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_visible;
183 /* Controllers blocked by the commandline in v1 */
184 static u16 cgroup_no_v1_mask;
186 /* some controllers are not supported in the default hierarchy */
187 static u16 cgrp_dfl_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 u16 have_fork_callback __read_mostly;
212 static u16 have_exit_callback __read_mostly;
213 static u16 have_free_callback __read_mostly;
215 /* Ditto for the can_fork callback. */
216 static u16 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, u16 ss_mask);
223 static void css_task_iter_advance(struct css_task_iter *it);
224 static int cgroup_destroy_locked(struct cgroup *cgrp);
225 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
226 struct cgroup_subsys *ss);
227 static void css_release(struct percpu_ref *ref);
228 static void kill_css(struct cgroup_subsys_state *css);
229 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
230 struct cgroup *cgrp, struct cftype cfts[],
234 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
235 * @ssid: subsys ID of interest
237 * cgroup_subsys_enabled() can only be used with literal subsys names which
238 * is fine for individual subsystems but unsuitable for cgroup core. This
239 * is slower static_key_enabled() based test indexed by @ssid.
241 static bool cgroup_ssid_enabled(int ssid)
243 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
246 static bool cgroup_ssid_no_v1(int ssid)
248 return cgroup_no_v1_mask & (1 << ssid);
252 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
253 * @cgrp: the cgroup of interest
255 * The default hierarchy is the v2 interface of cgroup and this function
256 * can be used to test whether a cgroup is on the default hierarchy for
257 * cases where a subsystem should behave differnetly depending on the
260 * The set of behaviors which change on the default hierarchy are still
261 * being determined and the mount option is prefixed with __DEVEL__.
263 * List of changed behaviors:
265 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
266 * and "name" are disallowed.
268 * - When mounting an existing superblock, mount options should match.
270 * - Remount is disallowed.
272 * - rename(2) is disallowed.
274 * - "tasks" is removed. Everything should be at process granularity. Use
275 * "cgroup.procs" instead.
277 * - "cgroup.procs" is not sorted. pids will be unique unless they got
278 * recycled inbetween reads.
280 * - "release_agent" and "notify_on_release" are removed. Replacement
281 * notification mechanism will be implemented.
283 * - "cgroup.clone_children" is removed.
285 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
286 * and its descendants contain no task; otherwise, 1. The file also
287 * generates kernfs notification which can be monitored through poll and
288 * [di]notify when the value of the file changes.
290 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
291 * take masks of ancestors with non-empty cpus/mems, instead of being
292 * moved to an ancestor.
294 * - cpuset: a task can be moved into an empty cpuset, and again it takes
295 * masks of ancestors.
297 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
300 * - blkcg: blk-throttle becomes properly hierarchical.
302 * - debug: disallowed on the default hierarchy.
304 static bool cgroup_on_dfl(const struct cgroup *cgrp)
306 return cgrp->root == &cgrp_dfl_root;
309 /* IDR wrappers which synchronize using cgroup_idr_lock */
310 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
315 idr_preload(gfp_mask);
316 spin_lock_bh(&cgroup_idr_lock);
317 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
318 spin_unlock_bh(&cgroup_idr_lock);
323 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
327 spin_lock_bh(&cgroup_idr_lock);
328 ret = idr_replace(idr, ptr, id);
329 spin_unlock_bh(&cgroup_idr_lock);
333 static void cgroup_idr_remove(struct idr *idr, int id)
335 spin_lock_bh(&cgroup_idr_lock);
337 spin_unlock_bh(&cgroup_idr_lock);
340 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
342 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
345 return container_of(parent_css, struct cgroup, self);
349 /* subsystems visibly enabled on a cgroup */
350 static u16 cgroup_control(struct cgroup *cgrp)
352 struct cgroup *parent = cgroup_parent(cgrp);
353 u16 root_ss_mask = cgrp->root->subsys_mask;
356 return parent->subtree_control;
358 if (cgroup_on_dfl(cgrp))
359 root_ss_mask &= ~cgrp_dfl_inhibit_ss_mask;
364 /* subsystems enabled on a cgroup */
365 static u16 cgroup_ss_mask(struct cgroup *cgrp)
367 struct cgroup *parent = cgroup_parent(cgrp);
370 return parent->subtree_ss_mask;
372 return cgrp->root->subsys_mask;
376 * cgroup_css - obtain a cgroup's css for the specified subsystem
377 * @cgrp: the cgroup of interest
378 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
380 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
381 * function must be called either under cgroup_mutex or rcu_read_lock() and
382 * the caller is responsible for pinning the returned css if it wants to
383 * keep accessing it outside the said locks. This function may return
384 * %NULL if @cgrp doesn't have @subsys_id enabled.
386 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
387 struct cgroup_subsys *ss)
390 return rcu_dereference_check(cgrp->subsys[ss->id],
391 lockdep_is_held(&cgroup_mutex));
397 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
401 * Similar to cgroup_css() but returns the effective css, which is defined
402 * as the matching css of the nearest ancestor including self which has @ss
403 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
404 * function is guaranteed to return non-NULL css.
406 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
407 struct cgroup_subsys *ss)
409 lockdep_assert_held(&cgroup_mutex);
415 * This function is used while updating css associations and thus
416 * can't test the csses directly. Test ss_mask.
418 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
419 cgrp = cgroup_parent(cgrp);
424 return cgroup_css(cgrp, ss);
428 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
429 * @cgrp: the cgroup of interest
430 * @ss: the subsystem of interest
432 * Find and get the effective css of @cgrp for @ss. The effective css is
433 * defined as the matching css of the nearest ancestor including self which
434 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
435 * the root css is returned, so this function always returns a valid css.
436 * The returned css must be put using css_put().
438 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
439 struct cgroup_subsys *ss)
441 struct cgroup_subsys_state *css;
446 css = cgroup_css(cgrp, ss);
448 if (css && css_tryget_online(css))
450 cgrp = cgroup_parent(cgrp);
453 css = init_css_set.subsys[ss->id];
460 /* convenient tests for these bits */
461 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
463 return !(cgrp->self.flags & CSS_ONLINE);
466 static void cgroup_get(struct cgroup *cgrp)
468 WARN_ON_ONCE(cgroup_is_dead(cgrp));
469 css_get(&cgrp->self);
472 static bool cgroup_tryget(struct cgroup *cgrp)
474 return css_tryget(&cgrp->self);
477 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
479 struct cgroup *cgrp = of->kn->parent->priv;
480 struct cftype *cft = of_cft(of);
483 * This is open and unprotected implementation of cgroup_css().
484 * seq_css() is only called from a kernfs file operation which has
485 * an active reference on the file. Because all the subsystem
486 * files are drained before a css is disassociated with a cgroup,
487 * the matching css from the cgroup's subsys table is guaranteed to
488 * be and stay valid until the enclosing operation is complete.
491 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
495 EXPORT_SYMBOL_GPL(of_css);
497 static int notify_on_release(const struct cgroup *cgrp)
499 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
503 * for_each_css - iterate all css's of a cgroup
504 * @css: the iteration cursor
505 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
506 * @cgrp: the target cgroup to iterate css's of
508 * Should be called under cgroup_[tree_]mutex.
510 #define for_each_css(css, ssid, cgrp) \
511 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
512 if (!((css) = rcu_dereference_check( \
513 (cgrp)->subsys[(ssid)], \
514 lockdep_is_held(&cgroup_mutex)))) { } \
518 * for_each_e_css - iterate all effective css's of a cgroup
519 * @css: the iteration cursor
520 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
521 * @cgrp: the target cgroup to iterate css's of
523 * Should be called under cgroup_[tree_]mutex.
525 #define for_each_e_css(css, ssid, cgrp) \
526 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
527 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
532 * for_each_subsys - iterate all enabled cgroup subsystems
533 * @ss: the iteration cursor
534 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
536 #define for_each_subsys(ss, ssid) \
537 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
538 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
541 * do_each_subsys_mask - filter for_each_subsys with a bitmask
542 * @ss: the iteration cursor
543 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
544 * @ss_mask: the bitmask
546 * The block will only run for cases where the ssid-th bit (1 << ssid) of
549 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
550 unsigned long __ss_mask = (ss_mask); \
551 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
555 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
556 (ss) = cgroup_subsys[ssid]; \
559 #define while_each_subsys_mask() \
564 /* iterate across the hierarchies */
565 #define for_each_root(root) \
566 list_for_each_entry((root), &cgroup_roots, root_list)
568 /* iterate over child cgrps, lock should be held throughout iteration */
569 #define cgroup_for_each_live_child(child, cgrp) \
570 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
571 if (({ lockdep_assert_held(&cgroup_mutex); \
572 cgroup_is_dead(child); })) \
576 static void cgroup_release_agent(struct work_struct *work);
577 static void check_for_release(struct cgroup *cgrp);
580 * A cgroup can be associated with multiple css_sets as different tasks may
581 * belong to different cgroups on different hierarchies. In the other
582 * direction, a css_set is naturally associated with multiple cgroups.
583 * This M:N relationship is represented by the following link structure
584 * which exists for each association and allows traversing the associations
587 struct cgrp_cset_link {
588 /* the cgroup and css_set this link associates */
590 struct css_set *cset;
592 /* list of cgrp_cset_links anchored at cgrp->cset_links */
593 struct list_head cset_link;
595 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
596 struct list_head cgrp_link;
600 * The default css_set - used by init and its children prior to any
601 * hierarchies being mounted. It contains a pointer to the root state
602 * for each subsystem. Also used to anchor the list of css_sets. Not
603 * reference-counted, to improve performance when child cgroups
604 * haven't been created.
606 struct css_set init_css_set = {
607 .refcount = ATOMIC_INIT(1),
608 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
609 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
610 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
611 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
612 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
613 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
616 static int css_set_count = 1; /* 1 for init_css_set */
619 * css_set_populated - does a css_set contain any tasks?
620 * @cset: target css_set
622 static bool css_set_populated(struct css_set *cset)
624 lockdep_assert_held(&css_set_lock);
626 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
630 * cgroup_update_populated - updated populated count of a cgroup
631 * @cgrp: the target cgroup
632 * @populated: inc or dec populated count
634 * One of the css_sets associated with @cgrp is either getting its first
635 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
636 * count is propagated towards root so that a given cgroup's populated_cnt
637 * is zero iff the cgroup and all its descendants don't contain any tasks.
639 * @cgrp's interface file "cgroup.populated" is zero if
640 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
641 * changes from or to zero, userland is notified that the content of the
642 * interface file has changed. This can be used to detect when @cgrp and
643 * its descendants become populated or empty.
645 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
647 lockdep_assert_held(&css_set_lock);
653 trigger = !cgrp->populated_cnt++;
655 trigger = !--cgrp->populated_cnt;
660 check_for_release(cgrp);
661 cgroup_file_notify(&cgrp->events_file);
663 cgrp = cgroup_parent(cgrp);
668 * css_set_update_populated - update populated state of a css_set
669 * @cset: target css_set
670 * @populated: whether @cset is populated or depopulated
672 * @cset is either getting the first task or losing the last. Update the
673 * ->populated_cnt of all associated cgroups accordingly.
675 static void css_set_update_populated(struct css_set *cset, bool populated)
677 struct cgrp_cset_link *link;
679 lockdep_assert_held(&css_set_lock);
681 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
682 cgroup_update_populated(link->cgrp, populated);
686 * css_set_move_task - move a task from one css_set to another
687 * @task: task being moved
688 * @from_cset: css_set @task currently belongs to (may be NULL)
689 * @to_cset: new css_set @task is being moved to (may be NULL)
690 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
692 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
693 * css_set, @from_cset can be NULL. If @task is being disassociated
694 * instead of moved, @to_cset can be NULL.
696 * This function automatically handles populated_cnt updates and
697 * css_task_iter adjustments but the caller is responsible for managing
698 * @from_cset and @to_cset's reference counts.
700 static void css_set_move_task(struct task_struct *task,
701 struct css_set *from_cset, struct css_set *to_cset,
704 lockdep_assert_held(&css_set_lock);
706 if (to_cset && !css_set_populated(to_cset))
707 css_set_update_populated(to_cset, true);
710 struct css_task_iter *it, *pos;
712 WARN_ON_ONCE(list_empty(&task->cg_list));
715 * @task is leaving, advance task iterators which are
716 * pointing to it so that they can resume at the next
717 * position. Advancing an iterator might remove it from
718 * the list, use safe walk. See css_task_iter_advance*()
721 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
723 if (it->task_pos == &task->cg_list)
724 css_task_iter_advance(it);
726 list_del_init(&task->cg_list);
727 if (!css_set_populated(from_cset))
728 css_set_update_populated(from_cset, false);
730 WARN_ON_ONCE(!list_empty(&task->cg_list));
735 * We are synchronized through cgroup_threadgroup_rwsem
736 * against PF_EXITING setting such that we can't race
737 * against cgroup_exit() changing the css_set to
738 * init_css_set and dropping the old one.
740 WARN_ON_ONCE(task->flags & PF_EXITING);
742 rcu_assign_pointer(task->cgroups, to_cset);
743 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
749 * hash table for cgroup groups. This improves the performance to find
750 * an existing css_set. This hash doesn't (currently) take into
751 * account cgroups in empty hierarchies.
753 #define CSS_SET_HASH_BITS 7
754 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
756 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
758 unsigned long key = 0UL;
759 struct cgroup_subsys *ss;
762 for_each_subsys(ss, i)
763 key += (unsigned long)css[i];
764 key = (key >> 16) ^ key;
769 static void put_css_set_locked(struct css_set *cset)
771 struct cgrp_cset_link *link, *tmp_link;
772 struct cgroup_subsys *ss;
775 lockdep_assert_held(&css_set_lock);
777 if (!atomic_dec_and_test(&cset->refcount))
780 /* This css_set is dead. unlink it and release cgroup and css refs */
781 for_each_subsys(ss, ssid) {
782 list_del(&cset->e_cset_node[ssid]);
783 css_put(cset->subsys[ssid]);
785 hash_del(&cset->hlist);
788 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
789 list_del(&link->cset_link);
790 list_del(&link->cgrp_link);
791 if (cgroup_parent(link->cgrp))
792 cgroup_put(link->cgrp);
796 kfree_rcu(cset, rcu_head);
799 static void put_css_set(struct css_set *cset)
802 * Ensure that the refcount doesn't hit zero while any readers
803 * can see it. Similar to atomic_dec_and_lock(), but for an
806 if (atomic_add_unless(&cset->refcount, -1, 1))
809 spin_lock_bh(&css_set_lock);
810 put_css_set_locked(cset);
811 spin_unlock_bh(&css_set_lock);
815 * refcounted get/put for css_set objects
817 static inline void get_css_set(struct css_set *cset)
819 atomic_inc(&cset->refcount);
823 * compare_css_sets - helper function for find_existing_css_set().
824 * @cset: candidate css_set being tested
825 * @old_cset: existing css_set for a task
826 * @new_cgrp: cgroup that's being entered by the task
827 * @template: desired set of css pointers in css_set (pre-calculated)
829 * Returns true if "cset" matches "old_cset" except for the hierarchy
830 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
832 static bool compare_css_sets(struct css_set *cset,
833 struct css_set *old_cset,
834 struct cgroup *new_cgrp,
835 struct cgroup_subsys_state *template[])
837 struct list_head *l1, *l2;
840 * On the default hierarchy, there can be csets which are
841 * associated with the same set of cgroups but different csses.
842 * Let's first ensure that csses match.
844 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
848 * Compare cgroup pointers in order to distinguish between
849 * different cgroups in hierarchies. As different cgroups may
850 * share the same effective css, this comparison is always
853 l1 = &cset->cgrp_links;
854 l2 = &old_cset->cgrp_links;
856 struct cgrp_cset_link *link1, *link2;
857 struct cgroup *cgrp1, *cgrp2;
861 /* See if we reached the end - both lists are equal length. */
862 if (l1 == &cset->cgrp_links) {
863 BUG_ON(l2 != &old_cset->cgrp_links);
866 BUG_ON(l2 == &old_cset->cgrp_links);
868 /* Locate the cgroups associated with these links. */
869 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
870 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
873 /* Hierarchies should be linked in the same order. */
874 BUG_ON(cgrp1->root != cgrp2->root);
877 * If this hierarchy is the hierarchy of the cgroup
878 * that's changing, then we need to check that this
879 * css_set points to the new cgroup; if it's any other
880 * hierarchy, then this css_set should point to the
881 * same cgroup as the old css_set.
883 if (cgrp1->root == new_cgrp->root) {
884 if (cgrp1 != new_cgrp)
895 * find_existing_css_set - init css array and find the matching css_set
896 * @old_cset: the css_set that we're using before the cgroup transition
897 * @cgrp: the cgroup that we're moving into
898 * @template: out param for the new set of csses, should be clear on entry
900 static struct css_set *find_existing_css_set(struct css_set *old_cset,
902 struct cgroup_subsys_state *template[])
904 struct cgroup_root *root = cgrp->root;
905 struct cgroup_subsys *ss;
906 struct css_set *cset;
911 * Build the set of subsystem state objects that we want to see in the
912 * new css_set. while subsystems can change globally, the entries here
913 * won't change, so no need for locking.
915 for_each_subsys(ss, i) {
916 if (root->subsys_mask & (1UL << i)) {
918 * @ss is in this hierarchy, so we want the
919 * effective css from @cgrp.
