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 /* cgroup namespace for init task */
211 struct cgroup_namespace init_cgroup_ns = {
212 .ns.count = REFCOUNT_INIT(2),
213 .user_ns = &init_user_ns,
214 .ns.ops = &cgroupns_operations,
215 .ns.inum = PROC_CGROUP_INIT_INO,
216 .root_cset = &init_css_set,
219 static struct file_system_type cgroup2_fs_type;
220 static struct cftype cgroup_base_files[];
221 static struct cftype cgroup_psi_files[];
223 /* cgroup optional features */
224 enum cgroup_opt_features {
226 OPT_FEATURE_PRESSURE,
231 static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
237 static u16 cgroup_feature_disable_mask __read_mostly;
239 static int cgroup_apply_control(struct cgroup *cgrp);
240 static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
241 static void css_task_iter_skip(struct css_task_iter *it,
242 struct task_struct *task);
243 static int cgroup_destroy_locked(struct cgroup *cgrp);
244 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
245 struct cgroup_subsys *ss);
246 static void css_release(struct percpu_ref *ref);
247 static void kill_css(struct cgroup_subsys_state *css);
248 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
249 struct cgroup *cgrp, struct cftype cfts[],
252 #ifdef CONFIG_DEBUG_CGROUP_REF
253 #define CGROUP_REF_FN_ATTRS noinline
254 #define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn);
255 #include <linux/cgroup_refcnt.h>
259 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
260 * @ssid: subsys ID of interest
262 * cgroup_subsys_enabled() can only be used with literal subsys names which
263 * is fine for individual subsystems but unsuitable for cgroup core. This
264 * is slower static_key_enabled() based test indexed by @ssid.
266 bool cgroup_ssid_enabled(int ssid)
268 if (!CGROUP_HAS_SUBSYS_CONFIG)
271 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
275 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
276 * @cgrp: the cgroup of interest
278 * The default hierarchy is the v2 interface of cgroup and this function
279 * can be used to test whether a cgroup is on the default hierarchy for
280 * cases where a subsystem should behave differently depending on the
283 * List of changed behaviors:
285 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
286 * and "name" are disallowed.
288 * - When mounting an existing superblock, mount options should match.
290 * - rename(2) is disallowed.
292 * - "tasks" is removed. Everything should be at process granularity. Use
293 * "cgroup.procs" instead.
295 * - "cgroup.procs" is not sorted. pids will be unique unless they got
296 * recycled in-between reads.
298 * - "release_agent" and "notify_on_release" are removed. Replacement
299 * notification mechanism will be implemented.
301 * - "cgroup.clone_children" is removed.
303 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
304 * and its descendants contain no task; otherwise, 1. The file also
305 * generates kernfs notification which can be monitored through poll and
306 * [di]notify when the value of the file changes.
308 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
309 * take masks of ancestors with non-empty cpus/mems, instead of being
310 * moved to an ancestor.
312 * - cpuset: a task can be moved into an empty cpuset, and again it takes
313 * masks of ancestors.
315 * - blkcg: blk-throttle becomes properly hierarchical.
317 bool cgroup_on_dfl(const struct cgroup *cgrp)
319 return cgrp->root == &cgrp_dfl_root;
322 /* IDR wrappers which synchronize using cgroup_idr_lock */
323 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
328 idr_preload(gfp_mask);
329 spin_lock_bh(&cgroup_idr_lock);
330 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
331 spin_unlock_bh(&cgroup_idr_lock);
336 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
340 spin_lock_bh(&cgroup_idr_lock);
341 ret = idr_replace(idr, ptr, id);
342 spin_unlock_bh(&cgroup_idr_lock);
346 static void cgroup_idr_remove(struct idr *idr, int id)
348 spin_lock_bh(&cgroup_idr_lock);
350 spin_unlock_bh(&cgroup_idr_lock);
353 static bool cgroup_has_tasks(struct cgroup *cgrp)
355 return cgrp->nr_populated_csets;
358 static bool cgroup_is_threaded(struct cgroup *cgrp)
360 return cgrp->dom_cgrp != cgrp;
363 /* can @cgrp host both domain and threaded children? */
364 static bool cgroup_is_mixable(struct cgroup *cgrp)
367 * Root isn't under domain level resource control exempting it from
368 * the no-internal-process constraint, so it can serve as a thread
369 * root and a parent of resource domains at the same time.
371 return !cgroup_parent(cgrp);
374 /* can @cgrp become a thread root? Should always be true for a thread root */
375 static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
377 /* mixables don't care */
378 if (cgroup_is_mixable(cgrp))
381 /* domain roots can't be nested under threaded */
382 if (cgroup_is_threaded(cgrp))
385 /* can only have either domain or threaded children */
386 if (cgrp->nr_populated_domain_children)
389 /* and no domain controllers can be enabled */
390 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
396 /* is @cgrp root of a threaded subtree? */
397 static bool cgroup_is_thread_root(struct cgroup *cgrp)
399 /* thread root should be a domain */
400 if (cgroup_is_threaded(cgrp))
403 /* a domain w/ threaded children is a thread root */
404 if (cgrp->nr_threaded_children)
408 * A domain which has tasks and explicit threaded controllers
409 * enabled is a thread root.
411 if (cgroup_has_tasks(cgrp) &&
412 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
418 /* a domain which isn't connected to the root w/o brekage can't be used */
419 static bool cgroup_is_valid_domain(struct cgroup *cgrp)
421 /* the cgroup itself can be a thread root */
422 if (cgroup_is_threaded(cgrp))
425 /* but the ancestors can't be unless mixable */
426 while ((cgrp = cgroup_parent(cgrp))) {
427 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
429 if (cgroup_is_threaded(cgrp))
436 /* subsystems visibly enabled on a cgroup */
437 static u16 cgroup_control(struct cgroup *cgrp)
439 struct cgroup *parent = cgroup_parent(cgrp);
440 u16 root_ss_mask = cgrp->root->subsys_mask;
443 u16 ss_mask = parent->subtree_control;
445 /* threaded cgroups can only have threaded controllers */
446 if (cgroup_is_threaded(cgrp))
447 ss_mask &= cgrp_dfl_threaded_ss_mask;
451 if (cgroup_on_dfl(cgrp))
452 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
453 cgrp_dfl_implicit_ss_mask);
457 /* subsystems enabled on a cgroup */
458 static u16 cgroup_ss_mask(struct cgroup *cgrp)
460 struct cgroup *parent = cgroup_parent(cgrp);
463 u16 ss_mask = parent->subtree_ss_mask;
465 /* threaded cgroups can only have threaded controllers */
466 if (cgroup_is_threaded(cgrp))
467 ss_mask &= cgrp_dfl_threaded_ss_mask;
471 return cgrp->root->subsys_mask;
475 * cgroup_css - obtain a cgroup's css for the specified subsystem
476 * @cgrp: the cgroup of interest
477 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
479 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
480 * function must be called either under cgroup_mutex or rcu_read_lock() and
481 * the caller is responsible for pinning the returned css if it wants to
482 * keep accessing it outside the said locks. This function may return
483 * %NULL if @cgrp doesn't have @subsys_id enabled.
485 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
486 struct cgroup_subsys *ss)
488 if (CGROUP_HAS_SUBSYS_CONFIG && ss)
489 return rcu_dereference_check(cgrp->subsys[ss->id],
490 lockdep_is_held(&cgroup_mutex));
496 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
497 * @cgrp: the cgroup of interest
498 * @ss: the subsystem of interest
500 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
501 * or is offline, %NULL is returned.
503 static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
504 struct cgroup_subsys *ss)
506 struct cgroup_subsys_state *css;
509 css = cgroup_css(cgrp, ss);
510 if (css && !css_tryget_online(css))
518 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
519 * @cgrp: the cgroup of interest
520 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
522 * Similar to cgroup_css() but returns the effective css, which is defined
523 * as the matching css of the nearest ancestor including self which has @ss
524 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
525 * function is guaranteed to return non-NULL css.
527 static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
528 struct cgroup_subsys *ss)
530 lockdep_assert_held(&cgroup_mutex);
536 * This function is used while updating css associations and thus
537 * can't test the csses directly. Test ss_mask.
539 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
540 cgrp = cgroup_parent(cgrp);
545 return cgroup_css(cgrp, ss);
549 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
550 * @cgrp: the cgroup of interest
551 * @ss: the subsystem of interest
553 * Find and get the effective css of @cgrp for @ss. The effective css is
554 * defined as the matching css of the nearest ancestor including self which
555 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
556 * the root css is returned, so this function always returns a valid css.
558 * The returned css is not guaranteed to be online, and therefore it is the
559 * callers responsibility to try get a reference for it.
561 struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
562 struct cgroup_subsys *ss)
564 struct cgroup_subsys_state *css;
566 if (!CGROUP_HAS_SUBSYS_CONFIG)
570 css = cgroup_css(cgrp, ss);
574 cgrp = cgroup_parent(cgrp);
577 return init_css_set.subsys[ss->id];
581 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
582 * @cgrp: the cgroup of interest
583 * @ss: the subsystem of interest
585 * Find and get the effective css of @cgrp for @ss. The effective css is
586 * defined as the matching css of the nearest ancestor including self which
587 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
588 * the root css is returned, so this function always returns a valid css.
589 * The returned css must be put using css_put().
591 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
592 struct cgroup_subsys *ss)
594 struct cgroup_subsys_state *css;
596 if (!CGROUP_HAS_SUBSYS_CONFIG)
602 css = cgroup_css(cgrp, ss);
604 if (css && css_tryget_online(css))
606 cgrp = cgroup_parent(cgrp);
609 css = init_css_set.subsys[ss->id];
615 EXPORT_SYMBOL_GPL(cgroup_get_e_css);
617 static void cgroup_get_live(struct cgroup *cgrp)
619 WARN_ON_ONCE(cgroup_is_dead(cgrp));
624 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
625 * is responsible for taking the css_set_lock.
626 * @cgrp: the cgroup in question
628 int __cgroup_task_count(const struct cgroup *cgrp)
631 struct cgrp_cset_link *link;
633 lockdep_assert_held(&css_set_lock);
635 list_for_each_entry(link, &cgrp->cset_links, cset_link)
636 count += link->cset->nr_tasks;
642 * cgroup_task_count - count the number of tasks in a cgroup.
643 * @cgrp: the cgroup in question
645 int cgroup_task_count(const struct cgroup *cgrp)
649 spin_lock_irq(&css_set_lock);
650 count = __cgroup_task_count(cgrp);
651 spin_unlock_irq(&css_set_lock);
656 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
658 struct cgroup *cgrp = of->kn->parent->priv;
659 struct cftype *cft = of_cft(of);
662 * This is open and unprotected implementation of cgroup_css().
663 * seq_css() is only called from a kernfs file operation which has
664 * an active reference on the file. Because all the subsystem
665 * files are drained before a css is disassociated with a cgroup,
666 * the matching css from the cgroup's subsys table is guaranteed to
667 * be and stay valid until the enclosing operation is complete.
669 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss)
670 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
674 EXPORT_SYMBOL_GPL(of_css);
677 * for_each_css - iterate all css's of a cgroup
678 * @css: the iteration cursor
679 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
680 * @cgrp: the target cgroup to iterate css's of
682 * Should be called under cgroup_[tree_]mutex.
684 #define for_each_css(css, ssid, cgrp) \
685 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
686 if (!((css) = rcu_dereference_check( \
687 (cgrp)->subsys[(ssid)], \
688 lockdep_is_held(&cgroup_mutex)))) { } \
692 * do_each_subsys_mask - filter for_each_subsys with a bitmask
693 * @ss: the iteration cursor
694 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
695 * @ss_mask: the bitmask
697 * The block will only run for cases where the ssid-th bit (1 << ssid) of
700 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
701 unsigned long __ss_mask = (ss_mask); \
702 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
706 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
707 (ss) = cgroup_subsys[ssid]; \
710 #define while_each_subsys_mask() \
715 /* iterate over child cgrps, lock should be held throughout iteration */
716 #define cgroup_for_each_live_child(child, cgrp) \
717 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
718 if (({ lockdep_assert_held(&cgroup_mutex); \
719 cgroup_is_dead(child); })) \
723 /* walk live descendants in pre order */
724 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
725 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
726 if (({ lockdep_assert_held(&cgroup_mutex); \
727 (dsct) = (d_css)->cgroup; \
728 cgroup_is_dead(dsct); })) \
732 /* walk live descendants in postorder */
733 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
734 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
735 if (({ lockdep_assert_held(&cgroup_mutex); \
736 (dsct) = (d_css)->cgroup; \
737 cgroup_is_dead(dsct); })) \
742 * The default css_set - used by init and its children prior to any
743 * hierarchies being mounted. It contains a pointer to the root state
744 * for each subsystem. Also used to anchor the list of css_sets. Not
745 * reference-counted, to improve performance when child cgroups
746 * haven't been created.
748 struct css_set init_css_set = {
749 .refcount = REFCOUNT_INIT(1),
750 .dom_cset = &init_css_set,
751 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
752 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
753 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks),
754 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters),
755 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets),
756 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
757 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node),
758 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node),
759 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
762 * The following field is re-initialized when this cset gets linked
763 * in cgroup_init(). However, let's initialize the field
764 * statically too so that the default cgroup can be accessed safely
767 .dfl_cgrp = &cgrp_dfl_root.cgrp,
770 static int css_set_count = 1; /* 1 for init_css_set */
772 static bool css_set_threaded(struct css_set *cset)
774 return cset->dom_cset != cset;
778 * css_set_populated - does a css_set contain any tasks?
779 * @cset: target css_set
781 * css_set_populated() should be the same as !!cset->nr_tasks at steady
782 * state. However, css_set_populated() can be called while a task is being
783 * added to or removed from the linked list before the nr_tasks is
784 * properly updated. Hence, we can't just look at ->nr_tasks here.
786 static bool css_set_populated(struct css_set *cset)
788 lockdep_assert_held(&css_set_lock);
790 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
794 * cgroup_update_populated - update the populated count of a cgroup
795 * @cgrp: the target cgroup
796 * @populated: inc or dec populated count
798 * One of the css_sets associated with @cgrp is either getting its first
799 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
800 * count is propagated towards root so that a given cgroup's
801 * nr_populated_children is zero iff none of its descendants contain any
804 * @cgrp's interface file "cgroup.populated" is zero if both
805 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
806 * 1 otherwise. When the sum changes from or to zero, userland is notified
807 * that the content of the interface file has changed. This can be used to
808 * detect when @cgrp and its descendants become populated or empty.
810 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
812 struct cgroup *child = NULL;
813 int adj = populated ? 1 : -1;
815 lockdep_assert_held(&css_set_lock);
818 bool was_populated = cgroup_is_populated(cgrp);
821 cgrp->nr_populated_csets += adj;
823 if (cgroup_is_threaded(child))
824 cgrp->nr_populated_threaded_children += adj;
826 cgrp->nr_populated_domain_children += adj;
829 if (was_populated == cgroup_is_populated(cgrp))
832 cgroup1_check_for_release(cgrp);
833 TRACE_CGROUP_PATH(notify_populated, cgrp,
834 cgroup_is_populated(cgrp));
835 cgroup_file_notify(&cgrp->events_file);
838 cgrp = cgroup_parent(cgrp);
843 * css_set_update_populated - update populated state of a css_set
844 * @cset: target css_set
845 * @populated: whether @cset is populated or depopulated
847 * @cset is either getting the first task or losing the last. Update the
848 * populated counters of all associated cgroups accordingly.
850 static void css_set_update_populated(struct css_set *cset, bool populated)
852 struct cgrp_cset_link *link;
854 lockdep_assert_held(&css_set_lock);
856 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
857 cgroup_update_populated(link->cgrp, populated);
861 * @task is leaving, advance task iterators which are pointing to it so
862 * that they can resume at the next position. Advancing an iterator might
863 * remove it from the list, use safe walk. See css_task_iter_skip() for
866 static void css_set_skip_task_iters(struct css_set *cset,
867 struct task_struct *task)
869 struct css_task_iter *it, *pos;
871 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
872 css_task_iter_skip(it, task);
876 * css_set_move_task - move a task from one css_set to another
877 * @task: task being moved
878 * @from_cset: css_set @task currently belongs to (may be NULL)
879 * @to_cset: new css_set @task is being moved to (may be NULL)
880 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
882 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
883 * css_set, @from_cset can be NULL. If @task is being disassociated
884 * instead of moved, @to_cset can be NULL.
886 * This function automatically handles populated counter updates and
887 * css_task_iter adjustments but the caller is responsible for managing
888 * @from_cset and @to_cset's reference counts.
890 static void css_set_move_task(struct task_struct *task,
891 struct css_set *from_cset, struct css_set *to_cset,
894 lockdep_assert_held(&css_set_lock);
896 if (to_cset && !css_set_populated(to_cset))
897 css_set_update_populated(to_cset, true);
900 WARN_ON_ONCE(list_empty(&task->cg_list));
902 css_set_skip_task_iters(from_cset, task);
903 list_del_init(&task->cg_list);
904 if (!css_set_populated(from_cset))
905 css_set_update_populated(from_cset, false);
907 WARN_ON_ONCE(!list_empty(&task->cg_list));
912 * We are synchronized through cgroup_threadgroup_rwsem
913 * against PF_EXITING setting such that we can't race
914 * against cgroup_exit()/cgroup_free() dropping the css_set.
916 WARN_ON_ONCE(task->flags & PF_EXITING);
918 cgroup_move_task(task, to_cset);
919 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
925 * hash table for cgroup groups. This improves the performance to find
926 * an existing css_set. This hash doesn't (currently) take into
927 * account cgroups in empty hierarchies.
929 #define CSS_SET_HASH_BITS 7
930 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
932 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
934 unsigned long key = 0UL;
935 struct cgroup_subsys *ss;
938 for_each_subsys(ss, i)
939 key += (unsigned long)css[i];
940 key = (key >> 16) ^ key;
945 void put_css_set_locked(struct css_set *cset)
947 struct cgrp_cset_link *link, *tmp_link;
948 struct cgroup_subsys *ss;
951 lockdep_assert_held(&css_set_lock);
953 if (!refcount_dec_and_test(&cset->refcount))
956 WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
958 /* This css_set is dead. Unlink it and release cgroup and css refs */
959 for_each_subsys(ss, ssid) {
960 list_del(&cset->e_cset_node[ssid]);
961 css_put(cset->subsys[ssid]);
963 hash_del(&cset->hlist);
966 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
967 list_del(&link->cset_link);
968 list_del(&link->cgrp_link);
969 if (cgroup_parent(link->cgrp))
970 cgroup_put(link->cgrp);
974 if (css_set_threaded(cset)) {
975 list_del(&cset->threaded_csets_node);
976 put_css_set_locked(cset->dom_cset);
979 kfree_rcu(cset, rcu_head);
983 * compare_css_sets - helper function for find_existing_css_set().
984 * @cset: candidate css_set being tested
985 * @old_cset: existing css_set for a task
986 * @new_cgrp: cgroup that's being entered by the task
987 * @template: desired set of css pointers in css_set (pre-calculated)
989 * Returns true if "cset" matches "old_cset" except for the hierarchy
990 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
992 static bool compare_css_sets(struct css_set *cset,
993 struct css_set *old_cset,
994 struct cgroup *new_cgrp,
995 struct cgroup_subsys_state *template[])
997 struct cgroup *new_dfl_cgrp;
998 struct list_head *l1, *l2;
1001 * On the default hierarchy, there can be csets which are
1002 * associated with the same set of cgroups but different csses.
1003 * Let's first ensure that csses match.
1005 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1009 /* @cset's domain should match the default cgroup's */
1010 if (cgroup_on_dfl(new_cgrp))
1011 new_dfl_cgrp = new_cgrp;
1013 new_dfl_cgrp = old_cset->dfl_cgrp;
1015 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1019 * Compare cgroup pointers in order to distinguish between
1020 * different cgroups in hierarchies. As different cgroups may
1021 * share the same effective css, this comparison is always
1024 l1 = &cset->cgrp_links;
1025 l2 = &old_cset->cgrp_links;
1027 struct cgrp_cset_link *link1, *link2;
1028 struct cgroup *cgrp1, *cgrp2;
1032 /* See if we reached the end - both lists are equal length. */
1033 if (l1 == &cset->cgrp_links) {
1034 BUG_ON(l2 != &old_cset->cgrp_links);
1037 BUG_ON(l2 == &old_cset->cgrp_links);
1039 /* Locate the cgroups associated with these links. */
1040 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1041 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1042 cgrp1 = link1->cgrp;
1043 cgrp2 = link2->cgrp;
1044 /* Hierarchies should be linked in the same order. */
1045 BUG_ON(cgrp1->root != cgrp2->root);
1048 * If this hierarchy is the hierarchy of the cgroup
1049 * that's changing, then we need to check that this
1050 * css_set points to the new cgroup; if it's any other
1051 * hierarchy, then this css_set should point to the
1052 * same cgroup as the old css_set.
