2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
60 #include <linux/atomic.h>
61 #include <linux/cpuset.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_lock protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DEFINE_SPINLOCK(css_set_lock);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_lock);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DEFINE_SPINLOCK(css_set_lock);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
103 * against file removal/re-creation across css hiding.
105 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
108 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
109 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
111 static DEFINE_SPINLOCK(release_agent_path_lock);
113 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct *cgroup_destroy_wq;
129 * pidlist destructions need to be flushed on cgroup destruction. Use a
130 * separate workqueue as flush domain.
132 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
134 /* generate an array of cgroup subsystem pointers */
135 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
136 static struct cgroup_subsys *cgroup_subsys[] = {
137 #include <linux/cgroup_subsys.h>
141 /* array of cgroup subsystem names */
142 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
143 static const char *cgroup_subsys_name[] = {
144 #include <linux/cgroup_subsys.h>
148 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
150 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
151 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
152 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
153 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
154 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
158 static struct static_key_true *cgroup_subsys_enabled_key[] = {
159 #include <linux/cgroup_subsys.h>
163 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
164 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
165 #include <linux/cgroup_subsys.h>
170 * The default hierarchy, reserved for the subsystems that are otherwise
171 * unattached - it never has more than a single cgroup, and all tasks are
172 * part of that cgroup.
174 struct cgroup_root cgrp_dfl_root;
175 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
178 * The default hierarchy always exists but is hidden until mounted for the
179 * first time. This is for backward compatibility.
181 static bool cgrp_dfl_visible;
183 /* Controllers blocked by the commandline in v1 */
184 static u16 cgroup_no_v1_mask;
186 /* some controllers are not supported in the default hierarchy */
187 static u16 cgrp_dfl_inhibit_ss_mask;
189 /* some controllers are implicitly enabled on the default hierarchy */
190 static unsigned long cgrp_dfl_implicit_ss_mask;
192 /* The list of hierarchy roots */
194 static LIST_HEAD(cgroup_roots);
195 static int cgroup_root_count;
197 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
198 static DEFINE_IDR(cgroup_hierarchy_idr);
201 * Assign a monotonically increasing serial number to csses. It guarantees
202 * cgroups with bigger numbers are newer than those with smaller numbers.
203 * Also, as csses are always appended to the parent's ->children list, it
204 * guarantees that sibling csses are always sorted in the ascending serial
205 * number order on the list. Protected by cgroup_mutex.
207 static u64 css_serial_nr_next = 1;
210 * These bitmask flags indicate whether tasks in the fork and exit paths have
211 * fork/exit handlers to call. This avoids us having to do extra work in the
212 * fork/exit path to check which subsystems have fork/exit callbacks.
214 static u16 have_fork_callback __read_mostly;
215 static u16 have_exit_callback __read_mostly;
216 static u16 have_free_callback __read_mostly;
218 /* Ditto for the can_fork callback. */
219 static u16 have_canfork_callback __read_mostly;
221 static struct file_system_type cgroup2_fs_type;
222 static struct cftype cgroup_dfl_base_files[];
223 static struct cftype cgroup_legacy_base_files[];
225 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask);
226 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp);
227 static int cgroup_apply_control(struct cgroup *cgrp);
228 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
229 static void css_task_iter_advance(struct css_task_iter *it);
230 static int cgroup_destroy_locked(struct cgroup *cgrp);
231 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
232 struct cgroup_subsys *ss);
233 static void css_release(struct percpu_ref *ref);
234 static void kill_css(struct cgroup_subsys_state *css);
235 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
236 struct cgroup *cgrp, struct cftype cfts[],
240 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
241 * @ssid: subsys ID of interest
243 * cgroup_subsys_enabled() can only be used with literal subsys names which
244 * is fine for individual subsystems but unsuitable for cgroup core. This
245 * is slower static_key_enabled() based test indexed by @ssid.
247 static bool cgroup_ssid_enabled(int ssid)
249 if (CGROUP_SUBSYS_COUNT == 0)
252 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
255 static bool cgroup_ssid_no_v1(int ssid)
257 return cgroup_no_v1_mask & (1 << ssid);
261 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
262 * @cgrp: the cgroup of interest
264 * The default hierarchy is the v2 interface of cgroup and this function
265 * can be used to test whether a cgroup is on the default hierarchy for
266 * cases where a subsystem should behave differnetly depending on the
269 * The set of behaviors which change on the default hierarchy are still
270 * being determined and the mount option is prefixed with __DEVEL__.
272 * List of changed behaviors:
274 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
275 * and "name" are disallowed.
277 * - When mounting an existing superblock, mount options should match.
279 * - Remount is disallowed.
281 * - rename(2) is disallowed.
283 * - "tasks" is removed. Everything should be at process granularity. Use
284 * "cgroup.procs" instead.
286 * - "cgroup.procs" is not sorted. pids will be unique unless they got
287 * recycled inbetween reads.
289 * - "release_agent" and "notify_on_release" are removed. Replacement
290 * notification mechanism will be implemented.
292 * - "cgroup.clone_children" is removed.
294 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
295 * and its descendants contain no task; otherwise, 1. The file also
296 * generates kernfs notification which can be monitored through poll and
297 * [di]notify when the value of the file changes.
299 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
300 * take masks of ancestors with non-empty cpus/mems, instead of being
301 * moved to an ancestor.
303 * - cpuset: a task can be moved into an empty cpuset, and again it takes
304 * masks of ancestors.
306 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
309 * - blkcg: blk-throttle becomes properly hierarchical.
311 * - debug: disallowed on the default hierarchy.
313 static bool cgroup_on_dfl(const struct cgroup *cgrp)
315 return cgrp->root == &cgrp_dfl_root;
318 /* IDR wrappers which synchronize using cgroup_idr_lock */
319 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
324 idr_preload(gfp_mask);
325 spin_lock_bh(&cgroup_idr_lock);
326 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
327 spin_unlock_bh(&cgroup_idr_lock);
332 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
336 spin_lock_bh(&cgroup_idr_lock);
337 ret = idr_replace(idr, ptr, id);
338 spin_unlock_bh(&cgroup_idr_lock);
342 static void cgroup_idr_remove(struct idr *idr, int id)
344 spin_lock_bh(&cgroup_idr_lock);
346 spin_unlock_bh(&cgroup_idr_lock);
349 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
351 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
354 return container_of(parent_css, struct cgroup, self);
358 /* subsystems visibly enabled on a cgroup */
359 static u16 cgroup_control(struct cgroup *cgrp)
361 struct cgroup *parent = cgroup_parent(cgrp);
362 u16 root_ss_mask = cgrp->root->subsys_mask;
365 return parent->subtree_control;
367 if (cgroup_on_dfl(cgrp))
368 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
369 cgrp_dfl_implicit_ss_mask);
373 /* subsystems enabled on a cgroup */
374 static u16 cgroup_ss_mask(struct cgroup *cgrp)
376 struct cgroup *parent = cgroup_parent(cgrp);
379 return parent->subtree_ss_mask;
381 return cgrp->root->subsys_mask;
385 * cgroup_css - obtain a cgroup's css for the specified subsystem
386 * @cgrp: the cgroup of interest
387 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
389 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
390 * function must be called either under cgroup_mutex or rcu_read_lock() and
391 * the caller is responsible for pinning the returned css if it wants to
392 * keep accessing it outside the said locks. This function may return
393 * %NULL if @cgrp doesn't have @subsys_id enabled.
395 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
396 struct cgroup_subsys *ss)
399 return rcu_dereference_check(cgrp->subsys[ss->id],
400 lockdep_is_held(&cgroup_mutex));
406 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
407 * @cgrp: the cgroup of interest
408 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
410 * Similar to cgroup_css() but returns the effective css, which is defined
411 * as the matching css of the nearest ancestor including self which has @ss
412 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
413 * function is guaranteed to return non-NULL css.
415 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
416 struct cgroup_subsys *ss)
418 lockdep_assert_held(&cgroup_mutex);
424 * This function is used while updating css associations and thus
425 * can't test the csses directly. Test ss_mask.
427 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
428 cgrp = cgroup_parent(cgrp);
433 return cgroup_css(cgrp, ss);
437 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
438 * @cgrp: the cgroup of interest
439 * @ss: the subsystem of interest
441 * Find and get the effective css of @cgrp for @ss. The effective css is
442 * defined as the matching css of the nearest ancestor including self which
443 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
444 * the root css is returned, so this function always returns a valid css.
445 * The returned css must be put using css_put().
447 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
448 struct cgroup_subsys *ss)
450 struct cgroup_subsys_state *css;
455 css = cgroup_css(cgrp, ss);
457 if (css && css_tryget_online(css))
459 cgrp = cgroup_parent(cgrp);
462 css = init_css_set.subsys[ss->id];
469 /* convenient tests for these bits */
470 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
472 return !(cgrp->self.flags & CSS_ONLINE);
475 static void cgroup_get(struct cgroup *cgrp)
477 WARN_ON_ONCE(cgroup_is_dead(cgrp));
478 css_get(&cgrp->self);
481 static bool cgroup_tryget(struct cgroup *cgrp)
483 return css_tryget(&cgrp->self);
486 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
488 struct cgroup *cgrp = of->kn->parent->priv;
489 struct cftype *cft = of_cft(of);
492 * This is open and unprotected implementation of cgroup_css().
493 * seq_css() is only called from a kernfs file operation which has
494 * an active reference on the file. Because all the subsystem
495 * files are drained before a css is disassociated with a cgroup,
496 * the matching css from the cgroup's subsys table is guaranteed to
497 * be and stay valid until the enclosing operation is complete.
500 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
504 EXPORT_SYMBOL_GPL(of_css);
506 static int notify_on_release(const struct cgroup *cgrp)
508 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
512 * for_each_css - iterate all css's of a cgroup
513 * @css: the iteration cursor
514 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
515 * @cgrp: the target cgroup to iterate css's of
517 * Should be called under cgroup_[tree_]mutex.
519 #define for_each_css(css, ssid, cgrp) \
520 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
521 if (!((css) = rcu_dereference_check( \
522 (cgrp)->subsys[(ssid)], \
523 lockdep_is_held(&cgroup_mutex)))) { } \
527 * for_each_e_css - iterate all effective css's of a cgroup
528 * @css: the iteration cursor
529 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
530 * @cgrp: the target cgroup to iterate css's of
532 * Should be called under cgroup_[tree_]mutex.
534 #define for_each_e_css(css, ssid, cgrp) \
535 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
536 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
541 * for_each_subsys - iterate all enabled cgroup subsystems
542 * @ss: the iteration cursor
543 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
545 #define for_each_subsys(ss, ssid) \
546 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
547 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
550 * do_each_subsys_mask - filter for_each_subsys with a bitmask
551 * @ss: the iteration cursor
552 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
553 * @ss_mask: the bitmask
555 * The block will only run for cases where the ssid-th bit (1 << ssid) of
558 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
559 unsigned long __ss_mask = (ss_mask); \
560 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
564 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
565 (ss) = cgroup_subsys[ssid]; \
568 #define while_each_subsys_mask() \
573 /* iterate across the hierarchies */
574 #define for_each_root(root) \
575 list_for_each_entry((root), &cgroup_roots, root_list)
577 /* iterate over child cgrps, lock should be held throughout iteration */
578 #define cgroup_for_each_live_child(child, cgrp) \
579 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
580 if (({ lockdep_assert_held(&cgroup_mutex); \
581 cgroup_is_dead(child); })) \
585 /* walk live descendants in preorder */
586 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
587 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
588 if (({ lockdep_assert_held(&cgroup_mutex); \
589 (dsct) = (d_css)->cgroup; \
590 cgroup_is_dead(dsct); })) \
594 /* walk live descendants in postorder */
595 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
596 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
597 if (({ lockdep_assert_held(&cgroup_mutex); \
598 (dsct) = (d_css)->cgroup; \
599 cgroup_is_dead(dsct); })) \
603 static void cgroup_release_agent(struct work_struct *work);
604 static void check_for_release(struct cgroup *cgrp);
607 * A cgroup can be associated with multiple css_sets as different tasks may
608 * belong to different cgroups on different hierarchies. In the other
609 * direction, a css_set is naturally associated with multiple cgroups.
610 * This M:N relationship is represented by the following link structure
611 * which exists for each association and allows traversing the associations
614 struct cgrp_cset_link {
615 /* the cgroup and css_set this link associates */
617 struct css_set *cset;
619 /* list of cgrp_cset_links anchored at cgrp->cset_links */
620 struct list_head cset_link;
622 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
623 struct list_head cgrp_link;
627 * The default css_set - used by init and its children prior to any
628 * hierarchies being mounted. It contains a pointer to the root state
629 * for each subsystem. Also used to anchor the list of css_sets. Not
630 * reference-counted, to improve performance when child cgroups
631 * haven't been created.
633 struct css_set init_css_set = {
634 .refcount = ATOMIC_INIT(1),
635 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
636 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
637 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
638 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
639 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
640 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
643 static int css_set_count = 1; /* 1 for init_css_set */
646 * css_set_populated - does a css_set contain any tasks?
647 * @cset: target css_set
649 static bool css_set_populated(struct css_set *cset)
651 lockdep_assert_held(&css_set_lock);
653 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
657 * cgroup_update_populated - updated populated count of a cgroup
658 * @cgrp: the target cgroup
659 * @populated: inc or dec populated count
661 * One of the css_sets associated with @cgrp is either getting its first
662 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
663 * count is propagated towards root so that a given cgroup's populated_cnt
664 * is zero iff the cgroup and all its descendants don't contain any tasks.
666 * @cgrp's interface file "cgroup.populated" is zero if
667 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
668 * changes from or to zero, userland is notified that the content of the
669 * interface file has changed. This can be used to detect when @cgrp and
670 * its descendants become populated or empty.
672 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
674 lockdep_assert_held(&css_set_lock);
680 trigger = !cgrp->populated_cnt++;
682 trigger = !--cgrp->populated_cnt;
687 check_for_release(cgrp);
688 cgroup_file_notify(&cgrp->events_file);
690 cgrp = cgroup_parent(cgrp);
695 * css_set_update_populated - update populated state of a css_set
696 * @cset: target css_set
697 * @populated: whether @cset is populated or depopulated
699 * @cset is either getting the first task or losing the last. Update the
700 * ->populated_cnt of all associated cgroups accordingly.
702 static void css_set_update_populated(struct css_set *cset, bool populated)
704 struct cgrp_cset_link *link;
706 lockdep_assert_held(&css_set_lock);
708 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
709 cgroup_update_populated(link->cgrp, populated);
713 * css_set_move_task - move a task from one css_set to another
714 * @task: task being moved
715 * @from_cset: css_set @task currently belongs to (may be NULL)
716 * @to_cset: new css_set @task is being moved to (may be NULL)
717 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
719 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
720 * css_set, @from_cset can be NULL. If @task is being disassociated
721 * instead of moved, @to_cset can be NULL.
723 * This function automatically handles populated_cnt updates and
724 * css_task_iter adjustments but the caller is responsible for managing
725 * @from_cset and @to_cset's reference counts.
727 static void css_set_move_task(struct task_struct *task,
728 struct css_set *from_cset, struct css_set *to_cset,
731 lockdep_assert_held(&css_set_lock);
733 if (to_cset && !css_set_populated(to_cset))
734 css_set_update_populated(to_cset, true);
737 struct css_task_iter *it, *pos;
739 WARN_ON_ONCE(list_empty(&task->cg_list));
742 * @task is leaving, advance task iterators which are
743 * pointing to it so that they can resume at the next
744 * position. Advancing an iterator might remove it from
745 * the list, use safe walk. See css_task_iter_advance*()
748 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
750 if (it->task_pos == &task->cg_list)
751 css_task_iter_advance(it);
753 list_del_init(&task->cg_list);
754 if (!css_set_populated(from_cset))
755 css_set_update_populated(from_cset, false);
757 WARN_ON_ONCE(!list_empty(&task->cg_list));
762 * We are synchronized through cgroup_threadgroup_rwsem
763 * against PF_EXITING setting such that we can't race
764 * against cgroup_exit() changing the css_set to
765 * init_css_set and dropping the old one.
767 WARN_ON_ONCE(task->flags & PF_EXITING);
769 rcu_assign_pointer(task->cgroups, to_cset);
770 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
776 * hash table for cgroup groups. This improves the performance to find
777 * an existing css_set. This hash doesn't (currently) take into
778 * account cgroups in empty hierarchies.
780 #define CSS_SET_HASH_BITS 7
781 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
783 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
785 unsigned long key = 0UL;
786 struct cgroup_subsys *ss;
789 for_each_subsys(ss, i)
790 key += (unsigned long)css[i];
791 key = (key >> 16) ^ key;
796 static void put_css_set_locked(struct css_set *cset)
798 struct cgrp_cset_link *link, *tmp_link;
799 struct cgroup_subsys *ss;
802 lockdep_assert_held(&css_set_lock);
804 if (!atomic_dec_and_test(&cset->refcount))
807 /* This css_set is dead. unlink it and release cgroup and css refs */
808 for_each_subsys(ss, ssid) {
809 list_del(&cset->e_cset_node[ssid]);
810 css_put(cset->subsys[ssid]);
812 hash_del(&cset->hlist);
815 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
816 list_del(&link->cset_link);
817 list_del(&link->cgrp_link);
818 if (cgroup_parent(link->cgrp))
819 cgroup_put(link->cgrp);
823 kfree_rcu(cset, rcu_head);
826 static void put_css_set(struct css_set *cset)
829 * Ensure that the refcount doesn't hit zero while any readers
830 * can see it. Similar to atomic_dec_and_lock(), but for an
833 if (atomic_add_unless(&cset->refcount, -1, 1))
836 spin_lock_bh(&css_set_lock);
837 put_css_set_locked(cset);
838 spin_unlock_bh(&css_set_lock);
842 * refcounted get/put for css_set objects
844 static inline void get_css_set(struct css_set *cset)
846 atomic_inc(&cset->refcount);
850 * compare_css_sets - helper function for find_existing_css_set().
