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>
62 #include <linux/proc_ns.h>
63 #include <linux/nsproxy.h>
64 #include <linux/file.h>
67 #define CREATE_TRACE_POINTS
68 #include <trace/events/cgroup.h>
71 * pidlists linger the following amount before being destroyed. The goal
72 * is avoiding frequent destruction in the middle of consecutive read calls
73 * Expiring in the middle is a performance problem not a correctness one.
74 * 1 sec should be enough.
76 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
78 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
82 * cgroup_mutex is the master lock. Any modification to cgroup or its
83 * hierarchy must be performed while holding it.
85 * css_set_lock protects task->cgroups pointer, the list of css_set
86 * objects, and the chain of tasks off each css_set.
88 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
89 * cgroup.h can use them for lockdep annotations.
91 #ifdef CONFIG_PROVE_RCU
92 DEFINE_MUTEX(cgroup_mutex);
93 DEFINE_SPINLOCK(css_set_lock);
94 EXPORT_SYMBOL_GPL(cgroup_mutex);
95 EXPORT_SYMBOL_GPL(css_set_lock);
97 static DEFINE_MUTEX(cgroup_mutex);
98 static DEFINE_SPINLOCK(css_set_lock);
102 * Protects cgroup_idr and css_idr so that IDs can be released without
103 * grabbing cgroup_mutex.
105 static DEFINE_SPINLOCK(cgroup_idr_lock);
108 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
109 * against file removal/re-creation across css hiding.
111 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
114 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
115 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
117 static DEFINE_SPINLOCK(release_agent_path_lock);
119 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
121 #define cgroup_assert_mutex_or_rcu_locked() \
122 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
123 !lockdep_is_held(&cgroup_mutex), \
124 "cgroup_mutex or RCU read lock required");
127 * cgroup destruction makes heavy use of work items and there can be a lot
128 * of concurrent destructions. Use a separate workqueue so that cgroup
129 * destruction work items don't end up filling up max_active of system_wq
130 * which may lead to deadlock.
132 static struct workqueue_struct *cgroup_destroy_wq;
135 * pidlist destructions need to be flushed on cgroup destruction. Use a
136 * separate workqueue as flush domain.
138 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
140 /* generate an array of cgroup subsystem pointers */
141 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
142 static struct cgroup_subsys *cgroup_subsys[] = {
143 #include <linux/cgroup_subsys.h>
147 /* array of cgroup subsystem names */
148 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
149 static const char *cgroup_subsys_name[] = {
150 #include <linux/cgroup_subsys.h>
154 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
156 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
157 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
158 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
159 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
160 #include <linux/cgroup_subsys.h>
163 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
164 static struct static_key_true *cgroup_subsys_enabled_key[] = {
165 #include <linux/cgroup_subsys.h>
169 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
170 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
171 #include <linux/cgroup_subsys.h>
176 * The default hierarchy, reserved for the subsystems that are otherwise
177 * unattached - it never has more than a single cgroup, and all tasks are
178 * part of that cgroup.
180 struct cgroup_root cgrp_dfl_root;
181 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
184 * The default hierarchy always exists but is hidden until mounted for the
185 * first time. This is for backward compatibility.
187 static bool cgrp_dfl_visible;
189 /* Controllers blocked by the commandline in v1 */
190 static u16 cgroup_no_v1_mask;
192 /* some controllers are not supported in the default hierarchy */
193 static u16 cgrp_dfl_inhibit_ss_mask;
195 /* some controllers are implicitly enabled on the default hierarchy */
196 static unsigned long cgrp_dfl_implicit_ss_mask;
198 /* The list of hierarchy roots */
200 static LIST_HEAD(cgroup_roots);
201 static int cgroup_root_count;
203 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
204 static DEFINE_IDR(cgroup_hierarchy_idr);
207 * Assign a monotonically increasing serial number to csses. It guarantees
208 * cgroups with bigger numbers are newer than those with smaller numbers.
209 * Also, as csses are always appended to the parent's ->children list, it
210 * guarantees that sibling csses are always sorted in the ascending serial
211 * number order on the list. Protected by cgroup_mutex.
213 static u64 css_serial_nr_next = 1;
216 * These bitmask flags indicate whether tasks in the fork and exit paths have
217 * fork/exit handlers to call. This avoids us having to do extra work in the
218 * fork/exit path to check which subsystems have fork/exit callbacks.
220 static u16 have_fork_callback __read_mostly;
221 static u16 have_exit_callback __read_mostly;
222 static u16 have_free_callback __read_mostly;
224 /* cgroup namespace for init task */
225 struct cgroup_namespace init_cgroup_ns = {
226 .count = { .counter = 2, },
227 .user_ns = &init_user_ns,
228 .ns.ops = &cgroupns_operations,
229 .ns.inum = PROC_CGROUP_INIT_INO,
230 .root_cset = &init_css_set,
233 /* Ditto for the can_fork callback. */
234 static u16 have_canfork_callback __read_mostly;
236 static struct file_system_type cgroup2_fs_type;
237 static struct cftype cgroup_dfl_base_files[];
238 static struct cftype cgroup_legacy_base_files[];
240 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
241 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
242 static int cgroup_apply_control(struct cgroup *cgrp);
243 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
244 static void css_task_iter_advance(struct css_task_iter *it);
245 static int cgroup_destroy_locked(struct cgroup *cgrp);
246 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
247 struct cgroup_subsys *ss);
248 static void css_release(struct percpu_ref *ref);
249 static void kill_css(struct cgroup_subsys_state *css);
250 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
251 struct cgroup *cgrp, struct cftype cfts[],
255 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
256 * @ssid: subsys ID of interest
258 * cgroup_subsys_enabled() can only be used with literal subsys names which
259 * is fine for individual subsystems but unsuitable for cgroup core. This
260 * is slower static_key_enabled() based test indexed by @ssid.
262 static bool cgroup_ssid_enabled(int ssid)
264 if (CGROUP_SUBSYS_COUNT == 0)
267 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
270 static bool cgroup_ssid_no_v1(int ssid)
272 return cgroup_no_v1_mask & (1 << ssid);
276 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
277 * @cgrp: the cgroup of interest
279 * The default hierarchy is the v2 interface of cgroup and this function
280 * can be used to test whether a cgroup is on the default hierarchy for
281 * cases where a subsystem should behave differnetly depending on the
284 * The set of behaviors which change on the default hierarchy are still
285 * being determined and the mount option is prefixed with __DEVEL__.
287 * List of changed behaviors:
289 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
290 * and "name" are disallowed.
292 * - When mounting an existing superblock, mount options should match.
294 * - Remount is disallowed.
296 * - rename(2) is disallowed.
298 * - "tasks" is removed. Everything should be at process granularity. Use
299 * "cgroup.procs" instead.
301 * - "cgroup.procs" is not sorted. pids will be unique unless they got
302 * recycled inbetween reads.
304 * - "release_agent" and "notify_on_release" are removed. Replacement
305 * notification mechanism will be implemented.
307 * - "cgroup.clone_children" is removed.
309 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
310 * and its descendants contain no task; otherwise, 1. The file also
311 * generates kernfs notification which can be monitored through poll and
312 * [di]notify when the value of the file changes.
314 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
315 * take masks of ancestors with non-empty cpus/mems, instead of being
316 * moved to an ancestor.
318 * - cpuset: a task can be moved into an empty cpuset, and again it takes
319 * masks of ancestors.
321 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
324 * - blkcg: blk-throttle becomes properly hierarchical.
326 * - debug: disallowed on the default hierarchy.
328 static bool cgroup_on_dfl(const struct cgroup *cgrp)
330 return cgrp->root == &cgrp_dfl_root;
333 /* IDR wrappers which synchronize using cgroup_idr_lock */
334 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
339 idr_preload(gfp_mask);
340 spin_lock_bh(&cgroup_idr_lock);
341 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
342 spin_unlock_bh(&cgroup_idr_lock);
347 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
351 spin_lock_bh(&cgroup_idr_lock);
352 ret = idr_replace(idr, ptr, id);
353 spin_unlock_bh(&cgroup_idr_lock);
357 static void cgroup_idr_remove(struct idr *idr, int id)
359 spin_lock_bh(&cgroup_idr_lock);
361 spin_unlock_bh(&cgroup_idr_lock);
364 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
366 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
369 return container_of(parent_css, struct cgroup, self);
373 /* subsystems visibly enabled on a cgroup */
374 static u16 cgroup_control(struct cgroup *cgrp)
376 struct cgroup *parent = cgroup_parent(cgrp);
377 u16 root_ss_mask = cgrp->root->subsys_mask;
380 return parent->subtree_control;
382 if (cgroup_on_dfl(cgrp))
383 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
384 cgrp_dfl_implicit_ss_mask);
388 /* subsystems enabled on a cgroup */
389 static u16 cgroup_ss_mask(struct cgroup *cgrp)
391 struct cgroup *parent = cgroup_parent(cgrp);
394 return parent->subtree_ss_mask;
396 return cgrp->root->subsys_mask;
400 * cgroup_css - obtain a cgroup's css for the specified subsystem
401 * @cgrp: the cgroup of interest
402 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
404 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
405 * function must be called either under cgroup_mutex or rcu_read_lock() and
406 * the caller is responsible for pinning the returned css if it wants to
407 * keep accessing it outside the said locks. This function may return
408 * %NULL if @cgrp doesn't have @subsys_id enabled.
410 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
411 struct cgroup_subsys *ss)
414 return rcu_dereference_check(cgrp->subsys[ss->id],
415 lockdep_is_held(&cgroup_mutex));
421 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
422 * @cgrp: the cgroup of interest
423 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
425 * Similar to cgroup_css() but returns the effective css, which is defined
426 * as the matching css of the nearest ancestor including self which has @ss
427 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
428 * function is guaranteed to return non-NULL css.
430 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
431 struct cgroup_subsys *ss)
433 lockdep_assert_held(&cgroup_mutex);
439 * This function is used while updating css associations and thus
440 * can't test the csses directly. Test ss_mask.
442 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
443 cgrp = cgroup_parent(cgrp);
448 return cgroup_css(cgrp, ss);
452 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
453 * @cgrp: the cgroup of interest
454 * @ss: the subsystem of interest
456 * Find and get the effective css of @cgrp for @ss. The effective css is
457 * defined as the matching css of the nearest ancestor including self which
458 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
459 * the root css is returned, so this function always returns a valid css.
460 * The returned css must be put using css_put().
462 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
463 struct cgroup_subsys *ss)
465 struct cgroup_subsys_state *css;
470 css = cgroup_css(cgrp, ss);
472 if (css && css_tryget_online(css))
474 cgrp = cgroup_parent(cgrp);
477 css = init_css_set.subsys[ss->id];
484 /* convenient tests for these bits */
485 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
487 return !(cgrp->self.flags & CSS_ONLINE);
490 static void cgroup_get(struct cgroup *cgrp)
492 WARN_ON_ONCE(cgroup_is_dead(cgrp));
493 css_get(&cgrp->self);
496 static bool cgroup_tryget(struct cgroup *cgrp)
498 return css_tryget(&cgrp->self);
501 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
503 struct cgroup *cgrp = of->kn->parent->priv;
504 struct cftype *cft = of_cft(of);
507 * This is open and unprotected implementation of cgroup_css().
508 * seq_css() is only called from a kernfs file operation which has
509 * an active reference on the file. Because all the subsystem
510 * files are drained before a css is disassociated with a cgroup,
511 * the matching css from the cgroup's subsys table is guaranteed to
512 * be and stay valid until the enclosing operation is complete.
515 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
519 EXPORT_SYMBOL_GPL(of_css);
521 static int notify_on_release(const struct cgroup *cgrp)
523 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
527 * for_each_css - iterate all css's of a cgroup
528 * @css: the iteration cursor
529 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
530 * @cgrp: the target cgroup to iterate css's of
532 * Should be called under cgroup_[tree_]mutex.
534 #define for_each_css(css, ssid, cgrp) \
535 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
536 if (!((css) = rcu_dereference_check( \
537 (cgrp)->subsys[(ssid)], \
538 lockdep_is_held(&cgroup_mutex)))) { } \
542 * for_each_e_css - iterate all effective css's of a cgroup
543 * @css: the iteration cursor
544 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
545 * @cgrp: the target cgroup to iterate css's of
547 * Should be called under cgroup_[tree_]mutex.
549 #define for_each_e_css(css, ssid, cgrp) \
550 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
551 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
556 * for_each_subsys - iterate all enabled cgroup subsystems
557 * @ss: the iteration cursor
558 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
560 #define for_each_subsys(ss, ssid) \
561 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
562 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
565 * do_each_subsys_mask - filter for_each_subsys with a bitmask
566 * @ss: the iteration cursor
567 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
568 * @ss_mask: the bitmask
570 * The block will only run for cases where the ssid-th bit (1 << ssid) of
573 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
574 unsigned long __ss_mask = (ss_mask); \
575 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
579 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
580 (ss) = cgroup_subsys[ssid]; \
583 #define while_each_subsys_mask() \
588 /* iterate across the hierarchies */
589 #define for_each_root(root) \
590 list_for_each_entry((root), &cgroup_roots, root_list)
592 /* iterate over child cgrps, lock should be held throughout iteration */
593 #define cgroup_for_each_live_child(child, cgrp) \
594 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
595 if (({ lockdep_assert_held(&cgroup_mutex); \
596 cgroup_is_dead(child); })) \
600 /* walk live descendants in preorder */
601 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
602 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
603 if (({ lockdep_assert_held(&cgroup_mutex); \
604 (dsct) = (d_css)->cgroup; \
605 cgroup_is_dead(dsct); })) \
609 /* walk live descendants in postorder */
610 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
611 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
612 if (({ lockdep_assert_held(&cgroup_mutex); \
613 (dsct) = (d_css)->cgroup; \
614 cgroup_is_dead(dsct); })) \
618 static void cgroup_release_agent(struct work_struct *work);
619 static void check_for_release(struct cgroup *cgrp);
622 * A cgroup can be associated with multiple css_sets as different tasks may
623 * belong to different cgroups on different hierarchies. In the other
624 * direction, a css_set is naturally associated with multiple cgroups.
625 * This M:N relationship is represented by the following link structure
626 * which exists for each association and allows traversing the associations
629 struct cgrp_cset_link {
630 /* the cgroup and css_set this link associates */
632 struct css_set *cset;
634 /* list of cgrp_cset_links anchored at cgrp->cset_links */
635 struct list_head cset_link;
637 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
638 struct list_head cgrp_link;
642 * The default css_set - used by init and its children prior to any
643 * hierarchies being mounted. It contains a pointer to the root state
644 * for each subsystem. Also used to anchor the list of css_sets. Not
645 * reference-counted, to improve performance when child cgroups
646 * haven't been created.
648 struct css_set init_css_set = {
649 .refcount = ATOMIC_INIT(1),
650 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
651 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
652 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
653 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
654 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
655 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
658 static int css_set_count = 1; /* 1 for init_css_set */
661 * css_set_populated - does a css_set contain any tasks?
662 * @cset: target css_set
664 static bool css_set_populated(struct css_set *cset)
666 lockdep_assert_held(&css_set_lock);
668 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
672 * cgroup_update_populated - updated populated count of a cgroup
673 * @cgrp: the target cgroup
674 * @populated: inc or dec populated count
676 * One of the css_sets associated with @cgrp is either getting its first
677 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
678 * count is propagated towards root so that a given cgroup's populated_cnt
679 * is zero iff the cgroup and all its descendants don't contain any tasks.
681 * @cgrp's interface file "cgroup.populated" is zero if
682 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
683 * changes from or to zero, userland is notified that the content of the
684 * interface file has changed. This can be used to detect when @cgrp and
685 * its descendants become populated or empty.
687 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
689 lockdep_assert_held(&css_set_lock);
695 trigger = !cgrp->populated_cnt++;
697 trigger = !--cgrp->populated_cnt;
702 check_for_release(cgrp);
703 cgroup_file_notify(&cgrp->events_file);
705 cgrp = cgroup_parent(cgrp);
710 * css_set_update_populated - update populated state of a css_set
711 * @cset: target css_set
712 * @populated: whether @cset is populated or depopulated
714 * @cset is either getting the first task or losing the last. Update the
715 * ->populated_cnt of all associated cgroups accordingly.
717 static void css_set_update_populated(struct css_set *cset, bool populated)
719 struct cgrp_cset_link *link;
721 lockdep_assert_held(&css_set_lock);
723 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
724 cgroup_update_populated(link->cgrp, populated);
728 * css_set_move_task - move a task from one css_set to another
729 * @task: task being moved
730 * @from_cset: css_set @task currently belongs to (may be NULL)
731 * @to_cset: new css_set @task is being moved to (may be NULL)
732 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
734 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
735 * css_set, @from_cset can be NULL. If @task is being disassociated
736 * instead of moved, @to_cset can be NULL.
738 * This function automatically handles populated_cnt updates and
739 * css_task_iter adjustments but the caller is responsible for managing
740 * @from_cset and @to_cset's reference counts.
742 static void css_set_move_task(struct task_struct *task,
743 struct css_set *from_cset, struct css_set *to_cset,
746 lockdep_assert_held(&css_set_lock);
748 if (to_cset && !css_set_populated(to_cset))
749 css_set_update_populated(to_cset, true);
752 struct css_task_iter *it, *pos;
754 WARN_ON_ONCE(list_empty(&task->cg_list));
757 * @task is leaving, advance task iterators which are
758 * pointing to it so that they can resume at the next
759 * position. Advancing an iterator might remove it from
760 * the list, use safe walk. See css_task_iter_advance*()
763 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
765 if (it->task_pos == &task->cg_list)
766 css_task_iter_advance(it);
768 list_del_init(&task->cg_list);
769 if (!css_set_populated(from_cset))
770 css_set_update_populated(from_cset, false);
772 WARN_ON_ONCE(!list_empty(&task->cg_list));
777 * We are synchronized through cgroup_threadgroup_rwsem
778 * against PF_EXITING setting such that we can't race
779 * against cgroup_exit() changing the css_set to
780 * init_css_set and dropping the old one.
782 WARN_ON_ONCE(task->flags & PF_EXITING);
784 rcu_assign_pointer(task->cgroups, to_cset);
785 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
791 * hash table for cgroup groups. This improves the performance to find
792 * an existing css_set. This hash doesn't (currently) take into
793 * account cgroups in empty hierarchies.
795 #define CSS_SET_HASH_BITS 7
796 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
798 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
800 unsigned long key = 0UL;
801 struct cgroup_subsys *ss;
804 for_each_subsys(ss, i)
805 key += (unsigned long)css[i];
806 key = (key >> 16) ^ key;
811 static void put_css_set_locked(struct css_set *cset)
813 struct cgrp_cset_link *link, *tmp_link;
814 struct cgroup_subsys *ss;
817 lockdep_assert_held(&css_set_lock);
819 if (!atomic_dec_and_test(&cset->refcount))
822 /* This css_set is dead. unlink it and release cgroup and css refs */
823 for_each_subsys(ss, ssid) {
824 list_del(&cset->e_cset_node[ssid]);
825 css_put(cset->subsys[ssid]);
827 hash_del(&cset->hlist);
830 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
831 list_del(&link->cset_link);
832 list_del(&link->cgrp_link);
833 if (cgroup_parent(link->cgrp))
834 cgroup_put(link->cgrp);
838 kfree_rcu(cset, rcu_head);
841 static void put_css_set(struct css_set *cset)
846 * Ensure that the refcount doesn't hit zero while any readers
847 * can see it. Similar to atomic_dec_and_lock(), but for an
850 if (atomic_add_unless(&cset->refcount, -1, 1))
853 spin_lock_irqsave(&css_set_lock, flags);
854 put_css_set_locked(cset);
855 spin_unlock_irqrestore(&css_set_lock, flags);
859 * refcounted get/put for css_set objects
861 static inline void get_css_set(struct css_set *cset)
863 atomic_inc(&cset->refcount);
867 * compare_css_sets - helper function for find_existing_css_set().