1054 if (cgrp1->root == new_cgrp->root) {
1055 if (cgrp1 != new_cgrp)
1066 * find_existing_css_set - init css array and find the matching css_set
1067 * @old_cset: the css_set that we're using before the cgroup transition
1068 * @cgrp: the cgroup that we're moving into
1069 * @template: out param for the new set of csses, should be clear on entry
1071 static struct css_set *find_existing_css_set(struct css_set *old_cset,
1072 struct cgroup *cgrp,
1073 struct cgroup_subsys_state *template[])
1075 struct cgroup_root *root = cgrp->root;
1076 struct cgroup_subsys *ss;
1077 struct css_set *cset;
1082 * Build the set of subsystem state objects that we want to see in the
1083 * new css_set. While subsystems can change globally, the entries here
1084 * won't change, so no need for locking.
1086 for_each_subsys(ss, i) {
1087 if (root->subsys_mask & (1UL << i)) {
1089 * @ss is in this hierarchy, so we want the
1090 * effective css from @cgrp.
1092 template[i] = cgroup_e_css_by_mask(cgrp, ss);
1095 * @ss is not in this hierarchy, so we don't want
1096 * to change the css.
1098 template[i] = old_cset->subsys[i];
1102 key = css_set_hash(template);
1103 hash_for_each_possible(css_set_table, cset, hlist, key) {
1104 if (!compare_css_sets(cset, old_cset, cgrp, template))
1107 /* This css_set matches what we need */
1111 /* No existing cgroup group matched */
1115 static void free_cgrp_cset_links(struct list_head *links_to_free)
1117 struct cgrp_cset_link *link, *tmp_link;
1119 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1120 list_del(&link->cset_link);
1126 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1127 * @count: the number of links to allocate
1128 * @tmp_links: list_head the allocated links are put on
1130 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1131 * through ->cset_link. Returns 0 on success or -errno.
1133 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1135 struct cgrp_cset_link *link;
1138 INIT_LIST_HEAD(tmp_links);
1140 for (i = 0; i < count; i++) {
1141 link = kzalloc(sizeof(*link), GFP_KERNEL);
1143 free_cgrp_cset_links(tmp_links);
1146 list_add(&link->cset_link, tmp_links);
1152 * link_css_set - a helper function to link a css_set to a cgroup
1153 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1154 * @cset: the css_set to be linked
1155 * @cgrp: the destination cgroup
1157 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1158 struct cgroup *cgrp)
1160 struct cgrp_cset_link *link;
1162 BUG_ON(list_empty(tmp_links));
1164 if (cgroup_on_dfl(cgrp))
1165 cset->dfl_cgrp = cgrp;
1167 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1172 * Always add links to the tail of the lists so that the lists are
1173 * in chronological order.
1175 list_move_tail(&link->cset_link, &cgrp->cset_links);
1176 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1178 if (cgroup_parent(cgrp))
1179 cgroup_get_live(cgrp);
1183 * find_css_set - return a new css_set with one cgroup updated
1184 * @old_cset: the baseline css_set
1185 * @cgrp: the cgroup to be updated
1187 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1188 * substituted into the appropriate hierarchy.
1190 static struct css_set *find_css_set(struct css_set *old_cset,
1191 struct cgroup *cgrp)
1193 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1194 struct css_set *cset;
1195 struct list_head tmp_links;
1196 struct cgrp_cset_link *link;
1197 struct cgroup_subsys *ss;
1201 lockdep_assert_held(&cgroup_mutex);
1203 /* First see if we already have a cgroup group that matches
1204 * the desired set */
1205 spin_lock_irq(&css_set_lock);
1206 cset = find_existing_css_set(old_cset, cgrp, template);
1209 spin_unlock_irq(&css_set_lock);
1214 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1218 /* Allocate all the cgrp_cset_link objects that we'll need */
1219 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1224 refcount_set(&cset->refcount, 1);
1225 cset->dom_cset = cset;
1226 INIT_LIST_HEAD(&cset->tasks);
1227 INIT_LIST_HEAD(&cset->mg_tasks);
1228 INIT_LIST_HEAD(&cset->dying_tasks);
1229 INIT_LIST_HEAD(&cset->task_iters);
1230 INIT_LIST_HEAD(&cset->threaded_csets);
1231 INIT_HLIST_NODE(&cset->hlist);
1232 INIT_LIST_HEAD(&cset->cgrp_links);
1233 INIT_LIST_HEAD(&cset->mg_src_preload_node);
1234 INIT_LIST_HEAD(&cset->mg_dst_preload_node);
1235 INIT_LIST_HEAD(&cset->mg_node);
1237 /* Copy the set of subsystem state objects generated in
1238 * find_existing_css_set() */
1239 memcpy(cset->subsys, template, sizeof(cset->subsys));
1241 spin_lock_irq(&css_set_lock);
1242 /* Add reference counts and links from the new css_set. */
1243 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1244 struct cgroup *c = link->cgrp;
1246 if (c->root == cgrp->root)
1248 link_css_set(&tmp_links, cset, c);
1251 BUG_ON(!list_empty(&tmp_links));
1255 /* Add @cset to the hash table */
1256 key = css_set_hash(cset->subsys);
1257 hash_add(css_set_table, &cset->hlist, key);
1259 for_each_subsys(ss, ssid) {
1260 struct cgroup_subsys_state *css = cset->subsys[ssid];
1262 list_add_tail(&cset->e_cset_node[ssid],
1263 &css->cgroup->e_csets[ssid]);
1267 spin_unlock_irq(&css_set_lock);
1270 * If @cset should be threaded, look up the matching dom_cset and
1271 * link them up. We first fully initialize @cset then look for the
1272 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1273 * to stay empty until we return.
1275 if (cgroup_is_threaded(cset->dfl_cgrp)) {
1276 struct css_set *dcset;
1278 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1284 spin_lock_irq(&css_set_lock);
1285 cset->dom_cset = dcset;
1286 list_add_tail(&cset->threaded_csets_node,
1287 &dcset->threaded_csets);
1288 spin_unlock_irq(&css_set_lock);
1294 struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1296 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv;
1298 return root_cgrp->root;
1301 void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
1303 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
1305 /* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1306 if (favor && !favoring) {
1307 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
1308 root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1309 } else if (!favor && favoring) {
1310 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
1311 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
1315 static int cgroup_init_root_id(struct cgroup_root *root)
1319 lockdep_assert_held(&cgroup_mutex);
1321 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1325 root->hierarchy_id = id;
1329 static void cgroup_exit_root_id(struct cgroup_root *root)
1331 lockdep_assert_held(&cgroup_mutex);
1333 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1336 void cgroup_free_root(struct cgroup_root *root)
1341 static void cgroup_destroy_root(struct cgroup_root *root)
1343 struct cgroup *cgrp = &root->cgrp;
1344 struct cgrp_cset_link *link, *tmp_link;
1346 trace_cgroup_destroy_root(root);
1348 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1350 BUG_ON(atomic_read(&root->nr_cgrps));
1351 BUG_ON(!list_empty(&cgrp->self.children));
1353 /* Rebind all subsystems back to the default hierarchy */
1354 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1357 * Release all the links from cset_links to this hierarchy's
1360 spin_lock_irq(&css_set_lock);
1362 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1363 list_del(&link->cset_link);
1364 list_del(&link->cgrp_link);
1368 spin_unlock_irq(&css_set_lock);
1370 if (!list_empty(&root->root_list)) {
1371 list_del(&root->root_list);
1372 cgroup_root_count--;
1375 cgroup_favor_dynmods(root, false);
1376 cgroup_exit_root_id(root);
1380 cgroup_rstat_exit(cgrp);
1381 kernfs_destroy_root(root->kf_root);
1382 cgroup_free_root(root);
1386 * Returned cgroup is without refcount but it's valid as long as cset pins it.
1388 static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
1389 struct cgroup_root *root)
1391 struct cgroup *res_cgroup = NULL;
1393 if (cset == &init_css_set) {
1394 res_cgroup = &root->cgrp;
1395 } else if (root == &cgrp_dfl_root) {
1396 res_cgroup = cset->dfl_cgrp;
1398 struct cgrp_cset_link *link;
1399 lockdep_assert_held(&css_set_lock);
1401 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1402 struct cgroup *c = link->cgrp;
1404 if (c->root == root) {
1411 BUG_ON(!res_cgroup);
1416 * look up cgroup associated with current task's cgroup namespace on the
1417 * specified hierarchy
1419 static struct cgroup *
1420 current_cgns_cgroup_from_root(struct cgroup_root *root)
1422 struct cgroup *res = NULL;
1423 struct css_set *cset;
1425 lockdep_assert_held(&css_set_lock);
1429 cset = current->nsproxy->cgroup_ns->root_cset;
1430 res = __cset_cgroup_from_root(cset, root);
1438 * Look up cgroup associated with current task's cgroup namespace on the default
1441 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
1442 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
1444 * - css_set_lock is not needed because we just read cset->dfl_cgrp.
1445 * - As a bonus returned cgrp is pinned with the current because it cannot
1446 * switch cgroup_ns asynchronously.
1448 static struct cgroup *current_cgns_cgroup_dfl(void)
1450 struct css_set *cset;
1452 if (current->nsproxy) {
1453 cset = current->nsproxy->cgroup_ns->root_cset;
1454 return __cset_cgroup_from_root(cset, &cgrp_dfl_root);
1457 * NOTE: This function may be called from bpf_cgroup_from_id()
1458 * on a task which has already passed exit_task_namespaces() and
1459 * nsproxy == NULL. Fall back to cgrp_dfl_root which will make all
1460 * cgroups visible for lookups.
1462 return &cgrp_dfl_root.cgrp;
1466 /* look up cgroup associated with given css_set on the specified hierarchy */
1467 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1468 struct cgroup_root *root)
1470 lockdep_assert_held(&cgroup_mutex);
1471 lockdep_assert_held(&css_set_lock);
1473 return __cset_cgroup_from_root(cset, root);
1477 * Return the cgroup for "task" from the given hierarchy. Must be
1478 * called with cgroup_mutex and css_set_lock held.
1480 struct cgroup *task_cgroup_from_root(struct task_struct *task,
1481 struct cgroup_root *root)
1484 * No need to lock the task - since we hold css_set_lock the
1485 * task can't change groups.
1487 return cset_cgroup_from_root(task_css_set(task), root);
1491 * A task must hold cgroup_mutex to modify cgroups.
1493 * Any task can increment and decrement the count field without lock.
1494 * So in general, code holding cgroup_mutex can't rely on the count
1495 * field not changing. However, if the count goes to zero, then only
1496 * cgroup_attach_task() can increment it again. Because a count of zero
1497 * means that no tasks are currently attached, therefore there is no
1498 * way a task attached to that cgroup can fork (the other way to
1499 * increment the count). So code holding cgroup_mutex can safely
1500 * assume that if the count is zero, it will stay zero. Similarly, if
1501 * a task holds cgroup_mutex on a cgroup with zero count, it
1502 * knows that the cgroup won't be removed, as cgroup_rmdir()
1505 * A cgroup can only be deleted if both its 'count' of using tasks
1506 * is zero, and its list of 'children' cgroups is empty. Since all
1507 * tasks in the system use _some_ cgroup, and since there is always at
1508 * least one task in the system (init, pid == 1), therefore, root cgroup
1509 * always has either children cgroups and/or using tasks. So we don't
1510 * need a special hack to ensure that root cgroup cannot be deleted.
1512 * P.S. One more locking exception. RCU is used to guard the
1513 * update of a tasks cgroup pointer by cgroup_attach_task()
1516 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1518 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1521 struct cgroup_subsys *ss = cft->ss;
1523 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1524 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1525 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1527 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1528 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1531 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1537 * cgroup_file_mode - deduce file mode of a control file
1538 * @cft: the control file in question
1540 * S_IRUGO for read, S_IWUSR for write.
1542 static umode_t cgroup_file_mode(const struct cftype *cft)
1546 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1549 if (cft->write_u64 || cft->write_s64 || cft->write) {
1550 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1560 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1561 * @subtree_control: the new subtree_control mask to consider
1562 * @this_ss_mask: available subsystems
1564 * On the default hierarchy, a subsystem may request other subsystems to be
1565 * enabled together through its ->depends_on mask. In such cases, more
1566 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1568 * This function calculates which subsystems need to be enabled if
1569 * @subtree_control is to be applied while restricted to @this_ss_mask.
1571 static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1573 u16 cur_ss_mask = subtree_control;
1574 struct cgroup_subsys *ss;
1577 lockdep_assert_held(&cgroup_mutex);
1579 cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1582 u16 new_ss_mask = cur_ss_mask;
1584 do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1585 new_ss_mask |= ss->depends_on;
1586 } while_each_subsys_mask();
1589 * Mask out subsystems which aren't available. This can
1590 * happen only if some depended-upon subsystems were bound
1591 * to non-default hierarchies.
1593 new_ss_mask &= this_ss_mask;
1595 if (new_ss_mask == cur_ss_mask)
1597 cur_ss_mask = new_ss_mask;
1604 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1605 * @kn: the kernfs_node being serviced
1607 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1608 * the method finishes if locking succeeded. Note that once this function
1609 * returns the cgroup returned by cgroup_kn_lock_live() may become
1610 * inaccessible any time. If the caller intends to continue to access the
1611 * cgroup, it should pin it before invoking this function.
1613 void cgroup_kn_unlock(struct kernfs_node *kn)
1615 struct cgroup *cgrp;
1617 if (kernfs_type(kn) == KERNFS_DIR)
1620 cgrp = kn->parent->priv;
1624 kernfs_unbreak_active_protection(kn);
1629 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1630 * @kn: the kernfs_node being serviced
1631 * @drain_offline: perform offline draining on the cgroup
1633 * This helper is to be used by a cgroup kernfs method currently servicing
1634 * @kn. It breaks the active protection, performs cgroup locking and
1635 * verifies that the associated cgroup is alive. Returns the cgroup if
1636 * alive; otherwise, %NULL. A successful return should be undone by a
1637 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1638 * cgroup is drained of offlining csses before return.
1640 * Any cgroup kernfs method implementation which requires locking the
1641 * associated cgroup should use this helper. It avoids nesting cgroup
1642 * locking under kernfs active protection and allows all kernfs operations
1643 * including self-removal.
1645 struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1647 struct cgroup *cgrp;
1649 if (kernfs_type(kn) == KERNFS_DIR)
1652 cgrp = kn->parent->priv;
1655 * We're gonna grab cgroup_mutex which nests outside kernfs
1656 * active_ref. cgroup liveliness check alone provides enough
1657 * protection against removal. Ensure @cgrp stays accessible and
1658 * break the active_ref protection.
1660 if (!cgroup_tryget(cgrp))
1662 kernfs_break_active_protection(kn);
1665 cgroup_lock_and_drain_offline(cgrp);
1669 if (!cgroup_is_dead(cgrp))
1672 cgroup_kn_unlock(kn);
1676 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1678 char name[CGROUP_FILE_NAME_MAX];
1680 lockdep_assert_held(&cgroup_mutex);
1682 if (cft->file_offset) {
1683 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1684 struct cgroup_file *cfile = (void *)css + cft->file_offset;
1686 spin_lock_irq(&cgroup_file_kn_lock);
1688 spin_unlock_irq(&cgroup_file_kn_lock);
1690 del_timer_sync(&cfile->notify_timer);
1693 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1697 * css_clear_dir - remove subsys files in a cgroup directory
1700 static void css_clear_dir(struct cgroup_subsys_state *css)
1702 struct cgroup *cgrp = css->cgroup;
1703 struct cftype *cfts;
1705 if (!(css->flags & CSS_VISIBLE))
1708 css->flags &= ~CSS_VISIBLE;
1711 if (cgroup_on_dfl(cgrp)) {
1712 cgroup_addrm_files(css, cgrp,
1713 cgroup_base_files, false);
1714 if (cgroup_psi_enabled())
1715 cgroup_addrm_files(css, cgrp,
1716 cgroup_psi_files, false);
1718 cgroup_addrm_files(css, cgrp,
1719 cgroup1_base_files, false);
1722 list_for_each_entry(cfts, &css->ss->cfts, node)
1723 cgroup_addrm_files(css, cgrp, cfts, false);
1728 * css_populate_dir - create subsys files in a cgroup directory
1731 * On failure, no file is added.
1733 static int css_populate_dir(struct cgroup_subsys_state *css)
1735 struct cgroup *cgrp = css->cgroup;
1736 struct cftype *cfts, *failed_cfts;
1739 if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1743 if (cgroup_on_dfl(cgrp)) {
1744 ret = cgroup_addrm_files(&cgrp->self, cgrp,
1745 cgroup_base_files, true);
1749 if (cgroup_psi_enabled()) {
1750 ret = cgroup_addrm_files(&cgrp->self, cgrp,
1751 cgroup_psi_files, true);
1756 cgroup_addrm_files(css, cgrp,
1757 cgroup1_base_files, true);
1760 list_for_each_entry(cfts, &css->ss->cfts, node) {
1761 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1769 css->flags |= CSS_VISIBLE;
1773 list_for_each_entry(cfts, &css->ss->cfts, node) {
1774 if (cfts == failed_cfts)
1776 cgroup_addrm_files(css, cgrp, cfts, false);
1781 int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1783 struct cgroup *dcgrp = &dst_root->cgrp;
1784 struct cgroup_subsys *ss;
1786 u16 dfl_disable_ss_mask = 0;
1788 lockdep_assert_held(&cgroup_mutex);
1790 do_each_subsys_mask(ss, ssid, ss_mask) {
1792 * If @ss has non-root csses attached to it, can't move.
1793 * If @ss is an implicit controller, it is exempt from this
1794 * rule and can be stolen.
1796 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1797 !ss->implicit_on_dfl)
1800 /* can't move between two non-dummy roots either */
1801 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1805 * Collect ssid's that need to be disabled from default
1808 if (ss->root == &cgrp_dfl_root)
1809 dfl_disable_ss_mask |= 1 << ssid;
1811 } while_each_subsys_mask();
1813 if (dfl_disable_ss_mask) {
1814 struct cgroup *scgrp = &cgrp_dfl_root.cgrp;
1817 * Controllers from default hierarchy that need to be rebound
1818 * are all disabled together in one go.
1820 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask;
1821 WARN_ON(cgroup_apply_control(scgrp));
1822 cgroup_finalize_control(scgrp, 0);
1825 do_each_subsys_mask(ss, ssid, ss_mask) {
1826 struct cgroup_root *src_root = ss->root;
1827 struct cgroup *scgrp = &src_root->cgrp;
1828 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1829 struct css_set *cset, *cset_pos;
1830 struct css_task_iter *it;
1832 WARN_ON(!css || cgroup_css(dcgrp, ss));
1834 if (src_root != &cgrp_dfl_root) {
1835 /* disable from the source */
1836 src_root->subsys_mask &= ~(1 << ssid);
1837 WARN_ON(cgroup_apply_control(scgrp));
1838 cgroup_finalize_control(scgrp, 0);
1842 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1843 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1844 ss->root = dst_root;
1845 css->cgroup = dcgrp;
1847 spin_lock_irq(&css_set_lock);
1848 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id]));
1849 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id],
1850 e_cset_node[ss->id]) {
1851 list_move_tail(&cset->e_cset_node[ss->id],
1852 &dcgrp->e_csets[ss->id]);
1854 * all css_sets of scgrp together in same order to dcgrp,
1855 * patch in-flight iterators to preserve correct iteration.
1856 * since the iterator is always advanced right away and
1857 * finished when it->cset_pos meets it->cset_head, so only
1858 * update it->cset_head is enough here.
