2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include "cgroup-internal.h"
33 #include <linux/cred.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/magic.h>
38 #include <linux/mutex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/sched/task.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/percpu-rwsem.h>
48 #include <linux/string.h>
49 #include <linux/hashtable.h>
50 #include <linux/idr.h>
51 #include <linux/kthread.h>
52 #include <linux/atomic.h>
53 #include <linux/cpuset.h>
54 #include <linux/proc_ns.h>
55 #include <linux/nsproxy.h>
56 #include <linux/file.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/cgroup.h>
62 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
66 * cgroup_mutex is the master lock. Any modification to cgroup or its
67 * hierarchy must be performed while holding it.
69 * css_set_lock protects task->cgroups pointer, the list of css_set
70 * objects, and the chain of tasks off each css_set.
72 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
73 * cgroup.h can use them for lockdep annotations.
75 DEFINE_MUTEX(cgroup_mutex
);
76 DEFINE_SPINLOCK(css_set_lock
);
78 #ifdef CONFIG_PROVE_RCU
79 EXPORT_SYMBOL_GPL(cgroup_mutex
);
80 EXPORT_SYMBOL_GPL(css_set_lock
);
84 * Protects cgroup_idr and css_idr so that IDs can be released without
85 * grabbing cgroup_mutex.
87 static DEFINE_SPINLOCK(cgroup_idr_lock
);
90 * Protects cgroup_file->kn for !self csses. It synchronizes notifications
91 * against file removal/re-creation across css hiding.
93 static DEFINE_SPINLOCK(cgroup_file_kn_lock
);
95 struct percpu_rw_semaphore cgroup_threadgroup_rwsem
;
97 #define cgroup_assert_mutex_or_rcu_locked() \
98 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
99 !lockdep_is_held(&cgroup_mutex), \
100 "cgroup_mutex or RCU read lock required");
103 * cgroup destruction makes heavy use of work items and there can be a lot
104 * of concurrent destructions. Use a separate workqueue so that cgroup
105 * destruction work items don't end up filling up max_active of system_wq
106 * which may lead to deadlock.
108 static struct workqueue_struct
*cgroup_destroy_wq
;
110 /* generate an array of cgroup subsystem pointers */
111 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
112 struct cgroup_subsys
*cgroup_subsys
[] = {
113 #include <linux/cgroup_subsys.h>
117 /* array of cgroup subsystem names */
118 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
119 static const char *cgroup_subsys_name
[] = {
120 #include <linux/cgroup_subsys.h>
124 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
126 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
127 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
128 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
129 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
130 #include <linux/cgroup_subsys.h>
133 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
134 static struct static_key_true
*cgroup_subsys_enabled_key
[] = {
135 #include <linux/cgroup_subsys.h>
139 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
140 static struct static_key_true
*cgroup_subsys_on_dfl_key
[] = {
141 #include <linux/cgroup_subsys.h>
145 static DEFINE_PER_CPU(struct cgroup_cpu_stat
, cgrp_dfl_root_cpu_stat
);
148 * The default hierarchy, reserved for the subsystems that are otherwise
149 * unattached - it never has more than a single cgroup, and all tasks are
150 * part of that cgroup.
152 struct cgroup_root cgrp_dfl_root
= { .cgrp
.cpu_stat
= &cgrp_dfl_root_cpu_stat
};
153 EXPORT_SYMBOL_GPL(cgrp_dfl_root
);
156 * The default hierarchy always exists but is hidden until mounted for the
157 * first time. This is for backward compatibility.
159 static bool cgrp_dfl_visible
;
161 /* some controllers are not supported in the default hierarchy */
162 static u16 cgrp_dfl_inhibit_ss_mask
;
164 /* some controllers are implicitly enabled on the default hierarchy */
165 static u16 cgrp_dfl_implicit_ss_mask
;
167 /* some controllers can be threaded on the default hierarchy */
168 static u16 cgrp_dfl_threaded_ss_mask
;
170 /* The list of hierarchy roots */
171 LIST_HEAD(cgroup_roots
);
172 static int cgroup_root_count
;
174 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
175 static DEFINE_IDR(cgroup_hierarchy_idr
);
178 * Assign a monotonically increasing serial number to csses. It guarantees
179 * cgroups with bigger numbers are newer than those with smaller numbers.
180 * Also, as csses are always appended to the parent's ->children list, it
181 * guarantees that sibling csses are always sorted in the ascending serial
182 * number order on the list. Protected by cgroup_mutex.
184 static u64 css_serial_nr_next
= 1;
187 * These bitmasks identify subsystems with specific features to avoid
188 * having to do iterative checks repeatedly.
190 static u16 have_fork_callback __read_mostly
;
191 static u16 have_exit_callback __read_mostly
;
192 static u16 have_release_callback __read_mostly
;
193 static u16 have_canfork_callback __read_mostly
;
195 /* cgroup namespace for init task */
196 struct cgroup_namespace init_cgroup_ns
= {
197 .count
= REFCOUNT_INIT(2),
198 .user_ns
= &init_user_ns
,
199 .ns
.ops
= &cgroupns_operations
,
200 .ns
.inum
= PROC_CGROUP_INIT_INO
,
201 .root_cset
= &init_css_set
,
204 static struct file_system_type cgroup2_fs_type
;
205 static struct cftype cgroup_base_files
[];
207 static int cgroup_apply_control(struct cgroup
*cgrp
);
208 static void cgroup_finalize_control(struct cgroup
*cgrp
, int ret
);
209 static void css_task_iter_skip(struct css_task_iter
*it
,
210 struct task_struct
*task
);
211 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
212 static struct cgroup_subsys_state
*css_create(struct cgroup
*cgrp
,
213 struct cgroup_subsys
*ss
);
214 static void css_release(struct percpu_ref
*ref
);
215 static void kill_css(struct cgroup_subsys_state
*css
);
216 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
217 struct cgroup
*cgrp
, struct cftype cfts
[],
221 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
222 * @ssid: subsys ID of interest
224 * cgroup_subsys_enabled() can only be used with literal subsys names which
225 * is fine for individual subsystems but unsuitable for cgroup core. This
226 * is slower static_key_enabled() based test indexed by @ssid.
228 bool cgroup_ssid_enabled(int ssid
)
230 if (CGROUP_SUBSYS_COUNT
== 0)
233 return static_key_enabled(cgroup_subsys_enabled_key
[ssid
]);
237 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
238 * @cgrp: the cgroup of interest
240 * The default hierarchy is the v2 interface of cgroup and this function
241 * can be used to test whether a cgroup is on the default hierarchy for
242 * cases where a subsystem should behave differnetly depending on the
245 * The set of behaviors which change on the default hierarchy are still
246 * being determined and the mount option is prefixed with __DEVEL__.
248 * List of changed behaviors:
250 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
251 * and "name" are disallowed.
253 * - When mounting an existing superblock, mount options should match.
255 * - Remount is disallowed.
257 * - rename(2) is disallowed.
259 * - "tasks" is removed. Everything should be at process granularity. Use
260 * "cgroup.procs" instead.
262 * - "cgroup.procs" is not sorted. pids will be unique unless they got
263 * recycled inbetween reads.
265 * - "release_agent" and "notify_on_release" are removed. Replacement
266 * notification mechanism will be implemented.
268 * - "cgroup.clone_children" is removed.
270 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
271 * and its descendants contain no task; otherwise, 1. The file also
272 * generates kernfs notification which can be monitored through poll and
273 * [di]notify when the value of the file changes.
275 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
276 * take masks of ancestors with non-empty cpus/mems, instead of being
277 * moved to an ancestor.
279 * - cpuset: a task can be moved into an empty cpuset, and again it takes
280 * masks of ancestors.
282 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
285 * - blkcg: blk-throttle becomes properly hierarchical.
287 * - debug: disallowed on the default hierarchy.
289 bool cgroup_on_dfl(const struct cgroup
*cgrp
)
291 return cgrp
->root
== &cgrp_dfl_root
;
294 /* IDR wrappers which synchronize using cgroup_idr_lock */
295 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
300 idr_preload(gfp_mask
);
301 spin_lock_bh(&cgroup_idr_lock
);
302 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
& ~__GFP_DIRECT_RECLAIM
);
303 spin_unlock_bh(&cgroup_idr_lock
);
308 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
312 spin_lock_bh(&cgroup_idr_lock
);
313 ret
= idr_replace(idr
, ptr
, id
);
314 spin_unlock_bh(&cgroup_idr_lock
);
318 static void cgroup_idr_remove(struct idr
*idr
, int id
)
320 spin_lock_bh(&cgroup_idr_lock
);
322 spin_unlock_bh(&cgroup_idr_lock
);
325 static bool cgroup_has_tasks(struct cgroup
*cgrp
)
327 return cgrp
->nr_populated_csets
;
330 bool cgroup_is_threaded(struct cgroup
*cgrp
)
332 return cgrp
->dom_cgrp
!= cgrp
;
335 /* can @cgrp host both domain and threaded children? */
336 static bool cgroup_is_mixable(struct cgroup
*cgrp
)
339 * Root isn't under domain level resource control exempting it from
340 * the no-internal-process constraint, so it can serve as a thread
341 * root and a parent of resource domains at the same time.
343 return !cgroup_parent(cgrp
);
346 /* can @cgrp become a thread root? should always be true for a thread root */
347 static bool cgroup_can_be_thread_root(struct cgroup
*cgrp
)
349 /* mixables don't care */
350 if (cgroup_is_mixable(cgrp
))
353 /* domain roots can't be nested under threaded */
354 if (cgroup_is_threaded(cgrp
))
357 /* can only have either domain or threaded children */
358 if (cgrp
->nr_populated_domain_children
)
361 /* and no domain controllers can be enabled */
362 if (cgrp
->subtree_control
& ~cgrp_dfl_threaded_ss_mask
)
368 /* is @cgrp root of a threaded subtree? */
369 bool cgroup_is_thread_root(struct cgroup
*cgrp
)
371 /* thread root should be a domain */
372 if (cgroup_is_threaded(cgrp
))
375 /* a domain w/ threaded children is a thread root */
376 if (cgrp
->nr_threaded_children
)
380 * A domain which has tasks and explicit threaded controllers
381 * enabled is a thread root.
383 if (cgroup_has_tasks(cgrp
) &&
384 (cgrp
->subtree_control
& cgrp_dfl_threaded_ss_mask
))
390 /* a domain which isn't connected to the root w/o brekage can't be used */
391 static bool cgroup_is_valid_domain(struct cgroup
*cgrp
)
393 /* the cgroup itself can be a thread root */
394 if (cgroup_is_threaded(cgrp
))
397 /* but the ancestors can't be unless mixable */
398 while ((cgrp
= cgroup_parent(cgrp
))) {
399 if (!cgroup_is_mixable(cgrp
) && cgroup_is_thread_root(cgrp
))
401 if (cgroup_is_threaded(cgrp
))
408 /* subsystems visibly enabled on a cgroup */
409 static u16
cgroup_control(struct cgroup
*cgrp
)
411 struct cgroup
*parent
= cgroup_parent(cgrp
);
412 u16 root_ss_mask
= cgrp
->root
->subsys_mask
;
415 u16 ss_mask
= parent
->subtree_control
;
417 /* threaded cgroups can only have threaded controllers */
418 if (cgroup_is_threaded(cgrp
))
419 ss_mask
&= cgrp_dfl_threaded_ss_mask
;
423 if (cgroup_on_dfl(cgrp
))
424 root_ss_mask
&= ~(cgrp_dfl_inhibit_ss_mask
|
425 cgrp_dfl_implicit_ss_mask
);
429 /* subsystems enabled on a cgroup */
430 static u16
cgroup_ss_mask(struct cgroup
*cgrp
)
432 struct cgroup
*parent
= cgroup_parent(cgrp
);
435 u16 ss_mask
= parent
->subtree_ss_mask
;
437 /* threaded cgroups can only have threaded controllers */
438 if (cgroup_is_threaded(cgrp
))
439 ss_mask
&= cgrp_dfl_threaded_ss_mask
;
443 return cgrp
->root
->subsys_mask
;
447 * cgroup_css - obtain a cgroup's css for the specified subsystem
448 * @cgrp: the cgroup of interest
449 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
451 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
452 * function must be called either under cgroup_mutex or rcu_read_lock() and
453 * the caller is responsible for pinning the returned css if it wants to
454 * keep accessing it outside the said locks. This function may return
455 * %NULL if @cgrp doesn't have @subsys_id enabled.
457 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
458 struct cgroup_subsys
*ss
)
461 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
462 lockdep_is_held(&cgroup_mutex
));
468 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
469 * @cgrp: the cgroup of interest
470 * @ss: the subsystem of interest
472 * Find and get @cgrp's css assocaited with @ss. If the css doesn't exist
473 * or is offline, %NULL is returned.
475 static struct cgroup_subsys_state
*cgroup_tryget_css(struct cgroup
*cgrp
,
476 struct cgroup_subsys
*ss
)
478 struct cgroup_subsys_state
*css
;
481 css
= cgroup_css(cgrp
, ss
);
482 if (!css
|| !css_tryget_online(css
))
490 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
491 * @cgrp: the cgroup of interest
492 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
494 * Similar to cgroup_css() but returns the effective css, which is defined
495 * as the matching css of the nearest ancestor including self which has @ss
496 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
497 * function is guaranteed to return non-NULL css.
499 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
500 struct cgroup_subsys
*ss
)
502 lockdep_assert_held(&cgroup_mutex
);
508 * This function is used while updating css associations and thus
509 * can't test the csses directly. Test ss_mask.
511 while (!(cgroup_ss_mask(cgrp
) & (1 << ss
->id
))) {
512 cgrp
= cgroup_parent(cgrp
);
517 return cgroup_css(cgrp
, ss
);
521 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
522 * @cgrp: the cgroup of interest
523 * @ss: the subsystem of interest
525 * Find and get the effective css of @cgrp for @ss. The effective css is
526 * defined as the matching css of the nearest ancestor including self which
527 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
528 * the root css is returned, so this function always returns a valid css.
529 * The returned css must be put using css_put().
531 struct cgroup_subsys_state
*cgroup_get_e_css(struct cgroup
*cgrp
,
532 struct cgroup_subsys
*ss
)
534 struct cgroup_subsys_state
*css
;
539 css
= cgroup_css(cgrp
, ss
);
541 if (css
&& css_tryget_online(css
))
543 cgrp
= cgroup_parent(cgrp
);
546 css
= init_css_set
.subsys
[ss
->id
];
553 static void cgroup_get_live(struct cgroup
*cgrp
)
555 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
556 css_get(&cgrp
->self
);
559 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
561 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
562 struct cftype
*cft
= of_cft(of
);
565 * This is open and unprotected implementation of cgroup_css().
566 * seq_css() is only called from a kernfs file operation which has
567 * an active reference on the file. Because all the subsystem
568 * files are drained before a css is disassociated with a cgroup,
569 * the matching css from the cgroup's subsys table is guaranteed to
570 * be and stay valid until the enclosing operation is complete.
573 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
577 EXPORT_SYMBOL_GPL(of_css
);
580 * for_each_css - iterate all css's of a cgroup
581 * @css: the iteration cursor
582 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
583 * @cgrp: the target cgroup to iterate css's of
585 * Should be called under cgroup_[tree_]mutex.
587 #define for_each_css(css, ssid, cgrp) \
588 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
589 if (!((css) = rcu_dereference_check( \
590 (cgrp)->subsys[(ssid)], \
591 lockdep_is_held(&cgroup_mutex)))) { } \
595 * for_each_e_css - iterate all effective css's of a cgroup
596 * @css: the iteration cursor
597 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
598 * @cgrp: the target cgroup to iterate css's of
600 * Should be called under cgroup_[tree_]mutex.
602 #define for_each_e_css(css, ssid, cgrp) \
603 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
604 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
609 * do_each_subsys_mask - filter for_each_subsys with a bitmask
610 * @ss: the iteration cursor
611 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
612 * @ss_mask: the bitmask
614 * The block will only run for cases where the ssid-th bit (1 << ssid) of
617 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
618 unsigned long __ss_mask = (ss_mask); \
619 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
623 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
624 (ss) = cgroup_subsys[ssid]; \
627 #define while_each_subsys_mask() \
632 /* iterate over child cgrps, lock should be held throughout iteration */
633 #define cgroup_for_each_live_child(child, cgrp) \
634 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
635 if (({ lockdep_assert_held(&cgroup_mutex); \
636 cgroup_is_dead(child); })) \
640 /* walk live descendants in preorder */
641 #define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \
642 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \
643 if (({ lockdep_assert_held(&cgroup_mutex); \
644 (dsct) = (d_css)->cgroup; \
645 cgroup_is_dead(dsct); })) \
649 /* walk live descendants in postorder */
650 #define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \
651 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
652 if (({ lockdep_assert_held(&cgroup_mutex); \
653 (dsct) = (d_css)->cgroup; \
654 cgroup_is_dead(dsct); })) \
659 * The default css_set - used by init and its children prior to any
660 * hierarchies being mounted. It contains a pointer to the root state
661 * for each subsystem. Also used to anchor the list of css_sets. Not
662 * reference-counted, to improve performance when child cgroups
663 * haven't been created.
665 struct css_set init_css_set
= {
666 .refcount
= REFCOUNT_INIT(1),
667 .dom_cset
= &init_css_set
,
668 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
669 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
670 .dying_tasks
= LIST_HEAD_INIT(init_css_set
.dying_tasks
),
671 .task_iters
= LIST_HEAD_INIT(init_css_set
.task_iters
),
672 .threaded_csets
= LIST_HEAD_INIT(init_css_set
.threaded_csets
),
673 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
674 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
675 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
678 * The following field is re-initialized when this cset gets linked
679 * in cgroup_init(). However, let's initialize the field
680 * statically too so that the default cgroup can be accessed safely
683 .dfl_cgrp
= &cgrp_dfl_root
.cgrp
,
686 static int css_set_count
= 1; /* 1 for init_css_set */
688 static bool css_set_threaded(struct css_set
*cset
)
690 return cset
->dom_cset
!= cset
;
694 * css_set_populated - does a css_set contain any tasks?
695 * @cset: target css_set
697 * css_set_populated() should be the same as !!cset->nr_tasks at steady
698 * state. However, css_set_populated() can be called while a task is being
699 * added to or removed from the linked list before the nr_tasks is
700 * properly updated. Hence, we can't just look at ->nr_tasks here.
702 static bool css_set_populated(struct css_set
*cset
)
704 lockdep_assert_held(&css_set_lock
);
706 return !list_empty(&cset
->tasks
) || !list_empty(&cset
->mg_tasks
);
710 * cgroup_update_populated - update the populated count of a cgroup
711 * @cgrp: the target cgroup
712 * @populated: inc or dec populated count
714 * One of the css_sets associated with @cgrp is either getting its first
715 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The
716 * count is propagated towards root so that a given cgroup's
717 * nr_populated_children is zero iff none of its descendants contain any
720 * @cgrp's interface file "cgroup.populated" is zero if both
721 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
722 * 1 otherwise. When the sum changes from or to zero, userland is notified
723 * that the content of the interface file has changed. This can be used to
724 * detect when @cgrp and its descendants become populated or empty.
726 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
728 struct cgroup
*child
= NULL
;
729 int adj
= populated
? 1 : -1;
731 lockdep_assert_held(&css_set_lock
);
734 bool was_populated
= cgroup_is_populated(cgrp
);
737 cgrp
->nr_populated_csets
+= adj
;
739 if (cgroup_is_threaded(child
))
740 cgrp
->nr_populated_threaded_children
+= adj
;
742 cgrp
->nr_populated_domain_children
+= adj
;
745 if (was_populated
== cgroup_is_populated(cgrp
))
748 cgroup1_check_for_release(cgrp
);
749 cgroup_file_notify(&cgrp
->events_file
);
752 cgrp
= cgroup_parent(cgrp
);
757 * css_set_update_populated - update populated state of a css_set
758 * @cset: target css_set
759 * @populated: whether @cset is populated or depopulated
761 * @cset is either getting the first task or losing the last. Update the
762 * populated counters of all associated cgroups accordingly.