868 * @cset: candidate css_set being tested
869 * @old_cset: existing css_set for a task
870 * @new_cgrp: cgroup that's being entered by the task
871 * @template: desired set of css pointers in css_set (pre-calculated)
873 * Returns true if "cset" matches "old_cset" except for the hierarchy
874 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
876 static bool compare_css_sets(struct css_set *cset,
877 struct css_set *old_cset,
878 struct cgroup *new_cgrp,
879 struct cgroup_subsys_state *template[])
881 struct list_head *l1, *l2;
884 * On the default hierarchy, there can be csets which are
885 * associated with the same set of cgroups but different csses.
886 * Let's first ensure that csses match.
888 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
892 * Compare cgroup pointers in order to distinguish between
893 * different cgroups in hierarchies. As different cgroups may
894 * share the same effective css, this comparison is always
897 l1 = &cset->cgrp_links;
898 l2 = &old_cset->cgrp_links;
900 struct cgrp_cset_link *link1, *link2;
901 struct cgroup *cgrp1, *cgrp2;
905 /* See if we reached the end - both lists are equal length. */
906 if (l1 == &cset->cgrp_links) {
907 BUG_ON(l2 != &old_cset->cgrp_links);
910 BUG_ON(l2 == &old_cset->cgrp_links);
912 /* Locate the cgroups associated with these links. */
913 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
914 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
917 /* Hierarchies should be linked in the same order. */
918 BUG_ON(cgrp1->root != cgrp2->root);
921 * If this hierarchy is the hierarchy of the cgroup
922 * that's changing, then we need to check that this
923 * css_set points to the new cgroup; if it's any other
924 * hierarchy, then this css_set should point to the
925 * same cgroup as the old css_set.
927 if (cgrp1->root == new_cgrp->root) {
928 if (cgrp1 != new_cgrp)
939 * find_existing_css_set - init css array and find the matching css_set
940 * @old_cset: the css_set that we're using before the cgroup transition
941 * @cgrp: the cgroup that we're moving into
942 * @template: out param for the new set of csses, should be clear on entry
944 static struct css_set *find_existing_css_set(struct css_set *old_cset,
946 struct cgroup_subsys_state *template[])
948 struct cgroup_root *root = cgrp->root;
949 struct cgroup_subsys *ss;
950 struct css_set *cset;
955 * Build the set of subsystem state objects that we want to see in the
956 * new css_set. while subsystems can change globally, the entries here
957 * won't change, so no need for locking.
959 for_each_subsys(ss, i) {
960 if (root->subsys_mask & (1UL << i)) {
962 * @ss is in this hierarchy, so we want the
963 * effective css from @cgrp.
965 template[i] = cgroup_e_css(cgrp, ss);
968 * @ss is not in this hierarchy, so we don't want
971 template[i] = old_cset->subsys[i];
975 key = css_set_hash(template);
976 hash_for_each_possible(css_set_table, cset, hlist, key) {
977 if (!compare_css_sets(cset, old_cset, cgrp, template))
980 /* This css_set matches what we need */
984 /* No existing cgroup group matched */
988 static void free_cgrp_cset_links(struct list_head *links_to_free)
990 struct cgrp_cset_link *link, *tmp_link;
992 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
993 list_del(&link->cset_link);
999 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1000 * @count: the number of links to allocate
1001 * @tmp_links: list_head the allocated links are put on
1003 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1004 * through ->cset_link. Returns 0 on success or -errno.
1006 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1008 struct cgrp_cset_link *link;
1011 INIT_LIST_HEAD(tmp_links);
1013 for (i = 0; i < count; i++) {
1014 link = kzalloc(sizeof(*link), GFP_KERNEL);
1016 free_cgrp_cset_links(tmp_links);
1019 list_add(&link->cset_link, tmp_links);
1025 * link_css_set - a helper function to link a css_set to a cgroup
1026 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1027 * @cset: the css_set to be linked
1028 * @cgrp: the destination cgroup
1030 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1031 struct cgroup *cgrp)
1033 struct cgrp_cset_link *link;
1035 BUG_ON(list_empty(tmp_links));
1037 if (cgroup_on_dfl(cgrp))
1038 cset->dfl_cgrp = cgrp;
1040 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1045 * Always add links to the tail of the lists so that the lists are
1046 * in choronological order.
1048 list_move_tail(&link->cset_link, &cgrp->cset_links);
1049 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1051 if (cgroup_parent(cgrp))
1056 * find_css_set - return a new css_set with one cgroup updated
1057 * @old_cset: the baseline css_set
1058 * @cgrp: the cgroup to be updated
1060 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1061 * substituted into the appropriate hierarchy.
1063 static struct css_set *find_css_set(struct css_set *old_cset,
1064 struct cgroup *cgrp)
1066 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1067 struct css_set *cset;
1068 struct list_head tmp_links;
1069 struct cgrp_cset_link *link;
1070 struct cgroup_subsys *ss;
1074 lockdep_assert_held(&cgroup_mutex);
1076 /* First see if we already have a cgroup group that matches
1077 * the desired set */
1078 spin_lock_irq(&css_set_lock);
1079 cset = find_existing_css_set(old_cset, cgrp, template);
1082 spin_unlock_irq(&css_set_lock);
1087 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1091 /* Allocate all the cgrp_cset_link objects that we'll need */
1092 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1097 atomic_set(&cset->refcount, 1);
1098 INIT_LIST_HEAD(&cset->cgrp_links);
1099 INIT_LIST_HEAD(&cset->tasks);
1100 INIT_LIST_HEAD(&cset->mg_tasks);
1101 INIT_LIST_HEAD(&cset->mg_preload_node);
1102 INIT_LIST_HEAD(&cset->mg_node);
1103 INIT_LIST_HEAD(&cset->task_iters);
1104 INIT_HLIST_NODE(&cset->hlist);
1106 /* Copy the set of subsystem state objects generated in
1107 * find_existing_css_set() */
1108 memcpy(cset->subsys, template, sizeof(cset->subsys));
1110 spin_lock_irq(&css_set_lock);
1111 /* Add reference counts and links from the new css_set. */
1112 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1113 struct cgroup *c = link->cgrp;
1115 if (c->root == cgrp->root)
1117 link_css_set(&tmp_links, cset, c);
1120 BUG_ON(!list_empty(&tmp_links));
1124 /* Add @cset to the hash table */
1125 key = css_set_hash(cset->subsys);
1126 hash_add(css_set_table, &cset->hlist, key);
1128 for_each_subsys(ss, ssid) {
1129 struct cgroup_subsys_state *css = cset->subsys[ssid];
1131 list_add_tail(&cset->e_cset_node[ssid],
1132 &css->cgroup->e_csets[ssid]);
1136 spin_unlock_irq(&css_set_lock);
1141 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1143 struct cgroup *root_cgrp = kf_root->kn->priv;
1145 return root_cgrp->root;
1148 static int cgroup_init_root_id(struct cgroup_root *root)
1152 lockdep_assert_held(&cgroup_mutex);
1154 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1158 root->hierarchy_id = id;
1162 static void cgroup_exit_root_id(struct cgroup_root *root)
1164 lockdep_assert_held(&cgroup_mutex);
1166 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1169 static void cgroup_free_root(struct cgroup_root *root)
1172 idr_destroy(&root->cgroup_idr);
1177 static void cgroup_destroy_root(struct cgroup_root *root)
1179 struct cgroup *cgrp = &root->cgrp;
1180 struct cgrp_cset_link *link, *tmp_link;
1182 trace_cgroup_destroy_root(root);
1184 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1186 BUG_ON(atomic_read(&root->nr_cgrps));
1187 BUG_ON(!list_empty(&cgrp->self.children));
1189 /* Rebind all subsystems back to the default hierarchy */
1190 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1193 * Release all the links from cset_links to this hierarchy's
1196 spin_lock_irq(&css_set_lock);
1198 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1199 list_del(&link->cset_link);
1200 list_del(&link->cgrp_link);
1204 spin_unlock_irq(&css_set_lock);
1206 if (!list_empty(&root->root_list)) {
1207 list_del(&root->root_list);
1208 cgroup_root_count--;
1211 cgroup_exit_root_id(root);
1213 mutex_unlock(&cgroup_mutex);
1215 kernfs_destroy_root(root->kf_root);
1216 cgroup_free_root(root);
1220 * look up cgroup associated with current task's cgroup namespace on the
1221 * specified hierarchy
1223 static struct cgroup *
1224 current_cgns_cgroup_from_root(struct cgroup_root *root)
1226 struct cgroup *res = NULL;
1227 struct css_set *cset;
1229 lockdep_assert_held(&css_set_lock);
1233 cset = current->nsproxy->cgroup_ns->root_cset;
1234 if (cset == &init_css_set) {
1237 struct cgrp_cset_link *link;
1239 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1240 struct cgroup *c = link->cgrp;
1242 if (c->root == root) {
1254 /* look up cgroup associated with given css_set on the specified hierarchy */
1255 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1256 struct cgroup_root *root)
1258 struct cgroup *res = NULL;
1260 lockdep_assert_held(&cgroup_mutex);
1261 lockdep_assert_held(&css_set_lock);
1263 if (cset == &init_css_set) {
1266 struct cgrp_cset_link *link;
1268 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1269 struct cgroup *c = link->cgrp;
1271 if (c->root == root) {
1283 * Return the cgroup for "task" from the given hierarchy. Must be
1284 * called with cgroup_mutex and css_set_lock held.
1286 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1287 struct cgroup_root *root)
1290 * No need to lock the task - since we hold cgroup_mutex the
1291 * task can't change groups, so the only thing that can happen
1292 * is that it exits and its css is set back to init_css_set.
1294 return cset_cgroup_from_root(task_css_set(task), root);
1298 * A task must hold cgroup_mutex to modify cgroups.
1300 * Any task can increment and decrement the count field without lock.
1301 * So in general, code holding cgroup_mutex can't rely on the count
1302 * field not changing. However, if the count goes to zero, then only
1303 * cgroup_attach_task() can increment it again. Because a count of zero
1304 * means that no tasks are currently attached, therefore there is no
1305 * way a task attached to that cgroup can fork (the other way to
1306 * increment the count). So code holding cgroup_mutex can safely
1307 * assume that if the count is zero, it will stay zero. Similarly, if
1308 * a task holds cgroup_mutex on a cgroup with zero count, it
1309 * knows that the cgroup won't be removed, as cgroup_rmdir()
1312 * A cgroup can only be deleted if both its 'count' of using tasks
1313 * is zero, and its list of 'children' cgroups is empty. Since all
1314 * tasks in the system use _some_ cgroup, and since there is always at
1315 * least one task in the system (init, pid == 1), therefore, root cgroup
1316 * always has either children cgroups and/or using tasks. So we don't
1317 * need a special hack to ensure that root cgroup cannot be deleted.
1319 * P.S. One more locking exception. RCU is used to guard the
1320 * update of a tasks cgroup pointer by cgroup_attach_task()
1323 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1324 static const struct file_operations proc_cgroupstats_operations;
1326 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1329 struct cgroup_subsys *ss = cft->ss;
1331 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1332 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1333 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1334 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1337 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1342 * cgroup_file_mode - deduce file mode of a control file
1343 * @cft: the control file in question
1345 * S_IRUGO for read, S_IWUSR for write.
1347 static umode_t cgroup_file_mode(const struct cftype *cft)
1351 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1354 if (cft->write_u64 || cft->write_s64 || cft->write) {
1355 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1365 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1366 * @subtree_control: the new subtree_control mask to consider
1367 * @this_ss_mask: available subsystems
1369 * On the default hierarchy, a subsystem may request other subsystems to be
1370 * enabled together through its ->depends_on mask. In such cases, more
1371 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1373 * This function calculates which subsystems need to be enabled if
1374 * @subtree_control is to be applied while restricted to @this_ss_mask.
1376 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1378 u16 cur_ss_mask = subtree_control;
1379 struct cgroup_subsys *ss;
1382 lockdep_assert_held(&cgroup_mutex);
1384 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1387 u16 new_ss_mask = cur_ss_mask;
1389 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1390 new_ss_mask |= ss->depends_on;
1391 } while_each_subsys_mask();
1394 * Mask out subsystems which aren't available. This can
1395 * happen only if some depended-upon subsystems were bound
1396 * to non-default hierarchies.
1398 new_ss_mask &= this_ss_mask;
1400 if (new_ss_mask == cur_ss_mask)
1402 cur_ss_mask = new_ss_mask;
1409 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1410 * @kn: the kernfs_node being serviced
1412 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1413 * the method finishes if locking succeeded. Note that once this function
1414 * returns the cgroup returned by cgroup_kn_lock_live() may become
1415 * inaccessible any time. If the caller intends to continue to access the
1416 * cgroup, it should pin it before invoking this function.
1418 static void cgroup_kn_unlock(struct kernfs_node *kn)
1420 struct cgroup *cgrp;
1422 if (kernfs_type(kn) == KERNFS_DIR)
1425 cgrp = kn->parent->priv;
1427 mutex_unlock(&cgroup_mutex);
1429 kernfs_unbreak_active_protection(kn);
1434 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1435 * @kn: the kernfs_node being serviced
1436 * @drain_offline: perform offline draining on the cgroup
1438 * This helper is to be used by a cgroup kernfs method currently servicing
1439 * @kn. It breaks the active protection, performs cgroup locking and
1440 * verifies that the associated cgroup is alive. Returns the cgroup if
1441 * alive; otherwise, %NULL. A successful return should be undone by a
1442 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1443 * cgroup is drained of offlining csses before return.
1445 * Any cgroup kernfs method implementation which requires locking the
1446 * associated cgroup should use this helper. It avoids nesting cgroup
1447 * locking under kernfs active protection and allows all kernfs operations
1448 * including self-removal.
1450 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1453 struct cgroup *cgrp;
1455 if (kernfs_type(kn) == KERNFS_DIR)
1458 cgrp = kn->parent->priv;
1461 * We're gonna grab cgroup_mutex which nests outside kernfs
1462 * active_ref. cgroup liveliness check alone provides enough
1463 * protection against removal. Ensure @cgrp stays accessible and
1464 * break the active_ref protection.
1466 if (!cgroup_tryget(cgrp))
1468 kernfs_break_active_protection(kn);
1471 cgroup_lock_and_drain_offline(cgrp);
1473 mutex_lock(&cgroup_mutex);
1475 if (!cgroup_is_dead(cgrp))
1478 cgroup_kn_unlock(kn);
1482 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1484 char name[CGROUP_FILE_NAME_MAX];
1486 lockdep_assert_held(&cgroup_mutex);
1488 if (cft->file_offset) {
1489 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1490 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1492 spin_lock_irq(&cgroup_file_kn_lock);
1494 spin_unlock_irq(&cgroup_file_kn_lock);
1497 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1501 * css_clear_dir - remove subsys files in a cgroup directory
1504 static void css_clear_dir(struct cgroup_subsys_state *css)
1506 struct cgroup *cgrp = css->cgroup;
1507 struct cftype *cfts;
1509 if (!(css->flags & CSS_VISIBLE))
1512 css->flags &= ~CSS_VISIBLE;
1514 list_for_each_entry(cfts, &css->ss->cfts, node)
1515 cgroup_addrm_files(css, cgrp, cfts, false);
1519 * css_populate_dir - create subsys files in a cgroup directory
1522 * On failure, no file is added.
1524 static int css_populate_dir(struct cgroup_subsys_state *css)
1526 struct cgroup *cgrp = css->cgroup;
1527 struct cftype *cfts, *failed_cfts;
1530 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1534 if (cgroup_on_dfl(cgrp))
1535 cfts = cgroup_dfl_base_files;
1537 cfts = cgroup_legacy_base_files;
1539 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1542 list_for_each_entry(cfts, &css->ss->cfts, node) {
1543 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1550 css->flags |= CSS_VISIBLE;
1554 list_for_each_entry(cfts, &css->ss->cfts, node) {
1555 if (cfts == failed_cfts)
1557 cgroup_addrm_files(css, cgrp, cfts, false);
1562 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1564 struct cgroup *dcgrp = &dst_root->cgrp;
1565 struct cgroup_subsys *ss;
1567 u16 dfl_disable_ss_mask = 0;
1569 lockdep_assert_held(&cgroup_mutex);
1571 do_each_subsys_mask(ss, ssid, ss_mask) {
1573 * If @ss has non-root csses attached to it, can't move.
1574 * If @ss is an implicit controller, it is exempt from this
1575 * rule and can be stolen.
1577 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1578 !ss->implicit_on_dfl)
1581 /* can't move between two non-dummy roots either */
1582 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1586 * Collect ssid's that need to be disabled from default
1589 if (ss->root == &cgrp_dfl_root)
1590 dfl_disable_ss_mask |= 1 << ssid;
1592 } while_each_subsys_mask();
1594 if (dfl_disable_ss_mask) {
1595 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1598 * Controllers from default hierarchy that need to be rebound
1599 * are all disabled together in one go.