1860 list_for_each_entry(it, &cset->task_iters, iters_node)
1861 if (it->cset_head == &scgrp->e_csets[ss->id])
1862 it->cset_head = &dcgrp->e_csets[ss->id];
1864 spin_unlock_irq(&css_set_lock);
1866 if (ss->css_rstat_flush) {
1867 list_del_rcu(&css->rstat_css_node);
1869 list_add_rcu(&css->rstat_css_node,
1870 &dcgrp->rstat_css_list);
1873 /* default hierarchy doesn't enable controllers by default */
1874 dst_root->subsys_mask |= 1 << ssid;
1875 if (dst_root == &cgrp_dfl_root) {
1876 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1878 dcgrp->subtree_control |= 1 << ssid;
1879 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1882 ret = cgroup_apply_control(dcgrp);
1884 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1889 } while_each_subsys_mask();
1891 kernfs_activate(dcgrp->kn);
1895 int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1896 struct kernfs_root *kf_root)
1900 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1901 struct cgroup *ns_cgroup;
1903 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1907 spin_lock_irq(&css_set_lock);
1908 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1909 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1910 spin_unlock_irq(&css_set_lock);
1912 if (len >= PATH_MAX)
1915 seq_escape(sf, buf, " \t\n\\");
1922 enum cgroup2_param {
1925 Opt_memory_localevents,
1926 Opt_memory_recursiveprot,
1930 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1931 fsparam_flag("nsdelegate", Opt_nsdelegate),
1932 fsparam_flag("favordynmods", Opt_favordynmods),
1933 fsparam_flag("memory_localevents", Opt_memory_localevents),
1934 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot),
1938 static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1940 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1941 struct fs_parse_result result;
1944 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1949 case Opt_nsdelegate:
1950 ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1952 case Opt_favordynmods:
1953 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
1955 case Opt_memory_localevents:
1956 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1958 case Opt_memory_recursiveprot:
1959 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1965 static void apply_cgroup_root_flags(unsigned int root_flags)
1967 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1968 if (root_flags & CGRP_ROOT_NS_DELEGATE)
1969 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1971 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1973 cgroup_favor_dynmods(&cgrp_dfl_root,
1974 root_flags & CGRP_ROOT_FAVOR_DYNMODS);
1976 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1977 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1979 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1981 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1982 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1984 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1988 static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1990 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1991 seq_puts(seq, ",nsdelegate");
1992 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
1993 seq_puts(seq, ",favordynmods");
1994 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1995 seq_puts(seq, ",memory_localevents");
1996 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1997 seq_puts(seq, ",memory_recursiveprot");
2001 static int cgroup_reconfigure(struct fs_context *fc)
2003 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2005 apply_cgroup_root_flags(ctx->flags);
2009 static void init_cgroup_housekeeping(struct cgroup *cgrp)
2011 struct cgroup_subsys *ss;
2014 INIT_LIST_HEAD(&cgrp->self.sibling);
2015 INIT_LIST_HEAD(&cgrp->self.children);
2016 INIT_LIST_HEAD(&cgrp->cset_links);
2017 INIT_LIST_HEAD(&cgrp->pidlists);
2018 mutex_init(&cgrp->pidlist_mutex);
2019 cgrp->self.cgroup = cgrp;
2020 cgrp->self.flags |= CSS_ONLINE;
2021 cgrp->dom_cgrp = cgrp;
2022 cgrp->max_descendants = INT_MAX;
2023 cgrp->max_depth = INT_MAX;
2024 INIT_LIST_HEAD(&cgrp->rstat_css_list);
2025 prev_cputime_init(&cgrp->prev_cputime);
2027 for_each_subsys(ss, ssid)
2028 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
2030 init_waitqueue_head(&cgrp->offline_waitq);
2031 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
2034 void init_cgroup_root(struct cgroup_fs_context *ctx)
2036 struct cgroup_root *root = ctx->root;
2037 struct cgroup *cgrp = &root->cgrp;
2039 INIT_LIST_HEAD(&root->root_list);
2040 atomic_set(&root->nr_cgrps, 1);
2042 init_cgroup_housekeeping(cgrp);
2044 /* DYNMODS must be modified through cgroup_favor_dynmods() */
2045 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
2046 if (ctx->release_agent)
2047 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
2049 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
2050 if (ctx->cpuset_clone_children)
2051 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
2054 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
2056 LIST_HEAD(tmp_links);
2057 struct cgroup *root_cgrp = &root->cgrp;
2058 struct kernfs_syscall_ops *kf_sops;
2059 struct css_set *cset;
2062 lockdep_assert_held(&cgroup_mutex);
2064 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
2070 * We're accessing css_set_count without locking css_set_lock here,
2071 * but that's OK - it can only be increased by someone holding
2072 * cgroup_lock, and that's us. Later rebinding may disable
2073 * controllers on the default hierarchy and thus create new csets,
2074 * which can't be more than the existing ones. Allocate 2x.
2076 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
2080 ret = cgroup_init_root_id(root);
2084 kf_sops = root == &cgrp_dfl_root ?
2085 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
2087 root->kf_root = kernfs_create_root(kf_sops,
2088 KERNFS_ROOT_CREATE_DEACTIVATED |
2089 KERNFS_ROOT_SUPPORT_EXPORTOP |
2090 KERNFS_ROOT_SUPPORT_USER_XATTR,
2092 if (IS_ERR(root->kf_root)) {
2093 ret = PTR_ERR(root->kf_root);
2096 root_cgrp->kn = kernfs_root_to_node(root->kf_root);
2097 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
2098 root_cgrp->ancestors[0] = root_cgrp;
2100 ret = css_populate_dir(&root_cgrp->self);
2104 ret = cgroup_rstat_init(root_cgrp);
2108 ret = rebind_subsystems(root, ss_mask);
2112 ret = cgroup_bpf_inherit(root_cgrp);
2115 trace_cgroup_setup_root(root);
2118 * There must be no failure case after here, since rebinding takes
2119 * care of subsystems' refcounts, which are explicitly dropped in
2120 * the failure exit path.
2122 list_add(&root->root_list, &cgroup_roots);
2123 cgroup_root_count++;
2126 * Link the root cgroup in this hierarchy into all the css_set
2129 spin_lock_irq(&css_set_lock);
2130 hash_for_each(css_set_table, i, cset, hlist) {
2131 link_css_set(&tmp_links, cset, root_cgrp);
2132 if (css_set_populated(cset))
2133 cgroup_update_populated(root_cgrp, true);
2135 spin_unlock_irq(&css_set_lock);
2137 BUG_ON(!list_empty(&root_cgrp->self.children));
2138 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2144 cgroup_rstat_exit(root_cgrp);
2146 kernfs_destroy_root(root->kf_root);
2147 root->kf_root = NULL;
2149 cgroup_exit_root_id(root);
2151 percpu_ref_exit(&root_cgrp->self.refcnt);
2153 free_cgrp_cset_links(&tmp_links);
2157 int cgroup_do_get_tree(struct fs_context *fc)
2159 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2162 ctx->kfc.root = ctx->root->kf_root;
2163 if (fc->fs_type == &cgroup2_fs_type)
2164 ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2166 ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2167 ret = kernfs_get_tree(fc);
2170 * In non-init cgroup namespace, instead of root cgroup's dentry,
2171 * we return the dentry corresponding to the cgroupns->root_cgrp.
2173 if (!ret && ctx->ns != &init_cgroup_ns) {
2174 struct dentry *nsdentry;
2175 struct super_block *sb = fc->root->d_sb;
2176 struct cgroup *cgrp;
2179 spin_lock_irq(&css_set_lock);
2181 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2183 spin_unlock_irq(&css_set_lock);
2186 nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2188 if (IS_ERR(nsdentry)) {
2189 deactivate_locked_super(sb);
2190 ret = PTR_ERR(nsdentry);
2193 fc->root = nsdentry;
2196 if (!ctx->kfc.new_sb_created)
2197 cgroup_put(&ctx->root->cgrp);
2203 * Destroy a cgroup filesystem context.
2205 static void cgroup_fs_context_free(struct fs_context *fc)
2207 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2210 kfree(ctx->release_agent);
2211 put_cgroup_ns(ctx->ns);
2212 kernfs_free_fs_context(fc);
2216 static int cgroup_get_tree(struct fs_context *fc)
2218 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2221 WRITE_ONCE(cgrp_dfl_visible, true);
2222 cgroup_get_live(&cgrp_dfl_root.cgrp);
2223 ctx->root = &cgrp_dfl_root;
2225 ret = cgroup_do_get_tree(fc);
2227 apply_cgroup_root_flags(ctx->flags);
2231 static const struct fs_context_operations cgroup_fs_context_ops = {
2232 .free = cgroup_fs_context_free,
2233 .parse_param = cgroup2_parse_param,
2234 .get_tree = cgroup_get_tree,
2235 .reconfigure = cgroup_reconfigure,
2238 static const struct fs_context_operations cgroup1_fs_context_ops = {
2239 .free = cgroup_fs_context_free,
2240 .parse_param = cgroup1_parse_param,
2241 .get_tree = cgroup1_get_tree,
2242 .reconfigure = cgroup1_reconfigure,
2246 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2247 * we select the namespace we're going to use.
2249 static int cgroup_init_fs_context(struct fs_context *fc)
2251 struct cgroup_fs_context *ctx;
2253 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2257 ctx->ns = current->nsproxy->cgroup_ns;
2258 get_cgroup_ns(ctx->ns);
2259 fc->fs_private = &ctx->kfc;
2260 if (fc->fs_type == &cgroup2_fs_type)
2261 fc->ops = &cgroup_fs_context_ops;
2263 fc->ops = &cgroup1_fs_context_ops;
2264 put_user_ns(fc->user_ns);
2265 fc->user_ns = get_user_ns(ctx->ns->user_ns);
2268 #ifdef CONFIG_CGROUP_FAVOR_DYNMODS
2269 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
2274 static void cgroup_kill_sb(struct super_block *sb)
2276 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2277 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2280 * If @root doesn't have any children, start killing it.
2281 * This prevents new mounts by disabling percpu_ref_tryget_live().
2283 * And don't kill the default root.
2285 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2286 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) {
2287 cgroup_bpf_offline(&root->cgrp);
2288 percpu_ref_kill(&root->cgrp.self.refcnt);
2290 cgroup_put(&root->cgrp);
2294 struct file_system_type cgroup_fs_type = {
2296 .init_fs_context = cgroup_init_fs_context,
2297 .parameters = cgroup1_fs_parameters,
2298 .kill_sb = cgroup_kill_sb,
2299 .fs_flags = FS_USERNS_MOUNT,
2302 static struct file_system_type cgroup2_fs_type = {
2304 .init_fs_context = cgroup_init_fs_context,
2305 .parameters = cgroup2_fs_parameters,
2306 .kill_sb = cgroup_kill_sb,
2307 .fs_flags = FS_USERNS_MOUNT,
2310 #ifdef CONFIG_CPUSETS
2311 static const struct fs_context_operations cpuset_fs_context_ops = {
2312 .get_tree = cgroup1_get_tree,
2313 .free = cgroup_fs_context_free,
2317 * This is ugly, but preserves the userspace API for existing cpuset
2318 * users. If someone tries to mount the "cpuset" filesystem, we
2319 * silently switch it to mount "cgroup" instead
2321 static int cpuset_init_fs_context(struct fs_context *fc)
2323 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2324 struct cgroup_fs_context *ctx;
2327 err = cgroup_init_fs_context(fc);
2333 fc->ops = &cpuset_fs_context_ops;
2335 ctx = cgroup_fc2context(fc);
2336 ctx->subsys_mask = 1 << cpuset_cgrp_id;
2337 ctx->flags |= CGRP_ROOT_NOPREFIX;
2338 ctx->release_agent = agent;
2340 get_filesystem(&cgroup_fs_type);
2341 put_filesystem(fc->fs_type);
2342 fc->fs_type = &cgroup_fs_type;
2347 static struct file_system_type cpuset_fs_type = {
2349 .init_fs_context = cpuset_init_fs_context,
2350 .fs_flags = FS_USERNS_MOUNT,
2354 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2355 struct cgroup_namespace *ns)
2357 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2359 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2362 int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2363 struct cgroup_namespace *ns)
2368 spin_lock_irq(&css_set_lock);
2370 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2372 spin_unlock_irq(&css_set_lock);
2377 EXPORT_SYMBOL_GPL(cgroup_path_ns);
2380 * cgroup_attach_lock - Lock for ->attach()
2381 * @lock_threadgroup: whether to down_write cgroup_threadgroup_rwsem
2383 * cgroup migration sometimes needs to stabilize threadgroups against forks and
2384 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach()
2385 * implementations (e.g. cpuset), also need to disable CPU hotplug.
2386 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can
2387 * lead to deadlocks.
2389 * Bringing up a CPU may involve creating and destroying tasks which requires
2390 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside
2391 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while
2392 * write-locking threadgroup_rwsem, the locking order is reversed and we end up
2393 * waiting for an on-going CPU hotplug operation which in turn is waiting for
2394 * the threadgroup_rwsem to be released to create new tasks. For more details:
2396 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu
2398 * Resolve the situation by always acquiring cpus_read_lock() before optionally
2399 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that
2400 * CPU hotplug is disabled on entry.
2402 void cgroup_attach_lock(bool lock_threadgroup)
2405 if (lock_threadgroup)
2406 percpu_down_write(&cgroup_threadgroup_rwsem);
2410 * cgroup_attach_unlock - Undo cgroup_attach_lock()
2411 * @lock_threadgroup: whether to up_write cgroup_threadgroup_rwsem
2413 void cgroup_attach_unlock(bool lock_threadgroup)
2415 if (lock_threadgroup)
2416 percpu_up_write(&cgroup_threadgroup_rwsem);
2421 * cgroup_migrate_add_task - add a migration target task to a migration context
2422 * @task: target task
2423 * @mgctx: target migration context
2425 * Add @task, which is a migration target, to @mgctx->tset. This function
2426 * becomes noop if @task doesn't need to be migrated. @task's css_set
2427 * should have been added as a migration source and @task->cg_list will be
2428 * moved from the css_set's tasks list to mg_tasks one.
2430 static void cgroup_migrate_add_task(struct task_struct *task,
2431 struct cgroup_mgctx *mgctx)
2433 struct css_set *cset;
2435 lockdep_assert_held(&css_set_lock);
2437 /* @task either already exited or can't exit until the end */
2438 if (task->flags & PF_EXITING)
2441 /* cgroup_threadgroup_rwsem protects racing against forks */
2442 WARN_ON_ONCE(list_empty(&task->cg_list));
2444 cset = task_css_set(task);
2445 if (!cset->mg_src_cgrp)
2448 mgctx->tset.nr_tasks++;
2450 list_move_tail(&task->cg_list, &cset->mg_tasks);
2451 if (list_empty(&cset->mg_node))
2452 list_add_tail(&cset->mg_node,
2453 &mgctx->tset.src_csets);
2454 if (list_empty(&cset->mg_dst_cset->mg_node))
2455 list_add_tail(&cset->mg_dst_cset->mg_node,
2456 &mgctx->tset.dst_csets);
2460 * cgroup_taskset_first - reset taskset and return the first task
2461 * @tset: taskset of interest
2462 * @dst_cssp: output variable for the destination css
2464 * @tset iteration is initialized and the first task is returned.
2466 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2467 struct cgroup_subsys_state **dst_cssp)
2469 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2470 tset->cur_task = NULL;
2472 return cgroup_taskset_next(tset, dst_cssp);
2476 * cgroup_taskset_next - iterate to the next task in taskset
2477 * @tset: taskset of interest
2478 * @dst_cssp: output variable for the destination css
2480 * Return the next task in @tset. Iteration must have been initialized
2481 * with cgroup_taskset_first().
2483 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2484 struct cgroup_subsys_state **dst_cssp)
2486 struct css_set *cset = tset->cur_cset;
2487 struct task_struct *task = tset->cur_task;
2489 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) {
2491 task = list_first_entry(&cset->mg_tasks,
2492 struct task_struct, cg_list);
2494 task = list_next_entry(task, cg_list);
2496 if (&task->cg_list != &cset->mg_tasks) {
2497 tset->cur_cset = cset;
2498 tset->cur_task = task;
2501 * This function may be called both before and
2502 * after cgroup_taskset_migrate(). The two cases
2503 * can be distinguished by looking at whether @cset
2504 * has its ->mg_dst_cset set.
2506 if (cset->mg_dst_cset)
2507 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2509 *dst_cssp = cset->subsys[tset->ssid];
2514 cset = list_next_entry(cset, mg_node);
2522 * cgroup_migrate_execute - migrate a taskset
2523 * @mgctx: migration context
2525 * Migrate tasks in @mgctx as setup by migration preparation functions.
2526 * This function fails iff one of the ->can_attach callbacks fails and
2527 * guarantees that either all or none of the tasks in @mgctx are migrated.
2528 * @mgctx is consumed regardless of success.
2530 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2532 struct cgroup_taskset *tset = &mgctx->tset;
2533 struct cgroup_subsys *ss;
2534 struct task_struct *task, *tmp_task;
2535 struct css_set *cset, *tmp_cset;
2536 int ssid, failed_ssid, ret;
2538 /* check that we can legitimately attach to the cgroup */
2539 if (tset->nr_tasks) {
2540 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2541 if (ss->can_attach) {
2543 ret = ss->can_attach(tset);
2546 goto out_cancel_attach;
2549 } while_each_subsys_mask();
2553 * Now that we're guaranteed success, proceed to move all tasks to
2554 * the new cgroup. There are no failure cases after here, so this
2555 * is the commit point.
2557 spin_lock_irq(&css_set_lock);
2558 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2559 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2560 struct css_set *from_cset = task_css_set(task);
2561 struct css_set *to_cset = cset->mg_dst_cset;
2563 get_css_set(to_cset);
2564 to_cset->nr_tasks++;
2565 css_set_move_task(task, from_cset, to_cset, true);
2566 from_cset->nr_tasks--;
2568 * If the source or destination cgroup is frozen,
2569 * the task might require to change its state.
2571 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2573 put_css_set_locked(from_cset);
2577 spin_unlock_irq(&css_set_lock);
2580 * Migration is committed, all target tasks are now on dst_csets.
2581 * Nothing is sensitive to fork() after this point. Notify
2582 * controllers that migration is complete.
2584 tset->csets = &tset->dst_csets;
2586 if (tset->nr_tasks) {
2587 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2592 } while_each_subsys_mask();
2596 goto out_release_tset;
2599 if (tset->nr_tasks) {
2600 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2601 if (ssid == failed_ssid)
2603 if (ss->cancel_attach) {
2605 ss->cancel_attach(tset);
2607 } while_each_subsys_mask();
2610 spin_lock_irq(&css_set_lock);
2611 list_splice_init(&tset->dst_csets, &tset->src_csets);
2612 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2613 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2614 list_del_init(&cset->mg_node);
2616 spin_unlock_irq(&css_set_lock);
2619 * Re-initialize the cgroup_taskset structure in case it is reused
2620 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2624 tset->csets = &tset->src_csets;
2629 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2630 * @dst_cgrp: destination cgroup to test
2632 * On the default hierarchy, except for the mixable, (possible) thread root
2633 * and threaded cgroups, subtree_control must be zero for migration
2634 * destination cgroups with tasks so that child cgroups don't compete
2637 int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2639 /* v1 doesn't have any restriction */
2640 if (!cgroup_on_dfl(dst_cgrp))
2643 /* verify @dst_cgrp can host resources */
2644 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2648 * If @dst_cgrp is already or can become a thread root or is
2649 * threaded, it doesn't matter.
2651 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2654 /* apply no-internal-process constraint */
2655 if (dst_cgrp->subtree_control)
2662 * cgroup_migrate_finish - cleanup after attach
2663 * @mgctx: migration context
2665 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2666 * those functions for details.
2668 void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2670 struct css_set *cset, *tmp_cset;
2672 lockdep_assert_held(&cgroup_mutex);
2674 spin_lock_irq(&css_set_lock);
2676 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets,
2677 mg_src_preload_node) {
2678 cset->mg_src_cgrp = NULL;
2679 cset->mg_dst_cgrp = NULL;
2680 cset->mg_dst_cset = NULL;
2681 list_del_init(&cset->mg_src_preload_node);
2682 put_css_set_locked(cset);
2685 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets,
2686 mg_dst_preload_node) {
2687 cset->mg_src_cgrp = NULL;
2688 cset->mg_dst_cgrp = NULL;
2689 cset->mg_dst_cset = NULL;
2690 list_del_init(&cset->mg_dst_preload_node);
2691 put_css_set_locked(cset);
2694 spin_unlock_irq(&css_set_lock);
2698 * cgroup_migrate_add_src - add a migration source css_set
2699 * @src_cset: the source css_set to add
2700 * @dst_cgrp: the destination cgroup
2701 * @mgctx: migration context
2703 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2704 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2705 * up by cgroup_migrate_finish().
