return should be undone by a
1354 * matching cgroup_kn_unlock() invocation.
1356 * Any cgroup kernfs method implementation which requires locking the
1357 * associated cgroup should use this helper. It avoids nesting cgroup
1358 * locking under kernfs active protection and allows all kernfs operations
1359 * including self-removal.
1361 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1363 struct cgroup *cgrp;
1365 if (kernfs_type(kn) == KERNFS_DIR)
1368 cgrp = kn->parent->priv;
1371 * We're gonna grab cgroup_mutex which nests outside kernfs
1372 * active_ref. cgroup liveliness check alone provides enough
1373 * protection against removal. Ensure @cgrp stays accessible and
1374 * break the active_ref protection.
1376 if (!cgroup_tryget(cgrp))
1378 kernfs_break_active_protection(kn);
1380 mutex_lock(&cgroup_mutex);
1382 if (!cgroup_is_dead(cgrp))
1385 cgroup_kn_unlock(kn);
1389 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1391 char name[CGROUP_FILE_NAME_MAX];
1393 lockdep_assert_held(&cgroup_mutex);
1395 if (cft->file_offset) {
1396 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1397 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1399 spin_lock_irq(&cgroup_file_kn_lock);
1401 spin_unlock_irq(&cgroup_file_kn_lock);
1404 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1408 * css_clear_dir - remove subsys files in a cgroup directory
1410 * @cgrp_override: specify if target cgroup is different from css->cgroup
1412 static void css_clear_dir(struct cgroup_subsys_state *css,
1413 struct cgroup *cgrp_override)
1415 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1416 struct cftype *cfts;
1418 list_for_each_entry(cfts, &css->ss->cfts, node)
1419 cgroup_addrm_files(css, cgrp, cfts, false);
1423 * css_populate_dir - create subsys files in a cgroup directory
1425 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1427 * On failure, no file is added.
1429 static int css_populate_dir(struct cgroup_subsys_state *css,
1430 struct cgroup *cgrp_override)
1432 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1433 struct cftype *cfts, *failed_cfts;
1437 if (cgroup_on_dfl(cgrp))
1438 cfts = cgroup_dfl_base_files;
1440 cfts = cgroup_legacy_base_files;
1442 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1445 list_for_each_entry(cfts, &css->ss->cfts, node) {
1446 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1454 list_for_each_entry(cfts, &css->ss->cfts, node) {
1455 if (cfts == failed_cfts)
1457 cgroup_addrm_files(css, cgrp, cfts, false);
1462 static int rebind_subsystems(struct cgroup_root *dst_root,
1463 unsigned long ss_mask)
1465 struct cgroup *dcgrp = &dst_root->cgrp;
1466 struct cgroup_subsys *ss;
1467 unsigned long tmp_ss_mask;
1470 lockdep_assert_held(&cgroup_mutex);
1472 for_each_subsys_which(ss, ssid, &ss_mask) {
1473 /* if @ss has non-root csses attached to it, can't move */
1474 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1477 /* can't move between two non-dummy roots either */
1478 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1482 /* skip creating root files on dfl_root for inhibited subsystems */
1483 tmp_ss_mask = ss_mask;
1484 if (dst_root == &cgrp_dfl_root)
1485 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1487 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1488 struct cgroup *scgrp = &ss->root->cgrp;
1491 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1496 * Rebinding back to the default root is not allowed to
1497 * fail. Using both default and non-default roots should
1498 * be rare. Moving subsystems back and forth even more so.
1499 * Just warn about it and continue.
1501 if (dst_root == &cgrp_dfl_root) {
1502 if (cgrp_dfl_root_visible) {
1503 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1505 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1510 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1513 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1519 * Nothing can fail from this point on. Remove files for the
1520 * removed subsystems and rebind each subsystem.
1522 for_each_subsys_which(ss, ssid, &ss_mask) {
1523 struct cgroup_root *src_root = ss->root;
1524 struct cgroup *scgrp = &src_root->cgrp;
1525 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1526 struct css_set *cset;
1528 WARN_ON(!css || cgroup_css(dcgrp, ss));
1530 css_clear_dir(css, NULL);
1532 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1533 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1534 ss->root = dst_root;
1535 css->cgroup = dcgrp;
1537 spin_lock_bh(&css_set_lock);
1538 hash_for_each(css_set_table, i, cset, hlist)
1539 list_move_tail(&cset->e_cset_node[ss->id],
1540 &dcgrp->e_csets[ss->id]);
1541 spin_unlock_bh(&css_set_lock);
1543 src_root->subsys_mask &= ~(1 << ssid);
1544 scgrp->subtree_control &= ~(1 << ssid);
1545 cgroup_refresh_subtree_ss_mask(scgrp);
1547 /* default hierarchy doesn't enable controllers by default */
1548 dst_root->subsys_mask |= 1 << ssid;
1549 if (dst_root == &cgrp_dfl_root) {
1550 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1552 dcgrp->subtree_control |= 1 << ssid;
1553 cgroup_refresh_subtree_ss_mask(dcgrp);
1554 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1561 kernfs_activate(dcgrp->kn);
1565 static int cgroup_show_options(struct seq_file *seq,
1566 struct kernfs_root *kf_root)
1568 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1569 struct cgroup_subsys *ss;
1572 if (root != &cgrp_dfl_root)
1573 for_each_subsys(ss, ssid)
1574 if (root->subsys_mask & (1 << ssid))
1575 seq_show_option(seq, ss->legacy_name, NULL);
1576 if (root->flags & CGRP_ROOT_NOPREFIX)
1577 seq_puts(seq, ",noprefix");
1578 if (root->flags & CGRP_ROOT_XATTR)
1579 seq_puts(seq, ",xattr");
1581 spin_lock(&release_agent_path_lock);
1582 if (strlen(root->release_agent_path))
1583 seq_show_option(seq, "release_agent",
1584 root->release_agent_path);
1585 spin_unlock(&release_agent_path_lock);
1587 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1588 seq_puts(seq, ",clone_children");
1589 if (strlen(root->name))
1590 seq_show_option(seq, "name", root->name);
1594 struct cgroup_sb_opts {
1595 unsigned long subsys_mask;
1597 char *release_agent;
1598 bool cpuset_clone_children;
1600 /* User explicitly requested empty subsystem */
1604 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1606 char *token, *o = data;
1607 bool all_ss = false, one_ss = false;
1608 unsigned long mask = -1UL;
1609 struct cgroup_subsys *ss;
1613 #ifdef CONFIG_CPUSETS
1614 mask = ~(1U << cpuset_cgrp_id);
1617 memset(opts, 0, sizeof(*opts));
1619 while ((token = strsep(&o, ",")) != NULL) {
1624 if (!strcmp(token, "none")) {
1625 /* Explicitly have no subsystems */
1629 if (!strcmp(token, "all")) {
1630 /* Mutually exclusive option 'all' + subsystem name */
1636 if (!strcmp(token, "noprefix")) {
1637 opts->flags |= CGRP_ROOT_NOPREFIX;
1640 if (!strcmp(token, "clone_children")) {
1641 opts->cpuset_clone_children = true;
1644 if (!strcmp(token, "xattr")) {
1645 opts->flags |= CGRP_ROOT_XATTR;
1648 if (!strncmp(token, "release_agent=", 14)) {
1649 /* Specifying two release agents is forbidden */
1650 if (opts->release_agent)
1652 opts->release_agent =
1653 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1654 if (!opts->release_agent)
1658 if (!strncmp(token, "name=", 5)) {
1659 const char *name = token + 5;
1660 /* Can't specify an empty name */
1663 /* Must match [\w.-]+ */
1664 for (i = 0; i < strlen(name); i++) {
1668 if ((c == '.') || (c == '-') || (c == '_'))
1672 /* Specifying two names is forbidden */
1675 opts->name = kstrndup(name,
1676 MAX_CGROUP_ROOT_NAMELEN - 1,
1684 for_each_subsys(ss, i) {
1685 if (strcmp(token, ss->legacy_name))
1687 if (!cgroup_ssid_enabled(i))
1689 if (cgroup_ssid_no_v1(i))
1692 /* Mutually exclusive option 'all' + subsystem name */
1695 opts->subsys_mask |= (1 << i);
1700 if (i == CGROUP_SUBSYS_COUNT)
1705 * If the 'all' option was specified select all the subsystems,
1706 * otherwise if 'none', 'name=' and a subsystem name options were
1707 * not specified, let's default to 'all'
1709 if (all_ss || (!one_ss && !opts->none && !opts->name))
1710 for_each_subsys(ss, i)
1711 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1712 opts->subsys_mask |= (1 << i);
1715 * We either have to specify by name or by subsystems. (So all
1716 * empty hierarchies must have a name).
1718 if (!opts->subsys_mask && !opts->name)
1722 * Option noprefix was introduced just for backward compatibility
1723 * with the old cpuset, so we allow noprefix only if mounting just
1724 * the cpuset subsystem.
1726 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1729 /* Can't specify "none" and some subsystems */
1730 if (opts->subsys_mask && opts->none)
1736 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1739 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1740 struct cgroup_sb_opts opts;
1741 unsigned long added_mask, removed_mask;
1743 if (root == &cgrp_dfl_root) {
1744 pr_err("remount is not allowed\n");
1748 mutex_lock(&cgroup_mutex);
1750 /* See what subsystems are wanted */
1751 ret = parse_cgroupfs_options(data, &opts);
1755 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1756 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1757 task_tgid_nr(current), current->comm);
1759 added_mask = opts.subsys_mask & ~root->subsys_mask;
1760 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1762 /* Don't allow flags or name to change at remount */
1763 if ((opts.flags ^ root->flags) ||
1764 (opts.name && strcmp(opts.name, root->name))) {
1765 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1766 opts.flags, opts.name ?: "", root->flags, root->name);
1771 /* remounting is not allowed for populated hierarchies */
1772 if (!list_empty(&root->cgrp.self.children)) {
1777 ret = rebind_subsystems(root, added_mask);
1781 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1783 if (opts.release_agent) {
1784 spin_lock(&release_agent_path_lock);
1785 strcpy(root->release_agent_path, opts.release_agent);
1786 spin_unlock(&release_agent_path_lock);
1789 kfree(opts.release_agent);
1791 mutex_unlock(&cgroup_mutex);
1796 * To reduce the fork() overhead for systems that are not actually using
1797 * their cgroups capability, we don't maintain the lists running through
1798 * each css_set to its tasks until we see the list actually used - in other
1799 * words after the first mount.
1801 static bool use_task_css_set_links __read_mostly;
1803 static void cgroup_enable_task_cg_lists(void)
1805 struct task_struct *p, *g;
1807 spin_lock_bh(&css_set_lock);
1809 if (use_task_css_set_links)
1812 use_task_css_set_links = true;
1815 * We need tasklist_lock because RCU is not safe against
1816 * while_each_thread(). Besides, a forking task that has passed
1817 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1818 * is not guaranteed to have its child immediately visible in the
1819 * tasklist if we walk through it with RCU.
1821 read_lock(&tasklist_lock);
1822 do_each_thread(g, p) {
1823 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1824 task_css_set(p) != &init_css_set);
1827 * We should check if the process is exiting, otherwise
1828 * it will race with cgroup_exit() in that the list
1829 * entry won't be deleted though the process has exited.
1830 * Do it while holding siglock so that we don't end up
1831 * racing against cgroup_exit().
1833 spin_lock_irq(&p->sighand->siglock);
1834 if (!(p->flags & PF_EXITING)) {
1835 struct css_set *cset = task_css_set(p);
1837 if (!css_set_populated(cset))
1838 css_set_update_populated(cset, true);
1839 list_add_tail(&p->cg_list, &cset->tasks);
1842 spin_unlock_irq(&p->sighand->siglock);
1843 } while_each_thread(g, p);
1844 read_unlock(&tasklist_lock);
1846 spin_unlock_bh(&css_set_lock);
1849 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1851 struct cgroup_subsys *ss;
1854 INIT_LIST_HEAD(&cgrp->self.sibling);
1855 INIT_LIST_HEAD(&cgrp->self.children);
1856 INIT_LIST_HEAD(&cgrp->cset_links);
1857 INIT_LIST_HEAD(&cgrp->pidlists);
1858 mutex_init(&cgrp->pidlist_mutex);
1859 cgrp->self.cgroup = cgrp;
1860 cgrp->self.flags |= CSS_ONLINE;
1862 for_each_subsys(ss, ssid)
1863 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1865 init_waitqueue_head(&cgrp->offline_waitq);
1866 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1869 static void init_cgroup_root(struct cgroup_root *root,
1870 struct cgroup_sb_opts *opts)
1872 struct cgroup *cgrp = &root->cgrp;
1874 INIT_LIST_HEAD(&root->root_list);
1875 atomic_set(&root->nr_cgrps, 1);
1877 init_cgroup_housekeeping(cgrp);
1878 idr_init(&root->cgroup_idr);
1880 root->flags = opts->flags;
1881 if (opts->release_agent)
1882 strcpy(root->release_agent_path, opts->release_agent);
1884 strcpy(root->name, opts->name);
1885 if (opts->cpuset_clone_children)
1886 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1889 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1891 LIST_HEAD(tmp_links);
1892 struct cgroup *root_cgrp = &root->cgrp;
1893 struct css_set *cset;
1896 lockdep_assert_held(&cgroup_mutex);
1898 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1901 root_cgrp->id = ret;
1902 root_cgrp->ancestor_ids[0] = ret;
1904 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1910 * We're accessing css_set_count without locking css_set_lock here,
1911 * but that's OK - it can only be increased by someone holding
1912 * cgroup_lock, and that's us. The worst that can happen is that we
1913 * have some link structures left over
1915 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1919 ret = cgroup_init_root_id(root);
1923 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1924 KERNFS_ROOT_CREATE_DEACTIVATED,
1926 if (IS_ERR(root->kf_root)) {
1927 ret = PTR_ERR(root->kf_root);
1930 root_cgrp->kn = root->kf_root->kn;
1932 ret = css_populate_dir(&root_cgrp->self, NULL);
1936 ret = rebind_subsystems(root, ss_mask);
1941 * There must be no failure case after here, since rebinding takes
1942 * care of subsystems' refcounts, which are explicitly dropped in
1943 * the failure exit path.