1601 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1602 WARN_ON(cgroup_apply_control(scgrp));
1603 cgroup_finalize_control(scgrp, 0);
1606 do_each_subsys_mask(ss, ssid, ss_mask) {
1607 struct cgroup_root *src_root = ss->root;
1608 struct cgroup *scgrp = &src_root->cgrp;
1609 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1610 struct css_set *cset;
1612 WARN_ON(!css || cgroup_css(dcgrp, ss));
1614 if (src_root != &cgrp_dfl_root) {
1615 /* disable from the source */
1616 src_root->subsys_mask &= ~(1 << ssid);
1617 WARN_ON(cgroup_apply_control(scgrp));
1618 cgroup_finalize_control(scgrp, 0);
1622 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1623 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1624 ss->root = dst_root;
1625 css->cgroup = dcgrp;
1627 spin_lock_irq(&css_set_lock);
1628 hash_for_each(css_set_table, i, cset, hlist)
1629 list_move_tail(&cset->e_cset_node[ss->id],
1630 &dcgrp->e_csets[ss->id]);
1631 spin_unlock_irq(&css_set_lock);
1633 /* default hierarchy doesn't enable controllers by default */
1634 dst_root->subsys_mask |= 1 << ssid;
1635 if (dst_root == &cgrp_dfl_root) {
1636 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1638 dcgrp->subtree_control |= 1 << ssid;
1639 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1642 ret = cgroup_apply_control(dcgrp);
1644 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1649 } while_each_subsys_mask();
1651 kernfs_activate(dcgrp->kn);
1655 static int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1656 struct kernfs_root *kf_root)
1660 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1661 struct cgroup *ns_cgroup;
1663 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1667 spin_lock_irq(&css_set_lock);
1668 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1669 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1670 spin_unlock_irq(&css_set_lock);
1672 if (len >= PATH_MAX)
1675 seq_escape(sf, buf, " \t\n\\");
1682 static int cgroup_show_options(struct seq_file *seq,
1683 struct kernfs_root *kf_root)
1685 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1686 struct cgroup_subsys *ss;
1689 if (root != &cgrp_dfl_root)
1690 for_each_subsys(ss, ssid)
1691 if (root->subsys_mask & (1 << ssid))
1692 seq_show_option(seq, ss->legacy_name, NULL);
1693 if (root->flags & CGRP_ROOT_NOPREFIX)
1694 seq_puts(seq, ",noprefix");
1695 if (root->flags & CGRP_ROOT_XATTR)
1696 seq_puts(seq, ",xattr");
1698 spin_lock(&release_agent_path_lock);
1699 if (strlen(root->release_agent_path))
1700 seq_show_option(seq, "release_agent",
1701 root->release_agent_path);
1702 spin_unlock(&release_agent_path_lock);
1704 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1705 seq_puts(seq, ",clone_children");
1706 if (strlen(root->name))
1707 seq_show_option(seq, "name", root->name);
1711 struct cgroup_sb_opts {
1714 char *release_agent;
1715 bool cpuset_clone_children;
1717 /* User explicitly requested empty subsystem */
1721 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1723 char *token, *o = data;
1724 bool all_ss = false, one_ss = false;
1726 struct cgroup_subsys *ss;
1730 #ifdef CONFIG_CPUSETS
1731 mask = ~((u16)1 << cpuset_cgrp_id);
1734 memset(opts, 0, sizeof(*opts));
1736 while ((token = strsep(&o, ",")) != NULL) {
1741 if (!strcmp(token, "none")) {
1742 /* Explicitly have no subsystems */
1746 if (!strcmp(token, "all")) {
1747 /* Mutually exclusive option 'all' + subsystem name */
1753 if (!strcmp(token, "noprefix")) {
1754 opts->flags |= CGRP_ROOT_NOPREFIX;
1757 if (!strcmp(token, "clone_children")) {
1758 opts->cpuset_clone_children = true;
1761 if (!strcmp(token, "xattr")) {
1762 opts->flags |= CGRP_ROOT_XATTR;
1765 if (!strncmp(token, "release_agent=", 14)) {
1766 /* Specifying two release agents is forbidden */
1767 if (opts->release_agent)
1769 opts->release_agent =
1770 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1771 if (!opts->release_agent)
1775 if (!strncmp(token, "name=", 5)) {
1776 const char *name = token + 5;
1777 /* Can't specify an empty name */
1780 /* Must match [\w.-]+ */
1781 for (i = 0; i < strlen(name); i++) {
1785 if ((c == '.') || (c == '-') || (c == '_'))
1789 /* Specifying two names is forbidden */
1792 opts->name = kstrndup(name,
1793 MAX_CGROUP_ROOT_NAMELEN - 1,
1801 for_each_subsys(ss, i) {
1802 if (strcmp(token, ss->legacy_name))
1804 if (!cgroup_ssid_enabled(i))
1806 if (cgroup_ssid_no_v1(i))
1809 /* Mutually exclusive option 'all' + subsystem name */
1812 opts->subsys_mask |= (1 << i);
1817 if (i == CGROUP_SUBSYS_COUNT)
1822 * If the 'all' option was specified select all the subsystems,
1823 * otherwise if 'none', 'name=' and a subsystem name options were
1824 * not specified, let's default to 'all'
1826 if (all_ss || (!one_ss && !opts->none && !opts->name))
1827 for_each_subsys(ss, i)
1828 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1829 opts->subsys_mask |= (1 << i);
1832 * We either have to specify by name or by subsystems. (So all
1833 * empty hierarchies must have a name).
1835 if (!opts->subsys_mask && !opts->name)
1839 * Option noprefix was introduced just for backward compatibility
1840 * with the old cpuset, so we allow noprefix only if mounting just
1841 * the cpuset subsystem.
1843 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1846 /* Can't specify "none" and some subsystems */
1847 if (opts->subsys_mask && opts->none)
1853 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1856 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1857 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
1858 struct cgroup_sb_opts opts;
1859 u16 added_mask, removed_mask;
1861 if (root == &cgrp_dfl_root) {
1862 pr_err("remount is not allowed\n");
1866 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1868 /* See what subsystems are wanted */
1869 ret = parse_cgroupfs_options(data, &opts);
1873 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1874 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1875 task_tgid_nr(current), current->comm);
1877 /* See cgroup_mount release_agent handling */
1878 if (opts.release_agent &&
1879 ((ns->user_ns != &init_user_ns) || !capable(CAP_SYS_ADMIN))) {
1884 added_mask = opts.subsys_mask & ~root->subsys_mask;
1885 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1887 /* Don't allow flags or name to change at remount */
1888 if ((opts.flags ^ root->flags) ||
1889 (opts.name && strcmp(opts.name, root->name))) {
1890 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1891 opts.flags, opts.name ?: "", root->flags, root->name);
1896 /* remounting is not allowed for populated hierarchies */
1897 if (!list_empty(&root->cgrp.self.children)) {
1902 ret = rebind_subsystems(root, added_mask);
1906 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1908 if (opts.release_agent) {
1909 spin_lock(&release_agent_path_lock);
1910 strcpy(root->release_agent_path, opts.release_agent);
1911 spin_unlock(&release_agent_path_lock);
1914 trace_cgroup_remount(root);
1917 kfree(opts.release_agent);
1919 mutex_unlock(&cgroup_mutex);
1924 * To reduce the fork() overhead for systems that are not actually using
1925 * their cgroups capability, we don't maintain the lists running through
1926 * each css_set to its tasks until we see the list actually used - in other
1927 * words after the first mount.
1929 static bool use_task_css_set_links __read_mostly;
1931 static void cgroup_enable_task_cg_lists(void)
1933 struct task_struct *p, *g;
1935 spin_lock_irq(&css_set_lock);
1937 if (use_task_css_set_links)
1940 use_task_css_set_links = true;
1943 * We need tasklist_lock because RCU is not safe against
1944 * while_each_thread(). Besides, a forking task that has passed
1945 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1946 * is not guaranteed to have its child immediately visible in the
1947 * tasklist if we walk through it with RCU.
1949 read_lock(&tasklist_lock);
1950 do_each_thread(g, p) {
1951 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1952 task_css_set(p) != &init_css_set);
1955 * We should check if the process is exiting, otherwise
1956 * it will race with cgroup_exit() in that the list
1957 * entry won't be deleted though the process has exited.
1958 * Do it while holding siglock so that we don't end up
1959 * racing against cgroup_exit().
1961 * Interrupts were already disabled while acquiring
1962 * the css_set_lock, so we do not need to disable it
1963 * again when acquiring the sighand->siglock here.
1965 spin_lock(&p->sighand->siglock);
1966 if (!(p->flags & PF_EXITING)) {
1967 struct css_set *cset = task_css_set(p);
1969 if (!css_set_populated(cset))
1970 css_set_update_populated(cset, true);
1971 list_add_tail(&p->cg_list, &cset->tasks);
1974 spin_unlock(&p->sighand->siglock);
1975 } while_each_thread(g, p);
1976 read_unlock(&tasklist_lock);
1978 spin_unlock_irq(&css_set_lock);
1981 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1983 struct cgroup_subsys *ss;
1986 INIT_LIST_HEAD(&cgrp->self.sibling);
1987 INIT_LIST_HEAD(&cgrp->self.children);
1988 INIT_LIST_HEAD(&cgrp->cset_links);
1989 INIT_LIST_HEAD(&cgrp->pidlists);
1990 mutex_init(&cgrp->pidlist_mutex);
1991 cgrp->self.cgroup = cgrp;
1992 cgrp->self.flags |= CSS_ONLINE;
1994 for_each_subsys(ss, ssid)
1995 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1997 init_waitqueue_head(&cgrp->offline_waitq);
1998 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
2001 static void init_cgroup_root(struct cgroup_root *root,
2002 struct cgroup_sb_opts *opts)
2004 struct cgroup *cgrp = &root->cgrp;
2006 INIT_LIST_HEAD(&root->root_list);
2007 atomic_set(&root->nr_cgrps, 1);
2009 init_cgroup_housekeeping(cgrp);
2010 idr_init(&root->cgroup_idr);
2012 root->flags = opts->flags;
2013 if (opts->release_agent)
2014 strcpy(root->release_agent_path, opts->release_agent);
2016 strcpy(root->name, opts->name);
2017 if (opts->cpuset_clone_children)
2018 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2021 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2023 LIST_HEAD(tmp_links);
2024 struct cgroup *root_cgrp = &root->cgrp;
2025 struct css_set *cset;
2028 lockdep_assert_held(&cgroup_mutex);
2030 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
2033 root_cgrp->id = ret;
2034 root_cgrp->ancestor_ids[0] = ret;
2036 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
2042 * We're accessing css_set_count without locking css_set_lock here,
2043 * but that's OK - it can only be increased by someone holding
2044 * cgroup_lock, and that's us. Later rebinding may disable
2045 * controllers on the default hierarchy and thus create new csets,
2046 * which can't be more than the existing ones. Allocate 2x.
2048 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2052 ret = cgroup_init_root_id(root);
2056 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
2057 KERNFS_ROOT_CREATE_DEACTIVATED,
2059 if (IS_ERR(root->kf_root)) {
2060 ret = PTR_ERR(root->kf_root);
2063 root_cgrp->kn = root->kf_root->kn;
2065 ret = css_populate_dir(&root_cgrp->self);
2069 ret = rebind_subsystems(root, ss_mask);
2073 trace_cgroup_setup_root(root);
2076 * There must be no failure case after here, since rebinding takes
2077 * care of subsystems' refcounts, which are explicitly dropped in
2078 * the failure exit path.
2080 list_add(&root->root_list, &cgroup_roots);
2081 cgroup_root_count++;
2084 * Link the root cgroup in this hierarchy into all the css_set
2087 spin_lock_irq(&css_set_lock);
2088 hash_for_each(css_set_table, i, cset, hlist) {
2089 link_css_set(&tmp_links, cset, root_cgrp);
2090 if (css_set_populated(cset))
2091 cgroup_update_populated(root_cgrp, true);
2093 spin_unlock_irq(&css_set_lock);
2095 BUG_ON(!list_empty(&root_cgrp->self.children));
2096 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2098 kernfs_activate(root_cgrp->kn);
2103 kernfs_destroy_root(root->kf_root);
2104 root->kf_root = NULL;
2106 cgroup_exit_root_id(root);
2108 percpu_ref_exit(&root_cgrp->self.refcnt);
2110 free_cgrp_cset_links(&tmp_links);
2114 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2115 int flags, const char *unused_dev_name,
2118 bool is_v2 = fs_type == &cgroup2_fs_type;
2119 struct super_block *pinned_sb = NULL;
2120 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2121 struct cgroup_subsys *ss;
2122 struct cgroup_root *root;
2123 struct cgroup_sb_opts opts;
2124 struct dentry *dentry;
2131 /* Check if the caller has permission to mount. */
2132 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2134 return ERR_PTR(-EPERM);
2138 * The first time anyone tries to mount a cgroup, enable the list
2139 * linking each css_set to its tasks and fix up all existing tasks.
2141 if (!use_task_css_set_links)
2142 cgroup_enable_task_cg_lists();
2146 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2148 return ERR_PTR(-EINVAL);
2150 cgrp_dfl_visible = true;
2151 root = &cgrp_dfl_root;
2152 cgroup_get(&root->cgrp);
2156 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2158 /* First find the desired set of subsystems */
2159 ret = parse_cgroupfs_options(data, &opts);
2164 * Destruction of cgroup root is asynchronous, so subsystems may
2165 * still be dying after the previous unmount. Let's drain the
2166 * dying subsystems. We just need to ensure that the ones
2167 * unmounted previously finish dying and don't care about new ones
2168 * starting. Testing ref liveliness is good enough.
2170 for_each_subsys(ss, i) {
2171 if (!(opts.subsys_mask & (1 << i)) ||
2172 ss->root == &cgrp_dfl_root)
2175 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2176 mutex_unlock(&cgroup_mutex);
2178 ret = restart_syscall();
2181 cgroup_put(&ss->root->cgrp);
2184 for_each_root(root) {
2185 bool name_match = false;
2187 if (root == &cgrp_dfl_root)
2191 * If we asked for a name then it must match. Also, if
2192 * name matches but sybsys_mask doesn't, we should fail.
2193 * Remember whether name matched.
2196 if (strcmp(opts.name, root->name))
2202 * If we asked for subsystems (or explicitly for no
2203 * subsystems) then they must match.
2205 if ((opts.subsys_mask || opts.none) &&
2206 (opts.subsys_mask != root->subsys_mask)) {
2213 if (root->flags ^ opts.flags)
2214 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2217 * We want to reuse @root whose lifetime is governed by its
2218 * ->cgrp. Let's check whether @root is alive and keep it
2219 * that way. As cgroup_kill_sb() can happen anytime, we
2220 * want to block it by pinning the sb so that @root doesn't
2221 * get killed before mount is complete.
2223 * With the sb pinned, tryget_live can reliably indicate
2224 * whether @root can be reused. If it's being killed,
2225 * drain it. We can use wait_queue for the wait but this
2226 * path is super cold. Let's just sleep a bit and retry.
2228 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2229 if (IS_ERR(pinned_sb) ||
2230 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2231 mutex_unlock(&cgroup_mutex);
2232 if (!IS_ERR_OR_NULL(pinned_sb))
2233 deactivate_super(pinned_sb);
2235 ret = restart_syscall();
2244 * No such thing, create a new one. name= matching without subsys
2245 * specification is allowed for already existing hierarchies but we
2246 * can't create new one without subsys specification.
2248 if (!opts.subsys_mask && !opts.none) {
2253 /* Hierarchies may only be created in the initial cgroup namespace. */
2254 if (ns != &init_cgroup_ns) {
2260 * Release agent gets called with all capabilities,
2261 * require capabilities to set release agent.
2263 if (opts.release_agent &&
2264 ((ns->user_ns != &init_user_ns) || !capable(CAP_SYS_ADMIN))) {
2269 root = kzalloc(sizeof(*root), GFP_KERNEL);
2275 init_cgroup_root(root, &opts);
2277 ret = cgroup_setup_root(root, opts.subsys_mask);
2279 cgroup_free_root(root);
2282 mutex_unlock(&cgroup_mutex);
2284 kfree(opts.release_agent);
2289 return ERR_PTR(ret);
2292 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2293 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2297 * In non-init cgroup namespace, instead of root cgroup's
2298 * dentry, we return the dentry corresponding to the
2299 * cgroupns->root_cgrp.
2301 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2302 struct dentry *nsdentry;
2303 struct cgroup *cgrp;
2305 mutex_lock(&cgroup_mutex);
2306 spin_lock_irq(&css_set_lock);
2308 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2310 spin_unlock_irq(&css_set_lock);
2311 mutex_unlock(&cgroup_mutex);
2313 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2318 if (IS_ERR(dentry) || !new_sb)
2319 cgroup_put(&root->cgrp);
2322 * If @pinned_sb, we're reusing an existing root and holding an
2323 * extra ref on its sb. Mount is complete. Put the extra ref.
2327 deactivate_super(pinned_sb);
2334 static void cgroup_kill_sb(struct super_block *sb)
2336 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2337 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2340 * If @root doesn't have any mounts or children, start killing it.
2341 * This prevents new mounts by disabling percpu_ref_tryget_live().
2342 * cgroup_mount() may wait for @root's release.
2344 * And don't kill the default root.
2346 if (!list_empty(&root->cgrp.self.children) ||
2347 root == &cgrp_dfl_root)
2348 cgroup_put(&root->cgrp);
2350 percpu_ref_kill(&root->cgrp.self.refcnt);
2355 static struct file_system_type cgroup_fs_type = {
2357 .mount = cgroup_mount,
2358 .kill_sb = cgroup_kill_sb,
2359 .fs_flags = FS_USERNS_MOUNT,
2362 static struct file_system_type cgroup2_fs_type = {
2364 .mount = cgroup_mount,
2365 .kill_sb = cgroup_kill_sb,
2366 .fs_flags = FS_USERNS_MOUNT,
2369 static int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2370 struct cgroup_namespace *ns)
2372 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2374 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2377 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2378 struct cgroup_namespace *ns)
2382 mutex_lock(&cgroup_mutex);
2383 spin_lock_irq(&css_set_lock);
2385 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2387 spin_unlock_irq(&css_set_lock);
2388 mutex_unlock(&cgroup_mutex);
2392 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2395 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2396 * @task: target task
2397 * @buf: the buffer to write the path into
2398 * @buflen: the length of the buffer
2400 * Determine @task's cgroup on the first (the one with the lowest non-zero
2401 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2402 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2403 * cgroup controller callbacks.
2405 * Return value is the same as kernfs_path().
2407 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2409 struct cgroup_root *root;
2410 struct cgroup *cgrp;
2411 int hierarchy_id = 1;
2414 mutex_lock(&cgroup_mutex);
2415 spin_lock_irq(&css_set_lock);
2417 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2420 cgrp = task_cgroup_from_root(task, root);
2421 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2423 /* if no hierarchy exists, everyone is in "/" */
2424 ret = strlcpy(buf, "/", buflen);
2427 spin_unlock_irq(&css_set_lock);
2428 mutex_unlock(&cgroup_mutex);
2431 EXPORT_SYMBOL_GPL(task_cgroup_path);
2433 /* used to track tasks and other necessary states during migration */
2434 struct cgroup_taskset {
2435 /* the src and dst cset list running through cset->mg_node */
2436 struct list_head src_csets;
2437 struct list_head dst_csets;
2439 /* the subsys currently being processed */
2443 * Fields for cgroup_taskset_*() iteration.
2445 * Before migration is committed, the target migration tasks are on
2446 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2447 * the csets on ->dst_csets. ->csets point to either ->src_csets
2448 * or ->dst_csets depending on whether migration is committed.
2450 * ->cur_csets and ->cur_task point to the current task position
2453 struct list_head *csets;
2454 struct css_set *cur_cset;
2455 struct task_struct *cur_task;
2458 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2459 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2460 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2461 .csets = &tset.src_csets, \
2465 * cgroup_taskset_add - try to add a migration target task to a taskset
2466 * @task: target task
2467 * @tset: target taskset
2469 * Add @task, which is a migration target, to @tset. This function becomes
2470 * noop if @task doesn't need to be migrated. @task's css_set should have
2471 * been added as a migration source and @task->cg_list will be moved from
2472 * the css_set's tasks list to mg_tasks one.
2474 static void cgroup_taskset_add(struct task_struct *task,
2475 struct cgroup_taskset *tset)
2477 struct css_set *cset;
2479 lockdep_assert_held(&css_set_lock);
2481 /* @task either already exited or can't exit until the end */
2482 if (task->flags & PF_EXITING)
2485 /* leave @task alone if post_fork() hasn't linked it yet */
2486 if (list_empty(&task->cg_list))
2489 cset = task_css_set(task);
2490 if (!cset->mg_src_cgrp)
2493 list_move_tail(&task->cg_list, &cset->mg_tasks);
2494 if (list_empty(&cset->mg_node))
2495 list_add_tail(&cset->mg_node, &tset->src_csets);
2496 if (list_empty(&cset->mg_dst_cset->mg_node))
2497 list_move_tail(&cset->mg_dst_cset->mg_node,
2502 * cgroup_taskset_first - reset taskset and return the first task
2503 * @tset: taskset of interest
2504 * @dst_cssp: output variable for the destination css
2506 * @tset iteration is initialized and the first task is returned.
2508 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2509 struct cgroup_subsys_state **dst_cssp)
2511 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2512 tset->cur_task = NULL;
2514 return cgroup_taskset_next(tset, dst_cssp);
2518 * cgroup_taskset_next - iterate to the next task in taskset
2519 * @tset: taskset of interest
2520 * @dst_cssp: output variable for the destination css
2522 * Return the next task in @tset. Iteration must have been initialized
2523 * with cgroup_taskset_first().
