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 <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex
);
87 DECLARE_RWSEM(css_set_rwsem
);
88 EXPORT_SYMBOL_GPL(cgroup_mutex
);
89 EXPORT_SYMBOL_GPL(css_set_rwsem
);
91 static DEFINE_MUTEX(cgroup_mutex
);
92 static DECLARE_RWSEM(css_set_rwsem
);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock
);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock
);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem
;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
111 !lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct
*cgroup_destroy_wq
;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct
*cgroup_pidlist_destroy_wq
;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys
*cgroup_subsys
[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name
[] = {
138 #include <linux/cgroup_subsys.h>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true
*cgroup_subsys_enabled_key
[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true
*cgroup_subsys_on_dfl_key
[] = {
159 #include <linux/cgroup_subsys.h>
164 * The default hierarchy, reserved for the subsystems that are otherwise
165 * unattached - it never has more than a single cgroup, and all tasks are
166 * part of that cgroup.
168 struct cgroup_root cgrp_dfl_root
;
169 EXPORT_SYMBOL_GPL(cgrp_dfl_root
);
172 * The default hierarchy always exists but is hidden until mounted for the
173 * first time. This is for backward compatibility.
175 static bool cgrp_dfl_root_visible
;
178 * Set by the boot param of the same name and makes subsystems with NULL
179 * ->dfl_files to use ->legacy_files on the default hierarchy.
181 static bool cgroup_legacy_files_on_dfl
;
183 /* some controllers are not supported in the default hierarchy */
184 static unsigned long cgrp_dfl_root_inhibit_ss_mask
;
186 /* The list of hierarchy roots */
188 static LIST_HEAD(cgroup_roots
);
189 static int cgroup_root_count
;
191 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
192 static DEFINE_IDR(cgroup_hierarchy_idr
);
195 * Assign a monotonically increasing serial number to csses. It guarantees
196 * cgroups with bigger numbers are newer than those with smaller numbers.
197 * Also, as csses are always appended to the parent's ->children list, it
198 * guarantees that sibling csses are always sorted in the ascending serial
199 * number order on the list. Protected by cgroup_mutex.
201 static u64 css_serial_nr_next
= 1;
204 * These bitmask flags indicate whether tasks in the fork and exit paths have
205 * fork/exit handlers to call. This avoids us having to do extra work in the
206 * fork/exit path to check which subsystems have fork/exit callbacks.
208 static unsigned long have_fork_callback __read_mostly
;
209 static unsigned long have_exit_callback __read_mostly
;
211 /* Ditto for the can_fork callback. */
212 static unsigned long have_canfork_callback __read_mostly
;
214 static struct cftype cgroup_dfl_base_files
[];
215 static struct cftype cgroup_legacy_base_files
[];
217 static int rebind_subsystems(struct cgroup_root
*dst_root
,
218 unsigned long ss_mask
);
219 static void css_task_iter_advance(struct css_task_iter
*it
);
220 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
221 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
223 static void css_release(struct percpu_ref
*ref
);
224 static void kill_css(struct cgroup_subsys_state
*css
);
225 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
226 struct cgroup
*cgrp
, struct cftype cfts
[],
230 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
231 * @ssid: subsys ID of interest
233 * cgroup_subsys_enabled() can only be used with literal subsys names which
234 * is fine for individual subsystems but unsuitable for cgroup core. This
235 * is slower static_key_enabled() based test indexed by @ssid.
237 static bool cgroup_ssid_enabled(int ssid
)
239 return static_key_enabled(cgroup_subsys_enabled_key
[ssid
]);
243 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
244 * @cgrp: the cgroup of interest
246 * The default hierarchy is the v2 interface of cgroup and this function
247 * can be used to test whether a cgroup is on the default hierarchy for
248 * cases where a subsystem should behave differnetly depending on the
251 * The set of behaviors which change on the default hierarchy are still
252 * being determined and the mount option is prefixed with __DEVEL__.
254 * List of changed behaviors:
256 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
257 * and "name" are disallowed.
259 * - When mounting an existing superblock, mount options should match.
261 * - Remount is disallowed.
263 * - rename(2) is disallowed.
265 * - "tasks" is removed. Everything should be at process granularity. Use
266 * "cgroup.procs" instead.
268 * - "cgroup.procs" is not sorted. pids will be unique unless they got
269 * recycled inbetween reads.
271 * - "release_agent" and "notify_on_release" are removed. Replacement
272 * notification mechanism will be implemented.
274 * - "cgroup.clone_children" is removed.
276 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
277 * and its descendants contain no task; otherwise, 1. The file also
278 * generates kernfs notification which can be monitored through poll and
279 * [di]notify when the value of the file changes.
281 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
282 * take masks of ancestors with non-empty cpus/mems, instead of being
283 * moved to an ancestor.
285 * - cpuset: a task can be moved into an empty cpuset, and again it takes
286 * masks of ancestors.
288 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
291 * - blkcg: blk-throttle becomes properly hierarchical.
293 * - debug: disallowed on the default hierarchy.
295 static bool cgroup_on_dfl(const struct cgroup
*cgrp
)
297 return cgrp
->root
== &cgrp_dfl_root
;
300 /* IDR wrappers which synchronize using cgroup_idr_lock */
301 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
306 idr_preload(gfp_mask
);
307 spin_lock_bh(&cgroup_idr_lock
);
308 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
& ~__GFP_WAIT
);
309 spin_unlock_bh(&cgroup_idr_lock
);
314 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
318 spin_lock_bh(&cgroup_idr_lock
);
319 ret
= idr_replace(idr
, ptr
, id
);
320 spin_unlock_bh(&cgroup_idr_lock
);
324 static void cgroup_idr_remove(struct idr
*idr
, int id
)
326 spin_lock_bh(&cgroup_idr_lock
);
328 spin_unlock_bh(&cgroup_idr_lock
);
331 static struct cgroup
*cgroup_parent(struct cgroup
*cgrp
)
333 struct cgroup_subsys_state
*parent_css
= cgrp
->self
.parent
;
336 return container_of(parent_css
, struct cgroup
, self
);
341 * cgroup_css - obtain a cgroup's css for the specified subsystem
342 * @cgrp: the cgroup of interest
343 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
345 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
346 * function must be called either under cgroup_mutex or rcu_read_lock() and
347 * the caller is responsible for pinning the returned css if it wants to
348 * keep accessing it outside the said locks. This function may return
349 * %NULL if @cgrp doesn't have @subsys_id enabled.
351 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
352 struct cgroup_subsys
*ss
)
355 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
356 lockdep_is_held(&cgroup_mutex
));
362 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
363 * @cgrp: the cgroup of interest
364 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
366 * Similar to cgroup_css() but returns the effective css, which is defined
367 * as the matching css of the nearest ancestor including self which has @ss
368 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
369 * function is guaranteed to return non-NULL css.
371 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
372 struct cgroup_subsys
*ss
)
374 lockdep_assert_held(&cgroup_mutex
);
379 if (!(cgrp
->root
->subsys_mask
& (1 << ss
->id
)))
383 * This function is used while updating css associations and thus
384 * can't test the csses directly. Use ->child_subsys_mask.
386 while (cgroup_parent(cgrp
) &&
387 !(cgroup_parent(cgrp
)->child_subsys_mask
& (1 << ss
->id
)))
388 cgrp
= cgroup_parent(cgrp
);
390 return cgroup_css(cgrp
, ss
);
394 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
395 * @cgrp: the cgroup of interest
396 * @ss: the subsystem of interest
398 * Find and get the effective css of @cgrp for @ss. The effective css is
399 * defined as the matching css of the nearest ancestor including self which
400 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
401 * the root css is returned, so this function always returns a valid css.
402 * The returned css must be put using css_put().
404 struct cgroup_subsys_state
*cgroup_get_e_css(struct cgroup
*cgrp
,
405 struct cgroup_subsys
*ss
)
407 struct cgroup_subsys_state
*css
;
412 css
= cgroup_css(cgrp
, ss
);
414 if (css
&& css_tryget_online(css
))
416 cgrp
= cgroup_parent(cgrp
);
419 css
= init_css_set
.subsys
[ss
->id
];
426 /* convenient tests for these bits */
427 static inline bool cgroup_is_dead(const struct cgroup
*cgrp
)
429 return !(cgrp
->self
.flags
& CSS_ONLINE
);
432 static void cgroup_get(struct cgroup
*cgrp
)
434 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
435 css_get(&cgrp
->self
);
438 static bool cgroup_tryget(struct cgroup
*cgrp
)
440 return css_tryget(&cgrp
->self
);
443 static void cgroup_put(struct cgroup
*cgrp
)
445 css_put(&cgrp
->self
);
448 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
450 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
451 struct cftype
*cft
= of_cft(of
);
454 * This is open and unprotected implementation of cgroup_css().
455 * seq_css() is only called from a kernfs file operation which has
456 * an active reference on the file. Because all the subsystem
457 * files are drained before a css is disassociated with a cgroup,
458 * the matching css from the cgroup's subsys table is guaranteed to
459 * be and stay valid until the enclosing operation is complete.
462 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
466 EXPORT_SYMBOL_GPL(of_css
);
469 * cgroup_is_descendant - test ancestry
470 * @cgrp: the cgroup to be tested
471 * @ancestor: possible ancestor of @cgrp
473 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
474 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
475 * and @ancestor are accessible.
477 bool cgroup_is_descendant(struct cgroup
*cgrp
, struct cgroup
*ancestor
)
480 if (cgrp
== ancestor
)
482 cgrp
= cgroup_parent(cgrp
);
487 static int notify_on_release(const struct cgroup
*cgrp
)
489 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
493 * for_each_css - iterate all css's of a cgroup
494 * @css: the iteration cursor
495 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
496 * @cgrp: the target cgroup to iterate css's of
498 * Should be called under cgroup_[tree_]mutex.
500 #define for_each_css(css, ssid, cgrp) \
501 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
502 if (!((css) = rcu_dereference_check( \
503 (cgrp)->subsys[(ssid)], \
504 lockdep_is_held(&cgroup_mutex)))) { } \
508 * for_each_e_css - iterate all effective css's of a cgroup
509 * @css: the iteration cursor
510 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
511 * @cgrp: the target cgroup to iterate css's of
513 * Should be called under cgroup_[tree_]mutex.
515 #define for_each_e_css(css, ssid, cgrp) \
516 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
517 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
522 * for_each_subsys - iterate all enabled cgroup subsystems
523 * @ss: the iteration cursor
524 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
526 #define for_each_subsys(ss, ssid) \
527 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
528 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
531 * for_each_subsys_which - filter for_each_subsys with a bitmask
532 * @ss: the iteration cursor
533 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
534 * @ss_maskp: a pointer to the bitmask
536 * The block will only run for cases where the ssid-th bit (1 << ssid) of
539 #define for_each_subsys_which(ss, ssid, ss_maskp) \
540 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
543 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
544 if (((ss) = cgroup_subsys[ssid]) && false) \
548 /* iterate across the hierarchies */
549 #define for_each_root(root) \
550 list_for_each_entry((root), &cgroup_roots, root_list)
552 /* iterate over child cgrps, lock should be held throughout iteration */
553 #define cgroup_for_each_live_child(child, cgrp) \
554 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
555 if (({ lockdep_assert_held(&cgroup_mutex); \
556 cgroup_is_dead(child); })) \
560 static void cgroup_release_agent(struct work_struct
*work
);
561 static void check_for_release(struct cgroup
*cgrp
);
564 * A cgroup can be associated with multiple css_sets as different tasks may
565 * belong to different cgroups on different hierarchies. In the other
566 * direction, a css_set is naturally associated with multiple cgroups.
567 * This M:N relationship is represented by the following link structure
568 * which exists for each association and allows traversing the associations
571 struct cgrp_cset_link
{
572 /* the cgroup and css_set this link associates */
574 struct css_set
*cset
;
576 /* list of cgrp_cset_links anchored at cgrp->cset_links */
577 struct list_head cset_link
;
579 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
580 struct list_head cgrp_link
;
584 * The default css_set - used by init and its children prior to any
585 * hierarchies being mounted. It contains a pointer to the root state
586 * for each subsystem. Also used to anchor the list of css_sets. Not
587 * reference-counted, to improve performance when child cgroups
588 * haven't been created.
590 struct css_set init_css_set
= {
591 .refcount
= ATOMIC_INIT(1),
592 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
593 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
594 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
595 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
596 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
597 .task_iters
= LIST_HEAD_INIT(init_css_set
.task_iters
),
600 static int css_set_count
= 1; /* 1 for init_css_set */
603 * css_set_populated - does a css_set contain any tasks?
604 * @cset: target css_set
606 static bool css_set_populated(struct css_set
*cset
)
608 lockdep_assert_held(&css_set_rwsem
);
610 return !list_empty(&cset
->tasks
) || !list_empty(&cset
->mg_tasks
);
614 * cgroup_update_populated - updated populated count of a cgroup
615 * @cgrp: the target cgroup
616 * @populated: inc or dec populated count
618 * One of the css_sets associated with @cgrp is either getting its first
619 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
620 * count is propagated towards root so that a given cgroup's populated_cnt
621 * is zero iff the cgroup and all its descendants don't contain any tasks.
623 * @cgrp's interface file "cgroup.populated" is zero if
624 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
625 * changes from or to zero, userland is notified that the content of the
626 * interface file has changed. This can be used to detect when @cgrp and
627 * its descendants become populated or empty.
629 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
631 lockdep_assert_held(&css_set_rwsem
);
637 trigger
= !cgrp
->populated_cnt
++;
639 trigger
= !--cgrp
->populated_cnt
;
644 check_for_release(cgrp
);
645 cgroup_file_notify(&cgrp
->events_file
);
647 cgrp
= cgroup_parent(cgrp
);
652 * css_set_update_populated - update populated state of a css_set
653 * @cset: target css_set
654 * @populated: whether @cset is populated or depopulated
656 * @cset is either getting the first task or losing the last. Update the
657 * ->populated_cnt of all associated cgroups accordingly.
659 static void css_set_update_populated(struct css_set
*cset
, bool populated
)
661 struct cgrp_cset_link
*link
;
663 lockdep_assert_held(&css_set_rwsem
);
665 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
)
666 cgroup_update_populated(link
->cgrp
, populated
);
670 * css_set_move_task - move a task from one css_set to another
671 * @task: task being moved
672 * @from_cset: css_set @task currently belongs to (may be NULL)
673 * @to_cset: new css_set @task is being moved to (may be NULL)
674 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
676 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
677 * css_set, @from_cset can be NULL. If @task is being disassociated
678 * instead of moved, @to_cset can be NULL.
680 * This function automatically handles populated_cnt updates and
681 * css_task_iter adjustments but the caller is responsible for managing
682 * @from_cset and @to_cset's reference counts.
684 static void css_set_move_task(struct task_struct
*task
,
685 struct css_set
*from_cset
, struct css_set
*to_cset
,
688 lockdep_assert_held(&css_set_rwsem
);
691 struct css_task_iter
*it
, *pos
;
693 WARN_ON_ONCE(list_empty(&task
->cg_list
));
696 * @task is leaving, advance task iterators which are
697 * pointing to it so that they can resume at the next
698 * position. Advancing an iterator might remove it from
699 * the list, use safe walk. See css_task_iter_advance*()
702 list_for_each_entry_safe(it
, pos
, &from_cset
->task_iters
,
704 if (it
->task_pos
== &task
->cg_list
)
705 css_task_iter_advance(it
);
707 list_del_init(&task
->cg_list
);
708 if (!css_set_populated(from_cset
))
709 css_set_update_populated(from_cset
, false);
711 WARN_ON_ONCE(!list_empty(&task
->cg_list
));
716 * We are synchronized through cgroup_threadgroup_rwsem
717 * against PF_EXITING setting such that we can't race
718 * against cgroup_exit() changing the css_set to
719 * init_css_set and dropping the old one.
721 WARN_ON_ONCE(task
->flags
& PF_EXITING
);
723 if (!css_set_populated(to_cset
))
724 css_set_update_populated(to_cset
, true);
725 rcu_assign_pointer(task
->cgroups
, to_cset
);
726 list_add_tail(&task
->cg_list
, use_mg_tasks
? &to_cset
->mg_tasks
:
732 * hash table for cgroup groups. This improves the performance to find
733 * an existing css_set. This hash doesn't (currently) take into
734 * account cgroups in empty hierarchies.
736 #define CSS_SET_HASH_BITS 7
737 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
739 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
741 unsigned long key
= 0UL;
742 struct cgroup_subsys
*ss
;
745 for_each_subsys(ss
, i
)
746 key
+= (unsigned long)css
[i
];
747 key
= (key
>> 16) ^ key
;
752 static void put_css_set_locked(struct css_set
*cset
)
754 struct cgrp_cset_link
*link
, *tmp_link
;
755 struct cgroup_subsys
*ss
;
758 lockdep_assert_held(&css_set_rwsem
);
760 if (!atomic_dec_and_test(&cset
->refcount
))
763 /* This css_set is dead. unlink it and release cgroup refcounts */
764 for_each_subsys(ss
, ssid
)
765 list_del(&cset
->e_cset_node
[ssid
]);
766 hash_del(&cset
->hlist
);
769 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
770 list_del(&link
->cset_link
);
771 list_del(&link
->cgrp_link
);
772 if (cgroup_parent(link
->cgrp
))
773 cgroup_put(link
->cgrp
);
777 kfree_rcu(cset
, rcu_head
);
780 static void put_css_set(struct css_set
*cset
)
783 * Ensure that the refcount doesn't hit zero while any readers
784 * can see it. Similar to atomic_dec_and_lock(), but for an
787 if (atomic_add_unless(&cset
->refcount
, -1, 1))
790 down_write(&css_set_rwsem
);
791 put_css_set_locked(cset
);
792 up_write(&css_set_rwsem
);
796 * refcounted get/put for css_set objects
798 static inline void get_css_set(struct css_set
*cset
)
800 atomic_inc(&cset
->refcount
);
804 * compare_css_sets - helper function for find_existing_css_set().
805 * @cset: candidate css_set being tested
806 * @old_cset: existing css_set for a task
807 * @new_cgrp: cgroup that's being entered by the task
808 * @template: desired set of css pointers in css_set (pre-calculated)
810 * Returns true if "cset" matches "old_cset" except for the hierarchy
811 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
813 static bool compare_css_sets(struct css_set
*cset
,
814 struct css_set
*old_cset
,
815 struct cgroup
*new_cgrp
,
816 struct cgroup_subsys_state
*template[])
818 struct list_head
*l1
, *l2
;
821 * On the default hierarchy, there can be csets which are
822 * associated with the same set of cgroups but different csses.
823 * Let's first ensure that csses match.
825 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
829 * Compare cgroup pointers in order to distinguish between
830 * different cgroups in hierarchies. As different cgroups may
831 * share the same effective css, this comparison is always
834 l1
= &cset
->cgrp_links
;
835 l2
= &old_cset
->cgrp_links
;
837 struct cgrp_cset_link
*link1
, *link2
;
838 struct cgroup
*cgrp1
, *cgrp2
;
842 /* See if we reached the end - both lists are equal length. */
843 if (l1
== &cset
->cgrp_links
) {
844 BUG_ON(l2
!= &old_cset
->cgrp_links
);
847 BUG_ON(l2
== &old_cset
->cgrp_links
);
849 /* Locate the cgroups associated with these links. */
850 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
851 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
854 /* Hierarchies should be linked in the same order. */
855 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
858 * If this hierarchy is the hierarchy of the cgroup
859 * that's changing, then we need to check that this
860 * css_set points to the new cgroup; if it's any other
861 * hierarchy, then this css_set should point to the
862 * same cgroup as the old css_set.