764 static void css_set_update_populated(struct css_set
*cset
, bool populated
)
766 struct cgrp_cset_link
*link
;
768 lockdep_assert_held(&css_set_lock
);
770 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
)
771 cgroup_update_populated(link
->cgrp
, populated
);
775 * @task is leaving, advance task iterators which are pointing to it so
776 * that they can resume at the next position. Advancing an iterator might
777 * remove it from the list, use safe walk. See css_task_iter_skip() for
780 static void css_set_skip_task_iters(struct css_set
*cset
,
781 struct task_struct
*task
)
783 struct css_task_iter
*it
, *pos
;
785 list_for_each_entry_safe(it
, pos
, &cset
->task_iters
, iters_node
)
786 css_task_iter_skip(it
, task
);
790 * css_set_move_task - move a task from one css_set to another
791 * @task: task being moved
792 * @from_cset: css_set @task currently belongs to (may be NULL)
793 * @to_cset: new css_set @task is being moved to (may be NULL)
794 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
796 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
797 * css_set, @from_cset can be NULL. If @task is being disassociated
798 * instead of moved, @to_cset can be NULL.
800 * This function automatically handles populated counter updates and
801 * css_task_iter adjustments but the caller is responsible for managing
802 * @from_cset and @to_cset's reference counts.
804 static void css_set_move_task(struct task_struct
*task
,
805 struct css_set
*from_cset
, struct css_set
*to_cset
,
808 lockdep_assert_held(&css_set_lock
);
810 if (to_cset
&& !css_set_populated(to_cset
))
811 css_set_update_populated(to_cset
, true);
814 WARN_ON_ONCE(list_empty(&task
->cg_list
));
816 css_set_skip_task_iters(from_cset
, task
);
817 list_del_init(&task
->cg_list
);
818 if (!css_set_populated(from_cset
))
819 css_set_update_populated(from_cset
, false);
821 WARN_ON_ONCE(!list_empty(&task
->cg_list
));
826 * We are synchronized through cgroup_threadgroup_rwsem
827 * against PF_EXITING setting such that we can't race
828 * against cgroup_exit() changing the css_set to
829 * init_css_set and dropping the old one.
831 WARN_ON_ONCE(task
->flags
& PF_EXITING
);
833 rcu_assign_pointer(task
->cgroups
, to_cset
);
834 list_add_tail(&task
->cg_list
, use_mg_tasks
? &to_cset
->mg_tasks
:
840 * hash table for cgroup groups. This improves the performance to find
841 * an existing css_set. This hash doesn't (currently) take into
842 * account cgroups in empty hierarchies.
844 #define CSS_SET_HASH_BITS 7
845 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
847 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
849 unsigned long key
= 0UL;
850 struct cgroup_subsys
*ss
;
853 for_each_subsys(ss
, i
)
854 key
+= (unsigned long)css
[i
];
855 key
= (key
>> 16) ^ key
;
860 void put_css_set_locked(struct css_set
*cset
)
862 struct cgrp_cset_link
*link
, *tmp_link
;
863 struct cgroup_subsys
*ss
;
866 lockdep_assert_held(&css_set_lock
);
868 if (!refcount_dec_and_test(&cset
->refcount
))
871 WARN_ON_ONCE(!list_empty(&cset
->threaded_csets
));
873 /* This css_set is dead. unlink it and release cgroup and css refs */
874 for_each_subsys(ss
, ssid
) {
875 list_del(&cset
->e_cset_node
[ssid
]);
876 css_put(cset
->subsys
[ssid
]);
878 hash_del(&cset
->hlist
);
881 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
882 list_del(&link
->cset_link
);
883 list_del(&link
->cgrp_link
);
884 if (cgroup_parent(link
->cgrp
))
885 cgroup_put(link
->cgrp
);
889 if (css_set_threaded(cset
)) {
890 list_del(&cset
->threaded_csets_node
);
891 put_css_set_locked(cset
->dom_cset
);
894 kfree_rcu(cset
, rcu_head
);
898 * compare_css_sets - helper function for find_existing_css_set().
899 * @cset: candidate css_set being tested
900 * @old_cset: existing css_set for a task
901 * @new_cgrp: cgroup that's being entered by the task
902 * @template: desired set of css pointers in css_set (pre-calculated)
904 * Returns true if "cset" matches "old_cset" except for the hierarchy
905 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
907 static bool compare_css_sets(struct css_set
*cset
,
908 struct css_set
*old_cset
,
909 struct cgroup
*new_cgrp
,
910 struct cgroup_subsys_state
*template[])
912 struct cgroup
*new_dfl_cgrp
;
913 struct list_head
*l1
, *l2
;
916 * On the default hierarchy, there can be csets which are
917 * associated with the same set of cgroups but different csses.
918 * Let's first ensure that csses match.
920 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
924 /* @cset's domain should match the default cgroup's */
925 if (cgroup_on_dfl(new_cgrp
))
926 new_dfl_cgrp
= new_cgrp
;
928 new_dfl_cgrp
= old_cset
->dfl_cgrp
;
930 if (new_dfl_cgrp
->dom_cgrp
!= cset
->dom_cset
->dfl_cgrp
)
934 * Compare cgroup pointers in order to distinguish between
935 * different cgroups in hierarchies. As different cgroups may
936 * share the same effective css, this comparison is always
939 l1
= &cset
->cgrp_links
;
940 l2
= &old_cset
->cgrp_links
;
942 struct cgrp_cset_link
*link1
, *link2
;
943 struct cgroup
*cgrp1
, *cgrp2
;
947 /* See if we reached the end - both lists are equal length. */
948 if (l1
== &cset
->cgrp_links
) {
949 BUG_ON(l2
!= &old_cset
->cgrp_links
);
952 BUG_ON(l2
== &old_cset
->cgrp_links
);
954 /* Locate the cgroups associated with these links. */
955 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
956 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
959 /* Hierarchies should be linked in the same order. */
960 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
963 * If this hierarchy is the hierarchy of the cgroup
964 * that's changing, then we need to check that this
965 * css_set points to the new cgroup; if it's any other
966 * hierarchy, then this css_set should point to the
967 * same cgroup as the old css_set.
969 if (cgrp1
->root
== new_cgrp
->root
) {
970 if (cgrp1
!= new_cgrp
)
981 * find_existing_css_set - init css array and find the matching css_set
982 * @old_cset: the css_set that we're using before the cgroup transition
983 * @cgrp: the cgroup that we're moving into
984 * @template: out param for the new set of csses, should be clear on entry
986 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
988 struct cgroup_subsys_state
*template[])
990 struct cgroup_root
*root
= cgrp
->root
;
991 struct cgroup_subsys
*ss
;
992 struct css_set
*cset
;
997 * Build the set of subsystem state objects that we want to see in the
998 * new css_set. while subsystems can change globally, the entries here
999 * won't change, so no need for locking.
1001 for_each_subsys(ss
, i
) {
1002 if (root
->subsys_mask
& (1UL << i
)) {
1004 * @ss is in this hierarchy, so we want the
1005 * effective css from @cgrp.
1007 template[i
] = cgroup_e_css(cgrp
, ss
);
1010 * @ss is not in this hierarchy, so we don't want
1011 * to change the css.
1013 template[i
] = old_cset
->subsys
[i
];
1017 key
= css_set_hash(template);
1018 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
1019 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
1022 /* This css_set matches what we need */
1026 /* No existing cgroup group matched */
1030 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
1032 struct cgrp_cset_link
*link
, *tmp_link
;
1034 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
1035 list_del(&link
->cset_link
);
1041 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1042 * @count: the number of links to allocate
1043 * @tmp_links: list_head the allocated links are put on
1045 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1046 * through ->cset_link. Returns 0 on success or -errno.
1048 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
1050 struct cgrp_cset_link
*link
;
1053 INIT_LIST_HEAD(tmp_links
);
1055 for (i
= 0; i
< count
; i
++) {
1056 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
1058 free_cgrp_cset_links(tmp_links
);
1061 list_add(&link
->cset_link
, tmp_links
);
1067 * link_css_set - a helper function to link a css_set to a cgroup
1068 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1069 * @cset: the css_set to be linked
1070 * @cgrp: the destination cgroup
1072 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
1073 struct cgroup
*cgrp
)
1075 struct cgrp_cset_link
*link
;
1077 BUG_ON(list_empty(tmp_links
));
1079 if (cgroup_on_dfl(cgrp
))
1080 cset
->dfl_cgrp
= cgrp
;
1082 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
1087 * Always add links to the tail of the lists so that the lists are
1088 * in choronological order.
1090 list_move_tail(&link
->cset_link
, &cgrp
->cset_links
);
1091 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
1093 if (cgroup_parent(cgrp
))
1094 cgroup_get_live(cgrp
);
1098 * find_css_set - return a new css_set with one cgroup updated
1099 * @old_cset: the baseline css_set
1100 * @cgrp: the cgroup to be updated
1102 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1103 * substituted into the appropriate hierarchy.
1105 static struct css_set
*find_css_set(struct css_set
*old_cset
,
1106 struct cgroup
*cgrp
)
1108 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
1109 struct css_set
*cset
;
1110 struct list_head tmp_links
;
1111 struct cgrp_cset_link
*link
;
1112 struct cgroup_subsys
*ss
;
1116 lockdep_assert_held(&cgroup_mutex
);
1118 /* First see if we already have a cgroup group that matches
1119 * the desired set */
1120 spin_lock_irq(&css_set_lock
);
1121 cset
= find_existing_css_set(old_cset
, cgrp
, template);
1124 spin_unlock_irq(&css_set_lock
);
1129 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
1133 /* Allocate all the cgrp_cset_link objects that we'll need */
1134 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
1139 refcount_set(&cset
->refcount
, 1);
1140 cset
->dom_cset
= cset
;
1141 INIT_LIST_HEAD(&cset
->tasks
);
1142 INIT_LIST_HEAD(&cset
->mg_tasks
);
1143 INIT_LIST_HEAD(&cset
->dying_tasks
);
1144 INIT_LIST_HEAD(&cset
->task_iters
);
1145 INIT_LIST_HEAD(&cset
->threaded_csets
);
1146 INIT_HLIST_NODE(&cset
->hlist
);
1147 INIT_LIST_HEAD(&cset
->cgrp_links
);
1148 INIT_LIST_HEAD(&cset
->mg_preload_node
);
1149 INIT_LIST_HEAD(&cset
->mg_node
);
1151 /* Copy the set of subsystem state objects generated in
1152 * find_existing_css_set() */
1153 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
1155 spin_lock_irq(&css_set_lock
);
1156 /* Add reference counts and links from the new css_set. */
1157 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
1158 struct cgroup
*c
= link
->cgrp
;
1160 if (c
->root
== cgrp
->root
)
1162 link_css_set(&tmp_links
, cset
, c
);
1165 BUG_ON(!list_empty(&tmp_links
));
1169 /* Add @cset to the hash table */
1170 key
= css_set_hash(cset
->subsys
);
1171 hash_add(css_set_table
, &cset
->hlist
, key
);
1173 for_each_subsys(ss
, ssid
) {
1174 struct cgroup_subsys_state
*css
= cset
->subsys
[ssid
];
1176 list_add_tail(&cset
->e_cset_node
[ssid
],
1177 &css
->cgroup
->e_csets
[ssid
]);
1181 spin_unlock_irq(&css_set_lock
);
1184 * If @cset should be threaded, look up the matching dom_cset and
1185 * link them up. We first fully initialize @cset then look for the
1186 * dom_cset. It's simpler this way and safe as @cset is guaranteed
1187 * to stay empty until we return.
1189 if (cgroup_is_threaded(cset
->dfl_cgrp
)) {
1190 struct css_set
*dcset
;
1192 dcset
= find_css_set(cset
, cset
->dfl_cgrp
->dom_cgrp
);
1198 spin_lock_irq(&css_set_lock
);
1199 cset
->dom_cset
= dcset
;
1200 list_add_tail(&cset
->threaded_csets_node
,
1201 &dcset
->threaded_csets
);
1202 spin_unlock_irq(&css_set_lock
);
1208 struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
1210 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
1212 return root_cgrp
->root
;
1215 static int cgroup_init_root_id(struct cgroup_root
*root
)
1219 lockdep_assert_held(&cgroup_mutex
);
1221 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
1225 root
->hierarchy_id
= id
;
1229 static void cgroup_exit_root_id(struct cgroup_root
*root
)
1231 lockdep_assert_held(&cgroup_mutex
);
1233 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
1236 void cgroup_free_root(struct cgroup_root
*root
)
1239 idr_destroy(&root
->cgroup_idr
);
1244 static void cgroup_destroy_root(struct cgroup_root
*root
)
1246 struct cgroup
*cgrp
= &root
->cgrp
;
1247 struct cgrp_cset_link
*link
, *tmp_link
;
1249 trace_cgroup_destroy_root(root
);
1251 cgroup_lock_and_drain_offline(&cgrp_dfl_root
.cgrp
);
1253 BUG_ON(atomic_read(&root
->nr_cgrps
));
1254 BUG_ON(!list_empty(&cgrp
->self
.children
));
1256 /* Rebind all subsystems back to the default hierarchy */
1257 WARN_ON(rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
));
1260 * Release all the links from cset_links to this hierarchy's
1263 spin_lock_irq(&css_set_lock
);
1265 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
1266 list_del(&link
->cset_link
);
1267 list_del(&link
->cgrp_link
);
1271 spin_unlock_irq(&css_set_lock
);
1273 if (!list_empty(&root
->root_list
)) {
1274 list_del(&root
->root_list
);
1275 cgroup_root_count
--;
1278 cgroup_exit_root_id(root
);
1280 mutex_unlock(&cgroup_mutex
);
1282 kernfs_destroy_root(root
->kf_root
);
1283 cgroup_free_root(root
);
1287 * look up cgroup associated with current task's cgroup namespace on the
1288 * specified hierarchy
1290 static struct cgroup
*
1291 current_cgns_cgroup_from_root(struct cgroup_root
*root
)
1293 struct cgroup
*res
= NULL
;
1294 struct css_set
*cset
;
1296 lockdep_assert_held(&css_set_lock
);
1300 cset
= current
->nsproxy
->cgroup_ns
->root_cset
;
1301 if (cset
== &init_css_set
) {
1304 struct cgrp_cset_link
*link
;
1306 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1307 struct cgroup
*c
= link
->cgrp
;
1309 if (c
->root
== root
) {
1321 /* look up cgroup associated with given css_set on the specified hierarchy */
1322 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
1323 struct cgroup_root
*root
)
1325 struct cgroup
*res
= NULL
;
1327 lockdep_assert_held(&cgroup_mutex
);
1328 lockdep_assert_held(&css_set_lock
);
1330 if (cset
== &init_css_set
) {
1332 } else if (root
== &cgrp_dfl_root
) {
1333 res
= cset
->dfl_cgrp
;
1335 struct cgrp_cset_link
*link
;
1337 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1338 struct cgroup
*c
= link
->cgrp
;
1340 if (c
->root
== root
) {
1352 * Return the cgroup for "task" from the given hierarchy. Must be
1353 * called with cgroup_mutex and css_set_lock held.
1355 struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
1356 struct cgroup_root
*root
)
1359 * No need to lock the task - since we hold cgroup_mutex the
1360 * task can't change groups, so the only thing that can happen
1361 * is that it exits and its css is set back to init_css_set.
1363 return cset_cgroup_from_root(task_css_set(task
), root
);
1367 * A task must hold cgroup_mutex to modify cgroups.
1369 * Any task can increment and decrement the count field without lock.
1370 * So in general, code holding cgroup_mutex can't rely on the count
1371 * field not changing. However, if the count goes to zero, then only
1372 * cgroup_attach_task() can increment it again. Because a count of zero
1373 * means that no tasks are currently attached, therefore there is no
1374 * way a task attached to that cgroup can fork (the other way to
1375 * increment the count). So code holding cgroup_mutex can safely
1376 * assume that if the count is zero, it will stay zero. Similarly, if
1377 * a task holds cgroup_mutex on a cgroup with zero count, it
1378 * knows that the cgroup won't be removed, as cgroup_rmdir()
1381 * A cgroup can only be deleted if both its 'count' of using tasks
1382 * is zero, and its list of 'children' cgroups is empty. Since all
1383 * tasks in the system use _some_ cgroup, and since there is always at
1384 * least one task in the system (init, pid == 1), therefore, root cgroup
1385 * always has either children cgroups and/or using tasks. So we don't
1386 * need a special hack to ensure that root cgroup cannot be deleted.
1388 * P.S. One more locking exception. RCU is used to guard the
1389 * update of a tasks cgroup pointer by cgroup_attach_task()
1392 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1394 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1397 struct cgroup_subsys
*ss
= cft
->ss
;
1399 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1400 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
1401 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
1402 cgroup_on_dfl(cgrp
) ? ss
->name
: ss
->legacy_name
,
1405 strlcpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1410 * cgroup_file_mode - deduce file mode of a control file
1411 * @cft: the control file in question
1413 * S_IRUGO for read, S_IWUSR for write.
1415 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1419 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1422 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
) {
1423 if (cft
->flags
& CFTYPE_WORLD_WRITABLE
)
1433 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1434 * @subtree_control: the new subtree_control mask to consider
1435 * @this_ss_mask: available subsystems
1437 * On the default hierarchy, a subsystem may request other subsystems to be
1438 * enabled together through its ->depends_on mask. In such cases, more
1439 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1441 * This function calculates which subsystems need to be enabled if
1442 * @subtree_control is to be applied while restricted to @this_ss_mask.
1444 static u16
cgroup_calc_subtree_ss_mask(u16 subtree_control
, u16 this_ss_mask
)
1446 u16 cur_ss_mask
= subtree_control
;
1447 struct cgroup_subsys
*ss
;
1450 lockdep_assert_held(&cgroup_mutex
);
1452 cur_ss_mask
|= cgrp_dfl_implicit_ss_mask
;
1455 u16 new_ss_mask
= cur_ss_mask
;
1457 do_each_subsys_mask(ss
, ssid
, cur_ss_mask
) {
1458 new_ss_mask
|= ss
->depends_on
;
1459 } while_each_subsys_mask();
1462 * Mask out subsystems which aren't available. This can
1463 * happen only if some depended-upon subsystems were bound
1464 * to non-default hierarchies.
1466 new_ss_mask
&= this_ss_mask
;
1468 if (new_ss_mask
== cur_ss_mask
)
1470 cur_ss_mask
= new_ss_mask
;
1477 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1478 * @kn: the kernfs_node being serviced
1480 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1481 * the method finishes if locking succeeded. Note that once this function
1482 * returns the cgroup returned by cgroup_kn_lock_live() may become
1483 * inaccessible any time. If the caller intends to continue to access the
1484 * cgroup, it should pin it before invoking this function.
1486 void cgroup_kn_unlock(struct kernfs_node
*kn
)
1488 struct cgroup
*cgrp
;
1490 if (kernfs_type(kn
) == KERNFS_DIR
)
1493 cgrp
= kn
->parent
->priv
;
1495 mutex_unlock(&cgroup_mutex
);
1497 kernfs_unbreak_active_protection(kn
);
1502 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1503 * @kn: the kernfs_node being serviced
1504 * @drain_offline: perform offline draining on the cgroup
1506 * This helper is to be used by a cgroup kernfs method currently servicing
1507 * @kn. It breaks the active protection, performs cgroup locking and
1508 * verifies that the associated cgroup is alive. Returns the cgroup if
1509 * alive; otherwise, %NULL. A successful return should be undone by a
1510 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the
1511 * cgroup is drained of offlining csses before return.
1513 * Any cgroup kernfs method implementation which requires locking the
1514 * associated cgroup should use this helper. It avoids nesting cgroup
1515 * locking under kernfs active protection and allows all kernfs operations
1516 * including self-removal.
1518 struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
, bool drain_offline
)
1520 struct cgroup
*cgrp
;
1522 if (kernfs_type(kn
) == KERNFS_DIR
)
1525 cgrp
= kn
->parent
->priv
;
1528 * We're gonna grab cgroup_mutex which nests outside kernfs
1529 * active_ref. cgroup liveliness check alone provides enough
1530 * protection against removal. Ensure @cgrp stays accessible and
1531 * break the active_ref protection.
1533 if (!cgroup_tryget(cgrp
))
1535 kernfs_break_active_protection(kn
);
1538 cgroup_lock_and_drain_offline(cgrp
);
1540 mutex_lock(&cgroup_mutex
);
1542 if (!cgroup_is_dead(cgrp
))
1545 cgroup_kn_unlock(kn
);
1549 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1551 char name
[CGROUP_FILE_NAME_MAX
];
1553 lockdep_assert_held(&cgroup_mutex
);
1555 if (cft
->file_offset
) {
1556 struct cgroup_subsys_state
*css
= cgroup_css(cgrp
, cft
->ss
);
1557 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
1559 spin_lock_irq(&cgroup_file_kn_lock
);
1561 spin_unlock_irq(&cgroup_file_kn_lock
);
1564 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1568 * css_clear_dir - remove subsys files in a cgroup directory
1571 static void css_clear_dir(struct cgroup_subsys_state
*css
)
1573 struct cgroup
*cgrp
= css
->cgroup
;
1574 struct cftype
*cfts
;
1576 if (!(css
->flags
& CSS_VISIBLE
))
1579 css
->flags
&= ~CSS_VISIBLE
;
1581 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
)
1582 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1586 * css_populate_dir - create subsys files in a cgroup directory
1589 * On failure, no file is added.
