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/proc_ns.h>
68 * pidlists linger the following amount before being destroyed. The goal
69 * is avoiding frequent destruction in the middle of consecutive read calls
70 * Expiring in the middle is a performance problem not a correctness one.
71 * 1 sec should be enough.
73 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
75 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
79 * cgroup_mutex is the master lock. Any modification to cgroup or its
80 * hierarchy must be performed while holding it.
82 * css_set_lock protects task->cgroups pointer, the list of css_set
83 * objects, and the chain of tasks off each css_set.
85 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
86 * cgroup.h can use them for lockdep annotations.
88 #ifdef CONFIG_PROVE_RCU
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
91 EXPORT_SYMBOL_GPL(cgroup_mutex);
92 EXPORT_SYMBOL_GPL(css_set_lock);
94 static DEFINE_MUTEX(cgroup_mutex);
95 static DEFINE_SPINLOCK(css_set_lock);
99 * Protects cgroup_idr and css_idr so that IDs can be released without
100 * grabbing cgroup_mutex.
102 static DEFINE_SPINLOCK(cgroup_idr_lock);
105 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
106 * against file removal/re-creation across css hiding.
108 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
111 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
112 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
114 static DEFINE_SPINLOCK(release_agent_path_lock);
116 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
118 #define cgroup_assert_mutex_or_rcu_locked() \
119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
120 !lockdep_is_held(&cgroup_mutex), \
121 "cgroup_mutex or RCU read lock required");
124 * cgroup destruction makes heavy use of work items and there can be a lot
125 * of concurrent destructions. Use a separate workqueue so that cgroup
126 * destruction work items don't end up filling up max_active of system_wq
127 * which may lead to deadlock.
129 static struct workqueue_struct *cgroup_destroy_wq;
132 * pidlist destructions need to be flushed on cgroup destruction. Use a
133 * separate workqueue as flush domain.
135 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
137 /* generate an array of cgroup subsystem pointers */
138 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
139 static struct cgroup_subsys *cgroup_subsys[] = {
140 #include <linux/cgroup_subsys.h>
144 /* array of cgroup subsystem names */
145 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
146 static const char *cgroup_subsys_name[] = {
147 #include <linux/cgroup_subsys.h>
151 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
153 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
154 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
155 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
156 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
157 #include <linux/cgroup_subsys.h>
160 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
161 static struct static_key_true *cgroup_subsys_enabled_key[] = {
162 #include <linux/cgroup_subsys.h>
166 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
167 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
168 #include <linux/cgroup_subsys.h>
173 * The default hierarchy, reserved for the subsystems that are otherwise
174 * unattached - it never has more than a single cgroup, and all tasks are
175 * part of that cgroup.
177 struct cgroup_root cgrp_dfl_root;
178 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
181 * The default hierarchy always exists but is hidden until mounted for the
182 * first time. This is for backward compatibility.
184 static bool cgrp_dfl_visible;
186 /* Controllers blocked by the commandline in v1 */
187 static u16 cgroup_no_v1_mask;
189 /* some controllers are not supported in the default hierarchy */
190 static u16 cgrp_dfl_inhibit_ss_mask;
192 /* some controllers are implicitly enabled on the default hierarchy */
193 static unsigned long cgrp_dfl_implicit_ss_mask;
195 /* The list of hierarchy roots */
197 static LIST_HEAD(cgroup_roots);
198 static int cgroup_root_count;
200 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
201 static DEFINE_IDR(cgroup_hierarchy_idr);
204 * Assign a monotonically increasing serial number to csses. It guarantees
205 * cgroups with bigger numbers are newer than those with smaller numbers.
206 * Also, as csses are always appended to the parent's ->children list, it
207 * guarantees that sibling csses are always sorted in the ascending serial
208 * number order on the list. Protected by cgroup_mutex.
210 static u64 css_serial_nr_next = 1;
213 * These bitmask flags indicate whether tasks in the fork and exit paths have
214 * fork/exit handlers to call. This avoids us having to do extra work in the
215 * fork/exit path to check which subsystems have fork/exit callbacks.
217 static u16 have_fork_callback __read_mostly;
218 static u16 have_exit_callback __read_mostly;
219 static u16 have_free_callback __read_mostly;
221 /* cgroup namespace for init task */
222 struct cgroup_namespace init_cgroup_ns = {
223 .count = { .counter = 2, },
224 .user_ns = &init_user_ns,
225 .ns.ops = &cgroupns_operations,
226 .ns.inum = PROC_CGROUP_INIT_INO,
227 .root_cset = &init_css_set,
230 /* Ditto for the can_fork callback. */
231 static u16 have_canfork_callback __read_mostly;
233 static struct file_system_type cgroup2_fs_type;
234 static struct cftype cgroup_dfl_base_files[];
235 static struct cftype cgroup_legacy_base_files[];
237 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
238 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
239 static int cgroup_apply_control(struct cgroup *cgrp);
240 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
241 static void css_task_iter_advance(struct css_task_iter *it);
242 static int cgroup_destroy_locked(struct cgroup *cgrp);
243 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 struct cgroup_subsys *ss);
245 static void css_release(struct percpu_ref *ref);
246 static void kill_css(struct cgroup_subsys_state *css);
247 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 struct cgroup *cgrp, struct cftype cfts[],
252 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
253 * @ssid: subsys ID of interest
255 * cgroup_subsys_enabled() can only be used with literal subsys names which
256 * is fine for individual subsystems but unsuitable for cgroup core. This
257 * is slower static_key_enabled() based test indexed by @ssid.
259 static bool cgroup_ssid_enabled(int ssid)
261 if (CGROUP_SUBSYS_COUNT == 0)
264 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
267 static bool cgroup_ssid_no_v1(int ssid)
269 return cgroup_no_v1_mask & (1 << ssid);
273 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
274 * @cgrp: the cgroup of interest
276 * The default hierarchy is the v2 interface of cgroup and this function
277 * can be used to test whether a cgroup is on the default hierarchy for
278 * cases where a subsystem should behave differnetly depending on the
281 * The set of behaviors which change on the default hierarchy are still
282 * being determined and the mount option is prefixed with __DEVEL__.
284 * List of changed behaviors:
286 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
287 * and "name" are disallowed.
289 * - When mounting an existing superblock, mount options should match.
291 * - Remount is disallowed.
293 * - rename(2) is disallowed.
295 * - "tasks" is removed. Everything should be at process granularity. Use
296 * "cgroup.procs" instead.
298 * - "cgroup.procs" is not sorted. pids will be unique unless they got
299 * recycled inbetween reads.
301 * - "release_agent" and "notify_on_release" are removed. Replacement
302 * notification mechanism will be implemented.
304 * - "cgroup.clone_children" is removed.
306 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
307 * and its descendants contain no task; otherwise, 1. The file also
308 * generates kernfs notification which can be monitored through poll and
309 * [di]notify when the value of the file changes.
311 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
312 * take masks of ancestors with non-empty cpus/mems, instead of being
313 * moved to an ancestor.
315 * - cpuset: a task can be moved into an empty cpuset, and again it takes
316 * masks of ancestors.
318 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
321 * - blkcg: blk-throttle becomes properly hierarchical.
323 * - debug: disallowed on the default hierarchy.
325 static bool cgroup_on_dfl(const struct cgroup *cgrp)
327 return cgrp->root == &cgrp_dfl_root;
330 /* IDR wrappers which synchronize using cgroup_idr_lock */
331 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
336 idr_preload(gfp_mask);
337 spin_lock_bh(&cgroup_idr_lock);
338 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
339 spin_unlock_bh(&cgroup_idr_lock);
344 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
348 spin_lock_bh(&cgroup_idr_lock);
349 ret = idr_replace(idr, ptr, id);
350 spin_unlock_bh(&cgroup_idr_lock);
354 static void cgroup_idr_remove(struct idr *idr, int id)
356 spin_lock_bh(&cgroup_idr_lock);
358 spin_unlock_bh(&cgroup_idr_lock);
361 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
363 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
366 return container_of(parent_css, struct cgroup, self);
370 /* subsystems visibly enabled on a cgroup */
371 static u16 cgroup_control(struct cgroup *cgrp)
373 struct cgroup *parent = cgroup_parent(cgrp);
374 u16 root_ss_mask = cgrp->root->subsys_mask;
377 return parent->subtree_control;
379 if (cgroup_on_dfl(cgrp))
380 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
381 cgrp_dfl_implicit_ss_mask);
385 /* subsystems enabled on a cgroup */
386 static u16 cgroup_ss_mask(struct cgroup *cgrp)
388 struct cgroup *parent = cgroup_parent(cgrp);
391 return parent->subtree_ss_mask;
393 return cgrp->root->subsys_mask;
397 * cgroup_css - obtain a cgroup's css for the specified subsystem
398 * @cgrp: the cgroup of interest
399 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
401 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
402 * function must be called either under cgroup_mutex or rcu_read_lock() and
403 * the caller is responsible for pinning the returned css if it wants to
404 * keep accessing it outside the said locks. This function may return
405 * %NULL if @cgrp doesn't have @subsys_id enabled.
407 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
408 struct cgroup_subsys *ss)
411 return rcu_dereference_check(cgrp->subsys[ss->id],
412 lockdep_is_held(&cgroup_mutex));
418 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
419 * @cgrp: the cgroup of interest
420 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
422 * Similar to cgroup_css() but returns the effective css, which is defined
423 * as the matching css of the nearest ancestor including self which has @ss
424 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
425 * function is guaranteed to return non-NULL css.
427 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
428 struct cgroup_subsys *ss)
430 lockdep_assert_held(&cgroup_mutex);
436 * This function is used while updating css associations and thus
437 * can't test the csses directly. Test ss_mask.
439 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
440 cgrp = cgroup_parent(cgrp);
445 return cgroup_css(cgrp, ss);
449 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
450 * @cgrp: the cgroup of interest
451 * @ss: the subsystem of interest
453 * Find and get the effective css of @cgrp for @ss. The effective css is
454 * defined as the matching css of the nearest ancestor including self which
455 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
456 * the root css is returned, so this function always returns a valid css.
457 * The returned css must be put using css_put().
459 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
460 struct cgroup_subsys *ss)
462 struct cgroup_subsys_state *css;
467 css = cgroup_css(cgrp, ss);
469 if (css && css_tryget_online(css))
471 cgrp = cgroup_parent(cgrp);
474 css = init_css_set.subsys[ss->id];
481 /* convenient tests for these bits */
482 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
484 return !(cgrp->self.flags & CSS_ONLINE);
487 static void cgroup_get(struct cgroup *cgrp)
489 WARN_ON_ONCE(cgroup_is_dead(cgrp));
490 css_get(&cgrp->self);
493 static bool cgroup_tryget(struct cgroup *cgrp)
495 return css_tryget(&cgrp->self);
498 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
500 struct cgroup *cgrp = of->kn->parent->priv;
501 struct cftype *cft = of_cft(of);
504 * This is open and unprotected implementation of cgroup_css().
505 * seq_css() is only called from a kernfs file operation which has
506 * an active reference on the file. Because all the subsystem
507 * files are drained before a css is disassociated with a cgroup,
508 * the matching css from the cgroup's subsys table is guaranteed to
509 * be and stay valid until the enclosing operation is complete.
512 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
516 EXPORT_SYMBOL_GPL(of_css);
518 static int notify_on_release(const struct cgroup *cgrp)
520 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
524 * for_each_css - iterate all css's of a cgroup
525 * @css: the iteration cursor
526 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
527 * @cgrp: the target cgroup to iterate css's of
529 * Should be called under cgroup_[tree_]mutex.
531 #define for_each_css(css, ssid, cgrp) \
532 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
533 if (!((css) = rcu_dereference_check( \
534 (cgrp)->subsys[(ssid)], \
535 lockdep_is_held(&cgroup_mutex)))) { } \
539 * for_each_e_css - iterate all effective css's of a cgroup
540 * @css: the iteration cursor
541 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
542 * @cgrp: the target cgroup to iterate css's of
544 * Should be called under cgroup_[tree_]mutex.
546 #define for_each_e_css(css, ssid, cgrp) \
547 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
548 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
553 * for_each_subsys - iterate all enabled cgroup subsystems
554 * @ss: the iteration cursor
555 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
557 #define for_each_subsys(ss, ssid) \
558 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
559 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
562 * do_each_subsys_mask - filter for_each_subsys with a bitmask
563 * @ss: the iteration cursor
564 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
565 * @ss_mask: the bitmask
567 * The block will only run for cases where the ssid-th bit (1 << ssid) of
570 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
571 unsigned long __ss_mask = (ss_mask); \
572 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
576 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
577 (ss) = cgroup_subsys[ssid]; \
580 #define while_each_subsys_mask() \
585 /* iterate across the hierarchies */
586 #define for_each_root(root) \
587 list_for_each_entry((root), &cgroup_roots, root_list)
589 /* iterate over child cgrps, lock should be held throughout iteration */
590 #define cgroup_for_each_live_child(child, cgrp) \
591 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
592 if (({ lockdep_assert_held(&cgroup_mutex); \
593 cgroup_is_dead(child); })) \
597 /* walk live descendants in preorder */
598 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
599 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
600 if (({ lockdep_assert_held(&cgroup_mutex); \
601 (dsct) = (d_css)->cgroup; \
602 cgroup_is_dead(dsct); })) \
606 /* walk live descendants in postorder */
607 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
608 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
609 if (({ lockdep_assert_held(&cgroup_mutex); \
610 (dsct) = (d_css)->cgroup; \
611 cgroup_is_dead(dsct); })) \
615 static void cgroup_release_agent(struct work_struct *work);
616 static void check_for_release(struct cgroup *cgrp);
619 * A cgroup can be associated with multiple css_sets as different tasks may
620 * belong to different cgroups on different hierarchies. In the other
621 * direction, a css_set is naturally associated with multiple cgroups.
622 * This M:N relationship is represented by the following link structure
623 * which exists for each association and allows traversing the associations
626 struct cgrp_cset_link {
627 /* the cgroup and css_set this link associates */
629 struct css_set *cset;
631 /* list of cgrp_cset_links anchored at cgrp->cset_links */
632 struct list_head cset_link;
634 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
635 struct list_head cgrp_link;
639 * The default css_set - used by init and its children prior to any
640 * hierarchies being mounted. It contains a pointer to the root state
641 * for each subsystem. Also used to anchor the list of css_sets. Not
642 * reference-counted, to improve performance when child cgroups
643 * haven't been created.
645 struct css_set init_css_set = {
646 .refcount = ATOMIC_INIT(1),
647 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
648 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
649 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
650 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
651 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
652 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
655 static int css_set_count = 1; /* 1 for init_css_set */
658 * css_set_populated - does a css_set contain any tasks?
659 * @cset: target css_set
661 static bool css_set_populated(struct css_set *cset)
663 lockdep_assert_held(&css_set_lock);
665 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
669 * cgroup_update_populated - updated populated count of a cgroup
670 * @cgrp: the target cgroup
671 * @populated: inc or dec populated count
673 * One of the css_sets associated with @cgrp is either getting its first
674 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
675 * count is propagated towards root so that a given cgroup's populated_cnt
676 * is zero iff the cgroup and all its descendants don't contain any tasks.
678 * @cgrp's interface file "cgroup.populated" is zero if
679 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
680 * changes from or to zero, userland is notified that the content of the
681 * interface file has changed. This can be used to detect when @cgrp and
682 * its descendants become populated or empty.
684 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
686 lockdep_assert_held(&css_set_lock);
692 trigger = !cgrp->populated_cnt++;
694 trigger = !--cgrp->populated_cnt;
699 check_for_release(cgrp);
700 cgroup_file_notify(&cgrp->events_file);
702 cgrp = cgroup_parent(cgrp);
707 * css_set_update_populated - update populated state of a css_set
708 * @cset: target css_set
709 * @populated: whether @cset is populated or depopulated
711 * @cset is either getting the first task or losing the last. Update the
712 * ->populated_cnt of all associated cgroups accordingly.
714 static void css_set_update_populated(struct css_set *cset, bool populated)
716 struct cgrp_cset_link *link;
718 lockdep_assert_held(&css_set_lock);
720 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
721 cgroup_update_populated(link->cgrp, populated);
725 * css_set_move_task - move a task from one css_set to another
726 * @task: task being moved
727 * @from_cset: css_set @task currently belongs to (may be NULL)
728 * @to_cset: new css_set @task is being moved to (may be NULL)
729 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
731 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
732 * css_set, @from_cset can be NULL. If @task is being disassociated
733 * instead of moved, @to_cset can be NULL.
735 * This function automatically handles populated_cnt updates and
736 * css_task_iter adjustments but the caller is responsible for managing
737 * @from_cset and @to_cset's reference counts.
739 static void css_set_move_task(struct task_struct *task,
740 struct css_set *from_cset, struct css_set *to_cset,
743 lockdep_assert_held(&css_set_lock);
745 if (to_cset && !css_set_populated(to_cset))
746 css_set_update_populated(to_cset, true);
749 struct css_task_iter *it, *pos;
751 WARN_ON_ONCE(list_empty(&task->cg_list));
754 * @task is leaving, advance task iterators which are
755 * pointing to it so that they can resume at the next
756 * position. Advancing an iterator might remove it from
757 * the list, use safe walk. See css_task_iter_advance*()
760 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
762 if (it->task_pos == &task->cg_list)
763 css_task_iter_advance(it);
765 list_del_init(&task->cg_list);
766 if (!css_set_populated(from_cset))
767 css_set_update_populated(from_cset, false);
769 WARN_ON_ONCE(!list_empty(&task->cg_list));
774 * We are synchronized through cgroup_threadgroup_rwsem
775 * against PF_EXITING setting such that we can't race
776 * against cgroup_exit() changing the css_set to
777 * init_css_set and dropping the old one.
779 WARN_ON_ONCE(task->flags & PF_EXITING);
781 rcu_assign_pointer(task->cgroups, to_cset);
782 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
788 * hash table for cgroup groups. This improves the performance to find
789 * an existing css_set. This hash doesn't (currently) take into
790 * account cgroups in empty hierarchies.
792 #define CSS_SET_HASH_BITS 7
793 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
795 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
797 unsigned long key = 0UL;
798 struct cgroup_subsys *ss;
801 for_each_subsys(ss, i)
802 key += (unsigned long)css[i];
803 key = (key >> 16) ^ key;
808 static void put_css_set_locked(struct css_set *cset)
810 struct cgrp_cset_link *link, *tmp_link;
811 struct cgroup_subsys *ss;
814 lockdep_assert_held(&css_set_lock);
816 if (!atomic_dec_and_test(&cset->refcount))
819 /* This css_set is dead. unlink it and release cgroup and css refs */
820 for_each_subsys(ss, ssid) {
821 list_del(&cset->e_cset_node[ssid]);
822 css_put(cset->subsys[ssid]);
824 hash_del(&cset->hlist);
827 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
828 list_del(&link->cset_link);
829 list_del(&link->cgrp_link);
830 if (cgroup_parent(link->cgrp))
831 cgroup_put(link->cgrp);
835 kfree_rcu(cset, rcu_head);
838 static void put_css_set(struct css_set *cset)
841 * Ensure that the refcount doesn't hit zero while any readers
842 * can see it. Similar to atomic_dec_and_lock(), but for an
845 if (atomic_add_unless(&cset->refcount, -1, 1))
848 spin_lock_bh(&css_set_lock);
849 put_css_set_locked(cset);
850 spin_unlock_bh(&css_set_lock);
854 * refcounted get/put for css_set objects
856 static inline void get_css_set(struct css_set *cset)
858 atomic_inc(&cset->refcount);
862 * compare_css_sets - helper function for find_existing_css_set().