864 if (cgrp1
->root
== new_cgrp
->root
) {
865 if (cgrp1
!= new_cgrp
)
876 * find_existing_css_set - init css array and find the matching css_set
877 * @old_cset: the css_set that we're using before the cgroup transition
878 * @cgrp: the cgroup that we're moving into
879 * @template: out param for the new set of csses, should be clear on entry
881 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
883 struct cgroup_subsys_state
*template[])
885 struct cgroup_root
*root
= cgrp
->root
;
886 struct cgroup_subsys
*ss
;
887 struct css_set
*cset
;
892 * Build the set of subsystem state objects that we want to see in the
893 * new css_set. while subsystems can change globally, the entries here
894 * won't change, so no need for locking.
896 for_each_subsys(ss
, i
) {
897 if (root
->subsys_mask
& (1UL << i
)) {
899 * @ss is in this hierarchy, so we want the
900 * effective css from @cgrp.
902 template[i
] = cgroup_e_css(cgrp
, ss
);
905 * @ss is not in this hierarchy, so we don't want
908 template[i
] = old_cset
->subsys
[i
];
912 key
= css_set_hash(template);
913 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
914 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
917 /* This css_set matches what we need */
921 /* No existing cgroup group matched */
925 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
927 struct cgrp_cset_link
*link
, *tmp_link
;
929 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
930 list_del(&link
->cset_link
);
936 * allocate_cgrp_cset_links - allocate cgrp_cset_links
937 * @count: the number of links to allocate
938 * @tmp_links: list_head the allocated links are put on
940 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
941 * through ->cset_link. Returns 0 on success or -errno.
943 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
945 struct cgrp_cset_link
*link
;
948 INIT_LIST_HEAD(tmp_links
);
950 for (i
= 0; i
< count
; i
++) {
951 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
953 free_cgrp_cset_links(tmp_links
);
956 list_add(&link
->cset_link
, tmp_links
);
962 * link_css_set - a helper function to link a css_set to a cgroup
963 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
964 * @cset: the css_set to be linked
965 * @cgrp: the destination cgroup
967 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
970 struct cgrp_cset_link
*link
;
972 BUG_ON(list_empty(tmp_links
));
974 if (cgroup_on_dfl(cgrp
))
975 cset
->dfl_cgrp
= cgrp
;
977 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
982 * Always add links to the tail of the lists so that the lists are
983 * in choronological order.
985 list_move_tail(&link
->cset_link
, &cgrp
->cset_links
);
986 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
988 if (cgroup_parent(cgrp
))
993 * find_css_set - return a new css_set with one cgroup updated
994 * @old_cset: the baseline css_set
995 * @cgrp: the cgroup to be updated
997 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
998 * substituted into the appropriate hierarchy.
1000 static struct css_set
*find_css_set(struct css_set
*old_cset
,
1001 struct cgroup
*cgrp
)
1003 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
1004 struct css_set
*cset
;
1005 struct list_head tmp_links
;
1006 struct cgrp_cset_link
*link
;
1007 struct cgroup_subsys
*ss
;
1011 lockdep_assert_held(&cgroup_mutex
);
1013 /* First see if we already have a cgroup group that matches
1014 * the desired set */
1015 down_read(&css_set_rwsem
);
1016 cset
= find_existing_css_set(old_cset
, cgrp
, template);
1019 up_read(&css_set_rwsem
);
1024 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
1028 /* Allocate all the cgrp_cset_link objects that we'll need */
1029 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
1034 atomic_set(&cset
->refcount
, 1);
1035 INIT_LIST_HEAD(&cset
->cgrp_links
);
1036 INIT_LIST_HEAD(&cset
->tasks
);
1037 INIT_LIST_HEAD(&cset
->mg_tasks
);
1038 INIT_LIST_HEAD(&cset
->mg_preload_node
);
1039 INIT_LIST_HEAD(&cset
->mg_node
);
1040 INIT_LIST_HEAD(&cset
->task_iters
);
1041 INIT_HLIST_NODE(&cset
->hlist
);
1043 /* Copy the set of subsystem state objects generated in
1044 * find_existing_css_set() */
1045 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
1047 down_write(&css_set_rwsem
);
1048 /* Add reference counts and links from the new css_set. */
1049 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
1050 struct cgroup
*c
= link
->cgrp
;
1052 if (c
->root
== cgrp
->root
)
1054 link_css_set(&tmp_links
, cset
, c
);
1057 BUG_ON(!list_empty(&tmp_links
));
1061 /* Add @cset to the hash table */
1062 key
= css_set_hash(cset
->subsys
);
1063 hash_add(css_set_table
, &cset
->hlist
, key
);
1065 for_each_subsys(ss
, ssid
)
1066 list_add_tail(&cset
->e_cset_node
[ssid
],
1067 &cset
->subsys
[ssid
]->cgroup
->e_csets
[ssid
]);
1069 up_write(&css_set_rwsem
);
1074 static struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
1076 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
1078 return root_cgrp
->root
;
1081 static int cgroup_init_root_id(struct cgroup_root
*root
)
1085 lockdep_assert_held(&cgroup_mutex
);
1087 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
1091 root
->hierarchy_id
= id
;
1095 static void cgroup_exit_root_id(struct cgroup_root
*root
)
1097 lockdep_assert_held(&cgroup_mutex
);
1099 if (root
->hierarchy_id
) {
1100 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
1101 root
->hierarchy_id
= 0;
1105 static void cgroup_free_root(struct cgroup_root
*root
)
1108 /* hierarchy ID should already have been released */
1109 WARN_ON_ONCE(root
->hierarchy_id
);
1111 idr_destroy(&root
->cgroup_idr
);
1116 static void cgroup_destroy_root(struct cgroup_root
*root
)
1118 struct cgroup
*cgrp
= &root
->cgrp
;
1119 struct cgrp_cset_link
*link
, *tmp_link
;
1121 mutex_lock(&cgroup_mutex
);
1123 BUG_ON(atomic_read(&root
->nr_cgrps
));
1124 BUG_ON(!list_empty(&cgrp
->self
.children
));
1126 /* Rebind all subsystems back to the default hierarchy */
1127 rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
);
1130 * Release all the links from cset_links to this hierarchy's
1133 down_write(&css_set_rwsem
);
1135 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
1136 list_del(&link
->cset_link
);
1137 list_del(&link
->cgrp_link
);
1140 up_write(&css_set_rwsem
);
1142 if (!list_empty(&root
->root_list
)) {
1143 list_del(&root
->root_list
);
1144 cgroup_root_count
--;
1147 cgroup_exit_root_id(root
);
1149 mutex_unlock(&cgroup_mutex
);
1151 kernfs_destroy_root(root
->kf_root
);
1152 cgroup_free_root(root
);
1155 /* look up cgroup associated with given css_set on the specified hierarchy */
1156 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
1157 struct cgroup_root
*root
)
1159 struct cgroup
*res
= NULL
;
1161 lockdep_assert_held(&cgroup_mutex
);
1162 lockdep_assert_held(&css_set_rwsem
);
1164 if (cset
== &init_css_set
) {
1167 struct cgrp_cset_link
*link
;
1169 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1170 struct cgroup
*c
= link
->cgrp
;
1172 if (c
->root
== root
) {
1184 * Return the cgroup for "task" from the given hierarchy. Must be
1185 * called with cgroup_mutex and css_set_rwsem held.
1187 static struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
1188 struct cgroup_root
*root
)
1191 * No need to lock the task - since we hold cgroup_mutex the
1192 * task can't change groups, so the only thing that can happen
1193 * is that it exits and its css is set back to init_css_set.
1195 return cset_cgroup_from_root(task_css_set(task
), root
);
1199 * A task must hold cgroup_mutex to modify cgroups.
1201 * Any task can increment and decrement the count field without lock.
1202 * So in general, code holding cgroup_mutex can't rely on the count
1203 * field not changing. However, if the count goes to zero, then only
1204 * cgroup_attach_task() can increment it again. Because a count of zero
1205 * means that no tasks are currently attached, therefore there is no
1206 * way a task attached to that cgroup can fork (the other way to
1207 * increment the count). So code holding cgroup_mutex can safely
1208 * assume that if the count is zero, it will stay zero. Similarly, if
1209 * a task holds cgroup_mutex on a cgroup with zero count, it
1210 * knows that the cgroup won't be removed, as cgroup_rmdir()
1213 * A cgroup can only be deleted if both its 'count' of using tasks
1214 * is zero, and its list of 'children' cgroups is empty. Since all
1215 * tasks in the system use _some_ cgroup, and since there is always at
1216 * least one task in the system (init, pid == 1), therefore, root cgroup
1217 * always has either children cgroups and/or using tasks. So we don't
1218 * need a special hack to ensure that root cgroup cannot be deleted.
1220 * P.S. One more locking exception. RCU is used to guard the
1221 * update of a tasks cgroup pointer by cgroup_attach_task()
1224 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1225 static const struct file_operations proc_cgroupstats_operations
;
1227 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1230 struct cgroup_subsys
*ss
= cft
->ss
;
1232 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1233 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
1234 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
1235 cgroup_on_dfl(cgrp
) ? ss
->name
: ss
->legacy_name
,
1238 strncpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1243 * cgroup_file_mode - deduce file mode of a control file
1244 * @cft: the control file in question
1246 * S_IRUGO for read, S_IWUSR for write.
1248 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1252 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1255 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
) {
1256 if (cft
->flags
& CFTYPE_WORLD_WRITABLE
)
1266 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1267 * @cgrp: the target cgroup
1268 * @subtree_control: the new subtree_control mask to consider
1270 * On the default hierarchy, a subsystem may request other subsystems to be
1271 * enabled together through its ->depends_on mask. In such cases, more
1272 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1274 * This function calculates which subsystems need to be enabled if
1275 * @subtree_control is to be applied to @cgrp. The returned mask is always
1276 * a superset of @subtree_control and follows the usual hierarchy rules.
1278 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup
*cgrp
,
1279 unsigned long subtree_control
)
1281 struct cgroup
*parent
= cgroup_parent(cgrp
);
1282 unsigned long cur_ss_mask
= subtree_control
;
1283 struct cgroup_subsys
*ss
;
1286 lockdep_assert_held(&cgroup_mutex
);
1288 if (!cgroup_on_dfl(cgrp
))
1292 unsigned long new_ss_mask
= cur_ss_mask
;
1294 for_each_subsys_which(ss
, ssid
, &cur_ss_mask
)
1295 new_ss_mask
|= ss
->depends_on
;
1298 * Mask out subsystems which aren't available. This can
1299 * happen only if some depended-upon subsystems were bound
1300 * to non-default hierarchies.
1303 new_ss_mask
&= parent
->child_subsys_mask
;
1305 new_ss_mask
&= cgrp
->root
->subsys_mask
;
1307 if (new_ss_mask
== cur_ss_mask
)
1309 cur_ss_mask
= new_ss_mask
;
1316 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1317 * @cgrp: the target cgroup
1319 * Update @cgrp->child_subsys_mask according to the current
1320 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1322 static void cgroup_refresh_child_subsys_mask(struct cgroup
*cgrp
)
1324 cgrp
->child_subsys_mask
=
1325 cgroup_calc_child_subsys_mask(cgrp
, cgrp
->subtree_control
);
1329 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1330 * @kn: the kernfs_node being serviced
1332 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1333 * the method finishes if locking succeeded. Note that once this function
1334 * returns the cgroup returned by cgroup_kn_lock_live() may become
1335 * inaccessible any time. If the caller intends to continue to access the
1336 * cgroup, it should pin it before invoking this function.
1338 static void cgroup_kn_unlock(struct kernfs_node
*kn
)
1340 struct cgroup
*cgrp
;
1342 if (kernfs_type(kn
) == KERNFS_DIR
)
1345 cgrp
= kn
->parent
->priv
;
1347 mutex_unlock(&cgroup_mutex
);
1349 kernfs_unbreak_active_protection(kn
);
1354 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1355 * @kn: the kernfs_node being serviced
1357 * This helper is to be used by a cgroup kernfs method currently servicing
1358 * @kn. It breaks the active protection, performs cgroup locking and
1359 * verifies that the associated cgroup is alive. Returns the cgroup if
1360 * alive; otherwise, %NULL. A successful return should be undone by a
1361 * matching cgroup_kn_unlock() invocation.
1363 * Any cgroup kernfs method implementation which requires locking the
1364 * associated cgroup should use this helper. It avoids nesting cgroup
1365 * locking under kernfs active protection and allows all kernfs operations
1366 * including self-removal.
1368 static struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
)
1370 struct cgroup
*cgrp
;
1372 if (kernfs_type(kn
) == KERNFS_DIR
)
1375 cgrp
= kn
->parent
->priv
;
1378 * We're gonna grab cgroup_mutex which nests outside kernfs
1379 * active_ref. cgroup liveliness check alone provides enough
1380 * protection against removal. Ensure @cgrp stays accessible and
1381 * break the active_ref protection.
1383 if (!cgroup_tryget(cgrp
))
1385 kernfs_break_active_protection(kn
);
1387 mutex_lock(&cgroup_mutex
);
1389 if (!cgroup_is_dead(cgrp
))
1392 cgroup_kn_unlock(kn
);
1396 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1398 char name
[CGROUP_FILE_NAME_MAX
];
1400 lockdep_assert_held(&cgroup_mutex
);
1401 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1405 * css_clear_dir - remove subsys files in a cgroup directory
1407 * @cgrp_override: specify if target cgroup is different from css->cgroup
1409 static void css_clear_dir(struct cgroup_subsys_state
*css
,
1410 struct cgroup
*cgrp_override
)
1412 struct cgroup
*cgrp
= cgrp_override
?: css
->cgroup
;
1413 struct cftype
*cfts
;
1415 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
)
1416 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1420 * css_populate_dir - create subsys files in a cgroup directory
1422 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1424 * On failure, no file is added.
1426 static int css_populate_dir(struct cgroup_subsys_state
*css
,
1427 struct cgroup
*cgrp_override
)
1429 struct cgroup
*cgrp
= cgrp_override
?: css
->cgroup
;
1430 struct cftype
*cfts
, *failed_cfts
;
1434 if (cgroup_on_dfl(cgrp
))
1435 cfts
= cgroup_dfl_base_files
;
1437 cfts
= cgroup_legacy_base_files
;
1439 return cgroup_addrm_files(&cgrp
->self
, cgrp
, cfts
, true);
1442 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1443 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, true);
1451 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1452 if (cfts
== failed_cfts
)
1454 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1459 static int rebind_subsystems(struct cgroup_root
*dst_root
,
1460 unsigned long ss_mask
)
1462 struct cgroup
*dcgrp
= &dst_root
->cgrp
;
1463 struct cgroup_subsys
*ss
;
1464 unsigned long tmp_ss_mask
;
1467 lockdep_assert_held(&cgroup_mutex
);
1469 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
1470 /* if @ss has non-root csses attached to it, can't move */
1471 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)))
1474 /* can't move between two non-dummy roots either */
1475 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1479 /* skip creating root files on dfl_root for inhibited subsystems */
1480 tmp_ss_mask
= ss_mask
;
1481 if (dst_root
== &cgrp_dfl_root
)
1482 tmp_ss_mask
&= ~cgrp_dfl_root_inhibit_ss_mask
;
1484 for_each_subsys_which(ss
, ssid
, &tmp_ss_mask
) {
1485 struct cgroup
*scgrp
= &ss
->root
->cgrp
;
1488 ret
= css_populate_dir(cgroup_css(scgrp
, ss
), dcgrp
);
1493 * Rebinding back to the default root is not allowed to
1494 * fail. Using both default and non-default roots should
1495 * be rare. Moving subsystems back and forth even more so.
1496 * Just warn about it and continue.
1498 if (dst_root
== &cgrp_dfl_root
) {
1499 if (cgrp_dfl_root_visible
) {
1500 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1502 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1507 for_each_subsys_which(ss
, tssid
, &tmp_ss_mask
) {
1510 css_clear_dir(cgroup_css(scgrp
, ss
), dcgrp
);
1516 * Nothing can fail from this point on. Remove files for the
1517 * removed subsystems and rebind each subsystem.