2525 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2526 struct cgroup_subsys_state **dst_cssp)
2528 struct css_set *cset = tset->cur_cset;
2529 struct task_struct *task = tset->cur_task;
2531 while (&cset->mg_node != tset->csets) {
2533 task = list_first_entry(&cset->mg_tasks,
2534 struct task_struct, cg_list);
2536 task = list_next_entry(task, cg_list);
2538 if (&task->cg_list != &cset->mg_tasks) {
2539 tset->cur_cset = cset;
2540 tset->cur_task = task;
2543 * This function may be called both before and
2544 * after cgroup_taskset_migrate(). The two cases
2545 * can be distinguished by looking at whether @cset
2546 * has its ->mg_dst_cset set.
2548 if (cset->mg_dst_cset)
2549 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2551 *dst_cssp = cset->subsys[tset->ssid];
2556 cset = list_next_entry(cset, mg_node);
2564 * cgroup_taskset_migrate - migrate a taskset
2565 * @tset: taget taskset
2566 * @root: cgroup root the migration is taking place on
2568 * Migrate tasks in @tset as setup by migration preparation functions.
2569 * This function fails iff one of the ->can_attach callbacks fails and
2570 * guarantees that either all or none of the tasks in @tset are migrated.
2571 * @tset is consumed regardless of success.
2573 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2574 struct cgroup_root *root)
2576 struct cgroup_subsys *ss;
2577 struct task_struct *task, *tmp_task;
2578 struct css_set *cset, *tmp_cset;
2579 int ssid, failed_ssid, ret;
2581 /* methods shouldn't be called if no task is actually migrating */
2582 if (list_empty(&tset->src_csets))
2585 /* check that we can legitimately attach to the cgroup */
2586 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2587 if (ss->can_attach) {
2589 ret = ss->can_attach(tset);
2592 goto out_cancel_attach;
2595 } while_each_subsys_mask();
2598 * Now that we're guaranteed success, proceed to move all tasks to
2599 * the new cgroup. There are no failure cases after here, so this
2600 * is the commit point.
2602 spin_lock_irq(&css_set_lock);
2603 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2604 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2605 struct css_set *from_cset = task_css_set(task);
2606 struct css_set *to_cset = cset->mg_dst_cset;
2608 get_css_set(to_cset);
2609 css_set_move_task(task, from_cset, to_cset, true);
2610 put_css_set_locked(from_cset);
2613 spin_unlock_irq(&css_set_lock);
2616 * Migration is committed, all target tasks are now on dst_csets.
2617 * Nothing is sensitive to fork() after this point. Notify
2618 * controllers that migration is complete.
2620 tset->csets = &tset->dst_csets;
2622 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2627 } while_each_subsys_mask();
2630 goto out_release_tset;
2633 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2634 if (ssid == failed_ssid)
2636 if (ss->cancel_attach) {
2638 ss->cancel_attach(tset);
2640 } while_each_subsys_mask();
2642 spin_lock_irq(&css_set_lock);
2643 list_splice_init(&tset->dst_csets, &tset->src_csets);
2644 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2645 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2646 list_del_init(&cset->mg_node);
2648 spin_unlock_irq(&css_set_lock);
2653 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2654 * @dst_cgrp: destination cgroup to test
2656 * On the default hierarchy, except for the root, subtree_control must be
2657 * zero for migration destination cgroups with tasks so that child cgroups
2658 * don't compete against tasks.
2660 static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2662 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2663 !dst_cgrp->subtree_control;
2667 * cgroup_migrate_finish - cleanup after attach
2668 * @preloaded_csets: list of preloaded css_sets
2670 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2671 * those functions for details.
2673 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2675 struct css_set *cset, *tmp_cset;
2677 lockdep_assert_held(&cgroup_mutex);
2679 spin_lock_irq(&css_set_lock);
2680 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2681 cset->mg_src_cgrp = NULL;
2682 cset->mg_dst_cgrp = NULL;
2683 cset->mg_dst_cset = NULL;
2684 list_del_init(&cset->mg_preload_node);
2685 put_css_set_locked(cset);
2687 spin_unlock_irq(&css_set_lock);
2691 * cgroup_migrate_add_src - add a migration source css_set
2692 * @src_cset: the source css_set to add
2693 * @dst_cgrp: the destination cgroup
2694 * @preloaded_csets: list of preloaded css_sets
2696 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2697 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2698 * up by cgroup_migrate_finish().
2700 * This function may be called without holding cgroup_threadgroup_rwsem
2701 * even if the target is a process. Threads may be created and destroyed
2702 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2703 * into play and the preloaded css_sets are guaranteed to cover all
2706 static void cgroup_migrate_add_src(struct css_set *src_cset,
2707 struct cgroup *dst_cgrp,
2708 struct list_head *preloaded_csets)
2710 struct cgroup *src_cgrp;
2712 lockdep_assert_held(&cgroup_mutex);
2713 lockdep_assert_held(&css_set_lock);
2716 * If ->dead, @src_set is associated with one or more dead cgroups
2717 * and doesn't contain any migratable tasks. Ignore it early so
2718 * that the rest of migration path doesn't get confused by it.
2723 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2725 if (!list_empty(&src_cset->mg_preload_node))
2728 WARN_ON(src_cset->mg_src_cgrp);
2729 WARN_ON(src_cset->mg_dst_cgrp);
2730 WARN_ON(!list_empty(&src_cset->mg_tasks));
2731 WARN_ON(!list_empty(&src_cset->mg_node));
2733 src_cset->mg_src_cgrp = src_cgrp;
2734 src_cset->mg_dst_cgrp = dst_cgrp;
2735 get_css_set(src_cset);
2736 list_add(&src_cset->mg_preload_node, preloaded_csets);
2740 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2741 * @preloaded_csets: list of preloaded source css_sets
2743 * Tasks are about to be moved and all the source css_sets have been
2744 * preloaded to @preloaded_csets. This function looks up and pins all
2745 * destination css_sets, links each to its source, and append them to
2748 * This function must be called after cgroup_migrate_add_src() has been
2749 * called on each migration source css_set. After migration is performed
2750 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2753 static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2756 struct css_set *src_cset, *tmp_cset;
2758 lockdep_assert_held(&cgroup_mutex);
2760 /* look up the dst cset for each src cset and link it to src */
2761 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2762 struct css_set *dst_cset;
2764 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2768 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2771 * If src cset equals dst, it's noop. Drop the src.
2772 * cgroup_migrate() will skip the cset too. Note that we
2773 * can't handle src == dst as some nodes are used by both.
2775 if (src_cset == dst_cset) {
2776 src_cset->mg_src_cgrp = NULL;
2777 src_cset->mg_dst_cgrp = NULL;
2778 list_del_init(&src_cset->mg_preload_node);
2779 put_css_set(src_cset);
2780 put_css_set(dst_cset);
2784 src_cset->mg_dst_cset = dst_cset;
2786 if (list_empty(&dst_cset->mg_preload_node))
2787 list_add(&dst_cset->mg_preload_node, &csets);
2789 put_css_set(dst_cset);
2792 list_splice_tail(&csets, preloaded_csets);
2795 cgroup_migrate_finish(&csets);
2800 * cgroup_migrate - migrate a process or task to a cgroup
2801 * @leader: the leader of the process or the task to migrate
2802 * @threadgroup: whether @leader points to the whole process or a single task
2803 * @root: cgroup root migration is taking place on
2805 * Migrate a process or task denoted by @leader. If migrating a process,
2806 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2807 * responsible for invoking cgroup_migrate_add_src() and
2808 * cgroup_migrate_prepare_dst() on the targets before invoking this
2809 * function and following up with cgroup_migrate_finish().
2811 * As long as a controller's ->can_attach() doesn't fail, this function is
2812 * guaranteed to succeed. This means that, excluding ->can_attach()
2813 * failure, when migrating multiple targets, the success or failure can be
2814 * decided for all targets by invoking group_migrate_prepare_dst() before
2815 * actually starting migrating.
2817 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2818 struct cgroup_root *root)
2820 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2821 struct task_struct *task;
2824 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2825 * already PF_EXITING could be freed from underneath us unless we
2826 * take an rcu_read_lock.
2828 spin_lock_irq(&css_set_lock);
2832 cgroup_taskset_add(task, &tset);
2835 } while_each_thread(leader, task);
2837 spin_unlock_irq(&css_set_lock);
2839 return cgroup_taskset_migrate(&tset, root);
2843 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2844 * @dst_cgrp: the cgroup to attach to
2845 * @leader: the task or the leader of the threadgroup to be attached
2846 * @threadgroup: attach the whole threadgroup?
2848 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2850 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2851 struct task_struct *leader, bool threadgroup)
2853 LIST_HEAD(preloaded_csets);
2854 struct task_struct *task;
2857 if (!cgroup_may_migrate_to(dst_cgrp))
2860 /* look up all src csets */
2861 spin_lock_irq(&css_set_lock);
2865 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2869 } while_each_thread(leader, task);
2871 spin_unlock_irq(&css_set_lock);
2873 /* prepare dst csets and commit */
2874 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2876 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2878 cgroup_migrate_finish(&preloaded_csets);
2881 trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2886 static int cgroup_procs_write_permission(struct task_struct *task,
2887 struct cgroup *dst_cgrp,
2888 struct kernfs_open_file *of)
2890 const struct cred *cred = current_cred();
2891 const struct cred *tcred = get_task_cred(task);
2895 * even if we're attaching all tasks in the thread group, we only
2896 * need to check permissions on one of them.
2898 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2899 !uid_eq(cred->euid, tcred->uid) &&
2900 !uid_eq(cred->euid, tcred->suid))
2903 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2904 struct super_block *sb = of->file->f_path.dentry->d_sb;
2905 struct cgroup *cgrp;
2906 struct inode *inode;
2908 spin_lock_irq(&css_set_lock);
2909 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2910 spin_unlock_irq(&css_set_lock);
2912 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2913 cgrp = cgroup_parent(cgrp);
2916 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2918 ret = inode_permission(inode, MAY_WRITE);
2928 * Find the task_struct of the task to attach by vpid and pass it along to the
2929 * function to attach either it or all tasks in its threadgroup. Will lock
2930 * cgroup_mutex and threadgroup.
2932 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2933 size_t nbytes, loff_t off, bool threadgroup)
2935 struct task_struct *tsk;
2936 struct cgroup_subsys *ss;
2937 struct cgroup *cgrp;
2941 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2944 cgrp = cgroup_kn_lock_live(of->kn, false);
2948 percpu_down_write(&cgroup_threadgroup_rwsem);
2951 tsk = find_task_by_vpid(pid);
2954 goto out_unlock_rcu;
2961 tsk = tsk->group_leader;
2964 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2965 * If userland migrates such a kthread to a non-root cgroup, it can
2966 * become trapped in a cpuset, or RT kthread may be born in a
2967 * cgroup with no rt_runtime allocated. Just say no.
2969 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2971 goto out_unlock_rcu;
2974 get_task_struct(tsk);
2977 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2979 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2981 put_task_struct(tsk);
2982 goto out_unlock_threadgroup;
2986 out_unlock_threadgroup:
2987 percpu_up_write(&cgroup_threadgroup_rwsem);
2988 for_each_subsys(ss, ssid)
2989 if (ss->post_attach)
2991 cgroup_kn_unlock(of->kn);
2992 return ret ?: nbytes;
2996 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2997 * @from: attach to all cgroups of a given task
2998 * @tsk: the task to be attached
3000 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
3002 struct cgroup_root *root;
3005 mutex_lock(&cgroup_mutex);
3006 percpu_down_write(&cgroup_threadgroup_rwsem);
3007 for_each_root(root) {
3008 struct cgroup *from_cgrp;
3010 if (root == &cgrp_dfl_root)
3013 spin_lock_irq(&css_set_lock);
3014 from_cgrp = task_cgroup_from_root(from, root);
3015 spin_unlock_irq(&css_set_lock);
3017 retval = cgroup_attach_task(from_cgrp, tsk, false);
3021 percpu_up_write(&cgroup_threadgroup_rwsem);
3022 mutex_unlock(&cgroup_mutex);
3026 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
3028 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
3029 char *buf, size_t nbytes, loff_t off)
3031 return __cgroup_procs_write(of, buf, nbytes, off, false);
3034 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
3035 char *buf, size_t nbytes, loff_t off)
3037 return __cgroup_procs_write(of, buf, nbytes, off, true);
3040 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
3041 char *buf, size_t nbytes, loff_t off)
3043 struct cgroup *cgrp;
3045 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
3048 * Release agent gets called with all capabilities,
3049 * require capabilities to set release agent.
3051 if ((of->file->f_cred->user_ns != &init_user_ns) ||
3052 !capable(CAP_SYS_ADMIN))
3055 cgrp = cgroup_kn_lock_live(of->kn, false);
3058 spin_lock(&release_agent_path_lock);
3059 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
3060 sizeof(cgrp->root->release_agent_path));
3061 spin_unlock(&release_agent_path_lock);
3062 cgroup_kn_unlock(of->kn);
3066 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
3068 struct cgroup *cgrp = seq_css(seq)->cgroup;
3070 spin_lock(&release_agent_path_lock);
3071 seq_puts(seq, cgrp->root->release_agent_path);
3072 spin_unlock(&release_agent_path_lock);
3073 seq_putc(seq, '\n');
3077 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
3079 seq_puts(seq, "0\n");
3083 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3085 struct cgroup_subsys *ss;
3086 bool printed = false;
3089 do_each_subsys_mask(ss, ssid, ss_mask) {
3092 seq_printf(seq, "%s", ss->name);
3094 } while_each_subsys_mask();
3096 seq_putc(seq, '\n');
3099 /* show controllers which are enabled from the parent */
3100 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3102 struct cgroup *cgrp = seq_css(seq)->cgroup;
3104 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3108 /* show controllers which are enabled for a given cgroup's children */
3109 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3111 struct cgroup *cgrp = seq_css(seq)->cgroup;
3113 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3118 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3119 * @cgrp: root of the subtree to update csses for
3121 * @cgrp's control masks have changed and its subtree's css associations
3122 * need to be updated accordingly. This function looks up all css_sets
3123 * which are attached to the subtree, creates the matching updated css_sets
3124 * and migrates the tasks to the new ones.
3126 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3128 LIST_HEAD(preloaded_csets);
3129 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3130 struct cgroup_subsys_state *d_css;
3131 struct cgroup *dsct;
3132 struct css_set *src_cset;
3135 lockdep_assert_held(&cgroup_mutex);
3137 percpu_down_write(&cgroup_threadgroup_rwsem);
3139 /* look up all csses currently attached to @cgrp's subtree */
3140 spin_lock_irq(&css_set_lock);
3141 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3142 struct cgrp_cset_link *link;
3144 list_for_each_entry(link, &dsct->cset_links, cset_link)
3145 cgroup_migrate_add_src(link->cset, dsct,
3148 spin_unlock_irq(&css_set_lock);
3150 /* NULL dst indicates self on default hierarchy */
3151 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3155 spin_lock_irq(&css_set_lock);
3156 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3157 struct task_struct *task, *ntask;
3159 /* src_csets precede dst_csets, break on the first dst_cset */
3160 if (!src_cset->mg_src_cgrp)
3163 /* all tasks in src_csets need to be migrated */
3164 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3165 cgroup_taskset_add(task, &tset);
3167 spin_unlock_irq(&css_set_lock);
3169 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3171 cgroup_migrate_finish(&preloaded_csets);
3172 percpu_up_write(&cgroup_threadgroup_rwsem);
3177 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3178 * @cgrp: root of the target subtree
3180 * Because css offlining is asynchronous, userland may try to re-enable a
3181 * controller while the previous css is still around. This function grabs
3182 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3184 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3185 __acquires(&cgroup_mutex)
3187 struct cgroup *dsct;
3188 struct cgroup_subsys_state *d_css;
3189 struct cgroup_subsys *ss;
3193 mutex_lock(&cgroup_mutex);
3195 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3196 for_each_subsys(ss, ssid) {
3197 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3200 if (!css || !percpu_ref_is_dying(&css->refcnt))
3204 prepare_to_wait(&dsct->offline_waitq, &wait,
3205 TASK_UNINTERRUPTIBLE);
3207 mutex_unlock(&cgroup_mutex);
3209 finish_wait(&dsct->offline_waitq, &wait);
3218 * cgroup_save_control - save control masks of a subtree
3219 * @cgrp: root of the target subtree
3221 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3222 * prefixed fields for @cgrp's subtree including @cgrp itself.
3224 static void cgroup_save_control(struct cgroup *cgrp)
3226 struct cgroup *dsct;
3227 struct cgroup_subsys_state *d_css;
3229 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3230 dsct->old_subtree_control = dsct->subtree_control;
3231 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3236 * cgroup_propagate_control - refresh control masks of a subtree
3237 * @cgrp: root of the target subtree
3239 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3240 * ->subtree_control and propagate controller availability through the
3241 * subtree so that descendants don't have unavailable controllers enabled.
3243 static void cgroup_propagate_control(struct cgroup *cgrp)
3245 struct cgroup *dsct;
3246 struct cgroup_subsys_state *d_css;
3248 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3249 dsct->subtree_control &= cgroup_control(dsct);
3250 dsct->subtree_ss_mask =
3251 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3252 cgroup_ss_mask(dsct));
3257 * cgroup_restore_control - restore control masks of a subtree
3258 * @cgrp: root of the target subtree
3260 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3261 * prefixed fields for @cgrp's subtree including @cgrp itself.
3263 static void cgroup_restore_control(struct cgroup *cgrp)
3265 struct cgroup *dsct;
3266 struct cgroup_subsys_state *d_css;
3268 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3269 dsct->subtree_control = dsct->old_subtree_control;
3270 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3274 static bool css_visible(struct cgroup_subsys_state *css)
3276 struct cgroup_subsys *ss = css->ss;
3277 struct cgroup *cgrp = css->cgroup;
3279 if (cgroup_control(cgrp) & (1 << ss->id))
3281 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3283 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3287 * cgroup_apply_control_enable - enable or show csses according to control
3288 * @cgrp: root of the target subtree
3290 * Walk @cgrp's subtree and create new csses or make the existing ones
3291 * visible. A css is created invisible if it's being implicitly enabled
3292 * through dependency. An invisible css is made visible when the userland
3293 * explicitly enables it.
3295 * Returns 0 on success, -errno on failure. On failure, csses which have
3296 * been processed already aren't cleaned up. The caller is responsible for
3297 * cleaning up with cgroup_apply_control_disble().
3299 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3301 struct cgroup *dsct;
3302 struct cgroup_subsys_state *d_css;
3303 struct cgroup_subsys *ss;
3306 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3307 for_each_subsys(ss, ssid) {
3308 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3310 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3312 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3316 css = css_create(dsct, ss);
3318 return PTR_ERR(css);
3321 if (css_visible(css)) {
3322 ret = css_populate_dir(css);
3333 * cgroup_apply_control_disable - kill or hide csses according to control
3334 * @cgrp: root of the target subtree
3336 * Walk @cgrp's subtree and kill and hide csses so that they match
3337 * cgroup_ss_mask() and cgroup_visible_mask().
3339 * A css is hidden when the userland requests it to be disabled while other
3340 * subsystems are still depending on it. The css must not actively control
3341 * resources and be in the vanilla state if it's made visible again later.