863 * @cset: candidate css_set being tested
864 * @old_cset: existing css_set for a task
865 * @new_cgrp: cgroup that's being entered by the task
866 * @template: desired set of css pointers in css_set (pre-calculated)
868 * Returns true if "cset" matches "old_cset" except for the hierarchy
869 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
871 static bool compare_css_sets(struct css_set *cset,
872 struct css_set *old_cset,
873 struct cgroup *new_cgrp,
874 struct cgroup_subsys_state *template[])
876 struct list_head *l1, *l2;
879 * On the default hierarchy, there can be csets which are
880 * associated with the same set of cgroups but different csses.
881 * Let's first ensure that csses match.
883 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
887 * Compare cgroup pointers in order to distinguish between
888 * different cgroups in hierarchies. As different cgroups may
889 * share the same effective css, this comparison is always
892 l1 = &cset->cgrp_links;
893 l2 = &old_cset->cgrp_links;
895 struct cgrp_cset_link *link1, *link2;
896 struct cgroup *cgrp1, *cgrp2;
900 /* See if we reached the end - both lists are equal length. */
901 if (l1 == &cset->cgrp_links) {
902 BUG_ON(l2 != &old_cset->cgrp_links);
905 BUG_ON(l2 == &old_cset->cgrp_links);
907 /* Locate the cgroups associated with these links. */
908 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
909 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
912 /* Hierarchies should be linked in the same order. */
913 BUG_ON(cgrp1->root != cgrp2->root);
916 * If this hierarchy is the hierarchy of the cgroup
917 * that's changing, then we need to check that this
918 * css_set points to the new cgroup; if it's any other
919 * hierarchy, then this css_set should point to the
920 * same cgroup as the old css_set.
922 if (cgrp1->root == new_cgrp->root) {
923 if (cgrp1 != new_cgrp)
934 * find_existing_css_set - init css array and find the matching css_set
935 * @old_cset: the css_set that we're using before the cgroup transition
936 * @cgrp: the cgroup that we're moving into
937 * @template: out param for the new set of csses, should be clear on entry
939 static struct css_set *find_existing_css_set(struct css_set *old_cset,
941 struct cgroup_subsys_state *template[])
943 struct cgroup_root *root = cgrp->root;
944 struct cgroup_subsys *ss;
945 struct css_set *cset;
950 * Build the set of subsystem state objects that we want to see in the
951 * new css_set. while subsystems can change globally, the entries here
952 * won't change, so no need for locking.
954 for_each_subsys(ss, i) {
955 if (root->subsys_mask & (1UL << i)) {
957 * @ss is in this hierarchy, so we want the
958 * effective css from @cgrp.
960 template[i] = cgroup_e_css(cgrp, ss);
963 * @ss is not in this hierarchy, so we don't want
966 template[i] = old_cset->subsys[i];
970 key = css_set_hash(template);
971 hash_for_each_possible(css_set_table, cset, hlist, key) {
972 if (!compare_css_sets(cset, old_cset, cgrp, template))
975 /* This css_set matches what we need */
979 /* No existing cgroup group matched */
983 static void free_cgrp_cset_links(struct list_head *links_to_free)
985 struct cgrp_cset_link *link, *tmp_link;
987 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
988 list_del(&link->cset_link);
994 * allocate_cgrp_cset_links - allocate cgrp_cset_links
995 * @count: the number of links to allocate
996 * @tmp_links: list_head the allocated links are put on
998 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
999 * through ->cset_link. Returns 0 on success or -errno.
1001 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1003 struct cgrp_cset_link *link;
1006 INIT_LIST_HEAD(tmp_links);
1008 for (i = 0; i < count; i++) {
1009 link = kzalloc(sizeof(*link), GFP_KERNEL);
1011 free_cgrp_cset_links(tmp_links);
1014 list_add(&link->cset_link, tmp_links);
1020 * link_css_set - a helper function to link a css_set to a cgroup
1021 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1022 * @cset: the css_set to be linked
1023 * @cgrp: the destination cgroup
1025 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1026 struct cgroup *cgrp)
1028 struct cgrp_cset_link *link;
1030 BUG_ON(list_empty(tmp_links));
1032 if (cgroup_on_dfl(cgrp))
1033 cset->dfl_cgrp = cgrp;
1035 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1040 * Always add links to the tail of the lists so that the lists are
1041 * in choronological order.
1043 list_move_tail(&link->cset_link, &cgrp->cset_links);
1044 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1046 if (cgroup_parent(cgrp))
1051 * find_css_set - return a new css_set with one cgroup updated
1052 * @old_cset: the baseline css_set
1053 * @cgrp: the cgroup to be updated
1055 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1056 * substituted into the appropriate hierarchy.
1058 static struct css_set *find_css_set(struct css_set *old_cset,
1059 struct cgroup *cgrp)
1061 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1062 struct css_set *cset;
1063 struct list_head tmp_links;
1064 struct cgrp_cset_link *link;
1065 struct cgroup_subsys *ss;
1069 lockdep_assert_held(&cgroup_mutex);
1071 /* First see if we already have a cgroup group that matches
1072 * the desired set */
1073 spin_lock_bh(&css_set_lock);
1074 cset = find_existing_css_set(old_cset, cgrp, template);
1077 spin_unlock_bh(&css_set_lock);
1082 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1086 /* Allocate all the cgrp_cset_link objects that we'll need */
1087 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1092 atomic_set(&cset->refcount, 1);
1093 INIT_LIST_HEAD(&cset->cgrp_links);
1094 INIT_LIST_HEAD(&cset->tasks);
1095 INIT_LIST_HEAD(&cset->mg_tasks);
1096 INIT_LIST_HEAD(&cset->mg_preload_node);
1097 INIT_LIST_HEAD(&cset->mg_node);
1098 INIT_LIST_HEAD(&cset->task_iters);
1099 INIT_HLIST_NODE(&cset->hlist);
1101 /* Copy the set of subsystem state objects generated in
1102 * find_existing_css_set() */
1103 memcpy(cset->subsys, template, sizeof(cset->subsys));
1105 spin_lock_bh(&css_set_lock);
1106 /* Add reference counts and links from the new css_set. */
1107 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1108 struct cgroup *c = link->cgrp;
1110 if (c->root == cgrp->root)
1112 link_css_set(&tmp_links, cset, c);
1115 BUG_ON(!list_empty(&tmp_links));
1119 /* Add @cset to the hash table */
1120 key = css_set_hash(cset->subsys);
1121 hash_add(css_set_table, &cset->hlist, key);
1123 for_each_subsys(ss, ssid) {
1124 struct cgroup_subsys_state *css = cset->subsys[ssid];
1126 list_add_tail(&cset->e_cset_node[ssid],
1127 &css->cgroup->e_csets[ssid]);
1131 spin_unlock_bh(&css_set_lock);
1136 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1138 struct cgroup *root_cgrp = kf_root->kn->priv;
1140 return root_cgrp->root;
1143 static int cgroup_init_root_id(struct cgroup_root *root)
1147 lockdep_assert_held(&cgroup_mutex);
1149 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1153 root->hierarchy_id = id;
1157 static void cgroup_exit_root_id(struct cgroup_root *root)
1159 lockdep_assert_held(&cgroup_mutex);
1161 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1164 static void cgroup_free_root(struct cgroup_root *root)
1167 idr_destroy(&root->cgroup_idr);
1172 static void cgroup_destroy_root(struct cgroup_root *root)
1174 struct cgroup *cgrp = &root->cgrp;
1175 struct cgrp_cset_link *link, *tmp_link;
1177 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1179 BUG_ON(atomic_read(&root->nr_cgrps));
1180 BUG_ON(!list_empty(&cgrp->self.children));
1182 /* Rebind all subsystems back to the default hierarchy */
1183 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1186 * Release all the links from cset_links to this hierarchy's
1189 spin_lock_bh(&css_set_lock);
1191 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1192 list_del(&link->cset_link);
1193 list_del(&link->cgrp_link);
1197 spin_unlock_bh(&css_set_lock);
1199 if (!list_empty(&root->root_list)) {
1200 list_del(&root->root_list);
1201 cgroup_root_count--;
1204 cgroup_exit_root_id(root);
1206 mutex_unlock(&cgroup_mutex);
1208 kernfs_destroy_root(root->kf_root);
1209 cgroup_free_root(root);
1213 * look up cgroup associated with current task's cgroup namespace on the
1214 * specified hierarchy
1216 static struct cgroup *
1217 current_cgns_cgroup_from_root(struct cgroup_root *root)
1219 struct cgroup *res = NULL;
1220 struct css_set *cset;
1222 lockdep_assert_held(&css_set_lock);
1226 cset = current->nsproxy->cgroup_ns->root_cset;
1227 if (cset == &init_css_set) {
1230 struct cgrp_cset_link *link;
1232 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1233 struct cgroup *c = link->cgrp;
1235 if (c->root == root) {
1247 /* look up cgroup associated with given css_set on the specified hierarchy */
1248 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1249 struct cgroup_root *root)
1251 struct cgroup *res = NULL;
1253 lockdep_assert_held(&cgroup_mutex);
1254 lockdep_assert_held(&css_set_lock);
1256 if (cset == &init_css_set) {
1259 struct cgrp_cset_link *link;
1261 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1262 struct cgroup *c = link->cgrp;
1264 if (c->root == root) {
1276 * Return the cgroup for "task" from the given hierarchy. Must be
1277 * called with cgroup_mutex and css_set_lock held.
1279 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1280 struct cgroup_root *root)
1283 * No need to lock the task - since we hold cgroup_mutex the
1284 * task can't change groups, so the only thing that can happen
1285 * is that it exits and its css is set back to init_css_set.
1287 return cset_cgroup_from_root(task_css_set(task), root);
1291 * A task must hold cgroup_mutex to modify cgroups.
1293 * Any task can increment and decrement the count field without lock.
1294 * So in general, code holding cgroup_mutex can't rely on the count
1295 * field not changing. However, if the count goes to zero, then only
1296 * cgroup_attach_task() can increment it again. Because a count of zero
1297 * means that no tasks are currently attached, therefore there is no
1298 * way a task attached to that cgroup can fork (the other way to
1299 * increment the count). So code holding cgroup_mutex can safely
1300 * assume that if the count is zero, it will stay zero. Similarly, if
1301 * a task holds cgroup_mutex on a cgroup with zero count, it
1302 * knows that the cgroup won't be removed, as cgroup_rmdir()
1305 * A cgroup can only be deleted if both its 'count' of using tasks
1306 * is zero, and its list of 'children' cgroups is empty. Since all
1307 * tasks in the system use _some_ cgroup, and since there is always at
1308 * least one task in the system (init, pid == 1), therefore, root cgroup
1309 * always has either children cgroups and/or using tasks. So we don't
1310 * need a special hack to ensure that root cgroup cannot be deleted.
1312 * P.S. One more locking exception. RCU is used to guard the
1313 * update of a tasks cgroup pointer by cgroup_attach_task()
1316 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1317 static const struct file_operations proc_cgroupstats_operations;
1319 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1322 struct cgroup_subsys *ss = cft->ss;
1324 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1325 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1326 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1327 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1330 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1335 * cgroup_file_mode - deduce file mode of a control file
1336 * @cft: the control file in question
1338 * S_IRUGO for read, S_IWUSR for write.
1340 static umode_t cgroup_file_mode(const struct cftype *cft)
1344 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1347 if (cft->write_u64 || cft->write_s64 || cft->write) {
1348 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1358 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1359 * @subtree_control: the new subtree_control mask to consider
1360 * @this_ss_mask: available subsystems
1362 * On the default hierarchy, a subsystem may request other subsystems to be
1363 * enabled together through its ->depends_on mask. In such cases, more
1364 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1366 * This function calculates which subsystems need to be enabled if
1367 * @subtree_control is to be applied while restricted to @this_ss_mask.
1369 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1371 u16 cur_ss_mask = subtree_control;
1372 struct cgroup_subsys *ss;
1375 lockdep_assert_held(&cgroup_mutex);
1377 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1380 u16 new_ss_mask = cur_ss_mask;
1382 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1383 new_ss_mask |= ss->depends_on;
1384 } while_each_subsys_mask();
1387 * Mask out subsystems which aren't available. This can
1388 * happen only if some depended-upon subsystems were bound
1389 * to non-default hierarchies.
1391 new_ss_mask &= this_ss_mask;
1393 if (new_ss_mask == cur_ss_mask)
1395 cur_ss_mask = new_ss_mask;
1402 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1403 * @kn: the kernfs_node being serviced
1405 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1406 * the method finishes if locking succeeded. Note that once this function
1407 * returns the cgroup returned by cgroup_kn_lock_live() may become
1408 * inaccessible any time. If the caller intends to continue to access the
1409 * cgroup, it should pin it before invoking this function.
1411 static void cgroup_kn_unlock(struct kernfs_node *kn)
1413 struct cgroup *cgrp;
1415 if (kernfs_type(kn) == KERNFS_DIR)
1418 cgrp = kn->parent->priv;
1420 mutex_unlock(&cgroup_mutex);
1422 kernfs_unbreak_active_protection(kn);
1427 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1428 * @kn: the kernfs_node being serviced
1429 * @drain_offline: perform offline draining on the cgroup
1431 * This helper is to be used by a cgroup kernfs method currently servicing
1432 * @kn. It breaks the active protection, performs cgroup locking and
1433 * verifies that the associated cgroup is alive. Returns the cgroup if
1434 * alive; otherwise, %NULL. A successful return should be undone by a
1435 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1436 * cgroup is drained of offlining csses before return.
1438 * Any cgroup kernfs method implementation which requires locking the
1439 * associated cgroup should use this helper. It avoids nesting cgroup
1440 * locking under kernfs active protection and allows all kernfs operations
1441 * including self-removal.
1443 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1446 struct cgroup *cgrp;
1448 if (kernfs_type(kn) == KERNFS_DIR)
1451 cgrp = kn->parent->priv;
1454 * We're gonna grab cgroup_mutex which nests outside kernfs
1455 * active_ref. cgroup liveliness check alone provides enough
1456 * protection against removal. Ensure @cgrp stays accessible and
1457 * break the active_ref protection.
1459 if (!cgroup_tryget(cgrp))
1461 kernfs_break_active_protection(kn);
1464 cgroup_lock_and_drain_offline(cgrp);
1466 mutex_lock(&cgroup_mutex);
1468 if (!cgroup_is_dead(cgrp))
1471 cgroup_kn_unlock(kn);
1475 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1477 char name[CGROUP_FILE_NAME_MAX];
1479 lockdep_assert_held(&cgroup_mutex);
1481 if (cft->file_offset) {
1482 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1483 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1485 spin_lock_irq(&cgroup_file_kn_lock);
1487 spin_unlock_irq(&cgroup_file_kn_lock);
1490 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1494 * css_clear_dir - remove subsys files in a cgroup directory
1497 static void css_clear_dir(struct cgroup_subsys_state *css)
1499 struct cgroup *cgrp = css->cgroup;
1500 struct cftype *cfts;
1502 if (!(css->flags & CSS_VISIBLE))
1505 css->flags &= ~CSS_VISIBLE;
1507 list_for_each_entry(cfts, &css->ss->cfts, node)
1508 cgroup_addrm_files(css, cgrp, cfts, false);
1512 * css_populate_dir - create subsys files in a cgroup directory
1515 * On failure, no file is added.
1517 static int css_populate_dir(struct cgroup_subsys_state *css)
1519 struct cgroup *cgrp = css->cgroup;
1520 struct cftype *cfts, *failed_cfts;
1523 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1527 if (cgroup_on_dfl(cgrp))
1528 cfts = cgroup_dfl_base_files;
1530 cfts = cgroup_legacy_base_files;
1532 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1535 list_for_each_entry(cfts, &css->ss->cfts, node) {
1536 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1543 css->flags |= CSS_VISIBLE;
1547 list_for_each_entry(cfts, &css->ss->cfts, node) {
1548 if (cfts == failed_cfts)
1550 cgroup_addrm_files(css, cgrp, cfts, false);
1555 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1557 struct cgroup *dcgrp = &dst_root->cgrp;
1558 struct cgroup_subsys *ss;
1561 lockdep_assert_held(&cgroup_mutex);
1563 do_each_subsys_mask(ss, ssid, ss_mask) {
1565 * If @ss has non-root csses attached to it, can't move.
1566 * If @ss is an implicit controller, it is exempt from this
1567 * rule and can be stolen.
