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 "cgroup-internal.h"
33 #include <linux/cred.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/magic.h>
38 #include <linux/mutex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/sched/task.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/percpu-rwsem.h>
48 #include <linux/string.h>
49 #include <linux/hashtable.h>
50 #include <linux/idr.h>
51 #include <linux/kthread.h>
52 #include <linux/atomic.h>
53 #include <linux/cpuset.h>
54 #include <linux/proc_ns.h>
55 #include <linux/nsproxy.h>
56 #include <linux/file.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/cgroup.h>
62 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
66 * cgroup_mutex is the master lock. Any modification to cgroup or its
67 * hierarchy must be performed while holding it.
69 * css_set_lock protects task->cgroups pointer, the list of css_set
70 * objects, and the chain of tasks off each css_set.
72 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
73 * cgroup.h can use them for lockdep annotations.
75 DEFINE_MUTEX(cgroup_mutex);
76 DEFINE_SPINLOCK(css_set_lock);
78 #ifdef CONFIG_PROVE_RCU
79 EXPORT_SYMBOL_GPL(cgroup_mutex);
80 EXPORT_SYMBOL_GPL(css_set_lock);
84 * Protects cgroup_idr and css_idr so that IDs can be released without
85 * grabbing cgroup_mutex.
87 static DEFINE_SPINLOCK(cgroup_idr_lock);
90 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
91 * against file removal/re-creation across css hiding.
93 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
95 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
97 #define cgroup_assert_mutex_or_rcu_locked() \
98 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
99 !lockdep_is_held(&cgroup_mutex), \
100 "cgroup_mutex or RCU read lock required");
103 * cgroup destruction makes heavy use of work items and there can be a lot
104 * of concurrent destructions. Use a separate workqueue so that cgroup
105 * destruction work items don't end up filling up max_active of system_wq
106 * which may lead to deadlock.
108 static struct workqueue_struct *cgroup_destroy_wq;
110 /* generate an array of cgroup subsystem pointers */
111 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
112 struct cgroup_subsys *cgroup_subsys[] = {
113 #include <linux/cgroup_subsys.h>
117 /* array of cgroup subsystem names */
118 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
119 static const char *cgroup_subsys_name[] = {
120 #include <linux/cgroup_subsys.h>
124 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
126 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
127 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
128 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
129 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
130 #include <linux/cgroup_subsys.h>
133 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
134 static struct static_key_true *cgroup_subsys_enabled_key[] = {
135 #include <linux/cgroup_subsys.h>
139 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
140 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
141 #include <linux/cgroup_subsys.h>
146 * The default hierarchy, reserved for the subsystems that are otherwise
147 * unattached - it never has more than a single cgroup, and all tasks are
148 * part of that cgroup.
150 struct cgroup_root cgrp_dfl_root;
151 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
154 * The default hierarchy always exists but is hidden until mounted for the
155 * first time. This is for backward compatibility.
157 static bool cgrp_dfl_visible;
159 /* some controllers are not supported in the default hierarchy */
160 static u16 cgrp_dfl_inhibit_ss_mask;
162 /* some controllers are implicitly enabled on the default hierarchy */
163 static u16 cgrp_dfl_implicit_ss_mask;
165 /* some controllers can be threaded on the default hierarchy */
166 static u16 cgrp_dfl_threaded_ss_mask;
168 /* The list of hierarchy roots */
169 LIST_HEAD(cgroup_roots);
170 static int cgroup_root_count;
172 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
173 static DEFINE_IDR(cgroup_hierarchy_idr);
176 * Assign a monotonically increasing serial number to csses. It guarantees
177 * cgroups with bigger numbers are newer than those with smaller numbers.
178 * Also, as csses are always appended to the parent's ->children list, it
179 * guarantees that sibling csses are always sorted in the ascending serial
180 * number order on the list. Protected by cgroup_mutex.
182 static u64 css_serial_nr_next = 1;
185 * These bitmasks identify subsystems with specific features to avoid
186 * having to do iterative checks repeatedly.
188 static u16 have_fork_callback __read_mostly;
189 static u16 have_exit_callback __read_mostly;
190 static u16 have_free_callback __read_mostly;
191 static u16 have_canfork_callback __read_mostly;
193 /* cgroup namespace for init task */
194 struct cgroup_namespace init_cgroup_ns = {
195 .count = REFCOUNT_INIT(2),
196 .user_ns = &init_user_ns,
197 .ns.ops = &cgroupns_operations,
198 .ns.inum = PROC_CGROUP_INIT_INO,
199 .root_cset = &init_css_set,
202 static struct file_system_type cgroup2_fs_type;
203 static struct cftype cgroup_base_files[];
205 static int cgroup_apply_control(struct cgroup *cgrp);
206 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
207 static void css_task_iter_advance(struct css_task_iter *it);
208 static int cgroup_destroy_locked(struct cgroup *cgrp);
209 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
210 struct cgroup_subsys *ss);
211 static void css_release(struct percpu_ref *ref);
212 static void kill_css(struct cgroup_subsys_state *css);
213 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
214 struct cgroup *cgrp, struct cftype cfts[],
218 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
219 * @ssid: subsys ID of interest
221 * cgroup_subsys_enabled() can only be used with literal subsys names which
222 * is fine for individual subsystems but unsuitable for cgroup core. This
223 * is slower static_key_enabled() based test indexed by @ssid.
225 bool cgroup_ssid_enabled(int ssid)
227 if (CGROUP_SUBSYS_COUNT == 0)
230 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
234 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
235 * @cgrp: the cgroup of interest
237 * The default hierarchy is the v2 interface of cgroup and this function
238 * can be used to test whether a cgroup is on the default hierarchy for
239 * cases where a subsystem should behave differnetly depending on the
242 * The set of behaviors which change on the default hierarchy are still
243 * being determined and the mount option is prefixed with __DEVEL__.
245 * List of changed behaviors:
247 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
248 * and "name" are disallowed.
250 * - When mounting an existing superblock, mount options should match.
252 * - Remount is disallowed.
254 * - rename(2) is disallowed.
256 * - "tasks" is removed. Everything should be at process granularity. Use
257 * "cgroup.procs" instead.
259 * - "cgroup.procs" is not sorted. pids will be unique unless they got
260 * recycled inbetween reads.
262 * - "release_agent" and "notify_on_release" are removed. Replacement
263 * notification mechanism will be implemented.
265 * - "cgroup.clone_children" is removed.
267 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
268 * and its descendants contain no task; otherwise, 1. The file also
269 * generates kernfs notification which can be monitored through poll and
270 * [di]notify when the value of the file changes.
272 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
273 * take masks of ancestors with non-empty cpus/mems, instead of being
274 * moved to an ancestor.
276 * - cpuset: a task can be moved into an empty cpuset, and again it takes
277 * masks of ancestors.
279 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
282 * - blkcg: blk-throttle becomes properly hierarchical.
284 * - debug: disallowed on the default hierarchy.
286 bool cgroup_on_dfl(const struct cgroup *cgrp)
288 return cgrp->root == &cgrp_dfl_root;
291 /* IDR wrappers which synchronize using cgroup_idr_lock */
292 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
297 idr_preload(gfp_mask);
298 spin_lock_bh(&cgroup_idr_lock);
299 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
300 spin_unlock_bh(&cgroup_idr_lock);
305 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
309 spin_lock_bh(&cgroup_idr_lock);
310 ret = idr_replace(idr, ptr, id);
311 spin_unlock_bh(&cgroup_idr_lock);
315 static void cgroup_idr_remove(struct idr *idr, int id)
317 spin_lock_bh(&cgroup_idr_lock);
319 spin_unlock_bh(&cgroup_idr_lock);
322 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
324 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
327 return container_of(parent_css, struct cgroup, self);
331 static bool cgroup_has_tasks(struct cgroup *cgrp)
333 return cgrp->nr_populated_csets;
336 bool cgroup_is_threaded(struct cgroup *cgrp)
338 return cgrp->dom_cgrp != cgrp;
341 /* can @cgrp host both domain and threaded children? */
342 static bool cgroup_is_mixable(struct cgroup *cgrp)
345 * Root isn't under domain level resource control exempting it from
346 * the no-internal-process constraint, so it can serve as a thread
347 * root and a parent of resource domains at the same time.
349 return !cgroup_parent(cgrp);
352 /* can @cgrp become a thread root? should always be true for a thread root */
353 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
355 /* mixables don't care */
356 if (cgroup_is_mixable(cgrp))
359 /* domain roots can't be nested under threaded */
360 if (cgroup_is_threaded(cgrp))
363 /* can only have either domain or threaded children */
364 if (cgrp->nr_populated_domain_children)
367 /* and no domain controllers can be enabled */
368 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
374 /* is @cgrp root of a threaded subtree? */
375 bool cgroup_is_thread_root(struct cgroup *cgrp)
377 /* thread root should be a domain */
378 if (cgroup_is_threaded(cgrp))
381 /* a domain w/ threaded children is a thread root */
382 if (cgrp->nr_threaded_children)
386 * A domain which has tasks and explicit threaded controllers
387 * enabled is a thread root.
389 if (cgroup_has_tasks(cgrp) &&
390 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
396 /* a domain which isn't connected to the root w/o brekage can't be used */
397 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
399 /* the cgroup itself can be a thread root */
400 if (cgroup_is_threaded(cgrp))
403 /* but the ancestors can't be unless mixable */
404 while ((cgrp = cgroup_parent(cgrp))) {
405 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
407 if (cgroup_is_threaded(cgrp))
414 /* subsystems visibly enabled on a cgroup */
415 static u16 cgroup_control(struct cgroup *cgrp)
417 struct cgroup *parent = cgroup_parent(cgrp);
418 u16 root_ss_mask = cgrp->root->subsys_mask;
421 u16 ss_mask = parent->subtree_control;
423 /* threaded cgroups can only have threaded controllers */
424 if (cgroup_is_threaded(cgrp))
425 ss_mask &= cgrp_dfl_threaded_ss_mask;
429 if (cgroup_on_dfl(cgrp))
430 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
431 cgrp_dfl_implicit_ss_mask);
435 /* subsystems enabled on a cgroup */
436 static u16 cgroup_ss_mask(struct cgroup *cgrp)
438 struct cgroup *parent = cgroup_parent(cgrp);
441 u16 ss_mask = parent->subtree_ss_mask;
443 /* threaded cgroups can only have threaded controllers */
444 if (cgroup_is_threaded(cgrp))
445 ss_mask &= cgrp_dfl_threaded_ss_mask;
449 return cgrp->root->subsys_mask;
453 * cgroup_css - obtain a cgroup's css for the specified subsystem
454 * @cgrp: the cgroup of interest
455 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
457 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
458 * function must be called either under cgroup_mutex or rcu_read_lock() and
459 * the caller is responsible for pinning the returned css if it wants to
460 * keep accessing it outside the said locks. This function may return
461 * %NULL if @cgrp doesn't have @subsys_id enabled.
463 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
464 struct cgroup_subsys *ss)
467 return rcu_dereference_check(cgrp->subsys[ss->id],
468 lockdep_is_held(&cgroup_mutex));
474 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
475 * @cgrp: the cgroup of interest
476 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
478 * Similar to cgroup_css() but returns the effective css, which is defined
479 * as the matching css of the nearest ancestor including self which has @ss
480 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
481 * function is guaranteed to return non-NULL css.
483 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
484 struct cgroup_subsys *ss)
486 lockdep_assert_held(&cgroup_mutex);
492 * This function is used while updating css associations and thus
493 * can't test the csses directly. Test ss_mask.
495 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
496 cgrp = cgroup_parent(cgrp);
501 return cgroup_css(cgrp, ss);
505 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
506 * @cgrp: the cgroup of interest
507 * @ss: the subsystem of interest
509 * Find and get the effective css of @cgrp for @ss. The effective css is
510 * defined as the matching css of the nearest ancestor including self which
511 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
512 * the root css is returned, so this function always returns a valid css.
513 * The returned css must be put using css_put().
515 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
516 struct cgroup_subsys *ss)
518 struct cgroup_subsys_state *css;
523 css = cgroup_css(cgrp, ss);
525 if (css && css_tryget_online(css))
527 cgrp = cgroup_parent(cgrp);
530 css = init_css_set.subsys[ss->id];
537 static void __maybe_unused cgroup_get(struct cgroup *cgrp)
539 css_get(&cgrp->self);
542 static void cgroup_get_live(struct cgroup *cgrp)
544 WARN_ON_ONCE(cgroup_is_dead(cgrp));
545 css_get(&cgrp->self);
548 static bool cgroup_tryget(struct cgroup *cgrp)
550 return css_tryget(&cgrp->self);
553 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
555 struct cgroup *cgrp = of->kn->parent->priv;
556 struct cftype *cft = of_cft(of);
559 * This is open and unprotected implementation of cgroup_css().
560 * seq_css() is only called from a kernfs file operation which has
561 * an active reference on the file. Because all the subsystem
562 * files are drained before a css is disassociated with a cgroup,
563 * the matching css from the cgroup's subsys table is guaranteed to
564 * be and stay valid until the enclosing operation is complete.
567 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
571 EXPORT_SYMBOL_GPL(of_css);
574 * for_each_css - iterate all css's of a cgroup
575 * @css: the iteration cursor
576 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
577 * @cgrp: the target cgroup to iterate css's of
579 * Should be called under cgroup_[tree_]mutex.
581 #define for_each_css(css, ssid, cgrp) \
582 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
583 if (!((css) = rcu_dereference_check( \
584 (cgrp)->subsys[(ssid)], \
585 lockdep_is_held(&cgroup_mutex)))) { } \
589 * for_each_e_css - iterate all effective css's of a cgroup
590 * @css: the iteration cursor
591 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
592 * @cgrp: the target cgroup to iterate css's of
594 * Should be called under cgroup_[tree_]mutex.
596 #define for_each_e_css(css, ssid, cgrp) \
597 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
598 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
603 * do_each_subsys_mask - filter for_each_subsys with a bitmask
604 * @ss: the iteration cursor
605 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
606 * @ss_mask: the bitmask
608 * The block will only run for cases where the ssid-th bit (1 << ssid) of
611 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
612 unsigned long __ss_mask = (ss_mask); \
613 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
617 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
618 (ss) = cgroup_subsys[ssid]; \
621 #define while_each_subsys_mask() \
626 /* iterate over child cgrps, lock should be held throughout iteration */
627 #define cgroup_for_each_live_child(child, cgrp) \
628 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
629 if (({ lockdep_assert_held(&cgroup_mutex); \
630 cgroup_is_dead(child); })) \
634 /* walk live descendants in preorder */
635 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
636 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
637 if (({ lockdep_assert_held(&cgroup_mutex); \
638 (dsct) = (d_css)->cgroup; \
639 cgroup_is_dead(dsct); })) \
643 /* walk live descendants in postorder */
644 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
645 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
646 if (({ lockdep_assert_held(&cgroup_mutex); \
647 (dsct) = (d_css)->cgroup; \
648 cgroup_is_dead(dsct); })) \
653 * The default css_set - used by init and its children prior to any
654 * hierarchies being mounted. It contains a pointer to the root state
655 * for each subsystem. Also used to anchor the list of css_sets. Not
656 * reference-counted, to improve performance when child cgroups
657 * haven't been created.
659 struct css_set init_css_set = {
660 .refcount = REFCOUNT_INIT(1),
661 .dom_cset = &init_css_set,
662 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
663 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
664 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
665 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
666 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
667 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
668 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
671 static int css_set_count = 1; /* 1 for init_css_set */
673 static bool css_set_threaded(struct css_set *cset)
675 return cset->dom_cset != cset;
679 * css_set_populated - does a css_set contain any tasks?
680 * @cset: target css_set
682 * css_set_populated() should be the same as !!cset->nr_tasks at steady
683 * state. However, css_set_populated() can be called while a task is being
684 * added to or removed from the linked list before the nr_tasks is
685 * properly updated. Hence, we can't just look at ->nr_tasks here.
687 static bool css_set_populated(struct css_set *cset)
689 lockdep_assert_held(&css_set_lock);
691 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
695 * cgroup_update_populated - update the populated count of a cgroup
696 * @cgrp: the target cgroup
697 * @populated: inc or dec populated count
699 * One of the css_sets associated with @cgrp is either getting its first
700 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
701 * count is propagated towards root so that a given cgroup's
702 * nr_populated_children is zero iff none of its descendants contain any
705 * @cgrp's interface file "cgroup.populated" is zero if both
706 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
707 * 1 otherwise. When the sum changes from or to zero, userland is notified
708 * that the content of the interface file has changed. This can be used to
709 * detect when @cgrp and its descendants become populated or empty.
