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>
57 #include <linux/sched/cputime.h>
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/cgroup.h>
63 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
65 /* let's not notify more than 100 times per second */
66 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
69 * cgroup_mutex is the master lock. Any modification to cgroup or its
70 * hierarchy must be performed while holding it.
72 * css_set_lock protects task->cgroups pointer, the list of css_set
73 * objects, and the chain of tasks off each css_set.
75 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
76 * cgroup.h can use them for lockdep annotations.
78 DEFINE_MUTEX(cgroup_mutex);
79 DEFINE_SPINLOCK(css_set_lock);
81 #ifdef CONFIG_PROVE_RCU
82 EXPORT_SYMBOL_GPL(cgroup_mutex);
83 EXPORT_SYMBOL_GPL(css_set_lock);
86 DEFINE_SPINLOCK(trace_cgroup_path_lock);
87 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
90 * Protects cgroup_idr and css_idr so that IDs can be released without
91 * grabbing cgroup_mutex.
93 static DEFINE_SPINLOCK(cgroup_idr_lock);
96 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
97 * against file removal/re-creation across css hiding.
99 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
101 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
103 #define cgroup_assert_mutex_or_rcu_locked() \
104 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
105 !lockdep_is_held(&cgroup_mutex), \
106 "cgroup_mutex or RCU read lock required");
109 * cgroup destruction makes heavy use of work items and there can be a lot
110 * of concurrent destructions. Use a separate workqueue so that cgroup
111 * destruction work items don't end up filling up max_active of system_wq
112 * which may lead to deadlock.
114 static struct workqueue_struct *cgroup_destroy_wq;
116 /* generate an array of cgroup subsystem pointers */
117 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
118 struct cgroup_subsys *cgroup_subsys[] = {
119 #include <linux/cgroup_subsys.h>
123 /* array of cgroup subsystem names */
124 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
125 static const char *cgroup_subsys_name[] = {
126 #include <linux/cgroup_subsys.h>
130 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
132 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
133 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
134 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
135 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
136 #include <linux/cgroup_subsys.h>
139 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
140 static struct static_key_true *cgroup_subsys_enabled_key[] = {
141 #include <linux/cgroup_subsys.h>
145 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
146 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
147 #include <linux/cgroup_subsys.h>
151 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
154 * The default hierarchy, reserved for the subsystems that are otherwise
155 * unattached - it never has more than a single cgroup, and all tasks are
156 * part of that cgroup.
158 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
159 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
162 * The default hierarchy always exists but is hidden until mounted for the
163 * first time. This is for backward compatibility.
165 static bool cgrp_dfl_visible;
167 /* some controllers are not supported in the default hierarchy */
168 static u16 cgrp_dfl_inhibit_ss_mask;
170 /* some controllers are implicitly enabled on the default hierarchy */
171 static u16 cgrp_dfl_implicit_ss_mask;
173 /* some controllers can be threaded on the default hierarchy */
174 static u16 cgrp_dfl_threaded_ss_mask;
176 /* The list of hierarchy roots */
177 LIST_HEAD(cgroup_roots);
178 static int cgroup_root_count;
180 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
181 static DEFINE_IDR(cgroup_hierarchy_idr);
184 * Assign a monotonically increasing serial number to csses. It guarantees
185 * cgroups with bigger numbers are newer than those with smaller numbers.
186 * Also, as csses are always appended to the parent's ->children list, it
187 * guarantees that sibling csses are always sorted in the ascending serial
188 * number order on the list. Protected by cgroup_mutex.
190 static u64 css_serial_nr_next = 1;
193 * These bitmasks identify subsystems with specific features to avoid
194 * having to do iterative checks repeatedly.
196 static u16 have_fork_callback __read_mostly;
197 static u16 have_exit_callback __read_mostly;
198 static u16 have_release_callback __read_mostly;
199 static u16 have_canfork_callback __read_mostly;
201 /* cgroup namespace for init task */
202 struct cgroup_namespace init_cgroup_ns = {
203 .count = REFCOUNT_INIT(2),
204 .user_ns = &init_user_ns,
205 .ns.ops = &cgroupns_operations,
206 .ns.inum = PROC_CGROUP_INIT_INO,
207 .root_cset = &init_css_set,
210 static struct file_system_type cgroup2_fs_type;
211 static struct cftype cgroup_base_files[];
213 static int cgroup_apply_control(struct cgroup *cgrp);
214 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
215 static void css_task_iter_skip(struct css_task_iter *it,
216 struct task_struct *task);
217 static int cgroup_destroy_locked(struct cgroup *cgrp);
218 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
219 struct cgroup_subsys *ss);
220 static void css_release(struct percpu_ref *ref);
221 static void kill_css(struct cgroup_subsys_state *css);
222 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
223 struct cgroup *cgrp, struct cftype cfts[],
227 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
228 * @ssid: subsys ID of interest
230 * cgroup_subsys_enabled() can only be used with literal subsys names which
231 * is fine for individual subsystems but unsuitable for cgroup core. This
232 * is slower static_key_enabled() based test indexed by @ssid.
234 bool cgroup_ssid_enabled(int ssid)
236 if (CGROUP_SUBSYS_COUNT == 0)
239 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
243 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
244 * @cgrp: the cgroup of interest
246 * The default hierarchy is the v2 interface of cgroup and this function
247 * can be used to test whether a cgroup is on the default hierarchy for
248 * cases where a subsystem should behave differnetly depending on the
251 * The set of behaviors which change on the default hierarchy are still
252 * being determined and the mount option is prefixed with __DEVEL__.
254 * List of changed behaviors:
256 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
257 * and "name" are disallowed.
259 * - When mounting an existing superblock, mount options should match.
261 * - Remount is disallowed.
263 * - rename(2) is disallowed.
265 * - "tasks" is removed. Everything should be at process granularity. Use
266 * "cgroup.procs" instead.
268 * - "cgroup.procs" is not sorted. pids will be unique unless they got
269 * recycled inbetween reads.
271 * - "release_agent" and "notify_on_release" are removed. Replacement
272 * notification mechanism will be implemented.
274 * - "cgroup.clone_children" is removed.
276 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
277 * and its descendants contain no task; otherwise, 1. The file also
278 * generates kernfs notification which can be monitored through poll and
279 * [di]notify when the value of the file changes.
281 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
282 * take masks of ancestors with non-empty cpus/mems, instead of being
283 * moved to an ancestor.
285 * - cpuset: a task can be moved into an empty cpuset, and again it takes
286 * masks of ancestors.
288 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
291 * - blkcg: blk-throttle becomes properly hierarchical.
293 * - debug: disallowed on the default hierarchy.
295 bool cgroup_on_dfl(const struct cgroup *cgrp)
297 return cgrp->root == &cgrp_dfl_root;
300 /* IDR wrappers which synchronize using cgroup_idr_lock */
301 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
306 idr_preload(gfp_mask);
307 spin_lock_bh(&cgroup_idr_lock);
308 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
309 spin_unlock_bh(&cgroup_idr_lock);
314 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
318 spin_lock_bh(&cgroup_idr_lock);
319 ret = idr_replace(idr, ptr, id);
320 spin_unlock_bh(&cgroup_idr_lock);
324 static void cgroup_idr_remove(struct idr *idr, int id)
326 spin_lock_bh(&cgroup_idr_lock);
328 spin_unlock_bh(&cgroup_idr_lock);
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_tryget_css - try to get a cgroup's css for the specified subsystem
475 * @cgrp: the cgroup of interest
476 * @ss: the subsystem of interest
478 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
479 * or is offline, %NULL is returned.
481 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
482 struct cgroup_subsys *ss)
484 struct cgroup_subsys_state *css;
487 css = cgroup_css(cgrp, ss);
488 if (!css || !css_tryget_online(css))
496 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
497 * @cgrp: the cgroup of interest
498 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
500 * Similar to cgroup_css() but returns the effective css, which is defined
501 * as the matching css of the nearest ancestor including self which has @ss
502 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
503 * function is guaranteed to return non-NULL css.
505 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
506 struct cgroup_subsys *ss)
508 lockdep_assert_held(&cgroup_mutex);
514 * This function is used while updating css associations and thus
515 * can't test the csses directly. Test ss_mask.
517 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
518 cgrp = cgroup_parent(cgrp);
523 return cgroup_css(cgrp, ss);
527 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
528 * @cgrp: the cgroup of interest
529 * @ss: the subsystem of interest
531 * Find and get the effective css of @cgrp for @ss. The effective css is
532 * defined as the matching css of the nearest ancestor including self which
533 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
534 * the root css is returned, so this function always returns a valid css.
535 * The returned css must be put using css_put().
537 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
538 struct cgroup_subsys *ss)
540 struct cgroup_subsys_state *css;
545 css = cgroup_css(cgrp, ss);
547 if (css && css_tryget_online(css))
549 cgrp = cgroup_parent(cgrp);
552 css = init_css_set.subsys[ss->id];
559 static void cgroup_get_live(struct cgroup *cgrp)
561 WARN_ON_ONCE(cgroup_is_dead(cgrp));
562 css_get(&cgrp->self);
565 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
567 struct cgroup *cgrp = of->kn->parent->priv;
568 struct cftype *cft = of_cft(of);
571 * This is open and unprotected implementation of cgroup_css().
572 * seq_css() is only called from a kernfs file operation which has
573 * an active reference on the file. Because all the subsystem
574 * files are drained before a css is disassociated with a cgroup,
575 * the matching css from the cgroup's subsys table is guaranteed to
576 * be and stay valid until the enclosing operation is complete.
579 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
583 EXPORT_SYMBOL_GPL(of_css);
586 * for_each_css - iterate all css's of a cgroup
587 * @css: the iteration cursor
588 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
589 * @cgrp: the target cgroup to iterate css's of
591 * Should be called under cgroup_[tree_]mutex.
593 #define for_each_css(css, ssid, cgrp) \
594 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
595 if (!((css) = rcu_dereference_check( \
596 (cgrp)->subsys[(ssid)], \
597 lockdep_is_held(&cgroup_mutex)))) { } \
601 * for_each_e_css - iterate all effective css's of a cgroup
602 * @css: the iteration cursor
603 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
604 * @cgrp: the target cgroup to iterate css's of
606 * Should be called under cgroup_[tree_]mutex.
608 #define for_each_e_css(css, ssid, cgrp) \
609 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
610 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
615 * do_each_subsys_mask - filter for_each_subsys with a bitmask
616 * @ss: the iteration cursor
617 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
618 * @ss_mask: the bitmask
620 * The block will only run for cases where the ssid-th bit (1 << ssid) of
623 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
624 unsigned long __ss_mask = (ss_mask); \
625 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
629 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
630 (ss) = cgroup_subsys[ssid]; \
633 #define while_each_subsys_mask() \
638 /* iterate over child cgrps, lock should be held throughout iteration */
639 #define cgroup_for_each_live_child(child, cgrp) \
640 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
641 if (({ lockdep_assert_held(&cgroup_mutex); \
642 cgroup_is_dead(child); })) \
646 /* walk live descendants in preorder */
647 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
648 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
649 if (({ lockdep_assert_held(&cgroup_mutex); \
650 (dsct) = (d_css)->cgroup; \
651 cgroup_is_dead(dsct); })) \
655 /* walk live descendants in postorder */
656 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
657 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
658 if (({ lockdep_assert_held(&cgroup_mutex); \
659 (dsct) = (d_css)->cgroup; \
660 cgroup_is_dead(dsct); })) \
665 * The default css_set - used by init and its children prior to any
666 * hierarchies being mounted. It contains a pointer to the root state
667 * for each subsystem. Also used to anchor the list of css_sets. Not
668 * reference-counted, to improve performance when child cgroups
669 * haven't been created.
671 struct css_set init_css_set = {
672 .refcount = REFCOUNT_INIT(1),
673 .dom_cset = &init_css_set,
674 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
675 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
676 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
677 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
678 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
679 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
680 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
681 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
684 * The following field is re-initialized when this cset gets linked
685 * in cgroup_init(). However, let's initialize the field
686 * statically too so that the default cgroup can be accessed safely
689 .dfl_cgrp = &cgrp_dfl_root.cgrp,
692 static int css_set_count = 1; /* 1 for init_css_set */
694 static bool css_set_threaded(struct css_set *cset)
696 return cset->dom_cset != cset;
700 * css_set_populated - does a css_set contain any tasks?
701 * @cset: target css_set
703 * css_set_populated() should be the same as !!cset->nr_tasks at steady
704 * state. However, css_set_populated() can be called while a task is being
705 * added to or removed from the linked list before the nr_tasks is
706 * properly updated. Hence, we can't just look at ->nr_tasks here.
708 static bool css_set_populated(struct css_set *cset)
710 lockdep_assert_held(&css_set_lock);
712 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
716 * cgroup_update_populated - update the populated count of a cgroup
717 * @cgrp: the target cgroup
718 * @populated: inc or dec populated count
720 * One of the css_sets associated with @cgrp is either getting its first
721 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
722 * count is propagated towards root so that a given cgroup's
723 * nr_populated_children is zero iff none of its descendants contain any
726 * @cgrp's interface file "cgroup.populated" is zero if both
727 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
728 * 1 otherwise. When the sum changes from or to zero, userland is notified
729 * that the content of the interface file has changed. This can be used to
730 * detect when @cgrp and its descendants become populated or empty.
732 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
734 struct cgroup *child = NULL;
735 int adj = populated ? 1 : -1;
737 lockdep_assert_held(&css_set_lock);
740 bool was_populated = cgroup_is_populated(cgrp);
743 cgrp->nr_populated_csets += adj;
745 if (cgroup_is_threaded(child))
746 cgrp->nr_populated_threaded_children += adj;
748 cgrp->nr_populated_domain_children += adj;
751 if (was_populated == cgroup_is_populated(cgrp))
754 cgroup1_check_for_release(cgrp);
755 cgroup_file_notify(&cgrp->events_file);
758 cgrp = cgroup_parent(cgrp);
763 * css_set_update_populated - update populated state of a css_set
764 * @cset: target css_set
765 * @populated: whether @cset is populated or depopulated
767 * @cset is either getting the first task or losing the last. Update the
768 * populated counters of all associated cgroups accordingly.
770 static void css_set_update_populated(struct css_set *cset, bool populated)
772 struct cgrp_cset_link *link;
774 lockdep_assert_held(&css_set_lock);
776 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
777 cgroup_update_populated(link->cgrp, populated);
781 * @task is leaving, advance task iterators which are pointing to it so
782 * that they can resume at the next position. Advancing an iterator might
783 * remove it from the list, use safe walk. See css_task_iter_skip() for
786 static void css_set_skip_task_iters(struct css_set *cset,
787 struct task_struct *task)
789 struct css_task_iter *it, *pos;
791 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
792 css_task_iter_skip(it, task);
796 * css_set_move_task - move a task from one css_set to another
797 * @task: task being moved
798 * @from_cset: css_set @task currently belongs to (may be NULL)
799 * @to_cset: new css_set @task is being moved to (may be NULL)
800 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
802 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
803 * css_set, @from_cset can be NULL. If @task is being disassociated
804 * instead of moved, @to_cset can be NULL.
806 * This function automatically handles populated counter updates and
807 * css_task_iter adjustments but the caller is responsible for managing
808 * @from_cset and @to_cset's reference counts.
810 static void css_set_move_task(struct task_struct *task,
811 struct css_set *from_cset, struct css_set *to_cset,
814 lockdep_assert_held(&css_set_lock);
816 if (to_cset && !css_set_populated(to_cset))
817 css_set_update_populated(to_cset, true);
820 WARN_ON_ONCE(list_empty(&task->cg_list));
822 css_set_skip_task_iters(from_cset, task);
823 list_del_init(&task->cg_list);
824 if (!css_set_populated(from_cset))
825 css_set_update_populated(from_cset, false);
827 WARN_ON_ONCE(!list_empty(&task->cg_list));
832 * We are synchronized through cgroup_threadgroup_rwsem
833 * against PF_EXITING setting such that we can't race
834 * against cgroup_exit() changing the css_set to
835 * init_css_set and dropping the old one.
837 WARN_ON_ONCE(task->flags & PF_EXITING);
839 rcu_assign_pointer(task->cgroups, to_cset);
840 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
846 * hash table for cgroup groups. This improves the performance to find
847 * an existing css_set. This hash doesn't (currently) take into
848 * account cgroups in empty hierarchies.
850 #define CSS_SET_HASH_BITS 7
851 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
853 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
855 unsigned long key = 0UL;
856 struct cgroup_subsys *ss;
859 for_each_subsys(ss, i)
860 key += (unsigned long)css[i];
861 key = (key >> 16) ^ key;
866 void put_css_set_locked(struct css_set *cset)
868 struct cgrp_cset_link *link, *tmp_link;
869 struct cgroup_subsys *ss;
872 lockdep_assert_held(&css_set_lock);
874 if (!refcount_dec_and_test(&cset->refcount))
877 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
879 /* This css_set is dead. unlink it and release cgroup and css refs */
880 for_each_subsys(ss, ssid) {
881 list_del(&cset->e_cset_node[ssid]);
882 css_put(cset->subsys[ssid]);
884 hash_del(&cset->hlist);
887 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
888 list_del(&link->cset_link);
889 list_del(&link->cgrp_link);
890 if (cgroup_parent(link->cgrp))
891 cgroup_put(link->cgrp);
895 if (css_set_threaded(cset)) {
896 list_del(&cset->threaded_csets_node);
897 put_css_set_locked(cset->dom_cset);
900 kfree_rcu(cset, rcu_head);
904 * compare_css_sets - helper function for find_existing_css_set().