1945 list_add(&root->root_list, &cgroup_roots);
1946 cgroup_root_count++;
1949 * Link the root cgroup in this hierarchy into all the css_set
1952 spin_lock_bh(&css_set_lock);
1953 hash_for_each(css_set_table, i, cset, hlist) {
1954 link_css_set(&tmp_links, cset, root_cgrp);
1955 if (css_set_populated(cset))
1956 cgroup_update_populated(root_cgrp, true);
1958 spin_unlock_bh(&css_set_lock);
1960 BUG_ON(!list_empty(&root_cgrp->self.children));
1961 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1963 kernfs_activate(root_cgrp->kn);
1968 kernfs_destroy_root(root->kf_root);
1969 root->kf_root = NULL;
1971 cgroup_exit_root_id(root);
1973 percpu_ref_exit(&root_cgrp->self.refcnt);
1975 free_cgrp_cset_links(&tmp_links);
1979 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1980 int flags, const char *unused_dev_name,
1983 bool is_v2 = fs_type == &cgroup2_fs_type;
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();
2002 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2003 return ERR_PTR(-EINVAL);
2005 cgrp_dfl_root_visible = true;
2006 root = &cgrp_dfl_root;
2007 cgroup_get(&root->cgrp);
2011 mutex_lock(&cgroup_mutex);
2013 /* First find the desired set of subsystems */
2014 ret = parse_cgroupfs_options(data, &opts);
2019 * Destruction of cgroup root is asynchronous, so subsystems may
2020 * still be dying after the previous unmount. Let's drain the
2021 * dying subsystems. We just need to ensure that the ones
2022 * unmounted previously finish dying and don't care about new ones
2023 * starting. Testing ref liveliness is good enough.
2025 for_each_subsys(ss, i) {
2026 if (!(opts.subsys_mask & (1 << i)) ||
2027 ss->root == &cgrp_dfl_root)
2030 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2031 mutex_unlock(&cgroup_mutex);
2033 ret = restart_syscall();
2036 cgroup_put(&ss->root->cgrp);
2039 for_each_root(root) {
2040 bool name_match = false;
2042 if (root == &cgrp_dfl_root)
2046 * If we asked for a name then it must match. Also, if
2047 * name matches but sybsys_mask doesn't, we should fail.
2048 * Remember whether name matched.
2051 if (strcmp(opts.name, root->name))
2057 * If we asked for subsystems (or explicitly for no
2058 * subsystems) then they must match.
2060 if ((opts.subsys_mask || opts.none) &&
2061 (opts.subsys_mask != root->subsys_mask)) {
2068 if (root->flags ^ opts.flags)
2069 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2072 * We want to reuse @root whose lifetime is governed by its
2073 * ->cgrp. Let's check whether @root is alive and keep it
2074 * that way. As cgroup_kill_sb() can happen anytime, we
2075 * want to block it by pinning the sb so that @root doesn't
2076 * get killed before mount is complete.
2078 * With the sb pinned, tryget_live can reliably indicate
2079 * whether @root can be reused. If it's being killed,
2080 * drain it. We can use wait_queue for the wait but this
2081 * path is super cold. Let's just sleep a bit and retry.
2083 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2084 if (IS_ERR(pinned_sb) ||
2085 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2086 mutex_unlock(&cgroup_mutex);
2087 if (!IS_ERR_OR_NULL(pinned_sb))
2088 deactivate_super(pinned_sb);
2090 ret = restart_syscall();
2099 * No such thing, create a new one. name= matching without subsys
2100 * specification is allowed for already existing hierarchies but we
2101 * can't create new one without subsys specification.
2103 if (!opts.subsys_mask && !opts.none) {
2108 root = kzalloc(sizeof(*root), GFP_KERNEL);
2114 init_cgroup_root(root, &opts);
2116 ret = cgroup_setup_root(root, opts.subsys_mask);
2118 cgroup_free_root(root);
2121 mutex_unlock(&cgroup_mutex);
2123 kfree(opts.release_agent);
2127 return ERR_PTR(ret);
2129 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2130 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2132 if (IS_ERR(dentry) || !new_sb)
2133 cgroup_put(&root->cgrp);
2136 * If @pinned_sb, we're reusing an existing root and holding an
2137 * extra ref on its sb. Mount is complete. Put the extra ref.
2141 deactivate_super(pinned_sb);
2147 static void cgroup_kill_sb(struct super_block *sb)
2149 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2150 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2153 * If @root doesn't have any mounts or children, start killing it.
2154 * This prevents new mounts by disabling percpu_ref_tryget_live().
2155 * cgroup_mount() may wait for @root's release.
2157 * And don't kill the default root.
2159 if (!list_empty(&root->cgrp.self.children) ||
2160 root == &cgrp_dfl_root)
2161 cgroup_put(&root->cgrp);
2163 percpu_ref_kill(&root->cgrp.self.refcnt);
2168 static struct file_system_type cgroup_fs_type = {
2170 .mount = cgroup_mount,
2171 .kill_sb = cgroup_kill_sb,
2174 static struct file_system_type cgroup2_fs_type = {
2176 .mount = cgroup_mount,
2177 .kill_sb = cgroup_kill_sb,
2181 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2182 * @task: target task
2183 * @buf: the buffer to write the path into
2184 * @buflen: the length of the buffer
2186 * Determine @task's cgroup on the first (the one with the lowest non-zero
2187 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2188 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2189 * cgroup controller callbacks.
2191 * Return value is the same as kernfs_path().
2193 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2195 struct cgroup_root *root;
2196 struct cgroup *cgrp;
2197 int hierarchy_id = 1;
2200 mutex_lock(&cgroup_mutex);
2201 spin_lock_bh(&css_set_lock);
2203 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2206 cgrp = task_cgroup_from_root(task, root);
2207 path = cgroup_path(cgrp, buf, buflen);
2209 /* if no hierarchy exists, everyone is in "/" */
2210 if (strlcpy(buf, "/", buflen) < buflen)
2214 spin_unlock_bh(&css_set_lock);
2215 mutex_unlock(&cgroup_mutex);
2218 EXPORT_SYMBOL_GPL(task_cgroup_path);
2220 /* used to track tasks and other necessary states during migration */
2221 struct cgroup_taskset {
2222 /* the src and dst cset list running through cset->mg_node */
2223 struct list_head src_csets;
2224 struct list_head dst_csets;
2226 /* the subsys currently being processed */
2230 * Fields for cgroup_taskset_*() iteration.
2232 * Before migration is committed, the target migration tasks are on
2233 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2234 * the csets on ->dst_csets. ->csets point to either ->src_csets
2235 * or ->dst_csets depending on whether migration is committed.
2237 * ->cur_csets and ->cur_task point to the current task position
2240 struct list_head *csets;
2241 struct css_set *cur_cset;
2242 struct task_struct *cur_task;
2245 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2246 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2247 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2248 .csets = &tset.src_csets, \
2252 * cgroup_taskset_add - try to add a migration target task to a taskset
2253 * @task: target task
2254 * @tset: target taskset
2256 * Add @task, which is a migration target, to @tset. This function becomes
2257 * noop if @task doesn't need to be migrated. @task's css_set should have
2258 * been added as a migration source and @task->cg_list will be moved from
2259 * the css_set's tasks list to mg_tasks one.
2261 static void cgroup_taskset_add(struct task_struct *task,
2262 struct cgroup_taskset *tset)
2264 struct css_set *cset;
2266 lockdep_assert_held(&css_set_lock);
2268 /* @task either already exited or can't exit until the end */
2269 if (task->flags & PF_EXITING)
2272 /* leave @task alone if post_fork() hasn't linked it yet */
2273 if (list_empty(&task->cg_list))
2276 cset = task_css_set(task);
2277 if (!cset->mg_src_cgrp)
2280 list_move_tail(&task->cg_list, &cset->mg_tasks);
2281 if (list_empty(&cset->mg_node))
2282 list_add_tail(&cset->mg_node, &tset->src_csets);
2283 if (list_empty(&cset->mg_dst_cset->mg_node))
2284 list_move_tail(&cset->mg_dst_cset->mg_node,
2289 * cgroup_taskset_first - reset taskset and return the first task
2290 * @tset: taskset of interest
2291 * @dst_cssp: output variable for the destination css
2293 * @tset iteration is initialized and the first task is returned.
2295 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2296 struct cgroup_subsys_state **dst_cssp)
2298 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2299 tset->cur_task = NULL;
2301 return cgroup_taskset_next(tset, dst_cssp);
2305 * cgroup_taskset_next - iterate to the next task in taskset
2306 * @tset: taskset of interest
2307 * @dst_cssp: output variable for the destination css
2309 * Return the next task in @tset. Iteration must have been initialized
2310 * with cgroup_taskset_first().
2312 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2313 struct cgroup_subsys_state **dst_cssp)
2315 struct css_set *cset = tset->cur_cset;
2316 struct task_struct *task = tset->cur_task;
2318 while (&cset->mg_node != tset->csets) {
2320 task = list_first_entry(&cset->mg_tasks,
2321 struct task_struct, cg_list);
2323 task = list_next_entry(task, cg_list);
2325 if (&task->cg_list != &cset->mg_tasks) {
2326 tset->cur_cset = cset;
2327 tset->cur_task = task;
2330 * This function may be called both before and
2331 * after cgroup_taskset_migrate(). The two cases
2332 * can be distinguished by looking at whether @cset
2333 * has its ->mg_dst_cset set.
2335 if (cset->mg_dst_cset)
2336 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2338 *dst_cssp = cset->subsys[tset->ssid];
2343 cset = list_next_entry(cset, mg_node);
2351 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2352 * @tset: taget taskset
2353 * @dst_cgrp: destination cgroup
2355 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2356 * ->can_attach callbacks fails and guarantees that either all or none of
2357 * the tasks in @tset are migrated. @tset is consumed regardless of
2360 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2361 struct cgroup *dst_cgrp)
2363 struct cgroup_subsys_state *css, *failed_css = NULL;
2364 struct task_struct *task, *tmp_task;
2365 struct css_set *cset, *tmp_cset;
2368 /* methods shouldn't be called if no task is actually migrating */
2369 if (list_empty(&tset->src_csets))
2372 /* check that we can legitimately attach to the cgroup */
2373 for_each_e_css(css, i, dst_cgrp) {
2374 if (css->ss->can_attach) {
2376 ret = css->ss->can_attach(tset);
2379 goto out_cancel_attach;
2385 * Now that we're guaranteed success, proceed to move all tasks to
2386 * the new cgroup. There are no failure cases after here, so this
2387 * is the commit point.
2389 spin_lock_bh(&css_set_lock);
2390 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2391 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2392 struct css_set *from_cset = task_css_set(task);
2393 struct css_set *to_cset = cset->mg_dst_cset;
2395 get_css_set(to_cset);
2396 css_set_move_task(task, from_cset, to_cset, true);
2397 put_css_set_locked(from_cset);
2400 spin_unlock_bh(&css_set_lock);
2403 * Migration is committed, all target tasks are now on dst_csets.
2404 * Nothing is sensitive to fork() after this point. Notify
2405 * controllers that migration is complete.
2407 tset->csets = &tset->dst_csets;
2409 for_each_e_css(css, i, dst_cgrp) {
2410 if (css->ss->attach) {
2412 css->ss->attach(tset);
2417 goto out_release_tset;
2420 for_each_e_css(css, i, dst_cgrp) {
2421 if (css == failed_css)
2423 if (css->ss->cancel_attach) {
2425 css->ss->cancel_attach(tset);
2429 spin_lock_bh(&css_set_lock);
2430 list_splice_init(&tset->dst_csets, &tset->src_csets);
2431 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2432 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2433 list_del_init(&cset->mg_node);
2435 spin_unlock_bh(&css_set_lock);
2440 * cgroup_migrate_finish - cleanup after attach
2441 * @preloaded_csets: list of preloaded css_sets
2443 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2444 * those functions for details.
2446 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2448 struct css_set *cset, *tmp_cset;
2450 lockdep_assert_held(&cgroup_mutex);
2452 spin_lock_bh(&css_set_lock);
2453 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2454 cset->mg_src_cgrp = NULL;
2455 cset->mg_dst_cset = NULL;
2456 list_del_init(&cset->mg_preload_node);
2457 put_css_set_locked(cset);
2459 spin_unlock_bh(&css_set_lock);
2463 * cgroup_migrate_add_src - add a migration source css_set
2464 * @src_cset: the source css_set to add
2465 * @dst_cgrp: the destination cgroup
2466 * @preloaded_csets: list of preloaded css_sets
2468 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2469 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2470 * up by cgroup_migrate_finish().
2472 * This function may be called without holding cgroup_threadgroup_rwsem
2473 * even if the target is a process. Threads may be created and destroyed
2474 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2475 * into play and the preloaded css_sets are guaranteed to cover all
2478 static void cgroup_migrate_add_src(struct css_set *src_cset,
2479 struct cgroup *dst_cgrp,
2480 struct list_head *preloaded_csets)
2482 struct cgroup *src_cgrp;
2484 lockdep_assert_held(&cgroup_mutex);
2485 lockdep_assert_held(&css_set_lock);
2487 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2489 if (!list_empty(&src_cset->mg_preload_node))
2492 WARN_ON(src_cset->mg_src_cgrp);
2493 WARN_ON(!list_empty(&src_cset->mg_tasks));
2494 WARN_ON(!list_empty(&src_cset->mg_node));
2496 src_cset->mg_src_cgrp = src_cgrp;
2497 get_css_set(src_cset);
2498 list_add(&src_cset->mg_preload_node, preloaded_csets);
2502 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2503 * @dst_cgrp: the destination cgroup (may be %NULL)
2504 * @preloaded_csets: list of preloaded source css_sets
2506 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2507 * have been preloaded to @preloaded_csets. This function looks up and
2508 * pins all destination css_sets, links each to its source, and append them
2509 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2510 * source css_set is assumed to be its cgroup on the default hierarchy.
