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) {
1315 struct cgrp_cset_link *link;
1317 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1318 struct cgroup *c = link->cgrp;
1320 if (c->root == root) {
1332 * Return the cgroup for "task" from the given hierarchy. Must be
1333 * called with cgroup_mutex and css_set_lock held.
1335 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1336 struct cgroup_root *root)
1339 * No need to lock the task - since we hold cgroup_mutex the
1340 * task can't change groups, so the only thing that can happen
1341 * is that it exits and its css is set back to init_css_set.
1343 return cset_cgroup_from_root(task_css_set(task), root);
1347 * A task must hold cgroup_mutex to modify cgroups.
1349 * Any task can increment and decrement the count field without lock.
1350 * So in general, code holding cgroup_mutex can't rely on the count
1351 * field not changing. However, if the count goes to zero, then only
1352 * cgroup_attach_task() can increment it again. Because a count of zero
1353 * means that no tasks are currently attached, therefore there is no
1354 * way a task attached to that cgroup can fork (the other way to
1355 * increment the count). So code holding cgroup_mutex can safely
1356 * assume that if the count is zero, it will stay zero. Similarly, if
1357 * a task holds cgroup_mutex on a cgroup with zero count, it
1358 * knows that the cgroup won't be removed, as cgroup_rmdir()
1361 * A cgroup can only be deleted if both its 'count' of using tasks
1362 * is zero, and its list of 'children' cgroups is empty. Since all
1363 * tasks in the system use _some_ cgroup, and since there is always at
1364 * least one task in the system (init, pid == 1), therefore, root cgroup
1365 * always has either children cgroups and/or using tasks. So we don't
1366 * need a special hack to ensure that root cgroup cannot be deleted.
1368 * P.S. One more locking exception. RCU is used to guard the
1369 * update of a tasks cgroup pointer by cgroup_attach_task()
1372 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1374 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1377 struct cgroup_subsys *ss = cft->ss;
1379 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1380 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1381 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1382 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1385 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1390 * cgroup_file_mode - deduce file mode of a control file
1391 * @cft: the control file in question
1393 * S_IRUGO for read, S_IWUSR for write.
1395 static umode_t cgroup_file_mode(const struct cftype *cft)
1399 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1402 if (cft->write_u64 || cft->write_s64 || cft->write) {
1403 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1413 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1414 * @subtree_control: the new subtree_control mask to consider
1415 * @this_ss_mask: available subsystems
1417 * On the default hierarchy, a subsystem may request other subsystems to be
1418 * enabled together through its ->depends_on mask. In such cases, more
1419 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1421 * This function calculates which subsystems need to be enabled if
1422 * @subtree_control is to be applied while restricted to @this_ss_mask.
1424 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1426 u16 cur_ss_mask = subtree_control;
1427 struct cgroup_subsys *ss;
1430 lockdep_assert_held(&cgroup_mutex);
1432 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1435 u16 new_ss_mask = cur_ss_mask;
1437 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1438 new_ss_mask |= ss->depends_on;
1439 } while_each_subsys_mask();
1442 * Mask out subsystems which aren't available. This can
1443 * happen only if some depended-upon subsystems were bound
1444 * to non-default hierarchies.
1446 new_ss_mask &= this_ss_mask;
1448 if (new_ss_mask == cur_ss_mask)
1450 cur_ss_mask = new_ss_mask;
1457 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1458 * @kn: the kernfs_node being serviced
1460 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1461 * the method finishes if locking succeeded. Note that once this function
1462 * returns the cgroup returned by cgroup_kn_lock_live() may become
1463 * inaccessible any time. If the caller intends to continue to access the
1464 * cgroup, it should pin it before invoking this function.
1466 void cgroup_kn_unlock(struct kernfs_node *kn)
1468 struct cgroup *cgrp;
1470 if (kernfs_type(kn) == KERNFS_DIR)
1473 cgrp = kn->parent->priv;
1475 mutex_unlock(&cgroup_mutex);
1477 kernfs_unbreak_active_protection(kn);
1482 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1483 * @kn: the kernfs_node being serviced
1484 * @drain_offline: perform offline draining on the cgroup
1486 * This helper is to be used by a cgroup kernfs method currently servicing
1487 * @kn. It breaks the active protection, performs cgroup locking and
1488 * verifies that the associated cgroup is alive. Returns the cgroup if
1489 * alive; otherwise, %NULL. A successful return should be undone by a
1490 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1491 * cgroup is drained of offlining csses before return.
1493 * Any cgroup kernfs method implementation which requires locking the
1494 * associated cgroup should use this helper. It avoids nesting cgroup
1495 * locking under kernfs active protection and allows all kernfs operations
1496 * including self-removal.
1498 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1500 struct cgroup *cgrp;
1502 if (kernfs_type(kn) == KERNFS_DIR)
1505 cgrp = kn->parent->priv;
1508 * We're gonna grab cgroup_mutex which nests outside kernfs
1509 * active_ref. cgroup liveliness check alone provides enough
1510 * protection against removal. Ensure @cgrp stays accessible and
1511 * break the active_ref protection.
1513 if (!cgroup_tryget(cgrp))
1515 kernfs_break_active_protection(kn);
1518 cgroup_lock_and_drain_offline(cgrp);
1520 mutex_lock(&cgroup_mutex);
1522 if (!cgroup_is_dead(cgrp))
1525 cgroup_kn_unlock(kn);
1529 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1531 char name[CGROUP_FILE_NAME_MAX];
1533 lockdep_assert_held(&cgroup_mutex);
1535 if (cft->file_offset) {
1536 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1537 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1539 spin_lock_irq(&cgroup_file_kn_lock);
1541 spin_unlock_irq(&cgroup_file_kn_lock);
1544 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1548 * css_clear_dir - remove subsys files in a cgroup directory
1551 static void css_clear_dir(struct cgroup_subsys_state *css)
1553 struct cgroup *cgrp = css->cgroup;
1554 struct cftype *cfts;
1556 if (!(css->flags & CSS_VISIBLE))
1559 css->flags &= ~CSS_VISIBLE;
1561 list_for_each_entry(cfts, &css->ss->cfts, node)
1562 cgroup_addrm_files(css, cgrp, cfts, false);
1566 * css_populate_dir - create subsys files in a cgroup directory
1569 * On failure, no file is added.
1571 static int css_populate_dir(struct cgroup_subsys_state *css)
1573 struct cgroup *cgrp = css->cgroup;
1574 struct cftype *cfts, *failed_cfts;
1577 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1581 if (cgroup_on_dfl(cgrp))
1582 cfts = cgroup_base_files;
1584 cfts = cgroup1_base_files;
1586 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1589 list_for_each_entry(cfts, &css->ss->cfts, node) {
1590 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1597 css->flags |= CSS_VISIBLE;
1601 list_for_each_entry(cfts, &css->ss->cfts, node) {
1602 if (cfts == failed_cfts)
1604 cgroup_addrm_files(css, cgrp, cfts, false);
1609 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1611 struct cgroup *dcgrp = &dst_root->cgrp;
1612 struct cgroup_subsys *ss;
1615 lockdep_assert_held(&cgroup_mutex);
1617 do_each_subsys_mask(ss, ssid, ss_mask) {
1619 * If @ss has non-root csses attached to it, can't move.
1620 * If @ss is an implicit controller, it is exempt from this
1621 * rule and can be stolen.
1623 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1624 !ss->implicit_on_dfl)
1627 /* can't move between two non-dummy roots either */
1628 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1630 } while_each_subsys_mask();
1632 do_each_subsys_mask(ss, ssid, ss_mask) {
1633 struct cgroup_root *src_root = ss->root;
1634 struct cgroup *scgrp = &src_root->cgrp;
1635 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1636 struct css_set *cset;
1638 WARN_ON(!css || cgroup_css(dcgrp, ss));
1640 /* disable from the source */
1641 src_root->subsys_mask &= ~(1 << ssid);
1642 WARN_ON(cgroup_apply_control(scgrp));
1643 cgroup_finalize_control(scgrp, 0);
1646 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1647 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1648 ss->root = dst_root;
1649 css->cgroup = dcgrp;
1651 spin_lock_irq(&css_set_lock);
1652 hash_for_each(css_set_table, i, cset, hlist)
1653 list_move_tail(&cset->e_cset_node[ss->id],
1654 &dcgrp->e_csets[ss->id]);
1655 spin_unlock_irq(&css_set_lock);
1657 /* default hierarchy doesn't enable controllers by default */
1658 dst_root->subsys_mask |= 1 << ssid;
1659 if (dst_root == &cgrp_dfl_root) {
1660 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1662 dcgrp->subtree_control |= 1 << ssid;
1663 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1666 ret = cgroup_apply_control(dcgrp);
1668 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1673 } while_each_subsys_mask();
1675 kernfs_activate(dcgrp->kn);
1679 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1680 struct kernfs_root *kf_root)
1684 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1685 struct cgroup *ns_cgroup;
1687 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1691 spin_lock_irq(&css_set_lock);
1692 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1693 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1694 spin_unlock_irq(&css_set_lock);
1696 if (len >= PATH_MAX)
1699 seq_escape(sf, buf, " \t\n\\");
1706 static int parse_cgroup_root_flags(char *data, unsigned int *root_flags)
1715 while ((token = strsep(&data, ",")) != NULL) {
1716 if (!strcmp(token, "nsdelegate")) {
1717 *root_flags |= CGRP_ROOT_NS_DELEGATE;
1721 pr_err("cgroup2: unknown option \"%s\"\n", token);
1728 static void apply_cgroup_root_flags(unsigned int root_flags)
1730 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1731 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1732 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1734 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1738 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1740 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1741 seq_puts(seq, ",nsdelegate");
1745 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1747 unsigned int root_flags;
1750 ret = parse_cgroup_root_flags(data, &root_flags);
1754 apply_cgroup_root_flags(root_flags);
1759 * To reduce the fork() overhead for systems that are not actually using
1760 * their cgroups capability, we don't maintain the lists running through
1761 * each css_set to its tasks until we see the list actually used - in other
1762 * words after the first mount.
1764 static bool use_task_css_set_links __read_mostly;
1766 static void cgroup_enable_task_cg_lists(void)
1768 struct task_struct *p, *g;
1770 spin_lock_irq(&css_set_lock);
1772 if (use_task_css_set_links)
1775 use_task_css_set_links = true;
1778 * We need tasklist_lock because RCU is not safe against
1779 * while_each_thread(). Besides, a forking task that has passed
1780 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1781 * is not guaranteed to have its child immediately visible in the
1782 * tasklist if we walk through it with RCU.
1784 read_lock(&tasklist_lock);
1785 do_each_thread(g, p) {
1786 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1787 task_css_set(p) != &init_css_set);
1790 * We should check if the process is exiting, otherwise
1791 * it will race with cgroup_exit() in that the list
1792 * entry won't be deleted though the process has exited.
1793 * Do it while holding siglock so that we don't end up
1794 * racing against cgroup_exit().
1796 * Interrupts were already disabled while acquiring
1797 * the css_set_lock, so we do not need to disable it
1798 * again when acquiring the sighand->siglock here.
1800 spin_lock(&p->sighand->siglock);
1801 if (!(p->flags & PF_EXITING)) {
1802 struct css_set *cset = task_css_set(p);
1804 if (!css_set_populated(cset))
1805 css_set_update_populated(cset, true);
1806 list_add_tail(&p->cg_list, &cset->tasks);
1810 spin_unlock(&p->sighand->siglock);
1811 } while_each_thread(g, p);
1812 read_unlock(&tasklist_lock);
1814 spin_unlock_irq(&css_set_lock);
1817 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1819 struct cgroup_subsys *ss;
1822 INIT_LIST_HEAD(&cgrp->self.sibling);
1823 INIT_LIST_HEAD(&cgrp->self.children);
1824 INIT_LIST_HEAD(&cgrp->cset_links);
1825 INIT_LIST_HEAD(&cgrp->pidlists);
1826 mutex_init(&cgrp->pidlist_mutex);
1827 cgrp->self.cgroup = cgrp;
1828 cgrp->self.flags |= CSS_ONLINE;
1829 cgrp->dom_cgrp = cgrp;
1830 cgrp->max_descendants = INT_MAX;
1831 cgrp->max_depth = INT_MAX;
1833 for_each_subsys(ss, ssid)
1834 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1836 init_waitqueue_head(&cgrp->offline_waitq);
1837 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1840 void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
1842 struct cgroup *cgrp = &root->cgrp;
1844 INIT_LIST_HEAD(&root->root_list);
1845 atomic_set(&root->nr_cgrps, 1);
1847 init_cgroup_housekeeping(cgrp);
1848 idr_init(&root->cgroup_idr);
1850 root->flags = opts->flags;
1851 if (opts->release_agent)
1852 strcpy(root->release_agent_path, opts->release_agent);
1854 strcpy(root->name, opts->name);
1855 if (opts->cpuset_clone_children)
1856 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1859 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
1861 LIST_HEAD(tmp_links);
1862 struct cgroup *root_cgrp = &root->cgrp;
1863 struct kernfs_syscall_ops *kf_sops;
1864 struct css_set *cset;
1867 lockdep_assert_held(&cgroup_mutex);
1869 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1872 root_cgrp->id = ret;
1873 root_cgrp->ancestor_ids[0] = ret;
1875 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1876 ref_flags, GFP_KERNEL);
1881 * We're accessing css_set_count without locking css_set_lock here,
1882 * but that's OK - it can only be increased by someone holding
1883 * cgroup_lock, and that's us. Later rebinding may disable
1884 * controllers on the default hierarchy and thus create new csets,
1885 * which can't be more than the existing ones. Allocate 2x.