905 * @cset: candidate css_set being tested
906 * @old_cset: existing css_set for a task
907 * @new_cgrp: cgroup that's being entered by the task
908 * @template: desired set of css pointers in css_set (pre-calculated)
910 * Returns true if "cset" matches "old_cset" except for the hierarchy
911 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
913 static bool compare_css_sets(struct css_set *cset,
914 struct css_set *old_cset,
915 struct cgroup *new_cgrp,
916 struct cgroup_subsys_state *template[])
918 struct cgroup *new_dfl_cgrp;
919 struct list_head *l1, *l2;
922 * On the default hierarchy, there can be csets which are
923 * associated with the same set of cgroups but different csses.
924 * Let's first ensure that csses match.
926 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
930 /* @cset's domain should match the default cgroup's */
931 if (cgroup_on_dfl(new_cgrp))
932 new_dfl_cgrp = new_cgrp;
934 new_dfl_cgrp = old_cset->dfl_cgrp;
936 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
940 * Compare cgroup pointers in order to distinguish between
941 * different cgroups in hierarchies. As different cgroups may
942 * share the same effective css, this comparison is always
945 l1 = &cset->cgrp_links;
946 l2 = &old_cset->cgrp_links;
948 struct cgrp_cset_link *link1, *link2;
949 struct cgroup *cgrp1, *cgrp2;
953 /* See if we reached the end - both lists are equal length. */
954 if (l1 == &cset->cgrp_links) {
955 BUG_ON(l2 != &old_cset->cgrp_links);
958 BUG_ON(l2 == &old_cset->cgrp_links);
960 /* Locate the cgroups associated with these links. */
961 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
962 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
965 /* Hierarchies should be linked in the same order. */
966 BUG_ON(cgrp1->root != cgrp2->root);
969 * If this hierarchy is the hierarchy of the cgroup
970 * that's changing, then we need to check that this
971 * css_set points to the new cgroup; if it's any other
972 * hierarchy, then this css_set should point to the
973 * same cgroup as the old css_set.
975 if (cgrp1->root == new_cgrp->root) {
976 if (cgrp1 != new_cgrp)
987 * find_existing_css_set - init css array and find the matching css_set
988 * @old_cset: the css_set that we're using before the cgroup transition
989 * @cgrp: the cgroup that we're moving into
990 * @template: out param for the new set of csses, should be clear on entry
992 static struct css_set *find_existing_css_set(struct css_set *old_cset,
994 struct cgroup_subsys_state *template[])
996 struct cgroup_root *root = cgrp->root;
997 struct cgroup_subsys *ss;
998 struct css_set *cset;
1003 * Build the set of subsystem state objects that we want to see in the
1004 * new css_set. while subsystems can change globally, the entries here
1005 * won't change, so no need for locking.
1007 for_each_subsys(ss, i) {
1008 if (root->subsys_mask & (1UL << i)) {
1010 * @ss is in this hierarchy, so we want the
1011 * effective css from @cgrp.
1013 template[i] = cgroup_e_css(cgrp, ss);
1016 * @ss is not in this hierarchy, so we don't want
1017 * to change the css.
1019 template[i] = old_cset->subsys[i];
1023 key = css_set_hash(template);
1024 hash_for_each_possible(css_set_table, cset, hlist, key) {
1025 if (!compare_css_sets(cset, old_cset, cgrp, template))
1028 /* This css_set matches what we need */
1032 /* No existing cgroup group matched */
1036 static void free_cgrp_cset_links(struct list_head *links_to_free)
1038 struct cgrp_cset_link *link, *tmp_link;
1040 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1041 list_del(&link->cset_link);
1047 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1048 * @count: the number of links to allocate
1049 * @tmp_links: list_head the allocated links are put on
1051 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1052 * through ->cset_link. Returns 0 on success or -errno.
1054 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1056 struct cgrp_cset_link *link;
1059 INIT_LIST_HEAD(tmp_links);
1061 for (i = 0; i < count; i++) {
1062 link = kzalloc(sizeof(*link), GFP_KERNEL);
1064 free_cgrp_cset_links(tmp_links);
1067 list_add(&link->cset_link, tmp_links);
1073 * link_css_set - a helper function to link a css_set to a cgroup
1074 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1075 * @cset: the css_set to be linked
1076 * @cgrp: the destination cgroup
1078 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1079 struct cgroup *cgrp)
1081 struct cgrp_cset_link *link;
1083 BUG_ON(list_empty(tmp_links));
1085 if (cgroup_on_dfl(cgrp))
1086 cset->dfl_cgrp = cgrp;
1088 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1093 * Always add links to the tail of the lists so that the lists are
1094 * in choronological order.
1096 list_move_tail(&link->cset_link, &cgrp->cset_links);
1097 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1099 if (cgroup_parent(cgrp))
1100 cgroup_get_live(cgrp);
1104 * find_css_set - return a new css_set with one cgroup updated
1105 * @old_cset: the baseline css_set
1106 * @cgrp: the cgroup to be updated
1108 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1109 * substituted into the appropriate hierarchy.
1111 static struct css_set *find_css_set(struct css_set *old_cset,
1112 struct cgroup *cgrp)
1114 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1115 struct css_set *cset;
1116 struct list_head tmp_links;
1117 struct cgrp_cset_link *link;
1118 struct cgroup_subsys *ss;
1122 lockdep_assert_held(&cgroup_mutex);
1124 /* First see if we already have a cgroup group that matches
1125 * the desired set */
1126 spin_lock_irq(&css_set_lock);
1127 cset = find_existing_css_set(old_cset, cgrp, template);
1130 spin_unlock_irq(&css_set_lock);
1135 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1139 /* Allocate all the cgrp_cset_link objects that we'll need */
1140 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1145 refcount_set(&cset->refcount, 1);
1146 cset->dom_cset = cset;
1147 INIT_LIST_HEAD(&cset->tasks);
1148 INIT_LIST_HEAD(&cset->mg_tasks);
1149 INIT_LIST_HEAD(&cset->dying_tasks);
1150 INIT_LIST_HEAD(&cset->task_iters);
1151 INIT_LIST_HEAD(&cset->threaded_csets);
1152 INIT_HLIST_NODE(&cset->hlist);
1153 INIT_LIST_HEAD(&cset->cgrp_links);
1154 INIT_LIST_HEAD(&cset->mg_preload_node);
1155 INIT_LIST_HEAD(&cset->mg_node);
1157 /* Copy the set of subsystem state objects generated in
1158 * find_existing_css_set() */
1159 memcpy(cset->subsys, template, sizeof(cset->subsys));
1161 spin_lock_irq(&css_set_lock);
1162 /* Add reference counts and links from the new css_set. */
1163 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1164 struct cgroup *c = link->cgrp;
1166 if (c->root == cgrp->root)
1168 link_css_set(&tmp_links, cset, c);
1171 BUG_ON(!list_empty(&tmp_links));
1175 /* Add @cset to the hash table */
1176 key = css_set_hash(cset->subsys);
1177 hash_add(css_set_table, &cset->hlist, key);
1179 for_each_subsys(ss, ssid) {
1180 struct cgroup_subsys_state *css = cset->subsys[ssid];
1182 list_add_tail(&cset->e_cset_node[ssid],
1183 &css->cgroup->e_csets[ssid]);
1187 spin_unlock_irq(&css_set_lock);
1190 * If @cset should be threaded, look up the matching dom_cset and
1191 * link them up. We first fully initialize @cset then look for the
1192 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1193 * to stay empty until we return.
1195 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1196 struct css_set *dcset;
1198 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1204 spin_lock_irq(&css_set_lock);
1205 cset->dom_cset = dcset;
1206 list_add_tail(&cset->threaded_csets_node,
1207 &dcset->threaded_csets);
1208 spin_unlock_irq(&css_set_lock);
1214 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1216 struct cgroup *root_cgrp = kf_root->kn->priv;
1218 return root_cgrp->root;
1221 static int cgroup_init_root_id(struct cgroup_root *root)
1225 lockdep_assert_held(&cgroup_mutex);
1227 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1231 root->hierarchy_id = id;
1235 static void cgroup_exit_root_id(struct cgroup_root *root)
1237 lockdep_assert_held(&cgroup_mutex);
1239 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1242 void cgroup_free_root(struct cgroup_root *root)
1245 idr_destroy(&root->cgroup_idr);
1250 static void cgroup_destroy_root(struct cgroup_root *root)
1252 struct cgroup *cgrp = &root->cgrp;
1253 struct cgrp_cset_link *link, *tmp_link;
1255 trace_cgroup_destroy_root(root);
1257 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1259 BUG_ON(atomic_read(&root->nr_cgrps));
1260 BUG_ON(!list_empty(&cgrp->self.children));
1262 /* Rebind all subsystems back to the default hierarchy */
1263 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1266 * Release all the links from cset_links to this hierarchy's
1269 spin_lock_irq(&css_set_lock);
1271 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1272 list_del(&link->cset_link);
1273 list_del(&link->cgrp_link);
1277 spin_unlock_irq(&css_set_lock);
1279 if (!list_empty(&root->root_list)) {
1280 list_del(&root->root_list);
1281 cgroup_root_count--;
1284 cgroup_exit_root_id(root);
1286 mutex_unlock(&cgroup_mutex);
1288 kernfs_destroy_root(root->kf_root);
1289 cgroup_free_root(root);
1293 * look up cgroup associated with current task's cgroup namespace on the
1294 * specified hierarchy
1296 static struct cgroup *
1297 current_cgns_cgroup_from_root(struct cgroup_root *root)
1299 struct cgroup *res = NULL;
1300 struct css_set *cset;
1302 lockdep_assert_held(&css_set_lock);
1306 cset = current->nsproxy->cgroup_ns->root_cset;
1307 if (cset == &init_css_set) {
1310 struct cgrp_cset_link *link;
1312 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1313 struct cgroup *c = link->cgrp;
1315 if (c->root == root) {
1327 /* look up cgroup associated with given css_set on the specified hierarchy */
1328 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1329 struct cgroup_root *root)
1331 struct cgroup *res = NULL;
1333 lockdep_assert_held(&cgroup_mutex);
1334 lockdep_assert_held(&css_set_lock);
1336 if (cset == &init_css_set) {
1338 } else if (root == &cgrp_dfl_root) {
1339 res = cset->dfl_cgrp;
1341 struct cgrp_cset_link *link;
1343 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1344 struct cgroup *c = link->cgrp;
1346 if (c->root == root) {
1358 * Return the cgroup for "task" from the given hierarchy. Must be
1359 * called with cgroup_mutex and css_set_lock held.
1361 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1362 struct cgroup_root *root)
1365 * No need to lock the task - since we hold cgroup_mutex the
1366 * task can't change groups, so the only thing that can happen
1367 * is that it exits and its css is set back to init_css_set.
1369 return cset_cgroup_from_root(task_css_set(task), root);
1373 * A task must hold cgroup_mutex to modify cgroups.
1375 * Any task can increment and decrement the count field without lock.
1376 * So in general, code holding cgroup_mutex can't rely on the count
1377 * field not changing. However, if the count goes to zero, then only
1378 * cgroup_attach_task() can increment it again. Because a count of zero
1379 * means that no tasks are currently attached, therefore there is no
1380 * way a task attached to that cgroup can fork (the other way to
1381 * increment the count). So code holding cgroup_mutex can safely
1382 * assume that if the count is zero, it will stay zero. Similarly, if
1383 * a task holds cgroup_mutex on a cgroup with zero count, it
1384 * knows that the cgroup won't be removed, as cgroup_rmdir()
1387 * A cgroup can only be deleted if both its 'count' of using tasks
1388 * is zero, and its list of 'children' cgroups is empty. Since all
1389 * tasks in the system use _some_ cgroup, and since there is always at
1390 * least one task in the system (init, pid == 1), therefore, root cgroup
1391 * always has either children cgroups and/or using tasks. So we don't
1392 * need a special hack to ensure that root cgroup cannot be deleted.
1394 * P.S. One more locking exception. RCU is used to guard the
1395 * update of a tasks cgroup pointer by cgroup_attach_task()
1398 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1400 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1403 struct cgroup_subsys *ss = cft->ss;
1405 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1406 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1407 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1408 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1411 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1416 * cgroup_file_mode - deduce file mode of a control file
1417 * @cft: the control file in question
1419 * S_IRUGO for read, S_IWUSR for write.
1421 static umode_t cgroup_file_mode(const struct cftype *cft)
1425 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1428 if (cft->write_u64 || cft->write_s64 || cft->write) {
1429 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1439 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1440 * @subtree_control: the new subtree_control mask to consider
1441 * @this_ss_mask: available subsystems
1443 * On the default hierarchy, a subsystem may request other subsystems to be
1444 * enabled together through its ->depends_on mask. In such cases, more
1445 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1447 * This function calculates which subsystems need to be enabled if
1448 * @subtree_control is to be applied while restricted to @this_ss_mask.
1450 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1452 u16 cur_ss_mask = subtree_control;
1453 struct cgroup_subsys *ss;
1456 lockdep_assert_held(&cgroup_mutex);
1458 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1461 u16 new_ss_mask = cur_ss_mask;
1463 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1464 new_ss_mask |= ss->depends_on;
1465 } while_each_subsys_mask();
1468 * Mask out subsystems which aren't available. This can
1469 * happen only if some depended-upon subsystems were bound
1470 * to non-default hierarchies.
1472 new_ss_mask &= this_ss_mask;
1474 if (new_ss_mask == cur_ss_mask)
1476 cur_ss_mask = new_ss_mask;
1483 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1484 * @kn: the kernfs_node being serviced
1486 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1487 * the method finishes if locking succeeded. Note that once this function
1488 * returns the cgroup returned by cgroup_kn_lock_live() may become
1489 * inaccessible any time. If the caller intends to continue to access the
1490 * cgroup, it should pin it before invoking this function.
1492 void cgroup_kn_unlock(struct kernfs_node *kn)
1494 struct cgroup *cgrp;
1496 if (kernfs_type(kn) == KERNFS_DIR)
1499 cgrp = kn->parent->priv;
1501 mutex_unlock(&cgroup_mutex);
1503 kernfs_unbreak_active_protection(kn);
1508 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1509 * @kn: the kernfs_node being serviced
1510 * @drain_offline: perform offline draining on the cgroup
1512 * This helper is to be used by a cgroup kernfs method currently servicing
1513 * @kn. It breaks the active protection, performs cgroup locking and
1514 * verifies that the associated cgroup is alive. Returns the cgroup if
1515 * alive; otherwise, %NULL. A successful return should be undone by a
1516 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1517 * cgroup is drained of offlining csses before return.
1519 * Any cgroup kernfs method implementation which requires locking the
1520 * associated cgroup should use this helper. It avoids nesting cgroup
1521 * locking under kernfs active protection and allows all kernfs operations
1522 * including self-removal.
1524 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1526 struct cgroup *cgrp;
1528 if (kernfs_type(kn) == KERNFS_DIR)
1531 cgrp = kn->parent->priv;
1534 * We're gonna grab cgroup_mutex which nests outside kernfs
1535 * active_ref. cgroup liveliness check alone provides enough
1536 * protection against removal. Ensure @cgrp stays accessible and
1537 * break the active_ref protection.
1539 if (!cgroup_tryget(cgrp))
1541 kernfs_break_active_protection(kn);
1544 cgroup_lock_and_drain_offline(cgrp);
1546 mutex_lock(&cgroup_mutex);
1548 if (!cgroup_is_dead(cgrp))
1551 cgroup_kn_unlock(kn);
1555 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1557 char name[CGROUP_FILE_NAME_MAX];
1559 lockdep_assert_held(&cgroup_mutex);
1561 if (cft->file_offset) {
1562 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1563 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1565 spin_lock_irq(&cgroup_file_kn_lock);
1567 spin_unlock_irq(&cgroup_file_kn_lock);
1569 del_timer_sync(&cfile->notify_timer);
1572 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1576 * css_clear_dir - remove subsys files in a cgroup directory
1579 static void css_clear_dir(struct cgroup_subsys_state *css)
1581 struct cgroup *cgrp = css->cgroup;
1582 struct cftype *cfts;
1584 if (!(css->flags & CSS_VISIBLE))
1587 css->flags &= ~CSS_VISIBLE;
1590 if (cgroup_on_dfl(cgrp))
1591 cfts = cgroup_base_files;
1593 cfts = cgroup1_base_files;
1595 cgroup_addrm_files(css, cgrp, cfts, false);
1597 list_for_each_entry(cfts, &css->ss->cfts, node)
1598 cgroup_addrm_files(css, cgrp, cfts, false);
1603 * css_populate_dir - create subsys files in a cgroup directory
1606 * On failure, no file is added.
1608 static int css_populate_dir(struct cgroup_subsys_state *css)
1610 struct cgroup *cgrp = css->cgroup;
1611 struct cftype *cfts, *failed_cfts;
1614 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1618 if (cgroup_on_dfl(cgrp))
1619 cfts = cgroup_base_files;
1621 cfts = cgroup1_base_files;
1623 ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1627 list_for_each_entry(cfts, &css->ss->cfts, node) {
1628 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1636 css->flags |= CSS_VISIBLE;
1640 list_for_each_entry(cfts, &css->ss->cfts, node) {
1641 if (cfts == failed_cfts)
1643 cgroup_addrm_files(css, cgrp, cfts, false);
1648 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1650 struct cgroup *dcgrp = &dst_root->cgrp;
1651 struct cgroup_subsys *ss;
1653 u16 dfl_disable_ss_mask = 0;
1655 lockdep_assert_held(&cgroup_mutex);
1657 do_each_subsys_mask(ss, ssid, ss_mask) {
1659 * If @ss has non-root csses attached to it, can't move.