2512 * This function must be called after cgroup_migrate_add_src() has been
2513 * called on each migration source css_set. After migration is performed
2514 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2517 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2518 struct list_head *preloaded_csets)
2521 struct css_set *src_cset, *tmp_cset;
2523 lockdep_assert_held(&cgroup_mutex);
2526 * Except for the root, subtree_ss_mask must be zero for a cgroup
2527 * with tasks so that child cgroups don't compete against tasks.
2529 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2530 dst_cgrp->subtree_ss_mask)
2533 /* look up the dst cset for each src cset and link it to src */
2534 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2535 struct css_set *dst_cset;
2537 dst_cset = find_css_set(src_cset,
2538 dst_cgrp ?: src_cset->dfl_cgrp);
2542 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2545 * If src cset equals dst, it's noop. Drop the src.
2546 * cgroup_migrate() will skip the cset too. Note that we
2547 * can't handle src == dst as some nodes are used by both.
2549 if (src_cset == dst_cset) {
2550 src_cset->mg_src_cgrp = NULL;
2551 list_del_init(&src_cset->mg_preload_node);
2552 put_css_set(src_cset);
2553 put_css_set(dst_cset);
2557 src_cset->mg_dst_cset = dst_cset;
2559 if (list_empty(&dst_cset->mg_preload_node))
2560 list_add(&dst_cset->mg_preload_node, &csets);
2562 put_css_set(dst_cset);
2565 list_splice_tail(&csets, preloaded_csets);
2568 cgroup_migrate_finish(&csets);
2573 * cgroup_migrate - migrate a process or task to a cgroup
2574 * @leader: the leader of the process or the task to migrate
2575 * @threadgroup: whether @leader points to the whole process or a single task
2576 * @cgrp: the destination cgroup
2578 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2579 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2580 * caller is also responsible for invoking cgroup_migrate_add_src() and
2581 * cgroup_migrate_prepare_dst() on the targets before invoking this
2582 * function and following up with cgroup_migrate_finish().
2584 * As long as a controller's ->can_attach() doesn't fail, this function is
2585 * guaranteed to succeed. This means that, excluding ->can_attach()
2586 * failure, when migrating multiple targets, the success or failure can be
2587 * decided for all targets by invoking group_migrate_prepare_dst() before
2588 * actually starting migrating.
2590 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2591 struct cgroup *cgrp)
2593 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2594 struct task_struct *task;
2597 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2598 * already PF_EXITING could be freed from underneath us unless we
2599 * take an rcu_read_lock.
2601 spin_lock_bh(&css_set_lock);
2605 cgroup_taskset_add(task, &tset);
2608 } while_each_thread(leader, task);
2610 spin_unlock_bh(&css_set_lock);
2612 return cgroup_taskset_migrate(&tset, cgrp);
2616 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2617 * @dst_cgrp: the cgroup to attach to
2618 * @leader: the task or the leader of the threadgroup to be attached
2619 * @threadgroup: attach the whole threadgroup?
2621 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2623 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2624 struct task_struct *leader, bool threadgroup)
2626 LIST_HEAD(preloaded_csets);
2627 struct task_struct *task;
2630 /* look up all src csets */
2631 spin_lock_bh(&css_set_lock);
2635 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2639 } while_each_thread(leader, task);
2641 spin_unlock_bh(&css_set_lock);
2643 /* prepare dst csets and commit */
2644 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2646 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2648 cgroup_migrate_finish(&preloaded_csets);
2652 static int cgroup_procs_write_permission(struct task_struct *task,
2653 struct cgroup *dst_cgrp,
2654 struct kernfs_open_file *of)
2656 const struct cred *cred = current_cred();
2657 const struct cred *tcred = get_task_cred(task);
2661 * even if we're attaching all tasks in the thread group, we only
2662 * need to check permissions on one of them.
2664 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2665 !uid_eq(cred->euid, tcred->uid) &&
2666 !uid_eq(cred->euid, tcred->suid))
2669 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2670 struct super_block *sb = of->file->f_path.dentry->d_sb;
2671 struct cgroup *cgrp;
2672 struct inode *inode;
2674 spin_lock_bh(&css_set_lock);
2675 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2676 spin_unlock_bh(&css_set_lock);
2678 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2679 cgrp = cgroup_parent(cgrp);
2682 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2684 ret = inode_permission(inode, MAY_WRITE);
2694 * Find the task_struct of the task to attach by vpid and pass it along to the
2695 * function to attach either it or all tasks in its threadgroup. Will lock
2696 * cgroup_mutex and threadgroup.
2698 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2699 size_t nbytes, loff_t off, bool threadgroup)
2701 struct task_struct *tsk;
2702 struct cgroup *cgrp;
2706 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2709 cgrp = cgroup_kn_lock_live(of->kn);
2713 percpu_down_write(&cgroup_threadgroup_rwsem);
2716 tsk = find_task_by_vpid(pid);
2719 goto out_unlock_rcu;
2726 tsk = tsk->group_leader;
2729 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2730 * trapped in a cpuset, or RT worker may be born in a cgroup
2731 * with no rt_runtime allocated. Just say no.
2733 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2735 goto out_unlock_rcu;
2738 get_task_struct(tsk);
2741 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2743 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2745 put_task_struct(tsk);
2746 goto out_unlock_threadgroup;
2750 out_unlock_threadgroup:
2751 percpu_up_write(&cgroup_threadgroup_rwsem);
2752 cgroup_kn_unlock(of->kn);
2753 cpuset_post_attach_flush();
2754 return ret ?: nbytes;
2758 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2759 * @from: attach to all cgroups of a given task
2760 * @tsk: the task to be attached
2762 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2764 struct cgroup_root *root;
2767 mutex_lock(&cgroup_mutex);
2768 for_each_root(root) {
2769 struct cgroup *from_cgrp;
2771 if (root == &cgrp_dfl_root)
2774 spin_lock_bh(&css_set_lock);
2775 from_cgrp = task_cgroup_from_root(from, root);
2776 spin_unlock_bh(&css_set_lock);
2778 retval = cgroup_attach_task(from_cgrp, tsk, false);
2782 mutex_unlock(&cgroup_mutex);
2786 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2788 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2789 char *buf, size_t nbytes, loff_t off)
2791 return __cgroup_procs_write(of, buf, nbytes, off, false);
2794 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2795 char *buf, size_t nbytes, loff_t off)
2797 return __cgroup_procs_write(of, buf, nbytes, off, true);
2800 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2801 char *buf, size_t nbytes, loff_t off)
2803 struct cgroup *cgrp;
2805 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2807 cgrp = cgroup_kn_lock_live(of->kn);
2810 spin_lock(&release_agent_path_lock);
2811 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2812 sizeof(cgrp->root->release_agent_path));
2813 spin_unlock(&release_agent_path_lock);
2814 cgroup_kn_unlock(of->kn);
2818 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2820 struct cgroup *cgrp = seq_css(seq)->cgroup;
2822 spin_lock(&release_agent_path_lock);
2823 seq_puts(seq, cgrp->root->release_agent_path);
2824 spin_unlock(&release_agent_path_lock);
2825 seq_putc(seq, '\n');
2829 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2831 seq_puts(seq, "0\n");
2835 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2837 struct cgroup_subsys *ss;
2838 bool printed = false;
2841 for_each_subsys_which(ss, ssid, &ss_mask) {
2844 seq_printf(seq, "%s", ss->name);
2848 seq_putc(seq, '\n');
2851 /* show controllers which are currently attached to the default hierarchy */
2852 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2854 struct cgroup *cgrp = seq_css(seq)->cgroup;
2856 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2857 ~cgrp_dfl_root_inhibit_ss_mask);
2861 /* show controllers which are enabled from the parent */
2862 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2864 struct cgroup *cgrp = seq_css(seq)->cgroup;
2866 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2870 /* show controllers which are enabled for a given cgroup's children */
2871 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2873 struct cgroup *cgrp = seq_css(seq)->cgroup;
2875 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2880 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2881 * @cgrp: root of the subtree to update csses for
2883 * @cgrp's subtree_ss_mask has changed and its subtree's (self excluded)
2884 * css associations need to be updated accordingly. This function looks up
2885 * all css_sets which are attached to the subtree, creates the matching
2886 * updated css_sets and migrates the tasks to the new ones.
2888 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2890 LIST_HEAD(preloaded_csets);
2891 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2892 struct cgroup_subsys_state *css;
2893 struct css_set *src_cset;
2896 lockdep_assert_held(&cgroup_mutex);
2898 percpu_down_write(&cgroup_threadgroup_rwsem);
2900 /* look up all csses currently attached to @cgrp's subtree */
2901 spin_lock_bh(&css_set_lock);
2902 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2903 struct cgrp_cset_link *link;
2905 /* self is not affected by subtree_ss_mask change */
2906 if (css->cgroup == cgrp)
2909 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2910 cgroup_migrate_add_src(link->cset, cgrp,
2913 spin_unlock_bh(&css_set_lock);
2915 /* NULL dst indicates self on default hierarchy */
2916 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2920 spin_lock_bh(&css_set_lock);
2921 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2922 struct task_struct *task, *ntask;
2924 /* src_csets precede dst_csets, break on the first dst_cset */
2925 if (!src_cset->mg_src_cgrp)
2928 /* all tasks in src_csets need to be migrated */
2929 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2930 cgroup_taskset_add(task, &tset);
2932 spin_unlock_bh(&css_set_lock);
2934 ret = cgroup_taskset_migrate(&tset, cgrp);
2936 cgroup_migrate_finish(&preloaded_csets);
2937 percpu_up_write(&cgroup_threadgroup_rwsem);
2941 /* change the enabled child controllers for a cgroup in the default hierarchy */
2942 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2943 char *buf, size_t nbytes,
2946 unsigned long enable = 0, disable = 0;
2947 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2948 struct cgroup *cgrp, *child;
2949 struct cgroup_subsys *ss;
2954 * Parse input - space separated list of subsystem names prefixed
2955 * with either + or -.
2957 buf = strstrip(buf);
2958 while ((tok = strsep(&buf, " "))) {
2959 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2963 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2964 if (!cgroup_ssid_enabled(ssid) ||
2965 strcmp(tok + 1, ss->name))
2969 enable |= 1 << ssid;
2970 disable &= ~(1 << ssid);
2971 } else if (*tok == '-') {
2972 disable |= 1 << ssid;
2973 enable &= ~(1 << ssid);
2979 if (ssid == CGROUP_SUBSYS_COUNT)
2983 cgrp = cgroup_kn_lock_live(of->kn);
2987 for_each_subsys(ss, ssid) {
2988 if (enable & (1 << ssid)) {
2989 if (cgrp->subtree_control & (1 << ssid)) {
2990 enable &= ~(1 << ssid);
2994 /* unavailable or not enabled on the parent? */
2995 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2996 (cgroup_parent(cgrp) &&
2997 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
3001 } else if (disable & (1 << ssid)) {
3002 if (!(cgrp->subtree_control & (1 << ssid))) {
3003 disable &= ~(1 << ssid);
3007 /* a child has it enabled? */
3008 cgroup_for_each_live_child(child, cgrp) {
3009 if (child->subtree_control & (1 << ssid)) {
3017 if (!enable && !disable) {
3023 * Except for the root, subtree_control must be zero for a cgroup
3024 * with tasks so that child cgroups don't compete against tasks.
3026 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3032 * Update subsys masks and calculate what needs to be done. More
3033 * subsystems than specified may need to be enabled or disabled
3034 * depending on subsystem dependencies.
3036 old_sc = cgrp->subtree_control;
3037 old_ss = cgrp->subtree_ss_mask;
3038 new_sc = (old_sc | enable) & ~disable;
3039 new_ss = cgroup_calc_subtree_ss_mask(cgrp, new_sc);
3041 css_enable = ~old_ss & new_ss;
3042 css_disable = old_ss & ~new_ss;
3043 enable |= css_enable;
3044 disable |= css_disable;
3047 * Because css offlining is asynchronous, userland might try to
3048 * re-enable the same controller while the previous instance is
3049 * still around. In such cases, wait till it's gone using
3052 for_each_subsys_which(ss, ssid, &css_enable) {
3053 cgroup_for_each_live_child(child, cgrp) {
3056 if (!cgroup_css(child, ss))
3060 prepare_to_wait(&child->offline_waitq, &wait,
3061 TASK_UNINTERRUPTIBLE);
3062 cgroup_kn_unlock(of->kn);
3064 finish_wait(&child->offline_waitq, &wait);
3067 return restart_syscall();
3071 cgrp->subtree_control = new_sc;
3072 cgrp->subtree_ss_mask = new_ss;
3075 * Create new csses or make the existing ones visible. A css is
3076 * created invisible if it's being implicitly enabled through
3077 * dependency. An invisible css is made visible when the userland
3078 * explicitly enables it.
3080 for_each_subsys(ss, ssid) {
3081 if (!(enable & (1 << ssid)))
3084 cgroup_for_each_live_child(child, cgrp) {
3085 if (css_enable & (1 << ssid))
3086 ret = create_css(child, ss,
3087 cgrp->subtree_control & (1 << ssid));
3089 ret = css_populate_dir(cgroup_css(child, ss),
3097 * At this point, cgroup_e_css() results reflect the new csses
3098 * making the following cgroup_update_dfl_csses() properly update
3099 * css associations of all tasks in the subtree.
