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/bpf-cgroup.h>
34 #include <linux/cred.h>
35 #include <linux/errno.h>
36 #include <linux/init_task.h>
37 #include <linux/kernel.h>
38 #include <linux/magic.h>
39 #include <linux/mutex.h>
40 #include <linux/mount.h>
41 #include <linux/pagemap.h>
42 #include <linux/proc_fs.h>
43 #include <linux/rcupdate.h>
44 #include <linux/sched.h>
45 #include <linux/sched/task.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/hashtable.h>
51 #include <linux/idr.h>
52 #include <linux/kthread.h>
53 #include <linux/atomic.h>
54 #include <linux/cpuset.h>
55 #include <linux/proc_ns.h>
56 #include <linux/nsproxy.h>
57 #include <linux/file.h>
58 #include <linux/fs_parser.h>
59 #include <linux/sched/cputime.h>
60 #include <linux/psi.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/cgroup.h>
66 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
68 /* let's not notify more than 100 times per second */
69 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
72 * To avoid confusing the compiler (and generating warnings) with code
73 * that attempts to access what would be a 0-element array (i.e. sized
74 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
75 * constant expression can be added.
77 #define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
80 * cgroup_mutex is the master lock. Any modification to cgroup or its
81 * hierarchy must be performed while holding it.
83 * css_set_lock protects task->cgroups pointer, the list of css_set
84 * objects, and the chain of tasks off each css_set.
86 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
87 * cgroup.h can use them for lockdep annotations.
89 DEFINE_MUTEX(cgroup_mutex);
90 DEFINE_SPINLOCK(css_set_lock);
92 #ifdef CONFIG_PROVE_RCU
93 EXPORT_SYMBOL_GPL(cgroup_mutex);
94 EXPORT_SYMBOL_GPL(css_set_lock);
97 DEFINE_SPINLOCK(trace_cgroup_path_lock);
98 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
99 static bool cgroup_debug __read_mostly;
102 * Protects cgroup_idr and css_idr so that IDs can be released without
103 * grabbing cgroup_mutex.
105 static DEFINE_SPINLOCK(cgroup_idr_lock);
108 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
109 * against file removal/re-creation across css hiding.
111 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
113 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct *cgroup_destroy_wq;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true *cgroup_subsys_enabled_key[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
159 #include <linux/cgroup_subsys.h>
163 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
165 /* the default hierarchy */
166 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
167 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
170 * The default hierarchy always exists but is hidden until mounted for the
171 * first time. This is for backward compatibility.
173 static bool cgrp_dfl_visible;
175 /* some controllers are not supported in the default hierarchy */
176 static u16 cgrp_dfl_inhibit_ss_mask;
178 /* some controllers are implicitly enabled on the default hierarchy */
179 static u16 cgrp_dfl_implicit_ss_mask;
181 /* some controllers can be threaded on the default hierarchy */
182 static u16 cgrp_dfl_threaded_ss_mask;
184 /* The list of hierarchy roots */
185 LIST_HEAD(cgroup_roots);
186 static int cgroup_root_count;
188 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
189 static DEFINE_IDR(cgroup_hierarchy_idr);
192 * Assign a monotonically increasing serial number to csses. It guarantees
193 * cgroups with bigger numbers are newer than those with smaller numbers.
194 * Also, as csses are always appended to the parent's ->children list, it
195 * guarantees that sibling csses are always sorted in the ascending serial
196 * number order on the list. Protected by cgroup_mutex.
198 static u64 css_serial_nr_next = 1;
201 * These bitmasks identify subsystems with specific features to avoid
202 * having to do iterative checks repeatedly.
204 static u16 have_fork_callback __read_mostly;
205 static u16 have_exit_callback __read_mostly;
206 static u16 have_release_callback __read_mostly;
207 static u16 have_canfork_callback __read_mostly;
209 /* cgroup namespace for init task */
210 struct cgroup_namespace init_cgroup_ns = {
211 .ns.count = REFCOUNT_INIT(2),
212 .user_ns = &init_user_ns,
213 .ns.ops = &cgroupns_operations,
214 .ns.inum = PROC_CGROUP_INIT_INO,
215 .root_cset = &init_css_set,
218 static struct file_system_type cgroup2_fs_type;
219 static struct cftype cgroup_base_files[];
220 static struct cftype cgroup_psi_files[];
222 /* cgroup optional features */
223 enum cgroup_opt_features {
225 OPT_FEATURE_PRESSURE,
230 static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
236 static u16 cgroup_feature_disable_mask __read_mostly;
238 static int cgroup_apply_control(struct cgroup *cgrp);
239 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
240 static void css_task_iter_skip(struct css_task_iter *it,
241 struct task_struct *task);
242 static int cgroup_destroy_locked(struct cgroup *cgrp);
243 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
244 struct cgroup_subsys *ss);
245 static void css_release(struct percpu_ref *ref);
246 static void kill_css(struct cgroup_subsys_state *css);
247 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
248 struct cgroup *cgrp, struct cftype cfts[],
251 #ifdef CONFIG_DEBUG_CGROUP_REF
252 #define CGROUP_REF_FN_ATTRS noinline
253 #define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
254 #include <linux/cgroup_refcnt.h>
258 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
259 * @ssid: subsys ID of interest
261 * cgroup_subsys_enabled() can only be used with literal subsys names which
262 * is fine for individual subsystems but unsuitable for cgroup core. This
263 * is slower static_key_enabled() based test indexed by @ssid.
265 bool cgroup_ssid_enabled(int ssid)
267 if (!CGROUP_HAS_SUBSYS_CONFIG)
270 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
274 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
275 * @cgrp: the cgroup of interest
277 * The default hierarchy is the v2 interface of cgroup and this function
278 * can be used to test whether a cgroup is on the default hierarchy for
279 * cases where a subsystem should behave differently depending on the
282 * List of changed behaviors:
284 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
285 * and "name" are disallowed.
287 * - When mounting an existing superblock, mount options should match.
289 * - rename(2) is disallowed.
291 * - "tasks" is removed. Everything should be at process granularity. Use
292 * "cgroup.procs" instead.
294 * - "cgroup.procs" is not sorted. pids will be unique unless they got
295 * recycled in-between reads.
297 * - "release_agent" and "notify_on_release" are removed. Replacement
298 * notification mechanism will be implemented.
300 * - "cgroup.clone_children" is removed.
302 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
303 * and its descendants contain no task; otherwise, 1. The file also
304 * generates kernfs notification which can be monitored through poll and
305 * [di]notify when the value of the file changes.
307 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
308 * take masks of ancestors with non-empty cpus/mems, instead of being
309 * moved to an ancestor.
311 * - cpuset: a task can be moved into an empty cpuset, and again it takes
312 * masks of ancestors.
314 * - blkcg: blk-throttle becomes properly hierarchical.
316 * - debug: disallowed on the default hierarchy.
318 bool cgroup_on_dfl(const struct cgroup *cgrp)
320 return cgrp->root == &cgrp_dfl_root;
323 /* IDR wrappers which synchronize using cgroup_idr_lock */
324 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
329 idr_preload(gfp_mask);
330 spin_lock_bh(&cgroup_idr_lock);
331 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
332 spin_unlock_bh(&cgroup_idr_lock);
337 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
341 spin_lock_bh(&cgroup_idr_lock);
342 ret = idr_replace(idr, ptr, id);
343 spin_unlock_bh(&cgroup_idr_lock);
347 static void cgroup_idr_remove(struct idr *idr, int id)
349 spin_lock_bh(&cgroup_idr_lock);
351 spin_unlock_bh(&cgroup_idr_lock);
354 static bool cgroup_has_tasks(struct cgroup *cgrp)
356 return cgrp->nr_populated_csets;
359 bool cgroup_is_threaded(struct cgroup *cgrp)
361 return cgrp->dom_cgrp != cgrp;
364 /* can @cgrp host both domain and threaded children? */
365 static bool cgroup_is_mixable(struct cgroup *cgrp)
368 * Root isn't under domain level resource control exempting it from
369 * the no-internal-process constraint, so it can serve as a thread
370 * root and a parent of resource domains at the same time.
372 return !cgroup_parent(cgrp);
375 /* can @cgrp become a thread root? Should always be true for a thread root */
376 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
378 /* mixables don't care */
379 if (cgroup_is_mixable(cgrp))
382 /* domain roots can't be nested under threaded */
383 if (cgroup_is_threaded(cgrp))
386 /* can only have either domain or threaded children */
387 if (cgrp->nr_populated_domain_children)
390 /* and no domain controllers can be enabled */
391 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
397 /* is @cgrp root of a threaded subtree? */
398 bool cgroup_is_thread_root(struct cgroup *cgrp)
400 /* thread root should be a domain */
401 if (cgroup_is_threaded(cgrp))
404 /* a domain w/ threaded children is a thread root */
405 if (cgrp->nr_threaded_children)
409 * A domain which has tasks and explicit threaded controllers
410 * enabled is a thread root.
412 if (cgroup_has_tasks(cgrp) &&
413 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
419 /* a domain which isn't connected to the root w/o brekage can't be used */
420 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
422 /* the cgroup itself can be a thread root */
423 if (cgroup_is_threaded(cgrp))
426 /* but the ancestors can't be unless mixable */
427 while ((cgrp = cgroup_parent(cgrp))) {
428 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
430 if (cgroup_is_threaded(cgrp))
437 /* subsystems visibly enabled on a cgroup */
438 static u16 cgroup_control(struct cgroup *cgrp)
440 struct cgroup *parent = cgroup_parent(cgrp);
441 u16 root_ss_mask = cgrp->root->subsys_mask;
444 u16 ss_mask = parent->subtree_control;
446 /* threaded cgroups can only have threaded controllers */
447 if (cgroup_is_threaded(cgrp))
448 ss_mask &= cgrp_dfl_threaded_ss_mask;
452 if (cgroup_on_dfl(cgrp))
453 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
454 cgrp_dfl_implicit_ss_mask);
458 /* subsystems enabled on a cgroup */
459 static u16 cgroup_ss_mask(struct cgroup *cgrp)
461 struct cgroup *parent = cgroup_parent(cgrp);
464 u16 ss_mask = parent->subtree_ss_mask;
466 /* threaded cgroups can only have threaded controllers */
467 if (cgroup_is_threaded(cgrp))
468 ss_mask &= cgrp_dfl_threaded_ss_mask;
472 return cgrp->root->subsys_mask;
476 * cgroup_css - obtain a cgroup's css for the specified subsystem
477 * @cgrp: the cgroup of interest
478 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
480 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
481 * function must be called either under cgroup_mutex or rcu_read_lock() and
482 * the caller is responsible for pinning the returned css if it wants to
483 * keep accessing it outside the said locks. This function may return
484 * %NULL if @cgrp doesn't have @subsys_id enabled.
486 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
487 struct cgroup_subsys *ss)
489 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
490 return rcu_dereference_check(cgrp->subsys[ss->id],
491 lockdep_is_held(&cgroup_mutex));
497 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
498 * @cgrp: the cgroup of interest
499 * @ss: the subsystem of interest
501 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
502 * or is offline, %NULL is returned.
504 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
505 struct cgroup_subsys *ss)
507 struct cgroup_subsys_state *css;
510 css = cgroup_css(cgrp, ss);
511 if (css && !css_tryget_online(css))
519 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
520 * @cgrp: the cgroup of interest
521 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
523 * Similar to cgroup_css() but returns the effective css, which is defined
524 * as the matching css of the nearest ancestor including self which has @ss
525 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
526 * function is guaranteed to return non-NULL css.
528 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
529 struct cgroup_subsys *ss)
531 lockdep_assert_held(&cgroup_mutex);
537 * This function is used while updating css associations and thus
538 * can't test the csses directly. Test ss_mask.
540 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
541 cgrp = cgroup_parent(cgrp);
546 return cgroup_css(cgrp, ss);
550 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
551 * @cgrp: the cgroup of interest
552 * @ss: the subsystem of interest
554 * Find and get the effective css of @cgrp for @ss. The effective css is
555 * defined as the matching css of the nearest ancestor including self which
556 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
557 * the root css is returned, so this function always returns a valid css.
559 * The returned css is not guaranteed to be online, and therefore it is the
560 * callers responsibility to try get a reference for it.
562 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
563 struct cgroup_subsys *ss)
565 struct cgroup_subsys_state *css;
567 if (!CGROUP_HAS_SUBSYS_CONFIG)
571 css = cgroup_css(cgrp, ss);
575 cgrp = cgroup_parent(cgrp);
578 return init_css_set.subsys[ss->id];
582 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
583 * @cgrp: the cgroup of interest
584 * @ss: the subsystem of interest
586 * Find and get the effective css of @cgrp for @ss. The effective css is
587 * defined as the matching css of the nearest ancestor including self which
588 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
589 * the root css is returned, so this function always returns a valid css.
590 * The returned css must be put using css_put().
592 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
593 struct cgroup_subsys *ss)
595 struct cgroup_subsys_state *css;
597 if (!CGROUP_HAS_SUBSYS_CONFIG)
603 css = cgroup_css(cgrp, ss);
605 if (css && css_tryget_online(css))
607 cgrp = cgroup_parent(cgrp);
610 css = init_css_set.subsys[ss->id];
616 EXPORT_SYMBOL_GPL(cgroup_get_e_css);
618 static void cgroup_get_live(struct cgroup *cgrp)
620 WARN_ON_ONCE(cgroup_is_dead(cgrp));
621 css_get(&cgrp->self);
625 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
626 * is responsible for taking the css_set_lock.
627 * @cgrp: the cgroup in question
629 int __cgroup_task_count(const struct cgroup *cgrp)
632 struct cgrp_cset_link *link;
634 lockdep_assert_held(&css_set_lock);
636 list_for_each_entry(link, &cgrp->cset_links, cset_link)
637 count += link->cset->nr_tasks;
643 * cgroup_task_count - count the number of tasks in a cgroup.
644 * @cgrp: the cgroup in question
646 int cgroup_task_count(const struct cgroup *cgrp)
650 spin_lock_irq(&css_set_lock);
651 count = __cgroup_task_count(cgrp);
652 spin_unlock_irq(&css_set_lock);
657 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
659 struct cgroup *cgrp = of->kn->parent->priv;
660 struct cftype *cft = of_cft(of);
663 * This is open and unprotected implementation of cgroup_css().
664 * seq_css() is only called from a kernfs file operation which has
665 * an active reference on the file. Because all the subsystem
666 * files are drained before a css is disassociated with a cgroup,
667 * the matching css from the cgroup's subsys table is guaranteed to
668 * be and stay valid until the enclosing operation is complete.
670 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
671 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
675 EXPORT_SYMBOL_GPL(of_css);
678 * for_each_css - iterate all css's of a cgroup
679 * @css: the iteration cursor
680 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
681 * @cgrp: the target cgroup to iterate css's of
683 * Should be called under cgroup_[tree_]mutex.
685 #define for_each_css(css, ssid, cgrp) \
686 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
687 if (!((css) = rcu_dereference_check( \
688 (cgrp)->subsys[(ssid)], \
689 lockdep_is_held(&cgroup_mutex)))) { } \
693 * for_each_e_css - iterate all effective css's of a cgroup
694 * @css: the iteration cursor
695 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
696 * @cgrp: the target cgroup to iterate css's of
698 * Should be called under cgroup_[tree_]mutex.
700 #define for_each_e_css(css, ssid, cgrp) \
701 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
702 if (!((css) = cgroup_e_css_by_mask(cgrp, \
703 cgroup_subsys[(ssid)]))) \
708 * do_each_subsys_mask - filter for_each_subsys with a bitmask
709 * @ss: the iteration cursor
710 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
711 * @ss_mask: the bitmask
713 * The block will only run for cases where the ssid-th bit (1 << ssid) of
716 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
717 unsigned long __ss_mask = (ss_mask); \
718 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
722 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
723 (ss) = cgroup_subsys[ssid]; \
726 #define while_each_subsys_mask() \
731 /* iterate over child cgrps, lock should be held throughout iteration */
732 #define cgroup_for_each_live_child(child, cgrp) \
733 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
734 if (({ lockdep_assert_held(&cgroup_mutex); \
735 cgroup_is_dead(child); })) \
739 /* walk live descendants in pre order */
740 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
741 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
742 if (({ lockdep_assert_held(&cgroup_mutex); \
743 (dsct) = (d_css)->cgroup; \
744 cgroup_is_dead(dsct); })) \
748 /* walk live descendants in postorder */
749 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
750 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
751 if (({ lockdep_assert_held(&cgroup_mutex); \
752 (dsct) = (d_css)->cgroup; \
753 cgroup_is_dead(dsct); })) \
758 * The default css_set - used by init and its children prior to any
759 * hierarchies being mounted. It contains a pointer to the root state
760 * for each subsystem. Also used to anchor the list of css_sets. Not
761 * reference-counted, to improve performance when child cgroups
762 * haven't been created.
764 struct css_set init_css_set = {
765 .refcount = REFCOUNT_INIT(1),
766 .dom_cset = &init_css_set,
767 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
768 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
769 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
770 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
771 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
772 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
773 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
774 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
775 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
778 * The following field is re-initialized when this cset gets linked
779 * in cgroup_init(). However, let's initialize the field
780 * statically too so that the default cgroup can be accessed safely
783 .dfl_cgrp = &cgrp_dfl_root.cgrp,
786 static int css_set_count = 1; /* 1 for init_css_set */
788 static bool css_set_threaded(struct css_set *cset)
790 return cset->dom_cset != cset;
794 * css_set_populated - does a css_set contain any tasks?
795 * @cset: target css_set
797 * css_set_populated() should be the same as !!cset->nr_tasks at steady
798 * state. However, css_set_populated() can be called while a task is being
799 * added to or removed from the linked list before the nr_tasks is
800 * properly updated. Hence, we can't just look at ->nr_tasks here.
802 static bool css_set_populated(struct css_set *cset)
804 lockdep_assert_held(&css_set_lock);
806 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
810 * cgroup_update_populated - update the populated count of a cgroup
811 * @cgrp: the target cgroup
812 * @populated: inc or dec populated count
814 * One of the css_sets associated with @cgrp is either getting its first
815 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
816 * count is propagated towards root so that a given cgroup's
817 * nr_populated_children is zero iff none of its descendants contain any
820 * @cgrp's interface file "cgroup.populated" is zero if both
821 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
822 * 1 otherwise. When the sum changes from or to zero, userland is notified
823 * that the content of the interface file has changed. This can be used to
824 * detect when @cgrp and its descendants become populated or empty.
826 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
828 struct cgroup *child = NULL;
829 int adj = populated ? 1 : -1;
831 lockdep_assert_held(&css_set_lock);
834 bool was_populated = cgroup_is_populated(cgrp);
837 cgrp->nr_populated_csets += adj;
839 if (cgroup_is_threaded(child))
840 cgrp->nr_populated_threaded_children += adj;
842 cgrp->nr_populated_domain_children += adj;
845 if (was_populated == cgroup_is_populated(cgrp))
848 cgroup1_check_for_release(cgrp);
849 TRACE_CGROUP_PATH(notify_populated, cgrp,
850 cgroup_is_populated(cgrp));
851 cgroup_file_notify(&cgrp->events_file);
854 cgrp = cgroup_parent(cgrp);
859 * css_set_update_populated - update populated state of a css_set
860 * @cset: target css_set
861 * @populated: whether @cset is populated or depopulated
863 * @cset is either getting the first task or losing the last. Update the
864 * populated counters of all associated cgroups accordingly.
866 static void css_set_update_populated(struct css_set *cset, bool populated)
868 struct cgrp_cset_link *link;
870 lockdep_assert_held(&css_set_lock);
872 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
873 cgroup_update_populated(link->cgrp, populated);
877 * @task is leaving, advance task iterators which are pointing to it so
878 * that they can resume at the next position. Advancing an iterator might
879 * remove it from the list, use safe walk. See css_task_iter_skip() for
882 static void css_set_skip_task_iters(struct css_set *cset,
883 struct task_struct *task)
885 struct css_task_iter *it, *pos;
887 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
888 css_task_iter_skip(it, task);
892 * css_set_move_task - move a task from one css_set to another
893 * @task: task being moved
894 * @from_cset: css_set @task currently belongs to (may be NULL)
895 * @to_cset: new css_set @task is being moved to (may be NULL)
896 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
898 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
899 * css_set, @from_cset can be NULL. If @task is being disassociated
900 * instead of moved, @to_cset can be NULL.