711 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
713 struct cgroup *child = NULL;
714 int adj = populated ? 1 : -1;
716 lockdep_assert_held(&css_set_lock);
719 bool was_populated = cgroup_is_populated(cgrp);
722 cgrp->nr_populated_csets += adj;
724 if (cgroup_is_threaded(child))
725 cgrp->nr_populated_threaded_children += adj;
727 cgrp->nr_populated_domain_children += adj;
730 if (was_populated == cgroup_is_populated(cgrp))
733 cgroup1_check_for_release(cgrp);
734 cgroup_file_notify(&cgrp->events_file);
737 cgrp = cgroup_parent(cgrp);
742 * css_set_update_populated - update populated state of a css_set
743 * @cset: target css_set
744 * @populated: whether @cset is populated or depopulated
746 * @cset is either getting the first task or losing the last. Update the
747 * populated counters of all associated cgroups accordingly.
749 static void css_set_update_populated(struct css_set *cset, bool populated)
751 struct cgrp_cset_link *link;
753 lockdep_assert_held(&css_set_lock);
755 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
756 cgroup_update_populated(link->cgrp, populated);
760 * css_set_move_task - move a task from one css_set to another
761 * @task: task being moved
762 * @from_cset: css_set @task currently belongs to (may be NULL)
763 * @to_cset: new css_set @task is being moved to (may be NULL)
764 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
766 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
767 * css_set, @from_cset can be NULL. If @task is being disassociated
768 * instead of moved, @to_cset can be NULL.
770 * This function automatically handles populated counter updates and
771 * css_task_iter adjustments but the caller is responsible for managing
772 * @from_cset and @to_cset's reference counts.
774 static void css_set_move_task(struct task_struct *task,
775 struct css_set *from_cset, struct css_set *to_cset,
778 lockdep_assert_held(&css_set_lock);
780 if (to_cset && !css_set_populated(to_cset))
781 css_set_update_populated(to_cset, true);
784 struct css_task_iter *it, *pos;
786 WARN_ON_ONCE(list_empty(&task->cg_list));
789 * @task is leaving, advance task iterators which are
790 * pointing to it so that they can resume at the next
791 * position. Advancing an iterator might remove it from
792 * the list, use safe walk. See css_task_iter_advance*()
795 list_for_each_entry_safe(it, pos, &from_cset->task_iters,
797 if (it->task_pos == &task->cg_list)
798 css_task_iter_advance(it);
800 list_del_init(&task->cg_list);
801 if (!css_set_populated(from_cset))
802 css_set_update_populated(from_cset, false);
804 WARN_ON_ONCE(!list_empty(&task->cg_list));
809 * We are synchronized through cgroup_threadgroup_rwsem
810 * against PF_EXITING setting such that we can't race
811 * against cgroup_exit() changing the css_set to
812 * init_css_set and dropping the old one.
814 WARN_ON_ONCE(task->flags & PF_EXITING);
816 rcu_assign_pointer(task->cgroups, to_cset);
817 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
823 * hash table for cgroup groups. This improves the performance to find
824 * an existing css_set. This hash doesn't (currently) take into
825 * account cgroups in empty hierarchies.
827 #define CSS_SET_HASH_BITS 7
828 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
830 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
832 unsigned long key = 0UL;
833 struct cgroup_subsys *ss;
836 for_each_subsys(ss, i)
837 key += (unsigned long)css[i];
838 key = (key >> 16) ^ key;
843 void put_css_set_locked(struct css_set *cset)
845 struct cgrp_cset_link *link, *tmp_link;
846 struct cgroup_subsys *ss;
849 lockdep_assert_held(&css_set_lock);
851 if (!refcount_dec_and_test(&cset->refcount))
854 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
856 /* This css_set is dead. unlink it and release cgroup and css refs */
857 for_each_subsys(ss, ssid) {
858 list_del(&cset->e_cset_node[ssid]);
859 css_put(cset->subsys[ssid]);
861 hash_del(&cset->hlist);
864 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
865 list_del(&link->cset_link);
866 list_del(&link->cgrp_link);
867 if (cgroup_parent(link->cgrp))
868 cgroup_put(link->cgrp);
872 if (css_set_threaded(cset)) {
873 list_del(&cset->threaded_csets_node);
874 put_css_set_locked(cset->dom_cset);
877 kfree_rcu(cset, rcu_head);
881 * compare_css_sets - helper function for find_existing_css_set().
882 * @cset: candidate css_set being tested
883 * @old_cset: existing css_set for a task
884 * @new_cgrp: cgroup that's being entered by the task
885 * @template: desired set of css pointers in css_set (pre-calculated)
887 * Returns true if "cset" matches "old_cset" except for the hierarchy
888 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
890 static bool compare_css_sets(struct css_set *cset,
891 struct css_set *old_cset,
892 struct cgroup *new_cgrp,
893 struct cgroup_subsys_state *template[])
895 struct cgroup *new_dfl_cgrp;
896 struct list_head *l1, *l2;
899 * On the default hierarchy, there can be csets which are
900 * associated with the same set of cgroups but different csses.
901 * Let's first ensure that csses match.
903 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
907 /* @cset's domain should match the default cgroup's */
908 if (cgroup_on_dfl(new_cgrp))
909 new_dfl_cgrp = new_cgrp;
911 new_dfl_cgrp = old_cset->dfl_cgrp;
913 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
917 * Compare cgroup pointers in order to distinguish between
918 * different cgroups in hierarchies. As different cgroups may
919 * share the same effective css, this comparison is always
922 l1 = &cset->cgrp_links;
923 l2 = &old_cset->cgrp_links;
925 struct cgrp_cset_link *link1, *link2;
926 struct cgroup *cgrp1, *cgrp2;
930 /* See if we reached the end - both lists are equal length. */
931 if (l1 == &cset->cgrp_links) {
932 BUG_ON(l2 != &old_cset->cgrp_links);
935 BUG_ON(l2 == &old_cset->cgrp_links);
937 /* Locate the cgroups associated with these links. */
938 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
939 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
942 /* Hierarchies should be linked in the same order. */
943 BUG_ON(cgrp1->root != cgrp2->root);
946 * If this hierarchy is the hierarchy of the cgroup
947 * that's changing, then we need to check that this
948 * css_set points to the new cgroup; if it's any other
949 * hierarchy, then this css_set should point to the
950 * same cgroup as the old css_set.
952 if (cgrp1->root == new_cgrp->root) {
953 if (cgrp1 != new_cgrp)
964 * find_existing_css_set - init css array and find the matching css_set
965 * @old_cset: the css_set that we're using before the cgroup transition
966 * @cgrp: the cgroup that we're moving into
967 * @template: out param for the new set of csses, should be clear on entry
969 static struct css_set *find_existing_css_set(struct css_set *old_cset,
971 struct cgroup_subsys_state *template[])
973 struct cgroup_root *root = cgrp->root;
974 struct cgroup_subsys *ss;
975 struct css_set *cset;
980 * Build the set of subsystem state objects that we want to see in the
981 * new css_set. while subsystems can change globally, the entries here
982 * won't change, so no need for locking.
984 for_each_subsys(ss, i) {
985 if (root->subsys_mask & (1UL << i)) {
987 * @ss is in this hierarchy, so we want the
988 * effective css from @cgrp.
990 template[i] = cgroup_e_css(cgrp, ss);
993 * @ss is not in this hierarchy, so we don't want
996 template[i] = old_cset->subsys[i];
1000 key = css_set_hash(template);
1001 hash_for_each_possible(css_set_table, cset, hlist, key) {
1002 if (!compare_css_sets(cset, old_cset, cgrp, template))
1005 /* This css_set matches what we need */
1009 /* No existing cgroup group matched */
1013 static void free_cgrp_cset_links(struct list_head *links_to_free)
1015 struct cgrp_cset_link *link, *tmp_link;
1017 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1018 list_del(&link->cset_link);
1024 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1025 * @count: the number of links to allocate
1026 * @tmp_links: list_head the allocated links are put on
1028 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1029 * through ->cset_link. Returns 0 on success or -errno.
1031 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1033 struct cgrp_cset_link *link;
1036 INIT_LIST_HEAD(tmp_links);
1038 for (i = 0; i < count; i++) {
1039 link = kzalloc(sizeof(*link), GFP_KERNEL);
1041 free_cgrp_cset_links(tmp_links);
1044 list_add(&link->cset_link, tmp_links);
1050 * link_css_set - a helper function to link a css_set to a cgroup
1051 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1052 * @cset: the css_set to be linked
1053 * @cgrp: the destination cgroup
1055 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1056 struct cgroup *cgrp)
1058 struct cgrp_cset_link *link;
1060 BUG_ON(list_empty(tmp_links));
1062 if (cgroup_on_dfl(cgrp))
1063 cset->dfl_cgrp = cgrp;
1065 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1070 * Always add links to the tail of the lists so that the lists are
1071 * in choronological order.
1073 list_move_tail(&link->cset_link, &cgrp->cset_links);
1074 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1076 if (cgroup_parent(cgrp))
1077 cgroup_get_live(cgrp);
1081 * find_css_set - return a new css_set with one cgroup updated
1082 * @old_cset: the baseline css_set
1083 * @cgrp: the cgroup to be updated
1085 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1086 * substituted into the appropriate hierarchy.
1088 static struct css_set *find_css_set(struct css_set *old_cset,
1089 struct cgroup *cgrp)
1091 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1092 struct css_set *cset;
1093 struct list_head tmp_links;
1094 struct cgrp_cset_link *link;
1095 struct cgroup_subsys *ss;
1099 lockdep_assert_held(&cgroup_mutex);
1101 /* First see if we already have a cgroup group that matches
1102 * the desired set */
1103 spin_lock_irq(&css_set_lock);
1104 cset = find_existing_css_set(old_cset, cgrp, template);
1107 spin_unlock_irq(&css_set_lock);
1112 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1116 /* Allocate all the cgrp_cset_link objects that we'll need */
1117 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1122 refcount_set(&cset->refcount, 1);
1123 cset->dom_cset = cset;
1124 INIT_LIST_HEAD(&cset->tasks);
1125 INIT_LIST_HEAD(&cset->mg_tasks);
1126 INIT_LIST_HEAD(&cset->task_iters);
1127 INIT_LIST_HEAD(&cset->threaded_csets);
1128 INIT_HLIST_NODE(&cset->hlist);
1129 INIT_LIST_HEAD(&cset->cgrp_links);
1130 INIT_LIST_HEAD(&cset->mg_preload_node);
1131 INIT_LIST_HEAD(&cset->mg_node);
1133 /* Copy the set of subsystem state objects generated in
1134 * find_existing_css_set() */
1135 memcpy(cset->subsys, template, sizeof(cset->subsys));
1137 spin_lock_irq(&css_set_lock);
1138 /* Add reference counts and links from the new css_set. */
1139 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1140 struct cgroup *c = link->cgrp;
1142 if (c->root == cgrp->root)
1144 link_css_set(&tmp_links, cset, c);
1147 BUG_ON(!list_empty(&tmp_links));
1151 /* Add @cset to the hash table */
1152 key = css_set_hash(cset->subsys);
1153 hash_add(css_set_table, &cset->hlist, key);
1155 for_each_subsys(ss, ssid) {
1156 struct cgroup_subsys_state *css = cset->subsys[ssid];
1158 list_add_tail(&cset->e_cset_node[ssid],
1159 &css->cgroup->e_csets[ssid]);
1163 spin_unlock_irq(&css_set_lock);
1166 * If @cset should be threaded, look up the matching dom_cset and
1167 * link them up. We first fully initialize @cset then look for the
1168 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1169 * to stay empty until we return.
1171 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1172 struct css_set *dcset;
1174 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1180 spin_lock_irq(&css_set_lock);
1181 cset->dom_cset = dcset;
1182 list_add_tail(&cset->threaded_csets_node,
1183 &dcset->threaded_csets);
1184 spin_unlock_irq(&css_set_lock);
1190 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1192 struct cgroup *root_cgrp = kf_root->kn->priv;
1194 return root_cgrp->root;
1197 static int cgroup_init_root_id(struct cgroup_root *root)
1201 lockdep_assert_held(&cgroup_mutex);
1203 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1207 root->hierarchy_id = id;
1211 static void cgroup_exit_root_id(struct cgroup_root *root)
1213 lockdep_assert_held(&cgroup_mutex);
1215 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1218 void cgroup_free_root(struct cgroup_root *root)
1221 idr_destroy(&root->cgroup_idr);
1226 static void cgroup_destroy_root(struct cgroup_root *root)
1228 struct cgroup *cgrp = &root->cgrp;
1229 struct cgrp_cset_link *link, *tmp_link;
1231 trace_cgroup_destroy_root(root);
1233 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1235 BUG_ON(atomic_read(&root->nr_cgrps));
1236 BUG_ON(!list_empty(&cgrp->self.children));
1238 /* Rebind all subsystems back to the default hierarchy */
1239 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1242 * Release all the links from cset_links to this hierarchy's
1245 spin_lock_irq(&css_set_lock);
1247 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1248 list_del(&link->cset_link);
1249 list_del(&link->cgrp_link);
1253 spin_unlock_irq(&css_set_lock);
1255 if (!list_empty(&root->root_list)) {
1256 list_del(&root->root_list);
1257 cgroup_root_count--;
1260 cgroup_exit_root_id(root);
1262 mutex_unlock(&cgroup_mutex);
1264 kernfs_destroy_root(root->kf_root);
1265 cgroup_free_root(root);
1269 * look up cgroup associated with current task's cgroup namespace on the
1270 * specified hierarchy
1272 static struct cgroup *
1273 current_cgns_cgroup_from_root(struct cgroup_root *root)
1275 struct cgroup *res = NULL;
1276 struct css_set *cset;
1278 lockdep_assert_held(&css_set_lock);
1282 cset = current->nsproxy->cgroup_ns->root_cset;
1283 if (cset == &init_css_set) {
1286 struct cgrp_cset_link *link;
1288 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1289 struct cgroup *c = link->cgrp;
1291 if (c->root == root) {
1303 /* look up cgroup associated with given css_set on the specified hierarchy */
1304 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1305 struct cgroup_root *root)
1307 struct cgroup *res = NULL;
1309 lockdep_assert_held(&cgroup_mutex);
1310 lockdep_assert_held(&css_set_lock);
1312 if (cset == &init_css_set) {
1314 } else if (root == &cgrp_dfl_root) {
1315 res = cset->dfl_cgrp;
1317 struct cgrp_cset_link *link;
1319 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1320 struct cgroup *c = link->cgrp;
1322 if (c->root == root) {
1334 * Return the cgroup for "task" from the given hierarchy. Must be
1335 * called with cgroup_mutex and css_set_lock held.
1337 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1338 struct cgroup_root *root)
1341 * No need to lock the task - since we hold cgroup_mutex the
1342 * task can't change groups, so the only thing that can happen
1343 * is that it exits and its css is set back to init_css_set.
1345 return cset_cgroup_from_root(task_css_set(task), root);
1349 * A task must hold cgroup_mutex to modify cgroups.
1351 * Any task can increment and decrement the count field without lock.
1352 * So in general, code holding cgroup_mutex can't rely on the count
1353 * field not changing. However, if the count goes to zero, then only
1354 * cgroup_attach_task() can increment it again. Because a count of zero
1355 * means that no tasks are currently attached, therefore there is no
1356 * way a task attached to that cgroup can fork (the other way to
1357 * increment the count). So code holding cgroup_mutex can safely
1358 * assume that if the count is zero, it will stay zero. Similarly, if
1359 * a task holds cgroup_mutex on a cgroup with zero count, it
1360 * knows that the cgroup won't be removed, as cgroup_rmdir()
1363 * A cgroup can only be deleted if both its 'count' of using tasks
1364 * is zero, and its list of 'children' cgroups is empty. Since all
1365 * tasks in the system use _some_ cgroup, and since there is always at
1366 * least one task in the system (init, pid == 1), therefore, root cgroup
1367 * always has either children cgroups and/or using tasks. So we don't
1368 * need a special hack to ensure that root cgroup cannot be deleted.
1370 * P.S. One more locking exception. RCU is used to guard the
1371 * update of a tasks cgroup pointer by cgroup_attach_task()
1374 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1376 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1379 struct cgroup_subsys *ss = cft->ss;
1381 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1382 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1383 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1384 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1387 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1392 * cgroup_file_mode - deduce file mode of a control file
1393 * @cft: the control file in question
1395 * S_IRUGO for read, S_IWUSR for write.
1397 static umode_t cgroup_file_mode(const struct cftype *cft)
1401 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1404 if (cft->write_u64 || cft->write_s64 || cft->write) {
1405 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1415 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1416 * @subtree_control: the new subtree_control mask to consider
1417 * @this_ss_mask: available subsystems
1419 * On the default hierarchy, a subsystem may request other subsystems to be
1420 * enabled together through its ->depends_on mask. In such cases, more
1421 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1423 * This function calculates which subsystems need to be enabled if
1424 * @subtree_control is to be applied while restricted to @this_ss_mask.