3101 ret = cgroup_update_dfl_csses(cgrp);
3106 * All tasks are migrated out of disabled csses. Kill or hide
3107 * them. A css is hidden when the userland requests it to be
3108 * disabled while other subsystems are still depending on it. The
3109 * css must not actively control resources and be in the vanilla
3110 * state if it's made visible again later. Controllers which may
3111 * be depended upon should provide ->css_reset() for this purpose.
3113 for_each_subsys(ss, ssid) {
3114 if (!(disable & (1 << ssid)))
3117 cgroup_for_each_live_child(child, cgrp) {
3118 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3120 if (css_disable & (1 << ssid)) {
3123 css_clear_dir(css, NULL);
3130 kernfs_activate(cgrp->kn);
3133 cgroup_kn_unlock(of->kn);
3134 return ret ?: nbytes;
3137 cgrp->subtree_control = old_sc;
3138 cgrp->subtree_ss_mask = old_ss;
3140 for_each_subsys(ss, ssid) {
3141 if (!(enable & (1 << ssid)))
3144 cgroup_for_each_live_child(child, cgrp) {
3145 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3150 if (css_enable & (1 << ssid))
3153 css_clear_dir(css, NULL);
3159 static int cgroup_events_show(struct seq_file *seq, void *v)
3161 seq_printf(seq, "populated %d\n",
3162 cgroup_is_populated(seq_css(seq)->cgroup));
3166 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3167 size_t nbytes, loff_t off)
3169 struct cgroup *cgrp = of->kn->parent->priv;
3170 struct cftype *cft = of->kn->priv;
3171 struct cgroup_subsys_state *css;
3175 return cft->write(of, buf, nbytes, off);
3178 * kernfs guarantees that a file isn't deleted with operations in
3179 * flight, which means that the matching css is and stays alive and
3180 * doesn't need to be pinned. The RCU locking is not necessary
3181 * either. It's just for the convenience of using cgroup_css().
3184 css = cgroup_css(cgrp, cft->ss);
3187 if (cft->write_u64) {
3188 unsigned long long v;
3189 ret = kstrtoull(buf, 0, &v);
3191 ret = cft->write_u64(css, cft, v);
3192 } else if (cft->write_s64) {
3194 ret = kstrtoll(buf, 0, &v);
3196 ret = cft->write_s64(css, cft, v);
3201 return ret ?: nbytes;
3204 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3206 return seq_cft(seq)->seq_start(seq, ppos);
3209 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3211 return seq_cft(seq)->seq_next(seq, v, ppos);
3214 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3216 seq_cft(seq)->seq_stop(seq, v);
3219 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3221 struct cftype *cft = seq_cft(m);
3222 struct cgroup_subsys_state *css = seq_css(m);
3225 return cft->seq_show(m, arg);
3228 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3229 else if (cft->read_s64)
3230 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3236 static struct kernfs_ops cgroup_kf_single_ops = {
3237 .atomic_write_len = PAGE_SIZE,
3238 .write = cgroup_file_write,
3239 .seq_show = cgroup_seqfile_show,
3242 static struct kernfs_ops cgroup_kf_ops = {
3243 .atomic_write_len = PAGE_SIZE,
3244 .write = cgroup_file_write,
3245 .seq_start = cgroup_seqfile_start,
3246 .seq_next = cgroup_seqfile_next,
3247 .seq_stop = cgroup_seqfile_stop,
3248 .seq_show = cgroup_seqfile_show,
3252 * cgroup_rename - Only allow simple rename of directories in place.
3254 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3255 const char *new_name_str)
3257 struct cgroup *cgrp = kn->priv;
3260 if (kernfs_type(kn) != KERNFS_DIR)
3262 if (kn->parent != new_parent)
3266 * This isn't a proper migration and its usefulness is very
3267 * limited. Disallow on the default hierarchy.
3269 if (cgroup_on_dfl(cgrp))
3273 * We're gonna grab cgroup_mutex which nests outside kernfs
3274 * active_ref. kernfs_rename() doesn't require active_ref
3275 * protection. Break them before grabbing cgroup_mutex.
3277 kernfs_break_active_protection(new_parent);
3278 kernfs_break_active_protection(kn);
3280 mutex_lock(&cgroup_mutex);
3282 ret = kernfs_rename(kn, new_parent, new_name_str);
3284 mutex_unlock(&cgroup_mutex);
3286 kernfs_unbreak_active_protection(kn);
3287 kernfs_unbreak_active_protection(new_parent);
3291 /* set uid and gid of cgroup dirs and files to that of the creator */
3292 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3294 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3295 .ia_uid = current_fsuid(),
3296 .ia_gid = current_fsgid(), };
3298 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3299 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3302 return kernfs_setattr(kn, &iattr);
3305 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3308 char name[CGROUP_FILE_NAME_MAX];
3309 struct kernfs_node *kn;
3310 struct lock_class_key *key = NULL;
3313 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3314 key = &cft->lockdep_key;
3316 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3317 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3322 ret = cgroup_kn_set_ugid(kn);
3328 if (cft->file_offset) {
3329 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3331 spin_lock_irq(&cgroup_file_kn_lock);
3333 spin_unlock_irq(&cgroup_file_kn_lock);
3340 * cgroup_addrm_files - add or remove files to a cgroup directory
3341 * @css: the target css
3342 * @cgrp: the target cgroup (usually css->cgroup)
3343 * @cfts: array of cftypes to be added
3344 * @is_add: whether to add or remove
3346 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3347 * For removals, this function never fails.
3349 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3350 struct cgroup *cgrp, struct cftype cfts[],
3353 struct cftype *cft, *cft_end = NULL;
3356 lockdep_assert_held(&cgroup_mutex);
3359 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3360 /* does cft->flags tell us to skip this file on @cgrp? */
3361 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3363 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3365 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3367 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3371 ret = cgroup_add_file(css, cgrp, cft);
3373 pr_warn("%s: failed to add %s, err=%d\n",
3374 __func__, cft->name, ret);
3380 cgroup_rm_file(cgrp, cft);
3386 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3389 struct cgroup_subsys *ss = cfts[0].ss;
3390 struct cgroup *root = &ss->root->cgrp;
3391 struct cgroup_subsys_state *css;
3394 lockdep_assert_held(&cgroup_mutex);
3396 /* add/rm files for all cgroups created before */
3397 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3398 struct cgroup *cgrp = css->cgroup;
3400 if (cgroup_is_dead(cgrp))
3403 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3409 kernfs_activate(root->kn);
3413 static void cgroup_exit_cftypes(struct cftype *cfts)
3417 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3418 /* free copy for custom atomic_write_len, see init_cftypes() */
3419 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3424 /* revert flags set by cgroup core while adding @cfts */
3425 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3429 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3433 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3434 struct kernfs_ops *kf_ops;
3436 WARN_ON(cft->ss || cft->kf_ops);
3439 kf_ops = &cgroup_kf_ops;
3441 kf_ops = &cgroup_kf_single_ops;
3444 * Ugh... if @cft wants a custom max_write_len, we need to
3445 * make a copy of kf_ops to set its atomic_write_len.
3447 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3448 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3450 cgroup_exit_cftypes(cfts);
3453 kf_ops->atomic_write_len = cft->max_write_len;
3456 cft->kf_ops = kf_ops;
3463 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3465 lockdep_assert_held(&cgroup_mutex);
3467 if (!cfts || !cfts[0].ss)
3470 list_del(&cfts->node);
3471 cgroup_apply_cftypes(cfts, false);
3472 cgroup_exit_cftypes(cfts);
3477 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3478 * @cfts: zero-length name terminated array of cftypes
3480 * Unregister @cfts. Files described by @cfts are removed from all
3481 * existing cgroups and all future cgroups won't have them either. This
3482 * function can be called anytime whether @cfts' subsys is attached or not.
3484 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3487 int cgroup_rm_cftypes(struct cftype *cfts)
3491 mutex_lock(&cgroup_mutex);
3492 ret = cgroup_rm_cftypes_locked(cfts);
3493 mutex_unlock(&cgroup_mutex);
3498 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3499 * @ss: target cgroup subsystem
3500 * @cfts: zero-length name terminated array of cftypes
3502 * Register @cfts to @ss. Files described by @cfts are created for all
3503 * existing cgroups to which @ss is attached and all future cgroups will
3504 * have them too. This function can be called anytime whether @ss is
3507 * Returns 0 on successful registration, -errno on failure. Note that this
3508 * function currently returns 0 as long as @cfts registration is successful
3509 * even if some file creation attempts on existing cgroups fail.
3511 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3515 if (!cgroup_ssid_enabled(ss->id))
3518 if (!cfts || cfts[0].name[0] == '\0')
3521 ret = cgroup_init_cftypes(ss, cfts);
3525 mutex_lock(&cgroup_mutex);
3527 list_add_tail(&cfts->node, &ss->cfts);
3528 ret = cgroup_apply_cftypes(cfts, true);
3530 cgroup_rm_cftypes_locked(cfts);
3532 mutex_unlock(&cgroup_mutex);
3537 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3538 * @ss: target cgroup subsystem
3539 * @cfts: zero-length name terminated array of cftypes
3541 * Similar to cgroup_add_cftypes() but the added files are only used for
3542 * the default hierarchy.
3544 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3548 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3549 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3550 return cgroup_add_cftypes(ss, cfts);
3554 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3555 * @ss: target cgroup subsystem
3556 * @cfts: zero-length name terminated array of cftypes
3558 * Similar to cgroup_add_cftypes() but the added files are only used for
3559 * the legacy hierarchies.
3561 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3565 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3566 cft->flags |= __CFTYPE_NOT_ON_DFL;
3567 return cgroup_add_cftypes(ss, cfts);
3571 * cgroup_file_notify - generate a file modified event for a cgroup_file
3572 * @cfile: target cgroup_file
3574 * @cfile must have been obtained by setting cftype->file_offset.
3576 void cgroup_file_notify(struct cgroup_file *cfile)
3578 unsigned long flags;
3580 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3582 kernfs_notify(cfile->kn);
3583 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3587 * cgroup_task_count - count the number of tasks in a cgroup.
3588 * @cgrp: the cgroup in question
3590 * Return the number of tasks in the cgroup.
3592 static int cgroup_task_count(const struct cgroup *cgrp)
3595 struct cgrp_cset_link *link;
3597 spin_lock_bh(&css_set_lock);
3598 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3599 count += atomic_read(&link->cset->refcount);
3600 spin_unlock_bh(&css_set_lock);
3605 * css_next_child - find the next child of a given css
3606 * @pos: the current position (%NULL to initiate traversal)
3607 * @parent: css whose children to walk
3609 * This function returns the next child of @parent and should be called
3610 * under either cgroup_mutex or RCU read lock. The only requirement is
3611 * that @parent and @pos are accessible. The next sibling is guaranteed to
3612 * be returned regardless of their states.
3614 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3615 * css which finished ->css_online() is guaranteed to be visible in the
3616 * future iterations and will stay visible until the last reference is put.
3617 * A css which hasn't finished ->css_online() or already finished
3618 * ->css_offline() may show up during traversal. It's each subsystem's
3619 * responsibility to synchronize against on/offlining.
3621 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3622 struct cgroup_subsys_state *parent)
3624 struct cgroup_subsys_state *next;
3626 cgroup_assert_mutex_or_rcu_locked();
3629 * @pos could already have been unlinked from the sibling list.
3630 * Once a cgroup is removed, its ->sibling.next is no longer
3631 * updated when its next sibling changes. CSS_RELEASED is set when
3632 * @pos is taken off list, at which time its next pointer is valid,
3633 * and, as releases are serialized, the one pointed to by the next
3634 * pointer is guaranteed to not have started release yet. This
3635 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3636 * critical section, the one pointed to by its next pointer is
3637 * guaranteed to not have finished its RCU grace period even if we
3638 * have dropped rcu_read_lock() inbetween iterations.
3640 * If @pos has CSS_RELEASED set, its next pointer can't be
3641 * dereferenced; however, as each css is given a monotonically
3642 * increasing unique serial number and always appended to the
3643 * sibling list, the next one can be found by walking the parent's
3644 * children until the first css with higher serial number than
3645 * @pos's. While this path can be slower, it happens iff iteration
3646 * races against release and the race window is very small.
3649 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3650 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3651 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3653 list_for_each_entry_rcu(next, &parent->children, sibling)
3654 if (next->serial_nr > pos->serial_nr)
3659 * @next, if not pointing to the head, can be dereferenced and is
3662 if (&next->sibling != &parent->children)
3668 * css_next_descendant_pre - find the next descendant for pre-order walk
3669 * @pos: the current position (%NULL to initiate traversal)
3670 * @root: css whose descendants to walk
3672 * To be used by css_for_each_descendant_pre(). Find the next descendant
3673 * to visit for pre-order traversal of @root's descendants. @root is
3674 * included in the iteration and the first node to be visited.
3676 * While this function requires cgroup_mutex or RCU read locking, it
3677 * doesn't require the whole traversal to be contained in a single critical
3678 * section. This function will return the correct next descendant as long
3679 * as both @pos and @root are accessible and @pos is a descendant of @root.
3681 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3682 * css which finished ->css_online() is guaranteed to be visible in the
3683 * future iterations and will stay visible until the last reference is put.
3684 * A css which hasn't finished ->css_online() or already finished
3685 * ->css_offline() may show up during traversal. It's each subsystem's
3686 * responsibility to synchronize against on/offlining.