851 * @cset: candidate css_set being tested
852 * @old_cset: existing css_set for a task
853 * @new_cgrp: cgroup that's being entered by the task
854 * @template: desired set of css pointers in css_set (pre-calculated)
856 * Returns true if "cset" matches "old_cset" except for the hierarchy
857 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
859 static bool compare_css_sets(struct css_set *cset,
860 struct css_set *old_cset,
861 struct cgroup *new_cgrp,
862 struct cgroup_subsys_state *template[])
864 struct list_head *l1, *l2;
867 * On the default hierarchy, there can be csets which are
868 * associated with the same set of cgroups but different csses.
869 * Let's first ensure that csses match.
871 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
875 * Compare cgroup pointers in order to distinguish between
876 * different cgroups in hierarchies. As different cgroups may
877 * share the same effective css, this comparison is always
880 l1 = &cset->cgrp_links;
881 l2 = &old_cset->cgrp_links;
883 struct cgrp_cset_link *link1, *link2;
884 struct cgroup *cgrp1, *cgrp2;
888 /* See if we reached the end - both lists are equal length. */
889 if (l1 == &cset->cgrp_links) {
890 BUG_ON(l2 != &old_cset->cgrp_links);
893 BUG_ON(l2 == &old_cset->cgrp_links);
895 /* Locate the cgroups associated with these links. */
896 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
897 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
900 /* Hierarchies should be linked in the same order. */
901 BUG_ON(cgrp1->root != cgrp2->root);
904 * If this hierarchy is the hierarchy of the cgroup
905 * that's changing, then we need to check that this
906 * css_set points to the new cgroup; if it's any other
907 * hierarchy, then this css_set should point to the
908 * same cgroup as the old css_set.
910 if (cgrp1->root == new_cgrp->root) {
911 if (cgrp1 != new_cgrp)
922 * find_existing_css_set - init css array and find the matching css_set
923 * @old_cset: the css_set that we're using before the cgroup transition
924 * @cgrp: the cgroup that we're moving into
925 * @template: out param for the new set of csses, should be clear on entry
927 static struct css_set *find_existing_css_set(struct css_set *old_cset,
929 struct cgroup_subsys_state *template[])
931 struct cgroup_root *root = cgrp->root;
932 struct cgroup_subsys *ss;
933 struct css_set *cset;
938 * Build the set of subsystem state objects that we want to see in the
939 * new css_set. while subsystems can change globally, the entries here
940 * won't change, so no need for locking.
942 for_each_subsys(ss, i) {
943 if (root->subsys_mask & (1UL << i)) {
945 * @ss is in this hierarchy, so we want the
946 * effective css from @cgrp.
948 template[i] = cgroup_e_css(cgrp, ss);
951 * @ss is not in this hierarchy, so we don't want
954 template[i] = old_cset->subsys[i];
958 key = css_set_hash(template);
959 hash_for_each_possible(css_set_table, cset, hlist, key) {
960 if (!compare_css_sets(cset, old_cset, cgrp, template))
963 /* This css_set matches what we need */
967 /* No existing cgroup group matched */
971 static void free_cgrp_cset_links(struct list_head *links_to_free)
973 struct cgrp_cset_link *link, *tmp_link;
975 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
976 list_del(&link->cset_link);
982 * allocate_cgrp_cset_links - allocate cgrp_cset_links
983 * @count: the number of links to allocate
984 * @tmp_links: list_head the allocated links are put on
986 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
987 * through ->cset_link. Returns 0 on success or -errno.
989 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
991 struct cgrp_cset_link *link;
994 INIT_LIST_HEAD(tmp_links);
996 for (i = 0; i < count; i++) {
997 link = kzalloc(sizeof(*link), GFP_KERNEL);
999 free_cgrp_cset_links(tmp_links);
1002 list_add(&link->cset_link, tmp_links);
1008 * link_css_set - a helper function to link a css_set to a cgroup
1009 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1010 * @cset: the css_set to be linked
1011 * @cgrp: the destination cgroup
1013 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1014 struct cgroup *cgrp)
1016 struct cgrp_cset_link *link;
1018 BUG_ON(list_empty(tmp_links));
1020 if (cgroup_on_dfl(cgrp))
1021 cset->dfl_cgrp = cgrp;
1023 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1028 * Always add links to the tail of the lists so that the lists are
1029 * in choronological order.
1031 list_move_tail(&link->cset_link, &cgrp->cset_links);
1032 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1034 if (cgroup_parent(cgrp))
1039 * find_css_set - return a new css_set with one cgroup updated
1040 * @old_cset: the baseline css_set
1041 * @cgrp: the cgroup to be updated
1043 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1044 * substituted into the appropriate hierarchy.
1046 static struct css_set *find_css_set(struct css_set *old_cset,
1047 struct cgroup *cgrp)
1049 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1050 struct css_set *cset;
1051 struct list_head tmp_links;
1052 struct cgrp_cset_link *link;
1053 struct cgroup_subsys *ss;
1057 lockdep_assert_held(&cgroup_mutex);
1059 /* First see if we already have a cgroup group that matches
1060 * the desired set */
1061 spin_lock_bh(&css_set_lock);
1062 cset = find_existing_css_set(old_cset, cgrp, template);
1065 spin_unlock_bh(&css_set_lock);
1070 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1074 /* Allocate all the cgrp_cset_link objects that we'll need */
1075 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1080 atomic_set(&cset->refcount, 1);
1081 INIT_LIST_HEAD(&cset->cgrp_links);
1082 INIT_LIST_HEAD(&cset->tasks);
1083 INIT_LIST_HEAD(&cset->mg_tasks);
1084 INIT_LIST_HEAD(&cset->mg_preload_node);
1085 INIT_LIST_HEAD(&cset->mg_node);
1086 INIT_LIST_HEAD(&cset->task_iters);
1087 INIT_HLIST_NODE(&cset->hlist);
1089 /* Copy the set of subsystem state objects generated in
1090 * find_existing_css_set() */
1091 memcpy(cset->subsys, template, sizeof(cset->subsys));
1093 spin_lock_bh(&css_set_lock);
1094 /* Add reference counts and links from the new css_set. */
1095 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1096 struct cgroup *c = link->cgrp;
1098 if (c->root == cgrp->root)
1100 link_css_set(&tmp_links, cset, c);
1103 BUG_ON(!list_empty(&tmp_links));
1107 /* Add @cset to the hash table */
1108 key = css_set_hash(cset->subsys);
1109 hash_add(css_set_table, &cset->hlist, key);
1111 for_each_subsys(ss, ssid) {
1112 struct cgroup_subsys_state *css = cset->subsys[ssid];
1114 list_add_tail(&cset->e_cset_node[ssid],
1115 &css->cgroup->e_csets[ssid]);
1119 spin_unlock_bh(&css_set_lock);
1124 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1126 struct cgroup *root_cgrp = kf_root->kn->priv;
1128 return root_cgrp->root;
1131 static int cgroup_init_root_id(struct cgroup_root *root)
1135 lockdep_assert_held(&cgroup_mutex);
1137 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1141 root->hierarchy_id = id;
1145 static void cgroup_exit_root_id(struct cgroup_root *root)
1147 lockdep_assert_held(&cgroup_mutex);
1149 if (root->hierarchy_id) {
1150 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1151 root->hierarchy_id = 0;
1155 static void cgroup_free_root(struct cgroup_root *root)
1158 /* hierarchy ID should already have been released */
1159 WARN_ON_ONCE(root->hierarchy_id);
1161 idr_destroy(&root->cgroup_idr);
1166 static void cgroup_destroy_root(struct cgroup_root *root)
1168 struct cgroup *cgrp = &root->cgrp;
1169 struct cgrp_cset_link *link, *tmp_link;
1171 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1173 BUG_ON(atomic_read(&root->nr_cgrps));
1174 BUG_ON(!list_empty(&cgrp->self.children));
1176 /* Rebind all subsystems back to the default hierarchy */
1177 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1180 * Release all the links from cset_links to this hierarchy's
1183 spin_lock_bh(&css_set_lock);
1185 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1186 list_del(&link->cset_link);
1187 list_del(&link->cgrp_link);
1191 spin_unlock_bh(&css_set_lock);
1193 if (!list_empty(&root->root_list)) {
1194 list_del(&root->root_list);
1195 cgroup_root_count--;
1198 cgroup_exit_root_id(root);
1200 mutex_unlock(&cgroup_mutex);
1202 kernfs_destroy_root(root->kf_root);
1203 cgroup_free_root(root);
1206 /* look up cgroup associated with given css_set on the specified hierarchy */
1207 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1208 struct cgroup_root *root)
1210 struct cgroup *res = NULL;
1212 lockdep_assert_held(&cgroup_mutex);
1213 lockdep_assert_held(&css_set_lock);
1215 if (cset == &init_css_set) {
1218 struct cgrp_cset_link *link;
1220 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1221 struct cgroup *c = link->cgrp;
1223 if (c->root == root) {
1235 * Return the cgroup for "task" from the given hierarchy. Must be
1236 * called with cgroup_mutex and css_set_lock held.
1238 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1239 struct cgroup_root *root)
1242 * No need to lock the task - since we hold cgroup_mutex the
1243 * task can't change groups, so the only thing that can happen
1244 * is that it exits and its css is set back to init_css_set.
1246 return cset_cgroup_from_root(task_css_set(task), root);
1250 * A task must hold cgroup_mutex to modify cgroups.
1252 * Any task can increment and decrement the count field without lock.
1253 * So in general, code holding cgroup_mutex can't rely on the count
1254 * field not changing. However, if the count goes to zero, then only
1255 * cgroup_attach_task() can increment it again. Because a count of zero
1256 * means that no tasks are currently attached, therefore there is no
1257 * way a task attached to that cgroup can fork (the other way to
1258 * increment the count). So code holding cgroup_mutex can safely
1259 * assume that if the count is zero, it will stay zero. Similarly, if
1260 * a task holds cgroup_mutex on a cgroup with zero count, it
1261 * knows that the cgroup won't be removed, as cgroup_rmdir()
1264 * A cgroup can only be deleted if both its 'count' of using tasks
1265 * is zero, and its list of 'children' cgroups is empty. Since all
1266 * tasks in the system use _some_ cgroup, and since there is always at
1267 * least one task in the system (init, pid == 1), therefore, root cgroup
1268 * always has either children cgroups and/or using tasks. So we don't
1269 * need a special hack to ensure that root cgroup cannot be deleted.
1271 * P.S. One more locking exception. RCU is used to guard the
1272 * update of a tasks cgroup pointer by cgroup_attach_task()
1275 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1276 static const struct file_operations proc_cgroupstats_operations;
1278 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1281 struct cgroup_subsys *ss = cft->ss;
1283 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1284 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1285 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1286 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1289 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1294 * cgroup_file_mode - deduce file mode of a control file
1295 * @cft: the control file in question
1297 * S_IRUGO for read, S_IWUSR for write.
1299 static umode_t cgroup_file_mode(const struct cftype *cft)
1303 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1306 if (cft->write_u64 || cft->write_s64 || cft->write) {
1307 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1317 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1318 * @subtree_control: the new subtree_control mask to consider
1319 * @this_ss_mask: available subsystems
1321 * On the default hierarchy, a subsystem may request other subsystems to be
1322 * enabled together through its ->depends_on mask. In such cases, more
1323 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1325 * This function calculates which subsystems need to be enabled if
1326 * @subtree_control is to be applied while restricted to @this_ss_mask.
1328 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1330 u16 cur_ss_mask = subtree_control;
1331 struct cgroup_subsys *ss;
1334 lockdep_assert_held(&cgroup_mutex);
1336 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1339 u16 new_ss_mask = cur_ss_mask;
1341 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1342 new_ss_mask |= ss->depends_on;
1343 } while_each_subsys_mask();
1346 * Mask out subsystems which aren't available. This can
1347 * happen only if some depended-upon subsystems were bound
1348 * to non-default hierarchies.
1350 new_ss_mask &= this_ss_mask;
1352 if (new_ss_mask == cur_ss_mask)
1354 cur_ss_mask = new_ss_mask;
1361 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1362 * @kn: the kernfs_node being serviced
1364 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1365 * the method finishes if locking succeeded. Note that once this function
1366 * returns the cgroup returned by cgroup_kn_lock_live() may become
1367 * inaccessible any time. If the caller intends to continue to access the
1368 * cgroup, it should pin it before invoking this function.
1370 static void cgroup_kn_unlock(struct kernfs_node *kn)
1372 struct cgroup *cgrp;
1374 if (kernfs_type(kn) == KERNFS_DIR)
1377 cgrp = kn->parent->priv;
1379 mutex_unlock(&cgroup_mutex);
1381 kernfs_unbreak_active_protection(kn);
1386 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1387 * @kn: the kernfs_node being serviced
1388 * @drain_offline: perform offline draining on the cgroup
1390 * This helper is to be used by a cgroup kernfs method currently servicing
1391 * @kn. It breaks the active protection, performs cgroup locking and
1392 * verifies that the associated cgroup is alive. Returns the cgroup if
1393 * alive; otherwise, %NULL. A successful return should be undone by a
1394 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1395 * cgroup is drained of offlining csses before return.
1397 * Any cgroup kernfs method implementation which requires locking the
1398 * associated cgroup should use this helper. It avoids nesting cgroup
1399 * locking under kernfs active protection and allows all kernfs operations
1400 * including self-removal.
1402 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn,
1405 struct cgroup *cgrp;
1407 if (kernfs_type(kn) == KERNFS_DIR)
1410 cgrp = kn->parent->priv;
1413 * We're gonna grab cgroup_mutex which nests outside kernfs
1414 * active_ref. cgroup liveliness check alone provides enough
1415 * protection against removal. Ensure @cgrp stays accessible and
1416 * break the active_ref protection.
1418 if (!cgroup_tryget(cgrp))
1420 kernfs_break_active_protection(kn);
1423 cgroup_lock_and_drain_offline(cgrp);
1425 mutex_lock(&cgroup_mutex);
1427 if (!cgroup_is_dead(cgrp))
1430 cgroup_kn_unlock(kn);
1434 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1436 char name[CGROUP_FILE_NAME_MAX];
1438 lockdep_assert_held(&cgroup_mutex);
1440 if (cft->file_offset) {
1441 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1442 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1444 spin_lock_irq(&cgroup_file_kn_lock);
1446 spin_unlock_irq(&cgroup_file_kn_lock);
1449 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1453 * css_clear_dir - remove subsys files in a cgroup directory
1456 static void css_clear_dir(struct cgroup_subsys_state *css)
1458 struct cgroup *cgrp = css->cgroup;
1459 struct cftype *cfts;
1461 if (!(css->flags & CSS_VISIBLE))
1464 css->flags &= ~CSS_VISIBLE;
1466 list_for_each_entry(cfts, &css->ss->cfts, node)
1467 cgroup_addrm_files(css, cgrp, cfts, false);
1471 * css_populate_dir - create subsys files in a cgroup directory
1474 * On failure, no file is added.
1476 static int css_populate_dir(struct cgroup_subsys_state *css)
1478 struct cgroup *cgrp = css->cgroup;
1479 struct cftype *cfts, *failed_cfts;
1482 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1486 if (cgroup_on_dfl(cgrp))
1487 cfts = cgroup_dfl_base_files;
1489 cfts = cgroup_legacy_base_files;
1491 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1494 list_for_each_entry(cfts, &css->ss->cfts, node) {
1495 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1502 css->flags |= CSS_VISIBLE;
1506 list_for_each_entry(cfts, &css->ss->cfts, node) {
1507 if (cfts == failed_cfts)
1509 cgroup_addrm_files(css, cgrp, cfts, false);
1514 static int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1516 struct cgroup *dcgrp = &dst_root->cgrp;
1517 struct cgroup_subsys *ss;
1520 lockdep_assert_held(&cgroup_mutex);
1522 do_each_subsys_mask(ss, ssid, ss_mask) {
1524 * If @ss has non-root csses attached to it, can't move.
1525 * If @ss is an implicit controller, it is exempt from this
1526 * rule and can be stolen.