1660 * If @ss is an implicit controller, it is exempt from this
1661 * rule and can be stolen.
1663 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1664 !ss->implicit_on_dfl)
1667 /* can't move between two non-dummy roots either */
1668 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1672 * Collect ssid's that need to be disabled from default
1675 if (ss->root == &cgrp_dfl_root)
1676 dfl_disable_ss_mask |= 1 << ssid;
1678 } while_each_subsys_mask();
1680 if (dfl_disable_ss_mask) {
1681 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1684 * Controllers from default hierarchy that need to be rebound
1685 * are all disabled together in one go.
1687 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1688 WARN_ON(cgroup_apply_control(scgrp));
1689 cgroup_finalize_control(scgrp, 0);
1692 do_each_subsys_mask(ss, ssid, ss_mask) {
1693 struct cgroup_root *src_root = ss->root;
1694 struct cgroup *scgrp = &src_root->cgrp;
1695 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1696 struct css_set *cset;
1698 WARN_ON(!css || cgroup_css(dcgrp, ss));
1700 if (src_root != &cgrp_dfl_root) {
1701 /* disable from the source */
1702 src_root->subsys_mask &= ~(1 << ssid);
1703 WARN_ON(cgroup_apply_control(scgrp));
1704 cgroup_finalize_control(scgrp, 0);
1708 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1709 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1710 ss->root = dst_root;
1711 css->cgroup = dcgrp;
1713 spin_lock_irq(&css_set_lock);
1714 hash_for_each(css_set_table, i, cset, hlist)
1715 list_move_tail(&cset->e_cset_node[ss->id],
1716 &dcgrp->e_csets[ss->id]);
1717 spin_unlock_irq(&css_set_lock);
1719 /* default hierarchy doesn't enable controllers by default */
1720 dst_root->subsys_mask |= 1 << ssid;
1721 if (dst_root == &cgrp_dfl_root) {
1722 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1724 dcgrp->subtree_control |= 1 << ssid;
1725 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1728 ret = cgroup_apply_control(dcgrp);
1730 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1735 } while_each_subsys_mask();
1737 kernfs_activate(dcgrp->kn);
1741 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1742 struct kernfs_root *kf_root)
1746 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1747 struct cgroup *ns_cgroup;
1749 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1753 spin_lock_irq(&css_set_lock);
1754 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1755 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1756 spin_unlock_irq(&css_set_lock);
1758 if (len >= PATH_MAX)
1761 seq_escape(sf, buf, " \t\n\\");
1768 static int parse_cgroup_root_flags(char *data, unsigned int *root_flags)
1774 if (!data || *data == '\0')
1777 while ((token = strsep(&data, ",")) != NULL) {
1778 if (!strcmp(token, "nsdelegate")) {
1779 *root_flags |= CGRP_ROOT_NS_DELEGATE;
1783 pr_err("cgroup2: unknown option \"%s\"\n", token);
1790 static void apply_cgroup_root_flags(unsigned int root_flags)
1792 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1793 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1794 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1796 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1800 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1802 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1803 seq_puts(seq, ",nsdelegate");
1807 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1809 unsigned int root_flags;
1812 ret = parse_cgroup_root_flags(data, &root_flags);
1816 apply_cgroup_root_flags(root_flags);
1821 * To reduce the fork() overhead for systems that are not actually using
1822 * their cgroups capability, we don't maintain the lists running through
1823 * each css_set to its tasks until we see the list actually used - in other
1824 * words after the first mount.
1826 static bool use_task_css_set_links __read_mostly;
1828 static void cgroup_enable_task_cg_lists(void)
1830 struct task_struct *p, *g;
1833 * We need tasklist_lock because RCU is not safe against
1834 * while_each_thread(). Besides, a forking task that has passed
1835 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1836 * is not guaranteed to have its child immediately visible in the
1837 * tasklist if we walk through it with RCU.
1839 read_lock(&tasklist_lock);
1840 spin_lock_irq(&css_set_lock);
1842 if (use_task_css_set_links)
1845 use_task_css_set_links = true;
1847 do_each_thread(g, p) {
1848 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1849 task_css_set(p) != &init_css_set);
1852 * We should check if the process is exiting, otherwise
1853 * it will race with cgroup_exit() in that the list
1854 * entry won't be deleted though the process has exited.
1855 * Do it while holding siglock so that we don't end up
1856 * racing against cgroup_exit().
1858 * Interrupts were already disabled while acquiring
1859 * the css_set_lock, so we do not need to disable it
1860 * again when acquiring the sighand->siglock here.
1862 spin_lock(&p->sighand->siglock);
1863 if (!(p->flags & PF_EXITING)) {
1864 struct css_set *cset = task_css_set(p);
1866 if (!css_set_populated(cset))
1867 css_set_update_populated(cset, true);
1868 list_add_tail(&p->cg_list, &cset->tasks);
1872 spin_unlock(&p->sighand->siglock);
1873 } while_each_thread(g, p);
1875 spin_unlock_irq(&css_set_lock);
1876 read_unlock(&tasklist_lock);
1879 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1881 struct cgroup_subsys *ss;
1884 INIT_LIST_HEAD(&cgrp->self.sibling);
1885 INIT_LIST_HEAD(&cgrp->self.children);
1886 INIT_LIST_HEAD(&cgrp->cset_links);
1887 INIT_LIST_HEAD(&cgrp->pidlists);
1888 mutex_init(&cgrp->pidlist_mutex);
1889 cgrp->self.cgroup = cgrp;
1890 cgrp->self.flags |= CSS_ONLINE;
1891 cgrp->dom_cgrp = cgrp;
1892 cgrp->max_descendants = INT_MAX;
1893 cgrp->max_depth = INT_MAX;
1894 INIT_LIST_HEAD(&cgrp->rstat_css_list);
1895 prev_cputime_init(&cgrp->prev_cputime);
1897 for_each_subsys(ss, ssid)
1898 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1900 init_waitqueue_head(&cgrp->offline_waitq);
1901 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1904 void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
1906 struct cgroup *cgrp = &root->cgrp;
1908 INIT_LIST_HEAD(&root->root_list);
1909 atomic_set(&root->nr_cgrps, 1);
1911 init_cgroup_housekeeping(cgrp);
1912 idr_init(&root->cgroup_idr);
1914 root->flags = opts->flags;
1915 if (opts->release_agent)
1916 strscpy(root->release_agent_path, opts->release_agent, PATH_MAX);
1918 strscpy(root->name, opts->name, MAX_CGROUP_ROOT_NAMELEN);
1919 if (opts->cpuset_clone_children)
1920 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1923 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
1925 LIST_HEAD(tmp_links);
1926 struct cgroup *root_cgrp = &root->cgrp;
1927 struct kernfs_syscall_ops *kf_sops;
1928 struct css_set *cset;
1931 lockdep_assert_held(&cgroup_mutex);
1933 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1936 root_cgrp->id = ret;
1937 root_cgrp->ancestor_ids[0] = ret;
1939 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1940 ref_flags, GFP_KERNEL);
1945 * We're accessing css_set_count without locking css_set_lock here,
1946 * but that's OK - it can only be increased by someone holding
1947 * cgroup_lock, and that's us. Later rebinding may disable
1948 * controllers on the default hierarchy and thus create new csets,
1949 * which can't be more than the existing ones. Allocate 2x.
1951 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1955 ret = cgroup_init_root_id(root);
1959 kf_sops = root == &cgrp_dfl_root ?
1960 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1962 root->kf_root = kernfs_create_root(kf_sops,
1963 KERNFS_ROOT_CREATE_DEACTIVATED |
1964 KERNFS_ROOT_SUPPORT_EXPORTOP,
1966 if (IS_ERR(root->kf_root)) {
1967 ret = PTR_ERR(root->kf_root);
1970 root_cgrp->kn = root->kf_root->kn;
1972 ret = css_populate_dir(&root_cgrp->self);
1976 ret = rebind_subsystems(root, ss_mask);
1980 ret = cgroup_bpf_inherit(root_cgrp);
1983 trace_cgroup_setup_root(root);
1986 * There must be no failure case after here, since rebinding takes
1987 * care of subsystems' refcounts, which are explicitly dropped in
1988 * the failure exit path.
1990 list_add(&root->root_list, &cgroup_roots);
1991 cgroup_root_count++;
1994 * Link the root cgroup in this hierarchy into all the css_set
1997 spin_lock_irq(&css_set_lock);
1998 hash_for_each(css_set_table, i, cset, hlist) {
1999 link_css_set(&tmp_links, cset, root_cgrp);
2000 if (css_set_populated(cset))
2001 cgroup_update_populated(root_cgrp, true);
2003 spin_unlock_irq(&css_set_lock);
2005 BUG_ON(!list_empty(&root_cgrp->self.children));
2006 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2008 kernfs_activate(root_cgrp->kn);
2013 kernfs_destroy_root(root->kf_root);
2014 root->kf_root = NULL;
2016 cgroup_exit_root_id(root);
2018 percpu_ref_exit(&root_cgrp->self.refcnt);
2020 free_cgrp_cset_links(&tmp_links);
2024 struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
2025 struct cgroup_root *root, unsigned long magic,
2026 struct cgroup_namespace *ns)
2028 struct dentry *dentry;
2029 bool new_sb = false;
2031 dentry = kernfs_mount(fs_type, flags, root->kf_root, magic, &new_sb);
2034 * In non-init cgroup namespace, instead of root cgroup's dentry,
2035 * we return the dentry corresponding to the cgroupns->root_cgrp.
2037 if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
2038 struct dentry *nsdentry;
2039 struct super_block *sb = dentry->d_sb;
2040 struct cgroup *cgrp;
2042 mutex_lock(&cgroup_mutex);
2043 spin_lock_irq(&css_set_lock);
2045 cgrp = cset_cgroup_from_root(ns->root_cset, root);
2047 spin_unlock_irq(&css_set_lock);
2048 mutex_unlock(&cgroup_mutex);
2050 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2052 if (IS_ERR(nsdentry))
2053 deactivate_locked_super(sb);
2058 cgroup_put(&root->cgrp);
2063 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
2064 int flags, const char *unused_dev_name,
2067 struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
2068 struct dentry *dentry;
2073 /* Check if the caller has permission to mount. */
2074 if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
2076 return ERR_PTR(-EPERM);
2080 * The first time anyone tries to mount a cgroup, enable the list
2081 * linking each css_set to its tasks and fix up all existing tasks.
2083 if (!use_task_css_set_links)
2084 cgroup_enable_task_cg_lists();
2086 if (fs_type == &cgroup2_fs_type) {
2087 unsigned int root_flags;
2089 ret = parse_cgroup_root_flags(data, &root_flags);
2092 return ERR_PTR(ret);
2095 cgrp_dfl_visible = true;
2096 cgroup_get_live(&cgrp_dfl_root.cgrp);
2098 dentry = cgroup_do_mount(&cgroup2_fs_type, flags, &cgrp_dfl_root,
2099 CGROUP2_SUPER_MAGIC, ns);
2100 if (!IS_ERR(dentry))
2101 apply_cgroup_root_flags(root_flags);
2103 dentry = cgroup1_mount(&cgroup_fs_type, flags, data,
2104 CGROUP_SUPER_MAGIC, ns);
2111 static void cgroup_kill_sb(struct super_block *sb)
2113 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2114 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2117 * If @root doesn't have any mounts or children, start killing it.
2118 * This prevents new mounts by disabling percpu_ref_tryget_live().
2119 * cgroup_mount() may wait for @root's release.
2121 * And don't kill the default root.
2123 if (!list_empty(&root->cgrp.self.children) ||
2124 root == &cgrp_dfl_root)
2125 cgroup_put(&root->cgrp);
2127 percpu_ref_kill(&root->cgrp.self.refcnt);
2132 struct file_system_type cgroup_fs_type = {
2134 .mount = cgroup_mount,
2135 .kill_sb = cgroup_kill_sb,
2136 .fs_flags = FS_USERNS_MOUNT,
2139 static struct file_system_type cgroup2_fs_type = {
2141 .mount = cgroup_mount,
2142 .kill_sb = cgroup_kill_sb,
2143 .fs_flags = FS_USERNS_MOUNT,
2146 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2147 struct cgroup_namespace *ns)
2149 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2151 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2154 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2155 struct cgroup_namespace *ns)
2159 mutex_lock(&cgroup_mutex);
2160 spin_lock_irq(&css_set_lock);
2162 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2164 spin_unlock_irq(&css_set_lock);
2165 mutex_unlock(&cgroup_mutex);
2169 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2172 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2173 * @task: target task
2174 * @buf: the buffer to write the path into
2175 * @buflen: the length of the buffer
2177 * Determine @task's cgroup on the first (the one with the lowest non-zero
2178 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2179 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2180 * cgroup controller callbacks.
2182 * Return value is the same as kernfs_path().
2184 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2186 struct cgroup_root *root;
2187 struct cgroup *cgrp;
2188 int hierarchy_id = 1;
2191 mutex_lock(&cgroup_mutex);
2192 spin_lock_irq(&css_set_lock);
2194 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2197 cgrp = task_cgroup_from_root(task, root);
2198 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2200 /* if no hierarchy exists, everyone is in "/" */
2201 ret = strlcpy(buf, "/", buflen);
2204 spin_unlock_irq(&css_set_lock);
2205 mutex_unlock(&cgroup_mutex);
2208 EXPORT_SYMBOL_GPL(task_cgroup_path);
2211 * cgroup_migrate_add_task - add a migration target task to a migration context
2212 * @task: target task
2213 * @mgctx: target migration context
2215 * Add @task, which is a migration target, to @mgctx->tset. This function
2216 * becomes noop if @task doesn't need to be migrated. @task's css_set
2217 * should have been added as a migration source and @task->cg_list will be
2218 * moved from the css_set's tasks list to mg_tasks one.
2220 static void cgroup_migrate_add_task(struct task_struct *task,
2221 struct cgroup_mgctx *mgctx)
2223 struct css_set *cset;
2225 lockdep_assert_held(&css_set_lock);
2227 /* @task either already exited or can't exit until the end */
2228 if (task->flags & PF_EXITING)
2231 /* leave @task alone if post_fork() hasn't linked it yet */
2232 if (list_empty(&task->cg_list))
2235 cset = task_css_set(task);
2236 if (!cset->mg_src_cgrp)
2239 mgctx->tset.nr_tasks++;
2241 list_move_tail(&task->cg_list, &cset->mg_tasks);
2242 if (list_empty(&cset->mg_node))
2243 list_add_tail(&cset->mg_node,
2244 &mgctx->tset.src_csets);
2245 if (list_empty(&cset->mg_dst_cset->mg_node))
2246 list_add_tail(&cset->mg_dst_cset->mg_node,
2247 &mgctx->tset.dst_csets);
2251 * cgroup_taskset_first - reset taskset and return the first task
2252 * @tset: taskset of interest
2253 * @dst_cssp: output variable for the destination css
2255 * @tset iteration is initialized and the first task is returned.
2257 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2258 struct cgroup_subsys_state **dst_cssp)
2260 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2261 tset->cur_task = NULL;
2263 return cgroup_taskset_next(tset, dst_cssp);
2267 * cgroup_taskset_next - iterate to the next task in taskset
2268 * @tset: taskset of interest
2269 * @dst_cssp: output variable for the destination css
2271 * Return the next task in @tset. Iteration must have been initialized
2272 * with cgroup_taskset_first().
2274 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2275 struct cgroup_subsys_state **dst_cssp)
2277 struct css_set *cset = tset->cur_cset;
2278 struct task_struct *task = tset->cur_task;
2280 while (&cset->mg_node != tset->csets) {
2282 task = list_first_entry(&cset->mg_tasks,
2283 struct task_struct, cg_list);
2285 task = list_next_entry(task, cg_list);
2287 if (&task->cg_list != &cset->mg_tasks) {
2288 tset->cur_cset = cset;
2289 tset->cur_task = task;
2292 * This function may be called both before and
2293 * after cgroup_taskset_migrate(). The two cases
2294 * can be distinguished by looking at whether @cset
2295 * has its ->mg_dst_cset set.
2297 if (cset->mg_dst_cset)
2298 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2300 *dst_cssp = cset->subsys[tset->ssid];
2305 cset = list_next_entry(cset, mg_node);
2313 * cgroup_taskset_migrate - migrate a taskset
2314 * @mgctx: migration context
2316 * Migrate tasks in @mgctx as setup by migration preparation functions.
2317 * This function fails iff one of the ->can_attach callbacks fails and
2318 * guarantees that either all or none of the tasks in @mgctx are migrated.
2319 * @mgctx is consumed regardless of success.
2321 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2323 struct cgroup_taskset *tset = &mgctx->tset;
2324 struct cgroup_subsys *ss;
2325 struct task_struct *task, *tmp_task;
2326 struct css_set *cset, *tmp_cset;
2327 int ssid, failed_ssid, ret;
2329 /* check that we can legitimately attach to the cgroup */
2330 if (tset->nr_tasks) {
2331 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2332 if (ss->can_attach) {
2334 ret = ss->can_attach(tset);
2337 goto out_cancel_attach;
2340 } while_each_subsys_mask();
2344 * Now that we're guaranteed success, proceed to move all tasks to
2345 * the new cgroup. There are no failure cases after here, so this
2346 * is the commit point.