3688 struct cgroup_subsys_state *
3689 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3690 struct cgroup_subsys_state *root)
3692 struct cgroup_subsys_state *next;
3694 cgroup_assert_mutex_or_rcu_locked();
3696 /* if first iteration, visit @root */
3700 /* visit the first child if exists */
3701 next = css_next_child(NULL, pos);
3705 /* no child, visit my or the closest ancestor's next sibling */
3706 while (pos != root) {
3707 next = css_next_child(pos, pos->parent);
3717 * css_rightmost_descendant - return the rightmost descendant of a css
3718 * @pos: css of interest
3720 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3721 * is returned. This can be used during pre-order traversal to skip
3724 * While this function requires cgroup_mutex or RCU read locking, it
3725 * doesn't require the whole traversal to be contained in a single critical
3726 * section. This function will return the correct rightmost descendant as
3727 * long as @pos is accessible.
3729 struct cgroup_subsys_state *
3730 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3732 struct cgroup_subsys_state *last, *tmp;
3734 cgroup_assert_mutex_or_rcu_locked();
3738 /* ->prev isn't RCU safe, walk ->next till the end */
3740 css_for_each_child(tmp, last)
3747 static struct cgroup_subsys_state *
3748 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3750 struct cgroup_subsys_state *last;
3754 pos = css_next_child(NULL, pos);
3761 * css_next_descendant_post - find the next descendant for post-order walk
3762 * @pos: the current position (%NULL to initiate traversal)
3763 * @root: css whose descendants to walk
3765 * To be used by css_for_each_descendant_post(). Find the next descendant
3766 * to visit for post-order traversal of @root's descendants. @root is
3767 * included in the iteration and the last node to be visited.
3769 * While this function requires cgroup_mutex or RCU read locking, it
3770 * doesn't require the whole traversal to be contained in a single critical
3771 * section. This function will return the correct next descendant as long
3772 * as both @pos and @cgroup are accessible and @pos is a descendant of
3775 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3776 * css which finished ->css_online() is guaranteed to be visible in the
3777 * future iterations and will stay visible until the last reference is put.
3778 * A css which hasn't finished ->css_online() or already finished
3779 * ->css_offline() may show up during traversal. It's each subsystem's
3780 * responsibility to synchronize against on/offlining.
3782 struct cgroup_subsys_state *
3783 css_next_descendant_post(struct cgroup_subsys_state *pos,
3784 struct cgroup_subsys_state *root)
3786 struct cgroup_subsys_state *next;
3788 cgroup_assert_mutex_or_rcu_locked();
3790 /* if first iteration, visit leftmost descendant which may be @root */
3792 return css_leftmost_descendant(root);
3794 /* if we visited @root, we're done */
3798 /* if there's an unvisited sibling, visit its leftmost descendant */
3799 next = css_next_child(pos, pos->parent);
3801 return css_leftmost_descendant(next);
3803 /* no sibling left, visit parent */
3808 * css_has_online_children - does a css have online children
3809 * @css: the target css
3811 * Returns %true if @css has any online children; otherwise, %false. This
3812 * function can be called from any context but the caller is responsible
3813 * for synchronizing against on/offlining as necessary.
3815 bool css_has_online_children(struct cgroup_subsys_state *css)
3817 struct cgroup_subsys_state *child;
3821 css_for_each_child(child, css) {
3822 if (child->flags & CSS_ONLINE) {
3832 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3833 * @it: the iterator to advance
3835 * Advance @it to the next css_set to walk.
3837 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3839 struct list_head *l = it->cset_pos;
3840 struct cgrp_cset_link *link;
3841 struct css_set *cset;
3843 lockdep_assert_held(&css_set_lock);
3845 /* Advance to the next non-empty css_set */
3848 if (l == it->cset_head) {
3849 it->cset_pos = NULL;
3850 it->task_pos = NULL;
3855 cset = container_of(l, struct css_set,
3856 e_cset_node[it->ss->id]);
3858 link = list_entry(l, struct cgrp_cset_link, cset_link);
3861 } while (!css_set_populated(cset));
3865 if (!list_empty(&cset->tasks))
3866 it->task_pos = cset->tasks.next;
3868 it->task_pos = cset->mg_tasks.next;
3870 it->tasks_head = &cset->tasks;
3871 it->mg_tasks_head = &cset->mg_tasks;
3874 * We don't keep css_sets locked across iteration steps and thus
3875 * need to take steps to ensure that iteration can be resumed after
3876 * the lock is re-acquired. Iteration is performed at two levels -
3877 * css_sets and tasks in them.
3879 * Once created, a css_set never leaves its cgroup lists, so a
3880 * pinned css_set is guaranteed to stay put and we can resume
3881 * iteration afterwards.
3883 * Tasks may leave @cset across iteration steps. This is resolved
3884 * by registering each iterator with the css_set currently being
3885 * walked and making css_set_move_task() advance iterators whose
3886 * next task is leaving.
3889 list_del(&it->iters_node);
3890 put_css_set_locked(it->cur_cset);
3893 it->cur_cset = cset;
3894 list_add(&it->iters_node, &cset->task_iters);
3897 static void css_task_iter_advance(struct css_task_iter *it)
3899 struct list_head *l = it->task_pos;
3901 lockdep_assert_held(&css_set_lock);
3905 * Advance iterator to find next entry. cset->tasks is consumed
3906 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3911 if (l == it->tasks_head)
3912 l = it->mg_tasks_head->next;
3914 if (l == it->mg_tasks_head)
3915 css_task_iter_advance_css_set(it);
3921 * css_task_iter_start - initiate task iteration
3922 * @css: the css to walk tasks of
3923 * @it: the task iterator to use
3925 * Initiate iteration through the tasks of @css. The caller can call
3926 * css_task_iter_next() to walk through the tasks until the function
3927 * returns NULL. On completion of iteration, css_task_iter_end() must be
3930 void css_task_iter_start(struct cgroup_subsys_state *css,
3931 struct css_task_iter *it)
3933 /* no one should try to iterate before mounting cgroups */
3934 WARN_ON_ONCE(!use_task_css_set_links);
3936 memset(it, 0, sizeof(*it));
3938 spin_lock_bh(&css_set_lock);
3943 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3945 it->cset_pos = &css->cgroup->cset_links;
3947 it->cset_head = it->cset_pos;
3949 css_task_iter_advance_css_set(it);
3951 spin_unlock_bh(&css_set_lock);
3955 * css_task_iter_next - return the next task for the iterator
3956 * @it: the task iterator being iterated
3958 * The "next" function for task iteration. @it should have been
3959 * initialized via css_task_iter_start(). Returns NULL when the iteration
3962 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3965 put_task_struct(it->cur_task);
3966 it->cur_task = NULL;
3969 spin_lock_bh(&css_set_lock);
3972 it->cur_task = list_entry(it->task_pos, struct task_struct,
3974 get_task_struct(it->cur_task);
3975 css_task_iter_advance(it);
3978 spin_unlock_bh(&css_set_lock);
3980 return it->cur_task;
3984 * css_task_iter_end - finish task iteration
3985 * @it: the task iterator to finish
3987 * Finish task iteration started by css_task_iter_start().
3989 void css_task_iter_end(struct css_task_iter *it)
3992 spin_lock_bh(&css_set_lock);
3993 list_del(&it->iters_node);
3994 put_css_set_locked(it->cur_cset);
3995 spin_unlock_bh(&css_set_lock);
3999 put_task_struct(it->cur_task);
4003 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4004 * @to: cgroup to which the tasks will be moved
4005 * @from: cgroup in which the tasks currently reside
4007 * Locking rules between cgroup_post_fork() and the migration path
4008 * guarantee that, if a task is forking while being migrated, the new child
4009 * is guaranteed to be either visible in the source cgroup after the
4010 * parent's migration is complete or put into the target cgroup. No task
4011 * can slip out of migration through forking.
4013 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4015 LIST_HEAD(preloaded_csets);
4016 struct cgrp_cset_link *link;
4017 struct css_task_iter it;
4018 struct task_struct *task;
4021 mutex_lock(&cgroup_mutex);
4023 /* all tasks in @from are being moved, all csets are source */
4024 spin_lock_bh(&css_set_lock);
4025 list_for_each_entry(link, &from->cset_links, cset_link)
4026 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4027 spin_unlock_bh(&css_set_lock);
4029 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4034 * Migrate tasks one-by-one until @from is empty. This fails iff
4035 * ->can_attach() fails.
4038 css_task_iter_start(&from->self, &it);
4039 task = css_task_iter_next(&it);
4041 get_task_struct(task);
4042 css_task_iter_end(&it);
4045 ret = cgroup_migrate(task, false, to);
4046 put_task_struct(task);
4048 } while (task && !ret);
4050 cgroup_migrate_finish(&preloaded_csets);
4051 mutex_unlock(&cgroup_mutex);
4056 * Stuff for reading the 'tasks'/'procs' files.
4058 * Reading this file can return large amounts of data if a cgroup has
4059 * *lots* of attached tasks. So it may need several calls to read(),
4060 * but we cannot guarantee that the information we produce is correct
4061 * unless we produce it entirely atomically.
4065 /* which pidlist file are we talking about? */
4066 enum cgroup_filetype {
4072 * A pidlist is a list of pids that virtually represents the contents of one
4073 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4074 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4077 struct cgroup_pidlist {
4079 * used to find which pidlist is wanted. doesn't change as long as
4080 * this particular list stays in the list.
4082 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4085 /* how many elements the above list has */
4087 /* each of these stored in a list by its cgroup */
4088 struct list_head links;
4089 /* pointer to the cgroup we belong to, for list removal purposes */
4090 struct cgroup *owner;
4091 /* for delayed destruction */
4092 struct delayed_work destroy_dwork;
4096 * The following two functions "fix" the issue where there are more pids
4097 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4098 * TODO: replace with a kernel-wide solution to this problem
4100 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4101 static void *pidlist_allocate(int count)
4103 if (PIDLIST_TOO_LARGE(count))
4104 return vmalloc(count * sizeof(pid_t));
4106 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4109 static void pidlist_free(void *p)
4115 * Used to destroy all pidlists lingering waiting for destroy timer. None
4116 * should be left afterwards.
4118 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4120 struct cgroup_pidlist *l, *tmp_l;
4122 mutex_lock(&cgrp->pidlist_mutex);
4123 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4124 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4125 mutex_unlock(&cgrp->pidlist_mutex);
4127 flush_workqueue(cgroup_pidlist_destroy_wq);
4128 BUG_ON(!list_empty(&cgrp->pidlists));
4131 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4133 struct delayed_work *dwork = to_delayed_work(work);
4134 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4136 struct cgroup_pidlist *tofree = NULL;
4138 mutex_lock(&l->owner->pidlist_mutex);
4141 * Destroy iff we didn't get queued again. The state won't change
4142 * as destroy_dwork can only be queued while locked.
4144 if (!delayed_work_pending(dwork)) {
4145 list_del(&l->links);
4146 pidlist_free(l->list);
4147 put_pid_ns(l->key.ns);
4151 mutex_unlock(&l->owner->pidlist_mutex);
4156 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4157 * Returns the number of unique elements.
4159 static int pidlist_uniq(pid_t *list, int length)
4164 * we presume the 0th element is unique, so i starts at 1. trivial
4165 * edge cases first; no work needs to be done for either
4167 if (length == 0 || length == 1)
4169 /* src and dest walk down the list; dest counts unique elements */
4170 for (src = 1; src < length; src++) {
4171 /* find next unique element */
4172 while (list[src] == list[src-1]) {
4177 /* dest always points to where the next unique element goes */
4178 list[dest] = list[src];
4186 * The two pid files - task and cgroup.procs - guaranteed that the result
4187 * is sorted, which forced this whole pidlist fiasco. As pid order is
4188 * different per namespace, each namespace needs differently sorted list,
4189 * making it impossible to use, for example, single rbtree of member tasks
4190 * sorted by task pointer. As pidlists can be fairly large, allocating one
4191 * per open file is dangerous, so cgroup had to implement shared pool of
4192 * pidlists keyed by cgroup and namespace.
4194 * All this extra complexity was caused by the original implementation
4195 * committing to an entirely unnecessary property. In the long term, we
4196 * want to do away with it. Explicitly scramble sort order if on the
4197 * default hierarchy so that no such expectation exists in the new
4200 * Scrambling is done by swapping every two consecutive bits, which is
4201 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4203 static pid_t pid_fry(pid_t pid)
4205 unsigned a = pid & 0x55555555;
4206 unsigned b = pid & 0xAAAAAAAA;
4208 return (a << 1) | (b >> 1);
4211 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4213 if (cgroup_on_dfl(cgrp))
4214 return pid_fry(pid);
4219 static int cmppid(const void *a, const void *b)
4221 return *(pid_t *)a - *(pid_t *)b;
4224 static int fried_cmppid(const void *a, const void *b)
4226 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4229 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4230 enum cgroup_filetype type)
4232 struct cgroup_pidlist *l;
4233 /* don't need task_nsproxy() if we're looking at ourself */
4234 struct pid_namespace *ns = task_active_pid_ns(current);
4236 lockdep_assert_held(&cgrp->pidlist_mutex);
4238 list_for_each_entry(l, &cgrp->pidlists, links)
4239 if (l->key.type == type && l->key.ns == ns)
4245 * find the appropriate pidlist for our purpose (given procs vs tasks)
4246 * returns with the lock on that pidlist already held, and takes care
4247 * of the use count, or returns NULL with no locks held if we're out of
4250 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4251 enum cgroup_filetype type)
4253 struct cgroup_pidlist *l;
4255 lockdep_assert_held(&cgrp->pidlist_mutex);
4257 l = cgroup_pidlist_find(cgrp, type);
4261 /* entry not found; create a new one */
4262 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4266 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4268 /* don't need task_nsproxy() if we're looking at ourself */
4269 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4271 list_add(&l->links, &cgrp->pidlists);
4276 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4278 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4279 struct cgroup_pidlist **lp)
4283 int pid, n = 0; /* used for populating the array */
4284 struct css_task_iter it;
4285 struct task_struct *tsk;
4286 struct cgroup_pidlist *l;
4288 lockdep_assert_held(&cgrp->pidlist_mutex);
4291 * If cgroup gets more users after we read count, we won't have
4292 * enough space - tough. This race is indistinguishable to the
4293 * caller from the case that the additional cgroup users didn't
4294 * show up until sometime later on.
