2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include "cgroup-internal.h"
33 #include <linux/bpf-cgroup.h>
34 #include <linux/cred.h>
35 #include <linux/errno.h>
36 #include <linux/init_task.h>
37 #include <linux/kernel.h>
38 #include <linux/magic.h>
39 #include <linux/mutex.h>
40 #include <linux/mount.h>
41 #include <linux/pagemap.h>
42 #include <linux/proc_fs.h>
43 #include <linux/rcupdate.h>
44 #include <linux/sched.h>
45 #include <linux/sched/task.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/hashtable.h>
51 #include <linux/idr.h>
52 #include <linux/kthread.h>
53 #include <linux/atomic.h>
54 #include <linux/cpuset.h>
55 #include <linux/proc_ns.h>
56 #include <linux/nsproxy.h>
57 #include <linux/file.h>
58 #include <linux/fs_parser.h>
59 #include <linux/sched/cputime.h>
60 #include <linux/psi.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/cgroup.h>
66 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
68 /* let's not notify more than 100 times per second */
69 #define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100)
72 * To avoid confusing the compiler (and generating warnings) with code
73 * that attempts to access what would be a 0-element array (i.e. sized
74 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this
75 * constant expression can be added.
77 #define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0)
80 * cgroup_mutex is the master lock. Any modification to cgroup or its
81 * hierarchy must be performed while holding it.
83 * css_set_lock protects task->cgroups pointer, the list of css_set
84 * objects, and the chain of tasks off each css_set.
86 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
87 * cgroup.h can use them for lockdep annotations.
89 DEFINE_MUTEX(cgroup_mutex
);
90 DEFINE_SPINLOCK(css_set_lock
);
92 #ifdef CONFIG_PROVE_RCU
93 EXPORT_SYMBOL_GPL(cgroup_mutex
);
94 EXPORT_SYMBOL_GPL(css_set_lock
);
97 DEFINE_SPINLOCK(trace_cgroup_path_lock
);
98 char trace_cgroup_path
[TRACE_CGROUP_PATH_LEN
];
99 static bool cgroup_debug __read_mostly
;
102 * Protects cgroup_idr and css_idr so that IDs can be released without
103 * grabbing cgroup_mutex.
105 static DEFINE_SPINLOCK(cgroup_idr_lock
);
108 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
109 * against file removal/re-creation across css hiding.
111 static DEFINE_SPINLOCK(cgroup_file_kn_lock
);
113 DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem
);
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct
*cgroup_destroy_wq
;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 struct cgroup_subsys
*cgroup_subsys
[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name
[] = {
138 #include <linux/cgroup_subsys.h>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true
*cgroup_subsys_enabled_key
[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true
*cgroup_subsys_on_dfl_key
[] = {
159 #include <linux/cgroup_subsys.h>
163 static DEFINE_PER_CPU(struct cgroup_rstat_cpu
, cgrp_dfl_root_rstat_cpu
);
165 /* the default hierarchy */
166 struct cgroup_root cgrp_dfl_root
= { .cgrp
.rstat_cpu
= &cgrp_dfl_root_rstat_cpu
};
167 EXPORT_SYMBOL_GPL(cgrp_dfl_root
);
170 * The default hierarchy always exists but is hidden until mounted for the
171 * first time. This is for backward compatibility.
173 static bool cgrp_dfl_visible
;
175 /* some controllers are not supported in the default hierarchy */
176 static u16 cgrp_dfl_inhibit_ss_mask
;
178 /* some controllers are implicitly enabled on the default hierarchy */
179 static u16 cgrp_dfl_implicit_ss_mask
;
181 /* some controllers can be threaded on the default hierarchy */
182 static u16 cgrp_dfl_threaded_ss_mask
;
184 /* The list of hierarchy roots */
185 LIST_HEAD(cgroup_roots
);
186 static int cgroup_root_count
;
188 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
189 static DEFINE_IDR(cgroup_hierarchy_idr
);
192 * Assign a monotonically increasing serial number to csses. It guarantees
193 * cgroups with bigger numbers are newer than those with smaller numbers.
194 * Also, as csses are always appended to the parent's ->children list, it
195 * guarantees that sibling csses are always sorted in the ascending serial
196 * number order on the list. Protected by cgroup_mutex.
198 static u64 css_serial_nr_next
= 1;
201 * These bitmasks identify subsystems with specific features to avoid
202 * having to do iterative checks repeatedly.
204 static u16 have_fork_callback __read_mostly
;
205 static u16 have_exit_callback __read_mostly
;
206 static u16 have_release_callback __read_mostly
;
207 static u16 have_canfork_callback __read_mostly
;
209 /* cgroup namespace for init task */
210 struct cgroup_namespace init_cgroup_ns
= {
211 .ns
.count
= REFCOUNT_INIT(2),
212 .user_ns
= &init_user_ns
,
213 .ns
.ops
= &cgroupns_operations
,
214 .ns
.inum
= PROC_CGROUP_INIT_INO
,
215 .root_cset
= &init_css_set
,
218 static struct file_system_type cgroup2_fs_type
;
219 static struct cftype cgroup_base_files
[];
221 /* cgroup optional features */
222 enum cgroup_opt_features
{
224 OPT_FEATURE_PRESSURE
,
229 static const char *cgroup_opt_feature_names
[OPT_FEATURE_COUNT
] = {
235 static u16 cgroup_feature_disable_mask __read_mostly
;
237 static int cgroup_apply_control(struct cgroup
*cgrp
);
238 static void cgroup_finalize_control(struct cgroup
*cgrp
, int ret
);
239 static void css_task_iter_skip(struct css_task_iter
*it
,
240 struct task_struct
*task
);
241 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
242 static struct cgroup_subsys_state
*css_create(struct cgroup
*cgrp
,
243 struct cgroup_subsys
*ss
);
244 static void css_release(struct percpu_ref
*ref
);
245 static void kill_css(struct cgroup_subsys_state
*css
);
246 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
247 struct cgroup
*cgrp
, struct cftype cfts
[],
251 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
252 * @ssid: subsys ID of interest
254 * cgroup_subsys_enabled() can only be used with literal subsys names which
255 * is fine for individual subsystems but unsuitable for cgroup core. This
256 * is slower static_key_enabled() based test indexed by @ssid.
258 bool cgroup_ssid_enabled(int ssid
)
260 if (!CGROUP_HAS_SUBSYS_CONFIG
)
263 return static_key_enabled(cgroup_subsys_enabled_key
[ssid
]);
267 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
268 * @cgrp: the cgroup of interest
270 * The default hierarchy is the v2 interface of cgroup and this function
271 * can be used to test whether a cgroup is on the default hierarchy for
272 * cases where a subsystem should behave differently depending on the
275 * List of changed behaviors:
277 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
278 * and "name" are disallowed.
280 * - When mounting an existing superblock, mount options should match.
282 * - rename(2) is disallowed.
284 * - "tasks" is removed. Everything should be at process granularity. Use
285 * "cgroup.procs" instead.
287 * - "cgroup.procs" is not sorted. pids will be unique unless they got
288 * recycled in-between reads.
290 * - "release_agent" and "notify_on_release" are removed. Replacement
291 * notification mechanism will be implemented.
293 * - "cgroup.clone_children" is removed.
295 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
296 * and its descendants contain no task; otherwise, 1. The file also
297 * generates kernfs notification which can be monitored through poll and
298 * [di]notify when the value of the file changes.
300 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
301 * take masks of ancestors with non-empty cpus/mems, instead of being
302 * moved to an ancestor.
304 * - cpuset: a task can be moved into an empty cpuset, and again it takes
305 * masks of ancestors.
307 * - blkcg: blk-throttle becomes properly hierarchical.
309 * - debug: disallowed on the default hierarchy.
311 bool cgroup_on_dfl(const struct cgroup
*cgrp
)
313 return cgrp
->root
== &cgrp_dfl_root
;
316 /* IDR wrappers which synchronize using cgroup_idr_lock */
317 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
322 idr_preload(gfp_mask
);
323 spin_lock_bh(&cgroup_idr_lock
);
324 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
& ~__GFP_DIRECT_RECLAIM
);
325 spin_unlock_bh(&cgroup_idr_lock
);
330 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
334 spin_lock_bh(&cgroup_idr_lock
);
335 ret
= idr_replace(idr
, ptr
, id
);
336 spin_unlock_bh(&cgroup_idr_lock
);
340 static void cgroup_idr_remove(struct idr
*idr
, int id
)
342 spin_lock_bh(&cgroup_idr_lock
);
344 spin_unlock_bh(&cgroup_idr_lock
);
347 static bool cgroup_has_tasks(struct cgroup
*cgrp
)
349 return cgrp
->nr_populated_csets
;
352 bool cgroup_is_threaded(struct cgroup
*cgrp
)
354 return cgrp
->dom_cgrp
!= cgrp
;
357 /* can @cgrp host both domain and threaded children? */
358 static bool cgroup_is_mixable(struct cgroup
*cgrp
)
361 * Root isn't under domain level resource control exempting it from
362 * the no-internal-process constraint, so it can serve as a thread
363 * root and a parent of resource domains at the same time.
365 return !cgroup_parent(cgrp
);
368 /* can @cgrp become a thread root? Should always be true for a thread root */
369 static bool cgroup_can_be_thread_root(struct cgroup
*cgrp
)
371 /* mixables don't care */
372 if (cgroup_is_mixable(cgrp
))
375 /* domain roots can't be nested under threaded */
376 if (cgroup_is_threaded(cgrp
))
379 /* can only have either domain or threaded children */
380 if (cgrp
->nr_populated_domain_children
)
383 /* and no domain controllers can be enabled */
384 if (cgrp
->subtree_control
& ~cgrp_dfl_threaded_ss_mask
)
390 /* is @cgrp root of a threaded subtree? */
391 bool cgroup_is_thread_root(struct cgroup
*cgrp
)
393 /* thread root should be a domain */
394 if (cgroup_is_threaded(cgrp
))
397 /* a domain w/ threaded children is a thread root */
398 if (cgrp
->nr_threaded_children
)
402 * A domain which has tasks and explicit threaded controllers
403 * enabled is a thread root.
405 if (cgroup_has_tasks(cgrp
) &&
406 (cgrp
->subtree_control
& cgrp_dfl_threaded_ss_mask
))
412 /* a domain which isn't connected to the root w/o brekage can't be used */
413 static bool cgroup_is_valid_domain(struct cgroup
*cgrp
)
415 /* the cgroup itself can be a thread root */
416 if (cgroup_is_threaded(cgrp
))
419 /* but the ancestors can't be unless mixable */
420 while ((cgrp
= cgroup_parent(cgrp
))) {
421 if (!cgroup_is_mixable(cgrp
) && cgroup_is_thread_root(cgrp
))
423 if (cgroup_is_threaded(cgrp
))
430 /* subsystems visibly enabled on a cgroup */
431 static u16
cgroup_control(struct cgroup
*cgrp
)
433 struct cgroup
*parent
= cgroup_parent(cgrp
);
434 u16 root_ss_mask
= cgrp
->root
->subsys_mask
;
437 u16 ss_mask
= parent
->subtree_control
;
439 /* threaded cgroups can only have threaded controllers */
440 if (cgroup_is_threaded(cgrp
))
441 ss_mask
&= cgrp_dfl_threaded_ss_mask
;
445 if (cgroup_on_dfl(cgrp
))
446 root_ss_mask
&= ~(cgrp_dfl_inhibit_ss_mask
|
447 cgrp_dfl_implicit_ss_mask
);
451 /* subsystems enabled on a cgroup */
452 static u16
cgroup_ss_mask(struct cgroup
*cgrp
)
454 struct cgroup
*parent
= cgroup_parent(cgrp
);
457 u16 ss_mask
= parent
->subtree_ss_mask
;
459 /* threaded cgroups can only have threaded controllers */
460 if (cgroup_is_threaded(cgrp
))
461 ss_mask
&= cgrp_dfl_threaded_ss_mask
;
465 return cgrp
->root
->subsys_mask
;
469 * cgroup_css - obtain a cgroup's css for the specified subsystem
470 * @cgrp: the cgroup of interest
471 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
473 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
474 * function must be called either under cgroup_mutex or rcu_read_lock() and
475 * the caller is responsible for pinning the returned css if it wants to
476 * keep accessing it outside the said locks. This function may return
477 * %NULL if @cgrp doesn't have @subsys_id enabled.
479 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
480 struct cgroup_subsys
*ss
)
482 if (CGROUP_HAS_SUBSYS_CONFIG
&& ss
)
483 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
484 lockdep_is_held(&cgroup_mutex
));
490 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
491 * @cgrp: the cgroup of interest
492 * @ss: the subsystem of interest
494 * Find and get @cgrp's css associated with @ss. If the css doesn't exist
495 * or is offline, %NULL is returned.
497 static struct cgroup_subsys_state
*cgroup_tryget_css(struct cgroup
*cgrp
,
498 struct cgroup_subsys
*ss
)
500 struct cgroup_subsys_state
*css
;
503 css
= cgroup_css(cgrp
, ss
);
504 if (css
&& !css_tryget_online(css
))
512 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
513 * @cgrp: the cgroup of interest
514 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
516 * Similar to cgroup_css() but returns the effective css, which is defined
517 * as the matching css of the nearest ancestor including self which has @ss
518 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
519 * function is guaranteed to return non-NULL css.
521 static struct cgroup_subsys_state
*cgroup_e_css_by_mask(struct cgroup
*cgrp
,
522 struct cgroup_subsys
*ss
)
524 lockdep_assert_held(&cgroup_mutex
);
530 * This function is used while updating css associations and thus
531 * can't test the csses directly. Test ss_mask.
533 while (!(cgroup_ss_mask(cgrp
) & (1 << ss
->id
))) {
534 cgrp
= cgroup_parent(cgrp
);
539 return cgroup_css(cgrp
, ss
);
543 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
544 * @cgrp: the cgroup of interest
545 * @ss: the subsystem of interest
547 * Find and get the effective css of @cgrp for @ss. The effective css is
548 * defined as the matching css of the nearest ancestor including self which
549 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
550 * the root css is returned, so this function always returns a valid css.
552 * The returned css is not guaranteed to be online, and therefore it is the
553 * callers responsibility to try get a reference for it.
555 struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
556 struct cgroup_subsys
*ss
)
558 struct cgroup_subsys_state
*css
;
560 if (!CGROUP_HAS_SUBSYS_CONFIG
)
564 css
= cgroup_css(cgrp
, ss
);
568 cgrp
= cgroup_parent(cgrp
);
571 return init_css_set
.subsys
[ss
->id
];
575 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
576 * @cgrp: the cgroup of interest
577 * @ss: the subsystem of interest
579 * Find and get the effective css of @cgrp for @ss. The effective css is
580 * defined as the matching css of the nearest ancestor including self which
581 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
582 * the root css is returned, so this function always returns a valid css.
583 * The returned css must be put using css_put().
585 struct cgroup_subsys_state
*cgroup_get_e_css(struct cgroup
*cgrp
,
586 struct cgroup_subsys
*ss
)
588 struct cgroup_subsys_state
*css
;
590 if (!CGROUP_HAS_SUBSYS_CONFIG
)
596 css
= cgroup_css(cgrp
, ss
);
598 if (css
&& css_tryget_online(css
))
600 cgrp
= cgroup_parent(cgrp
);
603 css
= init_css_set
.subsys
[ss
->id
];
609 EXPORT_SYMBOL_GPL(cgroup_get_e_css
);
611 static void cgroup_get_live(struct cgroup
*cgrp
)
613 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
614 css_get(&cgrp
->self
);
618 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
619 * is responsible for taking the css_set_lock.
620 * @cgrp: the cgroup in question
622 int __cgroup_task_count(const struct cgroup
*cgrp
)
625 struct cgrp_cset_link
*link
;
627 lockdep_assert_held(&css_set_lock
);
629 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
630 count
+= link
->cset
->nr_tasks
;
636 * cgroup_task_count - count the number of tasks in a cgroup.
637 * @cgrp: the cgroup in question
639 int cgroup_task_count(const struct cgroup
*cgrp
)
643 spin_lock_irq(&css_set_lock
);
644 count
= __cgroup_task_count(cgrp
);
645 spin_unlock_irq(&css_set_lock
);
650 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
652 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
653 struct cftype
*cft
= of_cft(of
);
656 * This is open and unprotected implementation of cgroup_css().
657 * seq_css() is only called from a kernfs file operation which has
658 * an active reference on the file. Because all the subsystem
659 * files are drained before a css is disassociated with a cgroup,
660 * the matching css from the cgroup's subsys table is guaranteed to
661 * be and stay valid until the enclosing operation is complete.
663 if (CGROUP_HAS_SUBSYS_CONFIG
&& cft
->ss
)
664 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
668 EXPORT_SYMBOL_GPL(of_css
);
671 * for_each_css - iterate all css's of a cgroup
672 * @css: the iteration cursor
673 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
674 * @cgrp: the target cgroup to iterate css's of
676 * Should be called under cgroup_[tree_]mutex.
678 #define for_each_css(css, ssid, cgrp) \
679 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
680 if (!((css) = rcu_dereference_check( \
681 (cgrp)->subsys[(ssid)], \
682 lockdep_is_held(&cgroup_mutex)))) { } \
686 * for_each_e_css - iterate all effective css's of a cgroup
687 * @css: the iteration cursor
688 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
689 * @cgrp: the target cgroup to iterate css's of
691 * Should be called under cgroup_[tree_]mutex.
693 #define for_each_e_css(css, ssid, cgrp) \
694 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
695 if (!((css) = cgroup_e_css_by_mask(cgrp, \
696 cgroup_subsys[(ssid)]))) \
701 * do_each_subsys_mask - filter for_each_subsys with a bitmask
702 * @ss: the iteration cursor
703 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
704 * @ss_mask: the bitmask
706 * The block will only run for cases where the ssid-th bit (1 << ssid) of
709 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
710 unsigned long __ss_mask = (ss_mask); \
711 if (!CGROUP_HAS_SUBSYS_CONFIG) { \
715 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
716 (ss) = cgroup_subsys[ssid]; \
719 #define while_each_subsys_mask() \
724 /* iterate over child cgrps, lock should be held throughout iteration */
725 #define cgroup_for_each_live_child(child, cgrp) \
726 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
727 if (({ lockdep_assert_held(&cgroup_mutex); \
728 cgroup_is_dead(child); })) \
732 /* walk live descendants in pre order */
733 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
734 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
735 if (({ lockdep_assert_held(&cgroup_mutex); \
736 (dsct) = (d_css)->cgroup; \
737 cgroup_is_dead(dsct); })) \
741 /* walk live descendants in postorder */
742 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
743 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
744 if (({ lockdep_assert_held(&cgroup_mutex); \
745 (dsct) = (d_css)->cgroup; \
746 cgroup_is_dead(dsct); })) \
751 * The default css_set - used by init and its children prior to any
752 * hierarchies being mounted. It contains a pointer to the root state
753 * for each subsystem. Also used to anchor the list of css_sets. Not
754 * reference-counted, to improve performance when child cgroups
755 * haven't been created.
757 struct css_set init_css_set
= {
758 .refcount
= REFCOUNT_INIT(1),
759 .dom_cset
= &init_css_set
,
760 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
761 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
762 .dying_tasks
= LIST_HEAD_INIT(init_css_set
.dying_tasks
),
763 .task_iters
= LIST_HEAD_INIT(init_css_set
.task_iters
),
764 .threaded_csets
= LIST_HEAD_INIT(init_css_set
.threaded_csets
),
765 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
766 .mg_src_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_src_preload_node
),
767 .mg_dst_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_dst_preload_node
),
768 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
771 * The following field is re-initialized when this cset gets linked
772 * in cgroup_init(). However, let's initialize the field
773 * statically too so that the default cgroup can be accessed safely
776 .dfl_cgrp
= &cgrp_dfl_root
.cgrp
,
779 static int css_set_count
= 1; /* 1 for init_css_set */
781 static bool css_set_threaded(struct css_set
*cset
)
783 return cset
->dom_cset
!= cset
;
787 * css_set_populated - does a css_set contain any tasks?
788 * @cset: target css_set
790 * css_set_populated() should be the same as !!cset->nr_tasks at steady
791 * state. However, css_set_populated() can be called while a task is being
792 * added to or removed from the linked list before the nr_tasks is
793 * properly updated. Hence, we can't just look at ->nr_tasks here.
795 static bool css_set_populated(struct css_set
*cset
)
797 lockdep_assert_held(&css_set_lock
);
799 return !list_empty(&cset
->tasks
) || !list_empty(&cset
->mg_tasks
);
803 * cgroup_update_populated - update the populated count of a cgroup
804 * @cgrp: the target cgroup
805 * @populated: inc or dec populated count
807 * One of the css_sets associated with @cgrp is either getting its first
808 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
809 * count is propagated towards root so that a given cgroup's
810 * nr_populated_children is zero iff none of its descendants contain any
813 * @cgrp's interface file "cgroup.populated" is zero if both
814 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
815 * 1 otherwise. When the sum changes from or to zero, userland is notified
816 * that the content of the interface file has changed. This can be used to
817 * detect when @cgrp and its descendants become populated or empty.
819 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
821 struct cgroup
*child
= NULL
;
822 int adj
= populated
? 1 : -1;
824 lockdep_assert_held(&css_set_lock
);
827 bool was_populated
= cgroup_is_populated(cgrp
);
830 cgrp
->nr_populated_csets
+= adj
;
832 if (cgroup_is_threaded(child
))
833 cgrp
->nr_populated_threaded_children
+= adj
;
835 cgrp
->nr_populated_domain_children
+= adj
;
838 if (was_populated
== cgroup_is_populated(cgrp
))
841 cgroup1_check_for_release(cgrp
);
842 TRACE_CGROUP_PATH(notify_populated
, cgrp
,
843 cgroup_is_populated(cgrp
));
844 cgroup_file_notify(&cgrp
->events_file
);
847 cgrp
= cgroup_parent(cgrp
);
852 * css_set_update_populated - update populated state of a css_set
853 * @cset: target css_set
854 * @populated: whether @cset is populated or depopulated
856 * @cset is either getting the first task or losing the last. Update the
857 * populated counters of all associated cgroups accordingly.
859 static void css_set_update_populated(struct css_set
*cset
, bool populated
)
861 struct cgrp_cset_link
*link
;
863 lockdep_assert_held(&css_set_lock
);
865 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
)
866 cgroup_update_populated(link
->cgrp
, populated
);
870 * @task is leaving, advance task iterators which are pointing to it so
871 * that they can resume at the next position. Advancing an iterator might
872 * remove it from the list, use safe walk. See css_task_iter_skip() for
875 static void css_set_skip_task_iters(struct css_set
*cset
,
876 struct task_struct
*task
)
878 struct css_task_iter
*it
, *pos
;
880 list_for_each_entry_safe(it
, pos
, &cset
->task_iters
, iters_node
)
881 css_task_iter_skip(it
, task
);
885 * css_set_move_task - move a task from one css_set to another
886 * @task: task being moved
887 * @from_cset: css_set @task currently belongs to (may be NULL)
888 * @to_cset: new css_set @task is being moved to (may be NULL)
889 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
891 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
892 * css_set, @from_cset can be NULL. If @task is being disassociated
893 * instead of moved, @to_cset can be NULL.
895 * This function automatically handles populated counter updates and
896 * css_task_iter adjustments but the caller is responsible for managing
897 * @from_cset and @to_cset's reference counts.
899 static void css_set_move_task(struct task_struct
*task
,
900 struct css_set
*from_cset
, struct css_set
*to_cset
,
903 lockdep_assert_held(&css_set_lock
);
905 if (to_cset
&& !css_set_populated(to_cset
))
906 css_set_update_populated(to_cset
, true);
909 WARN_ON_ONCE(list_empty(&task
->cg_list
));
911 css_set_skip_task_iters(from_cset
, task
);
912 list_del_init(&task
->cg_list
);
913 if (!css_set_populated(from_cset
))
914 css_set_update_populated(from_cset
, false);
916 WARN_ON_ONCE(!list_empty(&task
->cg_list
));
921 * We are synchronized through cgroup_threadgroup_rwsem
922 * against PF_EXITING setting such that we can't race
923 * against cgroup_exit()/cgroup_free() dropping the css_set.