1528 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1529 !ss->implicit_on_dfl)
1532 /* can't move between two non-dummy roots either */
1533 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1535 } while_each_subsys_mask();
1537 do_each_subsys_mask(ss, ssid, ss_mask) {
1538 struct cgroup_root *src_root = ss->root;
1539 struct cgroup *scgrp = &src_root->cgrp;
1540 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1541 struct css_set *cset;
1543 WARN_ON(!css || cgroup_css(dcgrp, ss));
1545 /* disable from the source */
1546 src_root->subsys_mask &= ~(1 << ssid);
1547 WARN_ON(cgroup_apply_control(scgrp));
1548 cgroup_finalize_control(scgrp, 0);
1551 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1552 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1553 ss->root = dst_root;
1554 css->cgroup = dcgrp;
1556 spin_lock_bh(&css_set_lock);
1557 hash_for_each(css_set_table, i, cset, hlist)
1558 list_move_tail(&cset->e_cset_node[ss->id],
1559 &dcgrp->e_csets[ss->id]);
1560 spin_unlock_bh(&css_set_lock);
1562 /* default hierarchy doesn't enable controllers by default */
1563 dst_root->subsys_mask |= 1 << ssid;
1564 if (dst_root == &cgrp_dfl_root) {
1565 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1567 dcgrp->subtree_control |= 1 << ssid;
1568 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1571 ret = cgroup_apply_control(dcgrp);
1573 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1578 } while_each_subsys_mask();
1580 kernfs_activate(dcgrp->kn);
1584 static int cgroup_show_options(struct seq_file *seq,
1585 struct kernfs_root *kf_root)
1587 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1588 struct cgroup_subsys *ss;
1591 if (root != &cgrp_dfl_root)
1592 for_each_subsys(ss, ssid)
1593 if (root->subsys_mask & (1 << ssid))
1594 seq_show_option(seq, ss->legacy_name, NULL);
1595 if (root->flags & CGRP_ROOT_NOPREFIX)
1596 seq_puts(seq, ",noprefix");
1597 if (root->flags & CGRP_ROOT_XATTR)
1598 seq_puts(seq, ",xattr");
1600 spin_lock(&release_agent_path_lock);
1601 if (strlen(root->release_agent_path))
1602 seq_show_option(seq, "release_agent",
1603 root->release_agent_path);
1604 spin_unlock(&release_agent_path_lock);
1606 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1607 seq_puts(seq, ",clone_children");
1608 if (strlen(root->name))
1609 seq_show_option(seq, "name", root->name);
1613 struct cgroup_sb_opts {
1616 char *release_agent;
1617 bool cpuset_clone_children;
1619 /* User explicitly requested empty subsystem */
1623 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1625 char *token, *o = data;
1626 bool all_ss = false, one_ss = false;
1628 struct cgroup_subsys *ss;
1632 #ifdef CONFIG_CPUSETS
1633 mask = ~((u16)1 << cpuset_cgrp_id);
1636 memset(opts, 0, sizeof(*opts));
1638 while ((token = strsep(&o, ",")) != NULL) {
1643 if (!strcmp(token, "none")) {
1644 /* Explicitly have no subsystems */
1648 if (!strcmp(token, "all")) {
1649 /* Mutually exclusive option 'all' + subsystem name */
1655 if (!strcmp(token, "noprefix")) {
1656 opts->flags |= CGRP_ROOT_NOPREFIX;
1659 if (!strcmp(token, "clone_children")) {
1660 opts->cpuset_clone_children = true;
1663 if (!strcmp(token, "xattr")) {
1664 opts->flags |= CGRP_ROOT_XATTR;
1667 if (!strncmp(token, "release_agent=", 14)) {
1668 /* Specifying two release agents is forbidden */
1669 if (opts->release_agent)
1671 opts->release_agent =
1672 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1673 if (!opts->release_agent)
1677 if (!strncmp(token, "name=", 5)) {
1678 const char *name = token + 5;
1679 /* Can't specify an empty name */
1682 /* Must match [\w.-]+ */
1683 for (i = 0; i < strlen(name); i++) {
1687 if ((c == '.') || (c == '-') || (c == '_'))
1691 /* Specifying two names is forbidden */
1694 opts->name = kstrndup(name,
1695 MAX_CGROUP_ROOT_NAMELEN - 1,
1703 for_each_subsys(ss, i) {
1704 if (strcmp(token, ss->legacy_name))
1706 if (!cgroup_ssid_enabled(i))
1708 if (cgroup_ssid_no_v1(i))
1711 /* Mutually exclusive option 'all' + subsystem name */
1714 opts->subsys_mask |= (1 << i);
1719 if (i == CGROUP_SUBSYS_COUNT)
1724 * If the 'all' option was specified select all the subsystems,
1725 * otherwise if 'none', 'name=' and a subsystem name options were
1726 * not specified, let's default to 'all'
1728 if (all_ss || (!one_ss && !opts->none && !opts->name))
1729 for_each_subsys(ss, i)
1730 if (cgroup_ssid_enabled(i) && !cgroup_ssid_no_v1(i))
1731 opts->subsys_mask |= (1 << i);
1734 * We either have to specify by name or by subsystems. (So all
1735 * empty hierarchies must have a name).
1737 if (!opts->subsys_mask && !opts->name)
1741 * Option noprefix was introduced just for backward compatibility
1742 * with the old cpuset, so we allow noprefix only if mounting just
1743 * the cpuset subsystem.
1745 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1748 /* Can't specify "none" and some subsystems */
1749 if (opts->subsys_mask && opts->none)
1755 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1758 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1759 struct cgroup_sb_opts opts;
1760 u16 added_mask, removed_mask;
1762 if (root == &cgrp_dfl_root) {
1763 pr_err("remount is not allowed\n");
1767 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1769 /* See what subsystems are wanted */
1770 ret = parse_cgroupfs_options(data, &opts);
1774 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1775 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1776 task_tgid_nr(current), current->comm);
1778 added_mask = opts.subsys_mask & ~root->subsys_mask;
1779 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1781 /* Don't allow flags or name to change at remount */
1782 if ((opts.flags ^ root->flags) ||
1783 (opts.name && strcmp(opts.name, root->name))) {
1784 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1785 opts.flags, opts.name ?: "", root->flags, root->name);
1790 /* remounting is not allowed for populated hierarchies */
1791 if (!list_empty(&root->cgrp.self.children)) {
1796 ret = rebind_subsystems(root, added_mask);
1800 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1802 if (opts.release_agent) {
1803 spin_lock(&release_agent_path_lock);
1804 strcpy(root->release_agent_path, opts.release_agent);
1805 spin_unlock(&release_agent_path_lock);
1808 kfree(opts.release_agent);
1810 mutex_unlock(&cgroup_mutex);
1815 * To reduce the fork() overhead for systems that are not actually using
1816 * their cgroups capability, we don't maintain the lists running through
1817 * each css_set to its tasks until we see the list actually used - in other
1818 * words after the first mount.
1820 static bool use_task_css_set_links __read_mostly;
1822 static void cgroup_enable_task_cg_lists(void)
1824 struct task_struct *p, *g;
1826 spin_lock_bh(&css_set_lock);
1828 if (use_task_css_set_links)
1831 use_task_css_set_links = true;
1834 * We need tasklist_lock because RCU is not safe against
1835 * while_each_thread(). Besides, a forking task that has passed
1836 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1837 * is not guaranteed to have its child immediately visible in the
1838 * tasklist if we walk through it with RCU.
1840 read_lock(&tasklist_lock);
1841 do_each_thread(g, p) {
1842 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1843 task_css_set(p) != &init_css_set);
1846 * We should check if the process is exiting, otherwise
1847 * it will race with cgroup_exit() in that the list
1848 * entry won't be deleted though the process has exited.
1849 * Do it while holding siglock so that we don't end up
1850 * racing against cgroup_exit().
1852 spin_lock_irq(&p->sighand->siglock);
1853 if (!(p->flags & PF_EXITING)) {
1854 struct css_set *cset = task_css_set(p);
1856 if (!css_set_populated(cset))
1857 css_set_update_populated(cset, true);
1858 list_add_tail(&p->cg_list, &cset->tasks);
1861 spin_unlock_irq(&p->sighand->siglock);
1862 } while_each_thread(g, p);
1863 read_unlock(&tasklist_lock);
1865 spin_unlock_bh(&css_set_lock);
1868 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1870 struct cgroup_subsys *ss;
1873 INIT_LIST_HEAD(&cgrp->self.sibling);
1874 INIT_LIST_HEAD(&cgrp->self.children);
1875 INIT_LIST_HEAD(&cgrp->cset_links);
1876 INIT_LIST_HEAD(&cgrp->pidlists);
1877 mutex_init(&cgrp->pidlist_mutex);
1878 cgrp->self.cgroup = cgrp;
1879 cgrp->self.flags |= CSS_ONLINE;
1881 for_each_subsys(ss, ssid)
1882 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1884 init_waitqueue_head(&cgrp->offline_waitq);
1885 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1888 static void init_cgroup_root(struct cgroup_root *root,
1889 struct cgroup_sb_opts *opts)
1891 struct cgroup *cgrp = &root->cgrp;
1893 INIT_LIST_HEAD(&root->root_list);
1894 atomic_set(&root->nr_cgrps, 1);
1896 init_cgroup_housekeeping(cgrp);
1897 idr_init(&root->cgroup_idr);
1899 root->flags = opts->flags;
1900 if (opts->release_agent)
1901 strcpy(root->release_agent_path, opts->release_agent);
1903 strcpy(root->name, opts->name);
1904 if (opts->cpuset_clone_children)
1905 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1908 static int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1910 LIST_HEAD(tmp_links);
1911 struct cgroup *root_cgrp = &root->cgrp;
1912 struct css_set *cset;
1915 lockdep_assert_held(&cgroup_mutex);
1917 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1920 root_cgrp->id = ret;
1921 root_cgrp->ancestor_ids[0] = ret;
1923 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1929 * We're accessing css_set_count without locking css_set_lock here,
1930 * but that's OK - it can only be increased by someone holding
1931 * cgroup_lock, and that's us. Later rebinding may disable
1932 * controllers on the default hierarchy and thus create new csets,
1933 * which can't be more than the existing ones. Allocate 2x.
1935 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1939 ret = cgroup_init_root_id(root);
1943 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1944 KERNFS_ROOT_CREATE_DEACTIVATED,
1946 if (IS_ERR(root->kf_root)) {
1947 ret = PTR_ERR(root->kf_root);
1950 root_cgrp->kn = root->kf_root->kn;
1952 ret = css_populate_dir(&root_cgrp->self);
1956 ret = rebind_subsystems(root, ss_mask);
1961 * There must be no failure case after here, since rebinding takes
1962 * care of subsystems' refcounts, which are explicitly dropped in
1963 * the failure exit path.
1965 list_add(&root->root_list, &cgroup_roots);
1966 cgroup_root_count++;
1969 * Link the root cgroup in this hierarchy into all the css_set
1972 spin_lock_bh(&css_set_lock);
1973 hash_for_each(css_set_table, i, cset, hlist) {
1974 link_css_set(&tmp_links, cset, root_cgrp);
1975 if (css_set_populated(cset))
1976 cgroup_update_populated(root_cgrp, true);
1978 spin_unlock_bh(&css_set_lock);
1980 BUG_ON(!list_empty(&root_cgrp->self.children));
1981 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1983 kernfs_activate(root_cgrp->kn);
1988 kernfs_destroy_root(root->kf_root);
1989 root->kf_root = NULL;
1991 cgroup_exit_root_id(root);
1993 percpu_ref_exit(&root_cgrp->self.refcnt);
1995 free_cgrp_cset_links(&tmp_links);
1999 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2000 int flags, const char *unused_dev_name,
2003 bool is_v2 = fs_type == &cgroup2_fs_type;
2004 struct super_block *pinned_sb = NULL;
2005 struct cgroup_subsys *ss;
2006 struct cgroup_root *root;
2007 struct cgroup_sb_opts opts;
2008 struct dentry *dentry;
2014 * The first time anyone tries to mount a cgroup, enable the list
2015 * linking each css_set to its tasks and fix up all existing tasks.
2017 if (!use_task_css_set_links)
2018 cgroup_enable_task_cg_lists();
2022 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data);
2023 return ERR_PTR(-EINVAL);
2025 cgrp_dfl_visible = true;
2026 root = &cgrp_dfl_root;
2027 cgroup_get(&root->cgrp);
2031 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
2033 /* First find the desired set of subsystems */
2034 ret = parse_cgroupfs_options(data, &opts);
2039 * Destruction of cgroup root is asynchronous, so subsystems may
2040 * still be dying after the previous unmount. Let's drain the
2041 * dying subsystems. We just need to ensure that the ones
2042 * unmounted previously finish dying and don't care about new ones
2043 * starting. Testing ref liveliness is good enough.
2045 for_each_subsys(ss, i) {
2046 if (!(opts.subsys_mask & (1 << i)) ||
2047 ss->root == &cgrp_dfl_root)
2050 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
2051 mutex_unlock(&cgroup_mutex);
2053 ret = restart_syscall();
2056 cgroup_put(&ss->root->cgrp);
2059 for_each_root(root) {
2060 bool name_match = false;
2062 if (root == &cgrp_dfl_root)
2066 * If we asked for a name then it must match. Also, if
2067 * name matches but sybsys_mask doesn't, we should fail.
2068 * Remember whether name matched.
2071 if (strcmp(opts.name, root->name))
2077 * If we asked for subsystems (or explicitly for no
2078 * subsystems) then they must match.
2080 if ((opts.subsys_mask || opts.none) &&
2081 (opts.subsys_mask != root->subsys_mask)) {
2088 if (root->flags ^ opts.flags)
2089 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2092 * We want to reuse @root whose lifetime is governed by its
2093 * ->cgrp. Let's check whether @root is alive and keep it
2094 * that way. As cgroup_kill_sb() can happen anytime, we
2095 * want to block it by pinning the sb so that @root doesn't
2096 * get killed before mount is complete.
2098 * With the sb pinned, tryget_live can reliably indicate
2099 * whether @root can be reused. If it's being killed,
2100 * drain it. We can use wait_queue for the wait but this
2101 * path is super cold. Let's just sleep a bit and retry.
2103 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2104 if (IS_ERR(pinned_sb) ||
2105 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2106 mutex_unlock(&cgroup_mutex);
2107 if (!IS_ERR_OR_NULL(pinned_sb))
2108 deactivate_super(pinned_sb);
2110 ret = restart_syscall();
2119 * No such thing, create a new one. name= matching without subsys
2120 * specification is allowed for already existing hierarchies but we
2121 * can't create new one without subsys specification.
2123 if (!opts.subsys_mask && !opts.none) {
2128 root = kzalloc(sizeof(*root), GFP_KERNEL);
2134 init_cgroup_root(root, &opts);
2136 ret = cgroup_setup_root(root, opts.subsys_mask);
2138 cgroup_free_root(root);
2141 mutex_unlock(&cgroup_mutex);
2143 kfree(opts.release_agent);
2147 return ERR_PTR(ret);
2149 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2150 is_v2 ? CGROUP2_SUPER_MAGIC : CGROUP_SUPER_MAGIC,
2152 if (IS_ERR(dentry) || !new_sb)
2153 cgroup_put(&root->cgrp);
2156 * If @pinned_sb, we're reusing an existing root and holding an
2157 * extra ref on its sb. Mount is complete. Put the extra ref.
2161 deactivate_super(pinned_sb);
2167 static void cgroup_kill_sb(struct super_block *sb)
2169 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2170 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2173 * If @root doesn't have any mounts or children, start killing it.
2174 * This prevents new mounts by disabling percpu_ref_tryget_live().
2175 * cgroup_mount() may wait for @root's release.
2177 * And don't kill the default root.
2179 if (!list_empty(&root->cgrp.self.children) ||
2180 root == &cgrp_dfl_root)
2181 cgroup_put(&root->cgrp);
2183 percpu_ref_kill(&root->cgrp.self.refcnt);
2188 static struct file_system_type cgroup_fs_type = {
2190 .mount = cgroup_mount,
2191 .kill_sb = cgroup_kill_sb,
2194 static struct file_system_type cgroup2_fs_type = {
2196 .mount = cgroup_mount,
2197 .kill_sb = cgroup_kill_sb,
2201 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2202 * @task: target task
2203 * @buf: the buffer to write the path into
2204 * @buflen: the length of the buffer
2206 * Determine @task's cgroup on the first (the one with the lowest non-zero
2207 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2208 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2209 * cgroup controller callbacks.
2211 * Return value is the same as kernfs_path().
2213 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2215 struct cgroup_root *root;
2216 struct cgroup *cgrp;
2217 int hierarchy_id = 1;
2220 mutex_lock(&cgroup_mutex);
2221 spin_lock_bh(&css_set_lock);
2223 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2226 cgrp = task_cgroup_from_root(task, root);
2227 path = cgroup_path(cgrp, buf, buflen);
2229 /* if no hierarchy exists, everyone is in "/" */
2230 if (strlcpy(buf, "/", buflen) < buflen)
2234 spin_unlock_bh(&css_set_lock);
2235 mutex_unlock(&cgroup_mutex);
2238 EXPORT_SYMBOL_GPL(task_cgroup_path);
2240 /* used to track tasks and other necessary states during migration */
2241 struct cgroup_taskset {
2242 /* the src and dst cset list running through cset->mg_node */
2243 struct list_head src_csets;
2244 struct list_head dst_csets;
2246 /* the subsys currently being processed */
2250 * Fields for cgroup_taskset_*() iteration.
2252 * Before migration is committed, the target migration tasks are on
2253 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2254 * the csets on ->dst_csets. ->csets point to either ->src_csets
2255 * or ->dst_csets depending on whether migration is committed.
2257 * ->cur_csets and ->cur_task point to the current task position
2260 struct list_head *csets;
2261 struct css_set *cur_cset;
2262 struct task_struct *cur_task;
2265 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2266 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2267 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2268 .csets = &tset.src_csets, \
2272 * cgroup_taskset_add - try to add a migration target task to a taskset
2273 * @task: target task
2274 * @tset: target taskset
2276 * Add @task, which is a migration target, to @tset. This function becomes
2277 * noop if @task doesn't need to be migrated. @task's css_set should have
2278 * been added as a migration source and @task->cg_list will be moved from
2279 * the css_set's tasks list to mg_tasks one.
2281 static void cgroup_taskset_add(struct task_struct *task,
2282 struct cgroup_taskset *tset)
2284 struct css_set *cset;
2286 lockdep_assert_held(&css_set_lock);
2288 /* @task either already exited or can't exit until the end */
2289 if (task->flags & PF_EXITING)
2292 /* leave @task alone if post_fork() hasn't linked it yet */
2293 if (list_empty(&task->cg_list))
2296 cset = task_css_set(task);
2297 if (!cset->mg_src_cgrp)
2300 list_move_tail(&task->cg_list, &cset->mg_tasks);
2301 if (list_empty(&cset->mg_node))
2302 list_add_tail(&cset->mg_node, &tset->src_csets);
2303 if (list_empty(&cset->mg_dst_cset->mg_node))
2304 list_move_tail(&cset->mg_dst_cset->mg_node,
2309 * cgroup_taskset_first - reset taskset and return the first task
2310 * @tset: taskset of interest
2311 * @dst_cssp: output variable for the destination css
2313 * @tset iteration is initialized and the first task is returned.
2315 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2316 struct cgroup_subsys_state **dst_cssp)
2318 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2319 tset->cur_task = NULL;
2321 return cgroup_taskset_next(tset, dst_cssp);
2325 * cgroup_taskset_next - iterate to the next task in taskset
2326 * @tset: taskset of interest
2327 * @dst_cssp: output variable for the destination css
2329 * Return the next task in @tset. Iteration must have been initialized
2330 * with cgroup_taskset_first().
2332 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2333 struct cgroup_subsys_state **dst_cssp)
2335 struct css_set *cset = tset->cur_cset;
2336 struct task_struct *task = tset->cur_task;
2338 while (&cset->mg_node != tset->csets) {
2340 task = list_first_entry(&cset->mg_tasks,
2341 struct task_struct, cg_list);
2343 task = list_next_entry(task, cg_list);
2345 if (&task->cg_list != &cset->mg_tasks) {
2346 tset->cur_cset = cset;
2347 tset->cur_task = task;
2350 * This function may be called both before and
2351 * after cgroup_taskset_migrate(). The two cases
2352 * can be distinguished by looking at whether @cset
2353 * has its ->mg_dst_cset set.