1426 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1428 u16 cur_ss_mask = subtree_control;
1429 struct cgroup_subsys *ss;
1432 lockdep_assert_held(&cgroup_mutex);
1434 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1437 u16 new_ss_mask = cur_ss_mask;
1439 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1440 new_ss_mask |= ss->depends_on;
1441 } while_each_subsys_mask();
1444 * Mask out subsystems which aren't available. This can
1445 * happen only if some depended-upon subsystems were bound
1446 * to non-default hierarchies.
1448 new_ss_mask &= this_ss_mask;
1450 if (new_ss_mask == cur_ss_mask)
1452 cur_ss_mask = new_ss_mask;
1459 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1460 * @kn: the kernfs_node being serviced
1462 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1463 * the method finishes if locking succeeded. Note that once this function
1464 * returns the cgroup returned by cgroup_kn_lock_live() may become
1465 * inaccessible any time. If the caller intends to continue to access the
1466 * cgroup, it should pin it before invoking this function.
1468 void cgroup_kn_unlock(struct kernfs_node *kn)
1470 struct cgroup *cgrp;
1472 if (kernfs_type(kn) == KERNFS_DIR)
1475 cgrp = kn->parent->priv;
1477 mutex_unlock(&cgroup_mutex);
1479 kernfs_unbreak_active_protection(kn);
1484 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1485 * @kn: the kernfs_node being serviced
1486 * @drain_offline: perform offline draining on the cgroup
1488 * This helper is to be used by a cgroup kernfs method currently servicing
1489 * @kn. It breaks the active protection, performs cgroup locking and
1490 * verifies that the associated cgroup is alive. Returns the cgroup if
1491 * alive; otherwise, %NULL. A successful return should be undone by a
1492 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1493 * cgroup is drained of offlining csses before return.
1495 * Any cgroup kernfs method implementation which requires locking the
1496 * associated cgroup should use this helper. It avoids nesting cgroup
1497 * locking under kernfs active protection and allows all kernfs operations
1498 * including self-removal.
1500 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1502 struct cgroup *cgrp;
1504 if (kernfs_type(kn) == KERNFS_DIR)
1507 cgrp = kn->parent->priv;
1510 * We're gonna grab cgroup_mutex which nests outside kernfs
1511 * active_ref. cgroup liveliness check alone provides enough
1512 * protection against removal. Ensure @cgrp stays accessible and
1513 * break the active_ref protection.
1515 if (!cgroup_tryget(cgrp))
1517 kernfs_break_active_protection(kn);
1520 cgroup_lock_and_drain_offline(cgrp);
1522 mutex_lock(&cgroup_mutex);
1524 if (!cgroup_is_dead(cgrp))
1527 cgroup_kn_unlock(kn);
1531 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1533 char name[CGROUP_FILE_NAME_MAX];
1535 lockdep_assert_held(&cgroup_mutex);
1537 if (cft->file_offset) {
1538 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1539 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1541 spin_lock_irq(&cgroup_file_kn_lock);
1543 spin_unlock_irq(&cgroup_file_kn_lock);
1546 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1550 * css_clear_dir - remove subsys files in a cgroup directory
1553 static void css_clear_dir(struct cgroup_subsys_state *css)
1555 struct cgroup *cgrp = css->cgroup;
1556 struct cftype *cfts;
1558 if (!(css->flags & CSS_VISIBLE))
1561 css->flags &= ~CSS_VISIBLE;
1563 list_for_each_entry(cfts, &css->ss->cfts, node)
1564 cgroup_addrm_files(css, cgrp, cfts, false);
1568 * css_populate_dir - create subsys files in a cgroup directory
1571 * On failure, no file is added.
1573 static int css_populate_dir(struct cgroup_subsys_state *css)
1575 struct cgroup *cgrp = css->cgroup;
1576 struct cftype *cfts, *failed_cfts;
1579 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1583 if (cgroup_on_dfl(cgrp))
1584 cfts = cgroup_base_files;
1586 cfts = cgroup1_base_files;
1588 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1591 list_for_each_entry(cfts, &css->ss->cfts, node) {
1592 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1599 css->flags |= CSS_VISIBLE;
1603 list_for_each_entry(cfts, &css->ss->cfts, node) {
1604 if (cfts == failed_cfts)
1606 cgroup_addrm_files(css, cgrp, cfts, false);
1611 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1613 struct cgroup *dcgrp = &dst_root->cgrp;
1614 struct cgroup_subsys *ss;
1617 lockdep_assert_held(&cgroup_mutex);
1619 do_each_subsys_mask(ss, ssid, ss_mask) {
1621 * If @ss has non-root csses attached to it, can't move.
1622 * If @ss is an implicit controller, it is exempt from this
1623 * rule and can be stolen.
1625 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1626 !ss->implicit_on_dfl)
1629 /* can't move between two non-dummy roots either */
1630 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1632 } while_each_subsys_mask();
1634 do_each_subsys_mask(ss, ssid, ss_mask) {
1635 struct cgroup_root *src_root = ss->root;
1636 struct cgroup *scgrp = &src_root->cgrp;
1637 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1638 struct css_set *cset;
1640 WARN_ON(!css || cgroup_css(dcgrp, ss));
1642 /* disable from the source */
1643 src_root->subsys_mask &= ~(1 << ssid);
1644 WARN_ON(cgroup_apply_control(scgrp));
1645 cgroup_finalize_control(scgrp, 0);
1648 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1649 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1650 ss->root = dst_root;
1651 css->cgroup = dcgrp;
1653 spin_lock_irq(&css_set_lock);
1654 hash_for_each(css_set_table, i, cset, hlist)
1655 list_move_tail(&cset->e_cset_node[ss->id],
1656 &dcgrp->e_csets[ss->id]);
1657 spin_unlock_irq(&css_set_lock);
1659 /* default hierarchy doesn't enable controllers by default */
1660 dst_root->subsys_mask |= 1 << ssid;
1661 if (dst_root == &cgrp_dfl_root) {
1662 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1664 dcgrp->subtree_control |= 1 << ssid;
1665 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1668 ret = cgroup_apply_control(dcgrp);
1670 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1675 } while_each_subsys_mask();
1677 kernfs_activate(dcgrp->kn);
1681 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1682 struct kernfs_root *kf_root)
1686 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1687 struct cgroup *ns_cgroup;
1689 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1693 spin_lock_irq(&css_set_lock);
1694 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1695 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1696 spin_unlock_irq(&css_set_lock);
1698 if (len >= PATH_MAX)
1701 seq_escape(sf, buf, " \t\n\\");
1708 static int parse_cgroup_root_flags(char *data, unsigned int *root_flags)
1717 while ((token = strsep(&data, ",")) != NULL) {
1718 if (!strcmp(token, "nsdelegate")) {
1719 *root_flags |= CGRP_ROOT_NS_DELEGATE;
1723 pr_err("cgroup2: unknown option \"%s\"\n", token);
1730 static void apply_cgroup_root_flags(unsigned int root_flags)
1732 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1733 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1734 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1736 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1740 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1742 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1743 seq_puts(seq, ",nsdelegate");
1747 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1749 unsigned int root_flags;
1752 ret = parse_cgroup_root_flags(data, &root_flags);
1756 apply_cgroup_root_flags(root_flags);
1761 * To reduce the fork() overhead for systems that are not actually using
1762 * their cgroups capability, we don't maintain the lists running through
1763 * each css_set to its tasks until we see the list actually used - in other
1764 * words after the first mount.
1766 static bool use_task_css_set_links __read_mostly;
1768 static void cgroup_enable_task_cg_lists(void)
1770 struct task_struct *p, *g;
1772 spin_lock_irq(&css_set_lock);
1774 if (use_task_css_set_links)
1777 use_task_css_set_links = true;
1780 * We need tasklist_lock because RCU is not safe against
1781 * while_each_thread(). Besides, a forking task that has passed
1782 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1783 * is not guaranteed to have its child immediately visible in the
1784 * tasklist if we walk through it with RCU.
1786 read_lock(&tasklist_lock);
1787 do_each_thread(g, p) {
1788 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1789 task_css_set(p) != &init_css_set);
1792 * We should check if the process is exiting, otherwise
1793 * it will race with cgroup_exit() in that the list
1794 * entry won't be deleted though the process has exited.
1795 * Do it while holding siglock so that we don't end up
1796 * racing against cgroup_exit().
1798 * Interrupts were already disabled while acquiring
1799 * the css_set_lock, so we do not need to disable it
1800 * again when acquiring the sighand->siglock here.
1802 spin_lock(&p->sighand->siglock);
1803 if (!(p->flags & PF_EXITING)) {
1804 struct css_set *cset = task_css_set(p);
1806 if (!css_set_populated(cset))
1807 css_set_update_populated(cset, true);
1808 list_add_tail(&p->cg_list, &cset->tasks);
1812 spin_unlock(&p->sighand->siglock);
1813 } while_each_thread(g, p);
1814 read_unlock(&tasklist_lock);
1816 spin_unlock_irq(&css_set_lock);
1819 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1821 struct cgroup_subsys *ss;
1824 INIT_LIST_HEAD(&cgrp->self.sibling);
1825 INIT_LIST_HEAD(&cgrp->self.children);
1826 INIT_LIST_HEAD(&cgrp->cset_links);
1827 INIT_LIST_HEAD(&cgrp->pidlists);
1828 mutex_init(&cgrp->pidlist_mutex);
1829 cgrp->self.cgroup = cgrp;
1830 cgrp->self.flags |= CSS_ONLINE;
1831 cgrp->dom_cgrp = cgrp;
1832 cgrp->max_descendants = INT_MAX;
1833 cgrp->max_depth = INT_MAX;
1835 for_each_subsys(ss, ssid)
1836 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1838 init_waitqueue_head(&cgrp->offline_waitq);
1839 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1842 void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
1844 struct cgroup *cgrp = &root->cgrp;
1846 INIT_LIST_HEAD(&root->root_list);
1847 atomic_set(&root->nr_cgrps, 1);
1849 init_cgroup_housekeeping(cgrp);
1850 idr_init(&root->cgroup_idr);
1852 root->flags = opts->flags;
1853 if (opts->release_agent)
1854 strcpy(root->release_agent_path, opts->release_agent);
1856 strcpy(root->name, opts->name);
1857 if (opts->cpuset_clone_children)
1858 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1861 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
1863 LIST_HEAD(tmp_links);
1864 struct cgroup *root_cgrp = &root->cgrp;
1865 struct kernfs_syscall_ops *kf_sops;
1866 struct css_set *cset;
1869 lockdep_assert_held(&cgroup_mutex);
1871 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1874 root_cgrp->id = ret;
1875 root_cgrp->ancestor_ids[0] = ret;
1877 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1878 ref_flags, GFP_KERNEL);
1883 * We're accessing css_set_count without locking css_set_lock here,
1884 * but that's OK - it can only be increased by someone holding
1885 * cgroup_lock, and that's us. Later rebinding may disable
1886 * controllers on the default hierarchy and thus create new csets,
1887 * which can't be more than the existing ones. Allocate 2x.
1889 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1893 ret = cgroup_init_root_id(root);
1897 kf_sops = root == &cgrp_dfl_root ?
1898 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1900 root->kf_root = kernfs_create_root(kf_sops,
1901 KERNFS_ROOT_CREATE_DEACTIVATED,
1903 if (IS_ERR(root->kf_root)) {
1904 ret = PTR_ERR(root->kf_root);
1907 root_cgrp->kn = root->kf_root->kn;
1909 ret = css_populate_dir(&root_cgrp->self);
1913 ret = rebind_subsystems(root, ss_mask);
1917 trace_cgroup_setup_root(root);
1920 * There must be no failure case after here, since rebinding takes
1921 * care of subsystems' refcounts, which are explicitly dropped in
1922 * the failure exit path.
1924 list_add(&root->root_list, &cgroup_roots);
1925 cgroup_root_count++;
1928 * Link the root cgroup in this hierarchy into all the css_set
1931 spin_lock_irq(&css_set_lock);
1932 hash_for_each(css_set_table, i, cset, hlist) {
1933 link_css_set(&tmp_links, cset, root_cgrp);
1934 if (css_set_populated(cset))
1935 cgroup_update_populated(root_cgrp, true);
1937 spin_unlock_irq(&css_set_lock);
1939 BUG_ON(!list_empty(&root_cgrp->self.children));
1940 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1942 kernfs_activate(root_cgrp->kn);
1947 kernfs_destroy_root(root->kf_root);
1948 root->kf_root = NULL;
1950 cgroup_exit_root_id(root);
1952 percpu_ref_exit(&root_cgrp->self.refcnt);
1954 free_cgrp_cset_links(&tmp_links);
1958 struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
1959 struct cgroup_root *root, unsigned long magic,
1960 struct cgroup_namespace *ns)
1962 struct dentry *dentry;
1965 dentry = kernfs_mount(fs_type, flags, root->kf_root, magic, &new_sb);
1968 * In non-init cgroup namespace, instead of root cgroup's dentry,
1969 * we return the dentry corresponding to the cgroupns->root_cgrp.
1971 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
1972 struct dentry *nsdentry;
1973 struct cgroup *cgrp;
1975 mutex_lock(&cgroup_mutex);
1976 spin_lock_irq(&css_set_lock);
1978 cgrp = cset_cgroup_from_root(ns->root_cset, root);
1980 spin_unlock_irq(&css_set_lock);
1981 mutex_unlock(&cgroup_mutex);
1983 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
1988 if (IS_ERR(dentry) || !new_sb)
1989 cgroup_put(&root->cgrp);
1994 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1995 int flags, const char *unused_dev_name,
1998 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
1999 struct dentry *dentry;
2004 /* Check if the caller has permission to mount. */
2005 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2007 return ERR_PTR(-EPERM);
2011 * The first time anyone tries to mount a cgroup, enable the list
2012 * linking each css_set to its tasks and fix up all existing tasks.
2014 if (!use_task_css_set_links)
2015 cgroup_enable_task_cg_lists();
2017 if (fs_type == &cgroup2_fs_type) {
2018 unsigned int root_flags;
2020 ret = parse_cgroup_root_flags(data, &root_flags);
2023 return ERR_PTR(ret);
2026 cgrp_dfl_visible = true;
2027 cgroup_get_live(&cgrp_dfl_root.cgrp);
2029 dentry = cgroup_do_mount(&cgroup2_fs_type, flags, &cgrp_dfl_root,
2030 CGROUP2_SUPER_MAGIC, ns);
2031 if (!IS_ERR(dentry))
2032 apply_cgroup_root_flags(root_flags);
2034 dentry = cgroup1_mount(&cgroup_fs_type, flags, data,
2035 CGROUP_SUPER_MAGIC, ns);
2042 static void cgroup_kill_sb(struct super_block *sb)
2044 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2045 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2048 * If @root doesn't have any mounts or children, start killing it.
2049 * This prevents new mounts by disabling percpu_ref_tryget_live().
2050 * cgroup_mount() may wait for @root's release.
2052 * And don't kill the default root.
2054 if (!list_empty(&root->cgrp.self.children) ||
2055 root == &cgrp_dfl_root)
2056 cgroup_put(&root->cgrp);
2058 percpu_ref_kill(&root->cgrp.self.refcnt);
2063 struct file_system_type cgroup_fs_type = {
2065 .mount = cgroup_mount,
2066 .kill_sb = cgroup_kill_sb,
2067 .fs_flags = FS_USERNS_MOUNT,
2070 static struct file_system_type cgroup2_fs_type = {
2072 .mount = cgroup_mount,
2073 .kill_sb = cgroup_kill_sb,
2074 .fs_flags = FS_USERNS_MOUNT,
2077 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2078 struct cgroup_namespace *ns)
2080 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2082 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2085 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2086 struct cgroup_namespace *ns)
2090 mutex_lock(&cgroup_mutex);
2091 spin_lock_irq(&css_set_lock);
2093 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2095 spin_unlock_irq(&css_set_lock);
2096 mutex_unlock(&cgroup_mutex);
2100 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2103 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2104 * @task: target task
2105 * @buf: the buffer to write the path into
2106 * @buflen: the length of the buffer
2108 * Determine @task's cgroup on the first (the one with the lowest non-zero
2109 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2110 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2111 * cgroup controller callbacks.
2113 * Return value is the same as kernfs_path().
2115 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2117 struct cgroup_root *root;
2118 struct cgroup *cgrp;
2119 int hierarchy_id = 1;
2122 mutex_lock(&cgroup_mutex);
2123 spin_lock_irq(&css_set_lock);
2125 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2128 cgrp = task_cgroup_from_root(task, root);
2129 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2131 /* if no hierarchy exists, everyone is in "/" */
2132 ret = strlcpy(buf, "/", buflen);
2135 spin_unlock_irq(&css_set_lock);
2136 mutex_unlock(&cgroup_mutex);
2139 EXPORT_SYMBOL_GPL(task_cgroup_path);
2142 * cgroup_migrate_add_task - add a migration target task to a migration context
2143 * @task: target task
2144 * @mgctx: target migration context
2146 * Add @task, which is a migration target, to @mgctx->tset. This function
2147 * becomes noop if @task doesn't need to be migrated. @task's css_set
2148 * should have been added as a migration source and @task->cg_list will be
2149 * moved from the css_set's tasks list to mg_tasks one.