2707 * This function may be called without holding cgroup_threadgroup_rwsem
2708 * even if the target is a process. Threads may be created and destroyed
2709 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2710 * into play and the preloaded css_sets are guaranteed to cover all
2713 void cgroup_migrate_add_src(struct css_set *src_cset,
2714 struct cgroup *dst_cgrp,
2715 struct cgroup_mgctx *mgctx)
2717 struct cgroup *src_cgrp;
2719 lockdep_assert_held(&cgroup_mutex);
2720 lockdep_assert_held(&css_set_lock);
2723 * If ->dead, @src_set is associated with one or more dead cgroups
2724 * and doesn't contain any migratable tasks. Ignore it early so
2725 * that the rest of migration path doesn't get confused by it.
2730 if (!list_empty(&src_cset->mg_src_preload_node))
2733 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2735 WARN_ON(src_cset->mg_src_cgrp);
2736 WARN_ON(src_cset->mg_dst_cgrp);
2737 WARN_ON(!list_empty(&src_cset->mg_tasks));
2738 WARN_ON(!list_empty(&src_cset->mg_node));
2740 src_cset->mg_src_cgrp = src_cgrp;
2741 src_cset->mg_dst_cgrp = dst_cgrp;
2742 get_css_set(src_cset);
2743 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets);
2747 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2748 * @mgctx: migration context
2750 * Tasks are about to be moved and all the source css_sets have been
2751 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2752 * pins all destination css_sets, links each to its source, and append them
2753 * to @mgctx->preloaded_dst_csets.
2755 * This function must be called after cgroup_migrate_add_src() has been
2756 * called on each migration source css_set. After migration is performed
2757 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2760 int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2762 struct css_set *src_cset, *tmp_cset;
2764 lockdep_assert_held(&cgroup_mutex);
2766 /* look up the dst cset for each src cset and link it to src */
2767 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2768 mg_src_preload_node) {
2769 struct css_set *dst_cset;
2770 struct cgroup_subsys *ss;
2773 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2777 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2780 * If src cset equals dst, it's noop. Drop the src.
2781 * cgroup_migrate() will skip the cset too. Note that we
2782 * can't handle src == dst as some nodes are used by both.
2784 if (src_cset == dst_cset) {
2785 src_cset->mg_src_cgrp = NULL;
2786 src_cset->mg_dst_cgrp = NULL;
2787 list_del_init(&src_cset->mg_src_preload_node);
2788 put_css_set(src_cset);
2789 put_css_set(dst_cset);
2793 src_cset->mg_dst_cset = dst_cset;
2795 if (list_empty(&dst_cset->mg_dst_preload_node))
2796 list_add_tail(&dst_cset->mg_dst_preload_node,
2797 &mgctx->preloaded_dst_csets);
2799 put_css_set(dst_cset);
2801 for_each_subsys(ss, ssid)
2802 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2803 mgctx->ss_mask |= 1 << ssid;
2810 * cgroup_migrate - migrate a process or task to a cgroup
2811 * @leader: the leader of the process or the task to migrate
2812 * @threadgroup: whether @leader points to the whole process or a single task
2813 * @mgctx: migration context
2815 * Migrate a process or task denoted by @leader. If migrating a process,
2816 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2817 * responsible for invoking cgroup_migrate_add_src() and
2818 * cgroup_migrate_prepare_dst() on the targets before invoking this
2819 * function and following up with cgroup_migrate_finish().
2821 * As long as a controller's ->can_attach() doesn't fail, this function is
2822 * guaranteed to succeed. This means that, excluding ->can_attach()
2823 * failure, when migrating multiple targets, the success or failure can be
2824 * decided for all targets by invoking group_migrate_prepare_dst() before
2825 * actually starting migrating.
2827 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2828 struct cgroup_mgctx *mgctx)
2830 struct task_struct *task;
2833 * The following thread iteration should be inside an RCU critical
2834 * section to prevent tasks from being freed while taking the snapshot.
2835 * spin_lock_irq() implies RCU critical section here.
2837 spin_lock_irq(&css_set_lock);
2840 cgroup_migrate_add_task(task, mgctx);
2843 } while_each_thread(leader, task);
2844 spin_unlock_irq(&css_set_lock);
2846 return cgroup_migrate_execute(mgctx);
2850 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2851 * @dst_cgrp: the cgroup to attach to
2852 * @leader: the task or the leader of the threadgroup to be attached
2853 * @threadgroup: attach the whole threadgroup?
2855 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2857 int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2860 DEFINE_CGROUP_MGCTX(mgctx);
2861 struct task_struct *task;
2864 /* look up all src csets */
2865 spin_lock_irq(&css_set_lock);
2869 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2872 } while_each_thread(leader, task);
2874 spin_unlock_irq(&css_set_lock);
2876 /* prepare dst csets and commit */
2877 ret = cgroup_migrate_prepare_dst(&mgctx);
2879 ret = cgroup_migrate(leader, threadgroup, &mgctx);
2881 cgroup_migrate_finish(&mgctx);
2884 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2889 struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2890 bool *threadgroup_locked)
2892 struct task_struct *tsk;
2895 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2896 return ERR_PTR(-EINVAL);
2899 * If we migrate a single thread, we don't care about threadgroup
2900 * stability. If the thread is `current`, it won't exit(2) under our
2901 * hands or change PID through exec(2). We exclude
2902 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2903 * callers by cgroup_mutex.
2904 * Therefore, we can skip the global lock.
2906 lockdep_assert_held(&cgroup_mutex);
2907 *threadgroup_locked = pid || threadgroup;
2908 cgroup_attach_lock(*threadgroup_locked);
2912 tsk = find_task_by_vpid(pid);
2914 tsk = ERR_PTR(-ESRCH);
2915 goto out_unlock_threadgroup;
2922 tsk = tsk->group_leader;
2925 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2926 * If userland migrates such a kthread to a non-root cgroup, it can
2927 * become trapped in a cpuset, or RT kthread may be born in a
2928 * cgroup with no rt_runtime allocated. Just say no.
2930 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2931 tsk = ERR_PTR(-EINVAL);
2932 goto out_unlock_threadgroup;
2935 get_task_struct(tsk);
2936 goto out_unlock_rcu;
2938 out_unlock_threadgroup:
2939 cgroup_attach_unlock(*threadgroup_locked);
2940 *threadgroup_locked = false;
2946 void cgroup_procs_write_finish(struct task_struct *task, bool threadgroup_locked)
2948 struct cgroup_subsys *ss;
2951 /* release reference from cgroup_procs_write_start() */
2952 put_task_struct(task);
2954 cgroup_attach_unlock(threadgroup_locked);
2956 for_each_subsys(ss, ssid)
2957 if (ss->post_attach)
2961 static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2963 struct cgroup_subsys *ss;
2964 bool printed = false;
2967 do_each_subsys_mask(ss, ssid, ss_mask) {
2970 seq_puts(seq, ss->name);
2972 } while_each_subsys_mask();
2974 seq_putc(seq, '\n');
2977 /* show controllers which are enabled from the parent */
2978 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2980 struct cgroup *cgrp = seq_css(seq)->cgroup;
2982 cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2986 /* show controllers which are enabled for a given cgroup's children */
2987 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2989 struct cgroup *cgrp = seq_css(seq)->cgroup;
2991 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2996 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2997 * @cgrp: root of the subtree to update csses for
2999 * @cgrp's control masks have changed and its subtree's css associations
3000 * need to be updated accordingly. This function looks up all css_sets
3001 * which are attached to the subtree, creates the matching updated css_sets
3002 * and migrates the tasks to the new ones.
3004 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
3006 DEFINE_CGROUP_MGCTX(mgctx);
3007 struct cgroup_subsys_state *d_css;
3008 struct cgroup *dsct;
3009 struct css_set *src_cset;
3013 lockdep_assert_held(&cgroup_mutex);
3015 /* look up all csses currently attached to @cgrp's subtree */
3016 spin_lock_irq(&css_set_lock);
3017 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3018 struct cgrp_cset_link *link;
3021 * As cgroup_update_dfl_csses() is only called by
3022 * cgroup_apply_control(). The csses associated with the
3023 * given cgrp will not be affected by changes made to
3024 * its subtree_control file. We can skip them.
3029 list_for_each_entry(link, &dsct->cset_links, cset_link)
3030 cgroup_migrate_add_src(link->cset, dsct, &mgctx);
3032 spin_unlock_irq(&css_set_lock);
3035 * We need to write-lock threadgroup_rwsem while migrating tasks.
3036 * However, if there are no source csets for @cgrp, changing its
3037 * controllers isn't gonna produce any task migrations and the
3038 * write-locking can be skipped safely.
3040 has_tasks = !list_empty(&mgctx.preloaded_src_csets);
3041 cgroup_attach_lock(has_tasks);
3043 /* NULL dst indicates self on default hierarchy */
3044 ret = cgroup_migrate_prepare_dst(&mgctx);
3048 spin_lock_irq(&css_set_lock);
3049 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets,
3050 mg_src_preload_node) {
3051 struct task_struct *task, *ntask;
3053 /* all tasks in src_csets need to be migrated */
3054 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
3055 cgroup_migrate_add_task(task, &mgctx);
3057 spin_unlock_irq(&css_set_lock);
3059 ret = cgroup_migrate_execute(&mgctx);
3061 cgroup_migrate_finish(&mgctx);
3062 cgroup_attach_unlock(has_tasks);
3067 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3068 * @cgrp: root of the target subtree
3070 * Because css offlining is asynchronous, userland may try to re-enable a
3071 * controller while the previous css is still around. This function grabs
3072 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3074 void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
3075 __acquires(&cgroup_mutex)
3077 struct cgroup *dsct;
3078 struct cgroup_subsys_state *d_css;
3079 struct cgroup_subsys *ss;
3085 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3086 for_each_subsys(ss, ssid) {
3087 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3090 if (!css || !percpu_ref_is_dying(&css->refcnt))
3093 cgroup_get_live(dsct);
3094 prepare_to_wait(&dsct->offline_waitq, &wait,
3095 TASK_UNINTERRUPTIBLE);
3099 finish_wait(&dsct->offline_waitq, &wait);
3108 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3109 * @cgrp: root of the target subtree
3111 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3112 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3115 static void cgroup_save_control(struct cgroup *cgrp)
3117 struct cgroup *dsct;
3118 struct cgroup_subsys_state *d_css;
3120 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3121 dsct->old_subtree_control = dsct->subtree_control;
3122 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3123 dsct->old_dom_cgrp = dsct->dom_cgrp;
3128 * cgroup_propagate_control - refresh control masks of a subtree
3129 * @cgrp: root of the target subtree
3131 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3132 * ->subtree_control and propagate controller availability through the
3133 * subtree so that descendants don't have unavailable controllers enabled.
3135 static void cgroup_propagate_control(struct cgroup *cgrp)
3137 struct cgroup *dsct;
3138 struct cgroup_subsys_state *d_css;
3140 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3141 dsct->subtree_control &= cgroup_control(dsct);
3142 dsct->subtree_ss_mask =
3143 cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3144 cgroup_ss_mask(dsct));
3149 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3150 * @cgrp: root of the target subtree
3152 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3153 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3156 static void cgroup_restore_control(struct cgroup *cgrp)
3158 struct cgroup *dsct;
3159 struct cgroup_subsys_state *d_css;
3161 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3162 dsct->subtree_control = dsct->old_subtree_control;
3163 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3164 dsct->dom_cgrp = dsct->old_dom_cgrp;
3168 static bool css_visible(struct cgroup_subsys_state *css)
3170 struct cgroup_subsys *ss = css->ss;
3171 struct cgroup *cgrp = css->cgroup;
3173 if (cgroup_control(cgrp) & (1 << ss->id))
3175 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3177 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3181 * cgroup_apply_control_enable - enable or show csses according to control
3182 * @cgrp: root of the target subtree
3184 * Walk @cgrp's subtree and create new csses or make the existing ones
3185 * visible. A css is created invisible if it's being implicitly enabled
3186 * through dependency. An invisible css is made visible when the userland
3187 * explicitly enables it.
3189 * Returns 0 on success, -errno on failure. On failure, csses which have
3190 * been processed already aren't cleaned up. The caller is responsible for
3191 * cleaning up with cgroup_apply_control_disable().
3193 static int cgroup_apply_control_enable(struct cgroup *cgrp)
3195 struct cgroup *dsct;
3196 struct cgroup_subsys_state *d_css;
3197 struct cgroup_subsys *ss;
3200 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3201 for_each_subsys(ss, ssid) {
3202 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3204 if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3208 css = css_create(dsct, ss);
3210 return PTR_ERR(css);
3213 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3215 if (css_visible(css)) {
3216 ret = css_populate_dir(css);
3227 * cgroup_apply_control_disable - kill or hide csses according to control
3228 * @cgrp: root of the target subtree
3230 * Walk @cgrp's subtree and kill and hide csses so that they match
3231 * cgroup_ss_mask() and cgroup_visible_mask().
3233 * A css is hidden when the userland requests it to be disabled while other
3234 * subsystems are still depending on it. The css must not actively control
3235 * resources and be in the vanilla state if it's made visible again later.
3236 * Controllers which may be depended upon should provide ->css_reset() for
3239 static void cgroup_apply_control_disable(struct cgroup *cgrp)
3241 struct cgroup *dsct;
3242 struct cgroup_subsys_state *d_css;
3243 struct cgroup_subsys *ss;
3246 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3247 for_each_subsys(ss, ssid) {
3248 struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3253 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3256 !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3258 } else if (!css_visible(css)) {
3268 * cgroup_apply_control - apply control mask updates to the subtree
3269 * @cgrp: root of the target subtree
3271 * subsystems can be enabled and disabled in a subtree using the following
3274 * 1. Call cgroup_save_control() to stash the current state.
3275 * 2. Update ->subtree_control masks in the subtree as desired.
3276 * 3. Call cgroup_apply_control() to apply the changes.
3277 * 4. Optionally perform other related operations.
3278 * 5. Call cgroup_finalize_control() to finish up.
3280 * This function implements step 3 and propagates the mask changes
3281 * throughout @cgrp's subtree, updates csses accordingly and perform
3282 * process migrations.
3284 static int cgroup_apply_control(struct cgroup *cgrp)
3288 cgroup_propagate_control(cgrp);
3290 ret = cgroup_apply_control_enable(cgrp);
3295 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3296 * making the following cgroup_update_dfl_csses() properly update
3297 * css associations of all tasks in the subtree.
3299 return cgroup_update_dfl_csses(cgrp);
3303 * cgroup_finalize_control - finalize control mask update
3304 * @cgrp: root of the target subtree
3305 * @ret: the result of the update
3307 * Finalize control mask update. See cgroup_apply_control() for more info.
3309 static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3312 cgroup_restore_control(cgrp);
3313 cgroup_propagate_control(cgrp);
3316 cgroup_apply_control_disable(cgrp);
3319 static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3321 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3323 /* if nothing is getting enabled, nothing to worry about */
3327 /* can @cgrp host any resources? */
3328 if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3331 /* mixables don't care */
3332 if (cgroup_is_mixable(cgrp))
3335 if (domain_enable) {
3336 /* can't enable domain controllers inside a thread subtree */
3337 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3341 * Threaded controllers can handle internal competitions
3342 * and are always allowed inside a (prospective) thread
3345 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3350 * Controllers can't be enabled for a cgroup with tasks to avoid
3351 * child cgroups competing against tasks.
3353 if (cgroup_has_tasks(cgrp))
3359 /* change the enabled child controllers for a cgroup in the default hierarchy */
3360 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3361 char *buf, size_t nbytes,
3364 u16 enable = 0, disable = 0;
3365 struct cgroup *cgrp, *child;
3366 struct cgroup_subsys *ss;
3371 * Parse input - space separated list of subsystem names prefixed
3372 * with either + or -.
3374 buf = strstrip(buf);
3375 while ((tok = strsep(&buf, " "))) {
3378 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3379 if (!cgroup_ssid_enabled(ssid) ||
3380 strcmp(tok + 1, ss->name))
3384 enable |= 1 << ssid;
3385 disable &= ~(1 << ssid);
3386 } else if (*tok == '-') {
3387 disable |= 1 << ssid;
3388 enable &= ~(1 << ssid);
3393 } while_each_subsys_mask();
3394 if (ssid == CGROUP_SUBSYS_COUNT)
3398 cgrp = cgroup_kn_lock_live(of->kn, true);
3402 for_each_subsys(ss, ssid) {
3403 if (enable & (1 << ssid)) {
3404 if (cgrp->subtree_control & (1 << ssid)) {
3405 enable &= ~(1 << ssid);
3409 if (!(cgroup_control(cgrp) & (1 << ssid))) {
3413 } else if (disable & (1 << ssid)) {
3414 if (!(cgrp->subtree_control & (1 << ssid))) {
3415 disable &= ~(1 << ssid);
3419 /* a child has it enabled? */
3420 cgroup_for_each_live_child(child, cgrp) {
3421 if (child->subtree_control & (1 << ssid)) {
3429 if (!enable && !disable) {
3434 ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3438 /* save and update control masks and prepare csses */
3439 cgroup_save_control(cgrp);
3441 cgrp->subtree_control |= enable;
3442 cgrp->subtree_control &= ~disable;
3444 ret = cgroup_apply_control(cgrp);
3445 cgroup_finalize_control(cgrp, ret);
3449 kernfs_activate(cgrp->kn);
3451 cgroup_kn_unlock(of->kn);
3452 return ret ?: nbytes;
3456 * cgroup_enable_threaded - make @cgrp threaded
3457 * @cgrp: the target cgroup
3459 * Called when "threaded" is written to the cgroup.type interface file and
3460 * tries to make @cgrp threaded and join the parent's resource domain.
3461 * This function is never called on the root cgroup as cgroup.type doesn't
3464 static int cgroup_enable_threaded(struct cgroup *cgrp)
3466 struct cgroup *parent = cgroup_parent(cgrp);
3467 struct cgroup *dom_cgrp = parent->dom_cgrp;
3468 struct cgroup *dsct;
3469 struct cgroup_subsys_state *d_css;
3472 lockdep_assert_held(&cgroup_mutex);
3474 /* noop if already threaded */
3475 if (cgroup_is_threaded(cgrp))
3479 * If @cgroup is populated or has domain controllers enabled, it
3480 * can't be switched. While the below cgroup_can_be_thread_root()
3481 * test can catch the same conditions, that's only when @parent is
3482 * not mixable, so let's check it explicitly.