1569 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1570 !ss->implicit_on_dfl)
1573 /* can't move between two non-dummy roots either */
1574 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1576 } while_each_subsys_mask();
1578 do_each_subsys_mask(ss, ssid, ss_mask) {
1579 struct cgroup_root *src_root = ss->root;
1580 struct cgroup *scgrp = &src_root->cgrp;
1581 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1582 struct css_set *cset;
1584 WARN_ON(!css || cgroup_css(dcgrp, ss));
1586 /* disable from the source */
1587 src_root->subsys_mask &= ~(1 << ssid);
1588 WARN_ON(cgroup_apply_control(scgrp));
1589 cgroup_finalize_control(scgrp, 0);
1592 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1593 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1594 ss->root = dst_root;
1595 css->cgroup = dcgrp;
1597 spin_lock_bh(&css_set_lock);
1598 hash_for_each(css_set_table, i, cset, hlist)
1599 list_move_tail(&cset->e_cset_node[ss->id],
1600 &dcgrp->e_csets[ss->id]);
1601 spin_unlock_bh(&css_set_lock);
1603 /* default hierarchy doesn't enable controllers by default */
1604 dst_root->subsys_mask |= 1 << ssid;
1605 if (dst_root == &cgrp_dfl_root) {
1606 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1608 dcgrp->subtree_control |= 1 << ssid;
1609 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1612 ret = cgroup_apply_control(dcgrp);
1614 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1619 } while_each_subsys_mask();
1621 kernfs_activate(dcgrp->kn);
1625 static int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1626 struct kernfs_root *kf_root)
1630 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1631 struct cgroup *ns_cgroup;
1633 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1637 spin_lock_bh(&css_set_lock);
1638 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1639 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1640 spin_unlock_bh(&css_set_lock);
1642 if (len >= PATH_MAX)
1645 seq_escape(sf, buf, " \t\n\\");
1652 static int cgroup_show_options(struct seq_file *seq,
1653 struct kernfs_root *kf_root)
1655 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1656 struct cgroup_subsys *ss;
1659 if (root != &cgrp_dfl_root)
1660 for_each_subsys(ss, ssid)
1661 if (root->subsys_mask & (1 << ssid))
1662 seq_show_option(seq, ss->legacy_name, NULL);
1663 if (root->flags & CGRP_ROOT_NOPREFIX)
1664 seq_puts(seq, ",noprefix");
1665 if (root->flags & CGRP_ROOT_XATTR)
1666 seq_puts(seq, ",xattr");
1668 spin_lock(&release_agent_path_lock);
1669 if (strlen(root->release_agent_path))
1670 seq_show_option(seq, "release_agent",
1671 root->release_agent_path);
1672 spin_unlock(&release_agent_path_lock);
1674 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1675 seq_puts(seq, ",clone_children");
1676 if (strlen(root->name))
1677 seq_show_option(seq, "name", root->name);
1681 struct cgroup_sb_opts {
1684 char *release_agent;
1685 bool cpuset_clone_children;
1687 /* User explicitly requested empty subsystem */
1691 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1693 char *token, *o = data;
1694 bool all_ss = false, one_ss = false;
1696 struct cgroup_subsys *ss;
1700 #ifdef CONFIG_CPUSETS
1701 mask = ~((u16)1 << cpuset_cgrp_id);
1704 memset(opts, 0, sizeof(*opts));
1706 while ((token = strsep(&o, ",")) != NULL) {
1711 if (!strcmp(token, "none")) {
1712 /* Explicitly have no subsystems */
1716 if (!strcmp(token, "all")) {
1717 /* Mutually exclusive option 'all' + subsystem name */
1723 if (!strcmp(token, "noprefix")) {
1724 opts->flags |= CGRP_ROOT_NOPREFIX;
1727 if (!strcmp(token, "clone_children")) {
1728 opts->cpuset_clone_children = true;
1731 if (!strcmp(token, "xattr")) {
1732 opts->flags |= CGRP_ROOT_XATTR;
1735 if (!strncmp(token, "release_agent=", 14)) {
1736 /* Specifying two release agents is forbidden */
1737 if (opts->release_agent)
1739 opts->release_agent =
1740 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1741 if (!opts->release_agent)
1745 if (!strncmp(token, "name=", 5)) {
1746 const char *name = token + 5;
1747 /* Can't specify an empty name */
1750 /* Must match [\w.-]+ */
1751 for (i = 0; i < strlen(name); i++) {
1755 if ((c == '.') || (c == '-') || (c == '_'))
1759 /* Specifying two names is forbidden */
1762 opts->name = kstrndup(name,
1763 MAX_CGROUP_ROOT_NAMELEN - 1,
1771 for_each_subsys(ss, i) {
1772 if (strcmp(token, ss->legacy_name))
1774 if (!cgroup_ssid_enabled(i))
1776 if (cgroup_ssid_no_v1(i))
1779 /* Mutually exclusive option 'all' + subsystem name */
1782 opts->subsys_mask |= (1 << i);
1787 if (i == CGROUP_SUBSYS_COUNT)
1792 * If the 'all' option was specified select all the subsystems,
1793 * otherwise if 'none', 'name=' and a subsystem name options were
1794 * not specified, let's default to 'all'
1796 if (all_ss || (!one_ss && !opts->none && !opts->name))
1797 for_each_subsys(ss, i)
1798 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1799 opts->subsys_mask |= (1 << i);
1802 * We either have to specify by name or by subsystems. (So all
1803 * empty hierarchies must have a name).
1805 if (!opts->subsys_mask && !opts->name)
1809 * Option noprefix was introduced just for backward compatibility
1810 * with the old cpuset, so we allow noprefix only if mounting just
1811 * the cpuset subsystem.
1813 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1816 /* Can't specify "none" and some subsystems */
1817 if (opts->subsys_mask && opts->none)
1823 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1826 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1827 struct cgroup_sb_opts opts;
1828 u16 added_mask, removed_mask;
1830 if (root == &cgrp_dfl_root) {
1831 pr_err("remount is not allowed\n");
1835 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1837 /* See what subsystems are wanted */
1838 ret = parse_cgroupfs_options(data, &opts);
1842 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1843 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1844 task_tgid_nr(current), current->comm);
1846 added_mask = opts.subsys_mask & ~root->subsys_mask;
1847 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1849 /* Don't allow flags or name to change at remount */
1850 if ((opts.flags ^ root->flags) ||
1851 (opts.name && strcmp(opts.name, root->name))) {
1852 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1853 opts.flags, opts.name ?: "", root->flags, root->name);
1858 /* remounting is not allowed for populated hierarchies */
1859 if (!list_empty(&root->cgrp.self.children)) {
1864 ret = rebind_subsystems(root, added_mask);
1868 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1870 if (opts.release_agent) {
1871 spin_lock(&release_agent_path_lock);
1872 strcpy(root->release_agent_path, opts.release_agent);
1873 spin_unlock(&release_agent_path_lock);
1876 kfree(opts.release_agent);
1878 mutex_unlock(&cgroup_mutex);
1883 * To reduce the fork() overhead for systems that are not actually using
1884 * their cgroups capability, we don't maintain the lists running through
1885 * each css_set to its tasks until we see the list actually used - in other
1886 * words after the first mount.
1888 static bool use_task_css_set_links __read_mostly;
1890 static void cgroup_enable_task_cg_lists(void)
1892 struct task_struct *p, *g;
1894 spin_lock_bh(&css_set_lock);
1896 if (use_task_css_set_links)
1899 use_task_css_set_links = true;
1902 * We need tasklist_lock because RCU is not safe against
1903 * while_each_thread(). Besides, a forking task that has passed
1904 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1905 * is not guaranteed to have its child immediately visible in the
1906 * tasklist if we walk through it with RCU.
1908 read_lock(&tasklist_lock);
1909 do_each_thread(g, p) {
1910 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1911 task_css_set(p) != &init_css_set);
1914 * We should check if the process is exiting, otherwise
1915 * it will race with cgroup_exit() in that the list
1916 * entry won't be deleted though the process has exited.
1917 * Do it while holding siglock so that we don't end up
1918 * racing against cgroup_exit().
1920 spin_lock_irq(&p->sighand->siglock);
1921 if (!(p->flags & PF_EXITING)) {
1922 struct css_set *cset = task_css_set(p);
1924 if (!css_set_populated(cset))
1925 css_set_update_populated(cset, true);
1926 list_add_tail(&p->cg_list, &cset->tasks);
1929 spin_unlock_irq(&p->sighand->siglock);
1930 } while_each_thread(g, p);
1931 read_unlock(&tasklist_lock);
1933 spin_unlock_bh(&css_set_lock);
1936 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1938 struct cgroup_subsys *ss;
1941 INIT_LIST_HEAD(&cgrp->self.sibling);
1942 INIT_LIST_HEAD(&cgrp->self.children);
1943 INIT_LIST_HEAD(&cgrp->cset_links);
1944 INIT_LIST_HEAD(&cgrp->pidlists);
1945 mutex_init(&cgrp->pidlist_mutex);
1946 cgrp->self.cgroup = cgrp;
1947 cgrp->self.flags |= CSS_ONLINE;
1949 for_each_subsys(ss, ssid)
1950 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1952 init_waitqueue_head(&cgrp->offline_waitq);
1953 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1956 static void init_cgroup_root(struct cgroup_root *root,
1957 struct cgroup_sb_opts *opts)
1959 struct cgroup *cgrp = &root->cgrp;
1961 INIT_LIST_HEAD(&root->root_list);
1962 atomic_set(&root->nr_cgrps, 1);
1964 init_cgroup_housekeeping(cgrp);
1965 idr_init(&root->cgroup_idr);
1967 root->flags = opts->flags;
1968 if (opts->release_agent)
1969 strcpy(root->release_agent_path, opts->release_agent);
1971 strcpy(root->name, opts->name);
1972 if (opts->cpuset_clone_children)
1973 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1976 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1978 LIST_HEAD(tmp_links);
1979 struct cgroup *root_cgrp = &root->cgrp;
1980 struct css_set *cset;
1983 lockdep_assert_held(&cgroup_mutex);
1985 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1988 root_cgrp->id = ret;
1989 root_cgrp->ancestor_ids[0] = ret;
1991 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1997 * We're accessing css_set_count without locking css_set_lock here,
1998 * but that's OK - it can only be increased by someone holding
1999 * cgroup_lock, and that's us. Later rebinding may disable
2000 * controllers on the default hierarchy and thus create new csets,
2001 * which can't be more than the existing ones. Allocate 2x.
2003 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2007 ret = cgroup_init_root_id(root);
2011 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
2012 KERNFS_ROOT_CREATE_DEACTIVATED,
2014 if (IS_ERR(root->kf_root)) {
2015 ret = PTR_ERR(root->kf_root);
2018 root_cgrp->kn = root->kf_root->kn;
2020 ret = css_populate_dir(&root_cgrp->self);
2024 ret = rebind_subsystems(root, ss_mask);
2029 * There must be no failure case after here, since rebinding takes
2030 * care of subsystems' refcounts, which are explicitly dropped in
2031 * the failure exit path.
2033 list_add(&root->root_list, &cgroup_roots);
2034 cgroup_root_count++;
2037 * Link the root cgroup in this hierarchy into all the css_set
2040 spin_lock_bh(&css_set_lock);
2041 hash_for_each(css_set_table, i, cset, hlist) {
2042 link_css_set(&tmp_links, cset, root_cgrp);
2043 if (css_set_populated(cset))
2044 cgroup_update_populated(root_cgrp, true);
2046 spin_unlock_bh(&css_set_lock);
2048 BUG_ON(!list_empty(&root_cgrp->self.children));
2049 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2051 kernfs_activate(root_cgrp->kn);
2056 kernfs_destroy_root(root->kf_root);
2057 root->kf_root = NULL;
2059 cgroup_exit_root_id(root);
2061 percpu_ref_exit(&root_cgrp->self.refcnt);
2063 free_cgrp_cset_links(&tmp_links);
2067 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2068 int flags, const char *unused_dev_name,
2071 bool is_v2 = fs_type == &cgroup2_fs_type;
2072 struct super_block *pinned_sb = NULL;
2073 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2074 struct cgroup_subsys *ss;
2075 struct cgroup_root *root;
2076 struct cgroup_sb_opts opts;
2077 struct dentry *dentry;
2084 /* Check if the caller has permission to mount. */
2085 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2087 return ERR_PTR(-EPERM);
2091 * The first time anyone tries to mount a cgroup, enable the list
2092 * linking each css_set to its tasks and fix up all existing tasks.
2094 if (!use_task_css_set_links)
2095 cgroup_enable_task_cg_lists();
2099 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2101 return ERR_PTR(-EINVAL);
2103 cgrp_dfl_visible = true;
2104 root = &cgrp_dfl_root;
2105 cgroup_get(&root->cgrp);
2109 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2111 /* First find the desired set of subsystems */
2112 ret = parse_cgroupfs_options(data, &opts);
2117 * Destruction of cgroup root is asynchronous, so subsystems may
2118 * still be dying after the previous unmount. Let's drain the
2119 * dying subsystems. We just need to ensure that the ones
2120 * unmounted previously finish dying and don't care about new ones
2121 * starting. Testing ref liveliness is good enough.
2123 for_each_subsys(ss, i) {
2124 if (!(opts.subsys_mask & (1 << i)) ||
2125 ss->root == &cgrp_dfl_root)
2128 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2129 mutex_unlock(&cgroup_mutex);
2131 ret = restart_syscall();
2134 cgroup_put(&ss->root->cgrp);
2137 for_each_root(root) {
2138 bool name_match = false;
2140 if (root == &cgrp_dfl_root)
2144 * If we asked for a name then it must match. Also, if
2145 * name matches but sybsys_mask doesn't, we should fail.
2146 * Remember whether name matched.
2149 if (strcmp(opts.name, root->name))
2155 * If we asked for subsystems (or explicitly for no
2156 * subsystems) then they must match.
2158 if ((opts.subsys_mask || opts.none) &&
2159 (opts.subsys_mask != root->subsys_mask)) {
2166 if (root->flags ^ opts.flags)
2167 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2170 * We want to reuse @root whose lifetime is governed by its
2171 * ->cgrp. Let's check whether @root is alive and keep it
2172 * that way. As cgroup_kill_sb() can happen anytime, we
2173 * want to block it by pinning the sb so that @root doesn't
2174 * get killed before mount is complete.
2176 * With the sb pinned, tryget_live can reliably indicate
2177 * whether @root can be reused. If it's being killed,
2178 * drain it. We can use wait_queue for the wait but this
2179 * path is super cold. Let's just sleep a bit and retry.
2181 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2182 if (IS_ERR(pinned_sb) ||
2183 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2184 mutex_unlock(&cgroup_mutex);
2185 if (!IS_ERR_OR_NULL(pinned_sb))
2186 deactivate_super(pinned_sb);
2188 ret = restart_syscall();
2197 * No such thing, create a new one. name= matching without subsys
2198 * specification is allowed for already existing hierarchies but we
2199 * can't create new one without subsys specification.
2201 if (!opts.subsys_mask && !opts.none) {
2207 * We know this subsystem has not yet been bound. Users in a non-init
2208 * user namespace may only mount hierarchies with no bound subsystems,
2209 * i.e. 'none,name=user1'
2211 if (!opts.none && !capable(CAP_SYS_ADMIN)) {
2216 root = kzalloc(sizeof(*root), GFP_KERNEL);
2222 init_cgroup_root(root, &opts);
2224 ret = cgroup_setup_root(root, opts.subsys_mask);
2226 cgroup_free_root(root);
2229 mutex_unlock(&cgroup_mutex);
2231 kfree(opts.release_agent);
2236 return ERR_PTR(ret);
2239 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2240 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2244 * In non-init cgroup namespace, instead of root cgroup's
2245 * dentry, we return the dentry corresponding to the
2246 * cgroupns->root_cgrp.
2248 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2249 struct dentry *nsdentry;
2250 struct cgroup *cgrp;
2252 mutex_lock(&cgroup_mutex);
2253 spin_lock_bh(&css_set_lock);
2255 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2257 spin_unlock_bh(&css_set_lock);
2258 mutex_unlock(&cgroup_mutex);
2260 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
2265 if (IS_ERR(dentry) || !new_sb)
2266 cgroup_put(&root->cgrp);
2269 * If @pinned_sb, we're reusing an existing root and holding an
2270 * extra ref on its sb. Mount is complete. Put the extra ref.
2274 deactivate_super(pinned_sb);
2281 static void cgroup_kill_sb(struct super_block *sb)
2283 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2284 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2287 * If @root doesn't have any mounts or children, start killing it.
2288 * This prevents new mounts by disabling percpu_ref_tryget_live().
2289 * cgroup_mount() may wait for @root's release.
2291 * And don't kill the default root.
2293 if (!list_empty(&root->cgrp.self.children) ||
2294 root == &cgrp_dfl_root)
2295 cgroup_put(&root->cgrp);
2297 percpu_ref_kill(&root->cgrp.self.refcnt);
2302 static struct file_system_type cgroup_fs_type = {
2304 .mount = cgroup_mount,
2305 .kill_sb = cgroup_kill_sb,
2306 .fs_flags = FS_USERNS_MOUNT,
2309 static struct file_system_type cgroup2_fs_type = {
2311 .mount = cgroup_mount,
2312 .kill_sb = cgroup_kill_sb,
2313 .fs_flags = FS_USERNS_MOUNT,
2316 static char *cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2317 struct cgroup_namespace *ns)
2319 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2322 ret = kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2323 if (ret < 0 || ret >= buflen)
2328 char *cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2329 struct cgroup_namespace *ns)
2333 mutex_lock(&cgroup_mutex);
2334 spin_lock_bh(&css_set_lock);
2336 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2338 spin_unlock_bh(&css_set_lock);
2339 mutex_unlock(&cgroup_mutex);
2343 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2346 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2347 * @task: target task
2348 * @buf: the buffer to write the path into
2349 * @buflen: the length of the buffer
2351 * Determine @task's cgroup on the first (the one with the lowest non-zero
2352 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2353 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2354 * cgroup controller callbacks.
2356 * Return value is the same as kernfs_path().
2358 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2360 struct cgroup_root *root;
2361 struct cgroup *cgrp;
2362 int hierarchy_id = 1;
2365 mutex_lock(&cgroup_mutex);
2366 spin_lock_bh(&css_set_lock);
2368 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2371 cgrp = task_cgroup_from_root(task, root);
2372 path = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2374 /* if no hierarchy exists, everyone is in "/" */
2375 if (strlcpy(buf, "/", buflen) < buflen)
2379 spin_unlock_bh(&css_set_lock);
2380 mutex_unlock(&cgroup_mutex);
2383 EXPORT_SYMBOL_GPL(task_cgroup_path);
2385 /* used to track tasks and other necessary states during migration */
2386 struct cgroup_taskset {
2387 /* the src and dst cset list running through cset->mg_node */
2388 struct list_head src_csets;
2389 struct list_head dst_csets;
2391 /* the subsys currently being processed */
2395 * Fields for cgroup_taskset_*() iteration.
2397 * Before migration is committed, the target migration tasks are on
2398 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2399 * the csets on ->dst_csets. ->csets point to either ->src_csets
2400 * or ->dst_csets depending on whether migration is committed.
2402 * ->cur_csets and ->cur_task point to the current task position
2405 struct list_head *csets;
2406 struct css_set *cur_cset;
2407 struct task_struct *cur_task;
2410 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2411 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2412 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2413 .csets = &tset.src_csets, \
2417 * cgroup_taskset_add - try to add a migration target task to a taskset
2418 * @task: target task
2419 * @tset: target taskset
2421 * Add @task, which is a migration target, to @tset. This function becomes
2422 * noop if @task doesn't need to be migrated. @task's css_set should have
2423 * been added as a migration source and @task->cg_list will be moved from
2424 * the css_set's tasks list to mg_tasks one.
2426 static void cgroup_taskset_add(struct task_struct *task,
2427 struct cgroup_taskset *tset)
2429 struct css_set *cset;
2431 lockdep_assert_held(&css_set_lock);
2433 /* @task either already exited or can't exit until the end */
2434 if (task->flags & PF_EXITING)
2437 /* leave @task alone if post_fork() hasn't linked it yet */
2438 if (list_empty(&task->cg_list))
2441 cset = task_css_set(task);
2442 if (!cset->mg_src_cgrp)
2445 list_move_tail(&task->cg_list, &cset->mg_tasks);
2446 if (list_empty(&cset->mg_node))
2447 list_add_tail(&cset->mg_node, &tset->src_csets);
2448 if (list_empty(&cset->mg_dst_cset->mg_node))
2449 list_move_tail(&cset->mg_dst_cset->mg_node,
2454 * cgroup_taskset_first - reset taskset and return the first task
2455 * @tset: taskset of interest
2456 * @dst_cssp: output variable for the destination css
2458 * @tset iteration is initialized and the first task is returned.