902 * This function automatically handles populated counter updates and
903 * css_task_iter adjustments but the caller is responsible for managing
904 * @from_cset and @to_cset's reference counts.
906 static void css_set_move_task(struct task_struct *task,
907 struct css_set *from_cset, struct css_set *to_cset,
910 lockdep_assert_held(&css_set_lock);
912 if (to_cset && !css_set_populated(to_cset))
913 css_set_update_populated(to_cset, true);
916 WARN_ON_ONCE(list_empty(&task->cg_list));
918 css_set_skip_task_iters(from_cset, task);
919 list_del_init(&task->cg_list);
920 if (!css_set_populated(from_cset))
921 css_set_update_populated(from_cset, false);
923 WARN_ON_ONCE(!list_empty(&task->cg_list));
928 * We are synchronized through cgroup_threadgroup_rwsem
929 * against PF_EXITING setting such that we can't race
930 * against cgroup_exit()/cgroup_free() dropping the css_set.
932 WARN_ON_ONCE(task->flags & PF_EXITING);
934 cgroup_move_task(task, to_cset);
935 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
941 * hash table for cgroup groups. This improves the performance to find
942 * an existing css_set. This hash doesn't (currently) take into
943 * account cgroups in empty hierarchies.
945 #define CSS_SET_HASH_BITS 7
946 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
948 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
950 unsigned long key = 0UL;
951 struct cgroup_subsys *ss;
954 for_each_subsys(ss, i)
955 key += (unsigned long)css[i];
956 key = (key >> 16) ^ key;
961 void put_css_set_locked(struct css_set *cset)
963 struct cgrp_cset_link *link, *tmp_link;
964 struct cgroup_subsys *ss;
967 lockdep_assert_held(&css_set_lock);
969 if (!refcount_dec_and_test(&cset->refcount))
972 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
974 /* This css_set is dead. Unlink it and release cgroup and css refs */
975 for_each_subsys(ss, ssid) {
976 list_del(&cset->e_cset_node[ssid]);
977 css_put(cset->subsys[ssid]);
979 hash_del(&cset->hlist);
982 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
983 list_del(&link->cset_link);
984 list_del(&link->cgrp_link);
985 if (cgroup_parent(link->cgrp))
986 cgroup_put(link->cgrp);
990 if (css_set_threaded(cset)) {
991 list_del(&cset->threaded_csets_node);
992 put_css_set_locked(cset->dom_cset);
995 kfree_rcu(cset, rcu_head);
999 * compare_css_sets - helper function for find_existing_css_set().
1000 * @cset: candidate css_set being tested
1001 * @old_cset: existing css_set for a task
1002 * @new_cgrp: cgroup that's being entered by the task
1003 * @template: desired set of css pointers in css_set (pre-calculated)
1005 * Returns true if "cset" matches "old_cset" except for the hierarchy
1006 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1008 static bool compare_css_sets(struct css_set *cset,
1009 struct css_set *old_cset,
1010 struct cgroup *new_cgrp,
1011 struct cgroup_subsys_state *template[])
1013 struct cgroup *new_dfl_cgrp;
1014 struct list_head *l1, *l2;
1017 * On the default hierarchy, there can be csets which are
1018 * associated with the same set of cgroups but different csses.
1019 * Let's first ensure that csses match.
1021 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1025 /* @cset's domain should match the default cgroup's */
1026 if (cgroup_on_dfl(new_cgrp))
1027 new_dfl_cgrp = new_cgrp;
1029 new_dfl_cgrp = old_cset->dfl_cgrp;
1031 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1035 * Compare cgroup pointers in order to distinguish between
1036 * different cgroups in hierarchies. As different cgroups may
1037 * share the same effective css, this comparison is always
1040 l1 = &cset->cgrp_links;
1041 l2 = &old_cset->cgrp_links;
1043 struct cgrp_cset_link *link1, *link2;
1044 struct cgroup *cgrp1, *cgrp2;
1048 /* See if we reached the end - both lists are equal length. */
1049 if (l1 == &cset->cgrp_links) {
1050 BUG_ON(l2 != &old_cset->cgrp_links);
1053 BUG_ON(l2 == &old_cset->cgrp_links);
1055 /* Locate the cgroups associated with these links. */
1056 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1057 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1058 cgrp1 = link1->cgrp;
1059 cgrp2 = link2->cgrp;
1060 /* Hierarchies should be linked in the same order. */
1061 BUG_ON(cgrp1->root != cgrp2->root);
1064 * If this hierarchy is the hierarchy of the cgroup
1065 * that's changing, then we need to check that this
1066 * css_set points to the new cgroup; if it's any other
1067 * hierarchy, then this css_set should point to the
1068 * same cgroup as the old css_set.
1070 if (cgrp1->root == new_cgrp->root) {
1071 if (cgrp1 != new_cgrp)
1082 * find_existing_css_set - init css array and find the matching css_set
1083 * @old_cset: the css_set that we're using before the cgroup transition
1084 * @cgrp: the cgroup that we're moving into
1085 * @template: out param for the new set of csses, should be clear on entry
1087 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1088 struct cgroup *cgrp,
1089 struct cgroup_subsys_state *template[])
1091 struct cgroup_root *root = cgrp->root;
1092 struct cgroup_subsys *ss;
1093 struct css_set *cset;
1098 * Build the set of subsystem state objects that we want to see in the
1099 * new css_set. While subsystems can change globally, the entries here
1100 * won't change, so no need for locking.
1102 for_each_subsys(ss, i) {
1103 if (root->subsys_mask & (1UL << i)) {
1105 * @ss is in this hierarchy, so we want the
1106 * effective css from @cgrp.
1108 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1111 * @ss is not in this hierarchy, so we don't want
1112 * to change the css.
1114 template[i] = old_cset->subsys[i];
1118 key = css_set_hash(template);
1119 hash_for_each_possible(css_set_table, cset, hlist, key) {
1120 if (!compare_css_sets(cset, old_cset, cgrp, template))
1123 /* This css_set matches what we need */
1127 /* No existing cgroup group matched */
1131 static void free_cgrp_cset_links(struct list_head *links_to_free)
1133 struct cgrp_cset_link *link, *tmp_link;
1135 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1136 list_del(&link->cset_link);
1142 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1143 * @count: the number of links to allocate
1144 * @tmp_links: list_head the allocated links are put on
1146 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1147 * through ->cset_link. Returns 0 on success or -errno.
1149 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1151 struct cgrp_cset_link *link;
1154 INIT_LIST_HEAD(tmp_links);
1156 for (i = 0; i < count; i++) {
1157 link = kzalloc(sizeof(*link), GFP_KERNEL);
1159 free_cgrp_cset_links(tmp_links);
1162 list_add(&link->cset_link, tmp_links);
1168 * link_css_set - a helper function to link a css_set to a cgroup
1169 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1170 * @cset: the css_set to be linked
1171 * @cgrp: the destination cgroup
1173 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1174 struct cgroup *cgrp)
1176 struct cgrp_cset_link *link;
1178 BUG_ON(list_empty(tmp_links));
1180 if (cgroup_on_dfl(cgrp))
1181 cset->dfl_cgrp = cgrp;
1183 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1188 * Always add links to the tail of the lists so that the lists are
1189 * in chronological order.
1191 list_move_tail(&link->cset_link, &cgrp->cset_links);
1192 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1194 if (cgroup_parent(cgrp))
1195 cgroup_get_live(cgrp);
1199 * find_css_set - return a new css_set with one cgroup updated
1200 * @old_cset: the baseline css_set
1201 * @cgrp: the cgroup to be updated
1203 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1204 * substituted into the appropriate hierarchy.
1206 static struct css_set *find_css_set(struct css_set *old_cset,
1207 struct cgroup *cgrp)
1209 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1210 struct css_set *cset;
1211 struct list_head tmp_links;
1212 struct cgrp_cset_link *link;
1213 struct cgroup_subsys *ss;
1217 lockdep_assert_held(&cgroup_mutex);
1219 /* First see if we already have a cgroup group that matches
1220 * the desired set */
1221 spin_lock_irq(&css_set_lock);
1222 cset = find_existing_css_set(old_cset, cgrp, template);
1225 spin_unlock_irq(&css_set_lock);
1230 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1234 /* Allocate all the cgrp_cset_link objects that we'll need */
1235 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1240 refcount_set(&cset->refcount, 1);
1241 cset->dom_cset = cset;
1242 INIT_LIST_HEAD(&cset->tasks);
1243 INIT_LIST_HEAD(&cset->mg_tasks);
1244 INIT_LIST_HEAD(&cset->dying_tasks);
1245 INIT_LIST_HEAD(&cset->task_iters);
1246 INIT_LIST_HEAD(&cset->threaded_csets);
1247 INIT_HLIST_NODE(&cset->hlist);
1248 INIT_LIST_HEAD(&cset->cgrp_links);
1249 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1250 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1251 INIT_LIST_HEAD(&cset->mg_node);
1253 /* Copy the set of subsystem state objects generated in
1254 * find_existing_css_set() */
1255 memcpy(cset->subsys, template, sizeof(cset->subsys));
1257 spin_lock_irq(&css_set_lock);
1258 /* Add reference counts and links from the new css_set. */
1259 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1260 struct cgroup *c = link->cgrp;
1262 if (c->root == cgrp->root)
1264 link_css_set(&tmp_links, cset, c);
1267 BUG_ON(!list_empty(&tmp_links));
1271 /* Add @cset to the hash table */
1272 key = css_set_hash(cset->subsys);
1273 hash_add(css_set_table, &cset->hlist, key);
1275 for_each_subsys(ss, ssid) {
1276 struct cgroup_subsys_state *css = cset->subsys[ssid];
1278 list_add_tail(&cset->e_cset_node[ssid],
1279 &css->cgroup->e_csets[ssid]);
1283 spin_unlock_irq(&css_set_lock);
1286 * If @cset should be threaded, look up the matching dom_cset and
1287 * link them up. We first fully initialize @cset then look for the
1288 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1289 * to stay empty until we return.
1291 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1292 struct css_set *dcset;
1294 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1300 spin_lock_irq(&css_set_lock);
1301 cset->dom_cset = dcset;
1302 list_add_tail(&cset->threaded_csets_node,
1303 &dcset->threaded_csets);
1304 spin_unlock_irq(&css_set_lock);
1310 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1312 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1314 return root_cgrp->root;
1317 void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1319 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1321 /* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1322 if (favor && !favoring) {
1323 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1324 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1325 } else if (!favor && favoring) {
1326 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1327 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1331 static int cgroup_init_root_id(struct cgroup_root *root)
1335 lockdep_assert_held(&cgroup_mutex);
1337 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1341 root->hierarchy_id = id;
1345 static void cgroup_exit_root_id(struct cgroup_root *root)
1347 lockdep_assert_held(&cgroup_mutex);
1349 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1352 void cgroup_free_root(struct cgroup_root *root)
1357 static void cgroup_destroy_root(struct cgroup_root *root)
1359 struct cgroup *cgrp = &root->cgrp;
1360 struct cgrp_cset_link *link, *tmp_link;
1362 trace_cgroup_destroy_root(root);
1364 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1366 BUG_ON(atomic_read(&root->nr_cgrps));
1367 BUG_ON(!list_empty(&cgrp->self.children));
1369 /* Rebind all subsystems back to the default hierarchy */
1370 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1373 * Release all the links from cset_links to this hierarchy's
1376 spin_lock_irq(&css_set_lock);
1378 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1379 list_del(&link->cset_link);
1380 list_del(&link->cgrp_link);
1384 spin_unlock_irq(&css_set_lock);
1386 if (!list_empty(&root->root_list)) {
1387 list_del(&root->root_list);
1388 cgroup_root_count--;
1391 cgroup_favor_dynmods(root, false);
1392 cgroup_exit_root_id(root);
1394 mutex_unlock(&cgroup_mutex);
1396 cgroup_rstat_exit(cgrp);
1397 kernfs_destroy_root(root->kf_root);
1398 cgroup_free_root(root);
1402 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1404 static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1405 struct cgroup_root *root)
1407 struct cgroup *res_cgroup = NULL;
1409 if (cset == &init_css_set) {
1410 res_cgroup = &root->cgrp;
1411 } else if (root == &cgrp_dfl_root) {
1412 res_cgroup = cset->dfl_cgrp;
1414 struct cgrp_cset_link *link;
1415 lockdep_assert_held(&css_set_lock);
1417 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1418 struct cgroup *c = link->cgrp;
1420 if (c->root == root) {
1427 BUG_ON(!res_cgroup);
1432 * look up cgroup associated with current task's cgroup namespace on the
1433 * specified hierarchy
1435 static struct cgroup *
1436 current_cgns_cgroup_from_root(struct cgroup_root *root)
1438 struct cgroup *res = NULL;
1439 struct css_set *cset;
1441 lockdep_assert_held(&css_set_lock);
1445 cset = current->nsproxy->cgroup_ns->root_cset;
1446 res = __cset_cgroup_from_root(cset, root);
1454 * Look up cgroup associated with current task's cgroup namespace on the default
1457 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1458 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1460 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1461 * - As a bonus returned cgrp is pinned with the current because it cannot
1462 * switch cgroup_ns asynchronously.
1464 static struct cgroup *current_cgns_cgroup_dfl(void)
1466 struct css_set *cset;
1468 cset = current->nsproxy->cgroup_ns->root_cset;
1469 return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1472 /* look up cgroup associated with given css_set on the specified hierarchy */
1473 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1474 struct cgroup_root *root)
1476 lockdep_assert_held(&cgroup_mutex);
1477 lockdep_assert_held(&css_set_lock);
1479 return __cset_cgroup_from_root(cset, root);
1483 * Return the cgroup for "task" from the given hierarchy. Must be
1484 * called with cgroup_mutex and css_set_lock held.
1486 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1487 struct cgroup_root *root)
1490 * No need to lock the task - since we hold css_set_lock the
1491 * task can't change groups.
1493 return cset_cgroup_from_root(task_css_set(task), root);
1497 * A task must hold cgroup_mutex to modify cgroups.
1499 * Any task can increment and decrement the count field without lock.
1500 * So in general, code holding cgroup_mutex can't rely on the count
1501 * field not changing. However, if the count goes to zero, then only
1502 * cgroup_attach_task() can increment it again. Because a count of zero
1503 * means that no tasks are currently attached, therefore there is no
1504 * way a task attached to that cgroup can fork (the other way to
1505 * increment the count). So code holding cgroup_mutex can safely
1506 * assume that if the count is zero, it will stay zero. Similarly, if
1507 * a task holds cgroup_mutex on a cgroup with zero count, it
1508 * knows that the cgroup won't be removed, as cgroup_rmdir()
1511 * A cgroup can only be deleted if both its 'count' of using tasks
1512 * is zero, and its list of 'children' cgroups is empty. Since all
1513 * tasks in the system use _some_ cgroup, and since there is always at
1514 * least one task in the system (init, pid == 1), therefore, root cgroup
1515 * always has either children cgroups and/or using tasks. So we don't
1516 * need a special hack to ensure that root cgroup cannot be deleted.
1518 * P.S. One more locking exception. RCU is used to guard the
1519 * update of a tasks cgroup pointer by cgroup_attach_task()
1522 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1524 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1527 struct cgroup_subsys *ss = cft->ss;
1529 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1530 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1531 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1533 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1534 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1537 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1543 * cgroup_file_mode - deduce file mode of a control file
1544 * @cft: the control file in question
1546 * S_IRUGO for read, S_IWUSR for write.
1548 static umode_t cgroup_file_mode(const struct cftype *cft)
1552 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1555 if (cft->write_u64 || cft->write_s64 || cft->write) {
1556 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1566 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1567 * @subtree_control: the new subtree_control mask to consider
1568 * @this_ss_mask: available subsystems
1570 * On the default hierarchy, a subsystem may request other subsystems to be
1571 * enabled together through its ->depends_on mask. In such cases, more
1572 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1574 * This function calculates which subsystems need to be enabled if
1575 * @subtree_control is to be applied while restricted to @this_ss_mask.
1577 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1579 u16 cur_ss_mask = subtree_control;
1580 struct cgroup_subsys *ss;
1583 lockdep_assert_held(&cgroup_mutex);
1585 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1588 u16 new_ss_mask = cur_ss_mask;
1590 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1591 new_ss_mask |= ss->depends_on;
1592 } while_each_subsys_mask();
1595 * Mask out subsystems which aren't available. This can
1596 * happen only if some depended-upon subsystems were bound
1597 * to non-default hierarchies.
1599 new_ss_mask &= this_ss_mask;
1601 if (new_ss_mask == cur_ss_mask)
1603 cur_ss_mask = new_ss_mask;
1610 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1611 * @kn: the kernfs_node being serviced
1613 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1614 * the method finishes if locking succeeded. Note that once this function
1615 * returns the cgroup returned by cgroup_kn_lock_live() may become
1616 * inaccessible any time. If the caller intends to continue to access the
1617 * cgroup, it should pin it before invoking this function.
1619 void cgroup_kn_unlock(struct kernfs_node *kn)
1621 struct cgroup *cgrp;
1623 if (kernfs_type(kn) == KERNFS_DIR)
1626 cgrp = kn->parent->priv;
1628 mutex_unlock(&cgroup_mutex);
1630 kernfs_unbreak_active_protection(kn);
1635 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1636 * @kn: the kernfs_node being serviced
1637 * @drain_offline: perform offline draining on the cgroup
1639 * This helper is to be used by a cgroup kernfs method currently servicing
1640 * @kn. It breaks the active protection, performs cgroup locking and
1641 * verifies that the associated cgroup is alive. Returns the cgroup if
1642 * alive; otherwise, %NULL. A successful return should be undone by a
1643 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1644 * cgroup is drained of offlining csses before return.
1646 * Any cgroup kernfs method implementation which requires locking the
1647 * associated cgroup should use this helper. It avoids nesting cgroup
1648 * locking under kernfs active protection and allows all kernfs operations
1649 * including self-removal.
1651 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1653 struct cgroup *cgrp;
1655 if (kernfs_type(kn) == KERNFS_DIR)
1658 cgrp = kn->parent->priv;
1661 * We're gonna grab cgroup_mutex which nests outside kernfs
1662 * active_ref. cgroup liveliness check alone provides enough
1663 * protection against removal. Ensure @cgrp stays accessible and
1664 * break the active_ref protection.
1666 if (!cgroup_tryget(cgrp))
1668 kernfs_break_active_protection(kn);
1671 cgroup_lock_and_drain_offline(cgrp);
1673 mutex_lock(&cgroup_mutex);
1675 if (!cgroup_is_dead(cgrp))
1678 cgroup_kn_unlock(kn);
1682 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1684 char name[CGROUP_FILE_NAME_MAX];
1686 lockdep_assert_held(&cgroup_mutex);
1688 if (cft->file_offset) {
1689 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1690 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1692 spin_lock_irq(&cgroup_file_kn_lock);
1694 spin_unlock_irq(&cgroup_file_kn_lock);
1696 del_timer_sync(&cfile->notify_timer);
1699 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1703 * css_clear_dir - remove subsys files in a cgroup directory
1706 static void css_clear_dir(struct cgroup_subsys_state *css)
1708 struct cgroup *cgrp = css->cgroup;
1709 struct cftype *cfts;
1711 if (!(css->flags & CSS_VISIBLE))
1714 css->flags &= ~CSS_VISIBLE;
1717 if (cgroup_on_dfl(cgrp)) {
1718 cgroup_addrm_files(css, cgrp,
1719 cgroup_base_files, false);
1720 if (cgroup_psi_enabled())
1721 cgroup_addrm_files(css, cgrp,
1722 cgroup_psi_files, false);
1724 cgroup_addrm_files(css, cgrp,
1725 cgroup1_base_files, false);
1728 list_for_each_entry(cfts, &css->ss->cfts, node)
1729 cgroup_addrm_files(css, cgrp, cfts, false);
1734 * css_populate_dir - create subsys files in a cgroup directory
1737 * On failure, no file is added.