1519 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
1520 struct cgroup_root
*src_root
= ss
->root
;
1521 struct cgroup
*scgrp
= &src_root
->cgrp
;
1522 struct cgroup_subsys_state
*css
= cgroup_css(scgrp
, ss
);
1523 struct css_set
*cset
;
1525 WARN_ON(!css
|| cgroup_css(dcgrp
, ss
));
1527 css_clear_dir(css
, NULL
);
1529 RCU_INIT_POINTER(scgrp
->subsys
[ssid
], NULL
);
1530 rcu_assign_pointer(dcgrp
->subsys
[ssid
], css
);
1531 ss
->root
= dst_root
;
1532 css
->cgroup
= dcgrp
;
1534 down_write(&css_set_rwsem
);
1535 hash_for_each(css_set_table
, i
, cset
, hlist
)
1536 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1537 &dcgrp
->e_csets
[ss
->id
]);
1538 up_write(&css_set_rwsem
);
1540 src_root
->subsys_mask
&= ~(1 << ssid
);
1541 scgrp
->subtree_control
&= ~(1 << ssid
);
1542 cgroup_refresh_child_subsys_mask(scgrp
);
1544 /* default hierarchy doesn't enable controllers by default */
1545 dst_root
->subsys_mask
|= 1 << ssid
;
1546 if (dst_root
== &cgrp_dfl_root
) {
1547 static_branch_enable(cgroup_subsys_on_dfl_key
[ssid
]);
1549 dcgrp
->subtree_control
|= 1 << ssid
;
1550 cgroup_refresh_child_subsys_mask(dcgrp
);
1551 static_branch_disable(cgroup_subsys_on_dfl_key
[ssid
]);
1558 kernfs_activate(dcgrp
->kn
);
1562 static int cgroup_show_options(struct seq_file
*seq
,
1563 struct kernfs_root
*kf_root
)
1565 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1566 struct cgroup_subsys
*ss
;
1569 if (root
!= &cgrp_dfl_root
)
1570 for_each_subsys(ss
, ssid
)
1571 if (root
->subsys_mask
& (1 << ssid
))
1572 seq_show_option(seq
, ss
->legacy_name
, NULL
);
1573 if (root
->flags
& CGRP_ROOT_NOPREFIX
)
1574 seq_puts(seq
, ",noprefix");
1575 if (root
->flags
& CGRP_ROOT_XATTR
)
1576 seq_puts(seq
, ",xattr");
1578 spin_lock(&release_agent_path_lock
);
1579 if (strlen(root
->release_agent_path
))
1580 seq_show_option(seq
, "release_agent",
1581 root
->release_agent_path
);
1582 spin_unlock(&release_agent_path_lock
);
1584 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
))
1585 seq_puts(seq
, ",clone_children");
1586 if (strlen(root
->name
))
1587 seq_show_option(seq
, "name", root
->name
);
1591 struct cgroup_sb_opts
{
1592 unsigned long subsys_mask
;
1594 char *release_agent
;
1595 bool cpuset_clone_children
;
1597 /* User explicitly requested empty subsystem */
1601 static int parse_cgroupfs_options(char *data
, struct cgroup_sb_opts
*opts
)
1603 char *token
, *o
= data
;
1604 bool all_ss
= false, one_ss
= false;
1605 unsigned long mask
= -1UL;
1606 struct cgroup_subsys
*ss
;
1610 #ifdef CONFIG_CPUSETS
1611 mask
= ~(1U << cpuset_cgrp_id
);
1614 memset(opts
, 0, sizeof(*opts
));
1616 while ((token
= strsep(&o
, ",")) != NULL
) {
1621 if (!strcmp(token
, "none")) {
1622 /* Explicitly have no subsystems */
1626 if (!strcmp(token
, "all")) {
1627 /* Mutually exclusive option 'all' + subsystem name */
1633 if (!strcmp(token
, "__DEVEL__sane_behavior")) {
1634 opts
->flags
|= CGRP_ROOT_SANE_BEHAVIOR
;
1637 if (!strcmp(token
, "noprefix")) {
1638 opts
->flags
|= CGRP_ROOT_NOPREFIX
;
1641 if (!strcmp(token
, "clone_children")) {
1642 opts
->cpuset_clone_children
= true;
1645 if (!strcmp(token
, "xattr")) {
1646 opts
->flags
|= CGRP_ROOT_XATTR
;
1649 if (!strncmp(token
, "release_agent=", 14)) {
1650 /* Specifying two release agents is forbidden */
1651 if (opts
->release_agent
)
1653 opts
->release_agent
=
1654 kstrndup(token
+ 14, PATH_MAX
- 1, GFP_KERNEL
);
1655 if (!opts
->release_agent
)
1659 if (!strncmp(token
, "name=", 5)) {
1660 const char *name
= token
+ 5;
1661 /* Can't specify an empty name */
1664 /* Must match [\w.-]+ */
1665 for (i
= 0; i
< strlen(name
); i
++) {
1669 if ((c
== '.') || (c
== '-') || (c
== '_'))
1673 /* Specifying two names is forbidden */
1676 opts
->name
= kstrndup(name
,
1677 MAX_CGROUP_ROOT_NAMELEN
- 1,
1685 for_each_subsys(ss
, i
) {
1686 if (strcmp(token
, ss
->legacy_name
))
1688 if (!cgroup_ssid_enabled(i
))
1691 /* Mutually exclusive option 'all' + subsystem name */
1694 opts
->subsys_mask
|= (1 << i
);
1699 if (i
== CGROUP_SUBSYS_COUNT
)
1703 if (opts
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1704 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1706 pr_err("sane_behavior: no other mount options allowed\n");
1713 * If the 'all' option was specified select all the subsystems,
1714 * otherwise if 'none', 'name=' and a subsystem name options were
1715 * not specified, let's default to 'all'
1717 if (all_ss
|| (!one_ss
&& !opts
->none
&& !opts
->name
))
1718 for_each_subsys(ss
, i
)
1719 if (cgroup_ssid_enabled(i
))
1720 opts
->subsys_mask
|= (1 << i
);
1723 * We either have to specify by name or by subsystems. (So all
1724 * empty hierarchies must have a name).
1726 if (!opts
->subsys_mask
&& !opts
->name
)
1730 * Option noprefix was introduced just for backward compatibility
1731 * with the old cpuset, so we allow noprefix only if mounting just
1732 * the cpuset subsystem.
1734 if ((opts
->flags
& CGRP_ROOT_NOPREFIX
) && (opts
->subsys_mask
& mask
))
1737 /* Can't specify "none" and some subsystems */
1738 if (opts
->subsys_mask
&& opts
->none
)
1744 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1747 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1748 struct cgroup_sb_opts opts
;
1749 unsigned long added_mask
, removed_mask
;
1751 if (root
== &cgrp_dfl_root
) {
1752 pr_err("remount is not allowed\n");
1756 mutex_lock(&cgroup_mutex
);
1758 /* See what subsystems are wanted */
1759 ret
= parse_cgroupfs_options(data
, &opts
);
1763 if (opts
.subsys_mask
!= root
->subsys_mask
|| opts
.release_agent
)
1764 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1765 task_tgid_nr(current
), current
->comm
);
1767 added_mask
= opts
.subsys_mask
& ~root
->subsys_mask
;
1768 removed_mask
= root
->subsys_mask
& ~opts
.subsys_mask
;
1770 /* Don't allow flags or name to change at remount */
1771 if ((opts
.flags
^ root
->flags
) ||
1772 (opts
.name
&& strcmp(opts
.name
, root
->name
))) {
1773 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1774 opts
.flags
, opts
.name
?: "", root
->flags
, root
->name
);
1779 /* remounting is not allowed for populated hierarchies */
1780 if (!list_empty(&root
->cgrp
.self
.children
)) {
1785 ret
= rebind_subsystems(root
, added_mask
);
1789 rebind_subsystems(&cgrp_dfl_root
, removed_mask
);
1791 if (opts
.release_agent
) {
1792 spin_lock(&release_agent_path_lock
);
1793 strcpy(root
->release_agent_path
, opts
.release_agent
);
1794 spin_unlock(&release_agent_path_lock
);
1797 kfree(opts
.release_agent
);
1799 mutex_unlock(&cgroup_mutex
);
1804 * To reduce the fork() overhead for systems that are not actually using
1805 * their cgroups capability, we don't maintain the lists running through
1806 * each css_set to its tasks until we see the list actually used - in other
1807 * words after the first mount.
1809 static bool use_task_css_set_links __read_mostly
;
1811 static void cgroup_enable_task_cg_lists(void)
1813 struct task_struct
*p
, *g
;
1815 down_write(&css_set_rwsem
);
1817 if (use_task_css_set_links
)
1820 use_task_css_set_links
= true;
1823 * We need tasklist_lock because RCU is not safe against
1824 * while_each_thread(). Besides, a forking task that has passed
1825 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1826 * is not guaranteed to have its child immediately visible in the
1827 * tasklist if we walk through it with RCU.
1829 read_lock(&tasklist_lock
);
1830 do_each_thread(g
, p
) {
1831 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1832 task_css_set(p
) != &init_css_set
);
1835 * We should check if the process is exiting, otherwise
1836 * it will race with cgroup_exit() in that the list
1837 * entry won't be deleted though the process has exited.
1838 * Do it while holding siglock so that we don't end up
1839 * racing against cgroup_exit().
1841 spin_lock_irq(&p
->sighand
->siglock
);
1842 if (!(p
->flags
& PF_EXITING
)) {
1843 struct css_set
*cset
= task_css_set(p
);
1845 if (!css_set_populated(cset
))
1846 css_set_update_populated(cset
, true);
1847 list_add_tail(&p
->cg_list
, &cset
->tasks
);
1850 spin_unlock_irq(&p
->sighand
->siglock
);
1851 } while_each_thread(g
, p
);
1852 read_unlock(&tasklist_lock
);
1854 up_write(&css_set_rwsem
);
1857 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1859 struct cgroup_subsys
*ss
;
1862 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1863 INIT_LIST_HEAD(&cgrp
->self
.children
);
1864 INIT_LIST_HEAD(&cgrp
->self
.files
);
1865 INIT_LIST_HEAD(&cgrp
->cset_links
);
1866 INIT_LIST_HEAD(&cgrp
->pidlists
);
1867 mutex_init(&cgrp
->pidlist_mutex
);
1868 cgrp
->self
.cgroup
= cgrp
;
1869 cgrp
->self
.flags
|= CSS_ONLINE
;
1871 for_each_subsys(ss
, ssid
)
1872 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1874 init_waitqueue_head(&cgrp
->offline_waitq
);
1875 INIT_WORK(&cgrp
->release_agent_work
, cgroup_release_agent
);
1878 static void init_cgroup_root(struct cgroup_root
*root
,
1879 struct cgroup_sb_opts
*opts
)
1881 struct cgroup
*cgrp
= &root
->cgrp
;
1883 INIT_LIST_HEAD(&root
->root_list
);
1884 atomic_set(&root
->nr_cgrps
, 1);
1886 init_cgroup_housekeeping(cgrp
);
1887 idr_init(&root
->cgroup_idr
);
1889 root
->flags
= opts
->flags
;
1890 if (opts
->release_agent
)
1891 strcpy(root
->release_agent_path
, opts
->release_agent
);
1893 strcpy(root
->name
, opts
->name
);
1894 if (opts
->cpuset_clone_children
)
1895 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1898 static int cgroup_setup_root(struct cgroup_root
*root
, unsigned long ss_mask
)
1900 LIST_HEAD(tmp_links
);
1901 struct cgroup
*root_cgrp
= &root
->cgrp
;
1902 struct css_set
*cset
;
1905 lockdep_assert_held(&cgroup_mutex
);
1907 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_KERNEL
);
1910 root_cgrp
->id
= ret
;
1912 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
, 0,
1918 * We're accessing css_set_count without locking css_set_rwsem here,
1919 * but that's OK - it can only be increased by someone holding
1920 * cgroup_lock, and that's us. The worst that can happen is that we
1921 * have some link structures left over
1923 ret
= allocate_cgrp_cset_links(css_set_count
, &tmp_links
);
1927 ret
= cgroup_init_root_id(root
);
1931 root
->kf_root
= kernfs_create_root(&cgroup_kf_syscall_ops
,
1932 KERNFS_ROOT_CREATE_DEACTIVATED
,
1934 if (IS_ERR(root
->kf_root
)) {
1935 ret
= PTR_ERR(root
->kf_root
);
1938 root_cgrp
->kn
= root
->kf_root
->kn
;
1940 ret
= css_populate_dir(&root_cgrp
->self
, NULL
);
1944 ret
= rebind_subsystems(root
, ss_mask
);
1949 * There must be no failure case after here, since rebinding takes
1950 * care of subsystems' refcounts, which are explicitly dropped in
1951 * the failure exit path.
1953 list_add(&root
->root_list
, &cgroup_roots
);
1954 cgroup_root_count
++;
1957 * Link the root cgroup in this hierarchy into all the css_set
1960 down_write(&css_set_rwsem
);
1961 hash_for_each(css_set_table
, i
, cset
, hlist
) {
1962 link_css_set(&tmp_links
, cset
, root_cgrp
);
1963 if (css_set_populated(cset
))
1964 cgroup_update_populated(root_cgrp
, true);
1966 up_write(&css_set_rwsem
);
1968 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
1969 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1971 kernfs_activate(root_cgrp
->kn
);
1976 kernfs_destroy_root(root
->kf_root
);
1977 root
->kf_root
= NULL
;
1979 cgroup_exit_root_id(root
);
1981 percpu_ref_exit(&root_cgrp
->self
.refcnt
);
1983 free_cgrp_cset_links(&tmp_links
);
1987 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
1988 int flags
, const char *unused_dev_name
,
1991 struct super_block
*pinned_sb
= NULL
;
1992 struct cgroup_subsys
*ss
;
1993 struct cgroup_root
*root
;
1994 struct cgroup_sb_opts opts
;
1995 struct dentry
*dentry
;
2001 * The first time anyone tries to mount a cgroup, enable the list
2002 * linking each css_set to its tasks and fix up all existing tasks.
2004 if (!use_task_css_set_links
)
2005 cgroup_enable_task_cg_lists();
2007 mutex_lock(&cgroup_mutex
);
2009 /* First find the desired set of subsystems */
2010 ret
= parse_cgroupfs_options(data
, &opts
);
2014 /* look for a matching existing root */
2015 if (opts
.flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
2016 cgrp_dfl_root_visible
= true;
2017 root
= &cgrp_dfl_root
;
2018 cgroup_get(&root
->cgrp
);
2024 * Destruction of cgroup root is asynchronous, so subsystems may
2025 * still be dying after the previous unmount. Let's drain the
2026 * dying subsystems. We just need to ensure that the ones
2027 * unmounted previously finish dying and don't care about new ones
2028 * starting. Testing ref liveliness is good enough.
2030 for_each_subsys(ss
, i
) {
2031 if (!(opts
.subsys_mask
& (1 << i
)) ||
2032 ss
->root
== &cgrp_dfl_root
)
2035 if (!percpu_ref_tryget_live(&ss
->root
->cgrp
.self
.refcnt
)) {
2036 mutex_unlock(&cgroup_mutex
);
2038 ret
= restart_syscall();
2041 cgroup_put(&ss
->root
->cgrp
);
2044 for_each_root(root
) {
2045 bool name_match
= false;
2047 if (root
== &cgrp_dfl_root
)
2051 * If we asked for a name then it must match. Also, if
2052 * name matches but sybsys_mask doesn't, we should fail.
2053 * Remember whether name matched.
2056 if (strcmp(opts
.name
, root
->name
))
2062 * If we asked for subsystems (or explicitly for no
2063 * subsystems) then they must match.
2065 if ((opts
.subsys_mask
|| opts
.none
) &&
2066 (opts
.subsys_mask
!= root
->subsys_mask
)) {
2073 if (root
->flags
^ opts
.flags
)
2074 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2077 * We want to reuse @root whose lifetime is governed by its
2078 * ->cgrp. Let's check whether @root is alive and keep it
2079 * that way. As cgroup_kill_sb() can happen anytime, we
2080 * want to block it by pinning the sb so that @root doesn't
2081 * get killed before mount is complete.
2083 * With the sb pinned, tryget_live can reliably indicate
2084 * whether @root can be reused. If it's being killed,
2085 * drain it. We can use wait_queue for the wait but this
2086 * path is super cold. Let's just sleep a bit and retry.
2088 pinned_sb
= kernfs_pin_sb(root
->kf_root
, NULL
);
2089 if (IS_ERR(pinned_sb
) ||
2090 !percpu_ref_tryget_live(&root
->cgrp
.self
.refcnt
)) {
2091 mutex_unlock(&cgroup_mutex
);
2092 if (!IS_ERR_OR_NULL(pinned_sb
))
2093 deactivate_super(pinned_sb
);
2095 ret
= restart_syscall();
2104 * No such thing, create a new one. name= matching without subsys
2105 * specification is allowed for already existing hierarchies but we
2106 * can't create new one without subsys specification.
2108 if (!opts
.subsys_mask
&& !opts
.none
) {
2113 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
2119 init_cgroup_root(root
, &opts
);
2121 ret
= cgroup_setup_root(root
, opts
.subsys_mask
);
2123 cgroup_free_root(root
);
2126 mutex_unlock(&cgroup_mutex
);
2128 kfree(opts
.release_agent
);
2132 return ERR_PTR(ret
);
2134 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
,
2135 CGROUP_SUPER_MAGIC
, &new_sb
);
2136 if (IS_ERR(dentry
) || !new_sb
)
2137 cgroup_put(&root
->cgrp
);
2140 * If @pinned_sb, we're reusing an existing root and holding an
2141 * extra ref on its sb. Mount is complete. Put the extra ref.
2145 deactivate_super(pinned_sb
);
2151 static void cgroup_kill_sb(struct super_block
*sb
)
2153 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
2154 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
2157 * If @root doesn't have any mounts or children, start killing it.
2158 * This prevents new mounts by disabling percpu_ref_tryget_live().
2159 * cgroup_mount() may wait for @root's release.
2161 * And don't kill the default root.
2163 if (!list_empty(&root
->cgrp
.self
.children
) ||
2164 root
== &cgrp_dfl_root
)
2165 cgroup_put(&root
->cgrp
);
2167 percpu_ref_kill(&root
->cgrp
.self
.refcnt
);
2172 static struct file_system_type cgroup_fs_type
= {
2174 .mount
= cgroup_mount
,
2175 .kill_sb
= cgroup_kill_sb
,
2179 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2180 * @task: target task
2181 * @buf: the buffer to write the path into
2182 * @buflen: the length of the buffer
2184 * Determine @task's cgroup on the first (the one with the lowest non-zero
2185 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2186 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2187 * cgroup controller callbacks.
2189 * Return value is the same as kernfs_path().
2191 char *task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
2193 struct cgroup_root
*root
;
2194 struct cgroup
*cgrp
;
2195 int hierarchy_id
= 1;
2198 mutex_lock(&cgroup_mutex
);
2199 down_read(&css_set_rwsem
);
2201 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
2204 cgrp
= task_cgroup_from_root(task
, root
);
2205 path
= cgroup_path(cgrp
, buf
, buflen
);
2207 /* if no hierarchy exists, everyone is in "/" */
2208 if (strlcpy(buf
, "/", buflen
) < buflen
)
2212 up_read(&css_set_rwsem
);
2213 mutex_unlock(&cgroup_mutex
);
2216 EXPORT_SYMBOL_GPL(task_cgroup_path
);
2218 /* used to track tasks and other necessary states during migration */
2219 struct cgroup_taskset
{
2220 /* the src and dst cset list running through cset->mg_node */
2221 struct list_head src_csets
;
2222 struct list_head dst_csets
;
2225 * Fields for cgroup_taskset_*() iteration.
2227 * Before migration is committed, the target migration tasks are on
2228 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2229 * the csets on ->dst_csets. ->csets point to either ->src_csets
2230 * or ->dst_csets depending on whether migration is committed.
2232 * ->cur_csets and ->cur_task point to the current task position
2235 struct list_head
*csets
;
2236 struct css_set
*cur_cset
;
2237 struct task_struct
*cur_task
;
2240 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2241 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2242 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2243 .csets = &tset.src_csets, \
2247 * cgroup_taskset_add - try to add a migration target task to a taskset
2248 * @task: target task
2249 * @tset: target taskset
2251 * Add @task, which is a migration target, to @tset. This function becomes
2252 * noop if @task doesn't need to be migrated. @task's css_set should have
2253 * been added as a migration source and @task->cg_list will be moved from
2254 * the css_set's tasks list to mg_tasks one.
2256 static void cgroup_taskset_add(struct task_struct
*task
,
2257 struct cgroup_taskset
*tset
)
2259 struct css_set
*cset
;
2261 lockdep_assert_held(&css_set_rwsem
);
2263 /* @task either already exited or can't exit until the end */
2264 if (task
->flags
& PF_EXITING
)
2267 /* leave @task alone if post_fork() hasn't linked it yet */
2268 if (list_empty(&task
->cg_list
))
2271 cset
= task_css_set(task
);
2272 if (!cset
->mg_src_cgrp
)
2275 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2276 if (list_empty(&cset
->mg_node
))
2277 list_add_tail(&cset
->mg_node
, &tset
->src_csets
);
2278 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2279 list_move_tail(&cset
->mg_dst_cset
->mg_node
,
2284 * cgroup_taskset_first - reset taskset and return the first task
2285 * @tset: taskset of interest
2287 * @tset iteration is initialized and the first task is returned.
2289 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
)
2291 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
2292 tset
->cur_task
= NULL
;
2294 return cgroup_taskset_next(tset
);
2298 * cgroup_taskset_next - iterate to the next task in taskset
2299 * @tset: taskset of interest
2301 * Return the next task in @tset. Iteration must have been initialized
2302 * with cgroup_taskset_first().