1591 static int css_populate_dir(struct cgroup_subsys_state
*css
)
1593 struct cgroup
*cgrp
= css
->cgroup
;
1594 struct cftype
*cfts
, *failed_cfts
;
1597 if ((css
->flags
& CSS_VISIBLE
) || !cgrp
->kn
)
1601 if (cgroup_on_dfl(cgrp
))
1602 cfts
= cgroup_base_files
;
1604 cfts
= cgroup1_base_files
;
1606 return cgroup_addrm_files(&cgrp
->self
, cgrp
, cfts
, true);
1609 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1610 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, true);
1617 css
->flags
|= CSS_VISIBLE
;
1621 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1622 if (cfts
== failed_cfts
)
1624 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1629 int rebind_subsystems(struct cgroup_root
*dst_root
, u16 ss_mask
)
1631 struct cgroup
*dcgrp
= &dst_root
->cgrp
;
1632 struct cgroup_subsys
*ss
;
1635 lockdep_assert_held(&cgroup_mutex
);
1637 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1639 * If @ss has non-root csses attached to it, can't move.
1640 * If @ss is an implicit controller, it is exempt from this
1641 * rule and can be stolen.
1643 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)) &&
1644 !ss
->implicit_on_dfl
)
1647 /* can't move between two non-dummy roots either */
1648 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1650 } while_each_subsys_mask();
1652 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1653 struct cgroup_root
*src_root
= ss
->root
;
1654 struct cgroup
*scgrp
= &src_root
->cgrp
;
1655 struct cgroup_subsys_state
*css
= cgroup_css(scgrp
, ss
);
1656 struct css_set
*cset
;
1658 WARN_ON(!css
|| cgroup_css(dcgrp
, ss
));
1660 /* disable from the source */
1661 src_root
->subsys_mask
&= ~(1 << ssid
);
1662 WARN_ON(cgroup_apply_control(scgrp
));
1663 cgroup_finalize_control(scgrp
, 0);
1666 RCU_INIT_POINTER(scgrp
->subsys
[ssid
], NULL
);
1667 rcu_assign_pointer(dcgrp
->subsys
[ssid
], css
);
1668 ss
->root
= dst_root
;
1669 css
->cgroup
= dcgrp
;
1671 spin_lock_irq(&css_set_lock
);
1672 hash_for_each(css_set_table
, i
, cset
, hlist
)
1673 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1674 &dcgrp
->e_csets
[ss
->id
]);
1675 spin_unlock_irq(&css_set_lock
);
1677 /* default hierarchy doesn't enable controllers by default */
1678 dst_root
->subsys_mask
|= 1 << ssid
;
1679 if (dst_root
== &cgrp_dfl_root
) {
1680 static_branch_enable(cgroup_subsys_on_dfl_key
[ssid
]);
1682 dcgrp
->subtree_control
|= 1 << ssid
;
1683 static_branch_disable(cgroup_subsys_on_dfl_key
[ssid
]);
1686 ret
= cgroup_apply_control(dcgrp
);
1688 pr_warn("partial failure to rebind %s controller (err=%d)\n",
1693 } while_each_subsys_mask();
1695 kernfs_activate(dcgrp
->kn
);
1699 int cgroup_show_path(struct seq_file
*sf
, struct kernfs_node
*kf_node
,
1700 struct kernfs_root
*kf_root
)
1704 struct cgroup_root
*kf_cgroot
= cgroup_root_from_kf(kf_root
);
1705 struct cgroup
*ns_cgroup
;
1707 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
1711 spin_lock_irq(&css_set_lock
);
1712 ns_cgroup
= current_cgns_cgroup_from_root(kf_cgroot
);
1713 len
= kernfs_path_from_node(kf_node
, ns_cgroup
->kn
, buf
, PATH_MAX
);
1714 spin_unlock_irq(&css_set_lock
);
1716 if (len
>= PATH_MAX
)
1719 seq_escape(sf
, buf
, " \t\n\\");
1726 static int parse_cgroup_root_flags(char *data
, unsigned int *root_flags
)
1732 if (!data
|| *data
== '\0')
1735 while ((token
= strsep(&data
, ",")) != NULL
) {
1736 if (!strcmp(token
, "nsdelegate")) {
1737 *root_flags
|= CGRP_ROOT_NS_DELEGATE
;
1741 pr_err("cgroup2: unknown option \"%s\"\n", token
);
1748 static void apply_cgroup_root_flags(unsigned int root_flags
)
1750 if (current
->nsproxy
->cgroup_ns
== &init_cgroup_ns
) {
1751 if (root_flags
& CGRP_ROOT_NS_DELEGATE
)
1752 cgrp_dfl_root
.flags
|= CGRP_ROOT_NS_DELEGATE
;
1754 cgrp_dfl_root
.flags
&= ~CGRP_ROOT_NS_DELEGATE
;
1758 static int cgroup_show_options(struct seq_file
*seq
, struct kernfs_root
*kf_root
)
1760 if (cgrp_dfl_root
.flags
& CGRP_ROOT_NS_DELEGATE
)
1761 seq_puts(seq
, ",nsdelegate");
1765 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1767 unsigned int root_flags
;
1770 ret
= parse_cgroup_root_flags(data
, &root_flags
);
1774 apply_cgroup_root_flags(root_flags
);
1779 * To reduce the fork() overhead for systems that are not actually using
1780 * their cgroups capability, we don't maintain the lists running through
1781 * each css_set to its tasks until we see the list actually used - in other
1782 * words after the first mount.
1784 static bool use_task_css_set_links __read_mostly
;
1786 static void cgroup_enable_task_cg_lists(void)
1788 struct task_struct
*p
, *g
;
1790 spin_lock_irq(&css_set_lock
);
1792 if (use_task_css_set_links
)
1795 use_task_css_set_links
= true;
1798 * We need tasklist_lock because RCU is not safe against
1799 * while_each_thread(). Besides, a forking task that has passed
1800 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1801 * is not guaranteed to have its child immediately visible in the
1802 * tasklist if we walk through it with RCU.
1804 read_lock(&tasklist_lock
);
1805 do_each_thread(g
, p
) {
1806 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1807 task_css_set(p
) != &init_css_set
);
1810 * We should check if the process is exiting, otherwise
1811 * it will race with cgroup_exit() in that the list
1812 * entry won't be deleted though the process has exited.
1813 * Do it while holding siglock so that we don't end up
1814 * racing against cgroup_exit().
1816 * Interrupts were already disabled while acquiring
1817 * the css_set_lock, so we do not need to disable it
1818 * again when acquiring the sighand->siglock here.
1820 spin_lock(&p
->sighand
->siglock
);
1821 if (!(p
->flags
& PF_EXITING
)) {
1822 struct css_set
*cset
= task_css_set(p
);
1824 if (!css_set_populated(cset
))
1825 css_set_update_populated(cset
, true);
1826 list_add_tail(&p
->cg_list
, &cset
->tasks
);
1830 spin_unlock(&p
->sighand
->siglock
);
1831 } while_each_thread(g
, p
);
1832 read_unlock(&tasklist_lock
);
1834 spin_unlock_irq(&css_set_lock
);
1837 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1839 struct cgroup_subsys
*ss
;
1842 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1843 INIT_LIST_HEAD(&cgrp
->self
.children
);
1844 INIT_LIST_HEAD(&cgrp
->cset_links
);
1845 INIT_LIST_HEAD(&cgrp
->pidlists
);
1846 mutex_init(&cgrp
->pidlist_mutex
);
1847 cgrp
->self
.cgroup
= cgrp
;
1848 cgrp
->self
.flags
|= CSS_ONLINE
;
1849 cgrp
->dom_cgrp
= cgrp
;
1850 cgrp
->max_descendants
= INT_MAX
;
1851 cgrp
->max_depth
= INT_MAX
;
1853 for_each_subsys(ss
, ssid
)
1854 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1856 init_waitqueue_head(&cgrp
->offline_waitq
);
1857 INIT_WORK(&cgrp
->release_agent_work
, cgroup1_release_agent
);
1860 void init_cgroup_root(struct cgroup_root
*root
, struct cgroup_sb_opts
*opts
)
1862 struct cgroup
*cgrp
= &root
->cgrp
;
1864 INIT_LIST_HEAD(&root
->root_list
);
1865 atomic_set(&root
->nr_cgrps
, 1);
1867 init_cgroup_housekeeping(cgrp
);
1868 idr_init(&root
->cgroup_idr
);
1870 root
->flags
= opts
->flags
;
1871 if (opts
->release_agent
)
1872 strlcpy(root
->release_agent_path
, opts
->release_agent
, PATH_MAX
);
1874 strlcpy(root
->name
, opts
->name
, MAX_CGROUP_ROOT_NAMELEN
);
1875 if (opts
->cpuset_clone_children
)
1876 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1879 int cgroup_setup_root(struct cgroup_root
*root
, u16 ss_mask
, int ref_flags
)
1881 LIST_HEAD(tmp_links
);
1882 struct cgroup
*root_cgrp
= &root
->cgrp
;
1883 struct kernfs_syscall_ops
*kf_sops
;
1884 struct css_set
*cset
;
1887 lockdep_assert_held(&cgroup_mutex
);
1889 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_KERNEL
);
1892 root_cgrp
->id
= ret
;
1893 root_cgrp
->ancestor_ids
[0] = ret
;
1895 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
,
1896 ref_flags
, GFP_KERNEL
);
1901 * We're accessing css_set_count without locking css_set_lock here,
1902 * but that's OK - it can only be increased by someone holding
1903 * cgroup_lock, and that's us. Later rebinding may disable
1904 * controllers on the default hierarchy and thus create new csets,
1905 * which can't be more than the existing ones. Allocate 2x.
1907 ret
= allocate_cgrp_cset_links(2 * css_set_count
, &tmp_links
);
1911 ret
= cgroup_init_root_id(root
);
1915 kf_sops
= root
== &cgrp_dfl_root
?
1916 &cgroup_kf_syscall_ops
: &cgroup1_kf_syscall_ops
;
1918 root
->kf_root
= kernfs_create_root(kf_sops
,
1919 KERNFS_ROOT_CREATE_DEACTIVATED
|
1920 KERNFS_ROOT_SUPPORT_EXPORTOP
,
1922 if (IS_ERR(root
->kf_root
)) {
1923 ret
= PTR_ERR(root
->kf_root
);
1926 root_cgrp
->kn
= root
->kf_root
->kn
;
1928 ret
= css_populate_dir(&root_cgrp
->self
);
1932 ret
= rebind_subsystems(root
, ss_mask
);
1936 ret
= cgroup_bpf_inherit(root_cgrp
);
1939 trace_cgroup_setup_root(root
);
1942 * There must be no failure case after here, since rebinding takes
1943 * care of subsystems' refcounts, which are explicitly dropped in
1944 * the failure exit path.
1946 list_add(&root
->root_list
, &cgroup_roots
);
1947 cgroup_root_count
++;
1950 * Link the root cgroup in this hierarchy into all the css_set
1953 spin_lock_irq(&css_set_lock
);
1954 hash_for_each(css_set_table
, i
, cset
, hlist
) {
1955 link_css_set(&tmp_links
, cset
, root_cgrp
);
1956 if (css_set_populated(cset
))
1957 cgroup_update_populated(root_cgrp
, true);
1959 spin_unlock_irq(&css_set_lock
);
1961 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
1962 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1964 kernfs_activate(root_cgrp
->kn
);
1969 kernfs_destroy_root(root
->kf_root
);
1970 root
->kf_root
= NULL
;
1972 cgroup_exit_root_id(root
);
1974 percpu_ref_exit(&root_cgrp
->self
.refcnt
);
1976 free_cgrp_cset_links(&tmp_links
);
1980 struct dentry
*cgroup_do_mount(struct file_system_type
*fs_type
, int flags
,
1981 struct cgroup_root
*root
, unsigned long magic
,
1982 struct cgroup_namespace
*ns
)
1984 struct dentry
*dentry
;
1985 bool new_sb
= false;
1987 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
, magic
, &new_sb
);
1990 * In non-init cgroup namespace, instead of root cgroup's dentry,
1991 * we return the dentry corresponding to the cgroupns->root_cgrp.
1993 if (!IS_ERR(dentry
) && ns
!= &init_cgroup_ns
) {
1994 struct dentry
*nsdentry
;
1995 struct super_block
*sb
= dentry
->d_sb
;
1996 struct cgroup
*cgrp
;
1998 mutex_lock(&cgroup_mutex
);
1999 spin_lock_irq(&css_set_lock
);
2001 cgrp
= cset_cgroup_from_root(ns
->root_cset
, root
);
2003 spin_unlock_irq(&css_set_lock
);
2004 mutex_unlock(&cgroup_mutex
);
2006 nsdentry
= kernfs_node_dentry(cgrp
->kn
, sb
);
2008 if (IS_ERR(nsdentry
))
2009 deactivate_locked_super(sb
);
2014 cgroup_put(&root
->cgrp
);
2019 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
2020 int flags
, const char *unused_dev_name
,
2023 struct cgroup_namespace
*ns
= current
->nsproxy
->cgroup_ns
;
2024 struct dentry
*dentry
;
2029 /* Check if the caller has permission to mount. */
2030 if (!ns_capable(ns
->user_ns
, CAP_SYS_ADMIN
)) {
2032 return ERR_PTR(-EPERM
);
2036 * The first time anyone tries to mount a cgroup, enable the list
2037 * linking each css_set to its tasks and fix up all existing tasks.
2039 if (!use_task_css_set_links
)
2040 cgroup_enable_task_cg_lists();
2042 if (fs_type
== &cgroup2_fs_type
) {
2043 unsigned int root_flags
;
2045 ret
= parse_cgroup_root_flags(data
, &root_flags
);
2048 return ERR_PTR(ret
);
2051 cgrp_dfl_visible
= true;
2052 cgroup_get_live(&cgrp_dfl_root
.cgrp
);
2054 dentry
= cgroup_do_mount(&cgroup2_fs_type
, flags
, &cgrp_dfl_root
,
2055 CGROUP2_SUPER_MAGIC
, ns
);
2056 if (!IS_ERR(dentry
))
2057 apply_cgroup_root_flags(root_flags
);
2059 dentry
= cgroup1_mount(&cgroup_fs_type
, flags
, data
,
2060 CGROUP_SUPER_MAGIC
, ns
);
2067 static void cgroup_kill_sb(struct super_block
*sb
)
2069 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
2070 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
2073 * If @root doesn't have any mounts or children, start killing it.
2074 * This prevents new mounts by disabling percpu_ref_tryget_live().
2075 * cgroup_mount() may wait for @root's release.
2077 * And don't kill the default root.
2079 if (!list_empty(&root
->cgrp
.self
.children
) ||
2080 root
== &cgrp_dfl_root
)
2081 cgroup_put(&root
->cgrp
);
2083 percpu_ref_kill(&root
->cgrp
.self
.refcnt
);
2088 struct file_system_type cgroup_fs_type
= {
2090 .mount
= cgroup_mount
,
2091 .kill_sb
= cgroup_kill_sb
,
2092 .fs_flags
= FS_USERNS_MOUNT
,
2095 static struct file_system_type cgroup2_fs_type
= {
2097 .mount
= cgroup_mount
,
2098 .kill_sb
= cgroup_kill_sb
,
2099 .fs_flags
= FS_USERNS_MOUNT
,
2102 int cgroup_path_ns_locked(struct cgroup
*cgrp
, char *buf
, size_t buflen
,
2103 struct cgroup_namespace
*ns
)
2105 struct cgroup
*root
= cset_cgroup_from_root(ns
->root_cset
, cgrp
->root
);
2107 return kernfs_path_from_node(cgrp
->kn
, root
->kn
, buf
, buflen
);
2110 int cgroup_path_ns(struct cgroup
*cgrp
, char *buf
, size_t buflen
,
2111 struct cgroup_namespace
*ns
)
2115 mutex_lock(&cgroup_mutex
);
2116 spin_lock_irq(&css_set_lock
);
2118 ret
= cgroup_path_ns_locked(cgrp
, buf
, buflen
, ns
);
2120 spin_unlock_irq(&css_set_lock
);
2121 mutex_unlock(&cgroup_mutex
);
2125 EXPORT_SYMBOL_GPL(cgroup_path_ns
);
2128 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2129 * @task: target task
2130 * @buf: the buffer to write the path into
2131 * @buflen: the length of the buffer
2133 * Determine @task's cgroup on the first (the one with the lowest non-zero
2134 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2135 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2136 * cgroup controller callbacks.
2138 * Return value is the same as kernfs_path().
2140 int task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
2142 struct cgroup_root
*root
;
2143 struct cgroup
*cgrp
;
2144 int hierarchy_id
= 1;
2147 mutex_lock(&cgroup_mutex
);
2148 spin_lock_irq(&css_set_lock
);
2150 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
2153 cgrp
= task_cgroup_from_root(task
, root
);
2154 ret
= cgroup_path_ns_locked(cgrp
, buf
, buflen
, &init_cgroup_ns
);
2156 /* if no hierarchy exists, everyone is in "/" */
2157 ret
= strlcpy(buf
, "/", buflen
);
2160 spin_unlock_irq(&css_set_lock
);
2161 mutex_unlock(&cgroup_mutex
);
2164 EXPORT_SYMBOL_GPL(task_cgroup_path
);
2167 * cgroup_migrate_add_task - add a migration target task to a migration context
2168 * @task: target task
2169 * @mgctx: target migration context
2171 * Add @task, which is a migration target, to @mgctx->tset. This function
2172 * becomes noop if @task doesn't need to be migrated. @task's css_set
2173 * should have been added as a migration source and @task->cg_list will be
2174 * moved from the css_set's tasks list to mg_tasks one.
2176 static void cgroup_migrate_add_task(struct task_struct
*task
,
2177 struct cgroup_mgctx
*mgctx
)
2179 struct css_set
*cset
;
2181 lockdep_assert_held(&css_set_lock
);
2183 /* @task either already exited or can't exit until the end */
2184 if (task
->flags
& PF_EXITING
)
2187 /* leave @task alone if post_fork() hasn't linked it yet */
2188 if (list_empty(&task
->cg_list
))
2191 cset
= task_css_set(task
);
2192 if (!cset
->mg_src_cgrp
)
2195 mgctx
->tset
.nr_tasks
++;
2197 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2198 if (list_empty(&cset
->mg_node
))
2199 list_add_tail(&cset
->mg_node
,
2200 &mgctx
->tset
.src_csets
);
2201 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2202 list_add_tail(&cset
->mg_dst_cset
->mg_node
,
2203 &mgctx
->tset
.dst_csets
);
2207 * cgroup_taskset_first - reset taskset and return the first task
2208 * @tset: taskset of interest
2209 * @dst_cssp: output variable for the destination css
2211 * @tset iteration is initialized and the first task is returned.
2213 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
,
2214 struct cgroup_subsys_state
**dst_cssp
)
2216 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
2217 tset
->cur_task
= NULL
;
2219 return cgroup_taskset_next(tset
, dst_cssp
);
2223 * cgroup_taskset_next - iterate to the next task in taskset
2224 * @tset: taskset of interest
2225 * @dst_cssp: output variable for the destination css
2227 * Return the next task in @tset. Iteration must have been initialized
2228 * with cgroup_taskset_first().
2230 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
,
2231 struct cgroup_subsys_state
**dst_cssp
)
2233 struct css_set
*cset
= tset
->cur_cset
;
2234 struct task_struct
*task
= tset
->cur_task
;
2236 while (&cset
->mg_node
!= tset
->csets
) {
2238 task
= list_first_entry(&cset
->mg_tasks
,
2239 struct task_struct
, cg_list
);
2241 task
= list_next_entry(task
, cg_list
);
2243 if (&task
->cg_list
!= &cset
->mg_tasks
) {
2244 tset
->cur_cset
= cset
;
2245 tset
->cur_task
= task
;
2248 * This function may be called both before and
2249 * after cgroup_taskset_migrate(). The two cases
2250 * can be distinguished by looking at whether @cset
2251 * has its ->mg_dst_cset set.