2151 static void cgroup_migrate_add_task(struct task_struct *task,
2152 struct cgroup_mgctx *mgctx)
2154 struct css_set *cset;
2156 lockdep_assert_held(&css_set_lock);
2158 /* @task either already exited or can't exit until the end */
2159 if (task->flags & PF_EXITING)
2162 /* leave @task alone if post_fork() hasn't linked it yet */
2163 if (list_empty(&task->cg_list))
2166 cset = task_css_set(task);
2167 if (!cset->mg_src_cgrp)
2170 list_move_tail(&task->cg_list, &cset->mg_tasks);
2171 if (list_empty(&cset->mg_node))
2172 list_add_tail(&cset->mg_node,
2173 &mgctx->tset.src_csets);
2174 if (list_empty(&cset->mg_dst_cset->mg_node))
2175 list_add_tail(&cset->mg_dst_cset->mg_node,
2176 &mgctx->tset.dst_csets);
2180 * cgroup_taskset_first - reset taskset and return the first task
2181 * @tset: taskset of interest
2182 * @dst_cssp: output variable for the destination css
2184 * @tset iteration is initialized and the first task is returned.
2186 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2187 struct cgroup_subsys_state **dst_cssp)
2189 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2190 tset->cur_task = NULL;
2192 return cgroup_taskset_next(tset, dst_cssp);
2196 * cgroup_taskset_next - iterate to the next task in taskset
2197 * @tset: taskset of interest
2198 * @dst_cssp: output variable for the destination css
2200 * Return the next task in @tset. Iteration must have been initialized
2201 * with cgroup_taskset_first().
2203 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2204 struct cgroup_subsys_state **dst_cssp)
2206 struct css_set *cset = tset->cur_cset;
2207 struct task_struct *task = tset->cur_task;
2209 while (&cset->mg_node != tset->csets) {
2211 task = list_first_entry(&cset->mg_tasks,
2212 struct task_struct, cg_list);
2214 task = list_next_entry(task, cg_list);
2216 if (&task->cg_list != &cset->mg_tasks) {
2217 tset->cur_cset = cset;
2218 tset->cur_task = task;
2221 * This function may be called both before and
2222 * after cgroup_taskset_migrate(). The two cases
2223 * can be distinguished by looking at whether @cset
2224 * has its ->mg_dst_cset set.
2226 if (cset->mg_dst_cset)
2227 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2229 *dst_cssp = cset->subsys[tset->ssid];
2234 cset = list_next_entry(cset, mg_node);
2242 * cgroup_taskset_migrate - migrate a taskset
2243 * @mgctx: migration context
2245 * Migrate tasks in @mgctx as setup by migration preparation functions.
2246 * This function fails iff one of the ->can_attach callbacks fails and
2247 * guarantees that either all or none of the tasks in @mgctx are migrated.
2248 * @mgctx is consumed regardless of success.
2250 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2252 struct cgroup_taskset *tset = &mgctx->tset;
2253 struct cgroup_subsys *ss;
2254 struct task_struct *task, *tmp_task;
2255 struct css_set *cset, *tmp_cset;
2256 int ssid, failed_ssid, ret;
2258 /* methods shouldn't be called if no task is actually migrating */
2259 if (list_empty(&tset->src_csets))
2262 /* check that we can legitimately attach to the cgroup */
2263 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2264 if (ss->can_attach) {
2266 ret = ss->can_attach(tset);
2269 goto out_cancel_attach;
2272 } while_each_subsys_mask();
2275 * Now that we're guaranteed success, proceed to move all tasks to
2276 * the new cgroup. There are no failure cases after here, so this
2277 * is the commit point.
2279 spin_lock_irq(&css_set_lock);
2280 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2281 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2282 struct css_set *from_cset = task_css_set(task);
2283 struct css_set *to_cset = cset->mg_dst_cset;
2285 get_css_set(to_cset);
2286 to_cset->nr_tasks++;
2287 css_set_move_task(task, from_cset, to_cset, true);
2288 put_css_set_locked(from_cset);
2289 from_cset->nr_tasks--;
2292 spin_unlock_irq(&css_set_lock);
2295 * Migration is committed, all target tasks are now on dst_csets.
2296 * Nothing is sensitive to fork() after this point. Notify
2297 * controllers that migration is complete.
2299 tset->csets = &tset->dst_csets;
2301 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2306 } while_each_subsys_mask();
2309 goto out_release_tset;
2312 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2313 if (ssid == failed_ssid)
2315 if (ss->cancel_attach) {
2317 ss->cancel_attach(tset);
2319 } while_each_subsys_mask();
2321 spin_lock_irq(&css_set_lock);
2322 list_splice_init(&tset->dst_csets, &tset->src_csets);
2323 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2324 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2325 list_del_init(&cset->mg_node);
2327 spin_unlock_irq(&css_set_lock);
2332 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2333 * @dst_cgrp: destination cgroup to test
2335 * On the default hierarchy, except for the mixable, (possible) thread root
2336 * and threaded cgroups, subtree_control must be zero for migration
2337 * destination cgroups with tasks so that child cgroups don't compete
2340 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2342 /* v1 doesn't have any restriction */
2343 if (!cgroup_on_dfl(dst_cgrp))
2346 /* verify @dst_cgrp can host resources */
2347 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2350 /* mixables don't care */
2351 if (cgroup_is_mixable(dst_cgrp))
2355 * If @dst_cgrp is already or can become a thread root or is
2356 * threaded, it doesn't matter.
2358 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2361 /* apply no-internal-process constraint */
2362 if (dst_cgrp->subtree_control)
2369 * cgroup_migrate_finish - cleanup after attach
2370 * @mgctx: migration context
2372 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2373 * those functions for details.
2375 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2377 LIST_HEAD(preloaded);
2378 struct css_set *cset, *tmp_cset;
2380 lockdep_assert_held(&cgroup_mutex);
2382 spin_lock_irq(&css_set_lock);
2384 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2385 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2387 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2388 cset->mg_src_cgrp = NULL;
2389 cset->mg_dst_cgrp = NULL;
2390 cset->mg_dst_cset = NULL;
2391 list_del_init(&cset->mg_preload_node);
2392 put_css_set_locked(cset);
2395 spin_unlock_irq(&css_set_lock);
2399 * cgroup_migrate_add_src - add a migration source css_set
2400 * @src_cset: the source css_set to add
2401 * @dst_cgrp: the destination cgroup
2402 * @mgctx: migration context
2404 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2405 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2406 * up by cgroup_migrate_finish().
2408 * This function may be called without holding cgroup_threadgroup_rwsem
2409 * even if the target is a process. Threads may be created and destroyed
2410 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2411 * into play and the preloaded css_sets are guaranteed to cover all
2414 void cgroup_migrate_add_src(struct css_set *src_cset,
2415 struct cgroup *dst_cgrp,
2416 struct cgroup_mgctx *mgctx)
2418 struct cgroup *src_cgrp;
2420 lockdep_assert_held(&cgroup_mutex);
2421 lockdep_assert_held(&css_set_lock);
2424 * If ->dead, @src_set is associated with one or more dead cgroups
2425 * and doesn't contain any migratable tasks. Ignore it early so
2426 * that the rest of migration path doesn't get confused by it.
2431 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2433 if (!list_empty(&src_cset->mg_preload_node))
2436 WARN_ON(src_cset->mg_src_cgrp);
2437 WARN_ON(src_cset->mg_dst_cgrp);
2438 WARN_ON(!list_empty(&src_cset->mg_tasks));
2439 WARN_ON(!list_empty(&src_cset->mg_node));
2441 src_cset->mg_src_cgrp = src_cgrp;
2442 src_cset->mg_dst_cgrp = dst_cgrp;
2443 get_css_set(src_cset);
2444 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2448 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2449 * @mgctx: migration context
2451 * Tasks are about to be moved and all the source css_sets have been
2452 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2453 * pins all destination css_sets, links each to its source, and append them
2454 * to @mgctx->preloaded_dst_csets.
2456 * This function must be called after cgroup_migrate_add_src() has been
2457 * called on each migration source css_set. After migration is performed
2458 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2461 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2463 struct css_set *src_cset, *tmp_cset;
2465 lockdep_assert_held(&cgroup_mutex);
2467 /* look up the dst cset for each src cset and link it to src */
2468 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2470 struct css_set *dst_cset;
2471 struct cgroup_subsys *ss;
2474 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2478 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2481 * If src cset equals dst, it's noop. Drop the src.
2482 * cgroup_migrate() will skip the cset too. Note that we
2483 * can't handle src == dst as some nodes are used by both.
2485 if (src_cset == dst_cset) {
2486 src_cset->mg_src_cgrp = NULL;
2487 src_cset->mg_dst_cgrp = NULL;
2488 list_del_init(&src_cset->mg_preload_node);
2489 put_css_set(src_cset);
2490 put_css_set(dst_cset);
2494 src_cset->mg_dst_cset = dst_cset;
2496 if (list_empty(&dst_cset->mg_preload_node))
2497 list_add_tail(&dst_cset->mg_preload_node,
2498 &mgctx->preloaded_dst_csets);
2500 put_css_set(dst_cset);
2502 for_each_subsys(ss, ssid)
2503 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2504 mgctx->ss_mask |= 1 << ssid;
2509 cgroup_migrate_finish(mgctx);
2514 * cgroup_migrate - migrate a process or task to a cgroup
2515 * @leader: the leader of the process or the task to migrate
2516 * @threadgroup: whether @leader points to the whole process or a single task
2517 * @mgctx: migration context
2519 * Migrate a process or task denoted by @leader. If migrating a process,
2520 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2521 * responsible for invoking cgroup_migrate_add_src() and
2522 * cgroup_migrate_prepare_dst() on the targets before invoking this
2523 * function and following up with cgroup_migrate_finish().
2525 * As long as a controller's ->can_attach() doesn't fail, this function is
2526 * guaranteed to succeed. This means that, excluding ->can_attach()
2527 * failure, when migrating multiple targets, the success or failure can be
2528 * decided for all targets by invoking group_migrate_prepare_dst() before
2529 * actually starting migrating.
2531 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2532 struct cgroup_mgctx *mgctx)
2534 struct task_struct *task;
2537 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2538 * already PF_EXITING could be freed from underneath us unless we
2539 * take an rcu_read_lock.
2541 spin_lock_irq(&css_set_lock);
2545 cgroup_migrate_add_task(task, mgctx);
2548 } while_each_thread(leader, task);
2550 spin_unlock_irq(&css_set_lock);
2552 return cgroup_migrate_execute(mgctx);
2556 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2557 * @dst_cgrp: the cgroup to attach to
2558 * @leader: the task or the leader of the threadgroup to be attached
2559 * @threadgroup: attach the whole threadgroup?
2561 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2563 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2566 DEFINE_CGROUP_MGCTX(mgctx);
2567 struct task_struct *task;
2570 ret = cgroup_migrate_vet_dst(dst_cgrp);
2574 /* look up all src csets */
2575 spin_lock_irq(&css_set_lock);
2579 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2582 } while_each_thread(leader, task);
2584 spin_unlock_irq(&css_set_lock);
2586 /* prepare dst csets and commit */
2587 ret = cgroup_migrate_prepare_dst(&mgctx);
2589 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2591 cgroup_migrate_finish(&mgctx);
2594 trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2599 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2600 __acquires(&cgroup_threadgroup_rwsem)
2602 struct task_struct *tsk;
2605 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2606 return ERR_PTR(-EINVAL);
2608 percpu_down_write(&cgroup_threadgroup_rwsem);
2612 tsk = find_task_by_vpid(pid);
2614 tsk = ERR_PTR(-ESRCH);
2615 goto out_unlock_threadgroup;
2622 tsk = tsk->group_leader;
2625 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2626 * If userland migrates such a kthread to a non-root cgroup, it can
2627 * become trapped in a cpuset, or RT kthread may be born in a
2628 * cgroup with no rt_runtime allocated. Just say no.
2630 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2631 tsk = ERR_PTR(-EINVAL);
2632 goto out_unlock_threadgroup;
2635 get_task_struct(tsk);
2636 goto out_unlock_rcu;
2638 out_unlock_threadgroup:
2639 percpu_up_write(&cgroup_threadgroup_rwsem);
2645 void cgroup_procs_write_finish(struct task_struct *task)
2646 __releases(&cgroup_threadgroup_rwsem)
2648 struct cgroup_subsys *ss;
2651 /* release reference from cgroup_procs_write_start() */
2652 put_task_struct(task);
2654 percpu_up_write(&cgroup_threadgroup_rwsem);
2655 for_each_subsys(ss, ssid)
2656 if (ss->post_attach)
2660 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2662 struct cgroup_subsys *ss;
2663 bool printed = false;
2666 do_each_subsys_mask(ss, ssid, ss_mask) {
2669 seq_printf(seq, "%s", ss->name);
2671 } while_each_subsys_mask();
2673 seq_putc(seq, '\n');
2676 /* show controllers which are enabled from the parent */
2677 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2679 struct cgroup *cgrp = seq_css(seq)->cgroup;
2681 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2685 /* show controllers which are enabled for a given cgroup's children */
2686 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2688 struct cgroup *cgrp = seq_css(seq)->cgroup;
2690 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2695 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2696 * @cgrp: root of the subtree to update csses for
2698 * @cgrp's control masks have changed and its subtree's css associations
2699 * need to be updated accordingly. This function looks up all css_sets
2700 * which are attached to the subtree, creates the matching updated css_sets
2701 * and migrates the tasks to the new ones.
2703 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2705 DEFINE_CGROUP_MGCTX(mgctx);
2706 struct cgroup_subsys_state *d_css;
2707 struct cgroup *dsct;
2708 struct css_set *src_cset;
2711 lockdep_assert_held(&cgroup_mutex);
2713 percpu_down_write(&cgroup_threadgroup_rwsem);
2715 /* look up all csses currently attached to @cgrp's subtree */
2716 spin_lock_irq(&css_set_lock);
2717 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2718 struct cgrp_cset_link *link;
2720 list_for_each_entry(link, &dsct->cset_links, cset_link)
2721 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2723 spin_unlock_irq(&css_set_lock);
2725 /* NULL dst indicates self on default hierarchy */
2726 ret = cgroup_migrate_prepare_dst(&mgctx);
2730 spin_lock_irq(&css_set_lock);
2731 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2732 struct task_struct *task, *ntask;
2734 /* all tasks in src_csets need to be migrated */
2735 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2736 cgroup_migrate_add_task(task, &mgctx);
2738 spin_unlock_irq(&css_set_lock);
2740 ret = cgroup_migrate_execute(&mgctx);
2742 cgroup_migrate_finish(&mgctx);
2743 percpu_up_write(&cgroup_threadgroup_rwsem);
2748 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2749 * @cgrp: root of the target subtree
2751 * Because css offlining is asynchronous, userland may try to re-enable a
2752 * controller while the previous css is still around. This function grabs
2753 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2755 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2756 __acquires(&cgroup_mutex)
2758 struct cgroup *dsct;
2759 struct cgroup_subsys_state *d_css;
2760 struct cgroup_subsys *ss;
2764 mutex_lock(&cgroup_mutex);
2766 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2767 for_each_subsys(ss, ssid) {
2768 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2771 if (!css || !percpu_ref_is_dying(&css->refcnt))
2774 cgroup_get_live(dsct);
2775 prepare_to_wait(&dsct->offline_waitq, &wait,
2776 TASK_UNINTERRUPTIBLE);
2778 mutex_unlock(&cgroup_mutex);
2780 finish_wait(&dsct->offline_waitq, &wait);
2789 * cgroup_save_control - save control masks of a subtree
2790 * @cgrp: root of the target subtree
2792 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
2793 * prefixed fields for @cgrp's subtree including @cgrp itself.
2795 static void cgroup_save_control(struct cgroup *cgrp)
2797 struct cgroup *dsct;
2798 struct cgroup_subsys_state *d_css;
2800 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2801 dsct->old_subtree_control = dsct->subtree_control;
2802 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2807 * cgroup_propagate_control - refresh control masks of a subtree
2808 * @cgrp: root of the target subtree
2810 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2811 * ->subtree_control and propagate controller availability through the
2812 * subtree so that descendants don't have unavailable controllers enabled.
2814 static void cgroup_propagate_control(struct cgroup *cgrp)
2816 struct cgroup *dsct;
2817 struct cgroup_subsys_state *d_css;
2819 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2820 dsct->subtree_control &= cgroup_control(dsct);
2821 dsct->subtree_ss_mask =
2822 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2823 cgroup_ss_mask(dsct));
2828 * cgroup_restore_control - restore control masks of a subtree
2829 * @cgrp: root of the target subtree
2831 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
2832 * prefixed fields for @cgrp's subtree including @cgrp itself.