925 WARN_ON_ONCE(task
->flags
& PF_EXITING
);
927 cgroup_move_task(task
, to_cset
);
928 list_add_tail(&task
->cg_list
, use_mg_tasks
? &to_cset
->mg_tasks
:
934 * hash table for cgroup groups. This improves the performance to find
935 * an existing css_set. This hash doesn't (currently) take into
936 * account cgroups in empty hierarchies.
938 #define CSS_SET_HASH_BITS 7
939 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
941 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
943 unsigned long key
= 0UL;
944 struct cgroup_subsys
*ss
;
947 for_each_subsys(ss
, i
)
948 key
+= (unsigned long)css
[i
];
949 key
= (key
>> 16) ^ key
;
954 void put_css_set_locked(struct css_set
*cset
)
956 struct cgrp_cset_link
*link
, *tmp_link
;
957 struct cgroup_subsys
*ss
;
960 lockdep_assert_held(&css_set_lock
);
962 if (!refcount_dec_and_test(&cset
->refcount
))
965 WARN_ON_ONCE(!list_empty(&cset
->threaded_csets
));
967 /* This css_set is dead. Unlink it and release cgroup and css refs */
968 for_each_subsys(ss
, ssid
) {
969 list_del(&cset
->e_cset_node
[ssid
]);
970 css_put(cset
->subsys
[ssid
]);
972 hash_del(&cset
->hlist
);
975 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
976 list_del(&link
->cset_link
);
977 list_del(&link
->cgrp_link
);
978 if (cgroup_parent(link
->cgrp
))
979 cgroup_put(link
->cgrp
);
983 if (css_set_threaded(cset
)) {
984 list_del(&cset
->threaded_csets_node
);
985 put_css_set_locked(cset
->dom_cset
);
988 kfree_rcu(cset
, rcu_head
);
992 * compare_css_sets - helper function for find_existing_css_set().
993 * @cset: candidate css_set being tested
994 * @old_cset: existing css_set for a task
995 * @new_cgrp: cgroup that's being entered by the task
996 * @template: desired set of css pointers in css_set (pre-calculated)
998 * Returns true if "cset" matches "old_cset" except for the hierarchy
999 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
1001 static bool compare_css_sets(struct css_set
*cset
,
1002 struct css_set
*old_cset
,
1003 struct cgroup
*new_cgrp
,
1004 struct cgroup_subsys_state
*template[])
1006 struct cgroup
*new_dfl_cgrp
;
1007 struct list_head
*l1
, *l2
;
1010 * On the default hierarchy, there can be csets which are
1011 * associated with the same set of cgroups but different csses.
1012 * Let's first ensure that csses match.
1014 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
1018 /* @cset's domain should match the default cgroup's */
1019 if (cgroup_on_dfl(new_cgrp
))
1020 new_dfl_cgrp
= new_cgrp
;
1022 new_dfl_cgrp
= old_cset
->dfl_cgrp
;
1024 if (new_dfl_cgrp
->dom_cgrp
!= cset
->dom_cset
->dfl_cgrp
)
1028 * Compare cgroup pointers in order to distinguish between
1029 * different cgroups in hierarchies. As different cgroups may
1030 * share the same effective css, this comparison is always
1033 l1
= &cset
->cgrp_links
;
1034 l2
= &old_cset
->cgrp_links
;
1036 struct cgrp_cset_link
*link1
, *link2
;
1037 struct cgroup
*cgrp1
, *cgrp2
;
1041 /* See if we reached the end - both lists are equal length. */
1042 if (l1
== &cset
->cgrp_links
) {
1043 BUG_ON(l2
!= &old_cset
->cgrp_links
);
1046 BUG_ON(l2
== &old_cset
->cgrp_links
);
1048 /* Locate the cgroups associated with these links. */
1049 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
1050 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
1051 cgrp1
= link1
->cgrp
;
1052 cgrp2
= link2
->cgrp
;
1053 /* Hierarchies should be linked in the same order. */
1054 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
1057 * If this hierarchy is the hierarchy of the cgroup
1058 * that's changing, then we need to check that this
1059 * css_set points to the new cgroup; if it's any other
1060 * hierarchy, then this css_set should point to the
1061 * same cgroup as the old css_set.
1063 if (cgrp1
->root
== new_cgrp
->root
) {
1064 if (cgrp1
!= new_cgrp
)
1075 * find_existing_css_set - init css array and find the matching css_set
1076 * @old_cset: the css_set that we're using before the cgroup transition
1077 * @cgrp: the cgroup that we're moving into
1078 * @template: out param for the new set of csses, should be clear on entry
1080 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
1081 struct cgroup
*cgrp
,
1082 struct cgroup_subsys_state
*template[])
1084 struct cgroup_root
*root
= cgrp
->root
;
1085 struct cgroup_subsys
*ss
;
1086 struct css_set
*cset
;
1091 * Build the set of subsystem state objects that we want to see in the
1092 * new css_set. While subsystems can change globally, the entries here
1093 * won't change, so no need for locking.
1095 for_each_subsys(ss
, i
) {
1096 if (root
->subsys_mask
& (1UL << i
)) {
1098 * @ss is in this hierarchy, so we want the
1099 * effective css from @cgrp.
1101 template[i
] = cgroup_e_css_by_mask(cgrp
, ss
);
1104 * @ss is not in this hierarchy, so we don't want
1105 * to change the css.
1107 template[i
] = old_cset
->subsys
[i
];
1111 key
= css_set_hash(template);
1112 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
1113 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
1116 /* This css_set matches what we need */
1120 /* No existing cgroup group matched */
1124 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
1126 struct cgrp_cset_link
*link
, *tmp_link
;
1128 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
1129 list_del(&link
->cset_link
);
1135 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1136 * @count: the number of links to allocate
1137 * @tmp_links: list_head the allocated links are put on
1139 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1140 * through ->cset_link. Returns 0 on success or -errno.
1142 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
1144 struct cgrp_cset_link
*link
;
1147 INIT_LIST_HEAD(tmp_links
);
1149 for (i
= 0; i
< count
; i
++) {
1150 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
1152 free_cgrp_cset_links(tmp_links
);
1155 list_add(&link
->cset_link
, tmp_links
);
1161 * link_css_set - a helper function to link a css_set to a cgroup
1162 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1163 * @cset: the css_set to be linked
1164 * @cgrp: the destination cgroup
1166 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
1167 struct cgroup
*cgrp
)
1169 struct cgrp_cset_link
*link
;
1171 BUG_ON(list_empty(tmp_links
));
1173 if (cgroup_on_dfl(cgrp
))
1174 cset
->dfl_cgrp
= cgrp
;
1176 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
1181 * Always add links to the tail of the lists so that the lists are
1182 * in chronological order.
1184 list_move_tail(&link
->cset_link
, &cgrp
->cset_links
);
1185 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
1187 if (cgroup_parent(cgrp
))
1188 cgroup_get_live(cgrp
);
1192 * find_css_set - return a new css_set with one cgroup updated
1193 * @old_cset: the baseline css_set
1194 * @cgrp: the cgroup to be updated
1196 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1197 * substituted into the appropriate hierarchy.
1199 static struct css_set
*find_css_set(struct css_set
*old_cset
,
1200 struct cgroup
*cgrp
)
1202 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
1203 struct css_set
*cset
;
1204 struct list_head tmp_links
;
1205 struct cgrp_cset_link
*link
;
1206 struct cgroup_subsys
*ss
;
1210 lockdep_assert_held(&cgroup_mutex
);
1212 /* First see if we already have a cgroup group that matches
1213 * the desired set */
1214 spin_lock_irq(&css_set_lock
);
1215 cset
= find_existing_css_set(old_cset
, cgrp
, template);
1218 spin_unlock_irq(&css_set_lock
);
1223 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
1227 /* Allocate all the cgrp_cset_link objects that we'll need */
1228 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
1233 refcount_set(&cset
->refcount
, 1);
1234 cset
->dom_cset
= cset
;
1235 INIT_LIST_HEAD(&cset
->tasks
);
1236 INIT_LIST_HEAD(&cset
->mg_tasks
);
1237 INIT_LIST_HEAD(&cset
->dying_tasks
);
1238 INIT_LIST_HEAD(&cset
->task_iters
);
1239 INIT_LIST_HEAD(&cset
->threaded_csets
);
1240 INIT_HLIST_NODE(&cset
->hlist
);
1241 INIT_LIST_HEAD(&cset
->cgrp_links
);
1242 INIT_LIST_HEAD(&cset
->mg_src_preload_node
);
1243 INIT_LIST_HEAD(&cset
->mg_dst_preload_node
);
1244 INIT_LIST_HEAD(&cset
->mg_node
);
1246 /* Copy the set of subsystem state objects generated in
1247 * find_existing_css_set() */
1248 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
1250 spin_lock_irq(&css_set_lock
);
1251 /* Add reference counts and links from the new css_set. */
1252 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
1253 struct cgroup
*c
= link
->cgrp
;
1255 if (c
->root
== cgrp
->root
)
1257 link_css_set(&tmp_links
, cset
, c
);
1260 BUG_ON(!list_empty(&tmp_links
));
1264 /* Add @cset to the hash table */
1265 key
= css_set_hash(cset
->subsys
);
1266 hash_add(css_set_table
, &cset
->hlist
, key
);
1268 for_each_subsys(ss
, ssid
) {
1269 struct cgroup_subsys_state
*css
= cset
->subsys
[ssid
];
1271 list_add_tail(&cset
->e_cset_node
[ssid
],
1272 &css
->cgroup
->e_csets
[ssid
]);
1276 spin_unlock_irq(&css_set_lock
);
1279 * If @cset should be threaded, look up the matching dom_cset and
1280 * link them up. We first fully initialize @cset then look for the
1281 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1282 * to stay empty until we return.
1284 if (cgroup_is_threaded(cset
->dfl_cgrp
)) {
1285 struct css_set
*dcset
;
1287 dcset
= find_css_set(cset
, cset
->dfl_cgrp
->dom_cgrp
);
1293 spin_lock_irq(&css_set_lock
);
1294 cset
->dom_cset
= dcset
;
1295 list_add_tail(&cset
->threaded_csets_node
,
1296 &dcset
->threaded_csets
);
1297 spin_unlock_irq(&css_set_lock
);
1303 struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
1305 struct cgroup
*root_cgrp
= kernfs_root_to_node(kf_root
)->priv
;
1307 return root_cgrp
->root
;
1310 void cgroup_favor_dynmods(struct cgroup_root
*root
, bool favor
)
1312 bool favoring
= root
->flags
& CGRP_ROOT_FAVOR_DYNMODS
;
1314 /* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
1315 if (favor
&& !favoring
) {
1316 rcu_sync_enter(&cgroup_threadgroup_rwsem
.rss
);
1317 root
->flags
|= CGRP_ROOT_FAVOR_DYNMODS
;
1318 } else if (!favor
&& favoring
) {
1319 rcu_sync_exit(&cgroup_threadgroup_rwsem
.rss
);
1320 root
->flags
&= ~CGRP_ROOT_FAVOR_DYNMODS
;
1324 static int cgroup_init_root_id(struct cgroup_root
*root
)
1328 lockdep_assert_held(&cgroup_mutex
);
1330 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
1334 root
->hierarchy_id
= id
;
1338 static void cgroup_exit_root_id(struct cgroup_root
*root
)
1340 lockdep_assert_held(&cgroup_mutex
);
1342 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
1345 void cgroup_free_root(struct cgroup_root
*root
)
1350 static void cgroup_destroy_root(struct cgroup_root
*root
)
1352 struct cgroup
*cgrp
= &root
->cgrp
;
1353 struct cgrp_cset_link
*link
, *tmp_link
;
1355 trace_cgroup_destroy_root(root
);
1357 cgroup_lock_and_drain_offline(&cgrp_dfl_root
.cgrp
);
1359 BUG_ON(atomic_read(&root
->nr_cgrps
));
1360 BUG_ON(!list_empty(&cgrp
->self
.children
));
1362 /* Rebind all subsystems back to the default hierarchy */
1363 WARN_ON(rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
));
1366 * Release all the links from cset_links to this hierarchy's
1369 spin_lock_irq(&css_set_lock
);
1371 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
1372 list_del(&link
->cset_link
);
1373 list_del(&link
->cgrp_link
);
1377 spin_unlock_irq(&css_set_lock
);
1379 if (!list_empty(&root
->root_list
)) {
1380 list_del(&root
->root_list
);
1381 cgroup_root_count
--;
1384 cgroup_favor_dynmods(root
, false);
1385 cgroup_exit_root_id(root
);
1387 mutex_unlock(&cgroup_mutex
);
1389 cgroup_rstat_exit(cgrp
);
1390 kernfs_destroy_root(root
->kf_root
);
1391 cgroup_free_root(root
);
1394 static inline struct cgroup
*__cset_cgroup_from_root(struct css_set
*cset
,
1395 struct cgroup_root
*root
)
1397 struct cgroup
*res_cgroup
= NULL
;
1399 if (cset
== &init_css_set
) {
1400 res_cgroup
= &root
->cgrp
;
1401 } else if (root
== &cgrp_dfl_root
) {
1402 res_cgroup
= cset
->dfl_cgrp
;
1404 struct cgrp_cset_link
*link
;
1406 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1407 struct cgroup
*c
= link
->cgrp
;
1409 if (c
->root
== root
) {
1420 * look up cgroup associated with current task's cgroup namespace on the
1421 * specified hierarchy
1423 static struct cgroup
*
1424 current_cgns_cgroup_from_root(struct cgroup_root
*root
)
1426 struct cgroup
*res
= NULL
;
1427 struct css_set
*cset
;
1429 lockdep_assert_held(&css_set_lock
);
1433 cset
= current
->nsproxy
->cgroup_ns
->root_cset
;
1434 res
= __cset_cgroup_from_root(cset
, root
);
1442 /* look up cgroup associated with given css_set on the specified hierarchy */
1443 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
1444 struct cgroup_root
*root
)
1446 struct cgroup
*res
= NULL
;
1448 lockdep_assert_held(&cgroup_mutex
);
1449 lockdep_assert_held(&css_set_lock
);
1451 res
= __cset_cgroup_from_root(cset
, root
);
1458 * Return the cgroup for "task" from the given hierarchy. Must be
1459 * called with cgroup_mutex and css_set_lock held.
1461 struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
1462 struct cgroup_root
*root
)
1465 * No need to lock the task - since we hold css_set_lock the
1466 * task can't change groups.
1468 return cset_cgroup_from_root(task_css_set(task
), root
);
1472 * A task must hold cgroup_mutex to modify cgroups.
1474 * Any task can increment and decrement the count field without lock.
1475 * So in general, code holding cgroup_mutex can't rely on the count
1476 * field not changing. However, if the count goes to zero, then only
1477 * cgroup_attach_task() can increment it again. Because a count of zero
1478 * means that no tasks are currently attached, therefore there is no
1479 * way a task attached to that cgroup can fork (the other way to
1480 * increment the count). So code holding cgroup_mutex can safely
1481 * assume that if the count is zero, it will stay zero. Similarly, if
1482 * a task holds cgroup_mutex on a cgroup with zero count, it
1483 * knows that the cgroup won't be removed, as cgroup_rmdir()
1486 * A cgroup can only be deleted if both its 'count' of using tasks
1487 * is zero, and its list of 'children' cgroups is empty. Since all
1488 * tasks in the system use _some_ cgroup, and since there is always at
1489 * least one task in the system (init, pid == 1), therefore, root cgroup
1490 * always has either children cgroups and/or using tasks. So we don't
1491 * need a special hack to ensure that root cgroup cannot be deleted.
1493 * P.S. One more locking exception. RCU is used to guard the
1494 * update of a tasks cgroup pointer by cgroup_attach_task()
1497 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1499 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1502 struct cgroup_subsys
*ss
= cft
->ss
;
1504 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1505 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
)) {
1506 const char *dbg
= (cft
->flags
& CFTYPE_DEBUG
) ? ".__DEBUG__." : "";
1508 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s%s.%s",
1509 dbg
, cgroup_on_dfl(cgrp
) ? ss
->name
: ss
->legacy_name
,
1512 strscpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1518 * cgroup_file_mode - deduce file mode of a control file
1519 * @cft: the control file in question
1521 * S_IRUGO for read, S_IWUSR for write.
1523 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1527 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1530 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
) {
1531 if (cft
->flags
& CFTYPE_WORLD_WRITABLE
)
1541 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1542 * @subtree_control: the new subtree_control mask to consider
1543 * @this_ss_mask: available subsystems
1545 * On the default hierarchy, a subsystem may request other subsystems to be
1546 * enabled together through its ->depends_on mask. In such cases, more
1547 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1549 * This function calculates which subsystems need to be enabled if
1550 * @subtree_control is to be applied while restricted to @this_ss_mask.
1552 static u16
cgroup_calc_subtree_ss_mask(u16 subtree_control
, u16 this_ss_mask
)
1554 u16 cur_ss_mask
= subtree_control
;
1555 struct cgroup_subsys
*ss
;
1558 lockdep_assert_held(&cgroup_mutex
);
1560 cur_ss_mask
|= cgrp_dfl_implicit_ss_mask
;
1563 u16 new_ss_mask
= cur_ss_mask
;
1565 do_each_subsys_mask(ss
, ssid
, cur_ss_mask
) {
1566 new_ss_mask
|= ss
->depends_on
;
1567 } while_each_subsys_mask();
1570 * Mask out subsystems which aren't available. This can
1571 * happen only if some depended-upon subsystems were bound
1572 * to non-default hierarchies.
1574 new_ss_mask
&= this_ss_mask
;
1576 if (new_ss_mask
== cur_ss_mask
)
1578 cur_ss_mask
= new_ss_mask
;
1585 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1586 * @kn: the kernfs_node being serviced
1588 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1589 * the method finishes if locking succeeded. Note that once this function
1590 * returns the cgroup returned by cgroup_kn_lock_live() may become
1591 * inaccessible any time. If the caller intends to continue to access the
1592 * cgroup, it should pin it before invoking this function.
1594 void cgroup_kn_unlock(struct kernfs_node
*kn
)
1596 struct cgroup
*cgrp
;
1598 if (kernfs_type(kn
) == KERNFS_DIR
)
1601 cgrp
= kn
->parent
->priv
;
1603 mutex_unlock(&cgroup_mutex
);
1605 kernfs_unbreak_active_protection(kn
);
1610 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1611 * @kn: the kernfs_node being serviced
1612 * @drain_offline: perform offline draining on the cgroup
1614 * This helper is to be used by a cgroup kernfs method currently servicing
1615 * @kn. It breaks the active protection, performs cgroup locking and
1616 * verifies that the associated cgroup is alive. Returns the cgroup if
1617 * alive; otherwise, %NULL. A successful return should be undone by a
1618 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1619 * cgroup is drained of offlining csses before return.
1621 * Any cgroup kernfs method implementation which requires locking the
1622 * associated cgroup should use this helper. It avoids nesting cgroup
1623 * locking under kernfs active protection and allows all kernfs operations
1624 * including self-removal.
1626 struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
, bool drain_offline
)
1628 struct cgroup
*cgrp
;
1630 if (kernfs_type(kn
) == KERNFS_DIR
)
1633 cgrp
= kn
->parent
->priv
;
1636 * We're gonna grab cgroup_mutex which nests outside kernfs
1637 * active_ref. cgroup liveliness check alone provides enough
1638 * protection against removal. Ensure @cgrp stays accessible and
1639 * break the active_ref protection.
1641 if (!cgroup_tryget(cgrp
))
1643 kernfs_break_active_protection(kn
);
1646 cgroup_lock_and_drain_offline(cgrp
);
1648 mutex_lock(&cgroup_mutex
);
1650 if (!cgroup_is_dead(cgrp
))
1653 cgroup_kn_unlock(kn
);
1657 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1659 char name
[CGROUP_FILE_NAME_MAX
];
1661 lockdep_assert_held(&cgroup_mutex
);
1663 if (cft
->file_offset
) {
1664 struct cgroup_subsys_state
*css
= cgroup_css(cgrp
, cft
->ss
);
1665 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
1667 spin_lock_irq(&cgroup_file_kn_lock
);
1669 spin_unlock_irq(&cgroup_file_kn_lock
);
1671 del_timer_sync(&cfile
->notify_timer
);
1674 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1678 * css_clear_dir - remove subsys files in a cgroup directory
1681 static void css_clear_dir(struct cgroup_subsys_state
*css
)
1683 struct cgroup
*cgrp
= css
->cgroup
;
1684 struct cftype
*cfts
;
1686 if (!(css
->flags
& CSS_VISIBLE
))
1689 css
->flags
&= ~CSS_VISIBLE
;
1692 if (cgroup_on_dfl(cgrp
))
1693 cfts
= cgroup_base_files
;
1695 cfts
= cgroup1_base_files
;
1697 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1699 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
)
1700 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1705 * css_populate_dir - create subsys files in a cgroup directory
1708 * On failure, no file is added.
1710 static int css_populate_dir(struct cgroup_subsys_state
*css
)
1712 struct cgroup
*cgrp
= css
->cgroup
;
1713 struct cftype
*cfts
, *failed_cfts
;
1716 if ((css
->flags
& CSS_VISIBLE
) || !cgrp
->kn
)
1720 if (cgroup_on_dfl(cgrp
))
1721 cfts
= cgroup_base_files
;
1723 cfts
= cgroup1_base_files
;
1725 ret
= cgroup_addrm_files(&cgrp
->self
, cgrp
, cfts
, true);
1729 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1730 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, true);
1738 css
->flags
|= CSS_VISIBLE
;
1742 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1743 if (cfts
== failed_cfts
)
1745 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1750 int rebind_subsystems(struct cgroup_root
*dst_root
, u16 ss_mask
)
1752 struct cgroup
*dcgrp
= &dst_root
->cgrp
;
1753 struct cgroup_subsys
*ss
;
1755 u16 dfl_disable_ss_mask
= 0;
1757 lockdep_assert_held(&cgroup_mutex
);
1759 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1761 * If @ss has non-root csses attached to it, can't move.
1762 * If @ss is an implicit controller, it is exempt from this
1763 * rule and can be stolen.
1765 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)) &&
1766 !ss
->implicit_on_dfl
)
1769 /* can't move between two non-dummy roots either */
1770 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1774 * Collect ssid's that need to be disabled from default
1777 if (ss
->root
== &cgrp_dfl_root
)
1778 dfl_disable_ss_mask
|= 1 << ssid
;
1780 } while_each_subsys_mask();
1782 if (dfl_disable_ss_mask
) {
1783 struct cgroup
*scgrp
= &cgrp_dfl_root
.cgrp
;
1786 * Controllers from default hierarchy that need to be rebound
1787 * are all disabled together in one go.