4296 length = cgroup_task_count(cgrp);
4297 array = pidlist_allocate(length);
4300 /* now, populate the array */
4301 css_task_iter_start(&cgrp->self, &it);
4302 while ((tsk = css_task_iter_next(&it))) {
4303 if (unlikely(n == length))
4305 /* get tgid or pid for procs or tasks file respectively */
4306 if (type == CGROUP_FILE_PROCS)
4307 pid = task_tgid_vnr(tsk);
4309 pid = task_pid_vnr(tsk);
4310 if (pid > 0) /* make sure to only use valid results */
4313 css_task_iter_end(&it);
4315 /* now sort & (if procs) strip out duplicates */
4316 if (cgroup_on_dfl(cgrp))
4317 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4319 sort(array, length, sizeof(pid_t), cmppid, NULL);
4320 if (type == CGROUP_FILE_PROCS)
4321 length = pidlist_uniq(array, length);
4323 l = cgroup_pidlist_find_create(cgrp, type);
4325 pidlist_free(array);
4329 /* store array, freeing old if necessary */
4330 pidlist_free(l->list);
4338 * cgroupstats_build - build and fill cgroupstats
4339 * @stats: cgroupstats to fill information into
4340 * @dentry: A dentry entry belonging to the cgroup for which stats have
4343 * Build and fill cgroupstats so that taskstats can export it to user
4346 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4348 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4349 struct cgroup *cgrp;
4350 struct css_task_iter it;
4351 struct task_struct *tsk;
4353 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4354 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4355 kernfs_type(kn) != KERNFS_DIR)
4358 mutex_lock(&cgroup_mutex);
4361 * We aren't being called from kernfs and there's no guarantee on
4362 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4363 * @kn->priv is RCU safe. Let's do the RCU dancing.
4366 cgrp = rcu_dereference(kn->priv);
4367 if (!cgrp || cgroup_is_dead(cgrp)) {
4369 mutex_unlock(&cgroup_mutex);
4374 css_task_iter_start(&cgrp->self, &it);
4375 while ((tsk = css_task_iter_next(&it))) {
4376 switch (tsk->state) {
4378 stats->nr_running++;
4380 case TASK_INTERRUPTIBLE:
4381 stats->nr_sleeping++;
4383 case TASK_UNINTERRUPTIBLE:
4384 stats->nr_uninterruptible++;
4387 stats->nr_stopped++;
4390 if (delayacct_is_task_waiting_on_io(tsk))
4391 stats->nr_io_wait++;
4395 css_task_iter_end(&it);
4397 mutex_unlock(&cgroup_mutex);
4403 * seq_file methods for the tasks/procs files. The seq_file position is the
4404 * next pid to display; the seq_file iterator is a pointer to the pid
4405 * in the cgroup->l->list array.
4408 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4411 * Initially we receive a position value that corresponds to
4412 * one more than the last pid shown (or 0 on the first call or
4413 * after a seek to the start). Use a binary-search to find the
4414 * next pid to display, if any
4416 struct kernfs_open_file *of = s->private;
4417 struct cgroup *cgrp = seq_css(s)->cgroup;
4418 struct cgroup_pidlist *l;
4419 enum cgroup_filetype type = seq_cft(s)->private;
4420 int index = 0, pid = *pos;
4423 mutex_lock(&cgrp->pidlist_mutex);
4426 * !NULL @of->priv indicates that this isn't the first start()
4427 * after open. If the matching pidlist is around, we can use that.
4428 * Look for it. Note that @of->priv can't be used directly. It
4429 * could already have been destroyed.
4432 of->priv = cgroup_pidlist_find(cgrp, type);
4435 * Either this is the first start() after open or the matching
4436 * pidlist has been destroyed inbetween. Create a new one.
4439 ret = pidlist_array_load(cgrp, type,
4440 (struct cgroup_pidlist **)&of->priv);
4442 return ERR_PTR(ret);
4447 int end = l->length;
4449 while (index < end) {
4450 int mid = (index + end) / 2;
4451 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4454 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4460 /* If we're off the end of the array, we're done */
4461 if (index >= l->length)
4463 /* Update the abstract position to be the actual pid that we found */
4464 iter = l->list + index;
4465 *pos = cgroup_pid_fry(cgrp, *iter);
4469 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4471 struct kernfs_open_file *of = s->private;
4472 struct cgroup_pidlist *l = of->priv;
4475 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4476 CGROUP_PIDLIST_DESTROY_DELAY);
4477 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4480 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4482 struct kernfs_open_file *of = s->private;
4483 struct cgroup_pidlist *l = of->priv;
4485 pid_t *end = l->list + l->length;
4487 * Advance to the next pid in the array. If this goes off the
4494 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4499 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4501 seq_printf(s, "%d\n", *(int *)v);
4506 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4509 return notify_on_release(css->cgroup);
4512 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4513 struct cftype *cft, u64 val)
4516 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4518 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4522 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4525 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4528 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4529 struct cftype *cft, u64 val)
4532 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4534 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4538 /* cgroup core interface files for the default hierarchy */
4539 static struct cftype cgroup_dfl_base_files[] = {
4541 .name = "cgroup.procs",
4542 .file_offset = offsetof(struct cgroup, procs_file),
4543 .seq_start = cgroup_pidlist_start,
4544 .seq_next = cgroup_pidlist_next,
4545 .seq_stop = cgroup_pidlist_stop,
4546 .seq_show = cgroup_pidlist_show,
4547 .private = CGROUP_FILE_PROCS,
4548 .write = cgroup_procs_write,
4551 .name = "cgroup.controllers",
4552 .flags = CFTYPE_ONLY_ON_ROOT,
4553 .seq_show = cgroup_root_controllers_show,
4556 .name = "cgroup.controllers",
4557 .flags = CFTYPE_NOT_ON_ROOT,
4558 .seq_show = cgroup_controllers_show,
4561 .name = "cgroup.subtree_control",
4562 .seq_show = cgroup_subtree_control_show,
4563 .write = cgroup_subtree_control_write,
4566 .name = "cgroup.events",
4567 .flags = CFTYPE_NOT_ON_ROOT,
4568 .file_offset = offsetof(struct cgroup, events_file),
4569 .seq_show = cgroup_events_show,
4574 /* cgroup core interface files for the legacy hierarchies */
4575 static struct cftype cgroup_legacy_base_files[] = {
4577 .name = "cgroup.procs",
4578 .seq_start = cgroup_pidlist_start,
4579 .seq_next = cgroup_pidlist_next,
4580 .seq_stop = cgroup_pidlist_stop,
4581 .seq_show = cgroup_pidlist_show,
4582 .private = CGROUP_FILE_PROCS,
4583 .write = cgroup_procs_write,
4586 .name = "cgroup.clone_children",
4587 .read_u64 = cgroup_clone_children_read,
4588 .write_u64 = cgroup_clone_children_write,
4591 .name = "cgroup.sane_behavior",
4592 .flags = CFTYPE_ONLY_ON_ROOT,
4593 .seq_show = cgroup_sane_behavior_show,
4597 .seq_start = cgroup_pidlist_start,
4598 .seq_next = cgroup_pidlist_next,
4599 .seq_stop = cgroup_pidlist_stop,
4600 .seq_show = cgroup_pidlist_show,
4601 .private = CGROUP_FILE_TASKS,
4602 .write = cgroup_tasks_write,
4605 .name = "notify_on_release",
4606 .read_u64 = cgroup_read_notify_on_release,
4607 .write_u64 = cgroup_write_notify_on_release,
4610 .name = "release_agent",
4611 .flags = CFTYPE_ONLY_ON_ROOT,
4612 .seq_show = cgroup_release_agent_show,
4613 .write = cgroup_release_agent_write,
4614 .max_write_len = PATH_MAX - 1,
4620 * css destruction is four-stage process.
4622 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4623 * Implemented in kill_css().
4625 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4626 * and thus css_tryget_online() is guaranteed to fail, the css can be
4627 * offlined by invoking offline_css(). After offlining, the base ref is
4628 * put. Implemented in css_killed_work_fn().
4630 * 3. When the percpu_ref reaches zero, the only possible remaining
4631 * accessors are inside RCU read sections. css_release() schedules the
4634 * 4. After the grace period, the css can be freed. Implemented in
4635 * css_free_work_fn().
4637 * It is actually hairier because both step 2 and 4 require process context
4638 * and thus involve punting to css->destroy_work adding two additional
4639 * steps to the already complex sequence.
4641 static void css_free_work_fn(struct work_struct *work)
4643 struct cgroup_subsys_state *css =
4644 container_of(work, struct cgroup_subsys_state, destroy_work);
4645 struct cgroup_subsys *ss = css->ss;
4646 struct cgroup *cgrp = css->cgroup;
4648 percpu_ref_exit(&css->refcnt);
4652 struct cgroup_subsys_state *parent = css->parent;
4656 cgroup_idr_remove(&ss->css_idr, id);
4662 /* cgroup free path */
4663 atomic_dec(&cgrp->root->nr_cgrps);
4664 cgroup_pidlist_destroy_all(cgrp);
4665 cancel_work_sync(&cgrp->release_agent_work);
4667 if (cgroup_parent(cgrp)) {
4669 * We get a ref to the parent, and put the ref when
4670 * this cgroup is being freed, so it's guaranteed
4671 * that the parent won't be destroyed before its
4674 cgroup_put(cgroup_parent(cgrp));
4675 kernfs_put(cgrp->kn);
4679 * This is root cgroup's refcnt reaching zero,
4680 * which indicates that the root should be
4683 cgroup_destroy_root(cgrp->root);
4688 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4690 struct cgroup_subsys_state *css =
4691 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4693 INIT_WORK(&css->destroy_work, css_free_work_fn);
4694 queue_work(cgroup_destroy_wq, &css->destroy_work);
4697 static void css_release_work_fn(struct work_struct *work)
4699 struct cgroup_subsys_state *css =
4700 container_of(work, struct cgroup_subsys_state, destroy_work);
4701 struct cgroup_subsys *ss = css->ss;
4702 struct cgroup *cgrp = css->cgroup;
4704 mutex_lock(&cgroup_mutex);
4706 css->flags |= CSS_RELEASED;
4707 list_del_rcu(&css->sibling);
4710 /* css release path */
4711 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4712 if (ss->css_released)
4713 ss->css_released(css);
4715 /* cgroup release path */
4716 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4720 * There are two control paths which try to determine
4721 * cgroup from dentry without going through kernfs -
4722 * cgroupstats_build() and css_tryget_online_from_dir().
4723 * Those are supported by RCU protecting clearing of
4724 * cgrp->kn->priv backpointer.
4726 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4729 mutex_unlock(&cgroup_mutex);
4731 call_rcu(&css->rcu_head, css_free_rcu_fn);
4734 static void css_release(struct percpu_ref *ref)
4736 struct cgroup_subsys_state *css =
4737 container_of(ref, struct cgroup_subsys_state, refcnt);
4739 INIT_WORK(&css->destroy_work, css_release_work_fn);
4740 queue_work(cgroup_destroy_wq, &css->destroy_work);
4743 static void init_and_link_css(struct cgroup_subsys_state *css,
4744 struct cgroup_subsys *ss, struct cgroup *cgrp)
4746 lockdep_assert_held(&cgroup_mutex);
4750 memset(css, 0, sizeof(*css));
4753 INIT_LIST_HEAD(&css->sibling);
4754 INIT_LIST_HEAD(&css->children);
4755 css->serial_nr = css_serial_nr_next++;
4756 atomic_set(&css->online_cnt, 0);
4758 if (cgroup_parent(cgrp)) {
4759 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4760 css_get(css->parent);
4763 BUG_ON(cgroup_css(cgrp, ss));
4766 /* invoke ->css_online() on a new CSS and mark it online if successful */
4767 static int online_css(struct cgroup_subsys_state *css)
4769 struct cgroup_subsys *ss = css->ss;
4772 lockdep_assert_held(&cgroup_mutex);
4775 ret = ss->css_online(css);
4777 css->flags |= CSS_ONLINE;
4778 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4780 atomic_inc(&css->online_cnt);
4782 atomic_inc(&css->parent->online_cnt);
4787 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4788 static void offline_css(struct cgroup_subsys_state *css)
4790 struct cgroup_subsys *ss = css->ss;
4792 lockdep_assert_held(&cgroup_mutex);
4794 if (!(css->flags & CSS_ONLINE))
4797 if (ss->css_offline)
4798 ss->css_offline(css);
4800 css->flags &= ~CSS_ONLINE;
4801 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4803 wake_up_all(&css->cgroup->offline_waitq);
4807 * create_css - create a cgroup_subsys_state
4808 * @cgrp: the cgroup new css will be associated with
4809 * @ss: the subsys of new css
4810 * @visible: whether to create control knobs for the new css or not
4812 * Create a new css associated with @cgrp - @ss pair. On success, the new
4813 * css is online and installed in @cgrp with all interface files created if
4814 * @visible. Returns 0 on success, -errno on failure.