1739 static int css_populate_dir(struct cgroup_subsys_state *css)
1741 struct cgroup *cgrp = css->cgroup;
1742 struct cftype *cfts, *failed_cfts;
1745 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1749 if (cgroup_on_dfl(cgrp)) {
1750 ret = cgroup_addrm_files(&cgrp->self, cgrp,
1751 cgroup_base_files, true);
1755 if (cgroup_psi_enabled()) {
1756 ret = cgroup_addrm_files(&cgrp->self, cgrp,
1757 cgroup_psi_files, true);
1762 cgroup_addrm_files(css, cgrp,
1763 cgroup1_base_files, true);
1766 list_for_each_entry(cfts, &css->ss->cfts, node) {
1767 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1775 css->flags |= CSS_VISIBLE;
1779 list_for_each_entry(cfts, &css->ss->cfts, node) {
1780 if (cfts == failed_cfts)
1782 cgroup_addrm_files(css, cgrp, cfts, false);
1787 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1789 struct cgroup *dcgrp = &dst_root->cgrp;
1790 struct cgroup_subsys *ss;
1792 u16 dfl_disable_ss_mask = 0;
1794 lockdep_assert_held(&cgroup_mutex);
1796 do_each_subsys_mask(ss, ssid, ss_mask) {
1798 * If @ss has non-root csses attached to it, can't move.
1799 * If @ss is an implicit controller, it is exempt from this
1800 * rule and can be stolen.
1802 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1803 !ss->implicit_on_dfl)
1806 /* can't move between two non-dummy roots either */
1807 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1811 * Collect ssid's that need to be disabled from default
1814 if (ss->root == &cgrp_dfl_root)
1815 dfl_disable_ss_mask |= 1 << ssid;
1817 } while_each_subsys_mask();
1819 if (dfl_disable_ss_mask) {
1820 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1823 * Controllers from default hierarchy that need to be rebound
1824 * are all disabled together in one go.
1826 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1827 WARN_ON(cgroup_apply_control(scgrp));
1828 cgroup_finalize_control(scgrp, 0);
1831 do_each_subsys_mask(ss, ssid, ss_mask) {
1832 struct cgroup_root *src_root = ss->root;
1833 struct cgroup *scgrp = &src_root->cgrp;
1834 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1835 struct css_set *cset;
1837 WARN_ON(!css || cgroup_css(dcgrp, ss));
1839 if (src_root != &cgrp_dfl_root) {
1840 /* disable from the source */
1841 src_root->subsys_mask &= ~(1 << ssid);
1842 WARN_ON(cgroup_apply_control(scgrp));
1843 cgroup_finalize_control(scgrp, 0);
1847 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1848 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1849 ss->root = dst_root;
1850 css->cgroup = dcgrp;
1852 spin_lock_irq(&css_set_lock);
1853 hash_for_each(css_set_table, i, cset, hlist)
1854 list_move_tail(&cset->e_cset_node[ss->id],
1855 &dcgrp->e_csets[ss->id]);
1856 spin_unlock_irq(&css_set_lock);
1858 if (ss->css_rstat_flush) {
1859 list_del_rcu(&css->rstat_css_node);
1861 list_add_rcu(&css->rstat_css_node,
1862 &dcgrp->rstat_css_list);
1865 /* default hierarchy doesn't enable controllers by default */
1866 dst_root->subsys_mask |= 1 << ssid;
1867 if (dst_root == &cgrp_dfl_root) {
1868 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1870 dcgrp->subtree_control |= 1 << ssid;
1871 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1874 ret = cgroup_apply_control(dcgrp);
1876 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1881 } while_each_subsys_mask();
1883 kernfs_activate(dcgrp->kn);
1887 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1888 struct kernfs_root *kf_root)
1892 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1893 struct cgroup *ns_cgroup;
1895 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1899 spin_lock_irq(&css_set_lock);
1900 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1901 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1902 spin_unlock_irq(&css_set_lock);
1904 if (len >= PATH_MAX)
1907 seq_escape(sf, buf, " \t\n\\");
1914 enum cgroup2_param {
1917 Opt_memory_localevents,
1918 Opt_memory_recursiveprot,
1922 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1923 fsparam_flag("nsdelegate", Opt_nsdelegate),
1924 fsparam_flag("favordynmods", Opt_favordynmods),
1925 fsparam_flag("memory_localevents", Opt_memory_localevents),
1926 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1930 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1932 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1933 struct fs_parse_result result;
1936 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1941 case Opt_nsdelegate:
1942 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1944 case Opt_favordynmods:
1945 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1947 case Opt_memory_localevents:
1948 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1950 case Opt_memory_recursiveprot:
1951 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1957 static void apply_cgroup_root_flags(unsigned int root_flags)
1959 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1960 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1961 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1963 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1965 cgroup_favor_dynmods(&cgrp_dfl_root,
1966 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1968 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1969 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1971 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1973 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1974 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1976 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1980 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1982 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1983 seq_puts(seq, ",nsdelegate");
1984 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
1985 seq_puts(seq, ",favordynmods");
1986 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1987 seq_puts(seq, ",memory_localevents");
1988 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1989 seq_puts(seq, ",memory_recursiveprot");
1993 static int cgroup_reconfigure(struct fs_context *fc)
1995 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1997 apply_cgroup_root_flags(ctx->flags);
2001 static void init_cgroup_housekeeping(struct cgroup *cgrp)
2003 struct cgroup_subsys *ss;
2006 INIT_LIST_HEAD(&cgrp->self.sibling);
2007 INIT_LIST_HEAD(&cgrp->self.children);
2008 INIT_LIST_HEAD(&cgrp->cset_links);
2009 INIT_LIST_HEAD(&cgrp->pidlists);
2010 mutex_init(&cgrp->pidlist_mutex);
2011 cgrp->self.cgroup = cgrp;
2012 cgrp->self.flags |= CSS_ONLINE;
2013 cgrp->dom_cgrp = cgrp;
2014 cgrp->max_descendants = INT_MAX;
2015 cgrp->max_depth = INT_MAX;
2016 INIT_LIST_HEAD(&cgrp->rstat_css_list);
2017 prev_cputime_init(&cgrp->prev_cputime);
2019 for_each_subsys(ss, ssid)
2020 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2022 init_waitqueue_head(&cgrp->offline_waitq);
2023 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2026 void init_cgroup_root(struct cgroup_fs_context *ctx)
2028 struct cgroup_root *root = ctx->root;
2029 struct cgroup *cgrp = &root->cgrp;
2031 INIT_LIST_HEAD(&root->root_list);
2032 atomic_set(&root->nr_cgrps, 1);
2034 init_cgroup_housekeeping(cgrp);
2036 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2037 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2038 if (ctx->release_agent)
2039 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2041 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2042 if (ctx->cpuset_clone_children)
2043 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2046 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2048 LIST_HEAD(tmp_links);
2049 struct cgroup *root_cgrp = &root->cgrp;
2050 struct kernfs_syscall_ops *kf_sops;
2051 struct css_set *cset;
2054 lockdep_assert_held(&cgroup_mutex);
2056 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2062 * We're accessing css_set_count without locking css_set_lock here,
2063 * but that's OK - it can only be increased by someone holding
2064 * cgroup_lock, and that's us. Later rebinding may disable
2065 * controllers on the default hierarchy and thus create new csets,
2066 * which can't be more than the existing ones. Allocate 2x.
2068 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2072 ret = cgroup_init_root_id(root);
2076 kf_sops = root == &cgrp_dfl_root ?
2077 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2079 root->kf_root = kernfs_create_root(kf_sops,
2080 KERNFS_ROOT_CREATE_DEACTIVATED |
2081 KERNFS_ROOT_SUPPORT_EXPORTOP |
2082 KERNFS_ROOT_SUPPORT_USER_XATTR,
2084 if (IS_ERR(root->kf_root)) {
2085 ret = PTR_ERR(root->kf_root);
2088 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2089 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2090 root_cgrp->ancestors[0] = root_cgrp;
2092 ret = css_populate_dir(&root_cgrp->self);
2096 ret = cgroup_rstat_init(root_cgrp);
2100 ret = rebind_subsystems(root, ss_mask);
2104 ret = cgroup_bpf_inherit(root_cgrp);
2107 trace_cgroup_setup_root(root);
2110 * There must be no failure case after here, since rebinding takes
2111 * care of subsystems' refcounts, which are explicitly dropped in
2112 * the failure exit path.
2114 list_add(&root->root_list, &cgroup_roots);
2115 cgroup_root_count++;
2118 * Link the root cgroup in this hierarchy into all the css_set
2121 spin_lock_irq(&css_set_lock);
2122 hash_for_each(css_set_table, i, cset, hlist) {
2123 link_css_set(&tmp_links, cset, root_cgrp);
2124 if (css_set_populated(cset))
2125 cgroup_update_populated(root_cgrp, true);
2127 spin_unlock_irq(&css_set_lock);
2129 BUG_ON(!list_empty(&root_cgrp->self.children));
2130 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2136 cgroup_rstat_exit(root_cgrp);
2138 kernfs_destroy_root(root->kf_root);
2139 root->kf_root = NULL;
2141 cgroup_exit_root_id(root);
2143 percpu_ref_exit(&root_cgrp->self.refcnt);
2145 free_cgrp_cset_links(&tmp_links);
2149 int cgroup_do_get_tree(struct fs_context *fc)
2151 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2154 ctx->kfc.root = ctx->root->kf_root;
2155 if (fc->fs_type == &cgroup2_fs_type)
2156 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2158 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2159 ret = kernfs_get_tree(fc);
2162 * In non-init cgroup namespace, instead of root cgroup's dentry,
2163 * we return the dentry corresponding to the cgroupns->root_cgrp.
2165 if (!ret && ctx->ns != &init_cgroup_ns) {
2166 struct dentry *nsdentry;
2167 struct super_block *sb = fc->root->d_sb;
2168 struct cgroup *cgrp;
2170 mutex_lock(&cgroup_mutex);
2171 spin_lock_irq(&css_set_lock);
2173 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2175 spin_unlock_irq(&css_set_lock);
2176 mutex_unlock(&cgroup_mutex);
2178 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2180 if (IS_ERR(nsdentry)) {
2181 deactivate_locked_super(sb);
2182 ret = PTR_ERR(nsdentry);
2185 fc->root = nsdentry;
2188 if (!ctx->kfc.new_sb_created)
2189 cgroup_put(&ctx->root->cgrp);
2195 * Destroy a cgroup filesystem context.
2197 static void cgroup_fs_context_free(struct fs_context *fc)
2199 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2202 kfree(ctx->release_agent);
2203 put_cgroup_ns(ctx->ns);
2204 kernfs_free_fs_context(fc);
2208 static int cgroup_get_tree(struct fs_context *fc)
2210 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2213 WRITE_ONCE(cgrp_dfl_visible, true);
2214 cgroup_get_live(&cgrp_dfl_root.cgrp);
2215 ctx->root = &cgrp_dfl_root;
2217 ret = cgroup_do_get_tree(fc);
2219 apply_cgroup_root_flags(ctx->flags);
2223 static const struct fs_context_operations cgroup_fs_context_ops = {
2224 .free = cgroup_fs_context_free,
2225 .parse_param = cgroup2_parse_param,
2226 .get_tree = cgroup_get_tree,
2227 .reconfigure = cgroup_reconfigure,
2230 static const struct fs_context_operations cgroup1_fs_context_ops = {
2231 .free = cgroup_fs_context_free,
2232 .parse_param = cgroup1_parse_param,
2233 .get_tree = cgroup1_get_tree,
2234 .reconfigure = cgroup1_reconfigure,
2238 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2239 * we select the namespace we're going to use.
2241 static int cgroup_init_fs_context(struct fs_context *fc)
2243 struct cgroup_fs_context *ctx;
2245 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2249 ctx->ns = current->nsproxy->cgroup_ns;
2250 get_cgroup_ns(ctx->ns);
2251 fc->fs_private = &ctx->kfc;
2252 if (fc->fs_type == &cgroup2_fs_type)
2253 fc->ops = &cgroup_fs_context_ops;
2255 fc->ops = &cgroup1_fs_context_ops;
2256 put_user_ns(fc->user_ns);
2257 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2260 #ifdef CONFIG_CGROUP_FAVOR_DYNMODS
2261 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2266 static void cgroup_kill_sb(struct super_block *sb)
2268 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2269 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2272 * If @root doesn't have any children, start killing it.
2273 * This prevents new mounts by disabling percpu_ref_tryget_live().
2275 * And don't kill the default root.
2277 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2278 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2279 cgroup_bpf_offline(&root->cgrp);
2280 percpu_ref_kill(&root->cgrp.self.refcnt);
2282 cgroup_put(&root->cgrp);
2286 struct file_system_type cgroup_fs_type = {
2288 .init_fs_context = cgroup_init_fs_context,
2289 .parameters = cgroup1_fs_parameters,
2290 .kill_sb = cgroup_kill_sb,
2291 .fs_flags = FS_USERNS_MOUNT,
2294 static struct file_system_type cgroup2_fs_type = {
2296 .init_fs_context = cgroup_init_fs_context,
2297 .parameters = cgroup2_fs_parameters,
2298 .kill_sb = cgroup_kill_sb,
2299 .fs_flags = FS_USERNS_MOUNT,
2302 #ifdef CONFIG_CPUSETS
2303 static const struct fs_context_operations cpuset_fs_context_ops = {
2304 .get_tree = cgroup1_get_tree,
2305 .free = cgroup_fs_context_free,
2309 * This is ugly, but preserves the userspace API for existing cpuset
2310 * users. If someone tries to mount the "cpuset" filesystem, we
2311 * silently switch it to mount "cgroup" instead
2313 static int cpuset_init_fs_context(struct fs_context *fc)
2315 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2316 struct cgroup_fs_context *ctx;
2319 err = cgroup_init_fs_context(fc);
2325 fc->ops = &cpuset_fs_context_ops;
2327 ctx = cgroup_fc2context(fc);
2328 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2329 ctx->flags |= CGRP_ROOT_NOPREFIX;
2330 ctx->release_agent = agent;
2332 get_filesystem(&cgroup_fs_type);
2333 put_filesystem(fc->fs_type);
2334 fc->fs_type = &cgroup_fs_type;
2339 static struct file_system_type cpuset_fs_type = {
2341 .init_fs_context = cpuset_init_fs_context,
2342 .fs_flags = FS_USERNS_MOUNT,
2346 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2347 struct cgroup_namespace *ns)
2349 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2351 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2354 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2355 struct cgroup_namespace *ns)
2359 mutex_lock(&cgroup_mutex);
2360 spin_lock_irq(&css_set_lock);
2362 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2364 spin_unlock_irq(&css_set_lock);
2365 mutex_unlock(&cgroup_mutex);
2369 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2372 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2373 * @task: target task
2374 * @buf: the buffer to write the path into
2375 * @buflen: the length of the buffer
2377 * Determine @task's cgroup on the first (the one with the lowest non-zero
2378 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2379 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2380 * cgroup controller callbacks.
2382 * Return value is the same as kernfs_path().
2384 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2386 struct cgroup_root *root;
2387 struct cgroup *cgrp;
2388 int hierarchy_id = 1;
2391 mutex_lock(&cgroup_mutex);
2392 spin_lock_irq(&css_set_lock);
2394 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2397 cgrp = task_cgroup_from_root(task, root);
2398 ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2400 /* if no hierarchy exists, everyone is in "/" */
2401 ret = strscpy(buf, "/", buflen);
2404 spin_unlock_irq(&css_set_lock);
2405 mutex_unlock(&cgroup_mutex);
2408 EXPORT_SYMBOL_GPL(task_cgroup_path);
2411 * cgroup_attach_lock - Lock for ->attach()
2412 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2414 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2415 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2416 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2417 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2418 * lead to deadlocks.
2420 * Bringing up a CPU may involve creating and destroying tasks which requires
2421 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2422 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2423 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2424 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2425 * the threadgroup_rwsem to be released to create new tasks. For more details:
2427 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2429 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2430 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2431 * CPU hotplug is disabled on entry.
2433 void cgroup_attach_lock(bool lock_threadgroup)
2436 if (lock_threadgroup)
2437 percpu_down_write(&cgroup_threadgroup_rwsem);
2441 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2442 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2444 void cgroup_attach_unlock(bool lock_threadgroup)
2446 if (lock_threadgroup)
2447 percpu_up_write(&cgroup_threadgroup_rwsem);
2452 * cgroup_migrate_add_task - add a migration target task to a migration context
2453 * @task: target task
2454 * @mgctx: target migration context
2456 * Add @task, which is a migration target, to @mgctx->tset. This function
2457 * becomes noop if @task doesn't need to be migrated. @task's css_set
2458 * should have been added as a migration source and @task->cg_list will be
2459 * moved from the css_set's tasks list to mg_tasks one.
2461 static void cgroup_migrate_add_task(struct task_struct *task,
2462 struct cgroup_mgctx *mgctx)
2464 struct css_set *cset;
2466 lockdep_assert_held(&css_set_lock);
2468 /* @task either already exited or can't exit until the end */
2469 if (task->flags & PF_EXITING)
2472 /* cgroup_threadgroup_rwsem protects racing against forks */
2473 WARN_ON_ONCE(list_empty(&task->cg_list));
2475 cset = task_css_set(task);
2476 if (!cset->mg_src_cgrp)
2479 mgctx->tset.nr_tasks++;
2481 list_move_tail(&task->cg_list, &cset->mg_tasks);
2482 if (list_empty(&cset->mg_node))
2483 list_add_tail(&cset->mg_node,
2484 &mgctx->tset.src_csets);
2485 if (list_empty(&cset->mg_dst_cset->mg_node))
2486 list_add_tail(&cset->mg_dst_cset->mg_node,
2487 &mgctx->tset.dst_csets);
2491 * cgroup_taskset_first - reset taskset and return the first task
2492 * @tset: taskset of interest
2493 * @dst_cssp: output variable for the destination css
2495 * @tset iteration is initialized and the first task is returned.
2497 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2498 struct cgroup_subsys_state **dst_cssp)
2500 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2501 tset->cur_task = NULL;
2503 return cgroup_taskset_next(tset, dst_cssp);
2507 * cgroup_taskset_next - iterate to the next task in taskset
2508 * @tset: taskset of interest
2509 * @dst_cssp: output variable for the destination css
2511 * Return the next task in @tset. Iteration must have been initialized
2512 * with cgroup_taskset_first().
2514 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2515 struct cgroup_subsys_state **dst_cssp)
2517 struct css_set *cset = tset->cur_cset;
2518 struct task_struct *task = tset->cur_task;
2520 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2522 task = list_first_entry(&cset->mg_tasks,
2523 struct task_struct, cg_list);
2525 task = list_next_entry(task, cg_list);
2527 if (&task->cg_list != &cset->mg_tasks) {
2528 tset->cur_cset = cset;
2529 tset->cur_task = task;
2532 * This function may be called both before and
2533 * after cgroup_taskset_migrate(). The two cases
2534 * can be distinguished by looking at whether @cset
2535 * has its ->mg_dst_cset set.
2537 if (cset->mg_dst_cset)
2538 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2540 *dst_cssp = cset->subsys[tset->ssid];
2545 cset = list_next_entry(cset, mg_node);
2553 * cgroup_migrate_execute - migrate a taskset
2554 * @mgctx: migration context
2556 * Migrate tasks in @mgctx as setup by migration preparation functions.
2557 * This function fails iff one of the ->can_attach callbacks fails and
2558 * guarantees that either all or none of the tasks in @mgctx are migrated.
2559 * @mgctx is consumed regardless of success.
2561 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2563 struct cgroup_taskset *tset = &mgctx->tset;
2564 struct cgroup_subsys *ss;
2565 struct task_struct *task, *tmp_task;
2566 struct css_set *cset, *tmp_cset;
2567 int ssid, failed_ssid, ret;
2569 /* check that we can legitimately attach to the cgroup */
2570 if (tset->nr_tasks) {
2571 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2572 if (ss->can_attach) {
2574 ret = ss->can_attach(tset);
2577 goto out_cancel_attach;
2580 } while_each_subsys_mask();
2584 * Now that we're guaranteed success, proceed to move all tasks to
2585 * the new cgroup. There are no failure cases after here, so this
2586 * is the commit point.