3484 if (cgroup_is_populated(cgrp) ||
3485 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3488 /* we're joining the parent's domain, ensure its validity */
3489 if (!cgroup_is_valid_domain(dom_cgrp) ||
3490 !cgroup_can_be_thread_root(dom_cgrp))
3494 * The following shouldn't cause actual migrations and should
3497 cgroup_save_control(cgrp);
3499 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3500 if (dsct == cgrp || cgroup_is_threaded(dsct))
3501 dsct->dom_cgrp = dom_cgrp;
3503 ret = cgroup_apply_control(cgrp);
3505 parent->nr_threaded_children++;
3507 cgroup_finalize_control(cgrp, ret);
3511 static int cgroup_type_show(struct seq_file *seq, void *v)
3513 struct cgroup *cgrp = seq_css(seq)->cgroup;
3515 if (cgroup_is_threaded(cgrp))
3516 seq_puts(seq, "threaded\n");
3517 else if (!cgroup_is_valid_domain(cgrp))
3518 seq_puts(seq, "domain invalid\n");
3519 else if (cgroup_is_thread_root(cgrp))
3520 seq_puts(seq, "domain threaded\n");
3522 seq_puts(seq, "domain\n");
3527 static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3528 size_t nbytes, loff_t off)
3530 struct cgroup *cgrp;
3533 /* only switching to threaded mode is supported */
3534 if (strcmp(strstrip(buf), "threaded"))
3537 /* drain dying csses before we re-apply (threaded) subtree control */
3538 cgrp = cgroup_kn_lock_live(of->kn, true);
3542 /* threaded can only be enabled */
3543 ret = cgroup_enable_threaded(cgrp);
3545 cgroup_kn_unlock(of->kn);
3546 return ret ?: nbytes;
3549 static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3551 struct cgroup *cgrp = seq_css(seq)->cgroup;
3552 int descendants = READ_ONCE(cgrp->max_descendants);
3554 if (descendants == INT_MAX)
3555 seq_puts(seq, "max\n");
3557 seq_printf(seq, "%d\n", descendants);
3562 static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3563 char *buf, size_t nbytes, loff_t off)
3565 struct cgroup *cgrp;
3569 buf = strstrip(buf);
3570 if (!strcmp(buf, "max")) {
3571 descendants = INT_MAX;
3573 ret = kstrtoint(buf, 0, &descendants);
3578 if (descendants < 0)
3581 cgrp = cgroup_kn_lock_live(of->kn, false);
3585 cgrp->max_descendants = descendants;
3587 cgroup_kn_unlock(of->kn);
3592 static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3594 struct cgroup *cgrp = seq_css(seq)->cgroup;
3595 int depth = READ_ONCE(cgrp->max_depth);
3597 if (depth == INT_MAX)
3598 seq_puts(seq, "max\n");
3600 seq_printf(seq, "%d\n", depth);
3605 static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3606 char *buf, size_t nbytes, loff_t off)
3608 struct cgroup *cgrp;
3612 buf = strstrip(buf);
3613 if (!strcmp(buf, "max")) {
3616 ret = kstrtoint(buf, 0, &depth);
3624 cgrp = cgroup_kn_lock_live(of->kn, false);
3628 cgrp->max_depth = depth;
3630 cgroup_kn_unlock(of->kn);
3635 static int cgroup_events_show(struct seq_file *seq, void *v)
3637 struct cgroup *cgrp = seq_css(seq)->cgroup;
3639 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3640 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3645 static int cgroup_stat_show(struct seq_file *seq, void *v)
3647 struct cgroup *cgroup = seq_css(seq)->cgroup;
3649 seq_printf(seq, "nr_descendants %d\n",
3650 cgroup->nr_descendants);
3651 seq_printf(seq, "nr_dying_descendants %d\n",
3652 cgroup->nr_dying_descendants);
3657 static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3658 struct cgroup *cgrp, int ssid)
3660 struct cgroup_subsys *ss = cgroup_subsys[ssid];
3661 struct cgroup_subsys_state *css;
3664 if (!ss->css_extra_stat_show)
3667 css = cgroup_tryget_css(cgrp, ss);
3671 ret = ss->css_extra_stat_show(seq, css);
3676 static int cpu_stat_show(struct seq_file *seq, void *v)
3678 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3681 cgroup_base_stat_cputime_show(seq);
3682 #ifdef CONFIG_CGROUP_SCHED
3683 ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3689 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3691 struct cgroup *cgrp = seq_css(seq)->cgroup;
3692 struct psi_group *psi = cgroup_psi(cgrp);
3694 return psi_show(seq, psi, PSI_IO);
3696 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3698 struct cgroup *cgrp = seq_css(seq)->cgroup;
3699 struct psi_group *psi = cgroup_psi(cgrp);
3701 return psi_show(seq, psi, PSI_MEM);
3703 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3705 struct cgroup *cgrp = seq_css(seq)->cgroup;
3706 struct psi_group *psi = cgroup_psi(cgrp);
3708 return psi_show(seq, psi, PSI_CPU);
3711 static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
3712 size_t nbytes, enum psi_res res)
3714 struct cgroup_file_ctx *ctx = of->priv;
3715 struct psi_trigger *new;
3716 struct cgroup *cgrp;
3717 struct psi_group *psi;
3719 cgrp = cgroup_kn_lock_live(of->kn, false);
3724 cgroup_kn_unlock(of->kn);
3726 /* Allow only one trigger per file descriptor */
3727 if (ctx->psi.trigger) {
3732 psi = cgroup_psi(cgrp);
3733 new = psi_trigger_create(psi, buf, res, of->file, of);
3736 return PTR_ERR(new);
3739 smp_store_release(&ctx->psi.trigger, new);
3745 static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3746 char *buf, size_t nbytes,
3749 return pressure_write(of, buf, nbytes, PSI_IO);
3752 static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3753 char *buf, size_t nbytes,
3756 return pressure_write(of, buf, nbytes, PSI_MEM);
3759 static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3760 char *buf, size_t nbytes,
3763 return pressure_write(of, buf, nbytes, PSI_CPU);
3766 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3767 static int cgroup_irq_pressure_show(struct seq_file *seq, void *v)
3769 struct cgroup *cgrp = seq_css(seq)->cgroup;
3770 struct psi_group *psi = cgroup_psi(cgrp);
3772 return psi_show(seq, psi, PSI_IRQ);
3775 static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of,
3776 char *buf, size_t nbytes,
3779 return pressure_write(of, buf, nbytes, PSI_IRQ);
3783 static int cgroup_pressure_show(struct seq_file *seq, void *v)
3785 struct cgroup *cgrp = seq_css(seq)->cgroup;
3786 struct psi_group *psi = cgroup_psi(cgrp);
3788 seq_printf(seq, "%d\n", psi->enabled);
3793 static ssize_t cgroup_pressure_write(struct kernfs_open_file *of,
3794 char *buf, size_t nbytes,
3799 struct cgroup *cgrp;
3800 struct psi_group *psi;
3802 ret = kstrtoint(strstrip(buf), 0, &enable);
3806 if (enable < 0 || enable > 1)
3809 cgrp = cgroup_kn_lock_live(of->kn, false);
3813 psi = cgroup_psi(cgrp);
3814 if (psi->enabled != enable) {
3817 /* show or hide {cpu,memory,io,irq}.pressure files */
3818 for (i = 0; i < NR_PSI_RESOURCES; i++)
3819 cgroup_file_show(&cgrp->psi_files[i], enable);
3821 psi->enabled = enable;
3823 psi_cgroup_restart(psi);
3826 cgroup_kn_unlock(of->kn);
3831 static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3834 struct cgroup_file_ctx *ctx = of->priv;
3836 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt);
3839 static int cgroup_pressure_open(struct kernfs_open_file *of)
3841 if (of->file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
3847 static void cgroup_pressure_release(struct kernfs_open_file *of)
3849 struct cgroup_file_ctx *ctx = of->priv;
3851 psi_trigger_destroy(ctx->psi.trigger);
3854 bool cgroup_psi_enabled(void)
3856 if (static_branch_likely(&psi_disabled))
3859 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3862 #else /* CONFIG_PSI */
3863 bool cgroup_psi_enabled(void)
3868 #endif /* CONFIG_PSI */
3870 static int cgroup_freeze_show(struct seq_file *seq, void *v)
3872 struct cgroup *cgrp = seq_css(seq)->cgroup;
3874 seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3879 static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3880 char *buf, size_t nbytes, loff_t off)
3882 struct cgroup *cgrp;
3886 ret = kstrtoint(strstrip(buf), 0, &freeze);
3890 if (freeze < 0 || freeze > 1)
3893 cgrp = cgroup_kn_lock_live(of->kn, false);
3897 cgroup_freeze(cgrp, freeze);
3899 cgroup_kn_unlock(of->kn);
3904 static void __cgroup_kill(struct cgroup *cgrp)
3906 struct css_task_iter it;
3907 struct task_struct *task;
3909 lockdep_assert_held(&cgroup_mutex);
3911 spin_lock_irq(&css_set_lock);
3912 set_bit(CGRP_KILL, &cgrp->flags);
3913 spin_unlock_irq(&css_set_lock);
3915 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3916 while ((task = css_task_iter_next(&it))) {
3917 /* Ignore kernel threads here. */
3918 if (task->flags & PF_KTHREAD)
3921 /* Skip tasks that are already dying. */
3922 if (__fatal_signal_pending(task))
3925 send_sig(SIGKILL, task, 0);
3927 css_task_iter_end(&it);
3929 spin_lock_irq(&css_set_lock);
3930 clear_bit(CGRP_KILL, &cgrp->flags);
3931 spin_unlock_irq(&css_set_lock);
3934 static void cgroup_kill(struct cgroup *cgrp)
3936 struct cgroup_subsys_state *css;
3937 struct cgroup *dsct;
3939 lockdep_assert_held(&cgroup_mutex);
3941 cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3942 __cgroup_kill(dsct);
3945 static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3946 size_t nbytes, loff_t off)
3950 struct cgroup *cgrp;
3952 ret = kstrtoint(strstrip(buf), 0, &kill);
3959 cgrp = cgroup_kn_lock_live(of->kn, false);
3964 * Killing is a process directed operation, i.e. the whole thread-group
3965 * is taken down so act like we do for cgroup.procs and only make this
3966 * writable in non-threaded cgroups.
3968 if (cgroup_is_threaded(cgrp))
3973 cgroup_kn_unlock(of->kn);
3975 return ret ?: nbytes;
3978 static int cgroup_file_open(struct kernfs_open_file *of)
3980 struct cftype *cft = of_cft(of);
3981 struct cgroup_file_ctx *ctx;
3984 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
3988 ctx->ns = current->nsproxy->cgroup_ns;
3989 get_cgroup_ns(ctx->ns);
3995 ret = cft->open(of);
3997 put_cgroup_ns(ctx->ns);
4003 static void cgroup_file_release(struct kernfs_open_file *of)
4005 struct cftype *cft = of_cft(of);
4006 struct cgroup_file_ctx *ctx = of->priv;
4010 put_cgroup_ns(ctx->ns);
4014 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
4015 size_t nbytes, loff_t off)
4017 struct cgroup_file_ctx *ctx = of->priv;
4018 struct cgroup *cgrp = of->kn->parent->priv;
4019 struct cftype *cft = of_cft(of);
4020 struct cgroup_subsys_state *css;
4027 * If namespaces are delegation boundaries, disallow writes to
4028 * files in an non-init namespace root from inside the namespace
4029 * except for the files explicitly marked delegatable -
4030 * cgroup.procs and cgroup.subtree_control.
4032 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
4033 !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
4034 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp)
4038 return cft->write(of, buf, nbytes, off);
4041 * kernfs guarantees that a file isn't deleted with operations in
4042 * flight, which means that the matching css is and stays alive and
4043 * doesn't need to be pinned. The RCU locking is not necessary
4044 * either. It's just for the convenience of using cgroup_css().
4047 css = cgroup_css(cgrp, cft->ss);
4050 if (cft->write_u64) {
4051 unsigned long long v;
4052 ret = kstrtoull(buf, 0, &v);
4054 ret = cft->write_u64(css, cft, v);
4055 } else if (cft->write_s64) {
4057 ret = kstrtoll(buf, 0, &v);
4059 ret = cft->write_s64(css, cft, v);
4064 return ret ?: nbytes;
4067 static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
4069 struct cftype *cft = of_cft(of);
4072 return cft->poll(of, pt);
4074 return kernfs_generic_poll(of, pt);
4077 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
4079 return seq_cft(seq)->seq_start(seq, ppos);
4082 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
4084 return seq_cft(seq)->seq_next(seq, v, ppos);
4087 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
4089 if (seq_cft(seq)->seq_stop)
4090 seq_cft(seq)->seq_stop(seq, v);
4093 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
4095 struct cftype *cft = seq_cft(m);
4096 struct cgroup_subsys_state *css = seq_css(m);
4099 return cft->seq_show(m, arg);
4102 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
4103 else if (cft->read_s64)
4104 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
4110 static struct kernfs_ops cgroup_kf_single_ops = {
4111 .atomic_write_len = PAGE_SIZE,
4112 .open = cgroup_file_open,
4113 .release = cgroup_file_release,
4114 .write = cgroup_file_write,
4115 .poll = cgroup_file_poll,
4116 .seq_show = cgroup_seqfile_show,
4119 static struct kernfs_ops cgroup_kf_ops = {
4120 .atomic_write_len = PAGE_SIZE,
4121 .open = cgroup_file_open,
4122 .release = cgroup_file_release,
4123 .write = cgroup_file_write,
4124 .poll = cgroup_file_poll,
4125 .seq_start = cgroup_seqfile_start,
4126 .seq_next = cgroup_seqfile_next,
4127 .seq_stop = cgroup_seqfile_stop,
4128 .seq_show = cgroup_seqfile_show,
4131 /* set uid and gid of cgroup dirs and files to that of the creator */
4132 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
4134 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
4135 .ia_uid = current_fsuid(),
4136 .ia_gid = current_fsgid(), };
4138 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
4139 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
4142 return kernfs_setattr(kn, &iattr);
4145 static void cgroup_file_notify_timer(struct timer_list *timer)
4147 cgroup_file_notify(container_of(timer, struct cgroup_file,
4151 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
4154 char name[CGROUP_FILE_NAME_MAX];
4155 struct kernfs_node *kn;
4156 struct lock_class_key *key = NULL;
4159 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4160 key = &cft->lockdep_key;
4162 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
4163 cgroup_file_mode(cft),
4164 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
4165 0, cft->kf_ops, cft,
4170 ret = cgroup_kn_set_ugid(kn);
4176 if (cft->file_offset) {
4177 struct cgroup_file *cfile = (void *)css + cft->file_offset;
4179 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
4181 spin_lock_irq(&cgroup_file_kn_lock);
4183 spin_unlock_irq(&cgroup_file_kn_lock);
4190 * cgroup_addrm_files - add or remove files to a cgroup directory
4191 * @css: the target css
4192 * @cgrp: the target cgroup (usually css->cgroup)
4193 * @cfts: array of cftypes to be added
4194 * @is_add: whether to add or remove
4196 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4197 * For removals, this function never fails.
4199 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
4200 struct cgroup *cgrp, struct cftype cfts[],
4203 struct cftype *cft, *cft_end = NULL;
4206 lockdep_assert_held(&cgroup_mutex);
4209 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
4210 /* does cft->flags tell us to skip this file on @cgrp? */
4211 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
4213 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
4215 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
4217 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
4219 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
4222 ret = cgroup_add_file(css, cgrp, cft);
4224 pr_warn("%s: failed to add %s, err=%d\n",
4225 __func__, cft->name, ret);
4231 cgroup_rm_file(cgrp, cft);
4237 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4239 struct cgroup_subsys *ss = cfts[0].ss;
4240 struct cgroup *root = &ss->root->cgrp;
4241 struct cgroup_subsys_state *css;
4244 lockdep_assert_held(&cgroup_mutex);
4246 /* add/rm files for all cgroups created before */
4247 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4248 struct cgroup *cgrp = css->cgroup;
4250 if (!(css->flags & CSS_VISIBLE))
4253 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4259 kernfs_activate(root->kn);
4263 static void cgroup_exit_cftypes(struct cftype *cfts)
4267 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4268 /* free copy for custom atomic_write_len, see init_cftypes() */
4269 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4274 /* revert flags set by cgroup core while adding @cfts */
4275 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL |
4280 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4285 for (cft = cfts; cft->name[0] != '\0'; cft++) {
4286 struct kernfs_ops *kf_ops;
4288 WARN_ON(cft->ss || cft->kf_ops);
4290 if (cft->flags & __CFTYPE_ADDED) {
4296 kf_ops = &cgroup_kf_ops;
4298 kf_ops = &cgroup_kf_single_ops;
4301 * Ugh... if @cft wants a custom max_write_len, we need to
4302 * make a copy of kf_ops to set its atomic_write_len.
4304 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4305 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4310 kf_ops->atomic_write_len = cft->max_write_len;
4313 cft->kf_ops = kf_ops;
4315 cft->flags |= __CFTYPE_ADDED;
4319 cgroup_exit_cftypes(cfts);
4323 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4325 lockdep_assert_held(&cgroup_mutex);
4327 list_del(&cfts->node);
4328 cgroup_apply_cftypes(cfts, false);
4329 cgroup_exit_cftypes(cfts);
4334 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4335 * @cfts: zero-length name terminated array of cftypes
4337 * Unregister @cfts. Files described by @cfts are removed from all
4338 * existing cgroups and all future cgroups won't have them either. This
4339 * function can be called anytime whether @cfts' subsys is attached or not.
4341 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4344 int cgroup_rm_cftypes(struct cftype *cfts)
4348 if (!cfts || cfts[0].name[0] == '\0')
4351 if (!(cfts[0].flags & __CFTYPE_ADDED))
4355 ret = cgroup_rm_cftypes_locked(cfts);
4361 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4362 * @ss: target cgroup subsystem
4363 * @cfts: zero-length name terminated array of cftypes
4365 * Register @cfts to @ss. Files described by @cfts are created for all
4366 * existing cgroups to which @ss is attached and all future cgroups will
4367 * have them too. This function can be called anytime whether @ss is
4370 * Returns 0 on successful registration, -errno on failure. Note that this
4371 * function currently returns 0 as long as @cfts registration is successful
4372 * even if some file creation attempts on existing cgroups fail.
4374 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4378 if (!cgroup_ssid_enabled(ss->id))
4381 if (!cfts || cfts[0].name[0] == '\0')
4384 ret = cgroup_init_cftypes(ss, cfts);
4390 list_add_tail(&cfts->node, &ss->cfts);
4391 ret = cgroup_apply_cftypes(cfts, true);
4393 cgroup_rm_cftypes_locked(cfts);
4400 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4401 * @ss: target cgroup subsystem
4402 * @cfts: zero-length name terminated array of cftypes
4404 * Similar to cgroup_add_cftypes() but the added files are only used for
4405 * the default hierarchy.
4407 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4411 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4412 cft->flags |= __CFTYPE_ONLY_ON_DFL;
4413 return cgroup_add_cftypes(ss, cfts);
4417 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4418 * @ss: target cgroup subsystem
4419 * @cfts: zero-length name terminated array of cftypes
4421 * Similar to cgroup_add_cftypes() but the added files are only used for
4422 * the legacy hierarchies.
4424 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4428 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4429 cft->flags |= __CFTYPE_NOT_ON_DFL;
4430 return cgroup_add_cftypes(ss, cfts);
4434 * cgroup_file_notify - generate a file modified event for a cgroup_file
4435 * @cfile: target cgroup_file
4437 * @cfile must have been obtained by setting cftype->file_offset.
4439 void cgroup_file_notify(struct cgroup_file *cfile)
4441 unsigned long flags;
4443 spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4445 unsigned long last = cfile->notified_at;
4446 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4448 if (time_in_range(jiffies, last, next)) {
4449 timer_reduce(&cfile->notify_timer, next);
4451 kernfs_notify(cfile->kn);
4452 cfile->notified_at = jiffies;
4455 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4459 * cgroup_file_show - show or hide a hidden cgroup file
4460 * @cfile: target cgroup_file obtained by setting cftype->file_offset
4461 * @show: whether to show or hide
4463 void cgroup_file_show(struct cgroup_file *cfile, bool show)
4465 struct kernfs_node *kn;
4467 spin_lock_irq(&cgroup_file_kn_lock);
4470 spin_unlock_irq(&cgroup_file_kn_lock);
4473 kernfs_show(kn, show);
4479 * css_next_child - find the next child of a given css
4480 * @pos: the current position (%NULL to initiate traversal)
4481 * @parent: css whose children to walk
4483 * This function returns the next child of @parent and should be called
4484 * under either cgroup_mutex or RCU read lock. The only requirement is
4485 * that @parent and @pos are accessible. The next sibling is guaranteed to
4486 * be returned regardless of their states.
4488 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4489 * css which finished ->css_online() is guaranteed to be visible in the
4490 * future iterations and will stay visible until the last reference is put.
4491 * A css which hasn't finished ->css_online() or already finished
4492 * ->css_offline() may show up during traversal. It's each subsystem's
4493 * responsibility to synchronize against on/offlining.
4495 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4496 struct cgroup_subsys_state *parent)
4498 struct cgroup_subsys_state *next;
4500 cgroup_assert_mutex_or_rcu_locked();
4503 * @pos could already have been unlinked from the sibling list.
4504 * Once a cgroup is removed, its ->sibling.next is no longer
4505 * updated when its next sibling changes. CSS_RELEASED is set when
4506 * @pos is taken off list, at which time its next pointer is valid,
4507 * and, as releases are serialized, the one pointed to by the next
4508 * pointer is guaranteed to not have started release yet. This
4509 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4510 * critical section, the one pointed to by its next pointer is
4511 * guaranteed to not have finished its RCU grace period even if we
4512 * have dropped rcu_read_lock() in-between iterations.