1887 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1891 ret = cgroup_init_root_id(root);
1895 kf_sops = root == &cgrp_dfl_root ?
1896 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1898 root->kf_root = kernfs_create_root(kf_sops,
1899 KERNFS_ROOT_CREATE_DEACTIVATED,
1901 if (IS_ERR(root->kf_root)) {
1902 ret = PTR_ERR(root->kf_root);
1905 root_cgrp->kn = root->kf_root->kn;
1907 ret = css_populate_dir(&root_cgrp->self);
1911 ret = rebind_subsystems(root, ss_mask);
1915 trace_cgroup_setup_root(root);
1918 * There must be no failure case after here, since rebinding takes
1919 * care of subsystems' refcounts, which are explicitly dropped in
1920 * the failure exit path.
1922 list_add(&root->root_list, &cgroup_roots);
1923 cgroup_root_count++;
1926 * Link the root cgroup in this hierarchy into all the css_set
1929 spin_lock_irq(&css_set_lock);
1930 hash_for_each(css_set_table, i, cset, hlist) {
1931 link_css_set(&tmp_links, cset, root_cgrp);
1932 if (css_set_populated(cset))
1933 cgroup_update_populated(root_cgrp, true);
1935 spin_unlock_irq(&css_set_lock);
1937 BUG_ON(!list_empty(&root_cgrp->self.children));
1938 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1940 kernfs_activate(root_cgrp->kn);
1945 kernfs_destroy_root(root->kf_root);
1946 root->kf_root = NULL;
1948 cgroup_exit_root_id(root);
1950 percpu_ref_exit(&root_cgrp->self.refcnt);
1952 free_cgrp_cset_links(&tmp_links);
1956 struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
1957 struct cgroup_root *root, unsigned long magic,
1958 struct cgroup_namespace *ns)
1960 struct dentry *dentry;
1963 dentry = kernfs_mount(fs_type, flags, root->kf_root, magic, &new_sb);
1966 * In non-init cgroup namespace, instead of root cgroup's dentry,
1967 * we return the dentry corresponding to the cgroupns->root_cgrp.
1969 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
1970 struct dentry *nsdentry;
1971 struct cgroup *cgrp;
1973 mutex_lock(&cgroup_mutex);
1974 spin_lock_irq(&css_set_lock);
1976 cgrp = cset_cgroup_from_root(ns->root_cset, root);
1978 spin_unlock_irq(&css_set_lock);
1979 mutex_unlock(&cgroup_mutex);
1981 nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
1986 if (IS_ERR(dentry) || !new_sb)
1987 cgroup_put(&root->cgrp);
1992 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1993 int flags, const char *unused_dev_name,
1996 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
1997 struct dentry *dentry;
2002 /* Check if the caller has permission to mount. */
2003 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2005 return ERR_PTR(-EPERM);
2009 * The first time anyone tries to mount a cgroup, enable the list
2010 * linking each css_set to its tasks and fix up all existing tasks.
2012 if (!use_task_css_set_links)
2013 cgroup_enable_task_cg_lists();
2015 if (fs_type == &cgroup2_fs_type) {
2016 unsigned int root_flags;
2018 ret = parse_cgroup_root_flags(data, &root_flags);
2021 return ERR_PTR(ret);
2024 cgrp_dfl_visible = true;
2025 cgroup_get_live(&cgrp_dfl_root.cgrp);
2027 dentry = cgroup_do_mount(&cgroup2_fs_type, flags, &cgrp_dfl_root,
2028 CGROUP2_SUPER_MAGIC, ns);
2029 if (!IS_ERR(dentry))
2030 apply_cgroup_root_flags(root_flags);
2032 dentry = cgroup1_mount(&cgroup_fs_type, flags, data,
2033 CGROUP_SUPER_MAGIC, ns);
2040 static void cgroup_kill_sb(struct super_block *sb)
2042 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2043 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2046 * If @root doesn't have any mounts or children, start killing it.
2047 * This prevents new mounts by disabling percpu_ref_tryget_live().
2048 * cgroup_mount() may wait for @root's release.
2050 * And don't kill the default root.
2052 if (!list_empty(&root->cgrp.self.children) ||
2053 root == &cgrp_dfl_root)
2054 cgroup_put(&root->cgrp);
2056 percpu_ref_kill(&root->cgrp.self.refcnt);
2061 struct file_system_type cgroup_fs_type = {
2063 .mount = cgroup_mount,
2064 .kill_sb = cgroup_kill_sb,
2065 .fs_flags = FS_USERNS_MOUNT,
2068 static struct file_system_type cgroup2_fs_type = {
2070 .mount = cgroup_mount,
2071 .kill_sb = cgroup_kill_sb,
2072 .fs_flags = FS_USERNS_MOUNT,
2075 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2076 struct cgroup_namespace *ns)
2078 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2080 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2083 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2084 struct cgroup_namespace *ns)
2088 mutex_lock(&cgroup_mutex);
2089 spin_lock_irq(&css_set_lock);
2091 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2093 spin_unlock_irq(&css_set_lock);
2094 mutex_unlock(&cgroup_mutex);
2098 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2101 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2102 * @task: target task
2103 * @buf: the buffer to write the path into
2104 * @buflen: the length of the buffer
2106 * Determine @task's cgroup on the first (the one with the lowest non-zero
2107 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2108 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2109 * cgroup controller callbacks.
2111 * Return value is the same as kernfs_path().
2113 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2115 struct cgroup_root *root;
2116 struct cgroup *cgrp;
2117 int hierarchy_id = 1;
2120 mutex_lock(&cgroup_mutex);
2121 spin_lock_irq(&css_set_lock);
2123 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2126 cgrp = task_cgroup_from_root(task, root);
2127 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2129 /* if no hierarchy exists, everyone is in "/" */
2130 ret = strlcpy(buf, "/", buflen);
2133 spin_unlock_irq(&css_set_lock);
2134 mutex_unlock(&cgroup_mutex);
2137 EXPORT_SYMBOL_GPL(task_cgroup_path);
2140 * cgroup_migrate_add_task - add a migration target task to a migration context
2141 * @task: target task
2142 * @mgctx: target migration context
2144 * Add @task, which is a migration target, to @mgctx->tset. This function
2145 * becomes noop if @task doesn't need to be migrated. @task's css_set
2146 * should have been added as a migration source and @task->cg_list will be
2147 * moved from the css_set's tasks list to mg_tasks one.
2149 static void cgroup_migrate_add_task(struct task_struct *task,
2150 struct cgroup_mgctx *mgctx)
2152 struct css_set *cset;
2154 lockdep_assert_held(&css_set_lock);
2156 /* @task either already exited or can't exit until the end */
2157 if (task->flags & PF_EXITING)
2160 /* leave @task alone if post_fork() hasn't linked it yet */
2161 if (list_empty(&task->cg_list))
2164 cset = task_css_set(task);
2165 if (!cset->mg_src_cgrp)
2168 list_move_tail(&task->cg_list, &cset->mg_tasks);
2169 if (list_empty(&cset->mg_node))
2170 list_add_tail(&cset->mg_node,
2171 &mgctx->tset.src_csets);
2172 if (list_empty(&cset->mg_dst_cset->mg_node))
2173 list_add_tail(&cset->mg_dst_cset->mg_node,
2174 &mgctx->tset.dst_csets);
2178 * cgroup_taskset_first - reset taskset and return the first task
2179 * @tset: taskset of interest
2180 * @dst_cssp: output variable for the destination css
2182 * @tset iteration is initialized and the first task is returned.
2184 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2185 struct cgroup_subsys_state **dst_cssp)
2187 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2188 tset->cur_task = NULL;
2190 return cgroup_taskset_next(tset, dst_cssp);
2194 * cgroup_taskset_next - iterate to the next task in taskset
2195 * @tset: taskset of interest
2196 * @dst_cssp: output variable for the destination css
2198 * Return the next task in @tset. Iteration must have been initialized
2199 * with cgroup_taskset_first().
2201 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2202 struct cgroup_subsys_state **dst_cssp)
2204 struct css_set *cset = tset->cur_cset;
2205 struct task_struct *task = tset->cur_task;
2207 while (&cset->mg_node != tset->csets) {
2209 task = list_first_entry(&cset->mg_tasks,
2210 struct task_struct, cg_list);
2212 task = list_next_entry(task, cg_list);
2214 if (&task->cg_list != &cset->mg_tasks) {
2215 tset->cur_cset = cset;
2216 tset->cur_task = task;
2219 * This function may be called both before and
2220 * after cgroup_taskset_migrate(). The two cases
2221 * can be distinguished by looking at whether @cset
2222 * has its ->mg_dst_cset set.
2224 if (cset->mg_dst_cset)
2225 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2227 *dst_cssp = cset->subsys[tset->ssid];
2232 cset = list_next_entry(cset, mg_node);
2240 * cgroup_taskset_migrate - migrate a taskset
2241 * @mgctx: migration context
2243 * Migrate tasks in @mgctx as setup by migration preparation functions.
2244 * This function fails iff one of the ->can_attach callbacks fails and
2245 * guarantees that either all or none of the tasks in @mgctx are migrated.
2246 * @mgctx is consumed regardless of success.
2248 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2250 struct cgroup_taskset *tset = &mgctx->tset;
2251 struct cgroup_subsys *ss;
2252 struct task_struct *task, *tmp_task;
2253 struct css_set *cset, *tmp_cset;
2254 int ssid, failed_ssid, ret;
2256 /* methods shouldn't be called if no task is actually migrating */
2257 if (list_empty(&tset->src_csets))
2260 /* check that we can legitimately attach to the cgroup */
2261 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2262 if (ss->can_attach) {
2264 ret = ss->can_attach(tset);
2267 goto out_cancel_attach;
2270 } while_each_subsys_mask();
2273 * Now that we're guaranteed success, proceed to move all tasks to
2274 * the new cgroup. There are no failure cases after here, so this
2275 * is the commit point.
2277 spin_lock_irq(&css_set_lock);
2278 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2279 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2280 struct css_set *from_cset = task_css_set(task);
2281 struct css_set *to_cset = cset->mg_dst_cset;
2283 get_css_set(to_cset);
2284 to_cset->nr_tasks++;
2285 css_set_move_task(task, from_cset, to_cset, true);
2286 put_css_set_locked(from_cset);
2287 from_cset->nr_tasks--;
2290 spin_unlock_irq(&css_set_lock);
2293 * Migration is committed, all target tasks are now on dst_csets.
2294 * Nothing is sensitive to fork() after this point. Notify
2295 * controllers that migration is complete.
2297 tset->csets = &tset->dst_csets;
2299 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2304 } while_each_subsys_mask();
2307 goto out_release_tset;
2310 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2311 if (ssid == failed_ssid)
2313 if (ss->cancel_attach) {
2315 ss->cancel_attach(tset);
2317 } while_each_subsys_mask();
2319 spin_lock_irq(&css_set_lock);
2320 list_splice_init(&tset->dst_csets, &tset->src_csets);
2321 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2322 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2323 list_del_init(&cset->mg_node);
2325 spin_unlock_irq(&css_set_lock);
2330 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2331 * @dst_cgrp: destination cgroup to test
2333 * On the default hierarchy, except for the mixable, (possible) thread root
2334 * and threaded cgroups, subtree_control must be zero for migration
2335 * destination cgroups with tasks so that child cgroups don't compete
2338 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2340 /* v1 doesn't have any restriction */
2341 if (!cgroup_on_dfl(dst_cgrp))
2344 /* verify @dst_cgrp can host resources */
2345 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2348 /* mixables don't care */
2349 if (cgroup_is_mixable(dst_cgrp))
2353 * If @dst_cgrp is already or can become a thread root or is
2354 * threaded, it doesn't matter.
2356 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2359 /* apply no-internal-process constraint */
2360 if (dst_cgrp->subtree_control)
2367 * cgroup_migrate_finish - cleanup after attach
2368 * @mgctx: migration context
2370 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2371 * those functions for details.
2373 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2375 LIST_HEAD(preloaded);
2376 struct css_set *cset, *tmp_cset;
2378 lockdep_assert_held(&cgroup_mutex);
2380 spin_lock_irq(&css_set_lock);
2382 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2383 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2385 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2386 cset->mg_src_cgrp = NULL;
2387 cset->mg_dst_cgrp = NULL;
2388 cset->mg_dst_cset = NULL;
2389 list_del_init(&cset->mg_preload_node);
2390 put_css_set_locked(cset);
2393 spin_unlock_irq(&css_set_lock);
2397 * cgroup_migrate_add_src - add a migration source css_set
2398 * @src_cset: the source css_set to add
2399 * @dst_cgrp: the destination cgroup
2400 * @mgctx: migration context
2402 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2403 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2404 * up by cgroup_migrate_finish().