3342 * Controllers which may be depended upon should provide ->css_reset() for
3345 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3347 struct cgroup *dsct;
3348 struct cgroup_subsys_state *d_css;
3349 struct cgroup_subsys *ss;
3352 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3353 for_each_subsys(ss, ssid) {
3354 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3356 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3362 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3364 } else if (!css_visible(css)) {
3374 * cgroup_apply_control - apply control mask updates to the subtree
3375 * @cgrp: root of the target subtree
3377 * subsystems can be enabled and disabled in a subtree using the following
3380 * 1. Call cgroup_save_control() to stash the current state.
3381 * 2. Update ->subtree_control masks in the subtree as desired.
3382 * 3. Call cgroup_apply_control() to apply the changes.
3383 * 4. Optionally perform other related operations.
3384 * 5. Call cgroup_finalize_control() to finish up.
3386 * This function implements step 3 and propagates the mask changes
3387 * throughout @cgrp's subtree, updates csses accordingly and perform
3388 * process migrations.
3390 static int cgroup_apply_control(struct cgroup *cgrp)
3394 cgroup_propagate_control(cgrp);
3396 ret = cgroup_apply_control_enable(cgrp);
3401 * At this point, cgroup_e_css() results reflect the new csses
3402 * making the following cgroup_update_dfl_csses() properly update
3403 * css associations of all tasks in the subtree.
3405 ret = cgroup_update_dfl_csses(cgrp);
3413 * cgroup_finalize_control - finalize control mask update
3414 * @cgrp: root of the target subtree
3415 * @ret: the result of the update
3417 * Finalize control mask update. See cgroup_apply_control() for more info.
3419 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3422 cgroup_restore_control(cgrp);
3423 cgroup_propagate_control(cgrp);
3426 cgroup_apply_control_disable(cgrp);
3429 /* change the enabled child controllers for a cgroup in the default hierarchy */
3430 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3431 char *buf, size_t nbytes,
3434 u16 enable = 0, disable = 0;
3435 struct cgroup *cgrp, *child;
3436 struct cgroup_subsys *ss;
3441 * Parse input - space separated list of subsystem names prefixed
3442 * with either + or -.
3444 buf = strstrip(buf);
3445 while ((tok = strsep(&buf, " "))) {
3448 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3449 if (!cgroup_ssid_enabled(ssid) ||
3450 strcmp(tok + 1, ss->name))
3454 enable |= 1 << ssid;
3455 disable &= ~(1 << ssid);
3456 } else if (*tok == '-') {
3457 disable |= 1 << ssid;
3458 enable &= ~(1 << ssid);
3463 } while_each_subsys_mask();
3464 if (ssid == CGROUP_SUBSYS_COUNT)
3468 cgrp = cgroup_kn_lock_live(of->kn, true);
3472 for_each_subsys(ss, ssid) {
3473 if (enable & (1 << ssid)) {
3474 if (cgrp->subtree_control & (1 << ssid)) {
3475 enable &= ~(1 << ssid);
3479 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3483 } else if (disable & (1 << ssid)) {
3484 if (!(cgrp->subtree_control & (1 << ssid))) {
3485 disable &= ~(1 << ssid);
3489 /* a child has it enabled? */
3490 cgroup_for_each_live_child(child, cgrp) {
3491 if (child->subtree_control & (1 << ssid)) {
3499 if (!enable && !disable) {
3505 * Except for the root, subtree_control must be zero for a cgroup
3506 * with tasks so that child cgroups don't compete against tasks.
3508 if (enable && cgroup_parent(cgrp)) {
3509 struct cgrp_cset_link *link;
3512 * Because namespaces pin csets too, @cgrp->cset_links
3513 * might not be empty even when @cgrp is empty. Walk and
3516 spin_lock_irq(&css_set_lock);
3519 list_for_each_entry(link, &cgrp->cset_links, cset_link) {
3520 if (css_set_populated(link->cset)) {
3526 spin_unlock_irq(&css_set_lock);
3532 /* save and update control masks and prepare csses */
3533 cgroup_save_control(cgrp);
3535 cgrp->subtree_control |= enable;
3536 cgrp->subtree_control &= ~disable;
3538 ret = cgroup_apply_control(cgrp);
3539 cgroup_finalize_control(cgrp, ret);
3543 kernfs_activate(cgrp->kn);
3545 cgroup_kn_unlock(of->kn);
3546 return ret ?: nbytes;
3549 static int cgroup_events_show(struct seq_file *seq, void *v)
3551 seq_printf(seq, "populated %d\n",
3552 cgroup_is_populated(seq_css(seq)->cgroup));
3556 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3557 size_t nbytes, loff_t off)
3559 struct cgroup *cgrp = of->kn->parent->priv;
3560 struct cftype *cft = of->kn->priv;
3561 struct cgroup_subsys_state *css;
3565 return cft->write(of, buf, nbytes, off);
3568 * kernfs guarantees that a file isn't deleted with operations in
3569 * flight, which means that the matching css is and stays alive and
3570 * doesn't need to be pinned. The RCU locking is not necessary
3571 * either. It's just for the convenience of using cgroup_css().
3574 css = cgroup_css(cgrp, cft->ss);
3577 if (cft->write_u64) {
3578 unsigned long long v;
3579 ret = kstrtoull(buf, 0, &v);
3581 ret = cft->write_u64(css, cft, v);
3582 } else if (cft->write_s64) {
3584 ret = kstrtoll(buf, 0, &v);
3586 ret = cft->write_s64(css, cft, v);
3591 return ret ?: nbytes;
3594 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3596 return seq_cft(seq)->seq_start(seq, ppos);
3599 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3601 return seq_cft(seq)->seq_next(seq, v, ppos);
3604 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3606 seq_cft(seq)->seq_stop(seq, v);
3609 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3611 struct cftype *cft = seq_cft(m);
3612 struct cgroup_subsys_state *css = seq_css(m);
3615 return cft->seq_show(m, arg);
3618 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3619 else if (cft->read_s64)
3620 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3626 static struct kernfs_ops cgroup_kf_single_ops = {
3627 .atomic_write_len = PAGE_SIZE,
3628 .write = cgroup_file_write,
3629 .seq_show = cgroup_seqfile_show,
3632 static struct kernfs_ops cgroup_kf_ops = {
3633 .atomic_write_len = PAGE_SIZE,
3634 .write = cgroup_file_write,
3635 .seq_start = cgroup_seqfile_start,
3636 .seq_next = cgroup_seqfile_next,
3637 .seq_stop = cgroup_seqfile_stop,
3638 .seq_show = cgroup_seqfile_show,
3642 * cgroup_rename - Only allow simple rename of directories in place.
3644 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3645 const char *new_name_str)
3647 struct cgroup *cgrp = kn->priv;
3650 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
3651 if (strchr(new_name_str, '\n'))
3654 if (kernfs_type(kn) != KERNFS_DIR)
3656 if (kn->parent != new_parent)
3660 * This isn't a proper migration and its usefulness is very
3661 * limited. Disallow on the default hierarchy.
3663 if (cgroup_on_dfl(cgrp))
3667 * We're gonna grab cgroup_mutex which nests outside kernfs
3668 * active_ref. kernfs_rename() doesn't require active_ref
3669 * protection. Break them before grabbing cgroup_mutex.
3671 kernfs_break_active_protection(new_parent);
3672 kernfs_break_active_protection(kn);
3674 mutex_lock(&cgroup_mutex);
3676 ret = kernfs_rename(kn, new_parent, new_name_str);
3678 trace_cgroup_rename(cgrp);
3680 mutex_unlock(&cgroup_mutex);
3682 kernfs_unbreak_active_protection(kn);
3683 kernfs_unbreak_active_protection(new_parent);
3687 /* set uid and gid of cgroup dirs and files to that of the creator */
3688 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3690 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3691 .ia_uid = current_fsuid(),
3692 .ia_gid = current_fsgid(), };
3694 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3695 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3698 return kernfs_setattr(kn, &iattr);
3701 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3704 char name[CGROUP_FILE_NAME_MAX];
3705 struct kernfs_node *kn;
3706 struct lock_class_key *key = NULL;
3709 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3710 key = &cft->lockdep_key;
3712 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3713 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3718 ret = cgroup_kn_set_ugid(kn);
3724 if (cft->file_offset) {
3725 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3727 spin_lock_irq(&cgroup_file_kn_lock);
3729 spin_unlock_irq(&cgroup_file_kn_lock);
3736 * cgroup_addrm_files - add or remove files to a cgroup directory
3737 * @css: the target css
3738 * @cgrp: the target cgroup (usually css->cgroup)
3739 * @cfts: array of cftypes to be added
3740 * @is_add: whether to add or remove
3742 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3743 * For removals, this function never fails.
3745 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3746 struct cgroup *cgrp, struct cftype cfts[],
3749 struct cftype *cft, *cft_end = NULL;
3752 lockdep_assert_held(&cgroup_mutex);
3755 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3756 /* does cft->flags tell us to skip this file on @cgrp? */
3757 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3759 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3761 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3763 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3767 ret = cgroup_add_file(css, cgrp, cft);
3769 pr_warn("%s: failed to add %s, err=%d\n",
3770 __func__, cft->name, ret);
3776 cgroup_rm_file(cgrp, cft);
3782 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3785 struct cgroup_subsys *ss = cfts[0].ss;
3786 struct cgroup *root = &ss->root->cgrp;
3787 struct cgroup_subsys_state *css;
3790 lockdep_assert_held(&cgroup_mutex);
3792 /* add/rm files for all cgroups created before */
3793 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3794 struct cgroup *cgrp = css->cgroup;
3796 if (!(css->flags & CSS_VISIBLE))
3799 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3805 kernfs_activate(root->kn);
3809 static void cgroup_exit_cftypes(struct cftype *cfts)
3813 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3814 /* free copy for custom atomic_write_len, see init_cftypes() */
3815 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3820 /* revert flags set by cgroup core while adding @cfts */
3821 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3825 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3829 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3830 struct kernfs_ops *kf_ops;
3832 WARN_ON(cft->ss || cft->kf_ops);
3835 kf_ops = &cgroup_kf_ops;
3837 kf_ops = &cgroup_kf_single_ops;
3840 * Ugh... if @cft wants a custom max_write_len, we need to
3841 * make a copy of kf_ops to set its atomic_write_len.
3843 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3844 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3846 cgroup_exit_cftypes(cfts);
3849 kf_ops->atomic_write_len = cft->max_write_len;
3852 cft->kf_ops = kf_ops;
3859 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3861 lockdep_assert_held(&cgroup_mutex);
3863 if (!cfts || !cfts[0].ss)
3866 list_del(&cfts->node);
3867 cgroup_apply_cftypes(cfts, false);
3868 cgroup_exit_cftypes(cfts);
3873 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3874 * @cfts: zero-length name terminated array of cftypes
3876 * Unregister @cfts. Files described by @cfts are removed from all
3877 * existing cgroups and all future cgroups won't have them either. This
3878 * function can be called anytime whether @cfts' subsys is attached or not.
3880 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3883 int cgroup_rm_cftypes(struct cftype *cfts)
3887 mutex_lock(&cgroup_mutex);
3888 ret = cgroup_rm_cftypes_locked(cfts);
3889 mutex_unlock(&cgroup_mutex);
3894 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3895 * @ss: target cgroup subsystem
3896 * @cfts: zero-length name terminated array of cftypes
3898 * Register @cfts to @ss. Files described by @cfts are created for all
3899 * existing cgroups to which @ss is attached and all future cgroups will
3900 * have them too. This function can be called anytime whether @ss is
3903 * Returns 0 on successful registration, -errno on failure. Note that this
3904 * function currently returns 0 as long as @cfts registration is successful
3905 * even if some file creation attempts on existing cgroups fail.
3907 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3911 if (!cgroup_ssid_enabled(ss->id))
3914 if (!cfts || cfts[0].name[0] == '\0')
3917 ret = cgroup_init_cftypes(ss, cfts);
3921 mutex_lock(&cgroup_mutex);
3923 list_add_tail(&cfts->node, &ss->cfts);
3924 ret = cgroup_apply_cftypes(cfts, true);
3926 cgroup_rm_cftypes_locked(cfts);
3928 mutex_unlock(&cgroup_mutex);
3933 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3934 * @ss: target cgroup subsystem
3935 * @cfts: zero-length name terminated array of cftypes
3937 * Similar to cgroup_add_cftypes() but the added files are only used for
3938 * the default hierarchy.
3940 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3944 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3945 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3946 return cgroup_add_cftypes(ss, cfts);
3950 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3951 * @ss: target cgroup subsystem
3952 * @cfts: zero-length name terminated array of cftypes
3954 * Similar to cgroup_add_cftypes() but the added files are only used for
3955 * the legacy hierarchies.
3957 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3961 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3962 cft->flags |= __CFTYPE_NOT_ON_DFL;
3963 return cgroup_add_cftypes(ss, cfts);
3967 * cgroup_file_notify - generate a file modified event for a cgroup_file
3968 * @cfile: target cgroup_file
3970 * @cfile must have been obtained by setting cftype->file_offset.
3972 void cgroup_file_notify(struct cgroup_file *cfile)
3974 unsigned long flags;
3976 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3978 kernfs_notify(cfile->kn);
3979 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3983 * cgroup_task_count - count the number of tasks in a cgroup.
3984 * @cgrp: the cgroup in question
3986 * Return the number of tasks in the cgroup. The returned number can be
3987 * higher than the actual number of tasks due to css_set references from
3988 * namespace roots and temporary usages.
3990 static int cgroup_task_count(const struct cgroup *cgrp)
3993 struct cgrp_cset_link *link;
3995 spin_lock_irq(&css_set_lock);
3996 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3997 count += atomic_read(&link->cset->refcount);
3998 spin_unlock_irq(&css_set_lock);
4003 * css_next_child - find the next child of a given css
4004 * @pos: the current position (%NULL to initiate traversal)
4005 * @parent: css whose children to walk
4007 * This function returns the next child of @parent and should be called
4008 * under either cgroup_mutex or RCU read lock. The only requirement is
4009 * that @parent and @pos are accessible. The next sibling is guaranteed to
4010 * be returned regardless of their states.
4012 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4013 * css which finished ->css_online() is guaranteed to be visible in the
4014 * future iterations and will stay visible until the last reference is put.
4015 * A css which hasn't finished ->css_online() or already finished
4016 * ->css_offline() may show up during traversal. It's each subsystem's
4017 * responsibility to synchronize against on/offlining.
4019 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4020 struct cgroup_subsys_state *parent)
4022 struct cgroup_subsys_state *next;
4024 cgroup_assert_mutex_or_rcu_locked();
4027 * @pos could already have been unlinked from the sibling list.
4028 * Once a cgroup is removed, its ->sibling.next is no longer
4029 * updated when its next sibling changes. CSS_RELEASED is set when
4030 * @pos is taken off list, at which time its next pointer is valid,
4031 * and, as releases are serialized, the one pointed to by the next
4032 * pointer is guaranteed to not have started release yet. This
4033 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4034 * critical section, the one pointed to by its next pointer is
4035 * guaranteed to not have finished its RCU grace period even if we
4036 * have dropped rcu_read_lock() inbetween iterations.
4038 * If @pos has CSS_RELEASED set, its next pointer can't be
4039 * dereferenced; however, as each css is given a monotonically
4040 * increasing unique serial number and always appended to the
4041 * sibling list, the next one can be found by walking the parent's
4042 * children until the first css with higher serial number than
4043 * @pos's. While this path can be slower, it happens iff iteration
4044 * races against release and the race window is very small.
4047 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4048 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4049 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4051 list_for_each_entry_rcu(next, &parent->children, sibling)
4052 if (next->serial_nr > pos->serial_nr)
4057 * @next, if not pointing to the head, can be dereferenced and is
4060 if (&next->sibling != &parent->children)
4066 * css_next_descendant_pre - find the next descendant for pre-order walk
4067 * @pos: the current position (%NULL to initiate traversal)
4068 * @root: css whose descendants to walk
4070 * To be used by css_for_each_descendant_pre(). Find the next descendant
4071 * to visit for pre-order traversal of @root's descendants. @root is
4072 * included in the iteration and the first node to be visited.
4074 * While this function requires cgroup_mutex or RCU read locking, it
4075 * doesn't require the whole traversal to be contained in a single critical
4076 * section. This function will return the correct next descendant as long
4077 * as both @pos and @root are accessible and @pos is a descendant of @root.
4079 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4080 * css which finished ->css_online() is guaranteed to be visible in the
4081 * future iterations and will stay visible until the last reference is put.
4082 * A css which hasn't finished ->css_online() or already finished
4083 * ->css_offline() may show up during traversal. It's each subsystem's
4084 * responsibility to synchronize against on/offlining.
4086 struct cgroup_subsys_state *
4087 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4088 struct cgroup_subsys_state *root)
4090 struct cgroup_subsys_state *next;
4092 cgroup_assert_mutex_or_rcu_locked();
4094 /* if first iteration, visit @root */
4098 /* visit the first child if exists */
4099 next = css_next_child(NULL, pos);
4103 /* no child, visit my or the closest ancestor's next sibling */
4104 while (pos != root) {
4105 next = css_next_child(pos, pos->parent);
4115 * css_rightmost_descendant - return the rightmost descendant of a css
4116 * @pos: css of interest
4118 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4119 * is returned. This can be used during pre-order traversal to skip
4122 * While this function requires cgroup_mutex or RCU read locking, it
4123 * doesn't require the whole traversal to be contained in a single critical
4124 * section. This function will return the correct rightmost descendant as
4125 * long as @pos is accessible.
4127 struct cgroup_subsys_state *
4128 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4130 struct cgroup_subsys_state *last, *tmp;
4132 cgroup_assert_mutex_or_rcu_locked();
4136 /* ->prev isn't RCU safe, walk ->next till the end */
4138 css_for_each_child(tmp, last)
4145 static struct cgroup_subsys_state *
4146 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4148 struct cgroup_subsys_state *last;
4152 pos = css_next_child(NULL, pos);
4159 * css_next_descendant_post - find the next descendant for post-order walk
4160 * @pos: the current position (%NULL to initiate traversal)
4161 * @root: css whose descendants to walk
4163 * To be used by css_for_each_descendant_post(). Find the next descendant
4164 * to visit for post-order traversal of @root's descendants. @root is
4165 * included in the iteration and the last node to be visited.
4167 * While this function requires cgroup_mutex or RCU read locking, it
4168 * doesn't require the whole traversal to be contained in a single critical
4169 * section. This function will return the correct next descendant as long
4170 * as both @pos and @cgroup are accessible and @pos is a descendant of
4173 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4174 * css which finished ->css_online() is guaranteed to be visible in the
4175 * future iterations and will stay visible until the last reference is put.
4176 * A css which hasn't finished ->css_online() or already finished
4177 * ->css_offline() may show up during traversal. It's each subsystem's
4178 * responsibility to synchronize against on/offlining.
4180 struct cgroup_subsys_state *
4181 css_next_descendant_post(struct cgroup_subsys_state *pos,
4182 struct cgroup_subsys_state *root)
4184 struct cgroup_subsys_state *next;
4186 cgroup_assert_mutex_or_rcu_locked();
4188 /* if first iteration, visit leftmost descendant which may be @root */
4190 return css_leftmost_descendant(root);
4192 /* if we visited @root, we're done */
4196 /* if there's an unvisited sibling, visit its leftmost descendant */
4197 next = css_next_child(pos, pos->parent);
4199 return css_leftmost_descendant(next);
4201 /* no sibling left, visit parent */
4206 * css_has_online_children - does a css have online children
4207 * @css: the target css
4209 * Returns %true if @css has any online children; otherwise, %false. This
4210 * function can be called from any context but the caller is responsible
4211 * for synchronizing against on/offlining as necessary.
