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/sched/deadline.h>
61 #include <linux/psi.h>
64 #define CREATE_TRACE_POINTS
65 #include <trace/events/cgroup.h>
67 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
69 /* let's not notify more than 100 times per second */
70 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
73 * To avoid confusing the compiler (and generating warnings) with code
74 * that attempts to access what would be a 0-element array (i.e. sized
75 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
76 * constant expression can be added.
78 #define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
81 * cgroup_mutex is the master lock. Any modification to cgroup or its
82 * hierarchy must be performed while holding it.
84 * css_set_lock protects task->cgroups pointer, the list of css_set
85 * objects, and the chain of tasks off each css_set.
87 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
88 * cgroup.h can use them for lockdep annotations.
90 DEFINE_MUTEX(cgroup_mutex);
91 DEFINE_SPINLOCK(css_set_lock);
93 #ifdef CONFIG_PROVE_RCU
94 EXPORT_SYMBOL_GPL(cgroup_mutex);
95 EXPORT_SYMBOL_GPL(css_set_lock);
98 DEFINE_SPINLOCK(trace_cgroup_path_lock);
99 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
100 static bool cgroup_debug __read_mostly;
103 * Protects cgroup_idr and css_idr so that IDs can be released without
104 * grabbing cgroup_mutex.
106 static DEFINE_SPINLOCK(cgroup_idr_lock);
109 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
110 * against file removal/re-creation across css hiding.
112 static DEFINE_SPINLOCK(cgroup_file_kn_lock);
114 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
116 #define cgroup_assert_mutex_or_rcu_locked() \
117 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
118 !lockdep_is_held(&cgroup_mutex), \
119 "cgroup_mutex or RCU read lock required");
122 * cgroup destruction makes heavy use of work items and there can be a lot
123 * of concurrent destructions. Use a separate workqueue so that cgroup
124 * destruction work items don't end up filling up max_active of system_wq
125 * which may lead to deadlock.
127 static struct workqueue_struct *cgroup_destroy_wq;
129 /* generate an array of cgroup subsystem pointers */
130 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
131 struct cgroup_subsys *cgroup_subsys[] = {
132 #include <linux/cgroup_subsys.h>
136 /* array of cgroup subsystem names */
137 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
138 static const char *cgroup_subsys_name[] = {
139 #include <linux/cgroup_subsys.h>
143 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
146 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
148 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
149 #include <linux/cgroup_subsys.h>
152 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
153 static struct static_key_true *cgroup_subsys_enabled_key[] = {
154 #include <linux/cgroup_subsys.h>
158 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
159 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
160 #include <linux/cgroup_subsys.h>
164 static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
166 /* the default hierarchy */
167 struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
168 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
171 * The default hierarchy always exists but is hidden until mounted for the
172 * first time. This is for backward compatibility.
174 static bool cgrp_dfl_visible;
176 /* some controllers are not supported in the default hierarchy */
177 static u16 cgrp_dfl_inhibit_ss_mask;
179 /* some controllers are implicitly enabled on the default hierarchy */
180 static u16 cgrp_dfl_implicit_ss_mask;
182 /* some controllers can be threaded on the default hierarchy */
183 static u16 cgrp_dfl_threaded_ss_mask;
185 /* The list of hierarchy roots */
186 LIST_HEAD(cgroup_roots);
187 static int cgroup_root_count;
189 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
190 static DEFINE_IDR(cgroup_hierarchy_idr);
193 * Assign a monotonically increasing serial number to csses. It guarantees
194 * cgroups with bigger numbers are newer than those with smaller numbers.
195 * Also, as csses are always appended to the parent's ->children list, it
196 * guarantees that sibling csses are always sorted in the ascending serial
197 * number order on the list. Protected by cgroup_mutex.
199 static u64 css_serial_nr_next = 1;
202 * These bitmasks identify subsystems with specific features to avoid
203 * having to do iterative checks repeatedly.
205 static u16 have_fork_callback __read_mostly;
206 static u16 have_exit_callback __read_mostly;
207 static u16 have_release_callback __read_mostly;
208 static u16 have_canfork_callback __read_mostly;
210 static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS);
212 /* cgroup namespace for init task */
213 struct cgroup_namespace init_cgroup_ns = {
214 .ns.count = REFCOUNT_INIT(2),
215 .user_ns = &init_user_ns,
216 .ns.ops = &cgroupns_operations,
217 .ns.inum = PROC_CGROUP_INIT_INO,
218 .root_cset = &init_css_set,
221 static struct file_system_type cgroup2_fs_type;
222 static struct cftype cgroup_base_files[];
223 static struct cftype cgroup_psi_files[];
225 /* cgroup optional features */
226 enum cgroup_opt_features {
228 OPT_FEATURE_PRESSURE,
233 static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
239 static u16 cgroup_feature_disable_mask __read_mostly;
241 static int cgroup_apply_control(struct cgroup *cgrp);
242 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
243 static void css_task_iter_skip(struct css_task_iter *it,
244 struct task_struct *task);
245 static int cgroup_destroy_locked(struct cgroup *cgrp);
246 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
247 struct cgroup_subsys *ss);
248 static void css_release(struct percpu_ref *ref);
249 static void kill_css(struct cgroup_subsys_state *css);
250 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
251 struct cgroup *cgrp, struct cftype cfts[],
254 #ifdef CONFIG_DEBUG_CGROUP_REF
255 #define CGROUP_REF_FN_ATTRS noinline
256 #define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
257 #include <linux/cgroup_refcnt.h>
261 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
262 * @ssid: subsys ID of interest
264 * cgroup_subsys_enabled() can only be used with literal subsys names which
265 * is fine for individual subsystems but unsuitable for cgroup core. This
266 * is slower static_key_enabled() based test indexed by @ssid.
268 bool cgroup_ssid_enabled(int ssid)
270 if (!CGROUP_HAS_SUBSYS_CONFIG)
273 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
277 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
278 * @cgrp: the cgroup of interest
280 * The default hierarchy is the v2 interface of cgroup and this function
281 * can be used to test whether a cgroup is on the default hierarchy for
282 * cases where a subsystem should behave differently depending on the
285 * List of changed behaviors:
287 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
288 * and "name" are disallowed.
290 * - When mounting an existing superblock, mount options should match.
292 * - rename(2) is disallowed.
294 * - "tasks" is removed. Everything should be at process granularity. Use
295 * "cgroup.procs" instead.
297 * - "cgroup.procs" is not sorted. pids will be unique unless they got
298 * recycled in-between reads.
300 * - "release_agent" and "notify_on_release" are removed. Replacement
301 * notification mechanism will be implemented.
303 * - "cgroup.clone_children" is removed.
305 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
306 * and its descendants contain no task; otherwise, 1. The file also
307 * generates kernfs notification which can be monitored through poll and
308 * [di]notify when the value of the file changes.
310 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
311 * take masks of ancestors with non-empty cpus/mems, instead of being
312 * moved to an ancestor.
314 * - cpuset: a task can be moved into an empty cpuset, and again it takes
315 * masks of ancestors.
317 * - blkcg: blk-throttle becomes properly hierarchical.
319 bool cgroup_on_dfl(const struct cgroup *cgrp)
321 return cgrp->root == &cgrp_dfl_root;
324 /* IDR wrappers which synchronize using cgroup_idr_lock */
325 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
330 idr_preload(gfp_mask);
331 spin_lock_bh(&cgroup_idr_lock);
332 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
333 spin_unlock_bh(&cgroup_idr_lock);
338 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
342 spin_lock_bh(&cgroup_idr_lock);
343 ret = idr_replace(idr, ptr, id);
344 spin_unlock_bh(&cgroup_idr_lock);
348 static void cgroup_idr_remove(struct idr *idr, int id)
350 spin_lock_bh(&cgroup_idr_lock);
352 spin_unlock_bh(&cgroup_idr_lock);
355 static bool cgroup_has_tasks(struct cgroup *cgrp)
357 return cgrp->nr_populated_csets;
360 static bool cgroup_is_threaded(struct cgroup *cgrp)
362 return cgrp->dom_cgrp != cgrp;
365 /* can @cgrp host both domain and threaded children? */
366 static bool cgroup_is_mixable(struct cgroup *cgrp)
369 * Root isn't under domain level resource control exempting it from
370 * the no-internal-process constraint, so it can serve as a thread
371 * root and a parent of resource domains at the same time.
373 return !cgroup_parent(cgrp);
376 /* can @cgrp become a thread root? Should always be true for a thread root */
377 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
379 /* mixables don't care */
380 if (cgroup_is_mixable(cgrp))
383 /* domain roots can't be nested under threaded */
384 if (cgroup_is_threaded(cgrp))
387 /* can only have either domain or threaded children */
388 if (cgrp->nr_populated_domain_children)
391 /* and no domain controllers can be enabled */
392 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
398 /* is @cgrp root of a threaded subtree? */
399 static bool cgroup_is_thread_root(struct cgroup *cgrp)
401 /* thread root should be a domain */
402 if (cgroup_is_threaded(cgrp))
405 /* a domain w/ threaded children is a thread root */
406 if (cgrp->nr_threaded_children)
410 * A domain which has tasks and explicit threaded controllers
411 * enabled is a thread root.
413 if (cgroup_has_tasks(cgrp) &&
414 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
420 /* a domain which isn't connected to the root w/o brekage can't be used */
421 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
423 /* the cgroup itself can be a thread root */
424 if (cgroup_is_threaded(cgrp))
427 /* but the ancestors can't be unless mixable */
428 while ((cgrp = cgroup_parent(cgrp))) {
429 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
431 if (cgroup_is_threaded(cgrp))
438 /* subsystems visibly enabled on a cgroup */
439 static u16 cgroup_control(struct cgroup *cgrp)
441 struct cgroup *parent = cgroup_parent(cgrp);
442 u16 root_ss_mask = cgrp->root->subsys_mask;
445 u16 ss_mask = parent->subtree_control;
447 /* threaded cgroups can only have threaded controllers */
448 if (cgroup_is_threaded(cgrp))
449 ss_mask &= cgrp_dfl_threaded_ss_mask;
453 if (cgroup_on_dfl(cgrp))
454 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
455 cgrp_dfl_implicit_ss_mask);
459 /* subsystems enabled on a cgroup */
460 static u16 cgroup_ss_mask(struct cgroup *cgrp)
462 struct cgroup *parent = cgroup_parent(cgrp);
465 u16 ss_mask = parent->subtree_ss_mask;
467 /* threaded cgroups can only have threaded controllers */
468 if (cgroup_is_threaded(cgrp))
469 ss_mask &= cgrp_dfl_threaded_ss_mask;
473 return cgrp->root->subsys_mask;
477 * cgroup_css - obtain a cgroup's css for the specified subsystem
478 * @cgrp: the cgroup of interest
479 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
481 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
482 * function must be called either under cgroup_mutex or rcu_read_lock() and
483 * the caller is responsible for pinning the returned css if it wants to
484 * keep accessing it outside the said locks. This function may return
485 * %NULL if @cgrp doesn't have @subsys_id enabled.
487 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
488 struct cgroup_subsys *ss)
490 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
491 return rcu_dereference_check(cgrp->subsys[ss->id],
492 lockdep_is_held(&cgroup_mutex));
498 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
499 * @cgrp: the cgroup of interest
500 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
502 * Similar to cgroup_css() but returns the effective css, which is defined
503 * as the matching css of the nearest ancestor including self which has @ss
504 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
505 * function is guaranteed to return non-NULL css.
507 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
508 struct cgroup_subsys *ss)
510 lockdep_assert_held(&cgroup_mutex);
516 * This function is used while updating css associations and thus
517 * can't test the csses directly. Test ss_mask.
519 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
520 cgrp = cgroup_parent(cgrp);
525 return cgroup_css(cgrp, ss);
529 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
530 * @cgrp: the cgroup of interest
531 * @ss: the subsystem of interest
533 * Find and get the effective css of @cgrp for @ss. The effective css is
534 * defined as the matching css of the nearest ancestor including self which
535 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
536 * the root css is returned, so this function always returns a valid css.
538 * The returned css is not guaranteed to be online, and therefore it is the
539 * callers responsibility to try get a reference for it.
541 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
542 struct cgroup_subsys *ss)
544 struct cgroup_subsys_state *css;
546 if (!CGROUP_HAS_SUBSYS_CONFIG)
550 css = cgroup_css(cgrp, ss);
554 cgrp = cgroup_parent(cgrp);
557 return init_css_set.subsys[ss->id];
561 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
562 * @cgrp: the cgroup of interest
563 * @ss: the subsystem of interest
565 * Find and get the effective css of @cgrp for @ss. The effective css is
566 * defined as the matching css of the nearest ancestor including self which
567 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
568 * the root css is returned, so this function always returns a valid css.
569 * The returned css must be put using css_put().
571 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
572 struct cgroup_subsys *ss)
574 struct cgroup_subsys_state *css;
576 if (!CGROUP_HAS_SUBSYS_CONFIG)
582 css = cgroup_css(cgrp, ss);
584 if (css && css_tryget_online(css))
586 cgrp = cgroup_parent(cgrp);
589 css = init_css_set.subsys[ss->id];
595 EXPORT_SYMBOL_GPL(cgroup_get_e_css);
597 static void cgroup_get_live(struct cgroup *cgrp)
599 WARN_ON_ONCE(cgroup_is_dead(cgrp));
604 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
605 * is responsible for taking the css_set_lock.
606 * @cgrp: the cgroup in question
608 int __cgroup_task_count(const struct cgroup *cgrp)
611 struct cgrp_cset_link *link;
613 lockdep_assert_held(&css_set_lock);
615 list_for_each_entry(link, &cgrp->cset_links, cset_link)
616 count += link->cset->nr_tasks;
622 * cgroup_task_count - count the number of tasks in a cgroup.
623 * @cgrp: the cgroup in question
625 int cgroup_task_count(const struct cgroup *cgrp)
629 spin_lock_irq(&css_set_lock);
630 count = __cgroup_task_count(cgrp);
631 spin_unlock_irq(&css_set_lock);
636 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
638 struct cgroup *cgrp = of->kn->parent->priv;
639 struct cftype *cft = of_cft(of);
642 * This is open and unprotected implementation of cgroup_css().
643 * seq_css() is only called from a kernfs file operation which has
644 * an active reference on the file. Because all the subsystem
645 * files are drained before a css is disassociated with a cgroup,
646 * the matching css from the cgroup's subsys table is guaranteed to
647 * be and stay valid until the enclosing operation is complete.
649 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
650 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
654 EXPORT_SYMBOL_GPL(of_css);
657 * for_each_css - iterate all css's of a cgroup
658 * @css: the iteration cursor
659 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
660 * @cgrp: the target cgroup to iterate css's of
662 * Should be called under cgroup_mutex.
664 #define for_each_css(css, ssid, cgrp) \
665 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
666 if (!((css) = rcu_dereference_check( \
667 (cgrp)->subsys[(ssid)], \
668 lockdep_is_held(&cgroup_mutex)))) { } \
672 * do_each_subsys_mask - filter for_each_subsys with a bitmask
673 * @ss: the iteration cursor
674 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
675 * @ss_mask: the bitmask
677 * The block will only run for cases where the ssid-th bit (1 << ssid) of
680 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
681 unsigned long __ss_mask = (ss_mask); \
682 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
686 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
687 (ss) = cgroup_subsys[ssid]; \
690 #define while_each_subsys_mask() \
695 /* iterate over child cgrps, lock should be held throughout iteration */
696 #define cgroup_for_each_live_child(child, cgrp) \
697 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
698 if (({ lockdep_assert_held(&cgroup_mutex); \
699 cgroup_is_dead(child); })) \
703 /* walk live descendants in pre order */
704 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
705 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
706 if (({ lockdep_assert_held(&cgroup_mutex); \
707 (dsct) = (d_css)->cgroup; \
708 cgroup_is_dead(dsct); })) \
712 /* walk live descendants in postorder */
713 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
714 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
715 if (({ lockdep_assert_held(&cgroup_mutex); \
716 (dsct) = (d_css)->cgroup; \
717 cgroup_is_dead(dsct); })) \
722 * The default css_set - used by init and its children prior to any
723 * hierarchies being mounted. It contains a pointer to the root state
724 * for each subsystem. Also used to anchor the list of css_sets. Not
725 * reference-counted, to improve performance when child cgroups
726 * haven't been created.
728 struct css_set init_css_set = {
729 .refcount = REFCOUNT_INIT(1),
730 .dom_cset = &init_css_set,
731 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
732 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
733 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
734 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
735 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
736 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
737 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
738 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
739 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
742 * The following field is re-initialized when this cset gets linked
743 * in cgroup_init(). However, let's initialize the field
744 * statically too so that the default cgroup can be accessed safely
747 .dfl_cgrp = &cgrp_dfl_root.cgrp,
750 static int css_set_count = 1; /* 1 for init_css_set */
752 static bool css_set_threaded(struct css_set *cset)
754 return cset->dom_cset != cset;
758 * css_set_populated - does a css_set contain any tasks?
759 * @cset: target css_set
761 * css_set_populated() should be the same as !!cset->nr_tasks at steady
762 * state. However, css_set_populated() can be called while a task is being
763 * added to or removed from the linked list before the nr_tasks is
764 * properly updated. Hence, we can't just look at ->nr_tasks here.
766 static bool css_set_populated(struct css_set *cset)
768 lockdep_assert_held(&css_set_lock);
770 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
774 * cgroup_update_populated - update the populated count of a cgroup
775 * @cgrp: the target cgroup
776 * @populated: inc or dec populated count
778 * One of the css_sets associated with @cgrp is either getting its first
779 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
780 * count is propagated towards root so that a given cgroup's
781 * nr_populated_children is zero iff none of its descendants contain any
784 * @cgrp's interface file "cgroup.populated" is zero if both
785 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
786 * 1 otherwise. When the sum changes from or to zero, userland is notified
787 * that the content of the interface file has changed. This can be used to
788 * detect when @cgrp and its descendants become populated or empty.
790 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
792 struct cgroup *child = NULL;
793 int adj = populated ? 1 : -1;
795 lockdep_assert_held(&css_set_lock);
798 bool was_populated = cgroup_is_populated(cgrp);
801 cgrp->nr_populated_csets += adj;
803 if (cgroup_is_threaded(child))
804 cgrp->nr_populated_threaded_children += adj;
806 cgrp->nr_populated_domain_children += adj;
809 if (was_populated == cgroup_is_populated(cgrp))
812 cgroup1_check_for_release(cgrp);
813 TRACE_CGROUP_PATH(notify_populated, cgrp,
814 cgroup_is_populated(cgrp));
815 cgroup_file_notify(&cgrp->events_file);
818 cgrp = cgroup_parent(cgrp);
823 * css_set_update_populated - update populated state of a css_set
824 * @cset: target css_set
825 * @populated: whether @cset is populated or depopulated
827 * @cset is either getting the first task or losing the last. Update the
828 * populated counters of all associated cgroups accordingly.
830 static void css_set_update_populated(struct css_set *cset, bool populated)
832 struct cgrp_cset_link *link;
834 lockdep_assert_held(&css_set_lock);
836 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
837 cgroup_update_populated(link->cgrp, populated);
841 * @task is leaving, advance task iterators which are pointing to it so
842 * that they can resume at the next position. Advancing an iterator might
843 * remove it from the list, use safe walk. See css_task_iter_skip() for
846 static void css_set_skip_task_iters(struct css_set *cset,
847 struct task_struct *task)
849 struct css_task_iter *it, *pos;
851 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
852 css_task_iter_skip(it, task);
856 * css_set_move_task - move a task from one css_set to another
857 * @task: task being moved
858 * @from_cset: css_set @task currently belongs to (may be NULL)
859 * @to_cset: new css_set @task is being moved to (may be NULL)
860 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
862 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
863 * css_set, @from_cset can be NULL. If @task is being disassociated
864 * instead of moved, @to_cset can be NULL.
866 * This function automatically handles populated counter updates and
867 * css_task_iter adjustments but the caller is responsible for managing
868 * @from_cset and @to_cset's reference counts.