2355 if (cset->mg_dst_cset)
2356 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2358 *dst_cssp = cset->subsys[tset->ssid];
2363 cset = list_next_entry(cset, mg_node);
2371 * cgroup_taskset_migrate - migrate a taskset
2372 * @tset: taget taskset
2373 * @root: cgroup root the migration is taking place on
2375 * Migrate tasks in @tset as setup by migration preparation functions.
2376 * This function fails iff one of the ->can_attach callbacks fails and
2377 * guarantees that either all or none of the tasks in @tset are migrated.
2378 * @tset is consumed regardless of success.
2380 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2381 struct cgroup_root *root)
2383 struct cgroup_subsys *ss;
2384 struct task_struct *task, *tmp_task;
2385 struct css_set *cset, *tmp_cset;
2386 int ssid, failed_ssid, ret;
2388 /* methods shouldn't be called if no task is actually migrating */
2389 if (list_empty(&tset->src_csets))
2392 /* check that we can legitimately attach to the cgroup */
2393 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2394 if (ss->can_attach) {
2396 ret = ss->can_attach(tset);
2399 goto out_cancel_attach;
2402 } while_each_subsys_mask();
2405 * Now that we're guaranteed success, proceed to move all tasks to
2406 * the new cgroup. There are no failure cases after here, so this
2407 * is the commit point.
2409 spin_lock_bh(&css_set_lock);
2410 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2411 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2412 struct css_set *from_cset = task_css_set(task);
2413 struct css_set *to_cset = cset->mg_dst_cset;
2415 get_css_set(to_cset);
2416 css_set_move_task(task, from_cset, to_cset, true);
2417 put_css_set_locked(from_cset);
2420 spin_unlock_bh(&css_set_lock);
2423 * Migration is committed, all target tasks are now on dst_csets.
2424 * Nothing is sensitive to fork() after this point. Notify
2425 * controllers that migration is complete.
2427 tset->csets = &tset->dst_csets;
2429 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2434 } while_each_subsys_mask();
2437 goto out_release_tset;
2440 do_each_subsys_mask(ss, ssid, root->subsys_mask) {
2441 if (ssid == failed_ssid)
2443 if (ss->cancel_attach) {
2445 ss->cancel_attach(tset);
2447 } while_each_subsys_mask();
2449 spin_lock_bh(&css_set_lock);
2450 list_splice_init(&tset->dst_csets, &tset->src_csets);
2451 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2452 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2453 list_del_init(&cset->mg_node);
2455 spin_unlock_bh(&css_set_lock);
2460 * cgroup_may_migrate_to - verify whether a cgroup can be migration destination
2461 * @dst_cgrp: destination cgroup to test
2463 * On the default hierarchy, except for the root, subtree_control must be
2464 * zero for migration destination cgroups with tasks so that child cgroups
2465 * don't compete against tasks.
2467 static bool cgroup_may_migrate_to(struct cgroup *dst_cgrp)
2469 return !cgroup_on_dfl(dst_cgrp) || !cgroup_parent(dst_cgrp) ||
2470 !dst_cgrp->subtree_control;
2474 * cgroup_migrate_finish - cleanup after attach
2475 * @preloaded_csets: list of preloaded css_sets
2477 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2478 * those functions for details.
2480 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2482 struct css_set *cset, *tmp_cset;
2484 lockdep_assert_held(&cgroup_mutex);
2486 spin_lock_bh(&css_set_lock);
2487 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2488 cset->mg_src_cgrp = NULL;
2489 cset->mg_dst_cgrp = NULL;
2490 cset->mg_dst_cset = NULL;
2491 list_del_init(&cset->mg_preload_node);
2492 put_css_set_locked(cset);
2494 spin_unlock_bh(&css_set_lock);
2498 * cgroup_migrate_add_src - add a migration source css_set
2499 * @src_cset: the source css_set to add
2500 * @dst_cgrp: the destination cgroup
2501 * @preloaded_csets: list of preloaded css_sets
2503 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2504 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2505 * up by cgroup_migrate_finish().
2507 * This function may be called without holding cgroup_threadgroup_rwsem
2508 * even if the target is a process. Threads may be created and destroyed
2509 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2510 * into play and the preloaded css_sets are guaranteed to cover all
2513 static void cgroup_migrate_add_src(struct css_set *src_cset,
2514 struct cgroup *dst_cgrp,
2515 struct list_head *preloaded_csets)
2517 struct cgroup *src_cgrp;
2519 lockdep_assert_held(&cgroup_mutex);
2520 lockdep_assert_held(&css_set_lock);
2523 * If ->dead, @src_set is associated with one or more dead cgroups
2524 * and doesn't contain any migratable tasks. Ignore it early so
2525 * that the rest of migration path doesn't get confused by it.
2530 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2532 if (!list_empty(&src_cset->mg_preload_node))
2535 WARN_ON(src_cset->mg_src_cgrp);
2536 WARN_ON(src_cset->mg_dst_cgrp);
2537 WARN_ON(!list_empty(&src_cset->mg_tasks));
2538 WARN_ON(!list_empty(&src_cset->mg_node));
2540 src_cset->mg_src_cgrp = src_cgrp;
2541 src_cset->mg_dst_cgrp = dst_cgrp;
2542 get_css_set(src_cset);
2543 list_add(&src_cset->mg_preload_node, preloaded_csets);
2547 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2548 * @preloaded_csets: list of preloaded source css_sets
2550 * Tasks are about to be moved and all the source css_sets have been
2551 * preloaded to @preloaded_csets. This function looks up and pins all
2552 * destination css_sets, links each to its source, and append them to
2555 * This function must be called after cgroup_migrate_add_src() has been
2556 * called on each migration source css_set. After migration is performed
2557 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2560 static int cgroup_migrate_prepare_dst(struct list_head *preloaded_csets)
2563 struct css_set *src_cset, *tmp_cset;
2565 lockdep_assert_held(&cgroup_mutex);
2567 /* look up the dst cset for each src cset and link it to src */
2568 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2569 struct css_set *dst_cset;
2571 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2575 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2578 * If src cset equals dst, it's noop. Drop the src.
2579 * cgroup_migrate() will skip the cset too. Note that we
2580 * can't handle src == dst as some nodes are used by both.
2582 if (src_cset == dst_cset) {
2583 src_cset->mg_src_cgrp = NULL;
2584 src_cset->mg_dst_cgrp = NULL;
2585 list_del_init(&src_cset->mg_preload_node);
2586 put_css_set(src_cset);
2587 put_css_set(dst_cset);
2591 src_cset->mg_dst_cset = dst_cset;
2593 if (list_empty(&dst_cset->mg_preload_node))
2594 list_add(&dst_cset->mg_preload_node, &csets);
2596 put_css_set(dst_cset);
2599 list_splice_tail(&csets, preloaded_csets);
2602 cgroup_migrate_finish(&csets);
2607 * cgroup_migrate - migrate a process or task to a cgroup
2608 * @leader: the leader of the process or the task to migrate
2609 * @threadgroup: whether @leader points to the whole process or a single task
2610 * @root: cgroup root migration is taking place on
2612 * Migrate a process or task denoted by @leader. If migrating a process,
2613 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2614 * responsible for invoking cgroup_migrate_add_src() and
2615 * cgroup_migrate_prepare_dst() on the targets before invoking this
2616 * function and following up with cgroup_migrate_finish().
2618 * As long as a controller's ->can_attach() doesn't fail, this function is
2619 * guaranteed to succeed. This means that, excluding ->can_attach()
2620 * failure, when migrating multiple targets, the success or failure can be
2621 * decided for all targets by invoking group_migrate_prepare_dst() before
2622 * actually starting migrating.
2624 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2625 struct cgroup_root *root)
2627 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2628 struct task_struct *task;
2631 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2632 * already PF_EXITING could be freed from underneath us unless we
2633 * take an rcu_read_lock.
2635 spin_lock_bh(&css_set_lock);
2639 cgroup_taskset_add(task, &tset);
2642 } while_each_thread(leader, task);
2644 spin_unlock_bh(&css_set_lock);
2646 return cgroup_taskset_migrate(&tset, root);
2650 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2651 * @dst_cgrp: the cgroup to attach to
2652 * @leader: the task or the leader of the threadgroup to be attached
2653 * @threadgroup: attach the whole threadgroup?
2655 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2657 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2658 struct task_struct *leader, bool threadgroup)
2660 LIST_HEAD(preloaded_csets);
2661 struct task_struct *task;
2664 if (!cgroup_may_migrate_to(dst_cgrp))
2667 /* look up all src csets */
2668 spin_lock_bh(&css_set_lock);
2672 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2676 } while_each_thread(leader, task);
2678 spin_unlock_bh(&css_set_lock);
2680 /* prepare dst csets and commit */
2681 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2683 ret = cgroup_migrate(leader, threadgroup, dst_cgrp->root);
2685 cgroup_migrate_finish(&preloaded_csets);
2689 static int cgroup_procs_write_permission(struct task_struct *task,
2690 struct cgroup *dst_cgrp,
2691 struct kernfs_open_file *of)
2693 const struct cred *cred = current_cred();
2694 const struct cred *tcred = get_task_cred(task);
2698 * even if we're attaching all tasks in the thread group, we only
2699 * need to check permissions on one of them.
2701 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2702 !uid_eq(cred->euid, tcred->uid) &&
2703 !uid_eq(cred->euid, tcred->suid))
2706 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2707 struct super_block *sb = of->file->f_path.dentry->d_sb;
2708 struct cgroup *cgrp;
2709 struct inode *inode;
2711 spin_lock_bh(&css_set_lock);
2712 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2713 spin_unlock_bh(&css_set_lock);
2715 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2716 cgrp = cgroup_parent(cgrp);
2719 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2721 ret = inode_permission(inode, MAY_WRITE);
2731 * Find the task_struct of the task to attach by vpid and pass it along to the
2732 * function to attach either it or all tasks in its threadgroup. Will lock
2733 * cgroup_mutex and threadgroup.
2735 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2736 size_t nbytes, loff_t off, bool threadgroup)
2738 struct task_struct *tsk;
2739 struct cgroup *cgrp;
2743 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2746 cgrp = cgroup_kn_lock_live(of->kn, false);
2750 percpu_down_write(&cgroup_threadgroup_rwsem);
2753 tsk = find_task_by_vpid(pid);
2756 goto out_unlock_rcu;
2763 tsk = tsk->group_leader;
2766 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2767 * trapped in a cpuset, or RT worker may be born in a cgroup
2768 * with no rt_runtime allocated. Just say no.
2770 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2772 goto out_unlock_rcu;
2775 get_task_struct(tsk);
2778 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2780 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2782 put_task_struct(tsk);
2783 goto out_unlock_threadgroup;
2787 out_unlock_threadgroup:
2788 percpu_up_write(&cgroup_threadgroup_rwsem);
2789 cgroup_kn_unlock(of->kn);
2790 cpuset_post_attach_flush();
2791 return ret ?: nbytes;
2795 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2796 * @from: attach to all cgroups of a given task
2797 * @tsk: the task to be attached
2799 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2801 struct cgroup_root *root;
2804 mutex_lock(&cgroup_mutex);
2805 for_each_root(root) {
2806 struct cgroup *from_cgrp;
2808 if (root == &cgrp_dfl_root)
2811 spin_lock_bh(&css_set_lock);
2812 from_cgrp = task_cgroup_from_root(from, root);
2813 spin_unlock_bh(&css_set_lock);
2815 retval = cgroup_attach_task(from_cgrp, tsk, false);
2819 mutex_unlock(&cgroup_mutex);
2823 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2825 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2826 char *buf, size_t nbytes, loff_t off)
2828 return __cgroup_procs_write(of, buf, nbytes, off, false);
2831 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2832 char *buf, size_t nbytes, loff_t off)
2834 return __cgroup_procs_write(of, buf, nbytes, off, true);
2837 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2838 char *buf, size_t nbytes, loff_t off)
2840 struct cgroup *cgrp;
2842 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2844 cgrp = cgroup_kn_lock_live(of->kn, false);
2847 spin_lock(&release_agent_path_lock);
2848 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2849 sizeof(cgrp->root->release_agent_path));
2850 spin_unlock(&release_agent_path_lock);
2851 cgroup_kn_unlock(of->kn);
2855 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2857 struct cgroup *cgrp = seq_css(seq)->cgroup;
2859 spin_lock(&release_agent_path_lock);
2860 seq_puts(seq, cgrp->root->release_agent_path);
2861 spin_unlock(&release_agent_path_lock);
2862 seq_putc(seq, '\n');
2866 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2868 seq_puts(seq, "0\n");
2872 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2874 struct cgroup_subsys *ss;
2875 bool printed = false;
2878 do_each_subsys_mask(ss, ssid, ss_mask) {
2881 seq_printf(seq, "%s", ss->name);
2883 } while_each_subsys_mask();
2885 seq_putc(seq, '\n');
2888 /* show controllers which are enabled from the parent */
2889 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2891 struct cgroup *cgrp = seq_css(seq)->cgroup;
2893 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2897 /* show controllers which are enabled for a given cgroup's children */
2898 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2900 struct cgroup *cgrp = seq_css(seq)->cgroup;
2902 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2907 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2908 * @cgrp: root of the subtree to update csses for
2910 * @cgrp's control masks have changed and its subtree's css associations
2911 * need to be updated accordingly. This function looks up all css_sets
2912 * which are attached to the subtree, creates the matching updated css_sets
2913 * and migrates the tasks to the new ones.
2915 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2917 LIST_HEAD(preloaded_csets);
2918 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2919 struct cgroup_subsys_state *d_css;
2920 struct cgroup *dsct;
2921 struct css_set *src_cset;
2924 lockdep_assert_held(&cgroup_mutex);
2926 percpu_down_write(&cgroup_threadgroup_rwsem);
2928 /* look up all csses currently attached to @cgrp's subtree */
2929 spin_lock_bh(&css_set_lock);
2930 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2931 struct cgrp_cset_link *link;
2933 list_for_each_entry(link, &dsct->cset_links, cset_link)
2934 cgroup_migrate_add_src(link->cset, dsct,
2937 spin_unlock_bh(&css_set_lock);
2939 /* NULL dst indicates self on default hierarchy */
2940 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
2944 spin_lock_bh(&css_set_lock);
2945 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2946 struct task_struct *task, *ntask;
2948 /* src_csets precede dst_csets, break on the first dst_cset */
2949 if (!src_cset->mg_src_cgrp)
2952 /* all tasks in src_csets need to be migrated */
2953 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2954 cgroup_taskset_add(task, &tset);
2956 spin_unlock_bh(&css_set_lock);
2958 ret = cgroup_taskset_migrate(&tset, cgrp->root);
2960 cgroup_migrate_finish(&preloaded_csets);
2961 percpu_up_write(&cgroup_threadgroup_rwsem);
2966 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2967 * @cgrp: root of the target subtree
2969 * Because css offlining is asynchronous, userland may try to re-enable a
2970 * controller while the previous css is still around. This function grabs
2971 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2973 static void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2974 __acquires(&cgroup_mutex)
2976 struct cgroup *dsct;
2977 struct cgroup_subsys_state *d_css;
2978 struct cgroup_subsys *ss;
2982 mutex_lock(&cgroup_mutex);
2984 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2985 for_each_subsys(ss, ssid) {
2986 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2989 if (!css || !percpu_ref_is_dying(&css->refcnt))
2993 prepare_to_wait(&dsct->offline_waitq, &wait,
2994 TASK_UNINTERRUPTIBLE);
2996 mutex_unlock(&cgroup_mutex);
2998 finish_wait(&dsct->offline_waitq, &wait);
3007 * cgroup_save_control - save control masks of a subtree
3008 * @cgrp: root of the target subtree
3010 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
3011 * prefixed fields for @cgrp's subtree including @cgrp itself.
3013 static void cgroup_save_control(struct cgroup *cgrp)
3015 struct cgroup *dsct;
3016 struct cgroup_subsys_state *d_css;
3018 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3019 dsct->old_subtree_control = dsct->subtree_control;
3020 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3025 * cgroup_propagate_control - refresh control masks of a subtree
3026 * @cgrp: root of the target subtree
3028 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3029 * ->subtree_control and propagate controller availability through the
3030 * subtree so that descendants don't have unavailable controllers enabled.
3032 static void cgroup_propagate_control(struct cgroup *cgrp)
3034 struct cgroup *dsct;
3035 struct cgroup_subsys_state *d_css;
3037 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3038 dsct->subtree_control &= cgroup_control(dsct);
3039 dsct->subtree_ss_mask =
3040 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3041 cgroup_ss_mask(dsct));
3046 * cgroup_restore_control - restore control masks of a subtree
3047 * @cgrp: root of the target subtree
3049 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
3050 * prefixed fields for @cgrp's subtree including @cgrp itself.
3052 static void cgroup_restore_control(struct cgroup *cgrp)
3054 struct cgroup *dsct;
3055 struct cgroup_subsys_state *d_css;
3057 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3058 dsct->subtree_control = dsct->old_subtree_control;
3059 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3063 static bool css_visible(struct cgroup_subsys_state *css)
3065 struct cgroup_subsys *ss = css->ss;
3066 struct cgroup *cgrp = css->cgroup;
3068 if (cgroup_control(cgrp) & (1 << ss->id))
3070 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3072 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3076 * cgroup_apply_control_enable - enable or show csses according to control
3077 * @cgrp: root of the target subtree
3079 * Walk @cgrp's subtree and create new csses or make the existing ones
3080 * visible. A css is created invisible if it's being implicitly enabled
3081 * through dependency. An invisible css is made visible when the userland
3082 * explicitly enables it.
3084 * Returns 0 on success, -errno on failure. On failure, csses which have
3085 * been processed already aren't cleaned up. The caller is responsible for
3086 * cleaning up with cgroup_apply_control_disble().
3088 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3090 struct cgroup *dsct;
3091 struct cgroup_subsys_state *d_css;
3092 struct cgroup_subsys *ss;
3095 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3096 for_each_subsys(ss, ssid) {
3097 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3099 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3101 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3105 css = css_create(dsct, ss);
3107 return PTR_ERR(css);
3110 if (css_visible(css)) {
3111 ret = css_populate_dir(css);
3122 * cgroup_apply_control_disable - kill or hide csses according to control
3123 * @cgrp: root of the target subtree
3125 * Walk @cgrp's subtree and kill and hide csses so that they match
3126 * cgroup_ss_mask() and cgroup_visible_mask().