1789 cgrp_dfl_root
.subsys_mask
&= ~dfl_disable_ss_mask
;
1790 WARN_ON(cgroup_apply_control(scgrp
));
1791 cgroup_finalize_control(scgrp
, 0);
1794 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1795 struct cgroup_root
*src_root
= ss
->root
;
1796 struct cgroup
*scgrp
= &src_root
->cgrp
;
1797 struct cgroup_subsys_state
*css
= cgroup_css(scgrp
, ss
);
1798 struct css_set
*cset
;
1800 WARN_ON(!css
|| cgroup_css(dcgrp
, ss
));
1802 if (src_root
!= &cgrp_dfl_root
) {
1803 /* disable from the source */
1804 src_root
->subsys_mask
&= ~(1 << ssid
);
1805 WARN_ON(cgroup_apply_control(scgrp
));
1806 cgroup_finalize_control(scgrp
, 0);
1810 RCU_INIT_POINTER(scgrp
->subsys
[ssid
], NULL
);
1811 rcu_assign_pointer(dcgrp
->subsys
[ssid
], css
);
1812 ss
->root
= dst_root
;
1813 css
->cgroup
= dcgrp
;
1815 spin_lock_irq(&css_set_lock
);
1816 hash_for_each(css_set_table
, i
, cset
, hlist
)
1817 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1818 &dcgrp
->e_csets
[ss
->id
]);
1819 spin_unlock_irq(&css_set_lock
);
1821 if (ss
->css_rstat_flush
) {
1822 list_del_rcu(&css
->rstat_css_node
);
1823 list_add_rcu(&css
->rstat_css_node
,
1824 &dcgrp
->rstat_css_list
);
1827 /* default hierarchy doesn't enable controllers by default */
1828 dst_root
->subsys_mask
|= 1 << ssid
;
1829 if (dst_root
== &cgrp_dfl_root
) {
1830 static_branch_enable(cgroup_subsys_on_dfl_key
[ssid
]);
1832 dcgrp
->subtree_control
|= 1 << ssid
;
1833 static_branch_disable(cgroup_subsys_on_dfl_key
[ssid
]);
1836 ret
= cgroup_apply_control(dcgrp
);
1838 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1843 } while_each_subsys_mask();
1845 kernfs_activate(dcgrp
->kn
);
1849 int cgroup_show_path(struct seq_file
*sf
, struct kernfs_node
*kf_node
,
1850 struct kernfs_root
*kf_root
)
1854 struct cgroup_root
*kf_cgroot
= cgroup_root_from_kf(kf_root
);
1855 struct cgroup
*ns_cgroup
;
1857 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
1861 spin_lock_irq(&css_set_lock
);
1862 ns_cgroup
= current_cgns_cgroup_from_root(kf_cgroot
);
1863 len
= kernfs_path_from_node(kf_node
, ns_cgroup
->kn
, buf
, PATH_MAX
);
1864 spin_unlock_irq(&css_set_lock
);
1866 if (len
>= PATH_MAX
)
1869 seq_escape(sf
, buf
, " \t\n\\");
1876 enum cgroup2_param
{
1879 Opt_memory_localevents
,
1880 Opt_memory_recursiveprot
,
1884 static const struct fs_parameter_spec cgroup2_fs_parameters
[] = {
1885 fsparam_flag("nsdelegate", Opt_nsdelegate
),
1886 fsparam_flag("favordynmods", Opt_favordynmods
),
1887 fsparam_flag("memory_localevents", Opt_memory_localevents
),
1888 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot
),
1892 static int cgroup2_parse_param(struct fs_context
*fc
, struct fs_parameter
*param
)
1894 struct cgroup_fs_context
*ctx
= cgroup_fc2context(fc
);
1895 struct fs_parse_result result
;
1898 opt
= fs_parse(fc
, cgroup2_fs_parameters
, param
, &result
);
1903 case Opt_nsdelegate
:
1904 ctx
->flags
|= CGRP_ROOT_NS_DELEGATE
;
1906 case Opt_favordynmods
:
1907 ctx
->flags
|= CGRP_ROOT_FAVOR_DYNMODS
;
1909 case Opt_memory_localevents
:
1910 ctx
->flags
|= CGRP_ROOT_MEMORY_LOCAL_EVENTS
;
1912 case Opt_memory_recursiveprot
:
1913 ctx
->flags
|= CGRP_ROOT_MEMORY_RECURSIVE_PROT
;
1919 static void apply_cgroup_root_flags(unsigned int root_flags
)
1921 if (current
->nsproxy
->cgroup_ns
== &init_cgroup_ns
) {
1922 if (root_flags
& CGRP_ROOT_NS_DELEGATE
)
1923 cgrp_dfl_root
.flags
|= CGRP_ROOT_NS_DELEGATE
;
1925 cgrp_dfl_root
.flags
&= ~CGRP_ROOT_NS_DELEGATE
;
1927 cgroup_favor_dynmods(&cgrp_dfl_root
,
1928 root_flags
& CGRP_ROOT_FAVOR_DYNMODS
);
1930 if (root_flags
& CGRP_ROOT_MEMORY_LOCAL_EVENTS
)
1931 cgrp_dfl_root
.flags
|= CGRP_ROOT_MEMORY_LOCAL_EVENTS
;
1933 cgrp_dfl_root
.flags
&= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS
;
1935 if (root_flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
)
1936 cgrp_dfl_root
.flags
|= CGRP_ROOT_MEMORY_RECURSIVE_PROT
;
1938 cgrp_dfl_root
.flags
&= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT
;
1942 static int cgroup_show_options(struct seq_file
*seq
, struct kernfs_root
*kf_root
)
1944 if (cgrp_dfl_root
.flags
& CGRP_ROOT_NS_DELEGATE
)
1945 seq_puts(seq
, ",nsdelegate");
1946 if (cgrp_dfl_root
.flags
& CGRP_ROOT_FAVOR_DYNMODS
)
1947 seq_puts(seq
, ",favordynmods");
1948 if (cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_LOCAL_EVENTS
)
1949 seq_puts(seq
, ",memory_localevents");
1950 if (cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
)
1951 seq_puts(seq
, ",memory_recursiveprot");
1955 static int cgroup_reconfigure(struct fs_context
*fc
)
1957 struct cgroup_fs_context
*ctx
= cgroup_fc2context(fc
);
1959 apply_cgroup_root_flags(ctx
->flags
);
1963 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1965 struct cgroup_subsys
*ss
;
1968 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1969 INIT_LIST_HEAD(&cgrp
->self
.children
);
1970 INIT_LIST_HEAD(&cgrp
->cset_links
);
1971 INIT_LIST_HEAD(&cgrp
->pidlists
);
1972 mutex_init(&cgrp
->pidlist_mutex
);
1973 cgrp
->self
.cgroup
= cgrp
;
1974 cgrp
->self
.flags
|= CSS_ONLINE
;
1975 cgrp
->dom_cgrp
= cgrp
;
1976 cgrp
->max_descendants
= INT_MAX
;
1977 cgrp
->max_depth
= INT_MAX
;
1978 INIT_LIST_HEAD(&cgrp
->rstat_css_list
);
1979 prev_cputime_init(&cgrp
->prev_cputime
);
1981 for_each_subsys(ss
, ssid
)
1982 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1984 init_waitqueue_head(&cgrp
->offline_waitq
);
1985 INIT_WORK(&cgrp
->release_agent_work
, cgroup1_release_agent
);
1988 void init_cgroup_root(struct cgroup_fs_context
*ctx
)
1990 struct cgroup_root
*root
= ctx
->root
;
1991 struct cgroup
*cgrp
= &root
->cgrp
;
1993 INIT_LIST_HEAD(&root
->root_list
);
1994 atomic_set(&root
->nr_cgrps
, 1);
1996 init_cgroup_housekeeping(cgrp
);
1998 /* DYNMODS must be modified through cgroup_favor_dynmods() */
1999 root
->flags
= ctx
->flags
& ~CGRP_ROOT_FAVOR_DYNMODS
;
2000 if (ctx
->release_agent
)
2001 strscpy(root
->release_agent_path
, ctx
->release_agent
, PATH_MAX
);
2003 strscpy(root
->name
, ctx
->name
, MAX_CGROUP_ROOT_NAMELEN
);
2004 if (ctx
->cpuset_clone_children
)
2005 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
2008 int cgroup_setup_root(struct cgroup_root
*root
, u16 ss_mask
)
2010 LIST_HEAD(tmp_links
);
2011 struct cgroup
*root_cgrp
= &root
->cgrp
;
2012 struct kernfs_syscall_ops
*kf_sops
;
2013 struct css_set
*cset
;
2016 lockdep_assert_held(&cgroup_mutex
);
2018 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
,
2024 * We're accessing css_set_count without locking css_set_lock here,
2025 * but that's OK - it can only be increased by someone holding
2026 * cgroup_lock, and that's us. Later rebinding may disable
2027 * controllers on the default hierarchy and thus create new csets,
2028 * which can't be more than the existing ones. Allocate 2x.
2030 ret
= allocate_cgrp_cset_links(2 * css_set_count
, &tmp_links
);
2034 ret
= cgroup_init_root_id(root
);
2038 kf_sops
= root
== &cgrp_dfl_root
?
2039 &cgroup_kf_syscall_ops
: &cgroup1_kf_syscall_ops
;
2041 root
->kf_root
= kernfs_create_root(kf_sops
,
2042 KERNFS_ROOT_CREATE_DEACTIVATED
|
2043 KERNFS_ROOT_SUPPORT_EXPORTOP
|
2044 KERNFS_ROOT_SUPPORT_USER_XATTR
,
2046 if (IS_ERR(root
->kf_root
)) {
2047 ret
= PTR_ERR(root
->kf_root
);
2050 root_cgrp
->kn
= kernfs_root_to_node(root
->kf_root
);
2051 WARN_ON_ONCE(cgroup_ino(root_cgrp
) != 1);
2052 root_cgrp
->ancestor_ids
[0] = cgroup_id(root_cgrp
);
2054 ret
= css_populate_dir(&root_cgrp
->self
);
2058 ret
= cgroup_rstat_init(root_cgrp
);
2062 ret
= rebind_subsystems(root
, ss_mask
);
2066 ret
= cgroup_bpf_inherit(root_cgrp
);
2069 trace_cgroup_setup_root(root
);
2072 * There must be no failure case after here, since rebinding takes
2073 * care of subsystems' refcounts, which are explicitly dropped in
2074 * the failure exit path.
2076 list_add(&root
->root_list
, &cgroup_roots
);
2077 cgroup_root_count
++;
2080 * Link the root cgroup in this hierarchy into all the css_set
2083 spin_lock_irq(&css_set_lock
);
2084 hash_for_each(css_set_table
, i
, cset
, hlist
) {
2085 link_css_set(&tmp_links
, cset
, root_cgrp
);
2086 if (css_set_populated(cset
))
2087 cgroup_update_populated(root_cgrp
, true);
2089 spin_unlock_irq(&css_set_lock
);
2091 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
2092 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
2098 cgroup_rstat_exit(root_cgrp
);
2100 kernfs_destroy_root(root
->kf_root
);
2101 root
->kf_root
= NULL
;
2103 cgroup_exit_root_id(root
);
2105 percpu_ref_exit(&root_cgrp
->self
.refcnt
);
2107 free_cgrp_cset_links(&tmp_links
);
2111 int cgroup_do_get_tree(struct fs_context
*fc
)
2113 struct cgroup_fs_context
*ctx
= cgroup_fc2context(fc
);
2116 ctx
->kfc
.root
= ctx
->root
->kf_root
;
2117 if (fc
->fs_type
== &cgroup2_fs_type
)
2118 ctx
->kfc
.magic
= CGROUP2_SUPER_MAGIC
;
2120 ctx
->kfc
.magic
= CGROUP_SUPER_MAGIC
;
2121 ret
= kernfs_get_tree(fc
);
2124 * In non-init cgroup namespace, instead of root cgroup's dentry,
2125 * we return the dentry corresponding to the cgroupns->root_cgrp.
2127 if (!ret
&& ctx
->ns
!= &init_cgroup_ns
) {
2128 struct dentry
*nsdentry
;
2129 struct super_block
*sb
= fc
->root
->d_sb
;
2130 struct cgroup
*cgrp
;
2132 mutex_lock(&cgroup_mutex
);
2133 spin_lock_irq(&css_set_lock
);
2135 cgrp
= cset_cgroup_from_root(ctx
->ns
->root_cset
, ctx
->root
);
2137 spin_unlock_irq(&css_set_lock
);
2138 mutex_unlock(&cgroup_mutex
);
2140 nsdentry
= kernfs_node_dentry(cgrp
->kn
, sb
);
2142 if (IS_ERR(nsdentry
)) {
2143 deactivate_locked_super(sb
);
2144 ret
= PTR_ERR(nsdentry
);
2147 fc
->root
= nsdentry
;
2150 if (!ctx
->kfc
.new_sb_created
)
2151 cgroup_put(&ctx
->root
->cgrp
);
2157 * Destroy a cgroup filesystem context.
2159 static void cgroup_fs_context_free(struct fs_context
*fc
)
2161 struct cgroup_fs_context
*ctx
= cgroup_fc2context(fc
);
2164 kfree(ctx
->release_agent
);
2165 put_cgroup_ns(ctx
->ns
);
2166 kernfs_free_fs_context(fc
);
2170 static int cgroup_get_tree(struct fs_context
*fc
)
2172 struct cgroup_fs_context
*ctx
= cgroup_fc2context(fc
);
2175 cgrp_dfl_visible
= true;
2176 cgroup_get_live(&cgrp_dfl_root
.cgrp
);
2177 ctx
->root
= &cgrp_dfl_root
;
2179 ret
= cgroup_do_get_tree(fc
);
2181 apply_cgroup_root_flags(ctx
->flags
);
2185 static const struct fs_context_operations cgroup_fs_context_ops
= {
2186 .free
= cgroup_fs_context_free
,
2187 .parse_param
= cgroup2_parse_param
,
2188 .get_tree
= cgroup_get_tree
,
2189 .reconfigure
= cgroup_reconfigure
,
2192 static const struct fs_context_operations cgroup1_fs_context_ops
= {
2193 .free
= cgroup_fs_context_free
,
2194 .parse_param
= cgroup1_parse_param
,
2195 .get_tree
= cgroup1_get_tree
,
2196 .reconfigure
= cgroup1_reconfigure
,
2200 * Initialise the cgroup filesystem creation/reconfiguration context. Notably,
2201 * we select the namespace we're going to use.
2203 static int cgroup_init_fs_context(struct fs_context
*fc
)
2205 struct cgroup_fs_context
*ctx
;
2207 ctx
= kzalloc(sizeof(struct cgroup_fs_context
), GFP_KERNEL
);
2211 ctx
->ns
= current
->nsproxy
->cgroup_ns
;
2212 get_cgroup_ns(ctx
->ns
);
2213 fc
->fs_private
= &ctx
->kfc
;
2214 if (fc
->fs_type
== &cgroup2_fs_type
)
2215 fc
->ops
= &cgroup_fs_context_ops
;
2217 fc
->ops
= &cgroup1_fs_context_ops
;
2218 put_user_ns(fc
->user_ns
);
2219 fc
->user_ns
= get_user_ns(ctx
->ns
->user_ns
);
2222 #ifdef CONFIG_CGROUP_FAVOR_DYNMODS
2223 ctx
->flags
|= CGRP_ROOT_FAVOR_DYNMODS
;
2228 static void cgroup_kill_sb(struct super_block
*sb
)
2230 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
2231 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
2234 * If @root doesn't have any children, start killing it.
2235 * This prevents new mounts by disabling percpu_ref_tryget_live().
2237 * And don't kill the default root.
2239 if (list_empty(&root
->cgrp
.self
.children
) && root
!= &cgrp_dfl_root
&&
2240 !percpu_ref_is_dying(&root
->cgrp
.self
.refcnt
)) {
2241 cgroup_bpf_offline(&root
->cgrp
);
2242 percpu_ref_kill(&root
->cgrp
.self
.refcnt
);
2244 cgroup_put(&root
->cgrp
);
2248 struct file_system_type cgroup_fs_type
= {
2250 .init_fs_context
= cgroup_init_fs_context
,
2251 .parameters
= cgroup1_fs_parameters
,
2252 .kill_sb
= cgroup_kill_sb
,
2253 .fs_flags
= FS_USERNS_MOUNT
,
2256 static struct file_system_type cgroup2_fs_type
= {
2258 .init_fs_context
= cgroup_init_fs_context
,
2259 .parameters
= cgroup2_fs_parameters
,
2260 .kill_sb
= cgroup_kill_sb
,
2261 .fs_flags
= FS_USERNS_MOUNT
,
2264 #ifdef CONFIG_CPUSETS
2265 static const struct fs_context_operations cpuset_fs_context_ops
= {
2266 .get_tree
= cgroup1_get_tree
,
2267 .free
= cgroup_fs_context_free
,
2271 * This is ugly, but preserves the userspace API for existing cpuset
2272 * users. If someone tries to mount the "cpuset" filesystem, we
2273 * silently switch it to mount "cgroup" instead
2275 static int cpuset_init_fs_context(struct fs_context
*fc
)
2277 char *agent
= kstrdup("/sbin/cpuset_release_agent", GFP_USER
);
2278 struct cgroup_fs_context
*ctx
;
2281 err
= cgroup_init_fs_context(fc
);
2287 fc
->ops
= &cpuset_fs_context_ops
;
2289 ctx
= cgroup_fc2context(fc
);
2290 ctx
->subsys_mask
= 1 << cpuset_cgrp_id
;
2291 ctx
->flags
|= CGRP_ROOT_NOPREFIX
;
2292 ctx
->release_agent
= agent
;
2294 get_filesystem(&cgroup_fs_type
);
2295 put_filesystem(fc
->fs_type
);
2296 fc
->fs_type
= &cgroup_fs_type
;
2301 static struct file_system_type cpuset_fs_type
= {
2303 .init_fs_context
= cpuset_init_fs_context
,
2304 .fs_flags
= FS_USERNS_MOUNT
,
2308 int cgroup_path_ns_locked(struct cgroup
*cgrp
, char *buf
, size_t buflen
,
2309 struct cgroup_namespace
*ns
)
2311 struct cgroup
*root
= cset_cgroup_from_root(ns
->root_cset
, cgrp
->root
);
2313 return kernfs_path_from_node(cgrp
->kn
, root
->kn
, buf
, buflen
);
2316 int cgroup_path_ns(struct cgroup
*cgrp
, char *buf
, size_t buflen
,
2317 struct cgroup_namespace
*ns
)
2321 mutex_lock(&cgroup_mutex
);
2322 spin_lock_irq(&css_set_lock
);
2324 ret
= cgroup_path_ns_locked(cgrp
, buf
, buflen
, ns
);
2326 spin_unlock_irq(&css_set_lock
);
2327 mutex_unlock(&cgroup_mutex
);
2331 EXPORT_SYMBOL_GPL(cgroup_path_ns
);
2334 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2335 * @task: target task
2336 * @buf: the buffer to write the path into
2337 * @buflen: the length of the buffer
2339 * Determine @task's cgroup on the first (the one with the lowest non-zero
2340 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2341 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2342 * cgroup controller callbacks.
2344 * Return value is the same as kernfs_path().
2346 int task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
2348 struct cgroup_root
*root
;
2349 struct cgroup
*cgrp
;
2350 int hierarchy_id
= 1;
2353 mutex_lock(&cgroup_mutex
);
2354 spin_lock_irq(&css_set_lock
);
2356 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
2359 cgrp
= task_cgroup_from_root(task
, root
);
2360 ret
= cgroup_path_ns_locked(cgrp
, buf
, buflen
, &init_cgroup_ns
);
2362 /* if no hierarchy exists, everyone is in "/" */
2363 ret
= strlcpy(buf
, "/", buflen
);
2366 spin_unlock_irq(&css_set_lock
);
2367 mutex_unlock(&cgroup_mutex
);
2370 EXPORT_SYMBOL_GPL(task_cgroup_path
);
2373 * cgroup_migrate_add_task - add a migration target task to a migration context
2374 * @task: target task
2375 * @mgctx: target migration context
2377 * Add @task, which is a migration target, to @mgctx->tset. This function
2378 * becomes noop if @task doesn't need to be migrated. @task's css_set
2379 * should have been added as a migration source and @task->cg_list will be
2380 * moved from the css_set's tasks list to mg_tasks one.
2382 static void cgroup_migrate_add_task(struct task_struct
*task
,
2383 struct cgroup_mgctx
*mgctx
)
2385 struct css_set
*cset
;
2387 lockdep_assert_held(&css_set_lock
);
2389 /* @task either already exited or can't exit until the end */
2390 if (task
->flags
& PF_EXITING
)
2393 /* cgroup_threadgroup_rwsem protects racing against forks */
2394 WARN_ON_ONCE(list_empty(&task
->cg_list
));
2396 cset
= task_css_set(task
);
2397 if (!cset
->mg_src_cgrp
)
2400 mgctx
->tset
.nr_tasks
++;
2402 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2403 if (list_empty(&cset
->mg_node
))
2404 list_add_tail(&cset
->mg_node
,
2405 &mgctx
->tset
.src_csets
);
2406 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2407 list_add_tail(&cset
->mg_dst_cset
->mg_node
,
2408 &mgctx
->tset
.dst_csets
);
2412 * cgroup_taskset_first - reset taskset and return the first task
2413 * @tset: taskset of interest
2414 * @dst_cssp: output variable for the destination css
2416 * @tset iteration is initialized and the first task is returned.
2418 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
,
2419 struct cgroup_subsys_state
**dst_cssp
)
2421 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
2422 tset
->cur_task
= NULL
;
2424 return cgroup_taskset_next(tset
, dst_cssp
);
2428 * cgroup_taskset_next - iterate to the next task in taskset
2429 * @tset: taskset of interest
2430 * @dst_cssp: output variable for the destination css
2432 * Return the next task in @tset. Iteration must have been initialized
2433 * with cgroup_taskset_first().
2435 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
,
2436 struct cgroup_subsys_state
**dst_cssp
)
2438 struct css_set
*cset
= tset
->cur_cset
;
2439 struct task_struct
*task
= tset
->cur_task
;
2441 while (CGROUP_HAS_SUBSYS_CONFIG
&& &cset
->mg_node
!= tset
->csets
) {
2443 task
= list_first_entry(&cset
->mg_tasks
,
2444 struct task_struct
, cg_list
);
2446 task
= list_next_entry(task
, cg_list
);
2448 if (&task
->cg_list
!= &cset
->mg_tasks
) {
2449 tset
->cur_cset
= cset
;
2450 tset
->cur_task
= task
;
2453 * This function may be called both before and
2454 * after cgroup_taskset_migrate(). The two cases
2455 * can be distinguished by looking at whether @cset
2456 * has its ->mg_dst_cset set.
2458 if (cset
->mg_dst_cset
)
2459 *dst_cssp
= cset
->mg_dst_cset
->subsys
[tset
->ssid
];
2461 *dst_cssp
= cset
->subsys
[tset
->ssid
];
2466 cset
= list_next_entry(cset
, mg_node
);
2474 * cgroup_migrate_execute - migrate a taskset
2475 * @mgctx: migration context
2477 * Migrate tasks in @mgctx as setup by migration preparation functions.
2478 * This function fails iff one of the ->can_attach callbacks fails and
2479 * guarantees that either all or none of the tasks in @mgctx are migrated.
2480 * @mgctx is consumed regardless of success.
2482 static int cgroup_migrate_execute(struct cgroup_mgctx
*mgctx
)
2484 struct cgroup_taskset
*tset
= &mgctx
->tset
;
2485 struct cgroup_subsys
*ss
;
2486 struct task_struct
*task
, *tmp_task
;
2487 struct css_set
*cset
, *tmp_cset
;
2488 int ssid
, failed_ssid
, ret
;
2490 /* check that we can legitimately attach to the cgroup */
2491 if (tset
->nr_tasks
) {
2492 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2493 if (ss
->can_attach
) {
2495 ret
= ss
->can_attach(tset
);
2498 goto out_cancel_attach
;
2501 } while_each_subsys_mask();
2505 * Now that we're guaranteed success, proceed to move all tasks to
2506 * the new cgroup. There are no failure cases after here, so this
2507 * is the commit point.
2509 spin_lock_irq(&css_set_lock
);
2510 list_for_each_entry(cset
, &tset
->src_csets
, mg_node
) {
2511 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
) {
2512 struct css_set
*from_cset
= task_css_set(task
);
2513 struct css_set
*to_cset
= cset
->mg_dst_cset
;
2515 get_css_set(to_cset
);
2516 to_cset
->nr_tasks
++;
2517 css_set_move_task(task
, from_cset
, to_cset
, true);
2518 from_cset
->nr_tasks
--;
2520 * If the source or destination cgroup is frozen,
2521 * the task might require to change its state.