870 static void css_set_move_task(struct task_struct *task,
871 struct css_set *from_cset, struct css_set *to_cset,
874 lockdep_assert_held(&css_set_lock);
876 if (to_cset && !css_set_populated(to_cset))
877 css_set_update_populated(to_cset, true);
880 WARN_ON_ONCE(list_empty(&task->cg_list));
882 css_set_skip_task_iters(from_cset, task);
883 list_del_init(&task->cg_list);
884 if (!css_set_populated(from_cset))
885 css_set_update_populated(from_cset, false);
887 WARN_ON_ONCE(!list_empty(&task->cg_list));
892 * We are synchronized through cgroup_threadgroup_rwsem
893 * against PF_EXITING setting such that we can't race
894 * against cgroup_exit()/cgroup_free() dropping the css_set.
896 WARN_ON_ONCE(task->flags & PF_EXITING);
898 cgroup_move_task(task, to_cset);
899 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
905 * hash table for cgroup groups. This improves the performance to find
906 * an existing css_set. This hash doesn't (currently) take into
907 * account cgroups in empty hierarchies.
909 #define CSS_SET_HASH_BITS 7
910 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
912 static unsigned long css_set_hash(struct cgroup_subsys_state **css)
914 unsigned long key = 0UL;
915 struct cgroup_subsys *ss;
918 for_each_subsys(ss, i)
919 key += (unsigned long)css[i];
920 key = (key >> 16) ^ key;
925 void put_css_set_locked(struct css_set *cset)
927 struct cgrp_cset_link *link, *tmp_link;
928 struct cgroup_subsys *ss;
931 lockdep_assert_held(&css_set_lock);
933 if (!refcount_dec_and_test(&cset->refcount))
936 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
938 /* This css_set is dead. Unlink it and release cgroup and css refs */
939 for_each_subsys(ss, ssid) {
940 list_del(&cset->e_cset_node[ssid]);
941 css_put(cset->subsys[ssid]);
943 hash_del(&cset->hlist);
946 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
947 list_del(&link->cset_link);
948 list_del(&link->cgrp_link);
949 if (cgroup_parent(link->cgrp))
950 cgroup_put(link->cgrp);
954 if (css_set_threaded(cset)) {
955 list_del(&cset->threaded_csets_node);
956 put_css_set_locked(cset->dom_cset);
959 kfree_rcu(cset, rcu_head);
963 * compare_css_sets - helper function for find_existing_css_set().
964 * @cset: candidate css_set being tested
965 * @old_cset: existing css_set for a task
966 * @new_cgrp: cgroup that's being entered by the task
967 * @template: desired set of css pointers in css_set (pre-calculated)
969 * Returns true if "cset" matches "old_cset" except for the hierarchy
970 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
972 static bool compare_css_sets(struct css_set *cset,
973 struct css_set *old_cset,
974 struct cgroup *new_cgrp,
975 struct cgroup_subsys_state *template[])
977 struct cgroup *new_dfl_cgrp;
978 struct list_head *l1, *l2;
981 * On the default hierarchy, there can be csets which are
982 * associated with the same set of cgroups but different csses.
983 * Let's first ensure that csses match.
985 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
989 /* @cset's domain should match the default cgroup's */
990 if (cgroup_on_dfl(new_cgrp))
991 new_dfl_cgrp = new_cgrp;
993 new_dfl_cgrp = old_cset->dfl_cgrp;
995 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
999 * Compare cgroup pointers in order to distinguish between
1000 * different cgroups in hierarchies. As different cgroups may
1001 * share the same effective css, this comparison is always
1004 l1 = &cset->cgrp_links;
1005 l2 = &old_cset->cgrp_links;
1007 struct cgrp_cset_link *link1, *link2;
1008 struct cgroup *cgrp1, *cgrp2;
1012 /* See if we reached the end - both lists are equal length. */
1013 if (l1 == &cset->cgrp_links) {
1014 BUG_ON(l2 != &old_cset->cgrp_links);
1017 BUG_ON(l2 == &old_cset->cgrp_links);
1019 /* Locate the cgroups associated with these links. */
1020 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1021 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1022 cgrp1 = link1->cgrp;
1023 cgrp2 = link2->cgrp;
1024 /* Hierarchies should be linked in the same order. */
1025 BUG_ON(cgrp1->root != cgrp2->root);
1028 * If this hierarchy is the hierarchy of the cgroup
1029 * that's changing, then we need to check that this
1030 * css_set points to the new cgroup; if it's any other
1031 * hierarchy, then this css_set should point to the
1032 * same cgroup as the old css_set.
1034 if (cgrp1->root == new_cgrp->root) {
1035 if (cgrp1 != new_cgrp)
1046 * find_existing_css_set - init css array and find the matching css_set
1047 * @old_cset: the css_set that we're using before the cgroup transition
1048 * @cgrp: the cgroup that we're moving into
1049 * @template: out param for the new set of csses, should be clear on entry
1051 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1052 struct cgroup *cgrp,
1053 struct cgroup_subsys_state **template)
1055 struct cgroup_root *root = cgrp->root;
1056 struct cgroup_subsys *ss;
1057 struct css_set *cset;
1062 * Build the set of subsystem state objects that we want to see in the
1063 * new css_set. While subsystems can change globally, the entries here
1064 * won't change, so no need for locking.
1066 for_each_subsys(ss, i) {
1067 if (root->subsys_mask & (1UL << i)) {
1069 * @ss is in this hierarchy, so we want the
1070 * effective css from @cgrp.
1072 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1075 * @ss is not in this hierarchy, so we don't want
1076 * to change the css.
1078 template[i] = old_cset->subsys[i];
1082 key = css_set_hash(template);
1083 hash_for_each_possible(css_set_table, cset, hlist, key) {
1084 if (!compare_css_sets(cset, old_cset, cgrp, template))
1087 /* This css_set matches what we need */
1091 /* No existing cgroup group matched */
1095 static void free_cgrp_cset_links(struct list_head *links_to_free)
1097 struct cgrp_cset_link *link, *tmp_link;
1099 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1100 list_del(&link->cset_link);
1106 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1107 * @count: the number of links to allocate
1108 * @tmp_links: list_head the allocated links are put on
1110 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1111 * through ->cset_link. Returns 0 on success or -errno.
1113 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1115 struct cgrp_cset_link *link;
1118 INIT_LIST_HEAD(tmp_links);
1120 for (i = 0; i < count; i++) {
1121 link = kzalloc(sizeof(*link), GFP_KERNEL);
1123 free_cgrp_cset_links(tmp_links);
1126 list_add(&link->cset_link, tmp_links);
1132 * link_css_set - a helper function to link a css_set to a cgroup
1133 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1134 * @cset: the css_set to be linked
1135 * @cgrp: the destination cgroup
1137 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1138 struct cgroup *cgrp)
1140 struct cgrp_cset_link *link;
1142 BUG_ON(list_empty(tmp_links));
1144 if (cgroup_on_dfl(cgrp))
1145 cset->dfl_cgrp = cgrp;
1147 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1152 * Always add links to the tail of the lists so that the lists are
1153 * in chronological order.
1155 list_move_tail(&link->cset_link, &cgrp->cset_links);
1156 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1158 if (cgroup_parent(cgrp))
1159 cgroup_get_live(cgrp);
1163 * find_css_set - return a new css_set with one cgroup updated
1164 * @old_cset: the baseline css_set
1165 * @cgrp: the cgroup to be updated
1167 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1168 * substituted into the appropriate hierarchy.
1170 static struct css_set *find_css_set(struct css_set *old_cset,
1171 struct cgroup *cgrp)
1173 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1174 struct css_set *cset;
1175 struct list_head tmp_links;
1176 struct cgrp_cset_link *link;
1177 struct cgroup_subsys *ss;
1181 lockdep_assert_held(&cgroup_mutex);
1183 /* First see if we already have a cgroup group that matches
1184 * the desired set */
1185 spin_lock_irq(&css_set_lock);
1186 cset = find_existing_css_set(old_cset, cgrp, template);
1189 spin_unlock_irq(&css_set_lock);
1194 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1198 /* Allocate all the cgrp_cset_link objects that we'll need */
1199 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1204 refcount_set(&cset->refcount, 1);
1205 cset->dom_cset = cset;
1206 INIT_LIST_HEAD(&cset->tasks);
1207 INIT_LIST_HEAD(&cset->mg_tasks);
1208 INIT_LIST_HEAD(&cset->dying_tasks);
1209 INIT_LIST_HEAD(&cset->task_iters);
1210 INIT_LIST_HEAD(&cset->threaded_csets);
1211 INIT_HLIST_NODE(&cset->hlist);
1212 INIT_LIST_HEAD(&cset->cgrp_links);
1213 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1214 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1215 INIT_LIST_HEAD(&cset->mg_node);
1217 /* Copy the set of subsystem state objects generated in
1218 * find_existing_css_set() */
1219 memcpy(cset->subsys, template, sizeof(cset->subsys));
1221 spin_lock_irq(&css_set_lock);
1222 /* Add reference counts and links from the new css_set. */
1223 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1224 struct cgroup *c = link->cgrp;
1226 if (c->root == cgrp->root)
1228 link_css_set(&tmp_links, cset, c);
1231 BUG_ON(!list_empty(&tmp_links));
1235 /* Add @cset to the hash table */
1236 key = css_set_hash(cset->subsys);
1237 hash_add(css_set_table, &cset->hlist, key);
1239 for_each_subsys(ss, ssid) {
1240 struct cgroup_subsys_state *css = cset->subsys[ssid];
1242 list_add_tail(&cset->e_cset_node[ssid],
1243 &css->cgroup->e_csets[ssid]);
1247 spin_unlock_irq(&css_set_lock);
1250 * If @cset should be threaded, look up the matching dom_cset and
1251 * link them up. We first fully initialize @cset then look for the
1252 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1253 * to stay empty until we return.
1255 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1256 struct css_set *dcset;
1258 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1264 spin_lock_irq(&css_set_lock);
1265 cset->dom_cset = dcset;
1266 list_add_tail(&cset->threaded_csets_node,
1267 &dcset->threaded_csets);
1268 spin_unlock_irq(&css_set_lock);
1274 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1276 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1278 return root_cgrp->root;
1281 void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1283 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1285 /* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1286 if (favor && !favoring) {
1287 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1288 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1289 } else if (!favor && favoring) {
1290 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1291 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1295 static int cgroup_init_root_id(struct cgroup_root *root)
1299 lockdep_assert_held(&cgroup_mutex);
1301 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1305 root->hierarchy_id = id;
1309 static void cgroup_exit_root_id(struct cgroup_root *root)
1311 lockdep_assert_held(&cgroup_mutex);
1313 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1316 void cgroup_free_root(struct cgroup_root *root)
1321 static void cgroup_destroy_root(struct cgroup_root *root)
1323 struct cgroup *cgrp = &root->cgrp;
1324 struct cgrp_cset_link *link, *tmp_link;
1326 trace_cgroup_destroy_root(root);
1328 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1330 BUG_ON(atomic_read(&root->nr_cgrps));
1331 BUG_ON(!list_empty(&cgrp->self.children));
1333 /* Rebind all subsystems back to the default hierarchy */
1334 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1337 * Release all the links from cset_links to this hierarchy's
1340 spin_lock_irq(&css_set_lock);
1342 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1343 list_del(&link->cset_link);
1344 list_del(&link->cgrp_link);
1348 spin_unlock_irq(&css_set_lock);
1350 if (!list_empty(&root->root_list)) {
1351 list_del(&root->root_list);
1352 cgroup_root_count--;
1355 if (!have_favordynmods)
1356 cgroup_favor_dynmods(root, false);
1358 cgroup_exit_root_id(root);
1362 cgroup_rstat_exit(cgrp);
1363 kernfs_destroy_root(root->kf_root);
1364 cgroup_free_root(root);
1368 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1370 static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1371 struct cgroup_root *root)
1373 struct cgroup *res_cgroup = NULL;
1375 if (cset == &init_css_set) {
1376 res_cgroup = &root->cgrp;
1377 } else if (root == &cgrp_dfl_root) {
1378 res_cgroup = cset->dfl_cgrp;
1380 struct cgrp_cset_link *link;
1381 lockdep_assert_held(&css_set_lock);
1383 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1384 struct cgroup *c = link->cgrp;
1386 if (c->root == root) {
1393 BUG_ON(!res_cgroup);
1398 * look up cgroup associated with current task's cgroup namespace on the
1399 * specified hierarchy
1401 static struct cgroup *
1402 current_cgns_cgroup_from_root(struct cgroup_root *root)
1404 struct cgroup *res = NULL;
1405 struct css_set *cset;
1407 lockdep_assert_held(&css_set_lock);
1411 cset = current->nsproxy->cgroup_ns->root_cset;
1412 res = __cset_cgroup_from_root(cset, root);
1420 * Look up cgroup associated with current task's cgroup namespace on the default
1423 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1424 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1426 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1427 * - As a bonus returned cgrp is pinned with the current because it cannot
1428 * switch cgroup_ns asynchronously.
1430 static struct cgroup *current_cgns_cgroup_dfl(void)
1432 struct css_set *cset;
1434 if (current->nsproxy) {
1435 cset = current->nsproxy->cgroup_ns->root_cset;
1436 return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1439 * NOTE: This function may be called from bpf_cgroup_from_id()
1440 * on a task which has already passed exit_task_namespaces() and
1441 * nsproxy == NULL. Fall back to cgrp_dfl_root which will make all
1442 * cgroups visible for lookups.
1444 return &cgrp_dfl_root.cgrp;
1448 /* look up cgroup associated with given css_set on the specified hierarchy */
1449 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1450 struct cgroup_root *root)
1452 lockdep_assert_held(&cgroup_mutex);
1453 lockdep_assert_held(&css_set_lock);
1455 return __cset_cgroup_from_root(cset, root);
1459 * Return the cgroup for "task" from the given hierarchy. Must be
1460 * called with cgroup_mutex and css_set_lock held.
1462 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1463 struct cgroup_root *root)
1466 * No need to lock the task - since we hold css_set_lock the
1467 * task can't change groups.
1469 return cset_cgroup_from_root(task_css_set(task), root);
1473 * A task must hold cgroup_mutex to modify cgroups.
1475 * Any task can increment and decrement the count field without lock.
1476 * So in general, code holding cgroup_mutex can't rely on the count
1477 * field not changing. However, if the count goes to zero, then only
1478 * cgroup_attach_task() can increment it again. Because a count of zero
1479 * means that no tasks are currently attached, therefore there is no
1480 * way a task attached to that cgroup can fork (the other way to
1481 * increment the count). So code holding cgroup_mutex can safely
1482 * assume that if the count is zero, it will stay zero. Similarly, if
1483 * a task holds cgroup_mutex on a cgroup with zero count, it
1484 * knows that the cgroup won't be removed, as cgroup_rmdir()
1487 * A cgroup can only be deleted if both its 'count' of using tasks
1488 * is zero, and its list of 'children' cgroups is empty. Since all
1489 * tasks in the system use _some_ cgroup, and since there is always at
1490 * least one task in the system (init, pid == 1), therefore, root cgroup
1491 * always has either children cgroups and/or using tasks. So we don't
1492 * need a special hack to ensure that root cgroup cannot be deleted.
1494 * P.S. One more locking exception. RCU is used to guard the
1495 * update of a tasks cgroup pointer by cgroup_attach_task()
1498 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1500 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1503 struct cgroup_subsys *ss = cft->ss;
1505 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1506 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1507 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1509 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1510 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1513 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1519 * cgroup_file_mode - deduce file mode of a control file
1520 * @cft: the control file in question
1522 * S_IRUGO for read, S_IWUSR for write.
1524 static umode_t cgroup_file_mode(const struct cftype *cft)
1528 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1531 if (cft->write_u64 || cft->write_s64 || cft->write) {
1532 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1542 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1543 * @subtree_control: the new subtree_control mask to consider
1544 * @this_ss_mask: available subsystems
1546 * On the default hierarchy, a subsystem may request other subsystems to be
1547 * enabled together through its ->depends_on mask. In such cases, more
1548 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1550 * This function calculates which subsystems need to be enabled if
1551 * @subtree_control is to be applied while restricted to @this_ss_mask.
1553 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1555 u16 cur_ss_mask = subtree_control;
1556 struct cgroup_subsys *ss;
1559 lockdep_assert_held(&cgroup_mutex);
1561 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1564 u16 new_ss_mask = cur_ss_mask;
1566 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1567 new_ss_mask |= ss->depends_on;
1568 } while_each_subsys_mask();
1571 * Mask out subsystems which aren't available. This can
1572 * happen only if some depended-upon subsystems were bound
1573 * to non-default hierarchies.
1575 new_ss_mask &= this_ss_mask;
1577 if (new_ss_mask == cur_ss_mask)
1579 cur_ss_mask = new_ss_mask;
1586 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1587 * @kn: the kernfs_node being serviced
1589 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1590 * the method finishes if locking succeeded. Note that once this function
1591 * returns the cgroup returned by cgroup_kn_lock_live() may become
1592 * inaccessible any time. If the caller intends to continue to access the
1593 * cgroup, it should pin it before invoking this function.
1595 void cgroup_kn_unlock(struct kernfs_node *kn)
1597 struct cgroup *cgrp;
1599 if (kernfs_type(kn) == KERNFS_DIR)
1602 cgrp = kn->parent->priv;
1606 kernfs_unbreak_active_protection(kn);
1611 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1612 * @kn: the kernfs_node being serviced
1613 * @drain_offline: perform offline draining on the cgroup
1615 * This helper is to be used by a cgroup kernfs method currently servicing
1616 * @kn. It breaks the active protection, performs cgroup locking and
1617 * verifies that the associated cgroup is alive. Returns the cgroup if
1618 * alive; otherwise, %NULL. A successful return should be undone by a
1619 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1620 * cgroup is drained of offlining csses before return.
1622 * Any cgroup kernfs method implementation which requires locking the
1623 * associated cgroup should use this helper. It avoids nesting cgroup
1624 * locking under kernfs active protection and allows all kernfs operations
1625 * including self-removal.
1627 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1629 struct cgroup *cgrp;
1631 if (kernfs_type(kn) == KERNFS_DIR)
1634 cgrp = kn->parent->priv;
1637 * We're gonna grab cgroup_mutex which nests outside kernfs
1638 * active_ref. cgroup liveliness check alone provides enough
1639 * protection against removal. Ensure @cgrp stays accessible and
1640 * break the active_ref protection.
1642 if (!cgroup_tryget(cgrp))
1644 kernfs_break_active_protection(kn);
1647 cgroup_lock_and_drain_offline(cgrp);
1651 if (!cgroup_is_dead(cgrp))
1654 cgroup_kn_unlock(kn);
1658 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1660 char name[CGROUP_FILE_NAME_MAX];
1662 lockdep_assert_held(&cgroup_mutex);
1664 if (cft->file_offset) {
1665 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1666 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1668 spin_lock_irq(&cgroup_file_kn_lock);
1670 spin_unlock_irq(&cgroup_file_kn_lock);
1672 del_timer_sync(&cfile->notify_timer);
1675 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1679 * css_clear_dir - remove subsys files in a cgroup directory
1682 static void css_clear_dir(struct cgroup_subsys_state *css)
1684 struct cgroup *cgrp = css->cgroup;
1685 struct cftype *cfts;
1687 if (!(css->flags & CSS_VISIBLE))
1690 css->flags &= ~CSS_VISIBLE;
1693 if (cgroup_on_dfl(cgrp)) {
1694 cgroup_addrm_files(css, cgrp,
1695 cgroup_base_files, false);
1696 if (cgroup_psi_enabled())
1697 cgroup_addrm_files(css, cgrp,
1698 cgroup_psi_files, false);
1700 cgroup_addrm_files(css, cgrp,
1701 cgroup1_base_files, false);
1704 list_for_each_entry(cfts, &css->ss->cfts, node)
1705 cgroup_addrm_files(css, cgrp, cfts, false);
1710 * css_populate_dir - create subsys files in a cgroup directory
1713 * On failure, no file is added.