2348 spin_lock_irq(&css_set_lock);
2349 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2350 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2351 struct css_set *from_cset = task_css_set(task);
2352 struct css_set *to_cset = cset->mg_dst_cset;
2354 get_css_set(to_cset);
2355 to_cset->nr_tasks++;
2356 css_set_move_task(task, from_cset, to_cset, true);
2357 put_css_set_locked(from_cset);
2358 from_cset->nr_tasks--;
2361 spin_unlock_irq(&css_set_lock);
2364 * Migration is committed, all target tasks are now on dst_csets.
2365 * Nothing is sensitive to fork() after this point. Notify
2366 * controllers that migration is complete.
2368 tset->csets = &tset->dst_csets;
2370 if (tset->nr_tasks) {
2371 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2376 } while_each_subsys_mask();
2380 goto out_release_tset;
2383 if (tset->nr_tasks) {
2384 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2385 if (ssid == failed_ssid)
2387 if (ss->cancel_attach) {
2389 ss->cancel_attach(tset);
2391 } while_each_subsys_mask();
2394 spin_lock_irq(&css_set_lock);
2395 list_splice_init(&tset->dst_csets, &tset->src_csets);
2396 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2397 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2398 list_del_init(&cset->mg_node);
2400 spin_unlock_irq(&css_set_lock);
2403 * Re-initialize the cgroup_taskset structure in case it is reused
2404 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2408 tset->csets = &tset->src_csets;
2413 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2414 * @dst_cgrp: destination cgroup to test
2416 * On the default hierarchy, except for the mixable, (possible) thread root
2417 * and threaded cgroups, subtree_control must be zero for migration
2418 * destination cgroups with tasks so that child cgroups don't compete
2421 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2423 /* v1 doesn't have any restriction */
2424 if (!cgroup_on_dfl(dst_cgrp))
2427 /* verify @dst_cgrp can host resources */
2428 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2431 /* mixables don't care */
2432 if (cgroup_is_mixable(dst_cgrp))
2436 * If @dst_cgrp is already or can become a thread root or is
2437 * threaded, it doesn't matter.
2439 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2442 /* apply no-internal-process constraint */
2443 if (dst_cgrp->subtree_control)
2450 * cgroup_migrate_finish - cleanup after attach
2451 * @mgctx: migration context
2453 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2454 * those functions for details.
2456 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2458 LIST_HEAD(preloaded);
2459 struct css_set *cset, *tmp_cset;
2461 lockdep_assert_held(&cgroup_mutex);
2463 spin_lock_irq(&css_set_lock);
2465 list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2466 list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2468 list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2469 cset->mg_src_cgrp = NULL;
2470 cset->mg_dst_cgrp = NULL;
2471 cset->mg_dst_cset = NULL;
2472 list_del_init(&cset->mg_preload_node);
2473 put_css_set_locked(cset);
2476 spin_unlock_irq(&css_set_lock);
2480 * cgroup_migrate_add_src - add a migration source css_set
2481 * @src_cset: the source css_set to add
2482 * @dst_cgrp: the destination cgroup
2483 * @mgctx: migration context
2485 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2486 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2487 * up by cgroup_migrate_finish().
2489 * This function may be called without holding cgroup_threadgroup_rwsem
2490 * even if the target is a process. Threads may be created and destroyed
2491 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2492 * into play and the preloaded css_sets are guaranteed to cover all
2495 void cgroup_migrate_add_src(struct css_set *src_cset,
2496 struct cgroup *dst_cgrp,
2497 struct cgroup_mgctx *mgctx)
2499 struct cgroup *src_cgrp;
2501 lockdep_assert_held(&cgroup_mutex);
2502 lockdep_assert_held(&css_set_lock);
2505 * If ->dead, @src_set is associated with one or more dead cgroups
2506 * and doesn't contain any migratable tasks. Ignore it early so
2507 * that the rest of migration path doesn't get confused by it.
2512 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2514 if (!list_empty(&src_cset->mg_preload_node))
2517 WARN_ON(src_cset->mg_src_cgrp);
2518 WARN_ON(src_cset->mg_dst_cgrp);
2519 WARN_ON(!list_empty(&src_cset->mg_tasks));
2520 WARN_ON(!list_empty(&src_cset->mg_node));
2522 src_cset->mg_src_cgrp = src_cgrp;
2523 src_cset->mg_dst_cgrp = dst_cgrp;
2524 get_css_set(src_cset);
2525 list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2529 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2530 * @mgctx: migration context
2532 * Tasks are about to be moved and all the source css_sets have been
2533 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2534 * pins all destination css_sets, links each to its source, and append them
2535 * to @mgctx->preloaded_dst_csets.
2537 * This function must be called after cgroup_migrate_add_src() has been
2538 * called on each migration source css_set. After migration is performed
2539 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2542 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2544 struct css_set *src_cset, *tmp_cset;
2546 lockdep_assert_held(&cgroup_mutex);
2548 /* look up the dst cset for each src cset and link it to src */
2549 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2551 struct css_set *dst_cset;
2552 struct cgroup_subsys *ss;
2555 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2559 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2562 * If src cset equals dst, it's noop. Drop the src.
2563 * cgroup_migrate() will skip the cset too. Note that we
2564 * can't handle src == dst as some nodes are used by both.
2566 if (src_cset == dst_cset) {
2567 src_cset->mg_src_cgrp = NULL;
2568 src_cset->mg_dst_cgrp = NULL;
2569 list_del_init(&src_cset->mg_preload_node);
2570 put_css_set(src_cset);
2571 put_css_set(dst_cset);
2575 src_cset->mg_dst_cset = dst_cset;
2577 if (list_empty(&dst_cset->mg_preload_node))
2578 list_add_tail(&dst_cset->mg_preload_node,
2579 &mgctx->preloaded_dst_csets);
2581 put_css_set(dst_cset);
2583 for_each_subsys(ss, ssid)
2584 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2585 mgctx->ss_mask |= 1 << ssid;
2590 cgroup_migrate_finish(mgctx);
2595 * cgroup_migrate - migrate a process or task to a cgroup
2596 * @leader: the leader of the process or the task to migrate
2597 * @threadgroup: whether @leader points to the whole process or a single task
2598 * @mgctx: migration context
2600 * Migrate a process or task denoted by @leader. If migrating a process,
2601 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2602 * responsible for invoking cgroup_migrate_add_src() and
2603 * cgroup_migrate_prepare_dst() on the targets before invoking this
2604 * function and following up with cgroup_migrate_finish().
2606 * As long as a controller's ->can_attach() doesn't fail, this function is
2607 * guaranteed to succeed. This means that, excluding ->can_attach()
2608 * failure, when migrating multiple targets, the success or failure can be
2609 * decided for all targets by invoking group_migrate_prepare_dst() before
2610 * actually starting migrating.
2612 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2613 struct cgroup_mgctx *mgctx)
2615 struct task_struct *task;
2618 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2619 * already PF_EXITING could be freed from underneath us unless we
2620 * take an rcu_read_lock.
2622 spin_lock_irq(&css_set_lock);
2626 cgroup_migrate_add_task(task, mgctx);
2629 } while_each_thread(leader, task);
2631 spin_unlock_irq(&css_set_lock);
2633 return cgroup_migrate_execute(mgctx);
2637 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2638 * @dst_cgrp: the cgroup to attach to
2639 * @leader: the task or the leader of the threadgroup to be attached
2640 * @threadgroup: attach the whole threadgroup?
2642 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2644 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2647 DEFINE_CGROUP_MGCTX(mgctx);
2648 struct task_struct *task;
2651 ret = cgroup_migrate_vet_dst(dst_cgrp);
2655 /* look up all src csets */
2656 spin_lock_irq(&css_set_lock);
2660 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2663 } while_each_thread(leader, task);
2665 spin_unlock_irq(&css_set_lock);
2667 /* prepare dst csets and commit */
2668 ret = cgroup_migrate_prepare_dst(&mgctx);
2670 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2672 cgroup_migrate_finish(&mgctx);
2675 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2680 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2681 __acquires(&cgroup_threadgroup_rwsem)
2683 struct task_struct *tsk;
2686 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2687 return ERR_PTR(-EINVAL);
2689 percpu_down_write(&cgroup_threadgroup_rwsem);
2693 tsk = find_task_by_vpid(pid);
2695 tsk = ERR_PTR(-ESRCH);
2696 goto out_unlock_threadgroup;
2703 tsk = tsk->group_leader;
2706 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2707 * If userland migrates such a kthread to a non-root cgroup, it can
2708 * become trapped in a cpuset, or RT kthread may be born in a
2709 * cgroup with no rt_runtime allocated. Just say no.
2711 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2712 tsk = ERR_PTR(-EINVAL);
2713 goto out_unlock_threadgroup;
2716 get_task_struct(tsk);
2717 goto out_unlock_rcu;
2719 out_unlock_threadgroup:
2720 percpu_up_write(&cgroup_threadgroup_rwsem);
2726 void cgroup_procs_write_finish(struct task_struct *task)
2727 __releases(&cgroup_threadgroup_rwsem)
2729 struct cgroup_subsys *ss;
2732 /* release reference from cgroup_procs_write_start() */
2733 put_task_struct(task);
2735 percpu_up_write(&cgroup_threadgroup_rwsem);
2736 for_each_subsys(ss, ssid)
2737 if (ss->post_attach)
2741 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2743 struct cgroup_subsys *ss;
2744 bool printed = false;
2747 do_each_subsys_mask(ss, ssid, ss_mask) {
2750 seq_printf(seq, "%s", ss->name);
2752 } while_each_subsys_mask();
2754 seq_putc(seq, '\n');
2757 /* show controllers which are enabled from the parent */
2758 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2760 struct cgroup *cgrp = seq_css(seq)->cgroup;
2762 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2766 /* show controllers which are enabled for a given cgroup's children */
2767 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2769 struct cgroup *cgrp = seq_css(seq)->cgroup;
2771 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2776 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2777 * @cgrp: root of the subtree to update csses for
2779 * @cgrp's control masks have changed and its subtree's css associations
2780 * need to be updated accordingly. This function looks up all css_sets
2781 * which are attached to the subtree, creates the matching updated css_sets
2782 * and migrates the tasks to the new ones.
2784 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2786 DEFINE_CGROUP_MGCTX(mgctx);
2787 struct cgroup_subsys_state *d_css;
2788 struct cgroup *dsct;
2789 struct css_set *src_cset;
2792 lockdep_assert_held(&cgroup_mutex);
2794 percpu_down_write(&cgroup_threadgroup_rwsem);
2796 /* look up all csses currently attached to @cgrp's subtree */
2797 spin_lock_irq(&css_set_lock);
2798 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2799 struct cgrp_cset_link *link;
2801 list_for_each_entry(link, &dsct->cset_links, cset_link)
2802 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2804 spin_unlock_irq(&css_set_lock);
2806 /* NULL dst indicates self on default hierarchy */
2807 ret = cgroup_migrate_prepare_dst(&mgctx);
2811 spin_lock_irq(&css_set_lock);
2812 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2813 struct task_struct *task, *ntask;
2815 /* all tasks in src_csets need to be migrated */
2816 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2817 cgroup_migrate_add_task(task, &mgctx);
2819 spin_unlock_irq(&css_set_lock);
2821 ret = cgroup_migrate_execute(&mgctx);
2823 cgroup_migrate_finish(&mgctx);
2824 percpu_up_write(&cgroup_threadgroup_rwsem);
2829 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2830 * @cgrp: root of the target subtree
2832 * Because css offlining is asynchronous, userland may try to re-enable a
2833 * controller while the previous css is still around. This function grabs
2834 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2836 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2837 __acquires(&cgroup_mutex)
2839 struct cgroup *dsct;
2840 struct cgroup_subsys_state *d_css;
2841 struct cgroup_subsys *ss;
2845 mutex_lock(&cgroup_mutex);
2847 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2848 for_each_subsys(ss, ssid) {
2849 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2852 if (!css || !percpu_ref_is_dying(&css->refcnt))
2855 cgroup_get_live(dsct);
2856 prepare_to_wait(&dsct->offline_waitq, &wait,
2857 TASK_UNINTERRUPTIBLE);
2859 mutex_unlock(&cgroup_mutex);
2861 finish_wait(&dsct->offline_waitq, &wait);
2870 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2871 * @cgrp: root of the target subtree
2873 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2874 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2877 static void cgroup_save_control(struct cgroup *cgrp)
2879 struct cgroup *dsct;
2880 struct cgroup_subsys_state *d_css;
2882 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2883 dsct->old_subtree_control = dsct->subtree_control;
2884 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2885 dsct->old_dom_cgrp = dsct->dom_cgrp;
2890 * cgroup_propagate_control - refresh control masks of a subtree
2891 * @cgrp: root of the target subtree
2893 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2894 * ->subtree_control and propagate controller availability through the
2895 * subtree so that descendants don't have unavailable controllers enabled.
2897 static void cgroup_propagate_control(struct cgroup *cgrp)
2899 struct cgroup *dsct;
2900 struct cgroup_subsys_state *d_css;
2902 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2903 dsct->subtree_control &= cgroup_control(dsct);
2904 dsct->subtree_ss_mask =
2905 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2906 cgroup_ss_mask(dsct));
2911 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
2912 * @cgrp: root of the target subtree
2914 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
2915 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2918 static void cgroup_restore_control(struct cgroup *cgrp)
2920 struct cgroup *dsct;
2921 struct cgroup_subsys_state *d_css;
2923 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2924 dsct->subtree_control = dsct->old_subtree_control;
2925 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
2926 dsct->dom_cgrp = dsct->old_dom_cgrp;
2930 static bool css_visible(struct cgroup_subsys_state *css)
2932 struct cgroup_subsys *ss = css->ss;
2933 struct cgroup *cgrp = css->cgroup;
2935 if (cgroup_control(cgrp) & (1 << ss->id))
2937 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
2939 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
2943 * cgroup_apply_control_enable - enable or show csses according to control
2944 * @cgrp: root of the target subtree
2946 * Walk @cgrp's subtree and create new csses or make the existing ones
2947 * visible. A css is created invisible if it's being implicitly enabled
2948 * through dependency. An invisible css is made visible when the userland
2949 * explicitly enables it.
2951 * Returns 0 on success, -errno on failure. On failure, csses which have
2952 * been processed already aren't cleaned up. The caller is responsible for
2953 * cleaning up with cgroup_apply_control_disable().
2955 static int cgroup_apply_control_enable(struct cgroup *cgrp)
2957 struct cgroup *dsct;
2958 struct cgroup_subsys_state *d_css;
2959 struct cgroup_subsys *ss;
2962 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2963 for_each_subsys(ss, ssid) {
2964 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2966 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
2970 css = css_create(dsct, ss);
2972 return PTR_ERR(css);
2975 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
2977 if (css_visible(css)) {
2978 ret = css_populate_dir(css);
2989 * cgroup_apply_control_disable - kill or hide csses according to control
2990 * @cgrp: root of the target subtree
2992 * Walk @cgrp's subtree and kill and hide csses so that they match
2993 * cgroup_ss_mask() and cgroup_visible_mask().
2995 * A css is hidden when the userland requests it to be disabled while other
2996 * subsystems are still depending on it. The css must not actively control
2997 * resources and be in the vanilla state if it's made visible again later.
2998 * Controllers which may be depended upon should provide ->css_reset() for
3001 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3003 struct cgroup *dsct;
3004 struct cgroup_subsys_state *d_css;
3005 struct cgroup_subsys *ss;
3008 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3009 for_each_subsys(ss, ssid) {
3010 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3015 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3018 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3020 } else if (!css_visible(css)) {
3030 * cgroup_apply_control - apply control mask updates to the subtree
3031 * @cgrp: root of the target subtree
3033 * subsystems can be enabled and disabled in a subtree using the following
3036 * 1. Call cgroup_save_control() to stash the current state.
3037 * 2. Update ->subtree_control masks in the subtree as desired.
3038 * 3. Call cgroup_apply_control() to apply the changes.
3039 * 4. Optionally perform other related operations.
3040 * 5. Call cgroup_finalize_control() to finish up.
3042 * This function implements step 3 and propagates the mask changes
3043 * throughout @cgrp's subtree, updates csses accordingly and perform
3044 * process migrations.
3046 static int cgroup_apply_control(struct cgroup *cgrp)
3050 cgroup_propagate_control(cgrp);
3052 ret = cgroup_apply_control_enable(cgrp);
3057 * At this point, cgroup_e_css() results reflect the new csses
3058 * making the following cgroup_update_dfl_csses() properly update
3059 * css associations of all tasks in the subtree.
3061 ret = cgroup_update_dfl_csses(cgrp);
3069 * cgroup_finalize_control - finalize control mask update
3070 * @cgrp: root of the target subtree
3071 * @ret: the result of the update
3073 * Finalize control mask update. See cgroup_apply_control() for more info.
3075 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3078 cgroup_restore_control(cgrp);
3079 cgroup_propagate_control(cgrp);
3082 cgroup_apply_control_disable(cgrp);
3085 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3087 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3089 /* if nothing is getting enabled, nothing to worry about */
3093 /* can @cgrp host any resources? */
3094 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3097 /* mixables don't care */
3098 if (cgroup_is_mixable(cgrp))
3101 if (domain_enable) {
3102 /* can't enable domain controllers inside a thread subtree */
3103 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3107 * Threaded controllers can handle internal competitions
3108 * and are always allowed inside a (prospective) thread
3111 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3116 * Controllers can't be enabled for a cgroup with tasks to avoid
3117 * child cgroups competing against tasks.
3119 if (cgroup_has_tasks(cgrp))
3125 /* change the enabled child controllers for a cgroup in the default hierarchy */
3126 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3127 char *buf, size_t nbytes,
3130 u16 enable = 0, disable = 0;
3131 struct cgroup *cgrp, *child;
3132 struct cgroup_subsys *ss;
3137 * Parse input - space separated list of subsystem names prefixed
3138 * with either + or -.