2253 if (cset
->mg_dst_cset
)
2254 *dst_cssp
= cset
->mg_dst_cset
->subsys
[tset
->ssid
];
2256 *dst_cssp
= cset
->subsys
[tset
->ssid
];
2261 cset
= list_next_entry(cset
, mg_node
);
2269 * cgroup_taskset_migrate - migrate a taskset
2270 * @mgctx: migration context
2272 * Migrate tasks in @mgctx as setup by migration preparation functions.
2273 * This function fails iff one of the ->can_attach callbacks fails and
2274 * guarantees that either all or none of the tasks in @mgctx are migrated.
2275 * @mgctx is consumed regardless of success.
2277 static int cgroup_migrate_execute(struct cgroup_mgctx
*mgctx
)
2279 struct cgroup_taskset
*tset
= &mgctx
->tset
;
2280 struct cgroup_subsys
*ss
;
2281 struct task_struct
*task
, *tmp_task
;
2282 struct css_set
*cset
, *tmp_cset
;
2283 int ssid
, failed_ssid
, ret
;
2285 /* check that we can legitimately attach to the cgroup */
2286 if (tset
->nr_tasks
) {
2287 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2288 if (ss
->can_attach
) {
2290 ret
= ss
->can_attach(tset
);
2293 goto out_cancel_attach
;
2296 } while_each_subsys_mask();
2300 * Now that we're guaranteed success, proceed to move all tasks to
2301 * the new cgroup. There are no failure cases after here, so this
2302 * is the commit point.
2304 spin_lock_irq(&css_set_lock
);
2305 list_for_each_entry(cset
, &tset
->src_csets
, mg_node
) {
2306 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
) {
2307 struct css_set
*from_cset
= task_css_set(task
);
2308 struct css_set
*to_cset
= cset
->mg_dst_cset
;
2310 get_css_set(to_cset
);
2311 to_cset
->nr_tasks
++;
2312 css_set_move_task(task
, from_cset
, to_cset
, true);
2313 put_css_set_locked(from_cset
);
2314 from_cset
->nr_tasks
--;
2317 spin_unlock_irq(&css_set_lock
);
2320 * Migration is committed, all target tasks are now on dst_csets.
2321 * Nothing is sensitive to fork() after this point. Notify
2322 * controllers that migration is complete.
2324 tset
->csets
= &tset
->dst_csets
;
2326 if (tset
->nr_tasks
) {
2327 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2332 } while_each_subsys_mask();
2336 goto out_release_tset
;
2339 if (tset
->nr_tasks
) {
2340 do_each_subsys_mask(ss
, ssid
, mgctx
->ss_mask
) {
2341 if (ssid
== failed_ssid
)
2343 if (ss
->cancel_attach
) {
2345 ss
->cancel_attach(tset
);
2347 } while_each_subsys_mask();
2350 spin_lock_irq(&css_set_lock
);
2351 list_splice_init(&tset
->dst_csets
, &tset
->src_csets
);
2352 list_for_each_entry_safe(cset
, tmp_cset
, &tset
->src_csets
, mg_node
) {
2353 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2354 list_del_init(&cset
->mg_node
);
2356 spin_unlock_irq(&css_set_lock
);
2359 * Re-initialize the cgroup_taskset structure in case it is reused
2360 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2364 tset
->csets
= &tset
->src_csets
;
2369 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2370 * @dst_cgrp: destination cgroup to test
2372 * On the default hierarchy, except for the mixable, (possible) thread root
2373 * and threaded cgroups, subtree_control must be zero for migration
2374 * destination cgroups with tasks so that child cgroups don't compete
2377 int cgroup_migrate_vet_dst(struct cgroup
*dst_cgrp
)
2379 /* v1 doesn't have any restriction */
2380 if (!cgroup_on_dfl(dst_cgrp
))
2383 /* verify @dst_cgrp can host resources */
2384 if (!cgroup_is_valid_domain(dst_cgrp
->dom_cgrp
))
2387 /* mixables don't care */
2388 if (cgroup_is_mixable(dst_cgrp
))
2392 * If @dst_cgrp is already or can become a thread root or is
2393 * threaded, it doesn't matter.
2395 if (cgroup_can_be_thread_root(dst_cgrp
) || cgroup_is_threaded(dst_cgrp
))
2398 /* apply no-internal-process constraint */
2399 if (dst_cgrp
->subtree_control
)
2406 * cgroup_migrate_finish - cleanup after attach
2407 * @mgctx: migration context
2409 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2410 * those functions for details.
2412 void cgroup_migrate_finish(struct cgroup_mgctx
*mgctx
)
2414 LIST_HEAD(preloaded
);
2415 struct css_set
*cset
, *tmp_cset
;
2417 lockdep_assert_held(&cgroup_mutex
);
2419 spin_lock_irq(&css_set_lock
);
2421 list_splice_tail_init(&mgctx
->preloaded_src_csets
, &preloaded
);
2422 list_splice_tail_init(&mgctx
->preloaded_dst_csets
, &preloaded
);
2424 list_for_each_entry_safe(cset
, tmp_cset
, &preloaded
, mg_preload_node
) {
2425 cset
->mg_src_cgrp
= NULL
;
2426 cset
->mg_dst_cgrp
= NULL
;
2427 cset
->mg_dst_cset
= NULL
;
2428 list_del_init(&cset
->mg_preload_node
);
2429 put_css_set_locked(cset
);
2432 spin_unlock_irq(&css_set_lock
);
2436 * cgroup_migrate_add_src - add a migration source css_set
2437 * @src_cset: the source css_set to add
2438 * @dst_cgrp: the destination cgroup
2439 * @mgctx: migration context
2441 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2442 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2443 * up by cgroup_migrate_finish().
2445 * This function may be called without holding cgroup_threadgroup_rwsem
2446 * even if the target is a process. Threads may be created and destroyed
2447 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2448 * into play and the preloaded css_sets are guaranteed to cover all
2451 void cgroup_migrate_add_src(struct css_set
*src_cset
,
2452 struct cgroup
*dst_cgrp
,
2453 struct cgroup_mgctx
*mgctx
)
2455 struct cgroup
*src_cgrp
;
2457 lockdep_assert_held(&cgroup_mutex
);
2458 lockdep_assert_held(&css_set_lock
);
2461 * If ->dead, @src_set is associated with one or more dead cgroups
2462 * and doesn't contain any migratable tasks. Ignore it early so
2463 * that the rest of migration path doesn't get confused by it.
2468 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2470 if (!list_empty(&src_cset
->mg_preload_node
))
2473 WARN_ON(src_cset
->mg_src_cgrp
);
2474 WARN_ON(src_cset
->mg_dst_cgrp
);
2475 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2476 WARN_ON(!list_empty(&src_cset
->mg_node
));
2478 src_cset
->mg_src_cgrp
= src_cgrp
;
2479 src_cset
->mg_dst_cgrp
= dst_cgrp
;
2480 get_css_set(src_cset
);
2481 list_add_tail(&src_cset
->mg_preload_node
, &mgctx
->preloaded_src_csets
);
2485 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2486 * @mgctx: migration context
2488 * Tasks are about to be moved and all the source css_sets have been
2489 * preloaded to @mgctx->preloaded_src_csets. This function looks up and
2490 * pins all destination css_sets, links each to its source, and append them
2491 * to @mgctx->preloaded_dst_csets.
2493 * This function must be called after cgroup_migrate_add_src() has been
2494 * called on each migration source css_set. After migration is performed
2495 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2498 int cgroup_migrate_prepare_dst(struct cgroup_mgctx
*mgctx
)
2500 struct css_set
*src_cset
, *tmp_cset
;
2502 lockdep_assert_held(&cgroup_mutex
);
2504 /* look up the dst cset for each src cset and link it to src */
2505 list_for_each_entry_safe(src_cset
, tmp_cset
, &mgctx
->preloaded_src_csets
,
2507 struct css_set
*dst_cset
;
2508 struct cgroup_subsys
*ss
;
2511 dst_cset
= find_css_set(src_cset
, src_cset
->mg_dst_cgrp
);
2515 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2518 * If src cset equals dst, it's noop. Drop the src.
2519 * cgroup_migrate() will skip the cset too. Note that we
2520 * can't handle src == dst as some nodes are used by both.
2522 if (src_cset
== dst_cset
) {
2523 src_cset
->mg_src_cgrp
= NULL
;
2524 src_cset
->mg_dst_cgrp
= NULL
;
2525 list_del_init(&src_cset
->mg_preload_node
);
2526 put_css_set(src_cset
);
2527 put_css_set(dst_cset
);
2531 src_cset
->mg_dst_cset
= dst_cset
;
2533 if (list_empty(&dst_cset
->mg_preload_node
))
2534 list_add_tail(&dst_cset
->mg_preload_node
,
2535 &mgctx
->preloaded_dst_csets
);
2537 put_css_set(dst_cset
);
2539 for_each_subsys(ss
, ssid
)
2540 if (src_cset
->subsys
[ssid
] != dst_cset
->subsys
[ssid
])
2541 mgctx
->ss_mask
|= 1 << ssid
;
2546 cgroup_migrate_finish(mgctx
);
2551 * cgroup_migrate - migrate a process or task to a cgroup
2552 * @leader: the leader of the process or the task to migrate
2553 * @threadgroup: whether @leader points to the whole process or a single task
2554 * @mgctx: migration context
2556 * Migrate a process or task denoted by @leader. If migrating a process,
2557 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also
2558 * responsible for invoking cgroup_migrate_add_src() and
2559 * cgroup_migrate_prepare_dst() on the targets before invoking this
2560 * function and following up with cgroup_migrate_finish().
2562 * As long as a controller's ->can_attach() doesn't fail, this function is
2563 * guaranteed to succeed. This means that, excluding ->can_attach()
2564 * failure, when migrating multiple targets, the success or failure can be
2565 * decided for all targets by invoking group_migrate_prepare_dst() before
2566 * actually starting migrating.
2568 int cgroup_migrate(struct task_struct
*leader
, bool threadgroup
,
2569 struct cgroup_mgctx
*mgctx
)
2571 struct task_struct
*task
;
2574 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2575 * already PF_EXITING could be freed from underneath us unless we
2576 * take an rcu_read_lock.
2578 spin_lock_irq(&css_set_lock
);
2582 cgroup_migrate_add_task(task
, mgctx
);
2585 } while_each_thread(leader
, task
);
2587 spin_unlock_irq(&css_set_lock
);
2589 return cgroup_migrate_execute(mgctx
);
2593 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2594 * @dst_cgrp: the cgroup to attach to
2595 * @leader: the task or the leader of the threadgroup to be attached
2596 * @threadgroup: attach the whole threadgroup?
2598 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2600 int cgroup_attach_task(struct cgroup
*dst_cgrp
, struct task_struct
*leader
,
2603 DEFINE_CGROUP_MGCTX(mgctx
);
2604 struct task_struct
*task
;
2607 ret
= cgroup_migrate_vet_dst(dst_cgrp
);
2611 /* look up all src csets */
2612 spin_lock_irq(&css_set_lock
);
2616 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
, &mgctx
);
2619 } while_each_thread(leader
, task
);
2621 spin_unlock_irq(&css_set_lock
);
2623 /* prepare dst csets and commit */
2624 ret
= cgroup_migrate_prepare_dst(&mgctx
);
2626 ret
= cgroup_migrate(leader
, threadgroup
, &mgctx
);
2628 cgroup_migrate_finish(&mgctx
);
2631 trace_cgroup_attach_task(dst_cgrp
, leader
, threadgroup
);
2636 struct task_struct
*cgroup_procs_write_start(char *buf
, bool threadgroup
)
2637 __acquires(&cgroup_threadgroup_rwsem
)
2639 struct task_struct
*tsk
;
2642 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2643 return ERR_PTR(-EINVAL
);
2645 percpu_down_write(&cgroup_threadgroup_rwsem
);
2649 tsk
= find_task_by_vpid(pid
);
2651 tsk
= ERR_PTR(-ESRCH
);
2652 goto out_unlock_threadgroup
;
2659 tsk
= tsk
->group_leader
;
2662 * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2663 * If userland migrates such a kthread to a non-root cgroup, it can
2664 * become trapped in a cpuset, or RT kthread may be born in a
2665 * cgroup with no rt_runtime allocated. Just say no.
2667 if (tsk
->no_cgroup_migration
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2668 tsk
= ERR_PTR(-EINVAL
);
2669 goto out_unlock_threadgroup
;
2672 get_task_struct(tsk
);
2673 goto out_unlock_rcu
;
2675 out_unlock_threadgroup
:
2676 percpu_up_write(&cgroup_threadgroup_rwsem
);
2682 void cgroup_procs_write_finish(struct task_struct
*task
)
2683 __releases(&cgroup_threadgroup_rwsem
)
2685 struct cgroup_subsys
*ss
;
2688 /* release reference from cgroup_procs_write_start() */
2689 put_task_struct(task
);
2691 percpu_up_write(&cgroup_threadgroup_rwsem
);
2692 for_each_subsys(ss
, ssid
)
2693 if (ss
->post_attach
)
2697 static void cgroup_print_ss_mask(struct seq_file
*seq
, u16 ss_mask
)
2699 struct cgroup_subsys
*ss
;
2700 bool printed
= false;
2703 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
2706 seq_printf(seq
, "%s", ss
->name
);
2708 } while_each_subsys_mask();
2710 seq_putc(seq
, '\n');
2713 /* show controllers which are enabled from the parent */
2714 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2716 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2718 cgroup_print_ss_mask(seq
, cgroup_control(cgrp
));
2722 /* show controllers which are enabled for a given cgroup's children */
2723 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2725 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2727 cgroup_print_ss_mask(seq
, cgrp
->subtree_control
);
2732 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2733 * @cgrp: root of the subtree to update csses for
2735 * @cgrp's control masks have changed and its subtree's css associations
2736 * need to be updated accordingly. This function looks up all css_sets
2737 * which are attached to the subtree, creates the matching updated css_sets
2738 * and migrates the tasks to the new ones.
2740 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2742 DEFINE_CGROUP_MGCTX(mgctx
);
2743 struct cgroup_subsys_state
*d_css
;
2744 struct cgroup
*dsct
;
2745 struct css_set
*src_cset
;
2748 lockdep_assert_held(&cgroup_mutex
);
2750 percpu_down_write(&cgroup_threadgroup_rwsem
);
2752 /* look up all csses currently attached to @cgrp's subtree */
2753 spin_lock_irq(&css_set_lock
);
2754 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
2755 struct cgrp_cset_link
*link
;
2757 list_for_each_entry(link
, &dsct
->cset_links
, cset_link
)
2758 cgroup_migrate_add_src(link
->cset
, dsct
, &mgctx
);
2760 spin_unlock_irq(&css_set_lock
);
2762 /* NULL dst indicates self on default hierarchy */
2763 ret
= cgroup_migrate_prepare_dst(&mgctx
);
2767 spin_lock_irq(&css_set_lock
);
2768 list_for_each_entry(src_cset
, &mgctx
.preloaded_src_csets
, mg_preload_node
) {
2769 struct task_struct
*task
, *ntask
;
2771 /* all tasks in src_csets need to be migrated */
2772 list_for_each_entry_safe(task
, ntask
, &src_cset
->tasks
, cg_list
)
2773 cgroup_migrate_add_task(task
, &mgctx
);
2775 spin_unlock_irq(&css_set_lock
);
2777 ret
= cgroup_migrate_execute(&mgctx
);
2779 cgroup_migrate_finish(&mgctx
);
2780 percpu_up_write(&cgroup_threadgroup_rwsem
);
2785 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2786 * @cgrp: root of the target subtree
2788 * Because css offlining is asynchronous, userland may try to re-enable a
2789 * controller while the previous css is still around. This function grabs
2790 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2792 void cgroup_lock_and_drain_offline(struct cgroup
*cgrp
)
2793 __acquires(&cgroup_mutex
)
2795 struct cgroup
*dsct
;
2796 struct cgroup_subsys_state
*d_css
;
2797 struct cgroup_subsys
*ss
;
2801 mutex_lock(&cgroup_mutex
);
2803 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
2804 for_each_subsys(ss
, ssid
) {
2805 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
2808 if (!css
|| !percpu_ref_is_dying(&css
->refcnt
))
2811 cgroup_get_live(dsct
);
2812 prepare_to_wait(&dsct
->offline_waitq
, &wait
,
2813 TASK_UNINTERRUPTIBLE
);
2815 mutex_unlock(&cgroup_mutex
);
2817 finish_wait(&dsct
->offline_waitq
, &wait
);
2826 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2827 * @cgrp: root of the target subtree
2829 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2830 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2833 static void cgroup_save_control(struct cgroup
*cgrp
)
2835 struct cgroup
*dsct
;
2836 struct cgroup_subsys_state
*d_css
;
2838 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
2839 dsct
->old_subtree_control
= dsct
->subtree_control
;
2840 dsct
->old_subtree_ss_mask
= dsct
->subtree_ss_mask
;
2841 dsct
->old_dom_cgrp
= dsct
->dom_cgrp
;
2846 * cgroup_propagate_control - refresh control masks of a subtree
2847 * @cgrp: root of the target subtree
2849 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2850 * ->subtree_control and propagate controller availability through the
2851 * subtree so that descendants don't have unavailable controllers enabled.
2853 static void cgroup_propagate_control(struct cgroup
*cgrp
)
2855 struct cgroup
*dsct
;
2856 struct cgroup_subsys_state
*d_css
;
2858 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
2859 dsct
->subtree_control
&= cgroup_control(dsct
);
2860 dsct
->subtree_ss_mask
=
2861 cgroup_calc_subtree_ss_mask(dsct
->subtree_control
,
2862 cgroup_ss_mask(dsct
));
2867 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
2868 * @cgrp: root of the target subtree
2870 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
2871 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2874 static void cgroup_restore_control(struct cgroup
*cgrp
)
2876 struct cgroup
*dsct
;
2877 struct cgroup_subsys_state
*d_css
;
2879 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
2880 dsct
->subtree_control
= dsct
->old_subtree_control
;
2881 dsct
->subtree_ss_mask
= dsct
->old_subtree_ss_mask
;
2882 dsct
->dom_cgrp
= dsct
->old_dom_cgrp
;
2886 static bool css_visible(struct cgroup_subsys_state
*css
)
2888 struct cgroup_subsys
*ss
= css
->ss
;
2889 struct cgroup
*cgrp
= css
->cgroup
;
2891 if (cgroup_control(cgrp
) & (1 << ss
->id
))
2893 if (!(cgroup_ss_mask(cgrp
) & (1 << ss
->id
)))
2895 return cgroup_on_dfl(cgrp
) && ss
->implicit_on_dfl
;
2899 * cgroup_apply_control_enable - enable or show csses according to control
2900 * @cgrp: root of the target subtree
2902 * Walk @cgrp's subtree and create new csses or make the existing ones
2903 * visible. A css is created invisible if it's being implicitly enabled
2904 * through dependency. An invisible css is made visible when the userland
2905 * explicitly enables it.
2907 * Returns 0 on success, -errno on failure. On failure, csses which have
2908 * been processed already aren't cleaned up. The caller is responsible for
2909 * cleaning up with cgroup_apply_control_disable().
2911 static int cgroup_apply_control_enable(struct cgroup
*cgrp
)
2913 struct cgroup
*dsct
;
2914 struct cgroup_subsys_state
*d_css
;
2915 struct cgroup_subsys
*ss
;
2918 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
) {
2919 for_each_subsys(ss
, ssid
) {
2920 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
2922 WARN_ON_ONCE(css
&& percpu_ref_is_dying(&css
->refcnt
));
2924 if (!(cgroup_ss_mask(dsct
) & (1 << ss
->id
)))
2928 css
= css_create(dsct
, ss
);
2930 return PTR_ERR(css
);
2933 if (css_visible(css
)) {
2934 ret
= css_populate_dir(css
);
2945 * cgroup_apply_control_disable - kill or hide csses according to control
2946 * @cgrp: root of the target subtree
2948 * Walk @cgrp's subtree and kill and hide csses so that they match
2949 * cgroup_ss_mask() and cgroup_visible_mask().
2951 * A css is hidden when the userland requests it to be disabled while other
2952 * subsystems are still depending on it. The css must not actively control
2953 * resources and be in the vanilla state if it's made visible again later.