4213 bool css_has_online_children(struct cgroup_subsys_state *css)
4215 struct cgroup_subsys_state *child;
4219 css_for_each_child(child, css) {
4220 if (child->flags & CSS_ONLINE) {
4230 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4231 * @it: the iterator to advance
4233 * Advance @it to the next css_set to walk.
4235 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4237 struct list_head *l = it->cset_pos;
4238 struct cgrp_cset_link *link;
4239 struct css_set *cset;
4241 lockdep_assert_held(&css_set_lock);
4243 /* Advance to the next non-empty css_set */
4246 if (l == it->cset_head) {
4247 it->cset_pos = NULL;
4248 it->task_pos = NULL;
4253 cset = container_of(l, struct css_set,
4254 e_cset_node[it->ss->id]);
4256 link = list_entry(l, struct cgrp_cset_link, cset_link);
4259 } while (!css_set_populated(cset));
4263 if (!list_empty(&cset->tasks))
4264 it->task_pos = cset->tasks.next;
4266 it->task_pos = cset->mg_tasks.next;
4268 it->tasks_head = &cset->tasks;
4269 it->mg_tasks_head = &cset->mg_tasks;
4272 * We don't keep css_sets locked across iteration steps and thus
4273 * need to take steps to ensure that iteration can be resumed after
4274 * the lock is re-acquired. Iteration is performed at two levels -
4275 * css_sets and tasks in them.
4277 * Once created, a css_set never leaves its cgroup lists, so a
4278 * pinned css_set is guaranteed to stay put and we can resume
4279 * iteration afterwards.
4281 * Tasks may leave @cset across iteration steps. This is resolved
4282 * by registering each iterator with the css_set currently being
4283 * walked and making css_set_move_task() advance iterators whose
4284 * next task is leaving.
4287 list_del(&it->iters_node);
4288 put_css_set_locked(it->cur_cset);
4291 it->cur_cset = cset;
4292 list_add(&it->iters_node, &cset->task_iters);
4295 static void css_task_iter_advance(struct css_task_iter *it)
4297 struct list_head *l = it->task_pos;
4299 lockdep_assert_held(&css_set_lock);
4303 * Advance iterator to find next entry. cset->tasks is consumed
4304 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4309 if (l == it->tasks_head)
4310 l = it->mg_tasks_head->next;
4312 if (l == it->mg_tasks_head)
4313 css_task_iter_advance_css_set(it);
4319 * css_task_iter_start - initiate task iteration
4320 * @css: the css to walk tasks of
4321 * @it: the task iterator to use
4323 * Initiate iteration through the tasks of @css. The caller can call
4324 * css_task_iter_next() to walk through the tasks until the function
4325 * returns NULL. On completion of iteration, css_task_iter_end() must be
4328 void css_task_iter_start(struct cgroup_subsys_state *css,
4329 struct css_task_iter *it)
4331 /* no one should try to iterate before mounting cgroups */
4332 WARN_ON_ONCE(!use_task_css_set_links);
4334 memset(it, 0, sizeof(*it));
4336 spin_lock_irq(&css_set_lock);
4341 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4343 it->cset_pos = &css->cgroup->cset_links;
4345 it->cset_head = it->cset_pos;
4347 css_task_iter_advance_css_set(it);
4349 spin_unlock_irq(&css_set_lock);
4353 * css_task_iter_next - return the next task for the iterator
4354 * @it: the task iterator being iterated
4356 * The "next" function for task iteration. @it should have been
4357 * initialized via css_task_iter_start(). Returns NULL when the iteration
4360 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4363 put_task_struct(it->cur_task);
4364 it->cur_task = NULL;
4367 spin_lock_irq(&css_set_lock);
4370 it->cur_task = list_entry(it->task_pos, struct task_struct,
4372 get_task_struct(it->cur_task);
4373 css_task_iter_advance(it);
4376 spin_unlock_irq(&css_set_lock);
4378 return it->cur_task;
4382 * css_task_iter_end - finish task iteration
4383 * @it: the task iterator to finish
4385 * Finish task iteration started by css_task_iter_start().
4387 void css_task_iter_end(struct css_task_iter *it)
4390 spin_lock_irq(&css_set_lock);
4391 list_del(&it->iters_node);
4392 put_css_set_locked(it->cur_cset);
4393 spin_unlock_irq(&css_set_lock);
4397 put_task_struct(it->cur_task);
4401 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4402 * @to: cgroup to which the tasks will be moved
4403 * @from: cgroup in which the tasks currently reside
4405 * Locking rules between cgroup_post_fork() and the migration path
4406 * guarantee that, if a task is forking while being migrated, the new child
4407 * is guaranteed to be either visible in the source cgroup after the
4408 * parent's migration is complete or put into the target cgroup. No task
4409 * can slip out of migration through forking.
4411 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4413 LIST_HEAD(preloaded_csets);
4414 struct cgrp_cset_link *link;
4415 struct css_task_iter it;
4416 struct task_struct *task;
4419 if (!cgroup_may_migrate_to(to))
4422 mutex_lock(&cgroup_mutex);
4424 percpu_down_write(&cgroup_threadgroup_rwsem);
4426 /* all tasks in @from are being moved, all csets are source */
4427 spin_lock_irq(&css_set_lock);
4428 list_for_each_entry(link, &from->cset_links, cset_link)
4429 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4430 spin_unlock_irq(&css_set_lock);
4432 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4437 * Migrate tasks one-by-one until @from is empty. This fails iff
4438 * ->can_attach() fails.
4441 css_task_iter_start(&from->self, &it);
4444 task = css_task_iter_next(&it);
4445 } while (task && (task->flags & PF_EXITING));
4448 get_task_struct(task);
4449 css_task_iter_end(&it);
4452 ret = cgroup_migrate(task, false, to->root);
4454 trace_cgroup_transfer_tasks(to, task, false);
4455 put_task_struct(task);
4457 } while (task && !ret);
4459 cgroup_migrate_finish(&preloaded_csets);
4460 percpu_up_write(&cgroup_threadgroup_rwsem);
4461 mutex_unlock(&cgroup_mutex);
4466 * Stuff for reading the 'tasks'/'procs' files.
4468 * Reading this file can return large amounts of data if a cgroup has
4469 * *lots* of attached tasks. So it may need several calls to read(),
4470 * but we cannot guarantee that the information we produce is correct
4471 * unless we produce it entirely atomically.
4475 /* which pidlist file are we talking about? */
4476 enum cgroup_filetype {
4482 * A pidlist is a list of pids that virtually represents the contents of one
4483 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4484 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4487 struct cgroup_pidlist {
4489 * used to find which pidlist is wanted. doesn't change as long as
4490 * this particular list stays in the list.
4492 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4495 /* how many elements the above list has */
4497 /* each of these stored in a list by its cgroup */
4498 struct list_head links;
4499 /* pointer to the cgroup we belong to, for list removal purposes */
4500 struct cgroup *owner;
4501 /* for delayed destruction */
4502 struct delayed_work destroy_dwork;
4506 * The following two functions "fix" the issue where there are more pids
4507 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4508 * TODO: replace with a kernel-wide solution to this problem
4510 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4511 static void *pidlist_allocate(int count)
4513 if (PIDLIST_TOO_LARGE(count))
4514 return vmalloc(count * sizeof(pid_t));
4516 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4519 static void pidlist_free(void *p)
4525 * Used to destroy all pidlists lingering waiting for destroy timer. None
4526 * should be left afterwards.
4528 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4530 struct cgroup_pidlist *l, *tmp_l;
4532 mutex_lock(&cgrp->pidlist_mutex);
4533 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4534 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4535 mutex_unlock(&cgrp->pidlist_mutex);
4537 flush_workqueue(cgroup_pidlist_destroy_wq);
4538 BUG_ON(!list_empty(&cgrp->pidlists));
4541 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4543 struct delayed_work *dwork = to_delayed_work(work);
4544 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4546 struct cgroup_pidlist *tofree = NULL;
4548 mutex_lock(&l->owner->pidlist_mutex);
4551 * Destroy iff we didn't get queued again. The state won't change
4552 * as destroy_dwork can only be queued while locked.
4554 if (!delayed_work_pending(dwork)) {
4555 list_del(&l->links);
4556 pidlist_free(l->list);
4557 put_pid_ns(l->key.ns);
4561 mutex_unlock(&l->owner->pidlist_mutex);
4566 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4567 * Returns the number of unique elements.
4569 static int pidlist_uniq(pid_t *list, int length)
4574 * we presume the 0th element is unique, so i starts at 1. trivial
4575 * edge cases first; no work needs to be done for either
4577 if (length == 0 || length == 1)
4579 /* src and dest walk down the list; dest counts unique elements */
4580 for (src = 1; src < length; src++) {
4581 /* find next unique element */
4582 while (list[src] == list[src-1]) {
4587 /* dest always points to where the next unique element goes */
4588 list[dest] = list[src];
4596 * The two pid files - task and cgroup.procs - guaranteed that the result
4597 * is sorted, which forced this whole pidlist fiasco. As pid order is
4598 * different per namespace, each namespace needs differently sorted list,
4599 * making it impossible to use, for example, single rbtree of member tasks
4600 * sorted by task pointer. As pidlists can be fairly large, allocating one
4601 * per open file is dangerous, so cgroup had to implement shared pool of
4602 * pidlists keyed by cgroup and namespace.
4604 * All this extra complexity was caused by the original implementation
4605 * committing to an entirely unnecessary property. In the long term, we
4606 * want to do away with it. Explicitly scramble sort order if on the
4607 * default hierarchy so that no such expectation exists in the new
4610 * Scrambling is done by swapping every two consecutive bits, which is
4611 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4613 static pid_t pid_fry(pid_t pid)
4615 unsigned a = pid & 0x55555555;
4616 unsigned b = pid & 0xAAAAAAAA;
4618 return (a << 1) | (b >> 1);
4621 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4623 if (cgroup_on_dfl(cgrp))
4624 return pid_fry(pid);
4629 static int cmppid(const void *a, const void *b)
4631 return *(pid_t *)a - *(pid_t *)b;
4634 static int fried_cmppid(const void *a, const void *b)
4636 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4639 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4640 enum cgroup_filetype type)
4642 struct cgroup_pidlist *l;
4643 /* don't need task_nsproxy() if we're looking at ourself */
4644 struct pid_namespace *ns = task_active_pid_ns(current);
4646 lockdep_assert_held(&cgrp->pidlist_mutex);
4648 list_for_each_entry(l, &cgrp->pidlists, links)
4649 if (l->key.type == type && l->key.ns == ns)
4655 * find the appropriate pidlist for our purpose (given procs vs tasks)
4656 * returns with the lock on that pidlist already held, and takes care
4657 * of the use count, or returns NULL with no locks held if we're out of
4660 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4661 enum cgroup_filetype type)
4663 struct cgroup_pidlist *l;
4665 lockdep_assert_held(&cgrp->pidlist_mutex);
4667 l = cgroup_pidlist_find(cgrp, type);
4671 /* entry not found; create a new one */
4672 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4676 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4678 /* don't need task_nsproxy() if we're looking at ourself */
4679 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4681 list_add(&l->links, &cgrp->pidlists);
4686 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4688 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4689 struct cgroup_pidlist **lp)
4693 int pid, n = 0; /* used for populating the array */
4694 struct css_task_iter it;
4695 struct task_struct *tsk;
4696 struct cgroup_pidlist *l;
4698 lockdep_assert_held(&cgrp->pidlist_mutex);
4701 * If cgroup gets more users after we read count, we won't have
4702 * enough space - tough. This race is indistinguishable to the
4703 * caller from the case that the additional cgroup users didn't
4704 * show up until sometime later on.
4706 length = cgroup_task_count(cgrp);
4707 array = pidlist_allocate(length);
4710 /* now, populate the array */
4711 css_task_iter_start(&cgrp->self, &it);
4712 while ((tsk = css_task_iter_next(&it))) {
4713 if (unlikely(n == length))
4715 /* get tgid or pid for procs or tasks file respectively */
4716 if (type == CGROUP_FILE_PROCS)
4717 pid = task_tgid_vnr(tsk);
4719 pid = task_pid_vnr(tsk);
4720 if (pid > 0) /* make sure to only use valid results */
4723 css_task_iter_end(&it);
4725 /* now sort & (if procs) strip out duplicates */
4726 if (cgroup_on_dfl(cgrp))
4727 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4729 sort(array, length, sizeof(pid_t), cmppid, NULL);
4730 if (type == CGROUP_FILE_PROCS)
4731 length = pidlist_uniq(array, length);
4733 l = cgroup_pidlist_find_create(cgrp, type);
4735 pidlist_free(array);
4739 /* store array, freeing old if necessary */
4740 pidlist_free(l->list);
4748 * cgroupstats_build - build and fill cgroupstats
4749 * @stats: cgroupstats to fill information into
4750 * @dentry: A dentry entry belonging to the cgroup for which stats have
4753 * Build and fill cgroupstats so that taskstats can export it to user
4756 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4758 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4759 struct cgroup *cgrp;
4760 struct css_task_iter it;
4761 struct task_struct *tsk;
4763 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4764 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4765 kernfs_type(kn) != KERNFS_DIR)
4768 mutex_lock(&cgroup_mutex);
4771 * We aren't being called from kernfs and there's no guarantee on
4772 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4773 * @kn->priv is RCU safe. Let's do the RCU dancing.
4776 cgrp = rcu_dereference(kn->priv);
4777 if (!cgrp || cgroup_is_dead(cgrp)) {
4779 mutex_unlock(&cgroup_mutex);
4784 css_task_iter_start(&cgrp->self, &it);
4785 while ((tsk = css_task_iter_next(&it))) {
4786 switch (tsk->state) {
4788 stats->nr_running++;
4790 case TASK_INTERRUPTIBLE:
4791 stats->nr_sleeping++;
4793 case TASK_UNINTERRUPTIBLE:
4794 stats->nr_uninterruptible++;
4797 stats->nr_stopped++;
4800 if (delayacct_is_task_waiting_on_io(tsk))
4801 stats->nr_io_wait++;
4805 css_task_iter_end(&it);
4807 mutex_unlock(&cgroup_mutex);
4813 * seq_file methods for the tasks/procs files. The seq_file position is the
4814 * next pid to display; the seq_file iterator is a pointer to the pid
4815 * in the cgroup->l->list array.
4818 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4821 * Initially we receive a position value that corresponds to
4822 * one more than the last pid shown (or 0 on the first call or
4823 * after a seek to the start). Use a binary-search to find the
4824 * next pid to display, if any
4826 struct kernfs_open_file *of = s->private;
4827 struct cgroup *cgrp = seq_css(s)->cgroup;
4828 struct cgroup_pidlist *l;
4829 enum cgroup_filetype type = seq_cft(s)->private;
4830 int index = 0, pid = *pos;
4833 mutex_lock(&cgrp->pidlist_mutex);
4836 * !NULL @of->priv indicates that this isn't the first start()
4837 * after open. If the matching pidlist is around, we can use that.
4838 * Look for it. Note that @of->priv can't be used directly. It
4839 * could already have been destroyed.
4842 of->priv = cgroup_pidlist_find(cgrp, type);
4845 * Either this is the first start() after open or the matching
4846 * pidlist has been destroyed inbetween. Create a new one.
4849 ret = pidlist_array_load(cgrp, type,
4850 (struct cgroup_pidlist **)&of->priv);
4852 return ERR_PTR(ret);
4857 int end = l->length;
4859 while (index < end) {
4860 int mid = (index + end) / 2;
4861 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4864 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4870 /* If we're off the end of the array, we're done */
4871 if (index >= l->length)
4873 /* Update the abstract position to be the actual pid that we found */
4874 iter = l->list + index;
4875 *pos = cgroup_pid_fry(cgrp, *iter);
4879 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4881 struct kernfs_open_file *of = s->private;
4882 struct cgroup_pidlist *l = of->priv;
4885 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4886 CGROUP_PIDLIST_DESTROY_DELAY);
4887 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4890 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4892 struct kernfs_open_file *of = s->private;
4893 struct cgroup_pidlist *l = of->priv;
4895 pid_t *end = l->list + l->length;
4897 * Advance to the next pid in the array. If this goes off the
4904 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4909 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4911 seq_printf(s, "%d\n", *(int *)v);
4916 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4919 return notify_on_release(css->cgroup);
4922 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4923 struct cftype *cft, u64 val)
4926 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4928 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4932 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4935 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4938 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4939 struct cftype *cft, u64 val)
4942 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4944 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4948 /* cgroup core interface files for the default hierarchy */
4949 static struct cftype cgroup_dfl_base_files[] = {
4951 .name = "cgroup.procs",
4952 .file_offset = offsetof(struct cgroup, procs_file),
4953 .seq_start = cgroup_pidlist_start,
4954 .seq_next = cgroup_pidlist_next,
4955 .seq_stop = cgroup_pidlist_stop,
4956 .seq_show = cgroup_pidlist_show,
4957 .private = CGROUP_FILE_PROCS,
4958 .write = cgroup_procs_write,
4961 .name = "cgroup.controllers",
4962 .seq_show = cgroup_controllers_show,
4965 .name = "cgroup.subtree_control",
4966 .seq_show = cgroup_subtree_control_show,
4967 .write = cgroup_subtree_control_write,
4970 .name = "cgroup.events",
4971 .flags = CFTYPE_NOT_ON_ROOT,
4972 .file_offset = offsetof(struct cgroup, events_file),
4973 .seq_show = cgroup_events_show,
4978 /* cgroup core interface files for the legacy hierarchies */
4979 static struct cftype cgroup_legacy_base_files[] = {
4981 .name = "cgroup.procs",
4982 .seq_start = cgroup_pidlist_start,
4983 .seq_next = cgroup_pidlist_next,
4984 .seq_stop = cgroup_pidlist_stop,
4985 .seq_show = cgroup_pidlist_show,
4986 .private = CGROUP_FILE_PROCS,
4987 .write = cgroup_procs_write,
4990 .name = "cgroup.clone_children",
4991 .read_u64 = cgroup_clone_children_read,
4992 .write_u64 = cgroup_clone_children_write,
4995 .name = "cgroup.sane_behavior",
4996 .flags = CFTYPE_ONLY_ON_ROOT,
4997 .seq_show = cgroup_sane_behavior_show,
5001 .seq_start = cgroup_pidlist_start,
5002 .seq_next = cgroup_pidlist_next,
5003 .seq_stop = cgroup_pidlist_stop,
5004 .seq_show = cgroup_pidlist_show,
5005 .private = CGROUP_FILE_TASKS,
5006 .write = cgroup_tasks_write,
5009 .name = "notify_on_release",
5010 .read_u64 = cgroup_read_notify_on_release,
5011 .write_u64 = cgroup_write_notify_on_release,
5014 .name = "release_agent",
5015 .flags = CFTYPE_ONLY_ON_ROOT,
5016 .seq_show = cgroup_release_agent_show,
5017 .write = cgroup_release_agent_write,
5018 .max_write_len = PATH_MAX - 1,
5024 * css destruction is four-stage process.
5026 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5027 * Implemented in kill_css().
5029 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5030 * and thus css_tryget_online() is guaranteed to fail, the css can be
5031 * offlined by invoking offline_css(). After offlining, the base ref is
5032 * put. Implemented in css_killed_work_fn().