2460 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2461 struct cgroup_subsys_state **dst_cssp)
2463 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2464 tset->cur_task = NULL;
2466 return cgroup_taskset_next(tset, dst_cssp);
2470 * cgroup_taskset_next - iterate to the next task in taskset
2471 * @tset: taskset of interest
2472 * @dst_cssp: output variable for the destination css
2474 * Return the next task in @tset. Iteration must have been initialized
2475 * with cgroup_taskset_first().
2477 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2478 struct cgroup_subsys_state **dst_cssp)
2480 struct css_set *cset = tset->cur_cset;
2481 struct task_struct *task = tset->cur_task;
2483 while (&cset->mg_node != tset->csets) {
2485 task = list_first_entry(&cset->mg_tasks,
2486 struct task_struct, cg_list);
2488 task = list_next_entry(task, cg_list);
2490 if (&task->cg_list != &cset->mg_tasks) {
2491 tset->cur_cset = cset;
2492 tset->cur_task = task;
2495 * This function may be called both before and
2496 * after cgroup_taskset_migrate(). The two cases
2497 * can be distinguished by looking at whether @cset
2498 * has its ->mg_dst_cset set.
2500 if (cset->mg_dst_cset)
2501 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2503 *dst_cssp = cset->subsys[tset->ssid];
2508 cset = list_next_entry(cset, mg_node);
2516 * cgroup_taskset_migrate - migrate a taskset
2517 * @tset: taget taskset
2518 * @root: cgroup root the migration is taking place on
2520 * Migrate tasks in @tset as setup by migration preparation functions.
2521 * This function fails iff one of the ->can_attach callbacks fails and
2522 * guarantees that either all or none of the tasks in @tset are migrated.
2523 * @tset is consumed regardless of success.
2525 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2526 struct cgroup_root *root)
2528 struct cgroup_subsys *ss;
2529 struct task_struct *task, *tmp_task;
2530 struct css_set *cset, *tmp_cset;
2531 int ssid, failed_ssid, ret;
2533 /* methods shouldn't be called if no task is actually migrating */
2534 if (list_empty(&tset->src_csets))
2537 /* check that we can legitimately attach to the cgroup */
2538 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2539 if (ss->can_attach) {
2541 ret = ss->can_attach(tset);
2544 goto out_cancel_attach;
2547 } while_each_subsys_mask();
2550 * Now that we're guaranteed success, proceed to move all tasks to
2551 * the new cgroup. There are no failure cases after here, so this
2552 * is the commit point.
2554 spin_lock_bh(&css_set_lock);
2555 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2556 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2557 struct css_set *from_cset = task_css_set(task);
2558 struct css_set *to_cset = cset->mg_dst_cset;
2560 get_css_set(to_cset);
2561 css_set_move_task(task, from_cset, to_cset, true);
2562 put_css_set_locked(from_cset);
2565 spin_unlock_bh(&css_set_lock);
2568 * Migration is committed, all target tasks are now on dst_csets.
2569 * Nothing is sensitive to fork() after this point. Notify
2570 * controllers that migration is complete.
2572 tset->csets = &tset->dst_csets;
2574 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2579 } while_each_subsys_mask();
2582 goto out_release_tset;
2585 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2586 if (ssid == failed_ssid)
2588 if (ss->cancel_attach) {
2590 ss->cancel_attach(tset);
2592 } while_each_subsys_mask();
2594 spin_lock_bh(&css_set_lock);
2595 list_splice_init(&tset->dst_csets, &tset->src_csets);
2596 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2597 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2598 list_del_init(&cset->mg_node);
2600 spin_unlock_bh(&css_set_lock);
2605 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2606 * @dst_cgrp: destination cgroup to test
2608 * On the default hierarchy, except for the root, subtree_control must be
2609 * zero for migration destination cgroups with tasks so that child cgroups
2610 * don't compete against tasks.
2612 static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2614 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2615 !dst_cgrp->subtree_control;
2619 * cgroup_migrate_finish - cleanup after attach
2620 * @preloaded_csets: list of preloaded css_sets
2622 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2623 * those functions for details.
2625 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2627 struct css_set *cset, *tmp_cset;
2629 lockdep_assert_held(&cgroup_mutex);
2631 spin_lock_bh(&css_set_lock);
2632 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2633 cset->mg_src_cgrp = NULL;
2634 cset->mg_dst_cgrp = NULL;
2635 cset->mg_dst_cset = NULL;
2636 list_del_init(&cset->mg_preload_node);
2637 put_css_set_locked(cset);
2639 spin_unlock_bh(&css_set_lock);
2643 * cgroup_migrate_add_src - add a migration source css_set
2644 * @src_cset: the source css_set to add
2645 * @dst_cgrp: the destination cgroup
2646 * @preloaded_csets: list of preloaded css_sets
2648 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2649 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2650 * up by cgroup_migrate_finish().
2652 * This function may be called without holding cgroup_threadgroup_rwsem
2653 * even if the target is a process. Threads may be created and destroyed
2654 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2655 * into play and the preloaded css_sets are guaranteed to cover all
2658 static void cgroup_migrate_add_src(struct css_set *src_cset,
2659 struct cgroup *dst_cgrp,
2660 struct list_head *preloaded_csets)
2662 struct cgroup *src_cgrp;
2664 lockdep_assert_held(&cgroup_mutex);
2665 lockdep_assert_held(&css_set_lock);
2668 * If ->dead, @src_set is associated with one or more dead cgroups
2669 * and doesn't contain any migratable tasks. Ignore it early so
2670 * that the rest of migration path doesn't get confused by it.
2675 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2677 if (!list_empty(&src_cset->mg_preload_node))
2680 WARN_ON(src_cset->mg_src_cgrp);
2681 WARN_ON(src_cset->mg_dst_cgrp);
2682 WARN_ON(!list_empty(&src_cset->mg_tasks));
2683 WARN_ON(!list_empty(&src_cset->mg_node));
2685 src_cset->mg_src_cgrp = src_cgrp;
2686 src_cset->mg_dst_cgrp = dst_cgrp;
2687 get_css_set(src_cset);
2688 list_add(&src_cset->mg_preload_node, preloaded_csets);
2692 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2693 * @preloaded_csets: list of preloaded source css_sets
2695 * Tasks are about to be moved and all the source css_sets have been
2696 * preloaded to @preloaded_csets. This function looks up and pins all
2697 * destination css_sets, links each to its source, and append them to
2700 * This function must be called after cgroup_migrate_add_src() has been
2701 * called on each migration source css_set. After migration is performed
2702 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2705 static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2708 struct css_set *src_cset, *tmp_cset;
2710 lockdep_assert_held(&cgroup_mutex);
2712 /* look up the dst cset for each src cset and link it to src */
2713 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2714 struct css_set *dst_cset;
2716 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2720 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2723 * If src cset equals dst, it's noop. Drop the src.
2724 * cgroup_migrate() will skip the cset too. Note that we
2725 * can't handle src == dst as some nodes are used by both.
2727 if (src_cset == dst_cset) {
2728 src_cset->mg_src_cgrp = NULL;
2729 src_cset->mg_dst_cgrp = NULL;
2730 list_del_init(&src_cset->mg_preload_node);
2731 put_css_set(src_cset);
2732 put_css_set(dst_cset);
2736 src_cset->mg_dst_cset = dst_cset;
2738 if (list_empty(&dst_cset->mg_preload_node))
2739 list_add(&dst_cset->mg_preload_node, &csets);
2741 put_css_set(dst_cset);
2744 list_splice_tail(&csets, preloaded_csets);
2747 cgroup_migrate_finish(&csets);
2752 * cgroup_migrate - migrate a process or task to a cgroup
2753 * @leader: the leader of the process or the task to migrate
2754 * @threadgroup: whether @leader points to the whole process or a single task
2755 * @root: cgroup root migration is taking place on
2757 * Migrate a process or task denoted by @leader. If migrating a process,
2758 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2759 * responsible for invoking cgroup_migrate_add_src() and
2760 * cgroup_migrate_prepare_dst() on the targets before invoking this
2761 * function and following up with cgroup_migrate_finish().
2763 * As long as a controller's ->can_attach() doesn't fail, this function is
2764 * guaranteed to succeed. This means that, excluding ->can_attach()
2765 * failure, when migrating multiple targets, the success or failure can be
2766 * decided for all targets by invoking group_migrate_prepare_dst() before
2767 * actually starting migrating.
2769 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2770 struct cgroup_root *root)
2772 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2773 struct task_struct *task;
2776 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2777 * already PF_EXITING could be freed from underneath us unless we
2778 * take an rcu_read_lock.
2780 spin_lock_bh(&css_set_lock);
2784 cgroup_taskset_add(task, &tset);
2787 } while_each_thread(leader, task);
2789 spin_unlock_bh(&css_set_lock);
2791 return cgroup_taskset_migrate(&tset, root);
2795 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2796 * @dst_cgrp: the cgroup to attach to
2797 * @leader: the task or the leader of the threadgroup to be attached
2798 * @threadgroup: attach the whole threadgroup?
2800 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2802 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2803 struct task_struct *leader, bool threadgroup)
2805 LIST_HEAD(preloaded_csets);
2806 struct task_struct *task;
2809 if (!cgroup_may_migrate_to(dst_cgrp))
2812 /* look up all src csets */
2813 spin_lock_bh(&css_set_lock);
2817 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2821 } while_each_thread(leader, task);
2823 spin_unlock_bh(&css_set_lock);
2825 /* prepare dst csets and commit */
2826 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2828 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2830 cgroup_migrate_finish(&preloaded_csets);
2834 static int cgroup_procs_write_permission(struct task_struct *task,
2835 struct cgroup *dst_cgrp,
2836 struct kernfs_open_file *of)
2838 const struct cred *cred = current_cred();
2839 const struct cred *tcred = get_task_cred(task);
2843 * even if we're attaching all tasks in the thread group, we only
2844 * need to check permissions on one of them.
2846 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2847 !uid_eq(cred->euid, tcred->uid) &&
2848 !uid_eq(cred->euid, tcred->suid))
2851 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2852 struct super_block *sb = of->file->f_path.dentry->d_sb;
2853 struct cgroup *cgrp;
2854 struct inode *inode;
2856 spin_lock_bh(&css_set_lock);
2857 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2858 spin_unlock_bh(&css_set_lock);
2860 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2861 cgrp = cgroup_parent(cgrp);
2864 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2866 ret = inode_permission(inode, MAY_WRITE);
2876 * Find the task_struct of the task to attach by vpid and pass it along to the
2877 * function to attach either it or all tasks in its threadgroup. Will lock
2878 * cgroup_mutex and threadgroup.
2880 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2881 size_t nbytes, loff_t off, bool threadgroup)
2883 struct task_struct *tsk;
2884 struct cgroup_subsys *ss;
2885 struct cgroup *cgrp;
2889 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2892 cgrp = cgroup_kn_lock_live(of->kn, false);
2896 percpu_down_write(&cgroup_threadgroup_rwsem);
2899 tsk = find_task_by_vpid(pid);
2902 goto out_unlock_rcu;
2909 tsk = tsk->group_leader;
2912 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2913 * trapped in a cpuset, or RT worker may be born in a cgroup
2914 * with no rt_runtime allocated. Just say no.
2916 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2918 goto out_unlock_rcu;
2921 get_task_struct(tsk);
2924 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2926 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2928 put_task_struct(tsk);
2929 goto out_unlock_threadgroup;
2933 out_unlock_threadgroup:
2934 percpu_up_write(&cgroup_threadgroup_rwsem);
2935 for_each_subsys(ss, ssid)
2936 if (ss->post_attach)
2938 cgroup_kn_unlock(of->kn);
2939 return ret ?: nbytes;
2943 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2944 * @from: attach to all cgroups of a given task
2945 * @tsk: the task to be attached
2947 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2949 struct cgroup_root *root;
2952 mutex_lock(&cgroup_mutex);
2953 for_each_root(root) {
2954 struct cgroup *from_cgrp;
2956 if (root == &cgrp_dfl_root)
2959 spin_lock_bh(&css_set_lock);
2960 from_cgrp = task_cgroup_from_root(from, root);
2961 spin_unlock_bh(&css_set_lock);
2963 retval = cgroup_attach_task(from_cgrp, tsk, false);
2967 mutex_unlock(&cgroup_mutex);
2971 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2973 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2974 char *buf, size_t nbytes, loff_t off)
2976 return __cgroup_procs_write(of, buf, nbytes, off, false);
2979 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2980 char *buf, size_t nbytes, loff_t off)
2982 return __cgroup_procs_write(of, buf, nbytes, off, true);
2985 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2986 char *buf, size_t nbytes, loff_t off)
2988 struct cgroup *cgrp;
2990 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2992 cgrp = cgroup_kn_lock_live(of->kn, false);
2995 spin_lock(&release_agent_path_lock);
2996 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2997 sizeof(cgrp->root->release_agent_path));
2998 spin_unlock(&release_agent_path_lock);
2999 cgroup_kn_unlock(of->kn);
3003 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
3005 struct cgroup *cgrp = seq_css(seq)->cgroup;
3007 spin_lock(&release_agent_path_lock);
3008 seq_puts(seq, cgrp->root->release_agent_path);
3009 spin_unlock(&release_agent_path_lock);
3010 seq_putc(seq, '\n');
3014 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
3016 seq_puts(seq, "0\n");
3020 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
3022 struct cgroup_subsys *ss;
3023 bool printed = false;
3026 do_each_subsys_mask(ss, ssid, ss_mask) {
3029 seq_printf(seq, "%s", ss->name);
3031 } while_each_subsys_mask();
3033 seq_putc(seq, '\n');
3036 /* show controllers which are enabled from the parent */
3037 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3039 struct cgroup *cgrp = seq_css(seq)->cgroup;
3041 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3045 /* show controllers which are enabled for a given cgroup's children */
3046 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3048 struct cgroup *cgrp = seq_css(seq)->cgroup;
3050 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3055 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3056 * @cgrp: root of the subtree to update csses for
3058 * @cgrp's control masks have changed and its subtree's css associations
3059 * need to be updated accordingly. This function looks up all css_sets
3060 * which are attached to the subtree, creates the matching updated css_sets
3061 * and migrates the tasks to the new ones.
3063 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3065 LIST_HEAD(preloaded_csets);
3066 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
3067 struct cgroup_subsys_state *d_css;
3068 struct cgroup *dsct;
3069 struct css_set *src_cset;
3072 lockdep_assert_held(&cgroup_mutex);
3074 percpu_down_write(&cgroup_threadgroup_rwsem);
3076 /* look up all csses currently attached to @cgrp's subtree */
3077 spin_lock_bh(&css_set_lock);
3078 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3079 struct cgrp_cset_link *link;
3081 list_for_each_entry(link, &dsct->cset_links, cset_link)
3082 cgroup_migrate_add_src(link->cset, dsct,
3085 spin_unlock_bh(&css_set_lock);
3087 /* NULL dst indicates self on default hierarchy */
3088 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
3092 spin_lock_bh(&css_set_lock);
3093 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
3094 struct task_struct *task, *ntask;
3096 /* src_csets precede dst_csets, break on the first dst_cset */
3097 if (!src_cset->mg_src_cgrp)
3100 /* all tasks in src_csets need to be migrated */
3101 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3102 cgroup_taskset_add(task, &tset);
3104 spin_unlock_bh(&css_set_lock);
3106 ret = cgroup_taskset_migrate(&tset, cgrp->root);
3108 cgroup_migrate_finish(&preloaded_csets);
3109 percpu_up_write(&cgroup_threadgroup_rwsem);
3114 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3115 * @cgrp: root of the target subtree
3117 * Because css offlining is asynchronous, userland may try to re-enable a
3118 * controller while the previous css is still around. This function grabs
3119 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3121 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3122 __acquires(&cgroup_mutex)
3124 struct cgroup *dsct;
3125 struct cgroup_subsys_state *d_css;
3126 struct cgroup_subsys *ss;
3130 mutex_lock(&cgroup_mutex);
3132 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3133 for_each_subsys(ss, ssid) {
3134 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3137 if (!css || !percpu_ref_is_dying(&css->refcnt))
3141 prepare_to_wait(&dsct->offline_waitq, &wait,
3142 TASK_UNINTERRUPTIBLE);
3144 mutex_unlock(&cgroup_mutex);
3146 finish_wait(&dsct->offline_waitq, &wait);
3155 * cgroup_save_control - save control masks of a subtree
3156 * @cgrp: root of the target subtree
3158 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3159 * prefixed fields for @cgrp's subtree including @cgrp itself.
3161 static void cgroup_save_control(struct cgroup *cgrp)
3163 struct cgroup *dsct;
3164 struct cgroup_subsys_state *d_css;
3166 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3167 dsct->old_subtree_control = dsct->subtree_control;
3168 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3173 * cgroup_propagate_control - refresh control masks of a subtree
3174 * @cgrp: root of the target subtree
3176 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3177 * ->subtree_control and propagate controller availability through the
3178 * subtree so that descendants don't have unavailable controllers enabled.
3180 static void cgroup_propagate_control(struct cgroup *cgrp)
3182 struct cgroup *dsct;
3183 struct cgroup_subsys_state *d_css;
3185 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3186 dsct->subtree_control &= cgroup_control(dsct);
3187 dsct->subtree_ss_mask =
3188 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3189 cgroup_ss_mask(dsct));
3194 * cgroup_restore_control - restore control masks of a subtree
3195 * @cgrp: root of the target subtree
3197 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3198 * prefixed fields for @cgrp's subtree including @cgrp itself.
3200 static void cgroup_restore_control(struct cgroup *cgrp)
3202 struct cgroup *dsct;
3203 struct cgroup_subsys_state *d_css;
3205 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3206 dsct->subtree_control = dsct->old_subtree_control;
3207 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3211 static bool css_visible(struct cgroup_subsys_state *css)
3213 struct cgroup_subsys *ss = css->ss;
3214 struct cgroup *cgrp = css->cgroup;
3216 if (cgroup_control(cgrp) & (1 << ss->id))
3218 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3220 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3224 * cgroup_apply_control_enable - enable or show csses according to control
3225 * @cgrp: root of the target subtree
3227 * Walk @cgrp's subtree and create new csses or make the existing ones
3228 * visible. A css is created invisible if it's being implicitly enabled
3229 * through dependency. An invisible css is made visible when the userland
3230 * explicitly enables it.
3232 * Returns 0 on success, -errno on failure. On failure, csses which have
3233 * been processed already aren't cleaned up. The caller is responsible for
3234 * cleaning up with cgroup_apply_control_disble().