2954 * Controllers which may be depended upon should provide ->css_reset() for
2957 static void cgroup_apply_control_disable(struct cgroup
*cgrp
)
2959 struct cgroup
*dsct
;
2960 struct cgroup_subsys_state
*d_css
;
2961 struct cgroup_subsys
*ss
;
2964 cgroup_for_each_live_descendant_post(dsct
, d_css
, cgrp
) {
2965 for_each_subsys(ss
, ssid
) {
2966 struct cgroup_subsys_state
*css
= cgroup_css(dsct
, ss
);
2968 WARN_ON_ONCE(css
&& percpu_ref_is_dying(&css
->refcnt
));
2974 !(cgroup_ss_mask(dsct
) & (1 << ss
->id
))) {
2976 } else if (!css_visible(css
)) {
2986 * cgroup_apply_control - apply control mask updates to the subtree
2987 * @cgrp: root of the target subtree
2989 * subsystems can be enabled and disabled in a subtree using the following
2992 * 1. Call cgroup_save_control() to stash the current state.
2993 * 2. Update ->subtree_control masks in the subtree as desired.
2994 * 3. Call cgroup_apply_control() to apply the changes.
2995 * 4. Optionally perform other related operations.
2996 * 5. Call cgroup_finalize_control() to finish up.
2998 * This function implements step 3 and propagates the mask changes
2999 * throughout @cgrp's subtree, updates csses accordingly and perform
3000 * process migrations.
3002 static int cgroup_apply_control(struct cgroup
*cgrp
)
3006 cgroup_propagate_control(cgrp
);
3008 ret
= cgroup_apply_control_enable(cgrp
);
3013 * At this point, cgroup_e_css() results reflect the new csses
3014 * making the following cgroup_update_dfl_csses() properly update
3015 * css associations of all tasks in the subtree.
3017 ret
= cgroup_update_dfl_csses(cgrp
);
3025 * cgroup_finalize_control - finalize control mask update
3026 * @cgrp: root of the target subtree
3027 * @ret: the result of the update
3029 * Finalize control mask update. See cgroup_apply_control() for more info.
3031 static void cgroup_finalize_control(struct cgroup
*cgrp
, int ret
)
3034 cgroup_restore_control(cgrp
);
3035 cgroup_propagate_control(cgrp
);
3038 cgroup_apply_control_disable(cgrp
);
3041 static int cgroup_vet_subtree_control_enable(struct cgroup
*cgrp
, u16 enable
)
3043 u16 domain_enable
= enable
& ~cgrp_dfl_threaded_ss_mask
;
3045 /* if nothing is getting enabled, nothing to worry about */
3049 /* can @cgrp host any resources? */
3050 if (!cgroup_is_valid_domain(cgrp
->dom_cgrp
))
3053 /* mixables don't care */
3054 if (cgroup_is_mixable(cgrp
))
3057 if (domain_enable
) {
3058 /* can't enable domain controllers inside a thread subtree */
3059 if (cgroup_is_thread_root(cgrp
) || cgroup_is_threaded(cgrp
))
3063 * Threaded controllers can handle internal competitions
3064 * and are always allowed inside a (prospective) thread
3067 if (cgroup_can_be_thread_root(cgrp
) || cgroup_is_threaded(cgrp
))
3072 * Controllers can't be enabled for a cgroup with tasks to avoid
3073 * child cgroups competing against tasks.
3075 if (cgroup_has_tasks(cgrp
))
3081 /* change the enabled child controllers for a cgroup in the default hierarchy */
3082 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
3083 char *buf
, size_t nbytes
,
3086 u16 enable
= 0, disable
= 0;
3087 struct cgroup
*cgrp
, *child
;
3088 struct cgroup_subsys
*ss
;
3093 * Parse input - space separated list of subsystem names prefixed
3094 * with either + or -.
3096 buf
= strstrip(buf
);
3097 while ((tok
= strsep(&buf
, " "))) {
3100 do_each_subsys_mask(ss
, ssid
, ~cgrp_dfl_inhibit_ss_mask
) {
3101 if (!cgroup_ssid_enabled(ssid
) ||
3102 strcmp(tok
+ 1, ss
->name
))
3106 enable
|= 1 << ssid
;
3107 disable
&= ~(1 << ssid
);
3108 } else if (*tok
== '-') {
3109 disable
|= 1 << ssid
;
3110 enable
&= ~(1 << ssid
);
3115 } while_each_subsys_mask();
3116 if (ssid
== CGROUP_SUBSYS_COUNT
)
3120 cgrp
= cgroup_kn_lock_live(of
->kn
, true);
3124 for_each_subsys(ss
, ssid
) {
3125 if (enable
& (1 << ssid
)) {
3126 if (cgrp
->subtree_control
& (1 << ssid
)) {
3127 enable
&= ~(1 << ssid
);
3131 if (!(cgroup_control(cgrp
) & (1 << ssid
))) {
3135 } else if (disable
& (1 << ssid
)) {
3136 if (!(cgrp
->subtree_control
& (1 << ssid
))) {
3137 disable
&= ~(1 << ssid
);
3141 /* a child has it enabled? */
3142 cgroup_for_each_live_child(child
, cgrp
) {
3143 if (child
->subtree_control
& (1 << ssid
)) {
3151 if (!enable
&& !disable
) {
3156 ret
= cgroup_vet_subtree_control_enable(cgrp
, enable
);
3160 /* save and update control masks and prepare csses */
3161 cgroup_save_control(cgrp
);
3163 cgrp
->subtree_control
|= enable
;
3164 cgrp
->subtree_control
&= ~disable
;
3166 ret
= cgroup_apply_control(cgrp
);
3167 cgroup_finalize_control(cgrp
, ret
);
3171 kernfs_activate(cgrp
->kn
);
3173 cgroup_kn_unlock(of
->kn
);
3174 return ret
?: nbytes
;
3178 * cgroup_enable_threaded - make @cgrp threaded
3179 * @cgrp: the target cgroup
3181 * Called when "threaded" is written to the cgroup.type interface file and
3182 * tries to make @cgrp threaded and join the parent's resource domain.
3183 * This function is never called on the root cgroup as cgroup.type doesn't
3186 static int cgroup_enable_threaded(struct cgroup
*cgrp
)
3188 struct cgroup
*parent
= cgroup_parent(cgrp
);
3189 struct cgroup
*dom_cgrp
= parent
->dom_cgrp
;
3190 struct cgroup
*dsct
;
3191 struct cgroup_subsys_state
*d_css
;
3194 lockdep_assert_held(&cgroup_mutex
);
3196 /* noop if already threaded */
3197 if (cgroup_is_threaded(cgrp
))
3201 * If @cgroup is populated or has domain controllers enabled, it
3202 * can't be switched. While the below cgroup_can_be_thread_root()
3203 * test can catch the same conditions, that's only when @parent is
3204 * not mixable, so let's check it explicitly.
3206 if (cgroup_is_populated(cgrp
) ||
3207 cgrp
->subtree_control
& ~cgrp_dfl_threaded_ss_mask
)
3210 /* we're joining the parent's domain, ensure its validity */
3211 if (!cgroup_is_valid_domain(dom_cgrp
) ||
3212 !cgroup_can_be_thread_root(dom_cgrp
))
3216 * The following shouldn't cause actual migrations and should
3219 cgroup_save_control(cgrp
);
3221 cgroup_for_each_live_descendant_pre(dsct
, d_css
, cgrp
)
3222 if (dsct
== cgrp
|| cgroup_is_threaded(dsct
))
3223 dsct
->dom_cgrp
= dom_cgrp
;
3225 ret
= cgroup_apply_control(cgrp
);
3227 parent
->nr_threaded_children
++;
3229 cgroup_finalize_control(cgrp
, ret
);
3233 static int cgroup_type_show(struct seq_file
*seq
, void *v
)
3235 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3237 if (cgroup_is_threaded(cgrp
))
3238 seq_puts(seq
, "threaded\n");
3239 else if (!cgroup_is_valid_domain(cgrp
))
3240 seq_puts(seq
, "domain invalid\n");
3241 else if (cgroup_is_thread_root(cgrp
))
3242 seq_puts(seq
, "domain threaded\n");
3244 seq_puts(seq
, "domain\n");
3249 static ssize_t
cgroup_type_write(struct kernfs_open_file
*of
, char *buf
,
3250 size_t nbytes
, loff_t off
)
3252 struct cgroup
*cgrp
;
3255 /* only switching to threaded mode is supported */
3256 if (strcmp(strstrip(buf
), "threaded"))
3259 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3263 /* threaded can only be enabled */
3264 ret
= cgroup_enable_threaded(cgrp
);
3266 cgroup_kn_unlock(of
->kn
);
3267 return ret
?: nbytes
;
3270 static int cgroup_max_descendants_show(struct seq_file
*seq
, void *v
)
3272 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3273 int descendants
= READ_ONCE(cgrp
->max_descendants
);
3275 if (descendants
== INT_MAX
)
3276 seq_puts(seq
, "max\n");
3278 seq_printf(seq
, "%d\n", descendants
);
3283 static ssize_t
cgroup_max_descendants_write(struct kernfs_open_file
*of
,
3284 char *buf
, size_t nbytes
, loff_t off
)
3286 struct cgroup
*cgrp
;
3290 buf
= strstrip(buf
);
3291 if (!strcmp(buf
, "max")) {
3292 descendants
= INT_MAX
;
3294 ret
= kstrtoint(buf
, 0, &descendants
);
3299 if (descendants
< 0)
3302 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3306 cgrp
->max_descendants
= descendants
;
3308 cgroup_kn_unlock(of
->kn
);
3313 static int cgroup_max_depth_show(struct seq_file
*seq
, void *v
)
3315 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
3316 int depth
= READ_ONCE(cgrp
->max_depth
);
3318 if (depth
== INT_MAX
)
3319 seq_puts(seq
, "max\n");
3321 seq_printf(seq
, "%d\n", depth
);
3326 static ssize_t
cgroup_max_depth_write(struct kernfs_open_file
*of
,
3327 char *buf
, size_t nbytes
, loff_t off
)
3329 struct cgroup
*cgrp
;
3333 buf
= strstrip(buf
);
3334 if (!strcmp(buf
, "max")) {
3337 ret
= kstrtoint(buf
, 0, &depth
);
3345 cgrp
= cgroup_kn_lock_live(of
->kn
, false);
3349 cgrp
->max_depth
= depth
;
3351 cgroup_kn_unlock(of
->kn
);
3356 static int cgroup_events_show(struct seq_file
*seq
, void *v
)
3358 seq_printf(seq
, "populated %d\n",
3359 cgroup_is_populated(seq_css(seq
)->cgroup
));
3363 static int cgroup_stat_show(struct seq_file
*seq
, void *v
)
3365 struct cgroup
*cgroup
= seq_css(seq
)->cgroup
;
3367 seq_printf(seq
, "nr_descendants %d\n",
3368 cgroup
->nr_descendants
);
3369 seq_printf(seq
, "nr_dying_descendants %d\n",
3370 cgroup
->nr_dying_descendants
);
3375 static int __maybe_unused
cgroup_extra_stat_show(struct seq_file
*seq
,
3376 struct cgroup
*cgrp
, int ssid
)
3378 struct cgroup_subsys
*ss
= cgroup_subsys
[ssid
];
3379 struct cgroup_subsys_state
*css
;
3382 if (!ss
->css_extra_stat_show
)
3385 css
= cgroup_tryget_css(cgrp
, ss
);
3389 ret
= ss
->css_extra_stat_show(seq
, css
);
3394 static int cpu_stat_show(struct seq_file
*seq
, void *v
)
3396 struct cgroup __maybe_unused
*cgrp
= seq_css(seq
)->cgroup
;
3399 cgroup_stat_show_cputime(seq
);
3400 #ifdef CONFIG_CGROUP_SCHED
3401 ret
= cgroup_extra_stat_show(seq
, cgrp
, cpu_cgrp_id
);
3406 static int cgroup_file_open(struct kernfs_open_file
*of
)
3408 struct cftype
*cft
= of
->kn
->priv
;
3411 return cft
->open(of
);
3415 static void cgroup_file_release(struct kernfs_open_file
*of
)
3417 struct cftype
*cft
= of
->kn
->priv
;
3423 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
3424 size_t nbytes
, loff_t off
)
3426 struct cgroup_namespace
*ns
= current
->nsproxy
->cgroup_ns
;
3427 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
3428 struct cftype
*cft
= of
->kn
->priv
;
3429 struct cgroup_subsys_state
*css
;
3433 * If namespaces are delegation boundaries, disallow writes to
3434 * files in an non-init namespace root from inside the namespace
3435 * except for the files explicitly marked delegatable -
3436 * cgroup.procs and cgroup.subtree_control.
3438 if ((cgrp
->root
->flags
& CGRP_ROOT_NS_DELEGATE
) &&
3439 !(cft
->flags
& CFTYPE_NS_DELEGATABLE
) &&
3440 ns
!= &init_cgroup_ns
&& ns
->root_cset
->dfl_cgrp
== cgrp
)
3444 return cft
->write(of
, buf
, nbytes
, off
);
3447 * kernfs guarantees that a file isn't deleted with operations in
3448 * flight, which means that the matching css is and stays alive and
3449 * doesn't need to be pinned. The RCU locking is not necessary
3450 * either. It's just for the convenience of using cgroup_css().
3453 css
= cgroup_css(cgrp
, cft
->ss
);
3456 if (cft
->write_u64
) {
3457 unsigned long long v
;
3458 ret
= kstrtoull(buf
, 0, &v
);
3460 ret
= cft
->write_u64(css
, cft
, v
);
3461 } else if (cft
->write_s64
) {
3463 ret
= kstrtoll(buf
, 0, &v
);
3465 ret
= cft
->write_s64(css
, cft
, v
);
3470 return ret
?: nbytes
;
3473 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
3475 return seq_cft(seq
)->seq_start(seq
, ppos
);
3478 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
3480 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
3483 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
3485 if (seq_cft(seq
)->seq_stop
)
3486 seq_cft(seq
)->seq_stop(seq
, v
);
3489 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
3491 struct cftype
*cft
= seq_cft(m
);
3492 struct cgroup_subsys_state
*css
= seq_css(m
);
3495 return cft
->seq_show(m
, arg
);
3498 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
3499 else if (cft
->read_s64
)
3500 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
3506 static struct kernfs_ops cgroup_kf_single_ops
= {
3507 .atomic_write_len
= PAGE_SIZE
,
3508 .open
= cgroup_file_open
,
3509 .release
= cgroup_file_release
,
3510 .write
= cgroup_file_write
,
3511 .seq_show
= cgroup_seqfile_show
,
3514 static struct kernfs_ops cgroup_kf_ops
= {
3515 .atomic_write_len
= PAGE_SIZE
,
3516 .open
= cgroup_file_open
,
3517 .release
= cgroup_file_release
,
3518 .write
= cgroup_file_write
,
3519 .seq_start
= cgroup_seqfile_start
,
3520 .seq_next
= cgroup_seqfile_next
,
3521 .seq_stop
= cgroup_seqfile_stop
,
3522 .seq_show
= cgroup_seqfile_show
,
3525 /* set uid and gid of cgroup dirs and files to that of the creator */
3526 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
3528 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
3529 .ia_uid
= current_fsuid(),
3530 .ia_gid
= current_fsgid(), };
3532 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
3533 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
3536 return kernfs_setattr(kn
, &iattr
);
3539 static int cgroup_add_file(struct cgroup_subsys_state
*css
, struct cgroup
*cgrp
,
3542 char name
[CGROUP_FILE_NAME_MAX
];
3543 struct kernfs_node
*kn
;
3544 struct lock_class_key
*key
= NULL
;
3547 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3548 key
= &cft
->lockdep_key
;
3550 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
3551 cgroup_file_mode(cft
),
3552 GLOBAL_ROOT_UID
, GLOBAL_ROOT_GID
,
3553 0, cft
->kf_ops
, cft
,
3558 ret
= cgroup_kn_set_ugid(kn
);
3564 if (cft
->file_offset
) {
3565 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
3567 spin_lock_irq(&cgroup_file_kn_lock
);
3569 spin_unlock_irq(&cgroup_file_kn_lock
);
3576 * cgroup_addrm_files - add or remove files to a cgroup directory
3577 * @css: the target css
3578 * @cgrp: the target cgroup (usually css->cgroup)
3579 * @cfts: array of cftypes to be added
3580 * @is_add: whether to add or remove
3582 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3583 * For removals, this function never fails.
3585 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
3586 struct cgroup
*cgrp
, struct cftype cfts
[],
3589 struct cftype
*cft
, *cft_end
= NULL
;
3592 lockdep_assert_held(&cgroup_mutex
);
3595 for (cft
= cfts
; cft
!= cft_end
&& cft
->name
[0] != '\0'; cft
++) {
3596 /* does cft->flags tell us to skip this file on @cgrp? */
3597 if ((cft
->flags
& __CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
3599 if ((cft
->flags
& __CFTYPE_NOT_ON_DFL
) && cgroup_on_dfl(cgrp
))
3601 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
3603 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
3607 ret
= cgroup_add_file(css
, cgrp
, cft
);
3609 pr_warn("%s: failed to add %s, err=%d\n",
3610 __func__
, cft
->name
, ret
);
3616 cgroup_rm_file(cgrp
, cft
);
3622 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
3624 struct cgroup_subsys
*ss
= cfts
[0].ss
;
3625 struct cgroup
*root
= &ss
->root
->cgrp
;
3626 struct cgroup_subsys_state
*css
;
3629 lockdep_assert_held(&cgroup_mutex
);
3631 /* add/rm files for all cgroups created before */
3632 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
3633 struct cgroup
*cgrp
= css
->cgroup
;
3635 if (!(css
->flags
& CSS_VISIBLE
))
3638 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, is_add
);
3644 kernfs_activate(root
->kn
);
3648 static void cgroup_exit_cftypes(struct cftype
*cfts
)
3652 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3653 /* free copy for custom atomic_write_len, see init_cftypes() */
3654 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
3659 /* revert flags set by cgroup core while adding @cfts */
3660 cft
->flags
&= ~(__CFTYPE_ONLY_ON_DFL
| __CFTYPE_NOT_ON_DFL
);
3664 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3668 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3669 struct kernfs_ops
*kf_ops
;
3671 WARN_ON(cft
->ss
|| cft
->kf_ops
);
3674 kf_ops
= &cgroup_kf_ops
;
3676 kf_ops
= &cgroup_kf_single_ops
;
3679 * Ugh... if @cft wants a custom max_write_len, we need to
3680 * make a copy of kf_ops to set its atomic_write_len.
3682 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
3683 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
3685 cgroup_exit_cftypes(cfts
);
3688 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3691 cft
->kf_ops
= kf_ops
;
3698 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3700 lockdep_assert_held(&cgroup_mutex
);
3702 if (!cfts
|| !cfts
[0].ss
)
3705 list_del(&cfts
->node
);
3706 cgroup_apply_cftypes(cfts
, false);
3707 cgroup_exit_cftypes(cfts
);
3712 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3713 * @cfts: zero-length name terminated array of cftypes
3715 * Unregister @cfts. Files described by @cfts are removed from all
3716 * existing cgroups and all future cgroups won't have them either. This
3717 * function can be called anytime whether @cfts' subsys is attached or not.
3719 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3722 int cgroup_rm_cftypes(struct cftype
*cfts
)
3726 mutex_lock(&cgroup_mutex
);
3727 ret
= cgroup_rm_cftypes_locked(cfts
);
3728 mutex_unlock(&cgroup_mutex
);
3733 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3734 * @ss: target cgroup subsystem
3735 * @cfts: zero-length name terminated array of cftypes
3737 * Register @cfts to @ss. Files described by @cfts are created for all
3738 * existing cgroups to which @ss is attached and all future cgroups will
3739 * have them too. This function can be called anytime whether @ss is
3742 * Returns 0 on successful registration, -errno on failure. Note that this
3743 * function currently returns 0 as long as @cfts registration is successful
3744 * even if some file creation attempts on existing cgroups fail.
3746 static int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3750 if (!cgroup_ssid_enabled(ss
->id
))
3753 if (!cfts
|| cfts
[0].name
[0] == '\0')
3756 ret
= cgroup_init_cftypes(ss
, cfts
);
3760 mutex_lock(&cgroup_mutex
);
3762 list_add_tail(&cfts
->node
, &ss
->cfts
);
3763 ret
= cgroup_apply_cftypes(cfts
, true);
3765 cgroup_rm_cftypes_locked(cfts
);
3767 mutex_unlock(&cgroup_mutex
);
3772 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3773 * @ss: target cgroup subsystem
3774 * @cfts: zero-length name terminated array of cftypes
3776 * Similar to cgroup_add_cftypes() but the added files are only used for
3777 * the default hierarchy.