1715 static int css_populate_dir(struct cgroup_subsys_state *css)
1717 struct cgroup *cgrp = css->cgroup;
1718 struct cftype *cfts, *failed_cfts;
1721 if (css->flags & CSS_VISIBLE)
1725 if (cgroup_on_dfl(cgrp)) {
1726 ret = cgroup_addrm_files(css, cgrp,
1727 cgroup_base_files, true);
1731 if (cgroup_psi_enabled()) {
1732 ret = cgroup_addrm_files(css, cgrp,
1733 cgroup_psi_files, true);
1738 ret = cgroup_addrm_files(css, cgrp,
1739 cgroup1_base_files, true);
1744 list_for_each_entry(cfts, &css->ss->cfts, node) {
1745 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1753 css->flags |= CSS_VISIBLE;
1757 list_for_each_entry(cfts, &css->ss->cfts, node) {
1758 if (cfts == failed_cfts)
1760 cgroup_addrm_files(css, cgrp, cfts, false);
1765 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1767 struct cgroup *dcgrp = &dst_root->cgrp;
1768 struct cgroup_subsys *ss;
1770 u16 dfl_disable_ss_mask = 0;
1772 lockdep_assert_held(&cgroup_mutex);
1774 do_each_subsys_mask(ss, ssid, ss_mask) {
1776 * If @ss has non-root csses attached to it, can't move.
1777 * If @ss is an implicit controller, it is exempt from this
1778 * rule and can be stolen.
1780 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1781 !ss->implicit_on_dfl)
1784 /* can't move between two non-dummy roots either */
1785 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1789 * Collect ssid's that need to be disabled from default
1792 if (ss->root == &cgrp_dfl_root)
1793 dfl_disable_ss_mask |= 1 << ssid;
1795 } while_each_subsys_mask();
1797 if (dfl_disable_ss_mask) {
1798 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1801 * Controllers from default hierarchy that need to be rebound
1802 * are all disabled together in one go.
1804 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1805 WARN_ON(cgroup_apply_control(scgrp));
1806 cgroup_finalize_control(scgrp, 0);
1809 do_each_subsys_mask(ss, ssid, ss_mask) {
1810 struct cgroup_root *src_root = ss->root;
1811 struct cgroup *scgrp = &src_root->cgrp;
1812 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1813 struct css_set *cset, *cset_pos;
1814 struct css_task_iter *it;
1816 WARN_ON(!css || cgroup_css(dcgrp, ss));
1818 if (src_root != &cgrp_dfl_root) {
1819 /* disable from the source */
1820 src_root->subsys_mask &= ~(1 << ssid);
1821 WARN_ON(cgroup_apply_control(scgrp));
1822 cgroup_finalize_control(scgrp, 0);
1826 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1827 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1828 ss->root = dst_root;
1829 css->cgroup = dcgrp;
1831 spin_lock_irq(&css_set_lock);
1832 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1833 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1834 e_cset_node[ss->id]) {
1835 list_move_tail(&cset->e_cset_node[ss->id],
1836 &dcgrp->e_csets[ss->id]);
1838 * all css_sets of scgrp together in same order to dcgrp,
1839 * patch in-flight iterators to preserve correct iteration.
1840 * since the iterator is always advanced right away and
1841 * finished when it->cset_pos meets it->cset_head, so only
1842 * update it->cset_head is enough here.
1844 list_for_each_entry(it, &cset->task_iters, iters_node)
1845 if (it->cset_head == &scgrp->e_csets[ss->id])
1846 it->cset_head = &dcgrp->e_csets[ss->id];
1848 spin_unlock_irq(&css_set_lock);
1850 if (ss->css_rstat_flush) {
1851 list_del_rcu(&css->rstat_css_node);
1853 list_add_rcu(&css->rstat_css_node,
1854 &dcgrp->rstat_css_list);
1857 /* default hierarchy doesn't enable controllers by default */
1858 dst_root->subsys_mask |= 1 << ssid;
1859 if (dst_root == &cgrp_dfl_root) {
1860 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1862 dcgrp->subtree_control |= 1 << ssid;
1863 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1866 ret = cgroup_apply_control(dcgrp);
1868 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1873 } while_each_subsys_mask();
1875 kernfs_activate(dcgrp->kn);
1879 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1880 struct kernfs_root *kf_root)
1884 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1885 struct cgroup *ns_cgroup;
1887 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1891 spin_lock_irq(&css_set_lock);
1892 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1893 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1894 spin_unlock_irq(&css_set_lock);
1896 if (len >= PATH_MAX)
1899 seq_escape(sf, buf, " \t\n\\");
1906 enum cgroup2_param {
1909 Opt_memory_localevents,
1910 Opt_memory_recursiveprot,
1911 Opt_memory_hugetlb_accounting,
1915 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1916 fsparam_flag("nsdelegate", Opt_nsdelegate),
1917 fsparam_flag("favordynmods", Opt_favordynmods),
1918 fsparam_flag("memory_localevents", Opt_memory_localevents),
1919 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1920 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting),
1924 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1926 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1927 struct fs_parse_result result;
1930 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1935 case Opt_nsdelegate:
1936 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1938 case Opt_favordynmods:
1939 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1941 case Opt_memory_localevents:
1942 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1944 case Opt_memory_recursiveprot:
1945 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1947 case Opt_memory_hugetlb_accounting:
1948 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1954 static void apply_cgroup_root_flags(unsigned int root_flags)
1956 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1957 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1958 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1960 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1962 cgroup_favor_dynmods(&cgrp_dfl_root,
1963 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1965 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1966 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1968 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1970 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1971 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1973 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1975 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
1976 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1978 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING;
1982 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1984 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1985 seq_puts(seq, ",nsdelegate");
1986 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
1987 seq_puts(seq, ",favordynmods");
1988 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1989 seq_puts(seq, ",memory_localevents");
1990 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1991 seq_puts(seq, ",memory_recursiveprot");
1992 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING)
1993 seq_puts(seq, ",memory_hugetlb_accounting");
1997 static int cgroup_reconfigure(struct fs_context *fc)
1999 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2001 apply_cgroup_root_flags(ctx->flags);
2005 static void init_cgroup_housekeeping(struct cgroup *cgrp)
2007 struct cgroup_subsys *ss;
2010 INIT_LIST_HEAD(&cgrp->self.sibling);
2011 INIT_LIST_HEAD(&cgrp->self.children);
2012 INIT_LIST_HEAD(&cgrp->cset_links);
2013 INIT_LIST_HEAD(&cgrp->pidlists);
2014 mutex_init(&cgrp->pidlist_mutex);
2015 cgrp->self.cgroup = cgrp;
2016 cgrp->self.flags |= CSS_ONLINE;
2017 cgrp->dom_cgrp = cgrp;
2018 cgrp->max_descendants = INT_MAX;
2019 cgrp->max_depth = INT_MAX;
2020 INIT_LIST_HEAD(&cgrp->rstat_css_list);
2021 prev_cputime_init(&cgrp->prev_cputime);
2023 for_each_subsys(ss, ssid)
2024 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2026 init_waitqueue_head(&cgrp->offline_waitq);
2027 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2030 void init_cgroup_root(struct cgroup_fs_context *ctx)
2032 struct cgroup_root *root = ctx->root;
2033 struct cgroup *cgrp = &root->cgrp;
2035 INIT_LIST_HEAD(&root->root_list);
2036 atomic_set(&root->nr_cgrps, 1);
2038 init_cgroup_housekeeping(cgrp);
2040 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2041 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2042 if (ctx->release_agent)
2043 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2045 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2046 if (ctx->cpuset_clone_children)
2047 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2050 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2052 LIST_HEAD(tmp_links);
2053 struct cgroup *root_cgrp = &root->cgrp;
2054 struct kernfs_syscall_ops *kf_sops;
2055 struct css_set *cset;
2058 lockdep_assert_held(&cgroup_mutex);
2060 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2066 * We're accessing css_set_count without locking css_set_lock here,
2067 * but that's OK - it can only be increased by someone holding
2068 * cgroup_lock, and that's us. Later rebinding may disable
2069 * controllers on the default hierarchy and thus create new csets,
2070 * which can't be more than the existing ones. Allocate 2x.
2072 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2076 ret = cgroup_init_root_id(root);
2080 kf_sops = root == &cgrp_dfl_root ?
2081 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2083 root->kf_root = kernfs_create_root(kf_sops,
2084 KERNFS_ROOT_CREATE_DEACTIVATED |
2085 KERNFS_ROOT_SUPPORT_EXPORTOP |
2086 KERNFS_ROOT_SUPPORT_USER_XATTR,
2088 if (IS_ERR(root->kf_root)) {
2089 ret = PTR_ERR(root->kf_root);
2092 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2093 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2094 root_cgrp->ancestors[0] = root_cgrp;
2096 ret = css_populate_dir(&root_cgrp->self);
2100 ret = cgroup_rstat_init(root_cgrp);
2104 ret = rebind_subsystems(root, ss_mask);
2108 ret = cgroup_bpf_inherit(root_cgrp);
2111 trace_cgroup_setup_root(root);
2114 * There must be no failure case after here, since rebinding takes
2115 * care of subsystems' refcounts, which are explicitly dropped in
2116 * the failure exit path.
2118 list_add(&root->root_list, &cgroup_roots);
2119 cgroup_root_count++;
2122 * Link the root cgroup in this hierarchy into all the css_set
2125 spin_lock_irq(&css_set_lock);
2126 hash_for_each(css_set_table, i, cset, hlist) {
2127 link_css_set(&tmp_links, cset, root_cgrp);
2128 if (css_set_populated(cset))
2129 cgroup_update_populated(root_cgrp, true);
2131 spin_unlock_irq(&css_set_lock);
2133 BUG_ON(!list_empty(&root_cgrp->self.children));
2134 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2140 cgroup_rstat_exit(root_cgrp);
2142 kernfs_destroy_root(root->kf_root);
2143 root->kf_root = NULL;
2145 cgroup_exit_root_id(root);
2147 percpu_ref_exit(&root_cgrp->self.refcnt);
2149 free_cgrp_cset_links(&tmp_links);
2153 int cgroup_do_get_tree(struct fs_context *fc)
2155 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2158 ctx->kfc.root = ctx->root->kf_root;
2159 if (fc->fs_type == &cgroup2_fs_type)
2160 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2162 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2163 ret = kernfs_get_tree(fc);
2166 * In non-init cgroup namespace, instead of root cgroup's dentry,
2167 * we return the dentry corresponding to the cgroupns->root_cgrp.
2169 if (!ret && ctx->ns != &init_cgroup_ns) {
2170 struct dentry *nsdentry;
2171 struct super_block *sb = fc->root->d_sb;
2172 struct cgroup *cgrp;
2175 spin_lock_irq(&css_set_lock);
2177 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2179 spin_unlock_irq(&css_set_lock);
2182 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2184 if (IS_ERR(nsdentry)) {
2185 deactivate_locked_super(sb);
2186 ret = PTR_ERR(nsdentry);
2189 fc->root = nsdentry;
2192 if (!ctx->kfc.new_sb_created)
2193 cgroup_put(&ctx->root->cgrp);
2199 * Destroy a cgroup filesystem context.
2201 static void cgroup_fs_context_free(struct fs_context *fc)
2203 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2206 kfree(ctx->release_agent);
2207 put_cgroup_ns(ctx->ns);
2208 kernfs_free_fs_context(fc);
2212 static int cgroup_get_tree(struct fs_context *fc)
2214 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2217 WRITE_ONCE(cgrp_dfl_visible, true);
2218 cgroup_get_live(&cgrp_dfl_root.cgrp);
2219 ctx->root = &cgrp_dfl_root;
2221 ret = cgroup_do_get_tree(fc);
2223 apply_cgroup_root_flags(ctx->flags);
2227 static const struct fs_context_operations cgroup_fs_context_ops = {
2228 .free = cgroup_fs_context_free,
2229 .parse_param = cgroup2_parse_param,
2230 .get_tree = cgroup_get_tree,
2231 .reconfigure = cgroup_reconfigure,
2234 static const struct fs_context_operations cgroup1_fs_context_ops = {
2235 .free = cgroup_fs_context_free,
2236 .parse_param = cgroup1_parse_param,
2237 .get_tree = cgroup1_get_tree,
2238 .reconfigure = cgroup1_reconfigure,
2242 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2243 * we select the namespace we're going to use.
2245 static int cgroup_init_fs_context(struct fs_context *fc)
2247 struct cgroup_fs_context *ctx;
2249 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2253 ctx->ns = current->nsproxy->cgroup_ns;
2254 get_cgroup_ns(ctx->ns);
2255 fc->fs_private = &ctx->kfc;
2256 if (fc->fs_type == &cgroup2_fs_type)
2257 fc->ops = &cgroup_fs_context_ops;
2259 fc->ops = &cgroup1_fs_context_ops;
2260 put_user_ns(fc->user_ns);
2261 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2264 if (have_favordynmods)
2265 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2270 static void cgroup_kill_sb(struct super_block *sb)
2272 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2273 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2276 * If @root doesn't have any children, start killing it.
2277 * This prevents new mounts by disabling percpu_ref_tryget_live().
2279 * And don't kill the default root.
2281 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2282 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2283 cgroup_bpf_offline(&root->cgrp);
2284 percpu_ref_kill(&root->cgrp.self.refcnt);
2286 cgroup_put(&root->cgrp);
2290 struct file_system_type cgroup_fs_type = {
2292 .init_fs_context = cgroup_init_fs_context,
2293 .parameters = cgroup1_fs_parameters,
2294 .kill_sb = cgroup_kill_sb,
2295 .fs_flags = FS_USERNS_MOUNT,
2298 static struct file_system_type cgroup2_fs_type = {
2300 .init_fs_context = cgroup_init_fs_context,
2301 .parameters = cgroup2_fs_parameters,
2302 .kill_sb = cgroup_kill_sb,
2303 .fs_flags = FS_USERNS_MOUNT,
2306 #ifdef CONFIG_CPUSETS
2307 static const struct fs_context_operations cpuset_fs_context_ops = {
2308 .get_tree = cgroup1_get_tree,
2309 .free = cgroup_fs_context_free,
2313 * This is ugly, but preserves the userspace API for existing cpuset
2314 * users. If someone tries to mount the "cpuset" filesystem, we
2315 * silently switch it to mount "cgroup" instead
2317 static int cpuset_init_fs_context(struct fs_context *fc)
2319 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2320 struct cgroup_fs_context *ctx;
2323 err = cgroup_init_fs_context(fc);
2329 fc->ops = &cpuset_fs_context_ops;
2331 ctx = cgroup_fc2context(fc);
2332 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2333 ctx->flags |= CGRP_ROOT_NOPREFIX;
2334 ctx->release_agent = agent;
2336 get_filesystem(&cgroup_fs_type);
2337 put_filesystem(fc->fs_type);
2338 fc->fs_type = &cgroup_fs_type;
2343 static struct file_system_type cpuset_fs_type = {
2345 .init_fs_context = cpuset_init_fs_context,
2346 .fs_flags = FS_USERNS_MOUNT,
2350 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2351 struct cgroup_namespace *ns)
2353 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2355 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2358 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2359 struct cgroup_namespace *ns)
2364 spin_lock_irq(&css_set_lock);
2366 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2368 spin_unlock_irq(&css_set_lock);
2373 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2376 * cgroup_attach_lock - Lock for ->attach()
2377 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2379 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2380 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2381 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2382 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2383 * lead to deadlocks.
2385 * Bringing up a CPU may involve creating and destroying tasks which requires
2386 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2387 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2388 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2389 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2390 * the threadgroup_rwsem to be released to create new tasks. For more details:
2392 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2394 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2395 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2396 * CPU hotplug is disabled on entry.
2398 void cgroup_attach_lock(bool lock_threadgroup)
2401 if (lock_threadgroup)
2402 percpu_down_write(&cgroup_threadgroup_rwsem);
2406 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2407 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2409 void cgroup_attach_unlock(bool lock_threadgroup)
2411 if (lock_threadgroup)
2412 percpu_up_write(&cgroup_threadgroup_rwsem);
2417 * cgroup_migrate_add_task - add a migration target task to a migration context
2418 * @task: target task
2419 * @mgctx: target migration context
2421 * Add @task, which is a migration target, to @mgctx->tset. This function
2422 * becomes noop if @task doesn't need to be migrated. @task's css_set
2423 * should have been added as a migration source and @task->cg_list will be
2424 * moved from the css_set's tasks list to mg_tasks one.
2426 static void cgroup_migrate_add_task(struct task_struct *task,
2427 struct cgroup_mgctx *mgctx)
2429 struct css_set *cset;
2431 lockdep_assert_held(&css_set_lock);
2433 /* @task either already exited or can't exit until the end */
2434 if (task->flags & PF_EXITING)
2437 /* cgroup_threadgroup_rwsem protects racing against forks */
2438 WARN_ON_ONCE(list_empty(&task->cg_list));
2440 cset = task_css_set(task);
2441 if (!cset->mg_src_cgrp)
2444 mgctx->tset.nr_tasks++;
2446 list_move_tail(&task->cg_list, &cset->mg_tasks);
2447 if (list_empty(&cset->mg_node))
2448 list_add_tail(&cset->mg_node,
2449 &mgctx->tset.src_csets);
2450 if (list_empty(&cset->mg_dst_cset->mg_node))
2451 list_add_tail(&cset->mg_dst_cset->mg_node,
2452 &mgctx->tset.dst_csets);
2456 * cgroup_taskset_first - reset taskset and return the first task
2457 * @tset: taskset of interest
2458 * @dst_cssp: output variable for the destination css
2460 * @tset iteration is initialized and the first task is returned.
2462 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2463 struct cgroup_subsys_state **dst_cssp)
2465 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2466 tset->cur_task = NULL;
2468 return cgroup_taskset_next(tset, dst_cssp);
2472 * cgroup_taskset_next - iterate to the next task in taskset
2473 * @tset: taskset of interest
2474 * @dst_cssp: output variable for the destination css
2476 * Return the next task in @tset. Iteration must have been initialized
2477 * with cgroup_taskset_first().
2479 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2480 struct cgroup_subsys_state **dst_cssp)
2482 struct css_set *cset = tset->cur_cset;
2483 struct task_struct *task = tset->cur_task;
2485 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2487 task = list_first_entry(&cset->mg_tasks,
2488 struct task_struct, cg_list);
2490 task = list_next_entry(task, cg_list);
2492 if (&task->cg_list != &cset->mg_tasks) {
2493 tset->cur_cset = cset;
2494 tset->cur_task = task;
2497 * This function may be called both before and
2498 * after cgroup_migrate_execute(). The two cases
2499 * can be distinguished by looking at whether @cset
2500 * has its ->mg_dst_cset set.
2502 if (cset->mg_dst_cset)
2503 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2505 *dst_cssp = cset->subsys[tset->ssid];
2510 cset = list_next_entry(cset, mg_node);
2518 * cgroup_migrate_execute - migrate a taskset
2519 * @mgctx: migration context
2521 * Migrate tasks in @mgctx as setup by migration preparation functions.
2522 * This function fails iff one of the ->can_attach callbacks fails and
2523 * guarantees that either all or none of the tasks in @mgctx are migrated.
2524 * @mgctx is consumed regardless of success.
2526 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2528 struct cgroup_taskset *tset = &mgctx->tset;
2529 struct cgroup_subsys *ss;
2530 struct task_struct *task, *tmp_task;
2531 struct css_set *cset, *tmp_cset;
2532 int ssid, failed_ssid, ret;
2534 /* check that we can legitimately attach to the cgroup */
2535 if (tset->nr_tasks) {
2536 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2537 if (ss->can_attach) {
2539 ret = ss->can_attach(tset);
2542 goto out_cancel_attach;
2545 } while_each_subsys_mask();
2549 * Now that we're guaranteed success, proceed to move all tasks to
2550 * the new cgroup. There are no failure cases after here, so this
2551 * is the commit point.
2553 spin_lock_irq(&css_set_lock);
2554 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2555 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2556 struct css_set *from_cset = task_css_set(task);
2557 struct css_set *to_cset = cset->mg_dst_cset;
2559 get_css_set(to_cset);
2560 to_cset->nr_tasks++;
2561 css_set_move_task(task, from_cset, to_cset, true);
2562 from_cset->nr_tasks--;
2564 * If the source or destination cgroup is frozen,
2565 * the task might require to change its state.
2567 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2569 put_css_set_locked(from_cset);
2573 spin_unlock_irq(&css_set_lock);
2576 * Migration is committed, all target tasks are now on dst_csets.