3140 buf = strstrip(buf);
3141 while ((tok = strsep(&buf, " "))) {
3144 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3145 if (!cgroup_ssid_enabled(ssid) ||
3146 strcmp(tok + 1, ss->name))
3150 enable |= 1 << ssid;
3151 disable &= ~(1 << ssid);
3152 } else if (*tok == '-') {
3153 disable |= 1 << ssid;
3154 enable &= ~(1 << ssid);
3159 } while_each_subsys_mask();
3160 if (ssid == CGROUP_SUBSYS_COUNT)
3164 cgrp = cgroup_kn_lock_live(of->kn, true);
3168 for_each_subsys(ss, ssid) {
3169 if (enable & (1 << ssid)) {
3170 if (cgrp->subtree_control & (1 << ssid)) {
3171 enable &= ~(1 << ssid);
3175 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3179 } else if (disable & (1 << ssid)) {
3180 if (!(cgrp->subtree_control & (1 << ssid))) {
3181 disable &= ~(1 << ssid);
3185 /* a child has it enabled? */
3186 cgroup_for_each_live_child(child, cgrp) {
3187 if (child->subtree_control & (1 << ssid)) {
3195 if (!enable && !disable) {
3200 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3204 /* save and update control masks and prepare csses */
3205 cgroup_save_control(cgrp);
3207 cgrp->subtree_control |= enable;
3208 cgrp->subtree_control &= ~disable;
3210 ret = cgroup_apply_control(cgrp);
3211 cgroup_finalize_control(cgrp, ret);
3215 kernfs_activate(cgrp->kn);
3217 cgroup_kn_unlock(of->kn);
3218 return ret ?: nbytes;
3222 * cgroup_enable_threaded - make @cgrp threaded
3223 * @cgrp: the target cgroup
3225 * Called when "threaded" is written to the cgroup.type interface file and
3226 * tries to make @cgrp threaded and join the parent's resource domain.
3227 * This function is never called on the root cgroup as cgroup.type doesn't
3230 static int cgroup_enable_threaded(struct cgroup *cgrp)
3232 struct cgroup *parent = cgroup_parent(cgrp);
3233 struct cgroup *dom_cgrp = parent->dom_cgrp;
3234 struct cgroup *dsct;
3235 struct cgroup_subsys_state *d_css;
3238 lockdep_assert_held(&cgroup_mutex);
3240 /* noop if already threaded */
3241 if (cgroup_is_threaded(cgrp))
3245 * If @cgroup is populated or has domain controllers enabled, it
3246 * can't be switched. While the below cgroup_can_be_thread_root()
3247 * test can catch the same conditions, that's only when @parent is
3248 * not mixable, so let's check it explicitly.
3250 if (cgroup_is_populated(cgrp) ||
3251 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3254 /* we're joining the parent's domain, ensure its validity */
3255 if (!cgroup_is_valid_domain(dom_cgrp) ||
3256 !cgroup_can_be_thread_root(dom_cgrp))
3260 * The following shouldn't cause actual migrations and should
3263 cgroup_save_control(cgrp);
3265 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3266 if (dsct == cgrp || cgroup_is_threaded(dsct))
3267 dsct->dom_cgrp = dom_cgrp;
3269 ret = cgroup_apply_control(cgrp);
3271 parent->nr_threaded_children++;
3273 cgroup_finalize_control(cgrp, ret);
3277 static int cgroup_type_show(struct seq_file *seq, void *v)
3279 struct cgroup *cgrp = seq_css(seq)->cgroup;
3281 if (cgroup_is_threaded(cgrp))
3282 seq_puts(seq, "threaded\n");
3283 else if (!cgroup_is_valid_domain(cgrp))
3284 seq_puts(seq, "domain invalid\n");
3285 else if (cgroup_is_thread_root(cgrp))
3286 seq_puts(seq, "domain threaded\n");
3288 seq_puts(seq, "domain\n");
3293 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3294 size_t nbytes, loff_t off)
3296 struct cgroup *cgrp;
3299 /* only switching to threaded mode is supported */
3300 if (strcmp(strstrip(buf), "threaded"))
3303 /* drain dying csses before we re-apply (threaded) subtree control */
3304 cgrp = cgroup_kn_lock_live(of->kn, true);
3308 /* threaded can only be enabled */
3309 ret = cgroup_enable_threaded(cgrp);
3311 cgroup_kn_unlock(of->kn);
3312 return ret ?: nbytes;
3315 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3317 struct cgroup *cgrp = seq_css(seq)->cgroup;
3318 int descendants = READ_ONCE(cgrp->max_descendants);
3320 if (descendants == INT_MAX)
3321 seq_puts(seq, "max\n");
3323 seq_printf(seq, "%d\n", descendants);
3328 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3329 char *buf, size_t nbytes, loff_t off)
3331 struct cgroup *cgrp;
3335 buf = strstrip(buf);
3336 if (!strcmp(buf, "max")) {
3337 descendants = INT_MAX;
3339 ret = kstrtoint(buf, 0, &descendants);
3344 if (descendants < 0)
3347 cgrp = cgroup_kn_lock_live(of->kn, false);
3351 cgrp->max_descendants = descendants;
3353 cgroup_kn_unlock(of->kn);
3358 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3360 struct cgroup *cgrp = seq_css(seq)->cgroup;
3361 int depth = READ_ONCE(cgrp->max_depth);
3363 if (depth == INT_MAX)
3364 seq_puts(seq, "max\n");
3366 seq_printf(seq, "%d\n", depth);
3371 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3372 char *buf, size_t nbytes, loff_t off)
3374 struct cgroup *cgrp;
3378 buf = strstrip(buf);
3379 if (!strcmp(buf, "max")) {
3382 ret = kstrtoint(buf, 0, &depth);
3390 cgrp = cgroup_kn_lock_live(of->kn, false);
3394 cgrp->max_depth = depth;
3396 cgroup_kn_unlock(of->kn);
3401 static int cgroup_events_show(struct seq_file *seq, void *v)
3403 seq_printf(seq, "populated %d\n",
3404 cgroup_is_populated(seq_css(seq)->cgroup));
3408 static int cgroup_stat_show(struct seq_file *seq, void *v)
3410 struct cgroup *cgroup = seq_css(seq)->cgroup;
3412 seq_printf(seq, "nr_descendants %d\n",
3413 cgroup->nr_descendants);
3414 seq_printf(seq, "nr_dying_descendants %d\n",
3415 cgroup->nr_dying_descendants);
3420 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3421 struct cgroup *cgrp, int ssid)
3423 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3424 struct cgroup_subsys_state *css;
3427 if (!ss->css_extra_stat_show)
3430 css = cgroup_tryget_css(cgrp, ss);
3434 ret = ss->css_extra_stat_show(seq, css);
3439 static int cpu_stat_show(struct seq_file *seq, void *v)
3441 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3444 cgroup_base_stat_cputime_show(seq);
3445 #ifdef CONFIG_CGROUP_SCHED
3446 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3451 static int cgroup_file_open(struct kernfs_open_file *of)
3453 struct cftype *cft = of->kn->priv;
3454 struct cgroup_file_ctx *ctx;
3457 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
3461 ctx->ns = current->nsproxy->cgroup_ns;
3462 get_cgroup_ns(ctx->ns);
3468 ret = cft->open(of);
3470 put_cgroup_ns(ctx->ns);
3476 static void cgroup_file_release(struct kernfs_open_file *of)
3478 struct cftype *cft = of->kn->priv;
3479 struct cgroup_file_ctx *ctx = of->priv;
3483 put_cgroup_ns(ctx->ns);
3487 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3488 size_t nbytes, loff_t off)
3490 struct cgroup_file_ctx *ctx = of->priv;
3491 struct cgroup *cgrp = of->kn->parent->priv;
3492 struct cftype *cft = of->kn->priv;
3493 struct cgroup_subsys_state *css;
3497 * If namespaces are delegation boundaries, disallow writes to
3498 * files in an non-init namespace root from inside the namespace
3499 * except for the files explicitly marked delegatable -
3500 * cgroup.procs and cgroup.subtree_control.
3502 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3503 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3504 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
3508 return cft->write(of, buf, nbytes, off);
3511 * kernfs guarantees that a file isn't deleted with operations in
3512 * flight, which means that the matching css is and stays alive and
3513 * doesn't need to be pinned. The RCU locking is not necessary
3514 * either. It's just for the convenience of using cgroup_css().
3517 css = cgroup_css(cgrp, cft->ss);
3520 if (cft->write_u64) {
3521 unsigned long long v;
3522 ret = kstrtoull(buf, 0, &v);
3524 ret = cft->write_u64(css, cft, v);
3525 } else if (cft->write_s64) {
3527 ret = kstrtoll(buf, 0, &v);
3529 ret = cft->write_s64(css, cft, v);
3534 return ret ?: nbytes;
3537 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3539 return seq_cft(seq)->seq_start(seq, ppos);
3542 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3544 return seq_cft(seq)->seq_next(seq, v, ppos);
3547 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3549 if (seq_cft(seq)->seq_stop)
3550 seq_cft(seq)->seq_stop(seq, v);
3553 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3555 struct cftype *cft = seq_cft(m);
3556 struct cgroup_subsys_state *css = seq_css(m);
3559 return cft->seq_show(m, arg);
3562 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3563 else if (cft->read_s64)
3564 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3570 static struct kernfs_ops cgroup_kf_single_ops = {
3571 .atomic_write_len = PAGE_SIZE,
3572 .open = cgroup_file_open,
3573 .release = cgroup_file_release,
3574 .write = cgroup_file_write,
3575 .seq_show = cgroup_seqfile_show,
3578 static struct kernfs_ops cgroup_kf_ops = {
3579 .atomic_write_len = PAGE_SIZE,
3580 .open = cgroup_file_open,
3581 .release = cgroup_file_release,
3582 .write = cgroup_file_write,
3583 .seq_start = cgroup_seqfile_start,
3584 .seq_next = cgroup_seqfile_next,
3585 .seq_stop = cgroup_seqfile_stop,
3586 .seq_show = cgroup_seqfile_show,
3589 /* set uid and gid of cgroup dirs and files to that of the creator */
3590 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3592 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3593 .ia_uid = current_fsuid(),
3594 .ia_gid = current_fsgid(), };
3596 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3597 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3600 return kernfs_setattr(kn, &iattr);
3603 static void cgroup_file_notify_timer(struct timer_list *timer)
3605 cgroup_file_notify(container_of(timer, struct cgroup_file,
3609 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3612 char name[CGROUP_FILE_NAME_MAX];
3613 struct kernfs_node *kn;
3614 struct lock_class_key *key = NULL;
3617 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3618 key = &cft->lockdep_key;
3620 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3621 cgroup_file_mode(cft),
3622 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3623 0, cft->kf_ops, cft,
3628 ret = cgroup_kn_set_ugid(kn);
3634 if (cft->file_offset) {
3635 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3637 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3639 spin_lock_irq(&cgroup_file_kn_lock);
3641 spin_unlock_irq(&cgroup_file_kn_lock);
3648 * cgroup_addrm_files - add or remove files to a cgroup directory
3649 * @css: the target css
3650 * @cgrp: the target cgroup (usually css->cgroup)
3651 * @cfts: array of cftypes to be added
3652 * @is_add: whether to add or remove
3654 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3655 * For removals, this function never fails.
3657 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3658 struct cgroup *cgrp, struct cftype cfts[],
3661 struct cftype *cft, *cft_end = NULL;
3664 lockdep_assert_held(&cgroup_mutex);
3667 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3668 /* does cft->flags tell us to skip this file on @cgrp? */
3669 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3671 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3673 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3675 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3679 ret = cgroup_add_file(css, cgrp, cft);
3681 pr_warn("%s: failed to add %s, err=%d\n",
3682 __func__, cft->name, ret);
3688 cgroup_rm_file(cgrp, cft);
3694 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3696 struct cgroup_subsys *ss = cfts[0].ss;
3697 struct cgroup *root = &ss->root->cgrp;
3698 struct cgroup_subsys_state *css;
3701 lockdep_assert_held(&cgroup_mutex);
3703 /* add/rm files for all cgroups created before */
3704 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3705 struct cgroup *cgrp = css->cgroup;
3707 if (!(css->flags & CSS_VISIBLE))
3710 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3716 kernfs_activate(root->kn);
3720 static void cgroup_exit_cftypes(struct cftype *cfts)
3724 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3725 /* free copy for custom atomic_write_len, see init_cftypes() */
3726 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3731 /* revert flags set by cgroup core while adding @cfts */
3732 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3736 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3740 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3741 struct kernfs_ops *kf_ops;
3743 WARN_ON(cft->ss || cft->kf_ops);
3746 kf_ops = &cgroup_kf_ops;
3748 kf_ops = &cgroup_kf_single_ops;
3751 * Ugh... if @cft wants a custom max_write_len, we need to
3752 * make a copy of kf_ops to set its atomic_write_len.
3754 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3755 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3757 cgroup_exit_cftypes(cfts);
3760 kf_ops->atomic_write_len = cft->max_write_len;
3763 cft->kf_ops = kf_ops;
3770 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3772 lockdep_assert_held(&cgroup_mutex);
3774 if (!cfts || !cfts[0].ss)
3777 list_del(&cfts->node);
3778 cgroup_apply_cftypes(cfts, false);
3779 cgroup_exit_cftypes(cfts);
3784 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3785 * @cfts: zero-length name terminated array of cftypes
3787 * Unregister @cfts. Files described by @cfts are removed from all
3788 * existing cgroups and all future cgroups won't have them either. This
3789 * function can be called anytime whether @cfts' subsys is attached or not.
3791 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3794 int cgroup_rm_cftypes(struct cftype *cfts)
3798 mutex_lock(&cgroup_mutex);
3799 ret = cgroup_rm_cftypes_locked(cfts);
3800 mutex_unlock(&cgroup_mutex);
3805 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3806 * @ss: target cgroup subsystem
3807 * @cfts: zero-length name terminated array of cftypes
3809 * Register @cfts to @ss. Files described by @cfts are created for all
3810 * existing cgroups to which @ss is attached and all future cgroups will
3811 * have them too. This function can be called anytime whether @ss is
3814 * Returns 0 on successful registration, -errno on failure. Note that this
3815 * function currently returns 0 as long as @cfts registration is successful
3816 * even if some file creation attempts on existing cgroups fail.
3818 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3822 if (!cgroup_ssid_enabled(ss->id))
3825 if (!cfts || cfts[0].name[0] == '\0')
3828 ret = cgroup_init_cftypes(ss, cfts);
3832 mutex_lock(&cgroup_mutex);
3834 list_add_tail(&cfts->node, &ss->cfts);
3835 ret = cgroup_apply_cftypes(cfts, true);
3837 cgroup_rm_cftypes_locked(cfts);
3839 mutex_unlock(&cgroup_mutex);
3844 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3845 * @ss: target cgroup subsystem
3846 * @cfts: zero-length name terminated array of cftypes
3848 * Similar to cgroup_add_cftypes() but the added files are only used for
3849 * the default hierarchy.
3851 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3855 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3856 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3857 return cgroup_add_cftypes(ss, cfts);
3861 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3862 * @ss: target cgroup subsystem
3863 * @cfts: zero-length name terminated array of cftypes
3865 * Similar to cgroup_add_cftypes() but the added files are only used for
3866 * the legacy hierarchies.
3868 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3872 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3873 cft->flags |= __CFTYPE_NOT_ON_DFL;
3874 return cgroup_add_cftypes(ss, cfts);
3878 * cgroup_file_notify - generate a file modified event for a cgroup_file
3879 * @cfile: target cgroup_file
3881 * @cfile must have been obtained by setting cftype->file_offset.
3883 void cgroup_file_notify(struct cgroup_file *cfile)
3885 unsigned long flags;
3887 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3889 unsigned long last = cfile->notified_at;
3890 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
3892 if (time_in_range(jiffies, last, next)) {
3893 timer_reduce(&cfile->notify_timer, next);
3895 kernfs_notify(cfile->kn);
3896 cfile->notified_at = jiffies;
3899 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3903 * css_next_child - find the next child of a given css
3904 * @pos: the current position (%NULL to initiate traversal)
3905 * @parent: css whose children to walk
3907 * This function returns the next child of @parent and should be called
3908 * under either cgroup_mutex or RCU read lock. The only requirement is
3909 * that @parent and @pos are accessible. The next sibling is guaranteed to
3910 * be returned regardless of their states.
3912 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3913 * css which finished ->css_online() is guaranteed to be visible in the
3914 * future iterations and will stay visible until the last reference is put.
3915 * A css which hasn't finished ->css_online() or already finished
3916 * ->css_offline() may show up during traversal. It's each subsystem's
3917 * responsibility to synchronize against on/offlining.
3919 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3920 struct cgroup_subsys_state *parent)
3922 struct cgroup_subsys_state *next;
3924 cgroup_assert_mutex_or_rcu_locked();
3927 * @pos could already have been unlinked from the sibling list.
3928 * Once a cgroup is removed, its ->sibling.next is no longer
3929 * updated when its next sibling changes. CSS_RELEASED is set when
3930 * @pos is taken off list, at which time its next pointer is valid,
3931 * and, as releases are serialized, the one pointed to by the next
3932 * pointer is guaranteed to not have started release yet. This
3933 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3934 * critical section, the one pointed to by its next pointer is
3935 * guaranteed to not have finished its RCU grace period even if we
3936 * have dropped rcu_read_lock() inbetween iterations.