4816 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4819 struct cgroup *parent = cgroup_parent(cgrp);
4820 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4821 struct cgroup_subsys_state *css;
4824 lockdep_assert_held(&cgroup_mutex);
4826 css = ss->css_alloc(parent_css);
4828 return PTR_ERR(css);
4830 init_and_link_css(css, ss, cgrp);
4832 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4836 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4838 goto err_free_percpu_ref;
4842 err = css_populate_dir(css, NULL);
4847 /* @css is ready to be brought online now, make it visible */
4848 list_add_tail_rcu(&css->sibling, &parent_css->children);
4849 cgroup_idr_replace(&ss->css_idr, css, css->id);
4851 err = online_css(css);
4855 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4856 cgroup_parent(parent)) {
4857 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4858 current->comm, current->pid, ss->name);
4859 if (!strcmp(ss->name, "memory"))
4860 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4861 ss->warned_broken_hierarchy = true;
4867 list_del_rcu(&css->sibling);
4868 css_clear_dir(css, NULL);
4870 cgroup_idr_remove(&ss->css_idr, css->id);
4871 err_free_percpu_ref:
4872 percpu_ref_exit(&css->refcnt);
4874 call_rcu(&css->rcu_head, css_free_rcu_fn);
4878 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4881 struct cgroup *parent, *cgrp, *tcgrp;
4882 struct cgroup_root *root;
4883 struct cgroup_subsys *ss;
4884 struct kernfs_node *kn;
4885 int level, ssid, ret;
4887 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4889 if (strchr(name, '\n'))
4892 parent = cgroup_kn_lock_live(parent_kn);
4895 root = parent->root;
4896 level = parent->level + 1;
4898 /* allocate the cgroup and its ID, 0 is reserved for the root */
4899 cgrp = kzalloc(sizeof(*cgrp) +
4900 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4906 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4911 * Temporarily set the pointer to NULL, so idr_find() won't return
4912 * a half-baked cgroup.
4914 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4917 goto out_cancel_ref;
4920 init_cgroup_housekeeping(cgrp);
4922 cgrp->self.parent = &parent->self;
4924 cgrp->level = level;
4926 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
4927 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4929 if (notify_on_release(parent))
4930 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4932 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4933 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4935 /* create the directory */
4936 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4944 * This extra ref will be put in cgroup_free_fn() and guarantees
4945 * that @cgrp->kn is always accessible.
4949 cgrp->self.serial_nr = css_serial_nr_next++;
4951 /* allocation complete, commit to creation */
4952 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4953 atomic_inc(&root->nr_cgrps);
4957 * @cgrp is now fully operational. If something fails after this
4958 * point, it'll be released via the normal destruction path.
4960 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4962 ret = cgroup_kn_set_ugid(kn);
4966 ret = css_populate_dir(&cgrp->self, NULL);
4970 /* let's create and online css's */
4971 for_each_subsys(ss, ssid) {
4972 if (parent->subtree_ss_mask & (1 << ssid)) {
4973 ret = create_css(cgrp, ss,
4974 parent->subtree_control & (1 << ssid));
4981 * On the default hierarchy, a child doesn't automatically inherit
4982 * subtree_control from the parent. Each is configured manually.
4984 if (!cgroup_on_dfl(cgrp)) {
4985 cgrp->subtree_control = parent->subtree_control;
4986 cgroup_refresh_subtree_ss_mask(cgrp);
4989 kernfs_activate(kn);
4995 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4997 percpu_ref_exit(&cgrp->self.refcnt);
5001 cgroup_kn_unlock(parent_kn);
5005 cgroup_destroy_locked(cgrp);
5010 * This is called when the refcnt of a css is confirmed to be killed.
5011 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5012 * initate destruction and put the css ref from kill_css().
5014 static void css_killed_work_fn(struct work_struct *work)
5016 struct cgroup_subsys_state *css =
5017 container_of(work, struct cgroup_subsys_state, destroy_work);
5019 mutex_lock(&cgroup_mutex);
5024 /* @css can't go away while we're holding cgroup_mutex */
5026 } while (css && atomic_dec_and_test(&css->online_cnt));
5028 mutex_unlock(&cgroup_mutex);
5031 /* css kill confirmation processing requires process context, bounce */
5032 static void css_killed_ref_fn(struct percpu_ref *ref)
5034 struct cgroup_subsys_state *css =
5035 container_of(ref, struct cgroup_subsys_state, refcnt);
5037 if (atomic_dec_and_test(&css->online_cnt)) {
5038 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5039 queue_work(cgroup_destroy_wq, &css->destroy_work);
5044 * kill_css - destroy a css
5045 * @css: css to destroy
5047 * This function initiates destruction of @css by removing cgroup interface
5048 * files and putting its base reference. ->css_offline() will be invoked
5049 * asynchronously once css_tryget_online() is guaranteed to fail and when
5050 * the reference count reaches zero, @css will be released.
5052 static void kill_css(struct cgroup_subsys_state *css)
5054 lockdep_assert_held(&cgroup_mutex);
5057 * This must happen before css is disassociated with its cgroup.
5058 * See seq_css() for details.
5060 css_clear_dir(css, NULL);
5063 * Killing would put the base ref, but we need to keep it alive
5064 * until after ->css_offline().
5069 * cgroup core guarantees that, by the time ->css_offline() is
5070 * invoked, no new css reference will be given out via
5071 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5072 * proceed to offlining css's because percpu_ref_kill() doesn't
5073 * guarantee that the ref is seen as killed on all CPUs on return.
5075 * Use percpu_ref_kill_and_confirm() to get notifications as each
5076 * css is confirmed to be seen as killed on all CPUs.
5078 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5082 * cgroup_destroy_locked - the first stage of cgroup destruction
5083 * @cgrp: cgroup to be destroyed
5085 * css's make use of percpu refcnts whose killing latency shouldn't be
5086 * exposed to userland and are RCU protected. Also, cgroup core needs to
5087 * guarantee that css_tryget_online() won't succeed by the time
5088 * ->css_offline() is invoked. To satisfy all the requirements,
5089 * destruction is implemented in the following two steps.
5091 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5092 * userland visible parts and start killing the percpu refcnts of
5093 * css's. Set up so that the next stage will be kicked off once all
5094 * the percpu refcnts are confirmed to be killed.
5096 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5097 * rest of destruction. Once all cgroup references are gone, the
5098 * cgroup is RCU-freed.
5100 * This function implements s1. After this step, @cgrp is gone as far as
5101 * the userland is concerned and a new cgroup with the same name may be
5102 * created. As cgroup doesn't care about the names internally, this
5103 * doesn't cause any problem.
5105 static int cgroup_destroy_locked(struct cgroup *cgrp)
5106 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5108 struct cgroup_subsys_state *css;
5111 lockdep_assert_held(&cgroup_mutex);
5114 * Only migration can raise populated from zero and we're already
5115 * holding cgroup_mutex.
5117 if (cgroup_is_populated(cgrp))
5121 * Make sure there's no live children. We can't test emptiness of
5122 * ->self.children as dead children linger on it while being
5123 * drained; otherwise, "rmdir parent/child parent" may fail.
5125 if (css_has_online_children(&cgrp->self))
5129 * Mark @cgrp dead. This prevents further task migration and child
5130 * creation by disabling cgroup_lock_live_group().
5132 cgrp->self.flags &= ~CSS_ONLINE;
5134 /* initiate massacre of all css's */
5135 for_each_css(css, ssid, cgrp)
5139 * Remove @cgrp directory along with the base files. @cgrp has an
5140 * extra ref on its kn.
5142 kernfs_remove(cgrp->kn);
5144 check_for_release(cgroup_parent(cgrp));
5146 /* put the base reference */
5147 percpu_ref_kill(&cgrp->self.refcnt);
5152 static int cgroup_rmdir(struct kernfs_node *kn)
5154 struct cgroup *cgrp;
5157 cgrp = cgroup_kn_lock_live(kn);
5161 ret = cgroup_destroy_locked(cgrp);
5163 cgroup_kn_unlock(kn);
5167 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5168 .remount_fs = cgroup_remount,
5169 .show_options = cgroup_show_options,
5170 .mkdir = cgroup_mkdir,
5171 .rmdir = cgroup_rmdir,
5172 .rename = cgroup_rename,
5175 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5177 struct cgroup_subsys_state *css;
5179 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5181 mutex_lock(&cgroup_mutex);
5183 idr_init(&ss->css_idr);
5184 INIT_LIST_HEAD(&ss->cfts);
5186 /* Create the root cgroup state for this subsystem */
5187 ss->root = &cgrp_dfl_root;
5188 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5189 /* We don't handle early failures gracefully */
5190 BUG_ON(IS_ERR(css));
5191 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5194 * Root csses are never destroyed and we can't initialize
5195 * percpu_ref during early init. Disable refcnting.
5197 css->flags |= CSS_NO_REF;
5200 /* allocation can't be done safely during early init */
5203 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5204 BUG_ON(css->id < 0);
5207 /* Update the init_css_set to contain a subsys
5208 * pointer to this state - since the subsystem is
5209 * newly registered, all tasks and hence the
5210 * init_css_set is in the subsystem's root cgroup. */
5211 init_css_set.subsys[ss->id] = css;
5213 have_fork_callback |= (bool)ss->fork << ss->id;
5214 have_exit_callback |= (bool)ss->exit << ss->id;
5215 have_free_callback |= (bool)ss->free << ss->id;
5216 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5218 /* At system boot, before all subsystems have been
5219 * registered, no tasks have been forked, so we don't
5220 * need to invoke fork callbacks here. */
5221 BUG_ON(!list_empty(&init_task.tasks));
5223 BUG_ON(online_css(css));
5225 mutex_unlock(&cgroup_mutex);
5229 * cgroup_init_early - cgroup initialization at system boot
5231 * Initialize cgroups at system boot, and initialize any
5232 * subsystems that request early init.
5234 int __init cgroup_init_early(void)
5236 static struct cgroup_sb_opts __initdata opts;
5237 struct cgroup_subsys *ss;
5240 init_cgroup_root(&cgrp_dfl_root, &opts);
5241 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5243 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5245 for_each_subsys(ss, i) {
5246 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5247 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5248 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5250 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5251 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5254 ss->name = cgroup_subsys_name[i];
5255 if (!ss->legacy_name)
5256 ss->legacy_name = cgroup_subsys_name[i];
5259 cgroup_init_subsys(ss, true);
5264 static unsigned long cgroup_disable_mask __initdata;
5267 * cgroup_init - cgroup initialization
5269 * Register cgroup filesystem and /proc file, and initialize
5270 * any subsystems that didn't request early init.
5272 int __init cgroup_init(void)
5274 struct cgroup_subsys *ss;
5278 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5279 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5280 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5282 mutex_lock(&cgroup_mutex);
5284 /* Add init_css_set to the hash table */
5285 key = css_set_hash(init_css_set.subsys);
5286 hash_add(css_set_table, &init_css_set.hlist, key);
5288 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5290 mutex_unlock(&cgroup_mutex);
5292 for_each_subsys(ss, ssid) {
5293 if (ss->early_init) {
5294 struct cgroup_subsys_state *css =
5295 init_css_set.subsys[ss->id];
5297 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5299 BUG_ON(css->id < 0);
5301 cgroup_init_subsys(ss, false);
5304 list_add_tail(&init_css_set.e_cset_node[ssid],
5305 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5308 * Setting dfl_root subsys_mask needs to consider the
5309 * disabled flag and cftype registration needs kmalloc,
5310 * both of which aren't available during early_init.
5312 if (cgroup_disable_mask & (1 << ssid)) {
5313 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5314 printk(KERN_INFO "Disabling %s control group subsystem\n",
5319 if (cgroup_ssid_no_v1(ssid))
5320 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5323 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5325 if (!ss->dfl_cftypes)
5326 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5328 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5329 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5331 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5332 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5336 ss->bind(init_css_set.subsys[ssid]);
5339 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5340 WARN_ON(register_filesystem(&cgroup_fs_type));
5341 WARN_ON(register_filesystem(&cgroup2_fs_type));
5342 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5347 static int __init cgroup_wq_init(void)
5350 * There isn't much point in executing destruction path in
5351 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5352 * Use 1 for @max_active.
5354 * We would prefer to do this in cgroup_init() above, but that
5355 * is called before init_workqueues(): so leave this until after.
5357 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5358 BUG_ON(!cgroup_destroy_wq);
5361 * Used to destroy pidlists and separate to serve as flush domain.
5362 * Cap @max_active to 1 too.
5364 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5366 BUG_ON(!cgroup_pidlist_destroy_wq);
5370 core_initcall(cgroup_wq_init);
5373 * proc_cgroup_show()
5374 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5375 * - Used for /proc/<pid>/cgroup.
5377 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5378 struct pid *pid, struct task_struct *tsk)
5382 struct cgroup_root *root;
5385 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5389 mutex_lock(&cgroup_mutex);
5390 spin_lock_bh(&css_set_lock);
5392 for_each_root(root) {
5393 struct cgroup_subsys *ss;
5394 struct cgroup *cgrp;
5395 int ssid, count = 0;
5397 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5400 seq_printf(m, "%d:", root->hierarchy_id);
5401 if (root != &cgrp_dfl_root)
5402 for_each_subsys(ss, ssid)
5403 if (root->subsys_mask & (1 << ssid))
5404 seq_printf(m, "%s%s", count++ ? "," : "",
5406 if (strlen(root->name))
5407 seq_printf(m, "%sname=%s", count ? "," : "",
5411 cgrp = task_cgroup_from_root(tsk, root);
5414 * On traditional hierarchies, all zombie tasks show up as
5415 * belonging to the root cgroup. On the default hierarchy,
5416 * while a zombie doesn't show up in "cgroup.procs" and
5417 * thus can't be migrated, its /proc/PID/cgroup keeps
5418 * reporting the cgroup it belonged to before exiting. If
5419 * the cgroup is removed before the zombie is reaped,
5420 * " (deleted)" is appended to the cgroup path.