5034 * 3. When the percpu_ref reaches zero, the only possible remaining
5035 * accessors are inside RCU read sections. css_release() schedules the
5038 * 4. After the grace period, the css can be freed. Implemented in
5039 * css_free_work_fn().
5041 * It is actually hairier because both step 2 and 4 require process context
5042 * and thus involve punting to css->destroy_work adding two additional
5043 * steps to the already complex sequence.
5045 static void css_free_work_fn(struct work_struct *work)
5047 struct cgroup_subsys_state *css =
5048 container_of(work, struct cgroup_subsys_state, destroy_work);
5049 struct cgroup_subsys *ss = css->ss;
5050 struct cgroup *cgrp = css->cgroup;
5052 percpu_ref_exit(&css->refcnt);
5056 struct cgroup_subsys_state *parent = css->parent;
5060 cgroup_idr_remove(&ss->css_idr, id);
5066 /* cgroup free path */
5067 atomic_dec(&cgrp->root->nr_cgrps);
5068 cgroup_pidlist_destroy_all(cgrp);
5069 cancel_work_sync(&cgrp->release_agent_work);
5071 if (cgroup_parent(cgrp)) {
5073 * We get a ref to the parent, and put the ref when
5074 * this cgroup is being freed, so it's guaranteed
5075 * that the parent won't be destroyed before its
5078 cgroup_put(cgroup_parent(cgrp));
5079 kernfs_put(cgrp->kn);
5083 * This is root cgroup's refcnt reaching zero,
5084 * which indicates that the root should be
5087 cgroup_destroy_root(cgrp->root);
5092 static void css_free_rcu_fn(struct rcu_head *rcu_head)
5094 struct cgroup_subsys_state *css =
5095 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
5097 INIT_WORK(&css->destroy_work, css_free_work_fn);
5098 queue_work(cgroup_destroy_wq, &css->destroy_work);
5101 static void css_release_work_fn(struct work_struct *work)
5103 struct cgroup_subsys_state *css =
5104 container_of(work, struct cgroup_subsys_state, destroy_work);
5105 struct cgroup_subsys *ss = css->ss;
5106 struct cgroup *cgrp = css->cgroup;
5108 mutex_lock(&cgroup_mutex);
5110 css->flags |= CSS_RELEASED;
5111 list_del_rcu(&css->sibling);
5114 /* css release path */
5115 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5116 if (ss->css_released)
5117 ss->css_released(css);
5119 /* cgroup release path */
5120 trace_cgroup_release(cgrp);
5122 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5126 * There are two control paths which try to determine
5127 * cgroup from dentry without going through kernfs -
5128 * cgroupstats_build() and css_tryget_online_from_dir().
5129 * Those are supported by RCU protecting clearing of
5130 * cgrp->kn->priv backpointer.
5133 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5137 mutex_unlock(&cgroup_mutex);
5139 call_rcu(&css->rcu_head, css_free_rcu_fn);
5142 static void css_release(struct percpu_ref *ref)
5144 struct cgroup_subsys_state *css =
5145 container_of(ref, struct cgroup_subsys_state, refcnt);
5147 INIT_WORK(&css->destroy_work, css_release_work_fn);
5148 queue_work(cgroup_destroy_wq, &css->destroy_work);
5151 static void init_and_link_css(struct cgroup_subsys_state *css,
5152 struct cgroup_subsys *ss, struct cgroup *cgrp)
5154 lockdep_assert_held(&cgroup_mutex);
5158 memset(css, 0, sizeof(*css));
5162 INIT_LIST_HEAD(&css->sibling);
5163 INIT_LIST_HEAD(&css->children);
5164 css->serial_nr = css_serial_nr_next++;
5165 atomic_set(&css->online_cnt, 0);
5167 if (cgroup_parent(cgrp)) {
5168 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5169 css_get(css->parent);
5172 BUG_ON(cgroup_css(cgrp, ss));
5175 /* invoke ->css_online() on a new CSS and mark it online if successful */
5176 static int online_css(struct cgroup_subsys_state *css)
5178 struct cgroup_subsys *ss = css->ss;
5181 lockdep_assert_held(&cgroup_mutex);
5184 ret = ss->css_online(css);
5186 css->flags |= CSS_ONLINE;
5187 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5189 atomic_inc(&css->online_cnt);
5191 atomic_inc(&css->parent->online_cnt);
5196 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5197 static void offline_css(struct cgroup_subsys_state *css)
5199 struct cgroup_subsys *ss = css->ss;
5201 lockdep_assert_held(&cgroup_mutex);
5203 if (!(css->flags & CSS_ONLINE))
5209 if (ss->css_offline)
5210 ss->css_offline(css);
5212 css->flags &= ~CSS_ONLINE;
5213 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5215 wake_up_all(&css->cgroup->offline_waitq);
5219 * css_create - create a cgroup_subsys_state
5220 * @cgrp: the cgroup new css will be associated with
5221 * @ss: the subsys of new css
5223 * Create a new css associated with @cgrp - @ss pair. On success, the new
5224 * css is online and installed in @cgrp. This function doesn't create the
5225 * interface files. Returns 0 on success, -errno on failure.
5227 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5228 struct cgroup_subsys *ss)
5230 struct cgroup *parent = cgroup_parent(cgrp);
5231 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5232 struct cgroup_subsys_state *css;
5235 lockdep_assert_held(&cgroup_mutex);
5237 css = ss->css_alloc(parent_css);
5239 css = ERR_PTR(-ENOMEM);
5243 init_and_link_css(css, ss, cgrp);
5245 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5249 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5254 /* @css is ready to be brought online now, make it visible */
5255 list_add_tail_rcu(&css->sibling, &parent_css->children);
5256 cgroup_idr_replace(&ss->css_idr, css, css->id);
5258 err = online_css(css);
5262 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5263 cgroup_parent(parent)) {
5264 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5265 current->comm, current->pid, ss->name);
5266 if (!strcmp(ss->name, "memory"))
5267 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5268 ss->warned_broken_hierarchy = true;
5274 list_del_rcu(&css->sibling);
5276 call_rcu(&css->rcu_head, css_free_rcu_fn);
5277 return ERR_PTR(err);
5281 * The returned cgroup is fully initialized including its control mask, but
5282 * it isn't associated with its kernfs_node and doesn't have the control
5285 static struct cgroup *cgroup_create(struct cgroup *parent)
5287 struct cgroup_root *root = parent->root;
5288 struct cgroup *cgrp, *tcgrp;
5289 int level = parent->level + 1;
5292 /* allocate the cgroup and its ID, 0 is reserved for the root */
5293 cgrp = kzalloc(sizeof(*cgrp) +
5294 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5296 return ERR_PTR(-ENOMEM);
5298 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5303 * Temporarily set the pointer to NULL, so idr_find() won't return
5304 * a half-baked cgroup.
5306 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5309 goto out_cancel_ref;
5312 init_cgroup_housekeeping(cgrp);
5314 cgrp->self.parent = &parent->self;
5316 cgrp->level = level;
5318 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5319 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5321 if (notify_on_release(parent))
5322 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5324 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5325 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5327 cgrp->self.serial_nr = css_serial_nr_next++;
5329 /* allocation complete, commit to creation */
5330 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5331 atomic_inc(&root->nr_cgrps);
5335 * @cgrp is now fully operational. If something fails after this
5336 * point, it'll be released via the normal destruction path.
5338 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5341 * On the default hierarchy, a child doesn't automatically inherit
5342 * subtree_control from the parent. Each is configured manually.
5344 if (!cgroup_on_dfl(cgrp))
5345 cgrp->subtree_control = cgroup_control(cgrp);
5347 cgroup_propagate_control(cgrp);
5352 percpu_ref_exit(&cgrp->self.refcnt);
5355 return ERR_PTR(ret);
5358 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5361 struct cgroup *parent, *cgrp;
5362 struct kernfs_node *kn;
5365 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5366 if (strchr(name, '\n'))
5369 parent = cgroup_kn_lock_live(parent_kn, false);
5373 cgrp = cgroup_create(parent);
5375 ret = PTR_ERR(cgrp);
5379 /* create the directory */
5380 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5388 * This extra ref will be put in cgroup_free_fn() and guarantees
5389 * that @cgrp->kn is always accessible.
5393 ret = cgroup_kn_set_ugid(kn);
5397 ret = css_populate_dir(&cgrp->self);
5401 ret = cgroup_apply_control_enable(cgrp);
5405 trace_cgroup_mkdir(cgrp);
5407 /* let's create and online css's */
5408 kernfs_activate(kn);
5414 cgroup_destroy_locked(cgrp);
5416 cgroup_kn_unlock(parent_kn);
5421 * This is called when the refcnt of a css is confirmed to be killed.
5422 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5423 * initate destruction and put the css ref from kill_css().
5425 static void css_killed_work_fn(struct work_struct *work)
5427 struct cgroup_subsys_state *css =
5428 container_of(work, struct cgroup_subsys_state, destroy_work);
5430 mutex_lock(&cgroup_mutex);
5435 /* @css can't go away while we're holding cgroup_mutex */
5437 } while (css && atomic_dec_and_test(&css->online_cnt));
5439 mutex_unlock(&cgroup_mutex);
5442 /* css kill confirmation processing requires process context, bounce */
5443 static void css_killed_ref_fn(struct percpu_ref *ref)
5445 struct cgroup_subsys_state *css =
5446 container_of(ref, struct cgroup_subsys_state, refcnt);
5448 if (atomic_dec_and_test(&css->online_cnt)) {
5449 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5450 queue_work(cgroup_destroy_wq, &css->destroy_work);
5455 * kill_css - destroy a css
5456 * @css: css to destroy
5458 * This function initiates destruction of @css by removing cgroup interface
5459 * files and putting its base reference. ->css_offline() will be invoked
5460 * asynchronously once css_tryget_online() is guaranteed to fail and when
5461 * the reference count reaches zero, @css will be released.
5463 static void kill_css(struct cgroup_subsys_state *css)
5465 lockdep_assert_held(&cgroup_mutex);
5467 if (css->flags & CSS_DYING)
5470 css->flags |= CSS_DYING;
5473 * This must happen before css is disassociated with its cgroup.
5474 * See seq_css() for details.
5479 * Killing would put the base ref, but we need to keep it alive
5480 * until after ->css_offline().
5485 * cgroup core guarantees that, by the time ->css_offline() is
5486 * invoked, no new css reference will be given out via
5487 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5488 * proceed to offlining css's because percpu_ref_kill() doesn't
5489 * guarantee that the ref is seen as killed on all CPUs on return.
5491 * Use percpu_ref_kill_and_confirm() to get notifications as each
5492 * css is confirmed to be seen as killed on all CPUs.
5494 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5498 * cgroup_destroy_locked - the first stage of cgroup destruction
5499 * @cgrp: cgroup to be destroyed
5501 * css's make use of percpu refcnts whose killing latency shouldn't be
5502 * exposed to userland and are RCU protected. Also, cgroup core needs to
5503 * guarantee that css_tryget_online() won't succeed by the time
5504 * ->css_offline() is invoked. To satisfy all the requirements,
5505 * destruction is implemented in the following two steps.
5507 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5508 * userland visible parts and start killing the percpu refcnts of
5509 * css's. Set up so that the next stage will be kicked off once all
5510 * the percpu refcnts are confirmed to be killed.
5512 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5513 * rest of destruction. Once all cgroup references are gone, the
5514 * cgroup is RCU-freed.
5516 * This function implements s1. After this step, @cgrp is gone as far as
5517 * the userland is concerned and a new cgroup with the same name may be
5518 * created. As cgroup doesn't care about the names internally, this
5519 * doesn't cause any problem.
5521 static int cgroup_destroy_locked(struct cgroup *cgrp)
5522 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5524 struct cgroup_subsys_state *css;
5525 struct cgrp_cset_link *link;
5528 lockdep_assert_held(&cgroup_mutex);
5531 * Only migration can raise populated from zero and we're already
5532 * holding cgroup_mutex.
5534 if (cgroup_is_populated(cgrp))
5538 * Make sure there's no live children. We can't test emptiness of
5539 * ->self.children as dead children linger on it while being
5540 * drained; otherwise, "rmdir parent/child parent" may fail.
5542 if (css_has_online_children(&cgrp->self))
5546 * Mark @cgrp and the associated csets dead. The former prevents
5547 * further task migration and child creation by disabling
5548 * cgroup_lock_live_group(). The latter makes the csets ignored by
5549 * the migration path.
5551 cgrp->self.flags &= ~CSS_ONLINE;
5553 spin_lock_irq(&css_set_lock);
5554 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5555 link->cset->dead = true;
5556 spin_unlock_irq(&css_set_lock);
5558 /* initiate massacre of all css's */
5559 for_each_css(css, ssid, cgrp)
5563 * Remove @cgrp directory along with the base files. @cgrp has an
5564 * extra ref on its kn.
5566 kernfs_remove(cgrp->kn);
5568 check_for_release(cgroup_parent(cgrp));
5570 /* put the base reference */
5571 percpu_ref_kill(&cgrp->self.refcnt);
5576 static int cgroup_rmdir(struct kernfs_node *kn)
5578 struct cgroup *cgrp;
5581 cgrp = cgroup_kn_lock_live(kn, false);
5585 ret = cgroup_destroy_locked(cgrp);
5588 trace_cgroup_rmdir(cgrp);
5590 cgroup_kn_unlock(kn);
5594 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5595 .remount_fs = cgroup_remount,
5596 .show_options = cgroup_show_options,
5597 .mkdir = cgroup_mkdir,
5598 .rmdir = cgroup_rmdir,
5599 .rename = cgroup_rename,
5600 .show_path = cgroup_show_path,
5603 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5605 struct cgroup_subsys_state *css;
5607 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5609 mutex_lock(&cgroup_mutex);
5611 idr_init(&ss->css_idr);
5612 INIT_LIST_HEAD(&ss->cfts);
5614 /* Create the root cgroup state for this subsystem */
5615 ss->root = &cgrp_dfl_root;
5616 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5617 /* We don't handle early failures gracefully */
5618 BUG_ON(IS_ERR(css));
5619 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5622 * Root csses are never destroyed and we can't initialize
5623 * percpu_ref during early init. Disable refcnting.
5625 css->flags |= CSS_NO_REF;
5628 /* allocation can't be done safely during early init */
5631 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5632 BUG_ON(css->id < 0);
5635 /* Update the init_css_set to contain a subsys
5636 * pointer to this state - since the subsystem is
5637 * newly registered, all tasks and hence the
5638 * init_css_set is in the subsystem's root cgroup. */
5639 init_css_set.subsys[ss->id] = css;
5641 have_fork_callback |= (bool)ss->fork << ss->id;
5642 have_exit_callback |= (bool)ss->exit << ss->id;
5643 have_free_callback |= (bool)ss->free << ss->id;
5644 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5646 /* At system boot, before all subsystems have been
5647 * registered, no tasks have been forked, so we don't
5648 * need to invoke fork callbacks here. */
5649 BUG_ON(!list_empty(&init_task.tasks));
5651 BUG_ON(online_css(css));
5653 mutex_unlock(&cgroup_mutex);
5657 * cgroup_init_early - cgroup initialization at system boot
5659 * Initialize cgroups at system boot, and initialize any
5660 * subsystems that request early init.
5662 int __init cgroup_init_early(void)
5664 static struct cgroup_sb_opts __initdata opts;
5665 struct cgroup_subsys *ss;
5668 init_cgroup_root(&cgrp_dfl_root, &opts);
5669 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5671 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5673 for_each_subsys(ss, i) {
5674 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5675 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5676 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5678 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5679 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5682 ss->name = cgroup_subsys_name[i];
5683 if (!ss->legacy_name)
5684 ss->legacy_name = cgroup_subsys_name[i];
5687 cgroup_init_subsys(ss, true);
5693 * cgroup_init - cgroup initialization
5695 * Register cgroup filesystem and /proc file, and initialize
5696 * any subsystems that didn't request early init.
5698 int __init cgroup_init(void)
5700 struct cgroup_subsys *ss;
5703 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5704 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5705 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5706 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5709 * The latency of the synchronize_sched() is too high for cgroups,
5710 * avoid it at the cost of forcing all readers into the slow path.
5712 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5714 get_user_ns(init_cgroup_ns.user_ns);
5716 mutex_lock(&cgroup_mutex);
5719 * Add init_css_set to the hash table so that dfl_root can link to
5722 hash_add(css_set_table, &init_css_set.hlist,
5723 css_set_hash(init_css_set.subsys));
5725 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5727 mutex_unlock(&cgroup_mutex);
5729 for_each_subsys(ss, ssid) {
5730 if (ss->early_init) {
5731 struct cgroup_subsys_state *css =
5732 init_css_set.subsys[ss->id];
5734 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5736 BUG_ON(css->id < 0);
5738 cgroup_init_subsys(ss, false);
5741 list_add_tail(&init_css_set.e_cset_node[ssid],
5742 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5745 * Setting dfl_root subsys_mask needs to consider the
5746 * disabled flag and cftype registration needs kmalloc,
5747 * both of which aren't available during early_init.
5749 if (!cgroup_ssid_enabled(ssid))
5752 if (cgroup_ssid_no_v1(ssid))
5753 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5756 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5758 if (ss->implicit_on_dfl)
5759 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5760 else if (!ss->dfl_cftypes)
5761 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5763 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5764 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5766 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5767 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5771 ss->bind(init_css_set.subsys[ssid]);
5773 mutex_lock(&cgroup_mutex);
5774 css_populate_dir(init_css_set.subsys[ssid]);
5775 mutex_unlock(&cgroup_mutex);
5778 /* init_css_set.subsys[] has been updated, re-hash */
5779 hash_del(&init_css_set.hlist);
5780 hash_add(css_set_table, &init_css_set.hlist,
5781 css_set_hash(init_css_set.subsys));
5783 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5784 WARN_ON(register_filesystem(&cgroup_fs_type));
5785 WARN_ON(register_filesystem(&cgroup2_fs_type));
5786 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5791 static int __init cgroup_wq_init(void)
5794 * There isn't much point in executing destruction path in
5795 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5796 * Use 1 for @max_active.
5798 * We would prefer to do this in cgroup_init() above, but that
5799 * is called before init_workqueues(): so leave this until after.
5801 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5802 BUG_ON(!cgroup_destroy_wq);
5805 * Used to destroy pidlists and separate to serve as flush domain.
5806 * Cap @max_active to 1 too.
5808 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5810 BUG_ON(!cgroup_pidlist_destroy_wq);
5814 core_initcall(cgroup_wq_init);
5817 * proc_cgroup_show()
5818 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5819 * - Used for /proc/<pid>/cgroup.
5821 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5822 struct pid *pid, struct task_struct *tsk)
5826 struct cgroup_root *root;
5829 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5833 mutex_lock(&cgroup_mutex);
5834 spin_lock_irq(&css_set_lock);
5836 for_each_root(root) {
5837 struct cgroup_subsys *ss;
5838 struct cgroup *cgrp;
5839 int ssid, count = 0;
5841 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5844 seq_printf(m, "%d:", root->hierarchy_id);
5845 if (root != &cgrp_dfl_root)
5846 for_each_subsys(ss, ssid)
5847 if (root->subsys_mask & (1 << ssid))
5848 seq_printf(m, "%s%s", count++ ? "," : "",
5850 if (strlen(root->name))
5851 seq_printf(m, "%sname=%s", count ? "," : "",
5855 cgrp = task_cgroup_from_root(tsk, root);
5858 * On traditional hierarchies, all zombie tasks show up as
5859 * belonging to the root cgroup. On the default hierarchy,
5860 * while a zombie doesn't show up in "cgroup.procs" and
5861 * thus can't be migrated, its /proc/PID/cgroup keeps
5862 * reporting the cgroup it belonged to before exiting. If
5863 * the cgroup is removed before the zombie is reaped,
5864 * " (deleted)" is appended to the cgroup path.