2406 * This function may be called without holding cgroup_threadgroup_rwsem
2407 * even if the target is a process. Threads may be created and destroyed
2408 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2409 * into play and the preloaded css_sets are guaranteed to cover all
2412 void cgroup_migrate_add_src(struct css_set *src_cset,
2413 struct cgroup *dst_cgrp,
2414 struct cgroup_mgctx *mgctx)
2416 struct cgroup *src_cgrp;
2418 lockdep_assert_held(&cgroup_mutex);
2419 lockdep_assert_held(&css_set_lock);
2422 * If ->dead, @src_set is associated with one or more dead cgroups
2423 * and doesn't contain any migratable tasks. Ignore it early so
2424 * that the rest of migration path doesn't get confused by it.
2429 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2431 if (!list_empty(&src_cset->mg_preload_node))
2434 WARN_ON(src_cset->mg_src_cgrp);
2435 WARN_ON(src_cset->mg_dst_cgrp);
2436 WARN_ON(!list_empty(&src_cset->mg_tasks));
2437 WARN_ON(!list_empty(&src_cset->mg_node));
2439 src_cset->mg_src_cgrp = src_cgrp;
2440 src_cset->mg_dst_cgrp = dst_cgrp;
2441 get_css_set(src_cset);
2442 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2446 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2447 * @mgctx: migration context
2449 * Tasks are about to be moved and all the source css_sets have been
2450 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2451 * pins all destination css_sets, links each to its source, and append them
2452 * to @mgctx->preloaded_dst_csets.
2454 * This function must be called after cgroup_migrate_add_src() has been
2455 * called on each migration source css_set. After migration is performed
2456 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2459 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2461 struct css_set *src_cset, *tmp_cset;
2463 lockdep_assert_held(&cgroup_mutex);
2465 /* look up the dst cset for each src cset and link it to src */
2466 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2468 struct css_set *dst_cset;
2469 struct cgroup_subsys *ss;
2472 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2476 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2479 * If src cset equals dst, it's noop. Drop the src.
2480 * cgroup_migrate() will skip the cset too. Note that we
2481 * can't handle src == dst as some nodes are used by both.
2483 if (src_cset == dst_cset) {
2484 src_cset->mg_src_cgrp = NULL;
2485 src_cset->mg_dst_cgrp = NULL;
2486 list_del_init(&src_cset->mg_preload_node);
2487 put_css_set(src_cset);
2488 put_css_set(dst_cset);
2492 src_cset->mg_dst_cset = dst_cset;
2494 if (list_empty(&dst_cset->mg_preload_node))
2495 list_add_tail(&dst_cset->mg_preload_node,
2496 &mgctx->preloaded_dst_csets);
2498 put_css_set(dst_cset);
2500 for_each_subsys(ss, ssid)
2501 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2502 mgctx->ss_mask |= 1 << ssid;
2507 cgroup_migrate_finish(mgctx);
2512 * cgroup_migrate - migrate a process or task to a cgroup
2513 * @leader: the leader of the process or the task to migrate
2514 * @threadgroup: whether @leader points to the whole process or a single task
2515 * @mgctx: migration context
2517 * Migrate a process or task denoted by @leader. If migrating a process,
2518 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2519 * responsible for invoking cgroup_migrate_add_src() and
2520 * cgroup_migrate_prepare_dst() on the targets before invoking this
2521 * function and following up with cgroup_migrate_finish().
2523 * As long as a controller's ->can_attach() doesn't fail, this function is
2524 * guaranteed to succeed. This means that, excluding ->can_attach()
2525 * failure, when migrating multiple targets, the success or failure can be
2526 * decided for all targets by invoking group_migrate_prepare_dst() before
2527 * actually starting migrating.
2529 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2530 struct cgroup_mgctx *mgctx)
2532 struct task_struct *task;
2535 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2536 * already PF_EXITING could be freed from underneath us unless we
2537 * take an rcu_read_lock.
2539 spin_lock_irq(&css_set_lock);
2543 cgroup_migrate_add_task(task, mgctx);
2546 } while_each_thread(leader, task);
2548 spin_unlock_irq(&css_set_lock);
2550 return cgroup_migrate_execute(mgctx);
2554 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2555 * @dst_cgrp: the cgroup to attach to
2556 * @leader: the task or the leader of the threadgroup to be attached
2557 * @threadgroup: attach the whole threadgroup?
2559 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2561 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2564 DEFINE_CGROUP_MGCTX(mgctx);
2565 struct task_struct *task;
2568 ret = cgroup_migrate_vet_dst(dst_cgrp);
2572 /* look up all src csets */
2573 spin_lock_irq(&css_set_lock);
2577 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2580 } while_each_thread(leader, task);
2582 spin_unlock_irq(&css_set_lock);
2584 /* prepare dst csets and commit */
2585 ret = cgroup_migrate_prepare_dst(&mgctx);
2587 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2589 cgroup_migrate_finish(&mgctx);
2592 trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2597 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2598 __acquires(&cgroup_threadgroup_rwsem)
2600 struct task_struct *tsk;
2603 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2604 return ERR_PTR(-EINVAL);
2606 percpu_down_write(&cgroup_threadgroup_rwsem);
2610 tsk = find_task_by_vpid(pid);
2612 tsk = ERR_PTR(-ESRCH);
2613 goto out_unlock_threadgroup;
2620 tsk = tsk->group_leader;
2623 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2624 * If userland migrates such a kthread to a non-root cgroup, it can
2625 * become trapped in a cpuset, or RT kthread may be born in a
2626 * cgroup with no rt_runtime allocated. Just say no.
2628 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2629 tsk = ERR_PTR(-EINVAL);
2630 goto out_unlock_threadgroup;
2633 get_task_struct(tsk);
2634 goto out_unlock_rcu;
2636 out_unlock_threadgroup:
2637 percpu_up_write(&cgroup_threadgroup_rwsem);
2643 void cgroup_procs_write_finish(struct task_struct *task)
2644 __releases(&cgroup_threadgroup_rwsem)
2646 struct cgroup_subsys *ss;
2649 /* release reference from cgroup_procs_write_start() */
2650 put_task_struct(task);
2652 percpu_up_write(&cgroup_threadgroup_rwsem);
2653 for_each_subsys(ss, ssid)
2654 if (ss->post_attach)
2658 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2660 struct cgroup_subsys *ss;
2661 bool printed = false;
2664 do_each_subsys_mask(ss, ssid, ss_mask) {
2667 seq_printf(seq, "%s", ss->name);
2669 } while_each_subsys_mask();
2671 seq_putc(seq, '\n');
2674 /* show controllers which are enabled from the parent */
2675 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2677 struct cgroup *cgrp = seq_css(seq)->cgroup;
2679 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2683 /* show controllers which are enabled for a given cgroup's children */
2684 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2686 struct cgroup *cgrp = seq_css(seq)->cgroup;
2688 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2693 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2694 * @cgrp: root of the subtree to update csses for
2696 * @cgrp's control masks have changed and its subtree's css associations
2697 * need to be updated accordingly. This function looks up all css_sets
2698 * which are attached to the subtree, creates the matching updated css_sets
2699 * and migrates the tasks to the new ones.
2701 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2703 DEFINE_CGROUP_MGCTX(mgctx);
2704 struct cgroup_subsys_state *d_css;
2705 struct cgroup *dsct;
2706 struct css_set *src_cset;
2709 lockdep_assert_held(&cgroup_mutex);
2711 percpu_down_write(&cgroup_threadgroup_rwsem);
2713 /* look up all csses currently attached to @cgrp's subtree */
2714 spin_lock_irq(&css_set_lock);
2715 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2716 struct cgrp_cset_link *link;
2718 list_for_each_entry(link, &dsct->cset_links, cset_link)
2719 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2721 spin_unlock_irq(&css_set_lock);
2723 /* NULL dst indicates self on default hierarchy */
2724 ret = cgroup_migrate_prepare_dst(&mgctx);
2728 spin_lock_irq(&css_set_lock);
2729 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2730 struct task_struct *task, *ntask;
2732 /* all tasks in src_csets need to be migrated */
2733 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2734 cgroup_migrate_add_task(task, &mgctx);
2736 spin_unlock_irq(&css_set_lock);
2738 ret = cgroup_migrate_execute(&mgctx);
2740 cgroup_migrate_finish(&mgctx);
2741 percpu_up_write(&cgroup_threadgroup_rwsem);
2746 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2747 * @cgrp: root of the target subtree
2749 * Because css offlining is asynchronous, userland may try to re-enable a
2750 * controller while the previous css is still around. This function grabs
2751 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2753 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2754 __acquires(&cgroup_mutex)
2756 struct cgroup *dsct;
2757 struct cgroup_subsys_state *d_css;
2758 struct cgroup_subsys *ss;
2762 mutex_lock(&cgroup_mutex);
2764 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2765 for_each_subsys(ss, ssid) {
2766 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2769 if (!css || !percpu_ref_is_dying(&css->refcnt))
2772 cgroup_get_live(dsct);
2773 prepare_to_wait(&dsct->offline_waitq, &wait,
2774 TASK_UNINTERRUPTIBLE);
2776 mutex_unlock(&cgroup_mutex);
2778 finish_wait(&dsct->offline_waitq, &wait);
2787 * cgroup_save_control - save control masks of a subtree
2788 * @cgrp: root of the target subtree
2790 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
2791 * prefixed fields for @cgrp's subtree including @cgrp itself.
2793 static void cgroup_save_control(struct cgroup *cgrp)
2795 struct cgroup *dsct;
2796 struct cgroup_subsys_state *d_css;
2798 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2799 dsct->old_subtree_control = dsct->subtree_control;
2800 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2805 * cgroup_propagate_control - refresh control masks of a subtree
2806 * @cgrp: root of the target subtree
2808 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2809 * ->subtree_control and propagate controller availability through the
2810 * subtree so that descendants don't have unavailable controllers enabled.
2812 static void cgroup_propagate_control(struct cgroup *cgrp)
2814 struct cgroup *dsct;
2815 struct cgroup_subsys_state *d_css;
2817 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2818 dsct->subtree_control &= cgroup_control(dsct);
2819 dsct->subtree_ss_mask =
2820 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2821 cgroup_ss_mask(dsct));
2826 * cgroup_restore_control - restore control masks of a subtree
2827 * @cgrp: root of the target subtree
2829 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
2830 * prefixed fields for @cgrp's subtree including @cgrp itself.
2832 static void cgroup_restore_control(struct cgroup *cgrp)
2834 struct cgroup *dsct;
2835 struct cgroup_subsys_state *d_css;
2837 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2838 dsct->subtree_control = dsct->old_subtree_control;
2839 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
2843 static bool css_visible(struct cgroup_subsys_state *css)
2845 struct cgroup_subsys *ss = css->ss;
2846 struct cgroup *cgrp = css->cgroup;
2848 if (cgroup_control(cgrp) & (1 << ss->id))
2850 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
2852 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
2856 * cgroup_apply_control_enable - enable or show csses according to control
2857 * @cgrp: root of the target subtree
2859 * Walk @cgrp's subtree and create new csses or make the existing ones
2860 * visible. A css is created invisible if it's being implicitly enabled
2861 * through dependency. An invisible css is made visible when the userland
2862 * explicitly enables it.
2864 * Returns 0 on success, -errno on failure. On failure, csses which have
2865 * been processed already aren't cleaned up. The caller is responsible for
2866 * cleaning up with cgroup_apply_control_disable().
2868 static int cgroup_apply_control_enable(struct cgroup *cgrp)
2870 struct cgroup *dsct;
2871 struct cgroup_subsys_state *d_css;
2872 struct cgroup_subsys *ss;
2875 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2876 for_each_subsys(ss, ssid) {
2877 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2879 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2881 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
2885 css = css_create(dsct, ss);
2887 return PTR_ERR(css);
2890 if (css_visible(css)) {
2891 ret = css_populate_dir(css);
2902 * cgroup_apply_control_disable - kill or hide csses according to control
2903 * @cgrp: root of the target subtree
2905 * Walk @cgrp's subtree and kill and hide csses so that they match
2906 * cgroup_ss_mask() and cgroup_visible_mask().
2908 * A css is hidden when the userland requests it to be disabled while other
2909 * subsystems are still depending on it. The css must not actively control
2910 * resources and be in the vanilla state if it's made visible again later.
2911 * Controllers which may be depended upon should provide ->css_reset() for
2914 static void cgroup_apply_control_disable(struct cgroup *cgrp)
2916 struct cgroup *dsct;
2917 struct cgroup_subsys_state *d_css;
2918 struct cgroup_subsys *ss;
2921 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2922 for_each_subsys(ss, ssid) {
2923 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2925 WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2931 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
2933 } else if (!css_visible(css)) {
2943 * cgroup_apply_control - apply control mask updates to the subtree
2944 * @cgrp: root of the target subtree
2946 * subsystems can be enabled and disabled in a subtree using the following
2949 * 1. Call cgroup_save_control() to stash the current state.