2834 static void cgroup_restore_control(struct cgroup *cgrp)
2836 struct cgroup *dsct;
2837 struct cgroup_subsys_state *d_css;
2839 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2840 dsct->subtree_control = dsct->old_subtree_control;
2841 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
2845 static bool css_visible(struct cgroup_subsys_state *css)
2847 struct cgroup_subsys *ss = css->ss;
2848 struct cgroup *cgrp = css->cgroup;
2850 if (cgroup_control(cgrp) & (1 << ss->id))
2852 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
2854 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
2858 * cgroup_apply_control_enable - enable or show csses according to control
2859 * @cgrp: root of the target subtree
2861 * Walk @cgrp's subtree and create new csses or make the existing ones
2862 * visible. A css is created invisible if it's being implicitly enabled
2863 * through dependency. An invisible css is made visible when the userland
2864 * explicitly enables it.
2866 * Returns 0 on success, -errno on failure. On failure, csses which have
2867 * been processed already aren't cleaned up. The caller is responsible for
2868 * cleaning up with cgroup_apply_control_disable().
2870 static int cgroup_apply_control_enable(struct cgroup *cgrp)
2872 struct cgroup *dsct;
2873 struct cgroup_subsys_state *d_css;
2874 struct cgroup_subsys *ss;
2877 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2878 for_each_subsys(ss, ssid) {
2879 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2881 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2883 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
2887 css = css_create(dsct, ss);
2889 return PTR_ERR(css);
2892 if (css_visible(css)) {
2893 ret = css_populate_dir(css);
2904 * cgroup_apply_control_disable - kill or hide csses according to control
2905 * @cgrp: root of the target subtree
2907 * Walk @cgrp's subtree and kill and hide csses so that they match
2908 * cgroup_ss_mask() and cgroup_visible_mask().
2910 * A css is hidden when the userland requests it to be disabled while other
2911 * subsystems are still depending on it. The css must not actively control
2912 * resources and be in the vanilla state if it's made visible again later.
2913 * Controllers which may be depended upon should provide ->css_reset() for
2916 static void cgroup_apply_control_disable(struct cgroup *cgrp)
2918 struct cgroup *dsct;
2919 struct cgroup_subsys_state *d_css;
2920 struct cgroup_subsys *ss;
2923 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2924 for_each_subsys(ss, ssid) {
2925 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2927 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2933 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
2935 } else if (!css_visible(css)) {
2945 * cgroup_apply_control - apply control mask updates to the subtree
2946 * @cgrp: root of the target subtree
2948 * subsystems can be enabled and disabled in a subtree using the following
2951 * 1. Call cgroup_save_control() to stash the current state.
2952 * 2. Update ->subtree_control masks in the subtree as desired.
2953 * 3. Call cgroup_apply_control() to apply the changes.
2954 * 4. Optionally perform other related operations.
2955 * 5. Call cgroup_finalize_control() to finish up.
2957 * This function implements step 3 and propagates the mask changes
2958 * throughout @cgrp's subtree, updates csses accordingly and perform
2959 * process migrations.
2961 static int cgroup_apply_control(struct cgroup *cgrp)
2965 cgroup_propagate_control(cgrp);
2967 ret = cgroup_apply_control_enable(cgrp);
2972 * At this point, cgroup_e_css() results reflect the new csses
2973 * making the following cgroup_update_dfl_csses() properly update
2974 * css associations of all tasks in the subtree.
2976 ret = cgroup_update_dfl_csses(cgrp);
2984 * cgroup_finalize_control - finalize control mask update
2985 * @cgrp: root of the target subtree
2986 * @ret: the result of the update
2988 * Finalize control mask update. See cgroup_apply_control() for more info.
2990 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
2993 cgroup_restore_control(cgrp);
2994 cgroup_propagate_control(cgrp);
2997 cgroup_apply_control_disable(cgrp);
3000 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3002 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3004 /* if nothing is getting enabled, nothing to worry about */
3008 /* can @cgrp host any resources? */
3009 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3012 /* mixables don't care */
3013 if (cgroup_is_mixable(cgrp))
3016 if (domain_enable) {
3017 /* can't enable domain controllers inside a thread subtree */
3018 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3022 * Threaded controllers can handle internal competitions
3023 * and are always allowed inside a (prospective) thread
3026 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3031 * Controllers can't be enabled for a cgroup with tasks to avoid
3032 * child cgroups competing against tasks.
3034 if (cgroup_has_tasks(cgrp))
3040 /* change the enabled child controllers for a cgroup in the default hierarchy */
3041 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3042 char *buf, size_t nbytes,
3045 u16 enable = 0, disable = 0;
3046 struct cgroup *cgrp, *child;
3047 struct cgroup_subsys *ss;
3052 * Parse input - space separated list of subsystem names prefixed
3053 * with either + or -.
3055 buf = strstrip(buf);
3056 while ((tok = strsep(&buf, " "))) {
3059 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3060 if (!cgroup_ssid_enabled(ssid) ||
3061 strcmp(tok + 1, ss->name))
3065 enable |= 1 << ssid;
3066 disable &= ~(1 << ssid);
3067 } else if (*tok == '-') {
3068 disable |= 1 << ssid;
3069 enable &= ~(1 << ssid);
3074 } while_each_subsys_mask();
3075 if (ssid == CGROUP_SUBSYS_COUNT)
3079 cgrp = cgroup_kn_lock_live(of->kn, true);
3083 for_each_subsys(ss, ssid) {
3084 if (enable & (1 << ssid)) {
3085 if (cgrp->subtree_control & (1 << ssid)) {
3086 enable &= ~(1 << ssid);
3090 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3094 } else if (disable & (1 << ssid)) {
3095 if (!(cgrp->subtree_control & (1 << ssid))) {
3096 disable &= ~(1 << ssid);
3100 /* a child has it enabled? */
3101 cgroup_for_each_live_child(child, cgrp) {
3102 if (child->subtree_control & (1 << ssid)) {
3110 if (!enable && !disable) {
3115 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3119 /* save and update control masks and prepare csses */
3120 cgroup_save_control(cgrp);
3122 cgrp->subtree_control |= enable;
3123 cgrp->subtree_control &= ~disable;
3125 ret = cgroup_apply_control(cgrp);
3127 cgroup_finalize_control(cgrp, ret);
3129 kernfs_activate(cgrp->kn);
3132 cgroup_kn_unlock(of->kn);
3133 return ret ?: nbytes;
3137 * cgroup_enable_threaded - make @cgrp threaded
3138 * @cgrp: the target cgroup
3140 * Called when "threaded" is written to the cgroup.type interface file and
3141 * tries to make @cgrp threaded and join the parent's resource domain.
3142 * This function is never called on the root cgroup as cgroup.type doesn't
3145 static int cgroup_enable_threaded(struct cgroup *cgrp)
3147 struct cgroup *parent = cgroup_parent(cgrp);
3148 struct cgroup *dom_cgrp = parent->dom_cgrp;
3151 lockdep_assert_held(&cgroup_mutex);
3153 /* noop if already threaded */
3154 if (cgroup_is_threaded(cgrp))
3157 /* we're joining the parent's domain, ensure its validity */
3158 if (!cgroup_is_valid_domain(dom_cgrp) ||
3159 !cgroup_can_be_thread_root(dom_cgrp))
3163 * The following shouldn't cause actual migrations and should
3166 cgroup_save_control(cgrp);
3168 cgrp->dom_cgrp = dom_cgrp;
3169 ret = cgroup_apply_control(cgrp);
3171 parent->nr_threaded_children++;
3173 cgrp->dom_cgrp = cgrp;
3175 cgroup_finalize_control(cgrp, ret);
3179 static int cgroup_type_show(struct seq_file *seq, void *v)
3181 struct cgroup *cgrp = seq_css(seq)->cgroup;
3183 if (cgroup_is_threaded(cgrp))
3184 seq_puts(seq, "threaded\n");
3185 else if (!cgroup_is_valid_domain(cgrp))
3186 seq_puts(seq, "domain invalid\n");
3187 else if (cgroup_is_thread_root(cgrp))
3188 seq_puts(seq, "domain threaded\n");
3190 seq_puts(seq, "domain\n");
3195 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3196 size_t nbytes, loff_t off)
3198 struct cgroup *cgrp;
3201 /* only switching to threaded mode is supported */
3202 if (strcmp(strstrip(buf), "threaded"))
3205 cgrp = cgroup_kn_lock_live(of->kn, false);
3209 /* threaded can only be enabled */
3210 ret = cgroup_enable_threaded(cgrp);
3212 cgroup_kn_unlock(of->kn);
3213 return ret ?: nbytes;
3216 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3218 struct cgroup *cgrp = seq_css(seq)->cgroup;
3219 int descendants = READ_ONCE(cgrp->max_descendants);
3221 if (descendants == INT_MAX)
3222 seq_puts(seq, "max\n");
3224 seq_printf(seq, "%d\n", descendants);
3229 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3230 char *buf, size_t nbytes, loff_t off)
3232 struct cgroup *cgrp;
3236 buf = strstrip(buf);
3237 if (!strcmp(buf, "max")) {
3238 descendants = INT_MAX;
3240 ret = kstrtoint(buf, 0, &descendants);
3245 if (descendants < 0 || descendants > INT_MAX)
3248 cgrp = cgroup_kn_lock_live(of->kn, false);
3252 cgrp->max_descendants = descendants;
3254 cgroup_kn_unlock(of->kn);
3259 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3261 struct cgroup *cgrp = seq_css(seq)->cgroup;
3262 int depth = READ_ONCE(cgrp->max_depth);
3264 if (depth == INT_MAX)
3265 seq_puts(seq, "max\n");
3267 seq_printf(seq, "%d\n", depth);
3272 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3273 char *buf, size_t nbytes, loff_t off)
3275 struct cgroup *cgrp;
3279 buf = strstrip(buf);
3280 if (!strcmp(buf, "max")) {
3283 ret = kstrtoint(buf, 0, &depth);
3288 if (depth < 0 || depth > INT_MAX)
3291 cgrp = cgroup_kn_lock_live(of->kn, false);
3295 cgrp->max_depth = depth;
3297 cgroup_kn_unlock(of->kn);
3302 static int cgroup_events_show(struct seq_file *seq, void *v)
3304 seq_printf(seq, "populated %d\n",
3305 cgroup_is_populated(seq_css(seq)->cgroup));
3309 static int cgroup_stats_show(struct seq_file *seq, void *v)
3311 struct cgroup *cgroup = seq_css(seq)->cgroup;
3313 seq_printf(seq, "nr_descendants %d\n",
3314 cgroup->nr_descendants);
3315 seq_printf(seq, "nr_dying_descendants %d\n",
3316 cgroup->nr_dying_descendants);
3321 static int cgroup_file_open(struct kernfs_open_file *of)
3323 struct cftype *cft = of->kn->priv;
3326 return cft->open(of);
3330 static void cgroup_file_release(struct kernfs_open_file *of)
3332 struct cftype *cft = of->kn->priv;
3338 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3339 size_t nbytes, loff_t off)
3341 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3342 struct cgroup *cgrp = of->kn->parent->priv;
3343 struct cftype *cft = of->kn->priv;
3344 struct cgroup_subsys_state *css;
3348 * If namespaces are delegation boundaries, disallow writes to
3349 * files in an non-init namespace root from inside the namespace
3350 * except for the files explicitly marked delegatable -
3351 * cgroup.procs and cgroup.subtree_control.
3353 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3354 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3355 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3359 return cft->write(of, buf, nbytes, off);
3362 * kernfs guarantees that a file isn't deleted with operations in
3363 * flight, which means that the matching css is and stays alive and
3364 * doesn't need to be pinned. The RCU locking is not necessary
3365 * either. It's just for the convenience of using cgroup_css().
3368 css = cgroup_css(cgrp, cft->ss);
3371 if (cft->write_u64) {
3372 unsigned long long v;
3373 ret = kstrtoull(buf, 0, &v);
3375 ret = cft->write_u64(css, cft, v);
3376 } else if (cft->write_s64) {
3378 ret = kstrtoll(buf, 0, &v);
3380 ret = cft->write_s64(css, cft, v);
3385 return ret ?: nbytes;
3388 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3390 return seq_cft(seq)->seq_start(seq, ppos);
3393 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3395 return seq_cft(seq)->seq_next(seq, v, ppos);
3398 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3400 if (seq_cft(seq)->seq_stop)
3401 seq_cft(seq)->seq_stop(seq, v);
3404 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3406 struct cftype *cft = seq_cft(m);
3407 struct cgroup_subsys_state *css = seq_css(m);
3410 return cft->seq_show(m, arg);
3413 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3414 else if (cft->read_s64)
3415 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3421 static struct kernfs_ops cgroup_kf_single_ops = {
3422 .atomic_write_len = PAGE_SIZE,
3423 .open = cgroup_file_open,
3424 .release = cgroup_file_release,
3425 .write = cgroup_file_write,
3426 .seq_show = cgroup_seqfile_show,
3429 static struct kernfs_ops cgroup_kf_ops = {
3430 .atomic_write_len = PAGE_SIZE,
3431 .open = cgroup_file_open,
3432 .release = cgroup_file_release,
3433 .write = cgroup_file_write,
3434 .seq_start = cgroup_seqfile_start,
3435 .seq_next = cgroup_seqfile_next,
3436 .seq_stop = cgroup_seqfile_stop,
3437 .seq_show = cgroup_seqfile_show,
3440 /* set uid and gid of cgroup dirs and files to that of the creator */
3441 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3443 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3444 .ia_uid = current_fsuid(),
3445 .ia_gid = current_fsgid(), };
3447 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3448 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3451 return kernfs_setattr(kn, &iattr);
3454 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3457 char name[CGROUP_FILE_NAME_MAX];
3458 struct kernfs_node *kn;
3459 struct lock_class_key *key = NULL;
3462 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3463 key = &cft->lockdep_key;
3465 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3466 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3471 ret = cgroup_kn_set_ugid(kn);
3477 if (cft->file_offset) {
3478 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3480 spin_lock_irq(&cgroup_file_kn_lock);
3482 spin_unlock_irq(&cgroup_file_kn_lock);
3489 * cgroup_addrm_files - add or remove files to a cgroup directory
3490 * @css: the target css
3491 * @cgrp: the target cgroup (usually css->cgroup)
3492 * @cfts: array of cftypes to be added
3493 * @is_add: whether to add or remove
3495 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3496 * For removals, this function never fails.
3498 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3499 struct cgroup *cgrp, struct cftype cfts[],
3502 struct cftype *cft, *cft_end = NULL;
3505 lockdep_assert_held(&cgroup_mutex);
3508 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3509 /* does cft->flags tell us to skip this file on @cgrp? */
3510 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3512 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3514 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3516 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3520 ret = cgroup_add_file(css, cgrp, cft);
3522 pr_warn("%s: failed to add %s, err=%d\n",
3523 __func__, cft->name, ret);
3529 cgroup_rm_file(cgrp, cft);
3535 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3537 struct cgroup_subsys *ss = cfts[0].ss;
3538 struct cgroup *root = &ss->root->cgrp;
3539 struct cgroup_subsys_state *css;
3542 lockdep_assert_held(&cgroup_mutex);
3544 /* add/rm files for all cgroups created before */
3545 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3546 struct cgroup *cgrp = css->cgroup;
3548 if (!(css->flags & CSS_VISIBLE))
3551 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3557 kernfs_activate(root->kn);
3561 static void cgroup_exit_cftypes(struct cftype *cfts)
3565 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3566 /* free copy for custom atomic_write_len, see init_cftypes() */
3567 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3572 /* revert flags set by cgroup core while adding @cfts */
3573 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3577 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3581 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3582 struct kernfs_ops *kf_ops;
3584 WARN_ON(cft->ss || cft->kf_ops);
3587 kf_ops = &cgroup_kf_ops;
3589 kf_ops = &cgroup_kf_single_ops;
3592 * Ugh... if @cft wants a custom max_write_len, we need to
3593 * make a copy of kf_ops to set its atomic_write_len.
3595 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3596 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3598 cgroup_exit_cftypes(cfts);
3601 kf_ops->atomic_write_len = cft->max_write_len;
3604 cft->kf_ops = kf_ops;
3611 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3613 lockdep_assert_held(&cgroup_mutex);
3615 if (!cfts || !cfts[0].ss)
3618 list_del(&cfts->node);
3619 cgroup_apply_cftypes(cfts, false);
3620 cgroup_exit_cftypes(cfts);
3625 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3626 * @cfts: zero-length name terminated array of cftypes
3628 * Unregister @cfts. Files described by @cfts are removed from all
3629 * existing cgroups and all future cgroups won't have them either. This
3630 * function can be called anytime whether @cfts' subsys is attached or not.