5866 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5867 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5868 current->nsproxy->cgroup_ns);
5869 if (retval >= PATH_MAX)
5870 retval = -ENAMETOOLONG;
5879 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5880 seq_puts(m, " (deleted)\n");
5887 spin_unlock_irq(&css_set_lock);
5888 mutex_unlock(&cgroup_mutex);
5894 /* Display information about each subsystem and each hierarchy */
5895 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5897 struct cgroup_subsys *ss;
5900 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5902 * ideally we don't want subsystems moving around while we do this.
5903 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5904 * subsys/hierarchy state.
5906 mutex_lock(&cgroup_mutex);
5908 for_each_subsys(ss, i)
5909 seq_printf(m, "%s\t%d\t%d\t%d\n",
5910 ss->legacy_name, ss->root->hierarchy_id,
5911 atomic_read(&ss->root->nr_cgrps),
5912 cgroup_ssid_enabled(i));
5914 mutex_unlock(&cgroup_mutex);
5918 static int cgroupstats_open(struct inode *inode, struct file *file)
5920 return single_open(file, proc_cgroupstats_show, NULL);
5923 static const struct file_operations proc_cgroupstats_operations = {
5924 .open = cgroupstats_open,
5926 .llseek = seq_lseek,
5927 .release = single_release,
5931 * cgroup_fork - initialize cgroup related fields during copy_process()
5932 * @child: pointer to task_struct of forking parent process.
5934 * A task is associated with the init_css_set until cgroup_post_fork()
5935 * attaches it to the parent's css_set. Empty cg_list indicates that
5936 * @child isn't holding reference to its css_set.
5938 void cgroup_fork(struct task_struct *child)
5940 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5941 INIT_LIST_HEAD(&child->cg_list);
5945 * cgroup_can_fork - called on a new task before the process is exposed
5946 * @child: the task in question.
5948 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5949 * returns an error, the fork aborts with that error code. This allows for
5950 * a cgroup subsystem to conditionally allow or deny new forks.
5952 int cgroup_can_fork(struct task_struct *child)
5954 struct cgroup_subsys *ss;
5957 do_each_subsys_mask(ss, i, have_canfork_callback) {
5958 ret = ss->can_fork(child);
5961 } while_each_subsys_mask();
5966 for_each_subsys(ss, j) {
5969 if (ss->cancel_fork)
5970 ss->cancel_fork(child);
5977 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5978 * @child: the task in question
5980 * This calls the cancel_fork() callbacks if a fork failed *after*
5981 * cgroup_can_fork() succeded.
5983 void cgroup_cancel_fork(struct task_struct *child)
5985 struct cgroup_subsys *ss;
5988 for_each_subsys(ss, i)
5989 if (ss->cancel_fork)
5990 ss->cancel_fork(child);
5994 * cgroup_post_fork - called on a new task after adding it to the task list
5995 * @child: the task in question
5997 * Adds the task to the list running through its css_set if necessary and
5998 * call the subsystem fork() callbacks. Has to be after the task is
5999 * visible on the task list in case we race with the first call to
6000 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
6003 void cgroup_post_fork(struct task_struct *child)
6005 struct cgroup_subsys *ss;
6009 * This may race against cgroup_enable_task_cg_lists(). As that
6010 * function sets use_task_css_set_links before grabbing
6011 * tasklist_lock and we just went through tasklist_lock to add
6012 * @child, it's guaranteed that either we see the set
6013 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
6014 * @child during its iteration.
6016 * If we won the race, @child is associated with %current's
6017 * css_set. Grabbing css_set_lock guarantees both that the
6018 * association is stable, and, on completion of the parent's
6019 * migration, @child is visible in the source of migration or
6020 * already in the destination cgroup. This guarantee is necessary
6021 * when implementing operations which need to migrate all tasks of
6022 * a cgroup to another.
6024 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
6025 * will remain in init_css_set. This is safe because all tasks are
6026 * in the init_css_set before cg_links is enabled and there's no
6027 * operation which transfers all tasks out of init_css_set.
6029 if (use_task_css_set_links) {
6030 struct css_set *cset;
6032 spin_lock_irq(&css_set_lock);
6033 cset = task_css_set(current);
6034 if (list_empty(&child->cg_list)) {
6036 css_set_move_task(child, NULL, cset, false);
6038 spin_unlock_irq(&css_set_lock);
6042 * Call ss->fork(). This must happen after @child is linked on
6043 * css_set; otherwise, @child might change state between ->fork()
6044 * and addition to css_set.
6046 do_each_subsys_mask(ss, i, have_fork_callback) {
6048 } while_each_subsys_mask();
6052 * cgroup_exit - detach cgroup from exiting task
6053 * @tsk: pointer to task_struct of exiting process
6055 * Description: Detach cgroup from @tsk and release it.
6057 * Note that cgroups marked notify_on_release force every task in
6058 * them to take the global cgroup_mutex mutex when exiting.
6059 * This could impact scaling on very large systems. Be reluctant to
6060 * use notify_on_release cgroups where very high task exit scaling
6061 * is required on large systems.
6063 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
6064 * call cgroup_exit() while the task is still competent to handle
6065 * notify_on_release(), then leave the task attached to the root cgroup in
6066 * each hierarchy for the remainder of its exit. No need to bother with
6067 * init_css_set refcnting. init_css_set never goes away and we can't race
6068 * with migration path - PF_EXITING is visible to migration path.
6070 void cgroup_exit(struct task_struct *tsk)
6072 struct cgroup_subsys *ss;
6073 struct css_set *cset;
6077 * Unlink from @tsk from its css_set. As migration path can't race
6078 * with us, we can check css_set and cg_list without synchronization.
6080 cset = task_css_set(tsk);
6082 if (!list_empty(&tsk->cg_list)) {
6083 spin_lock_irq(&css_set_lock);
6084 css_set_move_task(tsk, cset, NULL, false);
6085 spin_unlock_irq(&css_set_lock);
6090 /* see cgroup_post_fork() for details */
6091 do_each_subsys_mask(ss, i, have_exit_callback) {
6093 } while_each_subsys_mask();
6096 void cgroup_free(struct task_struct *task)
6098 struct css_set *cset = task_css_set(task);
6099 struct cgroup_subsys *ss;
6102 do_each_subsys_mask(ss, ssid, have_free_callback) {
6104 } while_each_subsys_mask();
6109 static void check_for_release(struct cgroup *cgrp)
6111 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
6112 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
6113 schedule_work(&cgrp->release_agent_work);
6117 * Notify userspace when a cgroup is released, by running the
6118 * configured release agent with the name of the cgroup (path
6119 * relative to the root of cgroup file system) as the argument.
6121 * Most likely, this user command will try to rmdir this cgroup.
6123 * This races with the possibility that some other task will be
6124 * attached to this cgroup before it is removed, or that some other
6125 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
6126 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
6127 * unused, and this cgroup will be reprieved from its death sentence,
6128 * to continue to serve a useful existence. Next time it's released,
6129 * we will get notified again, if it still has 'notify_on_release' set.
6131 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
6132 * means only wait until the task is successfully execve()'d. The
6133 * separate release agent task is forked by call_usermodehelper(),
6134 * then control in this thread returns here, without waiting for the
6135 * release agent task. We don't bother to wait because the caller of
6136 * this routine has no use for the exit status of the release agent
6137 * task, so no sense holding our caller up for that.
6139 static void cgroup_release_agent(struct work_struct *work)
6141 struct cgroup *cgrp =
6142 container_of(work, struct cgroup, release_agent_work);
6143 char *pathbuf = NULL, *agentbuf = NULL;
6144 char *argv[3], *envp[3];
6147 mutex_lock(&cgroup_mutex);
6149 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
6150 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
6151 if (!pathbuf || !agentbuf)
6154 spin_lock_irq(&css_set_lock);
6155 ret = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
6156 spin_unlock_irq(&css_set_lock);
6157 if (ret < 0 || ret >= PATH_MAX)
6164 /* minimal command environment */
6166 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
6169 mutex_unlock(&cgroup_mutex);
6170 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
6173 mutex_unlock(&cgroup_mutex);
6179 static int __init cgroup_disable(char *str)
6181 struct cgroup_subsys *ss;
6185 while ((token = strsep(&str, ",")) != NULL) {
6189 for_each_subsys(ss, i) {
6190 if (strcmp(token, ss->name) &&
6191 strcmp(token, ss->legacy_name))
6194 static_branch_disable(cgroup_subsys_enabled_key[i]);
6195 pr_info("Disabling %s control group subsystem\n",
6201 __setup("cgroup_disable=", cgroup_disable);
6203 static int __init cgroup_no_v1(char *str)
6205 struct cgroup_subsys *ss;
6209 while ((token = strsep(&str, ",")) != NULL) {
6213 if (!strcmp(token, "all")) {
6214 cgroup_no_v1_mask = U16_MAX;
6218 for_each_subsys(ss, i) {
6219 if (strcmp(token, ss->name) &&
6220 strcmp(token, ss->legacy_name))
6223 cgroup_no_v1_mask |= 1 << i;
6228 __setup("cgroup_no_v1=", cgroup_no_v1);
6231 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6232 * @dentry: directory dentry of interest
6233 * @ss: subsystem of interest
6235 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6236 * to get the corresponding css and return it. If such css doesn't exist
6237 * or can't be pinned, an ERR_PTR value is returned.
6239 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6240 struct cgroup_subsys *ss)
6242 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6243 struct file_system_type *s_type = dentry->d_sb->s_type;
6244 struct cgroup_subsys_state *css = NULL;
6245 struct cgroup *cgrp;
6247 /* is @dentry a cgroup dir? */
6248 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6249 !kn || kernfs_type(kn) != KERNFS_DIR)
6250 return ERR_PTR(-EBADF);
6255 * This path doesn't originate from kernfs and @kn could already
6256 * have been or be removed at any point. @kn->priv is RCU
6257 * protected for this access. See css_release_work_fn() for details.
6259 cgrp = rcu_dereference(kn->priv);
6261 css = cgroup_css(cgrp, ss);
6263 if (!css || !css_tryget_online(css))
6264 css = ERR_PTR(-ENOENT);
6271 * css_from_id - lookup css by id
6272 * @id: the cgroup id
6273 * @ss: cgroup subsys to be looked into
6275 * Returns the css if there's valid one with @id, otherwise returns NULL.
6276 * Should be called under rcu_read_lock().
6278 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6280 WARN_ON_ONCE(!rcu_read_lock_held());
6281 return idr_find(&ss->css_idr, id);
6285 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6286 * @path: path on the default hierarchy
6288 * Find the cgroup at @path on the default hierarchy, increment its
6289 * reference count and return it. Returns pointer to the found cgroup on
6290 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6291 * if @path points to a non-directory.
6293 struct cgroup *cgroup_get_from_path(const char *path)
6295 struct kernfs_node *kn;
6296 struct cgroup *cgrp;
6298 mutex_lock(&cgroup_mutex);
6300 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6302 if (kernfs_type(kn) == KERNFS_DIR) {
6306 cgrp = ERR_PTR(-ENOTDIR);
6310 cgrp = ERR_PTR(-ENOENT);
6313 mutex_unlock(&cgroup_mutex);
6316 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6319 * cgroup_get_from_fd - get a cgroup pointer from a fd
6320 * @fd: fd obtained by open(cgroup2_dir)
6322 * Find the cgroup from a fd which should be obtained
6323 * by opening a cgroup directory. Returns a pointer to the
6324 * cgroup on success. ERR_PTR is returned if the cgroup
6327 struct cgroup *cgroup_get_from_fd(int fd)
6329 struct cgroup_subsys_state *css;
6330 struct cgroup *cgrp;
6335 return ERR_PTR(-EBADF);
6337 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6340 return ERR_CAST(css);
6343 if (!cgroup_on_dfl(cgrp)) {
6345 return ERR_PTR(-EBADF);
6350 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6353 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6354 * definition in cgroup-defs.h.
6356 #ifdef CONFIG_SOCK_CGROUP_DATA
6358 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6360 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6361 static bool cgroup_sk_alloc_disabled __read_mostly;
6363 void cgroup_sk_alloc_disable(void)
6365 if (cgroup_sk_alloc_disabled)
6367 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6368 cgroup_sk_alloc_disabled = true;
6373 #define cgroup_sk_alloc_disabled false
6377 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6379 if (cgroup_sk_alloc_disabled) {
6380 skcd->no_refcnt = 1;
6384 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6391 struct css_set *cset;
6393 cset = task_css_set(current);
6394 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6395 skcd->val = (unsigned long)cset->dfl_cgrp;
6404 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6406 /* Socket clone path */
6408 if (skcd->no_refcnt)
6411 * We might be cloning a socket which is left in an empty
6412 * cgroup and the cgroup might have already been rmdir'd.
6413 * Don't use cgroup_get_live().
6415 cgroup_get(sock_cgroup_ptr(skcd));
6419 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6421 if (skcd->no_refcnt)
6424 cgroup_put(sock_cgroup_ptr(skcd));
6427 #endif /* CONFIG_SOCK_CGROUP_DATA */
6429 /* cgroup namespaces */
6431 static struct ucounts *inc_cgroup_namespaces(struct user_namespace *ns)
6433 return inc_ucount(ns, current_euid(), UCOUNT_CGROUP_NAMESPACES);
6436 static void dec_cgroup_namespaces(struct ucounts *ucounts)
6438 dec_ucount(ucounts, UCOUNT_CGROUP_NAMESPACES);
6441 static struct cgroup_namespace *alloc_cgroup_ns(void)
6443 struct cgroup_namespace *new_ns;
6446 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6448 return ERR_PTR(-ENOMEM);
6449 ret = ns_alloc_inum(&new_ns->ns);
6452 return ERR_PTR(ret);
6454 atomic_set(&new_ns->count, 1);
6455 new_ns->ns.ops = &cgroupns_operations;
6459 void free_cgroup_ns(struct cgroup_namespace *ns)
6461 put_css_set(ns->root_cset);
6462 dec_cgroup_namespaces(ns->ucounts);
6463 put_user_ns(ns->user_ns);
6464 ns_free_inum(&ns->ns);
6467 EXPORT_SYMBOL(free_cgroup_ns);
6469 struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6470 struct user_namespace *user_ns,
6471 struct cgroup_namespace *old_ns)
6473 struct cgroup_namespace *new_ns;
6474 struct ucounts *ucounts;
6475 struct css_set *cset;
6479 if (!(flags & CLONE_NEWCGROUP)) {
6480 get_cgroup_ns(old_ns);
6484 /* Allow only sysadmin to create cgroup namespace. */
6485 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6486 return ERR_PTR(-EPERM);
6488 ucounts = inc_cgroup_namespaces(user_ns);
6490 return ERR_PTR(-ENOSPC);
6492 /* It is not safe to take cgroup_mutex here */
6493 spin_lock_irq(&css_set_lock);
6494 cset = task_css_set(current);
6496 spin_unlock_irq(&css_set_lock);
6498 new_ns = alloc_cgroup_ns();
6499 if (IS_ERR(new_ns)) {
6501 dec_cgroup_namespaces(ucounts);
6505 new_ns->user_ns = get_user_ns(user_ns);
6506 new_ns->ucounts = ucounts;
6507 new_ns->root_cset = cset;
6512 static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6514 return container_of(ns, struct cgroup_namespace, ns);
6517 static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6519 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6521 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6522 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6525 /* Don't need to do anything if we are attaching to our own cgroupns. */
6526 if (cgroup_ns == nsproxy->cgroup_ns)
6529 get_cgroup_ns(cgroup_ns);
6530 put_cgroup_ns(nsproxy->cgroup_ns);
6531 nsproxy->cgroup_ns = cgroup_ns;
6536 static struct ns_common *cgroupns_get(struct task_struct *task)
6538 struct cgroup_namespace *ns = NULL;
6539 struct nsproxy *nsproxy;
6542 nsproxy = task->nsproxy;
6544 ns = nsproxy->cgroup_ns;
6549 return ns ? &ns->ns : NULL;
6552 static void cgroupns_put(struct ns_common *ns)
6554 put_cgroup_ns(to_cg_ns(ns));
6557 static struct user_namespace *cgroupns_owner(struct ns_common *ns)
6559 return to_cg_ns(ns)->user_ns;
6562 const struct proc_ns_operations cgroupns_operations = {
6564 .type = CLONE_NEWCGROUP,
6565 .get = cgroupns_get,
6566 .put = cgroupns_put,
6567 .install = cgroupns_install,
6568 .owner = cgroupns_owner,
6571 static __init int cgroup_namespaces_init(void)
6575 subsys_initcall(cgroup_namespaces_init);
6577 #ifdef CONFIG_CGROUP_DEBUG
6578 static struct cgroup_subsys_state *
6579 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6581 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6584 return ERR_PTR(-ENOMEM);
6589 static void debug_css_free(struct cgroup_subsys_state *css)
6594 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6597 return cgroup_task_count(css->cgroup);
6600 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6603 return (u64)(unsigned long)current->cgroups;
6606 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6612 count = atomic_read(&task_css_set(current)->refcount);
6617 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6619 struct cgrp_cset_link *link;
6620 struct css_set *cset;
6623 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6627 spin_lock_irq(&css_set_lock);
6629 cset = rcu_dereference(current->cgroups);
6630 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6631 struct cgroup *c = link->cgrp;
6633 cgroup_name(c, name_buf, NAME_MAX + 1);
6634 seq_printf(seq, "Root %d group %s\n",
6635 c->root->hierarchy_id, name_buf);
6638 spin_unlock_irq(&css_set_lock);
6643 #define MAX_TASKS_SHOWN_PER_CSS 25
6644 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6646 struct cgroup_subsys_state *css = seq_css(seq);
6647 struct cgrp_cset_link *link;
6649 spin_lock_irq(&css_set_lock);
6650 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6651 struct css_set *cset = link->cset;
6652 struct task_struct *task;
6655 seq_printf(seq, "css_set %p\n", cset);
6657 list_for_each_entry(task, &cset->tasks, cg_list) {
6658 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6660 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6663 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6664 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6666 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6670 seq_puts(seq, " ...\n");
6672 spin_unlock_irq(&css_set_lock);
6676 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6678 return (!cgroup_is_populated(css->cgroup) &&
6679 !css_has_online_children(&css->cgroup->self));
6682 static struct cftype debug_files[] = {
6684 .name = "taskcount",
6685 .read_u64 = debug_taskcount_read,
6689 .name = "current_css_set",
6690 .read_u64 = current_css_set_read,
6694 .name = "current_css_set_refcount",
6695 .read_u64 = current_css_set_refcount_read,
6699 .name = "current_css_set_cg_links",
6700 .seq_show = current_css_set_cg_links_read,
6704 .name = "cgroup_css_links",
6705 .seq_show = cgroup_css_links_read,
6709 .name = "releasable",
6710 .read_u64 = releasable_read,
6716 struct cgroup_subsys debug_cgrp_subsys = {
6717 .css_alloc = debug_css_alloc,
6718 .css_free = debug_css_free,
6719 .legacy_cftypes = debug_files,
6721 #endif /* CONFIG_CGROUP_DEBUG */