2588 spin_lock_irq(&css_set_lock);
2589 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2590 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2591 struct css_set *from_cset = task_css_set(task);
2592 struct css_set *to_cset = cset->mg_dst_cset;
2594 get_css_set(to_cset);
2595 to_cset->nr_tasks++;
2596 css_set_move_task(task, from_cset, to_cset, true);
2597 from_cset->nr_tasks--;
2599 * If the source or destination cgroup is frozen,
2600 * the task might require to change its state.
2602 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2604 put_css_set_locked(from_cset);
2608 spin_unlock_irq(&css_set_lock);
2611 * Migration is committed, all target tasks are now on dst_csets.
2612 * Nothing is sensitive to fork() after this point. Notify
2613 * controllers that migration is complete.
2615 tset->csets = &tset->dst_csets;
2617 if (tset->nr_tasks) {
2618 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2623 } while_each_subsys_mask();
2627 goto out_release_tset;
2630 if (tset->nr_tasks) {
2631 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2632 if (ssid == failed_ssid)
2634 if (ss->cancel_attach) {
2636 ss->cancel_attach(tset);
2638 } while_each_subsys_mask();
2641 spin_lock_irq(&css_set_lock);
2642 list_splice_init(&tset->dst_csets, &tset->src_csets);
2643 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2644 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2645 list_del_init(&cset->mg_node);
2647 spin_unlock_irq(&css_set_lock);
2650 * Re-initialize the cgroup_taskset structure in case it is reused
2651 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2655 tset->csets = &tset->src_csets;
2660 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2661 * @dst_cgrp: destination cgroup to test
2663 * On the default hierarchy, except for the mixable, (possible) thread root
2664 * and threaded cgroups, subtree_control must be zero for migration
2665 * destination cgroups with tasks so that child cgroups don't compete
2668 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2670 /* v1 doesn't have any restriction */
2671 if (!cgroup_on_dfl(dst_cgrp))
2674 /* verify @dst_cgrp can host resources */
2675 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2679 * If @dst_cgrp is already or can become a thread root or is
2680 * threaded, it doesn't matter.
2682 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2685 /* apply no-internal-process constraint */
2686 if (dst_cgrp->subtree_control)
2693 * cgroup_migrate_finish - cleanup after attach
2694 * @mgctx: migration context
2696 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2697 * those functions for details.
2699 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2701 struct css_set *cset, *tmp_cset;
2703 lockdep_assert_held(&cgroup_mutex);
2705 spin_lock_irq(&css_set_lock);
2707 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2708 mg_src_preload_node) {
2709 cset->mg_src_cgrp = NULL;
2710 cset->mg_dst_cgrp = NULL;
2711 cset->mg_dst_cset = NULL;
2712 list_del_init(&cset->mg_src_preload_node);
2713 put_css_set_locked(cset);
2716 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2717 mg_dst_preload_node) {
2718 cset->mg_src_cgrp = NULL;
2719 cset->mg_dst_cgrp = NULL;
2720 cset->mg_dst_cset = NULL;
2721 list_del_init(&cset->mg_dst_preload_node);
2722 put_css_set_locked(cset);
2725 spin_unlock_irq(&css_set_lock);
2729 * cgroup_migrate_add_src - add a migration source css_set
2730 * @src_cset: the source css_set to add
2731 * @dst_cgrp: the destination cgroup
2732 * @mgctx: migration context
2734 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2735 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2736 * up by cgroup_migrate_finish().
2738 * This function may be called without holding cgroup_threadgroup_rwsem
2739 * even if the target is a process. Threads may be created and destroyed
2740 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2741 * into play and the preloaded css_sets are guaranteed to cover all
2744 void cgroup_migrate_add_src(struct css_set *src_cset,
2745 struct cgroup *dst_cgrp,
2746 struct cgroup_mgctx *mgctx)
2748 struct cgroup *src_cgrp;
2750 lockdep_assert_held(&cgroup_mutex);
2751 lockdep_assert_held(&css_set_lock);
2754 * If ->dead, @src_set is associated with one or more dead cgroups
2755 * and doesn't contain any migratable tasks. Ignore it early so
2756 * that the rest of migration path doesn't get confused by it.
2761 if (!list_empty(&src_cset->mg_src_preload_node))
2764 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2766 WARN_ON(src_cset->mg_src_cgrp);
2767 WARN_ON(src_cset->mg_dst_cgrp);
2768 WARN_ON(!list_empty(&src_cset->mg_tasks));
2769 WARN_ON(!list_empty(&src_cset->mg_node));
2771 src_cset->mg_src_cgrp = src_cgrp;
2772 src_cset->mg_dst_cgrp = dst_cgrp;
2773 get_css_set(src_cset);
2774 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2778 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2779 * @mgctx: migration context
2781 * Tasks are about to be moved and all the source css_sets have been
2782 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2783 * pins all destination css_sets, links each to its source, and append them
2784 * to @mgctx->preloaded_dst_csets.
2786 * This function must be called after cgroup_migrate_add_src() has been
2787 * called on each migration source css_set. After migration is performed
2788 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2791 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2793 struct css_set *src_cset, *tmp_cset;
2795 lockdep_assert_held(&cgroup_mutex);
2797 /* look up the dst cset for each src cset and link it to src */
2798 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2799 mg_src_preload_node) {
2800 struct css_set *dst_cset;
2801 struct cgroup_subsys *ss;
2804 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2808 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2811 * If src cset equals dst, it's noop. Drop the src.
2812 * cgroup_migrate() will skip the cset too. Note that we
2813 * can't handle src == dst as some nodes are used by both.
2815 if (src_cset == dst_cset) {
2816 src_cset->mg_src_cgrp = NULL;
2817 src_cset->mg_dst_cgrp = NULL;
2818 list_del_init(&src_cset->mg_src_preload_node);
2819 put_css_set(src_cset);
2820 put_css_set(dst_cset);
2824 src_cset->mg_dst_cset = dst_cset;
2826 if (list_empty(&dst_cset->mg_dst_preload_node))
2827 list_add_tail(&dst_cset->mg_dst_preload_node,
2828 &mgctx->preloaded_dst_csets);
2830 put_css_set(dst_cset);
2832 for_each_subsys(ss, ssid)
2833 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2834 mgctx->ss_mask |= 1 << ssid;
2841 * cgroup_migrate - migrate a process or task to a cgroup
2842 * @leader: the leader of the process or the task to migrate
2843 * @threadgroup: whether @leader points to the whole process or a single task
2844 * @mgctx: migration context
2846 * Migrate a process or task denoted by @leader. If migrating a process,
2847 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2848 * responsible for invoking cgroup_migrate_add_src() and
2849 * cgroup_migrate_prepare_dst() on the targets before invoking this
2850 * function and following up with cgroup_migrate_finish().
2852 * As long as a controller's ->can_attach() doesn't fail, this function is
2853 * guaranteed to succeed. This means that, excluding ->can_attach()
2854 * failure, when migrating multiple targets, the success or failure can be
2855 * decided for all targets by invoking group_migrate_prepare_dst() before
2856 * actually starting migrating.
2858 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2859 struct cgroup_mgctx *mgctx)
2861 struct task_struct *task;
2864 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2865 * already PF_EXITING could be freed from underneath us unless we
2866 * take an rcu_read_lock.
2868 spin_lock_irq(&css_set_lock);
2871 cgroup_migrate_add_task(task, mgctx);
2874 } while_each_thread(leader, task);
2875 spin_unlock_irq(&css_set_lock);
2877 return cgroup_migrate_execute(mgctx);
2881 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2882 * @dst_cgrp: the cgroup to attach to
2883 * @leader: the task or the leader of the threadgroup to be attached
2884 * @threadgroup: attach the whole threadgroup?
2886 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2888 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2891 DEFINE_CGROUP_MGCTX(mgctx);
2892 struct task_struct *task;
2895 /* look up all src csets */
2896 spin_lock_irq(&css_set_lock);
2900 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2903 } while_each_thread(leader, task);
2905 spin_unlock_irq(&css_set_lock);
2907 /* prepare dst csets and commit */
2908 ret = cgroup_migrate_prepare_dst(&mgctx);
2910 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2912 cgroup_migrate_finish(&mgctx);
2915 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2920 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2921 bool *threadgroup_locked)
2923 struct task_struct *tsk;
2926 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2927 return ERR_PTR(-EINVAL);
2930 * If we migrate a single thread, we don't care about threadgroup
2931 * stability. If the thread is `current`, it won't exit(2) under our
2932 * hands or change PID through exec(2). We exclude
2933 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2934 * callers by cgroup_mutex.
2935 * Therefore, we can skip the global lock.
2937 lockdep_assert_held(&cgroup_mutex);
2938 *threadgroup_locked = pid || threadgroup;
2939 cgroup_attach_lock(*threadgroup_locked);
2943 tsk = find_task_by_vpid(pid);
2945 tsk = ERR_PTR(-ESRCH);
2946 goto out_unlock_threadgroup;
2953 tsk = tsk->group_leader;
2956 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2957 * If userland migrates such a kthread to a non-root cgroup, it can
2958 * become trapped in a cpuset, or RT kthread may be born in a
2959 * cgroup with no rt_runtime allocated. Just say no.
2961 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2962 tsk = ERR_PTR(-EINVAL);
2963 goto out_unlock_threadgroup;
2966 get_task_struct(tsk);
2967 goto out_unlock_rcu;
2969 out_unlock_threadgroup:
2970 cgroup_attach_unlock(*threadgroup_locked);
2971 *threadgroup_locked = false;
2977 void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2979 struct cgroup_subsys *ss;
2982 /* release reference from cgroup_procs_write_start() */
2983 put_task_struct(task);
2985 cgroup_attach_unlock(threadgroup_locked);
2987 for_each_subsys(ss, ssid)
2988 if (ss->post_attach)
2992 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2994 struct cgroup_subsys *ss;
2995 bool printed = false;
2998 do_each_subsys_mask(ss, ssid, ss_mask) {
3001 seq_puts(seq, ss->name);
3003 } while_each_subsys_mask();
3005 seq_putc(seq, '\n');
3008 /* show controllers which are enabled from the parent */
3009 static int cgroup_controllers_show(struct seq_file *seq, void *v)
3011 struct cgroup *cgrp = seq_css(seq)->cgroup;
3013 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
3017 /* show controllers which are enabled for a given cgroup's children */
3018 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
3020 struct cgroup *cgrp = seq_css(seq)->cgroup;
3022 cgroup_print_ss_mask(seq, cgrp->subtree_control);
3027 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
3028 * @cgrp: root of the subtree to update csses for
3030 * @cgrp's control masks have changed and its subtree's css associations
3031 * need to be updated accordingly. This function looks up all css_sets
3032 * which are attached to the subtree, creates the matching updated css_sets
3033 * and migrates the tasks to the new ones.
3035 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3037 DEFINE_CGROUP_MGCTX(mgctx);
3038 struct cgroup_subsys_state *d_css;
3039 struct cgroup *dsct;
3040 struct css_set *src_cset;
3044 lockdep_assert_held(&cgroup_mutex);
3046 /* look up all csses currently attached to @cgrp's subtree */
3047 spin_lock_irq(&css_set_lock);
3048 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3049 struct cgrp_cset_link *link;
3052 * As cgroup_update_dfl_csses() is only called by
3053 * cgroup_apply_control(). The csses associated with the
3054 * given cgrp will not be affected by changes made to
3055 * its subtree_control file. We can skip them.
3060 list_for_each_entry(link, &dsct->cset_links, cset_link)
3061 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3063 spin_unlock_irq(&css_set_lock);
3066 * We need to write-lock threadgroup_rwsem while migrating tasks.
3067 * However, if there are no source csets for @cgrp, changing its
3068 * controllers isn't gonna produce any task migrations and the
3069 * write-locking can be skipped safely.
3071 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3072 cgroup_attach_lock(has_tasks);
3074 /* NULL dst indicates self on default hierarchy */
3075 ret = cgroup_migrate_prepare_dst(&mgctx);
3079 spin_lock_irq(&css_set_lock);
3080 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3081 mg_src_preload_node) {
3082 struct task_struct *task, *ntask;
3084 /* all tasks in src_csets need to be migrated */
3085 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3086 cgroup_migrate_add_task(task, &mgctx);
3088 spin_unlock_irq(&css_set_lock);
3090 ret = cgroup_migrate_execute(&mgctx);
3092 cgroup_migrate_finish(&mgctx);
3093 cgroup_attach_unlock(has_tasks);
3098 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3099 * @cgrp: root of the target subtree
3101 * Because css offlining is asynchronous, userland may try to re-enable a
3102 * controller while the previous css is still around. This function grabs
3103 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3105 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3106 __acquires(&cgroup_mutex)
3108 struct cgroup *dsct;
3109 struct cgroup_subsys_state *d_css;
3110 struct cgroup_subsys *ss;
3114 mutex_lock(&cgroup_mutex);
3116 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3117 for_each_subsys(ss, ssid) {
3118 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3121 if (!css || !percpu_ref_is_dying(&css->refcnt))
3124 cgroup_get_live(dsct);
3125 prepare_to_wait(&dsct->offline_waitq, &wait,
3126 TASK_UNINTERRUPTIBLE);
3128 mutex_unlock(&cgroup_mutex);
3130 finish_wait(&dsct->offline_waitq, &wait);
3139 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3140 * @cgrp: root of the target subtree
3142 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3143 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3146 static void cgroup_save_control(struct cgroup *cgrp)
3148 struct cgroup *dsct;
3149 struct cgroup_subsys_state *d_css;
3151 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3152 dsct->old_subtree_control = dsct->subtree_control;
3153 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3154 dsct->old_dom_cgrp = dsct->dom_cgrp;
3159 * cgroup_propagate_control - refresh control masks of a subtree
3160 * @cgrp: root of the target subtree
3162 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3163 * ->subtree_control and propagate controller availability through the
3164 * subtree so that descendants don't have unavailable controllers enabled.
3166 static void cgroup_propagate_control(struct cgroup *cgrp)
3168 struct cgroup *dsct;
3169 struct cgroup_subsys_state *d_css;
3171 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3172 dsct->subtree_control &= cgroup_control(dsct);
3173 dsct->subtree_ss_mask =
3174 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3175 cgroup_ss_mask(dsct));
3180 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3181 * @cgrp: root of the target subtree
3183 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3184 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3187 static void cgroup_restore_control(struct cgroup *cgrp)
3189 struct cgroup *dsct;
3190 struct cgroup_subsys_state *d_css;
3192 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3193 dsct->subtree_control = dsct->old_subtree_control;
3194 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3195 dsct->dom_cgrp = dsct->old_dom_cgrp;
3199 static bool css_visible(struct cgroup_subsys_state *css)
3201 struct cgroup_subsys *ss = css->ss;
3202 struct cgroup *cgrp = css->cgroup;
3204 if (cgroup_control(cgrp) & (1 << ss->id))
3206 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3208 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3212 * cgroup_apply_control_enable - enable or show csses according to control
3213 * @cgrp: root of the target subtree
3215 * Walk @cgrp's subtree and create new csses or make the existing ones
3216 * visible. A css is created invisible if it's being implicitly enabled
3217 * through dependency. An invisible css is made visible when the userland
3218 * explicitly enables it.
3220 * Returns 0 on success, -errno on failure. On failure, csses which have
3221 * been processed already aren't cleaned up. The caller is responsible for
3222 * cleaning up with cgroup_apply_control_disable().
3224 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3226 struct cgroup *dsct;
3227 struct cgroup_subsys_state *d_css;
3228 struct cgroup_subsys *ss;
3231 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3232 for_each_subsys(ss, ssid) {
3233 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3235 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3239 css = css_create(dsct, ss);
3241 return PTR_ERR(css);
3244 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3246 if (css_visible(css)) {
3247 ret = css_populate_dir(css);
3258 * cgroup_apply_control_disable - kill or hide csses according to control
3259 * @cgrp: root of the target subtree
3261 * Walk @cgrp's subtree and kill and hide csses so that they match
3262 * cgroup_ss_mask() and cgroup_visible_mask().
3264 * A css is hidden when the userland requests it to be disabled while other
3265 * subsystems are still depending on it. The css must not actively control
3266 * resources and be in the vanilla state if it's made visible again later.
3267 * Controllers which may be depended upon should provide ->css_reset() for
3270 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3272 struct cgroup *dsct;
3273 struct cgroup_subsys_state *d_css;
3274 struct cgroup_subsys *ss;
3277 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3278 for_each_subsys(ss, ssid) {
3279 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3284 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3287 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3289 } else if (!css_visible(css)) {
3299 * cgroup_apply_control - apply control mask updates to the subtree
3300 * @cgrp: root of the target subtree
3302 * subsystems can be enabled and disabled in a subtree using the following
3305 * 1. Call cgroup_save_control() to stash the current state.
3306 * 2. Update ->subtree_control masks in the subtree as desired.
3307 * 3. Call cgroup_apply_control() to apply the changes.
3308 * 4. Optionally perform other related operations.
3309 * 5. Call cgroup_finalize_control() to finish up.
3311 * This function implements step 3 and propagates the mask changes
3312 * throughout @cgrp's subtree, updates csses accordingly and perform
3313 * process migrations.
3315 static int cgroup_apply_control(struct cgroup *cgrp)
3319 cgroup_propagate_control(cgrp);
3321 ret = cgroup_apply_control_enable(cgrp);
3326 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3327 * making the following cgroup_update_dfl_csses() properly update
3328 * css associations of all tasks in the subtree.
3330 return cgroup_update_dfl_csses(cgrp);
3334 * cgroup_finalize_control - finalize control mask update
3335 * @cgrp: root of the target subtree
3336 * @ret: the result of the update
3338 * Finalize control mask update. See cgroup_apply_control() for more info.
3340 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3343 cgroup_restore_control(cgrp);
3344 cgroup_propagate_control(cgrp);
3347 cgroup_apply_control_disable(cgrp);
3350 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3352 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3354 /* if nothing is getting enabled, nothing to worry about */
3358 /* can @cgrp host any resources? */
3359 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3362 /* mixables don't care */
3363 if (cgroup_is_mixable(cgrp))
3366 if (domain_enable) {
3367 /* can't enable domain controllers inside a thread subtree */
3368 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3372 * Threaded controllers can handle internal competitions
3373 * and are always allowed inside a (prospective) thread
3376 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3381 * Controllers can't be enabled for a cgroup with tasks to avoid
3382 * child cgroups competing against tasks.
3384 if (cgroup_has_tasks(cgrp))
3390 /* change the enabled child controllers for a cgroup in the default hierarchy */
3391 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3392 char *buf, size_t nbytes,
3395 u16 enable = 0, disable = 0;
3396 struct cgroup *cgrp, *child;
3397 struct cgroup_subsys *ss;
3402 * Parse input - space separated list of subsystem names prefixed
3403 * with either + or -.
3405 buf = strstrip(buf);
3406 while ((tok = strsep(&buf, " "))) {
3409 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3410 if (!cgroup_ssid_enabled(ssid) ||
3411 strcmp(tok + 1, ss->name))
3415 enable |= 1 << ssid;
3416 disable &= ~(1 << ssid);
3417 } else if (*tok == '-') {
3418 disable |= 1 << ssid;
3419 enable &= ~(1 << ssid);
3424 } while_each_subsys_mask();
3425 if (ssid == CGROUP_SUBSYS_COUNT)
3429 cgrp = cgroup_kn_lock_live(of->kn, true);
3433 for_each_subsys(ss, ssid) {
3434 if (enable & (1 << ssid)) {
3435 if (cgrp->subtree_control & (1 << ssid)) {
3436 enable &= ~(1 << ssid);
3440 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3444 } else if (disable & (1 << ssid)) {
3445 if (!(cgrp->subtree_control & (1 << ssid))) {
3446 disable &= ~(1 << ssid);
3450 /* a child has it enabled? */
3451 cgroup_for_each_live_child(child, cgrp) {
3452 if (child->subtree_control & (1 << ssid)) {
3460 if (!enable && !disable) {
3465 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3469 /* save and update control masks and prepare csses */
3470 cgroup_save_control(cgrp);
3472 cgrp->subtree_control |= enable;
3473 cgrp->subtree_control &= ~disable;
3475 ret = cgroup_apply_control(cgrp);
3476 cgroup_finalize_control(cgrp, ret);
3480 kernfs_activate(cgrp->kn);
3482 cgroup_kn_unlock(of->kn);
3483 return ret ?: nbytes;
3487 * cgroup_enable_threaded - make @cgrp threaded
3488 * @cgrp: the target cgroup
3490 * Called when "threaded" is written to the cgroup.type interface file and
3491 * tries to make @cgrp threaded and join the parent's resource domain.