3779 int cgroup_add_dfl_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3783 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3784 cft
->flags
|= __CFTYPE_ONLY_ON_DFL
;
3785 return cgroup_add_cftypes(ss
, cfts
);
3789 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3790 * @ss: target cgroup subsystem
3791 * @cfts: zero-length name terminated array of cftypes
3793 * Similar to cgroup_add_cftypes() but the added files are only used for
3794 * the legacy hierarchies.
3796 int cgroup_add_legacy_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3800 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3801 cft
->flags
|= __CFTYPE_NOT_ON_DFL
;
3802 return cgroup_add_cftypes(ss
, cfts
);
3806 * cgroup_file_notify - generate a file modified event for a cgroup_file
3807 * @cfile: target cgroup_file
3809 * @cfile must have been obtained by setting cftype->file_offset.
3811 void cgroup_file_notify(struct cgroup_file
*cfile
)
3813 unsigned long flags
;
3815 spin_lock_irqsave(&cgroup_file_kn_lock
, flags
);
3817 kernfs_notify(cfile
->kn
);
3818 spin_unlock_irqrestore(&cgroup_file_kn_lock
, flags
);
3822 * css_next_child - find the next child of a given css
3823 * @pos: the current position (%NULL to initiate traversal)
3824 * @parent: css whose children to walk
3826 * This function returns the next child of @parent and should be called
3827 * under either cgroup_mutex or RCU read lock. The only requirement is
3828 * that @parent and @pos are accessible. The next sibling is guaranteed to
3829 * be returned regardless of their states.
3831 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3832 * css which finished ->css_online() is guaranteed to be visible in the
3833 * future iterations and will stay visible until the last reference is put.
3834 * A css which hasn't finished ->css_online() or already finished
3835 * ->css_offline() may show up during traversal. It's each subsystem's
3836 * responsibility to synchronize against on/offlining.
3838 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
3839 struct cgroup_subsys_state
*parent
)
3841 struct cgroup_subsys_state
*next
;
3843 cgroup_assert_mutex_or_rcu_locked();
3846 * @pos could already have been unlinked from the sibling list.
3847 * Once a cgroup is removed, its ->sibling.next is no longer
3848 * updated when its next sibling changes. CSS_RELEASED is set when
3849 * @pos is taken off list, at which time its next pointer is valid,
3850 * and, as releases are serialized, the one pointed to by the next
3851 * pointer is guaranteed to not have started release yet. This
3852 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3853 * critical section, the one pointed to by its next pointer is
3854 * guaranteed to not have finished its RCU grace period even if we
3855 * have dropped rcu_read_lock() inbetween iterations.
3857 * If @pos has CSS_RELEASED set, its next pointer can't be
3858 * dereferenced; however, as each css is given a monotonically
3859 * increasing unique serial number and always appended to the
3860 * sibling list, the next one can be found by walking the parent's
3861 * children until the first css with higher serial number than
3862 * @pos's. While this path can be slower, it happens iff iteration
3863 * races against release and the race window is very small.
3866 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
3867 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
3868 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
3870 list_for_each_entry_rcu(next
, &parent
->children
, sibling
)
3871 if (next
->serial_nr
> pos
->serial_nr
)
3876 * @next, if not pointing to the head, can be dereferenced and is
3879 if (&next
->sibling
!= &parent
->children
)
3885 * css_next_descendant_pre - find the next descendant for pre-order walk
3886 * @pos: the current position (%NULL to initiate traversal)
3887 * @root: css whose descendants to walk
3889 * To be used by css_for_each_descendant_pre(). Find the next descendant
3890 * to visit for pre-order traversal of @root's descendants. @root is
3891 * included in the iteration and the first node to be visited.
3893 * While this function requires cgroup_mutex or RCU read locking, it
3894 * doesn't require the whole traversal to be contained in a single critical
3895 * section. This function will return the correct next descendant as long
3896 * as both @pos and @root are accessible and @pos is a descendant of @root.
3898 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3899 * css which finished ->css_online() is guaranteed to be visible in the
3900 * future iterations and will stay visible until the last reference is put.
3901 * A css which hasn't finished ->css_online() or already finished
3902 * ->css_offline() may show up during traversal. It's each subsystem's
3903 * responsibility to synchronize against on/offlining.
3905 struct cgroup_subsys_state
*
3906 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3907 struct cgroup_subsys_state
*root
)
3909 struct cgroup_subsys_state
*next
;
3911 cgroup_assert_mutex_or_rcu_locked();
3913 /* if first iteration, visit @root */
3917 /* visit the first child if exists */
3918 next
= css_next_child(NULL
, pos
);
3922 /* no child, visit my or the closest ancestor's next sibling */
3923 while (pos
!= root
) {
3924 next
= css_next_child(pos
, pos
->parent
);
3934 * css_rightmost_descendant - return the rightmost descendant of a css
3935 * @pos: css of interest
3937 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3938 * is returned. This can be used during pre-order traversal to skip
3941 * While this function requires cgroup_mutex or RCU read locking, it
3942 * doesn't require the whole traversal to be contained in a single critical
3943 * section. This function will return the correct rightmost descendant as
3944 * long as @pos is accessible.
3946 struct cgroup_subsys_state
*
3947 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3949 struct cgroup_subsys_state
*last
, *tmp
;
3951 cgroup_assert_mutex_or_rcu_locked();
3955 /* ->prev isn't RCU safe, walk ->next till the end */
3957 css_for_each_child(tmp
, last
)
3964 static struct cgroup_subsys_state
*
3965 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3967 struct cgroup_subsys_state
*last
;
3971 pos
= css_next_child(NULL
, pos
);
3978 * css_next_descendant_post - find the next descendant for post-order walk
3979 * @pos: the current position (%NULL to initiate traversal)
3980 * @root: css whose descendants to walk
3982 * To be used by css_for_each_descendant_post(). Find the next descendant
3983 * to visit for post-order traversal of @root's descendants. @root is
3984 * included in the iteration and the last node to be visited.
3986 * While this function requires cgroup_mutex or RCU read locking, it
3987 * doesn't require the whole traversal to be contained in a single critical
3988 * section. This function will return the correct next descendant as long
3989 * as both @pos and @cgroup are accessible and @pos is a descendant of
3992 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3993 * css which finished ->css_online() is guaranteed to be visible in the
3994 * future iterations and will stay visible until the last reference is put.
3995 * A css which hasn't finished ->css_online() or already finished
3996 * ->css_offline() may show up during traversal. It's each subsystem's
3997 * responsibility to synchronize against on/offlining.
3999 struct cgroup_subsys_state
*
4000 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
4001 struct cgroup_subsys_state
*root
)
4003 struct cgroup_subsys_state
*next
;
4005 cgroup_assert_mutex_or_rcu_locked();
4007 /* if first iteration, visit leftmost descendant which may be @root */
4009 return css_leftmost_descendant(root
);
4011 /* if we visited @root, we're done */
4015 /* if there's an unvisited sibling, visit its leftmost descendant */
4016 next
= css_next_child(pos
, pos
->parent
);
4018 return css_leftmost_descendant(next
);
4020 /* no sibling left, visit parent */
4025 * css_has_online_children - does a css have online children
4026 * @css: the target css
4028 * Returns %true if @css has any online children; otherwise, %false. This
4029 * function can be called from any context but the caller is responsible
4030 * for synchronizing against on/offlining as necessary.
4032 bool css_has_online_children(struct cgroup_subsys_state
*css
)
4034 struct cgroup_subsys_state
*child
;
4038 css_for_each_child(child
, css
) {
4039 if (child
->flags
& CSS_ONLINE
) {
4048 static struct css_set
*css_task_iter_next_css_set(struct css_task_iter
*it
)
4050 struct list_head
*l
;
4051 struct cgrp_cset_link
*link
;
4052 struct css_set
*cset
;
4054 lockdep_assert_held(&css_set_lock
);
4056 /* find the next threaded cset */
4057 if (it
->tcset_pos
) {
4058 l
= it
->tcset_pos
->next
;
4060 if (l
!= it
->tcset_head
) {
4062 return container_of(l
, struct css_set
,
4063 threaded_csets_node
);
4066 it
->tcset_pos
= NULL
;
4069 /* find the next cset */
4072 if (l
== it
->cset_head
) {
4073 it
->cset_pos
= NULL
;
4078 cset
= container_of(l
, struct css_set
, e_cset_node
[it
->ss
->id
]);
4080 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
4086 /* initialize threaded css_set walking */
4087 if (it
->flags
& CSS_TASK_ITER_THREADED
) {
4089 put_css_set_locked(it
->cur_dcset
);
4090 it
->cur_dcset
= cset
;
4093 it
->tcset_head
= &cset
->threaded_csets
;
4094 it
->tcset_pos
= &cset
->threaded_csets
;
4101 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4102 * @it: the iterator to advance
4104 * Advance @it to the next css_set to walk.
4106 static void css_task_iter_advance_css_set(struct css_task_iter
*it
)
4108 struct css_set
*cset
;
4110 lockdep_assert_held(&css_set_lock
);
4112 /* Advance to the next non-empty css_set */
4114 cset
= css_task_iter_next_css_set(it
);
4116 it
->task_pos
= NULL
;
4119 } while (!css_set_populated(cset
) && !list_empty(&cset
->dying_tasks
));
4121 if (!list_empty(&cset
->tasks
))
4122 it
->task_pos
= cset
->tasks
.next
;
4123 else if (!list_empty(&cset
->mg_tasks
))
4124 it
->task_pos
= cset
->mg_tasks
.next
;
4126 it
->task_pos
= cset
->dying_tasks
.next
;
4128 it
->tasks_head
= &cset
->tasks
;
4129 it
->mg_tasks_head
= &cset
->mg_tasks
;
4130 it
->dying_tasks_head
= &cset
->dying_tasks
;
4133 * We don't keep css_sets locked across iteration steps and thus
4134 * need to take steps to ensure that iteration can be resumed after
4135 * the lock is re-acquired. Iteration is performed at two levels -
4136 * css_sets and tasks in them.
4138 * Once created, a css_set never leaves its cgroup lists, so a
4139 * pinned css_set is guaranteed to stay put and we can resume
4140 * iteration afterwards.
4142 * Tasks may leave @cset across iteration steps. This is resolved
4143 * by registering each iterator with the css_set currently being
4144 * walked and making css_set_move_task() advance iterators whose
4145 * next task is leaving.
4148 list_del(&it
->iters_node
);
4149 put_css_set_locked(it
->cur_cset
);
4152 it
->cur_cset
= cset
;
4153 list_add(&it
->iters_node
, &cset
->task_iters
);
4156 static void css_task_iter_skip(struct css_task_iter
*it
,
4157 struct task_struct
*task
)
4159 lockdep_assert_held(&css_set_lock
);
4161 if (it
->task_pos
== &task
->cg_list
) {
4162 it
->task_pos
= it
->task_pos
->next
;
4163 it
->flags
|= CSS_TASK_ITER_SKIPPED
;
4167 static void css_task_iter_advance(struct css_task_iter
*it
)
4169 struct task_struct
*task
;
4171 lockdep_assert_held(&css_set_lock
);
4175 * Advance iterator to find next entry. cset->tasks is
4176 * consumed first and then ->mg_tasks. After ->mg_tasks,
4177 * we move onto the next cset.
4179 if (it
->flags
& CSS_TASK_ITER_SKIPPED
)
4180 it
->flags
&= ~CSS_TASK_ITER_SKIPPED
;
4182 it
->task_pos
= it
->task_pos
->next
;
4184 if (it
->task_pos
== it
->tasks_head
)
4185 it
->task_pos
= it
->mg_tasks_head
->next
;
4186 if (it
->task_pos
== it
->mg_tasks_head
)
4187 it
->task_pos
= it
->dying_tasks_head
->next
;
4188 if (it
->task_pos
== it
->dying_tasks_head
)
4189 css_task_iter_advance_css_set(it
);
4191 /* called from start, proceed to the first cset */
4192 css_task_iter_advance_css_set(it
);
4198 task
= list_entry(it
->task_pos
, struct task_struct
, cg_list
);
4200 if (it
->flags
& CSS_TASK_ITER_PROCS
) {
4201 /* if PROCS, skip over tasks which aren't group leaders */
4202 if (!thread_group_leader(task
))
4205 /* and dying leaders w/o live member threads */
4206 if (!atomic_read(&task
->signal
->live
))
4209 /* skip all dying ones */
4210 if (task
->flags
& PF_EXITING
)
4216 * css_task_iter_start - initiate task iteration
4217 * @css: the css to walk tasks of
4218 * @flags: CSS_TASK_ITER_* flags
4219 * @it: the task iterator to use
4221 * Initiate iteration through the tasks of @css. The caller can call
4222 * css_task_iter_next() to walk through the tasks until the function
4223 * returns NULL. On completion of iteration, css_task_iter_end() must be
4226 void css_task_iter_start(struct cgroup_subsys_state
*css
, unsigned int flags
,
4227 struct css_task_iter
*it
)
4229 /* no one should try to iterate before mounting cgroups */
4230 WARN_ON_ONCE(!use_task_css_set_links
);
4232 memset(it
, 0, sizeof(*it
));
4234 spin_lock_irq(&css_set_lock
);
4240 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
4242 it
->cset_pos
= &css
->cgroup
->cset_links
;
4244 it
->cset_head
= it
->cset_pos
;
4246 css_task_iter_advance(it
);
4248 spin_unlock_irq(&css_set_lock
);
4252 * css_task_iter_next - return the next task for the iterator
4253 * @it: the task iterator being iterated
4255 * The "next" function for task iteration. @it should have been
4256 * initialized via css_task_iter_start(). Returns NULL when the iteration
4259 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
4262 put_task_struct(it
->cur_task
);
4263 it
->cur_task
= NULL
;
4266 spin_lock_irq(&css_set_lock
);
4269 it
->cur_task
= list_entry(it
->task_pos
, struct task_struct
,
4271 get_task_struct(it
->cur_task
);
4272 css_task_iter_advance(it
);
4275 spin_unlock_irq(&css_set_lock
);
4277 return it
->cur_task
;
4281 * css_task_iter_end - finish task iteration
4282 * @it: the task iterator to finish
4284 * Finish task iteration started by css_task_iter_start().
4286 void css_task_iter_end(struct css_task_iter
*it
)
4289 spin_lock_irq(&css_set_lock
);
4290 list_del(&it
->iters_node
);
4291 put_css_set_locked(it
->cur_cset
);
4292 spin_unlock_irq(&css_set_lock
);
4296 put_css_set(it
->cur_dcset
);
4299 put_task_struct(it
->cur_task
);
4302 static void cgroup_procs_release(struct kernfs_open_file
*of
)
4305 css_task_iter_end(of
->priv
);
4310 static void *cgroup_procs_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
4312 struct kernfs_open_file
*of
= s
->private;
4313 struct css_task_iter
*it
= of
->priv
;
4315 return css_task_iter_next(it
);
4318 static void *__cgroup_procs_start(struct seq_file
*s
, loff_t
*pos
,
4319 unsigned int iter_flags
)
4321 struct kernfs_open_file
*of
= s
->private;
4322 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4323 struct css_task_iter
*it
= of
->priv
;
4326 * When a seq_file is seeked, it's always traversed sequentially
4327 * from position 0, so we can simply keep iterating on !0 *pos.
4330 if (WARN_ON_ONCE((*pos
)++))
4331 return ERR_PTR(-EINVAL
);
4333 it
= kzalloc(sizeof(*it
), GFP_KERNEL
);
4335 return ERR_PTR(-ENOMEM
);
4337 css_task_iter_start(&cgrp
->self
, iter_flags
, it
);
4338 } else if (!(*pos
)++) {
4339 css_task_iter_end(it
);
4340 css_task_iter_start(&cgrp
->self
, iter_flags
, it
);
4343 return cgroup_procs_next(s
, NULL
, NULL
);
4346 static void *cgroup_procs_start(struct seq_file
*s
, loff_t
*pos
)
4348 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4351 * All processes of a threaded subtree belong to the domain cgroup
4352 * of the subtree. Only threads can be distributed across the
4353 * subtree. Reject reads on cgroup.procs in the subtree proper.
4354 * They're always empty anyway.
4356 if (cgroup_is_threaded(cgrp
))
4357 return ERR_PTR(-EOPNOTSUPP
);
4359 return __cgroup_procs_start(s
, pos
, CSS_TASK_ITER_PROCS
|
4360 CSS_TASK_ITER_THREADED
);
4363 static int cgroup_procs_show(struct seq_file
*s
, void *v
)
4365 seq_printf(s
, "%d\n", task_pid_vnr(v
));
4369 static int cgroup_procs_write_permission(struct cgroup
*src_cgrp
,
4370 struct cgroup
*dst_cgrp
,
4371 struct super_block
*sb
)
4373 struct cgroup_namespace
*ns
= current
->nsproxy
->cgroup_ns
;
4374 struct cgroup
*com_cgrp
= src_cgrp
;
4375 struct inode
*inode
;
4378 lockdep_assert_held(&cgroup_mutex
);
4380 /* find the common ancestor */
4381 while (!cgroup_is_descendant(dst_cgrp
, com_cgrp
))
4382 com_cgrp
= cgroup_parent(com_cgrp
);
4384 /* %current should be authorized to migrate to the common ancestor */
4385 inode
= kernfs_get_inode(sb
, com_cgrp
->procs_file
.kn
);
4389 ret
= inode_permission(inode
, MAY_WRITE
);
4395 * If namespaces are delegation boundaries, %current must be able
4396 * to see both source and destination cgroups from its namespace.
4398 if ((cgrp_dfl_root
.flags
& CGRP_ROOT_NS_DELEGATE
) &&
4399 (!cgroup_is_descendant(src_cgrp
, ns
->root_cset
->dfl_cgrp
) ||
4400 !cgroup_is_descendant(dst_cgrp
, ns
->root_cset
->dfl_cgrp
)))
4406 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
4407 char *buf
, size_t nbytes
, loff_t off
)
4409 struct cgroup
*src_cgrp
, *dst_cgrp
;
4410 struct task_struct
*task
;
4413 dst_cgrp
= cgroup_kn_lock_live(of
->kn
, false);
4417 task
= cgroup_procs_write_start(buf
, true);
4418 ret
= PTR_ERR_OR_ZERO(task
);
4422 /* find the source cgroup */
4423 spin_lock_irq(&css_set_lock
);
4424 src_cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
4425 spin_unlock_irq(&css_set_lock
);
4427 ret
= cgroup_procs_write_permission(src_cgrp
, dst_cgrp
,
4428 of
->file
->f_path
.dentry
->d_sb
);
4432 ret
= cgroup_attach_task(dst_cgrp
, task
, true);
4435 cgroup_procs_write_finish(task
);
4437 cgroup_kn_unlock(of
->kn
);
4439 return ret
?: nbytes
;
4442 static void *cgroup_threads_start(struct seq_file
*s
, loff_t
*pos
)
4444 return __cgroup_procs_start(s
, pos
, 0);
4447 static ssize_t
cgroup_threads_write(struct kernfs_open_file
*of
,
4448 char *buf
, size_t nbytes
, loff_t off
)
4450 struct cgroup
*src_cgrp
, *dst_cgrp
;
4451 struct task_struct
*task
;
4454 buf
= strstrip(buf
);
4456 dst_cgrp
= cgroup_kn_lock_live(of
->kn
, false);
4460 task
= cgroup_procs_write_start(buf
, false);
4461 ret
= PTR_ERR_OR_ZERO(task
);
4465 /* find the source cgroup */
4466 spin_lock_irq(&css_set_lock
);
4467 src_cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
4468 spin_unlock_irq(&css_set_lock
);
4470 /* thread migrations follow the cgroup.procs delegation rule */
4471 ret
= cgroup_procs_write_permission(src_cgrp
, dst_cgrp
,
4472 of
->file
->f_path
.dentry
->d_sb
);
4476 /* and must be contained in the same domain */
4478 if (src_cgrp
->dom_cgrp
!= dst_cgrp
->dom_cgrp
)
4481 ret
= cgroup_attach_task(dst_cgrp
, task
, false);
4484 cgroup_procs_write_finish(task
);
4486 cgroup_kn_unlock(of
->kn
);
4488 return ret
?: nbytes
;
4491 /* cgroup core interface files for the default hierarchy */
4492 static struct cftype cgroup_base_files
[] = {
4494 .name
= "cgroup.type",
4495 .flags
= CFTYPE_NOT_ON_ROOT
,
4496 .seq_show
= cgroup_type_show
,
4497 .write
= cgroup_type_write
,
4500 .name
= "cgroup.procs",
4501 .flags
= CFTYPE_NS_DELEGATABLE
,
4502 .file_offset
= offsetof(struct cgroup
, procs_file
),
4503 .release
= cgroup_procs_release
,
4504 .seq_start
= cgroup_procs_start
,
4505 .seq_next
= cgroup_procs_next
,
4506 .seq_show
= cgroup_procs_show
,
4507 .write
= cgroup_procs_write
,
4510 .name
= "cgroup.threads",
4511 .flags
= CFTYPE_NS_DELEGATABLE
,
4512 .release
= cgroup_procs_release
,
4513 .seq_start
= cgroup_threads_start
,
4514 .seq_next
= cgroup_procs_next
,
4515 .seq_show
= cgroup_procs_show
,
4516 .write
= cgroup_threads_write
,
4519 .name
= "cgroup.controllers",
4520 .seq_show
= cgroup_controllers_show
,
4523 .name
= "cgroup.subtree_control",
4524 .flags
= CFTYPE_NS_DELEGATABLE
,
4525 .seq_show
= cgroup_subtree_control_show
,
4526 .write
= cgroup_subtree_control_write
,
4529 .name
= "cgroup.events",
4530 .flags
= CFTYPE_NOT_ON_ROOT
,
4531 .file_offset
= offsetof(struct cgroup
, events_file
),
4532 .seq_show
= cgroup_events_show
,
4535 .name
= "cgroup.max.descendants",
4536 .seq_show
= cgroup_max_descendants_show
,
4537 .write
= cgroup_max_descendants_write
,
4540 .name
= "cgroup.max.depth",
4541 .seq_show
= cgroup_max_depth_show
,
4542 .write
= cgroup_max_depth_write
,
4545 .name
= "cgroup.stat",
4546 .seq_show
= cgroup_stat_show
,
4550 .flags
= CFTYPE_NOT_ON_ROOT
,
4551 .seq_show
= cpu_stat_show
,
4557 * css destruction is four-stage process.