921 template[i] = cgroup_e_css(cgrp, ss);
924 * @ss is not in this hierarchy, so we don't want
927 template[i] = old_cset->subsys[i];
931 key = css_set_hash(template);
932 hash_for_each_possible(css_set_table, cset, hlist, key) {
933 if (!compare_css_sets(cset, old_cset, cgrp, template))
936 /* This css_set matches what we need */
940 /* No existing cgroup group matched */
944 static void free_cgrp_cset_links(struct list_head *links_to_free)
946 struct cgrp_cset_link *link, *tmp_link;
948 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
949 list_del(&link->cset_link);
955 * allocate_cgrp_cset_links - allocate cgrp_cset_links
956 * @count: the number of links to allocate
957 * @tmp_links: list_head the allocated links are put on
959 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
960 * through ->cset_link. Returns 0 on success or -errno.
962 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
964 struct cgrp_cset_link *link;
967 INIT_LIST_HEAD(tmp_links);
969 for (i = 0; i < count; i++) {
970 link = kzalloc(sizeof(*link), GFP_KERNEL);
972 free_cgrp_cset_links(tmp_links);
975 list_add(&link->cset_link, tmp_links);
981 * link_css_set - a helper function to link a css_set to a cgroup
982 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
983 * @cset: the css_set to be linked
984 * @cgrp: the destination cgroup
986 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
989 struct cgrp_cset_link *link;
991 BUG_ON(list_empty(tmp_links));
993 if (cgroup_on_dfl(cgrp))
994 cset->dfl_cgrp = cgrp;
996 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1001 * Always add links to the tail of the lists so that the lists are
1002 * in choronological order.
1004 list_move_tail(&link->cset_link, &cgrp->cset_links);
1005 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1007 if (cgroup_parent(cgrp))
1012 * find_css_set - return a new css_set with one cgroup updated
1013 * @old_cset: the baseline css_set
1014 * @cgrp: the cgroup to be updated
1016 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1017 * substituted into the appropriate hierarchy.
1019 static struct css_set *find_css_set(struct css_set *old_cset,
1020 struct cgroup *cgrp)
1022 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1023 struct css_set *cset;
1024 struct list_head tmp_links;
1025 struct cgrp_cset_link *link;
1026 struct cgroup_subsys *ss;
1030 lockdep_assert_held(&cgroup_mutex);
1032 /* First see if we already have a cgroup group that matches
1033 * the desired set */
1034 spin_lock_bh(&css_set_lock);
1035 cset = find_existing_css_set(old_cset, cgrp, template);
1038 spin_unlock_bh(&css_set_lock);
1043 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1047 /* Allocate all the cgrp_cset_link objects that we'll need */
1048 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1053 atomic_set(&cset->refcount, 1);
1054 INIT_LIST_HEAD(&cset->cgrp_links);
1055 INIT_LIST_HEAD(&cset->tasks);
1056 INIT_LIST_HEAD(&cset->mg_tasks);
1057 INIT_LIST_HEAD(&cset->mg_preload_node);
1058 INIT_LIST_HEAD(&cset->mg_node);
1059 INIT_LIST_HEAD(&cset->task_iters);
1060 INIT_HLIST_NODE(&cset->hlist);
1062 /* Copy the set of subsystem state objects generated in
1063 * find_existing_css_set() */
1064 memcpy(cset->subsys, template, sizeof(cset->subsys));
1066 spin_lock_bh(&css_set_lock);
1067 /* Add reference counts and links from the new css_set. */
1068 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1069 struct cgroup *c = link->cgrp;
1071 if (c->root == cgrp->root)
1073 link_css_set(&tmp_links, cset, c);
1076 BUG_ON(!list_empty(&tmp_links));
1080 /* Add @cset to the hash table */
1081 key = css_set_hash(cset->subsys);
1082 hash_add(css_set_table, &cset->hlist, key);
1084 for_each_subsys(ss, ssid) {
1085 struct cgroup_subsys_state *css = cset->subsys[ssid];
1087 list_add_tail(&cset->e_cset_node[ssid],
1088 &css->cgroup->e_csets[ssid]);
1092 spin_unlock_bh(&css_set_lock);
1097 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1099 struct cgroup *root_cgrp = kf_root->kn->priv;
1101 return root_cgrp->root;
1104 static int cgroup_init_root_id(struct cgroup_root *root)
1108 lockdep_assert_held(&cgroup_mutex);
1110 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1114 root->hierarchy_id = id;
1118 static void cgroup_exit_root_id(struct cgroup_root *root)
1120 lockdep_assert_held(&cgroup_mutex);
1122 if (root->hierarchy_id) {
1123 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1124 root->hierarchy_id = 0;
1128 static void cgroup_free_root(struct cgroup_root *root)
1131 /* hierarchy ID should already have been released */
1132 WARN_ON_ONCE(root->hierarchy_id);
1134 idr_destroy(&root->cgroup_idr);
1139 static void cgroup_destroy_root(struct cgroup_root *root)
1141 struct cgroup *cgrp = &root->cgrp;
1142 struct cgrp_cset_link *link, *tmp_link;
1144 mutex_lock(&cgroup_mutex);
1146 BUG_ON(atomic_read(&root->nr_cgrps));
1147 BUG_ON(!list_empty(&cgrp->self.children));
1149 /* Rebind all subsystems back to the default hierarchy */
1150 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1153 * Release all the links from cset_links to this hierarchy's
1156 spin_lock_bh(&css_set_lock);
1158 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1159 list_del(&link->cset_link);
1160 list_del(&link->cgrp_link);
1164 spin_unlock_bh(&css_set_lock);
1166 if (!list_empty(&root->root_list)) {
1167 list_del(&root->root_list);
1168 cgroup_root_count--;
1171 cgroup_exit_root_id(root);
1173 mutex_unlock(&cgroup_mutex);
1175 kernfs_destroy_root(root->kf_root);
1176 cgroup_free_root(root);
1179 /* look up cgroup associated with given css_set on the specified hierarchy */
1180 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1181 struct cgroup_root *root)
1183 struct cgroup *res = NULL;
1185 lockdep_assert_held(&cgroup_mutex);
1186 lockdep_assert_held(&css_set_lock);
1188 if (cset == &init_css_set) {
1191 struct cgrp_cset_link *link;
1193 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1194 struct cgroup *c = link->cgrp;
1196 if (c->root == root) {
1208 * Return the cgroup for "task" from the given hierarchy. Must be
1209 * called with cgroup_mutex and css_set_lock held.
1211 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1212 struct cgroup_root *root)
1215 * No need to lock the task - since we hold cgroup_mutex the
1216 * task can't change groups, so the only thing that can happen
1217 * is that it exits and its css is set back to init_css_set.
1219 return cset_cgroup_from_root(task_css_set(task), root);
1223 * A task must hold cgroup_mutex to modify cgroups.
1225 * Any task can increment and decrement the count field without lock.
1226 * So in general, code holding cgroup_mutex can't rely on the count
1227 * field not changing. However, if the count goes to zero, then only
1228 * cgroup_attach_task() can increment it again. Because a count of zero
1229 * means that no tasks are currently attached, therefore there is no
1230 * way a task attached to that cgroup can fork (the other way to
1231 * increment the count). So code holding cgroup_mutex can safely
1232 * assume that if the count is zero, it will stay zero. Similarly, if
1233 * a task holds cgroup_mutex on a cgroup with zero count, it
1234 * knows that the cgroup won't be removed, as cgroup_rmdir()
1237 * A cgroup can only be deleted if both its 'count' of using tasks
1238 * is zero, and its list of 'children' cgroups is empty. Since all
1239 * tasks in the system use _some_ cgroup, and since there is always at
1240 * least one task in the system (init, pid == 1), therefore, root cgroup
1241 * always has either children cgroups and/or using tasks. So we don't
1242 * need a special hack to ensure that root cgroup cannot be deleted.
1244 * P.S. One more locking exception. RCU is used to guard the
1245 * update of a tasks cgroup pointer by cgroup_attach_task()
1248 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1249 static const struct file_operations proc_cgroupstats_operations;
1251 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1254 struct cgroup_subsys *ss = cft->ss;
1256 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1257 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1258 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1259 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1262 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1267 * cgroup_file_mode - deduce file mode of a control file
1268 * @cft: the control file in question
1270 * S_IRUGO for read, S_IWUSR for write.
1272 static umode_t cgroup_file_mode(const struct cftype *cft)
1276 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1279 if (cft->write_u64 || cft->write_s64 || cft->write) {
1280 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1290 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1291 * @cgrp: the target cgroup
1292 * @subtree_control: the new subtree_control mask to consider
1294 * On the default hierarchy, a subsystem may request other subsystems to be
1295 * enabled together through its ->depends_on mask. In such cases, more
1296 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1298 * This function calculates which subsystems need to be enabled if
1299 * @subtree_control is to be applied to @cgrp. The returned mask is always
1300 * a superset of @subtree_control and follows the usual hierarchy rules.
1302 static u16 cgroup_calc_subtree_ss_mask(struct cgroup *cgrp, u16 subtree_control)
1304 u16 cur_ss_mask = subtree_control;
1305 struct cgroup_subsys *ss;
1308 lockdep_assert_held(&cgroup_mutex);
1310 if (!cgroup_on_dfl(cgrp))
1314 u16 new_ss_mask = cur_ss_mask;
1316 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1317 new_ss_mask |= ss->depends_on;
1318 } while_each_subsys_mask();
1321 * Mask out subsystems which aren't available. This can
1322 * happen only if some depended-upon subsystems were bound
1323 * to non-default hierarchies.
1325 new_ss_mask &= cgroup_ss_mask(cgrp);
1327 if (new_ss_mask == cur_ss_mask)
1329 cur_ss_mask = new_ss_mask;
1336 * cgroup_refresh_subtree_ss_mask - update subtree_ss_mask
1337 * @cgrp: the target cgroup
1339 * Update @cgrp->subtree_ss_mask according to the current
1340 * @cgrp->subtree_control using cgroup_calc_subtree_ss_mask().
1342 static void cgroup_refresh_subtree_ss_mask(struct cgroup *cgrp)
1344 cgrp->subtree_ss_mask =
1345 cgroup_calc_subtree_ss_mask(cgrp, cgrp->subtree_control);
1349 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1350 * @kn: the kernfs_node being serviced
1352 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1353 * the method finishes if locking succeeded. Note that once this function
1354 * returns the cgroup returned by cgroup_kn_lock_live() may become
1355 * inaccessible any time. If the caller intends to continue to access the
1356 * cgroup, it should pin it before invoking this function.
1358 static void cgroup_kn_unlock(struct kernfs_node *kn)
1360 struct cgroup *cgrp;
1362 if (kernfs_type(kn) == KERNFS_DIR)
1365 cgrp = kn->parent->priv;
1367 mutex_unlock(&cgroup_mutex);
1369 kernfs_unbreak_active_protection(kn);
1374 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1375 * @kn: the kernfs_node being serviced
1377 * This helper is to be used by a cgroup kernfs method currently servicing
1378 * @kn. It breaks the active protection, performs cgroup locking and
1379 * verifies that the associated cgroup is alive. Returns the cgroup if
1380 * alive; otherwise, %NULL. A successful return should be undone by a
1381 * matching cgroup_kn_unlock() invocation.
1383 * Any cgroup kernfs method implementation which requires locking the
1384 * associated cgroup should use this helper. It avoids nesting cgroup
1385 * locking under kernfs active protection and allows all kernfs operations
1386 * including self-removal.
1388 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1390 struct cgroup *cgrp;
1392 if (kernfs_type(kn) == KERNFS_DIR)
1395 cgrp = kn->parent->priv;
1398 * We're gonna grab cgroup_mutex which nests outside kernfs
1399 * active_ref. cgroup liveliness check alone provides enough
1400 * protection against removal. Ensure @cgrp stays accessible and
1401 * break the active_ref protection.
1403 if (!cgroup_tryget(cgrp))
1405 kernfs_break_active_protection(kn);
1407 mutex_lock(&cgroup_mutex);
1409 if (!cgroup_is_dead(cgrp))
1412 cgroup_kn_unlock(kn);
1416 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1418 char name[CGROUP_FILE_NAME_MAX];
1420 lockdep_assert_held(&cgroup_mutex);
1422 if (cft->file_offset) {
1423 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1424 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1426 spin_lock_irq(&cgroup_file_kn_lock);
1428 spin_unlock_irq(&cgroup_file_kn_lock);
1431 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1435 * css_clear_dir - remove subsys files in a cgroup directory
1437 * @cgrp_override: specify if target cgroup is different from css->cgroup
1439 static void css_clear_dir(struct cgroup_subsys_state *css,
1440 struct cgroup *cgrp_override)
1442 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1443 struct cftype *cfts;
1445 if (!(css->flags & CSS_VISIBLE))
1448 css->flags &= ~CSS_VISIBLE;
1450 list_for_each_entry(cfts, &css->ss->cfts, node)
1451 cgroup_addrm_files(css, cgrp, cfts, false);
1455 * css_populate_dir - create subsys files in a cgroup directory
1457 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1459 * On failure, no file is added.
1461 static int css_populate_dir(struct cgroup_subsys_state *css,
1462 struct cgroup *cgrp_override)
1464 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1465 struct cftype *cfts, *failed_cfts;
1468 if (css->flags & CSS_VISIBLE)
1472 if (cgroup_on_dfl(cgrp))
1473 cfts = cgroup_dfl_base_files;
1475 cfts = cgroup_legacy_base_files;
1477 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1480 list_for_each_entry(cfts, &css->ss->cfts, node) {
1481 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1488 css->flags |= CSS_VISIBLE;
1492 list_for_each_entry(cfts, &css->ss->cfts, node) {
1493 if (cfts == failed_cfts)
1495 cgroup_addrm_files(css, cgrp, cfts, false);
1500 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1502 struct cgroup *dcgrp = &dst_root->cgrp;
1503 struct cgroup_subsys *ss;
1507 lockdep_assert_held(&cgroup_mutex);
1509 do_each_subsys_mask(ss, ssid, ss_mask) {
1510 /* if @ss has non-root csses attached to it, can't move */
1511 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1514 /* can't move between two non-dummy roots either */
1515 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1517 } while_each_subsys_mask();
1519 /* skip creating root files on dfl_root for inhibited subsystems */
1520 tmp_ss_mask = ss_mask;
1521 if (dst_root == &cgrp_dfl_root)
1522 tmp_ss_mask &= ~cgrp_dfl_inhibit_ss_mask;
1524 do_each_subsys_mask(ss, ssid, tmp_ss_mask) {
1525 struct cgroup *scgrp = &ss->root->cgrp;
1528 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1533 * Rebinding back to the default root is not allowed to
1534 * fail. Using both default and non-default roots should
1535 * be rare. Moving subsystems back and forth even more so.
1536 * Just warn about it and continue.
1538 if (dst_root == &cgrp_dfl_root) {
1539 if (cgrp_dfl_visible) {
1540 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1542 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1547 do_each_subsys_mask(ss, tssid, tmp_ss_mask) {
1550 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1551 } while_each_subsys_mask();
1553 } while_each_subsys_mask();
1556 * Nothing can fail from this point on. Remove files for the
1557 * removed subsystems and rebind each subsystem.
1559 do_each_subsys_mask(ss, ssid, ss_mask) {
1560 struct cgroup_root *src_root = ss->root;
1561 struct cgroup *scgrp = &src_root->cgrp;
1562 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1563 struct css_set *cset;
1565 WARN_ON(!css || cgroup_css(dcgrp, ss));
1567 css_clear_dir(css, NULL);
1569 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1570 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1571 ss->root = dst_root;
1572 css->cgroup = dcgrp;
1574 spin_lock_bh(&css_set_lock);
1575 hash_for_each(css_set_table, i, cset, hlist)
1576 list_move_tail(&cset->e_cset_node[ss->id],
1577 &dcgrp->e_csets[ss->id]);
1578 spin_unlock_bh(&css_set_lock);
1580 src_root->subsys_mask &= ~(1 << ssid);
1581 scgrp->subtree_control &= ~(1 << ssid);
1582 cgroup_refresh_subtree_ss_mask(scgrp);
1584 /* default hierarchy doesn't enable controllers by default */
1585 dst_root->subsys_mask |= 1 << ssid;
1586 if (dst_root == &cgrp_dfl_root) {
1587 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1589 dcgrp->subtree_control |= 1 << ssid;
1590 cgroup_refresh_subtree_ss_mask(dcgrp);
1591 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1596 } while_each_subsys_mask();
1598 kernfs_activate(dcgrp->kn);
1602 static int cgroup_show_options(struct seq_file *seq,
1603 struct kernfs_root *kf_root)
1605 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1606 struct cgroup_subsys *ss;
1609 if (root != &cgrp_dfl_root)
1610 for_each_subsys(ss, ssid)
1611 if (root->subsys_mask & (1 << ssid))
1612 seq_show_option(seq, ss->legacy_name, NULL);
1613 if (root->flags & CGRP_ROOT_NOPREFIX)
1614 seq_puts(seq, ",noprefix");
1615 if (root->flags & CGRP_ROOT_XATTR)
1616 seq_puts(seq, ",xattr");
1618 spin_lock(&release_agent_path_lock);
1619 if (strlen(root->release_agent_path))
1620 seq_show_option(seq, "release_agent",
1621 root->release_agent_path);
1622 spin_unlock(&release_agent_path_lock);
1624 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1625 seq_puts(seq, ",clone_children");
1626 if (strlen(root->name))
1627 seq_show_option(seq, "name", root->name);
1631 struct cgroup_sb_opts {
1634 char *release_agent;
1635 bool cpuset_clone_children;
1637 /* User explicitly requested empty subsystem */
1641 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1643 char *token, *o = data;
1644 bool all_ss = false, one_ss = false;
1646 struct cgroup_subsys *ss;
1650 #ifdef CONFIG_CPUSETS
1651 mask = ~((u16)1 << cpuset_cgrp_id);
1654 memset(opts, 0, sizeof(*opts));
1656 while ((token = strsep(&o, ",")) != NULL) {
1661 if (!strcmp(token, "none")) {
1662 /* Explicitly have no subsystems */
1666 if (!strcmp(token, "all")) {
1667 /* Mutually exclusive option 'all' + subsystem name */
1673 if (!strcmp(token, "noprefix")) {
1674 opts->flags |= CGRP_ROOT_NOPREFIX;
1677 if (!strcmp(token, "clone_children")) {
1678 opts->cpuset_clone_children = true;
1681 if (!strcmp(token, "xattr")) {
1682 opts->flags |= CGRP_ROOT_XATTR;
1685 if (!strncmp(token, "release_agent=", 14)) {
1686 /* Specifying two release agents is forbidden */
1687 if (opts->release_agent)
1689 opts->release_agent =
1690 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1691 if (!opts->release_agent)
1695 if (!strncmp(token, "name=", 5)) {
1696 const char *name = token + 5;
1697 /* Can't specify an empty name */
1700 /* Must match [\w.-]+ */
1701 for (i = 0; i < strlen(name); i++) {
1705 if ((c == '.') || (c == '-') || (c == '_'))
1709 /* Specifying two names is forbidden */
1712 opts->name = kstrndup(name,
1713 MAX_CGROUP_ROOT_NAMELEN - 1,
1721 for_each_subsys(ss, i) {
1722 if (strcmp(token, ss->legacy_name))
1724 if (!cgroup_ssid_enabled(i))
1726 if (cgroup_ssid_no_v1(i))
1729 /* Mutually exclusive option 'all' + subsystem name */
1732 opts->subsys_mask |= (1 << i);
1737 if (i == CGROUP_SUBSYS_COUNT)
1742 * If the 'all' option was specified select all the subsystems,
1743 * otherwise if 'none', 'name=' and a subsystem name options were
1744 * not specified, let's default to 'all'
1746 if (all_ss || (!one_ss && !opts->none && !opts->name))
1747 for_each_subsys(ss, i)
1748 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1749 opts->subsys_mask |= (1 << i);
1752 * We either have to specify by name or by subsystems. (So all
1753 * empty hierarchies must have a name).