3128 * A css is hidden when the userland requests it to be disabled while other
3129 * subsystems are still depending on it. The css must not actively control
3130 * resources and be in the vanilla state if it's made visible again later.
3131 * Controllers which may be depended upon should provide ->css_reset() for
3134 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3136 struct cgroup *dsct;
3137 struct cgroup_subsys_state *d_css;
3138 struct cgroup_subsys *ss;
3141 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3142 for_each_subsys(ss, ssid) {
3143 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3145 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
3151 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3153 } else if (!css_visible(css)) {
3163 * cgroup_apply_control - apply control mask updates to the subtree
3164 * @cgrp: root of the target subtree
3166 * subsystems can be enabled and disabled in a subtree using the following
3169 * 1. Call cgroup_save_control() to stash the current state.
3170 * 2. Update ->subtree_control masks in the subtree as desired.
3171 * 3. Call cgroup_apply_control() to apply the changes.
3172 * 4. Optionally perform other related operations.
3173 * 5. Call cgroup_finalize_control() to finish up.
3175 * This function implements step 3 and propagates the mask changes
3176 * throughout @cgrp's subtree, updates csses accordingly and perform
3177 * process migrations.
3179 static int cgroup_apply_control(struct cgroup *cgrp)
3183 cgroup_propagate_control(cgrp);
3185 ret = cgroup_apply_control_enable(cgrp);
3190 * At this point, cgroup_e_css() results reflect the new csses
3191 * making the following cgroup_update_dfl_csses() properly update
3192 * css associations of all tasks in the subtree.
3194 ret = cgroup_update_dfl_csses(cgrp);
3202 * cgroup_finalize_control - finalize control mask update
3203 * @cgrp: root of the target subtree
3204 * @ret: the result of the update
3206 * Finalize control mask update. See cgroup_apply_control() for more info.
3208 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3211 cgroup_restore_control(cgrp);
3212 cgroup_propagate_control(cgrp);
3215 cgroup_apply_control_disable(cgrp);
3218 /* change the enabled child controllers for a cgroup in the default hierarchy */
3219 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3220 char *buf, size_t nbytes,
3223 u16 enable = 0, disable = 0;
3224 struct cgroup *cgrp, *child;
3225 struct cgroup_subsys *ss;
3230 * Parse input - space separated list of subsystem names prefixed
3231 * with either + or -.
3233 buf = strstrip(buf);
3234 while ((tok = strsep(&buf, " "))) {
3237 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3238 if (!cgroup_ssid_enabled(ssid) ||
3239 strcmp(tok + 1, ss->name))
3243 enable |= 1 << ssid;
3244 disable &= ~(1 << ssid);
3245 } else if (*tok == '-') {
3246 disable |= 1 << ssid;
3247 enable &= ~(1 << ssid);
3252 } while_each_subsys_mask();
3253 if (ssid == CGROUP_SUBSYS_COUNT)
3257 cgrp = cgroup_kn_lock_live(of->kn, true);
3261 for_each_subsys(ss, ssid) {
3262 if (enable & (1 << ssid)) {
3263 if (cgrp->subtree_control & (1 << ssid)) {
3264 enable &= ~(1 << ssid);
3268 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3272 } else if (disable & (1 << ssid)) {
3273 if (!(cgrp->subtree_control & (1 << ssid))) {
3274 disable &= ~(1 << ssid);
3278 /* a child has it enabled? */
3279 cgroup_for_each_live_child(child, cgrp) {
3280 if (child->subtree_control & (1 << ssid)) {
3288 if (!enable && !disable) {
3294 * Except for the root, subtree_control must be zero for a cgroup
3295 * with tasks so that child cgroups don't compete against tasks.
3297 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
3302 /* save and update control masks and prepare csses */
3303 cgroup_save_control(cgrp);
3305 cgrp->subtree_control |= enable;
3306 cgrp->subtree_control &= ~disable;
3308 ret = cgroup_apply_control(cgrp);
3310 cgroup_finalize_control(cgrp, ret);
3312 kernfs_activate(cgrp->kn);
3315 cgroup_kn_unlock(of->kn);
3316 return ret ?: nbytes;
3319 static int cgroup_events_show(struct seq_file *seq, void *v)
3321 seq_printf(seq, "populated %d\n",
3322 cgroup_is_populated(seq_css(seq)->cgroup));
3326 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3327 size_t nbytes, loff_t off)
3329 struct cgroup *cgrp = of->kn->parent->priv;
3330 struct cftype *cft = of->kn->priv;
3331 struct cgroup_subsys_state *css;
3335 return cft->write(of, buf, nbytes, off);
3338 * kernfs guarantees that a file isn't deleted with operations in
3339 * flight, which means that the matching css is and stays alive and
3340 * doesn't need to be pinned. The RCU locking is not necessary
3341 * either. It's just for the convenience of using cgroup_css().
3344 css = cgroup_css(cgrp, cft->ss);
3347 if (cft->write_u64) {
3348 unsigned long long v;
3349 ret = kstrtoull(buf, 0, &v);
3351 ret = cft->write_u64(css, cft, v);
3352 } else if (cft->write_s64) {
3354 ret = kstrtoll(buf, 0, &v);
3356 ret = cft->write_s64(css, cft, v);
3361 return ret ?: nbytes;
3364 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3366 return seq_cft(seq)->seq_start(seq, ppos);
3369 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3371 return seq_cft(seq)->seq_next(seq, v, ppos);
3374 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3376 seq_cft(seq)->seq_stop(seq, v);
3379 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3381 struct cftype *cft = seq_cft(m);
3382 struct cgroup_subsys_state *css = seq_css(m);
3385 return cft->seq_show(m, arg);
3388 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3389 else if (cft->read_s64)
3390 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3396 static struct kernfs_ops cgroup_kf_single_ops = {
3397 .atomic_write_len = PAGE_SIZE,
3398 .write = cgroup_file_write,
3399 .seq_show = cgroup_seqfile_show,
3402 static struct kernfs_ops cgroup_kf_ops = {
3403 .atomic_write_len = PAGE_SIZE,
3404 .write = cgroup_file_write,
3405 .seq_start = cgroup_seqfile_start,
3406 .seq_next = cgroup_seqfile_next,
3407 .seq_stop = cgroup_seqfile_stop,
3408 .seq_show = cgroup_seqfile_show,
3412 * cgroup_rename - Only allow simple rename of directories in place.
3414 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3415 const char *new_name_str)
3417 struct cgroup *cgrp = kn->priv;
3420 if (kernfs_type(kn) != KERNFS_DIR)
3422 if (kn->parent != new_parent)
3426 * This isn't a proper migration and its usefulness is very
3427 * limited. Disallow on the default hierarchy.
3429 if (cgroup_on_dfl(cgrp))
3433 * We're gonna grab cgroup_mutex which nests outside kernfs
3434 * active_ref. kernfs_rename() doesn't require active_ref
3435 * protection. Break them before grabbing cgroup_mutex.
3437 kernfs_break_active_protection(new_parent);
3438 kernfs_break_active_protection(kn);
3440 mutex_lock(&cgroup_mutex);
3442 ret = kernfs_rename(kn, new_parent, new_name_str);
3444 mutex_unlock(&cgroup_mutex);
3446 kernfs_unbreak_active_protection(kn);
3447 kernfs_unbreak_active_protection(new_parent);
3451 /* set uid and gid of cgroup dirs and files to that of the creator */
3452 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3454 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3455 .ia_uid = current_fsuid(),
3456 .ia_gid = current_fsgid(), };
3458 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3459 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3462 return kernfs_setattr(kn, &iattr);
3465 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3468 char name[CGROUP_FILE_NAME_MAX];
3469 struct kernfs_node *kn;
3470 struct lock_class_key *key = NULL;
3473 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3474 key = &cft->lockdep_key;
3476 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3477 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3482 ret = cgroup_kn_set_ugid(kn);
3488 if (cft->file_offset) {
3489 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3491 spin_lock_irq(&cgroup_file_kn_lock);
3493 spin_unlock_irq(&cgroup_file_kn_lock);
3500 * cgroup_addrm_files - add or remove files to a cgroup directory
3501 * @css: the target css
3502 * @cgrp: the target cgroup (usually css->cgroup)
3503 * @cfts: array of cftypes to be added
3504 * @is_add: whether to add or remove
3506 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3507 * For removals, this function never fails.
3509 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3510 struct cgroup *cgrp, struct cftype cfts[],
3513 struct cftype *cft, *cft_end = NULL;
3516 lockdep_assert_held(&cgroup_mutex);
3519 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3520 /* does cft->flags tell us to skip this file on @cgrp? */
3521 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3523 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3525 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3527 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3531 ret = cgroup_add_file(css, cgrp, cft);
3533 pr_warn("%s: failed to add %s, err=%d\n",
3534 __func__, cft->name, ret);
3540 cgroup_rm_file(cgrp, cft);
3546 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3549 struct cgroup_subsys *ss = cfts[0].ss;
3550 struct cgroup *root = &ss->root->cgrp;
3551 struct cgroup_subsys_state *css;
3554 lockdep_assert_held(&cgroup_mutex);
3556 /* add/rm files for all cgroups created before */
3557 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3558 struct cgroup *cgrp = css->cgroup;
3560 if (!(css->flags & CSS_VISIBLE))
3563 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3569 kernfs_activate(root->kn);
3573 static void cgroup_exit_cftypes(struct cftype *cfts)
3577 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3578 /* free copy for custom atomic_write_len, see init_cftypes() */
3579 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3584 /* revert flags set by cgroup core while adding @cfts */
3585 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3589 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3593 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3594 struct kernfs_ops *kf_ops;
3596 WARN_ON(cft->ss || cft->kf_ops);
3599 kf_ops = &cgroup_kf_ops;
3601 kf_ops = &cgroup_kf_single_ops;
3604 * Ugh... if @cft wants a custom max_write_len, we need to
3605 * make a copy of kf_ops to set its atomic_write_len.
3607 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3608 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3610 cgroup_exit_cftypes(cfts);
3613 kf_ops->atomic_write_len = cft->max_write_len;
3616 cft->kf_ops = kf_ops;
3623 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3625 lockdep_assert_held(&cgroup_mutex);
3627 if (!cfts || !cfts[0].ss)
3630 list_del(&cfts->node);
3631 cgroup_apply_cftypes(cfts, false);
3632 cgroup_exit_cftypes(cfts);
3637 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3638 * @cfts: zero-length name terminated array of cftypes
3640 * Unregister @cfts. Files described by @cfts are removed from all
3641 * existing cgroups and all future cgroups won't have them either. This
3642 * function can be called anytime whether @cfts' subsys is attached or not.
3644 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3647 int cgroup_rm_cftypes(struct cftype *cfts)
3651 mutex_lock(&cgroup_mutex);
3652 ret = cgroup_rm_cftypes_locked(cfts);
3653 mutex_unlock(&cgroup_mutex);
3658 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3659 * @ss: target cgroup subsystem
3660 * @cfts: zero-length name terminated array of cftypes
3662 * Register @cfts to @ss. Files described by @cfts are created for all
3663 * existing cgroups to which @ss is attached and all future cgroups will
3664 * have them too. This function can be called anytime whether @ss is
3667 * Returns 0 on successful registration, -errno on failure. Note that this
3668 * function currently returns 0 as long as @cfts registration is successful
3669 * even if some file creation attempts on existing cgroups fail.
3671 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3675 if (!cgroup_ssid_enabled(ss->id))
3678 if (!cfts || cfts[0].name[0] == '\0')
3681 ret = cgroup_init_cftypes(ss, cfts);
3685 mutex_lock(&cgroup_mutex);
3687 list_add_tail(&cfts->node, &ss->cfts);
3688 ret = cgroup_apply_cftypes(cfts, true);
3690 cgroup_rm_cftypes_locked(cfts);
3692 mutex_unlock(&cgroup_mutex);
3697 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3698 * @ss: target cgroup subsystem
3699 * @cfts: zero-length name terminated array of cftypes
3701 * Similar to cgroup_add_cftypes() but the added files are only used for
3702 * the default hierarchy.
3704 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3708 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3709 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3710 return cgroup_add_cftypes(ss, cfts);
3714 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3715 * @ss: target cgroup subsystem
3716 * @cfts: zero-length name terminated array of cftypes
3718 * Similar to cgroup_add_cftypes() but the added files are only used for
3719 * the legacy hierarchies.
3721 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3725 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3726 cft->flags |= __CFTYPE_NOT_ON_DFL;
3727 return cgroup_add_cftypes(ss, cfts);
3731 * cgroup_file_notify - generate a file modified event for a cgroup_file
3732 * @cfile: target cgroup_file
3734 * @cfile must have been obtained by setting cftype->file_offset.
3736 void cgroup_file_notify(struct cgroup_file *cfile)
3738 unsigned long flags;
3740 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3742 kernfs_notify(cfile->kn);
3743 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3747 * cgroup_task_count - count the number of tasks in a cgroup.
3748 * @cgrp: the cgroup in question
3750 * Return the number of tasks in the cgroup.
3752 static int cgroup_task_count(const struct cgroup *cgrp)
3755 struct cgrp_cset_link *link;
3757 spin_lock_bh(&css_set_lock);
3758 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3759 count += atomic_read(&link->cset->refcount);
3760 spin_unlock_bh(&css_set_lock);
3765 * css_next_child - find the next child of a given css
3766 * @pos: the current position (%NULL to initiate traversal)
3767 * @parent: css whose children to walk
3769 * This function returns the next child of @parent and should be called
3770 * under either cgroup_mutex or RCU read lock. The only requirement is
3771 * that @parent and @pos are accessible. The next sibling is guaranteed to
3772 * be returned regardless of their states.
3774 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3775 * css which finished ->css_online() is guaranteed to be visible in the
3776 * future iterations and will stay visible until the last reference is put.
3777 * A css which hasn't finished ->css_online() or already finished
3778 * ->css_offline() may show up during traversal. It's each subsystem's
3779 * responsibility to synchronize against on/offlining.
3781 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3782 struct cgroup_subsys_state *parent)
3784 struct cgroup_subsys_state *next;
3786 cgroup_assert_mutex_or_rcu_locked();
3789 * @pos could already have been unlinked from the sibling list.
3790 * Once a cgroup is removed, its ->sibling.next is no longer
3791 * updated when its next sibling changes. CSS_RELEASED is set when
3792 * @pos is taken off list, at which time its next pointer is valid,
3793 * and, as releases are serialized, the one pointed to by the next
3794 * pointer is guaranteed to not have started release yet. This
3795 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3796 * critical section, the one pointed to by its next pointer is
3797 * guaranteed to not have finished its RCU grace period even if we
3798 * have dropped rcu_read_lock() inbetween iterations.
3800 * If @pos has CSS_RELEASED set, its next pointer can't be
3801 * dereferenced; however, as each css is given a monotonically
3802 * increasing unique serial number and always appended to the
3803 * sibling list, the next one can be found by walking the parent's
3804 * children until the first css with higher serial number than
3805 * @pos's. While this path can be slower, it happens iff iteration
3806 * races against release and the race window is very small.
3809 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3810 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3811 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3813 list_for_each_entry_rcu(next, &parent->children, sibling)
3814 if (next->serial_nr > pos->serial_nr)
3819 * @next, if not pointing to the head, can be dereferenced and is
3822 if (&next->sibling != &parent->children)
3828 * css_next_descendant_pre - find the next descendant for pre-order walk
3829 * @pos: the current position (%NULL to initiate traversal)
3830 * @root: css whose descendants to walk
3832 * To be used by css_for_each_descendant_pre(). Find the next descendant
3833 * to visit for pre-order traversal of @root's descendants. @root is
3834 * included in the iteration and the first node to be visited.
3836 * While this function requires cgroup_mutex or RCU read locking, it
3837 * doesn't require the whole traversal to be contained in a single critical
3838 * section. This function will return the correct next descendant as long
3839 * as both @pos and @root are accessible and @pos is a descendant of @root.
3841 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3842 * css which finished ->css_online() is guaranteed to be visible in the
3843 * future iterations and will stay visible until the last reference is put.
3844 * A css which hasn't finished ->css_online() or already finished
3845 * ->css_offline() may show up during traversal. It's each subsystem's
3846 * responsibility to synchronize against on/offlining.
3848 struct cgroup_subsys_state *
3849 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3850 struct cgroup_subsys_state *root)
3852 struct cgroup_subsys_state *next;
3854 cgroup_assert_mutex_or_rcu_locked();
3856 /* if first iteration, visit @root */
3860 /* visit the first child if exists */
3861 next = css_next_child(NULL, pos);
3865 /* no child, visit my or the closest ancestor's next sibling */
3866 while (pos != root) {
3867 next = css_next_child(pos, pos->parent);
3877 * css_rightmost_descendant - return the rightmost descendant of a css
3878 * @pos: css of interest
3880 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3881 * is returned. This can be used during pre-order traversal to skip
3884 * While this function requires cgroup_mutex or RCU read locking, it
3885 * doesn't require the whole traversal to be contained in a single critical
3886 * section. This function will return the correct rightmost descendant as
3887 * long as @pos is accessible.
3889 struct cgroup_subsys_state *
3890 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3892 struct cgroup_subsys_state *last, *tmp;
3894 cgroup_assert_mutex_or_rcu_locked();
3898 /* ->prev isn't RCU safe, walk ->next till the end */
3900 css_for_each_child(tmp, last)
3907 static struct cgroup_subsys_state *
3908 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3910 struct cgroup_subsys_state *last;
3914 pos = css_next_child(NULL, pos);
3921 * css_next_descendant_post - find the next descendant for post-order walk
3922 * @pos: the current position (%NULL to initiate traversal)
3923 * @root: css whose descendants to walk
3925 * To be used by css_for_each_descendant_post(). Find the next descendant
3926 * to visit for post-order traversal of @root's descendants. @root is
3927 * included in the iteration and the last node to be visited.