2523 cgroup_freezer_migrate_task(task
, from_cset
->dfl_cgrp
,
2525 put_css_set_locked(from_cset
);
2529 spin_unlock_irq(&css_set_lock
);
2532 * Migration is committed, all target tasks are now on dst_csets.
2533 * Nothing is sensitive to fork() after this point. Notify
2534 * controllers that migration is complete.
2536 tset
->csets
= &tset
->dst_csets
;
2538 if (tset
->nr_tasks
) {
2539 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2544 } while_each_subsys_mask();
2548 goto out_release_tset
;
2551 if (tset
->nr_tasks
) {
2552 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2553 if (ssid
== failed_ssid
)
2555 if (ss
->cancel_attach
) {
2557 ss
->cancel_attach(tset
);
2559 } while_each_subsys_mask();
2562 spin_lock_irq(&css_set_lock
);
2563 list_splice_init(&tset
->dst_csets
, &tset
->src_csets
);
2564 list_for_each_entry_safe(cset
, tmp_cset
, &tset
->src_csets
, mg_node
) {
2565 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2566 list_del_init(&cset
->mg_node
);
2568 spin_unlock_irq(&css_set_lock
);
2571 * Re-initialize the cgroup_taskset structure in case it is reused
2572 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2576 tset
->csets
= &tset
->src_csets
;
2581 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2582 * @dst_cgrp: destination cgroup to test
2584 * On the default hierarchy, except for the mixable, (possible) thread root
2585 * and threaded cgroups, subtree_control must be zero for migration
2586 * destination cgroups with tasks so that child cgroups don't compete
2589 int cgroup_migrate_vet_dst(struct cgroup
*dst_cgrp
)
2591 /* v1 doesn't have any restriction */
2592 if (!cgroup_on_dfl(dst_cgrp
))
2595 /* verify @dst_cgrp can host resources */
2596 if (!cgroup_is_valid_domain(dst_cgrp
->dom_cgrp
))
2600 * If @dst_cgrp is already or can become a thread root or is
2601 * threaded, it doesn't matter.
2603 if (cgroup_can_be_thread_root(dst_cgrp
) || cgroup_is_threaded(dst_cgrp
))
2606 /* apply no-internal-process constraint */
2607 if (dst_cgrp
->subtree_control
)
2614 * cgroup_migrate_finish - cleanup after attach
2615 * @mgctx: migration context
2617 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2618 * those functions for details.
2620 void cgroup_migrate_finish(struct cgroup_mgctx
*mgctx
)
2622 struct css_set
*cset
, *tmp_cset
;
2624 lockdep_assert_held(&cgroup_mutex
);
2626 spin_lock_irq(&css_set_lock
);
2628 list_for_each_entry_safe(cset
, tmp_cset
, &mgctx
->preloaded_src_csets
,
2629 mg_src_preload_node
) {
2630 cset
->mg_src_cgrp
= NULL
;
2631 cset
->mg_dst_cgrp
= NULL
;
2632 cset
->mg_dst_cset
= NULL
;
2633 list_del_init(&cset
->mg_src_preload_node
);
2634 put_css_set_locked(cset
);
2637 list_for_each_entry_safe(cset
, tmp_cset
, &mgctx
->preloaded_dst_csets
,
2638 mg_dst_preload_node
) {
2639 cset
->mg_src_cgrp
= NULL
;
2640 cset
->mg_dst_cgrp
= NULL
;
2641 cset
->mg_dst_cset
= NULL
;
2642 list_del_init(&cset
->mg_dst_preload_node
);
2643 put_css_set_locked(cset
);
2646 spin_unlock_irq(&css_set_lock
);
2650 * cgroup_migrate_add_src - add a migration source css_set
2651 * @src_cset: the source css_set to add
2652 * @dst_cgrp: the destination cgroup
2653 * @mgctx: migration context
2655 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2656 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2657 * up by cgroup_migrate_finish().
2659 * This function may be called without holding cgroup_threadgroup_rwsem
2660 * even if the target is a process. Threads may be created and destroyed
2661 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2662 * into play and the preloaded css_sets are guaranteed to cover all
2665 void cgroup_migrate_add_src(struct css_set
*src_cset
,
2666 struct cgroup
*dst_cgrp
,
2667 struct cgroup_mgctx
*mgctx
)
2669 struct cgroup
*src_cgrp
;
2671 lockdep_assert_held(&cgroup_mutex
);
2672 lockdep_assert_held(&css_set_lock
);
2675 * If ->dead, @src_set is associated with one or more dead cgroups
2676 * and doesn't contain any migratable tasks. Ignore it early so
2677 * that the rest of migration path doesn't get confused by it.
2682 if (!list_empty(&src_cset
->mg_src_preload_node
))
2685 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2687 WARN_ON(src_cset
->mg_src_cgrp
);
2688 WARN_ON(src_cset
->mg_dst_cgrp
);
2689 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2690 WARN_ON(!list_empty(&src_cset
->mg_node
));
2692 src_cset
->mg_src_cgrp
= src_cgrp
;
2693 src_cset
->mg_dst_cgrp
= dst_cgrp
;
2694 get_css_set(src_cset
);
2695 list_add_tail(&src_cset
->mg_src_preload_node
, &mgctx
->preloaded_src_csets
);
2699 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2700 * @mgctx: migration context
2702 * Tasks are about to be moved and all the source css_sets have been
2703 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2704 * pins all destination css_sets, links each to its source, and append them
2705 * to @mgctx->preloaded_dst_csets.
2707 * This function must be called after cgroup_migrate_add_src() has been
2708 * called on each migration source css_set. After migration is performed
2709 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2712 int cgroup_migrate_prepare_dst(struct cgroup_mgctx
*mgctx
)
2714 struct css_set
*src_cset
, *tmp_cset
;
2716 lockdep_assert_held(&cgroup_mutex
);
2718 /* look up the dst cset for each src cset and link it to src */
2719 list_for_each_entry_safe(src_cset
, tmp_cset
, &mgctx
->preloaded_src_csets
,
2720 mg_src_preload_node
) {
2721 struct css_set
*dst_cset
;
2722 struct cgroup_subsys
*ss
;
2725 dst_cset
= find_css_set(src_cset
, src_cset
->mg_dst_cgrp
);
2729 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2732 * If src cset equals dst, it's noop. Drop the src.
2733 * cgroup_migrate() will skip the cset too. Note that we
2734 * can't handle src == dst as some nodes are used by both.
2736 if (src_cset
== dst_cset
) {
2737 src_cset
->mg_src_cgrp
= NULL
;
2738 src_cset
->mg_dst_cgrp
= NULL
;
2739 list_del_init(&src_cset
->mg_src_preload_node
);
2740 put_css_set(src_cset
);
2741 put_css_set(dst_cset
);
2745 src_cset
->mg_dst_cset
= dst_cset
;
2747 if (list_empty(&dst_cset
->mg_dst_preload_node
))
2748 list_add_tail(&dst_cset
->mg_dst_preload_node
,
2749 &mgctx
->preloaded_dst_csets
);
2751 put_css_set(dst_cset
);
2753 for_each_subsys(ss
, ssid
)
2754 if (src_cset
->subsys
[ssid
] != dst_cset
->subsys
[ssid
])
2755 mgctx
->ss_mask
|= 1 << ssid
;
2762 * cgroup_migrate - migrate a process or task to a cgroup
2763 * @leader: the leader of the process or the task to migrate
2764 * @threadgroup: whether @leader points to the whole process or a single task
2765 * @mgctx: migration context
2767 * Migrate a process or task denoted by @leader. If migrating a process,
2768 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2769 * responsible for invoking cgroup_migrate_add_src() and
2770 * cgroup_migrate_prepare_dst() on the targets before invoking this
2771 * function and following up with cgroup_migrate_finish().
2773 * As long as a controller's ->can_attach() doesn't fail, this function is
2774 * guaranteed to succeed. This means that, excluding ->can_attach()
2775 * failure, when migrating multiple targets, the success or failure can be
2776 * decided for all targets by invoking group_migrate_prepare_dst() before
2777 * actually starting migrating.
2779 int cgroup_migrate(struct task_struct
*leader
, bool threadgroup
,
2780 struct cgroup_mgctx
*mgctx
)
2782 struct task_struct
*task
;
2785 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2786 * already PF_EXITING could be freed from underneath us unless we
2787 * take an rcu_read_lock.
2789 spin_lock_irq(&css_set_lock
);
2793 cgroup_migrate_add_task(task
, mgctx
);
2796 } while_each_thread(leader
, task
);
2798 spin_unlock_irq(&css_set_lock
);
2800 return cgroup_migrate_execute(mgctx
);
2804 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2805 * @dst_cgrp: the cgroup to attach to
2806 * @leader: the task or the leader of the threadgroup to be attached
2807 * @threadgroup: attach the whole threadgroup?
2809 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2811 int cgroup_attach_task(struct cgroup
*dst_cgrp
, struct task_struct
*leader
,
2814 DEFINE_CGROUP_MGCTX(mgctx
);
2815 struct task_struct
*task
;
2818 /* look up all src csets */
2819 spin_lock_irq(&css_set_lock
);
2823 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
, &mgctx
);
2826 } while_each_thread(leader
, task
);
2828 spin_unlock_irq(&css_set_lock
);
2830 /* prepare dst csets and commit */
2831 ret
= cgroup_migrate_prepare_dst(&mgctx
);
2833 ret
= cgroup_migrate(leader
, threadgroup
, &mgctx
);
2835 cgroup_migrate_finish(&mgctx
);
2838 TRACE_CGROUP_PATH(attach_task
, dst_cgrp
, leader
, threadgroup
);
2843 struct task_struct
*cgroup_procs_write_start(char *buf
, bool threadgroup
,
2845 __acquires(&cgroup_threadgroup_rwsem
)
2847 struct task_struct
*tsk
;
2850 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2851 return ERR_PTR(-EINVAL
);
2854 * If we migrate a single thread, we don't care about threadgroup
2855 * stability. If the thread is `current`, it won't exit(2) under our
2856 * hands or change PID through exec(2). We exclude
2857 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2858 * callers by cgroup_mutex.
2859 * Therefore, we can skip the global lock.
2861 lockdep_assert_held(&cgroup_mutex
);
2862 if (pid
|| threadgroup
) {
2863 percpu_down_write(&cgroup_threadgroup_rwsem
);
2871 tsk
= find_task_by_vpid(pid
);
2873 tsk
= ERR_PTR(-ESRCH
);
2874 goto out_unlock_threadgroup
;
2881 tsk
= tsk
->group_leader
;
2884 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2885 * If userland migrates such a kthread to a non-root cgroup, it can
2886 * become trapped in a cpuset, or RT kthread may be born in a
2887 * cgroup with no rt_runtime allocated. Just say no.
2889 if (tsk
->no_cgroup_migration
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2890 tsk
= ERR_PTR(-EINVAL
);
2891 goto out_unlock_threadgroup
;
2894 get_task_struct(tsk
);
2895 goto out_unlock_rcu
;
2897 out_unlock_threadgroup
:
2899 percpu_up_write(&cgroup_threadgroup_rwsem
);
2907 void cgroup_procs_write_finish(struct task_struct
*task
, bool locked
)
2908 __releases(&cgroup_threadgroup_rwsem
)
2910 struct cgroup_subsys
*ss
;
2913 /* release reference from cgroup_procs_write_start() */
2914 put_task_struct(task
);
2917 percpu_up_write(&cgroup_threadgroup_rwsem
);
2918 for_each_subsys(ss
, ssid
)
2919 if (ss
->post_attach
)
2923 static void cgroup_print_ss_mask(struct seq_file
*seq
, u16 ss_mask
)
2925 struct cgroup_subsys
*ss
;
2926 bool printed
= false;
2929 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
2932 seq_puts(seq
, ss
->name
);
2934 } while_each_subsys_mask();
2936 seq_putc(seq
, '\n');
2939 /* show controllers which are enabled from the parent */
2940 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2942 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2944 cgroup_print_ss_mask(seq
, cgroup_control(cgrp
));
2948 /* show controllers which are enabled for a given cgroup's children */
2949 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2951 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2953 cgroup_print_ss_mask(seq
, cgrp
->subtree_control
);
2958 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2959 * @cgrp: root of the subtree to update csses for
2961 * @cgrp's control masks have changed and its subtree's css associations
2962 * need to be updated accordingly. This function looks up all css_sets
2963 * which are attached to the subtree, creates the matching updated css_sets
2964 * and migrates the tasks to the new ones.
2966 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2968 DEFINE_CGROUP_MGCTX(mgctx
);
2969 struct cgroup_subsys_state
*d_css
;
2970 struct cgroup
*dsct
;
2971 struct css_set
*src_cset
;
2975 lockdep_assert_held(&cgroup_mutex
);
2977 /* look up all csses currently attached to @cgrp's subtree */
2978 spin_lock_irq(&css_set_lock
);
2979 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
2980 struct cgrp_cset_link
*link
;
2983 * As cgroup_update_dfl_csses() is only called by
2984 * cgroup_apply_control(). The csses associated with the
2985 * given cgrp will not be affected by changes made to
2986 * its subtree_control file. We can skip them.
2991 list_for_each_entry(link
, &dsct
->cset_links
, cset_link
)
2992 cgroup_migrate_add_src(link
->cset
, dsct
, &mgctx
);
2994 spin_unlock_irq(&css_set_lock
);
2997 * We need to write-lock threadgroup_rwsem while migrating tasks.
2998 * However, if there are no source csets for @cgrp, changing its
2999 * controllers isn't gonna produce any task migrations and the
3000 * write-locking can be skipped safely.
3002 has_tasks
= !list_empty(&mgctx
.preloaded_src_csets
);
3004 percpu_down_write(&cgroup_threadgroup_rwsem
);
3006 /* NULL dst indicates self on default hierarchy */
3007 ret
= cgroup_migrate_prepare_dst(&mgctx
);
3011 spin_lock_irq(&css_set_lock
);
3012 list_for_each_entry(src_cset
, &mgctx
.preloaded_src_csets
,
3013 mg_src_preload_node
) {
3014 struct task_struct
*task
, *ntask
;
3016 /* all tasks in src_csets need to be migrated */
3017 list_for_each_entry_safe(task
, ntask
, &src_cset
->tasks
, cg_list
)
3018 cgroup_migrate_add_task(task
, &mgctx
);
3020 spin_unlock_irq(&css_set_lock
);
3022 ret
= cgroup_migrate_execute(&mgctx
);
3024 cgroup_migrate_finish(&mgctx
);
3026 percpu_up_write(&cgroup_threadgroup_rwsem
);
3031 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
3032 * @cgrp: root of the target subtree
3034 * Because css offlining is asynchronous, userland may try to re-enable a
3035 * controller while the previous css is still around. This function grabs
3036 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
3038 void cgroup_lock_and_drain_offline(struct cgroup
*cgrp
)
3039 __acquires(&cgroup_mutex
)
3041 struct cgroup
*dsct
;
3042 struct cgroup_subsys_state
*d_css
;
3043 struct cgroup_subsys
*ss
;
3047 mutex_lock(&cgroup_mutex
);
3049 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
3050 for_each_subsys(ss
, ssid
) {
3051 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
3054 if (!css
|| !percpu_ref_is_dying(&css
->refcnt
))
3057 cgroup_get_live(dsct
);
3058 prepare_to_wait(&dsct
->offline_waitq
, &wait
,
3059 TASK_UNINTERRUPTIBLE
);
3061 mutex_unlock(&cgroup_mutex
);
3063 finish_wait(&dsct
->offline_waitq
, &wait
);
3072 * cgroup_save_control - save control masks and dom_cgrp of a subtree
3073 * @cgrp: root of the target subtree
3075 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
3076 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3079 static void cgroup_save_control(struct cgroup
*cgrp
)
3081 struct cgroup
*dsct
;
3082 struct cgroup_subsys_state
*d_css
;
3084 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
3085 dsct
->old_subtree_control
= dsct
->subtree_control
;
3086 dsct
->old_subtree_ss_mask
= dsct
->subtree_ss_mask
;
3087 dsct
->old_dom_cgrp
= dsct
->dom_cgrp
;
3092 * cgroup_propagate_control - refresh control masks of a subtree
3093 * @cgrp: root of the target subtree
3095 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3096 * ->subtree_control and propagate controller availability through the
3097 * subtree so that descendants don't have unavailable controllers enabled.
3099 static void cgroup_propagate_control(struct cgroup
*cgrp
)
3101 struct cgroup
*dsct
;
3102 struct cgroup_subsys_state
*d_css
;
3104 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
3105 dsct
->subtree_control
&= cgroup_control(dsct
);
3106 dsct
->subtree_ss_mask
=
3107 cgroup_calc_subtree_ss_mask(dsct
->subtree_control
,
3108 cgroup_ss_mask(dsct
));
3113 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3114 * @cgrp: root of the target subtree
3116 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3117 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3120 static void cgroup_restore_control(struct cgroup
*cgrp
)
3122 struct cgroup
*dsct
;
3123 struct cgroup_subsys_state
*d_css
;
3125 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
3126 dsct
->subtree_control
= dsct
->old_subtree_control
;
3127 dsct
->subtree_ss_mask
= dsct
->old_subtree_ss_mask
;
3128 dsct
->dom_cgrp
= dsct
->old_dom_cgrp
;
3132 static bool css_visible(struct cgroup_subsys_state
*css
)
3134 struct cgroup_subsys
*ss
= css
->ss
;
3135 struct cgroup
*cgrp
= css
->cgroup
;
3137 if (cgroup_control(cgrp
) & (1 << ss
->id
))
3139 if (!(cgroup_ss_mask(cgrp
) & (1 << ss
->id
)))
3141 return cgroup_on_dfl(cgrp
) && ss
->implicit_on_dfl
;
3145 * cgroup_apply_control_enable - enable or show csses according to control
3146 * @cgrp: root of the target subtree
3148 * Walk @cgrp's subtree and create new csses or make the existing ones
3149 * visible. A css is created invisible if it's being implicitly enabled
3150 * through dependency. An invisible css is made visible when the userland
3151 * explicitly enables it.
3153 * Returns 0 on success, -errno on failure. On failure, csses which have
3154 * been processed already aren't cleaned up. The caller is responsible for
3155 * cleaning up with cgroup_apply_control_disable().
3157 static int cgroup_apply_control_enable(struct cgroup
*cgrp
)
3159 struct cgroup
*dsct
;
3160 struct cgroup_subsys_state
*d_css
;
3161 struct cgroup_subsys
*ss
;
3164 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
3165 for_each_subsys(ss
, ssid
) {
3166 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
3168 if (!(cgroup_ss_mask(dsct
) & (1 << ss
->id
)))
3172 css
= css_create(dsct
, ss
);
3174 return PTR_ERR(css
);
3177 WARN_ON_ONCE(percpu_ref_is_dying(&css
->refcnt
));
3179 if (css_visible(css
)) {
3180 ret
= css_populate_dir(css
);
3191 * cgroup_apply_control_disable - kill or hide csses according to control
3192 * @cgrp: root of the target subtree
3194 * Walk @cgrp's subtree and kill and hide csses so that they match
3195 * cgroup_ss_mask() and cgroup_visible_mask().
3197 * A css is hidden when the userland requests it to be disabled while other
3198 * subsystems are still depending on it. The css must not actively control
3199 * resources and be in the vanilla state if it's made visible again later.
3200 * Controllers which may be depended upon should provide ->css_reset() for
3203 static void cgroup_apply_control_disable(struct cgroup
*cgrp
)
3205 struct cgroup
*dsct
;
3206 struct cgroup_subsys_state
*d_css
;
3207 struct cgroup_subsys
*ss
;
3210 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
3211 for_each_subsys(ss
, ssid
) {
3212 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
3217 WARN_ON_ONCE(percpu_ref_is_dying(&css
->refcnt
));
3220 !(cgroup_ss_mask(dsct
) & (1 << ss
->id
))) {
3222 } else if (!css_visible(css
)) {
3232 * cgroup_apply_control - apply control mask updates to the subtree
3233 * @cgrp: root of the target subtree
3235 * subsystems can be enabled and disabled in a subtree using the following
3238 * 1. Call cgroup_save_control() to stash the current state.
3239 * 2. Update ->subtree_control masks in the subtree as desired.
3240 * 3. Call cgroup_apply_control() to apply the changes.
3241 * 4. Optionally perform other related operations.
3242 * 5. Call cgroup_finalize_control() to finish up.
3244 * This function implements step 3 and propagates the mask changes
3245 * throughout @cgrp's subtree, updates csses accordingly and perform
3246 * process migrations.
3248 static int cgroup_apply_control(struct cgroup
*cgrp
)
3252 cgroup_propagate_control(cgrp
);
3254 ret
= cgroup_apply_control_enable(cgrp
);
3259 * At this point, cgroup_e_css_by_mask() results reflect the new csses
3260 * making the following cgroup_update_dfl_csses() properly update
3261 * css associations of all tasks in the subtree.
3263 ret
= cgroup_update_dfl_csses(cgrp
);
3271 * cgroup_finalize_control - finalize control mask update
3272 * @cgrp: root of the target subtree
3273 * @ret: the result of the update
3275 * Finalize control mask update. See cgroup_apply_control() for more info.
3277 static void cgroup_finalize_control(struct cgroup
*cgrp
, int ret
)
3280 cgroup_restore_control(cgrp
);
3281 cgroup_propagate_control(cgrp
);
3284 cgroup_apply_control_disable(cgrp
);
3287 static int cgroup_vet_subtree_control_enable(struct cgroup
*cgrp
, u16 enable
)
3289 u16 domain_enable
= enable
& ~cgrp_dfl_threaded_ss_mask
;
3291 /* if nothing is getting enabled, nothing to worry about */
3295 /* can @cgrp host any resources? */
3296 if (!cgroup_is_valid_domain(cgrp
->dom_cgrp
))
3299 /* mixables don't care */
3300 if (cgroup_is_mixable(cgrp
))
3303 if (domain_enable
) {
3304 /* can't enable domain controllers inside a thread subtree */
3305 if (cgroup_is_thread_root(cgrp
) || cgroup_is_threaded(cgrp
))
3309 * Threaded controllers can handle internal competitions
3310 * and are always allowed inside a (prospective) thread
3313 if (cgroup_can_be_thread_root(cgrp
) || cgroup_is_threaded(cgrp
))
3318 * Controllers can't be enabled for a cgroup with tasks to avoid
3319 * child cgroups competing against tasks.
3321 if (cgroup_has_tasks(cgrp
))
3327 /* change the enabled child controllers for a cgroup in the default hierarchy */
3328 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
3329 char *buf
, size_t nbytes
,
3332 u16 enable
= 0, disable
= 0;
3333 struct cgroup
*cgrp
, *child
;
3334 struct cgroup_subsys
*ss
;
3339 * Parse input - space separated list of subsystem names prefixed
3340 * with either + or -.
3342 buf
= strstrip(buf
);
3343 while ((tok
= strsep(&buf
, " "))) {
3346 do_each_subsys_mask(ss
, ssid
, ~cgrp_dfl_inhibit_ss_mask
) {
3347 if (!cgroup_ssid_enabled(ssid
) ||
3348 strcmp(tok
+ 1, ss
->name
))
3352 enable
|= 1 << ssid
;
3353 disable
&= ~(1 << ssid
);
3354 } else if (*tok
== '-') {
3355 disable
|= 1 << ssid
;
3356 enable
&= ~(1 << ssid
);
3361 } while_each_subsys_mask();
3362 if (ssid
== CGROUP_SUBSYS_COUNT
)
3366 cgrp
= cgroup_kn_lock_live(of
->kn
, true);
3370 for_each_subsys(ss
, ssid
) {
3371 if (enable
& (1 << ssid
)) {
3372 if (cgrp
->subtree_control
& (1 << ssid
)) {
3373 enable
&= ~(1 << ssid
);
3377 if (!(cgroup_control(cgrp
) & (1 << ssid
))) {
3381 } else if (disable
& (1 << ssid
)) {
3382 if (!(cgrp
->subtree_control
& (1 << ssid
))) {
3383 disable
&= ~(1 << ssid
);
3387 /* a child has it enabled? */
3388 cgroup_for_each_live_child(child
, cgrp
) {
3389 if (child
->subtree_control
& (1 << ssid
)) {
3397 if (!enable
&& !disable
) {
3402 ret
= cgroup_vet_subtree_control_enable(cgrp
, enable
);
3406 /* save and update control masks and prepare csses */
3407 cgroup_save_control(cgrp
);
3409 cgrp
->subtree_control
|= enable
;
3410 cgrp
->subtree_control
&= ~disable
;
3412 ret
= cgroup_apply_control(cgrp
);
3413 cgroup_finalize_control(cgrp
, ret
);
3417 kernfs_activate(cgrp
->kn
);
3419 cgroup_kn_unlock(of
->kn
);
3420 return ret
?: nbytes
;
3424 * cgroup_enable_threaded - make @cgrp threaded
3425 * @cgrp: the target cgroup
3427 * Called when "threaded" is written to the cgroup.type interface file and
3428 * tries to make @cgrp threaded and join the parent's resource domain.