1755 if (!opts->subsys_mask && !opts->name)
1759 * Option noprefix was introduced just for backward compatibility
1760 * with the old cpuset, so we allow noprefix only if mounting just
1761 * the cpuset subsystem.
1763 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1766 /* Can't specify "none" and some subsystems */
1767 if (opts->subsys_mask && opts->none)
1773 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1776 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1777 struct cgroup_sb_opts opts;
1778 u16 added_mask, removed_mask;
1780 if (root == &cgrp_dfl_root) {
1781 pr_err("remount is not allowed\n");
1785 mutex_lock(&cgroup_mutex);
1787 /* See what subsystems are wanted */
1788 ret = parse_cgroupfs_options(data, &opts);
1792 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1793 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1794 task_tgid_nr(current), current->comm);
1796 added_mask = opts.subsys_mask & ~root->subsys_mask;
1797 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1799 /* Don't allow flags or name to change at remount */
1800 if ((opts.flags ^ root->flags) ||
1801 (opts.name && strcmp(opts.name, root->name))) {
1802 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1803 opts.flags, opts.name ?: "", root->flags, root->name);
1808 /* remounting is not allowed for populated hierarchies */
1809 if (!list_empty(&root->cgrp.self.children)) {
1814 ret = rebind_subsystems(root, added_mask);
1818 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1820 if (opts.release_agent) {
1821 spin_lock(&release_agent_path_lock);
1822 strcpy(root->release_agent_path, opts.release_agent);
1823 spin_unlock(&release_agent_path_lock);
1826 kfree(opts.release_agent);
1828 mutex_unlock(&cgroup_mutex);
1833 * To reduce the fork() overhead for systems that are not actually using
1834 * their cgroups capability, we don't maintain the lists running through
1835 * each css_set to its tasks until we see the list actually used - in other
1836 * words after the first mount.
1838 static bool use_task_css_set_links __read_mostly;
1840 static void cgroup_enable_task_cg_lists(void)
1842 struct task_struct *p, *g;
1844 spin_lock_bh(&css_set_lock);
1846 if (use_task_css_set_links)
1849 use_task_css_set_links = true;
1852 * We need tasklist_lock because RCU is not safe against
1853 * while_each_thread(). Besides, a forking task that has passed
1854 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1855 * is not guaranteed to have its child immediately visible in the
1856 * tasklist if we walk through it with RCU.
1858 read_lock(&tasklist_lock);
1859 do_each_thread(g, p) {
1860 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1861 task_css_set(p) != &init_css_set);
1864 * We should check if the process is exiting, otherwise
1865 * it will race with cgroup_exit() in that the list
1866 * entry won't be deleted though the process has exited.
1867 * Do it while holding siglock so that we don't end up
1868 * racing against cgroup_exit().
1870 spin_lock_irq(&p->sighand->siglock);
1871 if (!(p->flags & PF_EXITING)) {
1872 struct css_set *cset = task_css_set(p);
1874 if (!css_set_populated(cset))
1875 css_set_update_populated(cset, true);
1876 list_add_tail(&p->cg_list, &cset->tasks);
1879 spin_unlock_irq(&p->sighand->siglock);
1880 } while_each_thread(g, p);
1881 read_unlock(&tasklist_lock);
1883 spin_unlock_bh(&css_set_lock);
1886 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1888 struct cgroup_subsys *ss;
1891 INIT_LIST_HEAD(&cgrp->self.sibling);
1892 INIT_LIST_HEAD(&cgrp->self.children);
1893 INIT_LIST_HEAD(&cgrp->cset_links);
1894 INIT_LIST_HEAD(&cgrp->pidlists);
1895 mutex_init(&cgrp->pidlist_mutex);
1896 cgrp->self.cgroup = cgrp;
1897 cgrp->self.flags |= CSS_ONLINE;
1899 for_each_subsys(ss, ssid)
1900 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1902 init_waitqueue_head(&cgrp->offline_waitq);
1903 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1906 static void init_cgroup_root(struct cgroup_root *root,
1907 struct cgroup_sb_opts *opts)
1909 struct cgroup *cgrp = &root->cgrp;
1911 INIT_LIST_HEAD(&root->root_list);
1912 atomic_set(&root->nr_cgrps, 1);
1914 init_cgroup_housekeeping(cgrp);
1915 idr_init(&root->cgroup_idr);
1917 root->flags = opts->flags;
1918 if (opts->release_agent)
1919 strcpy(root->release_agent_path, opts->release_agent);
1921 strcpy(root->name, opts->name);
1922 if (opts->cpuset_clone_children)
1923 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1926 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1928 LIST_HEAD(tmp_links);
1929 struct cgroup *root_cgrp = &root->cgrp;
1930 struct css_set *cset;
1933 lockdep_assert_held(&cgroup_mutex);
1935 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1938 root_cgrp->id = ret;
1939 root_cgrp->ancestor_ids[0] = ret;
1941 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1947 * We're accessing css_set_count without locking css_set_lock here,
1948 * but that's OK - it can only be increased by someone holding
1949 * cgroup_lock, and that's us. The worst that can happen is that we
1950 * have some link structures left over
1952 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1956 ret = cgroup_init_root_id(root);
1960 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1961 KERNFS_ROOT_CREATE_DEACTIVATED,
1963 if (IS_ERR(root->kf_root)) {
1964 ret = PTR_ERR(root->kf_root);
1967 root_cgrp->kn = root->kf_root->kn;
1969 ret = css_populate_dir(&root_cgrp->self, NULL);
1973 ret = rebind_subsystems(root, ss_mask);
1978 * There must be no failure case after here, since rebinding takes
1979 * care of subsystems' refcounts, which are explicitly dropped in
1980 * the failure exit path.
1982 list_add(&root->root_list, &cgroup_roots);
1983 cgroup_root_count++;
1986 * Link the root cgroup in this hierarchy into all the css_set
1989 spin_lock_bh(&css_set_lock);
1990 hash_for_each(css_set_table, i, cset, hlist) {
1991 link_css_set(&tmp_links, cset, root_cgrp);
1992 if (css_set_populated(cset))
1993 cgroup_update_populated(root_cgrp, true);
1995 spin_unlock_bh(&css_set_lock);
1997 BUG_ON(!list_empty(&root_cgrp->self.children));
1998 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2000 kernfs_activate(root_cgrp->kn);
2005 kernfs_destroy_root(root->kf_root);
2006 root->kf_root = NULL;
2008 cgroup_exit_root_id(root);
2010 percpu_ref_exit(&root_cgrp->self.refcnt);
2012 free_cgrp_cset_links(&tmp_links);
2016 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2017 int flags, const char *unused_dev_name,
2020 bool is_v2 = fs_type == &cgroup2_fs_type;
2021 struct super_block *pinned_sb = NULL;
2022 struct cgroup_subsys *ss;
2023 struct cgroup_root *root;
2024 struct cgroup_sb_opts opts;
2025 struct dentry *dentry;
2031 * The first time anyone tries to mount a cgroup, enable the list
2032 * linking each css_set to its tasks and fix up all existing tasks.
2034 if (!use_task_css_set_links)
2035 cgroup_enable_task_cg_lists();
2039 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2040 return ERR_PTR(-EINVAL);
2042 cgrp_dfl_visible = true;
2043 root = &cgrp_dfl_root;
2044 cgroup_get(&root->cgrp);
2048 mutex_lock(&cgroup_mutex);
2050 /* First find the desired set of subsystems */
2051 ret = parse_cgroupfs_options(data, &opts);
2056 * Destruction of cgroup root is asynchronous, so subsystems may
2057 * still be dying after the previous unmount. Let's drain the
2058 * dying subsystems. We just need to ensure that the ones
2059 * unmounted previously finish dying and don't care about new ones
2060 * starting. Testing ref liveliness is good enough.
2062 for_each_subsys(ss, i) {
2063 if (!(opts.subsys_mask & (1 << i)) ||
2064 ss->root == &cgrp_dfl_root)
2067 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2068 mutex_unlock(&cgroup_mutex);
2070 ret = restart_syscall();
2073 cgroup_put(&ss->root->cgrp);
2076 for_each_root(root) {
2077 bool name_match = false;
2079 if (root == &cgrp_dfl_root)
2083 * If we asked for a name then it must match. Also, if
2084 * name matches but sybsys_mask doesn't, we should fail.
2085 * Remember whether name matched.
2088 if (strcmp(opts.name, root->name))
2094 * If we asked for subsystems (or explicitly for no
2095 * subsystems) then they must match.
2097 if ((opts.subsys_mask || opts.none) &&
2098 (opts.subsys_mask != root->subsys_mask)) {
2105 if (root->flags ^ opts.flags)
2106 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2109 * We want to reuse @root whose lifetime is governed by its
2110 * ->cgrp. Let's check whether @root is alive and keep it
2111 * that way. As cgroup_kill_sb() can happen anytime, we
2112 * want to block it by pinning the sb so that @root doesn't
2113 * get killed before mount is complete.
2115 * With the sb pinned, tryget_live can reliably indicate
2116 * whether @root can be reused. If it's being killed,
2117 * drain it. We can use wait_queue for the wait but this
2118 * path is super cold. Let's just sleep a bit and retry.
2120 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2121 if (IS_ERR(pinned_sb) ||
2122 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2123 mutex_unlock(&cgroup_mutex);
2124 if (!IS_ERR_OR_NULL(pinned_sb))
2125 deactivate_super(pinned_sb);
2127 ret = restart_syscall();
2136 * No such thing, create a new one. name= matching without subsys
2137 * specification is allowed for already existing hierarchies but we
2138 * can't create new one without subsys specification.
2140 if (!opts.subsys_mask && !opts.none) {
2145 root = kzalloc(sizeof(*root), GFP_KERNEL);
2151 init_cgroup_root(root, &opts);
2153 ret = cgroup_setup_root(root, opts.subsys_mask);
2155 cgroup_free_root(root);
2158 mutex_unlock(&cgroup_mutex);
2160 kfree(opts.release_agent);
2164 return ERR_PTR(ret);
2166 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2167 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2169 if (IS_ERR(dentry) || !new_sb)
2170 cgroup_put(&root->cgrp);
2173 * If @pinned_sb, we're reusing an existing root and holding an
2174 * extra ref on its sb. Mount is complete. Put the extra ref.
2178 deactivate_super(pinned_sb);
2184 static void cgroup_kill_sb(struct super_block *sb)
2186 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2187 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2190 * If @root doesn't have any mounts or children, start killing it.
2191 * This prevents new mounts by disabling percpu_ref_tryget_live().
2192 * cgroup_mount() may wait for @root's release.
2194 * And don't kill the default root.
2196 if (!list_empty(&root->cgrp.self.children) ||
2197 root == &cgrp_dfl_root)
2198 cgroup_put(&root->cgrp);
2200 percpu_ref_kill(&root->cgrp.self.refcnt);
2205 static struct file_system_type cgroup_fs_type = {
2207 .mount = cgroup_mount,
2208 .kill_sb = cgroup_kill_sb,
2211 static struct file_system_type cgroup2_fs_type = {
2213 .mount = cgroup_mount,
2214 .kill_sb = cgroup_kill_sb,
2218 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2219 * @task: target task
2220 * @buf: the buffer to write the path into
2221 * @buflen: the length of the buffer
2223 * Determine @task's cgroup on the first (the one with the lowest non-zero
2224 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2225 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2226 * cgroup controller callbacks.
2228 * Return value is the same as kernfs_path().
2230 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2232 struct cgroup_root *root;
2233 struct cgroup *cgrp;
2234 int hierarchy_id = 1;
2237 mutex_lock(&cgroup_mutex);
2238 spin_lock_bh(&css_set_lock);
2240 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2243 cgrp = task_cgroup_from_root(task, root);
2244 path = cgroup_path(cgrp, buf, buflen);
2246 /* if no hierarchy exists, everyone is in "/" */
2247 if (strlcpy(buf, "/", buflen) < buflen)
2251 spin_unlock_bh(&css_set_lock);
2252 mutex_unlock(&cgroup_mutex);
2255 EXPORT_SYMBOL_GPL(task_cgroup_path);
2257 /* used to track tasks and other necessary states during migration */
2258 struct cgroup_taskset {
2259 /* the src and dst cset list running through cset->mg_node */
2260 struct list_head src_csets;
2261 struct list_head dst_csets;
2263 /* the subsys currently being processed */
2267 * Fields for cgroup_taskset_*() iteration.
2269 * Before migration is committed, the target migration tasks are on
2270 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2271 * the csets on ->dst_csets. ->csets point to either ->src_csets
2272 * or ->dst_csets depending on whether migration is committed.
2274 * ->cur_csets and ->cur_task point to the current task position
2277 struct list_head *csets;
2278 struct css_set *cur_cset;
2279 struct task_struct *cur_task;
2282 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2283 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2284 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2285 .csets = &tset.src_csets, \
2289 * cgroup_taskset_add - try to add a migration target task to a taskset
2290 * @task: target task
2291 * @tset: target taskset
2293 * Add @task, which is a migration target, to @tset. This function becomes
2294 * noop if @task doesn't need to be migrated. @task's css_set should have
2295 * been added as a migration source and @task->cg_list will be moved from
2296 * the css_set's tasks list to mg_tasks one.
2298 static void cgroup_taskset_add(struct task_struct *task,
2299 struct cgroup_taskset *tset)
2301 struct css_set *cset;
2303 lockdep_assert_held(&css_set_lock);
2305 /* @task either already exited or can't exit until the end */
2306 if (task->flags & PF_EXITING)
2309 /* leave @task alone if post_fork() hasn't linked it yet */
2310 if (list_empty(&task->cg_list))
2313 cset = task_css_set(task);
2314 if (!cset->mg_src_cgrp)
2317 list_move_tail(&task->cg_list, &cset->mg_tasks);
2318 if (list_empty(&cset->mg_node))
2319 list_add_tail(&cset->mg_node, &tset->src_csets);
2320 if (list_empty(&cset->mg_dst_cset->mg_node))
2321 list_move_tail(&cset->mg_dst_cset->mg_node,
2326 * cgroup_taskset_first - reset taskset and return the first task
2327 * @tset: taskset of interest
2328 * @dst_cssp: output variable for the destination css
2330 * @tset iteration is initialized and the first task is returned.
2332 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2333 struct cgroup_subsys_state **dst_cssp)
2335 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2336 tset->cur_task = NULL;
2338 return cgroup_taskset_next(tset, dst_cssp);
2342 * cgroup_taskset_next - iterate to the next task in taskset
2343 * @tset: taskset of interest
2344 * @dst_cssp: output variable for the destination css
2346 * Return the next task in @tset. Iteration must have been initialized
2347 * with cgroup_taskset_first().
2349 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2350 struct cgroup_subsys_state **dst_cssp)
2352 struct css_set *cset = tset->cur_cset;
2353 struct task_struct *task = tset->cur_task;
2355 while (&cset->mg_node != tset->csets) {
2357 task = list_first_entry(&cset->mg_tasks,
2358 struct task_struct, cg_list);
2360 task = list_next_entry(task, cg_list);
2362 if (&task->cg_list != &cset->mg_tasks) {
2363 tset->cur_cset = cset;
2364 tset->cur_task = task;
2367 * This function may be called both before and
2368 * after cgroup_taskset_migrate(). The two cases
2369 * can be distinguished by looking at whether @cset
2370 * has its ->mg_dst_cset set.