3929 * While this function requires cgroup_mutex or RCU read locking, it
3930 * doesn't require the whole traversal to be contained in a single critical
3931 * section. This function will return the correct next descendant as long
3932 * as both @pos and @cgroup are accessible and @pos is a descendant of
3935 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3936 * css which finished ->css_online() is guaranteed to be visible in the
3937 * future iterations and will stay visible until the last reference is put.
3938 * A css which hasn't finished ->css_online() or already finished
3939 * ->css_offline() may show up during traversal. It's each subsystem's
3940 * responsibility to synchronize against on/offlining.
3942 struct cgroup_subsys_state *
3943 css_next_descendant_post(struct cgroup_subsys_state *pos,
3944 struct cgroup_subsys_state *root)
3946 struct cgroup_subsys_state *next;
3948 cgroup_assert_mutex_or_rcu_locked();
3950 /* if first iteration, visit leftmost descendant which may be @root */
3952 return css_leftmost_descendant(root);
3954 /* if we visited @root, we're done */
3958 /* if there's an unvisited sibling, visit its leftmost descendant */
3959 next = css_next_child(pos, pos->parent);
3961 return css_leftmost_descendant(next);
3963 /* no sibling left, visit parent */
3968 * css_has_online_children - does a css have online children
3969 * @css: the target css
3971 * Returns %true if @css has any online children; otherwise, %false. This
3972 * function can be called from any context but the caller is responsible
3973 * for synchronizing against on/offlining as necessary.
3975 bool css_has_online_children(struct cgroup_subsys_state *css)
3977 struct cgroup_subsys_state *child;
3981 css_for_each_child(child, css) {
3982 if (child->flags & CSS_ONLINE) {
3992 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3993 * @it: the iterator to advance
3995 * Advance @it to the next css_set to walk.
3997 static void css_task_iter_advance_css_set(struct css_task_iter *it)
3999 struct list_head *l = it->cset_pos;
4000 struct cgrp_cset_link *link;
4001 struct css_set *cset;
4003 lockdep_assert_held(&css_set_lock);
4005 /* Advance to the next non-empty css_set */
4008 if (l == it->cset_head) {
4009 it->cset_pos = NULL;
4010 it->task_pos = NULL;
4015 cset = container_of(l, struct css_set,
4016 e_cset_node[it->ss->id]);
4018 link = list_entry(l, struct cgrp_cset_link, cset_link);
4021 } while (!css_set_populated(cset));
4025 if (!list_empty(&cset->tasks))
4026 it->task_pos = cset->tasks.next;
4028 it->task_pos = cset->mg_tasks.next;
4030 it->tasks_head = &cset->tasks;
4031 it->mg_tasks_head = &cset->mg_tasks;
4034 * We don't keep css_sets locked across iteration steps and thus
4035 * need to take steps to ensure that iteration can be resumed after
4036 * the lock is re-acquired. Iteration is performed at two levels -
4037 * css_sets and tasks in them.
4039 * Once created, a css_set never leaves its cgroup lists, so a
4040 * pinned css_set is guaranteed to stay put and we can resume
4041 * iteration afterwards.
4043 * Tasks may leave @cset across iteration steps. This is resolved
4044 * by registering each iterator with the css_set currently being
4045 * walked and making css_set_move_task() advance iterators whose
4046 * next task is leaving.
4049 list_del(&it->iters_node);
4050 put_css_set_locked(it->cur_cset);
4053 it->cur_cset = cset;
4054 list_add(&it->iters_node, &cset->task_iters);
4057 static void css_task_iter_advance(struct css_task_iter *it)
4059 struct list_head *l = it->task_pos;
4061 lockdep_assert_held(&css_set_lock);
4065 * Advance iterator to find next entry. cset->tasks is consumed
4066 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4071 if (l == it->tasks_head)
4072 l = it->mg_tasks_head->next;
4074 if (l == it->mg_tasks_head)
4075 css_task_iter_advance_css_set(it);
4081 * css_task_iter_start - initiate task iteration
4082 * @css: the css to walk tasks of
4083 * @it: the task iterator to use
4085 * Initiate iteration through the tasks of @css. The caller can call
4086 * css_task_iter_next() to walk through the tasks until the function
4087 * returns NULL. On completion of iteration, css_task_iter_end() must be
4090 void css_task_iter_start(struct cgroup_subsys_state *css,
4091 struct css_task_iter *it)
4093 /* no one should try to iterate before mounting cgroups */
4094 WARN_ON_ONCE(!use_task_css_set_links);
4096 memset(it, 0, sizeof(*it));
4098 spin_lock_bh(&css_set_lock);
4103 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4105 it->cset_pos = &css->cgroup->cset_links;
4107 it->cset_head = it->cset_pos;
4109 css_task_iter_advance_css_set(it);
4111 spin_unlock_bh(&css_set_lock);
4115 * css_task_iter_next - return the next task for the iterator
4116 * @it: the task iterator being iterated
4118 * The "next" function for task iteration. @it should have been
4119 * initialized via css_task_iter_start(). Returns NULL when the iteration
4122 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4125 put_task_struct(it->cur_task);
4126 it->cur_task = NULL;
4129 spin_lock_bh(&css_set_lock);
4132 it->cur_task = list_entry(it->task_pos, struct task_struct,
4134 get_task_struct(it->cur_task);
4135 css_task_iter_advance(it);
4138 spin_unlock_bh(&css_set_lock);
4140 return it->cur_task;
4144 * css_task_iter_end - finish task iteration
4145 * @it: the task iterator to finish
4147 * Finish task iteration started by css_task_iter_start().
4149 void css_task_iter_end(struct css_task_iter *it)
4152 spin_lock_bh(&css_set_lock);
4153 list_del(&it->iters_node);
4154 put_css_set_locked(it->cur_cset);
4155 spin_unlock_bh(&css_set_lock);
4159 put_task_struct(it->cur_task);
4163 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
4164 * @to: cgroup to which the tasks will be moved
4165 * @from: cgroup in which the tasks currently reside
4167 * Locking rules between cgroup_post_fork() and the migration path
4168 * guarantee that, if a task is forking while being migrated, the new child
4169 * is guaranteed to be either visible in the source cgroup after the
4170 * parent's migration is complete or put into the target cgroup. No task
4171 * can slip out of migration through forking.
4173 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
4175 LIST_HEAD(preloaded_csets);
4176 struct cgrp_cset_link *link;
4177 struct css_task_iter it;
4178 struct task_struct *task;
4181 if (!cgroup_may_migrate_to(to))
4184 mutex_lock(&cgroup_mutex);
4186 /* all tasks in @from are being moved, all csets are source */
4187 spin_lock_bh(&css_set_lock);
4188 list_for_each_entry(link, &from->cset_links, cset_link)
4189 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
4190 spin_unlock_bh(&css_set_lock);
4192 ret = cgroup_migrate_prepare_dst(&preloaded_csets);
4197 * Migrate tasks one-by-one until @from is empty. This fails iff
4198 * ->can_attach() fails.
4201 css_task_iter_start(&from->self, &it);
4202 task = css_task_iter_next(&it);
4204 get_task_struct(task);
4205 css_task_iter_end(&it);
4208 ret = cgroup_migrate(task, false, to->root);
4209 put_task_struct(task);
4211 } while (task && !ret);
4213 cgroup_migrate_finish(&preloaded_csets);
4214 mutex_unlock(&cgroup_mutex);
4219 * Stuff for reading the 'tasks'/'procs' files.
4221 * Reading this file can return large amounts of data if a cgroup has
4222 * *lots* of attached tasks. So it may need several calls to read(),
4223 * but we cannot guarantee that the information we produce is correct
4224 * unless we produce it entirely atomically.
4228 /* which pidlist file are we talking about? */
4229 enum cgroup_filetype {
4235 * A pidlist is a list of pids that virtually represents the contents of one
4236 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4237 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4240 struct cgroup_pidlist {
4242 * used to find which pidlist is wanted. doesn't change as long as
4243 * this particular list stays in the list.
4245 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
4248 /* how many elements the above list has */
4250 /* each of these stored in a list by its cgroup */
4251 struct list_head links;
4252 /* pointer to the cgroup we belong to, for list removal purposes */
4253 struct cgroup *owner;
4254 /* for delayed destruction */
4255 struct delayed_work destroy_dwork;
4259 * The following two functions "fix" the issue where there are more pids
4260 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4261 * TODO: replace with a kernel-wide solution to this problem
4263 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4264 static void *pidlist_allocate(int count)
4266 if (PIDLIST_TOO_LARGE(count))
4267 return vmalloc(count * sizeof(pid_t));
4269 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4272 static void pidlist_free(void *p)
4278 * Used to destroy all pidlists lingering waiting for destroy timer. None
4279 * should be left afterwards.
4281 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4283 struct cgroup_pidlist *l, *tmp_l;
4285 mutex_lock(&cgrp->pidlist_mutex);
4286 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4287 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4288 mutex_unlock(&cgrp->pidlist_mutex);
4290 flush_workqueue(cgroup_pidlist_destroy_wq);
4291 BUG_ON(!list_empty(&cgrp->pidlists));
4294 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4296 struct delayed_work *dwork = to_delayed_work(work);
4297 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4299 struct cgroup_pidlist *tofree = NULL;
4301 mutex_lock(&l->owner->pidlist_mutex);
4304 * Destroy iff we didn't get queued again. The state won't change
4305 * as destroy_dwork can only be queued while locked.
4307 if (!delayed_work_pending(dwork)) {
4308 list_del(&l->links);
4309 pidlist_free(l->list);
4310 put_pid_ns(l->key.ns);
4314 mutex_unlock(&l->owner->pidlist_mutex);
4319 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4320 * Returns the number of unique elements.
4322 static int pidlist_uniq(pid_t *list, int length)
4327 * we presume the 0th element is unique, so i starts at 1. trivial
4328 * edge cases first; no work needs to be done for either
4330 if (length == 0 || length == 1)
4332 /* src and dest walk down the list; dest counts unique elements */
4333 for (src = 1; src < length; src++) {
4334 /* find next unique element */
4335 while (list[src] == list[src-1]) {
4340 /* dest always points to where the next unique element goes */
4341 list[dest] = list[src];
4349 * The two pid files - task and cgroup.procs - guaranteed that the result
4350 * is sorted, which forced this whole pidlist fiasco. As pid order is
4351 * different per namespace, each namespace needs differently sorted list,
4352 * making it impossible to use, for example, single rbtree of member tasks
4353 * sorted by task pointer. As pidlists can be fairly large, allocating one
4354 * per open file is dangerous, so cgroup had to implement shared pool of
4355 * pidlists keyed by cgroup and namespace.
4357 * All this extra complexity was caused by the original implementation
4358 * committing to an entirely unnecessary property. In the long term, we
4359 * want to do away with it. Explicitly scramble sort order if on the
4360 * default hierarchy so that no such expectation exists in the new
4363 * Scrambling is done by swapping every two consecutive bits, which is
4364 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4366 static pid_t pid_fry(pid_t pid)
4368 unsigned a = pid & 0x55555555;
4369 unsigned b = pid & 0xAAAAAAAA;
4371 return (a << 1) | (b >> 1);
4374 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4376 if (cgroup_on_dfl(cgrp))
4377 return pid_fry(pid);
4382 static int cmppid(const void *a, const void *b)
4384 return *(pid_t *)a - *(pid_t *)b;
4387 static int fried_cmppid(const void *a, const void *b)
4389 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4392 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4393 enum cgroup_filetype type)
4395 struct cgroup_pidlist *l;
4396 /* don't need task_nsproxy() if we're looking at ourself */
4397 struct pid_namespace *ns = task_active_pid_ns(current);
4399 lockdep_assert_held(&cgrp->pidlist_mutex);
4401 list_for_each_entry(l, &cgrp->pidlists, links)
4402 if (l->key.type == type && l->key.ns == ns)
4408 * find the appropriate pidlist for our purpose (given procs vs tasks)
4409 * returns with the lock on that pidlist already held, and takes care
4410 * of the use count, or returns NULL with no locks held if we're out of
4413 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4414 enum cgroup_filetype type)
4416 struct cgroup_pidlist *l;
4418 lockdep_assert_held(&cgrp->pidlist_mutex);
4420 l = cgroup_pidlist_find(cgrp, type);
4424 /* entry not found; create a new one */
4425 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4429 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4431 /* don't need task_nsproxy() if we're looking at ourself */
4432 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4434 list_add(&l->links, &cgrp->pidlists);
4439 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4441 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4442 struct cgroup_pidlist **lp)
4446 int pid, n = 0; /* used for populating the array */
4447 struct css_task_iter it;
4448 struct task_struct *tsk;
4449 struct cgroup_pidlist *l;
4451 lockdep_assert_held(&cgrp->pidlist_mutex);
4454 * If cgroup gets more users after we read count, we won't have
4455 * enough space - tough. This race is indistinguishable to the
4456 * caller from the case that the additional cgroup users didn't
4457 * show up until sometime later on.
4459 length = cgroup_task_count(cgrp);
4460 array = pidlist_allocate(length);
4463 /* now, populate the array */
4464 css_task_iter_start(&cgrp->self, &it);
4465 while ((tsk = css_task_iter_next(&it))) {
4466 if (unlikely(n == length))
4468 /* get tgid or pid for procs or tasks file respectively */
4469 if (type == CGROUP_FILE_PROCS)
4470 pid = task_tgid_vnr(tsk);
4472 pid = task_pid_vnr(tsk);
4473 if (pid > 0) /* make sure to only use valid results */
4476 css_task_iter_end(&it);
4478 /* now sort & (if procs) strip out duplicates */
4479 if (cgroup_on_dfl(cgrp))
4480 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4482 sort(array, length, sizeof(pid_t), cmppid, NULL);
4483 if (type == CGROUP_FILE_PROCS)
4484 length = pidlist_uniq(array, length);
4486 l = cgroup_pidlist_find_create(cgrp, type);
4488 pidlist_free(array);
4492 /* store array, freeing old if necessary */
4493 pidlist_free(l->list);
4501 * cgroupstats_build - build and fill cgroupstats
4502 * @stats: cgroupstats to fill information into
4503 * @dentry: A dentry entry belonging to the cgroup for which stats have
4506 * Build and fill cgroupstats so that taskstats can export it to user
4509 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4511 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4512 struct cgroup *cgrp;
4513 struct css_task_iter it;
4514 struct task_struct *tsk;
4516 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4517 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4518 kernfs_type(kn) != KERNFS_DIR)
4521 mutex_lock(&cgroup_mutex);
4524 * We aren't being called from kernfs and there's no guarantee on
4525 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4526 * @kn->priv is RCU safe. Let's do the RCU dancing.
4529 cgrp = rcu_dereference(kn->priv);
4530 if (!cgrp || cgroup_is_dead(cgrp)) {
4532 mutex_unlock(&cgroup_mutex);
4537 css_task_iter_start(&cgrp->self, &it);
4538 while ((tsk = css_task_iter_next(&it))) {
4539 switch (tsk->state) {
4541 stats->nr_running++;
4543 case TASK_INTERRUPTIBLE:
4544 stats->nr_sleeping++;
4546 case TASK_UNINTERRUPTIBLE:
4547 stats->nr_uninterruptible++;
4550 stats->nr_stopped++;
4553 if (delayacct_is_task_waiting_on_io(tsk))
4554 stats->nr_io_wait++;
4558 css_task_iter_end(&it);
4560 mutex_unlock(&cgroup_mutex);
4566 * seq_file methods for the tasks/procs files. The seq_file position is the
4567 * next pid to display; the seq_file iterator is a pointer to the pid
4568 * in the cgroup->l->list array.
4571 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4574 * Initially we receive a position value that corresponds to
4575 * one more than the last pid shown (or 0 on the first call or
4576 * after a seek to the start). Use a binary-search to find the
4577 * next pid to display, if any
4579 struct kernfs_open_file *of = s->private;
4580 struct cgroup *cgrp = seq_css(s)->cgroup;
4581 struct cgroup_pidlist *l;
4582 enum cgroup_filetype type = seq_cft(s)->private;
4583 int index = 0, pid = *pos;
4586 mutex_lock(&cgrp->pidlist_mutex);
4589 * !NULL @of->priv indicates that this isn't the first start()
4590 * after open. If the matching pidlist is around, we can use that.
4591 * Look for it. Note that @of->priv can't be used directly. It
4592 * could already have been destroyed.
4595 of->priv = cgroup_pidlist_find(cgrp, type);
4598 * Either this is the first start() after open or the matching
4599 * pidlist has been destroyed inbetween. Create a new one.