2950 * 2. Update ->subtree_control masks in the subtree as desired.
2951 * 3. Call cgroup_apply_control() to apply the changes.
2952 * 4. Optionally perform other related operations.
2953 * 5. Call cgroup_finalize_control() to finish up.
2955 * This function implements step 3 and propagates the mask changes
2956 * throughout @cgrp's subtree, updates csses accordingly and perform
2957 * process migrations.
2959 static int cgroup_apply_control(struct cgroup *cgrp)
2963 cgroup_propagate_control(cgrp);
2965 ret = cgroup_apply_control_enable(cgrp);
2970 * At this point, cgroup_e_css() results reflect the new csses
2971 * making the following cgroup_update_dfl_csses() properly update
2972 * css associations of all tasks in the subtree.
2974 ret = cgroup_update_dfl_csses(cgrp);
2982 * cgroup_finalize_control - finalize control mask update
2983 * @cgrp: root of the target subtree
2984 * @ret: the result of the update
2986 * Finalize control mask update. See cgroup_apply_control() for more info.
2988 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
2991 cgroup_restore_control(cgrp);
2992 cgroup_propagate_control(cgrp);
2995 cgroup_apply_control_disable(cgrp);
2998 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3000 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3002 /* if nothing is getting enabled, nothing to worry about */
3006 /* can @cgrp host any resources? */
3007 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3010 /* mixables don't care */
3011 if (cgroup_is_mixable(cgrp))
3014 if (domain_enable) {
3015 /* can't enable domain controllers inside a thread subtree */
3016 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3020 * Threaded controllers can handle internal competitions
3021 * and are always allowed inside a (prospective) thread
3024 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3029 * Controllers can't be enabled for a cgroup with tasks to avoid
3030 * child cgroups competing against tasks.
3032 if (cgroup_has_tasks(cgrp))
3038 /* change the enabled child controllers for a cgroup in the default hierarchy */
3039 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3040 char *buf, size_t nbytes,
3043 u16 enable = 0, disable = 0;
3044 struct cgroup *cgrp, *child;
3045 struct cgroup_subsys *ss;
3050 * Parse input - space separated list of subsystem names prefixed
3051 * with either + or -.
3053 buf = strstrip(buf);
3054 while ((tok = strsep(&buf, " "))) {
3057 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3058 if (!cgroup_ssid_enabled(ssid) ||
3059 strcmp(tok + 1, ss->name))
3063 enable |= 1 << ssid;
3064 disable &= ~(1 << ssid);
3065 } else if (*tok == '-') {
3066 disable |= 1 << ssid;
3067 enable &= ~(1 << ssid);
3072 } while_each_subsys_mask();
3073 if (ssid == CGROUP_SUBSYS_COUNT)
3077 cgrp = cgroup_kn_lock_live(of->kn, true);
3081 for_each_subsys(ss, ssid) {
3082 if (enable & (1 << ssid)) {
3083 if (cgrp->subtree_control & (1 << ssid)) {
3084 enable &= ~(1 << ssid);
3088 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3092 } else if (disable & (1 << ssid)) {
3093 if (!(cgrp->subtree_control & (1 << ssid))) {
3094 disable &= ~(1 << ssid);
3098 /* a child has it enabled? */
3099 cgroup_for_each_live_child(child, cgrp) {
3100 if (child->subtree_control & (1 << ssid)) {
3108 if (!enable && !disable) {
3113 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3117 /* save and update control masks and prepare csses */
3118 cgroup_save_control(cgrp);
3120 cgrp->subtree_control |= enable;
3121 cgrp->subtree_control &= ~disable;
3123 ret = cgroup_apply_control(cgrp);
3125 cgroup_finalize_control(cgrp, ret);
3127 kernfs_activate(cgrp->kn);
3130 cgroup_kn_unlock(of->kn);
3131 return ret ?: nbytes;
3135 * cgroup_enable_threaded - make @cgrp threaded
3136 * @cgrp: the target cgroup
3138 * Called when "threaded" is written to the cgroup.type interface file and
3139 * tries to make @cgrp threaded and join the parent's resource domain.
3140 * This function is never called on the root cgroup as cgroup.type doesn't
3143 static int cgroup_enable_threaded(struct cgroup *cgrp)
3145 struct cgroup *parent = cgroup_parent(cgrp);
3146 struct cgroup *dom_cgrp = parent->dom_cgrp;
3149 lockdep_assert_held(&cgroup_mutex);
3151 /* noop if already threaded */
3152 if (cgroup_is_threaded(cgrp))
3155 /* we're joining the parent's domain, ensure its validity */
3156 if (!cgroup_is_valid_domain(dom_cgrp) ||
3157 !cgroup_can_be_thread_root(dom_cgrp))
3161 * The following shouldn't cause actual migrations and should
3164 cgroup_save_control(cgrp);
3166 cgrp->dom_cgrp = dom_cgrp;
3167 ret = cgroup_apply_control(cgrp);
3169 parent->nr_threaded_children++;
3171 cgrp->dom_cgrp = cgrp;
3173 cgroup_finalize_control(cgrp, ret);
3177 static int cgroup_type_show(struct seq_file *seq, void *v)
3179 struct cgroup *cgrp = seq_css(seq)->cgroup;
3181 if (cgroup_is_threaded(cgrp))
3182 seq_puts(seq, "threaded\n");
3183 else if (!cgroup_is_valid_domain(cgrp))
3184 seq_puts(seq, "domain invalid\n");
3185 else if (cgroup_is_thread_root(cgrp))
3186 seq_puts(seq, "domain threaded\n");
3188 seq_puts(seq, "domain\n");
3193 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3194 size_t nbytes, loff_t off)
3196 struct cgroup *cgrp;
3199 /* only switching to threaded mode is supported */
3200 if (strcmp(strstrip(buf), "threaded"))
3203 cgrp = cgroup_kn_lock_live(of->kn, false);
3207 /* threaded can only be enabled */
3208 ret = cgroup_enable_threaded(cgrp);
3210 cgroup_kn_unlock(of->kn);
3211 return ret ?: nbytes;
3214 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3216 struct cgroup *cgrp = seq_css(seq)->cgroup;
3217 int descendants = READ_ONCE(cgrp->max_descendants);
3219 if (descendants == INT_MAX)
3220 seq_puts(seq, "max\n");
3222 seq_printf(seq, "%d\n", descendants);
3227 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3228 char *buf, size_t nbytes, loff_t off)
3230 struct cgroup *cgrp;
3234 buf = strstrip(buf);
3235 if (!strcmp(buf, "max")) {
3236 descendants = INT_MAX;
3238 ret = kstrtoint(buf, 0, &descendants);
3243 if (descendants < 0 || descendants > INT_MAX)
3246 cgrp = cgroup_kn_lock_live(of->kn, false);
3250 cgrp->max_descendants = descendants;
3252 cgroup_kn_unlock(of->kn);
3257 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3259 struct cgroup *cgrp = seq_css(seq)->cgroup;
3260 int depth = READ_ONCE(cgrp->max_depth);
3262 if (depth == INT_MAX)
3263 seq_puts(seq, "max\n");
3265 seq_printf(seq, "%d\n", depth);
3270 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3271 char *buf, size_t nbytes, loff_t off)
3273 struct cgroup *cgrp;
3277 buf = strstrip(buf);
3278 if (!strcmp(buf, "max")) {
3281 ret = kstrtoint(buf, 0, &depth);
3286 if (depth < 0 || depth > INT_MAX)
3289 cgrp = cgroup_kn_lock_live(of->kn, false);
3293 cgrp->max_depth = depth;
3295 cgroup_kn_unlock(of->kn);
3300 static int cgroup_events_show(struct seq_file *seq, void *v)
3302 seq_printf(seq, "populated %d\n",
3303 cgroup_is_populated(seq_css(seq)->cgroup));
3307 static int cgroup_stats_show(struct seq_file *seq, void *v)
3309 struct cgroup *cgroup = seq_css(seq)->cgroup;
3311 seq_printf(seq, "nr_descendants %d\n",
3312 cgroup->nr_descendants);
3313 seq_printf(seq, "nr_dying_descendants %d\n",
3314 cgroup->nr_dying_descendants);
3319 static int cgroup_file_open(struct kernfs_open_file *of)
3321 struct cftype *cft = of->kn->priv;
3324 return cft->open(of);
3328 static void cgroup_file_release(struct kernfs_open_file *of)
3330 struct cftype *cft = of->kn->priv;
3336 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3337 size_t nbytes, loff_t off)
3339 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3340 struct cgroup *cgrp = of->kn->parent->priv;
3341 struct cftype *cft = of->kn->priv;
3342 struct cgroup_subsys_state *css;
3346 * If namespaces are delegation boundaries, disallow writes to
3347 * files in an non-init namespace root from inside the namespace
3348 * except for the files explicitly marked delegatable -
3349 * cgroup.procs and cgroup.subtree_control.
3351 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3352 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3353 ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3357 return cft->write(of, buf, nbytes, off);
3360 * kernfs guarantees that a file isn't deleted with operations in
3361 * flight, which means that the matching css is and stays alive and
3362 * doesn't need to be pinned. The RCU locking is not necessary
3363 * either. It's just for the convenience of using cgroup_css().
3366 css = cgroup_css(cgrp, cft->ss);
3369 if (cft->write_u64) {
3370 unsigned long long v;
3371 ret = kstrtoull(buf, 0, &v);
3373 ret = cft->write_u64(css, cft, v);
3374 } else if (cft->write_s64) {
3376 ret = kstrtoll(buf, 0, &v);
3378 ret = cft->write_s64(css, cft, v);
3383 return ret ?: nbytes;
3386 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3388 return seq_cft(seq)->seq_start(seq, ppos);
3391 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3393 return seq_cft(seq)->seq_next(seq, v, ppos);
3396 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3398 if (seq_cft(seq)->seq_stop)
3399 seq_cft(seq)->seq_stop(seq, v);
3402 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3404 struct cftype *cft = seq_cft(m);
3405 struct cgroup_subsys_state *css = seq_css(m);
3408 return cft->seq_show(m, arg);
3411 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3412 else if (cft->read_s64)
3413 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3419 static struct kernfs_ops cgroup_kf_single_ops = {
3420 .atomic_write_len = PAGE_SIZE,
3421 .open = cgroup_file_open,
3422 .release = cgroup_file_release,
3423 .write = cgroup_file_write,
3424 .seq_show = cgroup_seqfile_show,
3427 static struct kernfs_ops cgroup_kf_ops = {
3428 .atomic_write_len = PAGE_SIZE,
3429 .open = cgroup_file_open,
3430 .release = cgroup_file_release,
3431 .write = cgroup_file_write,
3432 .seq_start = cgroup_seqfile_start,
3433 .seq_next = cgroup_seqfile_next,
3434 .seq_stop = cgroup_seqfile_stop,
3435 .seq_show = cgroup_seqfile_show,
3438 /* set uid and gid of cgroup dirs and files to that of the creator */
3439 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3441 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3442 .ia_uid = current_fsuid(),
3443 .ia_gid = current_fsgid(), };
3445 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3446 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3449 return kernfs_setattr(kn, &iattr);
3452 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3455 char name[CGROUP_FILE_NAME_MAX];
3456 struct kernfs_node *kn;
3457 struct lock_class_key *key = NULL;
3460 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3461 key = &cft->lockdep_key;
3463 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3464 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3469 ret = cgroup_kn_set_ugid(kn);
3475 if (cft->file_offset) {
3476 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3478 spin_lock_irq(&cgroup_file_kn_lock);
3480 spin_unlock_irq(&cgroup_file_kn_lock);
3487 * cgroup_addrm_files - add or remove files to a cgroup directory
3488 * @css: the target css
3489 * @cgrp: the target cgroup (usually css->cgroup)
3490 * @cfts: array of cftypes to be added
3491 * @is_add: whether to add or remove
3493 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3494 * For removals, this function never fails.
3496 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3497 struct cgroup *cgrp, struct cftype cfts[],
3500 struct cftype *cft, *cft_end = NULL;
3503 lockdep_assert_held(&cgroup_mutex);
3506 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3507 /* does cft->flags tell us to skip this file on @cgrp? */
3508 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3510 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3512 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3514 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3518 ret = cgroup_add_file(css, cgrp, cft);
3520 pr_warn("%s: failed to add %s, err=%d\n",
3521 __func__, cft->name, ret);
3527 cgroup_rm_file(cgrp, cft);
3533 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3535 struct cgroup_subsys *ss = cfts[0].ss;
3536 struct cgroup *root = &ss->root->cgrp;
3537 struct cgroup_subsys_state *css;
3540 lockdep_assert_held(&cgroup_mutex);
3542 /* add/rm files for all cgroups created before */
3543 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3544 struct cgroup *cgrp = css->cgroup;
3546 if (!(css->flags & CSS_VISIBLE))
3549 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3555 kernfs_activate(root->kn);
3559 static void cgroup_exit_cftypes(struct cftype *cfts)
3563 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3564 /* free copy for custom atomic_write_len, see init_cftypes() */
3565 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3570 /* revert flags set by cgroup core while adding @cfts */
3571 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3575 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3579 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3580 struct kernfs_ops *kf_ops;
3582 WARN_ON(cft->ss || cft->kf_ops);
3585 kf_ops = &cgroup_kf_ops;
3587 kf_ops = &cgroup_kf_single_ops;
3590 * Ugh... if @cft wants a custom max_write_len, we need to
3591 * make a copy of kf_ops to set its atomic_write_len.