3492 * This function is never called on the root cgroup as cgroup.type doesn't
3495 static int cgroup_enable_threaded(struct cgroup *cgrp)
3497 struct cgroup *parent = cgroup_parent(cgrp);
3498 struct cgroup *dom_cgrp = parent->dom_cgrp;
3499 struct cgroup *dsct;
3500 struct cgroup_subsys_state *d_css;
3503 lockdep_assert_held(&cgroup_mutex);
3505 /* noop if already threaded */
3506 if (cgroup_is_threaded(cgrp))
3510 * If @cgroup is populated or has domain controllers enabled, it
3511 * can't be switched. While the below cgroup_can_be_thread_root()
3512 * test can catch the same conditions, that's only when @parent is
3513 * not mixable, so let's check it explicitly.
3515 if (cgroup_is_populated(cgrp) ||
3516 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3519 /* we're joining the parent's domain, ensure its validity */
3520 if (!cgroup_is_valid_domain(dom_cgrp) ||
3521 !cgroup_can_be_thread_root(dom_cgrp))
3525 * The following shouldn't cause actual migrations and should
3528 cgroup_save_control(cgrp);
3530 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3531 if (dsct == cgrp || cgroup_is_threaded(dsct))
3532 dsct->dom_cgrp = dom_cgrp;
3534 ret = cgroup_apply_control(cgrp);
3536 parent->nr_threaded_children++;
3538 cgroup_finalize_control(cgrp, ret);
3542 static int cgroup_type_show(struct seq_file *seq, void *v)
3544 struct cgroup *cgrp = seq_css(seq)->cgroup;
3546 if (cgroup_is_threaded(cgrp))
3547 seq_puts(seq, "threaded\n");
3548 else if (!cgroup_is_valid_domain(cgrp))
3549 seq_puts(seq, "domain invalid\n");
3550 else if (cgroup_is_thread_root(cgrp))
3551 seq_puts(seq, "domain threaded\n");
3553 seq_puts(seq, "domain\n");
3558 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3559 size_t nbytes, loff_t off)
3561 struct cgroup *cgrp;
3564 /* only switching to threaded mode is supported */
3565 if (strcmp(strstrip(buf), "threaded"))
3568 /* drain dying csses before we re-apply (threaded) subtree control */
3569 cgrp = cgroup_kn_lock_live(of->kn, true);
3573 /* threaded can only be enabled */
3574 ret = cgroup_enable_threaded(cgrp);
3576 cgroup_kn_unlock(of->kn);
3577 return ret ?: nbytes;
3580 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3582 struct cgroup *cgrp = seq_css(seq)->cgroup;
3583 int descendants = READ_ONCE(cgrp->max_descendants);
3585 if (descendants == INT_MAX)
3586 seq_puts(seq, "max\n");
3588 seq_printf(seq, "%d\n", descendants);
3593 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3594 char *buf, size_t nbytes, loff_t off)
3596 struct cgroup *cgrp;
3600 buf = strstrip(buf);
3601 if (!strcmp(buf, "max")) {
3602 descendants = INT_MAX;
3604 ret = kstrtoint(buf, 0, &descendants);
3609 if (descendants < 0)
3612 cgrp = cgroup_kn_lock_live(of->kn, false);
3616 cgrp->max_descendants = descendants;
3618 cgroup_kn_unlock(of->kn);
3623 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3625 struct cgroup *cgrp = seq_css(seq)->cgroup;
3626 int depth = READ_ONCE(cgrp->max_depth);
3628 if (depth == INT_MAX)
3629 seq_puts(seq, "max\n");
3631 seq_printf(seq, "%d\n", depth);
3636 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3637 char *buf, size_t nbytes, loff_t off)
3639 struct cgroup *cgrp;
3643 buf = strstrip(buf);
3644 if (!strcmp(buf, "max")) {
3647 ret = kstrtoint(buf, 0, &depth);
3655 cgrp = cgroup_kn_lock_live(of->kn, false);
3659 cgrp->max_depth = depth;
3661 cgroup_kn_unlock(of->kn);
3666 static int cgroup_events_show(struct seq_file *seq, void *v)
3668 struct cgroup *cgrp = seq_css(seq)->cgroup;
3670 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3671 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3676 static int cgroup_stat_show(struct seq_file *seq, void *v)
3678 struct cgroup *cgroup = seq_css(seq)->cgroup;
3680 seq_printf(seq, "nr_descendants %d\n",
3681 cgroup->nr_descendants);
3682 seq_printf(seq, "nr_dying_descendants %d\n",
3683 cgroup->nr_dying_descendants);
3688 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3689 struct cgroup *cgrp, int ssid)
3691 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3692 struct cgroup_subsys_state *css;
3695 if (!ss->css_extra_stat_show)
3698 css = cgroup_tryget_css(cgrp, ss);
3702 ret = ss->css_extra_stat_show(seq, css);
3707 static int cpu_stat_show(struct seq_file *seq, void *v)
3709 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3712 cgroup_base_stat_cputime_show(seq);
3713 #ifdef CONFIG_CGROUP_SCHED
3714 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3720 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3722 struct cgroup *cgrp = seq_css(seq)->cgroup;
3723 struct psi_group *psi = cgroup_psi(cgrp);
3725 return psi_show(seq, psi, PSI_IO);
3727 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3729 struct cgroup *cgrp = seq_css(seq)->cgroup;
3730 struct psi_group *psi = cgroup_psi(cgrp);
3732 return psi_show(seq, psi, PSI_MEM);
3734 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3736 struct cgroup *cgrp = seq_css(seq)->cgroup;
3737 struct psi_group *psi = cgroup_psi(cgrp);
3739 return psi_show(seq, psi, PSI_CPU);
3742 static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3743 size_t nbytes, enum psi_res res)
3745 struct cgroup_file_ctx *ctx = of->priv;
3746 struct psi_trigger *new;
3747 struct cgroup *cgrp;
3748 struct psi_group *psi;
3750 cgrp = cgroup_kn_lock_live(of->kn, false);
3755 cgroup_kn_unlock(of->kn);
3757 /* Allow only one trigger per file descriptor */
3758 if (ctx->psi.trigger) {
3763 psi = cgroup_psi(cgrp);
3764 new = psi_trigger_create(psi, buf, res);
3767 return PTR_ERR(new);
3770 smp_store_release(&ctx->psi.trigger, new);
3776 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3777 char *buf, size_t nbytes,
3780 return pressure_write(of, buf, nbytes, PSI_IO);
3783 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3784 char *buf, size_t nbytes,
3787 return pressure_write(of, buf, nbytes, PSI_MEM);
3790 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3791 char *buf, size_t nbytes,
3794 return pressure_write(of, buf, nbytes, PSI_CPU);
3797 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3798 static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3800 struct cgroup *cgrp = seq_css(seq)->cgroup;
3801 struct psi_group *psi = cgroup_psi(cgrp);
3803 return psi_show(seq, psi, PSI_IRQ);
3806 static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3807 char *buf, size_t nbytes,
3810 return pressure_write(of, buf, nbytes, PSI_IRQ);
3814 static int cgroup_pressure_show(struct seq_file *seq, void *v)
3816 struct cgroup *cgrp = seq_css(seq)->cgroup;
3817 struct psi_group *psi = cgroup_psi(cgrp);
3819 seq_printf(seq, "%d\n", psi->enabled);
3824 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3825 char *buf, size_t nbytes,
3830 struct cgroup *cgrp;
3831 struct psi_group *psi;
3833 ret = kstrtoint(strstrip(buf), 0, &enable);
3837 if (enable < 0 || enable > 1)
3840 cgrp = cgroup_kn_lock_live(of->kn, false);
3844 psi = cgroup_psi(cgrp);
3845 if (psi->enabled != enable) {
3848 /* show or hide {cpu,memory,io,irq}.pressure files */
3849 for (i = 0; i < NR_PSI_RESOURCES; i++)
3850 cgroup_file_show(&cgrp->psi_files[i], enable);
3852 psi->enabled = enable;
3854 psi_cgroup_restart(psi);
3857 cgroup_kn_unlock(of->kn);
3862 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3865 struct cgroup_file_ctx *ctx = of->priv;
3867 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3870 static void cgroup_pressure_release(struct kernfs_open_file *of)
3872 struct cgroup_file_ctx *ctx = of->priv;
3874 psi_trigger_destroy(ctx->psi.trigger);
3877 bool cgroup_psi_enabled(void)
3879 if (static_branch_likely(&psi_disabled))
3882 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3885 #else /* CONFIG_PSI */
3886 bool cgroup_psi_enabled(void)
3891 #endif /* CONFIG_PSI */
3893 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3895 struct cgroup *cgrp = seq_css(seq)->cgroup;
3897 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3902 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3903 char *buf, size_t nbytes, loff_t off)
3905 struct cgroup *cgrp;
3909 ret = kstrtoint(strstrip(buf), 0, &freeze);
3913 if (freeze < 0 || freeze > 1)
3916 cgrp = cgroup_kn_lock_live(of->kn, false);
3920 cgroup_freeze(cgrp, freeze);
3922 cgroup_kn_unlock(of->kn);
3927 static void __cgroup_kill(struct cgroup *cgrp)
3929 struct css_task_iter it;
3930 struct task_struct *task;
3932 lockdep_assert_held(&cgroup_mutex);
3934 spin_lock_irq(&css_set_lock);
3935 set_bit(CGRP_KILL, &cgrp->flags);
3936 spin_unlock_irq(&css_set_lock);
3938 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3939 while ((task = css_task_iter_next(&it))) {
3940 /* Ignore kernel threads here. */
3941 if (task->flags & PF_KTHREAD)
3944 /* Skip tasks that are already dying. */
3945 if (__fatal_signal_pending(task))
3948 send_sig(SIGKILL, task, 0);
3950 css_task_iter_end(&it);
3952 spin_lock_irq(&css_set_lock);
3953 clear_bit(CGRP_KILL, &cgrp->flags);
3954 spin_unlock_irq(&css_set_lock);
3957 static void cgroup_kill(struct cgroup *cgrp)
3959 struct cgroup_subsys_state *css;
3960 struct cgroup *dsct;
3962 lockdep_assert_held(&cgroup_mutex);
3964 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3965 __cgroup_kill(dsct);
3968 static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3969 size_t nbytes, loff_t off)
3973 struct cgroup *cgrp;
3975 ret = kstrtoint(strstrip(buf), 0, &kill);
3982 cgrp = cgroup_kn_lock_live(of->kn, false);
3987 * Killing is a process directed operation, i.e. the whole thread-group
3988 * is taken down so act like we do for cgroup.procs and only make this
3989 * writable in non-threaded cgroups.
3991 if (cgroup_is_threaded(cgrp))
3996 cgroup_kn_unlock(of->kn);
3998 return ret ?: nbytes;
4001 static int cgroup_file_open(struct kernfs_open_file *of)
4003 struct cftype *cft = of_cft(of);
4004 struct cgroup_file_ctx *ctx;
4007 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4011 ctx->ns = current->nsproxy->cgroup_ns;
4012 get_cgroup_ns(ctx->ns);
4018 ret = cft->open(of);
4020 put_cgroup_ns(ctx->ns);
4026 static void cgroup_file_release(struct kernfs_open_file *of)
4028 struct cftype *cft = of_cft(of);
4029 struct cgroup_file_ctx *ctx = of->priv;
4033 put_cgroup_ns(ctx->ns);
4037 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4038 size_t nbytes, loff_t off)
4040 struct cgroup_file_ctx *ctx = of->priv;
4041 struct cgroup *cgrp = of->kn->parent->priv;
4042 struct cftype *cft = of_cft(of);
4043 struct cgroup_subsys_state *css;
4050 * If namespaces are delegation boundaries, disallow writes to
4051 * files in an non-init namespace root from inside the namespace
4052 * except for the files explicitly marked delegatable -
4053 * cgroup.procs and cgroup.subtree_control.
4055 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4056 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4057 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4061 return cft->write(of, buf, nbytes, off);
4064 * kernfs guarantees that a file isn't deleted with operations in
4065 * flight, which means that the matching css is and stays alive and
4066 * doesn't need to be pinned. The RCU locking is not necessary
4067 * either. It's just for the convenience of using cgroup_css().
4070 css = cgroup_css(cgrp, cft->ss);
4073 if (cft->write_u64) {
4074 unsigned long long v;
4075 ret = kstrtoull(buf, 0, &v);
4077 ret = cft->write_u64(css, cft, v);
4078 } else if (cft->write_s64) {
4080 ret = kstrtoll(buf, 0, &v);
4082 ret = cft->write_s64(css, cft, v);
4087 return ret ?: nbytes;
4090 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4092 struct cftype *cft = of_cft(of);
4095 return cft->poll(of, pt);
4097 return kernfs_generic_poll(of, pt);
4100 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4102 return seq_cft(seq)->seq_start(seq, ppos);
4105 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4107 return seq_cft(seq)->seq_next(seq, v, ppos);
4110 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4112 if (seq_cft(seq)->seq_stop)
4113 seq_cft(seq)->seq_stop(seq, v);
4116 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4118 struct cftype *cft = seq_cft(m);
4119 struct cgroup_subsys_state *css = seq_css(m);
4122 return cft->seq_show(m, arg);
4125 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4126 else if (cft->read_s64)
4127 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4133 static struct kernfs_ops cgroup_kf_single_ops = {
4134 .atomic_write_len = PAGE_SIZE,
4135 .open = cgroup_file_open,
4136 .release = cgroup_file_release,
4137 .write = cgroup_file_write,
4138 .poll = cgroup_file_poll,
4139 .seq_show = cgroup_seqfile_show,
4142 static struct kernfs_ops cgroup_kf_ops = {
4143 .atomic_write_len = PAGE_SIZE,
4144 .open = cgroup_file_open,
4145 .release = cgroup_file_release,
4146 .write = cgroup_file_write,
4147 .poll = cgroup_file_poll,
4148 .seq_start = cgroup_seqfile_start,
4149 .seq_next = cgroup_seqfile_next,
4150 .seq_stop = cgroup_seqfile_stop,
4151 .seq_show = cgroup_seqfile_show,
4154 /* set uid and gid of cgroup dirs and files to that of the creator */
4155 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
4157 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
4158 .ia_uid = current_fsuid(),
4159 .ia_gid = current_fsgid(), };
4161 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
4162 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
4165 return kernfs_setattr(kn, &iattr);
4168 static void cgroup_file_notify_timer(struct timer_list *timer)
4170 cgroup_file_notify(container_of(timer, struct cgroup_file,
4174 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4177 char name[CGROUP_FILE_NAME_MAX];
4178 struct kernfs_node *kn;
4179 struct lock_class_key *key = NULL;
4182 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4183 key = &cft->lockdep_key;
4185 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4186 cgroup_file_mode(cft),
4187 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
4188 0, cft->kf_ops, cft,
4193 ret = cgroup_kn_set_ugid(kn);
4199 if (cft->file_offset) {
4200 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4202 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4204 spin_lock_irq(&cgroup_file_kn_lock);
4206 spin_unlock_irq(&cgroup_file_kn_lock);
4213 * cgroup_addrm_files - add or remove files to a cgroup directory
4214 * @css: the target css
4215 * @cgrp: the target cgroup (usually css->cgroup)
4216 * @cfts: array of cftypes to be added
4217 * @is_add: whether to add or remove
4219 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4220 * For removals, this function never fails.
4222 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4223 struct cgroup *cgrp, struct cftype cfts[],
4226 struct cftype *cft, *cft_end = NULL;
4229 lockdep_assert_held(&cgroup_mutex);
4232 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4233 /* does cft->flags tell us to skip this file on @cgrp? */
4234 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4236 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4238 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4240 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4242 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4245 ret = cgroup_add_file(css, cgrp, cft);
4247 pr_warn("%s: failed to add %s, err=%d\n",
4248 __func__, cft->name, ret);
4254 cgroup_rm_file(cgrp, cft);
4260 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4262 struct cgroup_subsys *ss = cfts[0].ss;
4263 struct cgroup *root = &ss->root->cgrp;
4264 struct cgroup_subsys_state *css;
4267 lockdep_assert_held(&cgroup_mutex);
4269 /* add/rm files for all cgroups created before */
4270 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4271 struct cgroup *cgrp = css->cgroup;
4273 if (!(css->flags & CSS_VISIBLE))
4276 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4282 kernfs_activate(root->kn);
4286 static void cgroup_exit_cftypes(struct cftype *cfts)
4290 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4291 /* free copy for custom atomic_write_len, see init_cftypes() */
4292 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4297 /* revert flags set by cgroup core while adding @cfts */
4298 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4303 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4308 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4309 struct kernfs_ops *kf_ops;
4311 WARN_ON(cft->ss || cft->kf_ops);
4313 if (cft->flags & __CFTYPE_ADDED) {
4319 kf_ops = &cgroup_kf_ops;
4321 kf_ops = &cgroup_kf_single_ops;
4324 * Ugh... if @cft wants a custom max_write_len, we need to
4325 * make a copy of kf_ops to set its atomic_write_len.
4327 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4328 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4333 kf_ops->atomic_write_len = cft->max_write_len;
4336 cft->kf_ops = kf_ops;
4338 cft->flags |= __CFTYPE_ADDED;
4342 cgroup_exit_cftypes(cfts);
4346 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4348 lockdep_assert_held(&cgroup_mutex);
4350 list_del(&cfts->node);
4351 cgroup_apply_cftypes(cfts, false);
4352 cgroup_exit_cftypes(cfts);
4357 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4358 * @cfts: zero-length name terminated array of cftypes
4360 * Unregister @cfts. Files described by @cfts are removed from all
4361 * existing cgroups and all future cgroups won't have them either. This
4362 * function can be called anytime whether @cfts' subsys is attached or not.
4364 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4367 int cgroup_rm_cftypes(struct cftype *cfts)
4371 if (!cfts || cfts[0].name[0] == '\0')
4374 if (!(cfts[0].flags & __CFTYPE_ADDED))
4377 mutex_lock(&cgroup_mutex);
4378 ret = cgroup_rm_cftypes_locked(cfts);
4379 mutex_unlock(&cgroup_mutex);
4384 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4385 * @ss: target cgroup subsystem
4386 * @cfts: zero-length name terminated array of cftypes
4388 * Register @cfts to @ss. Files described by @cfts are created for all
4389 * existing cgroups to which @ss is attached and all future cgroups will
4390 * have them too. This function can be called anytime whether @ss is
4393 * Returns 0 on successful registration, -errno on failure. Note that this
4394 * function currently returns 0 as long as @cfts registration is successful
4395 * even if some file creation attempts on existing cgroups fail.
4397 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4401 if (!cgroup_ssid_enabled(ss->id))
4404 if (!cfts || cfts[0].name[0] == '\0')
4407 ret = cgroup_init_cftypes(ss, cfts);
4411 mutex_lock(&cgroup_mutex);
4413 list_add_tail(&cfts->node, &ss->cfts);
4414 ret = cgroup_apply_cftypes(cfts, true);
4416 cgroup_rm_cftypes_locked(cfts);
4418 mutex_unlock(&cgroup_mutex);
4423 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4424 * @ss: target cgroup subsystem
4425 * @cfts: zero-length name terminated array of cftypes
4427 * Similar to cgroup_add_cftypes() but the added files are only used for
4428 * the default hierarchy.