5422 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5423 path = cgroup_path(cgrp, buf, PATH_MAX);
5425 retval = -ENAMETOOLONG;
5434 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5435 seq_puts(m, " (deleted)\n");
5442 spin_unlock_bh(&css_set_lock);
5443 mutex_unlock(&cgroup_mutex);
5449 /* Display information about each subsystem and each hierarchy */
5450 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5452 struct cgroup_subsys *ss;
5455 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5457 * ideally we don't want subsystems moving around while we do this.
5458 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5459 * subsys/hierarchy state.
5461 mutex_lock(&cgroup_mutex);
5463 for_each_subsys(ss, i)
5464 seq_printf(m, "%s\t%d\t%d\t%d\n",
5465 ss->legacy_name, ss->root->hierarchy_id,
5466 atomic_read(&ss->root->nr_cgrps),
5467 cgroup_ssid_enabled(i));
5469 mutex_unlock(&cgroup_mutex);
5473 static int cgroupstats_open(struct inode *inode, struct file *file)
5475 return single_open(file, proc_cgroupstats_show, NULL);
5478 static const struct file_operations proc_cgroupstats_operations = {
5479 .open = cgroupstats_open,
5481 .llseek = seq_lseek,
5482 .release = single_release,
5486 * cgroup_fork - initialize cgroup related fields during copy_process()
5487 * @child: pointer to task_struct of forking parent process.
5489 * A task is associated with the init_css_set until cgroup_post_fork()
5490 * attaches it to the parent's css_set. Empty cg_list indicates that
5491 * @child isn't holding reference to its css_set.
5493 void cgroup_fork(struct task_struct *child)
5495 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5496 INIT_LIST_HEAD(&child->cg_list);
5500 * cgroup_can_fork - called on a new task before the process is exposed
5501 * @child: the task in question.
5503 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5504 * returns an error, the fork aborts with that error code. This allows for
5505 * a cgroup subsystem to conditionally allow or deny new forks.
5507 int cgroup_can_fork(struct task_struct *child)
5509 struct cgroup_subsys *ss;
5512 for_each_subsys_which(ss, i, &have_canfork_callback) {
5513 ret = ss->can_fork(child);
5521 for_each_subsys(ss, j) {
5524 if (ss->cancel_fork)
5525 ss->cancel_fork(child);
5532 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5533 * @child: the task in question
5535 * This calls the cancel_fork() callbacks if a fork failed *after*
5536 * cgroup_can_fork() succeded.
5538 void cgroup_cancel_fork(struct task_struct *child)
5540 struct cgroup_subsys *ss;
5543 for_each_subsys(ss, i)
5544 if (ss->cancel_fork)
5545 ss->cancel_fork(child);
5549 * cgroup_post_fork - called on a new task after adding it to the task list
5550 * @child: the task in question
5552 * Adds the task to the list running through its css_set if necessary and
5553 * call the subsystem fork() callbacks. Has to be after the task is
5554 * visible on the task list in case we race with the first call to
5555 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5558 void cgroup_post_fork(struct task_struct *child)
5560 struct cgroup_subsys *ss;
5564 * This may race against cgroup_enable_task_cg_lists(). As that
5565 * function sets use_task_css_set_links before grabbing
5566 * tasklist_lock and we just went through tasklist_lock to add
5567 * @child, it's guaranteed that either we see the set
5568 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5569 * @child during its iteration.
5571 * If we won the race, @child is associated with %current's
5572 * css_set. Grabbing css_set_lock guarantees both that the
5573 * association is stable, and, on completion of the parent's
5574 * migration, @child is visible in the source of migration or
5575 * already in the destination cgroup. This guarantee is necessary
5576 * when implementing operations which need to migrate all tasks of
5577 * a cgroup to another.
5579 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5580 * will remain in init_css_set. This is safe because all tasks are
5581 * in the init_css_set before cg_links is enabled and there's no
5582 * operation which transfers all tasks out of init_css_set.
5584 if (use_task_css_set_links) {
5585 struct css_set *cset;
5587 spin_lock_bh(&css_set_lock);
5588 cset = task_css_set(current);
5589 if (list_empty(&child->cg_list)) {
5591 css_set_move_task(child, NULL, cset, false);
5593 spin_unlock_bh(&css_set_lock);
5597 * Call ss->fork(). This must happen after @child is linked on
5598 * css_set; otherwise, @child might change state between ->fork()
5599 * and addition to css_set.
5601 for_each_subsys_which(ss, i, &have_fork_callback)
5606 * cgroup_exit - detach cgroup from exiting task
5607 * @tsk: pointer to task_struct of exiting process
5609 * Description: Detach cgroup from @tsk and release it.
5611 * Note that cgroups marked notify_on_release force every task in
5612 * them to take the global cgroup_mutex mutex when exiting.
5613 * This could impact scaling on very large systems. Be reluctant to
5614 * use notify_on_release cgroups where very high task exit scaling
5615 * is required on large systems.
5617 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5618 * call cgroup_exit() while the task is still competent to handle
5619 * notify_on_release(), then leave the task attached to the root cgroup in
5620 * each hierarchy for the remainder of its exit. No need to bother with
5621 * init_css_set refcnting. init_css_set never goes away and we can't race
5622 * with migration path - PF_EXITING is visible to migration path.
5624 void cgroup_exit(struct task_struct *tsk)
5626 struct cgroup_subsys *ss;
5627 struct css_set *cset;
5631 * Unlink from @tsk from its css_set. As migration path can't race
5632 * with us, we can check css_set and cg_list without synchronization.
5634 cset = task_css_set(tsk);
5636 if (!list_empty(&tsk->cg_list)) {
5637 spin_lock_bh(&css_set_lock);
5638 css_set_move_task(tsk, cset, NULL, false);
5639 spin_unlock_bh(&css_set_lock);
5644 /* see cgroup_post_fork() for details */
5645 for_each_subsys_which(ss, i, &have_exit_callback)
5649 void cgroup_free(struct task_struct *task)
5651 struct css_set *cset = task_css_set(task);
5652 struct cgroup_subsys *ss;
5655 for_each_subsys_which(ss, ssid, &have_free_callback)
5661 static void check_for_release(struct cgroup *cgrp)
5663 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5664 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5665 schedule_work(&cgrp->release_agent_work);
5669 * Notify userspace when a cgroup is released, by running the
5670 * configured release agent with the name of the cgroup (path
5671 * relative to the root of cgroup file system) as the argument.
5673 * Most likely, this user command will try to rmdir this cgroup.
5675 * This races with the possibility that some other task will be
5676 * attached to this cgroup before it is removed, or that some other
5677 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5678 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5679 * unused, and this cgroup will be reprieved from its death sentence,
5680 * to continue to serve a useful existence. Next time it's released,
5681 * we will get notified again, if it still has 'notify_on_release' set.
5683 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5684 * means only wait until the task is successfully execve()'d. The
5685 * separate release agent task is forked by call_usermodehelper(),
5686 * then control in this thread returns here, without waiting for the
5687 * release agent task. We don't bother to wait because the caller of
5688 * this routine has no use for the exit status of the release agent
5689 * task, so no sense holding our caller up for that.
5691 static void cgroup_release_agent(struct work_struct *work)
5693 struct cgroup *cgrp =
5694 container_of(work, struct cgroup, release_agent_work);
5695 char *pathbuf = NULL, *agentbuf = NULL, *path;
5696 char *argv[3], *envp[3];
5698 mutex_lock(&cgroup_mutex);
5700 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5701 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5702 if (!pathbuf || !agentbuf)
5705 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5713 /* minimal command environment */
5715 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5718 mutex_unlock(&cgroup_mutex);
5719 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5722 mutex_unlock(&cgroup_mutex);
5728 static int __init cgroup_disable(char *str)
5730 struct cgroup_subsys *ss;
5734 while ((token = strsep(&str, ",")) != NULL) {
5738 for_each_subsys(ss, i) {
5739 if (strcmp(token, ss->name) &&
5740 strcmp(token, ss->legacy_name))
5742 cgroup_disable_mask |= 1 << i;
5747 __setup("cgroup_disable=", cgroup_disable);
5749 static int __init cgroup_no_v1(char *str)
5751 struct cgroup_subsys *ss;
5755 while ((token = strsep(&str, ",")) != NULL) {
5759 if (!strcmp(token, "all")) {
5760 cgroup_no_v1_mask = ~0UL;
5764 for_each_subsys(ss, i) {
5765 if (strcmp(token, ss->name) &&
5766 strcmp(token, ss->legacy_name))
5769 cgroup_no_v1_mask |= 1 << i;
5774 __setup("cgroup_no_v1=", cgroup_no_v1);
5777 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5778 * @dentry: directory dentry of interest
5779 * @ss: subsystem of interest
5781 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5782 * to get the corresponding css and return it. If such css doesn't exist
5783 * or can't be pinned, an ERR_PTR value is returned.
5785 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5786 struct cgroup_subsys *ss)
5788 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5789 struct cgroup_subsys_state *css = NULL;
5790 struct cgroup *cgrp;
5792 /* is @dentry a cgroup dir? */
5793 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5794 kernfs_type(kn) != KERNFS_DIR)
5795 return ERR_PTR(-EBADF);
5800 * This path doesn't originate from kernfs and @kn could already
5801 * have been or be removed at any point. @kn->priv is RCU
5802 * protected for this access. See css_release_work_fn() for details.
5804 cgrp = rcu_dereference(kn->priv);
5806 css = cgroup_css(cgrp, ss);
5808 if (!css || !css_tryget_online(css))
5809 css = ERR_PTR(-ENOENT);
5816 * css_from_id - lookup css by id
5817 * @id: the cgroup id
5818 * @ss: cgroup subsys to be looked into
5820 * Returns the css if there's valid one with @id, otherwise returns NULL.
5821 * Should be called under rcu_read_lock().
5823 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5825 WARN_ON_ONCE(!rcu_read_lock_held());
5826 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5830 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5831 * @path: path on the default hierarchy
5833 * Find the cgroup at @path on the default hierarchy, increment its
5834 * reference count and return it. Returns pointer to the found cgroup on
5835 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5836 * if @path points to a non-directory.
5838 struct cgroup *cgroup_get_from_path(const char *path)
5840 struct kernfs_node *kn;
5841 struct cgroup *cgrp;
5843 mutex_lock(&cgroup_mutex);
5845 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5847 if (kernfs_type(kn) == KERNFS_DIR) {
5851 cgrp = ERR_PTR(-ENOTDIR);
5855 cgrp = ERR_PTR(-ENOENT);
5858 mutex_unlock(&cgroup_mutex);
5861 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5864 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5865 * definition in cgroup-defs.h.
5867 #ifdef CONFIG_SOCK_CGROUP_DATA
5869 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5871 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5872 static bool cgroup_sk_alloc_disabled __read_mostly;
5874 void cgroup_sk_alloc_disable(void)
5876 if (cgroup_sk_alloc_disabled)
5878 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5879 cgroup_sk_alloc_disabled = true;
5884 #define cgroup_sk_alloc_disabled false
5888 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5890 if (cgroup_sk_alloc_disabled)
5896 struct css_set *cset;
5898 cset = task_css_set(current);
5899 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5900 skcd->val = (unsigned long)cset->dfl_cgrp;
5909 void cgroup_sk_free(struct sock_cgroup_data *skcd)
5911 cgroup_put(sock_cgroup_ptr(skcd));
5914 #endif /* CONFIG_SOCK_CGROUP_DATA */
5916 #ifdef CONFIG_CGROUP_DEBUG
5917 static struct cgroup_subsys_state *
5918 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5920 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5923 return ERR_PTR(-ENOMEM);
5928 static void debug_css_free(struct cgroup_subsys_state *css)
5933 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5936 return cgroup_task_count(css->cgroup);
5939 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5942 return (u64)(unsigned long)current->cgroups;
5945 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5951 count = atomic_read(&task_css_set(current)->refcount);
5956 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5958 struct cgrp_cset_link *link;
5959 struct css_set *cset;
5962 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5966 spin_lock_bh(&css_set_lock);
5968 cset = rcu_dereference(current->cgroups);
5969 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5970 struct cgroup *c = link->cgrp;
5972 cgroup_name(c, name_buf, NAME_MAX + 1);
5973 seq_printf(seq, "Root %d group %s\n",
5974 c->root->hierarchy_id, name_buf);
5977 spin_unlock_bh(&css_set_lock);
5982 #define MAX_TASKS_SHOWN_PER_CSS 25
5983 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5985 struct cgroup_subsys_state *css = seq_css(seq);
5986 struct cgrp_cset_link *link;
5988 spin_lock_bh(&css_set_lock);
5989 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5990 struct css_set *cset = link->cset;
5991 struct task_struct *task;
5994 seq_printf(seq, "css_set %p\n", cset);
5996 list_for_each_entry(task, &cset->tasks, cg_list) {
5997 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5999 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6002 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6003 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6005 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6009 seq_puts(seq, " ...\n");
6011 spin_unlock_bh(&css_set_lock);
6015 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6017 return (!cgroup_is_populated(css->cgroup) &&
6018 !css_has_online_children(&css->cgroup->self));
6021 static struct cftype debug_files[] = {
6023 .name = "taskcount",
6024 .read_u64 = debug_taskcount_read,
6028 .name = "current_css_set",
6029 .read_u64 = current_css_set_read,
6033 .name = "current_css_set_refcount",
6034 .read_u64 = current_css_set_refcount_read,
6038 .name = "current_css_set_cg_links",
6039 .seq_show = current_css_set_cg_links_read,
6043 .name = "cgroup_css_links",
6044 .seq_show = cgroup_css_links_read,
6048 .name = "releasable",
6049 .read_u64 = releasable_read,
6055 struct cgroup_subsys debug_cgrp_subsys = {
6056 .css_alloc = debug_css_alloc,
6057 .css_free = debug_css_free,
6058 .legacy_cftypes = debug_files,
6060 #endif /* CONFIG_CGROUP_DEBUG */