3632 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3635 int cgroup_rm_cftypes(struct cftype *cfts)
3639 mutex_lock(&cgroup_mutex);
3640 ret = cgroup_rm_cftypes_locked(cfts);
3641 mutex_unlock(&cgroup_mutex);
3646 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3647 * @ss: target cgroup subsystem
3648 * @cfts: zero-length name terminated array of cftypes
3650 * Register @cfts to @ss. Files described by @cfts are created for all
3651 * existing cgroups to which @ss is attached and all future cgroups will
3652 * have them too. This function can be called anytime whether @ss is
3655 * Returns 0 on successful registration, -errno on failure. Note that this
3656 * function currently returns 0 as long as @cfts registration is successful
3657 * even if some file creation attempts on existing cgroups fail.
3659 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3663 if (!cgroup_ssid_enabled(ss->id))
3666 if (!cfts || cfts[0].name[0] == '\0')
3669 ret = cgroup_init_cftypes(ss, cfts);
3673 mutex_lock(&cgroup_mutex);
3675 list_add_tail(&cfts->node, &ss->cfts);
3676 ret = cgroup_apply_cftypes(cfts, true);
3678 cgroup_rm_cftypes_locked(cfts);
3680 mutex_unlock(&cgroup_mutex);
3685 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3686 * @ss: target cgroup subsystem
3687 * @cfts: zero-length name terminated array of cftypes
3689 * Similar to cgroup_add_cftypes() but the added files are only used for
3690 * the default hierarchy.
3692 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3696 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3697 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3698 return cgroup_add_cftypes(ss, cfts);
3702 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3703 * @ss: target cgroup subsystem
3704 * @cfts: zero-length name terminated array of cftypes
3706 * Similar to cgroup_add_cftypes() but the added files are only used for
3707 * the legacy hierarchies.
3709 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3713 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3714 cft->flags |= __CFTYPE_NOT_ON_DFL;
3715 return cgroup_add_cftypes(ss, cfts);
3719 * cgroup_file_notify - generate a file modified event for a cgroup_file
3720 * @cfile: target cgroup_file
3722 * @cfile must have been obtained by setting cftype->file_offset.
3724 void cgroup_file_notify(struct cgroup_file *cfile)
3726 unsigned long flags;
3728 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3730 kernfs_notify(cfile->kn);
3731 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3735 * css_next_child - find the next child of a given css
3736 * @pos: the current position (%NULL to initiate traversal)
3737 * @parent: css whose children to walk
3739 * This function returns the next child of @parent and should be called
3740 * under either cgroup_mutex or RCU read lock. The only requirement is
3741 * that @parent and @pos are accessible. The next sibling is guaranteed to
3742 * be returned regardless of their states.
3744 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3745 * css which finished ->css_online() is guaranteed to be visible in the
3746 * future iterations and will stay visible until the last reference is put.
3747 * A css which hasn't finished ->css_online() or already finished
3748 * ->css_offline() may show up during traversal. It's each subsystem's
3749 * responsibility to synchronize against on/offlining.
3751 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3752 struct cgroup_subsys_state *parent)
3754 struct cgroup_subsys_state *next;
3756 cgroup_assert_mutex_or_rcu_locked();
3759 * @pos could already have been unlinked from the sibling list.
3760 * Once a cgroup is removed, its ->sibling.next is no longer
3761 * updated when its next sibling changes. CSS_RELEASED is set when
3762 * @pos is taken off list, at which time its next pointer is valid,
3763 * and, as releases are serialized, the one pointed to by the next
3764 * pointer is guaranteed to not have started release yet. This
3765 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3766 * critical section, the one pointed to by its next pointer is
3767 * guaranteed to not have finished its RCU grace period even if we
3768 * have dropped rcu_read_lock() inbetween iterations.
3770 * If @pos has CSS_RELEASED set, its next pointer can't be
3771 * dereferenced; however, as each css is given a monotonically
3772 * increasing unique serial number and always appended to the
3773 * sibling list, the next one can be found by walking the parent's
3774 * children until the first css with higher serial number than
3775 * @pos's. While this path can be slower, it happens iff iteration
3776 * races against release and the race window is very small.
3779 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3780 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3781 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3783 list_for_each_entry_rcu(next, &parent->children, sibling)
3784 if (next->serial_nr > pos->serial_nr)
3789 * @next, if not pointing to the head, can be dereferenced and is
3792 if (&next->sibling != &parent->children)
3798 * css_next_descendant_pre - find the next descendant for pre-order walk
3799 * @pos: the current position (%NULL to initiate traversal)
3800 * @root: css whose descendants to walk
3802 * To be used by css_for_each_descendant_pre(). Find the next descendant
3803 * to visit for pre-order traversal of @root's descendants. @root is
3804 * included in the iteration and the first node to be visited.
3806 * While this function requires cgroup_mutex or RCU read locking, it
3807 * doesn't require the whole traversal to be contained in a single critical
3808 * section. This function will return the correct next descendant as long
3809 * as both @pos and @root are accessible and @pos is a descendant of @root.
3811 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3812 * css which finished ->css_online() is guaranteed to be visible in the
3813 * future iterations and will stay visible until the last reference is put.
3814 * A css which hasn't finished ->css_online() or already finished
3815 * ->css_offline() may show up during traversal. It's each subsystem's
3816 * responsibility to synchronize against on/offlining.
3818 struct cgroup_subsys_state *
3819 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3820 struct cgroup_subsys_state *root)
3822 struct cgroup_subsys_state *next;
3824 cgroup_assert_mutex_or_rcu_locked();
3826 /* if first iteration, visit @root */
3830 /* visit the first child if exists */
3831 next = css_next_child(NULL, pos);
3835 /* no child, visit my or the closest ancestor's next sibling */
3836 while (pos != root) {
3837 next = css_next_child(pos, pos->parent);
3847 * css_rightmost_descendant - return the rightmost descendant of a css
3848 * @pos: css of interest
3850 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3851 * is returned. This can be used during pre-order traversal to skip
3854 * While this function requires cgroup_mutex or RCU read locking, it
3855 * doesn't require the whole traversal to be contained in a single critical
3856 * section. This function will return the correct rightmost descendant as
3857 * long as @pos is accessible.
3859 struct cgroup_subsys_state *
3860 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3862 struct cgroup_subsys_state *last, *tmp;
3864 cgroup_assert_mutex_or_rcu_locked();
3868 /* ->prev isn't RCU safe, walk ->next till the end */
3870 css_for_each_child(tmp, last)
3877 static struct cgroup_subsys_state *
3878 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3880 struct cgroup_subsys_state *last;
3884 pos = css_next_child(NULL, pos);
3891 * css_next_descendant_post - find the next descendant for post-order walk
3892 * @pos: the current position (%NULL to initiate traversal)
3893 * @root: css whose descendants to walk
3895 * To be used by css_for_each_descendant_post(). Find the next descendant
3896 * to visit for post-order traversal of @root's descendants. @root is
3897 * included in the iteration and the last node to be visited.
3899 * While this function requires cgroup_mutex or RCU read locking, it
3900 * doesn't require the whole traversal to be contained in a single critical
3901 * section. This function will return the correct next descendant as long
3902 * as both @pos and @cgroup are accessible and @pos is a descendant of
3905 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3906 * css which finished ->css_online() is guaranteed to be visible in the
3907 * future iterations and will stay visible until the last reference is put.
3908 * A css which hasn't finished ->css_online() or already finished
3909 * ->css_offline() may show up during traversal. It's each subsystem's
3910 * responsibility to synchronize against on/offlining.
3912 struct cgroup_subsys_state *
3913 css_next_descendant_post(struct cgroup_subsys_state *pos,
3914 struct cgroup_subsys_state *root)
3916 struct cgroup_subsys_state *next;
3918 cgroup_assert_mutex_or_rcu_locked();
3920 /* if first iteration, visit leftmost descendant which may be @root */
3922 return css_leftmost_descendant(root);
3924 /* if we visited @root, we're done */
3928 /* if there's an unvisited sibling, visit its leftmost descendant */
3929 next = css_next_child(pos, pos->parent);
3931 return css_leftmost_descendant(next);
3933 /* no sibling left, visit parent */
3938 * css_has_online_children - does a css have online children
3939 * @css: the target css
3941 * Returns %true if @css has any online children; otherwise, %false. This
3942 * function can be called from any context but the caller is responsible
3943 * for synchronizing against on/offlining as necessary.
3945 bool css_has_online_children(struct cgroup_subsys_state *css)
3947 struct cgroup_subsys_state *child;
3951 css_for_each_child(child, css) {
3952 if (child->flags & CSS_ONLINE) {
3961 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
3963 struct list_head *l;
3964 struct cgrp_cset_link *link;
3965 struct css_set *cset;
3967 lockdep_assert_held(&css_set_lock);
3969 /* find the next threaded cset */
3970 if (it->tcset_pos) {
3971 l = it->tcset_pos->next;
3973 if (l != it->tcset_head) {
3975 return container_of(l, struct css_set,
3976 threaded_csets_node);
3979 it->tcset_pos = NULL;
3982 /* find the next cset */
3985 if (l == it->cset_head) {
3986 it->cset_pos = NULL;
3991 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
3993 link = list_entry(l, struct cgrp_cset_link, cset_link);
3999 /* initialize threaded css_set walking */
4000 if (it->flags & CSS_TASK_ITER_THREADED) {
4002 put_css_set_locked(it->cur_dcset);
4003 it->cur_dcset = cset;
4006 it->tcset_head = &cset->threaded_csets;
4007 it->tcset_pos = &cset->threaded_csets;
4014 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4015 * @it: the iterator to advance
4017 * Advance @it to the next css_set to walk.
4019 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4021 struct css_set *cset;
4023 lockdep_assert_held(&css_set_lock);
4025 /* Advance to the next non-empty css_set */
4027 cset = css_task_iter_next_css_set(it);
4029 it->task_pos = NULL;
4032 } while (!css_set_populated(cset));
4034 if (!list_empty(&cset->tasks))
4035 it->task_pos = cset->tasks.next;
4037 it->task_pos = cset->mg_tasks.next;
4039 it->tasks_head = &cset->tasks;
4040 it->mg_tasks_head = &cset->mg_tasks;
4043 * We don't keep css_sets locked across iteration steps and thus
4044 * need to take steps to ensure that iteration can be resumed after
4045 * the lock is re-acquired. Iteration is performed at two levels -
4046 * css_sets and tasks in them.
4048 * Once created, a css_set never leaves its cgroup lists, so a
4049 * pinned css_set is guaranteed to stay put and we can resume
4050 * iteration afterwards.
4052 * Tasks may leave @cset across iteration steps. This is resolved
4053 * by registering each iterator with the css_set currently being
4054 * walked and making css_set_move_task() advance iterators whose
4055 * next task is leaving.
4058 list_del(&it->iters_node);
4059 put_css_set_locked(it->cur_cset);
4062 it->cur_cset = cset;
4063 list_add(&it->iters_node, &cset->task_iters);
4066 static void css_task_iter_advance(struct css_task_iter *it)
4068 struct list_head *l = it->task_pos;
4070 lockdep_assert_held(&css_set_lock);
4075 * Advance iterator to find next entry. cset->tasks is consumed
4076 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4081 if (l == it->tasks_head)
4082 l = it->mg_tasks_head->next;
4084 if (l == it->mg_tasks_head)
4085 css_task_iter_advance_css_set(it);
4089 /* if PROCS, skip over tasks which aren't group leaders */
4090 if ((it->flags & CSS_TASK_ITER_PROCS) && it->task_pos &&
4091 !thread_group_leader(list_entry(it->task_pos, struct task_struct,
4097 * css_task_iter_start - initiate task iteration
4098 * @css: the css to walk tasks of
4099 * @flags: CSS_TASK_ITER_* flags
4100 * @it: the task iterator to use
4102 * Initiate iteration through the tasks of @css. The caller can call
4103 * css_task_iter_next() to walk through the tasks until the function
4104 * returns NULL. On completion of iteration, css_task_iter_end() must be
4107 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4108 struct css_task_iter *it)
4110 /* no one should try to iterate before mounting cgroups */
4111 WARN_ON_ONCE(!use_task_css_set_links);
4113 memset(it, 0, sizeof(*it));
4115 spin_lock_irq(&css_set_lock);
4121 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4123 it->cset_pos = &css->cgroup->cset_links;
4125 it->cset_head = it->cset_pos;
4127 css_task_iter_advance_css_set(it);
4129 spin_unlock_irq(&css_set_lock);
4133 * css_task_iter_next - return the next task for the iterator
4134 * @it: the task iterator being iterated
4136 * The "next" function for task iteration. @it should have been
4137 * initialized via css_task_iter_start(). Returns NULL when the iteration
4140 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4143 put_task_struct(it->cur_task);
4144 it->cur_task = NULL;
4147 spin_lock_irq(&css_set_lock);
4150 it->cur_task = list_entry(it->task_pos, struct task_struct,
4152 get_task_struct(it->cur_task);
4153 css_task_iter_advance(it);
4156 spin_unlock_irq(&css_set_lock);
4158 return it->cur_task;
4162 * css_task_iter_end - finish task iteration
4163 * @it: the task iterator to finish
4165 * Finish task iteration started by css_task_iter_start().
4167 void css_task_iter_end(struct css_task_iter *it)
4170 spin_lock_irq(&css_set_lock);
4171 list_del(&it->iters_node);
4172 put_css_set_locked(it->cur_cset);
4173 spin_unlock_irq(&css_set_lock);
4177 put_css_set(it->cur_dcset);
4180 put_task_struct(it->cur_task);
4183 static void cgroup_procs_release(struct kernfs_open_file *of)
4186 css_task_iter_end(of->priv);
4191 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4193 struct kernfs_open_file *of = s->private;
4194 struct css_task_iter *it = of->priv;
4196 return css_task_iter_next(it);
4199 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4200 unsigned int iter_flags)
4202 struct kernfs_open_file *of = s->private;
4203 struct cgroup *cgrp = seq_css(s)->cgroup;
4204 struct css_task_iter *it = of->priv;
4207 * When a seq_file is seeked, it's always traversed sequentially
4208 * from position 0, so we can simply keep iterating on !0 *pos.
4211 if (WARN_ON_ONCE((*pos)++))
4212 return ERR_PTR(-EINVAL);
4214 it = kzalloc(sizeof(*it), GFP_KERNEL);
4216 return ERR_PTR(-ENOMEM);
4218 css_task_iter_start(&cgrp->self, iter_flags, it);
4219 } else if (!(*pos)++) {
4220 css_task_iter_end(it);
4221 css_task_iter_start(&cgrp->self, iter_flags, it);
4224 return cgroup_procs_next(s, NULL, NULL);
4227 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4229 struct cgroup *cgrp = seq_css(s)->cgroup;
4232 * All processes of a threaded subtree belong to the domain cgroup
4233 * of the subtree. Only threads can be distributed across the
4234 * subtree. Reject reads on cgroup.procs in the subtree proper.
4235 * They're always empty anyway.
4237 if (cgroup_is_threaded(cgrp))
4238 return ERR_PTR(-EOPNOTSUPP);
4240 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4241 CSS_TASK_ITER_THREADED);
4244 static int cgroup_procs_show(struct seq_file *s, void *v)
4246 seq_printf(s, "%d\n", task_pid_vnr(v));
4250 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4251 struct cgroup *dst_cgrp,
4252 struct super_block *sb)
4254 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4255 struct cgroup *com_cgrp = src_cgrp;
4256 struct inode *inode;
4259 lockdep_assert_held(&cgroup_mutex);
4261 /* find the common ancestor */
4262 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4263 com_cgrp = cgroup_parent(com_cgrp);
4265 /* %current should be authorized to migrate to the common ancestor */
4266 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4270 ret = inode_permission(inode, MAY_WRITE);
4276 * If namespaces are delegation boundaries, %current must be able
4277 * to see both source and destination cgroups from its namespace.