2304 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
)
2306 struct css_set
*cset
= tset
->cur_cset
;
2307 struct task_struct
*task
= tset
->cur_task
;
2309 while (&cset
->mg_node
!= tset
->csets
) {
2311 task
= list_first_entry(&cset
->mg_tasks
,
2312 struct task_struct
, cg_list
);
2314 task
= list_next_entry(task
, cg_list
);
2316 if (&task
->cg_list
!= &cset
->mg_tasks
) {
2317 tset
->cur_cset
= cset
;
2318 tset
->cur_task
= task
;
2322 cset
= list_next_entry(cset
, mg_node
);
2330 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2331 * @tset: taget taskset
2332 * @dst_cgrp: destination cgroup
2334 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2335 * ->can_attach callbacks fails and guarantees that either all or none of
2336 * the tasks in @tset are migrated. @tset is consumed regardless of
2339 static int cgroup_taskset_migrate(struct cgroup_taskset
*tset
,
2340 struct cgroup
*dst_cgrp
)
2342 struct cgroup_subsys_state
*css
, *failed_css
= NULL
;
2343 struct task_struct
*task
, *tmp_task
;
2344 struct css_set
*cset
, *tmp_cset
;
2347 /* methods shouldn't be called if no task is actually migrating */
2348 if (list_empty(&tset
->src_csets
))
2351 /* check that we can legitimately attach to the cgroup */
2352 for_each_e_css(css
, i
, dst_cgrp
) {
2353 if (css
->ss
->can_attach
) {
2354 ret
= css
->ss
->can_attach(css
, tset
);
2357 goto out_cancel_attach
;
2363 * Now that we're guaranteed success, proceed to move all tasks to
2364 * the new cgroup. There are no failure cases after here, so this
2365 * is the commit point.
2367 down_write(&css_set_rwsem
);
2368 list_for_each_entry(cset
, &tset
->src_csets
, mg_node
) {
2369 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
) {
2370 struct css_set
*from_cset
= task_css_set(task
);
2371 struct css_set
*to_cset
= cset
->mg_dst_cset
;
2373 get_css_set(to_cset
);
2374 css_set_move_task(task
, from_cset
, to_cset
, true);
2375 put_css_set_locked(from_cset
);
2378 up_write(&css_set_rwsem
);
2381 * Migration is committed, all target tasks are now on dst_csets.
2382 * Nothing is sensitive to fork() after this point. Notify
2383 * controllers that migration is complete.
2385 tset
->csets
= &tset
->dst_csets
;
2387 for_each_e_css(css
, i
, dst_cgrp
)
2388 if (css
->ss
->attach
)
2389 css
->ss
->attach(css
, tset
);
2392 goto out_release_tset
;
2395 for_each_e_css(css
, i
, dst_cgrp
) {
2396 if (css
== failed_css
)
2398 if (css
->ss
->cancel_attach
)
2399 css
->ss
->cancel_attach(css
, tset
);
2402 down_write(&css_set_rwsem
);
2403 list_splice_init(&tset
->dst_csets
, &tset
->src_csets
);
2404 list_for_each_entry_safe(cset
, tmp_cset
, &tset
->src_csets
, mg_node
) {
2405 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2406 list_del_init(&cset
->mg_node
);
2408 up_write(&css_set_rwsem
);
2413 * cgroup_migrate_finish - cleanup after attach
2414 * @preloaded_csets: list of preloaded css_sets
2416 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2417 * those functions for details.
2419 static void cgroup_migrate_finish(struct list_head
*preloaded_csets
)
2421 struct css_set
*cset
, *tmp_cset
;
2423 lockdep_assert_held(&cgroup_mutex
);
2425 down_write(&css_set_rwsem
);
2426 list_for_each_entry_safe(cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2427 cset
->mg_src_cgrp
= NULL
;
2428 cset
->mg_dst_cset
= NULL
;
2429 list_del_init(&cset
->mg_preload_node
);
2430 put_css_set_locked(cset
);
2432 up_write(&css_set_rwsem
);
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 * @preloaded_csets: list of preloaded css_sets
2441 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2442 * @src_cset and add it to @preloaded_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 static void cgroup_migrate_add_src(struct css_set
*src_cset
,
2452 struct cgroup
*dst_cgrp
,
2453 struct list_head
*preloaded_csets
)
2455 struct cgroup
*src_cgrp
;
2457 lockdep_assert_held(&cgroup_mutex
);
2458 lockdep_assert_held(&css_set_rwsem
);
2460 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2462 if (!list_empty(&src_cset
->mg_preload_node
))
2465 WARN_ON(src_cset
->mg_src_cgrp
);
2466 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2467 WARN_ON(!list_empty(&src_cset
->mg_node
));
2469 src_cset
->mg_src_cgrp
= src_cgrp
;
2470 get_css_set(src_cset
);
2471 list_add(&src_cset
->mg_preload_node
, preloaded_csets
);
2475 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2476 * @dst_cgrp: the destination cgroup (may be %NULL)
2477 * @preloaded_csets: list of preloaded source css_sets
2479 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2480 * have been preloaded to @preloaded_csets. This function looks up and
2481 * pins all destination css_sets, links each to its source, and append them
2482 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2483 * source css_set is assumed to be its cgroup on the default hierarchy.
2485 * This function must be called after cgroup_migrate_add_src() has been
2486 * called on each migration source css_set. After migration is performed
2487 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2490 static int cgroup_migrate_prepare_dst(struct cgroup
*dst_cgrp
,
2491 struct list_head
*preloaded_csets
)
2494 struct css_set
*src_cset
, *tmp_cset
;
2496 lockdep_assert_held(&cgroup_mutex
);
2499 * Except for the root, child_subsys_mask must be zero for a cgroup
2500 * with tasks so that child cgroups don't compete against tasks.
2502 if (dst_cgrp
&& cgroup_on_dfl(dst_cgrp
) && cgroup_parent(dst_cgrp
) &&
2503 dst_cgrp
->child_subsys_mask
)
2506 /* look up the dst cset for each src cset and link it to src */
2507 list_for_each_entry_safe(src_cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2508 struct css_set
*dst_cset
;
2510 dst_cset
= find_css_set(src_cset
,
2511 dst_cgrp
?: src_cset
->dfl_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 list_del_init(&src_cset
->mg_preload_node
);
2525 put_css_set(src_cset
);
2526 put_css_set(dst_cset
);
2530 src_cset
->mg_dst_cset
= dst_cset
;
2532 if (list_empty(&dst_cset
->mg_preload_node
))
2533 list_add(&dst_cset
->mg_preload_node
, &csets
);
2535 put_css_set(dst_cset
);
2538 list_splice_tail(&csets
, preloaded_csets
);
2541 cgroup_migrate_finish(&csets
);
2546 * cgroup_migrate - migrate a process or task to a cgroup
2547 * @leader: the leader of the process or the task to migrate
2548 * @threadgroup: whether @leader points to the whole process or a single task
2549 * @cgrp: the destination cgroup
2551 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2552 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2553 * caller is also responsible for invoking cgroup_migrate_add_src() and
2554 * cgroup_migrate_prepare_dst() on the targets before invoking this
2555 * function and following up with cgroup_migrate_finish().
2557 * As long as a controller's ->can_attach() doesn't fail, this function is
2558 * guaranteed to succeed. This means that, excluding ->can_attach()
2559 * failure, when migrating multiple targets, the success or failure can be
2560 * decided for all targets by invoking group_migrate_prepare_dst() before
2561 * actually starting migrating.
2563 static int cgroup_migrate(struct task_struct
*leader
, bool threadgroup
,
2564 struct cgroup
*cgrp
)
2566 struct cgroup_taskset tset
= CGROUP_TASKSET_INIT(tset
);
2567 struct task_struct
*task
;
2570 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2571 * already PF_EXITING could be freed from underneath us unless we
2572 * take an rcu_read_lock.
2574 down_write(&css_set_rwsem
);
2578 cgroup_taskset_add(task
, &tset
);
2581 } while_each_thread(leader
, task
);
2583 up_write(&css_set_rwsem
);
2585 return cgroup_taskset_migrate(&tset
, cgrp
);
2589 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2590 * @dst_cgrp: the cgroup to attach to
2591 * @leader: the task or the leader of the threadgroup to be attached
2592 * @threadgroup: attach the whole threadgroup?
2594 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2596 static int cgroup_attach_task(struct cgroup
*dst_cgrp
,
2597 struct task_struct
*leader
, bool threadgroup
)
2599 LIST_HEAD(preloaded_csets
);
2600 struct task_struct
*task
;
2603 /* look up all src csets */
2604 down_read(&css_set_rwsem
);
2608 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
,
2612 } while_each_thread(leader
, task
);
2614 up_read(&css_set_rwsem
);
2616 /* prepare dst csets and commit */
2617 ret
= cgroup_migrate_prepare_dst(dst_cgrp
, &preloaded_csets
);
2619 ret
= cgroup_migrate(leader
, threadgroup
, dst_cgrp
);
2621 cgroup_migrate_finish(&preloaded_csets
);
2625 static int cgroup_procs_write_permission(struct task_struct
*task
,
2626 struct cgroup
*dst_cgrp
,
2627 struct kernfs_open_file
*of
)
2629 const struct cred
*cred
= current_cred();
2630 const struct cred
*tcred
= get_task_cred(task
);
2634 * even if we're attaching all tasks in the thread group, we only
2635 * need to check permissions on one of them.
2637 if (!uid_eq(cred
->euid
, GLOBAL_ROOT_UID
) &&
2638 !uid_eq(cred
->euid
, tcred
->uid
) &&
2639 !uid_eq(cred
->euid
, tcred
->suid
))
2642 if (!ret
&& cgroup_on_dfl(dst_cgrp
)) {
2643 struct super_block
*sb
= of
->file
->f_path
.dentry
->d_sb
;
2644 struct cgroup
*cgrp
;
2645 struct inode
*inode
;
2647 down_read(&css_set_rwsem
);
2648 cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
2649 up_read(&css_set_rwsem
);
2651 while (!cgroup_is_descendant(dst_cgrp
, cgrp
))
2652 cgrp
= cgroup_parent(cgrp
);
2655 inode
= kernfs_get_inode(sb
, cgrp
->procs_file
.kn
);
2657 ret
= inode_permission(inode
, MAY_WRITE
);
2667 * Find the task_struct of the task to attach by vpid and pass it along to the
2668 * function to attach either it or all tasks in its threadgroup. Will lock
2669 * cgroup_mutex and threadgroup.
2671 static ssize_t
__cgroup_procs_write(struct kernfs_open_file
*of
, char *buf
,
2672 size_t nbytes
, loff_t off
, bool threadgroup
)
2674 struct task_struct
*tsk
;
2675 struct cgroup
*cgrp
;
2679 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2682 cgrp
= cgroup_kn_lock_live(of
->kn
);
2686 percpu_down_write(&cgroup_threadgroup_rwsem
);
2689 tsk
= find_task_by_vpid(pid
);
2692 goto out_unlock_rcu
;
2699 tsk
= tsk
->group_leader
;
2702 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2703 * trapped in a cpuset, or RT worker may be born in a cgroup
2704 * with no rt_runtime allocated. Just say no.
2706 if (tsk
== kthreadd_task
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2708 goto out_unlock_rcu
;
2711 get_task_struct(tsk
);
2714 ret
= cgroup_procs_write_permission(tsk
, cgrp
, of
);
2716 ret
= cgroup_attach_task(cgrp
, tsk
, threadgroup
);
2718 put_task_struct(tsk
);
2719 goto out_unlock_threadgroup
;
2723 out_unlock_threadgroup
:
2724 percpu_up_write(&cgroup_threadgroup_rwsem
);
2725 cgroup_kn_unlock(of
->kn
);
2726 return ret
?: nbytes
;
2730 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2731 * @from: attach to all cgroups of a given task
2732 * @tsk: the task to be attached
2734 int cgroup_attach_task_all(struct task_struct
*from
, struct task_struct
*tsk
)
2736 struct cgroup_root
*root
;
2739 mutex_lock(&cgroup_mutex
);
2740 for_each_root(root
) {
2741 struct cgroup
*from_cgrp
;
2743 if (root
== &cgrp_dfl_root
)
2746 down_read(&css_set_rwsem
);
2747 from_cgrp
= task_cgroup_from_root(from
, root
);
2748 up_read(&css_set_rwsem
);
2750 retval
= cgroup_attach_task(from_cgrp
, tsk
, false);
2754 mutex_unlock(&cgroup_mutex
);
2758 EXPORT_SYMBOL_GPL(cgroup_attach_task_all
);
2760 static ssize_t
cgroup_tasks_write(struct kernfs_open_file
*of
,
2761 char *buf
, size_t nbytes
, loff_t off
)
2763 return __cgroup_procs_write(of
, buf
, nbytes
, off
, false);
2766 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
2767 char *buf
, size_t nbytes
, loff_t off
)
2769 return __cgroup_procs_write(of
, buf
, nbytes
, off
, true);
2772 static ssize_t
cgroup_release_agent_write(struct kernfs_open_file
*of
,
2773 char *buf
, size_t nbytes
, loff_t off
)
2775 struct cgroup
*cgrp
;
2777 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
2779 cgrp
= cgroup_kn_lock_live(of
->kn
);
2782 spin_lock(&release_agent_path_lock
);
2783 strlcpy(cgrp
->root
->release_agent_path
, strstrip(buf
),
2784 sizeof(cgrp
->root
->release_agent_path
));
2785 spin_unlock(&release_agent_path_lock
);
2786 cgroup_kn_unlock(of
->kn
);
2790 static int cgroup_release_agent_show(struct seq_file
*seq
, void *v
)
2792 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2794 spin_lock(&release_agent_path_lock
);
2795 seq_puts(seq
, cgrp
->root
->release_agent_path
);
2796 spin_unlock(&release_agent_path_lock
);
2797 seq_putc(seq
, '\n');
2801 static int cgroup_sane_behavior_show(struct seq_file
*seq
, void *v
)
2803 seq_puts(seq
, "0\n");
2807 static void cgroup_print_ss_mask(struct seq_file
*seq
, unsigned long ss_mask
)
2809 struct cgroup_subsys
*ss
;
2810 bool printed
= false;
2813 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
2816 seq_printf(seq
, "%s", ss
->name
);
2820 seq_putc(seq
, '\n');
2823 /* show controllers which are currently attached to the default hierarchy */
2824 static int cgroup_root_controllers_show(struct seq_file
*seq
, void *v
)
2826 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2828 cgroup_print_ss_mask(seq
, cgrp
->root
->subsys_mask
&
2829 ~cgrp_dfl_root_inhibit_ss_mask
);
2833 /* show controllers which are enabled from the parent */
2834 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2836 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2838 cgroup_print_ss_mask(seq
, cgroup_parent(cgrp
)->subtree_control
);
2842 /* show controllers which are enabled for a given cgroup's children */
2843 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2845 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2847 cgroup_print_ss_mask(seq
, cgrp
->subtree_control
);
2852 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2853 * @cgrp: root of the subtree to update csses for
2855 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2856 * css associations need to be updated accordingly. This function looks up
2857 * all css_sets which are attached to the subtree, creates the matching
2858 * updated css_sets and migrates the tasks to the new ones.
2860 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2862 LIST_HEAD(preloaded_csets
);
2863 struct cgroup_taskset tset
= CGROUP_TASKSET_INIT(tset
);
2864 struct cgroup_subsys_state
*css
;
2865 struct css_set
*src_cset
;
2868 lockdep_assert_held(&cgroup_mutex
);
2870 percpu_down_write(&cgroup_threadgroup_rwsem
);
2872 /* look up all csses currently attached to @cgrp's subtree */
2873 down_read(&css_set_rwsem
);
2874 css_for_each_descendant_pre(css
, cgroup_css(cgrp
, NULL
)) {
2875 struct cgrp_cset_link
*link
;
2877 /* self is not affected by child_subsys_mask change */
2878 if (css
->cgroup
== cgrp
)
2881 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
)
2882 cgroup_migrate_add_src(link
->cset
, cgrp
,
2885 up_read(&css_set_rwsem
);
2887 /* NULL dst indicates self on default hierarchy */
2888 ret
= cgroup_migrate_prepare_dst(NULL
, &preloaded_csets
);
2892 down_write(&css_set_rwsem
);
2893 list_for_each_entry(src_cset
, &preloaded_csets
, mg_preload_node
) {
2894 struct task_struct
*task
, *ntask
;
2896 /* src_csets precede dst_csets, break on the first dst_cset */
2897 if (!src_cset
->mg_src_cgrp
)
2900 /* all tasks in src_csets need to be migrated */
2901 list_for_each_entry_safe(task
, ntask
, &src_cset
->tasks
, cg_list
)
2902 cgroup_taskset_add(task
, &tset
);
2904 up_write(&css_set_rwsem
);
2906 ret
= cgroup_taskset_migrate(&tset
, cgrp
);
2908 cgroup_migrate_finish(&preloaded_csets
);
2909 percpu_up_write(&cgroup_threadgroup_rwsem
);
2913 /* change the enabled child controllers for a cgroup in the default hierarchy */
2914 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
2915 char *buf
, size_t nbytes
,
2918 unsigned long enable
= 0, disable
= 0;
2919 unsigned long css_enable
, css_disable
, old_sc
, new_sc
, old_ss
, new_ss
;
2920 struct cgroup
*cgrp
, *child
;
2921 struct cgroup_subsys
*ss
;
2926 * Parse input - space separated list of subsystem names prefixed
2927 * with either + or -.
2929 buf
= strstrip(buf
);
2930 while ((tok
= strsep(&buf
, " "))) {
2931 unsigned long tmp_ss_mask
= ~cgrp_dfl_root_inhibit_ss_mask
;
2935 for_each_subsys_which(ss
, ssid
, &tmp_ss_mask
) {
2936 if (!cgroup_ssid_enabled(ssid
) ||
2937 strcmp(tok
+ 1, ss
->name
))
2941 enable
|= 1 << ssid
;
2942 disable
&= ~(1 << ssid
);
2943 } else if (*tok
== '-') {
2944 disable
|= 1 << ssid
;
2945 enable
&= ~(1 << ssid
);
2951 if (ssid
== CGROUP_SUBSYS_COUNT
)
2955 cgrp
= cgroup_kn_lock_live(of
->kn
);
2959 for_each_subsys(ss
, ssid
) {
2960 if (enable
& (1 << ssid
)) {
2961 if (cgrp
->subtree_control
& (1 << ssid
)) {
2962 enable
&= ~(1 << ssid
);
2966 /* unavailable or not enabled on the parent? */
2967 if (!(cgrp_dfl_root
.subsys_mask
& (1 << ssid
)) ||
2968 (cgroup_parent(cgrp
) &&
2969 !(cgroup_parent(cgrp
)->subtree_control
& (1 << ssid
)))) {
2973 } else if (disable
& (1 << ssid
)) {
2974 if (!(cgrp
->subtree_control
& (1 << ssid
))) {
2975 disable
&= ~(1 << ssid
);
2979 /* a child has it enabled? */
2980 cgroup_for_each_live_child(child
, cgrp
) {
2981 if (child
->subtree_control
& (1 << ssid
)) {
2989 if (!enable
&& !disable
) {
2995 * Except for the root, subtree_control must be zero for a cgroup
2996 * with tasks so that child cgroups don't compete against tasks.