2372 if (cset->mg_dst_cset)
2373 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2375 *dst_cssp = cset->subsys[tset->ssid];
2380 cset = list_next_entry(cset, mg_node);
2388 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2389 * @tset: taget taskset
2390 * @dst_cgrp: destination cgroup
2392 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2393 * ->can_attach callbacks fails and guarantees that either all or none of
2394 * the tasks in @tset are migrated. @tset is consumed regardless of
2397 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2398 struct cgroup *dst_cgrp)
2400 struct cgroup_subsys_state *css, *failed_css = NULL;
2401 struct task_struct *task, *tmp_task;
2402 struct css_set *cset, *tmp_cset;
2405 /* methods shouldn't be called if no task is actually migrating */
2406 if (list_empty(&tset->src_csets))
2409 /* check that we can legitimately attach to the cgroup */
2410 for_each_e_css(css, i, dst_cgrp) {
2411 if (css->ss->can_attach) {
2413 ret = css->ss->can_attach(tset);
2416 goto out_cancel_attach;
2422 * Now that we're guaranteed success, proceed to move all tasks to
2423 * the new cgroup. There are no failure cases after here, so this
2424 * is the commit point.
2426 spin_lock_bh(&css_set_lock);
2427 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2428 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2429 struct css_set *from_cset = task_css_set(task);
2430 struct css_set *to_cset = cset->mg_dst_cset;
2432 get_css_set(to_cset);
2433 css_set_move_task(task, from_cset, to_cset, true);
2434 put_css_set_locked(from_cset);
2437 spin_unlock_bh(&css_set_lock);
2440 * Migration is committed, all target tasks are now on dst_csets.
2441 * Nothing is sensitive to fork() after this point. Notify
2442 * controllers that migration is complete.
2444 tset->csets = &tset->dst_csets;
2446 for_each_e_css(css, i, dst_cgrp) {
2447 if (css->ss->attach) {
2449 css->ss->attach(tset);
2454 goto out_release_tset;
2457 for_each_e_css(css, i, dst_cgrp) {
2458 if (css == failed_css)
2460 if (css->ss->cancel_attach) {
2462 css->ss->cancel_attach(tset);
2466 spin_lock_bh(&css_set_lock);
2467 list_splice_init(&tset->dst_csets, &tset->src_csets);
2468 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2469 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2470 list_del_init(&cset->mg_node);
2472 spin_unlock_bh(&css_set_lock);
2477 * cgroup_migrate_finish - cleanup after attach
2478 * @preloaded_csets: list of preloaded css_sets
2480 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2481 * those functions for details.
2483 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2485 struct css_set *cset, *tmp_cset;
2487 lockdep_assert_held(&cgroup_mutex);
2489 spin_lock_bh(&css_set_lock);
2490 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2491 cset->mg_src_cgrp = NULL;
2492 cset->mg_dst_cset = NULL;
2493 list_del_init(&cset->mg_preload_node);
2494 put_css_set_locked(cset);
2496 spin_unlock_bh(&css_set_lock);
2500 * cgroup_migrate_add_src - add a migration source css_set
2501 * @src_cset: the source css_set to add
2502 * @dst_cgrp: the destination cgroup
2503 * @preloaded_csets: list of preloaded css_sets
2505 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2506 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2507 * up by cgroup_migrate_finish().
2509 * This function may be called without holding cgroup_threadgroup_rwsem
2510 * even if the target is a process. Threads may be created and destroyed
2511 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2512 * into play and the preloaded css_sets are guaranteed to cover all
2515 static void cgroup_migrate_add_src(struct css_set *src_cset,
2516 struct cgroup *dst_cgrp,
2517 struct list_head *preloaded_csets)
2519 struct cgroup *src_cgrp;
2521 lockdep_assert_held(&cgroup_mutex);
2522 lockdep_assert_held(&css_set_lock);
2524 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2526 if (!list_empty(&src_cset->mg_preload_node))
2529 WARN_ON(src_cset->mg_src_cgrp);
2530 WARN_ON(!list_empty(&src_cset->mg_tasks));
2531 WARN_ON(!list_empty(&src_cset->mg_node));
2533 src_cset->mg_src_cgrp = src_cgrp;
2534 get_css_set(src_cset);
2535 list_add(&src_cset->mg_preload_node, preloaded_csets);
2539 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2540 * @dst_cgrp: the destination cgroup (may be %NULL)
2541 * @preloaded_csets: list of preloaded source css_sets
2543 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2544 * have been preloaded to @preloaded_csets. This function looks up and
2545 * pins all destination css_sets, links each to its source, and append them
2546 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2547 * source css_set is assumed to be its cgroup on the default hierarchy.
2549 * This function must be called after cgroup_migrate_add_src() has been
2550 * called on each migration source css_set. After migration is performed
2551 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2554 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2555 struct list_head *preloaded_csets)
2558 struct css_set *src_cset, *tmp_cset;
2560 lockdep_assert_held(&cgroup_mutex);
2563 * Except for the root, subtree_control must be zero for a cgroup
2564 * with tasks so that child cgroups don't compete against tasks.
2566 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2567 dst_cgrp->subtree_control)
2570 /* look up the dst cset for each src cset and link it to src */
2571 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2572 struct css_set *dst_cset;
2574 dst_cset = find_css_set(src_cset,
2575 dst_cgrp ?: src_cset->dfl_cgrp);
2579 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2582 * If src cset equals dst, it's noop. Drop the src.
2583 * cgroup_migrate() will skip the cset too. Note that we
2584 * can't handle src == dst as some nodes are used by both.
2586 if (src_cset == dst_cset) {
2587 src_cset->mg_src_cgrp = NULL;
2588 list_del_init(&src_cset->mg_preload_node);
2589 put_css_set(src_cset);
2590 put_css_set(dst_cset);
2594 src_cset->mg_dst_cset = dst_cset;
2596 if (list_empty(&dst_cset->mg_preload_node))
2597 list_add(&dst_cset->mg_preload_node, &csets);
2599 put_css_set(dst_cset);
2602 list_splice_tail(&csets, preloaded_csets);
2605 cgroup_migrate_finish(&csets);
2610 * cgroup_migrate - migrate a process or task to a cgroup
2611 * @leader: the leader of the process or the task to migrate
2612 * @threadgroup: whether @leader points to the whole process or a single task
2613 * @cgrp: the destination cgroup
2615 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2616 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2617 * caller is also responsible for invoking cgroup_migrate_add_src() and
2618 * cgroup_migrate_prepare_dst() on the targets before invoking this
2619 * function and following up with cgroup_migrate_finish().
2621 * As long as a controller's ->can_attach() doesn't fail, this function is
2622 * guaranteed to succeed. This means that, excluding ->can_attach()
2623 * failure, when migrating multiple targets, the success or failure can be
2624 * decided for all targets by invoking group_migrate_prepare_dst() before
2625 * actually starting migrating.
2627 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2628 struct cgroup *cgrp)
2630 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2631 struct task_struct *task;
2634 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2635 * already PF_EXITING could be freed from underneath us unless we
2636 * take an rcu_read_lock.
2638 spin_lock_bh(&css_set_lock);
2642 cgroup_taskset_add(task, &tset);
2645 } while_each_thread(leader, task);
2647 spin_unlock_bh(&css_set_lock);
2649 return cgroup_taskset_migrate(&tset, cgrp);
2653 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2654 * @dst_cgrp: the cgroup to attach to
2655 * @leader: the task or the leader of the threadgroup to be attached
2656 * @threadgroup: attach the whole threadgroup?
2658 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2660 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2661 struct task_struct *leader, bool threadgroup)
2663 LIST_HEAD(preloaded_csets);
2664 struct task_struct *task;
2667 /* look up all src csets */
2668 spin_lock_bh(&css_set_lock);
2672 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2676 } while_each_thread(leader, task);
2678 spin_unlock_bh(&css_set_lock);
2680 /* prepare dst csets and commit */
2681 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2683 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2685 cgroup_migrate_finish(&preloaded_csets);
2689 static int cgroup_procs_write_permission(struct task_struct *task,
2690 struct cgroup *dst_cgrp,
2691 struct kernfs_open_file *of)
2693 const struct cred *cred = current_cred();
2694 const struct cred *tcred = get_task_cred(task);
2698 * even if we're attaching all tasks in the thread group, we only
2699 * need to check permissions on one of them.
2701 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2702 !uid_eq(cred->euid, tcred->uid) &&
2703 !uid_eq(cred->euid, tcred->suid))
2706 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2707 struct super_block *sb = of->file->f_path.dentry->d_sb;
2708 struct cgroup *cgrp;
2709 struct inode *inode;
2711 spin_lock_bh(&css_set_lock);
2712 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2713 spin_unlock_bh(&css_set_lock);
2715 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2716 cgrp = cgroup_parent(cgrp);
2719 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2721 ret = inode_permission(inode, MAY_WRITE);
2731 * Find the task_struct of the task to attach by vpid and pass it along to the
2732 * function to attach either it or all tasks in its threadgroup. Will lock
2733 * cgroup_mutex and threadgroup.
2735 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2736 size_t nbytes, loff_t off, bool threadgroup)
2738 struct task_struct *tsk;
2739 struct cgroup *cgrp;
2743 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2746 cgrp = cgroup_kn_lock_live(of->kn);
2750 percpu_down_write(&cgroup_threadgroup_rwsem);
2753 tsk = find_task_by_vpid(pid);
2756 goto out_unlock_rcu;
2763 tsk = tsk->group_leader;
2766 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2767 * trapped in a cpuset, or RT worker may be born in a cgroup
2768 * with no rt_runtime allocated. Just say no.
2770 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2772 goto out_unlock_rcu;
2775 get_task_struct(tsk);
2778 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2780 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2782 put_task_struct(tsk);
2783 goto out_unlock_threadgroup;
2787 out_unlock_threadgroup:
2788 percpu_up_write(&cgroup_threadgroup_rwsem);
2789 cgroup_kn_unlock(of->kn);
2790 cpuset_post_attach_flush();
2791 return ret ?: nbytes;
2795 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2796 * @from: attach to all cgroups of a given task
2797 * @tsk: the task to be attached
2799 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2801 struct cgroup_root *root;
2804 mutex_lock(&cgroup_mutex);
2805 for_each_root(root) {
2806 struct cgroup *from_cgrp;
2808 if (root == &cgrp_dfl_root)
2811 spin_lock_bh(&css_set_lock);
2812 from_cgrp = task_cgroup_from_root(from, root);
2813 spin_unlock_bh(&css_set_lock);
2815 retval = cgroup_attach_task(from_cgrp, tsk, false);
2819 mutex_unlock(&cgroup_mutex);
2823 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2825 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2826 char *buf, size_t nbytes, loff_t off)
2828 return __cgroup_procs_write(of, buf, nbytes, off, false);
2831 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2832 char *buf, size_t nbytes, loff_t off)
2834 return __cgroup_procs_write(of, buf, nbytes, off, true);
2837 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2838 char *buf, size_t nbytes, loff_t off)
2840 struct cgroup *cgrp;
2842 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2844 cgrp = cgroup_kn_lock_live(of->kn);
2847 spin_lock(&release_agent_path_lock);
2848 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2849 sizeof(cgrp->root->release_agent_path));
2850 spin_unlock(&release_agent_path_lock);
2851 cgroup_kn_unlock(of->kn);
2855 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2857 struct cgroup *cgrp = seq_css(seq)->cgroup;
2859 spin_lock(&release_agent_path_lock);
2860 seq_puts(seq, cgrp->root->release_agent_path);
2861 spin_unlock(&release_agent_path_lock);
2862 seq_putc(seq, '\n');
2866 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2868 seq_puts(seq, "0\n");
2872 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2874 struct cgroup_subsys *ss;
2875 bool printed = false;
2878 do_each_subsys_mask(ss, ssid, ss_mask) {
2881 seq_printf(seq, "%s", ss->name);
2883 } while_each_subsys_mask();
2885 seq_putc(seq, '\n');
2888 /* show controllers which are enabled from the parent */
2889 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2891 struct cgroup *cgrp = seq_css(seq)->cgroup;
2893 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2897 /* show controllers which are enabled for a given cgroup's children */
2898 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2900 struct cgroup *cgrp = seq_css(seq)->cgroup;
2902 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2907 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2908 * @cgrp: root of the subtree to update csses for
2910 * @cgrp's subtree_ss_mask has changed and its subtree's (self excluded)
2911 * css associations need to be updated accordingly. This function looks up
2912 * all css_sets which are attached to the subtree, creates the matching
2913 * updated css_sets and migrates the tasks to the new ones.
2915 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2917 LIST_HEAD(preloaded_csets);
2918 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2919 struct cgroup_subsys_state *css;
2920 struct css_set *src_cset;
2923 lockdep_assert_held(&cgroup_mutex);
2925 percpu_down_write(&cgroup_threadgroup_rwsem);
2927 /* look up all csses currently attached to @cgrp's subtree */
2928 spin_lock_bh(&css_set_lock);
2929 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2930 struct cgrp_cset_link *link;
2932 /* self is not affected by subtree_ss_mask change */
2933 if (css->cgroup == cgrp)
2936 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2937 cgroup_migrate_add_src(link->cset, cgrp,
2940 spin_unlock_bh(&css_set_lock);
2942 /* NULL dst indicates self on default hierarchy */
2943 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2947 spin_lock_bh(&css_set_lock);
2948 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2949 struct task_struct *task, *ntask;
2951 /* src_csets precede dst_csets, break on the first dst_cset */
2952 if (!src_cset->mg_src_cgrp)
2955 /* all tasks in src_csets need to be migrated */
2956 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2957 cgroup_taskset_add(task, &tset);
2959 spin_unlock_bh(&css_set_lock);
2961 ret = cgroup_taskset_migrate(&tset, cgrp);
2963 cgroup_migrate_finish(&preloaded_csets);
2964 percpu_up_write(&cgroup_threadgroup_rwsem);
2969 * cgroup_drain_offline - wait for previously offlined csses to go away
2970 * @cgrp: parent of the target cgroups
2972 * Because css offlining is asynchronous, userland may try to re-enable a
2973 * controller while the previous css is still around. This function drains
2974 * the previous css instances of @cgrp's children.
2976 * Must be called with cgroup_mutex held. Returns %false if there were no
2977 * dying css instances. Returns %true if there were one or more and this
2978 * function waited. On %true return, cgroup_mutex has been dropped and
2979 * re-acquired inbetween which anything could have happened. The caller
2980 * typically would have to start over.
2982 static bool cgroup_drain_offline(struct cgroup *cgrp)
2984 struct cgroup *dsct;
2985 struct cgroup_subsys *ss;
2988 lockdep_assert_held(&cgroup_mutex);
2990 cgroup_for_each_live_child(dsct, cgrp) {
2991 for_each_subsys(ss, ssid) {
2992 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2999 prepare_to_wait(&dsct->offline_waitq, &wait,
3000 TASK_UNINTERRUPTIBLE);
3002 mutex_unlock(&cgroup_mutex);
3004 finish_wait(&dsct->offline_waitq, &wait);
3005 mutex_lock(&cgroup_mutex);
3016 * cgroup_apply_control_disable - kill or hide csses according to control
3017 * @cgrp: parent of the target cgroups
3019 * Walk @cgrp's children and kill and hide csses so that they match
3020 * cgroup_ss_mask() and cgroup_visible_mask().
3022 * A css is hidden when the userland requests it to be disabled while other
3023 * subsystems are still depending on it. The css must not actively control
3024 * resources and be in the vanilla state if it's made visible again later.
3025 * Controllers which may be depended upon should provide ->css_reset() for
3028 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3030 struct cgroup *dsct;
3031 struct cgroup_subsys *ss;
3034 cgroup_for_each_live_child(dsct, cgrp) {
3035 for_each_subsys(ss, ssid) {
3036 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3041 if (!(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3043 } else if (!(cgroup_control(dsct) & (1 << ss->id))) {
3044 css_clear_dir(css, NULL);
3052 /* change the enabled child controllers for a cgroup in the default hierarchy */
3053 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3054 char *buf, size_t nbytes,
3057 u16 enable = 0, disable = 0;
3058 u16 css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
3059 struct cgroup *cgrp, *child;
3060 struct cgroup_subsys *ss;
3065 * Parse input - space separated list of subsystem names prefixed
3066 * with either + or -.
3068 buf = strstrip(buf);
3069 while ((tok = strsep(&buf, " "))) {
3072 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3073 if (!cgroup_ssid_enabled(ssid) ||
3074 strcmp(tok + 1, ss->name))
3078 enable |= 1 << ssid;
3079 disable &= ~(1 << ssid);
3080 } else if (*tok == '-') {
3081 disable |= 1 << ssid;
3082 enable &= ~(1 << ssid);
3087 } while_each_subsys_mask();
3088 if (ssid == CGROUP_SUBSYS_COUNT)
3092 cgrp = cgroup_kn_lock_live(of->kn);
3096 for_each_subsys(ss, ssid) {
3097 if (enable & (1 << ssid)) {
3098 if (cgrp->subtree_control & (1 << ssid)) {
3099 enable &= ~(1 << ssid);
3103 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3107 } else if (disable & (1 << ssid)) {
3108 if (!(cgrp->subtree_control & (1 << ssid))) {
3109 disable &= ~(1 << ssid);
3113 /* a child has it enabled? */
3114 cgroup_for_each_live_child(child, cgrp) {
3115 if (child->subtree_control & (1 << ssid)) {
3123 if (!enable && !disable) {
3129 * Except for the root, subtree_control must be zero for a cgroup
3130 * with tasks so that child cgroups don't compete against tasks.