4279 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4280 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4281 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4287 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4288 char *buf, size_t nbytes, loff_t off)
4290 struct cgroup *src_cgrp, *dst_cgrp;
4291 struct task_struct *task;
4294 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4298 task = cgroup_procs_write_start(buf, true);
4299 ret = PTR_ERR_OR_ZERO(task);
4303 /* find the source cgroup */
4304 spin_lock_irq(&css_set_lock);
4305 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4306 spin_unlock_irq(&css_set_lock);
4308 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4309 of->file->f_path.dentry->d_sb);
4313 ret = cgroup_attach_task(dst_cgrp, task, true);
4316 cgroup_procs_write_finish(task);
4318 cgroup_kn_unlock(of->kn);
4320 return ret ?: nbytes;
4323 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4325 return __cgroup_procs_start(s, pos, 0);
4328 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4329 char *buf, size_t nbytes, loff_t off)
4331 struct cgroup *src_cgrp, *dst_cgrp;
4332 struct task_struct *task;
4335 buf = strstrip(buf);
4337 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4341 task = cgroup_procs_write_start(buf, false);
4342 ret = PTR_ERR_OR_ZERO(task);
4346 /* find the source cgroup */
4347 spin_lock_irq(&css_set_lock);
4348 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4349 spin_unlock_irq(&css_set_lock);
4351 /* thread migrations follow the cgroup.procs delegation rule */
4352 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4353 of->file->f_path.dentry->d_sb);
4357 /* and must be contained in the same domain */
4359 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4362 ret = cgroup_attach_task(dst_cgrp, task, false);
4365 cgroup_procs_write_finish(task);
4367 cgroup_kn_unlock(of->kn);
4369 return ret ?: nbytes;
4372 /* cgroup core interface files for the default hierarchy */
4373 static struct cftype cgroup_base_files[] = {
4375 .name = "cgroup.type",
4376 .flags = CFTYPE_NOT_ON_ROOT,
4377 .seq_show = cgroup_type_show,
4378 .write = cgroup_type_write,
4381 .name = "cgroup.procs",
4382 .flags = CFTYPE_NS_DELEGATABLE,
4383 .file_offset = offsetof(struct cgroup, procs_file),
4384 .release = cgroup_procs_release,
4385 .seq_start = cgroup_procs_start,
4386 .seq_next = cgroup_procs_next,
4387 .seq_show = cgroup_procs_show,
4388 .write = cgroup_procs_write,
4391 .name = "cgroup.threads",
4392 .release = cgroup_procs_release,
4393 .seq_start = cgroup_threads_start,
4394 .seq_next = cgroup_procs_next,
4395 .seq_show = cgroup_procs_show,
4396 .write = cgroup_threads_write,
4399 .name = "cgroup.controllers",
4400 .seq_show = cgroup_controllers_show,
4403 .name = "cgroup.subtree_control",
4404 .flags = CFTYPE_NS_DELEGATABLE,
4405 .seq_show = cgroup_subtree_control_show,
4406 .write = cgroup_subtree_control_write,
4409 .name = "cgroup.events",
4410 .flags = CFTYPE_NOT_ON_ROOT,
4411 .file_offset = offsetof(struct cgroup, events_file),
4412 .seq_show = cgroup_events_show,
4415 .name = "cgroup.max.descendants",
4416 .seq_show = cgroup_max_descendants_show,
4417 .write = cgroup_max_descendants_write,
4420 .name = "cgroup.max.depth",
4421 .seq_show = cgroup_max_depth_show,
4422 .write = cgroup_max_depth_write,
4425 .name = "cgroup.stat",
4426 .seq_show = cgroup_stats_show,
4432 * css destruction is four-stage process.
4434 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4435 * Implemented in kill_css().
4437 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4438 * and thus css_tryget_online() is guaranteed to fail, the css can be
4439 * offlined by invoking offline_css(). After offlining, the base ref is
4440 * put. Implemented in css_killed_work_fn().
4442 * 3. When the percpu_ref reaches zero, the only possible remaining
4443 * accessors are inside RCU read sections. css_release() schedules the
4446 * 4. After the grace period, the css can be freed. Implemented in
4447 * css_free_work_fn().
4449 * It is actually hairier because both step 2 and 4 require process context
4450 * and thus involve punting to css->destroy_work adding two additional
4451 * steps to the already complex sequence.
4453 static void css_free_work_fn(struct work_struct *work)
4455 struct cgroup_subsys_state *css =
4456 container_of(work, struct cgroup_subsys_state, destroy_work);
4457 struct cgroup_subsys *ss = css->ss;
4458 struct cgroup *cgrp = css->cgroup;
4460 percpu_ref_exit(&css->refcnt);
4464 struct cgroup_subsys_state *parent = css->parent;
4468 cgroup_idr_remove(&ss->css_idr, id);
4474 /* cgroup free path */
4475 atomic_dec(&cgrp->root->nr_cgrps);
4476 cgroup1_pidlist_destroy_all(cgrp);
4477 cancel_work_sync(&cgrp->release_agent_work);
4479 if (cgroup_parent(cgrp)) {
4481 * We get a ref to the parent, and put the ref when
4482 * this cgroup is being freed, so it's guaranteed
4483 * that the parent won't be destroyed before its
4486 cgroup_put(cgroup_parent(cgrp));
4487 kernfs_put(cgrp->kn);
4491 * This is root cgroup's refcnt reaching zero,
4492 * which indicates that the root should be
4495 cgroup_destroy_root(cgrp->root);
4500 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4502 struct cgroup_subsys_state *css =
4503 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4505 INIT_WORK(&css->destroy_work, css_free_work_fn);
4506 queue_work(cgroup_destroy_wq, &css->destroy_work);
4509 static void css_release_work_fn(struct work_struct *work)
4511 struct cgroup_subsys_state *css =
4512 container_of(work, struct cgroup_subsys_state, destroy_work);
4513 struct cgroup_subsys *ss = css->ss;
4514 struct cgroup *cgrp = css->cgroup;
4516 mutex_lock(&cgroup_mutex);
4518 css->flags |= CSS_RELEASED;
4519 list_del_rcu(&css->sibling);
4522 /* css release path */
4523 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4524 if (ss->css_released)
4525 ss->css_released(css);
4527 struct cgroup *tcgrp;
4529 /* cgroup release path */
4530 trace_cgroup_release(cgrp);
4532 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4533 tcgrp = cgroup_parent(tcgrp))
4534 tcgrp->nr_dying_descendants--;
4536 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4540 * There are two control paths which try to determine
4541 * cgroup from dentry without going through kernfs -
4542 * cgroupstats_build() and css_tryget_online_from_dir().
4543 * Those are supported by RCU protecting clearing of
4544 * cgrp->kn->priv backpointer.
4547 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4550 cgroup_bpf_put(cgrp);
4553 mutex_unlock(&cgroup_mutex);
4555 call_rcu(&css->rcu_head, css_free_rcu_fn);
4558 static void css_release(struct percpu_ref *ref)
4560 struct cgroup_subsys_state *css =
4561 container_of(ref, struct cgroup_subsys_state, refcnt);
4563 INIT_WORK(&css->destroy_work, css_release_work_fn);
4564 queue_work(cgroup_destroy_wq, &css->destroy_work);
4567 static void init_and_link_css(struct cgroup_subsys_state *css,
4568 struct cgroup_subsys *ss, struct cgroup *cgrp)
4570 lockdep_assert_held(&cgroup_mutex);
4572 cgroup_get_live(cgrp);
4574 memset(css, 0, sizeof(*css));
4578 INIT_LIST_HEAD(&css->sibling);
4579 INIT_LIST_HEAD(&css->children);
4580 css->serial_nr = css_serial_nr_next++;
4581 atomic_set(&css->online_cnt, 0);
4583 if (cgroup_parent(cgrp)) {
4584 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4585 css_get(css->parent);
4588 BUG_ON(cgroup_css(cgrp, ss));
4591 /* invoke ->css_online() on a new CSS and mark it online if successful */
4592 static int online_css(struct cgroup_subsys_state *css)
4594 struct cgroup_subsys *ss = css->ss;
4597 lockdep_assert_held(&cgroup_mutex);
4600 ret = ss->css_online(css);
4602 css->flags |= CSS_ONLINE;
4603 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4605 atomic_inc(&css->online_cnt);
4607 atomic_inc(&css->parent->online_cnt);
4612 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4613 static void offline_css(struct cgroup_subsys_state *css)
4615 struct cgroup_subsys *ss = css->ss;
4617 lockdep_assert_held(&cgroup_mutex);
4619 if (!(css->flags & CSS_ONLINE))
4625 if (ss->css_offline)
4626 ss->css_offline(css);
4628 css->flags &= ~CSS_ONLINE;
4629 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4631 wake_up_all(&css->cgroup->offline_waitq);
4635 * css_create - create a cgroup_subsys_state
4636 * @cgrp: the cgroup new css will be associated with
4637 * @ss: the subsys of new css
4639 * Create a new css associated with @cgrp - @ss pair. On success, the new
4640 * css is online and installed in @cgrp. This function doesn't create the
4641 * interface files. Returns 0 on success, -errno on failure.
4643 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4644 struct cgroup_subsys *ss)
4646 struct cgroup *parent = cgroup_parent(cgrp);
4647 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4648 struct cgroup_subsys_state *css;
4651 lockdep_assert_held(&cgroup_mutex);
4653 css = ss->css_alloc(parent_css);
4655 css = ERR_PTR(-ENOMEM);
4659 init_and_link_css(css, ss, cgrp);
4661 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4665 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4670 /* @css is ready to be brought online now, make it visible */
4671 list_add_tail_rcu(&css->sibling, &parent_css->children);
4672 cgroup_idr_replace(&ss->css_idr, css, css->id);
4674 err = online_css(css);
4678 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4679 cgroup_parent(parent)) {
4680 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4681 current->comm, current->pid, ss->name);
4682 if (!strcmp(ss->name, "memory"))
4683 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4684 ss->warned_broken_hierarchy = true;
4690 list_del_rcu(&css->sibling);
4692 call_rcu(&css->rcu_head, css_free_rcu_fn);
4693 return ERR_PTR(err);
4697 * The returned cgroup is fully initialized including its control mask, but
4698 * it isn't associated with its kernfs_node and doesn't have the control
4701 static struct cgroup *cgroup_create(struct cgroup *parent)
4703 struct cgroup_root *root = parent->root;
4704 struct cgroup *cgrp, *tcgrp;
4705 int level = parent->level + 1;
4708 /* allocate the cgroup and its ID, 0 is reserved for the root */
4709 cgrp = kzalloc(sizeof(*cgrp) +
4710 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4712 return ERR_PTR(-ENOMEM);
4714 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4719 * Temporarily set the pointer to NULL, so idr_find() won't return
4720 * a half-baked cgroup.
4722 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4725 goto out_cancel_ref;
4728 init_cgroup_housekeeping(cgrp);
4730 cgrp->self.parent = &parent->self;
4732 cgrp->level = level;
4734 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
4735 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4738 tcgrp->nr_descendants++;
4741 if (notify_on_release(parent))
4742 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4744 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4745 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4747 cgrp->self.serial_nr = css_serial_nr_next++;
4749 /* allocation complete, commit to creation */
4750 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4751 atomic_inc(&root->nr_cgrps);
4752 cgroup_get_live(parent);
4755 * @cgrp is now fully operational. If something fails after this
4756 * point, it'll be released via the normal destruction path.
4758 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4761 * On the default hierarchy, a child doesn't automatically inherit
4762 * subtree_control from the parent. Each is configured manually.
4764 if (!cgroup_on_dfl(cgrp))
4765 cgrp->subtree_control = cgroup_control(cgrp);
4768 cgroup_bpf_inherit(cgrp, parent);
4770 cgroup_propagate_control(cgrp);
4775 percpu_ref_exit(&cgrp->self.refcnt);
4778 return ERR_PTR(ret);
4781 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
4783 struct cgroup *cgroup;
4787 lockdep_assert_held(&cgroup_mutex);
4789 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
4790 if (cgroup->nr_descendants >= cgroup->max_descendants)
4793 if (level > cgroup->max_depth)
4804 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
4806 struct cgroup *parent, *cgrp;
4807 struct kernfs_node *kn;
4810 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
4811 if (strchr(name, '\n'))
4814 parent = cgroup_kn_lock_live(parent_kn, false);
4818 if (!cgroup_check_hierarchy_limits(parent)) {
4823 cgrp = cgroup_create(parent);
4825 ret = PTR_ERR(cgrp);
4829 /* create the directory */
4830 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4838 * This extra ref will be put in cgroup_free_fn() and guarantees
4839 * that @cgrp->kn is always accessible.
4843 ret = cgroup_kn_set_ugid(kn);
4847 ret = css_populate_dir(&cgrp->self);
4851 ret = cgroup_apply_control_enable(cgrp);
4855 trace_cgroup_mkdir(cgrp);
4857 /* let's create and online css's */
4858 kernfs_activate(kn);
4864 cgroup_destroy_locked(cgrp);
4866 cgroup_kn_unlock(parent_kn);
4871 * This is called when the refcnt of a css is confirmed to be killed.
4872 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4873 * initate destruction and put the css ref from kill_css().
4875 static void css_killed_work_fn(struct work_struct *work)
4877 struct cgroup_subsys_state *css =
4878 container_of(work, struct cgroup_subsys_state, destroy_work);
4880 mutex_lock(&cgroup_mutex);
4885 /* @css can't go away while we're holding cgroup_mutex */
4887 } while (css && atomic_dec_and_test(&css->online_cnt));
4889 mutex_unlock(&cgroup_mutex);
4892 /* css kill confirmation processing requires process context, bounce */
4893 static void css_killed_ref_fn(struct percpu_ref *ref)
4895 struct cgroup_subsys_state *css =
4896 container_of(ref, struct cgroup_subsys_state, refcnt);
4898 if (atomic_dec_and_test(&css->online_cnt)) {
4899 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4900 queue_work(cgroup_destroy_wq, &css->destroy_work);
4905 * kill_css - destroy a css
4906 * @css: css to destroy
4908 * This function initiates destruction of @css by removing cgroup interface
4909 * files and putting its base reference. ->css_offline() will be invoked
4910 * asynchronously once css_tryget_online() is guaranteed to fail and when
4911 * the reference count reaches zero, @css will be released.
4913 static void kill_css(struct cgroup_subsys_state *css)
4915 lockdep_assert_held(&cgroup_mutex);
4917 if (css->flags & CSS_DYING)
4920 css->flags |= CSS_DYING;
4923 * This must happen before css is disassociated with its cgroup.
4924 * See seq_css() for details.
4929 * Killing would put the base ref, but we need to keep it alive
4930 * until after ->css_offline().
4935 * cgroup core guarantees that, by the time ->css_offline() is
4936 * invoked, no new css reference will be given out via
4937 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4938 * proceed to offlining css's because percpu_ref_kill() doesn't
4939 * guarantee that the ref is seen as killed on all CPUs on return.
4941 * Use percpu_ref_kill_and_confirm() to get notifications as each
4942 * css is confirmed to be seen as killed on all CPUs.
4944 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4948 * cgroup_destroy_locked - the first stage of cgroup destruction
4949 * @cgrp: cgroup to be destroyed
4951 * css's make use of percpu refcnts whose killing latency shouldn't be
4952 * exposed to userland and are RCU protected. Also, cgroup core needs to
4953 * guarantee that css_tryget_online() won't succeed by the time
4954 * ->css_offline() is invoked. To satisfy all the requirements,
4955 * destruction is implemented in the following two steps.
4957 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4958 * userland visible parts and start killing the percpu refcnts of
4959 * css's. Set up so that the next stage will be kicked off once all
4960 * the percpu refcnts are confirmed to be killed.
4962 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4963 * rest of destruction. Once all cgroup references are gone, the
4964 * cgroup is RCU-freed.
4966 * This function implements s1. After this step, @cgrp is gone as far as
4967 * the userland is concerned and a new cgroup with the same name may be
4968 * created. As cgroup doesn't care about the names internally, this
4969 * doesn't cause any problem.
4971 static int cgroup_destroy_locked(struct cgroup *cgrp)
4972 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4974 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
4975 struct cgroup_subsys_state *css;
4976 struct cgrp_cset_link *link;
4979 lockdep_assert_held(&cgroup_mutex);
4982 * Only migration can raise populated from zero and we're already
4983 * holding cgroup_mutex.
4985 if (cgroup_is_populated(cgrp))
4989 * Make sure there's no live children. We can't test emptiness of
4990 * ->self.children as dead children linger on it while being
4991 * drained; otherwise, "rmdir parent/child parent" may fail.
4993 if (css_has_online_children(&cgrp->self))
4997 * Mark @cgrp and the associated csets dead. The former prevents
4998 * further task migration and child creation by disabling
4999 * cgroup_lock_live_group(). The latter makes the csets ignored by
5000 * the migration path.
5002 cgrp->self.flags &= ~CSS_ONLINE;
5004 spin_lock_irq(&css_set_lock);
5005 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5006 link->cset->dead = true;
5007 spin_unlock_irq(&css_set_lock);
5009 /* initiate massacre of all css's */
5010 for_each_css(css, ssid, cgrp)
5014 * Remove @cgrp directory along with the base files. @cgrp has an
5015 * extra ref on its kn.