4514 * If @pos has CSS_RELEASED set, its next pointer can't be
4515 * dereferenced; however, as each css is given a monotonically
4516 * increasing unique serial number and always appended to the
4517 * sibling list, the next one can be found by walking the parent's
4518 * children until the first css with higher serial number than
4519 * @pos's. While this path can be slower, it happens iff iteration
4520 * races against release and the race window is very small.
4523 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4524 } else if (likely(!(pos->flags & CSS_RELEASED))) {
4525 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4527 list_for_each_entry_rcu(next, &parent->children, sibling,
4528 lockdep_is_held(&cgroup_mutex))
4529 if (next->serial_nr > pos->serial_nr)
4534 * @next, if not pointing to the head, can be dereferenced and is
4537 if (&next->sibling != &parent->children)
4543 * css_next_descendant_pre - find the next descendant for pre-order walk
4544 * @pos: the current position (%NULL to initiate traversal)
4545 * @root: css whose descendants to walk
4547 * To be used by css_for_each_descendant_pre(). Find the next descendant
4548 * to visit for pre-order traversal of @root's descendants. @root is
4549 * included in the iteration and the first node to be visited.
4551 * While this function requires cgroup_mutex or RCU read locking, it
4552 * doesn't require the whole traversal to be contained in a single critical
4553 * section. This function will return the correct next descendant as long
4554 * as both @pos and @root are accessible and @pos is a descendant of @root.
4556 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4557 * css which finished ->css_online() is guaranteed to be visible in the
4558 * future iterations and will stay visible until the last reference is put.
4559 * A css which hasn't finished ->css_online() or already finished
4560 * ->css_offline() may show up during traversal. It's each subsystem's
4561 * responsibility to synchronize against on/offlining.
4563 struct cgroup_subsys_state *
4564 css_next_descendant_pre(struct cgroup_subsys_state *pos,
4565 struct cgroup_subsys_state *root)
4567 struct cgroup_subsys_state *next;
4569 cgroup_assert_mutex_or_rcu_locked();
4571 /* if first iteration, visit @root */
4575 /* visit the first child if exists */
4576 next = css_next_child(NULL, pos);
4580 /* no child, visit my or the closest ancestor's next sibling */
4581 while (pos != root) {
4582 next = css_next_child(pos, pos->parent);
4590 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4593 * css_rightmost_descendant - return the rightmost descendant of a css
4594 * @pos: css of interest
4596 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4597 * is returned. This can be used during pre-order traversal to skip
4600 * While this function requires cgroup_mutex or RCU read locking, it
4601 * doesn't require the whole traversal to be contained in a single critical
4602 * section. This function will return the correct rightmost descendant as
4603 * long as @pos is accessible.
4605 struct cgroup_subsys_state *
4606 css_rightmost_descendant(struct cgroup_subsys_state *pos)
4608 struct cgroup_subsys_state *last, *tmp;
4610 cgroup_assert_mutex_or_rcu_locked();
4614 /* ->prev isn't RCU safe, walk ->next till the end */
4616 css_for_each_child(tmp, last)
4623 static struct cgroup_subsys_state *
4624 css_leftmost_descendant(struct cgroup_subsys_state *pos)
4626 struct cgroup_subsys_state *last;
4630 pos = css_next_child(NULL, pos);
4637 * css_next_descendant_post - find the next descendant for post-order walk
4638 * @pos: the current position (%NULL to initiate traversal)
4639 * @root: css whose descendants to walk
4641 * To be used by css_for_each_descendant_post(). Find the next descendant
4642 * to visit for post-order traversal of @root's descendants. @root is
4643 * included in the iteration and the last node to be visited.
4645 * While this function requires cgroup_mutex or RCU read locking, it
4646 * doesn't require the whole traversal to be contained in a single critical
4647 * section. This function will return the correct next descendant as long
4648 * as both @pos and @cgroup are accessible and @pos is a descendant of
4651 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4652 * css which finished ->css_online() is guaranteed to be visible in the
4653 * future iterations and will stay visible until the last reference is put.
4654 * A css which hasn't finished ->css_online() or already finished
4655 * ->css_offline() may show up during traversal. It's each subsystem's
4656 * responsibility to synchronize against on/offlining.
4658 struct cgroup_subsys_state *
4659 css_next_descendant_post(struct cgroup_subsys_state *pos,
4660 struct cgroup_subsys_state *root)
4662 struct cgroup_subsys_state *next;
4664 cgroup_assert_mutex_or_rcu_locked();
4666 /* if first iteration, visit leftmost descendant which may be @root */
4668 return css_leftmost_descendant(root);
4670 /* if we visited @root, we're done */
4674 /* if there's an unvisited sibling, visit its leftmost descendant */
4675 next = css_next_child(pos, pos->parent);
4677 return css_leftmost_descendant(next);
4679 /* no sibling left, visit parent */
4684 * css_has_online_children - does a css have online children
4685 * @css: the target css
4687 * Returns %true if @css has any online children; otherwise, %false. This
4688 * function can be called from any context but the caller is responsible
4689 * for synchronizing against on/offlining as necessary.
4691 bool css_has_online_children(struct cgroup_subsys_state *css)
4693 struct cgroup_subsys_state *child;
4697 css_for_each_child(child, css) {
4698 if (child->flags & CSS_ONLINE) {
4707 static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4709 struct list_head *l;
4710 struct cgrp_cset_link *link;
4711 struct css_set *cset;
4713 lockdep_assert_held(&css_set_lock);
4715 /* find the next threaded cset */
4716 if (it->tcset_pos) {
4717 l = it->tcset_pos->next;
4719 if (l != it->tcset_head) {
4721 return container_of(l, struct css_set,
4722 threaded_csets_node);
4725 it->tcset_pos = NULL;
4728 /* find the next cset */
4731 if (l == it->cset_head) {
4732 it->cset_pos = NULL;
4737 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4739 link = list_entry(l, struct cgrp_cset_link, cset_link);
4745 /* initialize threaded css_set walking */
4746 if (it->flags & CSS_TASK_ITER_THREADED) {
4748 put_css_set_locked(it->cur_dcset);
4749 it->cur_dcset = cset;
4752 it->tcset_head = &cset->threaded_csets;
4753 it->tcset_pos = &cset->threaded_csets;
4760 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4761 * @it: the iterator to advance
4763 * Advance @it to the next css_set to walk.
4765 static void css_task_iter_advance_css_set(struct css_task_iter *it)
4767 struct css_set *cset;
4769 lockdep_assert_held(&css_set_lock);
4771 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4772 while ((cset = css_task_iter_next_css_set(it))) {
4773 if (!list_empty(&cset->tasks)) {
4774 it->cur_tasks_head = &cset->tasks;
4776 } else if (!list_empty(&cset->mg_tasks)) {
4777 it->cur_tasks_head = &cset->mg_tasks;
4779 } else if (!list_empty(&cset->dying_tasks)) {
4780 it->cur_tasks_head = &cset->dying_tasks;
4785 it->task_pos = NULL;
4788 it->task_pos = it->cur_tasks_head->next;
4791 * We don't keep css_sets locked across iteration steps and thus
4792 * need to take steps to ensure that iteration can be resumed after
4793 * the lock is re-acquired. Iteration is performed at two levels -
4794 * css_sets and tasks in them.
4796 * Once created, a css_set never leaves its cgroup lists, so a
4797 * pinned css_set is guaranteed to stay put and we can resume
4798 * iteration afterwards.
4800 * Tasks may leave @cset across iteration steps. This is resolved
4801 * by registering each iterator with the css_set currently being
4802 * walked and making css_set_move_task() advance iterators whose
4803 * next task is leaving.
4806 list_del(&it->iters_node);
4807 put_css_set_locked(it->cur_cset);
4810 it->cur_cset = cset;
4811 list_add(&it->iters_node, &cset->task_iters);
4814 static void css_task_iter_skip(struct css_task_iter *it,
4815 struct task_struct *task)
4817 lockdep_assert_held(&css_set_lock);
4819 if (it->task_pos == &task->cg_list) {
4820 it->task_pos = it->task_pos->next;
4821 it->flags |= CSS_TASK_ITER_SKIPPED;
4825 static void css_task_iter_advance(struct css_task_iter *it)
4827 struct task_struct *task;
4829 lockdep_assert_held(&css_set_lock);
4833 * Advance iterator to find next entry. We go through cset
4834 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4837 if (it->flags & CSS_TASK_ITER_SKIPPED)
4838 it->flags &= ~CSS_TASK_ITER_SKIPPED;
4840 it->task_pos = it->task_pos->next;
4842 if (it->task_pos == &it->cur_cset->tasks) {
4843 it->cur_tasks_head = &it->cur_cset->mg_tasks;
4844 it->task_pos = it->cur_tasks_head->next;
4846 if (it->task_pos == &it->cur_cset->mg_tasks) {
4847 it->cur_tasks_head = &it->cur_cset->dying_tasks;
4848 it->task_pos = it->cur_tasks_head->next;
4850 if (it->task_pos == &it->cur_cset->dying_tasks)
4851 css_task_iter_advance_css_set(it);
4853 /* called from start, proceed to the first cset */
4854 css_task_iter_advance_css_set(it);
4860 task = list_entry(it->task_pos, struct task_struct, cg_list);
4862 if (it->flags & CSS_TASK_ITER_PROCS) {
4863 /* if PROCS, skip over tasks which aren't group leaders */
4864 if (!thread_group_leader(task))
4867 /* and dying leaders w/o live member threads */
4868 if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4869 !atomic_read(&task->signal->live))
4872 /* skip all dying ones */
4873 if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4879 * css_task_iter_start - initiate task iteration
4880 * @css: the css to walk tasks of
4881 * @flags: CSS_TASK_ITER_* flags
4882 * @it: the task iterator to use
4884 * Initiate iteration through the tasks of @css. The caller can call
4885 * css_task_iter_next() to walk through the tasks until the function
4886 * returns NULL. On completion of iteration, css_task_iter_end() must be
4889 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4890 struct css_task_iter *it)
4892 memset(it, 0, sizeof(*it));
4894 spin_lock_irq(&css_set_lock);
4899 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss)
4900 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4902 it->cset_pos = &css->cgroup->cset_links;
4904 it->cset_head = it->cset_pos;
4906 css_task_iter_advance(it);
4908 spin_unlock_irq(&css_set_lock);
4912 * css_task_iter_next - return the next task for the iterator
4913 * @it: the task iterator being iterated
4915 * The "next" function for task iteration. @it should have been
4916 * initialized via css_task_iter_start(). Returns NULL when the iteration
4919 struct task_struct *css_task_iter_next(struct css_task_iter *it)
4922 put_task_struct(it->cur_task);
4923 it->cur_task = NULL;
4926 spin_lock_irq(&css_set_lock);
4928 /* @it may be half-advanced by skips, finish advancing */
4929 if (it->flags & CSS_TASK_ITER_SKIPPED)
4930 css_task_iter_advance(it);
4933 it->cur_task = list_entry(it->task_pos, struct task_struct,
4935 get_task_struct(it->cur_task);
4936 css_task_iter_advance(it);
4939 spin_unlock_irq(&css_set_lock);
4941 return it->cur_task;
4945 * css_task_iter_end - finish task iteration
4946 * @it: the task iterator to finish
4948 * Finish task iteration started by css_task_iter_start().
4950 void css_task_iter_end(struct css_task_iter *it)
4953 spin_lock_irq(&css_set_lock);
4954 list_del(&it->iters_node);
4955 put_css_set_locked(it->cur_cset);
4956 spin_unlock_irq(&css_set_lock);
4960 put_css_set(it->cur_dcset);
4963 put_task_struct(it->cur_task);
4966 static void cgroup_procs_release(struct kernfs_open_file *of)
4968 struct cgroup_file_ctx *ctx = of->priv;
4970 if (ctx->procs.started)
4971 css_task_iter_end(&ctx->procs.iter);
4974 static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4976 struct kernfs_open_file *of = s->private;
4977 struct cgroup_file_ctx *ctx = of->priv;
4982 return css_task_iter_next(&ctx->procs.iter);
4985 static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4986 unsigned int iter_flags)
4988 struct kernfs_open_file *of = s->private;
4989 struct cgroup *cgrp = seq_css(s)->cgroup;
4990 struct cgroup_file_ctx *ctx = of->priv;
4991 struct css_task_iter *it = &ctx->procs.iter;
4994 * When a seq_file is seeked, it's always traversed sequentially
4995 * from position 0, so we can simply keep iterating on !0 *pos.
4997 if (!ctx->procs.started) {
4998 if (WARN_ON_ONCE((*pos)))
4999 return ERR_PTR(-EINVAL);
5000 css_task_iter_start(&cgrp->self, iter_flags, it);
5001 ctx->procs.started = true;
5002 } else if (!(*pos)) {
5003 css_task_iter_end(it);
5004 css_task_iter_start(&cgrp->self, iter_flags, it);
5006 return it->cur_task;
5008 return cgroup_procs_next(s, NULL, NULL);
5011 static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
5013 struct cgroup *cgrp = seq_css(s)->cgroup;
5016 * All processes of a threaded subtree belong to the domain cgroup
5017 * of the subtree. Only threads can be distributed across the
5018 * subtree. Reject reads on cgroup.procs in the subtree proper.
5019 * They're always empty anyway.
5021 if (cgroup_is_threaded(cgrp))
5022 return ERR_PTR(-EOPNOTSUPP);
5024 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
5025 CSS_TASK_ITER_THREADED);
5028 static int cgroup_procs_show(struct seq_file *s, void *v)
5030 seq_printf(s, "%d\n", task_pid_vnr(v));
5034 static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
5037 struct inode *inode;
5039 lockdep_assert_held(&cgroup_mutex);
5041 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
5045 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE);
5050 static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
5051 struct cgroup *dst_cgrp,
5052 struct super_block *sb,
5053 struct cgroup_namespace *ns)
5055 struct cgroup *com_cgrp = src_cgrp;
5058 lockdep_assert_held(&cgroup_mutex);
5060 /* find the common ancestor */
5061 while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
5062 com_cgrp = cgroup_parent(com_cgrp);
5064 /* %current should be authorized to migrate to the common ancestor */
5065 ret = cgroup_may_write(com_cgrp, sb);
5070 * If namespaces are delegation boundaries, %current must be able
5071 * to see both source and destination cgroups from its namespace.
5073 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
5074 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
5075 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
5081 static int cgroup_attach_permissions(struct cgroup *src_cgrp,
5082 struct cgroup *dst_cgrp,
5083 struct super_block *sb, bool threadgroup,
5084 struct cgroup_namespace *ns)
5088 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns);
5092 ret = cgroup_migrate_vet_dst(dst_cgrp);
5096 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp))
5102 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
5105 struct cgroup_file_ctx *ctx = of->priv;
5106 struct cgroup *src_cgrp, *dst_cgrp;
5107 struct task_struct *task;
5108 const struct cred *saved_cred;
5110 bool threadgroup_locked;
5112 dst_cgrp = cgroup_kn_lock_live(of->kn, false);
5116 task = cgroup_procs_write_start(buf, threadgroup, &threadgroup_locked);
5117 ret = PTR_ERR_OR_ZERO(task);
5121 /* find the source cgroup */
5122 spin_lock_irq(&css_set_lock);
5123 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
5124 spin_unlock_irq(&css_set_lock);
5127 * Process and thread migrations follow same delegation rule. Check
5128 * permissions using the credentials from file open to protect against
5129 * inherited fd attacks.
5131 saved_cred = override_creds(of->file->f_cred);
5132 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp,
5133 of->file->f_path.dentry->d_sb,
5134 threadgroup, ctx->ns);
5135 revert_creds(saved_cred);
5139 ret = cgroup_attach_task(dst_cgrp, task, threadgroup);
5142 cgroup_procs_write_finish(task, threadgroup_locked);
5144 cgroup_kn_unlock(of->kn);
5149 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
5150 char *buf, size_t nbytes, loff_t off)
5152 return __cgroup_procs_write(of, buf, true) ?: nbytes;
5155 static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
5157 return __cgroup_procs_start(s, pos, 0);
5160 static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
5161 char *buf, size_t nbytes, loff_t off)
5163 return __cgroup_procs_write(of, buf, false) ?: nbytes;
5166 /* cgroup core interface files for the default hierarchy */
5167 static struct cftype cgroup_base_files[] = {
5169 .name = "cgroup.type",
5170 .flags = CFTYPE_NOT_ON_ROOT,
5171 .seq_show = cgroup_type_show,
5172 .write = cgroup_type_write,
5175 .name = "cgroup.procs",
5176 .flags = CFTYPE_NS_DELEGATABLE,
5177 .file_offset = offsetof(struct cgroup, procs_file),
5178 .release = cgroup_procs_release,
5179 .seq_start = cgroup_procs_start,
5180 .seq_next = cgroup_procs_next,
5181 .seq_show = cgroup_procs_show,
5182 .write = cgroup_procs_write,
5185 .name = "cgroup.threads",
5186 .flags = CFTYPE_NS_DELEGATABLE,
5187 .release = cgroup_procs_release,
5188 .seq_start = cgroup_threads_start,
5189 .seq_next = cgroup_procs_next,
5190 .seq_show = cgroup_procs_show,
5191 .write = cgroup_threads_write,
5194 .name = "cgroup.controllers",
5195 .seq_show = cgroup_controllers_show,
5198 .name = "cgroup.subtree_control",
5199 .flags = CFTYPE_NS_DELEGATABLE,
5200 .seq_show = cgroup_subtree_control_show,
5201 .write = cgroup_subtree_control_write,
5204 .name = "cgroup.events",
5205 .flags = CFTYPE_NOT_ON_ROOT,
5206 .file_offset = offsetof(struct cgroup, events_file),
5207 .seq_show = cgroup_events_show,
5210 .name = "cgroup.max.descendants",
5211 .seq_show = cgroup_max_descendants_show,
5212 .write = cgroup_max_descendants_write,
5215 .name = "cgroup.max.depth",
5216 .seq_show = cgroup_max_depth_show,
5217 .write = cgroup_max_depth_write,
5220 .name = "cgroup.stat",
5221 .seq_show = cgroup_stat_show,
5224 .name = "cgroup.freeze",
5225 .flags = CFTYPE_NOT_ON_ROOT,
5226 .seq_show = cgroup_freeze_show,
5227 .write = cgroup_freeze_write,
5230 .name = "cgroup.kill",
5231 .flags = CFTYPE_NOT_ON_ROOT,
5232 .write = cgroup_kill_write,
5236 .seq_show = cpu_stat_show,
5241 static struct cftype cgroup_psi_files[] = {
5244 .name = "io.pressure",
5245 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]),
5246 .open = cgroup_pressure_open,
5247 .seq_show = cgroup_io_pressure_show,
5248 .write = cgroup_io_pressure_write,
5249 .poll = cgroup_pressure_poll,
5250 .release = cgroup_pressure_release,
5253 .name = "memory.pressure",
5254 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]),
5255 .open = cgroup_pressure_open,
5256 .seq_show = cgroup_memory_pressure_show,
5257 .write = cgroup_memory_pressure_write,
5258 .poll = cgroup_pressure_poll,
5259 .release = cgroup_pressure_release,
5262 .name = "cpu.pressure",
5263 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]),
5264 .open = cgroup_pressure_open,
5265 .seq_show = cgroup_cpu_pressure_show,
5266 .write = cgroup_cpu_pressure_write,
5267 .poll = cgroup_pressure_poll,
5268 .release = cgroup_pressure_release,
5270 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
5272 .name = "irq.pressure",
5273 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]),
5274 .open = cgroup_pressure_open,
5275 .seq_show = cgroup_irq_pressure_show,
5276 .write = cgroup_irq_pressure_write,
5277 .poll = cgroup_pressure_poll,
5278 .release = cgroup_pressure_release,
5282 .name = "cgroup.pressure",
5283 .seq_show = cgroup_pressure_show,
5284 .write = cgroup_pressure_write,
5286 #endif /* CONFIG_PSI */
5291 * css destruction is four-stage process.
5293 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5294 * Implemented in kill_css().
5296 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5297 * and thus css_tryget_online() is guaranteed to fail, the css can be
5298 * offlined by invoking offline_css(). After offlining, the base ref is
5299 * put. Implemented in css_killed_work_fn().
5301 * 3. When the percpu_ref reaches zero, the only possible remaining
5302 * accessors are inside RCU read sections. css_release() schedules the
5305 * 4. After the grace period, the css can be freed. Implemented in
5306 * css_free_work_fn().