4430 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4434 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4435 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4436 return cgroup_add_cftypes(ss, cfts);
4440 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4441 * @ss: target cgroup subsystem
4442 * @cfts: zero-length name terminated array of cftypes
4444 * Similar to cgroup_add_cftypes() but the added files are only used for
4445 * the legacy hierarchies.
4447 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4451 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4452 cft->flags |= __CFTYPE_NOT_ON_DFL;
4453 return cgroup_add_cftypes(ss, cfts);
4457 * cgroup_file_notify - generate a file modified event for a cgroup_file
4458 * @cfile: target cgroup_file
4460 * @cfile must have been obtained by setting cftype->file_offset.
4462 void cgroup_file_notify(struct cgroup_file *cfile)
4464 unsigned long flags;
4466 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4468 unsigned long last = cfile->notified_at;
4469 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4471 if (time_in_range(jiffies, last, next)) {
4472 timer_reduce(&cfile->notify_timer, next);
4474 kernfs_notify(cfile->kn);
4475 cfile->notified_at = jiffies;
4478 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4482 * cgroup_file_show - show or hide a hidden cgroup file
4483 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4484 * @show: whether to show or hide
4486 void cgroup_file_show(struct cgroup_file *cfile, bool show)
4488 struct kernfs_node *kn;
4490 spin_lock_irq(&cgroup_file_kn_lock);
4493 spin_unlock_irq(&cgroup_file_kn_lock);
4496 kernfs_show(kn, show);
4502 * css_next_child - find the next child of a given css
4503 * @pos: the current position (%NULL to initiate traversal)
4504 * @parent: css whose children to walk
4506 * This function returns the next child of @parent and should be called
4507 * under either cgroup_mutex or RCU read lock. The only requirement is
4508 * that @parent and @pos are accessible. The next sibling is guaranteed to
4509 * be returned regardless of their states.
4511 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4512 * css which finished ->css_online() is guaranteed to be visible in the
4513 * future iterations and will stay visible until the last reference is put.
4514 * A css which hasn't finished ->css_online() or already finished
4515 * ->css_offline() may show up during traversal. It's each subsystem's
4516 * responsibility to synchronize against on/offlining.
4518 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4519 struct cgroup_subsys_state *parent)
4521 struct cgroup_subsys_state *next;
4523 cgroup_assert_mutex_or_rcu_locked();
4526 * @pos could already have been unlinked from the sibling list.
4527 * Once a cgroup is removed, its ->sibling.next is no longer
4528 * updated when its next sibling changes. CSS_RELEASED is set when
4529 * @pos is taken off list, at which time its next pointer is valid,
4530 * and, as releases are serialized, the one pointed to by the next
4531 * pointer is guaranteed to not have started release yet. This
4532 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4533 * critical section, the one pointed to by its next pointer is
4534 * guaranteed to not have finished its RCU grace period even if we
4535 * have dropped rcu_read_lock() in-between iterations.
4537 * If @pos has CSS_RELEASED set, its next pointer can't be
4538 * dereferenced; however, as each css is given a monotonically
4539 * increasing unique serial number and always appended to the
4540 * sibling list, the next one can be found by walking the parent's
4541 * children until the first css with higher serial number than
4542 * @pos's. While this path can be slower, it happens iff iteration
4543 * races against release and the race window is very small.
4546 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4547 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4548 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4550 list_for_each_entry_rcu(next, &parent->children, sibling,
4551 lockdep_is_held(&cgroup_mutex))
4552 if (next->serial_nr > pos->serial_nr)
4557 * @next, if not pointing to the head, can be dereferenced and is
4560 if (&next->sibling != &parent->children)
4566 * css_next_descendant_pre - find the next descendant for pre-order walk
4567 * @pos: the current position (%NULL to initiate traversal)
4568 * @root: css whose descendants to walk
4570 * To be used by css_for_each_descendant_pre(). Find the next descendant
4571 * to visit for pre-order traversal of @root's descendants. @root is
4572 * included in the iteration and the first node to be visited.
4574 * While this function requires cgroup_mutex or RCU read locking, it
4575 * doesn't require the whole traversal to be contained in a single critical
4576 * section. This function will return the correct next descendant as long
4577 * as both @pos and @root are accessible and @pos is a descendant of @root.
4579 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4580 * css which finished ->css_online() is guaranteed to be visible in the
4581 * future iterations and will stay visible until the last reference is put.
4582 * A css which hasn't finished ->css_online() or already finished
4583 * ->css_offline() may show up during traversal. It's each subsystem's
4584 * responsibility to synchronize against on/offlining.
4586 struct cgroup_subsys_state *
4587 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4588 struct cgroup_subsys_state *root)
4590 struct cgroup_subsys_state *next;
4592 cgroup_assert_mutex_or_rcu_locked();
4594 /* if first iteration, visit @root */
4598 /* visit the first child if exists */
4599 next = css_next_child(NULL, pos);
4603 /* no child, visit my or the closest ancestor's next sibling */
4604 while (pos != root) {
4605 next = css_next_child(pos, pos->parent);
4613 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4616 * css_rightmost_descendant - return the rightmost descendant of a css
4617 * @pos: css of interest
4619 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4620 * is returned. This can be used during pre-order traversal to skip
4623 * While this function requires cgroup_mutex or RCU read locking, it
4624 * doesn't require the whole traversal to be contained in a single critical
4625 * section. This function will return the correct rightmost descendant as
4626 * long as @pos is accessible.
4628 struct cgroup_subsys_state *
4629 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4631 struct cgroup_subsys_state *last, *tmp;
4633 cgroup_assert_mutex_or_rcu_locked();
4637 /* ->prev isn't RCU safe, walk ->next till the end */
4639 css_for_each_child(tmp, last)
4646 static struct cgroup_subsys_state *
4647 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4649 struct cgroup_subsys_state *last;
4653 pos = css_next_child(NULL, pos);
4660 * css_next_descendant_post - find the next descendant for post-order walk
4661 * @pos: the current position (%NULL to initiate traversal)
4662 * @root: css whose descendants to walk
4664 * To be used by css_for_each_descendant_post(). Find the next descendant
4665 * to visit for post-order traversal of @root's descendants. @root is
4666 * included in the iteration and the last node to be visited.
4668 * While this function requires cgroup_mutex or RCU read locking, it
4669 * doesn't require the whole traversal to be contained in a single critical
4670 * section. This function will return the correct next descendant as long
4671 * as both @pos and @cgroup are accessible and @pos is a descendant of
4674 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4675 * css which finished ->css_online() is guaranteed to be visible in the
4676 * future iterations and will stay visible until the last reference is put.
4677 * A css which hasn't finished ->css_online() or already finished
4678 * ->css_offline() may show up during traversal. It's each subsystem's
4679 * responsibility to synchronize against on/offlining.
4681 struct cgroup_subsys_state *
4682 css_next_descendant_post(struct cgroup_subsys_state *pos,
4683 struct cgroup_subsys_state *root)
4685 struct cgroup_subsys_state *next;
4687 cgroup_assert_mutex_or_rcu_locked();
4689 /* if first iteration, visit leftmost descendant which may be @root */
4691 return css_leftmost_descendant(root);
4693 /* if we visited @root, we're done */
4697 /* if there's an unvisited sibling, visit its leftmost descendant */
4698 next = css_next_child(pos, pos->parent);
4700 return css_leftmost_descendant(next);
4702 /* no sibling left, visit parent */
4707 * css_has_online_children - does a css have online children
4708 * @css: the target css
4710 * Returns %true if @css has any online children; otherwise, %false. This
4711 * function can be called from any context but the caller is responsible
4712 * for synchronizing against on/offlining as necessary.
4714 bool css_has_online_children(struct cgroup_subsys_state *css)
4716 struct cgroup_subsys_state *child;
4720 css_for_each_child(child, css) {
4721 if (child->flags & CSS_ONLINE) {
4730 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4732 struct list_head *l;
4733 struct cgrp_cset_link *link;
4734 struct css_set *cset;
4736 lockdep_assert_held(&css_set_lock);
4738 /* find the next threaded cset */
4739 if (it->tcset_pos) {
4740 l = it->tcset_pos->next;
4742 if (l != it->tcset_head) {
4744 return container_of(l, struct css_set,
4745 threaded_csets_node);
4748 it->tcset_pos = NULL;
4751 /* find the next cset */
4754 if (l == it->cset_head) {
4755 it->cset_pos = NULL;
4760 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4762 link = list_entry(l, struct cgrp_cset_link, cset_link);
4768 /* initialize threaded css_set walking */
4769 if (it->flags & CSS_TASK_ITER_THREADED) {
4771 put_css_set_locked(it->cur_dcset);
4772 it->cur_dcset = cset;
4775 it->tcset_head = &cset->threaded_csets;
4776 it->tcset_pos = &cset->threaded_csets;
4783 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4784 * @it: the iterator to advance
4786 * Advance @it to the next css_set to walk.
4788 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4790 struct css_set *cset;
4792 lockdep_assert_held(&css_set_lock);
4794 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4795 while ((cset = css_task_iter_next_css_set(it))) {
4796 if (!list_empty(&cset->tasks)) {
4797 it->cur_tasks_head = &cset->tasks;
4799 } else if (!list_empty(&cset->mg_tasks)) {
4800 it->cur_tasks_head = &cset->mg_tasks;
4802 } else if (!list_empty(&cset->dying_tasks)) {
4803 it->cur_tasks_head = &cset->dying_tasks;
4808 it->task_pos = NULL;
4811 it->task_pos = it->cur_tasks_head->next;
4814 * We don't keep css_sets locked across iteration steps and thus
4815 * need to take steps to ensure that iteration can be resumed after
4816 * the lock is re-acquired. Iteration is performed at two levels -
4817 * css_sets and tasks in them.
4819 * Once created, a css_set never leaves its cgroup lists, so a
4820 * pinned css_set is guaranteed to stay put and we can resume
4821 * iteration afterwards.
4823 * Tasks may leave @cset across iteration steps. This is resolved
4824 * by registering each iterator with the css_set currently being
4825 * walked and making css_set_move_task() advance iterators whose
4826 * next task is leaving.
4829 list_del(&it->iters_node);
4830 put_css_set_locked(it->cur_cset);
4833 it->cur_cset = cset;
4834 list_add(&it->iters_node, &cset->task_iters);
4837 static void css_task_iter_skip(struct css_task_iter *it,
4838 struct task_struct *task)
4840 lockdep_assert_held(&css_set_lock);
4842 if (it->task_pos == &task->cg_list) {
4843 it->task_pos = it->task_pos->next;
4844 it->flags |= CSS_TASK_ITER_SKIPPED;
4848 static void css_task_iter_advance(struct css_task_iter *it)
4850 struct task_struct *task;
4852 lockdep_assert_held(&css_set_lock);
4856 * Advance iterator to find next entry. We go through cset
4857 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4860 if (it->flags & CSS_TASK_ITER_SKIPPED)
4861 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4863 it->task_pos = it->task_pos->next;
4865 if (it->task_pos == &it->cur_cset->tasks) {
4866 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4867 it->task_pos = it->cur_tasks_head->next;
4869 if (it->task_pos == &it->cur_cset->mg_tasks) {
4870 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4871 it->task_pos = it->cur_tasks_head->next;
4873 if (it->task_pos == &it->cur_cset->dying_tasks)
4874 css_task_iter_advance_css_set(it);
4876 /* called from start, proceed to the first cset */
4877 css_task_iter_advance_css_set(it);
4883 task = list_entry(it->task_pos, struct task_struct, cg_list);
4885 if (it->flags & CSS_TASK_ITER_PROCS) {
4886 /* if PROCS, skip over tasks which aren't group leaders */
4887 if (!thread_group_leader(task))
4890 /* and dying leaders w/o live member threads */
4891 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4892 !atomic_read(&task->signal->live))
4895 /* skip all dying ones */
4896 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4902 * css_task_iter_start - initiate task iteration
4903 * @css: the css to walk tasks of
4904 * @flags: CSS_TASK_ITER_* flags
4905 * @it: the task iterator to use
4907 * Initiate iteration through the tasks of @css. The caller can call
4908 * css_task_iter_next() to walk through the tasks until the function
4909 * returns NULL. On completion of iteration, css_task_iter_end() must be
4912 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4913 struct css_task_iter *it)
4915 memset(it, 0, sizeof(*it));
4917 spin_lock_irq(&css_set_lock);
4922 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4923 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4925 it->cset_pos = &css->cgroup->cset_links;
4927 it->cset_head = it->cset_pos;
4929 css_task_iter_advance(it);
4931 spin_unlock_irq(&css_set_lock);
4935 * css_task_iter_next - return the next task for the iterator
4936 * @it: the task iterator being iterated
4938 * The "next" function for task iteration. @it should have been
4939 * initialized via css_task_iter_start(). Returns NULL when the iteration
4942 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4945 put_task_struct(it->cur_task);
4946 it->cur_task = NULL;
4949 spin_lock_irq(&css_set_lock);
4951 /* @it may be half-advanced by skips, finish advancing */
4952 if (it->flags & CSS_TASK_ITER_SKIPPED)
4953 css_task_iter_advance(it);
4956 it->cur_task = list_entry(it->task_pos, struct task_struct,
4958 get_task_struct(it->cur_task);
4959 css_task_iter_advance(it);
4962 spin_unlock_irq(&css_set_lock);
4964 return it->cur_task;
4968 * css_task_iter_end - finish task iteration
4969 * @it: the task iterator to finish
4971 * Finish task iteration started by css_task_iter_start().
4973 void css_task_iter_end(struct css_task_iter *it)
4976 spin_lock_irq(&css_set_lock);
4977 list_del(&it->iters_node);
4978 put_css_set_locked(it->cur_cset);
4979 spin_unlock_irq(&css_set_lock);
4983 put_css_set(it->cur_dcset);
4986 put_task_struct(it->cur_task);
4989 static void cgroup_procs_release(struct kernfs_open_file *of)
4991 struct cgroup_file_ctx *ctx = of->priv;
4993 if (ctx->procs.started)
4994 css_task_iter_end(&ctx->procs.iter);
4997 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4999 struct kernfs_open_file *of = s->private;
5000 struct cgroup_file_ctx *ctx = of->priv;
5005 return css_task_iter_next(&ctx->procs.iter);
5008 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5009 unsigned int iter_flags)
5011 struct kernfs_open_file *of = s->private;
5012 struct cgroup *cgrp = seq_css(s)->cgroup;
5013 struct cgroup_file_ctx *ctx = of->priv;
5014 struct css_task_iter *it = &ctx->procs.iter;
5017 * When a seq_file is seeked, it's always traversed sequentially
5018 * from position 0, so we can simply keep iterating on !0 *pos.
5020 if (!ctx->procs.started) {
5021 if (WARN_ON_ONCE((*pos)))
5022 return ERR_PTR(-EINVAL);
5023 css_task_iter_start(&cgrp->self, iter_flags, it);
5024 ctx->procs.started = true;
5025 } else if (!(*pos)) {
5026 css_task_iter_end(it);
5027 css_task_iter_start(&cgrp->self, iter_flags, it);
5029 return it->cur_task;
5031 return cgroup_procs_next(s, NULL, NULL);
5034 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5036 struct cgroup *cgrp = seq_css(s)->cgroup;
5039 * All processes of a threaded subtree belong to the domain cgroup
5040 * of the subtree. Only threads can be distributed across the
5041 * subtree. Reject reads on cgroup.procs in the subtree proper.
5042 * They're always empty anyway.
5044 if (cgroup_is_threaded(cgrp))
5045 return ERR_PTR(-EOPNOTSUPP);
5047 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5048 CSS_TASK_ITER_THREADED);
5051 static int cgroup_procs_show(struct seq_file *s, void *v)
5053 seq_printf(s, "%d\n", task_pid_vnr(v));
5057 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5060 struct inode *inode;
5062 lockdep_assert_held(&cgroup_mutex);
5064 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5068 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5073 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5074 struct cgroup *dst_cgrp,
5075 struct super_block *sb,
5076 struct cgroup_namespace *ns)
5078 struct cgroup *com_cgrp = src_cgrp;
5081 lockdep_assert_held(&cgroup_mutex);
5083 /* find the common ancestor */
5084 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5085 com_cgrp = cgroup_parent(com_cgrp);
5087 /* %current should be authorized to migrate to the common ancestor */
5088 ret = cgroup_may_write(com_cgrp, sb);
5093 * If namespaces are delegation boundaries, %current must be able
5094 * to see both source and destination cgroups from its namespace.
5096 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5097 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5098 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5104 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5105 struct cgroup *dst_cgrp,
5106 struct super_block *sb, bool threadgroup,
5107 struct cgroup_namespace *ns)
5111 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5115 ret = cgroup_migrate_vet_dst(dst_cgrp);
5119 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5125 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5128 struct cgroup_file_ctx *ctx = of->priv;
5129 struct cgroup *src_cgrp, *dst_cgrp;
5130 struct task_struct *task;
5131 const struct cred *saved_cred;
5133 bool threadgroup_locked;
5135 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5139 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5140 ret = PTR_ERR_OR_ZERO(task);
5144 /* find the source cgroup */
5145 spin_lock_irq(&css_set_lock);
5146 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5147 spin_unlock_irq(&css_set_lock);
5150 * Process and thread migrations follow same delegation rule. Check
5151 * permissions using the credentials from file open to protect against
5152 * inherited fd attacks.
5154 saved_cred = override_creds(of->file->f_cred);
5155 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5156 of->file->f_path.dentry->d_sb,
5157 threadgroup, ctx->ns);
5158 revert_creds(saved_cred);
5162 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5165 cgroup_procs_write_finish(task, threadgroup_locked);
5167 cgroup_kn_unlock(of->kn);
5172 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5173 char *buf, size_t nbytes, loff_t off)
5175 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5178 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5180 return __cgroup_procs_start(s, pos, 0);
5183 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5184 char *buf, size_t nbytes, loff_t off)
5186 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5189 /* cgroup core interface files for the default hierarchy */
5190 static struct cftype cgroup_base_files[] = {
5192 .name = "cgroup.type",
5193 .flags = CFTYPE_NOT_ON_ROOT,
5194 .seq_show = cgroup_type_show,
5195 .write = cgroup_type_write,
5198 .name = "cgroup.procs",
5199 .flags = CFTYPE_NS_DELEGATABLE,
5200 .file_offset = offsetof(struct cgroup, procs_file),
5201 .release = cgroup_procs_release,
5202 .seq_start = cgroup_procs_start,
5203 .seq_next = cgroup_procs_next,
5204 .seq_show = cgroup_procs_show,
5205 .write = cgroup_procs_write,
5208 .name = "cgroup.threads",
5209 .flags = CFTYPE_NS_DELEGATABLE,
5210 .release = cgroup_procs_release,
5211 .seq_start = cgroup_threads_start,
5212 .seq_next = cgroup_procs_next,
5213 .seq_show = cgroup_procs_show,
5214 .write = cgroup_threads_write,
5217 .name = "cgroup.controllers",
5218 .seq_show = cgroup_controllers_show,
5221 .name = "cgroup.subtree_control",
5222 .flags = CFTYPE_NS_DELEGATABLE,
5223 .seq_show = cgroup_subtree_control_show,
5224 .write = cgroup_subtree_control_write,
5227 .name = "cgroup.events",
5228 .flags = CFTYPE_NOT_ON_ROOT,
5229 .file_offset = offsetof(struct cgroup, events_file),
5230 .seq_show = cgroup_events_show,
5233 .name = "cgroup.max.descendants",
5234 .seq_show = cgroup_max_descendants_show,
5235 .write = cgroup_max_descendants_write,
5238 .name = "cgroup.max.depth",
5239 .seq_show = cgroup_max_depth_show,
5240 .write = cgroup_max_depth_write,
5243 .name = "cgroup.stat",
5244 .seq_show = cgroup_stat_show,
5247 .name = "cgroup.freeze",
5248 .flags = CFTYPE_NOT_ON_ROOT,
5249 .seq_show = cgroup_freeze_show,
5250 .write = cgroup_freeze_write,
5253 .name = "cgroup.kill",
5254 .flags = CFTYPE_NOT_ON_ROOT,
5255 .write = cgroup_kill_write,
5259 .seq_show = cpu_stat_show,
5264 static struct cftype cgroup_psi_files[] = {
5267 .name = "io.pressure",
5268 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5269 .seq_show = cgroup_io_pressure_show,
5270 .write = cgroup_io_pressure_write,
5271 .poll = cgroup_pressure_poll,
5272 .release = cgroup_pressure_release,
5275 .name = "memory.pressure",
5276 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5277 .seq_show = cgroup_memory_pressure_show,
5278 .write = cgroup_memory_pressure_write,
5279 .poll = cgroup_pressure_poll,
5280 .release = cgroup_pressure_release,
5283 .name = "cpu.pressure",
5284 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5285 .seq_show = cgroup_cpu_pressure_show,
5286 .write = cgroup_cpu_pressure_write,
5287 .poll = cgroup_pressure_poll,
5288 .release = cgroup_pressure_release,
5290 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
5292 .name = "irq.pressure",
5293 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5294 .seq_show = cgroup_irq_pressure_show,
5295 .write = cgroup_irq_pressure_write,
5296 .poll = cgroup_pressure_poll,
5297 .release = cgroup_pressure_release,
5301 .name = "cgroup.pressure",
5302 .seq_show = cgroup_pressure_show,
5303 .write = cgroup_pressure_write,
5305 #endif /* CONFIG_PSI */
5310 * css destruction is four-stage process.