5017 kernfs_remove(cgrp->kn);
5019 if (parent && cgroup_is_threaded(cgrp))
5020 parent->nr_threaded_children--;
5022 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5023 tcgrp->nr_descendants--;
5024 tcgrp->nr_dying_descendants++;
5027 cgroup1_check_for_release(parent);
5029 /* put the base reference */
5030 percpu_ref_kill(&cgrp->self.refcnt);
5035 int cgroup_rmdir(struct kernfs_node *kn)
5037 struct cgroup *cgrp;
5040 cgrp = cgroup_kn_lock_live(kn, false);
5044 ret = cgroup_destroy_locked(cgrp);
5047 trace_cgroup_rmdir(cgrp);
5049 cgroup_kn_unlock(kn);
5053 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5054 .show_options = cgroup_show_options,
5055 .remount_fs = cgroup_remount,
5056 .mkdir = cgroup_mkdir,
5057 .rmdir = cgroup_rmdir,
5058 .show_path = cgroup_show_path,
5061 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5063 struct cgroup_subsys_state *css;
5065 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5067 mutex_lock(&cgroup_mutex);
5069 idr_init(&ss->css_idr);
5070 INIT_LIST_HEAD(&ss->cfts);
5072 /* Create the root cgroup state for this subsystem */
5073 ss->root = &cgrp_dfl_root;
5074 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5075 /* We don't handle early failures gracefully */
5076 BUG_ON(IS_ERR(css));
5077 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5080 * Root csses are never destroyed and we can't initialize
5081 * percpu_ref during early init. Disable refcnting.
5083 css->flags |= CSS_NO_REF;
5086 /* allocation can't be done safely during early init */
5089 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5090 BUG_ON(css->id < 0);
5093 /* Update the init_css_set to contain a subsys
5094 * pointer to this state - since the subsystem is
5095 * newly registered, all tasks and hence the
5096 * init_css_set is in the subsystem's root cgroup. */
5097 init_css_set.subsys[ss->id] = css;
5099 have_fork_callback |= (bool)ss->fork << ss->id;
5100 have_exit_callback |= (bool)ss->exit << ss->id;
5101 have_free_callback |= (bool)ss->free << ss->id;
5102 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5104 /* At system boot, before all subsystems have been
5105 * registered, no tasks have been forked, so we don't
5106 * need to invoke fork callbacks here. */
5107 BUG_ON(!list_empty(&init_task.tasks));
5109 BUG_ON(online_css(css));
5111 mutex_unlock(&cgroup_mutex);
5115 * cgroup_init_early - cgroup initialization at system boot
5117 * Initialize cgroups at system boot, and initialize any
5118 * subsystems that request early init.
5120 int __init cgroup_init_early(void)
5122 static struct cgroup_sb_opts __initdata opts;
5123 struct cgroup_subsys *ss;
5126 init_cgroup_root(&cgrp_dfl_root, &opts);
5127 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5129 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5131 for_each_subsys(ss, i) {
5132 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5133 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5134 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5136 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5137 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5140 ss->name = cgroup_subsys_name[i];
5141 if (!ss->legacy_name)
5142 ss->legacy_name = cgroup_subsys_name[i];
5145 cgroup_init_subsys(ss, true);
5150 static u16 cgroup_disable_mask __initdata;
5153 * cgroup_init - cgroup initialization
5155 * Register cgroup filesystem and /proc file, and initialize
5156 * any subsystems that didn't request early init.
5158 int __init cgroup_init(void)
5160 struct cgroup_subsys *ss;
5163 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5164 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5165 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5166 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5169 * The latency of the synchronize_sched() is too high for cgroups,
5170 * avoid it at the cost of forcing all readers into the slow path.
5172 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5174 get_user_ns(init_cgroup_ns.user_ns);
5176 mutex_lock(&cgroup_mutex);
5179 * Add init_css_set to the hash table so that dfl_root can link to
5182 hash_add(css_set_table, &init_css_set.hlist,
5183 css_set_hash(init_css_set.subsys));
5185 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
5187 mutex_unlock(&cgroup_mutex);
5189 for_each_subsys(ss, ssid) {
5190 if (ss->early_init) {
5191 struct cgroup_subsys_state *css =
5192 init_css_set.subsys[ss->id];
5194 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5196 BUG_ON(css->id < 0);
5198 cgroup_init_subsys(ss, false);
5201 list_add_tail(&init_css_set.e_cset_node[ssid],
5202 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5205 * Setting dfl_root subsys_mask needs to consider the
5206 * disabled flag and cftype registration needs kmalloc,
5207 * both of which aren't available during early_init.
5209 if (cgroup_disable_mask & (1 << ssid)) {
5210 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5211 printk(KERN_INFO "Disabling %s control group subsystem\n",
5216 if (cgroup1_ssid_disabled(ssid))
5217 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5220 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5222 /* implicit controllers must be threaded too */
5223 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5225 if (ss->implicit_on_dfl)
5226 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5227 else if (!ss->dfl_cftypes)
5228 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5231 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5233 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5234 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5236 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5237 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5241 ss->bind(init_css_set.subsys[ssid]);
5244 /* init_css_set.subsys[] has been updated, re-hash */
5245 hash_del(&init_css_set.hlist);
5246 hash_add(css_set_table, &init_css_set.hlist,
5247 css_set_hash(init_css_set.subsys));
5249 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5250 WARN_ON(register_filesystem(&cgroup_fs_type));
5251 WARN_ON(register_filesystem(&cgroup2_fs_type));
5252 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5257 static int __init cgroup_wq_init(void)
5260 * There isn't much point in executing destruction path in
5261 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5262 * Use 1 for @max_active.
5264 * We would prefer to do this in cgroup_init() above, but that
5265 * is called before init_workqueues(): so leave this until after.
5267 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5268 BUG_ON(!cgroup_destroy_wq);
5271 core_initcall(cgroup_wq_init);
5274 * proc_cgroup_show()
5275 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5276 * - Used for /proc/<pid>/cgroup.
5278 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5279 struct pid *pid, struct task_struct *tsk)
5283 struct cgroup_root *root;
5286 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5290 mutex_lock(&cgroup_mutex);
5291 spin_lock_irq(&css_set_lock);
5293 for_each_root(root) {
5294 struct cgroup_subsys *ss;
5295 struct cgroup *cgrp;
5296 int ssid, count = 0;
5298 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5301 seq_printf(m, "%d:", root->hierarchy_id);
5302 if (root != &cgrp_dfl_root)
5303 for_each_subsys(ss, ssid)
5304 if (root->subsys_mask & (1 << ssid))
5305 seq_printf(m, "%s%s", count++ ? "," : "",
5307 if (strlen(root->name))
5308 seq_printf(m, "%sname=%s", count ? "," : "",
5312 cgrp = task_cgroup_from_root(tsk, root);
5315 * On traditional hierarchies, all zombie tasks show up as
5316 * belonging to the root cgroup. On the default hierarchy,
5317 * while a zombie doesn't show up in "cgroup.procs" and
5318 * thus can't be migrated, its /proc/PID/cgroup keeps
5319 * reporting the cgroup it belonged to before exiting. If
5320 * the cgroup is removed before the zombie is reaped,
5321 * " (deleted)" is appended to the cgroup path.
5323 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5324 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5325 current->nsproxy->cgroup_ns);
5326 if (retval >= PATH_MAX)
5327 retval = -ENAMETOOLONG;
5336 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5337 seq_puts(m, " (deleted)\n");
5344 spin_unlock_irq(&css_set_lock);
5345 mutex_unlock(&cgroup_mutex);
5352 * cgroup_fork - initialize cgroup related fields during copy_process()
5353 * @child: pointer to task_struct of forking parent process.
5355 * A task is associated with the init_css_set until cgroup_post_fork()
5356 * attaches it to the parent's css_set. Empty cg_list indicates that
5357 * @child isn't holding reference to its css_set.
5359 void cgroup_fork(struct task_struct *child)
5361 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5362 INIT_LIST_HEAD(&child->cg_list);
5366 * cgroup_can_fork - called on a new task before the process is exposed
5367 * @child: the task in question.
5369 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5370 * returns an error, the fork aborts with that error code. This allows for
5371 * a cgroup subsystem to conditionally allow or deny new forks.
5373 int cgroup_can_fork(struct task_struct *child)
5375 struct cgroup_subsys *ss;
5378 do_each_subsys_mask(ss, i, have_canfork_callback) {
5379 ret = ss->can_fork(child);
5382 } while_each_subsys_mask();
5387 for_each_subsys(ss, j) {
5390 if (ss->cancel_fork)
5391 ss->cancel_fork(child);
5398 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5399 * @child: the task in question
5401 * This calls the cancel_fork() callbacks if a fork failed *after*
5402 * cgroup_can_fork() succeded.
5404 void cgroup_cancel_fork(struct task_struct *child)
5406 struct cgroup_subsys *ss;
5409 for_each_subsys(ss, i)
5410 if (ss->cancel_fork)
5411 ss->cancel_fork(child);
5415 * cgroup_post_fork - called on a new task after adding it to the task list
5416 * @child: the task in question
5418 * Adds the task to the list running through its css_set if necessary and
5419 * call the subsystem fork() callbacks. Has to be after the task is
5420 * visible on the task list in case we race with the first call to
5421 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5424 void cgroup_post_fork(struct task_struct *child)
5426 struct cgroup_subsys *ss;
5430 * This may race against cgroup_enable_task_cg_lists(). As that
5431 * function sets use_task_css_set_links before grabbing
5432 * tasklist_lock and we just went through tasklist_lock to add
5433 * @child, it's guaranteed that either we see the set
5434 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5435 * @child during its iteration.
5437 * If we won the race, @child is associated with %current's
5438 * css_set. Grabbing css_set_lock guarantees both that the
5439 * association is stable, and, on completion of the parent's
5440 * migration, @child is visible in the source of migration or
5441 * already in the destination cgroup. This guarantee is necessary
5442 * when implementing operations which need to migrate all tasks of
5443 * a cgroup to another.
5445 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5446 * will remain in init_css_set. This is safe because all tasks are
5447 * in the init_css_set before cg_links is enabled and there's no
5448 * operation which transfers all tasks out of init_css_set.
5450 if (use_task_css_set_links) {
5451 struct css_set *cset;
5453 spin_lock_irq(&css_set_lock);
5454 cset = task_css_set(current);
5455 if (list_empty(&child->cg_list)) {
5458 css_set_move_task(child, NULL, cset, false);
5460 spin_unlock_irq(&css_set_lock);
5464 * Call ss->fork(). This must happen after @child is linked on
5465 * css_set; otherwise, @child might change state between ->fork()
5466 * and addition to css_set.
5468 do_each_subsys_mask(ss, i, have_fork_callback) {
5470 } while_each_subsys_mask();
5474 * cgroup_exit - detach cgroup from exiting task
5475 * @tsk: pointer to task_struct of exiting process
5477 * Description: Detach cgroup from @tsk and release it.
5479 * Note that cgroups marked notify_on_release force every task in
5480 * them to take the global cgroup_mutex mutex when exiting.
5481 * This could impact scaling on very large systems. Be reluctant to
5482 * use notify_on_release cgroups where very high task exit scaling
5483 * is required on large systems.
5485 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5486 * call cgroup_exit() while the task is still competent to handle
5487 * notify_on_release(), then leave the task attached to the root cgroup in
5488 * each hierarchy for the remainder of its exit. No need to bother with
5489 * init_css_set refcnting. init_css_set never goes away and we can't race
5490 * with migration path - PF_EXITING is visible to migration path.
5492 void cgroup_exit(struct task_struct *tsk)
5494 struct cgroup_subsys *ss;
5495 struct css_set *cset;
5499 * Unlink from @tsk from its css_set. As migration path can't race
5500 * with us, we can check css_set and cg_list without synchronization.
5502 cset = task_css_set(tsk);
5504 if (!list_empty(&tsk->cg_list)) {
5505 spin_lock_irq(&css_set_lock);
5506 css_set_move_task(tsk, cset, NULL, false);
5508 spin_unlock_irq(&css_set_lock);
5513 /* see cgroup_post_fork() for details */
5514 do_each_subsys_mask(ss, i, have_exit_callback) {
5516 } while_each_subsys_mask();
5519 void cgroup_free(struct task_struct *task)
5521 struct css_set *cset = task_css_set(task);
5522 struct cgroup_subsys *ss;
5525 do_each_subsys_mask(ss, ssid, have_free_callback) {
5527 } while_each_subsys_mask();
5532 static int __init cgroup_disable(char *str)
5534 struct cgroup_subsys *ss;
5538 while ((token = strsep(&str, ",")) != NULL) {
5542 for_each_subsys(ss, i) {
5543 if (strcmp(token, ss->name) &&
5544 strcmp(token, ss->legacy_name))
5546 cgroup_disable_mask |= 1 << i;
5551 __setup("cgroup_disable=", cgroup_disable);
5554 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5555 * @dentry: directory dentry of interest
5556 * @ss: subsystem of interest
5558 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5559 * to get the corresponding css and return it. If such css doesn't exist
5560 * or can't be pinned, an ERR_PTR value is returned.
5562 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5563 struct cgroup_subsys *ss)
5565 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5566 struct file_system_type *s_type = dentry->d_sb->s_type;
5567 struct cgroup_subsys_state *css = NULL;
5568 struct cgroup *cgrp;
5570 /* is @dentry a cgroup dir? */
5571 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5572 !kn || kernfs_type(kn) != KERNFS_DIR)
5573 return ERR_PTR(-EBADF);
5578 * This path doesn't originate from kernfs and @kn could already
5579 * have been or be removed at any point. @kn->priv is RCU
5580 * protected for this access. See css_release_work_fn() for details.
5582 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
5584 css = cgroup_css(cgrp, ss);
5586 if (!css || !css_tryget_online(css))
5587 css = ERR_PTR(-ENOENT);
5594 * css_from_id - lookup css by id
5595 * @id: the cgroup id
5596 * @ss: cgroup subsys to be looked into
5598 * Returns the css if there's valid one with @id, otherwise returns NULL.
5599 * Should be called under rcu_read_lock().
5601 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5603 WARN_ON_ONCE(!rcu_read_lock_held());
5604 return idr_find(&ss->css_idr, id);
5608 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5609 * @path: path on the default hierarchy
5611 * Find the cgroup at @path on the default hierarchy, increment its
5612 * reference count and return it. Returns pointer to the found cgroup on
5613 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5614 * if @path points to a non-directory.
5616 struct cgroup *cgroup_get_from_path(const char *path)
5618 struct kernfs_node *kn;
5619 struct cgroup *cgrp;
5621 mutex_lock(&cgroup_mutex);
5623 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5625 if (kernfs_type(kn) == KERNFS_DIR) {
5627 cgroup_get_live(cgrp);
5629 cgrp = ERR_PTR(-ENOTDIR);
5633 cgrp = ERR_PTR(-ENOENT);
5636 mutex_unlock(&cgroup_mutex);
5639 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5642 * cgroup_get_from_fd - get a cgroup pointer from a fd
5643 * @fd: fd obtained by open(cgroup2_dir)
5645 * Find the cgroup from a fd which should be obtained
5646 * by opening a cgroup directory. Returns a pointer to the
5647 * cgroup on success. ERR_PTR is returned if the cgroup
5650 struct cgroup *cgroup_get_from_fd(int fd)
5652 struct cgroup_subsys_state *css;
5653 struct cgroup *cgrp;
5658 return ERR_PTR(-EBADF);
5660 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5663 return ERR_CAST(css);
5666 if (!cgroup_on_dfl(cgrp)) {
5668 return ERR_PTR(-EBADF);
5673 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
5676 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5677 * definition in cgroup-defs.h.
5679 #ifdef CONFIG_SOCK_CGROUP_DATA
5681 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5683 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5684 static bool cgroup_sk_alloc_disabled __read_mostly;
5686 void cgroup_sk_alloc_disable(void)
5688 if (cgroup_sk_alloc_disabled)
5690 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5691 cgroup_sk_alloc_disabled = true;
5696 #define cgroup_sk_alloc_disabled false
5700 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5702 if (cgroup_sk_alloc_disabled)
5705 /* Socket clone path */
5708 * We might be cloning a socket which is left in an empty
5709 * cgroup and the cgroup might have already been rmdir'd.
5710 * Don't use cgroup_get_live().
5712 cgroup_get(sock_cgroup_ptr(skcd));
5719 struct css_set *cset;
5721 cset = task_css_set(current);
5722 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5723 skcd->val = (unsigned long)cset->dfl_cgrp;
5732 void cgroup_sk_free(struct sock_cgroup_data *skcd)
5734 cgroup_put(sock_cgroup_ptr(skcd));
5737 #endif /* CONFIG_SOCK_CGROUP_DATA */
5739 #ifdef CONFIG_CGROUP_BPF
5740 int cgroup_bpf_update(struct cgroup *cgrp, struct bpf_prog *prog,
5741 enum bpf_attach_type type, bool overridable)
5743 struct cgroup *parent = cgroup_parent(cgrp);
5746 mutex_lock(&cgroup_mutex);
5747 ret = __cgroup_bpf_update(cgrp, parent, prog, type, overridable);
5748 mutex_unlock(&cgroup_mutex);
5751 #endif /* CONFIG_CGROUP_BPF */