3236 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3238 struct cgroup *dsct;
3239 struct cgroup_subsys_state *d_css;
3240 struct cgroup_subsys *ss;
3243 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3244 for_each_subsys(ss, ssid) {
3245 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3247 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3249 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3253 css = css_create(dsct, ss);
3255 return PTR_ERR(css);
3258 if (css_visible(css)) {
3259 ret = css_populate_dir(css);
3270 * cgroup_apply_control_disable - kill or hide csses according to control
3271 * @cgrp: root of the target subtree
3273 * Walk @cgrp's subtree and kill and hide csses so that they match
3274 * cgroup_ss_mask() and cgroup_visible_mask().
3276 * A css is hidden when the userland requests it to be disabled while other
3277 * subsystems are still depending on it. The css must not actively control
3278 * resources and be in the vanilla state if it's made visible again later.
3279 * Controllers which may be depended upon should provide ->css_reset() for
3282 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3284 struct cgroup *dsct;
3285 struct cgroup_subsys_state *d_css;
3286 struct cgroup_subsys *ss;
3289 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3290 for_each_subsys(ss, ssid) {
3291 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3293 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3299 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3301 } else if (!css_visible(css)) {
3311 * cgroup_apply_control - apply control mask updates to the subtree
3312 * @cgrp: root of the target subtree
3314 * subsystems can be enabled and disabled in a subtree using the following
3317 * 1. Call cgroup_save_control() to stash the current state.
3318 * 2. Update ->subtree_control masks in the subtree as desired.
3319 * 3. Call cgroup_apply_control() to apply the changes.
3320 * 4. Optionally perform other related operations.
3321 * 5. Call cgroup_finalize_control() to finish up.
3323 * This function implements step 3 and propagates the mask changes
3324 * throughout @cgrp's subtree, updates csses accordingly and perform
3325 * process migrations.
3327 static int cgroup_apply_control(struct cgroup *cgrp)
3331 cgroup_propagate_control(cgrp);
3333 ret = cgroup_apply_control_enable(cgrp);
3338 * At this point, cgroup_e_css() results reflect the new csses
3339 * making the following cgroup_update_dfl_csses() properly update
3340 * css associations of all tasks in the subtree.
3342 ret = cgroup_update_dfl_csses(cgrp);
3350 * cgroup_finalize_control - finalize control mask update
3351 * @cgrp: root of the target subtree
3352 * @ret: the result of the update
3354 * Finalize control mask update. See cgroup_apply_control() for more info.
3356 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3359 cgroup_restore_control(cgrp);
3360 cgroup_propagate_control(cgrp);
3363 cgroup_apply_control_disable(cgrp);
3366 /* change the enabled child controllers for a cgroup in the default hierarchy */
3367 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3368 char *buf, size_t nbytes,
3371 u16 enable = 0, disable = 0;
3372 struct cgroup *cgrp, *child;
3373 struct cgroup_subsys *ss;
3378 * Parse input - space separated list of subsystem names prefixed
3379 * with either + or -.
3381 buf = strstrip(buf);
3382 while ((tok = strsep(&buf, " "))) {
3385 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3386 if (!cgroup_ssid_enabled(ssid) ||
3387 strcmp(tok + 1, ss->name))
3391 enable |= 1 << ssid;
3392 disable &= ~(1 << ssid);
3393 } else if (*tok == '-') {
3394 disable |= 1 << ssid;
3395 enable &= ~(1 << ssid);
3400 } while_each_subsys_mask();
3401 if (ssid == CGROUP_SUBSYS_COUNT)
3405 cgrp = cgroup_kn_lock_live(of->kn, true);
3409 for_each_subsys(ss, ssid) {
3410 if (enable & (1 << ssid)) {
3411 if (cgrp->subtree_control & (1 << ssid)) {
3412 enable &= ~(1 << ssid);
3416 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3420 } else if (disable & (1 << ssid)) {
3421 if (!(cgrp->subtree_control & (1 << ssid))) {
3422 disable &= ~(1 << ssid);
3426 /* a child has it enabled? */
3427 cgroup_for_each_live_child(child, cgrp) {
3428 if (child->subtree_control & (1 << ssid)) {
3436 if (!enable && !disable) {
3442 * Except for the root, subtree_control must be zero for a cgroup
3443 * with tasks so that child cgroups don't compete against tasks.
3445 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3450 /* save and update control masks and prepare csses */
3451 cgroup_save_control(cgrp);
3453 cgrp->subtree_control |= enable;
3454 cgrp->subtree_control &= ~disable;
3456 ret = cgroup_apply_control(cgrp);
3458 cgroup_finalize_control(cgrp, ret);
3460 kernfs_activate(cgrp->kn);
3463 cgroup_kn_unlock(of->kn);
3464 return ret ?: nbytes;
3467 static int cgroup_events_show(struct seq_file *seq, void *v)
3469 seq_printf(seq, "populated %d\n",
3470 cgroup_is_populated(seq_css(seq)->cgroup));
3474 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3475 size_t nbytes, loff_t off)
3477 struct cgroup *cgrp = of->kn->parent->priv;
3478 struct cftype *cft = of->kn->priv;
3479 struct cgroup_subsys_state *css;
3483 return cft->write(of, buf, nbytes, off);
3486 * kernfs guarantees that a file isn't deleted with operations in
3487 * flight, which means that the matching css is and stays alive and
3488 * doesn't need to be pinned. The RCU locking is not necessary
3489 * either. It's just for the convenience of using cgroup_css().
3492 css = cgroup_css(cgrp, cft->ss);
3495 if (cft->write_u64) {
3496 unsigned long long v;
3497 ret = kstrtoull(buf, 0, &v);
3499 ret = cft->write_u64(css, cft, v);
3500 } else if (cft->write_s64) {
3502 ret = kstrtoll(buf, 0, &v);
3504 ret = cft->write_s64(css, cft, v);
3509 return ret ?: nbytes;
3512 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3514 return seq_cft(seq)->seq_start(seq, ppos);
3517 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3519 return seq_cft(seq)->seq_next(seq, v, ppos);
3522 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3524 seq_cft(seq)->seq_stop(seq, v);
3527 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3529 struct cftype *cft = seq_cft(m);
3530 struct cgroup_subsys_state *css = seq_css(m);
3533 return cft->seq_show(m, arg);
3536 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3537 else if (cft->read_s64)
3538 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3544 static struct kernfs_ops cgroup_kf_single_ops = {
3545 .atomic_write_len = PAGE_SIZE,
3546 .write = cgroup_file_write,
3547 .seq_show = cgroup_seqfile_show,
3550 static struct kernfs_ops cgroup_kf_ops = {
3551 .atomic_write_len = PAGE_SIZE,
3552 .write = cgroup_file_write,
3553 .seq_start = cgroup_seqfile_start,
3554 .seq_next = cgroup_seqfile_next,
3555 .seq_stop = cgroup_seqfile_stop,
3556 .seq_show = cgroup_seqfile_show,
3560 * cgroup_rename - Only allow simple rename of directories in place.
3562 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3563 const char *new_name_str)
3565 struct cgroup *cgrp = kn->priv;
3568 if (kernfs_type(kn) != KERNFS_DIR)
3570 if (kn->parent != new_parent)
3574 * This isn't a proper migration and its usefulness is very
3575 * limited. Disallow on the default hierarchy.
3577 if (cgroup_on_dfl(cgrp))
3581 * We're gonna grab cgroup_mutex which nests outside kernfs
3582 * active_ref. kernfs_rename() doesn't require active_ref
3583 * protection. Break them before grabbing cgroup_mutex.
3585 kernfs_break_active_protection(new_parent);
3586 kernfs_break_active_protection(kn);
3588 mutex_lock(&cgroup_mutex);
3590 ret = kernfs_rename(kn, new_parent, new_name_str);
3592 mutex_unlock(&cgroup_mutex);
3594 kernfs_unbreak_active_protection(kn);
3595 kernfs_unbreak_active_protection(new_parent);
3599 /* set uid and gid of cgroup dirs and files to that of the creator */
3600 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3602 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3603 .ia_uid = current_fsuid(),
3604 .ia_gid = current_fsgid(), };
3606 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3607 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3610 return kernfs_setattr(kn, &iattr);
3613 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3616 char name[CGROUP_FILE_NAME_MAX];
3617 struct kernfs_node *kn;
3618 struct lock_class_key *key = NULL;
3621 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3622 key = &cft->lockdep_key;
3624 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3625 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3630 ret = cgroup_kn_set_ugid(kn);
3636 if (cft->file_offset) {
3637 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3639 spin_lock_irq(&cgroup_file_kn_lock);
3641 spin_unlock_irq(&cgroup_file_kn_lock);
3648 * cgroup_addrm_files - add or remove files to a cgroup directory
3649 * @css: the target css
3650 * @cgrp: the target cgroup (usually css->cgroup)
3651 * @cfts: array of cftypes to be added
3652 * @is_add: whether to add or remove
3654 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3655 * For removals, this function never fails.
3657 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3658 struct cgroup *cgrp, struct cftype cfts[],
3661 struct cftype *cft, *cft_end = NULL;
3664 lockdep_assert_held(&cgroup_mutex);
3667 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3668 /* does cft->flags tell us to skip this file on @cgrp? */
3669 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3671 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3673 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3675 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3679 ret = cgroup_add_file(css, cgrp, cft);
3681 pr_warn("%s: failed to add %s, err=%d\n",
3682 __func__, cft->name, ret);
3688 cgroup_rm_file(cgrp, cft);
3694 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3697 struct cgroup_subsys *ss = cfts[0].ss;
3698 struct cgroup *root = &ss->root->cgrp;
3699 struct cgroup_subsys_state *css;
3702 lockdep_assert_held(&cgroup_mutex);
3704 /* add/rm files for all cgroups created before */
3705 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3706 struct cgroup *cgrp = css->cgroup;
3708 if (!(css->flags & CSS_VISIBLE))
3711 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3717 kernfs_activate(root->kn);
3721 static void cgroup_exit_cftypes(struct cftype *cfts)
3725 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3726 /* free copy for custom atomic_write_len, see init_cftypes() */
3727 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3732 /* revert flags set by cgroup core while adding @cfts */
3733 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3737 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3741 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3742 struct kernfs_ops *kf_ops;
3744 WARN_ON(cft->ss || cft->kf_ops);
3747 kf_ops = &cgroup_kf_ops;
3749 kf_ops = &cgroup_kf_single_ops;
3752 * Ugh... if @cft wants a custom max_write_len, we need to
3753 * make a copy of kf_ops to set its atomic_write_len.
3755 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3756 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3758 cgroup_exit_cftypes(cfts);
3761 kf_ops->atomic_write_len = cft->max_write_len;
3764 cft->kf_ops = kf_ops;
3771 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3773 lockdep_assert_held(&cgroup_mutex);
3775 if (!cfts || !cfts[0].ss)
3778 list_del(&cfts->node);
3779 cgroup_apply_cftypes(cfts, false);
3780 cgroup_exit_cftypes(cfts);
3785 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3786 * @cfts: zero-length name terminated array of cftypes
3788 * Unregister @cfts. Files described by @cfts are removed from all
3789 * existing cgroups and all future cgroups won't have them either. This
3790 * function can be called anytime whether @cfts' subsys is attached or not.
3792 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3795 int cgroup_rm_cftypes(struct cftype *cfts)
3799 mutex_lock(&cgroup_mutex);
3800 ret = cgroup_rm_cftypes_locked(cfts);
3801 mutex_unlock(&cgroup_mutex);
3806 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3807 * @ss: target cgroup subsystem
3808 * @cfts: zero-length name terminated array of cftypes
3810 * Register @cfts to @ss. Files described by @cfts are created for all
3811 * existing cgroups to which @ss is attached and all future cgroups will
3812 * have them too. This function can be called anytime whether @ss is
3815 * Returns 0 on successful registration, -errno on failure. Note that this
3816 * function currently returns 0 as long as @cfts registration is successful
3817 * even if some file creation attempts on existing cgroups fail.
3819 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3823 if (!cgroup_ssid_enabled(ss->id))
3826 if (!cfts || cfts[0].name[0] == '\0')
3829 ret = cgroup_init_cftypes(ss, cfts);
3833 mutex_lock(&cgroup_mutex);
3835 list_add_tail(&cfts->node, &ss->cfts);
3836 ret = cgroup_apply_cftypes(cfts, true);
3838 cgroup_rm_cftypes_locked(cfts);
3840 mutex_unlock(&cgroup_mutex);
3845 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3846 * @ss: target cgroup subsystem
3847 * @cfts: zero-length name terminated array of cftypes
3849 * Similar to cgroup_add_cftypes() but the added files are only used for
3850 * the default hierarchy.
3852 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3856 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3857 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3858 return cgroup_add_cftypes(ss, cfts);
3862 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3863 * @ss: target cgroup subsystem
3864 * @cfts: zero-length name terminated array of cftypes
3866 * Similar to cgroup_add_cftypes() but the added files are only used for
3867 * the legacy hierarchies.
3869 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3873 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3874 cft->flags |= __CFTYPE_NOT_ON_DFL;
3875 return cgroup_add_cftypes(ss, cfts);
3879 * cgroup_file_notify - generate a file modified event for a cgroup_file
3880 * @cfile: target cgroup_file
3882 * @cfile must have been obtained by setting cftype->file_offset.
3884 void cgroup_file_notify(struct cgroup_file *cfile)
3886 unsigned long flags;
3888 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3890 kernfs_notify(cfile->kn);
3891 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3895 * cgroup_task_count - count the number of tasks in a cgroup.
3896 * @cgrp: the cgroup in question
3898 * Return the number of tasks in the cgroup.
3900 static int cgroup_task_count(const struct cgroup *cgrp)
3903 struct cgrp_cset_link *link;
3905 spin_lock_bh(&css_set_lock);
3906 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3907 count += atomic_read(&link->cset->refcount);
3908 spin_unlock_bh(&css_set_lock);
3913 * css_next_child - find the next child of a given css
3914 * @pos: the current position (%NULL to initiate traversal)
3915 * @parent: css whose children to walk
3917 * This function returns the next child of @parent and should be called
3918 * under either cgroup_mutex or RCU read lock. The only requirement is
3919 * that @parent and @pos are accessible. The next sibling is guaranteed to
3920 * be returned regardless of their states.
3922 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3923 * css which finished ->css_online() is guaranteed to be visible in the
3924 * future iterations and will stay visible until the last reference is put.
3925 * A css which hasn't finished ->css_online() or already finished
3926 * ->css_offline() may show up during traversal. It's each subsystem's
3927 * responsibility to synchronize against on/offlining.
3929 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3930 struct cgroup_subsys_state *parent)
3932 struct cgroup_subsys_state *next;
3934 cgroup_assert_mutex_or_rcu_locked();
3937 * @pos could already have been unlinked from the sibling list.
3938 * Once a cgroup is removed, its ->sibling.next is no longer
3939 * updated when its next sibling changes. CSS_RELEASED is set when
3940 * @pos is taken off list, at which time its next pointer is valid,
3941 * and, as releases are serialized, the one pointed to by the next
3942 * pointer is guaranteed to not have started release yet. This
3943 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3944 * critical section, the one pointed to by its next pointer is
3945 * guaranteed to not have finished its RCU grace period even if we
3946 * have dropped rcu_read_lock() inbetween iterations.
3948 * If @pos has CSS_RELEASED set, its next pointer can't be
3949 * dereferenced; however, as each css is given a monotonically
3950 * increasing unique serial number and always appended to the
3951 * sibling list, the next one can be found by walking the parent's
3952 * children until the first css with higher serial number than
3953 * @pos's. While this path can be slower, it happens iff iteration
3954 * races against release and the race window is very small.
3957 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3958 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3959 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3961 list_for_each_entry_rcu(next, &parent->children, sibling)
3962 if (next->serial_nr > pos->serial_nr)
3967 * @next, if not pointing to the head, can be dereferenced and is
3970 if (&next->sibling != &parent->children)
3976 * css_next_descendant_pre - find the next descendant for pre-order walk
3977 * @pos: the current position (%NULL to initiate traversal)
3978 * @root: css whose descendants to walk
3980 * To be used by css_for_each_descendant_pre(). Find the next descendant
3981 * to visit for pre-order traversal of @root's descendants. @root is
3982 * included in the iteration and the first node to be visited.
3984 * While this function requires cgroup_mutex or RCU read locking, it
3985 * doesn't require the whole traversal to be contained in a single critical
3986 * section. This function will return the correct next descendant as long
3987 * as both @pos and @root are accessible and @pos is a descendant of @root.
3989 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3990 * css which finished ->css_online() is guaranteed to be visible in the
3991 * future iterations and will stay visible until the last reference is put.
3992 * A css which hasn't finished ->css_online() or already finished
3993 * ->css_offline() may show up during traversal. It's each subsystem's
3994 * responsibility to synchronize against on/offlining.
3996 struct cgroup_subsys_state *
3997 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3998 struct cgroup_subsys_state *root)
4000 struct cgroup_subsys_state *next;
4002 cgroup_assert_mutex_or_rcu_locked();
4004 /* if first iteration, visit @root */
4008 /* visit the first child if exists */
4009 next = css_next_child(NULL, pos);
4013 /* no child, visit my or the closest ancestor's next sibling */
4014 while (pos != root) {
4015 next = css_next_child(pos, pos->parent);
4025 * css_rightmost_descendant - return the rightmost descendant of a css
4026 * @pos: css of interest
4028 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4029 * is returned. This can be used during pre-order traversal to skip
4032 * While this function requires cgroup_mutex or RCU read locking, it
4033 * doesn't require the whole traversal to be contained in a single critical
4034 * section. This function will return the correct rightmost descendant as
4035 * long as @pos is accessible.
4037 struct cgroup_subsys_state *
4038 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4040 struct cgroup_subsys_state *last, *tmp;
4042 cgroup_assert_mutex_or_rcu_locked();
4046 /* ->prev isn't RCU safe, walk ->next till the end */
4048 css_for_each_child(tmp, last)
4055 static struct cgroup_subsys_state *
4056 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4058 struct cgroup_subsys_state *last;
4062 pos = css_next_child(NULL, pos);
4069 * css_next_descendant_post - find the next descendant for post-order walk
4070 * @pos: the current position (%NULL to initiate traversal)
4071 * @root: css whose descendants to walk
4073 * To be used by css_for_each_descendant_post(). Find the next descendant
4074 * to visit for post-order traversal of @root's descendants. @root is
4075 * included in the iteration and the last node to be visited.
4077 * While this function requires cgroup_mutex or RCU read locking, it
4078 * doesn't require the whole traversal to be contained in a single critical
4079 * section. This function will return the correct next descendant as long
4080 * as both @pos and @cgroup are accessible and @pos is a descendant of
4083 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4084 * css which finished ->css_online() is guaranteed to be visible in the
4085 * future iterations and will stay visible until the last reference is put.
4086 * A css which hasn't finished ->css_online() or already finished
4087 * ->css_offline() may show up during traversal. It's each subsystem's
4088 * responsibility to synchronize against on/offlining.