3593 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3594 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3596 cgroup_exit_cftypes(cfts);
3599 kf_ops->atomic_write_len = cft->max_write_len;
3602 cft->kf_ops = kf_ops;
3609 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3611 lockdep_assert_held(&cgroup_mutex);
3613 if (!cfts || !cfts[0].ss)
3616 list_del(&cfts->node);
3617 cgroup_apply_cftypes(cfts, false);
3618 cgroup_exit_cftypes(cfts);
3623 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3624 * @cfts: zero-length name terminated array of cftypes
3626 * Unregister @cfts. Files described by @cfts are removed from all
3627 * existing cgroups and all future cgroups won't have them either. This
3628 * function can be called anytime whether @cfts' subsys is attached or not.
3630 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3633 int cgroup_rm_cftypes(struct cftype *cfts)
3637 mutex_lock(&cgroup_mutex);
3638 ret = cgroup_rm_cftypes_locked(cfts);
3639 mutex_unlock(&cgroup_mutex);
3644 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3645 * @ss: target cgroup subsystem
3646 * @cfts: zero-length name terminated array of cftypes
3648 * Register @cfts to @ss. Files described by @cfts are created for all
3649 * existing cgroups to which @ss is attached and all future cgroups will
3650 * have them too. This function can be called anytime whether @ss is
3653 * Returns 0 on successful registration, -errno on failure. Note that this
3654 * function currently returns 0 as long as @cfts registration is successful
3655 * even if some file creation attempts on existing cgroups fail.
3657 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3661 if (!cgroup_ssid_enabled(ss->id))
3664 if (!cfts || cfts[0].name[0] == '\0')
3667 ret = cgroup_init_cftypes(ss, cfts);
3671 mutex_lock(&cgroup_mutex);
3673 list_add_tail(&cfts->node, &ss->cfts);
3674 ret = cgroup_apply_cftypes(cfts, true);
3676 cgroup_rm_cftypes_locked(cfts);
3678 mutex_unlock(&cgroup_mutex);
3683 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3684 * @ss: target cgroup subsystem
3685 * @cfts: zero-length name terminated array of cftypes
3687 * Similar to cgroup_add_cftypes() but the added files are only used for
3688 * the default hierarchy.
3690 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3694 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3695 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3696 return cgroup_add_cftypes(ss, cfts);
3700 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3701 * @ss: target cgroup subsystem
3702 * @cfts: zero-length name terminated array of cftypes
3704 * Similar to cgroup_add_cftypes() but the added files are only used for
3705 * the legacy hierarchies.
3707 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3711 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3712 cft->flags |= __CFTYPE_NOT_ON_DFL;
3713 return cgroup_add_cftypes(ss, cfts);
3717 * cgroup_file_notify - generate a file modified event for a cgroup_file
3718 * @cfile: target cgroup_file
3720 * @cfile must have been obtained by setting cftype->file_offset.
3722 void cgroup_file_notify(struct cgroup_file *cfile)
3724 unsigned long flags;
3726 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3728 kernfs_notify(cfile->kn);
3729 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3733 * css_next_child - find the next child of a given css
3734 * @pos: the current position (%NULL to initiate traversal)
3735 * @parent: css whose children to walk
3737 * This function returns the next child of @parent and should be called
3738 * under either cgroup_mutex or RCU read lock. The only requirement is
3739 * that @parent and @pos are accessible. The next sibling is guaranteed to
3740 * be returned regardless of their states.
3742 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3743 * css which finished ->css_online() is guaranteed to be visible in the
3744 * future iterations and will stay visible until the last reference is put.
3745 * A css which hasn't finished ->css_online() or already finished
3746 * ->css_offline() may show up during traversal. It's each subsystem's
3747 * responsibility to synchronize against on/offlining.
3749 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3750 struct cgroup_subsys_state *parent)
3752 struct cgroup_subsys_state *next;
3754 cgroup_assert_mutex_or_rcu_locked();
3757 * @pos could already have been unlinked from the sibling list.
3758 * Once a cgroup is removed, its ->sibling.next is no longer
3759 * updated when its next sibling changes. CSS_RELEASED is set when
3760 * @pos is taken off list, at which time its next pointer is valid,
3761 * and, as releases are serialized, the one pointed to by the next
3762 * pointer is guaranteed to not have started release yet. This
3763 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3764 * critical section, the one pointed to by its next pointer is
3765 * guaranteed to not have finished its RCU grace period even if we
3766 * have dropped rcu_read_lock() inbetween iterations.
3768 * If @pos has CSS_RELEASED set, its next pointer can't be
3769 * dereferenced; however, as each css is given a monotonically
3770 * increasing unique serial number and always appended to the
3771 * sibling list, the next one can be found by walking the parent's
3772 * children until the first css with higher serial number than
3773 * @pos's. While this path can be slower, it happens iff iteration
3774 * races against release and the race window is very small.
3777 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3778 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3779 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3781 list_for_each_entry_rcu(next, &parent->children, sibling)
3782 if (next->serial_nr > pos->serial_nr)
3787 * @next, if not pointing to the head, can be dereferenced and is
3790 if (&next->sibling != &parent->children)
3796 * css_next_descendant_pre - find the next descendant for pre-order walk
3797 * @pos: the current position (%NULL to initiate traversal)
3798 * @root: css whose descendants to walk
3800 * To be used by css_for_each_descendant_pre(). Find the next descendant
3801 * to visit for pre-order traversal of @root's descendants. @root is
3802 * included in the iteration and the first node to be visited.
3804 * While this function requires cgroup_mutex or RCU read locking, it
3805 * doesn't require the whole traversal to be contained in a single critical
3806 * section. This function will return the correct next descendant as long
3807 * as both @pos and @root are accessible and @pos is a descendant of @root.
3809 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3810 * css which finished ->css_online() is guaranteed to be visible in the
3811 * future iterations and will stay visible until the last reference is put.
3812 * A css which hasn't finished ->css_online() or already finished
3813 * ->css_offline() may show up during traversal. It's each subsystem's
3814 * responsibility to synchronize against on/offlining.
3816 struct cgroup_subsys_state *
3817 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3818 struct cgroup_subsys_state *root)
3820 struct cgroup_subsys_state *next;
3822 cgroup_assert_mutex_or_rcu_locked();
3824 /* if first iteration, visit @root */
3828 /* visit the first child if exists */
3829 next = css_next_child(NULL, pos);
3833 /* no child, visit my or the closest ancestor's next sibling */
3834 while (pos != root) {
3835 next = css_next_child(pos, pos->parent);
3845 * css_rightmost_descendant - return the rightmost descendant of a css
3846 * @pos: css of interest
3848 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3849 * is returned. This can be used during pre-order traversal to skip
3852 * While this function requires cgroup_mutex or RCU read locking, it
3853 * doesn't require the whole traversal to be contained in a single critical
3854 * section. This function will return the correct rightmost descendant as
3855 * long as @pos is accessible.
3857 struct cgroup_subsys_state *
3858 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3860 struct cgroup_subsys_state *last, *tmp;
3862 cgroup_assert_mutex_or_rcu_locked();
3866 /* ->prev isn't RCU safe, walk ->next till the end */
3868 css_for_each_child(tmp, last)
3875 static struct cgroup_subsys_state *
3876 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3878 struct cgroup_subsys_state *last;
3882 pos = css_next_child(NULL, pos);
3889 * css_next_descendant_post - find the next descendant for post-order walk
3890 * @pos: the current position (%NULL to initiate traversal)
3891 * @root: css whose descendants to walk
3893 * To be used by css_for_each_descendant_post(). Find the next descendant
3894 * to visit for post-order traversal of @root's descendants. @root is
3895 * included in the iteration and the last node to be visited.
3897 * While this function requires cgroup_mutex or RCU read locking, it
3898 * doesn't require the whole traversal to be contained in a single critical
3899 * section. This function will return the correct next descendant as long
3900 * as both @pos and @cgroup are accessible and @pos is a descendant of
3903 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3904 * css which finished ->css_online() is guaranteed to be visible in the
3905 * future iterations and will stay visible until the last reference is put.
3906 * A css which hasn't finished ->css_online() or already finished
3907 * ->css_offline() may show up during traversal. It's each subsystem's
3908 * responsibility to synchronize against on/offlining.
3910 struct cgroup_subsys_state *
3911 css_next_descendant_post(struct cgroup_subsys_state *pos,
3912 struct cgroup_subsys_state *root)
3914 struct cgroup_subsys_state *next;
3916 cgroup_assert_mutex_or_rcu_locked();
3918 /* if first iteration, visit leftmost descendant which may be @root */
3920 return css_leftmost_descendant(root);
3922 /* if we visited @root, we're done */
3926 /* if there's an unvisited sibling, visit its leftmost descendant */
3927 next = css_next_child(pos, pos->parent);
3929 return css_leftmost_descendant(next);
3931 /* no sibling left, visit parent */
3936 * css_has_online_children - does a css have online children
3937 * @css: the target css
3939 * Returns %true if @css has any online children; otherwise, %false. This
3940 * function can be called from any context but the caller is responsible
3941 * for synchronizing against on/offlining as necessary.
3943 bool css_has_online_children(struct cgroup_subsys_state *css)
3945 struct cgroup_subsys_state *child;
3949 css_for_each_child(child, css) {
3950 if (child->flags & CSS_ONLINE) {
3959 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
3961 struct list_head *l;
3962 struct cgrp_cset_link *link;
3963 struct css_set *cset;
3965 lockdep_assert_held(&css_set_lock);
3967 /* find the next threaded cset */
3968 if (it->tcset_pos) {
3969 l = it->tcset_pos->next;
3971 if (l != it->tcset_head) {
3973 return container_of(l, struct css_set,
3974 threaded_csets_node);
3977 it->tcset_pos = NULL;
3980 /* find the next cset */
3983 if (l == it->cset_head) {
3984 it->cset_pos = NULL;
3989 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
3991 link = list_entry(l, struct cgrp_cset_link, cset_link);
3997 /* initialize threaded css_set walking */
3998 if (it->flags & CSS_TASK_ITER_THREADED) {
4000 put_css_set_locked(it->cur_dcset);
4001 it->cur_dcset = cset;
4004 it->tcset_head = &cset->threaded_csets;
4005 it->tcset_pos = &cset->threaded_csets;
4012 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4013 * @it: the iterator to advance
4015 * Advance @it to the next css_set to walk.
4017 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4019 struct css_set *cset;
4021 lockdep_assert_held(&css_set_lock);
4023 /* Advance to the next non-empty css_set */
4025 cset = css_task_iter_next_css_set(it);
4027 it->task_pos = NULL;
4030 } while (!css_set_populated(cset));
4032 if (!list_empty(&cset->tasks))
4033 it->task_pos = cset->tasks.next;
4035 it->task_pos = cset->mg_tasks.next;
4037 it->tasks_head = &cset->tasks;
4038 it->mg_tasks_head = &cset->mg_tasks;
4041 * We don't keep css_sets locked across iteration steps and thus
4042 * need to take steps to ensure that iteration can be resumed after
4043 * the lock is re-acquired. Iteration is performed at two levels -
4044 * css_sets and tasks in them.
4046 * Once created, a css_set never leaves its cgroup lists, so a
4047 * pinned css_set is guaranteed to stay put and we can resume
4048 * iteration afterwards.
4050 * Tasks may leave @cset across iteration steps. This is resolved
4051 * by registering each iterator with the css_set currently being
4052 * walked and making css_set_move_task() advance iterators whose
4053 * next task is leaving.