2577 * Nothing is sensitive to fork() after this point. Notify
2578 * controllers that migration is complete.
2580 tset->csets = &tset->dst_csets;
2582 if (tset->nr_tasks) {
2583 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2588 } while_each_subsys_mask();
2592 goto out_release_tset;
2595 if (tset->nr_tasks) {
2596 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2597 if (ssid == failed_ssid)
2599 if (ss->cancel_attach) {
2601 ss->cancel_attach(tset);
2603 } while_each_subsys_mask();
2606 spin_lock_irq(&css_set_lock);
2607 list_splice_init(&tset->dst_csets, &tset->src_csets);
2608 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2609 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2610 list_del_init(&cset->mg_node);
2612 spin_unlock_irq(&css_set_lock);
2615 * Re-initialize the cgroup_taskset structure in case it is reused
2616 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2620 tset->csets = &tset->src_csets;
2625 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2626 * @dst_cgrp: destination cgroup to test
2628 * On the default hierarchy, except for the mixable, (possible) thread root
2629 * and threaded cgroups, subtree_control must be zero for migration
2630 * destination cgroups with tasks so that child cgroups don't compete
2633 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2635 /* v1 doesn't have any restriction */
2636 if (!cgroup_on_dfl(dst_cgrp))
2639 /* verify @dst_cgrp can host resources */
2640 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2644 * If @dst_cgrp is already or can become a thread root or is
2645 * threaded, it doesn't matter.
2647 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2650 /* apply no-internal-process constraint */
2651 if (dst_cgrp->subtree_control)
2658 * cgroup_migrate_finish - cleanup after attach
2659 * @mgctx: migration context
2661 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2662 * those functions for details.
2664 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2666 struct css_set *cset, *tmp_cset;
2668 lockdep_assert_held(&cgroup_mutex);
2670 spin_lock_irq(&css_set_lock);
2672 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2673 mg_src_preload_node) {
2674 cset->mg_src_cgrp = NULL;
2675 cset->mg_dst_cgrp = NULL;
2676 cset->mg_dst_cset = NULL;
2677 list_del_init(&cset->mg_src_preload_node);
2678 put_css_set_locked(cset);
2681 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2682 mg_dst_preload_node) {
2683 cset->mg_src_cgrp = NULL;
2684 cset->mg_dst_cgrp = NULL;
2685 cset->mg_dst_cset = NULL;
2686 list_del_init(&cset->mg_dst_preload_node);
2687 put_css_set_locked(cset);
2690 spin_unlock_irq(&css_set_lock);
2694 * cgroup_migrate_add_src - add a migration source css_set
2695 * @src_cset: the source css_set to add
2696 * @dst_cgrp: the destination cgroup
2697 * @mgctx: migration context
2699 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2700 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2701 * up by cgroup_migrate_finish().
2703 * This function may be called without holding cgroup_threadgroup_rwsem
2704 * even if the target is a process. Threads may be created and destroyed
2705 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2706 * into play and the preloaded css_sets are guaranteed to cover all
2709 void cgroup_migrate_add_src(struct css_set *src_cset,
2710 struct cgroup *dst_cgrp,
2711 struct cgroup_mgctx *mgctx)
2713 struct cgroup *src_cgrp;
2715 lockdep_assert_held(&cgroup_mutex);
2716 lockdep_assert_held(&css_set_lock);
2719 * If ->dead, @src_set is associated with one or more dead cgroups
2720 * and doesn't contain any migratable tasks. Ignore it early so
2721 * that the rest of migration path doesn't get confused by it.
2726 if (!list_empty(&src_cset->mg_src_preload_node))
2729 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2731 WARN_ON(src_cset->mg_src_cgrp);
2732 WARN_ON(src_cset->mg_dst_cgrp);
2733 WARN_ON(!list_empty(&src_cset->mg_tasks));
2734 WARN_ON(!list_empty(&src_cset->mg_node));
2736 src_cset->mg_src_cgrp = src_cgrp;
2737 src_cset->mg_dst_cgrp = dst_cgrp;
2738 get_css_set(src_cset);
2739 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2743 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2744 * @mgctx: migration context
2746 * Tasks are about to be moved and all the source css_sets have been
2747 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2748 * pins all destination css_sets, links each to its source, and append them
2749 * to @mgctx->preloaded_dst_csets.
2751 * This function must be called after cgroup_migrate_add_src() has been
2752 * called on each migration source css_set. After migration is performed
2753 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2756 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2758 struct css_set *src_cset, *tmp_cset;
2760 lockdep_assert_held(&cgroup_mutex);
2762 /* look up the dst cset for each src cset and link it to src */
2763 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2764 mg_src_preload_node) {
2765 struct css_set *dst_cset;
2766 struct cgroup_subsys *ss;
2769 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2773 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2776 * If src cset equals dst, it's noop. Drop the src.
2777 * cgroup_migrate() will skip the cset too. Note that we
2778 * can't handle src == dst as some nodes are used by both.
2780 if (src_cset == dst_cset) {
2781 src_cset->mg_src_cgrp = NULL;
2782 src_cset->mg_dst_cgrp = NULL;
2783 list_del_init(&src_cset->mg_src_preload_node);
2784 put_css_set(src_cset);
2785 put_css_set(dst_cset);
2789 src_cset->mg_dst_cset = dst_cset;
2791 if (list_empty(&dst_cset->mg_dst_preload_node))
2792 list_add_tail(&dst_cset->mg_dst_preload_node,
2793 &mgctx->preloaded_dst_csets);
2795 put_css_set(dst_cset);
2797 for_each_subsys(ss, ssid)
2798 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2799 mgctx->ss_mask |= 1 << ssid;
2806 * cgroup_migrate - migrate a process or task to a cgroup
2807 * @leader: the leader of the process or the task to migrate
2808 * @threadgroup: whether @leader points to the whole process or a single task
2809 * @mgctx: migration context
2811 * Migrate a process or task denoted by @leader. If migrating a process,
2812 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2813 * responsible for invoking cgroup_migrate_add_src() and
2814 * cgroup_migrate_prepare_dst() on the targets before invoking this
2815 * function and following up with cgroup_migrate_finish().
2817 * As long as a controller's ->can_attach() doesn't fail, this function is
2818 * guaranteed to succeed. This means that, excluding ->can_attach()
2819 * failure, when migrating multiple targets, the success or failure can be
2820 * decided for all targets by invoking group_migrate_prepare_dst() before
2821 * actually starting migrating.
2823 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2824 struct cgroup_mgctx *mgctx)
2826 struct task_struct *task;
2829 * The following thread iteration should be inside an RCU critical
2830 * section to prevent tasks from being freed while taking the snapshot.
2831 * spin_lock_irq() implies RCU critical section here.
2833 spin_lock_irq(&css_set_lock);
2836 cgroup_migrate_add_task(task, mgctx);
2839 } while_each_thread(leader, task);
2840 spin_unlock_irq(&css_set_lock);
2842 return cgroup_migrate_execute(mgctx);
2846 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2847 * @dst_cgrp: the cgroup to attach to
2848 * @leader: the task or the leader of the threadgroup to be attached
2849 * @threadgroup: attach the whole threadgroup?
2851 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2853 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2856 DEFINE_CGROUP_MGCTX(mgctx);
2857 struct task_struct *task;
2860 /* look up all src csets */
2861 spin_lock_irq(&css_set_lock);
2865 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2868 } while_each_thread(leader, task);
2870 spin_unlock_irq(&css_set_lock);
2872 /* prepare dst csets and commit */
2873 ret = cgroup_migrate_prepare_dst(&mgctx);
2875 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2877 cgroup_migrate_finish(&mgctx);
2880 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2885 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2886 bool *threadgroup_locked)
2888 struct task_struct *tsk;
2891 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2892 return ERR_PTR(-EINVAL);
2895 * If we migrate a single thread, we don't care about threadgroup
2896 * stability. If the thread is `current`, it won't exit(2) under our
2897 * hands or change PID through exec(2). We exclude
2898 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2899 * callers by cgroup_mutex.
2900 * Therefore, we can skip the global lock.
2902 lockdep_assert_held(&cgroup_mutex);
2903 *threadgroup_locked = pid || threadgroup;
2904 cgroup_attach_lock(*threadgroup_locked);
2908 tsk = find_task_by_vpid(pid);
2910 tsk = ERR_PTR(-ESRCH);
2911 goto out_unlock_threadgroup;
2918 tsk = tsk->group_leader;
2921 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2922 * If userland migrates such a kthread to a non-root cgroup, it can
2923 * become trapped in a cpuset, or RT kthread may be born in a
2924 * cgroup with no rt_runtime allocated. Just say no.
2926 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2927 tsk = ERR_PTR(-EINVAL);
2928 goto out_unlock_threadgroup;
2931 get_task_struct(tsk);
2932 goto out_unlock_rcu;
2934 out_unlock_threadgroup:
2935 cgroup_attach_unlock(*threadgroup_locked);
2936 *threadgroup_locked = false;
2942 void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2944 struct cgroup_subsys *ss;
2947 /* release reference from cgroup_procs_write_start() */
2948 put_task_struct(task);
2950 cgroup_attach_unlock(threadgroup_locked);
2952 for_each_subsys(ss, ssid)
2953 if (ss->post_attach)
2957 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2959 struct cgroup_subsys *ss;
2960 bool printed = false;
2963 do_each_subsys_mask(ss, ssid, ss_mask) {
2966 seq_puts(seq, ss->name);
2968 } while_each_subsys_mask();
2970 seq_putc(seq, '\n');
2973 /* show controllers which are enabled from the parent */
2974 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2976 struct cgroup *cgrp = seq_css(seq)->cgroup;
2978 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2982 /* show controllers which are enabled for a given cgroup's children */
2983 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2985 struct cgroup *cgrp = seq_css(seq)->cgroup;
2987 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2992 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2993 * @cgrp: root of the subtree to update csses for
2995 * @cgrp's control masks have changed and its subtree's css associations
2996 * need to be updated accordingly. This function looks up all css_sets
2997 * which are attached to the subtree, creates the matching updated css_sets
2998 * and migrates the tasks to the new ones.
3000 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3002 DEFINE_CGROUP_MGCTX(mgctx);
3003 struct cgroup_subsys_state *d_css;
3004 struct cgroup *dsct;
3005 struct css_set *src_cset;
3009 lockdep_assert_held(&cgroup_mutex);
3011 /* look up all csses currently attached to @cgrp's subtree */
3012 spin_lock_irq(&css_set_lock);
3013 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3014 struct cgrp_cset_link *link;
3017 * As cgroup_update_dfl_csses() is only called by
3018 * cgroup_apply_control(). The csses associated with the
3019 * given cgrp will not be affected by changes made to
3020 * its subtree_control file. We can skip them.
3025 list_for_each_entry(link, &dsct->cset_links, cset_link)
3026 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3028 spin_unlock_irq(&css_set_lock);
3031 * We need to write-lock threadgroup_rwsem while migrating tasks.
3032 * However, if there are no source csets for @cgrp, changing its
3033 * controllers isn't gonna produce any task migrations and the
3034 * write-locking can be skipped safely.
3036 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3037 cgroup_attach_lock(has_tasks);
3039 /* NULL dst indicates self on default hierarchy */
3040 ret = cgroup_migrate_prepare_dst(&mgctx);
3044 spin_lock_irq(&css_set_lock);
3045 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3046 mg_src_preload_node) {
3047 struct task_struct *task, *ntask;
3049 /* all tasks in src_csets need to be migrated */
3050 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3051 cgroup_migrate_add_task(task, &mgctx);
3053 spin_unlock_irq(&css_set_lock);
3055 ret = cgroup_migrate_execute(&mgctx);
3057 cgroup_migrate_finish(&mgctx);
3058 cgroup_attach_unlock(has_tasks);
3063 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3064 * @cgrp: root of the target subtree
3066 * Because css offlining is asynchronous, userland may try to re-enable a
3067 * controller while the previous css is still around. This function grabs
3068 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3070 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3071 __acquires(&cgroup_mutex)
3073 struct cgroup *dsct;
3074 struct cgroup_subsys_state *d_css;
3075 struct cgroup_subsys *ss;
3081 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3082 for_each_subsys(ss, ssid) {
3083 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3086 if (!css || !percpu_ref_is_dying(&css->refcnt))
3089 cgroup_get_live(dsct);
3090 prepare_to_wait(&dsct->offline_waitq, &wait,
3091 TASK_UNINTERRUPTIBLE);
3095 finish_wait(&dsct->offline_waitq, &wait);
3104 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3105 * @cgrp: root of the target subtree
3107 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3108 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3111 static void cgroup_save_control(struct cgroup *cgrp)
3113 struct cgroup *dsct;
3114 struct cgroup_subsys_state *d_css;
3116 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3117 dsct->old_subtree_control = dsct->subtree_control;
3118 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3119 dsct->old_dom_cgrp = dsct->dom_cgrp;
3124 * cgroup_propagate_control - refresh control masks of a subtree
3125 * @cgrp: root of the target subtree
3127 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3128 * ->subtree_control and propagate controller availability through the
3129 * subtree so that descendants don't have unavailable controllers enabled.
3131 static void cgroup_propagate_control(struct cgroup *cgrp)
3133 struct cgroup *dsct;
3134 struct cgroup_subsys_state *d_css;
3136 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3137 dsct->subtree_control &= cgroup_control(dsct);
3138 dsct->subtree_ss_mask =
3139 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3140 cgroup_ss_mask(dsct));
3145 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3146 * @cgrp: root of the target subtree
3148 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3149 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3152 static void cgroup_restore_control(struct cgroup *cgrp)
3154 struct cgroup *dsct;
3155 struct cgroup_subsys_state *d_css;
3157 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3158 dsct->subtree_control = dsct->old_subtree_control;
3159 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3160 dsct->dom_cgrp = dsct->old_dom_cgrp;
3164 static bool css_visible(struct cgroup_subsys_state *css)
3166 struct cgroup_subsys *ss = css->ss;
3167 struct cgroup *cgrp = css->cgroup;
3169 if (cgroup_control(cgrp) & (1 << ss->id))
3171 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3173 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3177 * cgroup_apply_control_enable - enable or show csses according to control
3178 * @cgrp: root of the target subtree
3180 * Walk @cgrp's subtree and create new csses or make the existing ones
3181 * visible. A css is created invisible if it's being implicitly enabled
3182 * through dependency. An invisible css is made visible when the userland
3183 * explicitly enables it.
3185 * Returns 0 on success, -errno on failure. On failure, csses which have
3186 * been processed already aren't cleaned up. The caller is responsible for
3187 * cleaning up with cgroup_apply_control_disable().
3189 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3191 struct cgroup *dsct;
3192 struct cgroup_subsys_state *d_css;
3193 struct cgroup_subsys *ss;
3196 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3197 for_each_subsys(ss, ssid) {
3198 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3200 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3204 css = css_create(dsct, ss);
3206 return PTR_ERR(css);
3209 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3211 if (css_visible(css)) {
3212 ret = css_populate_dir(css);
3223 * cgroup_apply_control_disable - kill or hide csses according to control
3224 * @cgrp: root of the target subtree
3226 * Walk @cgrp's subtree and kill and hide csses so that they match
3227 * cgroup_ss_mask() and cgroup_visible_mask().
3229 * A css is hidden when the userland requests it to be disabled while other
3230 * subsystems are still depending on it. The css must not actively control
3231 * resources and be in the vanilla state if it's made visible again later.
3232 * Controllers which may be depended upon should provide ->css_reset() for
3235 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3237 struct cgroup *dsct;
3238 struct cgroup_subsys_state *d_css;
3239 struct cgroup_subsys *ss;
3242 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3243 for_each_subsys(ss, ssid) {
3244 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3249 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3252 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3254 } else if (!css_visible(css)) {
3264 * cgroup_apply_control - apply control mask updates to the subtree
3265 * @cgrp: root of the target subtree
3267 * subsystems can be enabled and disabled in a subtree using the following
3270 * 1. Call cgroup_save_control() to stash the current state.
3271 * 2. Update ->subtree_control masks in the subtree as desired.
3272 * 3. Call cgroup_apply_control() to apply the changes.
3273 * 4. Optionally perform other related operations.
3274 * 5. Call cgroup_finalize_control() to finish up.
3276 * This function implements step 3 and propagates the mask changes
3277 * throughout @cgrp's subtree, updates csses accordingly and perform
3278 * process migrations.
3280 static int cgroup_apply_control(struct cgroup *cgrp)
3284 cgroup_propagate_control(cgrp);
3286 ret = cgroup_apply_control_enable(cgrp);
3291 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3292 * making the following cgroup_update_dfl_csses() properly update
3293 * css associations of all tasks in the subtree.
3295 return cgroup_update_dfl_csses(cgrp);
3299 * cgroup_finalize_control - finalize control mask update
3300 * @cgrp: root of the target subtree
3301 * @ret: the result of the update
3303 * Finalize control mask update. See cgroup_apply_control() for more info.
3305 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3308 cgroup_restore_control(cgrp);
3309 cgroup_propagate_control(cgrp);
3312 cgroup_apply_control_disable(cgrp);
3315 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3317 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3319 /* if nothing is getting enabled, nothing to worry about */
3323 /* can @cgrp host any resources? */
3324 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3327 /* mixables don't care */
3328 if (cgroup_is_mixable(cgrp))
3331 if (domain_enable) {
3332 /* can't enable domain controllers inside a thread subtree */
3333 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3337 * Threaded controllers can handle internal competitions
3338 * and are always allowed inside a (prospective) thread
3341 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3346 * Controllers can't be enabled for a cgroup with tasks to avoid
3347 * child cgroups competing against tasks.
3349 if (cgroup_has_tasks(cgrp))
3355 /* change the enabled child controllers for a cgroup in the default hierarchy */
3356 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3357 char *buf, size_t nbytes,
3360 u16 enable = 0, disable = 0;
3361 struct cgroup *cgrp, *child;
3362 struct cgroup_subsys *ss;
3367 * Parse input - space separated list of subsystem names prefixed
3368 * with either + or -.
3370 buf = strstrip(buf);
3371 while ((tok = strsep(&buf, " "))) {
3374 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3375 if (!cgroup_ssid_enabled(ssid) ||
3376 strcmp(tok + 1, ss->name))
3380 enable |= 1 << ssid;
3381 disable &= ~(1 << ssid);
3382 } else if (*tok == '-') {
3383 disable |= 1 << ssid;
3384 enable &= ~(1 << ssid);
3389 } while_each_subsys_mask();
3390 if (ssid == CGROUP_SUBSYS_COUNT)
3394 cgrp = cgroup_kn_lock_live(of->kn, true);
3398 for_each_subsys(ss, ssid) {
3399 if (enable & (1 << ssid)) {
3400 if (cgrp->subtree_control & (1 << ssid)) {
3401 enable &= ~(1 << ssid);
3405 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3409 } else if (disable & (1 << ssid)) {
3410 if (!(cgrp->subtree_control & (1 << ssid))) {
3411 disable &= ~(1 << ssid);
3415 /* a child has it enabled? */
3416 cgroup_for_each_live_child(child, cgrp) {
3417 if (child->subtree_control & (1 << ssid)) {
3425 if (!enable && !disable) {
3430 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3434 /* save and update control masks and prepare csses */
3435 cgroup_save_control(cgrp);
3437 cgrp->subtree_control |= enable;
3438 cgrp->subtree_control &= ~disable;
3440 ret = cgroup_apply_control(cgrp);
3441 cgroup_finalize_control(cgrp, ret);
3445 kernfs_activate(cgrp->kn);
3447 cgroup_kn_unlock(of->kn);
3448 return ret ?: nbytes;
3452 * cgroup_enable_threaded - make @cgrp threaded
3453 * @cgrp: the target cgroup
3455 * Called when "threaded" is written to the cgroup.type interface file and
3456 * tries to make @cgrp threaded and join the parent's resource domain.