2998 if (enable
&& cgroup_parent(cgrp
) && !list_empty(&cgrp
->cset_links
)) {
3004 * Update subsys masks and calculate what needs to be done. More
3005 * subsystems than specified may need to be enabled or disabled
3006 * depending on subsystem dependencies.
3008 old_sc
= cgrp
->subtree_control
;
3009 old_ss
= cgrp
->child_subsys_mask
;
3010 new_sc
= (old_sc
| enable
) & ~disable
;
3011 new_ss
= cgroup_calc_child_subsys_mask(cgrp
, new_sc
);
3013 css_enable
= ~old_ss
& new_ss
;
3014 css_disable
= old_ss
& ~new_ss
;
3015 enable
|= css_enable
;
3016 disable
|= css_disable
;
3019 * Because css offlining is asynchronous, userland might try to
3020 * re-enable the same controller while the previous instance is
3021 * still around. In such cases, wait till it's gone using
3024 for_each_subsys_which(ss
, ssid
, &css_enable
) {
3025 cgroup_for_each_live_child(child
, cgrp
) {
3028 if (!cgroup_css(child
, ss
))
3032 prepare_to_wait(&child
->offline_waitq
, &wait
,
3033 TASK_UNINTERRUPTIBLE
);
3034 cgroup_kn_unlock(of
->kn
);
3036 finish_wait(&child
->offline_waitq
, &wait
);
3039 return restart_syscall();
3043 cgrp
->subtree_control
= new_sc
;
3044 cgrp
->child_subsys_mask
= new_ss
;
3047 * Create new csses or make the existing ones visible. A css is
3048 * created invisible if it's being implicitly enabled through
3049 * dependency. An invisible css is made visible when the userland
3050 * explicitly enables it.
3052 for_each_subsys(ss
, ssid
) {
3053 if (!(enable
& (1 << ssid
)))
3056 cgroup_for_each_live_child(child
, cgrp
) {
3057 if (css_enable
& (1 << ssid
))
3058 ret
= create_css(child
, ss
,
3059 cgrp
->subtree_control
& (1 << ssid
));
3061 ret
= css_populate_dir(cgroup_css(child
, ss
),
3069 * At this point, cgroup_e_css() results reflect the new csses
3070 * making the following cgroup_update_dfl_csses() properly update
3071 * css associations of all tasks in the subtree.
3073 ret
= cgroup_update_dfl_csses(cgrp
);
3078 * All tasks are migrated out of disabled csses. Kill or hide
3079 * them. A css is hidden when the userland requests it to be
3080 * disabled while other subsystems are still depending on it. The
3081 * css must not actively control resources and be in the vanilla
3082 * state if it's made visible again later. Controllers which may
3083 * be depended upon should provide ->css_reset() for this purpose.
3085 for_each_subsys(ss
, ssid
) {
3086 if (!(disable
& (1 << ssid
)))
3089 cgroup_for_each_live_child(child
, cgrp
) {
3090 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3092 if (css_disable
& (1 << ssid
)) {
3095 css_clear_dir(css
, NULL
);
3103 * The effective csses of all the descendants (excluding @cgrp) may
3104 * have changed. Subsystems can optionally subscribe to this event
3105 * by implementing ->css_e_css_changed() which is invoked if any of
3106 * the effective csses seen from the css's cgroup may have changed.
3108 for_each_subsys(ss
, ssid
) {
3109 struct cgroup_subsys_state
*this_css
= cgroup_css(cgrp
, ss
);
3110 struct cgroup_subsys_state
*css
;
3112 if (!ss
->css_e_css_changed
|| !this_css
)
3115 css_for_each_descendant_pre(css
, this_css
)
3116 if (css
!= this_css
)
3117 ss
->css_e_css_changed(css
);
3120 kernfs_activate(cgrp
->kn
);
3123 cgroup_kn_unlock(of
->kn
);
3124 return ret
?: nbytes
;
3127 cgrp
->subtree_control
= old_sc
;
3128 cgrp
->child_subsys_mask
= old_ss
;
3130 for_each_subsys(ss
, ssid
) {
3131 if (!(enable
& (1 << ssid
)))
3134 cgroup_for_each_live_child(child
, cgrp
) {
3135 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3140 if (css_enable
& (1 << ssid
))
3143 css_clear_dir(css
, NULL
);
3149 static int cgroup_events_show(struct seq_file
*seq
, void *v
)
3151 seq_printf(seq
, "populated %d\n",
3152 cgroup_is_populated(seq_css(seq
)->cgroup
));
3156 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
3157 size_t nbytes
, loff_t off
)
3159 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
3160 struct cftype
*cft
= of
->kn
->priv
;
3161 struct cgroup_subsys_state
*css
;
3165 return cft
->write(of
, buf
, nbytes
, off
);
3168 * kernfs guarantees that a file isn't deleted with operations in
3169 * flight, which means that the matching css is and stays alive and
3170 * doesn't need to be pinned. The RCU locking is not necessary
3171 * either. It's just for the convenience of using cgroup_css().
3174 css
= cgroup_css(cgrp
, cft
->ss
);
3177 if (cft
->write_u64
) {
3178 unsigned long long v
;
3179 ret
= kstrtoull(buf
, 0, &v
);
3181 ret
= cft
->write_u64(css
, cft
, v
);
3182 } else if (cft
->write_s64
) {
3184 ret
= kstrtoll(buf
, 0, &v
);
3186 ret
= cft
->write_s64(css
, cft
, v
);
3191 return ret
?: nbytes
;
3194 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
3196 return seq_cft(seq
)->seq_start(seq
, ppos
);
3199 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
3201 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
3204 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
3206 seq_cft(seq
)->seq_stop(seq
, v
);
3209 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
3211 struct cftype
*cft
= seq_cft(m
);
3212 struct cgroup_subsys_state
*css
= seq_css(m
);
3215 return cft
->seq_show(m
, arg
);
3218 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
3219 else if (cft
->read_s64
)
3220 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
3226 static struct kernfs_ops cgroup_kf_single_ops
= {
3227 .atomic_write_len
= PAGE_SIZE
,
3228 .write
= cgroup_file_write
,
3229 .seq_show
= cgroup_seqfile_show
,
3232 static struct kernfs_ops cgroup_kf_ops
= {
3233 .atomic_write_len
= PAGE_SIZE
,
3234 .write
= cgroup_file_write
,
3235 .seq_start
= cgroup_seqfile_start
,
3236 .seq_next
= cgroup_seqfile_next
,
3237 .seq_stop
= cgroup_seqfile_stop
,
3238 .seq_show
= cgroup_seqfile_show
,
3242 * cgroup_rename - Only allow simple rename of directories in place.
3244 static int cgroup_rename(struct kernfs_node
*kn
, struct kernfs_node
*new_parent
,
3245 const char *new_name_str
)
3247 struct cgroup
*cgrp
= kn
->priv
;
3250 if (kernfs_type(kn
) != KERNFS_DIR
)
3252 if (kn
->parent
!= new_parent
)
3256 * This isn't a proper migration and its usefulness is very
3257 * limited. Disallow on the default hierarchy.
3259 if (cgroup_on_dfl(cgrp
))
3263 * We're gonna grab cgroup_mutex which nests outside kernfs
3264 * active_ref. kernfs_rename() doesn't require active_ref
3265 * protection. Break them before grabbing cgroup_mutex.
3267 kernfs_break_active_protection(new_parent
);
3268 kernfs_break_active_protection(kn
);
3270 mutex_lock(&cgroup_mutex
);
3272 ret
= kernfs_rename(kn
, new_parent
, new_name_str
);
3274 mutex_unlock(&cgroup_mutex
);
3276 kernfs_unbreak_active_protection(kn
);
3277 kernfs_unbreak_active_protection(new_parent
);
3281 /* set uid and gid of cgroup dirs and files to that of the creator */
3282 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
3284 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
3285 .ia_uid
= current_fsuid(),
3286 .ia_gid
= current_fsgid(), };
3288 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
3289 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
3292 return kernfs_setattr(kn
, &iattr
);
3295 static int cgroup_add_file(struct cgroup_subsys_state
*css
, struct cgroup
*cgrp
,
3298 char name
[CGROUP_FILE_NAME_MAX
];
3299 struct kernfs_node
*kn
;
3300 struct lock_class_key
*key
= NULL
;
3303 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3304 key
= &cft
->lockdep_key
;
3306 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
3307 cgroup_file_mode(cft
), 0, cft
->kf_ops
, cft
,
3312 ret
= cgroup_kn_set_ugid(kn
);
3318 if (cft
->file_offset
) {
3319 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
3323 list_add(&cfile
->node
, &css
->files
);
3330 * cgroup_addrm_files - add or remove files to a cgroup directory
3331 * @css: the target css
3332 * @cgrp: the target cgroup (usually css->cgroup)
3333 * @cfts: array of cftypes to be added
3334 * @is_add: whether to add or remove
3336 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3337 * For removals, this function never fails.
3339 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
3340 struct cgroup
*cgrp
, struct cftype cfts
[],
3343 struct cftype
*cft
, *cft_end
= NULL
;
3346 lockdep_assert_held(&cgroup_mutex
);
3349 for (cft
= cfts
; cft
!= cft_end
&& cft
->name
[0] != '\0'; cft
++) {
3350 /* does cft->flags tell us to skip this file on @cgrp? */
3351 if ((cft
->flags
& __CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
3353 if ((cft
->flags
& __CFTYPE_NOT_ON_DFL
) && cgroup_on_dfl(cgrp
))
3355 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
3357 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
3361 ret
= cgroup_add_file(css
, cgrp
, cft
);
3363 pr_warn("%s: failed to add %s, err=%d\n",
3364 __func__
, cft
->name
, ret
);
3370 cgroup_rm_file(cgrp
, cft
);
3376 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
3379 struct cgroup_subsys
*ss
= cfts
[0].ss
;
3380 struct cgroup
*root
= &ss
->root
->cgrp
;
3381 struct cgroup_subsys_state
*css
;
3384 lockdep_assert_held(&cgroup_mutex
);
3386 /* add/rm files for all cgroups created before */
3387 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
3388 struct cgroup
*cgrp
= css
->cgroup
;
3390 if (cgroup_is_dead(cgrp
))
3393 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, is_add
);
3399 kernfs_activate(root
->kn
);
3403 static void cgroup_exit_cftypes(struct cftype
*cfts
)
3407 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3408 /* free copy for custom atomic_write_len, see init_cftypes() */
3409 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
3414 /* revert flags set by cgroup core while adding @cfts */
3415 cft
->flags
&= ~(__CFTYPE_ONLY_ON_DFL
| __CFTYPE_NOT_ON_DFL
);
3419 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3423 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3424 struct kernfs_ops
*kf_ops
;
3426 WARN_ON(cft
->ss
|| cft
->kf_ops
);
3429 kf_ops
= &cgroup_kf_ops
;
3431 kf_ops
= &cgroup_kf_single_ops
;
3434 * Ugh... if @cft wants a custom max_write_len, we need to
3435 * make a copy of kf_ops to set its atomic_write_len.
3437 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
3438 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
3440 cgroup_exit_cftypes(cfts
);
3443 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3446 cft
->kf_ops
= kf_ops
;
3453 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3455 lockdep_assert_held(&cgroup_mutex
);
3457 if (!cfts
|| !cfts
[0].ss
)
3460 list_del(&cfts
->node
);
3461 cgroup_apply_cftypes(cfts
, false);
3462 cgroup_exit_cftypes(cfts
);
3467 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3468 * @cfts: zero-length name terminated array of cftypes
3470 * Unregister @cfts. Files described by @cfts are removed from all
3471 * existing cgroups and all future cgroups won't have them either. This
3472 * function can be called anytime whether @cfts' subsys is attached or not.
3474 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3477 int cgroup_rm_cftypes(struct cftype
*cfts
)
3481 mutex_lock(&cgroup_mutex
);
3482 ret
= cgroup_rm_cftypes_locked(cfts
);
3483 mutex_unlock(&cgroup_mutex
);
3488 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3489 * @ss: target cgroup subsystem
3490 * @cfts: zero-length name terminated array of cftypes
3492 * Register @cfts to @ss. Files described by @cfts are created for all
3493 * existing cgroups to which @ss is attached and all future cgroups will
3494 * have them too. This function can be called anytime whether @ss is
3497 * Returns 0 on successful registration, -errno on failure. Note that this
3498 * function currently returns 0 as long as @cfts registration is successful
3499 * even if some file creation attempts on existing cgroups fail.
3501 static int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3505 if (!cgroup_ssid_enabled(ss
->id
))
3508 if (!cfts
|| cfts
[0].name
[0] == '\0')
3511 ret
= cgroup_init_cftypes(ss
, cfts
);
3515 mutex_lock(&cgroup_mutex
);
3517 list_add_tail(&cfts
->node
, &ss
->cfts
);
3518 ret
= cgroup_apply_cftypes(cfts
, true);
3520 cgroup_rm_cftypes_locked(cfts
);
3522 mutex_unlock(&cgroup_mutex
);
3527 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3528 * @ss: target cgroup subsystem
3529 * @cfts: zero-length name terminated array of cftypes
3531 * Similar to cgroup_add_cftypes() but the added files are only used for
3532 * the default hierarchy.
3534 int cgroup_add_dfl_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3538 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3539 cft
->flags
|= __CFTYPE_ONLY_ON_DFL
;
3540 return cgroup_add_cftypes(ss
, cfts
);
3544 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3545 * @ss: target cgroup subsystem
3546 * @cfts: zero-length name terminated array of cftypes
3548 * Similar to cgroup_add_cftypes() but the added files are only used for
3549 * the legacy hierarchies.
3551 int cgroup_add_legacy_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3556 * If legacy_flies_on_dfl, we want to show the legacy files on the
3557 * dfl hierarchy but iff the target subsystem hasn't been updated
3558 * for the dfl hierarchy yet.
3560 if (!cgroup_legacy_files_on_dfl
||
3561 ss
->dfl_cftypes
!= ss
->legacy_cftypes
) {
3562 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3563 cft
->flags
|= __CFTYPE_NOT_ON_DFL
;
3566 return cgroup_add_cftypes(ss
, cfts
);
3570 * cgroup_task_count - count the number of tasks in a cgroup.
3571 * @cgrp: the cgroup in question
3573 * Return the number of tasks in the cgroup.
3575 static int cgroup_task_count(const struct cgroup
*cgrp
)
3578 struct cgrp_cset_link
*link
;
3580 down_read(&css_set_rwsem
);
3581 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
3582 count
+= atomic_read(&link
->cset
->refcount
);
3583 up_read(&css_set_rwsem
);
3588 * css_next_child - find the next child of a given css
3589 * @pos: the current position (%NULL to initiate traversal)
3590 * @parent: css whose children to walk
3592 * This function returns the next child of @parent and should be called
3593 * under either cgroup_mutex or RCU read lock. The only requirement is
3594 * that @parent and @pos are accessible. The next sibling is guaranteed to
3595 * be returned regardless of their states.
3597 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3598 * css which finished ->css_online() is guaranteed to be visible in the
3599 * future iterations and will stay visible until the last reference is put.
3600 * A css which hasn't finished ->css_online() or already finished
3601 * ->css_offline() may show up during traversal. It's each subsystem's
3602 * responsibility to synchronize against on/offlining.
3604 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
3605 struct cgroup_subsys_state
*parent
)
3607 struct cgroup_subsys_state
*next
;
3609 cgroup_assert_mutex_or_rcu_locked();
3612 * @pos could already have been unlinked from the sibling list.
3613 * Once a cgroup is removed, its ->sibling.next is no longer
3614 * updated when its next sibling changes. CSS_RELEASED is set when
3615 * @pos is taken off list, at which time its next pointer is valid,
3616 * and, as releases are serialized, the one pointed to by the next
3617 * pointer is guaranteed to not have started release yet. This
3618 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3619 * critical section, the one pointed to by its next pointer is
3620 * guaranteed to not have finished its RCU grace period even if we
3621 * have dropped rcu_read_lock() inbetween iterations.
3623 * If @pos has CSS_RELEASED set, its next pointer can't be
3624 * dereferenced; however, as each css is given a monotonically
3625 * increasing unique serial number and always appended to the
3626 * sibling list, the next one can be found by walking the parent's
3627 * children until the first css with higher serial number than
3628 * @pos's. While this path can be slower, it happens iff iteration
3629 * races against release and the race window is very small.
3632 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
3633 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
3634 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
3636 list_for_each_entry_rcu(next
, &parent
->children
, sibling
)
3637 if (next
->serial_nr
> pos
->serial_nr
)
3642 * @next, if not pointing to the head, can be dereferenced and is
3645 if (&next
->sibling
!= &parent
->children
)
3651 * css_next_descendant_pre - find the next descendant for pre-order walk
3652 * @pos: the current position (%NULL to initiate traversal)
3653 * @root: css whose descendants to walk
3655 * To be used by css_for_each_descendant_pre(). Find the next descendant
3656 * to visit for pre-order traversal of @root's descendants. @root is
3657 * included in the iteration and the first node to be visited.
3659 * While this function requires cgroup_mutex or RCU read locking, it
3660 * doesn't require the whole traversal to be contained in a single critical
3661 * section. This function will return the correct next descendant as long
3662 * as both @pos and @root are accessible and @pos is a descendant of @root.
3664 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3665 * css which finished ->css_online() is guaranteed to be visible in the
3666 * future iterations and will stay visible until the last reference is put.
3667 * A css which hasn't finished ->css_online() or already finished
3668 * ->css_offline() may show up during traversal. It's each subsystem's
3669 * responsibility to synchronize against on/offlining.
3671 struct cgroup_subsys_state
*
3672 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3673 struct cgroup_subsys_state
*root
)
3675 struct cgroup_subsys_state
*next
;
3677 cgroup_assert_mutex_or_rcu_locked();
3679 /* if first iteration, visit @root */
3683 /* visit the first child if exists */
3684 next
= css_next_child(NULL
, pos
);
3688 /* no child, visit my or the closest ancestor's next sibling */
3689 while (pos
!= root
) {
3690 next
= css_next_child(pos
, pos
->parent
);
3700 * css_rightmost_descendant - return the rightmost descendant of a css
3701 * @pos: css of interest
3703 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3704 * is returned. This can be used during pre-order traversal to skip
3707 * While this function requires cgroup_mutex or RCU read locking, it
3708 * doesn't require the whole traversal to be contained in a single critical
3709 * section. This function will return the correct rightmost descendant as
3710 * long as @pos is accessible.
3712 struct cgroup_subsys_state
*
3713 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3715 struct cgroup_subsys_state
*last
, *tmp
;
3717 cgroup_assert_mutex_or_rcu_locked();
3721 /* ->prev isn't RCU safe, walk ->next till the end */
3723 css_for_each_child(tmp
, last
)
3730 static struct cgroup_subsys_state
*
3731 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3733 struct cgroup_subsys_state
*last
;
3737 pos
= css_next_child(NULL
, pos
);
3744 * css_next_descendant_post - find the next descendant for post-order walk
3745 * @pos: the current position (%NULL to initiate traversal)
3746 * @root: css whose descendants to walk
3748 * To be used by css_for_each_descendant_post(). Find the next descendant
3749 * to visit for post-order traversal of @root's descendants. @root is
3750 * included in the iteration and the last node to be visited.