4056 list_del(&it->iters_node);
4057 put_css_set_locked(it->cur_cset);
4060 it->cur_cset = cset;
4061 list_add(&it->iters_node, &cset->task_iters);
4064 static void css_task_iter_advance(struct css_task_iter *it)
4066 struct list_head *l = it->task_pos;
4068 lockdep_assert_held(&css_set_lock);
4073 * Advance iterator to find next entry. cset->tasks is consumed
4074 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
4079 if (l == it->tasks_head)
4080 l = it->mg_tasks_head->next;
4082 if (l == it->mg_tasks_head)
4083 css_task_iter_advance_css_set(it);
4087 /* if PROCS, skip over tasks which aren't group leaders */
4088 if ((it->flags & CSS_TASK_ITER_PROCS) && it->task_pos &&
4089 !thread_group_leader(list_entry(it->task_pos, struct task_struct,
4095 * css_task_iter_start - initiate task iteration
4096 * @css: the css to walk tasks of
4097 * @flags: CSS_TASK_ITER_* flags
4098 * @it: the task iterator to use
4100 * Initiate iteration through the tasks of @css. The caller can call
4101 * css_task_iter_next() to walk through the tasks until the function
4102 * returns NULL. On completion of iteration, css_task_iter_end() must be
4105 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4106 struct css_task_iter *it)
4108 /* no one should try to iterate before mounting cgroups */
4109 WARN_ON_ONCE(!use_task_css_set_links);
4111 memset(it, 0, sizeof(*it));
4113 spin_lock_irq(&css_set_lock);
4119 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4121 it->cset_pos = &css->cgroup->cset_links;
4123 it->cset_head = it->cset_pos;
4125 css_task_iter_advance_css_set(it);
4127 spin_unlock_irq(&css_set_lock);
4131 * css_task_iter_next - return the next task for the iterator
4132 * @it: the task iterator being iterated
4134 * The "next" function for task iteration. @it should have been
4135 * initialized via css_task_iter_start(). Returns NULL when the iteration
4138 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4141 put_task_struct(it->cur_task);
4142 it->cur_task = NULL;
4145 spin_lock_irq(&css_set_lock);
4148 it->cur_task = list_entry(it->task_pos, struct task_struct,
4150 get_task_struct(it->cur_task);
4151 css_task_iter_advance(it);
4154 spin_unlock_irq(&css_set_lock);
4156 return it->cur_task;
4160 * css_task_iter_end - finish task iteration
4161 * @it: the task iterator to finish
4163 * Finish task iteration started by css_task_iter_start().
4165 void css_task_iter_end(struct css_task_iter *it)
4168 spin_lock_irq(&css_set_lock);
4169 list_del(&it->iters_node);
4170 put_css_set_locked(it->cur_cset);
4171 spin_unlock_irq(&css_set_lock);
4175 put_css_set(it->cur_dcset);
4178 put_task_struct(it->cur_task);
4181 static void cgroup_procs_release(struct kernfs_open_file *of)
4184 css_task_iter_end(of->priv);
4189 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4191 struct kernfs_open_file *of = s->private;
4192 struct css_task_iter *it = of->priv;
4194 return css_task_iter_next(it);
4197 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4198 unsigned int iter_flags)
4200 struct kernfs_open_file *of = s->private;
4201 struct cgroup *cgrp = seq_css(s)->cgroup;
4202 struct css_task_iter *it = of->priv;
4205 * When a seq_file is seeked, it's always traversed sequentially
4206 * from position 0, so we can simply keep iterating on !0 *pos.
4209 if (WARN_ON_ONCE((*pos)++))
4210 return ERR_PTR(-EINVAL);
4212 it = kzalloc(sizeof(*it), GFP_KERNEL);
4214 return ERR_PTR(-ENOMEM);
4216 css_task_iter_start(&cgrp->self, iter_flags, it);
4217 } else if (!(*pos)++) {
4218 css_task_iter_end(it);
4219 css_task_iter_start(&cgrp->self, iter_flags, it);
4222 return cgroup_procs_next(s, NULL, NULL);
4225 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4227 struct cgroup *cgrp = seq_css(s)->cgroup;
4230 * All processes of a threaded subtree belong to the domain cgroup
4231 * of the subtree. Only threads can be distributed across the
4232 * subtree. Reject reads on cgroup.procs in the subtree proper.
4233 * They're always empty anyway.
4235 if (cgroup_is_threaded(cgrp))
4236 return ERR_PTR(-EOPNOTSUPP);
4238 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4239 CSS_TASK_ITER_THREADED);
4242 static int cgroup_procs_show(struct seq_file *s, void *v)
4244 seq_printf(s, "%d\n", task_pid_vnr(v));
4248 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4249 struct cgroup *dst_cgrp,
4250 struct super_block *sb)
4252 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4253 struct cgroup *com_cgrp = src_cgrp;
4254 struct inode *inode;
4257 lockdep_assert_held(&cgroup_mutex);
4259 /* find the common ancestor */
4260 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4261 com_cgrp = cgroup_parent(com_cgrp);
4263 /* %current should be authorized to migrate to the common ancestor */
4264 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4268 ret = inode_permission(inode, MAY_WRITE);
4274 * If namespaces are delegation boundaries, %current must be able
4275 * to see both source and destination cgroups from its namespace.
4277 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4278 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4279 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4285 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4286 char *buf, size_t nbytes, loff_t off)
4288 struct cgroup *src_cgrp, *dst_cgrp;
4289 struct task_struct *task;
4292 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4296 task = cgroup_procs_write_start(buf, true);
4297 ret = PTR_ERR_OR_ZERO(task);
4301 /* find the source cgroup */
4302 spin_lock_irq(&css_set_lock);
4303 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4304 spin_unlock_irq(&css_set_lock);
4306 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4307 of->file->f_path.dentry->d_sb);
4311 ret = cgroup_attach_task(dst_cgrp, task, true);
4314 cgroup_procs_write_finish(task);
4316 cgroup_kn_unlock(of->kn);
4318 return ret ?: nbytes;
4321 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4323 return __cgroup_procs_start(s, pos, 0);
4326 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4327 char *buf, size_t nbytes, loff_t off)
4329 struct cgroup *src_cgrp, *dst_cgrp;
4330 struct task_struct *task;
4333 buf = strstrip(buf);
4335 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4339 task = cgroup_procs_write_start(buf, false);
4340 ret = PTR_ERR_OR_ZERO(task);
4344 /* find the source cgroup */
4345 spin_lock_irq(&css_set_lock);
4346 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4347 spin_unlock_irq(&css_set_lock);
4349 /* thread migrations follow the cgroup.procs delegation rule */
4350 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4351 of->file->f_path.dentry->d_sb);
4355 /* and must be contained in the same domain */
4357 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4360 ret = cgroup_attach_task(dst_cgrp, task, false);
4363 cgroup_procs_write_finish(task);
4365 cgroup_kn_unlock(of->kn);
4367 return ret ?: nbytes;
4370 /* cgroup core interface files for the default hierarchy */
4371 static struct cftype cgroup_base_files[] = {
4373 .name = "cgroup.type",
4374 .flags = CFTYPE_NOT_ON_ROOT,
4375 .seq_show = cgroup_type_show,
4376 .write = cgroup_type_write,
4379 .name = "cgroup.procs",
4380 .flags = CFTYPE_NS_DELEGATABLE,
4381 .file_offset = offsetof(struct cgroup, procs_file),
4382 .release = cgroup_procs_release,
4383 .seq_start = cgroup_procs_start,
4384 .seq_next = cgroup_procs_next,
4385 .seq_show = cgroup_procs_show,
4386 .write = cgroup_procs_write,
4389 .name = "cgroup.threads",
4390 .release = cgroup_procs_release,
4391 .seq_start = cgroup_threads_start,
4392 .seq_next = cgroup_procs_next,
4393 .seq_show = cgroup_procs_show,
4394 .write = cgroup_threads_write,
4397 .name = "cgroup.controllers",
4398 .seq_show = cgroup_controllers_show,
4401 .name = "cgroup.subtree_control",
4402 .flags = CFTYPE_NS_DELEGATABLE,
4403 .seq_show = cgroup_subtree_control_show,
4404 .write = cgroup_subtree_control_write,
4407 .name = "cgroup.events",
4408 .flags = CFTYPE_NOT_ON_ROOT,
4409 .file_offset = offsetof(struct cgroup, events_file),
4410 .seq_show = cgroup_events_show,
4413 .name = "cgroup.max.descendants",
4414 .seq_show = cgroup_max_descendants_show,
4415 .write = cgroup_max_descendants_write,
4418 .name = "cgroup.max.depth",
4419 .seq_show = cgroup_max_depth_show,
4420 .write = cgroup_max_depth_write,
4423 .name = "cgroup.stat",
4424 .seq_show = cgroup_stats_show,
4430 * css destruction is four-stage process.
4432 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4433 * Implemented in kill_css().
4435 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4436 * and thus css_tryget_online() is guaranteed to fail, the css can be
4437 * offlined by invoking offline_css(). After offlining, the base ref is
4438 * put. Implemented in css_killed_work_fn().
4440 * 3. When the percpu_ref reaches zero, the only possible remaining
4441 * accessors are inside RCU read sections. css_release() schedules the
4444 * 4. After the grace period, the css can be freed. Implemented in
4445 * css_free_work_fn().
4447 * It is actually hairier because both step 2 and 4 require process context
4448 * and thus involve punting to css->destroy_work adding two additional
4449 * steps to the already complex sequence.
4451 static void css_free_work_fn(struct work_struct *work)
4453 struct cgroup_subsys_state *css =
4454 container_of(work, struct cgroup_subsys_state, destroy_work);
4455 struct cgroup_subsys *ss = css->ss;
4456 struct cgroup *cgrp = css->cgroup;
4458 percpu_ref_exit(&css->refcnt);
4462 struct cgroup_subsys_state *parent = css->parent;
4466 cgroup_idr_remove(&ss->css_idr, id);
4472 /* cgroup free path */
4473 atomic_dec(&cgrp->root->nr_cgrps);
4474 cgroup1_pidlist_destroy_all(cgrp);
4475 cancel_work_sync(&cgrp->release_agent_work);
4477 if (cgroup_parent(cgrp)) {
4479 * We get a ref to the parent, and put the ref when
4480 * this cgroup is being freed, so it's guaranteed
4481 * that the parent won't be destroyed before its
4484 cgroup_put(cgroup_parent(cgrp));
4485 kernfs_put(cgrp->kn);
4489 * This is root cgroup's refcnt reaching zero,
4490 * which indicates that the root should be
4493 cgroup_destroy_root(cgrp->root);
4498 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4500 struct cgroup_subsys_state *css =
4501 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4503 INIT_WORK(&css->destroy_work, css_free_work_fn);
4504 queue_work(cgroup_destroy_wq, &css->destroy_work);
4507 static void css_release_work_fn(struct work_struct *work)
4509 struct cgroup_subsys_state *css =
4510 container_of(work, struct cgroup_subsys_state, destroy_work);
4511 struct cgroup_subsys *ss = css->ss;
4512 struct cgroup *cgrp = css->cgroup;
4514 mutex_lock(&cgroup_mutex);
4516 css->flags |= CSS_RELEASED;
4517 list_del_rcu(&css->sibling);
4520 /* css release path */
4521 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4522 if (ss->css_released)
4523 ss->css_released(css);
4525 struct cgroup *tcgrp;
4527 /* cgroup release path */
4528 trace_cgroup_release(cgrp);
4530 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4531 tcgrp = cgroup_parent(tcgrp))
4532 tcgrp->nr_dying_descendants--;
4534 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4538 * There are two control paths which try to determine
4539 * cgroup from dentry without going through kernfs -
4540 * cgroupstats_build() and css_tryget_online_from_dir().
4541 * Those are supported by RCU protecting clearing of
4542 * cgrp->kn->priv backpointer.
4545 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4548 cgroup_bpf_put(cgrp);
4551 mutex_unlock(&cgroup_mutex);
4553 call_rcu(&css->rcu_head, css_free_rcu_fn);
4556 static void css_release(struct percpu_ref *ref)
4558 struct cgroup_subsys_state *css =
4559 container_of(ref, struct cgroup_subsys_state, refcnt);
4561 INIT_WORK(&css->destroy_work, css_release_work_fn);
4562 queue_work(cgroup_destroy_wq, &css->destroy_work);
4565 static void init_and_link_css(struct cgroup_subsys_state *css,
4566 struct cgroup_subsys *ss, struct cgroup *cgrp)
4568 lockdep_assert_held(&cgroup_mutex);
4570 cgroup_get_live(cgrp);
4572 memset(css, 0, sizeof(*css));
4576 INIT_LIST_HEAD(&css->sibling);
4577 INIT_LIST_HEAD(&css->children);
4578 css->serial_nr = css_serial_nr_next++;
4579 atomic_set(&css->online_cnt, 0);
4581 if (cgroup_parent(cgrp)) {
4582 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4583 css_get(css->parent);
4586 BUG_ON(cgroup_css(cgrp, ss));
4589 /* invoke ->css_online() on a new CSS and mark it online if successful */
4590 static int online_css(struct cgroup_subsys_state *css)
4592 struct cgroup_subsys *ss = css->ss;
4595 lockdep_assert_held(&cgroup_mutex);
4598 ret = ss->css_online(css);
4600 css->flags |= CSS_ONLINE;
4601 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4603 atomic_inc(&css->online_cnt);
4605 atomic_inc(&css->parent->online_cnt);
4610 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4611 static void offline_css(struct cgroup_subsys_state *css)
4613 struct cgroup_subsys *ss = css->ss;
4615 lockdep_assert_held(&cgroup_mutex);
4617 if (!(css->flags & CSS_ONLINE))
4623 if (ss->css_offline)
4624 ss->css_offline(css);
4626 css->flags &= ~CSS_ONLINE;
4627 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4629 wake_up_all(&css->cgroup->offline_waitq);
4633 * css_create - create a cgroup_subsys_state
4634 * @cgrp: the cgroup new css will be associated with
4635 * @ss: the subsys of new css
4637 * Create a new css associated with @cgrp - @ss pair. On success, the new
4638 * css is online and installed in @cgrp. This function doesn't create the
4639 * interface files. Returns 0 on success, -errno on failure.