3457 * This function is never called on the root cgroup as cgroup.type doesn't
3460 static int cgroup_enable_threaded(struct cgroup *cgrp)
3462 struct cgroup *parent = cgroup_parent(cgrp);
3463 struct cgroup *dom_cgrp = parent->dom_cgrp;
3464 struct cgroup *dsct;
3465 struct cgroup_subsys_state *d_css;
3468 lockdep_assert_held(&cgroup_mutex);
3470 /* noop if already threaded */
3471 if (cgroup_is_threaded(cgrp))
3475 * If @cgroup is populated or has domain controllers enabled, it
3476 * can't be switched. While the below cgroup_can_be_thread_root()
3477 * test can catch the same conditions, that's only when @parent is
3478 * not mixable, so let's check it explicitly.
3480 if (cgroup_is_populated(cgrp) ||
3481 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3484 /* we're joining the parent's domain, ensure its validity */
3485 if (!cgroup_is_valid_domain(dom_cgrp) ||
3486 !cgroup_can_be_thread_root(dom_cgrp))
3490 * The following shouldn't cause actual migrations and should
3493 cgroup_save_control(cgrp);
3495 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3496 if (dsct == cgrp || cgroup_is_threaded(dsct))
3497 dsct->dom_cgrp = dom_cgrp;
3499 ret = cgroup_apply_control(cgrp);
3501 parent->nr_threaded_children++;
3503 cgroup_finalize_control(cgrp, ret);
3507 static int cgroup_type_show(struct seq_file *seq, void *v)
3509 struct cgroup *cgrp = seq_css(seq)->cgroup;
3511 if (cgroup_is_threaded(cgrp))
3512 seq_puts(seq, "threaded\n");
3513 else if (!cgroup_is_valid_domain(cgrp))
3514 seq_puts(seq, "domain invalid\n");
3515 else if (cgroup_is_thread_root(cgrp))
3516 seq_puts(seq, "domain threaded\n");
3518 seq_puts(seq, "domain\n");
3523 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3524 size_t nbytes, loff_t off)
3526 struct cgroup *cgrp;
3529 /* only switching to threaded mode is supported */
3530 if (strcmp(strstrip(buf), "threaded"))
3533 /* drain dying csses before we re-apply (threaded) subtree control */
3534 cgrp = cgroup_kn_lock_live(of->kn, true);
3538 /* threaded can only be enabled */
3539 ret = cgroup_enable_threaded(cgrp);
3541 cgroup_kn_unlock(of->kn);
3542 return ret ?: nbytes;
3545 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3547 struct cgroup *cgrp = seq_css(seq)->cgroup;
3548 int descendants = READ_ONCE(cgrp->max_descendants);
3550 if (descendants == INT_MAX)
3551 seq_puts(seq, "max\n");
3553 seq_printf(seq, "%d\n", descendants);
3558 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3559 char *buf, size_t nbytes, loff_t off)
3561 struct cgroup *cgrp;
3565 buf = strstrip(buf);
3566 if (!strcmp(buf, "max")) {
3567 descendants = INT_MAX;
3569 ret = kstrtoint(buf, 0, &descendants);
3574 if (descendants < 0)
3577 cgrp = cgroup_kn_lock_live(of->kn, false);
3581 cgrp->max_descendants = descendants;
3583 cgroup_kn_unlock(of->kn);
3588 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3590 struct cgroup *cgrp = seq_css(seq)->cgroup;
3591 int depth = READ_ONCE(cgrp->max_depth);
3593 if (depth == INT_MAX)
3594 seq_puts(seq, "max\n");
3596 seq_printf(seq, "%d\n", depth);
3601 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3602 char *buf, size_t nbytes, loff_t off)
3604 struct cgroup *cgrp;
3608 buf = strstrip(buf);
3609 if (!strcmp(buf, "max")) {
3612 ret = kstrtoint(buf, 0, &depth);
3620 cgrp = cgroup_kn_lock_live(of->kn, false);
3624 cgrp->max_depth = depth;
3626 cgroup_kn_unlock(of->kn);
3631 static int cgroup_events_show(struct seq_file *seq, void *v)
3633 struct cgroup *cgrp = seq_css(seq)->cgroup;
3635 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3636 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3641 static int cgroup_stat_show(struct seq_file *seq, void *v)
3643 struct cgroup *cgroup = seq_css(seq)->cgroup;
3645 seq_printf(seq, "nr_descendants %d\n",
3646 cgroup->nr_descendants);
3647 seq_printf(seq, "nr_dying_descendants %d\n",
3648 cgroup->nr_dying_descendants);
3653 #ifdef CONFIG_CGROUP_SCHED
3655 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
3656 * @cgrp: the cgroup of interest
3657 * @ss: the subsystem of interest
3659 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
3660 * or is offline, %NULL is returned.
3662 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
3663 struct cgroup_subsys *ss)
3665 struct cgroup_subsys_state *css;
3668 css = cgroup_css(cgrp, ss);
3669 if (css && !css_tryget_online(css))
3676 static int cgroup_extra_stat_show(struct seq_file *seq, int ssid)
3678 struct cgroup *cgrp = seq_css(seq)->cgroup;
3679 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3680 struct cgroup_subsys_state *css;
3683 if (!ss->css_extra_stat_show)
3686 css = cgroup_tryget_css(cgrp, ss);
3690 ret = ss->css_extra_stat_show(seq, css);
3695 static int cgroup_local_stat_show(struct seq_file *seq,
3696 struct cgroup *cgrp, int ssid)
3698 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3699 struct cgroup_subsys_state *css;
3702 if (!ss->css_local_stat_show)
3705 css = cgroup_tryget_css(cgrp, ss);
3709 ret = ss->css_local_stat_show(seq, css);
3715 static int cpu_stat_show(struct seq_file *seq, void *v)
3719 cgroup_base_stat_cputime_show(seq);
3720 #ifdef CONFIG_CGROUP_SCHED
3721 ret = cgroup_extra_stat_show(seq, cpu_cgrp_id);
3726 static int cpu_local_stat_show(struct seq_file *seq, void *v)
3728 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3731 #ifdef CONFIG_CGROUP_SCHED
3732 ret = cgroup_local_stat_show(seq, cgrp, cpu_cgrp_id);
3738 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3740 struct cgroup *cgrp = seq_css(seq)->cgroup;
3741 struct psi_group *psi = cgroup_psi(cgrp);
3743 return psi_show(seq, psi, PSI_IO);
3745 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3747 struct cgroup *cgrp = seq_css(seq)->cgroup;
3748 struct psi_group *psi = cgroup_psi(cgrp);
3750 return psi_show(seq, psi, PSI_MEM);
3752 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3754 struct cgroup *cgrp = seq_css(seq)->cgroup;
3755 struct psi_group *psi = cgroup_psi(cgrp);
3757 return psi_show(seq, psi, PSI_CPU);
3760 static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3761 size_t nbytes, enum psi_res res)
3763 struct cgroup_file_ctx *ctx = of->priv;
3764 struct psi_trigger *new;
3765 struct cgroup *cgrp;
3766 struct psi_group *psi;
3768 cgrp = cgroup_kn_lock_live(of->kn, false);
3773 cgroup_kn_unlock(of->kn);
3775 /* Allow only one trigger per file descriptor */
3776 if (ctx->psi.trigger) {
3781 psi = cgroup_psi(cgrp);
3782 new = psi_trigger_create(psi, buf, res, of->file, of);
3785 return PTR_ERR(new);
3788 smp_store_release(&ctx->psi.trigger, new);
3794 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3795 char *buf, size_t nbytes,
3798 return pressure_write(of, buf, nbytes, PSI_IO);
3801 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3802 char *buf, size_t nbytes,
3805 return pressure_write(of, buf, nbytes, PSI_MEM);
3808 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3809 char *buf, size_t nbytes,
3812 return pressure_write(of, buf, nbytes, PSI_CPU);
3815 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3816 static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3818 struct cgroup *cgrp = seq_css(seq)->cgroup;
3819 struct psi_group *psi = cgroup_psi(cgrp);
3821 return psi_show(seq, psi, PSI_IRQ);
3824 static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3825 char *buf, size_t nbytes,
3828 return pressure_write(of, buf, nbytes, PSI_IRQ);
3832 static int cgroup_pressure_show(struct seq_file *seq, void *v)
3834 struct cgroup *cgrp = seq_css(seq)->cgroup;
3835 struct psi_group *psi = cgroup_psi(cgrp);
3837 seq_printf(seq, "%d\n", psi->enabled);
3842 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3843 char *buf, size_t nbytes,
3848 struct cgroup *cgrp;
3849 struct psi_group *psi;
3851 ret = kstrtoint(strstrip(buf), 0, &enable);
3855 if (enable < 0 || enable > 1)
3858 cgrp = cgroup_kn_lock_live(of->kn, false);
3862 psi = cgroup_psi(cgrp);
3863 if (psi->enabled != enable) {
3866 /* show or hide {cpu,memory,io,irq}.pressure files */
3867 for (i = 0; i < NR_PSI_RESOURCES; i++)
3868 cgroup_file_show(&cgrp->psi_files[i], enable);
3870 psi->enabled = enable;
3872 psi_cgroup_restart(psi);
3875 cgroup_kn_unlock(of->kn);
3880 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3883 struct cgroup_file_ctx *ctx = of->priv;
3885 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3888 static int cgroup_pressure_open(struct kernfs_open_file *of)
3890 if (of->file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
3896 static void cgroup_pressure_release(struct kernfs_open_file *of)
3898 struct cgroup_file_ctx *ctx = of->priv;
3900 psi_trigger_destroy(ctx->psi.trigger);
3903 bool cgroup_psi_enabled(void)
3905 if (static_branch_likely(&psi_disabled))
3908 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3911 #else /* CONFIG_PSI */
3912 bool cgroup_psi_enabled(void)
3917 #endif /* CONFIG_PSI */
3919 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3921 struct cgroup *cgrp = seq_css(seq)->cgroup;
3923 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3928 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3929 char *buf, size_t nbytes, loff_t off)
3931 struct cgroup *cgrp;
3935 ret = kstrtoint(strstrip(buf), 0, &freeze);
3939 if (freeze < 0 || freeze > 1)
3942 cgrp = cgroup_kn_lock_live(of->kn, false);
3946 cgroup_freeze(cgrp, freeze);
3948 cgroup_kn_unlock(of->kn);
3953 static void __cgroup_kill(struct cgroup *cgrp)
3955 struct css_task_iter it;
3956 struct task_struct *task;
3958 lockdep_assert_held(&cgroup_mutex);
3960 spin_lock_irq(&css_set_lock);
3961 set_bit(CGRP_KILL, &cgrp->flags);
3962 spin_unlock_irq(&css_set_lock);
3964 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3965 while ((task = css_task_iter_next(&it))) {
3966 /* Ignore kernel threads here. */
3967 if (task->flags & PF_KTHREAD)
3970 /* Skip tasks that are already dying. */
3971 if (__fatal_signal_pending(task))
3974 send_sig(SIGKILL, task, 0);
3976 css_task_iter_end(&it);
3978 spin_lock_irq(&css_set_lock);
3979 clear_bit(CGRP_KILL, &cgrp->flags);
3980 spin_unlock_irq(&css_set_lock);
3983 static void cgroup_kill(struct cgroup *cgrp)
3985 struct cgroup_subsys_state *css;
3986 struct cgroup *dsct;
3988 lockdep_assert_held(&cgroup_mutex);
3990 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3991 __cgroup_kill(dsct);
3994 static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3995 size_t nbytes, loff_t off)
3999 struct cgroup *cgrp;
4001 ret = kstrtoint(strstrip(buf), 0, &kill);
4008 cgrp = cgroup_kn_lock_live(of->kn, false);
4013 * Killing is a process directed operation, i.e. the whole thread-group
4014 * is taken down so act like we do for cgroup.procs and only make this
4015 * writable in non-threaded cgroups.
4017 if (cgroup_is_threaded(cgrp))
4022 cgroup_kn_unlock(of->kn);
4024 return ret ?: nbytes;
4027 static int cgroup_file_open(struct kernfs_open_file *of)
4029 struct cftype *cft = of_cft(of);
4030 struct cgroup_file_ctx *ctx;
4033 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
4037 ctx->ns = current->nsproxy->cgroup_ns;
4038 get_cgroup_ns(ctx->ns);
4044 ret = cft->open(of);
4046 put_cgroup_ns(ctx->ns);
4052 static void cgroup_file_release(struct kernfs_open_file *of)
4054 struct cftype *cft = of_cft(of);
4055 struct cgroup_file_ctx *ctx = of->priv;
4059 put_cgroup_ns(ctx->ns);
4063 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4064 size_t nbytes, loff_t off)
4066 struct cgroup_file_ctx *ctx = of->priv;
4067 struct cgroup *cgrp = of->kn->parent->priv;
4068 struct cftype *cft = of_cft(of);
4069 struct cgroup_subsys_state *css;
4076 * If namespaces are delegation boundaries, disallow writes to
4077 * files in an non-init namespace root from inside the namespace
4078 * except for the files explicitly marked delegatable -
4079 * cgroup.procs and cgroup.subtree_control.
4081 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4082 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4083 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4087 return cft->write(of, buf, nbytes, off);
4090 * kernfs guarantees that a file isn't deleted with operations in
4091 * flight, which means that the matching css is and stays alive and
4092 * doesn't need to be pinned. The RCU locking is not necessary
4093 * either. It's just for the convenience of using cgroup_css().
4096 css = cgroup_css(cgrp, cft->ss);
4099 if (cft->write_u64) {
4100 unsigned long long v;
4101 ret = kstrtoull(buf, 0, &v);
4103 ret = cft->write_u64(css, cft, v);
4104 } else if (cft->write_s64) {
4106 ret = kstrtoll(buf, 0, &v);
4108 ret = cft->write_s64(css, cft, v);
4113 return ret ?: nbytes;
4116 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4118 struct cftype *cft = of_cft(of);
4121 return cft->poll(of, pt);
4123 return kernfs_generic_poll(of, pt);
4126 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4128 return seq_cft(seq)->seq_start(seq, ppos);
4131 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4133 return seq_cft(seq)->seq_next(seq, v, ppos);
4136 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4138 if (seq_cft(seq)->seq_stop)
4139 seq_cft(seq)->seq_stop(seq, v);
4142 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4144 struct cftype *cft = seq_cft(m);
4145 struct cgroup_subsys_state *css = seq_css(m);
4148 return cft->seq_show(m, arg);
4151 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4152 else if (cft->read_s64)
4153 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4159 static struct kernfs_ops cgroup_kf_single_ops = {
4160 .atomic_write_len = PAGE_SIZE,
4161 .open = cgroup_file_open,
4162 .release = cgroup_file_release,
4163 .write = cgroup_file_write,
4164 .poll = cgroup_file_poll,
4165 .seq_show = cgroup_seqfile_show,
4168 static struct kernfs_ops cgroup_kf_ops = {
4169 .atomic_write_len = PAGE_SIZE,
4170 .open = cgroup_file_open,
4171 .release = cgroup_file_release,
4172 .write = cgroup_file_write,
4173 .poll = cgroup_file_poll,
4174 .seq_start = cgroup_seqfile_start,
4175 .seq_next = cgroup_seqfile_next,
4176 .seq_stop = cgroup_seqfile_stop,
4177 .seq_show = cgroup_seqfile_show,
4180 /* set uid and gid of cgroup dirs and files to that of the creator */
4181 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
4183 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
4184 .ia_uid = current_fsuid(),
4185 .ia_gid = current_fsgid(), };
4187 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
4188 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
4191 return kernfs_setattr(kn, &iattr);
4194 static void cgroup_file_notify_timer(struct timer_list *timer)
4196 cgroup_file_notify(container_of(timer, struct cgroup_file,
4200 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4203 char name[CGROUP_FILE_NAME_MAX];
4204 struct kernfs_node *kn;
4205 struct lock_class_key *key = NULL;
4208 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4209 key = &cft->lockdep_key;
4211 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4212 cgroup_file_mode(cft),
4213 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
4214 0, cft->kf_ops, cft,
4219 ret = cgroup_kn_set_ugid(kn);
4225 if (cft->file_offset) {
4226 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4228 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4230 spin_lock_irq(&cgroup_file_kn_lock);
4232 spin_unlock_irq(&cgroup_file_kn_lock);
4239 * cgroup_addrm_files - add or remove files to a cgroup directory
4240 * @css: the target css
4241 * @cgrp: the target cgroup (usually css->cgroup)
4242 * @cfts: array of cftypes to be added
4243 * @is_add: whether to add or remove
4245 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4246 * For removals, this function never fails.
4248 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4249 struct cgroup *cgrp, struct cftype cfts[],
4252 struct cftype *cft, *cft_end = NULL;
4255 lockdep_assert_held(&cgroup_mutex);
4258 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4259 /* does cft->flags tell us to skip this file on @cgrp? */
4260 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4262 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4264 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4266 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4268 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4271 ret = cgroup_add_file(css, cgrp, cft);
4273 pr_warn("%s: failed to add %s, err=%d\n",
4274 __func__, cft->name, ret);
4280 cgroup_rm_file(cgrp, cft);
4286 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4288 struct cgroup_subsys *ss = cfts[0].ss;
4289 struct cgroup *root = &ss->root->cgrp;
4290 struct cgroup_subsys_state *css;
4293 lockdep_assert_held(&cgroup_mutex);
4295 /* add/rm files for all cgroups created before */
4296 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4297 struct cgroup *cgrp = css->cgroup;
4299 if (!(css->flags & CSS_VISIBLE))
4302 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4308 kernfs_activate(root->kn);
4312 static void cgroup_exit_cftypes(struct cftype *cfts)
4316 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4317 /* free copy for custom atomic_write_len, see init_cftypes() */
4318 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4323 /* revert flags set by cgroup core while adding @cfts */
4324 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4329 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4334 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4335 struct kernfs_ops *kf_ops;
4337 WARN_ON(cft->ss || cft->kf_ops);
4339 if (cft->flags & __CFTYPE_ADDED) {
4345 kf_ops = &cgroup_kf_ops;
4347 kf_ops = &cgroup_kf_single_ops;
4350 * Ugh... if @cft wants a custom max_write_len, we need to
4351 * make a copy of kf_ops to set its atomic_write_len.
4353 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4354 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4359 kf_ops->atomic_write_len = cft->max_write_len;
4362 cft->kf_ops = kf_ops;
4364 cft->flags |= __CFTYPE_ADDED;
4368 cgroup_exit_cftypes(cfts);
4372 static void cgroup_rm_cftypes_locked(struct cftype *cfts)
4374 lockdep_assert_held(&cgroup_mutex);
4376 list_del(&cfts->node);
4377 cgroup_apply_cftypes(cfts, false);
4378 cgroup_exit_cftypes(cfts);
4382 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4383 * @cfts: zero-length name terminated array of cftypes
4385 * Unregister @cfts. Files described by @cfts are removed from all
4386 * existing cgroups and all future cgroups won't have them either. This
4387 * function can be called anytime whether @cfts' subsys is attached or not.
4389 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4392 int cgroup_rm_cftypes(struct cftype *cfts)
4394 if (!cfts || cfts[0].name[0] == '\0')
4397 if (!(cfts[0].flags & __CFTYPE_ADDED))
4401 cgroup_rm_cftypes_locked(cfts);
4407 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4408 * @ss: target cgroup subsystem
4409 * @cfts: zero-length name terminated array of cftypes
4411 * Register @cfts to @ss. Files described by @cfts are created for all
4412 * existing cgroups to which @ss is attached and all future cgroups will
4413 * have them too. This function can be called anytime whether @ss is
4416 * Returns 0 on successful registration, -errno on failure. Note that this
4417 * function currently returns 0 as long as @cfts registration is successful
4418 * even if some file creation attempts on existing cgroups fail.
4420 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4424 if (!cgroup_ssid_enabled(ss->id))
4427 if (!cfts || cfts[0].name[0] == '\0')
4430 ret = cgroup_init_cftypes(ss, cfts);
4436 list_add_tail(&cfts->node, &ss->cfts);
4437 ret = cgroup_apply_cftypes(cfts, true);
4439 cgroup_rm_cftypes_locked(cfts);
4446 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4447 * @ss: target cgroup subsystem
4448 * @cfts: zero-length name terminated array of cftypes
4450 * Similar to cgroup_add_cftypes() but the added files are only used for
4451 * the default hierarchy.
4453 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4457 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4458 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4459 return cgroup_add_cftypes(ss, cfts);
4463 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4464 * @ss: target cgroup subsystem
4465 * @cfts: zero-length name terminated array of cftypes
4467 * Similar to cgroup_add_cftypes() but the added files are only used for
4468 * the legacy hierarchies.