3752 * While this function requires cgroup_mutex or RCU read locking, it
3753 * doesn't require the whole traversal to be contained in a single critical
3754 * section. This function will return the correct next descendant as long
3755 * as both @pos and @cgroup are accessible and @pos is a descendant of
3758 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3759 * css which finished ->css_online() is guaranteed to be visible in the
3760 * future iterations and will stay visible until the last reference is put.
3761 * A css which hasn't finished ->css_online() or already finished
3762 * ->css_offline() may show up during traversal. It's each subsystem's
3763 * responsibility to synchronize against on/offlining.
3765 struct cgroup_subsys_state
*
3766 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
3767 struct cgroup_subsys_state
*root
)
3769 struct cgroup_subsys_state
*next
;
3771 cgroup_assert_mutex_or_rcu_locked();
3773 /* if first iteration, visit leftmost descendant which may be @root */
3775 return css_leftmost_descendant(root
);
3777 /* if we visited @root, we're done */
3781 /* if there's an unvisited sibling, visit its leftmost descendant */
3782 next
= css_next_child(pos
, pos
->parent
);
3784 return css_leftmost_descendant(next
);
3786 /* no sibling left, visit parent */
3791 * css_has_online_children - does a css have online children
3792 * @css: the target css
3794 * Returns %true if @css has any online children; otherwise, %false. This
3795 * function can be called from any context but the caller is responsible
3796 * for synchronizing against on/offlining as necessary.
3798 bool css_has_online_children(struct cgroup_subsys_state
*css
)
3800 struct cgroup_subsys_state
*child
;
3804 css_for_each_child(child
, css
) {
3805 if (child
->flags
& CSS_ONLINE
) {
3815 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3816 * @it: the iterator to advance
3818 * Advance @it to the next css_set to walk.
3820 static void css_task_iter_advance_css_set(struct css_task_iter
*it
)
3822 struct list_head
*l
= it
->cset_pos
;
3823 struct cgrp_cset_link
*link
;
3824 struct css_set
*cset
;
3826 lockdep_assert_held(&css_set_rwsem
);
3828 /* Advance to the next non-empty css_set */
3831 if (l
== it
->cset_head
) {
3832 it
->cset_pos
= NULL
;
3833 it
->task_pos
= NULL
;
3838 cset
= container_of(l
, struct css_set
,
3839 e_cset_node
[it
->ss
->id
]);
3841 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
3844 } while (!css_set_populated(cset
));
3848 if (!list_empty(&cset
->tasks
))
3849 it
->task_pos
= cset
->tasks
.next
;
3851 it
->task_pos
= cset
->mg_tasks
.next
;
3853 it
->tasks_head
= &cset
->tasks
;
3854 it
->mg_tasks_head
= &cset
->mg_tasks
;
3857 * We don't keep css_sets locked across iteration steps and thus
3858 * need to take steps to ensure that iteration can be resumed after
3859 * the lock is re-acquired. Iteration is performed at two levels -
3860 * css_sets and tasks in them.
3862 * Once created, a css_set never leaves its cgroup lists, so a
3863 * pinned css_set is guaranteed to stay put and we can resume
3864 * iteration afterwards.
3866 * Tasks may leave @cset across iteration steps. This is resolved
3867 * by registering each iterator with the css_set currently being
3868 * walked and making css_set_move_task() advance iterators whose
3869 * next task is leaving.
3872 list_del(&it
->iters_node
);
3873 put_css_set_locked(it
->cur_cset
);
3876 it
->cur_cset
= cset
;
3877 list_add(&it
->iters_node
, &cset
->task_iters
);
3880 static void css_task_iter_advance(struct css_task_iter
*it
)
3882 struct list_head
*l
= it
->task_pos
;
3884 lockdep_assert_held(&css_set_rwsem
);
3888 * Advance iterator to find next entry. cset->tasks is consumed
3889 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3894 if (l
== it
->tasks_head
)
3895 l
= it
->mg_tasks_head
->next
;
3897 if (l
== it
->mg_tasks_head
)
3898 css_task_iter_advance_css_set(it
);
3904 * css_task_iter_start - initiate task iteration
3905 * @css: the css to walk tasks of
3906 * @it: the task iterator to use
3908 * Initiate iteration through the tasks of @css. The caller can call
3909 * css_task_iter_next() to walk through the tasks until the function
3910 * returns NULL. On completion of iteration, css_task_iter_end() must be
3913 void css_task_iter_start(struct cgroup_subsys_state
*css
,
3914 struct css_task_iter
*it
)
3916 /* no one should try to iterate before mounting cgroups */
3917 WARN_ON_ONCE(!use_task_css_set_links
);
3919 memset(it
, 0, sizeof(*it
));
3921 down_write(&css_set_rwsem
);
3926 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
3928 it
->cset_pos
= &css
->cgroup
->cset_links
;
3930 it
->cset_head
= it
->cset_pos
;
3932 css_task_iter_advance_css_set(it
);
3934 up_write(&css_set_rwsem
);
3938 * css_task_iter_next - return the next task for the iterator
3939 * @it: the task iterator being iterated
3941 * The "next" function for task iteration. @it should have been
3942 * initialized via css_task_iter_start(). Returns NULL when the iteration
3945 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
3951 put_task_struct(it
->cur_task
);
3953 down_write(&css_set_rwsem
);
3955 it
->cur_task
= list_entry(it
->task_pos
, struct task_struct
, cg_list
);
3956 get_task_struct(it
->cur_task
);
3958 css_task_iter_advance(it
);
3960 up_write(&css_set_rwsem
);
3962 return it
->cur_task
;
3966 * css_task_iter_end - finish task iteration
3967 * @it: the task iterator to finish
3969 * Finish task iteration started by css_task_iter_start().
3971 void css_task_iter_end(struct css_task_iter
*it
)
3974 down_write(&css_set_rwsem
);
3975 list_del(&it
->iters_node
);
3976 put_css_set_locked(it
->cur_cset
);
3977 up_write(&css_set_rwsem
);
3981 put_task_struct(it
->cur_task
);
3985 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3986 * @to: cgroup to which the tasks will be moved
3987 * @from: cgroup in which the tasks currently reside
3989 * Locking rules between cgroup_post_fork() and the migration path
3990 * guarantee that, if a task is forking while being migrated, the new child
3991 * is guaranteed to be either visible in the source cgroup after the
3992 * parent's migration is complete or put into the target cgroup. No task
3993 * can slip out of migration through forking.
3995 int cgroup_transfer_tasks(struct cgroup
*to
, struct cgroup
*from
)
3997 LIST_HEAD(preloaded_csets
);
3998 struct cgrp_cset_link
*link
;
3999 struct css_task_iter it
;
4000 struct task_struct
*task
;
4003 mutex_lock(&cgroup_mutex
);
4005 /* all tasks in @from are being moved, all csets are source */
4006 down_read(&css_set_rwsem
);
4007 list_for_each_entry(link
, &from
->cset_links
, cset_link
)
4008 cgroup_migrate_add_src(link
->cset
, to
, &preloaded_csets
);
4009 up_read(&css_set_rwsem
);
4011 ret
= cgroup_migrate_prepare_dst(to
, &preloaded_csets
);
4016 * Migrate tasks one-by-one until @form is empty. This fails iff
4017 * ->can_attach() fails.
4020 css_task_iter_start(&from
->self
, &it
);
4021 task
= css_task_iter_next(&it
);
4023 get_task_struct(task
);
4024 css_task_iter_end(&it
);
4027 ret
= cgroup_migrate(task
, false, to
);
4028 put_task_struct(task
);
4030 } while (task
&& !ret
);
4032 cgroup_migrate_finish(&preloaded_csets
);
4033 mutex_unlock(&cgroup_mutex
);
4038 * Stuff for reading the 'tasks'/'procs' files.
4040 * Reading this file can return large amounts of data if a cgroup has
4041 * *lots* of attached tasks. So it may need several calls to read(),
4042 * but we cannot guarantee that the information we produce is correct
4043 * unless we produce it entirely atomically.
4047 /* which pidlist file are we talking about? */
4048 enum cgroup_filetype
{
4054 * A pidlist is a list of pids that virtually represents the contents of one
4055 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4056 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4059 struct cgroup_pidlist
{
4061 * used to find which pidlist is wanted. doesn't change as long as
4062 * this particular list stays in the list.
4064 struct { enum cgroup_filetype type
; struct pid_namespace
*ns
; } key
;
4067 /* how many elements the above list has */
4069 /* each of these stored in a list by its cgroup */
4070 struct list_head links
;
4071 /* pointer to the cgroup we belong to, for list removal purposes */
4072 struct cgroup
*owner
;
4073 /* for delayed destruction */
4074 struct delayed_work destroy_dwork
;
4078 * The following two functions "fix" the issue where there are more pids
4079 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4080 * TODO: replace with a kernel-wide solution to this problem
4082 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4083 static void *pidlist_allocate(int count
)
4085 if (PIDLIST_TOO_LARGE(count
))
4086 return vmalloc(count
* sizeof(pid_t
));
4088 return kmalloc(count
* sizeof(pid_t
), GFP_KERNEL
);
4091 static void pidlist_free(void *p
)
4097 * Used to destroy all pidlists lingering waiting for destroy timer. None
4098 * should be left afterwards.
4100 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
)
4102 struct cgroup_pidlist
*l
, *tmp_l
;
4104 mutex_lock(&cgrp
->pidlist_mutex
);
4105 list_for_each_entry_safe(l
, tmp_l
, &cgrp
->pidlists
, links
)
4106 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
, 0);
4107 mutex_unlock(&cgrp
->pidlist_mutex
);
4109 flush_workqueue(cgroup_pidlist_destroy_wq
);
4110 BUG_ON(!list_empty(&cgrp
->pidlists
));
4113 static void cgroup_pidlist_destroy_work_fn(struct work_struct
*work
)
4115 struct delayed_work
*dwork
= to_delayed_work(work
);
4116 struct cgroup_pidlist
*l
= container_of(dwork
, struct cgroup_pidlist
,
4118 struct cgroup_pidlist
*tofree
= NULL
;
4120 mutex_lock(&l
->owner
->pidlist_mutex
);
4123 * Destroy iff we didn't get queued again. The state won't change
4124 * as destroy_dwork can only be queued while locked.
4126 if (!delayed_work_pending(dwork
)) {
4127 list_del(&l
->links
);
4128 pidlist_free(l
->list
);
4129 put_pid_ns(l
->key
.ns
);
4133 mutex_unlock(&l
->owner
->pidlist_mutex
);
4138 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4139 * Returns the number of unique elements.
4141 static int pidlist_uniq(pid_t
*list
, int length
)
4146 * we presume the 0th element is unique, so i starts at 1. trivial
4147 * edge cases first; no work needs to be done for either
4149 if (length
== 0 || length
== 1)
4151 /* src and dest walk down the list; dest counts unique elements */
4152 for (src
= 1; src
< length
; src
++) {
4153 /* find next unique element */
4154 while (list
[src
] == list
[src
-1]) {
4159 /* dest always points to where the next unique element goes */
4160 list
[dest
] = list
[src
];
4168 * The two pid files - task and cgroup.procs - guaranteed that the result
4169 * is sorted, which forced this whole pidlist fiasco. As pid order is
4170 * different per namespace, each namespace needs differently sorted list,
4171 * making it impossible to use, for example, single rbtree of member tasks
4172 * sorted by task pointer. As pidlists can be fairly large, allocating one
4173 * per open file is dangerous, so cgroup had to implement shared pool of
4174 * pidlists keyed by cgroup and namespace.
4176 * All this extra complexity was caused by the original implementation
4177 * committing to an entirely unnecessary property. In the long term, we
4178 * want to do away with it. Explicitly scramble sort order if on the
4179 * default hierarchy so that no such expectation exists in the new
4182 * Scrambling is done by swapping every two consecutive bits, which is
4183 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4185 static pid_t
pid_fry(pid_t pid
)
4187 unsigned a
= pid
& 0x55555555;
4188 unsigned b
= pid
& 0xAAAAAAAA;
4190 return (a
<< 1) | (b
>> 1);
4193 static pid_t
cgroup_pid_fry(struct cgroup
*cgrp
, pid_t pid
)
4195 if (cgroup_on_dfl(cgrp
))
4196 return pid_fry(pid
);
4201 static int cmppid(const void *a
, const void *b
)
4203 return *(pid_t
*)a
- *(pid_t
*)b
;
4206 static int fried_cmppid(const void *a
, const void *b
)
4208 return pid_fry(*(pid_t
*)a
) - pid_fry(*(pid_t
*)b
);
4211 static struct cgroup_pidlist
*cgroup_pidlist_find(struct cgroup
*cgrp
,
4212 enum cgroup_filetype type
)
4214 struct cgroup_pidlist
*l
;
4215 /* don't need task_nsproxy() if we're looking at ourself */
4216 struct pid_namespace
*ns
= task_active_pid_ns(current
);
4218 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4220 list_for_each_entry(l
, &cgrp
->pidlists
, links
)
4221 if (l
->key
.type
== type
&& l
->key
.ns
== ns
)
4227 * find the appropriate pidlist for our purpose (given procs vs tasks)
4228 * returns with the lock on that pidlist already held, and takes care
4229 * of the use count, or returns NULL with no locks held if we're out of
4232 static struct cgroup_pidlist
*cgroup_pidlist_find_create(struct cgroup
*cgrp
,
4233 enum cgroup_filetype type
)
4235 struct cgroup_pidlist
*l
;
4237 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4239 l
= cgroup_pidlist_find(cgrp
, type
);
4243 /* entry not found; create a new one */
4244 l
= kzalloc(sizeof(struct cgroup_pidlist
), GFP_KERNEL
);
4248 INIT_DELAYED_WORK(&l
->destroy_dwork
, cgroup_pidlist_destroy_work_fn
);
4250 /* don't need task_nsproxy() if we're looking at ourself */
4251 l
->key
.ns
= get_pid_ns(task_active_pid_ns(current
));
4253 list_add(&l
->links
, &cgrp
->pidlists
);
4258 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4260 static int pidlist_array_load(struct cgroup
*cgrp
, enum cgroup_filetype type
,
4261 struct cgroup_pidlist
**lp
)
4265 int pid
, n
= 0; /* used for populating the array */
4266 struct css_task_iter it
;
4267 struct task_struct
*tsk
;
4268 struct cgroup_pidlist
*l
;
4270 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4273 * If cgroup gets more users after we read count, we won't have
4274 * enough space - tough. This race is indistinguishable to the
4275 * caller from the case that the additional cgroup users didn't
4276 * show up until sometime later on.
4278 length
= cgroup_task_count(cgrp
);
4279 array
= pidlist_allocate(length
);
4282 /* now, populate the array */
4283 css_task_iter_start(&cgrp
->self
, &it
);
4284 while ((tsk
= css_task_iter_next(&it
))) {
4285 if (unlikely(n
== length
))
4287 /* get tgid or pid for procs or tasks file respectively */
4288 if (type
== CGROUP_FILE_PROCS
)
4289 pid
= task_tgid_vnr(tsk
);
4291 pid
= task_pid_vnr(tsk
);
4292 if (pid
> 0) /* make sure to only use valid results */
4295 css_task_iter_end(&it
);
4297 /* now sort & (if procs) strip out duplicates */
4298 if (cgroup_on_dfl(cgrp
))
4299 sort(array
, length
, sizeof(pid_t
), fried_cmppid
, NULL
);
4301 sort(array
, length
, sizeof(pid_t
), cmppid
, NULL
);
4302 if (type
== CGROUP_FILE_PROCS
)
4303 length
= pidlist_uniq(array
, length
);
4305 l
= cgroup_pidlist_find_create(cgrp
, type
);
4307 pidlist_free(array
);
4311 /* store array, freeing old if necessary */
4312 pidlist_free(l
->list
);
4320 * cgroupstats_build - build and fill cgroupstats
4321 * @stats: cgroupstats to fill information into
4322 * @dentry: A dentry entry belonging to the cgroup for which stats have
4325 * Build and fill cgroupstats so that taskstats can export it to user
4328 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
4330 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
4331 struct cgroup
*cgrp
;
4332 struct css_task_iter it
;
4333 struct task_struct
*tsk
;
4335 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4336 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
4337 kernfs_type(kn
) != KERNFS_DIR
)
4340 mutex_lock(&cgroup_mutex
);
4343 * We aren't being called from kernfs and there's no guarantee on
4344 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4345 * @kn->priv is RCU safe. Let's do the RCU dancing.
4348 cgrp
= rcu_dereference(kn
->priv
);
4349 if (!cgrp
|| cgroup_is_dead(cgrp
)) {
4351 mutex_unlock(&cgroup_mutex
);
4356 css_task_iter_start(&cgrp
->self
, &it
);
4357 while ((tsk
= css_task_iter_next(&it
))) {
4358 switch (tsk
->state
) {
4360 stats
->nr_running
++;
4362 case TASK_INTERRUPTIBLE
:
4363 stats
->nr_sleeping
++;
4365 case TASK_UNINTERRUPTIBLE
:
4366 stats
->nr_uninterruptible
++;
4369 stats
->nr_stopped
++;
4372 if (delayacct_is_task_waiting_on_io(tsk
))
4373 stats
->nr_io_wait
++;
4377 css_task_iter_end(&it
);
4379 mutex_unlock(&cgroup_mutex
);
4385 * seq_file methods for the tasks/procs files. The seq_file position is the
4386 * next pid to display; the seq_file iterator is a pointer to the pid
4387 * in the cgroup->l->list array.
4390 static void *cgroup_pidlist_start(struct seq_file
*s
, loff_t
*pos
)
4393 * Initially we receive a position value that corresponds to
4394 * one more than the last pid shown (or 0 on the first call or
4395 * after a seek to the start). Use a binary-search to find the
4396 * next pid to display, if any
4398 struct kernfs_open_file
*of
= s
->private;
4399 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4400 struct cgroup_pidlist
*l
;
4401 enum cgroup_filetype type
= seq_cft(s
)->private;
4402 int index
= 0, pid
= *pos
;
4405 mutex_lock(&cgrp
->pidlist_mutex
);
4408 * !NULL @of->priv indicates that this isn't the first start()
4409 * after open. If the matching pidlist is around, we can use that.
4410 * Look for it. Note that @of->priv can't be used directly. It
4411 * could already have been destroyed.
4414 of
->priv
= cgroup_pidlist_find(cgrp
, type
);
4417 * Either this is the first start() after open or the matching
4418 * pidlist has been destroyed inbetween. Create a new one.