3429 * This function is never called on the root cgroup as cgroup.type doesn't
3432 static int cgroup_enable_threaded(struct cgroup
*cgrp
)
3434 struct cgroup
*parent
= cgroup_parent(cgrp
);
3435 struct cgroup
*dom_cgrp
= parent
->dom_cgrp
;
3436 struct cgroup
*dsct
;
3437 struct cgroup_subsys_state
*d_css
;
3440 lockdep_assert_held(&cgroup_mutex
);
3442 /* noop if already threaded */
3443 if (cgroup_is_threaded(cgrp
))
3447 * If @cgroup is populated or has domain controllers enabled, it
3448 * can't be switched. While the below cgroup_can_be_thread_root()
3449 * test can catch the same conditions, that's only when @parent is
3450 * not mixable, so let's check it explicitly.
3452 if (cgroup_is_populated(cgrp
) ||
3453 cgrp
->subtree_control
& ~cgrp_dfl_threaded_ss_mask
)
3456 /* we're joining the parent's domain, ensure its validity */
3457 if (!cgroup_is_valid_domain(dom_cgrp
) ||
3458 !cgroup_can_be_thread_root(dom_cgrp
))
3462 * The following shouldn't cause actual migrations and should
3465 cgroup_save_control(cgrp
);
3467 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
)
3468 if (dsct
== cgrp
|| cgroup_is_threaded(dsct
))
3469 dsct
->dom_cgrp
= dom_cgrp
;
3471 ret
= cgroup_apply_control(cgrp
);
3473 parent
->nr_threaded_children
++;
3475 cgroup_finalize_control(cgrp
, ret
);
3479 static int cgroup_type_show(struct seq_file
*seq
, void *v
)
3481 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3483 if (cgroup_is_threaded(cgrp
))
3484 seq_puts(seq
, "threaded\n");
3485 else if (!cgroup_is_valid_domain(cgrp
))
3486 seq_puts(seq
, "domain invalid\n");
3487 else if (cgroup_is_thread_root(cgrp
))
3488 seq_puts(seq
, "domain threaded\n");
3490 seq_puts(seq
, "domain\n");
3495 static ssize_t
cgroup_type_write(struct kernfs_open_file
*of
, char *buf
,
3496 size_t nbytes
, loff_t off
)
3498 struct cgroup
*cgrp
;
3501 /* only switching to threaded mode is supported */
3502 if (strcmp(strstrip(buf
), "threaded"))
3505 /* drain dying csses before we re-apply (threaded) subtree control */
3506 cgrp
= cgroup_kn_lock_live(of
->kn
, true);
3510 /* threaded can only be enabled */
3511 ret
= cgroup_enable_threaded(cgrp
);
3513 cgroup_kn_unlock(of
->kn
);
3514 return ret
?: nbytes
;
3517 static int cgroup_max_descendants_show(struct seq_file
*seq
, void *v
)
3519 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3520 int descendants
= READ_ONCE(cgrp
->max_descendants
);
3522 if (descendants
== INT_MAX
)
3523 seq_puts(seq
, "max\n");
3525 seq_printf(seq
, "%d\n", descendants
);
3530 static ssize_t
cgroup_max_descendants_write(struct kernfs_open_file
*of
,
3531 char *buf
, size_t nbytes
, loff_t off
)
3533 struct cgroup
*cgrp
;
3537 buf
= strstrip(buf
);
3538 if (!strcmp(buf
, "max")) {
3539 descendants
= INT_MAX
;
3541 ret
= kstrtoint(buf
, 0, &descendants
);
3546 if (descendants
< 0)
3549 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3553 cgrp
->max_descendants
= descendants
;
3555 cgroup_kn_unlock(of
->kn
);
3560 static int cgroup_max_depth_show(struct seq_file
*seq
, void *v
)
3562 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3563 int depth
= READ_ONCE(cgrp
->max_depth
);
3565 if (depth
== INT_MAX
)
3566 seq_puts(seq
, "max\n");
3568 seq_printf(seq
, "%d\n", depth
);
3573 static ssize_t
cgroup_max_depth_write(struct kernfs_open_file
*of
,
3574 char *buf
, size_t nbytes
, loff_t off
)
3576 struct cgroup
*cgrp
;
3580 buf
= strstrip(buf
);
3581 if (!strcmp(buf
, "max")) {
3584 ret
= kstrtoint(buf
, 0, &depth
);
3592 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3596 cgrp
->max_depth
= depth
;
3598 cgroup_kn_unlock(of
->kn
);
3603 static int cgroup_events_show(struct seq_file
*seq
, void *v
)
3605 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3607 seq_printf(seq
, "populated %d\n", cgroup_is_populated(cgrp
));
3608 seq_printf(seq
, "frozen %d\n", test_bit(CGRP_FROZEN
, &cgrp
->flags
));
3613 static int cgroup_stat_show(struct seq_file
*seq
, void *v
)
3615 struct cgroup
*cgroup
= seq_css(seq
)->cgroup
;
3617 seq_printf(seq
, "nr_descendants %d\n",
3618 cgroup
->nr_descendants
);
3619 seq_printf(seq
, "nr_dying_descendants %d\n",
3620 cgroup
->nr_dying_descendants
);
3625 static int __maybe_unused
cgroup_extra_stat_show(struct seq_file
*seq
,
3626 struct cgroup
*cgrp
, int ssid
)
3628 struct cgroup_subsys
*ss
= cgroup_subsys
[ssid
];
3629 struct cgroup_subsys_state
*css
;
3632 if (!ss
->css_extra_stat_show
)
3635 css
= cgroup_tryget_css(cgrp
, ss
);
3639 ret
= ss
->css_extra_stat_show(seq
, css
);
3644 static int cpu_stat_show(struct seq_file
*seq
, void *v
)
3646 struct cgroup __maybe_unused
*cgrp
= seq_css(seq
)->cgroup
;
3649 cgroup_base_stat_cputime_show(seq
);
3650 #ifdef CONFIG_CGROUP_SCHED
3651 ret
= cgroup_extra_stat_show(seq
, cgrp
, cpu_cgrp_id
);
3657 static int cgroup_io_pressure_show(struct seq_file
*seq
, void *v
)
3659 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3660 struct psi_group
*psi
= cgroup_ino(cgrp
) == 1 ? &psi_system
: cgrp
->psi
;
3662 return psi_show(seq
, psi
, PSI_IO
);
3664 static int cgroup_memory_pressure_show(struct seq_file
*seq
, void *v
)
3666 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3667 struct psi_group
*psi
= cgroup_ino(cgrp
) == 1 ? &psi_system
: cgrp
->psi
;
3669 return psi_show(seq
, psi
, PSI_MEM
);
3671 static int cgroup_cpu_pressure_show(struct seq_file
*seq
, void *v
)
3673 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3674 struct psi_group
*psi
= cgroup_ino(cgrp
) == 1 ? &psi_system
: cgrp
->psi
;
3676 return psi_show(seq
, psi
, PSI_CPU
);
3679 static ssize_t
cgroup_pressure_write(struct kernfs_open_file
*of
, char *buf
,
3680 size_t nbytes
, enum psi_res res
)
3682 struct cgroup_file_ctx
*ctx
= of
->priv
;
3683 struct psi_trigger
*new;
3684 struct cgroup
*cgrp
;
3685 struct psi_group
*psi
;
3687 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3692 cgroup_kn_unlock(of
->kn
);
3694 /* Allow only one trigger per file descriptor */
3695 if (ctx
->psi
.trigger
) {
3700 psi
= cgroup_ino(cgrp
) == 1 ? &psi_system
: cgrp
->psi
;
3701 new = psi_trigger_create(psi
, buf
, nbytes
, res
);
3704 return PTR_ERR(new);
3707 smp_store_release(&ctx
->psi
.trigger
, new);
3713 static ssize_t
cgroup_io_pressure_write(struct kernfs_open_file
*of
,
3714 char *buf
, size_t nbytes
,
3717 return cgroup_pressure_write(of
, buf
, nbytes
, PSI_IO
);
3720 static ssize_t
cgroup_memory_pressure_write(struct kernfs_open_file
*of
,
3721 char *buf
, size_t nbytes
,
3724 return cgroup_pressure_write(of
, buf
, nbytes
, PSI_MEM
);
3727 static ssize_t
cgroup_cpu_pressure_write(struct kernfs_open_file
*of
,
3728 char *buf
, size_t nbytes
,
3731 return cgroup_pressure_write(of
, buf
, nbytes
, PSI_CPU
);
3734 static __poll_t
cgroup_pressure_poll(struct kernfs_open_file
*of
,
3737 struct cgroup_file_ctx
*ctx
= of
->priv
;
3739 return psi_trigger_poll(&ctx
->psi
.trigger
, of
->file
, pt
);
3742 static void cgroup_pressure_release(struct kernfs_open_file
*of
)
3744 struct cgroup_file_ctx
*ctx
= of
->priv
;
3746 psi_trigger_destroy(ctx
->psi
.trigger
);
3749 bool cgroup_psi_enabled(void)
3751 return (cgroup_feature_disable_mask
& (1 << OPT_FEATURE_PRESSURE
)) == 0;
3754 #else /* CONFIG_PSI */
3755 bool cgroup_psi_enabled(void)
3760 #endif /* CONFIG_PSI */
3762 static int cgroup_freeze_show(struct seq_file
*seq
, void *v
)
3764 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3766 seq_printf(seq
, "%d\n", cgrp
->freezer
.freeze
);
3771 static ssize_t
cgroup_freeze_write(struct kernfs_open_file
*of
,
3772 char *buf
, size_t nbytes
, loff_t off
)
3774 struct cgroup
*cgrp
;
3778 ret
= kstrtoint(strstrip(buf
), 0, &freeze
);
3782 if (freeze
< 0 || freeze
> 1)
3785 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3789 cgroup_freeze(cgrp
, freeze
);
3791 cgroup_kn_unlock(of
->kn
);
3796 static void __cgroup_kill(struct cgroup
*cgrp
)
3798 struct css_task_iter it
;
3799 struct task_struct
*task
;
3801 lockdep_assert_held(&cgroup_mutex
);
3803 spin_lock_irq(&css_set_lock
);
3804 set_bit(CGRP_KILL
, &cgrp
->flags
);
3805 spin_unlock_irq(&css_set_lock
);
3807 css_task_iter_start(&cgrp
->self
, CSS_TASK_ITER_PROCS
| CSS_TASK_ITER_THREADED
, &it
);
3808 while ((task
= css_task_iter_next(&it
))) {
3809 /* Ignore kernel threads here. */
3810 if (task
->flags
& PF_KTHREAD
)
3813 /* Skip tasks that are already dying. */
3814 if (__fatal_signal_pending(task
))
3817 send_sig(SIGKILL
, task
, 0);
3819 css_task_iter_end(&it
);
3821 spin_lock_irq(&css_set_lock
);
3822 clear_bit(CGRP_KILL
, &cgrp
->flags
);
3823 spin_unlock_irq(&css_set_lock
);
3826 static void cgroup_kill(struct cgroup
*cgrp
)
3828 struct cgroup_subsys_state
*css
;
3829 struct cgroup
*dsct
;
3831 lockdep_assert_held(&cgroup_mutex
);
3833 cgroup_for_each_live_descendant_pre(dsct
, css
, cgrp
)
3834 __cgroup_kill(dsct
);
3837 static ssize_t
cgroup_kill_write(struct kernfs_open_file
*of
, char *buf
,
3838 size_t nbytes
, loff_t off
)
3842 struct cgroup
*cgrp
;
3844 ret
= kstrtoint(strstrip(buf
), 0, &kill
);
3851 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3856 * Killing is a process directed operation, i.e. the whole thread-group
3857 * is taken down so act like we do for cgroup.procs and only make this
3858 * writable in non-threaded cgroups.
3860 if (cgroup_is_threaded(cgrp
))
3865 cgroup_kn_unlock(of
->kn
);
3867 return ret
?: nbytes
;
3870 static int cgroup_file_open(struct kernfs_open_file
*of
)
3872 struct cftype
*cft
= of_cft(of
);
3873 struct cgroup_file_ctx
*ctx
;
3876 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
3880 ctx
->ns
= current
->nsproxy
->cgroup_ns
;
3881 get_cgroup_ns(ctx
->ns
);
3887 ret
= cft
->open(of
);
3889 put_cgroup_ns(ctx
->ns
);
3895 static void cgroup_file_release(struct kernfs_open_file
*of
)
3897 struct cftype
*cft
= of_cft(of
);
3898 struct cgroup_file_ctx
*ctx
= of
->priv
;
3902 put_cgroup_ns(ctx
->ns
);
3906 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
3907 size_t nbytes
, loff_t off
)
3909 struct cgroup_file_ctx
*ctx
= of
->priv
;
3910 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
3911 struct cftype
*cft
= of_cft(of
);
3912 struct cgroup_subsys_state
*css
;
3919 * If namespaces are delegation boundaries, disallow writes to
3920 * files in an non-init namespace root from inside the namespace
3921 * except for the files explicitly marked delegatable -
3922 * cgroup.procs and cgroup.subtree_control.
3924 if ((cgrp
->root
->flags
& CGRP_ROOT_NS_DELEGATE
) &&
3925 !(cft
->flags
& CFTYPE_NS_DELEGATABLE
) &&
3926 ctx
->ns
!= &init_cgroup_ns
&& ctx
->ns
->root_cset
->dfl_cgrp
== cgrp
)
3930 return cft
->write(of
, buf
, nbytes
, off
);
3933 * kernfs guarantees that a file isn't deleted with operations in
3934 * flight, which means that the matching css is and stays alive and
3935 * doesn't need to be pinned. The RCU locking is not necessary
3936 * either. It's just for the convenience of using cgroup_css().
3939 css
= cgroup_css(cgrp
, cft
->ss
);
3942 if (cft
->write_u64
) {
3943 unsigned long long v
;
3944 ret
= kstrtoull(buf
, 0, &v
);
3946 ret
= cft
->write_u64(css
, cft
, v
);
3947 } else if (cft
->write_s64
) {
3949 ret
= kstrtoll(buf
, 0, &v
);
3951 ret
= cft
->write_s64(css
, cft
, v
);
3956 return ret
?: nbytes
;
3959 static __poll_t
cgroup_file_poll(struct kernfs_open_file
*of
, poll_table
*pt
)
3961 struct cftype
*cft
= of_cft(of
);
3964 return cft
->poll(of
, pt
);
3966 return kernfs_generic_poll(of
, pt
);
3969 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
3971 return seq_cft(seq
)->seq_start(seq
, ppos
);
3974 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
3976 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
3979 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
3981 if (seq_cft(seq
)->seq_stop
)
3982 seq_cft(seq
)->seq_stop(seq
, v
);
3985 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
3987 struct cftype
*cft
= seq_cft(m
);
3988 struct cgroup_subsys_state
*css
= seq_css(m
);
3991 return cft
->seq_show(m
, arg
);
3994 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
3995 else if (cft
->read_s64
)
3996 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
4002 static struct kernfs_ops cgroup_kf_single_ops
= {
4003 .atomic_write_len
= PAGE_SIZE
,
4004 .open
= cgroup_file_open
,
4005 .release
= cgroup_file_release
,
4006 .write
= cgroup_file_write
,
4007 .poll
= cgroup_file_poll
,
4008 .seq_show
= cgroup_seqfile_show
,
4011 static struct kernfs_ops cgroup_kf_ops
= {
4012 .atomic_write_len
= PAGE_SIZE
,
4013 .open
= cgroup_file_open
,
4014 .release
= cgroup_file_release
,
4015 .write
= cgroup_file_write
,
4016 .poll
= cgroup_file_poll
,
4017 .seq_start
= cgroup_seqfile_start
,
4018 .seq_next
= cgroup_seqfile_next
,
4019 .seq_stop
= cgroup_seqfile_stop
,
4020 .seq_show
= cgroup_seqfile_show
,
4023 /* set uid and gid of cgroup dirs and files to that of the creator */
4024 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
4026 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
4027 .ia_uid
= current_fsuid(),
4028 .ia_gid
= current_fsgid(), };
4030 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
4031 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
4034 return kernfs_setattr(kn
, &iattr
);
4037 static void cgroup_file_notify_timer(struct timer_list
*timer
)
4039 cgroup_file_notify(container_of(timer
, struct cgroup_file
,
4043 static int cgroup_add_file(struct cgroup_subsys_state
*css
, struct cgroup
*cgrp
,
4046 char name
[CGROUP_FILE_NAME_MAX
];
4047 struct kernfs_node
*kn
;
4048 struct lock_class_key
*key
= NULL
;
4051 #ifdef CONFIG_DEBUG_LOCK_ALLOC
4052 key
= &cft
->lockdep_key
;
4054 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
4055 cgroup_file_mode(cft
),
4056 GLOBAL_ROOT_UID
, GLOBAL_ROOT_GID
,
4057 0, cft
->kf_ops
, cft
,
4062 ret
= cgroup_kn_set_ugid(kn
);
4068 if (cft
->file_offset
) {
4069 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
4071 timer_setup(&cfile
->notify_timer
, cgroup_file_notify_timer
, 0);
4073 spin_lock_irq(&cgroup_file_kn_lock
);
4075 spin_unlock_irq(&cgroup_file_kn_lock
);
4082 * cgroup_addrm_files - add or remove files to a cgroup directory
4083 * @css: the target css
4084 * @cgrp: the target cgroup (usually css->cgroup)
4085 * @cfts: array of cftypes to be added
4086 * @is_add: whether to add or remove
4088 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
4089 * For removals, this function never fails.
4091 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
4092 struct cgroup
*cgrp
, struct cftype cfts
[],
4095 struct cftype
*cft
, *cft_end
= NULL
;
4098 lockdep_assert_held(&cgroup_mutex
);
4101 for (cft
= cfts
; cft
!= cft_end
&& cft
->name
[0] != '\0'; cft
++) {
4102 /* does cft->flags tell us to skip this file on @cgrp? */
4103 if ((cft
->flags
& CFTYPE_PRESSURE
) && !cgroup_psi_enabled())
4105 if ((cft
->flags
& __CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
4107 if ((cft
->flags
& __CFTYPE_NOT_ON_DFL
) && cgroup_on_dfl(cgrp
))
4109 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
4111 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
4113 if ((cft
->flags
& CFTYPE_DEBUG
) && !cgroup_debug
)
4116 ret
= cgroup_add_file(css
, cgrp
, cft
);
4118 pr_warn("%s: failed to add %s, err=%d\n",
4119 __func__
, cft
->name
, ret
);
4125 cgroup_rm_file(cgrp
, cft
);
4131 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
4133 struct cgroup_subsys
*ss
= cfts
[0].ss
;
4134 struct cgroup
*root
= &ss
->root
->cgrp
;
4135 struct cgroup_subsys_state
*css
;
4138 lockdep_assert_held(&cgroup_mutex
);
4140 /* add/rm files for all cgroups created before */
4141 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
4142 struct cgroup
*cgrp
= css
->cgroup
;
4144 if (!(css
->flags
& CSS_VISIBLE
))
4147 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, is_add
);
4153 kernfs_activate(root
->kn
);
4157 static void cgroup_exit_cftypes(struct cftype
*cfts
)
4161 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
4162 /* free copy for custom atomic_write_len, see init_cftypes() */
4163 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
4168 /* revert flags set by cgroup core while adding @cfts */
4169 cft
->flags
&= ~(__CFTYPE_ONLY_ON_DFL
| __CFTYPE_NOT_ON_DFL
);
4173 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
4177 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
4178 struct kernfs_ops
*kf_ops
;
4180 WARN_ON(cft
->ss
|| cft
->kf_ops
);
4182 if ((cft
->flags
& CFTYPE_PRESSURE
) && !cgroup_psi_enabled())
4186 kf_ops
= &cgroup_kf_ops
;
4188 kf_ops
= &cgroup_kf_single_ops
;
4191 * Ugh... if @cft wants a custom max_write_len, we need to
4192 * make a copy of kf_ops to set its atomic_write_len.
4194 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
4195 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
4197 cgroup_exit_cftypes(cfts
);
4200 kf_ops
->atomic_write_len
= cft
->max_write_len
;
4203 cft
->kf_ops
= kf_ops
;
4210 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
4212 lockdep_assert_held(&cgroup_mutex
);
4214 if (!cfts
|| !cfts
[0].ss
)
4217 list_del(&cfts
->node
);
4218 cgroup_apply_cftypes(cfts
, false);
4219 cgroup_exit_cftypes(cfts
);
4224 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4225 * @cfts: zero-length name terminated array of cftypes
4227 * Unregister @cfts. Files described by @cfts are removed from all
4228 * existing cgroups and all future cgroups won't have them either. This
4229 * function can be called anytime whether @cfts' subsys is attached or not.
4231 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4234 int cgroup_rm_cftypes(struct cftype
*cfts
)
4238 mutex_lock(&cgroup_mutex
);
4239 ret
= cgroup_rm_cftypes_locked(cfts
);
4240 mutex_unlock(&cgroup_mutex
);
4245 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4246 * @ss: target cgroup subsystem
4247 * @cfts: zero-length name terminated array of cftypes
4249 * Register @cfts to @ss. Files described by @cfts are created for all
4250 * existing cgroups to which @ss is attached and all future cgroups will
4251 * have them too. This function can be called anytime whether @ss is
4254 * Returns 0 on successful registration, -errno on failure. Note that this
4255 * function currently returns 0 as long as @cfts registration is successful
4256 * even if some file creation attempts on existing cgroups fail.
4258 static int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
4262 if (!cgroup_ssid_enabled(ss
->id
))
4265 if (!cfts
|| cfts
[0].name
[0] == '\0')
4268 ret
= cgroup_init_cftypes(ss
, cfts
);
4272 mutex_lock(&cgroup_mutex
);
4274 list_add_tail(&cfts
->node
, &ss
->cfts
);
4275 ret
= cgroup_apply_cftypes(cfts
, true);
4277 cgroup_rm_cftypes_locked(cfts
);
4279 mutex_unlock(&cgroup_mutex
);
4284 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4285 * @ss: target cgroup subsystem
4286 * @cfts: zero-length name terminated array of cftypes
4288 * Similar to cgroup_add_cftypes() but the added files are only used for
4289 * the default hierarchy.
4291 int cgroup_add_dfl_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
4295 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
4296 cft
->flags
|= __CFTYPE_ONLY_ON_DFL
;
4297 return cgroup_add_cftypes(ss
, cfts
);
4301 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4302 * @ss: target cgroup subsystem
4303 * @cfts: zero-length name terminated array of cftypes
4305 * Similar to cgroup_add_cftypes() but the added files are only used for
4306 * the legacy hierarchies.
4308 int cgroup_add_legacy_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
4312 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
4313 cft
->flags
|= __CFTYPE_NOT_ON_DFL
;
4314 return cgroup_add_cftypes(ss
, cfts
);
4318 * cgroup_file_notify - generate a file modified event for a cgroup_file
4319 * @cfile: target cgroup_file
4321 * @cfile must have been obtained by setting cftype->file_offset.