4641 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4642 struct cgroup_subsys *ss)
4644 struct cgroup *parent = cgroup_parent(cgrp);
4645 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4646 struct cgroup_subsys_state *css;
4649 lockdep_assert_held(&cgroup_mutex);
4651 css = ss->css_alloc(parent_css);
4653 css = ERR_PTR(-ENOMEM);
4657 init_and_link_css(css, ss, cgrp);
4659 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4663 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4668 /* @css is ready to be brought online now, make it visible */
4669 list_add_tail_rcu(&css->sibling, &parent_css->children);
4670 cgroup_idr_replace(&ss->css_idr, css, css->id);
4672 err = online_css(css);
4676 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4677 cgroup_parent(parent)) {
4678 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4679 current->comm, current->pid, ss->name);
4680 if (!strcmp(ss->name, "memory"))
4681 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4682 ss->warned_broken_hierarchy = true;
4688 list_del_rcu(&css->sibling);
4690 call_rcu(&css->rcu_head, css_free_rcu_fn);
4691 return ERR_PTR(err);
4695 * The returned cgroup is fully initialized including its control mask, but
4696 * it isn't associated with its kernfs_node and doesn't have the control
4699 static struct cgroup *cgroup_create(struct cgroup *parent)
4701 struct cgroup_root *root = parent->root;
4702 struct cgroup *cgrp, *tcgrp;
4703 int level = parent->level + 1;
4706 /* allocate the cgroup and its ID, 0 is reserved for the root */
4707 cgrp = kzalloc(sizeof(*cgrp) +
4708 sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4710 return ERR_PTR(-ENOMEM);
4712 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4717 * Temporarily set the pointer to NULL, so idr_find() won't return
4718 * a half-baked cgroup.
4720 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4723 goto out_cancel_ref;
4726 init_cgroup_housekeeping(cgrp);
4728 cgrp->self.parent = &parent->self;
4730 cgrp->level = level;
4732 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
4733 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4736 tcgrp->nr_descendants++;
4739 if (notify_on_release(parent))
4740 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4742 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4743 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4745 cgrp->self.serial_nr = css_serial_nr_next++;
4747 /* allocation complete, commit to creation */
4748 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4749 atomic_inc(&root->nr_cgrps);
4750 cgroup_get_live(parent);
4753 * @cgrp is now fully operational. If something fails after this
4754 * point, it'll be released via the normal destruction path.
4756 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4759 * On the default hierarchy, a child doesn't automatically inherit
4760 * subtree_control from the parent. Each is configured manually.
4762 if (!cgroup_on_dfl(cgrp))
4763 cgrp->subtree_control = cgroup_control(cgrp);
4766 cgroup_bpf_inherit(cgrp, parent);
4768 cgroup_propagate_control(cgrp);
4773 percpu_ref_exit(&cgrp->self.refcnt);
4776 return ERR_PTR(ret);
4779 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
4781 struct cgroup *cgroup;
4785 lockdep_assert_held(&cgroup_mutex);
4787 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
4788 if (cgroup->nr_descendants >= cgroup->max_descendants)
4791 if (level > cgroup->max_depth)
4802 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
4804 struct cgroup *parent, *cgrp;
4805 struct kernfs_node *kn;
4808 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
4809 if (strchr(name, '\n'))
4812 parent = cgroup_kn_lock_live(parent_kn, false);
4816 if (!cgroup_check_hierarchy_limits(parent)) {
4821 cgrp = cgroup_create(parent);
4823 ret = PTR_ERR(cgrp);
4827 /* create the directory */
4828 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4836 * This extra ref will be put in cgroup_free_fn() and guarantees
4837 * that @cgrp->kn is always accessible.
4841 ret = cgroup_kn_set_ugid(kn);
4845 ret = css_populate_dir(&cgrp->self);
4849 ret = cgroup_apply_control_enable(cgrp);
4853 trace_cgroup_mkdir(cgrp);
4855 /* let's create and online css's */
4856 kernfs_activate(kn);
4862 cgroup_destroy_locked(cgrp);
4864 cgroup_kn_unlock(parent_kn);
4869 * This is called when the refcnt of a css is confirmed to be killed.
4870 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4871 * initate destruction and put the css ref from kill_css().
4873 static void css_killed_work_fn(struct work_struct *work)
4875 struct cgroup_subsys_state *css =
4876 container_of(work, struct cgroup_subsys_state, destroy_work);
4878 mutex_lock(&cgroup_mutex);
4883 /* @css can't go away while we're holding cgroup_mutex */
4885 } while (css && atomic_dec_and_test(&css->online_cnt));
4887 mutex_unlock(&cgroup_mutex);
4890 /* css kill confirmation processing requires process context, bounce */
4891 static void css_killed_ref_fn(struct percpu_ref *ref)
4893 struct cgroup_subsys_state *css =
4894 container_of(ref, struct cgroup_subsys_state, refcnt);
4896 if (atomic_dec_and_test(&css->online_cnt)) {
4897 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4898 queue_work(cgroup_destroy_wq, &css->destroy_work);
4903 * kill_css - destroy a css
4904 * @css: css to destroy
4906 * This function initiates destruction of @css by removing cgroup interface
4907 * files and putting its base reference. ->css_offline() will be invoked
4908 * asynchronously once css_tryget_online() is guaranteed to fail and when
4909 * the reference count reaches zero, @css will be released.
4911 static void kill_css(struct cgroup_subsys_state *css)
4913 lockdep_assert_held(&cgroup_mutex);
4915 if (css->flags & CSS_DYING)
4918 css->flags |= CSS_DYING;
4921 * This must happen before css is disassociated with its cgroup.
4922 * See seq_css() for details.
4927 * Killing would put the base ref, but we need to keep it alive
4928 * until after ->css_offline().
4933 * cgroup core guarantees that, by the time ->css_offline() is
4934 * invoked, no new css reference will be given out via
4935 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4936 * proceed to offlining css's because percpu_ref_kill() doesn't
4937 * guarantee that the ref is seen as killed on all CPUs on return.
4939 * Use percpu_ref_kill_and_confirm() to get notifications as each
4940 * css is confirmed to be seen as killed on all CPUs.
4942 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4946 * cgroup_destroy_locked - the first stage of cgroup destruction
4947 * @cgrp: cgroup to be destroyed
4949 * css's make use of percpu refcnts whose killing latency shouldn't be
4950 * exposed to userland and are RCU protected. Also, cgroup core needs to
4951 * guarantee that css_tryget_online() won't succeed by the time
4952 * ->css_offline() is invoked. To satisfy all the requirements,
4953 * destruction is implemented in the following two steps.
4955 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4956 * userland visible parts and start killing the percpu refcnts of
4957 * css's. Set up so that the next stage will be kicked off once all
4958 * the percpu refcnts are confirmed to be killed.
4960 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4961 * rest of destruction. Once all cgroup references are gone, the
4962 * cgroup is RCU-freed.
4964 * This function implements s1. After this step, @cgrp is gone as far as
4965 * the userland is concerned and a new cgroup with the same name may be
4966 * created. As cgroup doesn't care about the names internally, this
4967 * doesn't cause any problem.
4969 static int cgroup_destroy_locked(struct cgroup *cgrp)
4970 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4972 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
4973 struct cgroup_subsys_state *css;
4974 struct cgrp_cset_link *link;
4977 lockdep_assert_held(&cgroup_mutex);
4980 * Only migration can raise populated from zero and we're already
4981 * holding cgroup_mutex.
4983 if (cgroup_is_populated(cgrp))
4987 * Make sure there's no live children. We can't test emptiness of
4988 * ->self.children as dead children linger on it while being
4989 * drained; otherwise, "rmdir parent/child parent" may fail.
4991 if (css_has_online_children(&cgrp->self))
4995 * Mark @cgrp and the associated csets dead. The former prevents
4996 * further task migration and child creation by disabling
4997 * cgroup_lock_live_group(). The latter makes the csets ignored by
4998 * the migration path.
5000 cgrp->self.flags &= ~CSS_ONLINE;
5002 spin_lock_irq(&css_set_lock);
5003 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5004 link->cset->dead = true;
5005 spin_unlock_irq(&css_set_lock);
5007 /* initiate massacre of all css's */
5008 for_each_css(css, ssid, cgrp)
5012 * Remove @cgrp directory along with the base files. @cgrp has an
5013 * extra ref on its kn.
5015 kernfs_remove(cgrp->kn);
5017 if (parent && cgroup_is_threaded(cgrp))
5018 parent->nr_threaded_children--;
5020 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5021 tcgrp->nr_descendants--;
5022 tcgrp->nr_dying_descendants++;
5025 cgroup1_check_for_release(cgroup_parent(cgrp));
5027 /* put the base reference */
5028 percpu_ref_kill(&cgrp->self.refcnt);
5033 int cgroup_rmdir(struct kernfs_node *kn)
5035 struct cgroup *cgrp;
5038 cgrp = cgroup_kn_lock_live(kn, false);
5042 ret = cgroup_destroy_locked(cgrp);
5045 trace_cgroup_rmdir(cgrp);
5047 cgroup_kn_unlock(kn);
5051 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5052 .show_options = cgroup_show_options,
5053 .remount_fs = cgroup_remount,
5054 .mkdir = cgroup_mkdir,
5055 .rmdir = cgroup_rmdir,
5056 .show_path = cgroup_show_path,
5059 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5061 struct cgroup_subsys_state *css;
5063 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5065 mutex_lock(&cgroup_mutex);
5067 idr_init(&ss->css_idr);
5068 INIT_LIST_HEAD(&ss->cfts);
5070 /* Create the root cgroup state for this subsystem */
5071 ss->root = &cgrp_dfl_root;
5072 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5073 /* We don't handle early failures gracefully */
5074 BUG_ON(IS_ERR(css));
5075 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5078 * Root csses are never destroyed and we can't initialize
5079 * percpu_ref during early init. Disable refcnting.
5081 css->flags |= CSS_NO_REF;
5084 /* allocation can't be done safely during early init */
5087 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5088 BUG_ON(css->id < 0);
5091 /* Update the init_css_set to contain a subsys
5092 * pointer to this state - since the subsystem is
5093 * newly registered, all tasks and hence the
5094 * init_css_set is in the subsystem's root cgroup. */
5095 init_css_set.subsys[ss->id] = css;
5097 have_fork_callback |= (bool)ss->fork << ss->id;
5098 have_exit_callback |= (bool)ss->exit << ss->id;
5099 have_free_callback |= (bool)ss->free << ss->id;
5100 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5102 /* At system boot, before all subsystems have been
5103 * registered, no tasks have been forked, so we don't
5104 * need to invoke fork callbacks here. */
5105 BUG_ON(!list_empty(&init_task.tasks));
5107 BUG_ON(online_css(css));
5109 mutex_unlock(&cgroup_mutex);
5113 * cgroup_init_early - cgroup initialization at system boot
5115 * Initialize cgroups at system boot, and initialize any
5116 * subsystems that request early init.
5118 int __init cgroup_init_early(void)
5120 static struct cgroup_sb_opts __initdata opts;
5121 struct cgroup_subsys *ss;
5124 init_cgroup_root(&cgrp_dfl_root, &opts);
5125 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5127 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5129 for_each_subsys(ss, i) {
5130 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5131 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5132 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5134 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5135 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5138 ss->name = cgroup_subsys_name[i];
5139 if (!ss->legacy_name)
5140 ss->legacy_name = cgroup_subsys_name[i];
5143 cgroup_init_subsys(ss, true);
5148 static u16 cgroup_disable_mask __initdata;
5151 * cgroup_init - cgroup initialization
5153 * Register cgroup filesystem and /proc file, and initialize
5154 * any subsystems that didn't request early init.
5156 int __init cgroup_init(void)
5158 struct cgroup_subsys *ss;
5161 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5162 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5163 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5164 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5167 * The latency of the synchronize_sched() is too high for cgroups,
5168 * avoid it at the cost of forcing all readers into the slow path.
5170 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5172 get_user_ns(init_cgroup_ns.user_ns);
5174 mutex_lock(&cgroup_mutex);
5177 * Add init_css_set to the hash table so that dfl_root can link to
5180 hash_add(css_set_table, &init_css_set.hlist,
5181 css_set_hash(init_css_set.subsys));
5183 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
5185 mutex_unlock(&cgroup_mutex);
5187 for_each_subsys(ss, ssid) {
5188 if (ss->early_init) {
5189 struct cgroup_subsys_state *css =
5190 init_css_set.subsys[ss->id];
5192 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5194 BUG_ON(css->id < 0);
5196 cgroup_init_subsys(ss, false);
5199 list_add_tail(&init_css_set.e_cset_node[ssid],
5200 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5203 * Setting dfl_root subsys_mask needs to consider the
5204 * disabled flag and cftype registration needs kmalloc,
5205 * both of which aren't available during early_init.