4470 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4474 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4475 cft->flags |= __CFTYPE_NOT_ON_DFL;
4476 return cgroup_add_cftypes(ss, cfts);
4480 * cgroup_file_notify - generate a file modified event for a cgroup_file
4481 * @cfile: target cgroup_file
4483 * @cfile must have been obtained by setting cftype->file_offset.
4485 void cgroup_file_notify(struct cgroup_file *cfile)
4487 unsigned long flags;
4489 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4491 unsigned long last = cfile->notified_at;
4492 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4494 if (time_in_range(jiffies, last, next)) {
4495 timer_reduce(&cfile->notify_timer, next);
4497 kernfs_notify(cfile->kn);
4498 cfile->notified_at = jiffies;
4501 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4505 * cgroup_file_show - show or hide a hidden cgroup file
4506 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4507 * @show: whether to show or hide
4509 void cgroup_file_show(struct cgroup_file *cfile, bool show)
4511 struct kernfs_node *kn;
4513 spin_lock_irq(&cgroup_file_kn_lock);
4516 spin_unlock_irq(&cgroup_file_kn_lock);
4519 kernfs_show(kn, show);
4525 * css_next_child - find the next child of a given css
4526 * @pos: the current position (%NULL to initiate traversal)
4527 * @parent: css whose children to walk
4529 * This function returns the next child of @parent and should be called
4530 * under either cgroup_mutex or RCU read lock. The only requirement is
4531 * that @parent and @pos are accessible. The next sibling is guaranteed to
4532 * be returned regardless of their states.
4534 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4535 * css which finished ->css_online() is guaranteed to be visible in the
4536 * future iterations and will stay visible until the last reference is put.
4537 * A css which hasn't finished ->css_online() or already finished
4538 * ->css_offline() may show up during traversal. It's each subsystem's
4539 * responsibility to synchronize against on/offlining.
4541 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4542 struct cgroup_subsys_state *parent)
4544 struct cgroup_subsys_state *next;
4546 cgroup_assert_mutex_or_rcu_locked();
4549 * @pos could already have been unlinked from the sibling list.
4550 * Once a cgroup is removed, its ->sibling.next is no longer
4551 * updated when its next sibling changes. CSS_RELEASED is set when
4552 * @pos is taken off list, at which time its next pointer is valid,
4553 * and, as releases are serialized, the one pointed to by the next
4554 * pointer is guaranteed to not have started release yet. This
4555 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4556 * critical section, the one pointed to by its next pointer is
4557 * guaranteed to not have finished its RCU grace period even if we
4558 * have dropped rcu_read_lock() in-between iterations.
4560 * If @pos has CSS_RELEASED set, its next pointer can't be
4561 * dereferenced; however, as each css is given a monotonically
4562 * increasing unique serial number and always appended to the
4563 * sibling list, the next one can be found by walking the parent's
4564 * children until the first css with higher serial number than
4565 * @pos's. While this path can be slower, it happens iff iteration
4566 * races against release and the race window is very small.
4569 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4570 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4571 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4573 list_for_each_entry_rcu(next, &parent->children, sibling,
4574 lockdep_is_held(&cgroup_mutex))
4575 if (next->serial_nr > pos->serial_nr)
4580 * @next, if not pointing to the head, can be dereferenced and is
4583 if (&next->sibling != &parent->children)
4589 * css_next_descendant_pre - find the next descendant for pre-order walk
4590 * @pos: the current position (%NULL to initiate traversal)
4591 * @root: css whose descendants to walk
4593 * To be used by css_for_each_descendant_pre(). Find the next descendant
4594 * to visit for pre-order traversal of @root's descendants. @root is
4595 * included in the iteration and the first node to be visited.
4597 * While this function requires cgroup_mutex or RCU read locking, it
4598 * doesn't require the whole traversal to be contained in a single critical
4599 * section. This function will return the correct next descendant as long
4600 * as both @pos and @root are accessible and @pos is a descendant of @root.
4602 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4603 * css which finished ->css_online() is guaranteed to be visible in the
4604 * future iterations and will stay visible until the last reference is put.
4605 * A css which hasn't finished ->css_online() or already finished
4606 * ->css_offline() may show up during traversal. It's each subsystem's
4607 * responsibility to synchronize against on/offlining.
4609 struct cgroup_subsys_state *
4610 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4611 struct cgroup_subsys_state *root)
4613 struct cgroup_subsys_state *next;
4615 cgroup_assert_mutex_or_rcu_locked();
4617 /* if first iteration, visit @root */
4621 /* visit the first child if exists */
4622 next = css_next_child(NULL, pos);
4626 /* no child, visit my or the closest ancestor's next sibling */
4627 while (pos != root) {
4628 next = css_next_child(pos, pos->parent);
4636 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4639 * css_rightmost_descendant - return the rightmost descendant of a css
4640 * @pos: css of interest
4642 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4643 * is returned. This can be used during pre-order traversal to skip
4646 * While this function requires cgroup_mutex or RCU read locking, it
4647 * doesn't require the whole traversal to be contained in a single critical
4648 * section. This function will return the correct rightmost descendant as
4649 * long as @pos is accessible.
4651 struct cgroup_subsys_state *
4652 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4654 struct cgroup_subsys_state *last, *tmp;
4656 cgroup_assert_mutex_or_rcu_locked();
4660 /* ->prev isn't RCU safe, walk ->next till the end */
4662 css_for_each_child(tmp, last)
4669 static struct cgroup_subsys_state *
4670 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4672 struct cgroup_subsys_state *last;
4676 pos = css_next_child(NULL, pos);
4683 * css_next_descendant_post - find the next descendant for post-order walk
4684 * @pos: the current position (%NULL to initiate traversal)
4685 * @root: css whose descendants to walk
4687 * To be used by css_for_each_descendant_post(). Find the next descendant
4688 * to visit for post-order traversal of @root's descendants. @root is
4689 * included in the iteration and the last node to be visited.
4691 * While this function requires cgroup_mutex or RCU read locking, it
4692 * doesn't require the whole traversal to be contained in a single critical
4693 * section. This function will return the correct next descendant as long
4694 * as both @pos and @cgroup are accessible and @pos is a descendant of
4697 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4698 * css which finished ->css_online() is guaranteed to be visible in the
4699 * future iterations and will stay visible until the last reference is put.
4700 * A css which hasn't finished ->css_online() or already finished
4701 * ->css_offline() may show up during traversal. It's each subsystem's
4702 * responsibility to synchronize against on/offlining.
4704 struct cgroup_subsys_state *
4705 css_next_descendant_post(struct cgroup_subsys_state *pos,
4706 struct cgroup_subsys_state *root)
4708 struct cgroup_subsys_state *next;
4710 cgroup_assert_mutex_or_rcu_locked();
4712 /* if first iteration, visit leftmost descendant which may be @root */
4714 return css_leftmost_descendant(root);
4716 /* if we visited @root, we're done */
4720 /* if there's an unvisited sibling, visit its leftmost descendant */
4721 next = css_next_child(pos, pos->parent);
4723 return css_leftmost_descendant(next);
4725 /* no sibling left, visit parent */
4730 * css_has_online_children - does a css have online children
4731 * @css: the target css
4733 * Returns %true if @css has any online children; otherwise, %false. This
4734 * function can be called from any context but the caller is responsible
4735 * for synchronizing against on/offlining as necessary.
4737 bool css_has_online_children(struct cgroup_subsys_state *css)
4739 struct cgroup_subsys_state *child;
4743 css_for_each_child(child, css) {
4744 if (child->flags & CSS_ONLINE) {
4753 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4755 struct list_head *l;
4756 struct cgrp_cset_link *link;
4757 struct css_set *cset;
4759 lockdep_assert_held(&css_set_lock);
4761 /* find the next threaded cset */
4762 if (it->tcset_pos) {
4763 l = it->tcset_pos->next;
4765 if (l != it->tcset_head) {
4767 return container_of(l, struct css_set,
4768 threaded_csets_node);
4771 it->tcset_pos = NULL;
4774 /* find the next cset */
4777 if (l == it->cset_head) {
4778 it->cset_pos = NULL;
4783 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4785 link = list_entry(l, struct cgrp_cset_link, cset_link);
4791 /* initialize threaded css_set walking */
4792 if (it->flags & CSS_TASK_ITER_THREADED) {
4794 put_css_set_locked(it->cur_dcset);
4795 it->cur_dcset = cset;
4798 it->tcset_head = &cset->threaded_csets;
4799 it->tcset_pos = &cset->threaded_csets;
4806 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4807 * @it: the iterator to advance
4809 * Advance @it to the next css_set to walk.
4811 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4813 struct css_set *cset;
4815 lockdep_assert_held(&css_set_lock);
4817 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4818 while ((cset = css_task_iter_next_css_set(it))) {
4819 if (!list_empty(&cset->tasks)) {
4820 it->cur_tasks_head = &cset->tasks;
4822 } else if (!list_empty(&cset->mg_tasks)) {
4823 it->cur_tasks_head = &cset->mg_tasks;
4825 } else if (!list_empty(&cset->dying_tasks)) {
4826 it->cur_tasks_head = &cset->dying_tasks;
4831 it->task_pos = NULL;
4834 it->task_pos = it->cur_tasks_head->next;
4837 * We don't keep css_sets locked across iteration steps and thus
4838 * need to take steps to ensure that iteration can be resumed after
4839 * the lock is re-acquired. Iteration is performed at two levels -
4840 * css_sets and tasks in them.
4842 * Once created, a css_set never leaves its cgroup lists, so a
4843 * pinned css_set is guaranteed to stay put and we can resume
4844 * iteration afterwards.
4846 * Tasks may leave @cset across iteration steps. This is resolved
4847 * by registering each iterator with the css_set currently being
4848 * walked and making css_set_move_task() advance iterators whose
4849 * next task is leaving.
4852 list_del(&it->iters_node);
4853 put_css_set_locked(it->cur_cset);
4856 it->cur_cset = cset;
4857 list_add(&it->iters_node, &cset->task_iters);
4860 static void css_task_iter_skip(struct css_task_iter *it,
4861 struct task_struct *task)
4863 lockdep_assert_held(&css_set_lock);
4865 if (it->task_pos == &task->cg_list) {
4866 it->task_pos = it->task_pos->next;
4867 it->flags |= CSS_TASK_ITER_SKIPPED;
4871 static void css_task_iter_advance(struct css_task_iter *it)
4873 struct task_struct *task;
4875 lockdep_assert_held(&css_set_lock);
4879 * Advance iterator to find next entry. We go through cset
4880 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4883 if (it->flags & CSS_TASK_ITER_SKIPPED)
4884 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4886 it->task_pos = it->task_pos->next;
4888 if (it->task_pos == &it->cur_cset->tasks) {
4889 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4890 it->task_pos = it->cur_tasks_head->next;
4892 if (it->task_pos == &it->cur_cset->mg_tasks) {
4893 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4894 it->task_pos = it->cur_tasks_head->next;
4896 if (it->task_pos == &it->cur_cset->dying_tasks)
4897 css_task_iter_advance_css_set(it);
4899 /* called from start, proceed to the first cset */
4900 css_task_iter_advance_css_set(it);
4906 task = list_entry(it->task_pos, struct task_struct, cg_list);
4908 if (it->flags & CSS_TASK_ITER_PROCS) {
4909 /* if PROCS, skip over tasks which aren't group leaders */
4910 if (!thread_group_leader(task))
4913 /* and dying leaders w/o live member threads */
4914 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4915 !atomic_read(&task->signal->live))
4918 /* skip all dying ones */
4919 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4925 * css_task_iter_start - initiate task iteration
4926 * @css: the css to walk tasks of
4927 * @flags: CSS_TASK_ITER_* flags
4928 * @it: the task iterator to use
4930 * Initiate iteration through the tasks of @css. The caller can call
4931 * css_task_iter_next() to walk through the tasks until the function
4932 * returns NULL. On completion of iteration, css_task_iter_end() must be
4935 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4936 struct css_task_iter *it)
4938 unsigned long irqflags;
4940 memset(it, 0, sizeof(*it));
4942 spin_lock_irqsave(&css_set_lock, irqflags);
4947 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4948 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4950 it->cset_pos = &css->cgroup->cset_links;
4952 it->cset_head = it->cset_pos;
4954 css_task_iter_advance(it);
4956 spin_unlock_irqrestore(&css_set_lock, irqflags);
4960 * css_task_iter_next - return the next task for the iterator
4961 * @it: the task iterator being iterated
4963 * The "next" function for task iteration. @it should have been
4964 * initialized via css_task_iter_start(). Returns NULL when the iteration
4967 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4969 unsigned long irqflags;
4972 put_task_struct(it->cur_task);
4973 it->cur_task = NULL;
4976 spin_lock_irqsave(&css_set_lock, irqflags);
4978 /* @it may be half-advanced by skips, finish advancing */
4979 if (it->flags & CSS_TASK_ITER_SKIPPED)
4980 css_task_iter_advance(it);
4983 it->cur_task = list_entry(it->task_pos, struct task_struct,
4985 get_task_struct(it->cur_task);
4986 css_task_iter_advance(it);
4989 spin_unlock_irqrestore(&css_set_lock, irqflags);
4991 return it->cur_task;
4995 * css_task_iter_end - finish task iteration
4996 * @it: the task iterator to finish
4998 * Finish task iteration started by css_task_iter_start().
5000 void css_task_iter_end(struct css_task_iter *it)
5002 unsigned long irqflags;
5005 spin_lock_irqsave(&css_set_lock, irqflags);
5006 list_del(&it->iters_node);
5007 put_css_set_locked(it->cur_cset);
5008 spin_unlock_irqrestore(&css_set_lock, irqflags);
5012 put_css_set(it->cur_dcset);
5015 put_task_struct(it->cur_task);
5018 static void cgroup_procs_release(struct kernfs_open_file *of)
5020 struct cgroup_file_ctx *ctx = of->priv;
5022 if (ctx->procs.started)
5023 css_task_iter_end(&ctx->procs.iter);
5026 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
5028 struct kernfs_open_file *of = s->private;
5029 struct cgroup_file_ctx *ctx = of->priv;
5034 return css_task_iter_next(&ctx->procs.iter);
5037 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
5038 unsigned int iter_flags)
5040 struct kernfs_open_file *of = s->private;
5041 struct cgroup *cgrp = seq_css(s)->cgroup;
5042 struct cgroup_file_ctx *ctx = of->priv;
5043 struct css_task_iter *it = &ctx->procs.iter;
5046 * When a seq_file is seeked, it's always traversed sequentially
5047 * from position 0, so we can simply keep iterating on !0 *pos.
5049 if (!ctx->procs.started) {
5050 if (WARN_ON_ONCE((*pos)))
5051 return ERR_PTR(-EINVAL);
5052 css_task_iter_start(&cgrp->self, iter_flags, it);
5053 ctx->procs.started = true;
5054 } else if (!(*pos)) {
5055 css_task_iter_end(it);
5056 css_task_iter_start(&cgrp->self, iter_flags, it);
5058 return it->cur_task;
5060 return cgroup_procs_next(s, NULL, NULL);
5063 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5065 struct cgroup *cgrp = seq_css(s)->cgroup;
5068 * All processes of a threaded subtree belong to the domain cgroup
5069 * of the subtree. Only threads can be distributed across the
5070 * subtree. Reject reads on cgroup.procs in the subtree proper.
5071 * They're always empty anyway.
5073 if (cgroup_is_threaded(cgrp))
5074 return ERR_PTR(-EOPNOTSUPP);
5076 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5077 CSS_TASK_ITER_THREADED);
5080 static int cgroup_procs_show(struct seq_file *s, void *v)
5082 seq_printf(s, "%d\n", task_pid_vnr(v));
5086 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5089 struct inode *inode;
5091 lockdep_assert_held(&cgroup_mutex);
5093 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5097 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5102 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5103 struct cgroup *dst_cgrp,
5104 struct super_block *sb,
5105 struct cgroup_namespace *ns)
5107 struct cgroup *com_cgrp = src_cgrp;
5110 lockdep_assert_held(&cgroup_mutex);
5112 /* find the common ancestor */
5113 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5114 com_cgrp = cgroup_parent(com_cgrp);
5116 /* %current should be authorized to migrate to the common ancestor */
5117 ret = cgroup_may_write(com_cgrp, sb);
5122 * If namespaces are delegation boundaries, %current must be able
5123 * to see both source and destination cgroups from its namespace.
5125 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5126 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5127 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5133 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5134 struct cgroup *dst_cgrp,
5135 struct super_block *sb, bool threadgroup,
5136 struct cgroup_namespace *ns)
5140 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5144 ret = cgroup_migrate_vet_dst(dst_cgrp);
5148 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5154 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5157 struct cgroup_file_ctx *ctx = of->priv;
5158 struct cgroup *src_cgrp, *dst_cgrp;
5159 struct task_struct *task;
5160 const struct cred *saved_cred;
5162 bool threadgroup_locked;
5164 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5168 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5169 ret = PTR_ERR_OR_ZERO(task);
5173 /* find the source cgroup */
5174 spin_lock_irq(&css_set_lock);
5175 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5176 spin_unlock_irq(&css_set_lock);
5179 * Process and thread migrations follow same delegation rule. Check
5180 * permissions using the credentials from file open to protect against
5181 * inherited fd attacks.
5183 saved_cred = override_creds(of->file->f_cred);
5184 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5185 of->file->f_path.dentry->d_sb,
5186 threadgroup, ctx->ns);
5187 revert_creds(saved_cred);
5191 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5194 cgroup_procs_write_finish(task, threadgroup_locked);
5196 cgroup_kn_unlock(of->kn);
5201 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5202 char *buf, size_t nbytes, loff_t off)
5204 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5207 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5209 return __cgroup_procs_start(s, pos, 0);
5212 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5213 char *buf, size_t nbytes, loff_t off)
5215 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5218 /* cgroup core interface files for the default hierarchy */
5219 static struct cftype cgroup_base_files[] = {
5221 .name = "cgroup.type",
5222 .flags = CFTYPE_NOT_ON_ROOT,
5223 .seq_show = cgroup_type_show,
5224 .write = cgroup_type_write,
5227 .name = "cgroup.procs",
5228 .flags = CFTYPE_NS_DELEGATABLE,
5229 .file_offset = offsetof(struct cgroup, procs_file),
5230 .release = cgroup_procs_release,
5231 .seq_start = cgroup_procs_start,
5232 .seq_next = cgroup_procs_next,
5233 .seq_show = cgroup_procs_show,
5234 .write = cgroup_procs_write,
5237 .name = "cgroup.threads",
5238 .flags = CFTYPE_NS_DELEGATABLE,
5239 .release = cgroup_procs_release,
5240 .seq_start = cgroup_threads_start,
5241 .seq_next = cgroup_procs_next,
5242 .seq_show = cgroup_procs_show,
5243 .write = cgroup_threads_write,
5246 .name = "cgroup.controllers",
5247 .seq_show = cgroup_controllers_show,
5250 .name = "cgroup.subtree_control",
5251 .flags = CFTYPE_NS_DELEGATABLE,
5252 .seq_show = cgroup_subtree_control_show,
5253 .write = cgroup_subtree_control_write,
5256 .name = "cgroup.events",
5257 .flags = CFTYPE_NOT_ON_ROOT,
5258 .file_offset = offsetof(struct cgroup, events_file),
5259 .seq_show = cgroup_events_show,
5262 .name = "cgroup.max.descendants",
5263 .seq_show = cgroup_max_descendants_show,
5264 .write = cgroup_max_descendants_write,
5267 .name = "cgroup.max.depth",
5268 .seq_show = cgroup_max_depth_show,
5269 .write = cgroup_max_depth_write,
5272 .name = "cgroup.stat",
5273 .seq_show = cgroup_stat_show,
5276 .name = "cgroup.freeze",
5277 .flags = CFTYPE_NOT_ON_ROOT,
5278 .seq_show = cgroup_freeze_show,
5279 .write = cgroup_freeze_write,
5282 .name = "cgroup.kill",
5283 .flags = CFTYPE_NOT_ON_ROOT,
5284 .write = cgroup_kill_write,
5288 .seq_show = cpu_stat_show,
5291 .name = "cpu.stat.local",
5292 .seq_show = cpu_local_stat_show,
5297 static struct cftype cgroup_psi_files[] = {
5300 .name = "io.pressure",
5301 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5302 .open = cgroup_pressure_open,
5303 .seq_show = cgroup_io_pressure_show,
5304 .write = cgroup_io_pressure_write,
5305 .poll = cgroup_pressure_poll,
5306 .release = cgroup_pressure_release,
5309 .name = "memory.pressure",
5310 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5311 .open = cgroup_pressure_open,
5312 .seq_show = cgroup_memory_pressure_show,
5313 .write = cgroup_memory_pressure_write,
5314 .poll = cgroup_pressure_poll,
5315 .release = cgroup_pressure_release,
5318 .name = "cpu.pressure",
5319 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5320 .open = cgroup_pressure_open,
5321 .seq_show = cgroup_cpu_pressure_show,
5322 .write = cgroup_cpu_pressure_write,
5323 .poll = cgroup_pressure_poll,
5324 .release = cgroup_pressure_release,
5326 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
5328 .name = "irq.pressure",
5329 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5330 .open = cgroup_pressure_open,
5331 .seq_show = cgroup_irq_pressure_show,
5332 .write = cgroup_irq_pressure_write,
5333 .poll = cgroup_pressure_poll,
5334 .release = cgroup_pressure_release,
5338 .name = "cgroup.pressure",
5339 .seq_show = cgroup_pressure_show,
5340 .write = cgroup_pressure_write,
5342 #endif /* CONFIG_PSI */
5347 * css destruction is four-stage process.