4090 struct cgroup_subsys_state *
4091 css_next_descendant_post(struct cgroup_subsys_state *pos,
4092 struct cgroup_subsys_state *root)
4094 struct cgroup_subsys_state *next;
4096 cgroup_assert_mutex_or_rcu_locked();
4098 /* if first iteration, visit leftmost descendant which may be @root */
4100 return css_leftmost_descendant(root);
4102 /* if we visited @root, we're done */
4106 /* if there's an unvisited sibling, visit its leftmost descendant */
4107 next = css_next_child(pos, pos->parent);
4109 return css_leftmost_descendant(next);
4111 /* no sibling left, visit parent */
4116 * css_has_online_children - does a css have online children
4117 * @css: the target css
4119 * Returns %true if @css has any online children; otherwise, %false. This
4120 * function can be called from any context but the caller is responsible
4121 * for synchronizing against on/offlining as necessary.
4123 bool css_has_online_children(struct cgroup_subsys_state *css)
4125 struct cgroup_subsys_state *child;
4129 css_for_each_child(child, css) {
4130 if (child->flags & CSS_ONLINE) {
4140 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4141 * @it: the iterator to advance
4143 * Advance @it to the next css_set to walk.
4145 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4147 struct list_head *l = it->cset_pos;
4148 struct cgrp_cset_link *link;
4149 struct css_set *cset;
4151 lockdep_assert_held(&css_set_lock);
4153 /* Advance to the next non-empty css_set */
4156 if (l == it->cset_head) {
4157 it->cset_pos = NULL;
4158 it->task_pos = NULL;
4163 cset = container_of(l, struct css_set,
4164 e_cset_node[it->ss->id]);
4166 link = list_entry(l, struct cgrp_cset_link, cset_link);
4169 } while (!css_set_populated(cset));
4173 if (!list_empty(&cset->tasks))
4174 it->task_pos = cset->tasks.next;
4176 it->task_pos = cset->mg_tasks.next;
4178 it->tasks_head = &cset->tasks;
4179 it->mg_tasks_head = &cset->mg_tasks;
4182 * We don't keep css_sets locked across iteration steps and thus
4183 * need to take steps to ensure that iteration can be resumed after
4184 * the lock is re-acquired. Iteration is performed at two levels -
4185 * css_sets and tasks in them.
4187 * Once created, a css_set never leaves its cgroup lists, so a
4188 * pinned css_set is guaranteed to stay put and we can resume
4189 * iteration afterwards.
4191 * Tasks may leave @cset across iteration steps. This is resolved
4192 * by registering each iterator with the css_set currently being
4193 * walked and making css_set_move_task() advance iterators whose
4194 * next task is leaving.
4197 list_del(&it->iters_node);
4198 put_css_set_locked(it->cur_cset);
4201 it->cur_cset = cset;
4202 list_add(&it->iters_node, &cset->task_iters);
4205 static void css_task_iter_advance(struct css_task_iter *it)
4207 struct list_head *l = it->task_pos;
4209 lockdep_assert_held(&css_set_lock);
4213 * Advance iterator to find next entry. cset->tasks is consumed
4214 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4219 if (l == it->tasks_head)
4220 l = it->mg_tasks_head->next;
4222 if (l == it->mg_tasks_head)
4223 css_task_iter_advance_css_set(it);
4229 * css_task_iter_start - initiate task iteration
4230 * @css: the css to walk tasks of
4231 * @it: the task iterator to use
4233 * Initiate iteration through the tasks of @css. The caller can call
4234 * css_task_iter_next() to walk through the tasks until the function
4235 * returns NULL. On completion of iteration, css_task_iter_end() must be
4238 void css_task_iter_start(struct cgroup_subsys_state *css,
4239 struct css_task_iter *it)
4241 /* no one should try to iterate before mounting cgroups */
4242 WARN_ON_ONCE(!use_task_css_set_links);
4244 memset(it, 0, sizeof(*it));
4246 spin_lock_bh(&css_set_lock);
4251 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4253 it->cset_pos = &css->cgroup->cset_links;
4255 it->cset_head = it->cset_pos;
4257 css_task_iter_advance_css_set(it);
4259 spin_unlock_bh(&css_set_lock);
4263 * css_task_iter_next - return the next task for the iterator
4264 * @it: the task iterator being iterated
4266 * The "next" function for task iteration. @it should have been
4267 * initialized via css_task_iter_start(). Returns NULL when the iteration
4270 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4273 put_task_struct(it->cur_task);
4274 it->cur_task = NULL;
4277 spin_lock_bh(&css_set_lock);
4280 it->cur_task = list_entry(it->task_pos, struct task_struct,
4282 get_task_struct(it->cur_task);
4283 css_task_iter_advance(it);
4286 spin_unlock_bh(&css_set_lock);
4288 return it->cur_task;
4292 * css_task_iter_end - finish task iteration
4293 * @it: the task iterator to finish
4295 * Finish task iteration started by css_task_iter_start().
4297 void css_task_iter_end(struct css_task_iter *it)
4300 spin_lock_bh(&css_set_lock);
4301 list_del(&it->iters_node);
4302 put_css_set_locked(it->cur_cset);
4303 spin_unlock_bh(&css_set_lock);
4307 put_task_struct(it->cur_task);
4311 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4312 * @to: cgroup to which the tasks will be moved
4313 * @from: cgroup in which the tasks currently reside
4315 * Locking rules between cgroup_post_fork() and the migration path
4316 * guarantee that, if a task is forking while being migrated, the new child
4317 * is guaranteed to be either visible in the source cgroup after the
4318 * parent's migration is complete or put into the target cgroup. No task
4319 * can slip out of migration through forking.
4321 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4323 LIST_HEAD(preloaded_csets);
4324 struct cgrp_cset_link *link;
4325 struct css_task_iter it;
4326 struct task_struct *task;
4329 if (!cgroup_may_migrate_to(to))
4332 mutex_lock(&cgroup_mutex);
4334 /* all tasks in @from are being moved, all csets are source */
4335 spin_lock_bh(&css_set_lock);
4336 list_for_each_entry(link, &from->cset_links, cset_link)
4337 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4338 spin_unlock_bh(&css_set_lock);
4340 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4345 * Migrate tasks one-by-one until @from is empty. This fails iff
4346 * ->can_attach() fails.
4349 css_task_iter_start(&from->self, &it);
4350 task = css_task_iter_next(&it);
4352 get_task_struct(task);
4353 css_task_iter_end(&it);
4356 ret = cgroup_migrate(task, false, to->root);
4357 put_task_struct(task);
4359 } while (task && !ret);
4361 cgroup_migrate_finish(&preloaded_csets);
4362 mutex_unlock(&cgroup_mutex);
4367 * Stuff for reading the 'tasks'/'procs' files.
4369 * Reading this file can return large amounts of data if a cgroup has
4370 * *lots* of attached tasks. So it may need several calls to read(),
4371 * but we cannot guarantee that the information we produce is correct
4372 * unless we produce it entirely atomically.
4376 /* which pidlist file are we talking about? */
4377 enum cgroup_filetype {
4383 * A pidlist is a list of pids that virtually represents the contents of one
4384 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4385 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4388 struct cgroup_pidlist {
4390 * used to find which pidlist is wanted. doesn't change as long as
4391 * this particular list stays in the list.
4393 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4396 /* how many elements the above list has */
4398 /* each of these stored in a list by its cgroup */
4399 struct list_head links;
4400 /* pointer to the cgroup we belong to, for list removal purposes */
4401 struct cgroup *owner;
4402 /* for delayed destruction */
4403 struct delayed_work destroy_dwork;
4407 * The following two functions "fix" the issue where there are more pids
4408 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4409 * TODO: replace with a kernel-wide solution to this problem
4411 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4412 static void *pidlist_allocate(int count)
4414 if (PIDLIST_TOO_LARGE(count))
4415 return vmalloc(count * sizeof(pid_t));
4417 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4420 static void pidlist_free(void *p)
4426 * Used to destroy all pidlists lingering waiting for destroy timer. None
4427 * should be left afterwards.
4429 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4431 struct cgroup_pidlist *l, *tmp_l;
4433 mutex_lock(&cgrp->pidlist_mutex);
4434 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4435 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4436 mutex_unlock(&cgrp->pidlist_mutex);
4438 flush_workqueue(cgroup_pidlist_destroy_wq);
4439 BUG_ON(!list_empty(&cgrp->pidlists));
4442 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4444 struct delayed_work *dwork = to_delayed_work(work);
4445 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4447 struct cgroup_pidlist *tofree = NULL;
4449 mutex_lock(&l->owner->pidlist_mutex);
4452 * Destroy iff we didn't get queued again. The state won't change
4453 * as destroy_dwork can only be queued while locked.
4455 if (!delayed_work_pending(dwork)) {
4456 list_del(&l->links);
4457 pidlist_free(l->list);
4458 put_pid_ns(l->key.ns);
4462 mutex_unlock(&l->owner->pidlist_mutex);
4467 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4468 * Returns the number of unique elements.
4470 static int pidlist_uniq(pid_t *list, int length)
4475 * we presume the 0th element is unique, so i starts at 1. trivial
4476 * edge cases first; no work needs to be done for either
4478 if (length == 0 || length == 1)
4480 /* src and dest walk down the list; dest counts unique elements */
4481 for (src = 1; src < length; src++) {
4482 /* find next unique element */
4483 while (list[src] == list[src-1]) {
4488 /* dest always points to where the next unique element goes */
4489 list[dest] = list[src];
4497 * The two pid files - task and cgroup.procs - guaranteed that the result
4498 * is sorted, which forced this whole pidlist fiasco. As pid order is
4499 * different per namespace, each namespace needs differently sorted list,
4500 * making it impossible to use, for example, single rbtree of member tasks
4501 * sorted by task pointer. As pidlists can be fairly large, allocating one
4502 * per open file is dangerous, so cgroup had to implement shared pool of
4503 * pidlists keyed by cgroup and namespace.
4505 * All this extra complexity was caused by the original implementation
4506 * committing to an entirely unnecessary property. In the long term, we
4507 * want to do away with it. Explicitly scramble sort order if on the
4508 * default hierarchy so that no such expectation exists in the new
4511 * Scrambling is done by swapping every two consecutive bits, which is
4512 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4514 static pid_t pid_fry(pid_t pid)
4516 unsigned a = pid & 0x55555555;
4517 unsigned b = pid & 0xAAAAAAAA;
4519 return (a << 1) | (b >> 1);
4522 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4524 if (cgroup_on_dfl(cgrp))
4525 return pid_fry(pid);
4530 static int cmppid(const void *a, const void *b)
4532 return *(pid_t *)a - *(pid_t *)b;
4535 static int fried_cmppid(const void *a, const void *b)
4537 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4540 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4541 enum cgroup_filetype type)
4543 struct cgroup_pidlist *l;
4544 /* don't need task_nsproxy() if we're looking at ourself */
4545 struct pid_namespace *ns = task_active_pid_ns(current);
4547 lockdep_assert_held(&cgrp->pidlist_mutex);
4549 list_for_each_entry(l, &cgrp->pidlists, links)
4550 if (l->key.type == type && l->key.ns == ns)
4556 * find the appropriate pidlist for our purpose (given procs vs tasks)
4557 * returns with the lock on that pidlist already held, and takes care
4558 * of the use count, or returns NULL with no locks held if we're out of
4561 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4562 enum cgroup_filetype type)
4564 struct cgroup_pidlist *l;
4566 lockdep_assert_held(&cgrp->pidlist_mutex);
4568 l = cgroup_pidlist_find(cgrp, type);
4572 /* entry not found; create a new one */
4573 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4577 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4579 /* don't need task_nsproxy() if we're looking at ourself */
4580 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4582 list_add(&l->links, &cgrp->pidlists);
4587 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4589 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4590 struct cgroup_pidlist **lp)
4594 int pid, n = 0; /* used for populating the array */
4595 struct css_task_iter it;
4596 struct task_struct *tsk;
4597 struct cgroup_pidlist *l;
4599 lockdep_assert_held(&cgrp->pidlist_mutex);
4602 * If cgroup gets more users after we read count, we won't have
4603 * enough space - tough. This race is indistinguishable to the
4604 * caller from the case that the additional cgroup users didn't
4605 * show up until sometime later on.
4607 length = cgroup_task_count(cgrp);
4608 array = pidlist_allocate(length);
4611 /* now, populate the array */
4612 css_task_iter_start(&cgrp->self, &it);
4613 while ((tsk = css_task_iter_next(&it))) {
4614 if (unlikely(n == length))
4616 /* get tgid or pid for procs or tasks file respectively */
4617 if (type == CGROUP_FILE_PROCS)
4618 pid = task_tgid_vnr(tsk);
4620 pid = task_pid_vnr(tsk);
4621 if (pid > 0) /* make sure to only use valid results */
4624 css_task_iter_end(&it);
4626 /* now sort & (if procs) strip out duplicates */
4627 if (cgroup_on_dfl(cgrp))
4628 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4630 sort(array, length, sizeof(pid_t), cmppid, NULL);
4631 if (type == CGROUP_FILE_PROCS)
4632 length = pidlist_uniq(array, length);
4634 l = cgroup_pidlist_find_create(cgrp, type);
4636 pidlist_free(array);
4640 /* store array, freeing old if necessary */
4641 pidlist_free(l->list);
4649 * cgroupstats_build - build and fill cgroupstats
4650 * @stats: cgroupstats to fill information into
4651 * @dentry: A dentry entry belonging to the cgroup for which stats have
4654 * Build and fill cgroupstats so that taskstats can export it to user
4657 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4659 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4660 struct cgroup *cgrp;
4661 struct css_task_iter it;
4662 struct task_struct *tsk;
4664 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4665 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4666 kernfs_type(kn) != KERNFS_DIR)
4669 mutex_lock(&cgroup_mutex);
4672 * We aren't being called from kernfs and there's no guarantee on
4673 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4674 * @kn->priv is RCU safe. Let's do the RCU dancing.
4677 cgrp = rcu_dereference(kn->priv);
4678 if (!cgrp || cgroup_is_dead(cgrp)) {
4680 mutex_unlock(&cgroup_mutex);
4685 css_task_iter_start(&cgrp->self, &it);
4686 while ((tsk = css_task_iter_next(&it))) {
4687 switch (tsk->state) {
4689 stats->nr_running++;
4691 case TASK_INTERRUPTIBLE:
4692 stats->nr_sleeping++;
4694 case TASK_UNINTERRUPTIBLE:
4695 stats->nr_uninterruptible++;
4698 stats->nr_stopped++;
4701 if (delayacct_is_task_waiting_on_io(tsk))
4702 stats->nr_io_wait++;
4706 css_task_iter_end(&it);
4708 mutex_unlock(&cgroup_mutex);
4714 * seq_file methods for the tasks/procs files. The seq_file position is the
4715 * next pid to display; the seq_file iterator is a pointer to the pid
4716 * in the cgroup->l->list array.
4719 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4722 * Initially we receive a position value that corresponds to
4723 * one more than the last pid shown (or 0 on the first call or
4724 * after a seek to the start). Use a binary-search to find the
4725 * next pid to display, if any
4727 struct kernfs_open_file *of = s->private;
4728 struct cgroup *cgrp = seq_css(s)->cgroup;
4729 struct cgroup_pidlist *l;
4730 enum cgroup_filetype type = seq_cft(s)->private;
4731 int index = 0, pid = *pos;
4734 mutex_lock(&cgrp->pidlist_mutex);
4737 * !NULL @of->priv indicates that this isn't the first start()
4738 * after open. If the matching pidlist is around, we can use that.
4739 * Look for it. Note that @of->priv can't be used directly. It
4740 * could already have been destroyed.
4743 of->priv = cgroup_pidlist_find(cgrp, type);
4746 * Either this is the first start() after open or the matching
4747 * pidlist has been destroyed inbetween. Create a new one.
4750 ret = pidlist_array_load(cgrp, type,
4751 (struct cgroup_pidlist **)&of->priv);
4753 return ERR_PTR(ret);
4758 int end = l->length;
4760 while (index < end) {
4761 int mid = (index + end) / 2;
4762 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4765 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4771 /* If we're off the end of the array, we're done */
4772 if (index >= l->length)
4774 /* Update the abstract position to be the actual pid that we found */
4775 iter = l->list + index;
4776 *pos = cgroup_pid_fry(cgrp, *iter);
4780 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4782 struct kernfs_open_file *of = s->private;
4783 struct cgroup_pidlist *l = of->priv;
4786 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4787 CGROUP_PIDLIST_DESTROY_DELAY);
4788 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4791 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4793 struct kernfs_open_file *of = s->private;
4794 struct cgroup_pidlist *l = of->priv;
4796 pid_t *end = l->list + l->length;
4798 * Advance to the next pid in the array. If this goes off the
4805 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4810 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4812 seq_printf(s, "%d\n", *(int *)v);
4817 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4820 return notify_on_release(css->cgroup);
4823 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4824 struct cftype *cft, u64 val)
4827 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4829 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4833 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4836 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4839 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4840 struct cftype *cft, u64 val)
4843 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4845 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4849 /* cgroup core interface files for the default hierarchy */
4850 static struct cftype cgroup_dfl_base_files[] = {
4852 .name = "cgroup.procs",
4853 .file_offset = offsetof(struct cgroup, procs_file),
4854 .seq_start = cgroup_pidlist_start,
4855 .seq_next = cgroup_pidlist_next,
4856 .seq_stop = cgroup_pidlist_stop,
4857 .seq_show = cgroup_pidlist_show,
4858 .private = CGROUP_FILE_PROCS,
4859 .write = cgroup_procs_write,
4862 .name = "cgroup.controllers",
4863 .seq_show = cgroup_controllers_show,
4866 .name = "cgroup.subtree_control",
4867 .seq_show = cgroup_subtree_control_show,
4868 .write = cgroup_subtree_control_write,
4871 .name = "cgroup.events",
4872 .flags = CFTYPE_NOT_ON_ROOT,
4873 .file_offset = offsetof(struct cgroup, events_file),
4874 .seq_show = cgroup_events_show,
4879 /* cgroup core interface files for the legacy hierarchies */
4880 static struct cftype cgroup_legacy_base_files[] = {
4882 .name = "cgroup.procs",
4883 .seq_start = cgroup_pidlist_start,
4884 .seq_next = cgroup_pidlist_next,
4885 .seq_stop = cgroup_pidlist_stop,
4886 .seq_show = cgroup_pidlist_show,
4887 .private = CGROUP_FILE_PROCS,
4888 .write = cgroup_procs_write,
4891 .name = "cgroup.clone_children",
4892 .read_u64 = cgroup_clone_children_read,
4893 .write_u64 = cgroup_clone_children_write,
4896 .name = "cgroup.sane_behavior",
4897 .flags = CFTYPE_ONLY_ON_ROOT,
4898 .seq_show = cgroup_sane_behavior_show,
4902 .seq_start = cgroup_pidlist_start,
4903 .seq_next = cgroup_pidlist_next,
4904 .seq_stop = cgroup_pidlist_stop,
4905 .seq_show = cgroup_pidlist_show,
4906 .private = CGROUP_FILE_TASKS,
4907 .write = cgroup_tasks_write,
4910 .name = "notify_on_release",
4911 .read_u64 = cgroup_read_notify_on_release,
4912 .write_u64 = cgroup_write_notify_on_release,
4915 .name = "release_agent",
4916 .flags = CFTYPE_ONLY_ON_ROOT,
4917 .seq_show = cgroup_release_agent_show,
4918 .write = cgroup_release_agent_write,
4919 .max_write_len = PATH_MAX - 1,
4925 * css destruction is four-stage process.