5207 if (cgroup_disable_mask & (1 << ssid)) {
5208 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5209 printk(KERN_INFO "Disabling %s control group subsystem\n",
5214 if (cgroup1_ssid_disabled(ssid))
5215 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5218 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5220 /* implicit controllers must be threaded too */
5221 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5223 if (ss->implicit_on_dfl)
5224 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5225 else if (!ss->dfl_cftypes)
5226 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5229 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5231 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5232 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5234 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5235 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5239 ss->bind(init_css_set.subsys[ssid]);
5242 /* init_css_set.subsys[] has been updated, re-hash */
5243 hash_del(&init_css_set.hlist);
5244 hash_add(css_set_table, &init_css_set.hlist,
5245 css_set_hash(init_css_set.subsys));
5247 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5248 WARN_ON(register_filesystem(&cgroup_fs_type));
5249 WARN_ON(register_filesystem(&cgroup2_fs_type));
5250 WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5255 static int __init cgroup_wq_init(void)
5258 * There isn't much point in executing destruction path in
5259 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5260 * Use 1 for @max_active.
5262 * We would prefer to do this in cgroup_init() above, but that
5263 * is called before init_workqueues(): so leave this until after.
5265 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5266 BUG_ON(!cgroup_destroy_wq);
5269 core_initcall(cgroup_wq_init);
5272 * proc_cgroup_show()
5273 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5274 * - Used for /proc/<pid>/cgroup.
5276 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5277 struct pid *pid, struct task_struct *tsk)
5281 struct cgroup_root *root;
5284 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5288 mutex_lock(&cgroup_mutex);
5289 spin_lock_irq(&css_set_lock);
5291 for_each_root(root) {
5292 struct cgroup_subsys *ss;
5293 struct cgroup *cgrp;
5294 int ssid, count = 0;
5296 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5299 seq_printf(m, "%d:", root->hierarchy_id);
5300 if (root != &cgrp_dfl_root)
5301 for_each_subsys(ss, ssid)
5302 if (root->subsys_mask & (1 << ssid))
5303 seq_printf(m, "%s%s", count++ ? "," : "",
5305 if (strlen(root->name))
5306 seq_printf(m, "%sname=%s", count ? "," : "",
5310 cgrp = task_cgroup_from_root(tsk, root);
5313 * On traditional hierarchies, all zombie tasks show up as
5314 * belonging to the root cgroup. On the default hierarchy,
5315 * while a zombie doesn't show up in "cgroup.procs" and
5316 * thus can't be migrated, its /proc/PID/cgroup keeps
5317 * reporting the cgroup it belonged to before exiting. If
5318 * the cgroup is removed before the zombie is reaped,
5319 * " (deleted)" is appended to the cgroup path.
5321 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5322 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5323 current->nsproxy->cgroup_ns);
5324 if (retval >= PATH_MAX)
5325 retval = -ENAMETOOLONG;
5334 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5335 seq_puts(m, " (deleted)\n");
5342 spin_unlock_irq(&css_set_lock);
5343 mutex_unlock(&cgroup_mutex);
5350 * cgroup_fork - initialize cgroup related fields during copy_process()
5351 * @child: pointer to task_struct of forking parent process.
5353 * A task is associated with the init_css_set until cgroup_post_fork()
5354 * attaches it to the parent's css_set. Empty cg_list indicates that
5355 * @child isn't holding reference to its css_set.
5357 void cgroup_fork(struct task_struct *child)
5359 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5360 INIT_LIST_HEAD(&child->cg_list);
5364 * cgroup_can_fork - called on a new task before the process is exposed
5365 * @child: the task in question.
5367 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5368 * returns an error, the fork aborts with that error code. This allows for
5369 * a cgroup subsystem to conditionally allow or deny new forks.
5371 int cgroup_can_fork(struct task_struct *child)
5373 struct cgroup_subsys *ss;
5376 do_each_subsys_mask(ss, i, have_canfork_callback) {
5377 ret = ss->can_fork(child);
5380 } while_each_subsys_mask();
5385 for_each_subsys(ss, j) {
5388 if (ss->cancel_fork)
5389 ss->cancel_fork(child);
5396 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5397 * @child: the task in question
5399 * This calls the cancel_fork() callbacks if a fork failed *after*
5400 * cgroup_can_fork() succeded.
5402 void cgroup_cancel_fork(struct task_struct *child)
5404 struct cgroup_subsys *ss;
5407 for_each_subsys(ss, i)
5408 if (ss->cancel_fork)
5409 ss->cancel_fork(child);
5413 * cgroup_post_fork - called on a new task after adding it to the task list
5414 * @child: the task in question
5416 * Adds the task to the list running through its css_set if necessary and
5417 * call the subsystem fork() callbacks. Has to be after the task is
5418 * visible on the task list in case we race with the first call to
5419 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5422 void cgroup_post_fork(struct task_struct *child)
5424 struct cgroup_subsys *ss;
5428 * This may race against cgroup_enable_task_cg_lists(). As that
5429 * function sets use_task_css_set_links before grabbing
5430 * tasklist_lock and we just went through tasklist_lock to add
5431 * @child, it's guaranteed that either we see the set
5432 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5433 * @child during its iteration.
5435 * If we won the race, @child is associated with %current's
5436 * css_set. Grabbing css_set_lock guarantees both that the
5437 * association is stable, and, on completion of the parent's
5438 * migration, @child is visible in the source of migration or
5439 * already in the destination cgroup. This guarantee is necessary
5440 * when implementing operations which need to migrate all tasks of
5441 * a cgroup to another.
5443 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5444 * will remain in init_css_set. This is safe because all tasks are
5445 * in the init_css_set before cg_links is enabled and there's no
5446 * operation which transfers all tasks out of init_css_set.
5448 if (use_task_css_set_links) {
5449 struct css_set *cset;
5451 spin_lock_irq(&css_set_lock);
5452 cset = task_css_set(current);
5453 if (list_empty(&child->cg_list)) {
5456 css_set_move_task(child, NULL, cset, false);
5458 spin_unlock_irq(&css_set_lock);
5462 * Call ss->fork(). This must happen after @child is linked on
5463 * css_set; otherwise, @child might change state between ->fork()
5464 * and addition to css_set.
5466 do_each_subsys_mask(ss, i, have_fork_callback) {
5468 } while_each_subsys_mask();
5472 * cgroup_exit - detach cgroup from exiting task
5473 * @tsk: pointer to task_struct of exiting process
5475 * Description: Detach cgroup from @tsk and release it.
5477 * Note that cgroups marked notify_on_release force every task in
5478 * them to take the global cgroup_mutex mutex when exiting.
5479 * This could impact scaling on very large systems. Be reluctant to
5480 * use notify_on_release cgroups where very high task exit scaling
5481 * is required on large systems.
5483 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5484 * call cgroup_exit() while the task is still competent to handle
5485 * notify_on_release(), then leave the task attached to the root cgroup in
5486 * each hierarchy for the remainder of its exit. No need to bother with
5487 * init_css_set refcnting. init_css_set never goes away and we can't race
5488 * with migration path - PF_EXITING is visible to migration path.
5490 void cgroup_exit(struct task_struct *tsk)
5492 struct cgroup_subsys *ss;
5493 struct css_set *cset;
5497 * Unlink from @tsk from its css_set. As migration path can't race
5498 * with us, we can check css_set and cg_list without synchronization.
5500 cset = task_css_set(tsk);
5502 if (!list_empty(&tsk->cg_list)) {
5503 spin_lock_irq(&css_set_lock);
5504 css_set_move_task(tsk, cset, NULL, false);
5506 spin_unlock_irq(&css_set_lock);
5511 /* see cgroup_post_fork() for details */
5512 do_each_subsys_mask(ss, i, have_exit_callback) {
5514 } while_each_subsys_mask();
5517 void cgroup_free(struct task_struct *task)
5519 struct css_set *cset = task_css_set(task);
5520 struct cgroup_subsys *ss;
5523 do_each_subsys_mask(ss, ssid, have_free_callback) {
5525 } while_each_subsys_mask();
5530 static int __init cgroup_disable(char *str)
5532 struct cgroup_subsys *ss;
5536 while ((token = strsep(&str, ",")) != NULL) {
5540 for_each_subsys(ss, i) {
5541 if (strcmp(token, ss->name) &&
5542 strcmp(token, ss->legacy_name))
5544 cgroup_disable_mask |= 1 << i;
5549 __setup("cgroup_disable=", cgroup_disable);
5552 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5553 * @dentry: directory dentry of interest
5554 * @ss: subsystem of interest
5556 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5557 * to get the corresponding css and return it. If such css doesn't exist
5558 * or can't be pinned, an ERR_PTR value is returned.
5560 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5561 struct cgroup_subsys *ss)
5563 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5564 struct file_system_type *s_type = dentry->d_sb->s_type;
5565 struct cgroup_subsys_state *css = NULL;
5566 struct cgroup *cgrp;
5568 /* is @dentry a cgroup dir? */
5569 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5570 !kn || kernfs_type(kn) != KERNFS_DIR)
5571 return ERR_PTR(-EBADF);
5576 * This path doesn't originate from kernfs and @kn could already
5577 * have been or be removed at any point. @kn->priv is RCU
5578 * protected for this access. See css_release_work_fn() for details.
5580 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
5582 css = cgroup_css(cgrp, ss);
5584 if (!css || !css_tryget_online(css))
5585 css = ERR_PTR(-ENOENT);
5592 * css_from_id - lookup css by id
5593 * @id: the cgroup id
5594 * @ss: cgroup subsys to be looked into
5596 * Returns the css if there's valid one with @id, otherwise returns NULL.
5597 * Should be called under rcu_read_lock().
5599 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5601 WARN_ON_ONCE(!rcu_read_lock_held());
5602 return idr_find(&ss->css_idr, id);
5606 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5607 * @path: path on the default hierarchy
5609 * Find the cgroup at @path on the default hierarchy, increment its
5610 * reference count and return it. Returns pointer to the found cgroup on
5611 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5612 * if @path points to a non-directory.
5614 struct cgroup *cgroup_get_from_path(const char *path)
5616 struct kernfs_node *kn;
5617 struct cgroup *cgrp;
5619 mutex_lock(&cgroup_mutex);
5621 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5623 if (kernfs_type(kn) == KERNFS_DIR) {
5625 cgroup_get_live(cgrp);
5627 cgrp = ERR_PTR(-ENOTDIR);
5631 cgrp = ERR_PTR(-ENOENT);
5634 mutex_unlock(&cgroup_mutex);
5637 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5640 * cgroup_get_from_fd - get a cgroup pointer from a fd
5641 * @fd: fd obtained by open(cgroup2_dir)
5643 * Find the cgroup from a fd which should be obtained
5644 * by opening a cgroup directory. Returns a pointer to the
5645 * cgroup on success. ERR_PTR is returned if the cgroup
5648 struct cgroup *cgroup_get_from_fd(int fd)
5650 struct cgroup_subsys_state *css;
5651 struct cgroup *cgrp;
5656 return ERR_PTR(-EBADF);
5658 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5661 return ERR_CAST(css);
5664 if (!cgroup_on_dfl(cgrp)) {
5666 return ERR_PTR(-EBADF);
5671 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
5674 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5675 * definition in cgroup-defs.h.
5677 #ifdef CONFIG_SOCK_CGROUP_DATA
5679 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5681 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5682 static bool cgroup_sk_alloc_disabled __read_mostly;
5684 void cgroup_sk_alloc_disable(void)
5686 if (cgroup_sk_alloc_disabled)
5688 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5689 cgroup_sk_alloc_disabled = true;
5694 #define cgroup_sk_alloc_disabled false
5698 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5700 if (cgroup_sk_alloc_disabled)
5703 /* Socket clone path */
5706 * We might be cloning a socket which is left in an empty
5707 * cgroup and the cgroup might have already been rmdir'd.
5708 * Don't use cgroup_get_live().
5710 cgroup_get(sock_cgroup_ptr(skcd));
5717 struct css_set *cset;
5719 cset = task_css_set(current);
5720 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5721 skcd->val = (unsigned long)cset->dfl_cgrp;
5730 void cgroup_sk_free(struct sock_cgroup_data *skcd)
5732 cgroup_put(sock_cgroup_ptr(skcd));
5735 #endif /* CONFIG_SOCK_CGROUP_DATA */
5737 #ifdef CONFIG_CGROUP_BPF
5738 int cgroup_bpf_update(struct cgroup *cgrp, struct bpf_prog *prog,
5739 enum bpf_attach_type type, bool overridable)
5741 struct cgroup *parent = cgroup_parent(cgrp);
5744 mutex_lock(&cgroup_mutex);
5745 ret = __cgroup_bpf_update(cgrp, parent, prog, type, overridable);
5746 mutex_unlock(&cgroup_mutex);
5749 #endif /* CONFIG_CGROUP_BPF */