5349 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5350 * Implemented in kill_css().
5352 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5353 * and thus css_tryget_online() is guaranteed to fail, the css can be
5354 * offlined by invoking offline_css(). After offlining, the base ref is
5355 * put. Implemented in css_killed_work_fn().
5357 * 3. When the percpu_ref reaches zero, the only possible remaining
5358 * accessors are inside RCU read sections. css_release() schedules the
5361 * 4. After the grace period, the css can be freed. Implemented in
5362 * css_free_rwork_fn().
5364 * It is actually hairier because both step 2 and 4 require process context
5365 * and thus involve punting to css->destroy_work adding two additional
5366 * steps to the already complex sequence.
5368 static void css_free_rwork_fn(struct work_struct *work)
5370 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5371 struct cgroup_subsys_state, destroy_rwork);
5372 struct cgroup_subsys *ss = css->ss;
5373 struct cgroup *cgrp = css->cgroup;
5375 percpu_ref_exit(&css->refcnt);
5379 struct cgroup_subsys_state *parent = css->parent;
5383 cgroup_idr_remove(&ss->css_idr, id);
5389 /* cgroup free path */
5390 atomic_dec(&cgrp->root->nr_cgrps);
5391 cgroup1_pidlist_destroy_all(cgrp);
5392 cancel_work_sync(&cgrp->release_agent_work);
5393 bpf_cgrp_storage_free(cgrp);
5395 if (cgroup_parent(cgrp)) {
5397 * We get a ref to the parent, and put the ref when
5398 * this cgroup is being freed, so it's guaranteed
5399 * that the parent won't be destroyed before its
5402 cgroup_put(cgroup_parent(cgrp));
5403 kernfs_put(cgrp->kn);
5404 psi_cgroup_free(cgrp);
5405 cgroup_rstat_exit(cgrp);
5409 * This is root cgroup's refcnt reaching zero,
5410 * which indicates that the root should be
5413 cgroup_destroy_root(cgrp->root);
5418 static void css_release_work_fn(struct work_struct *work)
5420 struct cgroup_subsys_state *css =
5421 container_of(work, struct cgroup_subsys_state, destroy_work);
5422 struct cgroup_subsys *ss = css->ss;
5423 struct cgroup *cgrp = css->cgroup;
5427 css->flags |= CSS_RELEASED;
5428 list_del_rcu(&css->sibling);
5431 /* css release path */
5432 if (!list_empty(&css->rstat_css_node)) {
5433 cgroup_rstat_flush(cgrp);
5434 list_del_rcu(&css->rstat_css_node);
5437 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5438 if (ss->css_released)
5439 ss->css_released(css);
5441 struct cgroup *tcgrp;
5443 /* cgroup release path */
5444 TRACE_CGROUP_PATH(release, cgrp);
5446 cgroup_rstat_flush(cgrp);
5448 spin_lock_irq(&css_set_lock);
5449 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5450 tcgrp = cgroup_parent(tcgrp))
5451 tcgrp->nr_dying_descendants--;
5452 spin_unlock_irq(&css_set_lock);
5455 * There are two control paths which try to determine
5456 * cgroup from dentry without going through kernfs -
5457 * cgroupstats_build() and css_tryget_online_from_dir().
5458 * Those are supported by RCU protecting clearing of
5459 * cgrp->kn->priv backpointer.
5462 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5468 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5469 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5472 static void css_release(struct percpu_ref *ref)
5474 struct cgroup_subsys_state *css =
5475 container_of(ref, struct cgroup_subsys_state, refcnt);
5477 INIT_WORK(&css->destroy_work, css_release_work_fn);
5478 queue_work(cgroup_destroy_wq, &css->destroy_work);
5481 static void init_and_link_css(struct cgroup_subsys_state *css,
5482 struct cgroup_subsys *ss, struct cgroup *cgrp)
5484 lockdep_assert_held(&cgroup_mutex);
5486 cgroup_get_live(cgrp);
5488 memset(css, 0, sizeof(*css));
5492 INIT_LIST_HEAD(&css->sibling);
5493 INIT_LIST_HEAD(&css->children);
5494 INIT_LIST_HEAD(&css->rstat_css_node);
5495 css->serial_nr = css_serial_nr_next++;
5496 atomic_set(&css->online_cnt, 0);
5498 if (cgroup_parent(cgrp)) {
5499 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5500 css_get(css->parent);
5503 if (ss->css_rstat_flush)
5504 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5506 BUG_ON(cgroup_css(cgrp, ss));
5509 /* invoke ->css_online() on a new CSS and mark it online if successful */
5510 static int online_css(struct cgroup_subsys_state *css)
5512 struct cgroup_subsys *ss = css->ss;
5515 lockdep_assert_held(&cgroup_mutex);
5518 ret = ss->css_online(css);
5520 css->flags |= CSS_ONLINE;
5521 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5523 atomic_inc(&css->online_cnt);
5525 atomic_inc(&css->parent->online_cnt);
5530 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5531 static void offline_css(struct cgroup_subsys_state *css)
5533 struct cgroup_subsys *ss = css->ss;
5535 lockdep_assert_held(&cgroup_mutex);
5537 if (!(css->flags & CSS_ONLINE))
5540 if (ss->css_offline)
5541 ss->css_offline(css);
5543 css->flags &= ~CSS_ONLINE;
5544 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5546 wake_up_all(&css->cgroup->offline_waitq);
5550 * css_create - create a cgroup_subsys_state
5551 * @cgrp: the cgroup new css will be associated with
5552 * @ss: the subsys of new css
5554 * Create a new css associated with @cgrp - @ss pair. On success, the new
5555 * css is online and installed in @cgrp. This function doesn't create the
5556 * interface files. Returns 0 on success, -errno on failure.
5558 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5559 struct cgroup_subsys *ss)
5561 struct cgroup *parent = cgroup_parent(cgrp);
5562 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5563 struct cgroup_subsys_state *css;
5566 lockdep_assert_held(&cgroup_mutex);
5568 css = ss->css_alloc(parent_css);
5570 css = ERR_PTR(-ENOMEM);
5574 init_and_link_css(css, ss, cgrp);
5576 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5580 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5585 /* @css is ready to be brought online now, make it visible */
5586 list_add_tail_rcu(&css->sibling, &parent_css->children);
5587 cgroup_idr_replace(&ss->css_idr, css, css->id);
5589 err = online_css(css);
5596 list_del_rcu(&css->sibling);
5598 list_del_rcu(&css->rstat_css_node);
5599 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5600 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5601 return ERR_PTR(err);
5605 * The returned cgroup is fully initialized including its control mask, but
5606 * it doesn't have the control mask applied.
5608 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5611 struct cgroup_root *root = parent->root;
5612 struct cgroup *cgrp, *tcgrp;
5613 struct kernfs_node *kn;
5614 int level = parent->level + 1;
5617 /* allocate the cgroup and its ID, 0 is reserved for the root */
5618 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5620 return ERR_PTR(-ENOMEM);
5622 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5626 ret = cgroup_rstat_init(cgrp);
5628 goto out_cancel_ref;
5630 /* create the directory */
5631 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5638 init_cgroup_housekeeping(cgrp);
5640 cgrp->self.parent = &parent->self;
5642 cgrp->level = level;
5644 ret = psi_cgroup_alloc(cgrp);
5646 goto out_kernfs_remove;
5648 ret = cgroup_bpf_inherit(cgrp);
5653 * New cgroup inherits effective freeze counter, and
5654 * if the parent has to be frozen, the child has too.
5656 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5657 if (cgrp->freezer.e_freeze) {
5659 * Set the CGRP_FREEZE flag, so when a process will be
5660 * attached to the child cgroup, it will become frozen.
5661 * At this point the new cgroup is unpopulated, so we can
5662 * consider it frozen immediately.
5664 set_bit(CGRP_FREEZE, &cgrp->flags);
5665 set_bit(CGRP_FROZEN, &cgrp->flags);
5668 spin_lock_irq(&css_set_lock);
5669 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5670 cgrp->ancestors[tcgrp->level] = tcgrp;
5672 if (tcgrp != cgrp) {
5673 tcgrp->nr_descendants++;
5676 * If the new cgroup is frozen, all ancestor cgroups
5677 * get a new frozen descendant, but their state can't
5678 * change because of this.
5680 if (cgrp->freezer.e_freeze)
5681 tcgrp->freezer.nr_frozen_descendants++;
5684 spin_unlock_irq(&css_set_lock);
5686 if (notify_on_release(parent))
5687 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5689 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5690 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5692 cgrp->self.serial_nr = css_serial_nr_next++;
5694 /* allocation complete, commit to creation */
5695 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5696 atomic_inc(&root->nr_cgrps);
5697 cgroup_get_live(parent);
5700 * On the default hierarchy, a child doesn't automatically inherit
5701 * subtree_control from the parent. Each is configured manually.
5703 if (!cgroup_on_dfl(cgrp))
5704 cgrp->subtree_control = cgroup_control(cgrp);
5706 cgroup_propagate_control(cgrp);
5711 psi_cgroup_free(cgrp);
5713 kernfs_remove(cgrp->kn);
5715 cgroup_rstat_exit(cgrp);
5717 percpu_ref_exit(&cgrp->self.refcnt);
5720 return ERR_PTR(ret);
5723 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5725 struct cgroup *cgroup;
5729 lockdep_assert_held(&cgroup_mutex);
5731 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5732 if (cgroup->nr_descendants >= cgroup->max_descendants)
5735 if (level > cgroup->max_depth)
5746 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5748 struct cgroup *parent, *cgrp;
5751 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5752 if (strchr(name, '\n'))
5755 parent = cgroup_kn_lock_live(parent_kn, false);
5759 if (!cgroup_check_hierarchy_limits(parent)) {
5764 cgrp = cgroup_create(parent, name, mode);
5766 ret = PTR_ERR(cgrp);
5771 * This extra ref will be put in cgroup_free_fn() and guarantees
5772 * that @cgrp->kn is always accessible.
5774 kernfs_get(cgrp->kn);
5776 ret = cgroup_kn_set_ugid(cgrp->kn);
5780 ret = css_populate_dir(&cgrp->self);
5784 ret = cgroup_apply_control_enable(cgrp);
5788 TRACE_CGROUP_PATH(mkdir, cgrp);
5790 /* let's create and online css's */
5791 kernfs_activate(cgrp->kn);
5797 cgroup_destroy_locked(cgrp);
5799 cgroup_kn_unlock(parent_kn);
5804 * This is called when the refcnt of a css is confirmed to be killed.
5805 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5806 * initiate destruction and put the css ref from kill_css().
5808 static void css_killed_work_fn(struct work_struct *work)
5810 struct cgroup_subsys_state *css =
5811 container_of(work, struct cgroup_subsys_state, destroy_work);
5818 /* @css can't go away while we're holding cgroup_mutex */
5820 } while (css && atomic_dec_and_test(&css->online_cnt));
5825 /* css kill confirmation processing requires process context, bounce */
5826 static void css_killed_ref_fn(struct percpu_ref *ref)
5828 struct cgroup_subsys_state *css =
5829 container_of(ref, struct cgroup_subsys_state, refcnt);
5831 if (atomic_dec_and_test(&css->online_cnt)) {
5832 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5833 queue_work(cgroup_destroy_wq, &css->destroy_work);
5838 * kill_css - destroy a css
5839 * @css: css to destroy
5841 * This function initiates destruction of @css by removing cgroup interface
5842 * files and putting its base reference. ->css_offline() will be invoked
5843 * asynchronously once css_tryget_online() is guaranteed to fail and when
5844 * the reference count reaches zero, @css will be released.
5846 static void kill_css(struct cgroup_subsys_state *css)
5848 lockdep_assert_held(&cgroup_mutex);
5850 if (css->flags & CSS_DYING)
5853 css->flags |= CSS_DYING;
5856 * This must happen before css is disassociated with its cgroup.
5857 * See seq_css() for details.
5862 * Killing would put the base ref, but we need to keep it alive
5863 * until after ->css_offline().
5868 * cgroup core guarantees that, by the time ->css_offline() is
5869 * invoked, no new css reference will be given out via
5870 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5871 * proceed to offlining css's because percpu_ref_kill() doesn't
5872 * guarantee that the ref is seen as killed on all CPUs on return.
5874 * Use percpu_ref_kill_and_confirm() to get notifications as each
5875 * css is confirmed to be seen as killed on all CPUs.
5877 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5881 * cgroup_destroy_locked - the first stage of cgroup destruction
5882 * @cgrp: cgroup to be destroyed
5884 * css's make use of percpu refcnts whose killing latency shouldn't be
5885 * exposed to userland and are RCU protected. Also, cgroup core needs to
5886 * guarantee that css_tryget_online() won't succeed by the time
5887 * ->css_offline() is invoked. To satisfy all the requirements,
5888 * destruction is implemented in the following two steps.
5890 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5891 * userland visible parts and start killing the percpu refcnts of
5892 * css's. Set up so that the next stage will be kicked off once all
5893 * the percpu refcnts are confirmed to be killed.
5895 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5896 * rest of destruction. Once all cgroup references are gone, the
5897 * cgroup is RCU-freed.
5899 * This function implements s1. After this step, @cgrp is gone as far as
5900 * the userland is concerned and a new cgroup with the same name may be
5901 * created. As cgroup doesn't care about the names internally, this
5902 * doesn't cause any problem.
5904 static int cgroup_destroy_locked(struct cgroup *cgrp)
5905 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5907 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5908 struct cgroup_subsys_state *css;
5909 struct cgrp_cset_link *link;
5912 lockdep_assert_held(&cgroup_mutex);
5915 * Only migration can raise populated from zero and we're already
5916 * holding cgroup_mutex.
5918 if (cgroup_is_populated(cgrp))
5922 * Make sure there's no live children. We can't test emptiness of
5923 * ->self.children as dead children linger on it while being
5924 * drained; otherwise, "rmdir parent/child parent" may fail.
5926 if (css_has_online_children(&cgrp->self))
5930 * Mark @cgrp and the associated csets dead. The former prevents
5931 * further task migration and child creation by disabling
5932 * cgroup_kn_lock_live(). The latter makes the csets ignored by
5933 * the migration path.
5935 cgrp->self.flags &= ~CSS_ONLINE;
5937 spin_lock_irq(&css_set_lock);
5938 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5939 link->cset->dead = true;
5940 spin_unlock_irq(&css_set_lock);
5942 /* initiate massacre of all css's */
5943 for_each_css(css, ssid, cgrp)
5946 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5947 css_clear_dir(&cgrp->self);
5948 kernfs_remove(cgrp->kn);
5950 if (cgroup_is_threaded(cgrp))
5951 parent->nr_threaded_children--;
5953 spin_lock_irq(&css_set_lock);
5954 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5955 tcgrp->nr_descendants--;
5956 tcgrp->nr_dying_descendants++;
5958 * If the dying cgroup is frozen, decrease frozen descendants
5959 * counters of ancestor cgroups.
5961 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5962 tcgrp->freezer.nr_frozen_descendants--;
5964 spin_unlock_irq(&css_set_lock);
5966 cgroup1_check_for_release(parent);
5968 cgroup_bpf_offline(cgrp);
5970 /* put the base reference */
5971 percpu_ref_kill(&cgrp->self.refcnt);
5976 int cgroup_rmdir(struct kernfs_node *kn)
5978 struct cgroup *cgrp;
5981 cgrp = cgroup_kn_lock_live(kn, false);
5985 ret = cgroup_destroy_locked(cgrp);
5987 TRACE_CGROUP_PATH(rmdir, cgrp);
5989 cgroup_kn_unlock(kn);
5993 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5994 .show_options = cgroup_show_options,
5995 .mkdir = cgroup_mkdir,
5996 .rmdir = cgroup_rmdir,
5997 .show_path = cgroup_show_path,
6000 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
6002 struct cgroup_subsys_state *css;
6004 pr_debug("Initializing cgroup subsys %s\n", ss->name);
6008 idr_init(&ss->css_idr);
6009 INIT_LIST_HEAD(&ss->cfts);
6011 /* Create the root cgroup state for this subsystem */
6012 ss->root = &cgrp_dfl_root;
6013 css = ss->css_alloc(NULL);
6014 /* We don't handle early failures gracefully */
6015 BUG_ON(IS_ERR(css));
6016 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
6019 * Root csses are never destroyed and we can't initialize
6020 * percpu_ref during early init. Disable refcnting.
6022 css->flags |= CSS_NO_REF;
6025 /* allocation can't be done safely during early init */
6028 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
6029 BUG_ON(css->id < 0);
6032 /* Update the init_css_set to contain a subsys
6033 * pointer to this state - since the subsystem is
6034 * newly registered, all tasks and hence the
6035 * init_css_set is in the subsystem's root cgroup. */
6036 init_css_set.subsys[ss->id] = css;
6038 have_fork_callback |= (bool)ss->fork << ss->id;
6039 have_exit_callback |= (bool)ss->exit << ss->id;
6040 have_release_callback |= (bool)ss->release << ss->id;
6041 have_canfork_callback |= (bool)ss->can_fork << ss->id;
6043 /* At system boot, before all subsystems have been
6044 * registered, no tasks have been forked, so we don't
6045 * need to invoke fork callbacks here. */
6046 BUG_ON(!list_empty(&init_task.tasks));
6048 BUG_ON(online_css(css));
6054 * cgroup_init_early - cgroup initialization at system boot
6056 * Initialize cgroups at system boot, and initialize any
6057 * subsystems that request early init.
6059 int __init cgroup_init_early(void)
6061 static struct cgroup_fs_context __initdata ctx;
6062 struct cgroup_subsys *ss;
6065 ctx.root = &cgrp_dfl_root;
6066 init_cgroup_root(&ctx);
6067 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6069 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6071 for_each_subsys(ss, i) {
6072 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6073 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6074 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6076 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6077 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6080 ss->name = cgroup_subsys_name[i];
6081 if (!ss->legacy_name)
6082 ss->legacy_name = cgroup_subsys_name[i];
6085 cgroup_init_subsys(ss, true);
6091 * cgroup_init - cgroup initialization
6093 * Register cgroup filesystem and /proc file, and initialize
6094 * any subsystems that didn't request early init.