3132 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3137 if (cgroup_drain_offline(cgrp)) {
3138 cgroup_kn_unlock(of->kn);
3139 return restart_syscall();
3143 * Update subsys masks and calculate what needs to be done. More
3144 * subsystems than specified may need to be enabled or disabled
3145 * depending on subsystem dependencies.
3147 old_sc = cgrp->subtree_control;
3148 old_ss = cgrp->subtree_ss_mask;
3149 new_sc = (old_sc | enable) & ~disable;
3150 new_ss = cgroup_calc_subtree_ss_mask(cgrp, new_sc);
3152 css_enable = ~old_ss & new_ss;
3153 css_disable = old_ss & ~new_ss;
3154 enable |= css_enable;
3155 disable |= css_disable;
3157 cgrp->subtree_control = new_sc;
3158 cgrp->subtree_ss_mask = new_ss;
3161 * Create new csses or make the existing ones visible. A css is
3162 * created invisible if it's being implicitly enabled through
3163 * dependency. An invisible css is made visible when the userland
3164 * explicitly enables it.
3166 do_each_subsys_mask(ss, ssid, enable) {
3167 cgroup_for_each_live_child(child, cgrp) {
3168 if (css_enable & (1 << ssid)) {
3169 struct cgroup_subsys_state *css;
3171 css = css_create(child, ss);
3177 if (cgrp->subtree_control & (1 << ssid)) {
3178 ret = css_populate_dir(css, NULL);
3183 ret = css_populate_dir(cgroup_css(child, ss),
3189 } while_each_subsys_mask();
3192 * At this point, cgroup_e_css() results reflect the new csses
3193 * making the following cgroup_update_dfl_csses() properly update
3194 * css associations of all tasks in the subtree.
3196 ret = cgroup_update_dfl_csses(cgrp);
3200 /* all tasks are migrated out of disabled csses, commit disable */
3201 cgroup_apply_control_disable(cgrp);
3203 kernfs_activate(cgrp->kn);
3206 cgroup_kn_unlock(of->kn);
3207 return ret ?: nbytes;
3210 /* restore masks and shoot down new csses */
3211 cgrp->subtree_control = old_sc;
3212 cgrp->subtree_ss_mask = old_ss;
3214 cgroup_apply_control_disable(cgrp);
3219 static int cgroup_events_show(struct seq_file *seq, void *v)
3221 seq_printf(seq, "populated %d\n",
3222 cgroup_is_populated(seq_css(seq)->cgroup));
3226 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3227 size_t nbytes, loff_t off)
3229 struct cgroup *cgrp = of->kn->parent->priv;
3230 struct cftype *cft = of->kn->priv;
3231 struct cgroup_subsys_state *css;
3235 return cft->write(of, buf, nbytes, off);
3238 * kernfs guarantees that a file isn't deleted with operations in
3239 * flight, which means that the matching css is and stays alive and
3240 * doesn't need to be pinned. The RCU locking is not necessary
3241 * either. It's just for the convenience of using cgroup_css().
3244 css = cgroup_css(cgrp, cft->ss);
3247 if (cft->write_u64) {
3248 unsigned long long v;
3249 ret = kstrtoull(buf, 0, &v);
3251 ret = cft->write_u64(css, cft, v);
3252 } else if (cft->write_s64) {
3254 ret = kstrtoll(buf, 0, &v);
3256 ret = cft->write_s64(css, cft, v);
3261 return ret ?: nbytes;
3264 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3266 return seq_cft(seq)->seq_start(seq, ppos);
3269 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3271 return seq_cft(seq)->seq_next(seq, v, ppos);
3274 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3276 seq_cft(seq)->seq_stop(seq, v);
3279 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3281 struct cftype *cft = seq_cft(m);
3282 struct cgroup_subsys_state *css = seq_css(m);
3285 return cft->seq_show(m, arg);
3288 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3289 else if (cft->read_s64)
3290 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3296 static struct kernfs_ops cgroup_kf_single_ops = {
3297 .atomic_write_len = PAGE_SIZE,
3298 .write = cgroup_file_write,
3299 .seq_show = cgroup_seqfile_show,
3302 static struct kernfs_ops cgroup_kf_ops = {
3303 .atomic_write_len = PAGE_SIZE,
3304 .write = cgroup_file_write,
3305 .seq_start = cgroup_seqfile_start,
3306 .seq_next = cgroup_seqfile_next,
3307 .seq_stop = cgroup_seqfile_stop,
3308 .seq_show = cgroup_seqfile_show,
3312 * cgroup_rename - Only allow simple rename of directories in place.
3314 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3315 const char *new_name_str)
3317 struct cgroup *cgrp = kn->priv;
3320 if (kernfs_type(kn) != KERNFS_DIR)
3322 if (kn->parent != new_parent)
3326 * This isn't a proper migration and its usefulness is very
3327 * limited. Disallow on the default hierarchy.
3329 if (cgroup_on_dfl(cgrp))
3333 * We're gonna grab cgroup_mutex which nests outside kernfs
3334 * active_ref. kernfs_rename() doesn't require active_ref
3335 * protection. Break them before grabbing cgroup_mutex.
3337 kernfs_break_active_protection(new_parent);
3338 kernfs_break_active_protection(kn);
3340 mutex_lock(&cgroup_mutex);
3342 ret = kernfs_rename(kn, new_parent, new_name_str);
3344 mutex_unlock(&cgroup_mutex);
3346 kernfs_unbreak_active_protection(kn);
3347 kernfs_unbreak_active_protection(new_parent);
3351 /* set uid and gid of cgroup dirs and files to that of the creator */
3352 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3354 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3355 .ia_uid = current_fsuid(),
3356 .ia_gid = current_fsgid(), };
3358 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3359 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3362 return kernfs_setattr(kn, &iattr);
3365 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3368 char name[CGROUP_FILE_NAME_MAX];
3369 struct kernfs_node *kn;
3370 struct lock_class_key *key = NULL;
3373 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3374 key = &cft->lockdep_key;
3376 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3377 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3382 ret = cgroup_kn_set_ugid(kn);
3388 if (cft->file_offset) {
3389 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3391 spin_lock_irq(&cgroup_file_kn_lock);
3393 spin_unlock_irq(&cgroup_file_kn_lock);
3400 * cgroup_addrm_files - add or remove files to a cgroup directory
3401 * @css: the target css
3402 * @cgrp: the target cgroup (usually css->cgroup)
3403 * @cfts: array of cftypes to be added
3404 * @is_add: whether to add or remove
3406 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3407 * For removals, this function never fails.
3409 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3410 struct cgroup *cgrp, struct cftype cfts[],
3413 struct cftype *cft, *cft_end = NULL;
3416 lockdep_assert_held(&cgroup_mutex);
3419 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3420 /* does cft->flags tell us to skip this file on @cgrp? */
3421 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3423 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3425 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3427 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3431 ret = cgroup_add_file(css, cgrp, cft);
3433 pr_warn("%s: failed to add %s, err=%d\n",
3434 __func__, cft->name, ret);
3440 cgroup_rm_file(cgrp, cft);
3446 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3449 struct cgroup_subsys *ss = cfts[0].ss;
3450 struct cgroup *root = &ss->root->cgrp;
3451 struct cgroup_subsys_state *css;
3454 lockdep_assert_held(&cgroup_mutex);
3456 /* add/rm files for all cgroups created before */
3457 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3458 struct cgroup *cgrp = css->cgroup;
3460 if (!(css->flags & CSS_VISIBLE))
3463 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3469 kernfs_activate(root->kn);
3473 static void cgroup_exit_cftypes(struct cftype *cfts)
3477 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3478 /* free copy for custom atomic_write_len, see init_cftypes() */
3479 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3484 /* revert flags set by cgroup core while adding @cfts */
3485 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3489 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3493 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3494 struct kernfs_ops *kf_ops;
3496 WARN_ON(cft->ss || cft->kf_ops);
3499 kf_ops = &cgroup_kf_ops;
3501 kf_ops = &cgroup_kf_single_ops;
3504 * Ugh... if @cft wants a custom max_write_len, we need to
3505 * make a copy of kf_ops to set its atomic_write_len.
3507 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3508 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3510 cgroup_exit_cftypes(cfts);
3513 kf_ops->atomic_write_len = cft->max_write_len;
3516 cft->kf_ops = kf_ops;
3523 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3525 lockdep_assert_held(&cgroup_mutex);
3527 if (!cfts || !cfts[0].ss)
3530 list_del(&cfts->node);
3531 cgroup_apply_cftypes(cfts, false);
3532 cgroup_exit_cftypes(cfts);
3537 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3538 * @cfts: zero-length name terminated array of cftypes
3540 * Unregister @cfts. Files described by @cfts are removed from all
3541 * existing cgroups and all future cgroups won't have them either. This
3542 * function can be called anytime whether @cfts' subsys is attached or not.
3544 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3547 int cgroup_rm_cftypes(struct cftype *cfts)
3551 mutex_lock(&cgroup_mutex);
3552 ret = cgroup_rm_cftypes_locked(cfts);
3553 mutex_unlock(&cgroup_mutex);
3558 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3559 * @ss: target cgroup subsystem
3560 * @cfts: zero-length name terminated array of cftypes
3562 * Register @cfts to @ss. Files described by @cfts are created for all
3563 * existing cgroups to which @ss is attached and all future cgroups will
3564 * have them too. This function can be called anytime whether @ss is
3567 * Returns 0 on successful registration, -errno on failure. Note that this
3568 * function currently returns 0 as long as @cfts registration is successful
3569 * even if some file creation attempts on existing cgroups fail.
3571 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3575 if (!cgroup_ssid_enabled(ss->id))
3578 if (!cfts || cfts[0].name[0] == '\0')
3581 ret = cgroup_init_cftypes(ss, cfts);
3585 mutex_lock(&cgroup_mutex);
3587 list_add_tail(&cfts->node, &ss->cfts);
3588 ret = cgroup_apply_cftypes(cfts, true);
3590 cgroup_rm_cftypes_locked(cfts);
3592 mutex_unlock(&cgroup_mutex);
3597 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3598 * @ss: target cgroup subsystem
3599 * @cfts: zero-length name terminated array of cftypes
3601 * Similar to cgroup_add_cftypes() but the added files are only used for
3602 * the default hierarchy.
3604 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3608 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3609 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3610 return cgroup_add_cftypes(ss, cfts);
3614 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3615 * @ss: target cgroup subsystem
3616 * @cfts: zero-length name terminated array of cftypes
3618 * Similar to cgroup_add_cftypes() but the added files are only used for
3619 * the legacy hierarchies.
3621 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3625 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3626 cft->flags |= __CFTYPE_NOT_ON_DFL;
3627 return cgroup_add_cftypes(ss, cfts);
3631 * cgroup_file_notify - generate a file modified event for a cgroup_file
3632 * @cfile: target cgroup_file
3634 * @cfile must have been obtained by setting cftype->file_offset.
3636 void cgroup_file_notify(struct cgroup_file *cfile)
3638 unsigned long flags;
3640 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3642 kernfs_notify(cfile->kn);
3643 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3647 * cgroup_task_count - count the number of tasks in a cgroup.
3648 * @cgrp: the cgroup in question
3650 * Return the number of tasks in the cgroup.
3652 static int cgroup_task_count(const struct cgroup *cgrp)
3655 struct cgrp_cset_link *link;
3657 spin_lock_bh(&css_set_lock);
3658 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3659 count += atomic_read(&link->cset->refcount);
3660 spin_unlock_bh(&css_set_lock);
3665 * css_next_child - find the next child of a given css
3666 * @pos: the current position (%NULL to initiate traversal)
3667 * @parent: css whose children to walk
3669 * This function returns the next child of @parent and should be called
3670 * under either cgroup_mutex or RCU read lock. The only requirement is
3671 * that @parent and @pos are accessible. The next sibling is guaranteed to
3672 * be returned regardless of their states.
3674 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3675 * css which finished ->css_online() is guaranteed to be visible in the
3676 * future iterations and will stay visible until the last reference is put.
3677 * A css which hasn't finished ->css_online() or already finished
3678 * ->css_offline() may show up during traversal. It's each subsystem's
3679 * responsibility to synchronize against on/offlining.
3681 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3682 struct cgroup_subsys_state *parent)
3684 struct cgroup_subsys_state *next;
3686 cgroup_assert_mutex_or_rcu_locked();
3689 * @pos could already have been unlinked from the sibling list.
3690 * Once a cgroup is removed, its ->sibling.next is no longer
3691 * updated when its next sibling changes. CSS_RELEASED is set when
3692 * @pos is taken off list, at which time its next pointer is valid,
3693 * and, as releases are serialized, the one pointed to by the next
3694 * pointer is guaranteed to not have started release yet. This
3695 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3696 * critical section, the one pointed to by its next pointer is
3697 * guaranteed to not have finished its RCU grace period even if we
3698 * have dropped rcu_read_lock() inbetween iterations.
3700 * If @pos has CSS_RELEASED set, its next pointer can't be
3701 * dereferenced; however, as each css is given a monotonically
3702 * increasing unique serial number and always appended to the
3703 * sibling list, the next one can be found by walking the parent's
3704 * children until the first css with higher serial number than
3705 * @pos's. While this path can be slower, it happens iff iteration
3706 * races against release and the race window is very small.
3709 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3710 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3711 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3713 list_for_each_entry_rcu(next, &parent->children, sibling)
3714 if (next->serial_nr > pos->serial_nr)
3719 * @next, if not pointing to the head, can be dereferenced and is
3722 if (&next->sibling != &parent->children)
3728 * css_next_descendant_pre - find the next descendant for pre-order walk
3729 * @pos: the current position (%NULL to initiate traversal)
3730 * @root: css whose descendants to walk
3732 * To be used by css_for_each_descendant_pre(). Find the next descendant
3733 * to visit for pre-order traversal of @root's descendants. @root is
3734 * included in the iteration and the first node to be visited.
3736 * While this function requires cgroup_mutex or RCU read locking, it
3737 * doesn't require the whole traversal to be contained in a single critical
3738 * section. This function will return the correct next descendant as long
3739 * as both @pos and @root are accessible and @pos is a descendant of @root.
3741 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3742 * css which finished ->css_online() is guaranteed to be visible in the
3743 * future iterations and will stay visible until the last reference is put.
3744 * A css which hasn't finished ->css_online() or already finished
3745 * ->css_offline() may show up during traversal. It's each subsystem's
3746 * responsibility to synchronize against on/offlining.
3748 struct cgroup_subsys_state *
3749 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3750 struct cgroup_subsys_state *root)
3752 struct cgroup_subsys_state *next;
3754 cgroup_assert_mutex_or_rcu_locked();
3756 /* if first iteration, visit @root */
3760 /* visit the first child if exists */
3761 next = css_next_child(NULL, pos);
3765 /* no child, visit my or the closest ancestor's next sibling */
3766 while (pos != root) {
3767 next = css_next_child(pos, pos->parent);
3777 * css_rightmost_descendant - return the rightmost descendant of a css
3778 * @pos: css of interest
3780 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3781 * is returned. This can be used during pre-order traversal to skip
3784 * While this function requires cgroup_mutex or RCU read locking, it
3785 * doesn't require the whole traversal to be contained in a single critical
3786 * section. This function will return the correct rightmost descendant as
3787 * long as @pos is accessible.
3789 struct cgroup_subsys_state *
3790 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3792 struct cgroup_subsys_state *last, *tmp;
3794 cgroup_assert_mutex_or_rcu_locked();
3798 /* ->prev isn't RCU safe, walk ->next till the end */
3800 css_for_each_child(tmp, last)
3807 static struct cgroup_subsys_state *
3808 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3810 struct cgroup_subsys_state *last;
3814 pos = css_next_child(NULL, pos);
3821 * css_next_descendant_post - find the next descendant for post-order walk
3822 * @pos: the current position (%NULL to initiate traversal)
3823 * @root: css whose descendants to walk
3825 * To be used by css_for_each_descendant_post(). Find the next descendant
3826 * to visit for post-order traversal of @root's descendants. @root is
3827 * included in the iteration and the last node to be visited.
3829 * While this function requires cgroup_mutex or RCU read locking, it
3830 * doesn't require the whole traversal to be contained in a single critical
3831 * section. This function will return the correct next descendant as long
3832 * as both @pos and @cgroup are accessible and @pos is a descendant of
3835 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3836 * css which finished ->css_online() is guaranteed to be visible in the
3837 * future iterations and will stay visible until the last reference is put.
3838 * A css which hasn't finished ->css_online() or already finished
3839 * ->css_offline() may show up during traversal. It's each subsystem's
3840 * responsibility to synchronize against on/offlining.
3842 struct cgroup_subsys_state *
3843 css_next_descendant_post(struct cgroup_subsys_state *pos,
3844 struct cgroup_subsys_state *root)
3846 struct cgroup_subsys_state *next;
3848 cgroup_assert_mutex_or_rcu_locked();
3850 /* if first iteration, visit leftmost descendant which may be @root */
3852 return css_leftmost_descendant(root);
3854 /* if we visited @root, we're done */
3858 /* if there's an unvisited sibling, visit its leftmost descendant */
3859 next = css_next_child(pos, pos->parent);
3861 return css_leftmost_descendant(next);
3863 /* no sibling left, visit parent */
3868 * css_has_online_children - does a css have online children
3869 * @css: the target css
3871 * Returns %true if @css has any online children; otherwise, %false. This
3872 * function can be called from any context but the caller is responsible
3873 * for synchronizing against on/offlining as necessary.