4323 void cgroup_file_notify(struct cgroup_file
*cfile
)
4325 unsigned long flags
;
4327 spin_lock_irqsave(&cgroup_file_kn_lock
, flags
);
4329 unsigned long last
= cfile
->notified_at
;
4330 unsigned long next
= last
+ CGROUP_FILE_NOTIFY_MIN_INTV
;
4332 if (time_in_range(jiffies
, last
, next
)) {
4333 timer_reduce(&cfile
->notify_timer
, next
);
4335 kernfs_notify(cfile
->kn
);
4336 cfile
->notified_at
= jiffies
;
4339 spin_unlock_irqrestore(&cgroup_file_kn_lock
, flags
);
4343 * css_next_child - find the next child of a given css
4344 * @pos: the current position (%NULL to initiate traversal)
4345 * @parent: css whose children to walk
4347 * This function returns the next child of @parent and should be called
4348 * under either cgroup_mutex or RCU read lock. The only requirement is
4349 * that @parent and @pos are accessible. The next sibling is guaranteed to
4350 * be returned regardless of their states.
4352 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4353 * css which finished ->css_online() is guaranteed to be visible in the
4354 * future iterations and will stay visible until the last reference is put.
4355 * A css which hasn't finished ->css_online() or already finished
4356 * ->css_offline() may show up during traversal. It's each subsystem's
4357 * responsibility to synchronize against on/offlining.
4359 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
4360 struct cgroup_subsys_state
*parent
)
4362 struct cgroup_subsys_state
*next
;
4364 cgroup_assert_mutex_or_rcu_locked();
4367 * @pos could already have been unlinked from the sibling list.
4368 * Once a cgroup is removed, its ->sibling.next is no longer
4369 * updated when its next sibling changes. CSS_RELEASED is set when
4370 * @pos is taken off list, at which time its next pointer is valid,
4371 * and, as releases are serialized, the one pointed to by the next
4372 * pointer is guaranteed to not have started release yet. This
4373 * implies that if we observe !CSS_RELEASED on @pos in this RCU
4374 * critical section, the one pointed to by its next pointer is
4375 * guaranteed to not have finished its RCU grace period even if we
4376 * have dropped rcu_read_lock() in-between iterations.
4378 * If @pos has CSS_RELEASED set, its next pointer can't be
4379 * dereferenced; however, as each css is given a monotonically
4380 * increasing unique serial number and always appended to the
4381 * sibling list, the next one can be found by walking the parent's
4382 * children until the first css with higher serial number than
4383 * @pos's. While this path can be slower, it happens iff iteration
4384 * races against release and the race window is very small.
4387 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
4388 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
4389 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
4391 list_for_each_entry_rcu(next
, &parent
->children
, sibling
,
4392 lockdep_is_held(&cgroup_mutex
))
4393 if (next
->serial_nr
> pos
->serial_nr
)
4398 * @next, if not pointing to the head, can be dereferenced and is
4401 if (&next
->sibling
!= &parent
->children
)
4407 * css_next_descendant_pre - find the next descendant for pre-order walk
4408 * @pos: the current position (%NULL to initiate traversal)
4409 * @root: css whose descendants to walk
4411 * To be used by css_for_each_descendant_pre(). Find the next descendant
4412 * to visit for pre-order traversal of @root's descendants. @root is
4413 * included in the iteration and the first node to be visited.
4415 * While this function requires cgroup_mutex or RCU read locking, it
4416 * doesn't require the whole traversal to be contained in a single critical
4417 * section. This function will return the correct next descendant as long
4418 * as both @pos and @root are accessible and @pos is a descendant of @root.
4420 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4421 * css which finished ->css_online() is guaranteed to be visible in the
4422 * future iterations and will stay visible until the last reference is put.
4423 * A css which hasn't finished ->css_online() or already finished
4424 * ->css_offline() may show up during traversal. It's each subsystem's
4425 * responsibility to synchronize against on/offlining.
4427 struct cgroup_subsys_state
*
4428 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
4429 struct cgroup_subsys_state
*root
)
4431 struct cgroup_subsys_state
*next
;
4433 cgroup_assert_mutex_or_rcu_locked();
4435 /* if first iteration, visit @root */
4439 /* visit the first child if exists */
4440 next
= css_next_child(NULL
, pos
);
4444 /* no child, visit my or the closest ancestor's next sibling */
4445 while (pos
!= root
) {
4446 next
= css_next_child(pos
, pos
->parent
);
4454 EXPORT_SYMBOL_GPL(css_next_descendant_pre
);
4457 * css_rightmost_descendant - return the rightmost descendant of a css
4458 * @pos: css of interest
4460 * Return the rightmost descendant of @pos. If there's no descendant, @pos
4461 * is returned. This can be used during pre-order traversal to skip
4464 * While this function requires cgroup_mutex or RCU read locking, it
4465 * doesn't require the whole traversal to be contained in a single critical
4466 * section. This function will return the correct rightmost descendant as
4467 * long as @pos is accessible.
4469 struct cgroup_subsys_state
*
4470 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
4472 struct cgroup_subsys_state
*last
, *tmp
;
4474 cgroup_assert_mutex_or_rcu_locked();
4478 /* ->prev isn't RCU safe, walk ->next till the end */
4480 css_for_each_child(tmp
, last
)
4487 static struct cgroup_subsys_state
*
4488 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
4490 struct cgroup_subsys_state
*last
;
4494 pos
= css_next_child(NULL
, pos
);
4501 * css_next_descendant_post - find the next descendant for post-order walk
4502 * @pos: the current position (%NULL to initiate traversal)
4503 * @root: css whose descendants to walk
4505 * To be used by css_for_each_descendant_post(). Find the next descendant
4506 * to visit for post-order traversal of @root's descendants. @root is
4507 * included in the iteration and the last node to be visited.
4509 * While this function requires cgroup_mutex or RCU read locking, it
4510 * doesn't require the whole traversal to be contained in a single critical
4511 * section. This function will return the correct next descendant as long
4512 * as both @pos and @cgroup are accessible and @pos is a descendant of
4515 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4516 * css which finished ->css_online() is guaranteed to be visible in the
4517 * future iterations and will stay visible until the last reference is put.
4518 * A css which hasn't finished ->css_online() or already finished
4519 * ->css_offline() may show up during traversal. It's each subsystem's
4520 * responsibility to synchronize against on/offlining.
4522 struct cgroup_subsys_state
*
4523 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
4524 struct cgroup_subsys_state
*root
)
4526 struct cgroup_subsys_state
*next
;
4528 cgroup_assert_mutex_or_rcu_locked();
4530 /* if first iteration, visit leftmost descendant which may be @root */
4532 return css_leftmost_descendant(root
);
4534 /* if we visited @root, we're done */
4538 /* if there's an unvisited sibling, visit its leftmost descendant */
4539 next
= css_next_child(pos
, pos
->parent
);
4541 return css_leftmost_descendant(next
);
4543 /* no sibling left, visit parent */
4548 * css_has_online_children - does a css have online children
4549 * @css: the target css
4551 * Returns %true if @css has any online children; otherwise, %false. This
4552 * function can be called from any context but the caller is responsible
4553 * for synchronizing against on/offlining as necessary.
4555 bool css_has_online_children(struct cgroup_subsys_state
*css
)
4557 struct cgroup_subsys_state
*child
;
4561 css_for_each_child(child
, css
) {
4562 if (child
->flags
& CSS_ONLINE
) {
4571 static struct css_set
*css_task_iter_next_css_set(struct css_task_iter
*it
)
4573 struct list_head
*l
;
4574 struct cgrp_cset_link
*link
;
4575 struct css_set
*cset
;
4577 lockdep_assert_held(&css_set_lock
);
4579 /* find the next threaded cset */
4580 if (it
->tcset_pos
) {
4581 l
= it
->tcset_pos
->next
;
4583 if (l
!= it
->tcset_head
) {
4585 return container_of(l
, struct css_set
,
4586 threaded_csets_node
);
4589 it
->tcset_pos
= NULL
;
4592 /* find the next cset */
4595 if (l
== it
->cset_head
) {
4596 it
->cset_pos
= NULL
;
4601 cset
= container_of(l
, struct css_set
, e_cset_node
[it
->ss
->id
]);
4603 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
4609 /* initialize threaded css_set walking */
4610 if (it
->flags
& CSS_TASK_ITER_THREADED
) {
4612 put_css_set_locked(it
->cur_dcset
);
4613 it
->cur_dcset
= cset
;
4616 it
->tcset_head
= &cset
->threaded_csets
;
4617 it
->tcset_pos
= &cset
->threaded_csets
;
4624 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4625 * @it: the iterator to advance
4627 * Advance @it to the next css_set to walk.
4629 static void css_task_iter_advance_css_set(struct css_task_iter
*it
)
4631 struct css_set
*cset
;
4633 lockdep_assert_held(&css_set_lock
);
4635 /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4636 while ((cset
= css_task_iter_next_css_set(it
))) {
4637 if (!list_empty(&cset
->tasks
)) {
4638 it
->cur_tasks_head
= &cset
->tasks
;
4640 } else if (!list_empty(&cset
->mg_tasks
)) {
4641 it
->cur_tasks_head
= &cset
->mg_tasks
;
4643 } else if (!list_empty(&cset
->dying_tasks
)) {
4644 it
->cur_tasks_head
= &cset
->dying_tasks
;
4649 it
->task_pos
= NULL
;
4652 it
->task_pos
= it
->cur_tasks_head
->next
;
4655 * We don't keep css_sets locked across iteration steps and thus
4656 * need to take steps to ensure that iteration can be resumed after
4657 * the lock is re-acquired. Iteration is performed at two levels -
4658 * css_sets and tasks in them.
4660 * Once created, a css_set never leaves its cgroup lists, so a
4661 * pinned css_set is guaranteed to stay put and we can resume
4662 * iteration afterwards.
4664 * Tasks may leave @cset across iteration steps. This is resolved
4665 * by registering each iterator with the css_set currently being
4666 * walked and making css_set_move_task() advance iterators whose
4667 * next task is leaving.
4670 list_del(&it
->iters_node
);
4671 put_css_set_locked(it
->cur_cset
);
4674 it
->cur_cset
= cset
;
4675 list_add(&it
->iters_node
, &cset
->task_iters
);
4678 static void css_task_iter_skip(struct css_task_iter
*it
,
4679 struct task_struct
*task
)
4681 lockdep_assert_held(&css_set_lock
);
4683 if (it
->task_pos
== &task
->cg_list
) {
4684 it
->task_pos
= it
->task_pos
->next
;
4685 it
->flags
|= CSS_TASK_ITER_SKIPPED
;
4689 static void css_task_iter_advance(struct css_task_iter
*it
)
4691 struct task_struct
*task
;
4693 lockdep_assert_held(&css_set_lock
);
4697 * Advance iterator to find next entry. We go through cset
4698 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4701 if (it
->flags
& CSS_TASK_ITER_SKIPPED
)
4702 it
->flags
&= ~CSS_TASK_ITER_SKIPPED
;
4704 it
->task_pos
= it
->task_pos
->next
;
4706 if (it
->task_pos
== &it
->cur_cset
->tasks
) {
4707 it
->cur_tasks_head
= &it
->cur_cset
->mg_tasks
;
4708 it
->task_pos
= it
->cur_tasks_head
->next
;
4710 if (it
->task_pos
== &it
->cur_cset
->mg_tasks
) {
4711 it
->cur_tasks_head
= &it
->cur_cset
->dying_tasks
;
4712 it
->task_pos
= it
->cur_tasks_head
->next
;
4714 if (it
->task_pos
== &it
->cur_cset
->dying_tasks
)
4715 css_task_iter_advance_css_set(it
);
4717 /* called from start, proceed to the first cset */
4718 css_task_iter_advance_css_set(it
);
4724 task
= list_entry(it
->task_pos
, struct task_struct
, cg_list
);
4726 if (it
->flags
& CSS_TASK_ITER_PROCS
) {
4727 /* if PROCS, skip over tasks which aren't group leaders */
4728 if (!thread_group_leader(task
))
4731 /* and dying leaders w/o live member threads */
4732 if (it
->cur_tasks_head
== &it
->cur_cset
->dying_tasks
&&
4733 !atomic_read(&task
->signal
->live
))
4736 /* skip all dying ones */
4737 if (it
->cur_tasks_head
== &it
->cur_cset
->dying_tasks
)
4743 * css_task_iter_start - initiate task iteration
4744 * @css: the css to walk tasks of
4745 * @flags: CSS_TASK_ITER_* flags
4746 * @it: the task iterator to use
4748 * Initiate iteration through the tasks of @css. The caller can call
4749 * css_task_iter_next() to walk through the tasks until the function
4750 * returns NULL. On completion of iteration, css_task_iter_end() must be
4753 void css_task_iter_start(struct cgroup_subsys_state
*css
, unsigned int flags
,
4754 struct css_task_iter
*it
)
4756 memset(it
, 0, sizeof(*it
));
4758 spin_lock_irq(&css_set_lock
);
4763 if (CGROUP_HAS_SUBSYS_CONFIG
&& it
->ss
)
4764 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
4766 it
->cset_pos
= &css
->cgroup
->cset_links
;
4768 it
->cset_head
= it
->cset_pos
;
4770 css_task_iter_advance(it
);
4772 spin_unlock_irq(&css_set_lock
);
4776 * css_task_iter_next - return the next task for the iterator
4777 * @it: the task iterator being iterated
4779 * The "next" function for task iteration. @it should have been
4780 * initialized via css_task_iter_start(). Returns NULL when the iteration
4783 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
4786 put_task_struct(it
->cur_task
);
4787 it
->cur_task
= NULL
;
4790 spin_lock_irq(&css_set_lock
);
4792 /* @it may be half-advanced by skips, finish advancing */
4793 if (it
->flags
& CSS_TASK_ITER_SKIPPED
)
4794 css_task_iter_advance(it
);
4797 it
->cur_task
= list_entry(it
->task_pos
, struct task_struct
,
4799 get_task_struct(it
->cur_task
);
4800 css_task_iter_advance(it
);
4803 spin_unlock_irq(&css_set_lock
);
4805 return it
->cur_task
;
4809 * css_task_iter_end - finish task iteration
4810 * @it: the task iterator to finish
4812 * Finish task iteration started by css_task_iter_start().
4814 void css_task_iter_end(struct css_task_iter
*it
)
4817 spin_lock_irq(&css_set_lock
);
4818 list_del(&it
->iters_node
);
4819 put_css_set_locked(it
->cur_cset
);
4820 spin_unlock_irq(&css_set_lock
);
4824 put_css_set(it
->cur_dcset
);
4827 put_task_struct(it
->cur_task
);
4830 static void cgroup_procs_release(struct kernfs_open_file
*of
)
4832 struct cgroup_file_ctx
*ctx
= of
->priv
;
4834 if (ctx
->procs
.started
)
4835 css_task_iter_end(&ctx
->procs
.iter
);
4838 static void *cgroup_procs_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
4840 struct kernfs_open_file
*of
= s
->private;
4841 struct cgroup_file_ctx
*ctx
= of
->priv
;
4846 return css_task_iter_next(&ctx
->procs
.iter
);
4849 static void *__cgroup_procs_start(struct seq_file
*s
, loff_t
*pos
,
4850 unsigned int iter_flags
)
4852 struct kernfs_open_file
*of
= s
->private;
4853 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4854 struct cgroup_file_ctx
*ctx
= of
->priv
;
4855 struct css_task_iter
*it
= &ctx
->procs
.iter
;
4858 * When a seq_file is seeked, it's always traversed sequentially
4859 * from position 0, so we can simply keep iterating on !0 *pos.
4861 if (!ctx
->procs
.started
) {
4862 if (WARN_ON_ONCE((*pos
)))
4863 return ERR_PTR(-EINVAL
);
4864 css_task_iter_start(&cgrp
->self
, iter_flags
, it
);
4865 ctx
->procs
.started
= true;
4866 } else if (!(*pos
)) {
4867 css_task_iter_end(it
);
4868 css_task_iter_start(&cgrp
->self
, iter_flags
, it
);
4870 return it
->cur_task
;
4872 return cgroup_procs_next(s
, NULL
, NULL
);
4875 static void *cgroup_procs_start(struct seq_file
*s
, loff_t
*pos
)
4877 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4880 * All processes of a threaded subtree belong to the domain cgroup
4881 * of the subtree. Only threads can be distributed across the
4882 * subtree. Reject reads on cgroup.procs in the subtree proper.
4883 * They're always empty anyway.
4885 if (cgroup_is_threaded(cgrp
))
4886 return ERR_PTR(-EOPNOTSUPP
);
4888 return __cgroup_procs_start(s
, pos
, CSS_TASK_ITER_PROCS
|
4889 CSS_TASK_ITER_THREADED
);
4892 static int cgroup_procs_show(struct seq_file
*s
, void *v
)
4894 seq_printf(s
, "%d\n", task_pid_vnr(v
));
4898 static int cgroup_may_write(const struct cgroup
*cgrp
, struct super_block
*sb
)
4901 struct inode
*inode
;
4903 lockdep_assert_held(&cgroup_mutex
);
4905 inode
= kernfs_get_inode(sb
, cgrp
->procs_file
.kn
);
4909 ret
= inode_permission(&init_user_ns
, inode
, MAY_WRITE
);
4914 static int cgroup_procs_write_permission(struct cgroup
*src_cgrp
,
4915 struct cgroup
*dst_cgrp
,
4916 struct super_block
*sb
,
4917 struct cgroup_namespace
*ns
)
4919 struct cgroup
*com_cgrp
= src_cgrp
;
4922 lockdep_assert_held(&cgroup_mutex
);
4924 /* find the common ancestor */
4925 while (!cgroup_is_descendant(dst_cgrp
, com_cgrp
))
4926 com_cgrp
= cgroup_parent(com_cgrp
);
4928 /* %current should be authorized to migrate to the common ancestor */
4929 ret
= cgroup_may_write(com_cgrp
, sb
);
4934 * If namespaces are delegation boundaries, %current must be able
4935 * to see both source and destination cgroups from its namespace.
4937 if ((cgrp_dfl_root
.flags
& CGRP_ROOT_NS_DELEGATE
) &&
4938 (!cgroup_is_descendant(src_cgrp
, ns
->root_cset
->dfl_cgrp
) ||
4939 !cgroup_is_descendant(dst_cgrp
, ns
->root_cset
->dfl_cgrp
)))
4945 static int cgroup_attach_permissions(struct cgroup
*src_cgrp
,
4946 struct cgroup
*dst_cgrp
,
4947 struct super_block
*sb
, bool threadgroup
,
4948 struct cgroup_namespace
*ns
)
4952 ret
= cgroup_procs_write_permission(src_cgrp
, dst_cgrp
, sb
, ns
);
4956 ret
= cgroup_migrate_vet_dst(dst_cgrp
);
4960 if (!threadgroup
&& (src_cgrp
->dom_cgrp
!= dst_cgrp
->dom_cgrp
))
4966 static ssize_t
__cgroup_procs_write(struct kernfs_open_file
*of
, char *buf
,
4969 struct cgroup_file_ctx
*ctx
= of
->priv
;
4970 struct cgroup
*src_cgrp
, *dst_cgrp
;
4971 struct task_struct
*task
;
4972 const struct cred
*saved_cred
;
4976 dst_cgrp
= cgroup_kn_lock_live(of
->kn
, false);
4980 task
= cgroup_procs_write_start(buf
, threadgroup
, &locked
);
4981 ret
= PTR_ERR_OR_ZERO(task
);
4985 /* find the source cgroup */
4986 spin_lock_irq(&css_set_lock
);
4987 src_cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
4988 spin_unlock_irq(&css_set_lock
);
4991 * Process and thread migrations follow same delegation rule. Check
4992 * permissions using the credentials from file open to protect against
4993 * inherited fd attacks.
4995 saved_cred
= override_creds(of
->file
->f_cred
);
4996 ret
= cgroup_attach_permissions(src_cgrp
, dst_cgrp
,
4997 of
->file
->f_path
.dentry
->d_sb
,
4998 threadgroup
, ctx
->ns
);
4999 revert_creds(saved_cred
);
5003 ret
= cgroup_attach_task(dst_cgrp
, task
, threadgroup
);
5006 cgroup_procs_write_finish(task
, locked
);
5008 cgroup_kn_unlock(of
->kn
);
5013 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
5014 char *buf
, size_t nbytes
, loff_t off
)
5016 return __cgroup_procs_write(of
, buf
, true) ?: nbytes
;
5019 static void *cgroup_threads_start(struct seq_file
*s
, loff_t
*pos
)
5021 return __cgroup_procs_start(s
, pos
, 0);
5024 static ssize_t
cgroup_threads_write(struct kernfs_open_file
*of
,
5025 char *buf
, size_t nbytes
, loff_t off
)
5027 return __cgroup_procs_write(of
, buf
, false) ?: nbytes
;
5030 /* cgroup core interface files for the default hierarchy */
5031 static struct cftype cgroup_base_files
[] = {
5033 .name
= "cgroup.type",
5034 .flags
= CFTYPE_NOT_ON_ROOT
,
5035 .seq_show
= cgroup_type_show
,
5036 .write
= cgroup_type_write
,
5039 .name
= "cgroup.procs",
5040 .flags
= CFTYPE_NS_DELEGATABLE
,
5041 .file_offset
= offsetof(struct cgroup
, procs_file
),
5042 .release
= cgroup_procs_release
,
5043 .seq_start
= cgroup_procs_start
,
5044 .seq_next
= cgroup_procs_next
,
5045 .seq_show
= cgroup_procs_show
,
5046 .write
= cgroup_procs_write
,
5049 .name
= "cgroup.threads",
5050 .flags
= CFTYPE_NS_DELEGATABLE
,
5051 .release
= cgroup_procs_release
,
5052 .seq_start
= cgroup_threads_start
,
5053 .seq_next
= cgroup_procs_next
,
5054 .seq_show
= cgroup_procs_show
,
5055 .write
= cgroup_threads_write
,
5058 .name
= "cgroup.controllers",
5059 .seq_show
= cgroup_controllers_show
,
5062 .name
= "cgroup.subtree_control",
5063 .flags
= CFTYPE_NS_DELEGATABLE
,
5064 .seq_show
= cgroup_subtree_control_show
,
5065 .write
= cgroup_subtree_control_write
,
5068 .name
= "cgroup.events",
5069 .flags
= CFTYPE_NOT_ON_ROOT
,
5070 .file_offset
= offsetof(struct cgroup
, events_file
),
5071 .seq_show
= cgroup_events_show
,
5074 .name
= "cgroup.max.descendants",
5075 .seq_show
= cgroup_max_descendants_show
,
5076 .write
= cgroup_max_descendants_write
,
5079 .name
= "cgroup.max.depth",
5080 .seq_show
= cgroup_max_depth_show
,
5081 .write
= cgroup_max_depth_write
,
5084 .name
= "cgroup.stat",
5085 .seq_show
= cgroup_stat_show
,
5088 .name
= "cgroup.freeze",
5089 .flags
= CFTYPE_NOT_ON_ROOT
,
5090 .seq_show
= cgroup_freeze_show
,
5091 .write
= cgroup_freeze_write
,
5094 .name
= "cgroup.kill",
5095 .flags
= CFTYPE_NOT_ON_ROOT
,
5096 .write
= cgroup_kill_write
,
5100 .seq_show
= cpu_stat_show
,
5104 .name
= "io.pressure",
5105 .flags
= CFTYPE_PRESSURE
,
5106 .seq_show
= cgroup_io_pressure_show
,
5107 .write
= cgroup_io_pressure_write
,
5108 .poll
= cgroup_pressure_poll
,
5109 .release
= cgroup_pressure_release
,
5112 .name
= "memory.pressure",
5113 .flags
= CFTYPE_PRESSURE
,
5114 .seq_show
= cgroup_memory_pressure_show
,
5115 .write
= cgroup_memory_pressure_write
,
5116 .poll
= cgroup_pressure_poll
,
5117 .release
= cgroup_pressure_release
,
5120 .name
= "cpu.pressure",
5121 .flags
= CFTYPE_PRESSURE
,
5122 .seq_show
= cgroup_cpu_pressure_show
,
5123 .write
= cgroup_cpu_pressure_write
,
5124 .poll
= cgroup_pressure_poll
,
5125 .release
= cgroup_pressure_release
,
5127 #endif /* CONFIG_PSI */
5132 * css destruction is four-stage process.