4559 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4560 * Implemented in kill_css().
4562 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4563 * and thus css_tryget_online() is guaranteed to fail, the css can be
4564 * offlined by invoking offline_css(). After offlining, the base ref is
4565 * put. Implemented in css_killed_work_fn().
4567 * 3. When the percpu_ref reaches zero, the only possible remaining
4568 * accessors are inside RCU read sections. css_release() schedules the
4571 * 4. After the grace period, the css can be freed. Implemented in
4572 * css_free_work_fn().
4574 * It is actually hairier because both step 2 and 4 require process context
4575 * and thus involve punting to css->destroy_work adding two additional
4576 * steps to the already complex sequence.
4578 static void css_free_work_fn(struct work_struct
*work
)
4580 struct cgroup_subsys_state
*css
=
4581 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4582 struct cgroup_subsys
*ss
= css
->ss
;
4583 struct cgroup
*cgrp
= css
->cgroup
;
4585 percpu_ref_exit(&css
->refcnt
);
4589 struct cgroup_subsys_state
*parent
= css
->parent
;
4593 cgroup_idr_remove(&ss
->css_idr
, id
);
4599 /* cgroup free path */
4600 atomic_dec(&cgrp
->root
->nr_cgrps
);
4601 cgroup1_pidlist_destroy_all(cgrp
);
4602 cancel_work_sync(&cgrp
->release_agent_work
);
4604 if (cgroup_parent(cgrp
)) {
4606 * We get a ref to the parent, and put the ref when
4607 * this cgroup is being freed, so it's guaranteed
4608 * that the parent won't be destroyed before its
4611 cgroup_put(cgroup_parent(cgrp
));
4612 kernfs_put(cgrp
->kn
);
4613 if (cgroup_on_dfl(cgrp
))
4614 cgroup_stat_exit(cgrp
);
4618 * This is root cgroup's refcnt reaching zero,
4619 * which indicates that the root should be
4622 cgroup_destroy_root(cgrp
->root
);
4627 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4629 struct cgroup_subsys_state
*css
=
4630 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4632 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4633 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4636 static void css_release_work_fn(struct work_struct
*work
)
4638 struct cgroup_subsys_state
*css
=
4639 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4640 struct cgroup_subsys
*ss
= css
->ss
;
4641 struct cgroup
*cgrp
= css
->cgroup
;
4643 mutex_lock(&cgroup_mutex
);
4645 css
->flags
|= CSS_RELEASED
;
4646 list_del_rcu(&css
->sibling
);
4649 /* css release path */
4650 cgroup_idr_replace(&ss
->css_idr
, NULL
, css
->id
);
4651 if (ss
->css_released
)
4652 ss
->css_released(css
);
4654 struct cgroup
*tcgrp
;
4656 /* cgroup release path */
4657 trace_cgroup_release(cgrp
);
4659 if (cgroup_on_dfl(cgrp
))
4660 cgroup_stat_flush(cgrp
);
4662 spin_lock_irq(&css_set_lock
);
4663 for (tcgrp
= cgroup_parent(cgrp
); tcgrp
;
4664 tcgrp
= cgroup_parent(tcgrp
))
4665 tcgrp
->nr_dying_descendants
--;
4666 spin_unlock_irq(&css_set_lock
);
4668 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
4672 * There are two control paths which try to determine
4673 * cgroup from dentry without going through kernfs -
4674 * cgroupstats_build() and css_tryget_online_from_dir().
4675 * Those are supported by RCU protecting clearing of
4676 * cgrp->kn->priv backpointer.
4679 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
,
4682 cgroup_bpf_put(cgrp
);
4685 mutex_unlock(&cgroup_mutex
);
4687 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4690 static void css_release(struct percpu_ref
*ref
)
4692 struct cgroup_subsys_state
*css
=
4693 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4695 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
4696 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4699 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4700 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4702 lockdep_assert_held(&cgroup_mutex
);
4704 cgroup_get_live(cgrp
);
4706 memset(css
, 0, sizeof(*css
));
4710 INIT_LIST_HEAD(&css
->sibling
);
4711 INIT_LIST_HEAD(&css
->children
);
4712 css
->serial_nr
= css_serial_nr_next
++;
4713 atomic_set(&css
->online_cnt
, 0);
4715 if (cgroup_parent(cgrp
)) {
4716 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
4717 css_get(css
->parent
);
4720 BUG_ON(cgroup_css(cgrp
, ss
));
4723 /* invoke ->css_online() on a new CSS and mark it online if successful */
4724 static int online_css(struct cgroup_subsys_state
*css
)
4726 struct cgroup_subsys
*ss
= css
->ss
;
4729 lockdep_assert_held(&cgroup_mutex
);
4732 ret
= ss
->css_online(css
);
4734 css
->flags
|= CSS_ONLINE
;
4735 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4737 atomic_inc(&css
->online_cnt
);
4739 atomic_inc(&css
->parent
->online_cnt
);
4744 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4745 static void offline_css(struct cgroup_subsys_state
*css
)
4747 struct cgroup_subsys
*ss
= css
->ss
;
4749 lockdep_assert_held(&cgroup_mutex
);
4751 if (!(css
->flags
& CSS_ONLINE
))
4754 if (ss
->css_offline
)
4755 ss
->css_offline(css
);
4757 css
->flags
&= ~CSS_ONLINE
;
4758 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4760 wake_up_all(&css
->cgroup
->offline_waitq
);
4764 * css_create - create a cgroup_subsys_state
4765 * @cgrp: the cgroup new css will be associated with
4766 * @ss: the subsys of new css
4768 * Create a new css associated with @cgrp - @ss pair. On success, the new
4769 * css is online and installed in @cgrp. This function doesn't create the
4770 * interface files. Returns 0 on success, -errno on failure.
4772 static struct cgroup_subsys_state
*css_create(struct cgroup
*cgrp
,
4773 struct cgroup_subsys
*ss
)
4775 struct cgroup
*parent
= cgroup_parent(cgrp
);
4776 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
4777 struct cgroup_subsys_state
*css
;
4780 lockdep_assert_held(&cgroup_mutex
);
4782 css
= ss
->css_alloc(parent_css
);
4784 css
= ERR_PTR(-ENOMEM
);
4788 init_and_link_css(css
, ss
, cgrp
);
4790 err
= percpu_ref_init(&css
->refcnt
, css_release
, 0, GFP_KERNEL
);
4794 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_KERNEL
);
4799 /* @css is ready to be brought online now, make it visible */
4800 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
4801 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4803 err
= online_css(css
);
4807 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4808 cgroup_parent(parent
)) {
4809 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4810 current
->comm
, current
->pid
, ss
->name
);
4811 if (!strcmp(ss
->name
, "memory"))
4812 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4813 ss
->warned_broken_hierarchy
= true;
4819 list_del_rcu(&css
->sibling
);
4821 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4822 return ERR_PTR(err
);
4826 * The returned cgroup is fully initialized including its control mask, but
4827 * it isn't associated with its kernfs_node and doesn't have the control
4830 static struct cgroup
*cgroup_create(struct cgroup
*parent
)
4832 struct cgroup_root
*root
= parent
->root
;
4833 struct cgroup
*cgrp
, *tcgrp
;
4834 int level
= parent
->level
+ 1;
4837 /* allocate the cgroup and its ID, 0 is reserved for the root */
4838 cgrp
= kzalloc(sizeof(*cgrp
) +
4839 sizeof(cgrp
->ancestor_ids
[0]) * (level
+ 1), GFP_KERNEL
);
4841 return ERR_PTR(-ENOMEM
);
4843 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
, 0, GFP_KERNEL
);
4847 if (cgroup_on_dfl(parent
)) {
4848 ret
= cgroup_stat_init(cgrp
);
4850 goto out_cancel_ref
;
4854 * Temporarily set the pointer to NULL, so idr_find() won't return
4855 * a half-baked cgroup.
4857 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_KERNEL
);
4863 init_cgroup_housekeeping(cgrp
);
4865 cgrp
->self
.parent
= &parent
->self
;
4867 cgrp
->level
= level
;
4868 ret
= cgroup_bpf_inherit(cgrp
);
4872 spin_lock_irq(&css_set_lock
);
4873 for (tcgrp
= cgrp
; tcgrp
; tcgrp
= cgroup_parent(tcgrp
)) {
4874 cgrp
->ancestor_ids
[tcgrp
->level
] = tcgrp
->id
;
4877 tcgrp
->nr_descendants
++;
4879 spin_unlock_irq(&css_set_lock
);
4881 if (notify_on_release(parent
))
4882 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4884 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4885 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4887 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
4889 /* allocation complete, commit to creation */
4890 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
4891 atomic_inc(&root
->nr_cgrps
);
4892 cgroup_get_live(parent
);
4895 * @cgrp is now fully operational. If something fails after this
4896 * point, it'll be released via the normal destruction path.
4898 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4901 * On the default hierarchy, a child doesn't automatically inherit
4902 * subtree_control from the parent. Each is configured manually.
4904 if (!cgroup_on_dfl(cgrp
))
4905 cgrp
->subtree_control
= cgroup_control(cgrp
);
4907 cgroup_propagate_control(cgrp
);
4912 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4914 if (cgroup_on_dfl(parent
))
4915 cgroup_stat_exit(cgrp
);
4917 percpu_ref_exit(&cgrp
->self
.refcnt
);
4920 return ERR_PTR(ret
);
4923 static bool cgroup_check_hierarchy_limits(struct cgroup
*parent
)
4925 struct cgroup
*cgroup
;
4929 lockdep_assert_held(&cgroup_mutex
);
4931 for (cgroup
= parent
; cgroup
; cgroup
= cgroup_parent(cgroup
)) {
4932 if (cgroup
->nr_descendants
>= cgroup
->max_descendants
)
4935 if (level
> cgroup
->max_depth
)
4946 int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
, umode_t mode
)
4948 struct cgroup
*parent
, *cgrp
;
4949 struct kernfs_node
*kn
;
4952 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
4953 if (strchr(name
, '\n'))
4956 parent
= cgroup_kn_lock_live(parent_kn
, false);
4960 if (!cgroup_check_hierarchy_limits(parent
)) {
4965 cgrp
= cgroup_create(parent
);
4967 ret
= PTR_ERR(cgrp
);
4971 /* create the directory */
4972 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4980 * This extra ref will be put in cgroup_free_fn() and guarantees
4981 * that @cgrp->kn is always accessible.
4985 ret
= cgroup_kn_set_ugid(kn
);
4989 ret
= css_populate_dir(&cgrp
->self
);
4993 ret
= cgroup_apply_control_enable(cgrp
);
4997 trace_cgroup_mkdir(cgrp
);
4999 /* let's create and online css's */
5000 kernfs_activate(kn
);
5006 cgroup_destroy_locked(cgrp
);
5008 cgroup_kn_unlock(parent_kn
);
5013 * This is called when the refcnt of a css is confirmed to be killed.
5014 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5015 * initate destruction and put the css ref from kill_css().
5017 static void css_killed_work_fn(struct work_struct
*work
)
5019 struct cgroup_subsys_state
*css
=
5020 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
5022 mutex_lock(&cgroup_mutex
);
5027 /* @css can't go away while we're holding cgroup_mutex */
5029 } while (css
&& atomic_dec_and_test(&css
->online_cnt
));
5031 mutex_unlock(&cgroup_mutex
);
5034 /* css kill confirmation processing requires process context, bounce */
5035 static void css_killed_ref_fn(struct percpu_ref
*ref
)
5037 struct cgroup_subsys_state
*css
=
5038 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
5040 if (atomic_dec_and_test(&css
->online_cnt
)) {
5041 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
5042 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
5047 * kill_css - destroy a css
5048 * @css: css to destroy
5050 * This function initiates destruction of @css by removing cgroup interface
5051 * files and putting its base reference. ->css_offline() will be invoked
5052 * asynchronously once css_tryget_online() is guaranteed to fail and when
5053 * the reference count reaches zero, @css will be released.
5055 static void kill_css(struct cgroup_subsys_state
*css
)
5057 lockdep_assert_held(&cgroup_mutex
);
5059 if (css
->flags
& CSS_DYING
)
5062 css
->flags
|= CSS_DYING
;
5065 * This must happen before css is disassociated with its cgroup.
5066 * See seq_css() for details.
5071 * Killing would put the base ref, but we need to keep it alive
5072 * until after ->css_offline().
5077 * cgroup core guarantees that, by the time ->css_offline() is
5078 * invoked, no new css reference will be given out via
5079 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5080 * proceed to offlining css's because percpu_ref_kill() doesn't
5081 * guarantee that the ref is seen as killed on all CPUs on return.
5083 * Use percpu_ref_kill_and_confirm() to get notifications as each
5084 * css is confirmed to be seen as killed on all CPUs.
5086 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
5090 * cgroup_destroy_locked - the first stage of cgroup destruction
5091 * @cgrp: cgroup to be destroyed
5093 * css's make use of percpu refcnts whose killing latency shouldn't be
5094 * exposed to userland and are RCU protected. Also, cgroup core needs to
5095 * guarantee that css_tryget_online() won't succeed by the time
5096 * ->css_offline() is invoked. To satisfy all the requirements,
5097 * destruction is implemented in the following two steps.
5099 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5100 * userland visible parts and start killing the percpu refcnts of
5101 * css's. Set up so that the next stage will be kicked off once all
5102 * the percpu refcnts are confirmed to be killed.
5104 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5105 * rest of destruction. Once all cgroup references are gone, the
5106 * cgroup is RCU-freed.
5108 * This function implements s1. After this step, @cgrp is gone as far as
5109 * the userland is concerned and a new cgroup with the same name may be
5110 * created. As cgroup doesn't care about the names internally, this
5111 * doesn't cause any problem.
5113 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
5114 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
5116 struct cgroup
*tcgrp
, *parent
= cgroup_parent(cgrp
);
5117 struct cgroup_subsys_state
*css
;
5118 struct cgrp_cset_link
*link
;
5121 lockdep_assert_held(&cgroup_mutex
);
5124 * Only migration can raise populated from zero and we're already
5125 * holding cgroup_mutex.
5127 if (cgroup_is_populated(cgrp
))
5131 * Make sure there's no live children. We can't test emptiness of
5132 * ->self.children as dead children linger on it while being
5133 * drained; otherwise, "rmdir parent/child parent" may fail.
5135 if (css_has_online_children(&cgrp
->self
))
5139 * Mark @cgrp and the associated csets dead. The former prevents
5140 * further task migration and child creation by disabling
5141 * cgroup_lock_live_group(). The latter makes the csets ignored by
5142 * the migration path.
5144 cgrp
->self
.flags
&= ~CSS_ONLINE
;
5146 spin_lock_irq(&css_set_lock
);
5147 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
5148 link
->cset
->dead
= true;
5149 spin_unlock_irq(&css_set_lock
);
5151 /* initiate massacre of all css's */
5152 for_each_css(css
, ssid
, cgrp
)
5156 * Remove @cgrp directory along with the base files. @cgrp has an
5157 * extra ref on its kn.
5159 kernfs_remove(cgrp
->kn
);
5161 if (parent
&& cgroup_is_threaded(cgrp
))
5162 parent
->nr_threaded_children
--;
5164 spin_lock_irq(&css_set_lock
);
5165 for (tcgrp
= cgroup_parent(cgrp
); tcgrp
; tcgrp
= cgroup_parent(tcgrp
)) {
5166 tcgrp
->nr_descendants
--;
5167 tcgrp
->nr_dying_descendants
++;
5169 spin_unlock_irq(&css_set_lock
);
5171 cgroup1_check_for_release(parent
);
5173 /* put the base reference */
5174 percpu_ref_kill(&cgrp
->self
.refcnt
);
5179 int cgroup_rmdir(struct kernfs_node
*kn
)
5181 struct cgroup
*cgrp
;
5184 cgrp
= cgroup_kn_lock_live(kn
, false);
5188 ret
= cgroup_destroy_locked(cgrp
);
5191 trace_cgroup_rmdir(cgrp
);
5193 cgroup_kn_unlock(kn
);
5197 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
5198 .show_options
= cgroup_show_options
,
5199 .remount_fs
= cgroup_remount
,
5200 .mkdir
= cgroup_mkdir
,
5201 .rmdir
= cgroup_rmdir
,
5202 .show_path
= cgroup_show_path
,
5205 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
5207 struct cgroup_subsys_state
*css
;
5209 pr_debug("Initializing cgroup subsys %s\n", ss
->name
);
5211 mutex_lock(&cgroup_mutex
);
5213 idr_init(&ss
->css_idr
);
5214 INIT_LIST_HEAD(&ss
->cfts
);
5216 /* Create the root cgroup state for this subsystem */
5217 ss
->root
= &cgrp_dfl_root
;
5218 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
5219 /* We don't handle early failures gracefully */
5220 BUG_ON(IS_ERR(css
));
5221 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
5224 * Root csses are never destroyed and we can't initialize
5225 * percpu_ref during early init. Disable refcnting.
5227 css
->flags
|= CSS_NO_REF
;
5230 /* allocation can't be done safely during early init */
5233 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
5234 BUG_ON(css
->id
< 0);
5237 /* Update the init_css_set to contain a subsys
5238 * pointer to this state - since the subsystem is
5239 * newly registered, all tasks and hence the
5240 * init_css_set is in the subsystem's root cgroup. */
5241 init_css_set
.subsys
[ss
->id
] = css
;
5243 have_fork_callback
|= (bool)ss
->fork
<< ss
->id
;
5244 have_exit_callback
|= (bool)ss
->exit
<< ss
->id
;
5245 have_release_callback
|= (bool)ss
->release
<< ss
->id
;
5246 have_canfork_callback
|= (bool)ss
->can_fork
<< ss
->id
;
5248 /* At system boot, before all subsystems have been
5249 * registered, no tasks have been forked, so we don't
5250 * need to invoke fork callbacks here. */
5251 BUG_ON(!list_empty(&init_task
.tasks
));
5253 BUG_ON(online_css(css
));
5255 mutex_unlock(&cgroup_mutex
);
5259 * cgroup_init_early - cgroup initialization at system boot
5261 * Initialize cgroups at system boot, and initialize any
5262 * subsystems that request early init.
5264 int __init
cgroup_init_early(void)
5266 static struct cgroup_sb_opts __initdata opts
;
5267 struct cgroup_subsys
*ss
;
5270 init_cgroup_root(&cgrp_dfl_root
, &opts
);
5271 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
5273 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
5275 for_each_subsys(ss
, i
) {
5276 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
5277 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5278 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
5280 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
5281 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
5284 ss
->name
= cgroup_subsys_name
[i
];
5285 if (!ss
->legacy_name
)
5286 ss
->legacy_name
= cgroup_subsys_name
[i
];
5289 cgroup_init_subsys(ss
, true);
5294 static u16 cgroup_disable_mask __initdata
;
5297 * cgroup_init - cgroup initialization
5299 * Register cgroup filesystem and /proc file, and initialize
5300 * any subsystems that didn't request early init.