4421 ret
= pidlist_array_load(cgrp
, type
,
4422 (struct cgroup_pidlist
**)&of
->priv
);
4424 return ERR_PTR(ret
);
4429 int end
= l
->length
;
4431 while (index
< end
) {
4432 int mid
= (index
+ end
) / 2;
4433 if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) == pid
) {
4436 } else if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) <= pid
)
4442 /* If we're off the end of the array, we're done */
4443 if (index
>= l
->length
)
4445 /* Update the abstract position to be the actual pid that we found */
4446 iter
= l
->list
+ index
;
4447 *pos
= cgroup_pid_fry(cgrp
, *iter
);
4451 static void cgroup_pidlist_stop(struct seq_file
*s
, void *v
)
4453 struct kernfs_open_file
*of
= s
->private;
4454 struct cgroup_pidlist
*l
= of
->priv
;
4457 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
,
4458 CGROUP_PIDLIST_DESTROY_DELAY
);
4459 mutex_unlock(&seq_css(s
)->cgroup
->pidlist_mutex
);
4462 static void *cgroup_pidlist_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
4464 struct kernfs_open_file
*of
= s
->private;
4465 struct cgroup_pidlist
*l
= of
->priv
;
4467 pid_t
*end
= l
->list
+ l
->length
;
4469 * Advance to the next pid in the array. If this goes off the
4476 *pos
= cgroup_pid_fry(seq_css(s
)->cgroup
, *p
);
4481 static int cgroup_pidlist_show(struct seq_file
*s
, void *v
)
4483 seq_printf(s
, "%d\n", *(int *)v
);
4488 static u64
cgroup_read_notify_on_release(struct cgroup_subsys_state
*css
,
4491 return notify_on_release(css
->cgroup
);
4494 static int cgroup_write_notify_on_release(struct cgroup_subsys_state
*css
,
4495 struct cftype
*cft
, u64 val
)
4498 set_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4500 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4504 static u64
cgroup_clone_children_read(struct cgroup_subsys_state
*css
,
4507 return test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4510 static int cgroup_clone_children_write(struct cgroup_subsys_state
*css
,
4511 struct cftype
*cft
, u64 val
)
4514 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4516 clear_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4520 /* cgroup core interface files for the default hierarchy */
4521 static struct cftype cgroup_dfl_base_files
[] = {
4523 .name
= "cgroup.procs",
4524 .file_offset
= offsetof(struct cgroup
, procs_file
),
4525 .seq_start
= cgroup_pidlist_start
,
4526 .seq_next
= cgroup_pidlist_next
,
4527 .seq_stop
= cgroup_pidlist_stop
,
4528 .seq_show
= cgroup_pidlist_show
,
4529 .private = CGROUP_FILE_PROCS
,
4530 .write
= cgroup_procs_write
,
4533 .name
= "cgroup.controllers",
4534 .flags
= CFTYPE_ONLY_ON_ROOT
,
4535 .seq_show
= cgroup_root_controllers_show
,
4538 .name
= "cgroup.controllers",
4539 .flags
= CFTYPE_NOT_ON_ROOT
,
4540 .seq_show
= cgroup_controllers_show
,
4543 .name
= "cgroup.subtree_control",
4544 .seq_show
= cgroup_subtree_control_show
,
4545 .write
= cgroup_subtree_control_write
,
4548 .name
= "cgroup.events",
4549 .flags
= CFTYPE_NOT_ON_ROOT
,
4550 .file_offset
= offsetof(struct cgroup
, events_file
),
4551 .seq_show
= cgroup_events_show
,
4556 /* cgroup core interface files for the legacy hierarchies */
4557 static struct cftype cgroup_legacy_base_files
[] = {
4559 .name
= "cgroup.procs",
4560 .seq_start
= cgroup_pidlist_start
,
4561 .seq_next
= cgroup_pidlist_next
,
4562 .seq_stop
= cgroup_pidlist_stop
,
4563 .seq_show
= cgroup_pidlist_show
,
4564 .private = CGROUP_FILE_PROCS
,
4565 .write
= cgroup_procs_write
,
4568 .name
= "cgroup.clone_children",
4569 .read_u64
= cgroup_clone_children_read
,
4570 .write_u64
= cgroup_clone_children_write
,
4573 .name
= "cgroup.sane_behavior",
4574 .flags
= CFTYPE_ONLY_ON_ROOT
,
4575 .seq_show
= cgroup_sane_behavior_show
,
4579 .seq_start
= cgroup_pidlist_start
,
4580 .seq_next
= cgroup_pidlist_next
,
4581 .seq_stop
= cgroup_pidlist_stop
,
4582 .seq_show
= cgroup_pidlist_show
,
4583 .private = CGROUP_FILE_TASKS
,
4584 .write
= cgroup_tasks_write
,
4587 .name
= "notify_on_release",
4588 .read_u64
= cgroup_read_notify_on_release
,
4589 .write_u64
= cgroup_write_notify_on_release
,
4592 .name
= "release_agent",
4593 .flags
= CFTYPE_ONLY_ON_ROOT
,
4594 .seq_show
= cgroup_release_agent_show
,
4595 .write
= cgroup_release_agent_write
,
4596 .max_write_len
= PATH_MAX
- 1,
4602 * css destruction is four-stage process.
4604 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4605 * Implemented in kill_css().
4607 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4608 * and thus css_tryget_online() is guaranteed to fail, the css can be
4609 * offlined by invoking offline_css(). After offlining, the base ref is
4610 * put. Implemented in css_killed_work_fn().
4612 * 3. When the percpu_ref reaches zero, the only possible remaining
4613 * accessors are inside RCU read sections. css_release() schedules the
4616 * 4. After the grace period, the css can be freed. Implemented in
4617 * css_free_work_fn().
4619 * It is actually hairier because both step 2 and 4 require process context
4620 * and thus involve punting to css->destroy_work adding two additional
4621 * steps to the already complex sequence.
4623 static void css_free_work_fn(struct work_struct
*work
)
4625 struct cgroup_subsys_state
*css
=
4626 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4627 struct cgroup_subsys
*ss
= css
->ss
;
4628 struct cgroup
*cgrp
= css
->cgroup
;
4629 struct cgroup_file
*cfile
;
4631 percpu_ref_exit(&css
->refcnt
);
4633 list_for_each_entry(cfile
, &css
->files
, node
)
4634 kernfs_put(cfile
->kn
);
4641 css_put(css
->parent
);
4644 cgroup_idr_remove(&ss
->css_idr
, id
);
4647 /* cgroup free path */
4648 atomic_dec(&cgrp
->root
->nr_cgrps
);
4649 cgroup_pidlist_destroy_all(cgrp
);
4650 cancel_work_sync(&cgrp
->release_agent_work
);
4652 if (cgroup_parent(cgrp
)) {
4654 * We get a ref to the parent, and put the ref when
4655 * this cgroup is being freed, so it's guaranteed
4656 * that the parent won't be destroyed before its
4659 cgroup_put(cgroup_parent(cgrp
));
4660 kernfs_put(cgrp
->kn
);
4664 * This is root cgroup's refcnt reaching zero,
4665 * which indicates that the root should be
4668 cgroup_destroy_root(cgrp
->root
);
4673 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4675 struct cgroup_subsys_state
*css
=
4676 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4678 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4679 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4682 static void css_release_work_fn(struct work_struct
*work
)
4684 struct cgroup_subsys_state
*css
=
4685 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4686 struct cgroup_subsys
*ss
= css
->ss
;
4687 struct cgroup
*cgrp
= css
->cgroup
;
4689 mutex_lock(&cgroup_mutex
);
4691 css
->flags
|= CSS_RELEASED
;
4692 list_del_rcu(&css
->sibling
);
4695 /* css release path */
4696 cgroup_idr_replace(&ss
->css_idr
, NULL
, css
->id
);
4697 if (ss
->css_released
)
4698 ss
->css_released(css
);
4700 /* cgroup release path */
4701 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
4705 * There are two control paths which try to determine
4706 * cgroup from dentry without going through kernfs -
4707 * cgroupstats_build() and css_tryget_online_from_dir().
4708 * Those are supported by RCU protecting clearing of
4709 * cgrp->kn->priv backpointer.
4711 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
, NULL
);
4714 mutex_unlock(&cgroup_mutex
);
4716 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4719 static void css_release(struct percpu_ref
*ref
)
4721 struct cgroup_subsys_state
*css
=
4722 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4724 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
4725 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4728 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4729 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4731 lockdep_assert_held(&cgroup_mutex
);
4735 memset(css
, 0, sizeof(*css
));
4738 INIT_LIST_HEAD(&css
->sibling
);
4739 INIT_LIST_HEAD(&css
->children
);
4740 INIT_LIST_HEAD(&css
->files
);
4741 css
->serial_nr
= css_serial_nr_next
++;
4743 if (cgroup_parent(cgrp
)) {
4744 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
4745 css_get(css
->parent
);
4748 BUG_ON(cgroup_css(cgrp
, ss
));
4751 /* invoke ->css_online() on a new CSS and mark it online if successful */
4752 static int online_css(struct cgroup_subsys_state
*css
)
4754 struct cgroup_subsys
*ss
= css
->ss
;
4757 lockdep_assert_held(&cgroup_mutex
);
4760 ret
= ss
->css_online(css
);
4762 css
->flags
|= CSS_ONLINE
;
4763 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4768 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4769 static void offline_css(struct cgroup_subsys_state
*css
)
4771 struct cgroup_subsys
*ss
= css
->ss
;
4773 lockdep_assert_held(&cgroup_mutex
);
4775 if (!(css
->flags
& CSS_ONLINE
))
4778 if (ss
->css_offline
)
4779 ss
->css_offline(css
);
4781 css
->flags
&= ~CSS_ONLINE
;
4782 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4784 wake_up_all(&css
->cgroup
->offline_waitq
);
4788 * create_css - create a cgroup_subsys_state
4789 * @cgrp: the cgroup new css will be associated with
4790 * @ss: the subsys of new css
4791 * @visible: whether to create control knobs for the new css or not
4793 * Create a new css associated with @cgrp - @ss pair. On success, the new
4794 * css is online and installed in @cgrp with all interface files created if
4795 * @visible. Returns 0 on success, -errno on failure.
4797 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
4800 struct cgroup
*parent
= cgroup_parent(cgrp
);
4801 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
4802 struct cgroup_subsys_state
*css
;
4805 lockdep_assert_held(&cgroup_mutex
);
4807 css
= ss
->css_alloc(parent_css
);
4809 return PTR_ERR(css
);
4811 init_and_link_css(css
, ss
, cgrp
);
4813 err
= percpu_ref_init(&css
->refcnt
, css_release
, 0, GFP_KERNEL
);
4817 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_KERNEL
);
4819 goto err_free_percpu_ref
;
4823 err
= css_populate_dir(css
, NULL
);
4828 /* @css is ready to be brought online now, make it visible */
4829 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
4830 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4832 err
= online_css(css
);
4836 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4837 cgroup_parent(parent
)) {
4838 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4839 current
->comm
, current
->pid
, ss
->name
);
4840 if (!strcmp(ss
->name
, "memory"))
4841 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4842 ss
->warned_broken_hierarchy
= true;
4848 list_del_rcu(&css
->sibling
);
4849 css_clear_dir(css
, NULL
);
4851 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4852 err_free_percpu_ref
:
4853 percpu_ref_exit(&css
->refcnt
);
4855 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4859 static int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
,
4862 struct cgroup
*parent
, *cgrp
;
4863 struct cgroup_root
*root
;
4864 struct cgroup_subsys
*ss
;
4865 struct kernfs_node
*kn
;
4868 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4870 if (strchr(name
, '\n'))
4873 parent
= cgroup_kn_lock_live(parent_kn
);
4876 root
= parent
->root
;
4878 /* allocate the cgroup and its ID, 0 is reserved for the root */
4879 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
4885 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
, 0, GFP_KERNEL
);
4890 * Temporarily set the pointer to NULL, so idr_find() won't return
4891 * a half-baked cgroup.
4893 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_KERNEL
);
4896 goto out_cancel_ref
;
4899 init_cgroup_housekeeping(cgrp
);
4901 cgrp
->self
.parent
= &parent
->self
;
4904 if (notify_on_release(parent
))
4905 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4907 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4908 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4910 /* create the directory */
4911 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4919 * This extra ref will be put in cgroup_free_fn() and guarantees
4920 * that @cgrp->kn is always accessible.
4924 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
4926 /* allocation complete, commit to creation */
4927 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
4928 atomic_inc(&root
->nr_cgrps
);
4932 * @cgrp is now fully operational. If something fails after this
4933 * point, it'll be released via the normal destruction path.
4935 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4937 ret
= cgroup_kn_set_ugid(kn
);
4941 ret
= css_populate_dir(&cgrp
->self
, NULL
);
4945 /* let's create and online css's */
4946 for_each_subsys(ss
, ssid
) {
4947 if (parent
->child_subsys_mask
& (1 << ssid
)) {
4948 ret
= create_css(cgrp
, ss
,
4949 parent
->subtree_control
& (1 << ssid
));
4956 * On the default hierarchy, a child doesn't automatically inherit
4957 * subtree_control from the parent. Each is configured manually.
4959 if (!cgroup_on_dfl(cgrp
)) {
4960 cgrp
->subtree_control
= parent
->subtree_control
;
4961 cgroup_refresh_child_subsys_mask(cgrp
);
4964 kernfs_activate(kn
);
4970 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4972 percpu_ref_exit(&cgrp
->self
.refcnt
);
4976 cgroup_kn_unlock(parent_kn
);
4980 cgroup_destroy_locked(cgrp
);
4985 * This is called when the refcnt of a css is confirmed to be killed.
4986 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4987 * initate destruction and put the css ref from kill_css().
4989 static void css_killed_work_fn(struct work_struct
*work
)
4991 struct cgroup_subsys_state
*css
=
4992 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4994 mutex_lock(&cgroup_mutex
);
4996 mutex_unlock(&cgroup_mutex
);
5001 /* css kill confirmation processing requires process context, bounce */
5002 static void css_killed_ref_fn(struct percpu_ref
*ref
)
5004 struct cgroup_subsys_state
*css
=
5005 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
5007 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
5008 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
5012 * kill_css - destroy a css
5013 * @css: css to destroy
5015 * This function initiates destruction of @css by removing cgroup interface
5016 * files and putting its base reference. ->css_offline() will be invoked
5017 * asynchronously once css_tryget_online() is guaranteed to fail and when
5018 * the reference count reaches zero, @css will be released.
5020 static void kill_css(struct cgroup_subsys_state
*css
)
5022 lockdep_assert_held(&cgroup_mutex
);
5025 * This must happen before css is disassociated with its cgroup.
5026 * See seq_css() for details.
5028 css_clear_dir(css
, NULL
);
5031 * Killing would put the base ref, but we need to keep it alive
5032 * until after ->css_offline().
5037 * cgroup core guarantees that, by the time ->css_offline() is
5038 * invoked, no new css reference will be given out via
5039 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5040 * proceed to offlining css's because percpu_ref_kill() doesn't
5041 * guarantee that the ref is seen as killed on all CPUs on return.
5043 * Use percpu_ref_kill_and_confirm() to get notifications as each
5044 * css is confirmed to be seen as killed on all CPUs.
5046 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
5050 * cgroup_destroy_locked - the first stage of cgroup destruction
5051 * @cgrp: cgroup to be destroyed
5053 * css's make use of percpu refcnts whose killing latency shouldn't be
5054 * exposed to userland and are RCU protected. Also, cgroup core needs to
5055 * guarantee that css_tryget_online() won't succeed by the time
5056 * ->css_offline() is invoked. To satisfy all the requirements,
5057 * destruction is implemented in the following two steps.
5059 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5060 * userland visible parts and start killing the percpu refcnts of
5061 * css's. Set up so that the next stage will be kicked off once all
5062 * the percpu refcnts are confirmed to be killed.
5064 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5065 * rest of destruction. Once all cgroup references are gone, the
5066 * cgroup is RCU-freed.
5068 * This function implements s1. After this step, @cgrp is gone as far as
5069 * the userland is concerned and a new cgroup with the same name may be
5070 * created. As cgroup doesn't care about the names internally, this
5071 * doesn't cause any problem.
5073 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
5074 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
5076 struct cgroup_subsys_state
*css
;
5079 lockdep_assert_held(&cgroup_mutex
);
5082 * Only migration can raise populated from zero and we're already
5083 * holding cgroup_mutex.
5085 if (cgroup_is_populated(cgrp
))
5089 * Make sure there's no live children. We can't test emptiness of
5090 * ->self.children as dead children linger on it while being
5091 * drained; otherwise, "rmdir parent/child parent" may fail.
5093 if (css_has_online_children(&cgrp
->self
))
5097 * Mark @cgrp dead. This prevents further task migration and child
5098 * creation by disabling cgroup_lock_live_group().
5100 cgrp
->self
.flags
&= ~CSS_ONLINE
;
5102 /* initiate massacre of all css's */
5103 for_each_css(css
, ssid
, cgrp
)
5107 * Remove @cgrp directory along with the base files. @cgrp has an
5108 * extra ref on its kn.
5110 kernfs_remove(cgrp
->kn
);
5112 check_for_release(cgroup_parent(cgrp
));
5114 /* put the base reference */
5115 percpu_ref_kill(&cgrp
->self
.refcnt
);
5120 static int cgroup_rmdir(struct kernfs_node
*kn
)
5122 struct cgroup
*cgrp
;
5125 cgrp
= cgroup_kn_lock_live(kn
);
5129 ret
= cgroup_destroy_locked(cgrp
);
5131 cgroup_kn_unlock(kn
);
5135 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
5136 .remount_fs
= cgroup_remount
,
5137 .show_options
= cgroup_show_options
,
5138 .mkdir
= cgroup_mkdir
,
5139 .rmdir
= cgroup_rmdir
,
5140 .rename
= cgroup_rename
,
5143 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
5145 struct cgroup_subsys_state
*css
;
5147 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
5149 mutex_lock(&cgroup_mutex
);
5151 idr_init(&ss
->css_idr
);
5152 INIT_LIST_HEAD(&ss
->cfts
);
5154 /* Create the root cgroup state for this subsystem */
5155 ss
->root
= &cgrp_dfl_root
;
5156 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
5157 /* We don't handle early failures gracefully */
5158 BUG_ON(IS_ERR(css
));
5159 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
5162 * Root csses are never destroyed and we can't initialize
5163 * percpu_ref during early init. Disable refcnting.
5165 css
->flags
|= CSS_NO_REF
;
5168 /* allocation can't be done safely during early init */
5171 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
5172 BUG_ON(css
->id
< 0);
5175 /* Update the init_css_set to contain a subsys
5176 * pointer to this state - since the subsystem is
5177 * newly registered, all tasks and hence the
5178 * init_css_set is in the subsystem's root cgroup. */
5179 init_css_set
.subsys
[ss
->id
] = css
;
5181 have_fork_callback
|= (bool)ss
->fork
<< ss
->id
;
5182 have_exit_callback
|= (bool)ss
->exit
<< ss
->id
;
5183 have_canfork_callback
|= (bool)ss
->can_fork
<< ss
->id
;
5185 /* At system boot, before all subsystems have been
5186 * registered, no tasks have been forked, so we don't
5187 * need to invoke fork callbacks here. */
5188 BUG_ON(!list_empty(&init_task
.tasks
));
5190 BUG_ON(online_css(css
));
5192 mutex_unlock(&cgroup_mutex
);
5196 * cgroup_init_early - cgroup initialization at system boot
5198 * Initialize cgroups at system boot, and initialize any
5199 * subsystems that request early init.