5134 * 1. Destruction starts. Killing of the percpu_ref is initiated.
5135 * Implemented in kill_css().
5137 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
5138 * and thus css_tryget_online() is guaranteed to fail, the css can be
5139 * offlined by invoking offline_css(). After offlining, the base ref is
5140 * put. Implemented in css_killed_work_fn().
5142 * 3. When the percpu_ref reaches zero, the only possible remaining
5143 * accessors are inside RCU read sections. css_release() schedules the
5146 * 4. After the grace period, the css can be freed. Implemented in
5147 * css_free_work_fn().
5149 * It is actually hairier because both step 2 and 4 require process context
5150 * and thus involve punting to css->destroy_work adding two additional
5151 * steps to the already complex sequence.
5153 static void css_free_rwork_fn(struct work_struct
*work
)
5155 struct cgroup_subsys_state
*css
= container_of(to_rcu_work(work
),
5156 struct cgroup_subsys_state
, destroy_rwork
);
5157 struct cgroup_subsys
*ss
= css
->ss
;
5158 struct cgroup
*cgrp
= css
->cgroup
;
5160 percpu_ref_exit(&css
->refcnt
);
5164 struct cgroup_subsys_state
*parent
= css
->parent
;
5168 cgroup_idr_remove(&ss
->css_idr
, id
);
5174 /* cgroup free path */
5175 atomic_dec(&cgrp
->root
->nr_cgrps
);
5176 cgroup1_pidlist_destroy_all(cgrp
);
5177 cancel_work_sync(&cgrp
->release_agent_work
);
5179 if (cgroup_parent(cgrp
)) {
5181 * We get a ref to the parent, and put the ref when
5182 * this cgroup is being freed, so it's guaranteed
5183 * that the parent won't be destroyed before its
5186 cgroup_put(cgroup_parent(cgrp
));
5187 kernfs_put(cgrp
->kn
);
5188 psi_cgroup_free(cgrp
);
5189 cgroup_rstat_exit(cgrp
);
5193 * This is root cgroup's refcnt reaching zero,
5194 * which indicates that the root should be
5197 cgroup_destroy_root(cgrp
->root
);
5202 static void css_release_work_fn(struct work_struct
*work
)
5204 struct cgroup_subsys_state
*css
=
5205 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
5206 struct cgroup_subsys
*ss
= css
->ss
;
5207 struct cgroup
*cgrp
= css
->cgroup
;
5209 mutex_lock(&cgroup_mutex
);
5211 css
->flags
|= CSS_RELEASED
;
5212 list_del_rcu(&css
->sibling
);
5215 /* css release path */
5216 if (!list_empty(&css
->rstat_css_node
)) {
5217 cgroup_rstat_flush(cgrp
);
5218 list_del_rcu(&css
->rstat_css_node
);
5221 cgroup_idr_replace(&ss
->css_idr
, NULL
, css
->id
);
5222 if (ss
->css_released
)
5223 ss
->css_released(css
);
5225 struct cgroup
*tcgrp
;
5227 /* cgroup release path */
5228 TRACE_CGROUP_PATH(release
, cgrp
);
5230 cgroup_rstat_flush(cgrp
);
5232 spin_lock_irq(&css_set_lock
);
5233 for (tcgrp
= cgroup_parent(cgrp
); tcgrp
;
5234 tcgrp
= cgroup_parent(tcgrp
))
5235 tcgrp
->nr_dying_descendants
--;
5236 spin_unlock_irq(&css_set_lock
);
5239 * There are two control paths which try to determine
5240 * cgroup from dentry without going through kernfs -
5241 * cgroupstats_build() and css_tryget_online_from_dir().
5242 * Those are supported by RCU protecting clearing of
5243 * cgrp->kn->priv backpointer.
5246 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
,
5250 mutex_unlock(&cgroup_mutex
);
5252 INIT_RCU_WORK(&css
->destroy_rwork
, css_free_rwork_fn
);
5253 queue_rcu_work(cgroup_destroy_wq
, &css
->destroy_rwork
);
5256 static void css_release(struct percpu_ref
*ref
)
5258 struct cgroup_subsys_state
*css
=
5259 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
5261 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
5262 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
5265 static void init_and_link_css(struct cgroup_subsys_state
*css
,
5266 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
5268 lockdep_assert_held(&cgroup_mutex
);
5270 cgroup_get_live(cgrp
);
5272 memset(css
, 0, sizeof(*css
));
5276 INIT_LIST_HEAD(&css
->sibling
);
5277 INIT_LIST_HEAD(&css
->children
);
5278 INIT_LIST_HEAD(&css
->rstat_css_node
);
5279 css
->serial_nr
= css_serial_nr_next
++;
5280 atomic_set(&css
->online_cnt
, 0);
5282 if (cgroup_parent(cgrp
)) {
5283 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
5284 css_get(css
->parent
);
5287 if (ss
->css_rstat_flush
)
5288 list_add_rcu(&css
->rstat_css_node
, &cgrp
->rstat_css_list
);
5290 BUG_ON(cgroup_css(cgrp
, ss
));
5293 /* invoke ->css_online() on a new CSS and mark it online if successful */
5294 static int online_css(struct cgroup_subsys_state
*css
)
5296 struct cgroup_subsys
*ss
= css
->ss
;
5299 lockdep_assert_held(&cgroup_mutex
);
5302 ret
= ss
->css_online(css
);
5304 css
->flags
|= CSS_ONLINE
;
5305 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
5307 atomic_inc(&css
->online_cnt
);
5309 atomic_inc(&css
->parent
->online_cnt
);
5314 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
5315 static void offline_css(struct cgroup_subsys_state
*css
)
5317 struct cgroup_subsys
*ss
= css
->ss
;
5319 lockdep_assert_held(&cgroup_mutex
);
5321 if (!(css
->flags
& CSS_ONLINE
))
5324 if (ss
->css_offline
)
5325 ss
->css_offline(css
);
5327 css
->flags
&= ~CSS_ONLINE
;
5328 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
5330 wake_up_all(&css
->cgroup
->offline_waitq
);
5334 * css_create - create a cgroup_subsys_state
5335 * @cgrp: the cgroup new css will be associated with
5336 * @ss: the subsys of new css
5338 * Create a new css associated with @cgrp - @ss pair. On success, the new
5339 * css is online and installed in @cgrp. This function doesn't create the
5340 * interface files. Returns 0 on success, -errno on failure.
5342 static struct cgroup_subsys_state
*css_create(struct cgroup
*cgrp
,
5343 struct cgroup_subsys
*ss
)
5345 struct cgroup
*parent
= cgroup_parent(cgrp
);
5346 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
5347 struct cgroup_subsys_state
*css
;
5350 lockdep_assert_held(&cgroup_mutex
);
5352 css
= ss
->css_alloc(parent_css
);
5354 css
= ERR_PTR(-ENOMEM
);
5358 init_and_link_css(css
, ss
, cgrp
);
5360 err
= percpu_ref_init(&css
->refcnt
, css_release
, 0, GFP_KERNEL
);
5364 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_KERNEL
);
5369 /* @css is ready to be brought online now, make it visible */
5370 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
5371 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
5373 err
= online_css(css
);
5380 list_del_rcu(&css
->sibling
);
5382 list_del_rcu(&css
->rstat_css_node
);
5383 INIT_RCU_WORK(&css
->destroy_rwork
, css_free_rwork_fn
);
5384 queue_rcu_work(cgroup_destroy_wq
, &css
->destroy_rwork
);
5385 return ERR_PTR(err
);
5389 * The returned cgroup is fully initialized including its control mask, but
5390 * it isn't associated with its kernfs_node and doesn't have the control
5393 static struct cgroup
*cgroup_create(struct cgroup
*parent
, const char *name
,
5396 struct cgroup_root
*root
= parent
->root
;
5397 struct cgroup
*cgrp
, *tcgrp
;
5398 struct kernfs_node
*kn
;
5399 int level
= parent
->level
+ 1;
5402 /* allocate the cgroup and its ID, 0 is reserved for the root */
5403 cgrp
= kzalloc(struct_size(cgrp
, ancestor_ids
, (level
+ 1)),
5406 return ERR_PTR(-ENOMEM
);
5408 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
, 0, GFP_KERNEL
);
5412 ret
= cgroup_rstat_init(cgrp
);
5414 goto out_cancel_ref
;
5416 /* create the directory */
5417 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
5424 init_cgroup_housekeeping(cgrp
);
5426 cgrp
->self
.parent
= &parent
->self
;
5428 cgrp
->level
= level
;
5430 ret
= psi_cgroup_alloc(cgrp
);
5432 goto out_kernfs_remove
;
5434 ret
= cgroup_bpf_inherit(cgrp
);
5439 * New cgroup inherits effective freeze counter, and
5440 * if the parent has to be frozen, the child has too.
5442 cgrp
->freezer
.e_freeze
= parent
->freezer
.e_freeze
;
5443 if (cgrp
->freezer
.e_freeze
) {
5445 * Set the CGRP_FREEZE flag, so when a process will be
5446 * attached to the child cgroup, it will become frozen.
5447 * At this point the new cgroup is unpopulated, so we can
5448 * consider it frozen immediately.
5450 set_bit(CGRP_FREEZE
, &cgrp
->flags
);
5451 set_bit(CGRP_FROZEN
, &cgrp
->flags
);
5454 spin_lock_irq(&css_set_lock
);
5455 for (tcgrp
= cgrp
; tcgrp
; tcgrp
= cgroup_parent(tcgrp
)) {
5456 cgrp
->ancestor_ids
[tcgrp
->level
] = cgroup_id(tcgrp
);
5458 if (tcgrp
!= cgrp
) {
5459 tcgrp
->nr_descendants
++;
5462 * If the new cgroup is frozen, all ancestor cgroups
5463 * get a new frozen descendant, but their state can't
5464 * change because of this.
5466 if (cgrp
->freezer
.e_freeze
)
5467 tcgrp
->freezer
.nr_frozen_descendants
++;
5470 spin_unlock_irq(&css_set_lock
);
5472 if (notify_on_release(parent
))
5473 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
5475 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
5476 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
5478 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
5480 /* allocation complete, commit to creation */
5481 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
5482 atomic_inc(&root
->nr_cgrps
);
5483 cgroup_get_live(parent
);
5486 * On the default hierarchy, a child doesn't automatically inherit
5487 * subtree_control from the parent. Each is configured manually.
5489 if (!cgroup_on_dfl(cgrp
))
5490 cgrp
->subtree_control
= cgroup_control(cgrp
);
5492 cgroup_propagate_control(cgrp
);
5497 psi_cgroup_free(cgrp
);
5499 kernfs_remove(cgrp
->kn
);
5501 cgroup_rstat_exit(cgrp
);
5503 percpu_ref_exit(&cgrp
->self
.refcnt
);
5506 return ERR_PTR(ret
);
5509 static bool cgroup_check_hierarchy_limits(struct cgroup
*parent
)
5511 struct cgroup
*cgroup
;
5515 lockdep_assert_held(&cgroup_mutex
);
5517 for (cgroup
= parent
; cgroup
; cgroup
= cgroup_parent(cgroup
)) {
5518 if (cgroup
->nr_descendants
>= cgroup
->max_descendants
)
5521 if (level
> cgroup
->max_depth
)
5532 int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
, umode_t mode
)
5534 struct cgroup
*parent
, *cgrp
;
5537 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
5538 if (strchr(name
, '\n'))
5541 parent
= cgroup_kn_lock_live(parent_kn
, false);
5545 if (!cgroup_check_hierarchy_limits(parent
)) {
5550 cgrp
= cgroup_create(parent
, name
, mode
);
5552 ret
= PTR_ERR(cgrp
);
5557 * This extra ref will be put in cgroup_free_fn() and guarantees
5558 * that @cgrp->kn is always accessible.
5560 kernfs_get(cgrp
->kn
);
5562 ret
= cgroup_kn_set_ugid(cgrp
->kn
);
5566 ret
= css_populate_dir(&cgrp
->self
);
5570 ret
= cgroup_apply_control_enable(cgrp
);
5574 TRACE_CGROUP_PATH(mkdir
, cgrp
);
5576 /* let's create and online css's */
5577 kernfs_activate(cgrp
->kn
);
5583 cgroup_destroy_locked(cgrp
);
5585 cgroup_kn_unlock(parent_kn
);
5590 * This is called when the refcnt of a css is confirmed to be killed.
5591 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5592 * initiate destruction and put the css ref from kill_css().
5594 static void css_killed_work_fn(struct work_struct
*work
)
5596 struct cgroup_subsys_state
*css
=
5597 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
5599 mutex_lock(&cgroup_mutex
);
5604 /* @css can't go away while we're holding cgroup_mutex */
5606 } while (css
&& atomic_dec_and_test(&css
->online_cnt
));
5608 mutex_unlock(&cgroup_mutex
);
5611 /* css kill confirmation processing requires process context, bounce */
5612 static void css_killed_ref_fn(struct percpu_ref
*ref
)
5614 struct cgroup_subsys_state
*css
=
5615 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
5617 if (atomic_dec_and_test(&css
->online_cnt
)) {
5618 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
5619 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
5624 * kill_css - destroy a css
5625 * @css: css to destroy
5627 * This function initiates destruction of @css by removing cgroup interface
5628 * files and putting its base reference. ->css_offline() will be invoked
5629 * asynchronously once css_tryget_online() is guaranteed to fail and when
5630 * the reference count reaches zero, @css will be released.
5632 static void kill_css(struct cgroup_subsys_state
*css
)
5634 lockdep_assert_held(&cgroup_mutex
);
5636 if (css
->flags
& CSS_DYING
)
5639 css
->flags
|= CSS_DYING
;
5642 * This must happen before css is disassociated with its cgroup.
5643 * See seq_css() for details.
5648 * Killing would put the base ref, but we need to keep it alive
5649 * until after ->css_offline().
5654 * cgroup core guarantees that, by the time ->css_offline() is
5655 * invoked, no new css reference will be given out via
5656 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5657 * proceed to offlining css's because percpu_ref_kill() doesn't
5658 * guarantee that the ref is seen as killed on all CPUs on return.
5660 * Use percpu_ref_kill_and_confirm() to get notifications as each
5661 * css is confirmed to be seen as killed on all CPUs.
5663 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
5667 * cgroup_destroy_locked - the first stage of cgroup destruction
5668 * @cgrp: cgroup to be destroyed
5670 * css's make use of percpu refcnts whose killing latency shouldn't be
5671 * exposed to userland and are RCU protected. Also, cgroup core needs to
5672 * guarantee that css_tryget_online() won't succeed by the time
5673 * ->css_offline() is invoked. To satisfy all the requirements,
5674 * destruction is implemented in the following two steps.
5676 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5677 * userland visible parts and start killing the percpu refcnts of
5678 * css's. Set up so that the next stage will be kicked off once all
5679 * the percpu refcnts are confirmed to be killed.
5681 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5682 * rest of destruction. Once all cgroup references are gone, the
5683 * cgroup is RCU-freed.
5685 * This function implements s1. After this step, @cgrp is gone as far as
5686 * the userland is concerned and a new cgroup with the same name may be
5687 * created. As cgroup doesn't care about the names internally, this
5688 * doesn't cause any problem.
5690 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
5691 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
5693 struct cgroup
*tcgrp
, *parent
= cgroup_parent(cgrp
);
5694 struct cgroup_subsys_state
*css
;
5695 struct cgrp_cset_link
*link
;
5698 lockdep_assert_held(&cgroup_mutex
);
5701 * Only migration can raise populated from zero and we're already
5702 * holding cgroup_mutex.
5704 if (cgroup_is_populated(cgrp
))
5708 * Make sure there's no live children. We can't test emptiness of
5709 * ->self.children as dead children linger on it while being
5710 * drained; otherwise, "rmdir parent/child parent" may fail.
5712 if (css_has_online_children(&cgrp
->self
))
5716 * Mark @cgrp and the associated csets dead. The former prevents
5717 * further task migration and child creation by disabling
5718 * cgroup_lock_live_group(). The latter makes the csets ignored by
5719 * the migration path.
5721 cgrp
->self
.flags
&= ~CSS_ONLINE
;
5723 spin_lock_irq(&css_set_lock
);
5724 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
5725 link
->cset
->dead
= true;
5726 spin_unlock_irq(&css_set_lock
);
5728 /* initiate massacre of all css's */
5729 for_each_css(css
, ssid
, cgrp
)
5732 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */
5733 css_clear_dir(&cgrp
->self
);
5734 kernfs_remove(cgrp
->kn
);
5736 if (cgroup_is_threaded(cgrp
))
5737 parent
->nr_threaded_children
--;
5739 spin_lock_irq(&css_set_lock
);
5740 for (tcgrp
= cgroup_parent(cgrp
); tcgrp
; tcgrp
= cgroup_parent(tcgrp
)) {
5741 tcgrp
->nr_descendants
--;
5742 tcgrp
->nr_dying_descendants
++;
5744 * If the dying cgroup is frozen, decrease frozen descendants
5745 * counters of ancestor cgroups.
5747 if (test_bit(CGRP_FROZEN
, &cgrp
->flags
))
5748 tcgrp
->freezer
.nr_frozen_descendants
--;
5750 spin_unlock_irq(&css_set_lock
);
5752 cgroup1_check_for_release(parent
);
5754 cgroup_bpf_offline(cgrp
);
5756 /* put the base reference */
5757 percpu_ref_kill(&cgrp
->self
.refcnt
);
5762 int cgroup_rmdir(struct kernfs_node
*kn
)
5764 struct cgroup
*cgrp
;
5767 cgrp
= cgroup_kn_lock_live(kn
, false);
5771 ret
= cgroup_destroy_locked(cgrp
);
5773 TRACE_CGROUP_PATH(rmdir
, cgrp
);
5775 cgroup_kn_unlock(kn
);
5779 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
5780 .show_options
= cgroup_show_options
,
5781 .mkdir
= cgroup_mkdir
,
5782 .rmdir
= cgroup_rmdir
,
5783 .show_path
= cgroup_show_path
,
5786 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
5788 struct cgroup_subsys_state
*css
;
5790 pr_debug("Initializing cgroup subsys %s\n", ss
->name
);
5792 mutex_lock(&cgroup_mutex
);
5794 idr_init(&ss
->css_idr
);
5795 INIT_LIST_HEAD(&ss
->cfts
);
5797 /* Create the root cgroup state for this subsystem */
5798 ss
->root
= &cgrp_dfl_root
;
5799 css
= ss
->css_alloc(NULL
);
5800 /* We don't handle early failures gracefully */
5801 BUG_ON(IS_ERR(css
));
5802 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
5805 * Root csses are never destroyed and we can't initialize
5806 * percpu_ref during early init. Disable refcnting.
5808 css
->flags
|= CSS_NO_REF
;
5811 /* allocation can't be done safely during early init */
5814 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
5815 BUG_ON(css
->id
< 0);
5818 /* Update the init_css_set to contain a subsys
5819 * pointer to this state - since the subsystem is
5820 * newly registered, all tasks and hence the
5821 * init_css_set is in the subsystem's root cgroup. */
5822 init_css_set
.subsys
[ss
->id
] = css
;
5824 have_fork_callback
|= (bool)ss
->fork
<< ss
->id
;
5825 have_exit_callback
|= (bool)ss
->exit
<< ss
->id
;
5826 have_release_callback
|= (bool)ss
->release
<< ss
->id
;
5827 have_canfork_callback
|= (bool)ss
->can_fork
<< ss
->id
;
5829 /* At system boot, before all subsystems have been
5830 * registered, no tasks have been forked, so we don't
5831 * need to invoke fork callbacks here. */
5832 BUG_ON(!list_empty(&init_task
.tasks
));
5834 BUG_ON(online_css(css
));
5836 mutex_unlock(&cgroup_mutex
);
5840 * cgroup_init_early - cgroup initialization at system boot
5842 * Initialize cgroups at system boot, and initialize any
5843 * subsystems that request early init.
5845 int __init
cgroup_init_early(void)
5847 static struct cgroup_fs_context __initdata ctx
;
5848 struct cgroup_subsys
*ss
;
5851 ctx
.root
= &cgrp_dfl_root
;
5852 init_cgroup_root(&ctx
);
5853 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
5855 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
5857 for_each_subsys(ss
, i
) {
5858 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
5859 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5860 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
5862 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
5863 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
5866 ss
->name
= cgroup_subsys_name
[i
];
5867 if (!ss
->legacy_name
)
5868 ss
->legacy_name
= cgroup_subsys_name
[i
];
5871 cgroup_init_subsys(ss
, true);
5877 * cgroup_init - cgroup initialization
5879 * Register cgroup filesystem and /proc file, and initialize
5880 * any subsystems that didn't request early init.
5882 int __init
cgroup_init(void)
5884 struct cgroup_subsys
*ss
;
5887 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT
> 16);
5888 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_base_files
));
5889 BUG_ON(cgroup_init_cftypes(NULL
, cgroup1_base_files
));
5891 cgroup_rstat_boot();
5893 get_user_ns(init_cgroup_ns
.user_ns
);
5895 mutex_lock(&cgroup_mutex
);
5898 * Add init_css_set to the hash table so that dfl_root can link to
5901 hash_add(css_set_table
, &init_css_set
.hlist
,
5902 css_set_hash(init_css_set
.subsys
));
5904 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
5906 mutex_unlock(&cgroup_mutex
);
5908 for_each_subsys(ss
, ssid
) {
5909 if (ss
->early_init
) {
5910 struct cgroup_subsys_state
*css
=
5911 init_css_set
.subsys
[ss
->id
];
5913 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
5915 BUG_ON(css
->id
< 0);
5917 cgroup_init_subsys(ss
, false);
5920 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
5921 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
5924 * Setting dfl_root subsys_mask needs to consider the
5925 * disabled flag and cftype registration needs kmalloc,
5926 * both of which aren't available during early_init.
5928 if (!cgroup_ssid_enabled(ssid
))
5931 if (cgroup1_ssid_disabled(ssid
))
5932 printk(KERN_INFO
"Disabling %s control group subsystem in v1 mounts\n",
5935 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
5937 /* implicit controllers must be threaded too */
5938 WARN_ON(ss
->implicit_on_dfl
&& !ss
->threaded
);
5940 if (ss
->implicit_on_dfl
)
5941 cgrp_dfl_implicit_ss_mask
|= 1 << ss
->id
;
5942 else if (!ss
->dfl_cftypes
)
5943 cgrp_dfl_inhibit_ss_mask
|= 1 << ss
->id
;
5946 cgrp_dfl_threaded_ss_mask
|= 1 << ss
->id
;
5948 if (ss
->dfl_cftypes
== ss
->legacy_cftypes
) {
5949 WARN_ON(cgroup_add_cftypes(ss
, ss
->dfl_cftypes
));
5951 WARN_ON(cgroup_add_dfl_cftypes(ss
, ss
->dfl_cftypes
));
5952 WARN_ON(cgroup_add_legacy_cftypes(ss
, ss
->legacy_cftypes
));
5956 ss
->bind(init_css_set
.subsys
[ssid
]);
5958 mutex_lock(&cgroup_mutex
);
5959 css_populate_dir(init_css_set
.subsys
[ssid
]);
5960 mutex_unlock(&cgroup_mutex
);
5963 /* init_css_set.subsys[] has been updated, re-hash */
5964 hash_del(&init_css_set
.hlist
);
5965 hash_add(css_set_table
, &init_css_set
.hlist
,
5966 css_set_hash(init_css_set
.subsys
));
5968 WARN_ON(sysfs_create_mount_point(fs_kobj
, "cgroup"));
5969 WARN_ON(register_filesystem(&cgroup_fs_type
));
5970 WARN_ON(register_filesystem(&cgroup2_fs_type
));
5971 WARN_ON(!proc_create_single("cgroups", 0, NULL
, proc_cgroupstats_show
));
5972 #ifdef CONFIG_CPUSETS
5973 WARN_ON(register_filesystem(&cpuset_fs_type
));
5979 static int __init
cgroup_wq_init(void)
5982 * There isn't much point in executing destruction path in
5983 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5984 * Use 1 for @max_active.