3875 bool css_has_online_children(struct cgroup_subsys_state *css)
3877 struct cgroup_subsys_state *child;
3881 css_for_each_child(child, css) {
3882 if (child->flags & CSS_ONLINE) {
3892 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3893 * @it: the iterator to advance
3895 * Advance @it to the next css_set to walk.
3897 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3899 struct list_head *l = it->cset_pos;
3900 struct cgrp_cset_link *link;
3901 struct css_set *cset;
3903 lockdep_assert_held(&css_set_lock);
3905 /* Advance to the next non-empty css_set */
3908 if (l == it->cset_head) {
3909 it->cset_pos = NULL;
3910 it->task_pos = NULL;
3915 cset = container_of(l, struct css_set,
3916 e_cset_node[it->ss->id]);
3918 link = list_entry(l, struct cgrp_cset_link, cset_link);
3921 } while (!css_set_populated(cset));
3925 if (!list_empty(&cset->tasks))
3926 it->task_pos = cset->tasks.next;
3928 it->task_pos = cset->mg_tasks.next;
3930 it->tasks_head = &cset->tasks;
3931 it->mg_tasks_head = &cset->mg_tasks;
3934 * We don't keep css_sets locked across iteration steps and thus
3935 * need to take steps to ensure that iteration can be resumed after
3936 * the lock is re-acquired. Iteration is performed at two levels -
3937 * css_sets and tasks in them.
3939 * Once created, a css_set never leaves its cgroup lists, so a
3940 * pinned css_set is guaranteed to stay put and we can resume
3941 * iteration afterwards.
3943 * Tasks may leave @cset across iteration steps. This is resolved
3944 * by registering each iterator with the css_set currently being
3945 * walked and making css_set_move_task() advance iterators whose
3946 * next task is leaving.
3949 list_del(&it->iters_node);
3950 put_css_set_locked(it->cur_cset);
3953 it->cur_cset = cset;
3954 list_add(&it->iters_node, &cset->task_iters);
3957 static void css_task_iter_advance(struct css_task_iter *it)
3959 struct list_head *l = it->task_pos;
3961 lockdep_assert_held(&css_set_lock);
3965 * Advance iterator to find next entry. cset->tasks is consumed
3966 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3971 if (l == it->tasks_head)
3972 l = it->mg_tasks_head->next;
3974 if (l == it->mg_tasks_head)
3975 css_task_iter_advance_css_set(it);
3981 * css_task_iter_start - initiate task iteration
3982 * @css: the css to walk tasks of
3983 * @it: the task iterator to use
3985 * Initiate iteration through the tasks of @css. The caller can call
3986 * css_task_iter_next() to walk through the tasks until the function
3987 * returns NULL. On completion of iteration, css_task_iter_end() must be
3990 void css_task_iter_start(struct cgroup_subsys_state *css,
3991 struct css_task_iter *it)
3993 /* no one should try to iterate before mounting cgroups */
3994 WARN_ON_ONCE(!use_task_css_set_links);
3996 memset(it, 0, sizeof(*it));
3998 spin_lock_bh(&css_set_lock);
4003 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4005 it->cset_pos = &css->cgroup->cset_links;
4007 it->cset_head = it->cset_pos;
4009 css_task_iter_advance_css_set(it);
4011 spin_unlock_bh(&css_set_lock);
4015 * css_task_iter_next - return the next task for the iterator
4016 * @it: the task iterator being iterated
4018 * The "next" function for task iteration. @it should have been
4019 * initialized via css_task_iter_start(). Returns NULL when the iteration
4022 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4025 put_task_struct(it->cur_task);
4026 it->cur_task = NULL;
4029 spin_lock_bh(&css_set_lock);
4032 it->cur_task = list_entry(it->task_pos, struct task_struct,
4034 get_task_struct(it->cur_task);
4035 css_task_iter_advance(it);
4038 spin_unlock_bh(&css_set_lock);
4040 return it->cur_task;
4044 * css_task_iter_end - finish task iteration
4045 * @it: the task iterator to finish
4047 * Finish task iteration started by css_task_iter_start().
4049 void css_task_iter_end(struct css_task_iter *it)
4052 spin_lock_bh(&css_set_lock);
4053 list_del(&it->iters_node);
4054 put_css_set_locked(it->cur_cset);
4055 spin_unlock_bh(&css_set_lock);
4059 put_task_struct(it->cur_task);
4063 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4064 * @to: cgroup to which the tasks will be moved
4065 * @from: cgroup in which the tasks currently reside
4067 * Locking rules between cgroup_post_fork() and the migration path
4068 * guarantee that, if a task is forking while being migrated, the new child
4069 * is guaranteed to be either visible in the source cgroup after the
4070 * parent's migration is complete or put into the target cgroup. No task
4071 * can slip out of migration through forking.
4073 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4075 LIST_HEAD(preloaded_csets);
4076 struct cgrp_cset_link *link;
4077 struct css_task_iter it;
4078 struct task_struct *task;
4081 mutex_lock(&cgroup_mutex);
4083 /* all tasks in @from are being moved, all csets are source */
4084 spin_lock_bh(&css_set_lock);
4085 list_for_each_entry(link, &from->cset_links, cset_link)
4086 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4087 spin_unlock_bh(&css_set_lock);
4089 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
4094 * Migrate tasks one-by-one until @from is empty. This fails iff
4095 * ->can_attach() fails.
4098 css_task_iter_start(&from->self, &it);
4099 task = css_task_iter_next(&it);
4101 get_task_struct(task);
4102 css_task_iter_end(&it);
4105 ret = cgroup_migrate(task, false, to);
4106 put_task_struct(task);
4108 } while (task && !ret);
4110 cgroup_migrate_finish(&preloaded_csets);
4111 mutex_unlock(&cgroup_mutex);
4116 * Stuff for reading the 'tasks'/'procs' files.
4118 * Reading this file can return large amounts of data if a cgroup has
4119 * *lots* of attached tasks. So it may need several calls to read(),
4120 * but we cannot guarantee that the information we produce is correct
4121 * unless we produce it entirely atomically.
4125 /* which pidlist file are we talking about? */
4126 enum cgroup_filetype {
4132 * A pidlist is a list of pids that virtually represents the contents of one
4133 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4134 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4137 struct cgroup_pidlist {
4139 * used to find which pidlist is wanted. doesn't change as long as
4140 * this particular list stays in the list.
4142 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4145 /* how many elements the above list has */
4147 /* each of these stored in a list by its cgroup */
4148 struct list_head links;
4149 /* pointer to the cgroup we belong to, for list removal purposes */
4150 struct cgroup *owner;
4151 /* for delayed destruction */
4152 struct delayed_work destroy_dwork;
4156 * The following two functions "fix" the issue where there are more pids
4157 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4158 * TODO: replace with a kernel-wide solution to this problem
4160 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4161 static void *pidlist_allocate(int count)
4163 if (PIDLIST_TOO_LARGE(count))
4164 return vmalloc(count * sizeof(pid_t));
4166 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4169 static void pidlist_free(void *p)
4175 * Used to destroy all pidlists lingering waiting for destroy timer. None
4176 * should be left afterwards.
4178 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4180 struct cgroup_pidlist *l, *tmp_l;
4182 mutex_lock(&cgrp->pidlist_mutex);
4183 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4184 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4185 mutex_unlock(&cgrp->pidlist_mutex);
4187 flush_workqueue(cgroup_pidlist_destroy_wq);
4188 BUG_ON(!list_empty(&cgrp->pidlists));
4191 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4193 struct delayed_work *dwork = to_delayed_work(work);
4194 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4196 struct cgroup_pidlist *tofree = NULL;
4198 mutex_lock(&l->owner->pidlist_mutex);
4201 * Destroy iff we didn't get queued again. The state won't change
4202 * as destroy_dwork can only be queued while locked.
4204 if (!delayed_work_pending(dwork)) {
4205 list_del(&l->links);
4206 pidlist_free(l->list);
4207 put_pid_ns(l->key.ns);
4211 mutex_unlock(&l->owner->pidlist_mutex);
4216 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4217 * Returns the number of unique elements.
4219 static int pidlist_uniq(pid_t *list, int length)
4224 * we presume the 0th element is unique, so i starts at 1. trivial
4225 * edge cases first; no work needs to be done for either
4227 if (length == 0 || length == 1)
4229 /* src and dest walk down the list; dest counts unique elements */
4230 for (src = 1; src < length; src++) {
4231 /* find next unique element */
4232 while (list[src] == list[src-1]) {
4237 /* dest always points to where the next unique element goes */
4238 list[dest] = list[src];
4246 * The two pid files - task and cgroup.procs - guaranteed that the result
4247 * is sorted, which forced this whole pidlist fiasco. As pid order is
4248 * different per namespace, each namespace needs differently sorted list,
4249 * making it impossible to use, for example, single rbtree of member tasks
4250 * sorted by task pointer. As pidlists can be fairly large, allocating one
4251 * per open file is dangerous, so cgroup had to implement shared pool of
4252 * pidlists keyed by cgroup and namespace.
4254 * All this extra complexity was caused by the original implementation
4255 * committing to an entirely unnecessary property. In the long term, we
4256 * want to do away with it. Explicitly scramble sort order if on the
4257 * default hierarchy so that no such expectation exists in the new
4260 * Scrambling is done by swapping every two consecutive bits, which is
4261 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4263 static pid_t pid_fry(pid_t pid)
4265 unsigned a = pid & 0x55555555;
4266 unsigned b = pid & 0xAAAAAAAA;
4268 return (a << 1) | (b >> 1);
4271 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4273 if (cgroup_on_dfl(cgrp))
4274 return pid_fry(pid);
4279 static int cmppid(const void *a, const void *b)
4281 return *(pid_t *)a - *(pid_t *)b;
4284 static int fried_cmppid(const void *a, const void *b)
4286 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4289 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4290 enum cgroup_filetype type)
4292 struct cgroup_pidlist *l;
4293 /* don't need task_nsproxy() if we're looking at ourself */
4294 struct pid_namespace *ns = task_active_pid_ns(current);
4296 lockdep_assert_held(&cgrp->pidlist_mutex);
4298 list_for_each_entry(l, &cgrp->pidlists, links)
4299 if (l->key.type == type && l->key.ns == ns)
4305 * find the appropriate pidlist for our purpose (given procs vs tasks)
4306 * returns with the lock on that pidlist already held, and takes care
4307 * of the use count, or returns NULL with no locks held if we're out of
4310 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4311 enum cgroup_filetype type)
4313 struct cgroup_pidlist *l;
4315 lockdep_assert_held(&cgrp->pidlist_mutex);
4317 l = cgroup_pidlist_find(cgrp, type);
4321 /* entry not found; create a new one */
4322 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4326 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4328 /* don't need task_nsproxy() if we're looking at ourself */
4329 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4331 list_add(&l->links, &cgrp->pidlists);
4336 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4338 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4339 struct cgroup_pidlist **lp)
4343 int pid, n = 0; /* used for populating the array */
4344 struct css_task_iter it;
4345 struct task_struct *tsk;
4346 struct cgroup_pidlist *l;
4348 lockdep_assert_held(&cgrp->pidlist_mutex);
4351 * If cgroup gets more users after we read count, we won't have
4352 * enough space - tough. This race is indistinguishable to the
4353 * caller from the case that the additional cgroup users didn't
4354 * show up until sometime later on.
4356 length = cgroup_task_count(cgrp);
4357 array = pidlist_allocate(length);
4360 /* now, populate the array */
4361 css_task_iter_start(&cgrp->self, &it);
4362 while ((tsk = css_task_iter_next(&it))) {
4363 if (unlikely(n == length))
4365 /* get tgid or pid for procs or tasks file respectively */
4366 if (type == CGROUP_FILE_PROCS)
4367 pid = task_tgid_vnr(tsk);
4369 pid = task_pid_vnr(tsk);
4370 if (pid > 0) /* make sure to only use valid results */
4373 css_task_iter_end(&it);
4375 /* now sort & (if procs) strip out duplicates */
4376 if (cgroup_on_dfl(cgrp))
4377 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4379 sort(array, length, sizeof(pid_t), cmppid, NULL);
4380 if (type == CGROUP_FILE_PROCS)
4381 length = pidlist_uniq(array, length);
4383 l = cgroup_pidlist_find_create(cgrp, type);
4385 pidlist_free(array);
4389 /* store array, freeing old if necessary */
4390 pidlist_free(l->list);
4398 * cgroupstats_build - build and fill cgroupstats
4399 * @stats: cgroupstats to fill information into
4400 * @dentry: A dentry entry belonging to the cgroup for which stats have
4403 * Build and fill cgroupstats so that taskstats can export it to user
4406 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4408 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4409 struct cgroup *cgrp;
4410 struct css_task_iter it;
4411 struct task_struct *tsk;
4413 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4414 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4415 kernfs_type(kn) != KERNFS_DIR)
4418 mutex_lock(&cgroup_mutex);
4421 * We aren't being called from kernfs and there's no guarantee on
4422 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4423 * @kn->priv is RCU safe. Let's do the RCU dancing.
4426 cgrp = rcu_dereference(kn->priv);
4427 if (!cgrp || cgroup_is_dead(cgrp)) {
4429 mutex_unlock(&cgroup_mutex);
4434 css_task_iter_start(&cgrp->self, &it);
4435 while ((tsk = css_task_iter_next(&it))) {
4436 switch (tsk->state) {
4438 stats->nr_running++;
4440 case TASK_INTERRUPTIBLE:
4441 stats->nr_sleeping++;
4443 case TASK_UNINTERRUPTIBLE:
4444 stats->nr_uninterruptible++;
4447 stats->nr_stopped++;
4450 if (delayacct_is_task_waiting_on_io(tsk))
4451 stats->nr_io_wait++;
4455 css_task_iter_end(&it);
4457 mutex_unlock(&cgroup_mutex);
4463 * seq_file methods for the tasks/procs files. The seq_file position is the
4464 * next pid to display; the seq_file iterator is a pointer to the pid
4465 * in the cgroup->l->list array.
4468 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4471 * Initially we receive a position value that corresponds to
4472 * one more than the last pid shown (or 0 on the first call or
4473 * after a seek to the start). Use a binary-search to find the
4474 * next pid to display, if any
4476 struct kernfs_open_file *of = s->private;
4477 struct cgroup *cgrp = seq_css(s)->cgroup;
4478 struct cgroup_pidlist *l;
4479 enum cgroup_filetype type = seq_cft(s)->private;
4480 int index = 0, pid = *pos;
4483 mutex_lock(&cgrp->pidlist_mutex);
4486 * !NULL @of->priv indicates that this isn't the first start()
4487 * after open. If the matching pidlist is around, we can use that.
4488 * Look for it. Note that @of->priv can't be used directly. It
4489 * could already have been destroyed.
4492 of->priv = cgroup_pidlist_find(cgrp, type);
4495 * Either this is the first start() after open or the matching
4496 * pidlist has been destroyed inbetween. Create a new one.
4499 ret = pidlist_array_load(cgrp, type,
4500 (struct cgroup_pidlist **)&of->priv);
4502 return ERR_PTR(ret);
4507 int end = l->length;
4509 while (index < end) {
4510 int mid = (index + end) / 2;
4511 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4514 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4520 /* If we're off the end of the array, we're done */
4521 if (index >= l->length)
4523 /* Update the abstract position to be the actual pid that we found */
4524 iter = l->list + index;
4525 *pos = cgroup_pid_fry(cgrp, *iter);
4529 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4531 struct kernfs_open_file *of = s->private;
4532 struct cgroup_pidlist *l = of->priv;
4535 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4536 CGROUP_PIDLIST_DESTROY_DELAY);
4537 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4540 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4542 struct kernfs_open_file *of = s->private;
4543 struct cgroup_pidlist *l = of->priv;
4545 pid_t *end = l->list + l->length;
4547 * Advance to the next pid in the array. If this goes off the
4554 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4559 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4561 seq_printf(s, "%d\n", *(int *)v);
4566 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4569 return notify_on_release(css->cgroup);
4572 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4573 struct cftype *cft, u64 val)
4576 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4578 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4582 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4585 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4588 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4589 struct cftype *cft, u64 val)
4592 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4594 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4598 /* cgroup core interface files for the default hierarchy */
4599 static struct cftype cgroup_dfl_base_files[] = {
4601 .name = "cgroup.procs",
4602 .file_offset = offsetof(struct cgroup, procs_file),
4603 .seq_start = cgroup_pidlist_start,
4604 .seq_next = cgroup_pidlist_next,
4605 .seq_stop = cgroup_pidlist_stop,
4606 .seq_show = cgroup_pidlist_show,
4607 .private = CGROUP_FILE_PROCS,
4608 .write = cgroup_procs_write,
4611 .name = "cgroup.controllers",
4612 .seq_show = cgroup_controllers_show,
4615 .name = "cgroup.subtree_control",
4616 .seq_show = cgroup_subtree_control_show,
4617 .write = cgroup_subtree_control_write,
4620 .name = "cgroup.events",
4621 .flags = CFTYPE_NOT_ON_ROOT,
4622 .file_offset = offsetof(struct cgroup, events_file),
4623 .seq_show = cgroup_events_show,
4628 /* cgroup core interface files for the legacy hierarchies */
4629 static struct cftype cgroup_legacy_base_files[] = {
4631 .name = "cgroup.procs",
4632 .seq_start = cgroup_pidlist_start,
4633 .seq_next = cgroup_pidlist_next,
4634 .seq_stop = cgroup_pidlist_stop,
4635 .seq_show = cgroup_pidlist_show,
4636 .private = CGROUP_FILE_PROCS,
4637 .write = cgroup_procs_write,
4640 .name = "cgroup.clone_children",
4641 .read_u64 = cgroup_clone_children_read,
4642 .write_u64 = cgroup_clone_children_write,
4645 .name = "cgroup.sane_behavior",
4646 .flags = CFTYPE_ONLY_ON_ROOT,
4647 .seq_show = cgroup_sane_behavior_show,
4651 .seq_start = cgroup_pidlist_start,
4652 .seq_next = cgroup_pidlist_next,
4653 .seq_stop = cgroup_pidlist_stop,
4654 .seq_show = cgroup_pidlist_show,
4655 .private = CGROUP_FILE_TASKS,
4656 .write = cgroup_tasks_write,
4659 .name = "notify_on_release",
4660 .read_u64 = cgroup_read_notify_on_release,
4661 .write_u64 = cgroup_write_notify_on_release,
4664 .name = "release_agent",
4665 .flags = CFTYPE_ONLY_ON_ROOT,
4666 .seq_show = cgroup_release_agent_show,
4667 .write = cgroup_release_agent_write,
4668 .max_write_len = PATH_MAX - 1,
4674 * css destruction is four-stage process.