5308 * It is actually hairier because both step 2 and 4 require process context
5309 * and thus involve punting to css->destroy_work adding two additional
5310 * steps to the already complex sequence.
5312 static void css_free_rwork_fn(struct work_struct *work)
5314 struct cgroup_subsys_state *css = container_of(to_rcu_work(work),
5315 struct cgroup_subsys_state, destroy_rwork);
5316 struct cgroup_subsys *ss = css->ss;
5317 struct cgroup *cgrp = css->cgroup;
5319 percpu_ref_exit(&css->refcnt);
5323 struct cgroup_subsys_state *parent = css->parent;
5327 cgroup_idr_remove(&ss->css_idr, id);
5333 /* cgroup free path */
5334 atomic_dec(&cgrp->root->nr_cgrps);
5335 cgroup1_pidlist_destroy_all(cgrp);
5336 cancel_work_sync(&cgrp->release_agent_work);
5337 bpf_cgrp_storage_free(cgrp);
5339 if (cgroup_parent(cgrp)) {
5341 * We get a ref to the parent, and put the ref when
5342 * this cgroup is being freed, so it's guaranteed
5343 * that the parent won't be destroyed before its
5346 cgroup_put(cgroup_parent(cgrp));
5347 kernfs_put(cgrp->kn);
5348 psi_cgroup_free(cgrp);
5349 cgroup_rstat_exit(cgrp);
5353 * This is root cgroup's refcnt reaching zero,
5354 * which indicates that the root should be
5357 cgroup_destroy_root(cgrp->root);
5362 static void css_release_work_fn(struct work_struct *work)
5364 struct cgroup_subsys_state *css =
5365 container_of(work, struct cgroup_subsys_state, destroy_work);
5366 struct cgroup_subsys *ss = css->ss;
5367 struct cgroup *cgrp = css->cgroup;
5371 css->flags |= CSS_RELEASED;
5372 list_del_rcu(&css->sibling);
5375 /* css release path */
5376 if (!list_empty(&css->rstat_css_node)) {
5377 cgroup_rstat_flush(cgrp);
5378 list_del_rcu(&css->rstat_css_node);
5381 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
5382 if (ss->css_released)
5383 ss->css_released(css);
5385 struct cgroup *tcgrp;
5387 /* cgroup release path */
5388 TRACE_CGROUP_PATH(release, cgrp);
5390 cgroup_rstat_flush(cgrp);
5392 spin_lock_irq(&css_set_lock);
5393 for (tcgrp = cgroup_parent(cgrp); tcgrp;
5394 tcgrp = cgroup_parent(tcgrp))
5395 tcgrp->nr_dying_descendants--;
5396 spin_unlock_irq(&css_set_lock);
5399 * There are two control paths which try to determine
5400 * cgroup from dentry without going through kernfs -
5401 * cgroupstats_build() and css_tryget_online_from_dir().
5402 * Those are supported by RCU protecting clearing of
5403 * cgrp->kn->priv backpointer.
5406 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
5412 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5413 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5416 static void css_release(struct percpu_ref *ref)
5418 struct cgroup_subsys_state *css =
5419 container_of(ref, struct cgroup_subsys_state, refcnt);
5421 INIT_WORK(&css->destroy_work, css_release_work_fn);
5422 queue_work(cgroup_destroy_wq, &css->destroy_work);
5425 static void init_and_link_css(struct cgroup_subsys_state *css,
5426 struct cgroup_subsys *ss, struct cgroup *cgrp)
5428 lockdep_assert_held(&cgroup_mutex);
5430 cgroup_get_live(cgrp);
5432 memset(css, 0, sizeof(*css));
5436 INIT_LIST_HEAD(&css->sibling);
5437 INIT_LIST_HEAD(&css->children);
5438 INIT_LIST_HEAD(&css->rstat_css_node);
5439 css->serial_nr = css_serial_nr_next++;
5440 atomic_set(&css->online_cnt, 0);
5442 if (cgroup_parent(cgrp)) {
5443 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
5444 css_get(css->parent);
5447 if (ss->css_rstat_flush)
5448 list_add_rcu(&css->rstat_css_node, &cgrp->rstat_css_list);
5450 BUG_ON(cgroup_css(cgrp, ss));
5453 /* invoke ->css_online() on a new CSS and mark it online if successful */
5454 static int online_css(struct cgroup_subsys_state *css)
5456 struct cgroup_subsys *ss = css->ss;
5459 lockdep_assert_held(&cgroup_mutex);
5462 ret = ss->css_online(css);
5464 css->flags |= CSS_ONLINE;
5465 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
5467 atomic_inc(&css->online_cnt);
5469 atomic_inc(&css->parent->online_cnt);
5474 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5475 static void offline_css(struct cgroup_subsys_state *css)
5477 struct cgroup_subsys *ss = css->ss;
5479 lockdep_assert_held(&cgroup_mutex);
5481 if (!(css->flags & CSS_ONLINE))
5484 if (ss->css_offline)
5485 ss->css_offline(css);
5487 css->flags &= ~CSS_ONLINE;
5488 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
5490 wake_up_all(&css->cgroup->offline_waitq);
5494 * css_create - create a cgroup_subsys_state
5495 * @cgrp: the cgroup new css will be associated with
5496 * @ss: the subsys of new css
5498 * Create a new css associated with @cgrp - @ss pair. On success, the new
5499 * css is online and installed in @cgrp. This function doesn't create the
5500 * interface files. Returns 0 on success, -errno on failure.
5502 static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
5503 struct cgroup_subsys *ss)
5505 struct cgroup *parent = cgroup_parent(cgrp);
5506 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
5507 struct cgroup_subsys_state *css;
5510 lockdep_assert_held(&cgroup_mutex);
5512 css = ss->css_alloc(parent_css);
5514 css = ERR_PTR(-ENOMEM);
5518 init_and_link_css(css, ss, cgrp);
5520 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
5524 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
5529 /* @css is ready to be brought online now, make it visible */
5530 list_add_tail_rcu(&css->sibling, &parent_css->children);
5531 cgroup_idr_replace(&ss->css_idr, css, css->id);
5533 err = online_css(css);
5540 list_del_rcu(&css->sibling);
5542 list_del_rcu(&css->rstat_css_node);
5543 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn);
5544 queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
5545 return ERR_PTR(err);
5549 * The returned cgroup is fully initialized including its control mask, but
5550 * it isn't associated with its kernfs_node and doesn't have the control
5553 static struct cgroup *cgroup_create(struct cgroup *parent, const char *name,
5556 struct cgroup_root *root = parent->root;
5557 struct cgroup *cgrp, *tcgrp;
5558 struct kernfs_node *kn;
5559 int level = parent->level + 1;
5562 /* allocate the cgroup and its ID, 0 is reserved for the root */
5563 cgrp = kzalloc(struct_size(cgrp, ancestors, (level + 1)), GFP_KERNEL);
5565 return ERR_PTR(-ENOMEM);
5567 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
5571 ret = cgroup_rstat_init(cgrp);
5573 goto out_cancel_ref;
5575 /* create the directory */
5576 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
5583 init_cgroup_housekeeping(cgrp);
5585 cgrp->self.parent = &parent->self;
5587 cgrp->level = level;
5589 ret = psi_cgroup_alloc(cgrp);
5591 goto out_kernfs_remove;
5593 ret = cgroup_bpf_inherit(cgrp);
5598 * New cgroup inherits effective freeze counter, and
5599 * if the parent has to be frozen, the child has too.
5601 cgrp->freezer.e_freeze = parent->freezer.e_freeze;
5602 if (cgrp->freezer.e_freeze) {
5604 * Set the CGRP_FREEZE flag, so when a process will be
5605 * attached to the child cgroup, it will become frozen.
5606 * At this point the new cgroup is unpopulated, so we can
5607 * consider it frozen immediately.
5609 set_bit(CGRP_FREEZE, &cgrp->flags);
5610 set_bit(CGRP_FROZEN, &cgrp->flags);
5613 spin_lock_irq(&css_set_lock);
5614 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5615 cgrp->ancestors[tcgrp->level] = tcgrp;
5617 if (tcgrp != cgrp) {
5618 tcgrp->nr_descendants++;
5621 * If the new cgroup is frozen, all ancestor cgroups
5622 * get a new frozen descendant, but their state can't
5623 * change because of this.
5625 if (cgrp->freezer.e_freeze)
5626 tcgrp->freezer.nr_frozen_descendants++;
5629 spin_unlock_irq(&css_set_lock);
5631 if (notify_on_release(parent))
5632 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
5634 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
5635 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
5637 cgrp->self.serial_nr = css_serial_nr_next++;
5639 /* allocation complete, commit to creation */
5640 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
5641 atomic_inc(&root->nr_cgrps);
5642 cgroup_get_live(parent);
5645 * On the default hierarchy, a child doesn't automatically inherit
5646 * subtree_control from the parent. Each is configured manually.
5648 if (!cgroup_on_dfl(cgrp))
5649 cgrp->subtree_control = cgroup_control(cgrp);
5651 cgroup_propagate_control(cgrp);
5656 psi_cgroup_free(cgrp);
5658 kernfs_remove(cgrp->kn);
5660 cgroup_rstat_exit(cgrp);
5662 percpu_ref_exit(&cgrp->self.refcnt);
5665 return ERR_PTR(ret);
5668 static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
5670 struct cgroup *cgroup;
5674 lockdep_assert_held(&cgroup_mutex);
5676 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
5677 if (cgroup->nr_descendants >= cgroup->max_descendants)
5680 if (level > cgroup->max_depth)
5691 int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
5693 struct cgroup *parent, *cgrp;
5696 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5697 if (strchr(name, '\n'))
5700 parent = cgroup_kn_lock_live(parent_kn, false);
5704 if (!cgroup_check_hierarchy_limits(parent)) {
5709 cgrp = cgroup_create(parent, name, mode);
5711 ret = PTR_ERR(cgrp);
5716 * This extra ref will be put in cgroup_free_fn() and guarantees
5717 * that @cgrp->kn is always accessible.
5719 kernfs_get(cgrp->kn);
5721 ret = cgroup_kn_set_ugid(cgrp->kn);
5725 ret = css_populate_dir(&cgrp->self);
5729 ret = cgroup_apply_control_enable(cgrp);
5733 TRACE_CGROUP_PATH(mkdir, cgrp);
5735 /* let's create and online css's */
5736 kernfs_activate(cgrp->kn);
5742 cgroup_destroy_locked(cgrp);
5744 cgroup_kn_unlock(parent_kn);
5749 * This is called when the refcnt of a css is confirmed to be killed.
5750 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5751 * initiate destruction and put the css ref from kill_css().
5753 static void css_killed_work_fn(struct work_struct *work)
5755 struct cgroup_subsys_state *css =
5756 container_of(work, struct cgroup_subsys_state, destroy_work);
5763 /* @css can't go away while we're holding cgroup_mutex */
5765 } while (css && atomic_dec_and_test(&css->online_cnt));
5770 /* css kill confirmation processing requires process context, bounce */
5771 static void css_killed_ref_fn(struct percpu_ref *ref)
5773 struct cgroup_subsys_state *css =
5774 container_of(ref, struct cgroup_subsys_state, refcnt);
5776 if (atomic_dec_and_test(&css->online_cnt)) {
5777 INIT_WORK(&css->destroy_work, css_killed_work_fn);
5778 queue_work(cgroup_destroy_wq, &css->destroy_work);
5783 * kill_css - destroy a css
5784 * @css: css to destroy
5786 * This function initiates destruction of @css by removing cgroup interface
5787 * files and putting its base reference. ->css_offline() will be invoked
5788 * asynchronously once css_tryget_online() is guaranteed to fail and when
5789 * the reference count reaches zero, @css will be released.
5791 static void kill_css(struct cgroup_subsys_state *css)
5793 lockdep_assert_held(&cgroup_mutex);
5795 if (css->flags & CSS_DYING)
5798 css->flags |= CSS_DYING;
5801 * This must happen before css is disassociated with its cgroup.
5802 * See seq_css() for details.
5807 * Killing would put the base ref, but we need to keep it alive
5808 * until after ->css_offline().
5813 * cgroup core guarantees that, by the time ->css_offline() is
5814 * invoked, no new css reference will be given out via
5815 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5816 * proceed to offlining css's because percpu_ref_kill() doesn't
5817 * guarantee that the ref is seen as killed on all CPUs on return.
5819 * Use percpu_ref_kill_and_confirm() to get notifications as each
5820 * css is confirmed to be seen as killed on all CPUs.
5822 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
5826 * cgroup_destroy_locked - the first stage of cgroup destruction
5827 * @cgrp: cgroup to be destroyed
5829 * css's make use of percpu refcnts whose killing latency shouldn't be
5830 * exposed to userland and are RCU protected. Also, cgroup core needs to
5831 * guarantee that css_tryget_online() won't succeed by the time
5832 * ->css_offline() is invoked. To satisfy all the requirements,
5833 * destruction is implemented in the following two steps.
5835 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5836 * userland visible parts and start killing the percpu refcnts of
5837 * css's. Set up so that the next stage will be kicked off once all
5838 * the percpu refcnts are confirmed to be killed.
5840 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5841 * rest of destruction. Once all cgroup references are gone, the
5842 * cgroup is RCU-freed.
5844 * This function implements s1. After this step, @cgrp is gone as far as
5845 * the userland is concerned and a new cgroup with the same name may be
5846 * created. As cgroup doesn't care about the names internally, this
5847 * doesn't cause any problem.
5849 static int cgroup_destroy_locked(struct cgroup *cgrp)
5850 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
5852 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
5853 struct cgroup_subsys_state *css;
5854 struct cgrp_cset_link *link;
5857 lockdep_assert_held(&cgroup_mutex);
5860 * Only migration can raise populated from zero and we're already
5861 * holding cgroup_mutex.
5863 if (cgroup_is_populated(cgrp))
5867 * Make sure there's no live children. We can't test emptiness of
5868 * ->self.children as dead children linger on it while being
5869 * drained; otherwise, "rmdir parent/child parent" may fail.
5871 if (css_has_online_children(&cgrp->self))
5875 * Mark @cgrp and the associated csets dead. The former prevents
5876 * further task migration and child creation by disabling
5877 * cgroup_lock_live_group(). The latter makes the csets ignored by
5878 * the migration path.
5880 cgrp->self.flags &= ~CSS_ONLINE;
5882 spin_lock_irq(&css_set_lock);
5883 list_for_each_entry(link, &cgrp->cset_links, cset_link)
5884 link->cset->dead = true;
5885 spin_unlock_irq(&css_set_lock);
5887 /* initiate massacre of all css's */
5888 for_each_css(css, ssid, cgrp)
5891 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5892 css_clear_dir(&cgrp->self);
5893 kernfs_remove(cgrp->kn);
5895 if (cgroup_is_threaded(cgrp))
5896 parent->nr_threaded_children--;
5898 spin_lock_irq(&css_set_lock);
5899 for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5900 tcgrp->nr_descendants--;
5901 tcgrp->nr_dying_descendants++;
5903 * If the dying cgroup is frozen, decrease frozen descendants
5904 * counters of ancestor cgroups.
5906 if (test_bit(CGRP_FROZEN, &cgrp->flags))
5907 tcgrp->freezer.nr_frozen_descendants--;
5909 spin_unlock_irq(&css_set_lock);
5911 cgroup1_check_for_release(parent);
5913 cgroup_bpf_offline(cgrp);
5915 /* put the base reference */
5916 percpu_ref_kill(&cgrp->self.refcnt);
5921 int cgroup_rmdir(struct kernfs_node *kn)
5923 struct cgroup *cgrp;
5926 cgrp = cgroup_kn_lock_live(kn, false);
5930 ret = cgroup_destroy_locked(cgrp);
5932 TRACE_CGROUP_PATH(rmdir, cgrp);
5934 cgroup_kn_unlock(kn);
5938 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5939 .show_options = cgroup_show_options,
5940 .mkdir = cgroup_mkdir,
5941 .rmdir = cgroup_rmdir,
5942 .show_path = cgroup_show_path,
5945 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5947 struct cgroup_subsys_state *css;
5949 pr_debug("Initializing cgroup subsys %s\n", ss->name);
5953 idr_init(&ss->css_idr);
5954 INIT_LIST_HEAD(&ss->cfts);
5956 /* Create the root cgroup state for this subsystem */
5957 ss->root = &cgrp_dfl_root;
5958 css = ss->css_alloc(NULL);
5959 /* We don't handle early failures gracefully */
5960 BUG_ON(IS_ERR(css));
5961 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5964 * Root csses are never destroyed and we can't initialize
5965 * percpu_ref during early init. Disable refcnting.
5967 css->flags |= CSS_NO_REF;
5970 /* allocation can't be done safely during early init */
5973 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5974 BUG_ON(css->id < 0);
5977 /* Update the init_css_set to contain a subsys
5978 * pointer to this state - since the subsystem is
5979 * newly registered, all tasks and hence the
5980 * init_css_set is in the subsystem's root cgroup. */
5981 init_css_set.subsys[ss->id] = css;
5983 have_fork_callback |= (bool)ss->fork << ss->id;
5984 have_exit_callback |= (bool)ss->exit << ss->id;
5985 have_release_callback |= (bool)ss->release << ss->id;
5986 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5988 /* At system boot, before all subsystems have been
5989 * registered, no tasks have been forked, so we don't
5990 * need to invoke fork callbacks here. */
5991 BUG_ON(!list_empty(&init_task.tasks));
5993 BUG_ON(online_css(css));
5999 * cgroup_init_early - cgroup initialization at system boot
6001 * Initialize cgroups at system boot, and initialize any
6002 * subsystems that request early init.
6004 int __init cgroup_init_early(void)
6006 static struct cgroup_fs_context __initdata ctx;
6007 struct cgroup_subsys *ss;
6010 ctx.root = &cgrp_dfl_root;
6011 init_cgroup_root(&ctx);
6012 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
6014 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
6016 for_each_subsys(ss, i) {
6017 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
6018 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
6019 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
6021 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
6022 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
6025 ss->name = cgroup_subsys_name[i];
6026 if (!ss->legacy_name)
6027 ss->legacy_name = cgroup_subsys_name[i];
6030 cgroup_init_subsys(ss, true);
6036 * cgroup_init - cgroup initialization
6038 * Register cgroup filesystem and /proc file, and initialize
6039 * any subsystems that didn't request early init.
6041 int __init cgroup_init(void)
6043 struct cgroup_subsys *ss;
6046 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
6047 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
6048 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files));
6049 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
6051 cgroup_rstat_boot();
6053 get_user_ns(init_cgroup_ns.user_ns);
6058 * Add init_css_set to the hash table so that dfl_root can link to
6061 hash_add(css_set_table, &init_css_set.hlist,
6062 css_set_hash(init_css_set.subsys));
6064 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
6068 for_each_subsys(ss, ssid) {
6069 if (ss->early_init) {
6070 struct cgroup_subsys_state *css =
6071 init_css_set.subsys[ss->id];
6073 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
6075 BUG_ON(css->id < 0);
6077 cgroup_init_subsys(ss, false);
6080 list_add_tail(&init_css_set.e_cset_node[ssid],
6081 &cgrp_dfl_root.cgrp.e_csets[ssid]);
6084 * Setting dfl_root subsys_mask needs to consider the
6085 * disabled flag and cftype registration needs kmalloc,
6086 * both of which aren't available during early_init.
6088 if (!cgroup_ssid_enabled(ssid))
6091 if (cgroup1_ssid_disabled(ssid))
6092 printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
6095 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
6097 /* implicit controllers must be threaded too */
6098 WARN_ON(ss->implicit_on_dfl && !ss->threaded);
6100 if (ss->implicit_on_dfl)
6101 cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
6102 else if (!ss->dfl_cftypes)
6103 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
6106 cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
6108 if (ss->dfl_cftypes == ss->legacy_cftypes) {
6109 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
6111 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
6112 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
6116 ss->bind(init_css_set.subsys[ssid]);
6119 css_populate_dir(init_css_set.subsys[ssid]);
6123 /* init_css_set.subsys[] has been updated, re-hash */
6124 hash_del(&init_css_set.hlist);
6125 hash_add(css_set_table, &init_css_set.hlist,
6126 css_set_hash(init_css_set.subsys));
6128 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
6129 WARN_ON(register_filesystem(&cgroup_fs_type));
6130 WARN_ON(register_filesystem(&cgroup2_fs_type));
6131 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show));
6132 #ifdef CONFIG_CPUSETS
6133 WARN_ON(register_filesystem(&cpuset_fs_type));
6139 static int __init cgroup_wq_init(void)
6142 * There isn't much point in executing destruction path in
6143 * parallel. Good chunk is serialized with cgroup_mutex anyway.