4927 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4928 * Implemented in kill_css().
4930 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4931 * and thus css_tryget_online() is guaranteed to fail, the css can be
4932 * offlined by invoking offline_css(). After offlining, the base ref is
4933 * put. Implemented in css_killed_work_fn().
4935 * 3. When the percpu_ref reaches zero, the only possible remaining
4936 * accessors are inside RCU read sections. css_release() schedules the
4939 * 4. After the grace period, the css can be freed. Implemented in
4940 * css_free_work_fn().
4942 * It is actually hairier because both step 2 and 4 require process context
4943 * and thus involve punting to css->destroy_work adding two additional
4944 * steps to the already complex sequence.
4946 static void css_free_work_fn(struct work_struct *work)
4948 struct cgroup_subsys_state *css =
4949 container_of(work, struct cgroup_subsys_state, destroy_work);
4950 struct cgroup_subsys *ss = css->ss;
4951 struct cgroup *cgrp = css->cgroup;
4953 percpu_ref_exit(&css->refcnt);
4957 struct cgroup_subsys_state *parent = css->parent;
4961 cgroup_idr_remove(&ss->css_idr, id);
4967 /* cgroup free path */
4968 atomic_dec(&cgrp->root->nr_cgrps);
4969 cgroup_pidlist_destroy_all(cgrp);
4970 cancel_work_sync(&cgrp->release_agent_work);
4972 if (cgroup_parent(cgrp)) {
4974 * We get a ref to the parent, and put the ref when
4975 * this cgroup is being freed, so it's guaranteed
4976 * that the parent won't be destroyed before its
4979 cgroup_put(cgroup_parent(cgrp));
4980 kernfs_put(cgrp->kn);
4984 * This is root cgroup's refcnt reaching zero,
4985 * which indicates that the root should be
4988 cgroup_destroy_root(cgrp->root);
4993 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4995 struct cgroup_subsys_state *css =
4996 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4998 INIT_WORK(&css->destroy_work, css_free_work_fn);
4999 queue_work(cgroup_destroy_wq, &css->destroy_work);
5002 static void css_release_work_fn(struct work_struct *work)
5004 struct cgroup_subsys_state *css =
5005 container_of(work, struct cgroup_subsys_state, destroy_work);
5006 struct cgroup_subsys *ss = css->ss;
5007 struct cgroup *cgrp = css->cgroup;
5009 mutex_lock(&cgroup_mutex);
5011 css->flags |= CSS_RELEASED;
5012 list_del_rcu(&css->sibling);
5015 /* css release path */
5016 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5017 if (ss->css_released)
5018 ss->css_released(css);
5020 /* cgroup release path */
5021 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
5025 * There are two control paths which try to determine
5026 * cgroup from dentry without going through kernfs -
5027 * cgroupstats_build() and css_tryget_online_from_dir().
5028 * Those are supported by RCU protecting clearing of
5029 * cgrp->kn->priv backpointer.
5032 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5036 mutex_unlock(&cgroup_mutex);
5038 call_rcu(&css->rcu_head, css_free_rcu_fn);
5041 static void css_release(struct percpu_ref *ref)
5043 struct cgroup_subsys_state *css =
5044 container_of(ref, struct cgroup_subsys_state, refcnt);
5046 INIT_WORK(&css->destroy_work, css_release_work_fn);
5047 queue_work(cgroup_destroy_wq, &css->destroy_work);
5050 static void init_and_link_css(struct cgroup_subsys_state *css,
5051 struct cgroup_subsys *ss, struct cgroup *cgrp)
5053 lockdep_assert_held(&cgroup_mutex);
5057 memset(css, 0, sizeof(*css));
5060 INIT_LIST_HEAD(&css->sibling);
5061 INIT_LIST_HEAD(&css->children);
5062 css->serial_nr = css_serial_nr_next++;
5063 atomic_set(&css->online_cnt, 0);
5065 if (cgroup_parent(cgrp)) {
5066 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5067 css_get(css->parent);
5070 BUG_ON(cgroup_css(cgrp, ss));
5073 /* invoke ->css_online() on a new CSS and mark it online if successful */
5074 static int online_css(struct cgroup_subsys_state *css)
5076 struct cgroup_subsys *ss = css->ss;
5079 lockdep_assert_held(&cgroup_mutex);
5082 ret = ss->css_online(css);
5084 css->flags |= CSS_ONLINE;
5085 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5087 atomic_inc(&css->online_cnt);
5089 atomic_inc(&css->parent->online_cnt);
5094 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5095 static void offline_css(struct cgroup_subsys_state *css)
5097 struct cgroup_subsys *ss = css->ss;
5099 lockdep_assert_held(&cgroup_mutex);
5101 if (!(css->flags & CSS_ONLINE))
5107 if (ss->css_offline)
5108 ss->css_offline(css);
5110 css->flags &= ~CSS_ONLINE;
5111 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5113 wake_up_all(&css->cgroup->offline_waitq);
5117 * css_create - create a cgroup_subsys_state
5118 * @cgrp: the cgroup new css will be associated with
5119 * @ss: the subsys of new css
5121 * Create a new css associated with @cgrp - @ss pair. On success, the new
5122 * css is online and installed in @cgrp. This function doesn't create the
5123 * interface files. Returns 0 on success, -errno on failure.
5125 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5126 struct cgroup_subsys *ss)
5128 struct cgroup *parent = cgroup_parent(cgrp);
5129 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5130 struct cgroup_subsys_state *css;
5133 lockdep_assert_held(&cgroup_mutex);
5135 css = ss->css_alloc(parent_css);
5139 init_and_link_css(css, ss, cgrp);
5141 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5145 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5147 goto err_free_percpu_ref;
5150 /* @css is ready to be brought online now, make it visible */
5151 list_add_tail_rcu(&css->sibling, &parent_css->children);
5152 cgroup_idr_replace(&ss->css_idr, css, css->id);
5154 err = online_css(css);
5158 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5159 cgroup_parent(parent)) {
5160 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5161 current->comm, current->pid, ss->name);
5162 if (!strcmp(ss->name, "memory"))
5163 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5164 ss->warned_broken_hierarchy = true;
5170 list_del_rcu(&css->sibling);
5171 cgroup_idr_remove(&ss->css_idr, css->id);
5172 err_free_percpu_ref:
5173 percpu_ref_exit(&css->refcnt);
5175 call_rcu(&css->rcu_head, css_free_rcu_fn);
5176 return ERR_PTR(err);
5179 static struct cgroup *cgroup_create(struct cgroup *parent)
5181 struct cgroup_root *root = parent->root;
5182 struct cgroup *cgrp, *tcgrp;
5183 int level = parent->level + 1;
5186 /* allocate the cgroup and its ID, 0 is reserved for the root */
5187 cgrp = kzalloc(sizeof(*cgrp) +
5188 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5190 return ERR_PTR(-ENOMEM);
5192 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5197 * Temporarily set the pointer to NULL, so idr_find() won't return
5198 * a half-baked cgroup.
5200 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5203 goto out_cancel_ref;
5206 init_cgroup_housekeeping(cgrp);
5208 cgrp->self.parent = &parent->self;
5210 cgrp->level = level;
5212 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5213 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5215 if (notify_on_release(parent))
5216 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5218 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5219 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5221 cgrp->self.serial_nr = css_serial_nr_next++;
5223 /* allocation complete, commit to creation */
5224 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5225 atomic_inc(&root->nr_cgrps);
5229 * @cgrp is now fully operational. If something fails after this
5230 * point, it'll be released via the normal destruction path.
5232 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5235 * On the default hierarchy, a child doesn't automatically inherit
5236 * subtree_control from the parent. Each is configured manually.
5238 if (!cgroup_on_dfl(cgrp))
5239 cgrp->subtree_control = cgroup_control(cgrp);
5241 cgroup_propagate_control(cgrp);
5243 /* @cgrp doesn't have dir yet so the following will only create csses */
5244 ret = cgroup_apply_control_enable(cgrp);
5251 percpu_ref_exit(&cgrp->self.refcnt);
5254 return ERR_PTR(ret);
5256 cgroup_destroy_locked(cgrp);
5257 return ERR_PTR(ret);
5260 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5263 struct cgroup *parent, *cgrp;
5264 struct kernfs_node *kn;
5267 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5268 if (strchr(name, '\n'))
5271 parent = cgroup_kn_lock_live(parent_kn, false);
5275 cgrp = cgroup_create(parent);
5277 ret = PTR_ERR(cgrp);
5281 /* create the directory */
5282 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5290 * This extra ref will be put in cgroup_free_fn() and guarantees
5291 * that @cgrp->kn is always accessible.
5295 ret = cgroup_kn_set_ugid(kn);
5299 ret = css_populate_dir(&cgrp->self);
5303 ret = cgroup_apply_control_enable(cgrp);
5307 /* let's create and online css's */
5308 kernfs_activate(kn);
5314 cgroup_destroy_locked(cgrp);
5316 cgroup_kn_unlock(parent_kn);
5321 * This is called when the refcnt of a css is confirmed to be killed.
5322 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5323 * initate destruction and put the css ref from kill_css().
5325 static void css_killed_work_fn(struct work_struct *work)
5327 struct cgroup_subsys_state *css =
5328 container_of(work, struct cgroup_subsys_state, destroy_work);
5330 mutex_lock(&cgroup_mutex);
5335 /* @css can't go away while we're holding cgroup_mutex */
5337 } while (css && atomic_dec_and_test(&css->online_cnt));
5339 mutex_unlock(&cgroup_mutex);
5342 /* css kill confirmation processing requires process context, bounce */
5343 static void css_killed_ref_fn(struct percpu_ref *ref)
5345 struct cgroup_subsys_state *css =
5346 container_of(ref, struct cgroup_subsys_state, refcnt);
5348 if (atomic_dec_and_test(&css->online_cnt)) {
5349 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5350 queue_work(cgroup_destroy_wq, &css->destroy_work);
5355 * kill_css - destroy a css
5356 * @css: css to destroy
5358 * This function initiates destruction of @css by removing cgroup interface
5359 * files and putting its base reference. ->css_offline() will be invoked
5360 * asynchronously once css_tryget_online() is guaranteed to fail and when
5361 * the reference count reaches zero, @css will be released.
5363 static void kill_css(struct cgroup_subsys_state *css)
5365 lockdep_assert_held(&cgroup_mutex);
5368 * This must happen before css is disassociated with its cgroup.
5369 * See seq_css() for details.
5374 * Killing would put the base ref, but we need to keep it alive
5375 * until after ->css_offline().
5380 * cgroup core guarantees that, by the time ->css_offline() is
5381 * invoked, no new css reference will be given out via
5382 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5383 * proceed to offlining css's because percpu_ref_kill() doesn't
5384 * guarantee that the ref is seen as killed on all CPUs on return.
5386 * Use percpu_ref_kill_and_confirm() to get notifications as each
5387 * css is confirmed to be seen as killed on all CPUs.
5389 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5393 * cgroup_destroy_locked - the first stage of cgroup destruction
5394 * @cgrp: cgroup to be destroyed
5396 * css's make use of percpu refcnts whose killing latency shouldn't be
5397 * exposed to userland and are RCU protected. Also, cgroup core needs to
5398 * guarantee that css_tryget_online() won't succeed by the time
5399 * ->css_offline() is invoked. To satisfy all the requirements,
5400 * destruction is implemented in the following two steps.
5402 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5403 * userland visible parts and start killing the percpu refcnts of
5404 * css's. Set up so that the next stage will be kicked off once all
5405 * the percpu refcnts are confirmed to be killed.
5407 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5408 * rest of destruction. Once all cgroup references are gone, the
5409 * cgroup is RCU-freed.
5411 * This function implements s1. After this step, @cgrp is gone as far as
5412 * the userland is concerned and a new cgroup with the same name may be
5413 * created. As cgroup doesn't care about the names internally, this
5414 * doesn't cause any problem.
5416 static int cgroup_destroy_locked(struct cgroup *cgrp)
5417 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5419 struct cgroup_subsys_state *css;
5420 struct cgrp_cset_link *link;
5423 lockdep_assert_held(&cgroup_mutex);
5426 * Only migration can raise populated from zero and we're already
5427 * holding cgroup_mutex.
5429 if (cgroup_is_populated(cgrp))
5433 * Make sure there's no live children. We can't test emptiness of
5434 * ->self.children as dead children linger on it while being
5435 * drained; otherwise, "rmdir parent/child parent" may fail.
5437 if (css_has_online_children(&cgrp->self))
5441 * Mark @cgrp and the associated csets dead. The former prevents
5442 * further task migration and child creation by disabling
5443 * cgroup_lock_live_group(). The latter makes the csets ignored by
5444 * the migration path.
5446 cgrp->self.flags &= ~CSS_ONLINE;
5448 spin_lock_bh(&css_set_lock);
5449 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5450 link->cset->dead = true;
5451 spin_unlock_bh(&css_set_lock);
5453 /* initiate massacre of all css's */
5454 for_each_css(css, ssid, cgrp)
5458 * Remove @cgrp directory along with the base files. @cgrp has an
5459 * extra ref on its kn.
5461 kernfs_remove(cgrp->kn);
5463 check_for_release(cgroup_parent(cgrp));
5465 /* put the base reference */
5466 percpu_ref_kill(&cgrp->self.refcnt);
5471 static int cgroup_rmdir(struct kernfs_node *kn)
5473 struct cgroup *cgrp;
5476 cgrp = cgroup_kn_lock_live(kn, false);
5480 ret = cgroup_destroy_locked(cgrp);
5482 cgroup_kn_unlock(kn);
5486 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5487 .remount_fs = cgroup_remount,
5488 .show_options = cgroup_show_options,
5489 .mkdir = cgroup_mkdir,
5490 .rmdir = cgroup_rmdir,
5491 .rename = cgroup_rename,
5492 .show_path = cgroup_show_path,
5495 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5497 struct cgroup_subsys_state *css;
5499 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5501 mutex_lock(&cgroup_mutex);
5503 idr_init(&ss->css_idr);
5504 INIT_LIST_HEAD(&ss->cfts);
5506 /* Create the root cgroup state for this subsystem */
5507 ss->root = &cgrp_dfl_root;
5508 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5509 /* We don't handle early failures gracefully */
5510 BUG_ON(IS_ERR(css));
5511 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5514 * Root csses are never destroyed and we can't initialize
5515 * percpu_ref during early init. Disable refcnting.
5517 css->flags |= CSS_NO_REF;
5520 /* allocation can't be done safely during early init */
5523 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5524 BUG_ON(css->id < 0);
5527 /* Update the init_css_set to contain a subsys
5528 * pointer to this state - since the subsystem is
5529 * newly registered, all tasks and hence the
5530 * init_css_set is in the subsystem's root cgroup. */
5531 init_css_set.subsys[ss->id] = css;
5533 have_fork_callback |= (bool)ss->fork << ss->id;
5534 have_exit_callback |= (bool)ss->exit << ss->id;
5535 have_free_callback |= (bool)ss->free << ss->id;
5536 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5538 /* At system boot, before all subsystems have been
5539 * registered, no tasks have been forked, so we don't
5540 * need to invoke fork callbacks here. */
5541 BUG_ON(!list_empty(&init_task.tasks));
5543 BUG_ON(online_css(css));
5545 mutex_unlock(&cgroup_mutex);
5549 * cgroup_init_early - cgroup initialization at system boot
5551 * Initialize cgroups at system boot, and initialize any
5552 * subsystems that request early init.
5554 int __init cgroup_init_early(void)
5556 static struct cgroup_sb_opts __initdata opts;
5557 struct cgroup_subsys *ss;
5560 init_cgroup_root(&cgrp_dfl_root, &opts);
5561 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5563 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5565 for_each_subsys(ss, i) {
5566 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5567 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5568 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5570 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5571 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5574 ss->name = cgroup_subsys_name[i];
5575 if (!ss->legacy_name)
5576 ss->legacy_name = cgroup_subsys_name[i];
5579 cgroup_init_subsys(ss, true);
5584 static u16 cgroup_disable_mask __initdata;
5587 * cgroup_init - cgroup initialization
5589 * Register cgroup filesystem and /proc file, and initialize
5590 * any subsystems that didn't request early init.
5592 int __init cgroup_init(void)
5594 struct cgroup_subsys *ss;
5597 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5598 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5599 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5600 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5602 get_user_ns(init_cgroup_ns.user_ns);
5604 mutex_lock(&cgroup_mutex);
5607 * Add init_css_set to the hash table so that dfl_root can link to
5610 hash_add(css_set_table, &init_css_set.hlist,
5611 css_set_hash(init_css_set.subsys));
5613 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5615 mutex_unlock(&cgroup_mutex);
5617 for_each_subsys(ss, ssid) {
5618 if (ss->early_init) {
5619 struct cgroup_subsys_state *css =
5620 init_css_set.subsys[ss->id];
5622 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5624 BUG_ON(css->id < 0);
5626 cgroup_init_subsys(ss, false);
5629 list_add_tail(&init_css_set.e_cset_node[ssid],
5630 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5633 * Setting dfl_root subsys_mask needs to consider the
5634 * disabled flag and cftype registration needs kmalloc,
5635 * both of which aren't available during early_init.
5637 if (cgroup_disable_mask & (1 << ssid)) {
5638 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5639 printk(KERN_INFO "Disabling %s control group subsystem\n",
5644 if (cgroup_ssid_no_v1(ssid))
5645 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5648 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5650 if (ss->implicit_on_dfl)
5651 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5652 else if (!ss->dfl_cftypes)
5653 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5655 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5656 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5658 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5659 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5663 ss->bind(init_css_set.subsys[ssid]);
5666 /* init_css_set.subsys[] has been updated, re-hash */
5667 hash_del(&init_css_set.hlist);
5668 hash_add(css_set_table, &init_css_set.hlist,
5669 css_set_hash(init_css_set.subsys));
5671 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5672 WARN_ON(register_filesystem(&cgroup_fs_type));
5673 WARN_ON(register_filesystem(&cgroup2_fs_type));
5674 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5679 static int __init cgroup_wq_init(void)
5682 * There isn't much point in executing destruction path in
5683 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5684 * Use 1 for @max_active.
5686 * We would prefer to do this in cgroup_init() above, but that
5687 * is called before init_workqueues(): so leave this until after.
5689 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5690 BUG_ON(!cgroup_destroy_wq);
5693 * Used to destroy pidlists and separate to serve as flush domain.
5694 * Cap @max_active to 1 too.