5312 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5313 * Implemented in kill_css().
5315 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5316 * and thus css_tryget_online() is guaranteed to fail, the css can be
5317 * offlined by invoking offline_css(). After offlining, the base ref is
5318 * put. Implemented in css_killed_work_fn().
5320 * 3. When the percpu_ref reaches zero, the only possible remaining
5321 * accessors are inside RCU read sections. css_release() schedules the
5324 * 4. After the grace period, the css can be freed. Implemented in
5325 * css_free_work_fn().
5327 * It is actually hairier because both step 2 and 4 require process context
5328 * and thus involve punting to css->destroy_work adding two additional
5329 * steps to the already complex sequence.
5331 static void css_free_rwork_fn(struct work_struct *work)
5333 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5334 struct cgroup_subsys_state, destroy_rwork);
5335 struct cgroup_subsys *ss = css->ss;
5336 struct cgroup *cgrp = css->cgroup;
5338 percpu_ref_exit(&css->refcnt);
5342 struct cgroup_subsys_state *parent = css->parent;
5346 cgroup_idr_remove(&ss->css_idr, id);
5352 /* cgroup free path */
5353 atomic_dec(&cgrp->root->nr_cgrps);
5354 cgroup1_pidlist_destroy_all(cgrp);
5355 cancel_work_sync(&cgrp->release_agent_work);
5356 bpf_cgrp_storage_free(cgrp);
5358 if (cgroup_parent(cgrp)) {
5360 * We get a ref to the parent, and put the ref when
5361 * this cgroup is being freed, so it's guaranteed
5362 * that the parent won't be destroyed before its
5365 cgroup_put(cgroup_parent(cgrp));
5366 kernfs_put(cgrp->kn);
5367 psi_cgroup_free(cgrp);
5368 cgroup_rstat_exit(cgrp);
5372 * This is root cgroup's refcnt reaching zero,
5373 * which indicates that the root should be
5376 cgroup_destroy_root(cgrp->root);
5381 static void css_release_work_fn(struct work_struct *work)
5383 struct cgroup_subsys_state *css =
5384 container_of(work, struct cgroup_subsys_state, destroy_work);
5385 struct cgroup_subsys *ss = css->ss;
5386 struct cgroup *cgrp = css->cgroup;
5388 mutex_lock(&cgroup_mutex);
5390 css->flags |= CSS_RELEASED;
5391 list_del_rcu(&css->sibling);
5394 /* css release path */
5395 if (!list_empty(&css->rstat_css_node)) {
5396 cgroup_rstat_flush(cgrp);
5397 list_del_rcu(&css->rstat_css_node);
5400 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5401 if (ss->css_released)
5402 ss->css_released(css);
5404 struct cgroup *tcgrp;
5406 /* cgroup release path */
5407 TRACE_CGROUP_PATH(release, cgrp);
5409 cgroup_rstat_flush(cgrp);
5411 spin_lock_irq(&css_set_lock);
5412 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5413 tcgrp = cgroup_parent(tcgrp))
5414 tcgrp->nr_dying_descendants--;
5415 spin_unlock_irq(&css_set_lock);
5418 * There are two control paths which try to determine
5419 * cgroup from dentry without going through kernfs -
5420 * cgroupstats_build() and css_tryget_online_from_dir().
5421 * Those are supported by RCU protecting clearing of
5422 * cgrp->kn->priv backpointer.
5425 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5429 mutex_unlock(&cgroup_mutex);
5431 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5432 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5435 static void css_release(struct percpu_ref *ref)
5437 struct cgroup_subsys_state *css =
5438 container_of(ref, struct cgroup_subsys_state, refcnt);
5440 INIT_WORK(&css->destroy_work, css_release_work_fn);
5441 queue_work(cgroup_destroy_wq, &css->destroy_work);
5444 static void init_and_link_css(struct cgroup_subsys_state *css,
5445 struct cgroup_subsys *ss, struct cgroup *cgrp)
5447 lockdep_assert_held(&cgroup_mutex);
5449 cgroup_get_live(cgrp);
5451 memset(css, 0, sizeof(*css));
5455 INIT_LIST_HEAD(&css->sibling);
5456 INIT_LIST_HEAD(&css->children);
5457 INIT_LIST_HEAD(&css->rstat_css_node);
5458 css->serial_nr = css_serial_nr_next++;
5459 atomic_set(&css->online_cnt, 0);
5461 if (cgroup_parent(cgrp)) {
5462 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5463 css_get(css->parent);
5466 if (ss->css_rstat_flush)
5467 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5469 BUG_ON(cgroup_css(cgrp, ss));
5472 /* invoke ->css_online() on a new CSS and mark it online if successful */
5473 static int online_css(struct cgroup_subsys_state *css)
5475 struct cgroup_subsys *ss = css->ss;
5478 lockdep_assert_held(&cgroup_mutex);
5481 ret = ss->css_online(css);
5483 css->flags |= CSS_ONLINE;
5484 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5486 atomic_inc(&css->online_cnt);
5488 atomic_inc(&css->parent->online_cnt);
5493 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5494 static void offline_css(struct cgroup_subsys_state *css)
5496 struct cgroup_subsys *ss = css->ss;
5498 lockdep_assert_held(&cgroup_mutex);
5500 if (!(css->flags & CSS_ONLINE))
5503 if (ss->css_offline)
5504 ss->css_offline(css);
5506 css->flags &= ~CSS_ONLINE;
5507 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5509 wake_up_all(&css->cgroup->offline_waitq);
5513 * css_create - create a cgroup_subsys_state
5514 * @cgrp: the cgroup new css will be associated with
5515 * @ss: the subsys of new css
5517 * Create a new css associated with @cgrp - @ss pair. On success, the new
5518 * css is online and installed in @cgrp. This function doesn't create the
5519 * interface files. Returns 0 on success, -errno on failure.
5521 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5522 struct cgroup_subsys *ss)
5524 struct cgroup *parent = cgroup_parent(cgrp);
5525 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5526 struct cgroup_subsys_state *css;
5529 lockdep_assert_held(&cgroup_mutex);
5531 css = ss->css_alloc(parent_css);
5533 css = ERR_PTR(-ENOMEM);
5537 init_and_link_css(css, ss, cgrp);
5539 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5543 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5548 /* @css is ready to be brought online now, make it visible */
5549 list_add_tail_rcu(&css->sibling, &parent_css->children);
5550 cgroup_idr_replace(&ss->css_idr, css, css->id);
5552 err = online_css(css);
5559 list_del_rcu(&css->sibling);
5561 list_del_rcu(&css->rstat_css_node);
5562 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5563 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5564 return ERR_PTR(err);
5568 * The returned cgroup is fully initialized including its control mask, but
5569 * it isn't associated with its kernfs_node and doesn't have the control
5572 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5575 struct cgroup_root *root = parent->root;
5576 struct cgroup *cgrp, *tcgrp;
5577 struct kernfs_node *kn;
5578 int level = parent->level + 1;
5581 /* allocate the cgroup and its ID, 0 is reserved for the root */
5582 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5584 return ERR_PTR(-ENOMEM);
5586 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5590 ret = cgroup_rstat_init(cgrp);
5592 goto out_cancel_ref;
5594 /* create the directory */
5595 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5602 init_cgroup_housekeeping(cgrp);
5604 cgrp->self.parent = &parent->self;
5606 cgrp->level = level;
5608 ret = psi_cgroup_alloc(cgrp);
5610 goto out_kernfs_remove;
5612 ret = cgroup_bpf_inherit(cgrp);
5617 * New cgroup inherits effective freeze counter, and
5618 * if the parent has to be frozen, the child has too.
5620 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5621 if (cgrp->freezer.e_freeze) {
5623 * Set the CGRP_FREEZE flag, so when a process will be
5624 * attached to the child cgroup, it will become frozen.
5625 * At this point the new cgroup is unpopulated, so we can
5626 * consider it frozen immediately.
5628 set_bit(CGRP_FREEZE, &cgrp->flags);
5629 set_bit(CGRP_FROZEN, &cgrp->flags);
5632 spin_lock_irq(&css_set_lock);
5633 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5634 cgrp->ancestors[tcgrp->level] = tcgrp;
5636 if (tcgrp != cgrp) {
5637 tcgrp->nr_descendants++;
5640 * If the new cgroup is frozen, all ancestor cgroups
5641 * get a new frozen descendant, but their state can't
5642 * change because of this.
5644 if (cgrp->freezer.e_freeze)
5645 tcgrp->freezer.nr_frozen_descendants++;
5648 spin_unlock_irq(&css_set_lock);
5650 if (notify_on_release(parent))
5651 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5653 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5654 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5656 cgrp->self.serial_nr = css_serial_nr_next++;
5658 /* allocation complete, commit to creation */
5659 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5660 atomic_inc(&root->nr_cgrps);
5661 cgroup_get_live(parent);
5664 * On the default hierarchy, a child doesn't automatically inherit
5665 * subtree_control from the parent. Each is configured manually.
5667 if (!cgroup_on_dfl(cgrp))
5668 cgrp->subtree_control = cgroup_control(cgrp);
5670 cgroup_propagate_control(cgrp);
5675 psi_cgroup_free(cgrp);
5677 kernfs_remove(cgrp->kn);
5679 cgroup_rstat_exit(cgrp);
5681 percpu_ref_exit(&cgrp->self.refcnt);
5684 return ERR_PTR(ret);
5687 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5689 struct cgroup *cgroup;
5693 lockdep_assert_held(&cgroup_mutex);
5695 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5696 if (cgroup->nr_descendants >= cgroup->max_descendants)
5699 if (level > cgroup->max_depth)
5710 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5712 struct cgroup *parent, *cgrp;
5715 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5716 if (strchr(name, '\n'))
5719 parent = cgroup_kn_lock_live(parent_kn, false);
5723 if (!cgroup_check_hierarchy_limits(parent)) {
5728 cgrp = cgroup_create(parent, name, mode);
5730 ret = PTR_ERR(cgrp);
5735 * This extra ref will be put in cgroup_free_fn() and guarantees
5736 * that @cgrp->kn is always accessible.
5738 kernfs_get(cgrp->kn);
5740 ret = cgroup_kn_set_ugid(cgrp->kn);
5744 ret = css_populate_dir(&cgrp->self);
5748 ret = cgroup_apply_control_enable(cgrp);
5752 TRACE_CGROUP_PATH(mkdir, cgrp);
5754 /* let's create and online css's */
5755 kernfs_activate(cgrp->kn);
5761 cgroup_destroy_locked(cgrp);
5763 cgroup_kn_unlock(parent_kn);
5768 * This is called when the refcnt of a css is confirmed to be killed.
5769 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5770 * initiate destruction and put the css ref from kill_css().
5772 static void css_killed_work_fn(struct work_struct *work)
5774 struct cgroup_subsys_state *css =
5775 container_of(work, struct cgroup_subsys_state, destroy_work);
5777 mutex_lock(&cgroup_mutex);
5782 /* @css can't go away while we're holding cgroup_mutex */
5784 } while (css && atomic_dec_and_test(&css->online_cnt));
5786 mutex_unlock(&cgroup_mutex);
5789 /* css kill confirmation processing requires process context, bounce */
5790 static void css_killed_ref_fn(struct percpu_ref *ref)
5792 struct cgroup_subsys_state *css =
5793 container_of(ref, struct cgroup_subsys_state, refcnt);
5795 if (atomic_dec_and_test(&css->online_cnt)) {
5796 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5797 queue_work(cgroup_destroy_wq, &css->destroy_work);
5802 * kill_css - destroy a css
5803 * @css: css to destroy
5805 * This function initiates destruction of @css by removing cgroup interface
5806 * files and putting its base reference. ->css_offline() will be invoked
5807 * asynchronously once css_tryget_online() is guaranteed to fail and when
5808 * the reference count reaches zero, @css will be released.
5810 static void kill_css(struct cgroup_subsys_state *css)
5812 lockdep_assert_held(&cgroup_mutex);
5814 if (css->flags & CSS_DYING)
5817 css->flags |= CSS_DYING;
5820 * This must happen before css is disassociated with its cgroup.
5821 * See seq_css() for details.
5826 * Killing would put the base ref, but we need to keep it alive
5827 * until after ->css_offline().
5832 * cgroup core guarantees that, by the time ->css_offline() is
5833 * invoked, no new css reference will be given out via
5834 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5835 * proceed to offlining css's because percpu_ref_kill() doesn't
5836 * guarantee that the ref is seen as killed on all CPUs on return.
5838 * Use percpu_ref_kill_and_confirm() to get notifications as each
5839 * css is confirmed to be seen as killed on all CPUs.
5841 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5845 * cgroup_destroy_locked - the first stage of cgroup destruction
5846 * @cgrp: cgroup to be destroyed
5848 * css's make use of percpu refcnts whose killing latency shouldn't be
5849 * exposed to userland and are RCU protected. Also, cgroup core needs to
5850 * guarantee that css_tryget_online() won't succeed by the time
5851 * ->css_offline() is invoked. To satisfy all the requirements,
5852 * destruction is implemented in the following two steps.
5854 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5855 * userland visible parts and start killing the percpu refcnts of
5856 * css's. Set up so that the next stage will be kicked off once all
5857 * the percpu refcnts are confirmed to be killed.
5859 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5860 * rest of destruction. Once all cgroup references are gone, the
5861 * cgroup is RCU-freed.
5863 * This function implements s1. After this step, @cgrp is gone as far as
5864 * the userland is concerned and a new cgroup with the same name may be
5865 * created. As cgroup doesn't care about the names internally, this
5866 * doesn't cause any problem.
5868 static int cgroup_destroy_locked(struct cgroup *cgrp)
5869 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5871 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5872 struct cgroup_subsys_state *css;
5873 struct cgrp_cset_link *link;
5876 lockdep_assert_held(&cgroup_mutex);
5879 * Only migration can raise populated from zero and we're already
5880 * holding cgroup_mutex.
5882 if (cgroup_is_populated(cgrp))
5886 * Make sure there's no live children. We can't test emptiness of
5887 * ->self.children as dead children linger on it while being
5888 * drained; otherwise, "rmdir parent/child parent" may fail.
5890 if (css_has_online_children(&cgrp->self))
5894 * Mark @cgrp and the associated csets dead. The former prevents
5895 * further task migration and child creation by disabling
5896 * cgroup_lock_live_group(). The latter makes the csets ignored by
5897 * the migration path.
5899 cgrp->self.flags &= ~CSS_ONLINE;
5901 spin_lock_irq(&css_set_lock);
5902 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5903 link->cset->dead = true;
5904 spin_unlock_irq(&css_set_lock);
5906 /* initiate massacre of all css's */
5907 for_each_css(css, ssid, cgrp)
5910 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5911 css_clear_dir(&cgrp->self);
5912 kernfs_remove(cgrp->kn);
5914 if (cgroup_is_threaded(cgrp))
5915 parent->nr_threaded_children--;
5917 spin_lock_irq(&css_set_lock);
5918 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5919 tcgrp->nr_descendants--;
5920 tcgrp->nr_dying_descendants++;
5922 * If the dying cgroup is frozen, decrease frozen descendants
5923 * counters of ancestor cgroups.
5925 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5926 tcgrp->freezer.nr_frozen_descendants--;
5928 spin_unlock_irq(&css_set_lock);
5930 cgroup1_check_for_release(parent);
5932 cgroup_bpf_offline(cgrp);
5934 /* put the base reference */
5935 percpu_ref_kill(&cgrp->self.refcnt);
5940 int cgroup_rmdir(struct kernfs_node *kn)
5942 struct cgroup *cgrp;
5945 cgrp = cgroup_kn_lock_live(kn, false);
5949 ret = cgroup_destroy_locked(cgrp);
5951 TRACE_CGROUP_PATH(rmdir, cgrp);
5953 cgroup_kn_unlock(kn);
5957 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5958 .show_options = cgroup_show_options,
5959 .mkdir = cgroup_mkdir,
5960 .rmdir = cgroup_rmdir,
5961 .show_path = cgroup_show_path,
5964 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5966 struct cgroup_subsys_state *css;
5968 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5970 mutex_lock(&cgroup_mutex);
5972 idr_init(&ss->css_idr);
5973 INIT_LIST_HEAD(&ss->cfts);
5975 /* Create the root cgroup state for this subsystem */
5976 ss->root = &cgrp_dfl_root;
5977 css = ss->css_alloc(NULL);
5978 /* We don't handle early failures gracefully */
5979 BUG_ON(IS_ERR(css));
5980 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5983 * Root csses are never destroyed and we can't initialize
5984 * percpu_ref during early init. Disable refcnting.
5986 css->flags |= CSS_NO_REF;
5989 /* allocation can't be done safely during early init */
5992 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5993 BUG_ON(css->id < 0);
5996 /* Update the init_css_set to contain a subsys
5997 * pointer to this state - since the subsystem is
5998 * newly registered, all tasks and hence the
5999 * init_css_set is in the subsystem's root cgroup. */
6000 init_css_set.subsys[ss->id] = css;
6002 have_fork_callback |= (bool)ss->fork << ss->id;
6003 have_exit_callback |= (bool)ss->exit << ss->id;
6004 have_release_callback |= (bool)ss->release << ss->id;
6005 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6007 /* At system boot, before all subsystems have been
6008 * registered, no tasks have been forked, so we don't
6009 * need to invoke fork callbacks here. */
6010 BUG_ON(!list_empty(&init_task.tasks));
6012 BUG_ON(online_css(css));
6014 mutex_unlock(&cgroup_mutex);
6018 * cgroup_init_early - cgroup initialization at system boot
6020 * Initialize cgroups at system boot, and initialize any
6021 * subsystems that request early init.
6023 int __init cgroup_init_early(void)
6025 static struct cgroup_fs_context __initdata ctx;
6026 struct cgroup_subsys *ss;
6029 ctx.root = &cgrp_dfl_root;
6030 init_cgroup_root(&ctx);
6031 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6033 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6035 for_each_subsys(ss, i) {
6036 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6037 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6038 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6040 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6041 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6044 ss->name = cgroup_subsys_name[i];
6045 if (!ss->legacy_name)
6046 ss->legacy_name = cgroup_subsys_name[i];
6049 cgroup_init_subsys(ss, true);
6055 * cgroup_init - cgroup initialization
6057 * Register cgroup filesystem and /proc file, and initialize
6058 * any subsystems that didn't request early init.
6060 int __init cgroup_init(void)
6062 struct cgroup_subsys *ss;
6065 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6066 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6067 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6068 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6070 cgroup_rstat_boot();
6072 get_user_ns(init_cgroup_ns.user_ns);
6074 mutex_lock(&cgroup_mutex);
6077 * Add init_css_set to the hash table so that dfl_root can link to
6080 hash_add(css_set_table, &init_css_set.hlist,
6081 css_set_hash(init_css_set.subsys));
6083 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6085 mutex_unlock(&cgroup_mutex);
6087 for_each_subsys(ss, ssid) {
6088 if (ss->early_init) {
6089 struct cgroup_subsys_state *css =
6090 init_css_set.subsys[ss->id];
6092 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6094 BUG_ON(css->id < 0);
6096 cgroup_init_subsys(ss, false);
6099 list_add_tail(&init_css_set.e_cset_node[ssid],
6100 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6103 * Setting dfl_root subsys_mask needs to consider the
6104 * disabled flag and cftype registration needs kmalloc,
6105 * both of which aren't available during early_init.