4676 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4677 * Implemented in kill_css().
4679 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4680 * and thus css_tryget_online() is guaranteed to fail, the css can be
4681 * offlined by invoking offline_css(). After offlining, the base ref is
4682 * put. Implemented in css_killed_work_fn().
4684 * 3. When the percpu_ref reaches zero, the only possible remaining
4685 * accessors are inside RCU read sections. css_release() schedules the
4688 * 4. After the grace period, the css can be freed. Implemented in
4689 * css_free_work_fn().
4691 * It is actually hairier because both step 2 and 4 require process context
4692 * and thus involve punting to css->destroy_work adding two additional
4693 * steps to the already complex sequence.
4695 static void css_free_work_fn(struct work_struct *work)
4697 struct cgroup_subsys_state *css =
4698 container_of(work, struct cgroup_subsys_state, destroy_work);
4699 struct cgroup_subsys *ss = css->ss;
4700 struct cgroup *cgrp = css->cgroup;
4702 percpu_ref_exit(&css->refcnt);
4706 struct cgroup_subsys_state *parent = css->parent;
4710 cgroup_idr_remove(&ss->css_idr, id);
4716 /* cgroup free path */
4717 atomic_dec(&cgrp->root->nr_cgrps);
4718 cgroup_pidlist_destroy_all(cgrp);
4719 cancel_work_sync(&cgrp->release_agent_work);
4721 if (cgroup_parent(cgrp)) {
4723 * We get a ref to the parent, and put the ref when
4724 * this cgroup is being freed, so it's guaranteed
4725 * that the parent won't be destroyed before its
4728 cgroup_put(cgroup_parent(cgrp));
4729 kernfs_put(cgrp->kn);
4733 * This is root cgroup's refcnt reaching zero,
4734 * which indicates that the root should be
4737 cgroup_destroy_root(cgrp->root);
4742 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4744 struct cgroup_subsys_state *css =
4745 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4747 INIT_WORK(&css->destroy_work, css_free_work_fn);
4748 queue_work(cgroup_destroy_wq, &css->destroy_work);
4751 static void css_release_work_fn(struct work_struct *work)
4753 struct cgroup_subsys_state *css =
4754 container_of(work, struct cgroup_subsys_state, destroy_work);
4755 struct cgroup_subsys *ss = css->ss;
4756 struct cgroup *cgrp = css->cgroup;
4758 mutex_lock(&cgroup_mutex);
4760 css->flags |= CSS_RELEASED;
4761 list_del_rcu(&css->sibling);
4764 /* css release path */
4765 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4766 if (ss->css_released)
4767 ss->css_released(css);
4769 /* cgroup release path */
4770 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4774 * There are two control paths which try to determine
4775 * cgroup from dentry without going through kernfs -
4776 * cgroupstats_build() and css_tryget_online_from_dir().
4777 * Those are supported by RCU protecting clearing of
4778 * cgrp->kn->priv backpointer.
4781 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4785 mutex_unlock(&cgroup_mutex);
4787 call_rcu(&css->rcu_head, css_free_rcu_fn);
4790 static void css_release(struct percpu_ref *ref)
4792 struct cgroup_subsys_state *css =
4793 container_of(ref, struct cgroup_subsys_state, refcnt);
4795 INIT_WORK(&css->destroy_work, css_release_work_fn);
4796 queue_work(cgroup_destroy_wq, &css->destroy_work);
4799 static void init_and_link_css(struct cgroup_subsys_state *css,
4800 struct cgroup_subsys *ss, struct cgroup *cgrp)
4802 lockdep_assert_held(&cgroup_mutex);
4806 memset(css, 0, sizeof(*css));
4809 INIT_LIST_HEAD(&css->sibling);
4810 INIT_LIST_HEAD(&css->children);
4811 css->serial_nr = css_serial_nr_next++;
4812 atomic_set(&css->online_cnt, 0);
4814 if (cgroup_parent(cgrp)) {
4815 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4816 css_get(css->parent);
4819 BUG_ON(cgroup_css(cgrp, ss));
4822 /* invoke ->css_online() on a new CSS and mark it online if successful */
4823 static int online_css(struct cgroup_subsys_state *css)
4825 struct cgroup_subsys *ss = css->ss;
4828 lockdep_assert_held(&cgroup_mutex);
4831 ret = ss->css_online(css);
4833 css->flags |= CSS_ONLINE;
4834 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4836 atomic_inc(&css->online_cnt);
4838 atomic_inc(&css->parent->online_cnt);
4843 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4844 static void offline_css(struct cgroup_subsys_state *css)
4846 struct cgroup_subsys *ss = css->ss;
4848 lockdep_assert_held(&cgroup_mutex);
4850 if (!(css->flags & CSS_ONLINE))
4856 if (ss->css_offline)
4857 ss->css_offline(css);
4859 css->flags &= ~CSS_ONLINE;
4860 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4862 wake_up_all(&css->cgroup->offline_waitq);
4866 * css_create - create a cgroup_subsys_state
4867 * @cgrp: the cgroup new css will be associated with
4868 * @ss: the subsys of new css
4870 * Create a new css associated with @cgrp - @ss pair. On success, the new
4871 * css is online and installed in @cgrp. This function doesn't create the
4872 * interface files. Returns 0 on success, -errno on failure.
4874 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4875 struct cgroup_subsys *ss)
4877 struct cgroup *parent = cgroup_parent(cgrp);
4878 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4879 struct cgroup_subsys_state *css;
4882 lockdep_assert_held(&cgroup_mutex);
4884 css = ss->css_alloc(parent_css);
4888 init_and_link_css(css, ss, cgrp);
4890 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4894 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4896 goto err_free_percpu_ref;
4899 /* @css is ready to be brought online now, make it visible */
4900 list_add_tail_rcu(&css->sibling, &parent_css->children);
4901 cgroup_idr_replace(&ss->css_idr, css, css->id);
4903 err = online_css(css);
4907 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4908 cgroup_parent(parent)) {
4909 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4910 current->comm, current->pid, ss->name);
4911 if (!strcmp(ss->name, "memory"))
4912 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4913 ss->warned_broken_hierarchy = true;
4919 list_del_rcu(&css->sibling);
4920 cgroup_idr_remove(&ss->css_idr, css->id);
4921 err_free_percpu_ref:
4922 percpu_ref_exit(&css->refcnt);
4924 call_rcu(&css->rcu_head, css_free_rcu_fn);
4925 return ERR_PTR(err);
4928 static struct cgroup *cgroup_create(struct cgroup *parent)
4930 struct cgroup_root *root = parent->root;
4931 struct cgroup_subsys *ss;
4932 struct cgroup *cgrp, *tcgrp;
4933 int level = parent->level + 1;
4936 /* allocate the cgroup and its ID, 0 is reserved for the root */
4937 cgrp = kzalloc(sizeof(*cgrp) +
4938 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4940 return ERR_PTR(-ENOMEM);
4942 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4947 * Temporarily set the pointer to NULL, so idr_find() won't return
4948 * a half-baked cgroup.
4950 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4953 goto out_cancel_ref;
4956 init_cgroup_housekeeping(cgrp);
4958 cgrp->self.parent = &parent->self;
4960 cgrp->level = level;
4962 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
4963 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4965 if (notify_on_release(parent))
4966 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4968 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4969 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4971 cgrp->self.serial_nr = css_serial_nr_next++;
4973 /* allocation complete, commit to creation */
4974 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4975 atomic_inc(&root->nr_cgrps);
4979 * @cgrp is now fully operational. If something fails after this
4980 * point, it'll be released via the normal destruction path.
4982 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4984 /* create the csses */
4985 do_each_subsys_mask(ss, ssid, cgroup_ss_mask(cgrp)) {
4986 struct cgroup_subsys_state *css;
4988 css = css_create(cgrp, ss);
4993 } while_each_subsys_mask();
4996 * On the default hierarchy, a child doesn't automatically inherit
4997 * subtree_control from the parent. Each is configured manually.
4999 if (!cgroup_on_dfl(cgrp)) {
5000 cgrp->subtree_control = cgroup_control(cgrp);
5001 cgroup_refresh_subtree_ss_mask(cgrp);
5007 percpu_ref_exit(&cgrp->self.refcnt);
5010 return ERR_PTR(ret);
5012 cgroup_destroy_locked(cgrp);
5013 return ERR_PTR(ret);
5016 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5019 struct cgroup *parent, *cgrp;
5020 struct cgroup_subsys *ss;
5021 struct kernfs_node *kn;
5024 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5025 if (strchr(name, '\n'))
5028 parent = cgroup_kn_lock_live(parent_kn);
5032 cgrp = cgroup_create(parent);
5034 ret = PTR_ERR(cgrp);
5038 /* create the directory */
5039 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5047 * This extra ref will be put in cgroup_free_fn() and guarantees
5048 * that @cgrp->kn is always accessible.
5052 ret = cgroup_kn_set_ugid(kn);
5056 ret = css_populate_dir(&cgrp->self, NULL);
5060 do_each_subsys_mask(ss, ssid, cgroup_control(cgrp)) {
5061 ret = css_populate_dir(cgroup_css(cgrp, ss), NULL);
5064 } while_each_subsys_mask();
5066 /* let's create and online css's */
5067 kernfs_activate(kn);
5073 cgroup_destroy_locked(cgrp);
5075 cgroup_kn_unlock(parent_kn);
5080 * This is called when the refcnt of a css is confirmed to be killed.
5081 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5082 * initate destruction and put the css ref from kill_css().
5084 static void css_killed_work_fn(struct work_struct *work)
5086 struct cgroup_subsys_state *css =
5087 container_of(work, struct cgroup_subsys_state, destroy_work);
5089 mutex_lock(&cgroup_mutex);
5094 /* @css can't go away while we're holding cgroup_mutex */
5096 } while (css && atomic_dec_and_test(&css->online_cnt));
5098 mutex_unlock(&cgroup_mutex);
5101 /* css kill confirmation processing requires process context, bounce */
5102 static void css_killed_ref_fn(struct percpu_ref *ref)
5104 struct cgroup_subsys_state *css =
5105 container_of(ref, struct cgroup_subsys_state, refcnt);
5107 if (atomic_dec_and_test(&css->online_cnt)) {
5108 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5109 queue_work(cgroup_destroy_wq, &css->destroy_work);
5114 * kill_css - destroy a css
5115 * @css: css to destroy
5117 * This function initiates destruction of @css by removing cgroup interface
5118 * files and putting its base reference. ->css_offline() will be invoked
5119 * asynchronously once css_tryget_online() is guaranteed to fail and when
5120 * the reference count reaches zero, @css will be released.
5122 static void kill_css(struct cgroup_subsys_state *css)
5124 lockdep_assert_held(&cgroup_mutex);
5127 * This must happen before css is disassociated with its cgroup.
5128 * See seq_css() for details.
5130 css_clear_dir(css, NULL);
5133 * Killing would put the base ref, but we need to keep it alive
5134 * until after ->css_offline().
5139 * cgroup core guarantees that, by the time ->css_offline() is
5140 * invoked, no new css reference will be given out via
5141 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5142 * proceed to offlining css's because percpu_ref_kill() doesn't
5143 * guarantee that the ref is seen as killed on all CPUs on return.
5145 * Use percpu_ref_kill_and_confirm() to get notifications as each
5146 * css is confirmed to be seen as killed on all CPUs.
5148 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5152 * cgroup_destroy_locked - the first stage of cgroup destruction
5153 * @cgrp: cgroup to be destroyed
5155 * css's make use of percpu refcnts whose killing latency shouldn't be
5156 * exposed to userland and are RCU protected. Also, cgroup core needs to
5157 * guarantee that css_tryget_online() won't succeed by the time
5158 * ->css_offline() is invoked. To satisfy all the requirements,
5159 * destruction is implemented in the following two steps.
5161 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5162 * userland visible parts and start killing the percpu refcnts of
5163 * css's. Set up so that the next stage will be kicked off once all
5164 * the percpu refcnts are confirmed to be killed.
5166 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5167 * rest of destruction. Once all cgroup references are gone, the
5168 * cgroup is RCU-freed.
5170 * This function implements s1. After this step, @cgrp is gone as far as
5171 * the userland is concerned and a new cgroup with the same name may be
5172 * created. As cgroup doesn't care about the names internally, this
5173 * doesn't cause any problem.
5175 static int cgroup_destroy_locked(struct cgroup *cgrp)
5176 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5178 struct cgroup_subsys_state *css;
5181 lockdep_assert_held(&cgroup_mutex);
5184 * Only migration can raise populated from zero and we're already
5185 * holding cgroup_mutex.
5187 if (cgroup_is_populated(cgrp))
5191 * Make sure there's no live children. We can't test emptiness of
5192 * ->self.children as dead children linger on it while being
5193 * drained; otherwise, "rmdir parent/child parent" may fail.
5195 if (css_has_online_children(&cgrp->self))
5199 * Mark @cgrp dead. This prevents further task migration and child
5200 * creation by disabling cgroup_lock_live_group().
5202 cgrp->self.flags &= ~CSS_ONLINE;
5204 /* initiate massacre of all css's */
5205 for_each_css(css, ssid, cgrp)
5209 * Remove @cgrp directory along with the base files. @cgrp has an
5210 * extra ref on its kn.
5212 kernfs_remove(cgrp->kn);
5214 check_for_release(cgroup_parent(cgrp));
5216 /* put the base reference */
5217 percpu_ref_kill(&cgrp->self.refcnt);
5222 static int cgroup_rmdir(struct kernfs_node *kn)
5224 struct cgroup *cgrp;
5227 cgrp = cgroup_kn_lock_live(kn);
5231 ret = cgroup_destroy_locked(cgrp);
5233 cgroup_kn_unlock(kn);
5237 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5238 .remount_fs = cgroup_remount,
5239 .show_options = cgroup_show_options,
5240 .mkdir = cgroup_mkdir,
5241 .rmdir = cgroup_rmdir,
5242 .rename = cgroup_rename,
5245 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5247 struct cgroup_subsys_state *css;
5249 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5251 mutex_lock(&cgroup_mutex);
5253 idr_init(&ss->css_idr);
5254 INIT_LIST_HEAD(&ss->cfts);
5256 /* Create the root cgroup state for this subsystem */
5257 ss->root = &cgrp_dfl_root;
5258 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5259 /* We don't handle early failures gracefully */
5260 BUG_ON(IS_ERR(css));
5261 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5264 * Root csses are never destroyed and we can't initialize
5265 * percpu_ref during early init. Disable refcnting.
5267 css->flags |= CSS_NO_REF;
5270 /* allocation can't be done safely during early init */
5273 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5274 BUG_ON(css->id < 0);
5277 /* Update the init_css_set to contain a subsys
5278 * pointer to this state - since the subsystem is
5279 * newly registered, all tasks and hence the
5280 * init_css_set is in the subsystem's root cgroup. */
5281 init_css_set.subsys[ss->id] = css;
5283 have_fork_callback |= (bool)ss->fork << ss->id;
5284 have_exit_callback |= (bool)ss->exit << ss->id;
5285 have_free_callback |= (bool)ss->free << ss->id;
5286 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5288 /* At system boot, before all subsystems have been
5289 * registered, no tasks have been forked, so we don't
5290 * need to invoke fork callbacks here. */
5291 BUG_ON(!list_empty(&init_task.tasks));
5293 BUG_ON(online_css(css));
5295 mutex_unlock(&cgroup_mutex);
5299 * cgroup_init_early - cgroup initialization at system boot
5301 * Initialize cgroups at system boot, and initialize any
5302 * subsystems that request early init.
5304 int __init cgroup_init_early(void)
5306 static struct cgroup_sb_opts __initdata opts;
5307 struct cgroup_subsys *ss;
5310 init_cgroup_root(&cgrp_dfl_root, &opts);
5311 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5313 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5315 for_each_subsys(ss, i) {
5316 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5317 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5318 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5320 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5321 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5324 ss->name = cgroup_subsys_name[i];
5325 if (!ss->legacy_name)
5326 ss->legacy_name = cgroup_subsys_name[i];
5329 cgroup_init_subsys(ss, true);
5334 static u16 cgroup_disable_mask __initdata;
5337 * cgroup_init - cgroup initialization
5339 * Register cgroup filesystem and /proc file, and initialize
5340 * any subsystems that didn't request early init.