5696 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5698 BUG_ON(!cgroup_pidlist_destroy_wq);
5702 core_initcall(cgroup_wq_init);
5705 * proc_cgroup_show()
5706 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5707 * - Used for /proc/<pid>/cgroup.
5709 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5710 struct pid *pid, struct task_struct *tsk)
5714 struct cgroup_root *root;
5717 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5721 mutex_lock(&cgroup_mutex);
5722 spin_lock_bh(&css_set_lock);
5724 for_each_root(root) {
5725 struct cgroup_subsys *ss;
5726 struct cgroup *cgrp;
5727 int ssid, count = 0;
5729 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5732 seq_printf(m, "%d:", root->hierarchy_id);
5733 if (root != &cgrp_dfl_root)
5734 for_each_subsys(ss, ssid)
5735 if (root->subsys_mask & (1 << ssid))
5736 seq_printf(m, "%s%s", count++ ? "," : "",
5738 if (strlen(root->name))
5739 seq_printf(m, "%sname=%s", count ? "," : "",
5743 cgrp = task_cgroup_from_root(tsk, root);
5746 * On traditional hierarchies, all zombie tasks show up as
5747 * belonging to the root cgroup. On the default hierarchy,
5748 * while a zombie doesn't show up in "cgroup.procs" and
5749 * thus can't be migrated, its /proc/PID/cgroup keeps
5750 * reporting the cgroup it belonged to before exiting. If
5751 * the cgroup is removed before the zombie is reaped,
5752 * " (deleted)" is appended to the cgroup path.
5754 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5755 path = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5756 current->nsproxy->cgroup_ns);
5758 retval = -ENAMETOOLONG;
5767 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5768 seq_puts(m, " (deleted)\n");
5775 spin_unlock_bh(&css_set_lock);
5776 mutex_unlock(&cgroup_mutex);
5782 /* Display information about each subsystem and each hierarchy */
5783 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5785 struct cgroup_subsys *ss;
5788 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5790 * ideally we don't want subsystems moving around while we do this.
5791 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5792 * subsys/hierarchy state.
5794 mutex_lock(&cgroup_mutex);
5796 for_each_subsys(ss, i)
5797 seq_printf(m, "%s\t%d\t%d\t%d\n",
5798 ss->legacy_name, ss->root->hierarchy_id,
5799 atomic_read(&ss->root->nr_cgrps),
5800 cgroup_ssid_enabled(i));
5802 mutex_unlock(&cgroup_mutex);
5806 static int cgroupstats_open(struct inode *inode, struct file *file)
5808 return single_open(file, proc_cgroupstats_show, NULL);
5811 static const struct file_operations proc_cgroupstats_operations = {
5812 .open = cgroupstats_open,
5814 .llseek = seq_lseek,
5815 .release = single_release,
5819 * cgroup_fork - initialize cgroup related fields during copy_process()
5820 * @child: pointer to task_struct of forking parent process.
5822 * A task is associated with the init_css_set until cgroup_post_fork()
5823 * attaches it to the parent's css_set. Empty cg_list indicates that
5824 * @child isn't holding reference to its css_set.
5826 void cgroup_fork(struct task_struct *child)
5828 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5829 INIT_LIST_HEAD(&child->cg_list);
5833 * cgroup_can_fork - called on a new task before the process is exposed
5834 * @child: the task in question.
5836 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5837 * returns an error, the fork aborts with that error code. This allows for
5838 * a cgroup subsystem to conditionally allow or deny new forks.
5840 int cgroup_can_fork(struct task_struct *child)
5842 struct cgroup_subsys *ss;
5845 do_each_subsys_mask(ss, i, have_canfork_callback) {
5846 ret = ss->can_fork(child);
5849 } while_each_subsys_mask();
5854 for_each_subsys(ss, j) {
5857 if (ss->cancel_fork)
5858 ss->cancel_fork(child);
5865 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5866 * @child: the task in question
5868 * This calls the cancel_fork() callbacks if a fork failed *after*
5869 * cgroup_can_fork() succeded.
5871 void cgroup_cancel_fork(struct task_struct *child)
5873 struct cgroup_subsys *ss;
5876 for_each_subsys(ss, i)
5877 if (ss->cancel_fork)
5878 ss->cancel_fork(child);
5882 * cgroup_post_fork - called on a new task after adding it to the task list
5883 * @child: the task in question
5885 * Adds the task to the list running through its css_set if necessary and
5886 * call the subsystem fork() callbacks. Has to be after the task is
5887 * visible on the task list in case we race with the first call to
5888 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5891 void cgroup_post_fork(struct task_struct *child)
5893 struct cgroup_subsys *ss;
5897 * This may race against cgroup_enable_task_cg_lists(). As that
5898 * function sets use_task_css_set_links before grabbing
5899 * tasklist_lock and we just went through tasklist_lock to add
5900 * @child, it's guaranteed that either we see the set
5901 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5902 * @child during its iteration.
5904 * If we won the race, @child is associated with %current's
5905 * css_set. Grabbing css_set_lock guarantees both that the
5906 * association is stable, and, on completion of the parent's
5907 * migration, @child is visible in the source of migration or
5908 * already in the destination cgroup. This guarantee is necessary
5909 * when implementing operations which need to migrate all tasks of
5910 * a cgroup to another.
5912 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5913 * will remain in init_css_set. This is safe because all tasks are
5914 * in the init_css_set before cg_links is enabled and there's no
5915 * operation which transfers all tasks out of init_css_set.
5917 if (use_task_css_set_links) {
5918 struct css_set *cset;
5920 spin_lock_bh(&css_set_lock);
5921 cset = task_css_set(current);
5922 if (list_empty(&child->cg_list)) {
5924 css_set_move_task(child, NULL, cset, false);
5926 spin_unlock_bh(&css_set_lock);
5930 * Call ss->fork(). This must happen after @child is linked on
5931 * css_set; otherwise, @child might change state between ->fork()
5932 * and addition to css_set.
5934 do_each_subsys_mask(ss, i, have_fork_callback) {
5936 } while_each_subsys_mask();
5940 * cgroup_exit - detach cgroup from exiting task
5941 * @tsk: pointer to task_struct of exiting process
5943 * Description: Detach cgroup from @tsk and release it.
5945 * Note that cgroups marked notify_on_release force every task in
5946 * them to take the global cgroup_mutex mutex when exiting.
5947 * This could impact scaling on very large systems. Be reluctant to
5948 * use notify_on_release cgroups where very high task exit scaling
5949 * is required on large systems.
5951 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5952 * call cgroup_exit() while the task is still competent to handle
5953 * notify_on_release(), then leave the task attached to the root cgroup in
5954 * each hierarchy for the remainder of its exit. No need to bother with
5955 * init_css_set refcnting. init_css_set never goes away and we can't race
5956 * with migration path - PF_EXITING is visible to migration path.
5958 void cgroup_exit(struct task_struct *tsk)
5960 struct cgroup_subsys *ss;
5961 struct css_set *cset;
5965 * Unlink from @tsk from its css_set. As migration path can't race
5966 * with us, we can check css_set and cg_list without synchronization.
5968 cset = task_css_set(tsk);
5970 if (!list_empty(&tsk->cg_list)) {
5971 spin_lock_bh(&css_set_lock);
5972 css_set_move_task(tsk, cset, NULL, false);
5973 spin_unlock_bh(&css_set_lock);
5978 /* see cgroup_post_fork() for details */
5979 do_each_subsys_mask(ss, i, have_exit_callback) {
5981 } while_each_subsys_mask();
5984 void cgroup_free(struct task_struct *task)
5986 struct css_set *cset = task_css_set(task);
5987 struct cgroup_subsys *ss;
5990 do_each_subsys_mask(ss, ssid, have_free_callback) {
5992 } while_each_subsys_mask();
5997 static void check_for_release(struct cgroup *cgrp)
5999 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
6000 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
6001 schedule_work(&cgrp->release_agent_work);
6005 * Notify userspace when a cgroup is released, by running the
6006 * configured release agent with the name of the cgroup (path
6007 * relative to the root of cgroup file system) as the argument.
6009 * Most likely, this user command will try to rmdir this cgroup.
6011 * This races with the possibility that some other task will be
6012 * attached to this cgroup before it is removed, or that some other
6013 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
6014 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
6015 * unused, and this cgroup will be reprieved from its death sentence,
6016 * to continue to serve a useful existence. Next time it's released,
6017 * we will get notified again, if it still has 'notify_on_release' set.
6019 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
6020 * means only wait until the task is successfully execve()'d. The
6021 * separate release agent task is forked by call_usermodehelper(),
6022 * then control in this thread returns here, without waiting for the
6023 * release agent task. We don't bother to wait because the caller of
6024 * this routine has no use for the exit status of the release agent
6025 * task, so no sense holding our caller up for that.
6027 static void cgroup_release_agent(struct work_struct *work)
6029 struct cgroup *cgrp =
6030 container_of(work, struct cgroup, release_agent_work);
6031 char *pathbuf = NULL, *agentbuf = NULL, *path;
6032 char *argv[3], *envp[3];
6034 mutex_lock(&cgroup_mutex);
6036 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
6037 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
6038 if (!pathbuf || !agentbuf)
6041 spin_lock_bh(&css_set_lock);
6042 path = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
6043 spin_unlock_bh(&css_set_lock);
6051 /* minimal command environment */
6053 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
6056 mutex_unlock(&cgroup_mutex);
6057 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
6060 mutex_unlock(&cgroup_mutex);
6066 static int __init cgroup_disable(char *str)
6068 struct cgroup_subsys *ss;
6072 while ((token = strsep(&str, ",")) != NULL) {
6076 for_each_subsys(ss, i) {
6077 if (strcmp(token, ss->name) &&
6078 strcmp(token, ss->legacy_name))
6080 cgroup_disable_mask |= 1 << i;
6085 __setup("cgroup_disable=", cgroup_disable);
6087 static int __init cgroup_no_v1(char *str)
6089 struct cgroup_subsys *ss;
6093 while ((token = strsep(&str, ",")) != NULL) {
6097 if (!strcmp(token, "all")) {
6098 cgroup_no_v1_mask = U16_MAX;
6102 for_each_subsys(ss, i) {
6103 if (strcmp(token, ss->name) &&
6104 strcmp(token, ss->legacy_name))
6107 cgroup_no_v1_mask |= 1 << i;
6112 __setup("cgroup_no_v1=", cgroup_no_v1);
6115 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6116 * @dentry: directory dentry of interest
6117 * @ss: subsystem of interest
6119 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6120 * to get the corresponding css and return it. If such css doesn't exist
6121 * or can't be pinned, an ERR_PTR value is returned.
6123 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6124 struct cgroup_subsys *ss)
6126 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6127 struct file_system_type *s_type = dentry->d_sb->s_type;
6128 struct cgroup_subsys_state *css = NULL;
6129 struct cgroup *cgrp;
6131 /* is @dentry a cgroup dir? */
6132 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6133 !kn || kernfs_type(kn) != KERNFS_DIR)
6134 return ERR_PTR(-EBADF);
6139 * This path doesn't originate from kernfs and @kn could already
6140 * have been or be removed at any point. @kn->priv is RCU
6141 * protected for this access. See css_release_work_fn() for details.
6143 cgrp = rcu_dereference(kn->priv);
6145 css = cgroup_css(cgrp, ss);
6147 if (!css || !css_tryget_online(css))
6148 css = ERR_PTR(-ENOENT);
6155 * css_from_id - lookup css by id
6156 * @id: the cgroup id
6157 * @ss: cgroup subsys to be looked into
6159 * Returns the css if there's valid one with @id, otherwise returns NULL.
6160 * Should be called under rcu_read_lock().
6162 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6164 WARN_ON_ONCE(!rcu_read_lock_held());
6165 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
6169 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6170 * @path: path on the default hierarchy
6172 * Find the cgroup at @path on the default hierarchy, increment its
6173 * reference count and return it. Returns pointer to the found cgroup on
6174 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6175 * if @path points to a non-directory.
6177 struct cgroup *cgroup_get_from_path(const char *path)
6179 struct kernfs_node *kn;
6180 struct cgroup *cgrp;
6182 mutex_lock(&cgroup_mutex);
6184 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6186 if (kernfs_type(kn) == KERNFS_DIR) {
6190 cgrp = ERR_PTR(-ENOTDIR);
6194 cgrp = ERR_PTR(-ENOENT);
6197 mutex_unlock(&cgroup_mutex);
6200 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6203 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6204 * definition in cgroup-defs.h.
6206 #ifdef CONFIG_SOCK_CGROUP_DATA
6208 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6210 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6211 static bool cgroup_sk_alloc_disabled __read_mostly;
6213 void cgroup_sk_alloc_disable(void)
6215 if (cgroup_sk_alloc_disabled)
6217 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6218 cgroup_sk_alloc_disabled = true;
6223 #define cgroup_sk_alloc_disabled false
6227 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6229 if (cgroup_sk_alloc_disabled)
6235 struct css_set *cset;
6237 cset = task_css_set(current);
6238 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6239 skcd->val = (unsigned long)cset->dfl_cgrp;
6248 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6250 cgroup_put(sock_cgroup_ptr(skcd));
6253 #endif /* CONFIG_SOCK_CGROUP_DATA */
6255 /* cgroup namespaces */
6257 static struct cgroup_namespace *alloc_cgroup_ns(void)
6259 struct cgroup_namespace *new_ns;
6262 new_ns = kzalloc(sizeof(struct cgroup_namespace), GFP_KERNEL);
6264 return ERR_PTR(-ENOMEM);
6265 ret = ns_alloc_inum(&new_ns->ns);
6268 return ERR_PTR(ret);
6270 atomic_set(&new_ns->count, 1);
6271 new_ns->ns.ops = &cgroupns_operations;
6275 void free_cgroup_ns(struct cgroup_namespace *ns)
6277 put_css_set(ns->root_cset);
6278 put_user_ns(ns->user_ns);
6279 ns_free_inum(&ns->ns);
6282 EXPORT_SYMBOL(free_cgroup_ns);
6284 struct cgroup_namespace *copy_cgroup_ns(unsigned long flags,
6285 struct user_namespace *user_ns,
6286 struct cgroup_namespace *old_ns)
6288 struct cgroup_namespace *new_ns;
6289 struct css_set *cset;
6293 if (!(flags & CLONE_NEWCGROUP)) {
6294 get_cgroup_ns(old_ns);
6298 /* Allow only sysadmin to create cgroup namespace. */
6299 if (!ns_capable(user_ns, CAP_SYS_ADMIN))
6300 return ERR_PTR(-EPERM);
6302 mutex_lock(&cgroup_mutex);
6303 spin_lock_bh(&css_set_lock);
6305 cset = task_css_set(current);
6308 spin_unlock_bh(&css_set_lock);
6309 mutex_unlock(&cgroup_mutex);
6311 new_ns = alloc_cgroup_ns();
6312 if (IS_ERR(new_ns)) {
6317 new_ns->user_ns = get_user_ns(user_ns);
6318 new_ns->root_cset = cset;
6323 static inline struct cgroup_namespace *to_cg_ns(struct ns_common *ns)
6325 return container_of(ns, struct cgroup_namespace, ns);
6328 static int cgroupns_install(struct nsproxy *nsproxy, struct ns_common *ns)
6330 struct cgroup_namespace *cgroup_ns = to_cg_ns(ns);
6332 if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN) ||
6333 !ns_capable(cgroup_ns->user_ns, CAP_SYS_ADMIN))
6336 /* Don't need to do anything if we are attaching to our own cgroupns. */
6337 if (cgroup_ns == nsproxy->cgroup_ns)
6340 get_cgroup_ns(cgroup_ns);
6341 put_cgroup_ns(nsproxy->cgroup_ns);
6342 nsproxy->cgroup_ns = cgroup_ns;
6347 static struct ns_common *cgroupns_get(struct task_struct *task)
6349 struct cgroup_namespace *ns = NULL;
6350 struct nsproxy *nsproxy;
6353 nsproxy = task->nsproxy;
6355 ns = nsproxy->cgroup_ns;
6360 return ns ? &ns->ns : NULL;
6363 static void cgroupns_put(struct ns_common *ns)
6365 put_cgroup_ns(to_cg_ns(ns));
6368 const struct proc_ns_operations cgroupns_operations = {
6370 .type = CLONE_NEWCGROUP,
6371 .get = cgroupns_get,
6372 .put = cgroupns_put,
6373 .install = cgroupns_install,
6376 static __init int cgroup_namespaces_init(void)
6380 subsys_initcall(cgroup_namespaces_init);
6382 #ifdef CONFIG_CGROUP_DEBUG
6383 static struct cgroup_subsys_state *
6384 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6386 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6389 return ERR_PTR(-ENOMEM);
6394 static void debug_css_free(struct cgroup_subsys_state *css)
6399 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6402 return cgroup_task_count(css->cgroup);
6405 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6408 return (u64)(unsigned long)current->cgroups;
6411 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6417 count = atomic_read(&task_css_set(current)->refcount);
6422 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6424 struct cgrp_cset_link *link;
6425 struct css_set *cset;
6428 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6432 spin_lock_bh(&css_set_lock);
6434 cset = rcu_dereference(current->cgroups);
6435 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6436 struct cgroup *c = link->cgrp;
6438 cgroup_name(c, name_buf, NAME_MAX + 1);
6439 seq_printf(seq, "Root %d group %s\n",
6440 c->root->hierarchy_id, name_buf);
6443 spin_unlock_bh(&css_set_lock);
6448 #define MAX_TASKS_SHOWN_PER_CSS 25
6449 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6451 struct cgroup_subsys_state *css = seq_css(seq);
6452 struct cgrp_cset_link *link;
6454 spin_lock_bh(&css_set_lock);
6455 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6456 struct css_set *cset = link->cset;
6457 struct task_struct *task;
6460 seq_printf(seq, "css_set %p\n", cset);
6462 list_for_each_entry(task, &cset->tasks, cg_list) {
6463 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6465 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6468 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6469 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6471 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6475 seq_puts(seq, " ...\n");
6477 spin_unlock_bh(&css_set_lock);
6481 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6483 return (!cgroup_is_populated(css->cgroup) &&
6484 !css_has_online_children(&css->cgroup->self));
6487 static struct cftype debug_files[] = {
6489 .name = "taskcount",
6490 .read_u64 = debug_taskcount_read,
6494 .name = "current_css_set",
6495 .read_u64 = current_css_set_read,
6499 .name = "current_css_set_refcount",
6500 .read_u64 = current_css_set_refcount_read,
6504 .name = "current_css_set_cg_links",
6505 .seq_show = current_css_set_cg_links_read,
6509 .name = "cgroup_css_links",
6510 .seq_show = cgroup_css_links_read,
6514 .name = "releasable",
6515 .read_u64 = releasable_read,
6521 struct cgroup_subsys debug_cgrp_subsys = {
6522 .css_alloc = debug_css_alloc,
6523 .css_free = debug_css_free,
6524 .legacy_cftypes = debug_files,
6526 #endif /* CONFIG_CGROUP_DEBUG */