5201 int __init
cgroup_init_early(void)
5203 static struct cgroup_sb_opts __initdata opts
;
5204 struct cgroup_subsys
*ss
;
5207 init_cgroup_root(&cgrp_dfl_root
, &opts
);
5208 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
5210 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
5212 for_each_subsys(ss
, i
) {
5213 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
5214 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5215 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
5217 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
5218 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
5221 ss
->name
= cgroup_subsys_name
[i
];
5222 if (!ss
->legacy_name
)
5223 ss
->legacy_name
= cgroup_subsys_name
[i
];
5226 cgroup_init_subsys(ss
, true);
5231 static unsigned long cgroup_disable_mask __initdata
;
5234 * cgroup_init - cgroup initialization
5236 * Register cgroup filesystem and /proc file, and initialize
5237 * any subsystems that didn't request early init.
5239 int __init
cgroup_init(void)
5241 struct cgroup_subsys
*ss
;
5245 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem
));
5246 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_dfl_base_files
));
5247 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_legacy_base_files
));
5249 mutex_lock(&cgroup_mutex
);
5251 /* Add init_css_set to the hash table */
5252 key
= css_set_hash(init_css_set
.subsys
);
5253 hash_add(css_set_table
, &init_css_set
.hlist
, key
);
5255 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
5257 mutex_unlock(&cgroup_mutex
);
5259 for_each_subsys(ss
, ssid
) {
5260 if (ss
->early_init
) {
5261 struct cgroup_subsys_state
*css
=
5262 init_css_set
.subsys
[ss
->id
];
5264 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
5266 BUG_ON(css
->id
< 0);
5268 cgroup_init_subsys(ss
, false);
5271 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
5272 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
5275 * Setting dfl_root subsys_mask needs to consider the
5276 * disabled flag and cftype registration needs kmalloc,
5277 * both of which aren't available during early_init.
5279 if (cgroup_disable_mask
& (1 << ssid
)) {
5280 static_branch_disable(cgroup_subsys_enabled_key
[ssid
]);
5281 printk(KERN_INFO
"Disabling %s control group subsystem\n",
5286 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
5288 if (cgroup_legacy_files_on_dfl
&& !ss
->dfl_cftypes
)
5289 ss
->dfl_cftypes
= ss
->legacy_cftypes
;
5291 if (!ss
->dfl_cftypes
)
5292 cgrp_dfl_root_inhibit_ss_mask
|= 1 << ss
->id
;
5294 if (ss
->dfl_cftypes
== ss
->legacy_cftypes
) {
5295 WARN_ON(cgroup_add_cftypes(ss
, ss
->dfl_cftypes
));
5297 WARN_ON(cgroup_add_dfl_cftypes(ss
, ss
->dfl_cftypes
));
5298 WARN_ON(cgroup_add_legacy_cftypes(ss
, ss
->legacy_cftypes
));
5302 ss
->bind(init_css_set
.subsys
[ssid
]);
5305 err
= sysfs_create_mount_point(fs_kobj
, "cgroup");
5309 err
= register_filesystem(&cgroup_fs_type
);
5311 sysfs_remove_mount_point(fs_kobj
, "cgroup");
5315 proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
);
5319 static int __init
cgroup_wq_init(void)
5322 * There isn't much point in executing destruction path in
5323 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5324 * Use 1 for @max_active.
5326 * We would prefer to do this in cgroup_init() above, but that
5327 * is called before init_workqueues(): so leave this until after.
5329 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
5330 BUG_ON(!cgroup_destroy_wq
);
5333 * Used to destroy pidlists and separate to serve as flush domain.
5334 * Cap @max_active to 1 too.
5336 cgroup_pidlist_destroy_wq
= alloc_workqueue("cgroup_pidlist_destroy",
5338 BUG_ON(!cgroup_pidlist_destroy_wq
);
5342 core_initcall(cgroup_wq_init
);
5345 * proc_cgroup_show()
5346 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5347 * - Used for /proc/<pid>/cgroup.
5349 int proc_cgroup_show(struct seq_file
*m
, struct pid_namespace
*ns
,
5350 struct pid
*pid
, struct task_struct
*tsk
)
5354 struct cgroup_root
*root
;
5357 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5361 mutex_lock(&cgroup_mutex
);
5362 down_read(&css_set_rwsem
);
5364 for_each_root(root
) {
5365 struct cgroup_subsys
*ss
;
5366 struct cgroup
*cgrp
;
5367 int ssid
, count
= 0;
5369 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_root_visible
)
5372 seq_printf(m
, "%d:", root
->hierarchy_id
);
5373 if (root
!= &cgrp_dfl_root
)
5374 for_each_subsys(ss
, ssid
)
5375 if (root
->subsys_mask
& (1 << ssid
))
5376 seq_printf(m
, "%s%s", count
++ ? "," : "",
5378 if (strlen(root
->name
))
5379 seq_printf(m
, "%sname=%s", count
? "," : "",
5382 cgrp
= task_cgroup_from_root(tsk
, root
);
5383 path
= cgroup_path(cgrp
, buf
, PATH_MAX
);
5385 retval
= -ENAMETOOLONG
;
5394 up_read(&css_set_rwsem
);
5395 mutex_unlock(&cgroup_mutex
);
5401 /* Display information about each subsystem and each hierarchy */
5402 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
5404 struct cgroup_subsys
*ss
;
5407 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5409 * ideally we don't want subsystems moving around while we do this.
5410 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5411 * subsys/hierarchy state.
5413 mutex_lock(&cgroup_mutex
);
5415 for_each_subsys(ss
, i
)
5416 seq_printf(m
, "%s\t%d\t%d\t%d\n",
5417 ss
->legacy_name
, ss
->root
->hierarchy_id
,
5418 atomic_read(&ss
->root
->nr_cgrps
),
5419 cgroup_ssid_enabled(i
));
5421 mutex_unlock(&cgroup_mutex
);
5425 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
5427 return single_open(file
, proc_cgroupstats_show
, NULL
);
5430 static const struct file_operations proc_cgroupstats_operations
= {
5431 .open
= cgroupstats_open
,
5433 .llseek
= seq_lseek
,
5434 .release
= single_release
,
5437 static void **subsys_canfork_priv_p(void *ss_priv
[CGROUP_CANFORK_COUNT
], int i
)
5439 if (CGROUP_CANFORK_START
<= i
&& i
< CGROUP_CANFORK_END
)
5440 return &ss_priv
[i
- CGROUP_CANFORK_START
];
5444 static void *subsys_canfork_priv(void *ss_priv
[CGROUP_CANFORK_COUNT
], int i
)
5446 void **private = subsys_canfork_priv_p(ss_priv
, i
);
5447 return private ? *private : NULL
;
5451 * cgroup_fork - initialize cgroup related fields during copy_process()
5452 * @child: pointer to task_struct of forking parent process.
5454 * A task is associated with the init_css_set until cgroup_post_fork()
5455 * attaches it to the parent's css_set. Empty cg_list indicates that
5456 * @child isn't holding reference to its css_set.
5458 void cgroup_fork(struct task_struct
*child
)
5460 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
5461 INIT_LIST_HEAD(&child
->cg_list
);
5465 * cgroup_can_fork - called on a new task before the process is exposed
5466 * @child: the task in question.
5468 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5469 * returns an error, the fork aborts with that error code. This allows for
5470 * a cgroup subsystem to conditionally allow or deny new forks.
5472 int cgroup_can_fork(struct task_struct
*child
,
5473 void *ss_priv
[CGROUP_CANFORK_COUNT
])
5475 struct cgroup_subsys
*ss
;
5478 for_each_subsys_which(ss
, i
, &have_canfork_callback
) {
5479 ret
= ss
->can_fork(child
, subsys_canfork_priv_p(ss_priv
, i
));
5487 for_each_subsys(ss
, j
) {
5490 if (ss
->cancel_fork
)
5491 ss
->cancel_fork(child
, subsys_canfork_priv(ss_priv
, j
));
5498 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5499 * @child: the task in question
5501 * This calls the cancel_fork() callbacks if a fork failed *after*
5502 * cgroup_can_fork() succeded.
5504 void cgroup_cancel_fork(struct task_struct
*child
,
5505 void *ss_priv
[CGROUP_CANFORK_COUNT
])
5507 struct cgroup_subsys
*ss
;
5510 for_each_subsys(ss
, i
)
5511 if (ss
->cancel_fork
)
5512 ss
->cancel_fork(child
, subsys_canfork_priv(ss_priv
, i
));
5516 * cgroup_post_fork - called on a new task after adding it to the task list
5517 * @child: the task in question
5519 * Adds the task to the list running through its css_set if necessary and
5520 * call the subsystem fork() callbacks. Has to be after the task is
5521 * visible on the task list in case we race with the first call to
5522 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5525 void cgroup_post_fork(struct task_struct
*child
,
5526 void *old_ss_priv
[CGROUP_CANFORK_COUNT
])
5528 struct cgroup_subsys
*ss
;
5532 * This may race against cgroup_enable_task_cg_lists(). As that
5533 * function sets use_task_css_set_links before grabbing
5534 * tasklist_lock and we just went through tasklist_lock to add
5535 * @child, it's guaranteed that either we see the set
5536 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5537 * @child during its iteration.
5539 * If we won the race, @child is associated with %current's
5540 * css_set. Grabbing css_set_rwsem guarantees both that the
5541 * association is stable, and, on completion of the parent's
5542 * migration, @child is visible in the source of migration or
5543 * already in the destination cgroup. This guarantee is necessary
5544 * when implementing operations which need to migrate all tasks of
5545 * a cgroup to another.
5547 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5548 * will remain in init_css_set. This is safe because all tasks are
5549 * in the init_css_set before cg_links is enabled and there's no
5550 * operation which transfers all tasks out of init_css_set.
5552 if (use_task_css_set_links
) {
5553 struct css_set
*cset
;
5555 down_write(&css_set_rwsem
);
5556 cset
= task_css_set(current
);
5557 if (list_empty(&child
->cg_list
)) {
5559 css_set_move_task(child
, NULL
, cset
, false);
5561 up_write(&css_set_rwsem
);
5565 * Call ss->fork(). This must happen after @child is linked on
5566 * css_set; otherwise, @child might change state between ->fork()
5567 * and addition to css_set.
5569 for_each_subsys_which(ss
, i
, &have_fork_callback
)
5570 ss
->fork(child
, subsys_canfork_priv(old_ss_priv
, i
));
5574 * cgroup_exit - detach cgroup from exiting task
5575 * @tsk: pointer to task_struct of exiting process
5577 * Description: Detach cgroup from @tsk and release it.
5579 * Note that cgroups marked notify_on_release force every task in
5580 * them to take the global cgroup_mutex mutex when exiting.
5581 * This could impact scaling on very large systems. Be reluctant to
5582 * use notify_on_release cgroups where very high task exit scaling
5583 * is required on large systems.
5585 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5586 * call cgroup_exit() while the task is still competent to handle
5587 * notify_on_release(), then leave the task attached to the root cgroup in
5588 * each hierarchy for the remainder of its exit. No need to bother with
5589 * init_css_set refcnting. init_css_set never goes away and we can't race
5590 * with migration path - PF_EXITING is visible to migration path.
5592 void cgroup_exit(struct task_struct
*tsk
)
5594 struct cgroup_subsys
*ss
;
5595 struct css_set
*cset
;
5596 bool put_cset
= false;
5600 * Unlink from @tsk from its css_set. As migration path can't race
5601 * with us, we can check css_set and cg_list without synchronization.
5603 cset
= task_css_set(tsk
);
5605 if (!list_empty(&tsk
->cg_list
)) {
5606 down_write(&css_set_rwsem
);
5607 css_set_move_task(tsk
, cset
, NULL
, false);
5608 up_write(&css_set_rwsem
);
5612 /* Reassign the task to the init_css_set. */
5613 RCU_INIT_POINTER(tsk
->cgroups
, &init_css_set
);
5615 /* see cgroup_post_fork() for details */
5616 for_each_subsys_which(ss
, i
, &have_exit_callback
) {
5617 struct cgroup_subsys_state
*old_css
= cset
->subsys
[i
];
5618 struct cgroup_subsys_state
*css
= task_css(tsk
, i
);
5620 ss
->exit(css
, old_css
, tsk
);
5627 static void check_for_release(struct cgroup
*cgrp
)
5629 if (notify_on_release(cgrp
) && !cgroup_is_populated(cgrp
) &&
5630 !css_has_online_children(&cgrp
->self
) && !cgroup_is_dead(cgrp
))
5631 schedule_work(&cgrp
->release_agent_work
);
5635 * Notify userspace when a cgroup is released, by running the
5636 * configured release agent with the name of the cgroup (path
5637 * relative to the root of cgroup file system) as the argument.
5639 * Most likely, this user command will try to rmdir this cgroup.
5641 * This races with the possibility that some other task will be
5642 * attached to this cgroup before it is removed, or that some other
5643 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5644 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5645 * unused, and this cgroup will be reprieved from its death sentence,
5646 * to continue to serve a useful existence. Next time it's released,
5647 * we will get notified again, if it still has 'notify_on_release' set.
5649 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5650 * means only wait until the task is successfully execve()'d. The
5651 * separate release agent task is forked by call_usermodehelper(),
5652 * then control in this thread returns here, without waiting for the
5653 * release agent task. We don't bother to wait because the caller of
5654 * this routine has no use for the exit status of the release agent
5655 * task, so no sense holding our caller up for that.
5657 static void cgroup_release_agent(struct work_struct
*work
)
5659 struct cgroup
*cgrp
=
5660 container_of(work
, struct cgroup
, release_agent_work
);
5661 char *pathbuf
= NULL
, *agentbuf
= NULL
, *path
;
5662 char *argv
[3], *envp
[3];
5664 mutex_lock(&cgroup_mutex
);
5666 pathbuf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5667 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
5668 if (!pathbuf
|| !agentbuf
)
5671 path
= cgroup_path(cgrp
, pathbuf
, PATH_MAX
);
5679 /* minimal command environment */
5681 envp
[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5684 mutex_unlock(&cgroup_mutex
);
5685 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
5688 mutex_unlock(&cgroup_mutex
);
5694 static int __init
cgroup_disable(char *str
)
5696 struct cgroup_subsys
*ss
;
5700 while ((token
= strsep(&str
, ",")) != NULL
) {
5704 for_each_subsys(ss
, i
) {
5705 if (strcmp(token
, ss
->name
) &&
5706 strcmp(token
, ss
->legacy_name
))
5708 cgroup_disable_mask
|= 1 << i
;
5713 __setup("cgroup_disable=", cgroup_disable
);
5715 static int __init
cgroup_set_legacy_files_on_dfl(char *str
)
5717 printk("cgroup: using legacy files on the default hierarchy\n");
5718 cgroup_legacy_files_on_dfl
= true;
5721 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl
);
5724 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5725 * @dentry: directory dentry of interest
5726 * @ss: subsystem of interest
5728 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5729 * to get the corresponding css and return it. If such css doesn't exist
5730 * or can't be pinned, an ERR_PTR value is returned.
5732 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
5733 struct cgroup_subsys
*ss
)
5735 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5736 struct cgroup_subsys_state
*css
= NULL
;
5737 struct cgroup
*cgrp
;
5739 /* is @dentry a cgroup dir? */
5740 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
5741 kernfs_type(kn
) != KERNFS_DIR
)
5742 return ERR_PTR(-EBADF
);
5747 * This path doesn't originate from kernfs and @kn could already
5748 * have been or be removed at any point. @kn->priv is RCU
5749 * protected for this access. See css_release_work_fn() for details.
5751 cgrp
= rcu_dereference(kn
->priv
);
5753 css
= cgroup_css(cgrp
, ss
);
5755 if (!css
|| !css_tryget_online(css
))
5756 css
= ERR_PTR(-ENOENT
);
5763 * css_from_id - lookup css by id
5764 * @id: the cgroup id
5765 * @ss: cgroup subsys to be looked into
5767 * Returns the css if there's valid one with @id, otherwise returns NULL.
5768 * Should be called under rcu_read_lock().
5770 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5772 WARN_ON_ONCE(!rcu_read_lock_held());
5773 return id
> 0 ? idr_find(&ss
->css_idr
, id
) : NULL
;
5776 #ifdef CONFIG_CGROUP_DEBUG
5777 static struct cgroup_subsys_state
*
5778 debug_css_alloc(struct cgroup_subsys_state
*parent_css
)
5780 struct cgroup_subsys_state
*css
= kzalloc(sizeof(*css
), GFP_KERNEL
);
5783 return ERR_PTR(-ENOMEM
);
5788 static void debug_css_free(struct cgroup_subsys_state
*css
)
5793 static u64
debug_taskcount_read(struct cgroup_subsys_state
*css
,
5796 return cgroup_task_count(css
->cgroup
);
5799 static u64
current_css_set_read(struct cgroup_subsys_state
*css
,
5802 return (u64
)(unsigned long)current
->cgroups
;
5805 static u64
current_css_set_refcount_read(struct cgroup_subsys_state
*css
,
5811 count
= atomic_read(&task_css_set(current
)->refcount
);
5816 static int current_css_set_cg_links_read(struct seq_file
*seq
, void *v
)
5818 struct cgrp_cset_link
*link
;
5819 struct css_set
*cset
;
5822 name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
5826 down_read(&css_set_rwsem
);
5828 cset
= rcu_dereference(current
->cgroups
);
5829 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
5830 struct cgroup
*c
= link
->cgrp
;
5832 cgroup_name(c
, name_buf
, NAME_MAX
+ 1);
5833 seq_printf(seq
, "Root %d group %s\n",
5834 c
->root
->hierarchy_id
, name_buf
);
5837 up_read(&css_set_rwsem
);
5842 #define MAX_TASKS_SHOWN_PER_CSS 25
5843 static int cgroup_css_links_read(struct seq_file
*seq
, void *v
)
5845 struct cgroup_subsys_state
*css
= seq_css(seq
);
5846 struct cgrp_cset_link
*link
;
5848 down_read(&css_set_rwsem
);
5849 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
) {
5850 struct css_set
*cset
= link
->cset
;
5851 struct task_struct
*task
;
5854 seq_printf(seq
, "css_set %p\n", cset
);
5856 list_for_each_entry(task
, &cset
->tasks
, cg_list
) {
5857 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5859 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5862 list_for_each_entry(task
, &cset
->mg_tasks
, cg_list
) {
5863 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5865 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5869 seq_puts(seq
, " ...\n");
5871 up_read(&css_set_rwsem
);
5875 static u64
releasable_read(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
5877 return (!cgroup_is_populated(css
->cgroup
) &&
5878 !css_has_online_children(&css
->cgroup
->self
));
5881 static struct cftype debug_files
[] = {
5883 .name
= "taskcount",
5884 .read_u64
= debug_taskcount_read
,
5888 .name
= "current_css_set",
5889 .read_u64
= current_css_set_read
,
5893 .name
= "current_css_set_refcount",
5894 .read_u64
= current_css_set_refcount_read
,
5898 .name
= "current_css_set_cg_links",
5899 .seq_show
= current_css_set_cg_links_read
,
5903 .name
= "cgroup_css_links",
5904 .seq_show
= cgroup_css_links_read
,
5908 .name
= "releasable",
5909 .read_u64
= releasable_read
,
5915 struct cgroup_subsys debug_cgrp_subsys
= {
5916 .css_alloc
= debug_css_alloc
,
5917 .css_free
= debug_css_free
,
5918 .legacy_cftypes
= debug_files
,
5920 #endif /* CONFIG_CGROUP_DEBUG */