3938 * If @pos has CSS_RELEASED set, its next pointer can't be
3939 * dereferenced; however, as each css is given a monotonically
3940 * increasing unique serial number and always appended to the
3941 * sibling list, the next one can be found by walking the parent's
3942 * children until the first css with higher serial number than
3943 * @pos's. While this path can be slower, it happens iff iteration
3944 * races against release and the race window is very small.
3947 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3948 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3949 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3951 list_for_each_entry_rcu(next, &parent->children, sibling)
3952 if (next->serial_nr > pos->serial_nr)
3957 * @next, if not pointing to the head, can be dereferenced and is
3960 if (&next->sibling != &parent->children)
3966 * css_next_descendant_pre - find the next descendant for pre-order walk
3967 * @pos: the current position (%NULL to initiate traversal)
3968 * @root: css whose descendants to walk
3970 * To be used by css_for_each_descendant_pre(). Find the next descendant
3971 * to visit for pre-order traversal of @root's descendants. @root is
3972 * included in the iteration and the first node to be visited.
3974 * While this function requires cgroup_mutex or RCU read locking, it
3975 * doesn't require the whole traversal to be contained in a single critical
3976 * section. This function will return the correct next descendant as long
3977 * as both @pos and @root are accessible and @pos is a descendant of @root.
3979 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3980 * css which finished ->css_online() is guaranteed to be visible in the
3981 * future iterations and will stay visible until the last reference is put.
3982 * A css which hasn't finished ->css_online() or already finished
3983 * ->css_offline() may show up during traversal. It's each subsystem's
3984 * responsibility to synchronize against on/offlining.
3986 struct cgroup_subsys_state *
3987 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3988 struct cgroup_subsys_state *root)
3990 struct cgroup_subsys_state *next;
3992 cgroup_assert_mutex_or_rcu_locked();
3994 /* if first iteration, visit @root */
3998 /* visit the first child if exists */
3999 next = css_next_child(NULL, pos);
4003 /* no child, visit my or the closest ancestor's next sibling */
4004 while (pos != root) {
4005 next = css_next_child(pos, pos->parent);
4015 * css_rightmost_descendant - return the rightmost descendant of a css
4016 * @pos: css of interest
4018 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4019 * is returned. This can be used during pre-order traversal to skip
4022 * While this function requires cgroup_mutex or RCU read locking, it
4023 * doesn't require the whole traversal to be contained in a single critical
4024 * section. This function will return the correct rightmost descendant as
4025 * long as @pos is accessible.
4027 struct cgroup_subsys_state *
4028 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4030 struct cgroup_subsys_state *last, *tmp;
4032 cgroup_assert_mutex_or_rcu_locked();
4036 /* ->prev isn't RCU safe, walk ->next till the end */
4038 css_for_each_child(tmp, last)
4045 static struct cgroup_subsys_state *
4046 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4048 struct cgroup_subsys_state *last;
4052 pos = css_next_child(NULL, pos);
4059 * css_next_descendant_post - find the next descendant for post-order walk
4060 * @pos: the current position (%NULL to initiate traversal)
4061 * @root: css whose descendants to walk
4063 * To be used by css_for_each_descendant_post(). Find the next descendant
4064 * to visit for post-order traversal of @root's descendants. @root is
4065 * included in the iteration and the last node to be visited.
4067 * While this function requires cgroup_mutex or RCU read locking, it
4068 * doesn't require the whole traversal to be contained in a single critical
4069 * section. This function will return the correct next descendant as long
4070 * as both @pos and @cgroup are accessible and @pos is a descendant of
4073 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4074 * css which finished ->css_online() is guaranteed to be visible in the
4075 * future iterations and will stay visible until the last reference is put.
4076 * A css which hasn't finished ->css_online() or already finished
4077 * ->css_offline() may show up during traversal. It's each subsystem's
4078 * responsibility to synchronize against on/offlining.
4080 struct cgroup_subsys_state *
4081 css_next_descendant_post(struct cgroup_subsys_state *pos,
4082 struct cgroup_subsys_state *root)
4084 struct cgroup_subsys_state *next;
4086 cgroup_assert_mutex_or_rcu_locked();
4088 /* if first iteration, visit leftmost descendant which may be @root */
4090 return css_leftmost_descendant(root);
4092 /* if we visited @root, we're done */
4096 /* if there's an unvisited sibling, visit its leftmost descendant */
4097 next = css_next_child(pos, pos->parent);
4099 return css_leftmost_descendant(next);
4101 /* no sibling left, visit parent */
4106 * css_has_online_children - does a css have online children
4107 * @css: the target css
4109 * Returns %true if @css has any online children; otherwise, %false. This
4110 * function can be called from any context but the caller is responsible
4111 * for synchronizing against on/offlining as necessary.
4113 bool css_has_online_children(struct cgroup_subsys_state *css)
4115 struct cgroup_subsys_state *child;
4119 css_for_each_child(child, css) {
4120 if (child->flags & CSS_ONLINE) {
4129 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4131 struct list_head *l;
4132 struct cgrp_cset_link *link;
4133 struct css_set *cset;
4135 lockdep_assert_held(&css_set_lock);
4137 /* find the next threaded cset */
4138 if (it->tcset_pos) {
4139 l = it->tcset_pos->next;
4141 if (l != it->tcset_head) {
4143 return container_of(l, struct css_set,
4144 threaded_csets_node);
4147 it->tcset_pos = NULL;
4150 /* find the next cset */
4153 if (l == it->cset_head) {
4154 it->cset_pos = NULL;
4159 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4161 link = list_entry(l, struct cgrp_cset_link, cset_link);
4167 /* initialize threaded css_set walking */
4168 if (it->flags & CSS_TASK_ITER_THREADED) {
4170 put_css_set_locked(it->cur_dcset);
4171 it->cur_dcset = cset;
4174 it->tcset_head = &cset->threaded_csets;
4175 it->tcset_pos = &cset->threaded_csets;
4182 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4183 * @it: the iterator to advance
4185 * Advance @it to the next css_set to walk.
4187 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4189 struct css_set *cset;
4191 lockdep_assert_held(&css_set_lock);
4193 /* Advance to the next non-empty css_set */
4195 cset = css_task_iter_next_css_set(it);
4197 it->task_pos = NULL;
4200 } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
4202 if (!list_empty(&cset->tasks)) {
4203 it->task_pos = cset->tasks.next;
4204 it->cur_tasks_head = &cset->tasks;
4205 } else if (!list_empty(&cset->mg_tasks)) {
4206 it->task_pos = cset->mg_tasks.next;
4207 it->cur_tasks_head = &cset->mg_tasks;
4209 it->task_pos = cset->dying_tasks.next;
4210 it->cur_tasks_head = &cset->dying_tasks;
4213 it->tasks_head = &cset->tasks;
4214 it->mg_tasks_head = &cset->mg_tasks;
4215 it->dying_tasks_head = &cset->dying_tasks;
4218 * We don't keep css_sets locked across iteration steps and thus
4219 * need to take steps to ensure that iteration can be resumed after
4220 * the lock is re-acquired. Iteration is performed at two levels -
4221 * css_sets and tasks in them.
4223 * Once created, a css_set never leaves its cgroup lists, so a
4224 * pinned css_set is guaranteed to stay put and we can resume
4225 * iteration afterwards.
4227 * Tasks may leave @cset across iteration steps. This is resolved
4228 * by registering each iterator with the css_set currently being
4229 * walked and making css_set_move_task() advance iterators whose
4230 * next task is leaving.
4233 list_del(&it->iters_node);
4234 put_css_set_locked(it->cur_cset);
4237 it->cur_cset = cset;
4238 list_add(&it->iters_node, &cset->task_iters);
4241 static void css_task_iter_skip(struct css_task_iter *it,
4242 struct task_struct *task)
4244 lockdep_assert_held(&css_set_lock);
4246 if (it->task_pos == &task->cg_list) {
4247 it->task_pos = it->task_pos->next;
4248 it->flags |= CSS_TASK_ITER_SKIPPED;
4252 static void css_task_iter_advance(struct css_task_iter *it)
4254 struct task_struct *task;
4256 lockdep_assert_held(&css_set_lock);
4260 * Advance iterator to find next entry. cset->tasks is
4261 * consumed first and then ->mg_tasks. After ->mg_tasks,
4262 * we move onto the next cset.
4264 if (it->flags & CSS_TASK_ITER_SKIPPED)
4265 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4267 it->task_pos = it->task_pos->next;
4269 if (it->task_pos == it->tasks_head) {
4270 it->task_pos = it->mg_tasks_head->next;
4271 it->cur_tasks_head = it->mg_tasks_head;
4273 if (it->task_pos == it->mg_tasks_head) {
4274 it->task_pos = it->dying_tasks_head->next;
4275 it->cur_tasks_head = it->dying_tasks_head;
4277 if (it->task_pos == it->dying_tasks_head)
4278 css_task_iter_advance_css_set(it);
4280 /* called from start, proceed to the first cset */
4281 css_task_iter_advance_css_set(it);
4287 task = list_entry(it->task_pos, struct task_struct, cg_list);
4289 if (it->flags & CSS_TASK_ITER_PROCS) {
4290 /* if PROCS, skip over tasks which aren't group leaders */
4291 if (!thread_group_leader(task))
4294 /* and dying leaders w/o live member threads */
4295 if (it->cur_tasks_head == it->dying_tasks_head &&
4296 !atomic_read(&task->signal->live))
4299 /* skip all dying ones */
4300 if (it->cur_tasks_head == it->dying_tasks_head)
4306 * css_task_iter_start - initiate task iteration
4307 * @css: the css to walk tasks of
4308 * @flags: CSS_TASK_ITER_* flags
4309 * @it: the task iterator to use
4311 * Initiate iteration through the tasks of @css. The caller can call
4312 * css_task_iter_next() to walk through the tasks until the function
4313 * returns NULL. On completion of iteration, css_task_iter_end() must be
4316 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4317 struct css_task_iter *it)
4319 /* no one should try to iterate before mounting cgroups */
4320 WARN_ON_ONCE(!use_task_css_set_links);
4322 memset(it, 0, sizeof(*it));
4324 spin_lock_irq(&css_set_lock);
4330 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4332 it->cset_pos = &css->cgroup->cset_links;
4334 it->cset_head = it->cset_pos;
4336 css_task_iter_advance(it);
4338 spin_unlock_irq(&css_set_lock);
4342 * css_task_iter_next - return the next task for the iterator
4343 * @it: the task iterator being iterated
4345 * The "next" function for task iteration. @it should have been
4346 * initialized via css_task_iter_start(). Returns NULL when the iteration
4349 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4352 put_task_struct(it->cur_task);
4353 it->cur_task = NULL;
4356 spin_lock_irq(&css_set_lock);
4358 /* @it may be half-advanced by skips, finish advancing */
4359 if (it->flags & CSS_TASK_ITER_SKIPPED)
4360 css_task_iter_advance(it);
4363 it->cur_task = list_entry(it->task_pos, struct task_struct,
4365 get_task_struct(it->cur_task);
4366 css_task_iter_advance(it);
4369 spin_unlock_irq(&css_set_lock);
4371 return it->cur_task;
4375 * css_task_iter_end - finish task iteration
4376 * @it: the task iterator to finish
4378 * Finish task iteration started by css_task_iter_start().
4380 void css_task_iter_end(struct css_task_iter *it)
4383 spin_lock_irq(&css_set_lock);
4384 list_del(&it->iters_node);
4385 put_css_set_locked(it->cur_cset);
4386 spin_unlock_irq(&css_set_lock);
4390 put_css_set(it->cur_dcset);
4393 put_task_struct(it->cur_task);
4396 static void cgroup_procs_release(struct kernfs_open_file *of)
4398 struct cgroup_file_ctx *ctx = of->priv;
4400 if (ctx->procs.started)
4401 css_task_iter_end(&ctx->procs.iter);
4404 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4406 struct kernfs_open_file *of = s->private;
4407 struct cgroup_file_ctx *ctx = of->priv;
4412 return css_task_iter_next(&ctx->procs.iter);
4415 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4416 unsigned int iter_flags)
4418 struct kernfs_open_file *of = s->private;
4419 struct cgroup *cgrp = seq_css(s)->cgroup;
4420 struct cgroup_file_ctx *ctx = of->priv;
4421 struct css_task_iter *it = &ctx->procs.iter;
4424 * When a seq_file is seeked, it's always traversed sequentially
4425 * from position 0, so we can simply keep iterating on !0 *pos.
4427 if (!ctx->procs.started) {
4428 if (WARN_ON_ONCE((*pos)))
4429 return ERR_PTR(-EINVAL);
4430 css_task_iter_start(&cgrp->self, iter_flags, it);
4431 ctx->procs.started = true;
4432 } else if (!(*pos)) {
4433 css_task_iter_end(it);
4434 css_task_iter_start(&cgrp->self, iter_flags, it);
4436 return it->cur_task;
4438 return cgroup_procs_next(s, NULL, NULL);
4441 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4443 struct cgroup *cgrp = seq_css(s)->cgroup;
4446 * All processes of a threaded subtree belong to the domain cgroup
4447 * of the subtree. Only threads can be distributed across the
4448 * subtree. Reject reads on cgroup.procs in the subtree proper.
4449 * They're always empty anyway.
4451 if (cgroup_is_threaded(cgrp))
4452 return ERR_PTR(-EOPNOTSUPP);
4454 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4455 CSS_TASK_ITER_THREADED);
4458 static int cgroup_procs_show(struct seq_file *s, void *v)
4460 seq_printf(s, "%d\n", task_pid_vnr(v));
4464 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4465 struct cgroup *dst_cgrp,
4466 struct super_block *sb,
4467 struct cgroup_namespace *ns)
4469 struct cgroup *com_cgrp = src_cgrp;
4470 struct inode *inode;
4473 lockdep_assert_held(&cgroup_mutex);
4475 /* find the common ancestor */
4476 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4477 com_cgrp = cgroup_parent(com_cgrp);
4479 /* %current should be authorized to migrate to the common ancestor */
4480 inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4484 ret = inode_permission(inode, MAY_WRITE);
4490 * If namespaces are delegation boundaries, %current must be able
4491 * to see both source and destination cgroups from its namespace.
4493 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4494 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4495 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4501 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4502 char *buf, size_t nbytes, loff_t off)
4504 struct cgroup_file_ctx *ctx = of->priv;
4505 struct cgroup *src_cgrp, *dst_cgrp;
4506 struct task_struct *task;
4507 const struct cred *saved_cred;
4510 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4514 task = cgroup_procs_write_start(buf, true);
4515 ret = PTR_ERR_OR_ZERO(task);
4519 /* find the source cgroup */
4520 spin_lock_irq(&css_set_lock);
4521 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4522 spin_unlock_irq(&css_set_lock);
4525 * Process and thread migrations follow same delegation rule. Check
4526 * permissions using the credentials from file open to protect against
4527 * inherited fd attacks.
4529 saved_cred = override_creds(of->file->f_cred);
4530 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4531 of->file->f_path.dentry->d_sb,
4533 revert_creds(saved_cred);
4537 ret = cgroup_attach_task(dst_cgrp, task, true);
4540 cgroup_procs_write_finish(task);
4542 cgroup_kn_unlock(of->kn);
4544 return ret ?: nbytes;
4547 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4549 return __cgroup_procs_start(s, pos, 0);
4552 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4553 char *buf, size_t nbytes, loff_t off)
4555 struct cgroup_file_ctx *ctx = of->priv;
4556 struct cgroup *src_cgrp, *dst_cgrp;
4557 struct task_struct *task;
4558 const struct cred *saved_cred;
4561 buf = strstrip(buf);
4563 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4567 task = cgroup_procs_write_start(buf, false);
4568 ret = PTR_ERR_OR_ZERO(task);
4572 /* find the source cgroup */
4573 spin_lock_irq(&css_set_lock);
4574 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4575 spin_unlock_irq(&css_set_lock);
4578 * Process and thread migrations follow same delegation rule. Check
4579 * permissions using the credentials from file open to protect against
4580 * inherited fd attacks.
4582 saved_cred = override_creds(of->file->f_cred);
4583 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4584 of->file->f_path.dentry->d_sb,
4586 revert_creds(saved_cred);
4590 /* and must be contained in the same domain */
4592 if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4595 ret = cgroup_attach_task(dst_cgrp, task, false);
4598 cgroup_procs_write_finish(task);
4600 cgroup_kn_unlock(of->kn);
4602 return ret ?: nbytes;
4605 /* cgroup core interface files for the default hierarchy */
4606 static struct cftype cgroup_base_files[] = {
4608 .name = "cgroup.type",
4609 .flags = CFTYPE_NOT_ON_ROOT,
4610 .seq_show = cgroup_type_show,
4611 .write = cgroup_type_write,
4614 .name = "cgroup.procs",
4615 .flags = CFTYPE_NS_DELEGATABLE,
4616 .file_offset = offsetof(struct cgroup, procs_file),
4617 .release = cgroup_procs_release,
4618 .seq_start = cgroup_procs_start,
4619 .seq_next = cgroup_procs_next,
4620 .seq_show = cgroup_procs_show,
4621 .write = cgroup_procs_write,
4624 .name = "cgroup.threads",
4625 .flags = CFTYPE_NS_DELEGATABLE,
4626 .release = cgroup_procs_release,
4627 .seq_start = cgroup_threads_start,
4628 .seq_next = cgroup_procs_next,
4629 .seq_show = cgroup_procs_show,
4630 .write = cgroup_threads_write,
4633 .name = "cgroup.controllers",
4634 .seq_show = cgroup_controllers_show,
4637 .name = "cgroup.subtree_control",
4638 .flags = CFTYPE_NS_DELEGATABLE,
4639 .seq_show = cgroup_subtree_control_show,
4640 .write = cgroup_subtree_control_write,
4643 .name = "cgroup.events",
4644 .flags = CFTYPE_NOT_ON_ROOT,
4645 .file_offset = offsetof(struct cgroup, events_file),
4646 .seq_show = cgroup_events_show,
4649 .name = "cgroup.max.descendants",
4650 .seq_show = cgroup_max_descendants_show,
4651 .write = cgroup_max_descendants_write,
4654 .name = "cgroup.max.depth",
4655 .seq_show = cgroup_max_depth_show,
4656 .write = cgroup_max_depth_write,
4659 .name = "cgroup.stat",
4660 .seq_show = cgroup_stat_show,
4664 .flags = CFTYPE_NOT_ON_ROOT,
4665 .seq_show = cpu_stat_show,
4671 * css destruction is four-stage process.