6096 int __init cgroup_init(void)
6098 struct cgroup_subsys *ss;
6101 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6102 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6103 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6104 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6106 cgroup_rstat_boot();
6108 get_user_ns(init_cgroup_ns.user_ns);
6113 * Add init_css_set to the hash table so that dfl_root can link to
6116 hash_add(css_set_table, &init_css_set.hlist,
6117 css_set_hash(init_css_set.subsys));
6119 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6123 for_each_subsys(ss, ssid) {
6124 if (ss->early_init) {
6125 struct cgroup_subsys_state *css =
6126 init_css_set.subsys[ss->id];
6128 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6130 BUG_ON(css->id < 0);
6132 cgroup_init_subsys(ss, false);
6135 list_add_tail(&init_css_set.e_cset_node[ssid],
6136 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6139 * Setting dfl_root subsys_mask needs to consider the
6140 * disabled flag and cftype registration needs kmalloc,
6141 * both of which aren't available during early_init.
6143 if (!cgroup_ssid_enabled(ssid))
6146 if (cgroup1_ssid_disabled(ssid))
6147 pr_info("Disabling %s control group subsystem in v1 mounts\n",
6150 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6152 /* implicit controllers must be threaded too */
6153 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6155 if (ss->implicit_on_dfl)
6156 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6157 else if (!ss->dfl_cftypes)
6158 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6161 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6163 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6164 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6166 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6167 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6171 ss->bind(init_css_set.subsys[ssid]);
6174 css_populate_dir(init_css_set.subsys[ssid]);
6178 /* init_css_set.subsys[] has been updated, re-hash */
6179 hash_del(&init_css_set.hlist);
6180 hash_add(css_set_table, &init_css_set.hlist,
6181 css_set_hash(init_css_set.subsys));
6183 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6184 WARN_ON(register_filesystem(&cgroup_fs_type));
6185 WARN_ON(register_filesystem(&cgroup2_fs_type));
6186 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6187 #ifdef CONFIG_CPUSETS
6188 WARN_ON(register_filesystem(&cpuset_fs_type));
6194 static int __init cgroup_wq_init(void)
6197 * There isn't much point in executing destruction path in
6198 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6199 * Use 1 for @max_active.
6201 * We would prefer to do this in cgroup_init() above, but that
6202 * is called before init_workqueues(): so leave this until after.
6204 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6205 BUG_ON(!cgroup_destroy_wq);
6208 core_initcall(cgroup_wq_init);
6210 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6212 struct kernfs_node *kn;
6214 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6217 kernfs_path(kn, buf, buflen);
6222 * cgroup_get_from_id : get the cgroup associated with cgroup id
6224 * On success return the cgrp or ERR_PTR on failure
6225 * Only cgroups within current task's cgroup NS are valid.
6227 struct cgroup *cgroup_get_from_id(u64 id)
6229 struct kernfs_node *kn;
6230 struct cgroup *cgrp, *root_cgrp;
6232 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6234 return ERR_PTR(-ENOENT);
6236 if (kernfs_type(kn) != KERNFS_DIR) {
6238 return ERR_PTR(-ENOENT);
6243 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6244 if (cgrp && !cgroup_tryget(cgrp))
6251 return ERR_PTR(-ENOENT);
6253 root_cgrp = current_cgns_cgroup_dfl();
6254 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6256 return ERR_PTR(-ENOENT);
6261 EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6264 * proc_cgroup_show()
6265 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6266 * - Used for /proc/<pid>/cgroup.
6268 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6269 struct pid *pid, struct task_struct *tsk)
6273 struct cgroup_root *root;
6276 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6281 spin_lock_irq(&css_set_lock);
6283 for_each_root(root) {
6284 struct cgroup_subsys *ss;
6285 struct cgroup *cgrp;
6286 int ssid, count = 0;
6288 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6291 seq_printf(m, "%d:", root->hierarchy_id);
6292 if (root != &cgrp_dfl_root)
6293 for_each_subsys(ss, ssid)
6294 if (root->subsys_mask & (1 << ssid))
6295 seq_printf(m, "%s%s", count++ ? "," : "",
6297 if (strlen(root->name))
6298 seq_printf(m, "%sname=%s", count ? "," : "",
6302 cgrp = task_cgroup_from_root(tsk, root);
6305 * On traditional hierarchies, all zombie tasks show up as
6306 * belonging to the root cgroup. On the default hierarchy,
6307 * while a zombie doesn't show up in "cgroup.procs" and
6308 * thus can't be migrated, its /proc/PID/cgroup keeps
6309 * reporting the cgroup it belonged to before exiting. If
6310 * the cgroup is removed before the zombie is reaped,
6311 * " (deleted)" is appended to the cgroup path.
6313 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6314 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6315 current->nsproxy->cgroup_ns);
6316 if (retval >= PATH_MAX)
6317 retval = -ENAMETOOLONG;
6326 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6327 seq_puts(m, " (deleted)\n");
6334 spin_unlock_irq(&css_set_lock);
6342 * cgroup_fork - initialize cgroup related fields during copy_process()
6343 * @child: pointer to task_struct of forking parent process.
6345 * A task is associated with the init_css_set until cgroup_post_fork()
6346 * attaches it to the target css_set.
6348 void cgroup_fork(struct task_struct *child)
6350 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6351 INIT_LIST_HEAD(&child->cg_list);
6355 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6356 * @f: file corresponding to cgroup_dir
6358 * Find the cgroup from a file pointer associated with a cgroup directory.
6359 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6360 * cgroup cannot be found.
6362 static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6364 struct cgroup_subsys_state *css;
6366 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6368 return ERR_CAST(css);
6374 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6376 * @f: file corresponding to cgroup2_dir
6378 static struct cgroup *cgroup_get_from_file(struct file *f)
6380 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6383 return ERR_CAST(cgrp);
6385 if (!cgroup_on_dfl(cgrp)) {
6387 return ERR_PTR(-EBADF);
6394 * cgroup_css_set_fork - find or create a css_set for a child process
6395 * @kargs: the arguments passed to create the child process
6397 * This functions finds or creates a new css_set which the child
6398 * process will be attached to in cgroup_post_fork(). By default,
6399 * the child process will be given the same css_set as its parent.
6401 * If CLONE_INTO_CGROUP is specified this function will try to find an
6402 * existing css_set which includes the requested cgroup and if not create
6403 * a new css_set that the child will be attached to later. If this function
6404 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6405 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6406 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6407 * to the target cgroup.
6409 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6410 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6413 struct cgroup *dst_cgrp = NULL;
6414 struct css_set *cset;
6415 struct super_block *sb;
6418 if (kargs->flags & CLONE_INTO_CGROUP)
6421 cgroup_threadgroup_change_begin(current);
6423 spin_lock_irq(&css_set_lock);
6424 cset = task_css_set(current);
6426 spin_unlock_irq(&css_set_lock);
6428 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6433 f = fget_raw(kargs->cgroup);
6438 sb = f->f_path.dentry->d_sb;
6440 dst_cgrp = cgroup_get_from_file(f);
6441 if (IS_ERR(dst_cgrp)) {
6442 ret = PTR_ERR(dst_cgrp);
6447 if (cgroup_is_dead(dst_cgrp)) {
6453 * Verify that we the target cgroup is writable for us. This is
6454 * usually done by the vfs layer but since we're not going through
6455 * the vfs layer here we need to do it "manually".
6457 ret = cgroup_may_write(dst_cgrp, sb);
6462 * Spawning a task directly into a cgroup works by passing a file
6463 * descriptor to the target cgroup directory. This can even be an O_PATH
6464 * file descriptor. But it can never be a cgroup.procs file descriptor.
6465 * This was done on purpose so spawning into a cgroup could be
6466 * conceptualized as an atomic
6468 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6469 * write(fd, <child-pid>, ...);
6471 * sequence, i.e. it's a shorthand for the caller opening and writing
6472 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6473 * to always use the caller's credentials.
6475 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6476 !(kargs->flags & CLONE_THREAD),
6477 current->nsproxy->cgroup_ns);
6481 kargs->cset = find_css_set(cset, dst_cgrp);
6489 kargs->cgrp = dst_cgrp;
6493 cgroup_threadgroup_change_end(current);
6498 cgroup_put(dst_cgrp);
6501 put_css_set(kargs->cset);
6506 * cgroup_css_set_put_fork - drop references we took during fork
6507 * @kargs: the arguments passed to create the child process
6509 * Drop references to the prepared css_set and target cgroup if
6510 * CLONE_INTO_CGROUP was requested.
6512 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6513 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6515 struct cgroup *cgrp = kargs->cgrp;
6516 struct css_set *cset = kargs->cset;
6518 cgroup_threadgroup_change_end(current);
6525 if (kargs->flags & CLONE_INTO_CGROUP) {
6535 * cgroup_can_fork - called on a new task before the process is exposed
6536 * @child: the child process
6537 * @kargs: the arguments passed to create the child process
6539 * This prepares a new css_set for the child process which the child will
6540 * be attached to in cgroup_post_fork().
6541 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6542 * callback returns an error, the fork aborts with that error code. This
6543 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6545 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6547 struct cgroup_subsys *ss;
6550 ret = cgroup_css_set_fork(kargs);
6554 do_each_subsys_mask(ss, i, have_canfork_callback) {
6555 ret = ss->can_fork(child, kargs->cset);
6558 } while_each_subsys_mask();
6563 for_each_subsys(ss, j) {
6566 if (ss->cancel_fork)
6567 ss->cancel_fork(child, kargs->cset);
6570 cgroup_css_set_put_fork(kargs);
6576 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6577 * @child: the child process
6578 * @kargs: the arguments passed to create the child process
6580 * This calls the cancel_fork() callbacks if a fork failed *after*
6581 * cgroup_can_fork() succeeded and cleans up references we took to
6582 * prepare a new css_set for the child process in cgroup_can_fork().
6584 void cgroup_cancel_fork(struct task_struct *child,
6585 struct kernel_clone_args *kargs)
6587 struct cgroup_subsys *ss;
6590 for_each_subsys(ss, i)
6591 if (ss->cancel_fork)
6592 ss->cancel_fork(child, kargs->cset);
6594 cgroup_css_set_put_fork(kargs);
6598 * cgroup_post_fork - finalize cgroup setup for the child process
6599 * @child: the child process
6600 * @kargs: the arguments passed to create the child process
6602 * Attach the child process to its css_set calling the subsystem fork()
6605 void cgroup_post_fork(struct task_struct *child,
6606 struct kernel_clone_args *kargs)
6607 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6609 unsigned long cgrp_flags = 0;
6611 struct cgroup_subsys *ss;
6612 struct css_set *cset;
6618 spin_lock_irq(&css_set_lock);
6620 /* init tasks are special, only link regular threads */
6621 if (likely(child->pid)) {
6623 cgrp_flags = kargs->cgrp->flags;
6625 cgrp_flags = cset->dfl_cgrp->flags;
6627 WARN_ON_ONCE(!list_empty(&child->cg_list));
6629 css_set_move_task(child, NULL, cset, false);
6635 if (!(child->flags & PF_KTHREAD)) {
6636 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6638 * If the cgroup has to be frozen, the new task has
6639 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6640 * get the task into the frozen state.
6642 spin_lock(&child->sighand->siglock);
6643 WARN_ON_ONCE(child->frozen);
6644 child->jobctl |= JOBCTL_TRAP_FREEZE;
6645 spin_unlock(&child->sighand->siglock);
6648 * Calling cgroup_update_frozen() isn't required here,
6649 * because it will be called anyway a bit later from
6650 * do_freezer_trap(). So we avoid cgroup's transient
6651 * switch from the frozen state and back.
6656 * If the cgroup is to be killed notice it now and take the
6657 * child down right after we finished preparing it for
6660 kill = test_bit(CGRP_KILL, &cgrp_flags);
6663 spin_unlock_irq(&css_set_lock);
6666 * Call ss->fork(). This must happen after @child is linked on
6667 * css_set; otherwise, @child might change state between ->fork()
6668 * and addition to css_set.
6670 do_each_subsys_mask(ss, i, have_fork_callback) {
6672 } while_each_subsys_mask();
6674 /* Make the new cset the root_cset of the new cgroup namespace. */
6675 if (kargs->flags & CLONE_NEWCGROUP) {
6676 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6679 child->nsproxy->cgroup_ns->root_cset = cset;
6683 /* Cgroup has to be killed so take down child immediately. */
6685 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6687 cgroup_css_set_put_fork(kargs);
6691 * cgroup_exit - detach cgroup from exiting task
6692 * @tsk: pointer to task_struct of exiting process
6694 * Description: Detach cgroup from @tsk.
6697 void cgroup_exit(struct task_struct *tsk)
6699 struct cgroup_subsys *ss;
6700 struct css_set *cset;
6703 spin_lock_irq(&css_set_lock);
6705 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6706 cset = task_css_set(tsk);
6707 css_set_move_task(tsk, cset, NULL, false);
6708 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6712 dec_dl_tasks_cs(tsk);
6714 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6715 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6716 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6717 cgroup_update_frozen(task_dfl_cgroup(tsk));
6719 spin_unlock_irq(&css_set_lock);
6721 /* see cgroup_post_fork() for details */
6722 do_each_subsys_mask(ss, i, have_exit_callback) {
6724 } while_each_subsys_mask();
6727 void cgroup_release(struct task_struct *task)
6729 struct cgroup_subsys *ss;
6732 do_each_subsys_mask(ss, ssid, have_release_callback) {
6734 } while_each_subsys_mask();
6736 spin_lock_irq(&css_set_lock);
6737 css_set_skip_task_iters(task_css_set(task), task);
6738 list_del_init(&task->cg_list);
6739 spin_unlock_irq(&css_set_lock);
6742 void cgroup_free(struct task_struct *task)
6744 struct css_set *cset = task_css_set(task);
6748 static int __init cgroup_disable(char *str)
6750 struct cgroup_subsys *ss;
6754 while ((token = strsep(&str, ",")) != NULL) {
6758 for_each_subsys(ss, i) {
6759 if (strcmp(token, ss->name) &&
6760 strcmp(token, ss->legacy_name))
6763 static_branch_disable(cgroup_subsys_enabled_key[i]);
6764 pr_info("Disabling %s control group subsystem\n",
6768 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6769 if (strcmp(token, cgroup_opt_feature_names[i]))
6771 cgroup_feature_disable_mask |= 1 << i;
6772 pr_info("Disabling %s control group feature\n",
6773 cgroup_opt_feature_names[i]);
6779 __setup("cgroup_disable=", cgroup_disable);
6781 void __init __weak enable_debug_cgroup(void) { }
6783 static int __init enable_cgroup_debug(char *str)
6785 cgroup_debug = true;
6786 enable_debug_cgroup();
6789 __setup("cgroup_debug", enable_cgroup_debug);
6791 static int __init cgroup_favordynmods_setup(char *str)
6793 return (kstrtobool(str, &have_favordynmods) == 0);
6795 __setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
6798 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6799 * @dentry: directory dentry of interest
6800 * @ss: subsystem of interest
6802 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6803 * to get the corresponding css and return it. If such css doesn't exist
6804 * or can't be pinned, an ERR_PTR value is returned.
6806 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6807 struct cgroup_subsys *ss)
6809 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6810 struct file_system_type *s_type = dentry->d_sb->s_type;
6811 struct cgroup_subsys_state *css = NULL;
6812 struct cgroup *cgrp;
6814 /* is @dentry a cgroup dir? */
6815 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6816 !kn || kernfs_type(kn) != KERNFS_DIR)
6817 return ERR_PTR(-EBADF);
6822 * This path doesn't originate from kernfs and @kn could already
6823 * have been or be removed at any point. @kn->priv is RCU
6824 * protected for this access. See css_release_work_fn() for details.
6826 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6828 css = cgroup_css(cgrp, ss);
6830 if (!css || !css_tryget_online(css))
6831 css = ERR_PTR(-ENOENT);
6838 * css_from_id - lookup css by id
6839 * @id: the cgroup id
6840 * @ss: cgroup subsys to be looked into
6842 * Returns the css if there's valid one with @id, otherwise returns NULL.
6843 * Should be called under rcu_read_lock().
6845 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6847 WARN_ON_ONCE(!rcu_read_lock_held());
6848 return idr_find(&ss->css_idr, id);
6852 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6853 * @path: path on the default hierarchy
6855 * Find the cgroup at @path on the default hierarchy, increment its
6856 * reference count and return it. Returns pointer to the found cgroup on
6857 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6858 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6860 struct cgroup *cgroup_get_from_path(const char *path)
6862 struct kernfs_node *kn;
6863 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6864 struct cgroup *root_cgrp;
6866 root_cgrp = current_cgns_cgroup_dfl();
6867 kn = kernfs_walk_and_get(root_cgrp->kn, path);
6871 if (kernfs_type(kn) != KERNFS_DIR) {
6872 cgrp = ERR_PTR(-ENOTDIR);
6878 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6879 if (!cgrp || !cgroup_tryget(cgrp))
6880 cgrp = ERR_PTR(-ENOENT);
6889 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6892 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6893 * @fd: fd obtained by open(cgroup_dir)
6895 * Find the cgroup from a fd which should be obtained
6896 * by opening a cgroup directory. Returns a pointer to the
6897 * cgroup on success. ERR_PTR is returned if the cgroup
6900 struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6902 struct cgroup *cgrp;
6903 struct fd f = fdget_raw(fd);
6905 return ERR_PTR(-EBADF);
6907 cgrp = cgroup_v1v2_get_from_file(f.file);
6913 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6915 * @fd: fd obtained by open(cgroup2_dir)
6917 struct cgroup *cgroup_get_from_fd(int fd)
6919 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6922 return ERR_CAST(cgrp);
6924 if (!cgroup_on_dfl(cgrp)) {
6926 return ERR_PTR(-EBADF);
6930 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6932 static u64 power_of_ten(int power)
6941 * cgroup_parse_float - parse a floating number
6942 * @input: input string
6943 * @dec_shift: number of decimal digits to shift
6946 * Parse a decimal floating point number in @input and store the result in
6947 * @v with decimal point right shifted @dec_shift times. For example, if
6948 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6949 * Returns 0 on success, -errno otherwise.
6951 * There's nothing cgroup specific about this function except that it's
6952 * currently the only user.
6954 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6956 s64 whole, frac = 0;
6957 int fstart = 0, fend = 0, flen;
6959 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6964 flen = fend > fstart ? fend - fstart : 0;
6965 if (flen < dec_shift)
6966 frac *= power_of_ten(dec_shift - flen);
6968 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6970 *v = whole * power_of_ten(dec_shift) + frac;
6975 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6976 * definition in cgroup-defs.h.
6978 #ifdef CONFIG_SOCK_CGROUP_DATA
6980 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6982 struct cgroup *cgroup;
6985 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6986 if (in_interrupt()) {
6987 cgroup = &cgrp_dfl_root.cgrp;
6993 struct css_set *cset;
6995 cset = task_css_set(current);
6996 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6997 cgroup = cset->dfl_cgrp;
7003 skcd->cgroup = cgroup;
7004 cgroup_bpf_get(cgroup);
7008 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
7010 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7013 * We might be cloning a socket which is left in an empty
7014 * cgroup and the cgroup might have already been rmdir'd.
7015 * Don't use cgroup_get_live().
7018 cgroup_bpf_get(cgrp);
7021 void cgroup_sk_free(struct sock_cgroup_data *skcd)
7023 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
7025 cgroup_bpf_put(cgrp);
7029 #endif /* CONFIG_SOCK_CGROUP_DATA */
7032 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
7033 ssize_t size, const char *prefix)
7038 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
7039 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
7043 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
7045 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
7047 if (WARN_ON(ret >= size))
7054 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
7057 struct cgroup_subsys *ss;
7061 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7062 PAGE_SIZE - ret, NULL);
7063 if (cgroup_psi_enabled())
7064 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7065 PAGE_SIZE - ret, NULL);
7067 for_each_subsys(ss, ssid)
7068 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7070 cgroup_subsys_name[ssid]);
7074 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7076 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7079 return snprintf(buf, PAGE_SIZE,
7082 "memory_localevents\n"
7083 "memory_recursiveprot\n"
7084 "memory_hugetlb_accounting\n");
7086 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7088 static struct attribute *cgroup_sysfs_attrs[] = {
7089 &cgroup_delegate_attr.attr,
7090 &cgroup_features_attr.attr,
7094 static const struct attribute_group cgroup_sysfs_attr_group = {
7095 .attrs = cgroup_sysfs_attrs,
7099 static int __init cgroup_sysfs_init(void)
7101 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7103 subsys_initcall(cgroup_sysfs_init);
7105 #endif /* CONFIG_SYSFS */