4602 ret = pidlist_array_load(cgrp, type,
4603 (struct cgroup_pidlist **)&of->priv);
4605 return ERR_PTR(ret);
4610 int end = l->length;
4612 while (index < end) {
4613 int mid = (index + end) / 2;
4614 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4617 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4623 /* If we're off the end of the array, we're done */
4624 if (index >= l->length)
4626 /* Update the abstract position to be the actual pid that we found */
4627 iter = l->list + index;
4628 *pos = cgroup_pid_fry(cgrp, *iter);
4632 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4634 struct kernfs_open_file *of = s->private;
4635 struct cgroup_pidlist *l = of->priv;
4638 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4639 CGROUP_PIDLIST_DESTROY_DELAY);
4640 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4643 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4645 struct kernfs_open_file *of = s->private;
4646 struct cgroup_pidlist *l = of->priv;
4648 pid_t *end = l->list + l->length;
4650 * Advance to the next pid in the array. If this goes off the
4657 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4662 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4664 seq_printf(s, "%d\n", *(int *)v);
4669 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4672 return notify_on_release(css->cgroup);
4675 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4676 struct cftype *cft, u64 val)
4679 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4681 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4685 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4688 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4691 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4692 struct cftype *cft, u64 val)
4695 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4697 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4701 /* cgroup core interface files for the default hierarchy */
4702 static struct cftype cgroup_dfl_base_files[] = {
4704 .name = "cgroup.procs",
4705 .file_offset = offsetof(struct cgroup, procs_file),
4706 .seq_start = cgroup_pidlist_start,
4707 .seq_next = cgroup_pidlist_next,
4708 .seq_stop = cgroup_pidlist_stop,
4709 .seq_show = cgroup_pidlist_show,
4710 .private = CGROUP_FILE_PROCS,
4711 .write = cgroup_procs_write,
4714 .name = "cgroup.controllers",
4715 .seq_show = cgroup_controllers_show,
4718 .name = "cgroup.subtree_control",
4719 .seq_show = cgroup_subtree_control_show,
4720 .write = cgroup_subtree_control_write,
4723 .name = "cgroup.events",
4724 .flags = CFTYPE_NOT_ON_ROOT,
4725 .file_offset = offsetof(struct cgroup, events_file),
4726 .seq_show = cgroup_events_show,
4731 /* cgroup core interface files for the legacy hierarchies */
4732 static struct cftype cgroup_legacy_base_files[] = {
4734 .name = "cgroup.procs",
4735 .seq_start = cgroup_pidlist_start,
4736 .seq_next = cgroup_pidlist_next,
4737 .seq_stop = cgroup_pidlist_stop,
4738 .seq_show = cgroup_pidlist_show,
4739 .private = CGROUP_FILE_PROCS,
4740 .write = cgroup_procs_write,
4743 .name = "cgroup.clone_children",
4744 .read_u64 = cgroup_clone_children_read,
4745 .write_u64 = cgroup_clone_children_write,
4748 .name = "cgroup.sane_behavior",
4749 .flags = CFTYPE_ONLY_ON_ROOT,
4750 .seq_show = cgroup_sane_behavior_show,
4754 .seq_start = cgroup_pidlist_start,
4755 .seq_next = cgroup_pidlist_next,
4756 .seq_stop = cgroup_pidlist_stop,
4757 .seq_show = cgroup_pidlist_show,
4758 .private = CGROUP_FILE_TASKS,
4759 .write = cgroup_tasks_write,
4762 .name = "notify_on_release",
4763 .read_u64 = cgroup_read_notify_on_release,
4764 .write_u64 = cgroup_write_notify_on_release,
4767 .name = "release_agent",
4768 .flags = CFTYPE_ONLY_ON_ROOT,
4769 .seq_show = cgroup_release_agent_show,
4770 .write = cgroup_release_agent_write,
4771 .max_write_len = PATH_MAX - 1,
4777 * css destruction is four-stage process.
4779 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4780 * Implemented in kill_css().
4782 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4783 * and thus css_tryget_online() is guaranteed to fail, the css can be
4784 * offlined by invoking offline_css(). After offlining, the base ref is
4785 * put. Implemented in css_killed_work_fn().
4787 * 3. When the percpu_ref reaches zero, the only possible remaining
4788 * accessors are inside RCU read sections. css_release() schedules the
4791 * 4. After the grace period, the css can be freed. Implemented in
4792 * css_free_work_fn().
4794 * It is actually hairier because both step 2 and 4 require process context
4795 * and thus involve punting to css->destroy_work adding two additional
4796 * steps to the already complex sequence.
4798 static void css_free_work_fn(struct work_struct *work)
4800 struct cgroup_subsys_state *css =
4801 container_of(work, struct cgroup_subsys_state, destroy_work);
4802 struct cgroup_subsys *ss = css->ss;
4803 struct cgroup *cgrp = css->cgroup;
4805 percpu_ref_exit(&css->refcnt);
4809 struct cgroup_subsys_state *parent = css->parent;
4813 cgroup_idr_remove(&ss->css_idr, id);
4819 /* cgroup free path */
4820 atomic_dec(&cgrp->root->nr_cgrps);
4821 cgroup_pidlist_destroy_all(cgrp);
4822 cancel_work_sync(&cgrp->release_agent_work);
4824 if (cgroup_parent(cgrp)) {
4826 * We get a ref to the parent, and put the ref when
4827 * this cgroup is being freed, so it's guaranteed
4828 * that the parent won't be destroyed before its
4831 cgroup_put(cgroup_parent(cgrp));
4832 kernfs_put(cgrp->kn);
4836 * This is root cgroup's refcnt reaching zero,
4837 * which indicates that the root should be
4840 cgroup_destroy_root(cgrp->root);
4845 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4847 struct cgroup_subsys_state *css =
4848 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4850 INIT_WORK(&css->destroy_work, css_free_work_fn);
4851 queue_work(cgroup_destroy_wq, &css->destroy_work);
4854 static void css_release_work_fn(struct work_struct *work)
4856 struct cgroup_subsys_state *css =
4857 container_of(work, struct cgroup_subsys_state, destroy_work);
4858 struct cgroup_subsys *ss = css->ss;
4859 struct cgroup *cgrp = css->cgroup;
4861 mutex_lock(&cgroup_mutex);
4863 css->flags |= CSS_RELEASED;
4864 list_del_rcu(&css->sibling);
4867 /* css release path */
4868 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4869 if (ss->css_released)
4870 ss->css_released(css);
4872 /* cgroup release path */
4873 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4877 * There are two control paths which try to determine
4878 * cgroup from dentry without going through kernfs -
4879 * cgroupstats_build() and css_tryget_online_from_dir().
4880 * Those are supported by RCU protecting clearing of
4881 * cgrp->kn->priv backpointer.
4884 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4888 mutex_unlock(&cgroup_mutex);
4890 call_rcu(&css->rcu_head, css_free_rcu_fn);
4893 static void css_release(struct percpu_ref *ref)
4895 struct cgroup_subsys_state *css =
4896 container_of(ref, struct cgroup_subsys_state, refcnt);
4898 INIT_WORK(&css->destroy_work, css_release_work_fn);
4899 queue_work(cgroup_destroy_wq, &css->destroy_work);
4902 static void init_and_link_css(struct cgroup_subsys_state *css,
4903 struct cgroup_subsys *ss, struct cgroup *cgrp)
4905 lockdep_assert_held(&cgroup_mutex);
4909 memset(css, 0, sizeof(*css));
4912 INIT_LIST_HEAD(&css->sibling);
4913 INIT_LIST_HEAD(&css->children);
4914 css->serial_nr = css_serial_nr_next++;
4915 atomic_set(&css->online_cnt, 0);
4917 if (cgroup_parent(cgrp)) {
4918 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4919 css_get(css->parent);
4922 BUG_ON(cgroup_css(cgrp, ss));
4925 /* invoke ->css_online() on a new CSS and mark it online if successful */
4926 static int online_css(struct cgroup_subsys_state *css)
4928 struct cgroup_subsys *ss = css->ss;
4931 lockdep_assert_held(&cgroup_mutex);
4934 ret = ss->css_online(css);
4936 css->flags |= CSS_ONLINE;
4937 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4939 atomic_inc(&css->online_cnt);
4941 atomic_inc(&css->parent->online_cnt);
4946 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4947 static void offline_css(struct cgroup_subsys_state *css)
4949 struct cgroup_subsys *ss = css->ss;
4951 lockdep_assert_held(&cgroup_mutex);
4953 if (!(css->flags & CSS_ONLINE))
4959 if (ss->css_offline)
4960 ss->css_offline(css);
4962 css->flags &= ~CSS_ONLINE;
4963 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4965 wake_up_all(&css->cgroup->offline_waitq);
4969 * css_create - create a cgroup_subsys_state
4970 * @cgrp: the cgroup new css will be associated with
4971 * @ss: the subsys of new css
4973 * Create a new css associated with @cgrp - @ss pair. On success, the new
4974 * css is online and installed in @cgrp. This function doesn't create the
4975 * interface files. Returns 0 on success, -errno on failure.
4977 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4978 struct cgroup_subsys *ss)
4980 struct cgroup *parent = cgroup_parent(cgrp);
4981 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4982 struct cgroup_subsys_state *css;
4985 lockdep_assert_held(&cgroup_mutex);
4987 css = ss->css_alloc(parent_css);
4991 init_and_link_css(css, ss, cgrp);
4993 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4997 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4999 goto err_free_percpu_ref;
5002 /* @css is ready to be brought online now, make it visible */
5003 list_add_tail_rcu(&css->sibling, &parent_css->children);
5004 cgroup_idr_replace(&ss->css_idr, css, css->id);
5006 err = online_css(css);
5010 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
5011 cgroup_parent(parent)) {
5012 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
5013 current->comm, current->pid, ss->name);
5014 if (!strcmp(ss->name, "memory"))
5015 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
5016 ss->warned_broken_hierarchy = true;
5022 list_del_rcu(&css->sibling);
5023 cgroup_idr_remove(&ss->css_idr, css->id);
5024 err_free_percpu_ref:
5025 percpu_ref_exit(&css->refcnt);
5027 call_rcu(&css->rcu_head, css_free_rcu_fn);
5028 return ERR_PTR(err);
5031 static struct cgroup *cgroup_create(struct cgroup *parent)
5033 struct cgroup_root *root = parent->root;
5034 struct cgroup *cgrp, *tcgrp;
5035 int level = parent->level + 1;
5038 /* allocate the cgroup and its ID, 0 is reserved for the root */
5039 cgrp = kzalloc(sizeof(*cgrp) +
5040 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
5042 return ERR_PTR(-ENOMEM);
5044 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5049 * Temporarily set the pointer to NULL, so idr_find() won't return
5050 * a half-baked cgroup.
5052 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
5055 goto out_cancel_ref;
5058 init_cgroup_housekeeping(cgrp);
5060 cgrp->self.parent = &parent->self;
5062 cgrp->level = level;
5064 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp))
5065 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
5067 if (notify_on_release(parent))
5068 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5070 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5071 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5073 cgrp->self.serial_nr = css_serial_nr_next++;
5075 /* allocation complete, commit to creation */
5076 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5077 atomic_inc(&root->nr_cgrps);
5081 * @cgrp is now fully operational. If something fails after this
5082 * point, it'll be released via the normal destruction path.
5084 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5087 * On the default hierarchy, a child doesn't automatically inherit
5088 * subtree_control from the parent. Each is configured manually.
5090 if (!cgroup_on_dfl(cgrp))
5091 cgrp->subtree_control = cgroup_control(cgrp);
5093 cgroup_propagate_control(cgrp);
5095 /* @cgrp doesn't have dir yet so the following will only create csses */
5096 ret = cgroup_apply_control_enable(cgrp);
5103 percpu_ref_exit(&cgrp->self.refcnt);
5106 return ERR_PTR(ret);
5108 cgroup_destroy_locked(cgrp);
5109 return ERR_PTR(ret);
5112 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
5115 struct cgroup *parent, *cgrp;
5116 struct kernfs_node *kn;
5119 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5120 if (strchr(name, '\n'))
5123 parent = cgroup_kn_lock_live(parent_kn, false);
5127 cgrp = cgroup_create(parent);
5129 ret = PTR_ERR(cgrp);
5133 /* create the directory */
5134 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5142 * This extra ref will be put in cgroup_free_fn() and guarantees
5143 * that @cgrp->kn is always accessible.
5147 ret = cgroup_kn_set_ugid(kn);
5151 ret = css_populate_dir(&cgrp->self);
5155 ret = cgroup_apply_control_enable(cgrp);
5159 /* let's create and online css's */
5160 kernfs_activate(kn);
5166 cgroup_destroy_locked(cgrp);
5168 cgroup_kn_unlock(parent_kn);
5173 * This is called when the refcnt of a css is confirmed to be killed.
5174 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5175 * initate destruction and put the css ref from kill_css().
5177 static void css_killed_work_fn(struct work_struct *work)
5179 struct cgroup_subsys_state *css =
5180 container_of(work, struct cgroup_subsys_state, destroy_work);
5182 mutex_lock(&cgroup_mutex);
5187 /* @css can't go away while we're holding cgroup_mutex */
5189 } while (css && atomic_dec_and_test(&css->online_cnt));
5191 mutex_unlock(&cgroup_mutex);
5194 /* css kill confirmation processing requires process context, bounce */
5195 static void css_killed_ref_fn(struct percpu_ref *ref)
5197 struct cgroup_subsys_state *css =
5198 container_of(ref, struct cgroup_subsys_state, refcnt);
5200 if (atomic_dec_and_test(&css->online_cnt)) {
5201 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5202 queue_work(cgroup_destroy_wq, &css->destroy_work);
5207 * kill_css - destroy a css
5208 * @css: css to destroy
5210 * This function initiates destruction of @css by removing cgroup interface
5211 * files and putting its base reference. ->css_offline() will be invoked
5212 * asynchronously once css_tryget_online() is guaranteed to fail and when
5213 * the reference count reaches zero, @css will be released.
5215 static void kill_css(struct cgroup_subsys_state *css)
5217 lockdep_assert_held(&cgroup_mutex);
5220 * This must happen before css is disassociated with its cgroup.
5221 * See seq_css() for details.
5226 * Killing would put the base ref, but we need to keep it alive
5227 * until after ->css_offline().
5232 * cgroup core guarantees that, by the time ->css_offline() is
5233 * invoked, no new css reference will be given out via
5234 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5235 * proceed to offlining css's because percpu_ref_kill() doesn't
5236 * guarantee that the ref is seen as killed on all CPUs on return.
5238 * Use percpu_ref_kill_and_confirm() to get notifications as each
5239 * css is confirmed to be seen as killed on all CPUs.
5241 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5245 * cgroup_destroy_locked - the first stage of cgroup destruction
5246 * @cgrp: cgroup to be destroyed
5248 * css's make use of percpu refcnts whose killing latency shouldn't be
5249 * exposed to userland and are RCU protected. Also, cgroup core needs to
5250 * guarantee that css_tryget_online() won't succeed by the time
5251 * ->css_offline() is invoked. To satisfy all the requirements,
5252 * destruction is implemented in the following two steps.
5254 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5255 * userland visible parts and start killing the percpu refcnts of
5256 * css's. Set up so that the next stage will be kicked off once all
5257 * the percpu refcnts are confirmed to be killed.
5259 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5260 * rest of destruction. Once all cgroup references are gone, the
5261 * cgroup is RCU-freed.
5263 * This function implements s1. After this step, @cgrp is gone as far as
5264 * the userland is concerned and a new cgroup with the same name may be
5265 * created. As cgroup doesn't care about the names internally, this
5266 * doesn't cause any problem.
5268 static int cgroup_destroy_locked(struct cgroup *cgrp)
5269 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5271 struct cgroup_subsys_state *css;
5272 struct cgrp_cset_link *link;
5275 lockdep_assert_held(&cgroup_mutex);
5278 * Only migration can raise populated from zero and we're already
5279 * holding cgroup_mutex.
5281 if (cgroup_is_populated(cgrp))
5285 * Make sure there's no live children. We can't test emptiness of
5286 * ->self.children as dead children linger on it while being
5287 * drained; otherwise, "rmdir parent/child parent" may fail.
5289 if (css_has_online_children(&cgrp->self))
5293 * Mark @cgrp and the associated csets dead. The former prevents
5294 * further task migration and child creation by disabling
5295 * cgroup_lock_live_group(). The latter makes the csets ignored by
5296 * the migration path.
5298 cgrp->self.flags &= ~CSS_ONLINE;
5300 spin_lock_bh(&css_set_lock);
5301 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5302 link->cset->dead = true;
5303 spin_unlock_bh(&css_set_lock);
5305 /* initiate massacre of all css's */
5306 for_each_css(css, ssid, cgrp)
5310 * Remove @cgrp directory along with the base files. @cgrp has an
5311 * extra ref on its kn.
5313 kernfs_remove(cgrp->kn);
5315 check_for_release(cgroup_parent(cgrp));
5317 /* put the base reference */
5318 percpu_ref_kill(&cgrp->self.refcnt);
5323 static int cgroup_rmdir(struct kernfs_node *kn)
5325 struct cgroup *cgrp;
5328 cgrp = cgroup_kn_lock_live(kn, false);
5332 ret = cgroup_destroy_locked(cgrp);
5334 cgroup_kn_unlock(kn);
5338 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5339 .remount_fs = cgroup_remount,
5340 .show_options = cgroup_show_options,
5341 .mkdir = cgroup_mkdir,
5342 .rmdir = cgroup_rmdir,
5343 .rename = cgroup_rename,
5346 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5348 struct cgroup_subsys_state *css;
5350 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5352 mutex_lock(&cgroup_mutex);
5354 idr_init(&ss->css_idr);
5355 INIT_LIST_HEAD(&ss->cfts);
5357 /* Create the root cgroup state for this subsystem */
5358 ss->root = &cgrp_dfl_root;
5359 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5360 /* We don't handle early failures gracefully */
5361 BUG_ON(IS_ERR(css));
5362 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5365 * Root csses are never destroyed and we can't initialize
5366 * percpu_ref during early init. Disable refcnting.
5368 css->flags |= CSS_NO_REF;
5371 /* allocation can't be done safely during early init */
5374 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5375 BUG_ON(css->id < 0);
5378 /* Update the init_css_set to contain a subsys
5379 * pointer to this state - since the subsystem is
5380 * newly registered, all tasks and hence the
5381 * init_css_set is in the subsystem's root cgroup. */
5382 init_css_set.subsys[ss->id] = css;
5384 have_fork_callback |= (bool)ss->fork << ss->id;
5385 have_exit_callback |= (bool)ss->exit << ss->id;
5386 have_free_callback |= (bool)ss->free << ss->id;
5387 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5389 /* At system boot, before all subsystems have been
5390 * registered, no tasks have been forked, so we don't
5391 * need to invoke fork callbacks here. */
5392 BUG_ON(!list_empty(&init_task.tasks));
5394 BUG_ON(online_css(css));
5396 mutex_unlock(&cgroup_mutex);
5400 * cgroup_init_early - cgroup initialization at system boot
5402 * Initialize cgroups at system boot, and initialize any
5403 * subsystems that request early init.
5405 int __init cgroup_init_early(void)
5407 static struct cgroup_sb_opts __initdata opts;
5408 struct cgroup_subsys *ss;
5411 init_cgroup_root(&cgrp_dfl_root, &opts);
5412 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5414 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5416 for_each_subsys(ss, i) {
5417 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5418 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5419 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5421 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5422 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5425 ss->name = cgroup_subsys_name[i];
5426 if (!ss->legacy_name)
5427 ss->legacy_name = cgroup_subsys_name[i];
5430 cgroup_init_subsys(ss, true);
5435 static u16 cgroup_disable_mask __initdata;
5438 * cgroup_init - cgroup initialization
5440 * Register cgroup filesystem and /proc file, and initialize
5441 * any subsystems that didn't request early init.