4673 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4674 * Implemented in kill_css().
4676 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4677 * and thus css_tryget_online() is guaranteed to fail, the css can be
4678 * offlined by invoking offline_css(). After offlining, the base ref is
4679 * put. Implemented in css_killed_work_fn().
4681 * 3. When the percpu_ref reaches zero, the only possible remaining
4682 * accessors are inside RCU read sections. css_release() schedules the
4685 * 4. After the grace period, the css can be freed. Implemented in
4686 * css_free_work_fn().
4688 * It is actually hairier because both step 2 and 4 require process context
4689 * and thus involve punting to css->destroy_work adding two additional
4690 * steps to the already complex sequence.
4692 static void css_free_rwork_fn(struct work_struct *work)
4694 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
4695 struct cgroup_subsys_state, destroy_rwork);
4696 struct cgroup_subsys *ss = css->ss;
4697 struct cgroup *cgrp = css->cgroup;
4699 percpu_ref_exit(&css->refcnt);
4703 struct cgroup_subsys_state *parent = css->parent;
4707 cgroup_idr_remove(&ss->css_idr, id);
4713 /* cgroup free path */
4714 atomic_dec(&cgrp->root->nr_cgrps);
4715 cgroup1_pidlist_destroy_all(cgrp);
4716 cancel_work_sync(&cgrp->release_agent_work);
4718 if (cgroup_parent(cgrp)) {
4720 * We get a ref to the parent, and put the ref when
4721 * this cgroup is being freed, so it's guaranteed
4722 * that the parent won't be destroyed before its
4725 cgroup_put(cgroup_parent(cgrp));
4726 kernfs_put(cgrp->kn);
4727 if (cgroup_on_dfl(cgrp))
4728 cgroup_rstat_exit(cgrp);
4732 * This is root cgroup's refcnt reaching zero,
4733 * which indicates that the root should be
4736 cgroup_destroy_root(cgrp->root);
4741 static void css_release_work_fn(struct work_struct *work)
4743 struct cgroup_subsys_state *css =
4744 container_of(work, struct cgroup_subsys_state, destroy_work);
4745 struct cgroup_subsys *ss = css->ss;
4746 struct cgroup *cgrp = css->cgroup;
4748 mutex_lock(&cgroup_mutex);
4750 css->flags |= CSS_RELEASED;
4751 list_del_rcu(&css->sibling);
4754 /* css release path */
4755 if (!list_empty(&css->rstat_css_node)) {
4756 cgroup_rstat_flush(cgrp);
4757 list_del_rcu(&css->rstat_css_node);
4760 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4761 if (ss->css_released)
4762 ss->css_released(css);
4764 struct cgroup *tcgrp;
4766 /* cgroup release path */
4767 TRACE_CGROUP_PATH(release, cgrp);
4769 if (cgroup_on_dfl(cgrp))
4770 cgroup_rstat_flush(cgrp);
4772 spin_lock_irq(&css_set_lock);
4773 for (tcgrp = cgroup_parent(cgrp); tcgrp;
4774 tcgrp = cgroup_parent(tcgrp))
4775 tcgrp->nr_dying_descendants--;
4776 spin_unlock_irq(&css_set_lock);
4778 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4782 * There are two control paths which try to determine
4783 * cgroup from dentry without going through kernfs -
4784 * cgroupstats_build() and css_tryget_online_from_dir().
4785 * Those are supported by RCU protecting clearing of
4786 * cgrp->kn->priv backpointer.
4789 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4792 cgroup_bpf_put(cgrp);
4795 mutex_unlock(&cgroup_mutex);
4797 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4798 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4801 static void css_release(struct percpu_ref *ref)
4803 struct cgroup_subsys_state *css =
4804 container_of(ref, struct cgroup_subsys_state, refcnt);
4806 INIT_WORK(&css->destroy_work, css_release_work_fn);
4807 queue_work(cgroup_destroy_wq, &css->destroy_work);
4810 static void init_and_link_css(struct cgroup_subsys_state *css,
4811 struct cgroup_subsys *ss, struct cgroup *cgrp)
4813 lockdep_assert_held(&cgroup_mutex);
4815 cgroup_get_live(cgrp);
4817 memset(css, 0, sizeof(*css));
4821 INIT_LIST_HEAD(&css->sibling);
4822 INIT_LIST_HEAD(&css->children);
4823 INIT_LIST_HEAD(&css->rstat_css_node);
4824 css->serial_nr = css_serial_nr_next++;
4825 atomic_set(&css->online_cnt, 0);
4827 if (cgroup_parent(cgrp)) {
4828 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4829 css_get(css->parent);
4832 if (cgroup_on_dfl(cgrp) && ss->css_rstat_flush)
4833 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
4835 BUG_ON(cgroup_css(cgrp, ss));
4838 /* invoke ->css_online() on a new CSS and mark it online if successful */
4839 static int online_css(struct cgroup_subsys_state *css)
4841 struct cgroup_subsys *ss = css->ss;
4844 lockdep_assert_held(&cgroup_mutex);
4847 ret = ss->css_online(css);
4849 css->flags |= CSS_ONLINE;
4850 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4852 atomic_inc(&css->online_cnt);
4854 atomic_inc(&css->parent->online_cnt);
4859 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4860 static void offline_css(struct cgroup_subsys_state *css)
4862 struct cgroup_subsys *ss = css->ss;
4864 lockdep_assert_held(&cgroup_mutex);
4866 if (!(css->flags & CSS_ONLINE))
4869 if (ss->css_offline)
4870 ss->css_offline(css);
4872 css->flags &= ~CSS_ONLINE;
4873 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4875 wake_up_all(&css->cgroup->offline_waitq);
4879 * css_create - create a cgroup_subsys_state
4880 * @cgrp: the cgroup new css will be associated with
4881 * @ss: the subsys of new css
4883 * Create a new css associated with @cgrp - @ss pair. On success, the new
4884 * css is online and installed in @cgrp. This function doesn't create the
4885 * interface files. Returns 0 on success, -errno on failure.
4887 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4888 struct cgroup_subsys *ss)
4890 struct cgroup *parent = cgroup_parent(cgrp);
4891 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4892 struct cgroup_subsys_state *css;
4895 lockdep_assert_held(&cgroup_mutex);
4897 css = ss->css_alloc(parent_css);
4899 css = ERR_PTR(-ENOMEM);
4903 init_and_link_css(css, ss, cgrp);
4905 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4909 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4914 /* @css is ready to be brought online now, make it visible */
4915 list_add_tail_rcu(&css->sibling, &parent_css->children);
4916 cgroup_idr_replace(&ss->css_idr, css, css->id);
4918 err = online_css(css);
4922 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4923 cgroup_parent(parent)) {
4924 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4925 current->comm, current->pid, ss->name);
4926 if (!strcmp(ss->name, "memory"))
4927 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4928 ss->warned_broken_hierarchy = true;
4934 list_del_rcu(&css->sibling);
4936 list_del_rcu(&css->rstat_css_node);
4937 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
4938 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
4939 return ERR_PTR(err);
4943 * The returned cgroup is fully initialized including its control mask, but
4944 * it isn't associated with its kernfs_node and doesn't have the control
4947 static struct cgroup *cgroup_create(struct cgroup *parent)
4949 struct cgroup_root *root = parent->root;
4950 struct cgroup *cgrp, *tcgrp;
4951 int level = parent->level + 1;
4954 /* allocate the cgroup and its ID, 0 is reserved for the root */
4955 cgrp = kzalloc(struct_size(cgrp, ancestor_ids, (level + 1)),
4958 return ERR_PTR(-ENOMEM);
4960 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4964 if (cgroup_on_dfl(parent)) {
4965 ret = cgroup_rstat_init(cgrp);
4967 goto out_cancel_ref;
4971 * Temporarily set the pointer to NULL, so idr_find() won't return
4972 * a half-baked cgroup.
4974 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4980 init_cgroup_housekeeping(cgrp);
4982 cgrp->self.parent = &parent->self;
4984 cgrp->level = level;
4985 ret = cgroup_bpf_inherit(cgrp);
4989 spin_lock_irq(&css_set_lock);
4990 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
4991 cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4994 tcgrp->nr_descendants++;
4996 spin_unlock_irq(&css_set_lock);
4998 if (notify_on_release(parent))
4999 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5001 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5002 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5004 cgrp->self.serial_nr = css_serial_nr_next++;
5006 /* allocation complete, commit to creation */
5007 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5008 atomic_inc(&root->nr_cgrps);
5009 cgroup_get_live(parent);
5012 * @cgrp is now fully operational. If something fails after this
5013 * point, it'll be released via the normal destruction path.
5015 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
5018 * On the default hierarchy, a child doesn't automatically inherit
5019 * subtree_control from the parent. Each is configured manually.
5021 if (!cgroup_on_dfl(cgrp))
5022 cgrp->subtree_control = cgroup_control(cgrp);
5024 cgroup_propagate_control(cgrp);
5029 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
5031 if (cgroup_on_dfl(parent))
5032 cgroup_rstat_exit(cgrp);
5034 percpu_ref_exit(&cgrp->self.refcnt);
5037 return ERR_PTR(ret);
5040 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5042 struct cgroup *cgroup;
5046 lockdep_assert_held(&cgroup_mutex);
5048 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5049 if (cgroup->nr_descendants >= cgroup->max_descendants)
5052 if (level > cgroup->max_depth)
5063 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5065 struct cgroup *parent, *cgrp;
5066 struct kernfs_node *kn;
5069 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5070 if (strchr(name, '\n'))
5073 parent = cgroup_kn_lock_live(parent_kn, false);
5077 if (!cgroup_check_hierarchy_limits(parent)) {
5082 cgrp = cgroup_create(parent);
5084 ret = PTR_ERR(cgrp);
5088 /* create the directory */
5089 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5097 * This extra ref will be put in cgroup_free_fn() and guarantees
5098 * that @cgrp->kn is always accessible.
5102 ret = cgroup_kn_set_ugid(kn);
5106 ret = css_populate_dir(&cgrp->self);
5110 ret = cgroup_apply_control_enable(cgrp);
5114 TRACE_CGROUP_PATH(mkdir, cgrp);
5116 /* let's create and online css's */
5117 kernfs_activate(kn);
5123 cgroup_destroy_locked(cgrp);
5125 cgroup_kn_unlock(parent_kn);
5130 * This is called when the refcnt of a css is confirmed to be killed.
5131 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5132 * initate destruction and put the css ref from kill_css().
5134 static void css_killed_work_fn(struct work_struct *work)
5136 struct cgroup_subsys_state *css =
5137 container_of(work, struct cgroup_subsys_state, destroy_work);
5139 mutex_lock(&cgroup_mutex);
5144 /* @css can't go away while we're holding cgroup_mutex */
5146 } while (css && atomic_dec_and_test(&css->online_cnt));
5148 mutex_unlock(&cgroup_mutex);
5151 /* css kill confirmation processing requires process context, bounce */
5152 static void css_killed_ref_fn(struct percpu_ref *ref)
5154 struct cgroup_subsys_state *css =
5155 container_of(ref, struct cgroup_subsys_state, refcnt);
5157 if (atomic_dec_and_test(&css->online_cnt)) {
5158 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5159 queue_work(cgroup_destroy_wq, &css->destroy_work);
5164 * kill_css - destroy a css
5165 * @css: css to destroy
5167 * This function initiates destruction of @css by removing cgroup interface
5168 * files and putting its base reference. ->css_offline() will be invoked
5169 * asynchronously once css_tryget_online() is guaranteed to fail and when
5170 * the reference count reaches zero, @css will be released.
5172 static void kill_css(struct cgroup_subsys_state *css)
5174 lockdep_assert_held(&cgroup_mutex);
5176 if (css->flags & CSS_DYING)
5179 css->flags |= CSS_DYING;
5182 * This must happen before css is disassociated with its cgroup.
5183 * See seq_css() for details.
5188 * Killing would put the base ref, but we need to keep it alive
5189 * until after ->css_offline().
5194 * cgroup core guarantees that, by the time ->css_offline() is
5195 * invoked, no new css reference will be given out via
5196 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5197 * proceed to offlining css's because percpu_ref_kill() doesn't
5198 * guarantee that the ref is seen as killed on all CPUs on return.
5200 * Use percpu_ref_kill_and_confirm() to get notifications as each
5201 * css is confirmed to be seen as killed on all CPUs.
5203 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5207 * cgroup_destroy_locked - the first stage of cgroup destruction
5208 * @cgrp: cgroup to be destroyed
5210 * css's make use of percpu refcnts whose killing latency shouldn't be
5211 * exposed to userland and are RCU protected. Also, cgroup core needs to
5212 * guarantee that css_tryget_online() won't succeed by the time
5213 * ->css_offline() is invoked. To satisfy all the requirements,
5214 * destruction is implemented in the following two steps.
5216 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5217 * userland visible parts and start killing the percpu refcnts of
5218 * css's. Set up so that the next stage will be kicked off once all
5219 * the percpu refcnts are confirmed to be killed.
5221 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5222 * rest of destruction. Once all cgroup references are gone, the
5223 * cgroup is RCU-freed.
5225 * This function implements s1. After this step, @cgrp is gone as far as
5226 * the userland is concerned and a new cgroup with the same name may be
5227 * created. As cgroup doesn't care about the names internally, this
5228 * doesn't cause any problem.
5230 static int cgroup_destroy_locked(struct cgroup *cgrp)
5231 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5233 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5234 struct cgroup_subsys_state *css;
5235 struct cgrp_cset_link *link;
5238 lockdep_assert_held(&cgroup_mutex);
5241 * Only migration can raise populated from zero and we're already
5242 * holding cgroup_mutex.
5244 if (cgroup_is_populated(cgrp))
5248 * Make sure there's no live children. We can't test emptiness of
5249 * ->self.children as dead children linger on it while being
5250 * drained; otherwise, "rmdir parent/child parent" may fail.
5252 if (css_has_online_children(&cgrp->self))
5256 * Mark @cgrp and the associated csets dead. The former prevents
5257 * further task migration and child creation by disabling
5258 * cgroup_lock_live_group(). The latter makes the csets ignored by
5259 * the migration path.
5261 cgrp->self.flags &= ~CSS_ONLINE;
5263 spin_lock_irq(&css_set_lock);
5264 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5265 link->cset->dead = true;
5266 spin_unlock_irq(&css_set_lock);
5268 /* initiate massacre of all css's */
5269 for_each_css(css, ssid, cgrp)
5272 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5273 css_clear_dir(&cgrp->self);
5274 kernfs_remove(cgrp->kn);
5276 if (parent && cgroup_is_threaded(cgrp))
5277 parent->nr_threaded_children--;
5279 spin_lock_irq(&css_set_lock);
5280 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5281 tcgrp->nr_descendants--;
5282 tcgrp->nr_dying_descendants++;
5284 spin_unlock_irq(&css_set_lock);
5286 cgroup1_check_for_release(parent);
5288 /* put the base reference */
5289 percpu_ref_kill(&cgrp->self.refcnt);
5294 int cgroup_rmdir(struct kernfs_node *kn)
5296 struct cgroup *cgrp;
5299 cgrp = cgroup_kn_lock_live(kn, false);
5303 ret = cgroup_destroy_locked(cgrp);
5305 TRACE_CGROUP_PATH(rmdir, cgrp);
5307 cgroup_kn_unlock(kn);
5311 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5312 .show_options = cgroup_show_options,
5313 .remount_fs = cgroup_remount,
5314 .mkdir = cgroup_mkdir,
5315 .rmdir = cgroup_rmdir,
5316 .show_path = cgroup_show_path,
5319 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5321 struct cgroup_subsys_state *css;
5323 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5325 mutex_lock(&cgroup_mutex);
5327 idr_init(&ss->css_idr);
5328 INIT_LIST_HEAD(&ss->cfts);
5330 /* Create the root cgroup state for this subsystem */
5331 ss->root = &cgrp_dfl_root;
5332 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5333 /* We don't handle early failures gracefully */
5334 BUG_ON(IS_ERR(css));
5335 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5338 * Root csses are never destroyed and we can't initialize
5339 * percpu_ref during early init. Disable refcnting.