5342 int __init cgroup_init(void)
5344 struct cgroup_subsys *ss;
5347 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5348 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5349 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5350 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5352 mutex_lock(&cgroup_mutex);
5355 * Add init_css_set to the hash table so that dfl_root can link to
5358 hash_add(css_set_table, &init_css_set.hlist,
5359 css_set_hash(init_css_set.subsys));
5361 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5363 mutex_unlock(&cgroup_mutex);
5365 for_each_subsys(ss, ssid) {
5366 if (ss->early_init) {
5367 struct cgroup_subsys_state *css =
5368 init_css_set.subsys[ss->id];
5370 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5372 BUG_ON(css->id < 0);
5374 cgroup_init_subsys(ss, false);
5377 list_add_tail(&init_css_set.e_cset_node[ssid],
5378 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5381 * Setting dfl_root subsys_mask needs to consider the
5382 * disabled flag and cftype registration needs kmalloc,
5383 * both of which aren't available during early_init.
5385 if (cgroup_disable_mask & (1 << ssid)) {
5386 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5387 printk(KERN_INFO "Disabling %s control group subsystem\n",
5392 if (cgroup_ssid_no_v1(ssid))
5393 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5396 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5398 if (!ss->dfl_cftypes)
5399 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5401 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5402 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5404 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5405 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5409 ss->bind(init_css_set.subsys[ssid]);
5412 /* init_css_set.subsys[] has been updated, re-hash */
5413 hash_del(&init_css_set.hlist);
5414 hash_add(css_set_table, &init_css_set.hlist,
5415 css_set_hash(init_css_set.subsys));
5417 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5418 WARN_ON(register_filesystem(&cgroup_fs_type));
5419 WARN_ON(register_filesystem(&cgroup2_fs_type));
5420 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5425 static int __init cgroup_wq_init(void)
5428 * There isn't much point in executing destruction path in
5429 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5430 * Use 1 for @max_active.
5432 * We would prefer to do this in cgroup_init() above, but that
5433 * is called before init_workqueues(): so leave this until after.
5435 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5436 BUG_ON(!cgroup_destroy_wq);
5439 * Used to destroy pidlists and separate to serve as flush domain.
5440 * Cap @max_active to 1 too.
5442 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5444 BUG_ON(!cgroup_pidlist_destroy_wq);
5448 core_initcall(cgroup_wq_init);
5451 * proc_cgroup_show()
5452 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5453 * - Used for /proc/<pid>/cgroup.
5455 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5456 struct pid *pid, struct task_struct *tsk)
5460 struct cgroup_root *root;
5463 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5467 mutex_lock(&cgroup_mutex);
5468 spin_lock_bh(&css_set_lock);
5470 for_each_root(root) {
5471 struct cgroup_subsys *ss;
5472 struct cgroup *cgrp;
5473 int ssid, count = 0;
5475 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5478 seq_printf(m, "%d:", root->hierarchy_id);
5479 if (root != &cgrp_dfl_root)
5480 for_each_subsys(ss, ssid)
5481 if (root->subsys_mask & (1 << ssid))
5482 seq_printf(m, "%s%s", count++ ? "," : "",
5484 if (strlen(root->name))
5485 seq_printf(m, "%sname=%s", count ? "," : "",
5489 cgrp = task_cgroup_from_root(tsk, root);
5492 * On traditional hierarchies, all zombie tasks show up as
5493 * belonging to the root cgroup. On the default hierarchy,
5494 * while a zombie doesn't show up in "cgroup.procs" and
5495 * thus can't be migrated, its /proc/PID/cgroup keeps
5496 * reporting the cgroup it belonged to before exiting. If
5497 * the cgroup is removed before the zombie is reaped,
5498 * " (deleted)" is appended to the cgroup path.
5500 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5501 path = cgroup_path(cgrp, buf, PATH_MAX);
5503 retval = -ENAMETOOLONG;
5512 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5513 seq_puts(m, " (deleted)\n");
5520 spin_unlock_bh(&css_set_lock);
5521 mutex_unlock(&cgroup_mutex);
5527 /* Display information about each subsystem and each hierarchy */
5528 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5530 struct cgroup_subsys *ss;
5533 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5535 * ideally we don't want subsystems moving around while we do this.
5536 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5537 * subsys/hierarchy state.
5539 mutex_lock(&cgroup_mutex);
5541 for_each_subsys(ss, i)
5542 seq_printf(m, "%s\t%d\t%d\t%d\n",
5543 ss->legacy_name, ss->root->hierarchy_id,
5544 atomic_read(&ss->root->nr_cgrps),
5545 cgroup_ssid_enabled(i));
5547 mutex_unlock(&cgroup_mutex);
5551 static int cgroupstats_open(struct inode *inode, struct file *file)
5553 return single_open(file, proc_cgroupstats_show, NULL);
5556 static const struct file_operations proc_cgroupstats_operations = {
5557 .open = cgroupstats_open,
5559 .llseek = seq_lseek,
5560 .release = single_release,
5564 * cgroup_fork - initialize cgroup related fields during copy_process()
5565 * @child: pointer to task_struct of forking parent process.
5567 * A task is associated with the init_css_set until cgroup_post_fork()
5568 * attaches it to the parent's css_set. Empty cg_list indicates that
5569 * @child isn't holding reference to its css_set.
5571 void cgroup_fork(struct task_struct *child)
5573 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5574 INIT_LIST_HEAD(&child->cg_list);
5578 * cgroup_can_fork - called on a new task before the process is exposed
5579 * @child: the task in question.
5581 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5582 * returns an error, the fork aborts with that error code. This allows for
5583 * a cgroup subsystem to conditionally allow or deny new forks.
5585 int cgroup_can_fork(struct task_struct *child)
5587 struct cgroup_subsys *ss;
5590 do_each_subsys_mask(ss, i, have_canfork_callback) {
5591 ret = ss->can_fork(child);
5594 } while_each_subsys_mask();
5599 for_each_subsys(ss, j) {
5602 if (ss->cancel_fork)
5603 ss->cancel_fork(child);
5610 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5611 * @child: the task in question
5613 * This calls the cancel_fork() callbacks if a fork failed *after*
5614 * cgroup_can_fork() succeded.
5616 void cgroup_cancel_fork(struct task_struct *child)
5618 struct cgroup_subsys *ss;
5621 for_each_subsys(ss, i)
5622 if (ss->cancel_fork)
5623 ss->cancel_fork(child);
5627 * cgroup_post_fork - called on a new task after adding it to the task list
5628 * @child: the task in question
5630 * Adds the task to the list running through its css_set if necessary and
5631 * call the subsystem fork() callbacks. Has to be after the task is
5632 * visible on the task list in case we race with the first call to
5633 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5636 void cgroup_post_fork(struct task_struct *child)
5638 struct cgroup_subsys *ss;
5642 * This may race against cgroup_enable_task_cg_lists(). As that
5643 * function sets use_task_css_set_links before grabbing
5644 * tasklist_lock and we just went through tasklist_lock to add
5645 * @child, it's guaranteed that either we see the set
5646 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5647 * @child during its iteration.
5649 * If we won the race, @child is associated with %current's
5650 * css_set. Grabbing css_set_lock guarantees both that the
5651 * association is stable, and, on completion of the parent's
5652 * migration, @child is visible in the source of migration or
5653 * already in the destination cgroup. This guarantee is necessary
5654 * when implementing operations which need to migrate all tasks of
5655 * a cgroup to another.
5657 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5658 * will remain in init_css_set. This is safe because all tasks are
5659 * in the init_css_set before cg_links is enabled and there's no
5660 * operation which transfers all tasks out of init_css_set.
5662 if (use_task_css_set_links) {
5663 struct css_set *cset;
5665 spin_lock_bh(&css_set_lock);
5666 cset = task_css_set(current);
5667 if (list_empty(&child->cg_list)) {
5669 css_set_move_task(child, NULL, cset, false);
5671 spin_unlock_bh(&css_set_lock);
5675 * Call ss->fork(). This must happen after @child is linked on
5676 * css_set; otherwise, @child might change state between ->fork()
5677 * and addition to css_set.
5679 do_each_subsys_mask(ss, i, have_fork_callback) {
5681 } while_each_subsys_mask();
5685 * cgroup_exit - detach cgroup from exiting task
5686 * @tsk: pointer to task_struct of exiting process
5688 * Description: Detach cgroup from @tsk and release it.
5690 * Note that cgroups marked notify_on_release force every task in
5691 * them to take the global cgroup_mutex mutex when exiting.
5692 * This could impact scaling on very large systems. Be reluctant to
5693 * use notify_on_release cgroups where very high task exit scaling
5694 * is required on large systems.
5696 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5697 * call cgroup_exit() while the task is still competent to handle
5698 * notify_on_release(), then leave the task attached to the root cgroup in
5699 * each hierarchy for the remainder of its exit. No need to bother with
5700 * init_css_set refcnting. init_css_set never goes away and we can't race
5701 * with migration path - PF_EXITING is visible to migration path.
5703 void cgroup_exit(struct task_struct *tsk)
5705 struct cgroup_subsys *ss;
5706 struct css_set *cset;
5710 * Unlink from @tsk from its css_set. As migration path can't race
5711 * with us, we can check css_set and cg_list without synchronization.
5713 cset = task_css_set(tsk);
5715 if (!list_empty(&tsk->cg_list)) {
5716 spin_lock_bh(&css_set_lock);
5717 css_set_move_task(tsk, cset, NULL, false);
5718 spin_unlock_bh(&css_set_lock);
5723 /* see cgroup_post_fork() for details */
5724 do_each_subsys_mask(ss, i, have_exit_callback) {
5726 } while_each_subsys_mask();
5729 void cgroup_free(struct task_struct *task)
5731 struct css_set *cset = task_css_set(task);
5732 struct cgroup_subsys *ss;
5735 do_each_subsys_mask(ss, ssid, have_free_callback) {
5737 } while_each_subsys_mask();
5742 static void check_for_release(struct cgroup *cgrp)
5744 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5745 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5746 schedule_work(&cgrp->release_agent_work);
5750 * Notify userspace when a cgroup is released, by running the
5751 * configured release agent with the name of the cgroup (path
5752 * relative to the root of cgroup file system) as the argument.
5754 * Most likely, this user command will try to rmdir this cgroup.
5756 * This races with the possibility that some other task will be
5757 * attached to this cgroup before it is removed, or that some other
5758 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5759 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5760 * unused, and this cgroup will be reprieved from its death sentence,
5761 * to continue to serve a useful existence. Next time it's released,
5762 * we will get notified again, if it still has 'notify_on_release' set.
5764 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5765 * means only wait until the task is successfully execve()'d. The
5766 * separate release agent task is forked by call_usermodehelper(),
5767 * then control in this thread returns here, without waiting for the
5768 * release agent task. We don't bother to wait because the caller of
5769 * this routine has no use for the exit status of the release agent
5770 * task, so no sense holding our caller up for that.
5772 static void cgroup_release_agent(struct work_struct *work)
5774 struct cgroup *cgrp =
5775 container_of(work, struct cgroup, release_agent_work);
5776 char *pathbuf = NULL, *agentbuf = NULL, *path;
5777 char *argv[3], *envp[3];
5779 mutex_lock(&cgroup_mutex);
5781 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5782 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5783 if (!pathbuf || !agentbuf)
5786 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5794 /* minimal command environment */
5796 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5799 mutex_unlock(&cgroup_mutex);
5800 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5803 mutex_unlock(&cgroup_mutex);
5809 static int __init cgroup_disable(char *str)
5811 struct cgroup_subsys *ss;
5815 while ((token = strsep(&str, ",")) != NULL) {
5819 for_each_subsys(ss, i) {
5820 if (strcmp(token, ss->name) &&
5821 strcmp(token, ss->legacy_name))
5823 cgroup_disable_mask |= 1 << i;
5828 __setup("cgroup_disable=", cgroup_disable);
5830 static int __init cgroup_no_v1(char *str)
5832 struct cgroup_subsys *ss;
5836 while ((token = strsep(&str, ",")) != NULL) {
5840 if (!strcmp(token, "all")) {
5841 cgroup_no_v1_mask = U16_MAX;
5845 for_each_subsys(ss, i) {
5846 if (strcmp(token, ss->name) &&
5847 strcmp(token, ss->legacy_name))
5850 cgroup_no_v1_mask |= 1 << i;
5855 __setup("cgroup_no_v1=", cgroup_no_v1);
5858 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5859 * @dentry: directory dentry of interest
5860 * @ss: subsystem of interest
5862 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5863 * to get the corresponding css and return it. If such css doesn't exist
5864 * or can't be pinned, an ERR_PTR value is returned.
5866 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5867 struct cgroup_subsys *ss)
5869 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5870 struct file_system_type *s_type = dentry->d_sb->s_type;
5871 struct cgroup_subsys_state *css = NULL;
5872 struct cgroup *cgrp;
5874 /* is @dentry a cgroup dir? */
5875 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5876 !kn || kernfs_type(kn) != KERNFS_DIR)
5877 return ERR_PTR(-EBADF);
5882 * This path doesn't originate from kernfs and @kn could already
5883 * have been or be removed at any point. @kn->priv is RCU
5884 * protected for this access. See css_release_work_fn() for details.
5886 cgrp = rcu_dereference(kn->priv);
5888 css = cgroup_css(cgrp, ss);
5890 if (!css || !css_tryget_online(css))
5891 css = ERR_PTR(-ENOENT);
5898 * css_from_id - lookup css by id
5899 * @id: the cgroup id
5900 * @ss: cgroup subsys to be looked into
5902 * Returns the css if there's valid one with @id, otherwise returns NULL.
5903 * Should be called under rcu_read_lock().
5905 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5907 WARN_ON_ONCE(!rcu_read_lock_held());
5908 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5912 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5913 * @path: path on the default hierarchy
5915 * Find the cgroup at @path on the default hierarchy, increment its
5916 * reference count and return it. Returns pointer to the found cgroup on
5917 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5918 * if @path points to a non-directory.
5920 struct cgroup *cgroup_get_from_path(const char *path)
5922 struct kernfs_node *kn;
5923 struct cgroup *cgrp;
5925 mutex_lock(&cgroup_mutex);
5927 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5929 if (kernfs_type(kn) == KERNFS_DIR) {
5933 cgrp = ERR_PTR(-ENOTDIR);
5937 cgrp = ERR_PTR(-ENOENT);
5940 mutex_unlock(&cgroup_mutex);
5943 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5946 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5947 * definition in cgroup-defs.h.
5949 #ifdef CONFIG_SOCK_CGROUP_DATA
5951 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5953 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5954 static bool cgroup_sk_alloc_disabled __read_mostly;
5956 void cgroup_sk_alloc_disable(void)
5958 if (cgroup_sk_alloc_disabled)
5960 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5961 cgroup_sk_alloc_disabled = true;
5966 #define cgroup_sk_alloc_disabled false
5970 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5972 if (cgroup_sk_alloc_disabled)
5978 struct css_set *cset;
5980 cset = task_css_set(current);
5981 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5982 skcd->val = (unsigned long)cset->dfl_cgrp;
5991 void cgroup_sk_free(struct sock_cgroup_data *skcd)
5993 cgroup_put(sock_cgroup_ptr(skcd));
5996 #endif /* CONFIG_SOCK_CGROUP_DATA */
5998 #ifdef CONFIG_CGROUP_DEBUG
5999 static struct cgroup_subsys_state *
6000 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6002 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6005 return ERR_PTR(-ENOMEM);
6010 static void debug_css_free(struct cgroup_subsys_state *css)
6015 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6018 return cgroup_task_count(css->cgroup);
6021 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6024 return (u64)(unsigned long)current->cgroups;
6027 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6033 count = atomic_read(&task_css_set(current)->refcount);
6038 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6040 struct cgrp_cset_link *link;
6041 struct css_set *cset;
6044 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6048 spin_lock_bh(&css_set_lock);
6050 cset = rcu_dereference(current->cgroups);
6051 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6052 struct cgroup *c = link->cgrp;
6054 cgroup_name(c, name_buf, NAME_MAX + 1);
6055 seq_printf(seq, "Root %d group %s\n",
6056 c->root->hierarchy_id, name_buf);
6059 spin_unlock_bh(&css_set_lock);
6064 #define MAX_TASKS_SHOWN_PER_CSS 25
6065 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6067 struct cgroup_subsys_state *css = seq_css(seq);
6068 struct cgrp_cset_link *link;
6070 spin_lock_bh(&css_set_lock);
6071 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6072 struct css_set *cset = link->cset;
6073 struct task_struct *task;
6076 seq_printf(seq, "css_set %p\n", cset);
6078 list_for_each_entry(task, &cset->tasks, cg_list) {
6079 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6081 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6084 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6085 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6087 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6091 seq_puts(seq, " ...\n");
6093 spin_unlock_bh(&css_set_lock);
6097 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6099 return (!cgroup_is_populated(css->cgroup) &&
6100 !css_has_online_children(&css->cgroup->self));
6103 static struct cftype debug_files[] = {
6105 .name = "taskcount",
6106 .read_u64 = debug_taskcount_read,
6110 .name = "current_css_set",
6111 .read_u64 = current_css_set_read,
6115 .name = "current_css_set_refcount",
6116 .read_u64 = current_css_set_refcount_read,
6120 .name = "current_css_set_cg_links",
6121 .seq_show = current_css_set_cg_links_read,
6125 .name = "cgroup_css_links",
6126 .seq_show = cgroup_css_links_read,
6130 .name = "releasable",
6131 .read_u64 = releasable_read,
6137 struct cgroup_subsys debug_cgrp_subsys = {
6138 .css_alloc = debug_css_alloc,
6139 .css_free = debug_css_free,
6140 .legacy_cftypes = debug_files,
6142 #endif /* CONFIG_CGROUP_DEBUG */