6107 if (!cgroup_ssid_enabled(ssid))
6110 if (cgroup1_ssid_disabled(ssid))
6111 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
6114 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6116 /* implicit controllers must be threaded too */
6117 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6119 if (ss->implicit_on_dfl)
6120 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6121 else if (!ss->dfl_cftypes)
6122 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6125 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6127 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6128 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6130 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6131 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6135 ss->bind(init_css_set.subsys[ssid]);
6137 mutex_lock(&cgroup_mutex);
6138 css_populate_dir(init_css_set.subsys[ssid]);
6139 mutex_unlock(&cgroup_mutex);
6142 /* init_css_set.subsys[] has been updated, re-hash */
6143 hash_del(&init_css_set.hlist);
6144 hash_add(css_set_table, &init_css_set.hlist,
6145 css_set_hash(init_css_set.subsys));
6147 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6148 WARN_ON(register_filesystem(&cgroup_fs_type));
6149 WARN_ON(register_filesystem(&cgroup2_fs_type));
6150 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6151 #ifdef CONFIG_CPUSETS
6152 WARN_ON(register_filesystem(&cpuset_fs_type));
6158 static int __init cgroup_wq_init(void)
6161 * There isn't much point in executing destruction path in
6162 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6163 * Use 1 for @max_active.
6165 * We would prefer to do this in cgroup_init() above, but that
6166 * is called before init_workqueues(): so leave this until after.
6168 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6169 BUG_ON(!cgroup_destroy_wq);
6172 core_initcall(cgroup_wq_init);
6174 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6176 struct kernfs_node *kn;
6178 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6181 kernfs_path(kn, buf, buflen);
6186 * cgroup_get_from_id : get the cgroup associated with cgroup id
6188 * On success return the cgrp or ERR_PTR on failure
6189 * Only cgroups within current task's cgroup NS are valid.
6191 struct cgroup *cgroup_get_from_id(u64 id)
6193 struct kernfs_node *kn;
6194 struct cgroup *cgrp, *root_cgrp;
6196 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6198 return ERR_PTR(-ENOENT);
6200 if (kernfs_type(kn) != KERNFS_DIR) {
6202 return ERR_PTR(-ENOENT);
6207 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6208 if (cgrp && !cgroup_tryget(cgrp))
6215 return ERR_PTR(-ENOENT);
6217 root_cgrp = current_cgns_cgroup_dfl();
6218 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6220 return ERR_PTR(-ENOENT);
6225 EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6228 * proc_cgroup_show()
6229 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6230 * - Used for /proc/<pid>/cgroup.
6232 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6233 struct pid *pid, struct task_struct *tsk)
6237 struct cgroup_root *root;
6240 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6244 mutex_lock(&cgroup_mutex);
6245 spin_lock_irq(&css_set_lock);
6247 for_each_root(root) {
6248 struct cgroup_subsys *ss;
6249 struct cgroup *cgrp;
6250 int ssid, count = 0;
6252 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6255 seq_printf(m, "%d:", root->hierarchy_id);
6256 if (root != &cgrp_dfl_root)
6257 for_each_subsys(ss, ssid)
6258 if (root->subsys_mask & (1 << ssid))
6259 seq_printf(m, "%s%s", count++ ? "," : "",
6261 if (strlen(root->name))
6262 seq_printf(m, "%sname=%s", count ? "," : "",
6266 cgrp = task_cgroup_from_root(tsk, root);
6269 * On traditional hierarchies, all zombie tasks show up as
6270 * belonging to the root cgroup. On the default hierarchy,
6271 * while a zombie doesn't show up in "cgroup.procs" and
6272 * thus can't be migrated, its /proc/PID/cgroup keeps
6273 * reporting the cgroup it belonged to before exiting. If
6274 * the cgroup is removed before the zombie is reaped,
6275 * " (deleted)" is appended to the cgroup path.
6277 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6278 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6279 current->nsproxy->cgroup_ns);
6280 if (retval >= PATH_MAX)
6281 retval = -ENAMETOOLONG;
6290 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6291 seq_puts(m, " (deleted)\n");
6298 spin_unlock_irq(&css_set_lock);
6299 mutex_unlock(&cgroup_mutex);
6306 * cgroup_fork - initialize cgroup related fields during copy_process()
6307 * @child: pointer to task_struct of forking parent process.
6309 * A task is associated with the init_css_set until cgroup_post_fork()
6310 * attaches it to the target css_set.
6312 void cgroup_fork(struct task_struct *child)
6314 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6315 INIT_LIST_HEAD(&child->cg_list);
6319 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6320 * @f: file corresponding to cgroup_dir
6322 * Find the cgroup from a file pointer associated with a cgroup directory.
6323 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6324 * cgroup cannot be found.
6326 static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6328 struct cgroup_subsys_state *css;
6330 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6332 return ERR_CAST(css);
6338 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6340 * @f: file corresponding to cgroup2_dir
6342 static struct cgroup *cgroup_get_from_file(struct file *f)
6344 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6347 return ERR_CAST(cgrp);
6349 if (!cgroup_on_dfl(cgrp)) {
6351 return ERR_PTR(-EBADF);
6358 * cgroup_css_set_fork - find or create a css_set for a child process
6359 * @kargs: the arguments passed to create the child process
6361 * This functions finds or creates a new css_set which the child
6362 * process will be attached to in cgroup_post_fork(). By default,
6363 * the child process will be given the same css_set as its parent.
6365 * If CLONE_INTO_CGROUP is specified this function will try to find an
6366 * existing css_set which includes the requested cgroup and if not create
6367 * a new css_set that the child will be attached to later. If this function
6368 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6369 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6370 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6371 * to the target cgroup.
6373 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6374 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6377 struct cgroup *dst_cgrp = NULL;
6378 struct css_set *cset;
6379 struct super_block *sb;
6382 if (kargs->flags & CLONE_INTO_CGROUP)
6383 mutex_lock(&cgroup_mutex);
6385 cgroup_threadgroup_change_begin(current);
6387 spin_lock_irq(&css_set_lock);
6388 cset = task_css_set(current);
6390 spin_unlock_irq(&css_set_lock);
6392 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6397 f = fget_raw(kargs->cgroup);
6402 sb = f->f_path.dentry->d_sb;
6404 dst_cgrp = cgroup_get_from_file(f);
6405 if (IS_ERR(dst_cgrp)) {
6406 ret = PTR_ERR(dst_cgrp);
6411 if (cgroup_is_dead(dst_cgrp)) {
6417 * Verify that we the target cgroup is writable for us. This is
6418 * usually done by the vfs layer but since we're not going through
6419 * the vfs layer here we need to do it "manually".
6421 ret = cgroup_may_write(dst_cgrp, sb);
6426 * Spawning a task directly into a cgroup works by passing a file
6427 * descriptor to the target cgroup directory. This can even be an O_PATH
6428 * file descriptor. But it can never be a cgroup.procs file descriptor.
6429 * This was done on purpose so spawning into a cgroup could be
6430 * conceptualized as an atomic
6432 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6433 * write(fd, <child-pid>, ...);
6435 * sequence, i.e. it's a shorthand for the caller opening and writing
6436 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6437 * to always use the caller's credentials.
6439 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6440 !(kargs->flags & CLONE_THREAD),
6441 current->nsproxy->cgroup_ns);
6445 kargs->cset = find_css_set(cset, dst_cgrp);
6453 kargs->cgrp = dst_cgrp;
6457 cgroup_threadgroup_change_end(current);
6458 mutex_unlock(&cgroup_mutex);
6462 cgroup_put(dst_cgrp);
6465 put_css_set(kargs->cset);
6470 * cgroup_css_set_put_fork - drop references we took during fork
6471 * @kargs: the arguments passed to create the child process
6473 * Drop references to the prepared css_set and target cgroup if
6474 * CLONE_INTO_CGROUP was requested.
6476 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6477 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6479 cgroup_threadgroup_change_end(current);
6481 if (kargs->flags & CLONE_INTO_CGROUP) {
6482 struct cgroup *cgrp = kargs->cgrp;
6483 struct css_set *cset = kargs->cset;
6485 mutex_unlock(&cgroup_mutex);
6500 * cgroup_can_fork - called on a new task before the process is exposed
6501 * @child: the child process
6502 * @kargs: the arguments passed to create the child process
6504 * This prepares a new css_set for the child process which the child will
6505 * be attached to in cgroup_post_fork().
6506 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6507 * callback returns an error, the fork aborts with that error code. This
6508 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6510 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6512 struct cgroup_subsys *ss;
6515 ret = cgroup_css_set_fork(kargs);
6519 do_each_subsys_mask(ss, i, have_canfork_callback) {
6520 ret = ss->can_fork(child, kargs->cset);
6523 } while_each_subsys_mask();
6528 for_each_subsys(ss, j) {
6531 if (ss->cancel_fork)
6532 ss->cancel_fork(child, kargs->cset);
6535 cgroup_css_set_put_fork(kargs);
6541 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6542 * @child: the child process
6543 * @kargs: the arguments passed to create the child process
6545 * This calls the cancel_fork() callbacks if a fork failed *after*
6546 * cgroup_can_fork() succeeded and cleans up references we took to
6547 * prepare a new css_set for the child process in cgroup_can_fork().
6549 void cgroup_cancel_fork(struct task_struct *child,
6550 struct kernel_clone_args *kargs)
6552 struct cgroup_subsys *ss;
6555 for_each_subsys(ss, i)
6556 if (ss->cancel_fork)
6557 ss->cancel_fork(child, kargs->cset);
6559 cgroup_css_set_put_fork(kargs);
6563 * cgroup_post_fork - finalize cgroup setup for the child process
6564 * @child: the child process
6565 * @kargs: the arguments passed to create the child process
6567 * Attach the child process to its css_set calling the subsystem fork()
6570 void cgroup_post_fork(struct task_struct *child,
6571 struct kernel_clone_args *kargs)
6572 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6574 unsigned long cgrp_flags = 0;
6576 struct cgroup_subsys *ss;
6577 struct css_set *cset;
6583 spin_lock_irq(&css_set_lock);
6585 /* init tasks are special, only link regular threads */
6586 if (likely(child->pid)) {
6588 cgrp_flags = kargs->cgrp->flags;
6590 cgrp_flags = cset->dfl_cgrp->flags;
6592 WARN_ON_ONCE(!list_empty(&child->cg_list));
6594 css_set_move_task(child, NULL, cset, false);
6600 if (!(child->flags & PF_KTHREAD)) {
6601 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6603 * If the cgroup has to be frozen, the new task has
6604 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6605 * get the task into the frozen state.
6607 spin_lock(&child->sighand->siglock);
6608 WARN_ON_ONCE(child->frozen);
6609 child->jobctl |= JOBCTL_TRAP_FREEZE;
6610 spin_unlock(&child->sighand->siglock);
6613 * Calling cgroup_update_frozen() isn't required here,
6614 * because it will be called anyway a bit later from
6615 * do_freezer_trap(). So we avoid cgroup's transient
6616 * switch from the frozen state and back.
6621 * If the cgroup is to be killed notice it now and take the
6622 * child down right after we finished preparing it for
6625 kill = test_bit(CGRP_KILL, &cgrp_flags);
6628 spin_unlock_irq(&css_set_lock);
6631 * Call ss->fork(). This must happen after @child is linked on
6632 * css_set; otherwise, @child might change state between ->fork()
6633 * and addition to css_set.
6635 do_each_subsys_mask(ss, i, have_fork_callback) {
6637 } while_each_subsys_mask();
6639 /* Make the new cset the root_cset of the new cgroup namespace. */
6640 if (kargs->flags & CLONE_NEWCGROUP) {
6641 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6644 child->nsproxy->cgroup_ns->root_cset = cset;
6648 /* Cgroup has to be killed so take down child immediately. */
6650 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6652 cgroup_css_set_put_fork(kargs);
6656 * cgroup_exit - detach cgroup from exiting task
6657 * @tsk: pointer to task_struct of exiting process
6659 * Description: Detach cgroup from @tsk.
6662 void cgroup_exit(struct task_struct *tsk)
6664 struct cgroup_subsys *ss;
6665 struct css_set *cset;
6668 spin_lock_irq(&css_set_lock);
6670 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6671 cset = task_css_set(tsk);
6672 css_set_move_task(tsk, cset, NULL, false);
6673 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6676 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6677 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6678 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6679 cgroup_update_frozen(task_dfl_cgroup(tsk));
6681 spin_unlock_irq(&css_set_lock);
6683 /* see cgroup_post_fork() for details */
6684 do_each_subsys_mask(ss, i, have_exit_callback) {
6686 } while_each_subsys_mask();
6689 void cgroup_release(struct task_struct *task)
6691 struct cgroup_subsys *ss;
6694 do_each_subsys_mask(ss, ssid, have_release_callback) {
6696 } while_each_subsys_mask();
6698 spin_lock_irq(&css_set_lock);
6699 css_set_skip_task_iters(task_css_set(task), task);
6700 list_del_init(&task->cg_list);
6701 spin_unlock_irq(&css_set_lock);
6704 void cgroup_free(struct task_struct *task)
6706 struct css_set *cset = task_css_set(task);
6710 static int __init cgroup_disable(char *str)
6712 struct cgroup_subsys *ss;
6716 while ((token = strsep(&str, ",")) != NULL) {
6720 for_each_subsys(ss, i) {
6721 if (strcmp(token, ss->name) &&
6722 strcmp(token, ss->legacy_name))
6725 static_branch_disable(cgroup_subsys_enabled_key[i]);
6726 pr_info("Disabling %s control group subsystem\n",
6730 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6731 if (strcmp(token, cgroup_opt_feature_names[i]))
6733 cgroup_feature_disable_mask |= 1 << i;
6734 pr_info("Disabling %s control group feature\n",
6735 cgroup_opt_feature_names[i]);
6741 __setup("cgroup_disable=", cgroup_disable);
6743 void __init __weak enable_debug_cgroup(void) { }
6745 static int __init enable_cgroup_debug(char *str)
6747 cgroup_debug = true;
6748 enable_debug_cgroup();
6751 __setup("cgroup_debug", enable_cgroup_debug);
6754 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6755 * @dentry: directory dentry of interest
6756 * @ss: subsystem of interest
6758 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6759 * to get the corresponding css and return it. If such css doesn't exist
6760 * or can't be pinned, an ERR_PTR value is returned.
6762 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6763 struct cgroup_subsys *ss)
6765 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6766 struct file_system_type *s_type = dentry->d_sb->s_type;
6767 struct cgroup_subsys_state *css = NULL;
6768 struct cgroup *cgrp;
6770 /* is @dentry a cgroup dir? */
6771 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6772 !kn || kernfs_type(kn) != KERNFS_DIR)
6773 return ERR_PTR(-EBADF);
6778 * This path doesn't originate from kernfs and @kn could already
6779 * have been or be removed at any point. @kn->priv is RCU
6780 * protected for this access. See css_release_work_fn() for details.
6782 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6784 css = cgroup_css(cgrp, ss);
6786 if (!css || !css_tryget_online(css))
6787 css = ERR_PTR(-ENOENT);
6794 * css_from_id - lookup css by id
6795 * @id: the cgroup id
6796 * @ss: cgroup subsys to be looked into
6798 * Returns the css if there's valid one with @id, otherwise returns NULL.
6799 * Should be called under rcu_read_lock().
6801 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6803 WARN_ON_ONCE(!rcu_read_lock_held());
6804 return idr_find(&ss->css_idr, id);
6808 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6809 * @path: path on the default hierarchy
6811 * Find the cgroup at @path on the default hierarchy, increment its
6812 * reference count and return it. Returns pointer to the found cgroup on
6813 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6814 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6816 struct cgroup *cgroup_get_from_path(const char *path)
6818 struct kernfs_node *kn;
6819 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6820 struct cgroup *root_cgrp;
6822 root_cgrp = current_cgns_cgroup_dfl();
6823 kn = kernfs_walk_and_get(root_cgrp->kn, path);
6827 if (kernfs_type(kn) != KERNFS_DIR) {
6828 cgrp = ERR_PTR(-ENOTDIR);
6834 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6835 if (!cgrp || !cgroup_tryget(cgrp))
6836 cgrp = ERR_PTR(-ENOENT);
6845 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6848 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6849 * @fd: fd obtained by open(cgroup_dir)
6851 * Find the cgroup from a fd which should be obtained
6852 * by opening a cgroup directory. Returns a pointer to the
6853 * cgroup on success. ERR_PTR is returned if the cgroup
6856 struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6858 struct cgroup *cgrp;
6863 return ERR_PTR(-EBADF);
6865 cgrp = cgroup_v1v2_get_from_file(f);
6871 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6873 * @fd: fd obtained by open(cgroup2_dir)
6875 struct cgroup *cgroup_get_from_fd(int fd)
6877 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6880 return ERR_CAST(cgrp);
6882 if (!cgroup_on_dfl(cgrp)) {
6884 return ERR_PTR(-EBADF);
6888 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6890 static u64 power_of_ten(int power)
6899 * cgroup_parse_float - parse a floating number
6900 * @input: input string
6901 * @dec_shift: number of decimal digits to shift
6904 * Parse a decimal floating point number in @input and store the result in
6905 * @v with decimal point right shifted @dec_shift times. For example, if
6906 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6907 * Returns 0 on success, -errno otherwise.
6909 * There's nothing cgroup specific about this function except that it's
6910 * currently the only user.
6912 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6914 s64 whole, frac = 0;
6915 int fstart = 0, fend = 0, flen;
6917 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6922 flen = fend > fstart ? fend - fstart : 0;
6923 if (flen < dec_shift)
6924 frac *= power_of_ten(dec_shift - flen);
6926 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6928 *v = whole * power_of_ten(dec_shift) + frac;
6933 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6934 * definition in cgroup-defs.h.
6936 #ifdef CONFIG_SOCK_CGROUP_DATA
6938 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6940 struct cgroup *cgroup;
6943 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6944 if (in_interrupt()) {
6945 cgroup = &cgrp_dfl_root.cgrp;
6951 struct css_set *cset;
6953 cset = task_css_set(current);
6954 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6955 cgroup = cset->dfl_cgrp;
6961 skcd->cgroup = cgroup;
6962 cgroup_bpf_get(cgroup);
6966 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6968 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6971 * We might be cloning a socket which is left in an empty
6972 * cgroup and the cgroup might have already been rmdir'd.
6973 * Don't use cgroup_get_live().
6976 cgroup_bpf_get(cgrp);
6979 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6981 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6983 cgroup_bpf_put(cgrp);
6987 #endif /* CONFIG_SOCK_CGROUP_DATA */
6990 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6991 ssize_t size, const char *prefix)
6996 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6997 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7001 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
7003 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
7005 if (WARN_ON(ret >= size))
7012 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7015 struct cgroup_subsys *ss;
7019 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7020 PAGE_SIZE - ret, NULL);
7021 if (cgroup_psi_enabled())
7022 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7023 PAGE_SIZE - ret, NULL);
7025 for_each_subsys(ss, ssid)
7026 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7028 cgroup_subsys_name[ssid]);
7032 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7034 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7037 return snprintf(buf, PAGE_SIZE,
7040 "memory_localevents\n"
7041 "memory_recursiveprot\n");
7043 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7045 static struct attribute *cgroup_sysfs_attrs[] = {
7046 &cgroup_delegate_attr.attr,
7047 &cgroup_features_attr.attr,
7051 static const struct attribute_group cgroup_sysfs_attr_group = {
7052 .attrs = cgroup_sysfs_attrs,
7056 static int __init cgroup_sysfs_init(void)
7058 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7060 subsys_initcall(cgroup_sysfs_init);
7062 #endif /* CONFIG_SYSFS */