5302 int __init
cgroup_init(void)
5304 struct cgroup_subsys
*ss
;
5307 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT
> 16);
5308 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem
));
5309 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_base_files
));
5310 BUG_ON(cgroup_init_cftypes(NULL
, cgroup1_base_files
));
5315 * The latency of the synchronize_sched() is too high for cgroups,
5316 * avoid it at the cost of forcing all readers into the slow path.
5318 rcu_sync_enter_start(&cgroup_threadgroup_rwsem
.rss
);
5320 get_user_ns(init_cgroup_ns
.user_ns
);
5322 mutex_lock(&cgroup_mutex
);
5325 * Add init_css_set to the hash table so that dfl_root can link to
5328 hash_add(css_set_table
, &init_css_set
.hlist
,
5329 css_set_hash(init_css_set
.subsys
));
5331 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0, 0));
5333 mutex_unlock(&cgroup_mutex
);
5335 for_each_subsys(ss
, ssid
) {
5336 if (ss
->early_init
) {
5337 struct cgroup_subsys_state
*css
=
5338 init_css_set
.subsys
[ss
->id
];
5340 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
5342 BUG_ON(css
->id
< 0);
5344 cgroup_init_subsys(ss
, false);
5347 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
5348 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
5351 * Setting dfl_root subsys_mask needs to consider the
5352 * disabled flag and cftype registration needs kmalloc,
5353 * both of which aren't available during early_init.
5355 if (cgroup_disable_mask
& (1 << ssid
)) {
5356 static_branch_disable(cgroup_subsys_enabled_key
[ssid
]);
5357 printk(KERN_INFO
"Disabling %s control group subsystem\n",
5362 if (cgroup1_ssid_disabled(ssid
))
5363 printk(KERN_INFO
"Disabling %s control group subsystem in v1 mounts\n",
5366 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
5368 /* implicit controllers must be threaded too */
5369 WARN_ON(ss
->implicit_on_dfl
&& !ss
->threaded
);
5371 if (ss
->implicit_on_dfl
)
5372 cgrp_dfl_implicit_ss_mask
|= 1 << ss
->id
;
5373 else if (!ss
->dfl_cftypes
)
5374 cgrp_dfl_inhibit_ss_mask
|= 1 << ss
->id
;
5377 cgrp_dfl_threaded_ss_mask
|= 1 << ss
->id
;
5379 if (ss
->dfl_cftypes
== ss
->legacy_cftypes
) {
5380 WARN_ON(cgroup_add_cftypes(ss
, ss
->dfl_cftypes
));
5382 WARN_ON(cgroup_add_dfl_cftypes(ss
, ss
->dfl_cftypes
));
5383 WARN_ON(cgroup_add_legacy_cftypes(ss
, ss
->legacy_cftypes
));
5387 ss
->bind(init_css_set
.subsys
[ssid
]);
5389 mutex_lock(&cgroup_mutex
);
5390 css_populate_dir(init_css_set
.subsys
[ssid
]);
5391 mutex_unlock(&cgroup_mutex
);
5394 /* init_css_set.subsys[] has been updated, re-hash */
5395 hash_del(&init_css_set
.hlist
);
5396 hash_add(css_set_table
, &init_css_set
.hlist
,
5397 css_set_hash(init_css_set
.subsys
));
5399 WARN_ON(sysfs_create_mount_point(fs_kobj
, "cgroup"));
5400 WARN_ON(register_filesystem(&cgroup_fs_type
));
5401 WARN_ON(register_filesystem(&cgroup2_fs_type
));
5402 WARN_ON(!proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
));
5407 static int __init
cgroup_wq_init(void)
5410 * There isn't much point in executing destruction path in
5411 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5412 * Use 1 for @max_active.
5414 * We would prefer to do this in cgroup_init() above, but that
5415 * is called before init_workqueues(): so leave this until after.
5417 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
5418 BUG_ON(!cgroup_destroy_wq
);
5421 core_initcall(cgroup_wq_init
);
5423 void cgroup_path_from_kernfs_id(const union kernfs_node_id
*id
,
5424 char *buf
, size_t buflen
)
5426 struct kernfs_node
*kn
;
5428 kn
= kernfs_get_node_by_id(cgrp_dfl_root
.kf_root
, id
);
5431 kernfs_path(kn
, buf
, buflen
);
5436 * proc_cgroup_show()
5437 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5438 * - Used for /proc/<pid>/cgroup.
5440 int proc_cgroup_show(struct seq_file
*m
, struct pid_namespace
*ns
,
5441 struct pid
*pid
, struct task_struct
*tsk
)
5445 struct cgroup_root
*root
;
5448 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5452 mutex_lock(&cgroup_mutex
);
5453 spin_lock_irq(&css_set_lock
);
5455 for_each_root(root
) {
5456 struct cgroup_subsys
*ss
;
5457 struct cgroup
*cgrp
;
5458 int ssid
, count
= 0;
5460 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_visible
)
5463 seq_printf(m
, "%d:", root
->hierarchy_id
);
5464 if (root
!= &cgrp_dfl_root
)
5465 for_each_subsys(ss
, ssid
)
5466 if (root
->subsys_mask
& (1 << ssid
))
5467 seq_printf(m
, "%s%s", count
++ ? "," : "",
5469 if (strlen(root
->name
))
5470 seq_printf(m
, "%sname=%s", count
? "," : "",
5474 cgrp
= task_cgroup_from_root(tsk
, root
);
5477 * On traditional hierarchies, all zombie tasks show up as
5478 * belonging to the root cgroup. On the default hierarchy,
5479 * while a zombie doesn't show up in "cgroup.procs" and
5480 * thus can't be migrated, its /proc/PID/cgroup keeps
5481 * reporting the cgroup it belonged to before exiting. If
5482 * the cgroup is removed before the zombie is reaped,
5483 * " (deleted)" is appended to the cgroup path.
5485 if (cgroup_on_dfl(cgrp
) || !(tsk
->flags
& PF_EXITING
)) {
5486 retval
= cgroup_path_ns_locked(cgrp
, buf
, PATH_MAX
,
5487 current
->nsproxy
->cgroup_ns
);
5488 if (retval
>= PATH_MAX
)
5489 retval
= -ENAMETOOLONG
;
5498 if (cgroup_on_dfl(cgrp
) && cgroup_is_dead(cgrp
))
5499 seq_puts(m
, " (deleted)\n");
5506 spin_unlock_irq(&css_set_lock
);
5507 mutex_unlock(&cgroup_mutex
);
5514 * cgroup_fork - initialize cgroup related fields during copy_process()
5515 * @child: pointer to task_struct of forking parent process.
5517 * A task is associated with the init_css_set until cgroup_post_fork()
5518 * attaches it to the parent's css_set. Empty cg_list indicates that
5519 * @child isn't holding reference to its css_set.
5521 void cgroup_fork(struct task_struct
*child
)
5523 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
5524 INIT_LIST_HEAD(&child
->cg_list
);
5528 * cgroup_can_fork - called on a new task before the process is exposed
5529 * @child: the task in question.
5531 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5532 * returns an error, the fork aborts with that error code. This allows for
5533 * a cgroup subsystem to conditionally allow or deny new forks.
5535 int cgroup_can_fork(struct task_struct
*child
)
5537 struct cgroup_subsys
*ss
;
5540 do_each_subsys_mask(ss
, i
, have_canfork_callback
) {
5541 ret
= ss
->can_fork(child
);
5544 } while_each_subsys_mask();
5549 for_each_subsys(ss
, j
) {
5552 if (ss
->cancel_fork
)
5553 ss
->cancel_fork(child
);
5560 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5561 * @child: the task in question
5563 * This calls the cancel_fork() callbacks if a fork failed *after*
5564 * cgroup_can_fork() succeded.
5566 void cgroup_cancel_fork(struct task_struct
*child
)
5568 struct cgroup_subsys
*ss
;
5571 for_each_subsys(ss
, i
)
5572 if (ss
->cancel_fork
)
5573 ss
->cancel_fork(child
);
5577 * cgroup_post_fork - called on a new task after adding it to the task list
5578 * @child: the task in question
5580 * Adds the task to the list running through its css_set if necessary and
5581 * call the subsystem fork() callbacks. Has to be after the task is
5582 * visible on the task list in case we race with the first call to
5583 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5586 void cgroup_post_fork(struct task_struct
*child
)
5588 struct cgroup_subsys
*ss
;
5592 * This may race against cgroup_enable_task_cg_lists(). As that
5593 * function sets use_task_css_set_links before grabbing
5594 * tasklist_lock and we just went through tasklist_lock to add
5595 * @child, it's guaranteed that either we see the set
5596 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5597 * @child during its iteration.
5599 * If we won the race, @child is associated with %current's
5600 * css_set. Grabbing css_set_lock guarantees both that the
5601 * association is stable, and, on completion of the parent's
5602 * migration, @child is visible in the source of migration or
5603 * already in the destination cgroup. This guarantee is necessary
5604 * when implementing operations which need to migrate all tasks of
5605 * a cgroup to another.
5607 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5608 * will remain in init_css_set. This is safe because all tasks are
5609 * in the init_css_set before cg_links is enabled and there's no
5610 * operation which transfers all tasks out of init_css_set.
5612 if (use_task_css_set_links
) {
5613 struct css_set
*cset
;
5615 spin_lock_irq(&css_set_lock
);
5616 cset
= task_css_set(current
);
5617 if (list_empty(&child
->cg_list
)) {
5620 css_set_move_task(child
, NULL
, cset
, false);
5622 spin_unlock_irq(&css_set_lock
);
5626 * Call ss->fork(). This must happen after @child is linked on
5627 * css_set; otherwise, @child might change state between ->fork()
5628 * and addition to css_set.
5630 do_each_subsys_mask(ss
, i
, have_fork_callback
) {
5632 } while_each_subsys_mask();
5636 * cgroup_exit - detach cgroup from exiting task
5637 * @tsk: pointer to task_struct of exiting process
5639 * Description: Detach cgroup from @tsk and release it.
5641 * Note that cgroups marked notify_on_release force every task in
5642 * them to take the global cgroup_mutex mutex when exiting.
5643 * This could impact scaling on very large systems. Be reluctant to
5644 * use notify_on_release cgroups where very high task exit scaling
5645 * is required on large systems.
5647 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5648 * call cgroup_exit() while the task is still competent to handle
5649 * notify_on_release(), then leave the task attached to the root cgroup in
5650 * each hierarchy for the remainder of its exit. No need to bother with
5651 * init_css_set refcnting. init_css_set never goes away and we can't race
5652 * with migration path - PF_EXITING is visible to migration path.
5654 void cgroup_exit(struct task_struct
*tsk
)
5656 struct cgroup_subsys
*ss
;
5657 struct css_set
*cset
;
5661 * Unlink from @tsk from its css_set. As migration path can't race
5662 * with us, we can check css_set and cg_list without synchronization.
5664 cset
= task_css_set(tsk
);
5666 if (!list_empty(&tsk
->cg_list
)) {
5667 spin_lock_irq(&css_set_lock
);
5668 css_set_move_task(tsk
, cset
, NULL
, false);
5669 list_add_tail(&tsk
->cg_list
, &cset
->dying_tasks
);
5671 spin_unlock_irq(&css_set_lock
);
5676 /* see cgroup_post_fork() for details */
5677 do_each_subsys_mask(ss
, i
, have_exit_callback
) {
5679 } while_each_subsys_mask();
5682 void cgroup_release(struct task_struct
*task
)
5684 struct cgroup_subsys
*ss
;
5687 do_each_subsys_mask(ss
, ssid
, have_release_callback
) {
5689 } while_each_subsys_mask();
5691 if (use_task_css_set_links
) {
5692 spin_lock_irq(&css_set_lock
);
5693 css_set_skip_task_iters(task_css_set(task
), task
);
5694 list_del_init(&task
->cg_list
);
5695 spin_unlock_irq(&css_set_lock
);
5699 void cgroup_free(struct task_struct
*task
)
5701 struct css_set
*cset
= task_css_set(task
);
5705 static int __init
cgroup_disable(char *str
)
5707 struct cgroup_subsys
*ss
;
5711 while ((token
= strsep(&str
, ",")) != NULL
) {
5715 for_each_subsys(ss
, i
) {
5716 if (strcmp(token
, ss
->name
) &&
5717 strcmp(token
, ss
->legacy_name
))
5719 cgroup_disable_mask
|= 1 << i
;
5724 __setup("cgroup_disable=", cgroup_disable
);
5727 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5728 * @dentry: directory dentry of interest
5729 * @ss: subsystem of interest
5731 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5732 * to get the corresponding css and return it. If such css doesn't exist
5733 * or can't be pinned, an ERR_PTR value is returned.
5735 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
5736 struct cgroup_subsys
*ss
)
5738 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5739 struct file_system_type
*s_type
= dentry
->d_sb
->s_type
;
5740 struct cgroup_subsys_state
*css
= NULL
;
5741 struct cgroup
*cgrp
;
5743 /* is @dentry a cgroup dir? */
5744 if ((s_type
!= &cgroup_fs_type
&& s_type
!= &cgroup2_fs_type
) ||
5745 !kn
|| kernfs_type(kn
) != KERNFS_DIR
)
5746 return ERR_PTR(-EBADF
);
5751 * This path doesn't originate from kernfs and @kn could already
5752 * have been or be removed at any point. @kn->priv is RCU
5753 * protected for this access. See css_release_work_fn() for details.
5755 cgrp
= rcu_dereference(*(void __rcu __force
**)&kn
->priv
);
5757 css
= cgroup_css(cgrp
, ss
);
5759 if (!css
|| !css_tryget_online(css
))
5760 css
= ERR_PTR(-ENOENT
);
5767 * css_from_id - lookup css by id
5768 * @id: the cgroup id
5769 * @ss: cgroup subsys to be looked into
5771 * Returns the css if there's valid one with @id, otherwise returns NULL.
5772 * Should be called under rcu_read_lock().
5774 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5776 WARN_ON_ONCE(!rcu_read_lock_held());
5777 return idr_find(&ss
->css_idr
, id
);
5781 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5782 * @path: path on the default hierarchy
5784 * Find the cgroup at @path on the default hierarchy, increment its
5785 * reference count and return it. Returns pointer to the found cgroup on
5786 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5787 * if @path points to a non-directory.
5789 struct cgroup
*cgroup_get_from_path(const char *path
)
5791 struct kernfs_node
*kn
;
5792 struct cgroup
*cgrp
;
5794 mutex_lock(&cgroup_mutex
);
5796 kn
= kernfs_walk_and_get(cgrp_dfl_root
.cgrp
.kn
, path
);
5798 if (kernfs_type(kn
) == KERNFS_DIR
) {
5800 cgroup_get_live(cgrp
);
5802 cgrp
= ERR_PTR(-ENOTDIR
);
5806 cgrp
= ERR_PTR(-ENOENT
);
5809 mutex_unlock(&cgroup_mutex
);
5812 EXPORT_SYMBOL_GPL(cgroup_get_from_path
);
5815 * cgroup_get_from_fd - get a cgroup pointer from a fd
5816 * @fd: fd obtained by open(cgroup2_dir)
5818 * Find the cgroup from a fd which should be obtained
5819 * by opening a cgroup directory. Returns a pointer to the
5820 * cgroup on success. ERR_PTR is returned if the cgroup
5823 struct cgroup
*cgroup_get_from_fd(int fd
)
5825 struct cgroup_subsys_state
*css
;
5826 struct cgroup
*cgrp
;
5831 return ERR_PTR(-EBADF
);
5833 css
= css_tryget_online_from_dir(f
->f_path
.dentry
, NULL
);
5836 return ERR_CAST(css
);
5839 if (!cgroup_on_dfl(cgrp
)) {
5841 return ERR_PTR(-EBADF
);
5846 EXPORT_SYMBOL_GPL(cgroup_get_from_fd
);
5849 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5850 * definition in cgroup-defs.h.
5852 #ifdef CONFIG_SOCK_CGROUP_DATA
5854 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5856 DEFINE_SPINLOCK(cgroup_sk_update_lock
);
5857 static bool cgroup_sk_alloc_disabled __read_mostly
;
5859 void cgroup_sk_alloc_disable(void)
5861 if (cgroup_sk_alloc_disabled
)
5863 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5864 cgroup_sk_alloc_disabled
= true;
5869 #define cgroup_sk_alloc_disabled false
5873 void cgroup_sk_alloc(struct sock_cgroup_data
*skcd
)
5875 if (cgroup_sk_alloc_disabled
)
5878 /* Socket clone path */
5881 * We might be cloning a socket which is left in an empty
5882 * cgroup and the cgroup might have already been rmdir'd.
5883 * Don't use cgroup_get_live().
5885 cgroup_get(sock_cgroup_ptr(skcd
));
5892 struct css_set
*cset
;
5894 cset
= task_css_set(current
);
5895 if (likely(cgroup_tryget(cset
->dfl_cgrp
))) {
5896 skcd
->val
= (unsigned long)cset
->dfl_cgrp
;
5905 void cgroup_sk_free(struct sock_cgroup_data
*skcd
)
5907 cgroup_put(sock_cgroup_ptr(skcd
));
5910 #endif /* CONFIG_SOCK_CGROUP_DATA */
5912 #ifdef CONFIG_CGROUP_BPF
5913 int cgroup_bpf_attach(struct cgroup
*cgrp
, struct bpf_prog
*prog
,
5914 enum bpf_attach_type type
, u32 flags
)
5918 mutex_lock(&cgroup_mutex
);
5919 ret
= __cgroup_bpf_attach(cgrp
, prog
, type
, flags
);
5920 mutex_unlock(&cgroup_mutex
);
5923 int cgroup_bpf_detach(struct cgroup
*cgrp
, struct bpf_prog
*prog
,
5924 enum bpf_attach_type type
, u32 flags
)
5928 mutex_lock(&cgroup_mutex
);
5929 ret
= __cgroup_bpf_detach(cgrp
, prog
, type
, flags
);
5930 mutex_unlock(&cgroup_mutex
);
5933 int cgroup_bpf_query(struct cgroup
*cgrp
, const union bpf_attr
*attr
,
5934 union bpf_attr __user
*uattr
)
5938 mutex_lock(&cgroup_mutex
);
5939 ret
= __cgroup_bpf_query(cgrp
, attr
, uattr
);
5940 mutex_unlock(&cgroup_mutex
);
5943 #endif /* CONFIG_CGROUP_BPF */
5946 static ssize_t
show_delegatable_files(struct cftype
*files
, char *buf
,
5947 ssize_t size
, const char *prefix
)
5952 for (cft
= files
; cft
&& cft
->name
[0] != '\0'; cft
++) {
5953 if (!(cft
->flags
& CFTYPE_NS_DELEGATABLE
))
5957 ret
+= snprintf(buf
+ ret
, size
- ret
, "%s.", prefix
);
5959 ret
+= snprintf(buf
+ ret
, size
- ret
, "%s\n", cft
->name
);
5961 if (unlikely(ret
>= size
)) {
5970 static ssize_t
delegate_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
5973 struct cgroup_subsys
*ss
;
5977 ret
= show_delegatable_files(cgroup_base_files
, buf
, PAGE_SIZE
- ret
,
5980 for_each_subsys(ss
, ssid
)
5981 ret
+= show_delegatable_files(ss
->dfl_cftypes
, buf
+ ret
,
5983 cgroup_subsys_name
[ssid
]);
5987 static struct kobj_attribute cgroup_delegate_attr
= __ATTR_RO(delegate
);
5989 static ssize_t
features_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
5992 return snprintf(buf
, PAGE_SIZE
, "nsdelegate\n");
5994 static struct kobj_attribute cgroup_features_attr
= __ATTR_RO(features
);
5996 static struct attribute
*cgroup_sysfs_attrs
[] = {
5997 &cgroup_delegate_attr
.attr
,
5998 &cgroup_features_attr
.attr
,
6002 static const struct attribute_group cgroup_sysfs_attr_group
= {
6003 .attrs
= cgroup_sysfs_attrs
,
6007 static int __init
cgroup_sysfs_init(void)
6009 return sysfs_create_group(kernel_kobj
, &cgroup_sysfs_attr_group
);
6011 subsys_initcall(cgroup_sysfs_init
);
6012 #endif /* CONFIG_SYSFS */