5443 int __init cgroup_init(void)
5445 struct cgroup_subsys *ss;
5448 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5449 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5450 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5451 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5453 mutex_lock(&cgroup_mutex);
5456 * Add init_css_set to the hash table so that dfl_root can link to
5459 hash_add(css_set_table, &init_css_set.hlist,
5460 css_set_hash(init_css_set.subsys));
5462 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5464 mutex_unlock(&cgroup_mutex);
5466 for_each_subsys(ss, ssid) {
5467 if (ss->early_init) {
5468 struct cgroup_subsys_state *css =
5469 init_css_set.subsys[ss->id];
5471 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5473 BUG_ON(css->id < 0);
5475 cgroup_init_subsys(ss, false);
5478 list_add_tail(&init_css_set.e_cset_node[ssid],
5479 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5482 * Setting dfl_root subsys_mask needs to consider the
5483 * disabled flag and cftype registration needs kmalloc,
5484 * both of which aren't available during early_init.
5486 if (cgroup_disable_mask & (1 << ssid)) {
5487 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5488 printk(KERN_INFO "Disabling %s control group subsystem\n",
5493 if (cgroup_ssid_no_v1(ssid))
5494 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5497 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5499 if (ss->implicit_on_dfl)
5500 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5501 else if (!ss->dfl_cftypes)
5502 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5504 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5505 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5507 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5508 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5512 ss->bind(init_css_set.subsys[ssid]);
5515 /* init_css_set.subsys[] has been updated, re-hash */
5516 hash_del(&init_css_set.hlist);
5517 hash_add(css_set_table, &init_css_set.hlist,
5518 css_set_hash(init_css_set.subsys));
5520 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5521 WARN_ON(register_filesystem(&cgroup_fs_type));
5522 WARN_ON(register_filesystem(&cgroup2_fs_type));
5523 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5528 static int __init cgroup_wq_init(void)
5531 * There isn't much point in executing destruction path in
5532 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5533 * Use 1 for @max_active.
5535 * We would prefer to do this in cgroup_init() above, but that
5536 * is called before init_workqueues(): so leave this until after.
5538 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5539 BUG_ON(!cgroup_destroy_wq);
5542 * Used to destroy pidlists and separate to serve as flush domain.
5543 * Cap @max_active to 1 too.
5545 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5547 BUG_ON(!cgroup_pidlist_destroy_wq);
5551 core_initcall(cgroup_wq_init);
5554 * proc_cgroup_show()
5555 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5556 * - Used for /proc/<pid>/cgroup.
5558 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5559 struct pid *pid, struct task_struct *tsk)
5563 struct cgroup_root *root;
5566 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5570 mutex_lock(&cgroup_mutex);
5571 spin_lock_bh(&css_set_lock);
5573 for_each_root(root) {
5574 struct cgroup_subsys *ss;
5575 struct cgroup *cgrp;
5576 int ssid, count = 0;
5578 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5581 seq_printf(m, "%d:", root->hierarchy_id);
5582 if (root != &cgrp_dfl_root)
5583 for_each_subsys(ss, ssid)
5584 if (root->subsys_mask & (1 << ssid))
5585 seq_printf(m, "%s%s", count++ ? "," : "",
5587 if (strlen(root->name))
5588 seq_printf(m, "%sname=%s", count ? "," : "",
5592 cgrp = task_cgroup_from_root(tsk, root);
5595 * On traditional hierarchies, all zombie tasks show up as
5596 * belonging to the root cgroup. On the default hierarchy,
5597 * while a zombie doesn't show up in "cgroup.procs" and
5598 * thus can't be migrated, its /proc/PID/cgroup keeps
5599 * reporting the cgroup it belonged to before exiting. If
5600 * the cgroup is removed before the zombie is reaped,
5601 * " (deleted)" is appended to the cgroup path.
5603 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5604 path = cgroup_path(cgrp, buf, PATH_MAX);
5606 retval = -ENAMETOOLONG;
5615 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5616 seq_puts(m, " (deleted)\n");
5623 spin_unlock_bh(&css_set_lock);
5624 mutex_unlock(&cgroup_mutex);
5630 /* Display information about each subsystem and each hierarchy */
5631 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5633 struct cgroup_subsys *ss;
5636 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5638 * ideally we don't want subsystems moving around while we do this.
5639 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5640 * subsys/hierarchy state.
5642 mutex_lock(&cgroup_mutex);
5644 for_each_subsys(ss, i)
5645 seq_printf(m, "%s\t%d\t%d\t%d\n",
5646 ss->legacy_name, ss->root->hierarchy_id,
5647 atomic_read(&ss->root->nr_cgrps),
5648 cgroup_ssid_enabled(i));
5650 mutex_unlock(&cgroup_mutex);
5654 static int cgroupstats_open(struct inode *inode, struct file *file)
5656 return single_open(file, proc_cgroupstats_show, NULL);
5659 static const struct file_operations proc_cgroupstats_operations = {
5660 .open = cgroupstats_open,
5662 .llseek = seq_lseek,
5663 .release = single_release,
5667 * cgroup_fork - initialize cgroup related fields during copy_process()
5668 * @child: pointer to task_struct of forking parent process.
5670 * A task is associated with the init_css_set until cgroup_post_fork()
5671 * attaches it to the parent's css_set. Empty cg_list indicates that
5672 * @child isn't holding reference to its css_set.
5674 void cgroup_fork(struct task_struct *child)
5676 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5677 INIT_LIST_HEAD(&child->cg_list);
5681 * cgroup_can_fork - called on a new task before the process is exposed
5682 * @child: the task in question.
5684 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5685 * returns an error, the fork aborts with that error code. This allows for
5686 * a cgroup subsystem to conditionally allow or deny new forks.
5688 int cgroup_can_fork(struct task_struct *child)
5690 struct cgroup_subsys *ss;
5693 do_each_subsys_mask(ss, i, have_canfork_callback) {
5694 ret = ss->can_fork(child);
5697 } while_each_subsys_mask();
5702 for_each_subsys(ss, j) {
5705 if (ss->cancel_fork)
5706 ss->cancel_fork(child);
5713 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5714 * @child: the task in question
5716 * This calls the cancel_fork() callbacks if a fork failed *after*
5717 * cgroup_can_fork() succeded.
5719 void cgroup_cancel_fork(struct task_struct *child)
5721 struct cgroup_subsys *ss;
5724 for_each_subsys(ss, i)
5725 if (ss->cancel_fork)
5726 ss->cancel_fork(child);
5730 * cgroup_post_fork - called on a new task after adding it to the task list
5731 * @child: the task in question
5733 * Adds the task to the list running through its css_set if necessary and
5734 * call the subsystem fork() callbacks. Has to be after the task is
5735 * visible on the task list in case we race with the first call to
5736 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5739 void cgroup_post_fork(struct task_struct *child)
5741 struct cgroup_subsys *ss;
5745 * This may race against cgroup_enable_task_cg_lists(). As that
5746 * function sets use_task_css_set_links before grabbing
5747 * tasklist_lock and we just went through tasklist_lock to add
5748 * @child, it's guaranteed that either we see the set
5749 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5750 * @child during its iteration.
5752 * If we won the race, @child is associated with %current's
5753 * css_set. Grabbing css_set_lock guarantees both that the
5754 * association is stable, and, on completion of the parent's
5755 * migration, @child is visible in the source of migration or
5756 * already in the destination cgroup. This guarantee is necessary
5757 * when implementing operations which need to migrate all tasks of
5758 * a cgroup to another.
5760 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5761 * will remain in init_css_set. This is safe because all tasks are
5762 * in the init_css_set before cg_links is enabled and there's no
5763 * operation which transfers all tasks out of init_css_set.
5765 if (use_task_css_set_links) {
5766 struct css_set *cset;
5768 spin_lock_bh(&css_set_lock);
5769 cset = task_css_set(current);
5770 if (list_empty(&child->cg_list)) {
5772 css_set_move_task(child, NULL, cset, false);
5774 spin_unlock_bh(&css_set_lock);
5778 * Call ss->fork(). This must happen after @child is linked on
5779 * css_set; otherwise, @child might change state between ->fork()
5780 * and addition to css_set.
5782 do_each_subsys_mask(ss, i, have_fork_callback) {
5784 } while_each_subsys_mask();
5788 * cgroup_exit - detach cgroup from exiting task
5789 * @tsk: pointer to task_struct of exiting process
5791 * Description: Detach cgroup from @tsk and release it.
5793 * Note that cgroups marked notify_on_release force every task in
5794 * them to take the global cgroup_mutex mutex when exiting.
5795 * This could impact scaling on very large systems. Be reluctant to
5796 * use notify_on_release cgroups where very high task exit scaling
5797 * is required on large systems.
5799 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5800 * call cgroup_exit() while the task is still competent to handle
5801 * notify_on_release(), then leave the task attached to the root cgroup in
5802 * each hierarchy for the remainder of its exit. No need to bother with
5803 * init_css_set refcnting. init_css_set never goes away and we can't race
5804 * with migration path - PF_EXITING is visible to migration path.
5806 void cgroup_exit(struct task_struct *tsk)
5808 struct cgroup_subsys *ss;
5809 struct css_set *cset;
5813 * Unlink from @tsk from its css_set. As migration path can't race
5814 * with us, we can check css_set and cg_list without synchronization.
5816 cset = task_css_set(tsk);
5818 if (!list_empty(&tsk->cg_list)) {
5819 spin_lock_bh(&css_set_lock);
5820 css_set_move_task(tsk, cset, NULL, false);
5821 spin_unlock_bh(&css_set_lock);
5826 /* see cgroup_post_fork() for details */
5827 do_each_subsys_mask(ss, i, have_exit_callback) {
5829 } while_each_subsys_mask();
5832 void cgroup_free(struct task_struct *task)
5834 struct css_set *cset = task_css_set(task);
5835 struct cgroup_subsys *ss;
5838 do_each_subsys_mask(ss, ssid, have_free_callback) {
5840 } while_each_subsys_mask();
5845 static void check_for_release(struct cgroup *cgrp)
5847 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5848 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5849 schedule_work(&cgrp->release_agent_work);
5853 * Notify userspace when a cgroup is released, by running the
5854 * configured release agent with the name of the cgroup (path
5855 * relative to the root of cgroup file system) as the argument.
5857 * Most likely, this user command will try to rmdir this cgroup.
5859 * This races with the possibility that some other task will be
5860 * attached to this cgroup before it is removed, or that some other
5861 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5862 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5863 * unused, and this cgroup will be reprieved from its death sentence,
5864 * to continue to serve a useful existence. Next time it's released,
5865 * we will get notified again, if it still has 'notify_on_release' set.
5867 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5868 * means only wait until the task is successfully execve()'d. The
5869 * separate release agent task is forked by call_usermodehelper(),
5870 * then control in this thread returns here, without waiting for the
5871 * release agent task. We don't bother to wait because the caller of
5872 * this routine has no use for the exit status of the release agent
5873 * task, so no sense holding our caller up for that.
5875 static void cgroup_release_agent(struct work_struct *work)
5877 struct cgroup *cgrp =
5878 container_of(work, struct cgroup, release_agent_work);
5879 char *pathbuf = NULL, *agentbuf = NULL, *path;
5880 char *argv[3], *envp[3];
5882 mutex_lock(&cgroup_mutex);
5884 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5885 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5886 if (!pathbuf || !agentbuf)
5889 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5897 /* minimal command environment */
5899 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5902 mutex_unlock(&cgroup_mutex);
5903 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5906 mutex_unlock(&cgroup_mutex);
5912 static int __init cgroup_disable(char *str)
5914 struct cgroup_subsys *ss;
5918 while ((token = strsep(&str, ",")) != NULL) {
5922 for_each_subsys(ss, i) {
5923 if (strcmp(token, ss->name) &&
5924 strcmp(token, ss->legacy_name))
5926 cgroup_disable_mask |= 1 << i;
5931 __setup("cgroup_disable=", cgroup_disable);
5933 static int __init cgroup_no_v1(char *str)
5935 struct cgroup_subsys *ss;
5939 while ((token = strsep(&str, ",")) != NULL) {
5943 if (!strcmp(token, "all")) {
5944 cgroup_no_v1_mask = U16_MAX;
5948 for_each_subsys(ss, i) {
5949 if (strcmp(token, ss->name) &&
5950 strcmp(token, ss->legacy_name))
5953 cgroup_no_v1_mask |= 1 << i;
5958 __setup("cgroup_no_v1=", cgroup_no_v1);
5961 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5962 * @dentry: directory dentry of interest
5963 * @ss: subsystem of interest
5965 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5966 * to get the corresponding css and return it. If such css doesn't exist
5967 * or can't be pinned, an ERR_PTR value is returned.
5969 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5970 struct cgroup_subsys *ss)
5972 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5973 struct file_system_type *s_type = dentry->d_sb->s_type;
5974 struct cgroup_subsys_state *css = NULL;
5975 struct cgroup *cgrp;
5977 /* is @dentry a cgroup dir? */
5978 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5979 !kn || kernfs_type(kn) != KERNFS_DIR)
5980 return ERR_PTR(-EBADF);
5985 * This path doesn't originate from kernfs and @kn could already
5986 * have been or be removed at any point. @kn->priv is RCU
5987 * protected for this access. See css_release_work_fn() for details.
5989 cgrp = rcu_dereference(kn->priv);
5991 css = cgroup_css(cgrp, ss);
5993 if (!css || !css_tryget_online(css))
5994 css = ERR_PTR(-ENOENT);
6001 * css_from_id - lookup css by id
6002 * @id: the cgroup id
6003 * @ss: cgroup subsys to be looked into
6005 * Returns the css if there's valid one with @id, otherwise returns NULL.
6006 * Should be called under rcu_read_lock().
6008 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6010 WARN_ON_ONCE(!rcu_read_lock_held());
6011 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
6015 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6016 * @path: path on the default hierarchy
6018 * Find the cgroup at @path on the default hierarchy, increment its
6019 * reference count and return it. Returns pointer to the found cgroup on
6020 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
6021 * if @path points to a non-directory.
6023 struct cgroup *cgroup_get_from_path(const char *path)
6025 struct kernfs_node *kn;
6026 struct cgroup *cgrp;
6028 mutex_lock(&cgroup_mutex);
6030 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
6032 if (kernfs_type(kn) == KERNFS_DIR) {
6036 cgrp = ERR_PTR(-ENOTDIR);
6040 cgrp = ERR_PTR(-ENOENT);
6043 mutex_unlock(&cgroup_mutex);
6046 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6049 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6050 * definition in cgroup-defs.h.
6052 #ifdef CONFIG_SOCK_CGROUP_DATA
6054 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
6056 DEFINE_SPINLOCK(cgroup_sk_update_lock);
6057 static bool cgroup_sk_alloc_disabled __read_mostly;
6059 void cgroup_sk_alloc_disable(void)
6061 if (cgroup_sk_alloc_disabled)
6063 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
6064 cgroup_sk_alloc_disabled = true;
6069 #define cgroup_sk_alloc_disabled false
6073 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6075 if (cgroup_sk_alloc_disabled)
6081 struct css_set *cset;
6083 cset = task_css_set(current);
6084 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6085 skcd->val = (unsigned long)cset->dfl_cgrp;
6094 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6096 cgroup_put(sock_cgroup_ptr(skcd));
6099 #endif /* CONFIG_SOCK_CGROUP_DATA */
6101 #ifdef CONFIG_CGROUP_DEBUG
6102 static struct cgroup_subsys_state *
6103 debug_css_alloc(struct cgroup_subsys_state *parent_css)
6105 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
6108 return ERR_PTR(-ENOMEM);
6113 static void debug_css_free(struct cgroup_subsys_state *css)
6118 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
6121 return cgroup_task_count(css->cgroup);
6124 static u64 current_css_set_read(struct cgroup_subsys_state *css,
6127 return (u64)(unsigned long)current->cgroups;
6130 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
6136 count = atomic_read(&task_css_set(current)->refcount);
6141 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
6143 struct cgrp_cset_link *link;
6144 struct css_set *cset;
6147 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
6151 spin_lock_bh(&css_set_lock);
6153 cset = rcu_dereference(current->cgroups);
6154 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
6155 struct cgroup *c = link->cgrp;
6157 cgroup_name(c, name_buf, NAME_MAX + 1);
6158 seq_printf(seq, "Root %d group %s\n",
6159 c->root->hierarchy_id, name_buf);
6162 spin_unlock_bh(&css_set_lock);
6167 #define MAX_TASKS_SHOWN_PER_CSS 25
6168 static int cgroup_css_links_read(struct seq_file *seq, void *v)
6170 struct cgroup_subsys_state *css = seq_css(seq);
6171 struct cgrp_cset_link *link;
6173 spin_lock_bh(&css_set_lock);
6174 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
6175 struct css_set *cset = link->cset;
6176 struct task_struct *task;
6179 seq_printf(seq, "css_set %p\n", cset);
6181 list_for_each_entry(task, &cset->tasks, cg_list) {
6182 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6184 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6187 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
6188 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
6190 seq_printf(seq, " task %d\n", task_pid_vnr(task));
6194 seq_puts(seq, " ...\n");
6196 spin_unlock_bh(&css_set_lock);
6200 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
6202 return (!cgroup_is_populated(css->cgroup) &&
6203 !css_has_online_children(&css->cgroup->self));
6206 static struct cftype debug_files[] = {
6208 .name = "taskcount",
6209 .read_u64 = debug_taskcount_read,
6213 .name = "current_css_set",
6214 .read_u64 = current_css_set_read,
6218 .name = "current_css_set_refcount",
6219 .read_u64 = current_css_set_refcount_read,
6223 .name = "current_css_set_cg_links",
6224 .seq_show = current_css_set_cg_links_read,
6228 .name = "cgroup_css_links",
6229 .seq_show = cgroup_css_links_read,
6233 .name = "releasable",
6234 .read_u64 = releasable_read,
6240 struct cgroup_subsys debug_cgrp_subsys = {
6241 .css_alloc = debug_css_alloc,
6242 .css_free = debug_css_free,
6243 .legacy_cftypes = debug_files,
6245 #endif /* CONFIG_CGROUP_DEBUG */