6144 * Use 1 for @max_active.
6146 * We would prefer to do this in cgroup_init() above, but that
6147 * is called before init_workqueues(): so leave this until after.
6149 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
6150 BUG_ON(!cgroup_destroy_wq);
6153 core_initcall(cgroup_wq_init);
6155 void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen)
6157 struct kernfs_node *kn;
6159 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6162 kernfs_path(kn, buf, buflen);
6167 * cgroup_get_from_id : get the cgroup associated with cgroup id
6169 * On success return the cgrp or ERR_PTR on failure
6170 * Only cgroups within current task's cgroup NS are valid.
6172 struct cgroup *cgroup_get_from_id(u64 id)
6174 struct kernfs_node *kn;
6175 struct cgroup *cgrp, *root_cgrp;
6177 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id);
6179 return ERR_PTR(-ENOENT);
6181 if (kernfs_type(kn) != KERNFS_DIR) {
6183 return ERR_PTR(-ENOENT);
6188 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6189 if (cgrp && !cgroup_tryget(cgrp))
6196 return ERR_PTR(-ENOENT);
6198 root_cgrp = current_cgns_cgroup_dfl();
6199 if (!cgroup_is_descendant(cgrp, root_cgrp)) {
6201 return ERR_PTR(-ENOENT);
6206 EXPORT_SYMBOL_GPL(cgroup_get_from_id);
6209 * proc_cgroup_show()
6210 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6211 * - Used for /proc/<pid>/cgroup.
6213 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
6214 struct pid *pid, struct task_struct *tsk)
6218 struct cgroup_root *root;
6221 buf = kmalloc(PATH_MAX, GFP_KERNEL);
6226 spin_lock_irq(&css_set_lock);
6228 for_each_root(root) {
6229 struct cgroup_subsys *ss;
6230 struct cgroup *cgrp;
6231 int ssid, count = 0;
6233 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible))
6236 seq_printf(m, "%d:", root->hierarchy_id);
6237 if (root != &cgrp_dfl_root)
6238 for_each_subsys(ss, ssid)
6239 if (root->subsys_mask & (1 << ssid))
6240 seq_printf(m, "%s%s", count++ ? "," : "",
6242 if (strlen(root->name))
6243 seq_printf(m, "%sname=%s", count ? "," : "",
6247 cgrp = task_cgroup_from_root(tsk, root);
6250 * On traditional hierarchies, all zombie tasks show up as
6251 * belonging to the root cgroup. On the default hierarchy,
6252 * while a zombie doesn't show up in "cgroup.procs" and
6253 * thus can't be migrated, its /proc/PID/cgroup keeps
6254 * reporting the cgroup it belonged to before exiting. If
6255 * the cgroup is removed before the zombie is reaped,
6256 * " (deleted)" is appended to the cgroup path.
6258 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
6259 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
6260 current->nsproxy->cgroup_ns);
6261 if (retval >= PATH_MAX)
6262 retval = -ENAMETOOLONG;
6271 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
6272 seq_puts(m, " (deleted)\n");
6279 spin_unlock_irq(&css_set_lock);
6287 * cgroup_fork - initialize cgroup related fields during copy_process()
6288 * @child: pointer to task_struct of forking parent process.
6290 * A task is associated with the init_css_set until cgroup_post_fork()
6291 * attaches it to the target css_set.
6293 void cgroup_fork(struct task_struct *child)
6295 RCU_INIT_POINTER(child->cgroups, &init_css_set);
6296 INIT_LIST_HEAD(&child->cg_list);
6300 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer
6301 * @f: file corresponding to cgroup_dir
6303 * Find the cgroup from a file pointer associated with a cgroup directory.
6304 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the
6305 * cgroup cannot be found.
6307 static struct cgroup *cgroup_v1v2_get_from_file(struct file *f)
6309 struct cgroup_subsys_state *css;
6311 css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
6313 return ERR_CAST(css);
6319 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports
6321 * @f: file corresponding to cgroup2_dir
6323 static struct cgroup *cgroup_get_from_file(struct file *f)
6325 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f);
6328 return ERR_CAST(cgrp);
6330 if (!cgroup_on_dfl(cgrp)) {
6332 return ERR_PTR(-EBADF);
6339 * cgroup_css_set_fork - find or create a css_set for a child process
6340 * @kargs: the arguments passed to create the child process
6342 * This functions finds or creates a new css_set which the child
6343 * process will be attached to in cgroup_post_fork(). By default,
6344 * the child process will be given the same css_set as its parent.
6346 * If CLONE_INTO_CGROUP is specified this function will try to find an
6347 * existing css_set which includes the requested cgroup and if not create
6348 * a new css_set that the child will be attached to later. If this function
6349 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6350 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6351 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6352 * to the target cgroup.
6354 static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
6355 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem)
6358 struct cgroup *dst_cgrp = NULL;
6359 struct css_set *cset;
6360 struct super_block *sb;
6363 if (kargs->flags & CLONE_INTO_CGROUP)
6366 cgroup_threadgroup_change_begin(current);
6368 spin_lock_irq(&css_set_lock);
6369 cset = task_css_set(current);
6371 spin_unlock_irq(&css_set_lock);
6373 if (!(kargs->flags & CLONE_INTO_CGROUP)) {
6378 f = fget_raw(kargs->cgroup);
6383 sb = f->f_path.dentry->d_sb;
6385 dst_cgrp = cgroup_get_from_file(f);
6386 if (IS_ERR(dst_cgrp)) {
6387 ret = PTR_ERR(dst_cgrp);
6392 if (cgroup_is_dead(dst_cgrp)) {
6398 * Verify that we the target cgroup is writable for us. This is
6399 * usually done by the vfs layer but since we're not going through
6400 * the vfs layer here we need to do it "manually".
6402 ret = cgroup_may_write(dst_cgrp, sb);
6407 * Spawning a task directly into a cgroup works by passing a file
6408 * descriptor to the target cgroup directory. This can even be an O_PATH
6409 * file descriptor. But it can never be a cgroup.procs file descriptor.
6410 * This was done on purpose so spawning into a cgroup could be
6411 * conceptualized as an atomic
6413 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6414 * write(fd, <child-pid>, ...);
6416 * sequence, i.e. it's a shorthand for the caller opening and writing
6417 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6418 * to always use the caller's credentials.
6420 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb,
6421 !(kargs->flags & CLONE_THREAD),
6422 current->nsproxy->cgroup_ns);
6426 kargs->cset = find_css_set(cset, dst_cgrp);
6434 kargs->cgrp = dst_cgrp;
6438 cgroup_threadgroup_change_end(current);
6443 cgroup_put(dst_cgrp);
6446 put_css_set(kargs->cset);
6451 * cgroup_css_set_put_fork - drop references we took during fork
6452 * @kargs: the arguments passed to create the child process
6454 * Drop references to the prepared css_set and target cgroup if
6455 * CLONE_INTO_CGROUP was requested.
6457 static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs)
6458 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6460 struct cgroup *cgrp = kargs->cgrp;
6461 struct css_set *cset = kargs->cset;
6463 cgroup_threadgroup_change_end(current);
6470 if (kargs->flags & CLONE_INTO_CGROUP) {
6480 * cgroup_can_fork - called on a new task before the process is exposed
6481 * @child: the child process
6482 * @kargs: the arguments passed to create the child process
6484 * This prepares a new css_set for the child process which the child will
6485 * be attached to in cgroup_post_fork().
6486 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6487 * callback returns an error, the fork aborts with that error code. This
6488 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6490 int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs)
6492 struct cgroup_subsys *ss;
6495 ret = cgroup_css_set_fork(kargs);
6499 do_each_subsys_mask(ss, i, have_canfork_callback) {
6500 ret = ss->can_fork(child, kargs->cset);
6503 } while_each_subsys_mask();
6508 for_each_subsys(ss, j) {
6511 if (ss->cancel_fork)
6512 ss->cancel_fork(child, kargs->cset);
6515 cgroup_css_set_put_fork(kargs);
6521 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6522 * @child: the child process
6523 * @kargs: the arguments passed to create the child process
6525 * This calls the cancel_fork() callbacks if a fork failed *after*
6526 * cgroup_can_fork() succeeded and cleans up references we took to
6527 * prepare a new css_set for the child process in cgroup_can_fork().
6529 void cgroup_cancel_fork(struct task_struct *child,
6530 struct kernel_clone_args *kargs)
6532 struct cgroup_subsys *ss;
6535 for_each_subsys(ss, i)
6536 if (ss->cancel_fork)
6537 ss->cancel_fork(child, kargs->cset);
6539 cgroup_css_set_put_fork(kargs);
6543 * cgroup_post_fork - finalize cgroup setup for the child process
6544 * @child: the child process
6545 * @kargs: the arguments passed to create the child process
6547 * Attach the child process to its css_set calling the subsystem fork()
6550 void cgroup_post_fork(struct task_struct *child,
6551 struct kernel_clone_args *kargs)
6552 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex)
6554 unsigned long cgrp_flags = 0;
6556 struct cgroup_subsys *ss;
6557 struct css_set *cset;
6563 spin_lock_irq(&css_set_lock);
6565 /* init tasks are special, only link regular threads */
6566 if (likely(child->pid)) {
6568 cgrp_flags = kargs->cgrp->flags;
6570 cgrp_flags = cset->dfl_cgrp->flags;
6572 WARN_ON_ONCE(!list_empty(&child->cg_list));
6574 css_set_move_task(child, NULL, cset, false);
6580 if (!(child->flags & PF_KTHREAD)) {
6581 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) {
6583 * If the cgroup has to be frozen, the new task has
6584 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6585 * get the task into the frozen state.
6587 spin_lock(&child->sighand->siglock);
6588 WARN_ON_ONCE(child->frozen);
6589 child->jobctl |= JOBCTL_TRAP_FREEZE;
6590 spin_unlock(&child->sighand->siglock);
6593 * Calling cgroup_update_frozen() isn't required here,
6594 * because it will be called anyway a bit later from
6595 * do_freezer_trap(). So we avoid cgroup's transient
6596 * switch from the frozen state and back.
6601 * If the cgroup is to be killed notice it now and take the
6602 * child down right after we finished preparing it for
6605 kill = test_bit(CGRP_KILL, &cgrp_flags);
6608 spin_unlock_irq(&css_set_lock);
6611 * Call ss->fork(). This must happen after @child is linked on
6612 * css_set; otherwise, @child might change state between ->fork()
6613 * and addition to css_set.
6615 do_each_subsys_mask(ss, i, have_fork_callback) {
6617 } while_each_subsys_mask();
6619 /* Make the new cset the root_cset of the new cgroup namespace. */
6620 if (kargs->flags & CLONE_NEWCGROUP) {
6621 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset;
6624 child->nsproxy->cgroup_ns->root_cset = cset;
6628 /* Cgroup has to be killed so take down child immediately. */
6630 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID);
6632 cgroup_css_set_put_fork(kargs);
6636 * cgroup_exit - detach cgroup from exiting task
6637 * @tsk: pointer to task_struct of exiting process
6639 * Description: Detach cgroup from @tsk.
6642 void cgroup_exit(struct task_struct *tsk)
6644 struct cgroup_subsys *ss;
6645 struct css_set *cset;
6648 spin_lock_irq(&css_set_lock);
6650 WARN_ON_ONCE(list_empty(&tsk->cg_list));
6651 cset = task_css_set(tsk);
6652 css_set_move_task(tsk, cset, NULL, false);
6653 list_add_tail(&tsk->cg_list, &cset->dying_tasks);
6657 dec_dl_tasks_cs(tsk);
6659 WARN_ON_ONCE(cgroup_task_frozen(tsk));
6660 if (unlikely(!(tsk->flags & PF_KTHREAD) &&
6661 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags)))
6662 cgroup_update_frozen(task_dfl_cgroup(tsk));
6664 spin_unlock_irq(&css_set_lock);
6666 /* see cgroup_post_fork() for details */
6667 do_each_subsys_mask(ss, i, have_exit_callback) {
6669 } while_each_subsys_mask();
6672 void cgroup_release(struct task_struct *task)
6674 struct cgroup_subsys *ss;
6677 do_each_subsys_mask(ss, ssid, have_release_callback) {
6679 } while_each_subsys_mask();
6681 spin_lock_irq(&css_set_lock);
6682 css_set_skip_task_iters(task_css_set(task), task);
6683 list_del_init(&task->cg_list);
6684 spin_unlock_irq(&css_set_lock);
6687 void cgroup_free(struct task_struct *task)
6689 struct css_set *cset = task_css_set(task);
6693 static int __init cgroup_disable(char *str)
6695 struct cgroup_subsys *ss;
6699 while ((token = strsep(&str, ",")) != NULL) {
6703 for_each_subsys(ss, i) {
6704 if (strcmp(token, ss->name) &&
6705 strcmp(token, ss->legacy_name))
6708 static_branch_disable(cgroup_subsys_enabled_key[i]);
6709 pr_info("Disabling %s control group subsystem\n",
6713 for (i = 0; i < OPT_FEATURE_COUNT; i++) {
6714 if (strcmp(token, cgroup_opt_feature_names[i]))
6716 cgroup_feature_disable_mask |= 1 << i;
6717 pr_info("Disabling %s control group feature\n",
6718 cgroup_opt_feature_names[i]);
6724 __setup("cgroup_disable=", cgroup_disable);
6726 void __init __weak enable_debug_cgroup(void) { }
6728 static int __init enable_cgroup_debug(char *str)
6730 cgroup_debug = true;
6731 enable_debug_cgroup();
6734 __setup("cgroup_debug", enable_cgroup_debug);
6737 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6738 * @dentry: directory dentry of interest
6739 * @ss: subsystem of interest
6741 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6742 * to get the corresponding css and return it. If such css doesn't exist
6743 * or can't be pinned, an ERR_PTR value is returned.
6745 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
6746 struct cgroup_subsys *ss)
6748 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
6749 struct file_system_type *s_type = dentry->d_sb->s_type;
6750 struct cgroup_subsys_state *css = NULL;
6751 struct cgroup *cgrp;
6753 /* is @dentry a cgroup dir? */
6754 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
6755 !kn || kernfs_type(kn) != KERNFS_DIR)
6756 return ERR_PTR(-EBADF);
6761 * This path doesn't originate from kernfs and @kn could already
6762 * have been or be removed at any point. @kn->priv is RCU
6763 * protected for this access. See css_release_work_fn() for details.
6765 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6767 css = cgroup_css(cgrp, ss);
6769 if (!css || !css_tryget_online(css))
6770 css = ERR_PTR(-ENOENT);
6777 * css_from_id - lookup css by id
6778 * @id: the cgroup id
6779 * @ss: cgroup subsys to be looked into
6781 * Returns the css if there's valid one with @id, otherwise returns NULL.
6782 * Should be called under rcu_read_lock().
6784 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
6786 WARN_ON_ONCE(!rcu_read_lock_held());
6787 return idr_find(&ss->css_idr, id);
6791 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6792 * @path: path on the default hierarchy
6794 * Find the cgroup at @path on the default hierarchy, increment its
6795 * reference count and return it. Returns pointer to the found cgroup on
6796 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6797 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6799 struct cgroup *cgroup_get_from_path(const char *path)
6801 struct kernfs_node *kn;
6802 struct cgroup *cgrp = ERR_PTR(-ENOENT);
6803 struct cgroup *root_cgrp;
6805 root_cgrp = current_cgns_cgroup_dfl();
6806 kn = kernfs_walk_and_get(root_cgrp->kn, path);
6810 if (kernfs_type(kn) != KERNFS_DIR) {
6811 cgrp = ERR_PTR(-ENOTDIR);
6817 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
6818 if (!cgrp || !cgroup_tryget(cgrp))
6819 cgrp = ERR_PTR(-ENOENT);
6828 EXPORT_SYMBOL_GPL(cgroup_get_from_path);
6831 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd
6832 * @fd: fd obtained by open(cgroup_dir)
6834 * Find the cgroup from a fd which should be obtained
6835 * by opening a cgroup directory. Returns a pointer to the
6836 * cgroup on success. ERR_PTR is returned if the cgroup
6839 struct cgroup *cgroup_v1v2_get_from_fd(int fd)
6841 struct cgroup *cgrp;
6842 struct fd f = fdget_raw(fd);
6844 return ERR_PTR(-EBADF);
6846 cgrp = cgroup_v1v2_get_from_file(f.file);
6852 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports
6854 * @fd: fd obtained by open(cgroup2_dir)
6856 struct cgroup *cgroup_get_from_fd(int fd)
6858 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd);
6861 return ERR_CAST(cgrp);
6863 if (!cgroup_on_dfl(cgrp)) {
6865 return ERR_PTR(-EBADF);
6869 EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
6871 static u64 power_of_ten(int power)
6880 * cgroup_parse_float - parse a floating number
6881 * @input: input string
6882 * @dec_shift: number of decimal digits to shift
6885 * Parse a decimal floating point number in @input and store the result in
6886 * @v with decimal point right shifted @dec_shift times. For example, if
6887 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6888 * Returns 0 on success, -errno otherwise.
6890 * There's nothing cgroup specific about this function except that it's
6891 * currently the only user.
6893 int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v)
6895 s64 whole, frac = 0;
6896 int fstart = 0, fend = 0, flen;
6898 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend))
6903 flen = fend > fstart ? fend - fstart : 0;
6904 if (flen < dec_shift)
6905 frac *= power_of_ten(dec_shift - flen);
6907 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift));
6909 *v = whole * power_of_ten(dec_shift) + frac;
6914 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6915 * definition in cgroup-defs.h.
6917 #ifdef CONFIG_SOCK_CGROUP_DATA
6919 void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
6921 struct cgroup *cgroup;
6924 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6925 if (in_interrupt()) {
6926 cgroup = &cgrp_dfl_root.cgrp;
6932 struct css_set *cset;
6934 cset = task_css_set(current);
6935 if (likely(cgroup_tryget(cset->dfl_cgrp))) {
6936 cgroup = cset->dfl_cgrp;
6942 skcd->cgroup = cgroup;
6943 cgroup_bpf_get(cgroup);
6947 void cgroup_sk_clone(struct sock_cgroup_data *skcd)
6949 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6952 * We might be cloning a socket which is left in an empty
6953 * cgroup and the cgroup might have already been rmdir'd.
6954 * Don't use cgroup_get_live().
6957 cgroup_bpf_get(cgrp);
6960 void cgroup_sk_free(struct sock_cgroup_data *skcd)
6962 struct cgroup *cgrp = sock_cgroup_ptr(skcd);
6964 cgroup_bpf_put(cgrp);
6968 #endif /* CONFIG_SOCK_CGROUP_DATA */
6971 static ssize_t show_delegatable_files(struct cftype *files, char *buf,
6972 ssize_t size, const char *prefix)
6977 for (cft = files; cft && cft->name[0] != '\0'; cft++) {
6978 if (!(cft->flags & CFTYPE_NS_DELEGATABLE))
6982 ret += snprintf(buf + ret, size - ret, "%s.", prefix);
6984 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name);
6986 if (WARN_ON(ret >= size))
6993 static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr,
6996 struct cgroup_subsys *ss;
7000 ret = show_delegatable_files(cgroup_base_files, buf + ret,
7001 PAGE_SIZE - ret, NULL);
7002 if (cgroup_psi_enabled())
7003 ret += show_delegatable_files(cgroup_psi_files, buf + ret,
7004 PAGE_SIZE - ret, NULL);
7006 for_each_subsys(ss, ssid)
7007 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret,
7009 cgroup_subsys_name[ssid]);
7013 static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate);
7015 static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
7018 return snprintf(buf, PAGE_SIZE,
7021 "memory_localevents\n"
7022 "memory_recursiveprot\n");
7024 static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features);
7026 static struct attribute *cgroup_sysfs_attrs[] = {
7027 &cgroup_delegate_attr.attr,
7028 &cgroup_features_attr.attr,
7032 static const struct attribute_group cgroup_sysfs_attr_group = {
7033 .attrs = cgroup_sysfs_attrs,
7037 static int __init cgroup_sysfs_init(void)
7039 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group);
7041 subsys_initcall(cgroup_sysfs_init);
7043 #endif /* CONFIG_SYSFS */