5986 * We would prefer to do this in cgroup_init() above, but that
5987 * is called before init_workqueues(): so leave this until after.
5989 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
5990 BUG_ON(!cgroup_destroy_wq
);
5993 core_initcall(cgroup_wq_init
);
5995 void cgroup_path_from_kernfs_id(u64 id
, char *buf
, size_t buflen
)
5997 struct kernfs_node
*kn
;
5999 kn
= kernfs_find_and_get_node_by_id(cgrp_dfl_root
.kf_root
, id
);
6002 kernfs_path(kn
, buf
, buflen
);
6007 * cgroup_get_from_id : get the cgroup associated with cgroup id
6009 * On success return the cgrp, on failure return NULL
6011 struct cgroup
*cgroup_get_from_id(u64 id
)
6013 struct kernfs_node
*kn
;
6014 struct cgroup
*cgrp
= NULL
;
6016 kn
= kernfs_find_and_get_node_by_id(cgrp_dfl_root
.kf_root
, id
);
6022 cgrp
= rcu_dereference(*(void __rcu __force
**)&kn
->priv
);
6023 if (cgrp
&& !cgroup_tryget(cgrp
))
6032 EXPORT_SYMBOL_GPL(cgroup_get_from_id
);
6035 * proc_cgroup_show()
6036 * - Print task's cgroup paths into seq_file, one line for each hierarchy
6037 * - Used for /proc/<pid>/cgroup.
6039 int proc_cgroup_show(struct seq_file
*m
, struct pid_namespace
*ns
,
6040 struct pid
*pid
, struct task_struct
*tsk
)
6044 struct cgroup_root
*root
;
6047 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
6051 mutex_lock(&cgroup_mutex
);
6052 spin_lock_irq(&css_set_lock
);
6054 for_each_root(root
) {
6055 struct cgroup_subsys
*ss
;
6056 struct cgroup
*cgrp
;
6057 int ssid
, count
= 0;
6059 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_visible
)
6062 seq_printf(m
, "%d:", root
->hierarchy_id
);
6063 if (root
!= &cgrp_dfl_root
)
6064 for_each_subsys(ss
, ssid
)
6065 if (root
->subsys_mask
& (1 << ssid
))
6066 seq_printf(m
, "%s%s", count
++ ? "," : "",
6068 if (strlen(root
->name
))
6069 seq_printf(m
, "%sname=%s", count
? "," : "",
6073 cgrp
= task_cgroup_from_root(tsk
, root
);
6076 * On traditional hierarchies, all zombie tasks show up as
6077 * belonging to the root cgroup. On the default hierarchy,
6078 * while a zombie doesn't show up in "cgroup.procs" and
6079 * thus can't be migrated, its /proc/PID/cgroup keeps
6080 * reporting the cgroup it belonged to before exiting. If
6081 * the cgroup is removed before the zombie is reaped,
6082 * " (deleted)" is appended to the cgroup path.
6084 if (cgroup_on_dfl(cgrp
) || !(tsk
->flags
& PF_EXITING
)) {
6085 retval
= cgroup_path_ns_locked(cgrp
, buf
, PATH_MAX
,
6086 current
->nsproxy
->cgroup_ns
);
6087 if (retval
>= PATH_MAX
)
6088 retval
= -ENAMETOOLONG
;
6097 if (cgroup_on_dfl(cgrp
) && cgroup_is_dead(cgrp
))
6098 seq_puts(m
, " (deleted)\n");
6105 spin_unlock_irq(&css_set_lock
);
6106 mutex_unlock(&cgroup_mutex
);
6113 * cgroup_fork - initialize cgroup related fields during copy_process()
6114 * @child: pointer to task_struct of forking parent process.
6116 * A task is associated with the init_css_set until cgroup_post_fork()
6117 * attaches it to the target css_set.
6119 void cgroup_fork(struct task_struct
*child
)
6121 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
6122 INIT_LIST_HEAD(&child
->cg_list
);
6125 static struct cgroup
*cgroup_get_from_file(struct file
*f
)
6127 struct cgroup_subsys_state
*css
;
6128 struct cgroup
*cgrp
;
6130 css
= css_tryget_online_from_dir(f
->f_path
.dentry
, NULL
);
6132 return ERR_CAST(css
);
6135 if (!cgroup_on_dfl(cgrp
)) {
6137 return ERR_PTR(-EBADF
);
6144 * cgroup_css_set_fork - find or create a css_set for a child process
6145 * @kargs: the arguments passed to create the child process
6147 * This functions finds or creates a new css_set which the child
6148 * process will be attached to in cgroup_post_fork(). By default,
6149 * the child process will be given the same css_set as its parent.
6151 * If CLONE_INTO_CGROUP is specified this function will try to find an
6152 * existing css_set which includes the requested cgroup and if not create
6153 * a new css_set that the child will be attached to later. If this function
6154 * succeeds it will hold cgroup_threadgroup_rwsem on return. If
6155 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex
6156 * before grabbing cgroup_threadgroup_rwsem and will hold a reference
6157 * to the target cgroup.
6159 static int cgroup_css_set_fork(struct kernel_clone_args
*kargs
)
6160 __acquires(&cgroup_mutex
) __acquires(&cgroup_threadgroup_rwsem
)
6163 struct cgroup
*dst_cgrp
= NULL
;
6164 struct css_set
*cset
;
6165 struct super_block
*sb
;
6168 if (kargs
->flags
& CLONE_INTO_CGROUP
)
6169 mutex_lock(&cgroup_mutex
);
6171 cgroup_threadgroup_change_begin(current
);
6173 spin_lock_irq(&css_set_lock
);
6174 cset
= task_css_set(current
);
6176 spin_unlock_irq(&css_set_lock
);
6178 if (!(kargs
->flags
& CLONE_INTO_CGROUP
)) {
6183 f
= fget_raw(kargs
->cgroup
);
6188 sb
= f
->f_path
.dentry
->d_sb
;
6190 dst_cgrp
= cgroup_get_from_file(f
);
6191 if (IS_ERR(dst_cgrp
)) {
6192 ret
= PTR_ERR(dst_cgrp
);
6197 if (cgroup_is_dead(dst_cgrp
)) {
6203 * Verify that we the target cgroup is writable for us. This is
6204 * usually done by the vfs layer but since we're not going through
6205 * the vfs layer here we need to do it "manually".
6207 ret
= cgroup_may_write(dst_cgrp
, sb
);
6212 * Spawning a task directly into a cgroup works by passing a file
6213 * descriptor to the target cgroup directory. This can even be an O_PATH
6214 * file descriptor. But it can never be a cgroup.procs file descriptor.
6215 * This was done on purpose so spawning into a cgroup could be
6216 * conceptualized as an atomic
6218 * fd = openat(dfd_cgroup, "cgroup.procs", ...);
6219 * write(fd, <child-pid>, ...);
6221 * sequence, i.e. it's a shorthand for the caller opening and writing
6222 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us
6223 * to always use the caller's credentials.
6225 ret
= cgroup_attach_permissions(cset
->dfl_cgrp
, dst_cgrp
, sb
,
6226 !(kargs
->flags
& CLONE_THREAD
),
6227 current
->nsproxy
->cgroup_ns
);
6231 kargs
->cset
= find_css_set(cset
, dst_cgrp
);
6239 kargs
->cgrp
= dst_cgrp
;
6243 cgroup_threadgroup_change_end(current
);
6244 mutex_unlock(&cgroup_mutex
);
6248 cgroup_put(dst_cgrp
);
6251 put_css_set(kargs
->cset
);
6256 * cgroup_css_set_put_fork - drop references we took during fork
6257 * @kargs: the arguments passed to create the child process
6259 * Drop references to the prepared css_set and target cgroup if
6260 * CLONE_INTO_CGROUP was requested.
6262 static void cgroup_css_set_put_fork(struct kernel_clone_args
*kargs
)
6263 __releases(&cgroup_threadgroup_rwsem
) __releases(&cgroup_mutex
)
6265 cgroup_threadgroup_change_end(current
);
6267 if (kargs
->flags
& CLONE_INTO_CGROUP
) {
6268 struct cgroup
*cgrp
= kargs
->cgrp
;
6269 struct css_set
*cset
= kargs
->cset
;
6271 mutex_unlock(&cgroup_mutex
);
6286 * cgroup_can_fork - called on a new task before the process is exposed
6287 * @child: the child process
6288 * @kargs: the arguments passed to create the child process
6290 * This prepares a new css_set for the child process which the child will
6291 * be attached to in cgroup_post_fork().
6292 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork()
6293 * callback returns an error, the fork aborts with that error code. This
6294 * allows for a cgroup subsystem to conditionally allow or deny new forks.
6296 int cgroup_can_fork(struct task_struct
*child
, struct kernel_clone_args
*kargs
)
6298 struct cgroup_subsys
*ss
;
6301 ret
= cgroup_css_set_fork(kargs
);
6305 do_each_subsys_mask(ss
, i
, have_canfork_callback
) {
6306 ret
= ss
->can_fork(child
, kargs
->cset
);
6309 } while_each_subsys_mask();
6314 for_each_subsys(ss
, j
) {
6317 if (ss
->cancel_fork
)
6318 ss
->cancel_fork(child
, kargs
->cset
);
6321 cgroup_css_set_put_fork(kargs
);
6327 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
6328 * @child: the child process
6329 * @kargs: the arguments passed to create the child process
6331 * This calls the cancel_fork() callbacks if a fork failed *after*
6332 * cgroup_can_fork() succeeded and cleans up references we took to
6333 * prepare a new css_set for the child process in cgroup_can_fork().
6335 void cgroup_cancel_fork(struct task_struct
*child
,
6336 struct kernel_clone_args
*kargs
)
6338 struct cgroup_subsys
*ss
;
6341 for_each_subsys(ss
, i
)
6342 if (ss
->cancel_fork
)
6343 ss
->cancel_fork(child
, kargs
->cset
);
6345 cgroup_css_set_put_fork(kargs
);
6349 * cgroup_post_fork - finalize cgroup setup for the child process
6350 * @child: the child process
6351 * @kargs: the arguments passed to create the child process
6353 * Attach the child process to its css_set calling the subsystem fork()
6356 void cgroup_post_fork(struct task_struct
*child
,
6357 struct kernel_clone_args
*kargs
)
6358 __releases(&cgroup_threadgroup_rwsem
) __releases(&cgroup_mutex
)
6360 unsigned long cgrp_flags
= 0;
6362 struct cgroup_subsys
*ss
;
6363 struct css_set
*cset
;
6369 spin_lock_irq(&css_set_lock
);
6371 /* init tasks are special, only link regular threads */
6372 if (likely(child
->pid
)) {
6374 cgrp_flags
= kargs
->cgrp
->flags
;
6376 cgrp_flags
= cset
->dfl_cgrp
->flags
;
6378 WARN_ON_ONCE(!list_empty(&child
->cg_list
));
6380 css_set_move_task(child
, NULL
, cset
, false);
6386 if (!(child
->flags
& PF_KTHREAD
)) {
6387 if (unlikely(test_bit(CGRP_FREEZE
, &cgrp_flags
))) {
6389 * If the cgroup has to be frozen, the new task has
6390 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to
6391 * get the task into the frozen state.
6393 spin_lock(&child
->sighand
->siglock
);
6394 WARN_ON_ONCE(child
->frozen
);
6395 child
->jobctl
|= JOBCTL_TRAP_FREEZE
;
6396 spin_unlock(&child
->sighand
->siglock
);
6399 * Calling cgroup_update_frozen() isn't required here,
6400 * because it will be called anyway a bit later from
6401 * do_freezer_trap(). So we avoid cgroup's transient
6402 * switch from the frozen state and back.
6407 * If the cgroup is to be killed notice it now and take the
6408 * child down right after we finished preparing it for
6411 kill
= test_bit(CGRP_KILL
, &cgrp_flags
);
6414 spin_unlock_irq(&css_set_lock
);
6417 * Call ss->fork(). This must happen after @child is linked on
6418 * css_set; otherwise, @child might change state between ->fork()
6419 * and addition to css_set.
6421 do_each_subsys_mask(ss
, i
, have_fork_callback
) {
6423 } while_each_subsys_mask();
6425 /* Make the new cset the root_cset of the new cgroup namespace. */
6426 if (kargs
->flags
& CLONE_NEWCGROUP
) {
6427 struct css_set
*rcset
= child
->nsproxy
->cgroup_ns
->root_cset
;
6430 child
->nsproxy
->cgroup_ns
->root_cset
= cset
;
6434 /* Cgroup has to be killed so take down child immediately. */
6436 do_send_sig_info(SIGKILL
, SEND_SIG_NOINFO
, child
, PIDTYPE_TGID
);
6438 cgroup_css_set_put_fork(kargs
);
6442 * cgroup_exit - detach cgroup from exiting task
6443 * @tsk: pointer to task_struct of exiting process
6445 * Description: Detach cgroup from @tsk.
6448 void cgroup_exit(struct task_struct
*tsk
)
6450 struct cgroup_subsys
*ss
;
6451 struct css_set
*cset
;
6454 spin_lock_irq(&css_set_lock
);
6456 WARN_ON_ONCE(list_empty(&tsk
->cg_list
));
6457 cset
= task_css_set(tsk
);
6458 css_set_move_task(tsk
, cset
, NULL
, false);
6459 list_add_tail(&tsk
->cg_list
, &cset
->dying_tasks
);
6462 WARN_ON_ONCE(cgroup_task_frozen(tsk
));
6463 if (unlikely(!(tsk
->flags
& PF_KTHREAD
) &&
6464 test_bit(CGRP_FREEZE
, &task_dfl_cgroup(tsk
)->flags
)))
6465 cgroup_update_frozen(task_dfl_cgroup(tsk
));
6467 spin_unlock_irq(&css_set_lock
);
6469 /* see cgroup_post_fork() for details */
6470 do_each_subsys_mask(ss
, i
, have_exit_callback
) {
6472 } while_each_subsys_mask();
6475 void cgroup_release(struct task_struct
*task
)
6477 struct cgroup_subsys
*ss
;
6480 do_each_subsys_mask(ss
, ssid
, have_release_callback
) {
6482 } while_each_subsys_mask();
6484 spin_lock_irq(&css_set_lock
);
6485 css_set_skip_task_iters(task_css_set(task
), task
);
6486 list_del_init(&task
->cg_list
);
6487 spin_unlock_irq(&css_set_lock
);
6490 void cgroup_free(struct task_struct
*task
)
6492 struct css_set
*cset
= task_css_set(task
);
6496 static int __init
cgroup_disable(char *str
)
6498 struct cgroup_subsys
*ss
;
6502 while ((token
= strsep(&str
, ",")) != NULL
) {
6506 for_each_subsys(ss
, i
) {
6507 if (strcmp(token
, ss
->name
) &&
6508 strcmp(token
, ss
->legacy_name
))
6511 static_branch_disable(cgroup_subsys_enabled_key
[i
]);
6512 pr_info("Disabling %s control group subsystem\n",
6516 for (i
= 0; i
< OPT_FEATURE_COUNT
; i
++) {
6517 if (strcmp(token
, cgroup_opt_feature_names
[i
]))
6519 cgroup_feature_disable_mask
|= 1 << i
;
6520 pr_info("Disabling %s control group feature\n",
6521 cgroup_opt_feature_names
[i
]);
6527 __setup("cgroup_disable=", cgroup_disable
);
6529 void __init __weak
enable_debug_cgroup(void) { }
6531 static int __init
enable_cgroup_debug(char *str
)
6533 cgroup_debug
= true;
6534 enable_debug_cgroup();
6537 __setup("cgroup_debug", enable_cgroup_debug
);
6540 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
6541 * @dentry: directory dentry of interest
6542 * @ss: subsystem of interest
6544 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
6545 * to get the corresponding css and return it. If such css doesn't exist
6546 * or can't be pinned, an ERR_PTR value is returned.
6548 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
6549 struct cgroup_subsys
*ss
)
6551 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
6552 struct file_system_type
*s_type
= dentry
->d_sb
->s_type
;
6553 struct cgroup_subsys_state
*css
= NULL
;
6554 struct cgroup
*cgrp
;
6556 /* is @dentry a cgroup dir? */
6557 if ((s_type
!= &cgroup_fs_type
&& s_type
!= &cgroup2_fs_type
) ||
6558 !kn
|| kernfs_type(kn
) != KERNFS_DIR
)
6559 return ERR_PTR(-EBADF
);
6564 * This path doesn't originate from kernfs and @kn could already
6565 * have been or be removed at any point. @kn->priv is RCU
6566 * protected for this access. See css_release_work_fn() for details.
6568 cgrp
= rcu_dereference(*(void __rcu __force
**)&kn
->priv
);
6570 css
= cgroup_css(cgrp
, ss
);
6572 if (!css
|| !css_tryget_online(css
))
6573 css
= ERR_PTR(-ENOENT
);
6580 * css_from_id - lookup css by id
6581 * @id: the cgroup id
6582 * @ss: cgroup subsys to be looked into
6584 * Returns the css if there's valid one with @id, otherwise returns NULL.
6585 * Should be called under rcu_read_lock().
6587 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
6589 WARN_ON_ONCE(!rcu_read_lock_held());
6590 return idr_find(&ss
->css_idr
, id
);
6594 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
6595 * @path: path on the default hierarchy
6597 * Find the cgroup at @path on the default hierarchy, increment its
6598 * reference count and return it. Returns pointer to the found cgroup on
6599 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already
6600 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory.
6602 struct cgroup
*cgroup_get_from_path(const char *path
)
6604 struct kernfs_node
*kn
;
6605 struct cgroup
*cgrp
= ERR_PTR(-ENOENT
);
6607 kn
= kernfs_walk_and_get(cgrp_dfl_root
.cgrp
.kn
, path
);
6611 if (kernfs_type(kn
) != KERNFS_DIR
) {
6612 cgrp
= ERR_PTR(-ENOTDIR
);
6618 cgrp
= rcu_dereference(*(void __rcu __force
**)&kn
->priv
);
6619 if (!cgrp
|| !cgroup_tryget(cgrp
))
6620 cgrp
= ERR_PTR(-ENOENT
);
6629 EXPORT_SYMBOL_GPL(cgroup_get_from_path
);
6632 * cgroup_get_from_fd - get a cgroup pointer from a fd
6633 * @fd: fd obtained by open(cgroup2_dir)
6635 * Find the cgroup from a fd which should be obtained
6636 * by opening a cgroup directory. Returns a pointer to the
6637 * cgroup on success. ERR_PTR is returned if the cgroup
6640 struct cgroup
*cgroup_get_from_fd(int fd
)
6642 struct cgroup
*cgrp
;
6647 return ERR_PTR(-EBADF
);
6649 cgrp
= cgroup_get_from_file(f
);
6653 EXPORT_SYMBOL_GPL(cgroup_get_from_fd
);
6655 static u64
power_of_ten(int power
)
6664 * cgroup_parse_float - parse a floating number
6665 * @input: input string
6666 * @dec_shift: number of decimal digits to shift
6669 * Parse a decimal floating point number in @input and store the result in
6670 * @v with decimal point right shifted @dec_shift times. For example, if
6671 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345.
6672 * Returns 0 on success, -errno otherwise.
6674 * There's nothing cgroup specific about this function except that it's
6675 * currently the only user.
6677 int cgroup_parse_float(const char *input
, unsigned dec_shift
, s64
*v
)
6679 s64 whole
, frac
= 0;
6680 int fstart
= 0, fend
= 0, flen
;
6682 if (!sscanf(input
, "%lld.%n%lld%n", &whole
, &fstart
, &frac
, &fend
))
6687 flen
= fend
> fstart
? fend
- fstart
: 0;
6688 if (flen
< dec_shift
)
6689 frac
*= power_of_ten(dec_shift
- flen
);
6691 frac
= DIV_ROUND_CLOSEST_ULL(frac
, power_of_ten(flen
- dec_shift
));
6693 *v
= whole
* power_of_ten(dec_shift
) + frac
;
6698 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
6699 * definition in cgroup-defs.h.
6701 #ifdef CONFIG_SOCK_CGROUP_DATA
6703 void cgroup_sk_alloc(struct sock_cgroup_data
*skcd
)
6705 struct cgroup
*cgroup
;
6708 /* Don't associate the sock with unrelated interrupted task's cgroup. */
6709 if (in_interrupt()) {
6710 cgroup
= &cgrp_dfl_root
.cgrp
;
6716 struct css_set
*cset
;
6718 cset
= task_css_set(current
);
6719 if (likely(cgroup_tryget(cset
->dfl_cgrp
))) {
6720 cgroup
= cset
->dfl_cgrp
;
6726 skcd
->cgroup
= cgroup
;
6727 cgroup_bpf_get(cgroup
);
6731 void cgroup_sk_clone(struct sock_cgroup_data
*skcd
)
6733 struct cgroup
*cgrp
= sock_cgroup_ptr(skcd
);
6736 * We might be cloning a socket which is left in an empty
6737 * cgroup and the cgroup might have already been rmdir'd.
6738 * Don't use cgroup_get_live().
6741 cgroup_bpf_get(cgrp
);
6744 void cgroup_sk_free(struct sock_cgroup_data
*skcd
)
6746 struct cgroup
*cgrp
= sock_cgroup_ptr(skcd
);
6748 cgroup_bpf_put(cgrp
);
6752 #endif /* CONFIG_SOCK_CGROUP_DATA */
6755 static ssize_t
show_delegatable_files(struct cftype
*files
, char *buf
,
6756 ssize_t size
, const char *prefix
)
6761 for (cft
= files
; cft
&& cft
->name
[0] != '\0'; cft
++) {
6762 if (!(cft
->flags
& CFTYPE_NS_DELEGATABLE
))
6765 if ((cft
->flags
& CFTYPE_PRESSURE
) && !cgroup_psi_enabled())
6769 ret
+= snprintf(buf
+ ret
, size
- ret
, "%s.", prefix
);
6771 ret
+= snprintf(buf
+ ret
, size
- ret
, "%s\n", cft
->name
);
6773 if (WARN_ON(ret
>= size
))
6780 static ssize_t
delegate_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
6783 struct cgroup_subsys
*ss
;
6787 ret
= show_delegatable_files(cgroup_base_files
, buf
, PAGE_SIZE
- ret
,
6790 for_each_subsys(ss
, ssid
)
6791 ret
+= show_delegatable_files(ss
->dfl_cftypes
, buf
+ ret
,
6793 cgroup_subsys_name
[ssid
]);
6797 static struct kobj_attribute cgroup_delegate_attr
= __ATTR_RO(delegate
);
6799 static ssize_t
features_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
6802 return snprintf(buf
, PAGE_SIZE
,
6805 "memory_localevents\n"
6806 "memory_recursiveprot\n");
6808 static struct kobj_attribute cgroup_features_attr
= __ATTR_RO(features
);
6810 static struct attribute
*cgroup_sysfs_attrs
[] = {
6811 &cgroup_delegate_attr
.attr
,
6812 &cgroup_features_attr
.attr
,
6816 static const struct attribute_group cgroup_sysfs_attr_group
= {
6817 .attrs
= cgroup_sysfs_attrs
,
6821 static int __init
cgroup_sysfs_init(void)
6823 return sysfs_create_group(kernel_kobj
, &cgroup_sysfs_attr_group
);
6825 subsys_initcall(cgroup_sysfs_init
);
6827 #endif /* CONFIG_SYSFS */