5341 css->flags |= CSS_NO_REF;
5344 /* allocation can't be done safely during early init */
5347 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5348 BUG_ON(css->id < 0);
5351 /* Update the init_css_set to contain a subsys
5352 * pointer to this state - since the subsystem is
5353 * newly registered, all tasks and hence the
5354 * init_css_set is in the subsystem's root cgroup. */
5355 init_css_set.subsys[ss->id] = css;
5357 have_fork_callback |= (bool)ss->fork << ss->id;
5358 have_exit_callback |= (bool)ss->exit << ss->id;
5359 have_release_callback |= (bool)ss->release << ss->id;
5360 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5362 /* At system boot, before all subsystems have been
5363 * registered, no tasks have been forked, so we don't
5364 * need to invoke fork callbacks here. */
5365 BUG_ON(!list_empty(&init_task.tasks));
5367 BUG_ON(online_css(css));
5369 mutex_unlock(&cgroup_mutex);
5373 * cgroup_init_early - cgroup initialization at system boot
5375 * Initialize cgroups at system boot, and initialize any
5376 * subsystems that request early init.
5378 int __init cgroup_init_early(void)
5380 static struct cgroup_sb_opts __initdata opts;
5381 struct cgroup_subsys *ss;
5384 init_cgroup_root(&cgrp_dfl_root, &opts);
5385 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5387 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5389 for_each_subsys(ss, i) {
5390 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5391 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5392 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5394 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5395 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5398 ss->name = cgroup_subsys_name[i];
5399 if (!ss->legacy_name)
5400 ss->legacy_name = cgroup_subsys_name[i];
5403 cgroup_init_subsys(ss, true);
5409 * cgroup_init - cgroup initialization
5411 * Register cgroup filesystem and /proc file, and initialize
5412 * any subsystems that didn't request early init.
5414 int __init cgroup_init(void)
5416 struct cgroup_subsys *ss;
5419 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5420 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5421 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5422 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5424 cgroup_rstat_boot();
5427 * The latency of the synchronize_sched() is too high for cgroups,
5428 * avoid it at the cost of forcing all readers into the slow path.
5430 rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5432 get_user_ns(init_cgroup_ns.user_ns);
5434 mutex_lock(&cgroup_mutex);
5437 * Add init_css_set to the hash table so that dfl_root can link to
5440 hash_add(css_set_table, &init_css_set.hlist,
5441 css_set_hash(init_css_set.subsys));
5443 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
5445 mutex_unlock(&cgroup_mutex);
5447 for_each_subsys(ss, ssid) {
5448 if (ss->early_init) {
5449 struct cgroup_subsys_state *css =
5450 init_css_set.subsys[ss->id];
5452 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5454 BUG_ON(css->id < 0);
5456 cgroup_init_subsys(ss, false);
5459 list_add_tail(&init_css_set.e_cset_node[ssid],
5460 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5463 * Setting dfl_root subsys_mask needs to consider the
5464 * disabled flag and cftype registration needs kmalloc,
5465 * both of which aren't available during early_init.
5467 if (!cgroup_ssid_enabled(ssid))
5470 if (cgroup1_ssid_disabled(ssid))
5471 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5474 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5476 /* implicit controllers must be threaded too */
5477 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5479 if (ss->implicit_on_dfl)
5480 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5481 else if (!ss->dfl_cftypes)
5482 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5485 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5487 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5488 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5490 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5491 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5495 ss->bind(init_css_set.subsys[ssid]);
5497 mutex_lock(&cgroup_mutex);
5498 css_populate_dir(init_css_set.subsys[ssid]);
5499 mutex_unlock(&cgroup_mutex);
5502 /* init_css_set.subsys[] has been updated, re-hash */
5503 hash_del(&init_css_set.hlist);
5504 hash_add(css_set_table, &init_css_set.hlist,
5505 css_set_hash(init_css_set.subsys));
5507 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5508 WARN_ON(register_filesystem(&cgroup_fs_type));
5509 WARN_ON(register_filesystem(&cgroup2_fs_type));
5510 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
5515 static int __init cgroup_wq_init(void)
5518 * There isn't much point in executing destruction path in
5519 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5520 * Use 1 for @max_active.
5522 * We would prefer to do this in cgroup_init() above, but that
5523 * is called before init_workqueues(): so leave this until after.
5525 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5526 BUG_ON(!cgroup_destroy_wq);
5529 core_initcall(cgroup_wq_init);
5531 void cgroup_path_from_kernfs_id(const union kernfs_node_id *id,
5532 char *buf, size_t buflen)
5534 struct kernfs_node *kn;
5536 kn = kernfs_get_node_by_id(cgrp_dfl_root.kf_root, id);
5539 kernfs_path(kn, buf, buflen);
5544 * proc_cgroup_show()
5545 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5546 * - Used for /proc/<pid>/cgroup.
5548 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5549 struct pid *pid, struct task_struct *tsk)
5553 struct cgroup_root *root;
5556 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5560 mutex_lock(&cgroup_mutex);
5561 spin_lock_irq(&css_set_lock);
5563 for_each_root(root) {
5564 struct cgroup_subsys *ss;
5565 struct cgroup *cgrp;
5566 int ssid, count = 0;
5568 if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5571 seq_printf(m, "%d:", root->hierarchy_id);
5572 if (root != &cgrp_dfl_root)
5573 for_each_subsys(ss, ssid)
5574 if (root->subsys_mask & (1 << ssid))
5575 seq_printf(m, "%s%s", count++ ? "," : "",
5577 if (strlen(root->name))
5578 seq_printf(m, "%sname=%s", count ? "," : "",
5582 cgrp = task_cgroup_from_root(tsk, root);
5585 * On traditional hierarchies, all zombie tasks show up as
5586 * belonging to the root cgroup. On the default hierarchy,
5587 * while a zombie doesn't show up in "cgroup.procs" and
5588 * thus can't be migrated, its /proc/PID/cgroup keeps
5589 * reporting the cgroup it belonged to before exiting. If
5590 * the cgroup is removed before the zombie is reaped,
5591 * " (deleted)" is appended to the cgroup path.
5593 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5594 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5595 current->nsproxy->cgroup_ns);
5596 if (retval >= PATH_MAX)
5597 retval = -ENAMETOOLONG;
5606 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5607 seq_puts(m, " (deleted)\n");
5614 spin_unlock_irq(&css_set_lock);
5615 mutex_unlock(&cgroup_mutex);
5622 * cgroup_fork - initialize cgroup related fields during copy_process()
5623 * @child: pointer to task_struct of forking parent process.
5625 * A task is associated with the init_css_set until cgroup_post_fork()
5626 * attaches it to the parent's css_set. Empty cg_list indicates that
5627 * @child isn't holding reference to its css_set.
5629 void cgroup_fork(struct task_struct *child)
5631 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5632 INIT_LIST_HEAD(&child->cg_list);
5636 * cgroup_can_fork - called on a new task before the process is exposed
5637 * @child: the task in question.
5639 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5640 * returns an error, the fork aborts with that error code. This allows for
5641 * a cgroup subsystem to conditionally allow or deny new forks.
5643 int cgroup_can_fork(struct task_struct *child)
5645 struct cgroup_subsys *ss;
5648 do_each_subsys_mask(ss, i, have_canfork_callback) {
5649 ret = ss->can_fork(child);
5652 } while_each_subsys_mask();
5657 for_each_subsys(ss, j) {
5660 if (ss->cancel_fork)
5661 ss->cancel_fork(child);
5668 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5669 * @child: the task in question
5671 * This calls the cancel_fork() callbacks if a fork failed *after*
5672 * cgroup_can_fork() succeded.
5674 void cgroup_cancel_fork(struct task_struct *child)
5676 struct cgroup_subsys *ss;
5679 for_each_subsys(ss, i)
5680 if (ss->cancel_fork)
5681 ss->cancel_fork(child);
5685 * cgroup_post_fork - called on a new task after adding it to the task list
5686 * @child: the task in question
5688 * Adds the task to the list running through its css_set if necessary and
5689 * call the subsystem fork() callbacks. Has to be after the task is
5690 * visible on the task list in case we race with the first call to
5691 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5694 void cgroup_post_fork(struct task_struct *child)
5696 struct cgroup_subsys *ss;
5700 * This may race against cgroup_enable_task_cg_lists(). As that
5701 * function sets use_task_css_set_links before grabbing
5702 * tasklist_lock and we just went through tasklist_lock to add
5703 * @child, it's guaranteed that either we see the set
5704 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5705 * @child during its iteration.
5707 * If we won the race, @child is associated with %current's
5708 * css_set. Grabbing css_set_lock guarantees both that the
5709 * association is stable, and, on completion of the parent's
5710 * migration, @child is visible in the source of migration or
5711 * already in the destination cgroup. This guarantee is necessary
5712 * when implementing operations which need to migrate all tasks of
5713 * a cgroup to another.
5715 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5716 * will remain in init_css_set. This is safe because all tasks are
5717 * in the init_css_set before cg_links is enabled and there's no
5718 * operation which transfers all tasks out of init_css_set.
5720 if (use_task_css_set_links) {
5721 struct css_set *cset;
5723 spin_lock_irq(&css_set_lock);
5724 cset = task_css_set(current);
5725 if (list_empty(&child->cg_list)) {
5728 css_set_move_task(child, NULL, cset, false);
5730 spin_unlock_irq(&css_set_lock);
5734 * Call ss->fork(). This must happen after @child is linked on
5735 * css_set; otherwise, @child might change state between ->fork()
5736 * and addition to css_set.
5738 do_each_subsys_mask(ss, i, have_fork_callback) {
5740 } while_each_subsys_mask();
5744 * cgroup_exit - detach cgroup from exiting task
5745 * @tsk: pointer to task_struct of exiting process
5747 * Description: Detach cgroup from @tsk and release it.
5749 * Note that cgroups marked notify_on_release force every task in
5750 * them to take the global cgroup_mutex mutex when exiting.
5751 * This could impact scaling on very large systems. Be reluctant to
5752 * use notify_on_release cgroups where very high task exit scaling
5753 * is required on large systems.
5755 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5756 * call cgroup_exit() while the task is still competent to handle
5757 * notify_on_release(), then leave the task attached to the root cgroup in
5758 * each hierarchy for the remainder of its exit. No need to bother with
5759 * init_css_set refcnting. init_css_set never goes away and we can't race
5760 * with migration path - PF_EXITING is visible to migration path.
5762 void cgroup_exit(struct task_struct *tsk)
5764 struct cgroup_subsys *ss;
5765 struct css_set *cset;
5769 * Unlink from @tsk from its css_set. As migration path can't race
5770 * with us, we can check css_set and cg_list without synchronization.
5772 cset = task_css_set(tsk);
5774 if (!list_empty(&tsk->cg_list)) {
5775 spin_lock_irq(&css_set_lock);
5776 css_set_move_task(tsk, cset, NULL, false);
5777 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
5779 spin_unlock_irq(&css_set_lock);
5784 /* see cgroup_post_fork() for details */
5785 do_each_subsys_mask(ss, i, have_exit_callback) {
5787 } while_each_subsys_mask();
5790 void cgroup_release(struct task_struct *task)
5792 struct cgroup_subsys *ss;
5795 do_each_subsys_mask(ss, ssid, have_release_callback) {
5797 } while_each_subsys_mask();
5799 if (use_task_css_set_links) {
5800 spin_lock_irq(&css_set_lock);
5801 css_set_skip_task_iters(task_css_set(task), task);
5802 list_del_init(&task->cg_list);
5803 spin_unlock_irq(&css_set_lock);
5807 void cgroup_free(struct task_struct *task)
5809 struct css_set *cset = task_css_set(task);
5813 static int __init cgroup_disable(char *str)
5815 struct cgroup_subsys *ss;
5819 while ((token = strsep(&str, ",")) != NULL) {
5823 for_each_subsys(ss, i) {
5824 if (strcmp(token, ss->name) &&
5825 strcmp(token, ss->legacy_name))
5828 static_branch_disable(cgroup_subsys_enabled_key[i]);
5829 pr_info("Disabling %s control group subsystem\n",
5835 __setup("cgroup_disable=", cgroup_disable);
5838 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5839 * @dentry: directory dentry of interest
5840 * @ss: subsystem of interest
5842 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5843 * to get the corresponding css and return it. If such css doesn't exist
5844 * or can't be pinned, an ERR_PTR value is returned.
5846 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5847 struct cgroup_subsys *ss)
5849 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5850 struct file_system_type *s_type = dentry->d_sb->s_type;
5851 struct cgroup_subsys_state *css = NULL;
5852 struct cgroup *cgrp;
5854 /* is @dentry a cgroup dir? */
5855 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5856 !kn || kernfs_type(kn) != KERNFS_DIR)
5857 return ERR_PTR(-EBADF);
5862 * This path doesn't originate from kernfs and @kn could already
5863 * have been or be removed at any point. @kn->priv is RCU
5864 * protected for this access. See css_release_work_fn() for details.
5866 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
5868 css = cgroup_css(cgrp, ss);
5870 if (!css || !css_tryget_online(css))
5871 css = ERR_PTR(-ENOENT);
5878 * css_from_id - lookup css by id
5879 * @id: the cgroup id
5880 * @ss: cgroup subsys to be looked into
5882 * Returns the css if there's valid one with @id, otherwise returns NULL.
5883 * Should be called under rcu_read_lock().
5885 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5887 WARN_ON_ONCE(!rcu_read_lock_held());
5888 return idr_find(&ss->css_idr, id);
5892 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5893 * @path: path on the default hierarchy
5895 * Find the cgroup at @path on the default hierarchy, increment its
5896 * reference count and return it. Returns pointer to the found cgroup on
5897 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5898 * if @path points to a non-directory.
5900 struct cgroup *cgroup_get_from_path(const char *path)
5902 struct kernfs_node *kn;
5903 struct cgroup *cgrp;
5905 mutex_lock(&cgroup_mutex);
5907 kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5909 if (kernfs_type(kn) == KERNFS_DIR) {
5911 cgroup_get_live(cgrp);
5913 cgrp = ERR_PTR(-ENOTDIR);
5917 cgrp = ERR_PTR(-ENOENT);
5920 mutex_unlock(&cgroup_mutex);
5923 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5926 * cgroup_get_from_fd - get a cgroup pointer from a fd
5927 * @fd: fd obtained by open(cgroup2_dir)
5929 * Find the cgroup from a fd which should be obtained
5930 * by opening a cgroup directory. Returns a pointer to the
5931 * cgroup on success. ERR_PTR is returned if the cgroup
5934 struct cgroup *cgroup_get_from_fd(int fd)
5936 struct cgroup_subsys_state *css;
5937 struct cgroup *cgrp;
5942 return ERR_PTR(-EBADF);
5944 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5947 return ERR_CAST(css);
5950 if (!cgroup_on_dfl(cgrp)) {
5952 return ERR_PTR(-EBADF);
5957 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
5960 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5961 * definition in cgroup-defs.h.
5963 #ifdef CONFIG_SOCK_CGROUP_DATA
5965 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5967 DEFINE_SPINLOCK(cgroup_sk_update_lock);
5968 static bool cgroup_sk_alloc_disabled __read_mostly;
5970 void cgroup_sk_alloc_disable(void)
5972 if (cgroup_sk_alloc_disabled)
5974 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5975 cgroup_sk_alloc_disabled = true;
5980 #define cgroup_sk_alloc_disabled false
5984 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5986 if (cgroup_sk_alloc_disabled) {
5987 skcd->no_refcnt = 1;
5991 /* Don't associate the sock with unrelated interrupted task's cgroup. */
5998 struct css_set *cset;
6000 cset = task_css_set(current);
6001 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6002 skcd->val = (unsigned long)cset->dfl_cgrp;
6011 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6013 /* Socket clone path */
6015 if (skcd->no_refcnt)
6018 * We might be cloning a socket which is left in an empty
6019 * cgroup and the cgroup might have already been rmdir'd.
6020 * Don't use cgroup_get_live().
6022 cgroup_get(sock_cgroup_ptr(skcd));
6026 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6028 if (skcd->no_refcnt)
6031 cgroup_put(sock_cgroup_ptr(skcd));
6034 #endif /* CONFIG_SOCK_CGROUP_DATA */
6036 #ifdef CONFIG_CGROUP_BPF
6037 int cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
6038 enum bpf_attach_type type, u32 flags)
6042 mutex_lock(&cgroup_mutex);
6043 ret = __cgroup_bpf_attach(cgrp, prog, type, flags);
6044 mutex_unlock(&cgroup_mutex);
6047 int cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
6048 enum bpf_attach_type type, u32 flags)
6052 mutex_lock(&cgroup_mutex);
6053 ret = __cgroup_bpf_detach(cgrp, prog, type, flags);
6054 mutex_unlock(&cgroup_mutex);
6057 int cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
6058 union bpf_attr __user *uattr)
6062 mutex_lock(&cgroup_mutex);
6063 ret = __cgroup_bpf_query(cgrp, attr, uattr);
6064 mutex_unlock(&cgroup_mutex);
6067 #endif /* CONFIG_CGROUP_BPF */
6070 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6071 ssize_t size, const char *prefix)
6076 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6077 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6081 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6083 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6085 if (unlikely(ret >= size)) {
6094 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6097 struct cgroup_subsys *ss;
6101 ret = show_delegatable_files(cgroup_base_files, buf, PAGE_SIZE - ret,
6104 for_each_subsys(ss, ssid)
6105 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
6107 cgroup_subsys_name[ssid]);
6111 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
6113 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
6116 return snprintf(buf, PAGE_SIZE, "nsdelegate\n");
6118 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
6120 static struct attribute *cgroup_sysfs_attrs[] = {
6121 &cgroup_delegate_attr.attr,
6122 &cgroup_features_attr.attr,
6126 static const struct attribute_group cgroup_sysfs_attr_group = {
6127 .attrs = cgroup_sysfs_attrs,
6131 static int __init cgroup_sysfs_init(void)
6133 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
6135 subsys_initcall(cgroup_sysfs_init);
6136 #endif /* CONFIG_SYSFS */