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/percpu-rwsem.h>
49 #include <linux/string.h>
50 #include <linux/sort.h>
51 #include <linux/kmod.h>
52 #include <linux/delayacct.h>
53 #include <linux/cgroupstats.h>
54 #include <linux/hashtable.h>
55 #include <linux/pid_namespace.h>
56 #include <linux/idr.h>
57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
58 #include <linux/kthread.h>
59 #include <linux/delay.h>
60 #include <linux/atomic.h>
61 #include <linux/cpuset.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_lock 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 DEFINE_SPINLOCK(css_set_lock
);
88 EXPORT_SYMBOL_GPL(cgroup_mutex
);
89 EXPORT_SYMBOL_GPL(css_set_lock
);
91 static DEFINE_MUTEX(cgroup_mutex
);
92 static DEFINE_SPINLOCK(css_set_lock
);
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_file->kn for !self csses. It synchronizes notifications
103 * against file removal/re-creation across css hiding.
105 static DEFINE_SPINLOCK(cgroup_file_kn_lock
);
108 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
109 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
111 static DEFINE_SPINLOCK(release_agent_path_lock
);
113 struct percpu_rw_semaphore cgroup_threadgroup_rwsem
;
115 #define cgroup_assert_mutex_or_rcu_locked() \
116 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
117 !lockdep_is_held(&cgroup_mutex), \
118 "cgroup_mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct
*cgroup_destroy_wq
;
129 * pidlist destructions need to be flushed on cgroup destruction. Use a
130 * separate workqueue as flush domain.
132 static struct workqueue_struct
*cgroup_pidlist_destroy_wq
;
134 /* generate an array of cgroup subsystem pointers */
135 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
136 static struct cgroup_subsys
*cgroup_subsys
[] = {
137 #include <linux/cgroup_subsys.h>
141 /* array of cgroup subsystem names */
142 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
143 static const char *cgroup_subsys_name
[] = {
144 #include <linux/cgroup_subsys.h>
148 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
150 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
151 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
152 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
153 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
154 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
158 static struct static_key_true
*cgroup_subsys_enabled_key
[] = {
159 #include <linux/cgroup_subsys.h>
163 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
164 static struct static_key_true
*cgroup_subsys_on_dfl_key
[] = {
165 #include <linux/cgroup_subsys.h>
170 * The default hierarchy, reserved for the subsystems that are otherwise
171 * unattached - it never has more than a single cgroup, and all tasks are
172 * part of that cgroup.
174 struct cgroup_root cgrp_dfl_root
;
175 EXPORT_SYMBOL_GPL(cgrp_dfl_root
);
178 * The default hierarchy always exists but is hidden until mounted for the
179 * first time. This is for backward compatibility.
181 static bool cgrp_dfl_visible
;
183 /* Controllers blocked by the commandline in v1 */
184 static u16 cgroup_no_v1_mask
;
186 /* some controllers are not supported in the default hierarchy */
187 static u16 cgrp_dfl_inhibit_ss_mask
;
189 /* The list of hierarchy roots */
191 static LIST_HEAD(cgroup_roots
);
192 static int cgroup_root_count
;
194 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
195 static DEFINE_IDR(cgroup_hierarchy_idr
);
198 * Assign a monotonically increasing serial number to csses. It guarantees
199 * cgroups with bigger numbers are newer than those with smaller numbers.
200 * Also, as csses are always appended to the parent's ->children list, it
201 * guarantees that sibling csses are always sorted in the ascending serial
202 * number order on the list. Protected by cgroup_mutex.
204 static u64 css_serial_nr_next
= 1;
207 * These bitmask flags indicate whether tasks in the fork and exit paths have
208 * fork/exit handlers to call. This avoids us having to do extra work in the
209 * fork/exit path to check which subsystems have fork/exit callbacks.
211 static u16 have_fork_callback __read_mostly
;
212 static u16 have_exit_callback __read_mostly
;
213 static u16 have_free_callback __read_mostly
;
215 /* Ditto for the can_fork callback. */
216 static u16 have_canfork_callback __read_mostly
;
218 static struct file_system_type cgroup2_fs_type
;
219 static struct cftype cgroup_dfl_base_files
[];
220 static struct cftype cgroup_legacy_base_files
[];
222 static int rebind_subsystems(struct cgroup_root
*dst_root
, u16 ss_mask
);
223 static void css_task_iter_advance(struct css_task_iter
*it
);
224 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
225 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
227 static void css_release(struct percpu_ref
*ref
);
228 static void kill_css(struct cgroup_subsys_state
*css
);
229 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
230 struct cgroup
*cgrp
, struct cftype cfts
[],
234 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
235 * @ssid: subsys ID of interest
237 * cgroup_subsys_enabled() can only be used with literal subsys names which
238 * is fine for individual subsystems but unsuitable for cgroup core. This
239 * is slower static_key_enabled() based test indexed by @ssid.
241 static bool cgroup_ssid_enabled(int ssid
)
243 return static_key_enabled(cgroup_subsys_enabled_key
[ssid
]);
246 static bool cgroup_ssid_no_v1(int ssid
)
248 return cgroup_no_v1_mask
& (1 << ssid
);
252 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
253 * @cgrp: the cgroup of interest
255 * The default hierarchy is the v2 interface of cgroup and this function
256 * can be used to test whether a cgroup is on the default hierarchy for
257 * cases where a subsystem should behave differnetly depending on the
260 * The set of behaviors which change on the default hierarchy are still
261 * being determined and the mount option is prefixed with __DEVEL__.
263 * List of changed behaviors:
265 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
266 * and "name" are disallowed.
268 * - When mounting an existing superblock, mount options should match.
270 * - Remount is disallowed.
272 * - rename(2) is disallowed.
274 * - "tasks" is removed. Everything should be at process granularity. Use
275 * "cgroup.procs" instead.
277 * - "cgroup.procs" is not sorted. pids will be unique unless they got
278 * recycled inbetween reads.
280 * - "release_agent" and "notify_on_release" are removed. Replacement
281 * notification mechanism will be implemented.
283 * - "cgroup.clone_children" is removed.
285 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
286 * and its descendants contain no task; otherwise, 1. The file also
287 * generates kernfs notification which can be monitored through poll and
288 * [di]notify when the value of the file changes.
290 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
291 * take masks of ancestors with non-empty cpus/mems, instead of being
292 * moved to an ancestor.
294 * - cpuset: a task can be moved into an empty cpuset, and again it takes
295 * masks of ancestors.
297 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
300 * - blkcg: blk-throttle becomes properly hierarchical.
302 * - debug: disallowed on the default hierarchy.
304 static bool cgroup_on_dfl(const struct cgroup
*cgrp
)
306 return cgrp
->root
== &cgrp_dfl_root
;
309 /* IDR wrappers which synchronize using cgroup_idr_lock */
310 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
315 idr_preload(gfp_mask
);
316 spin_lock_bh(&cgroup_idr_lock
);
317 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
& ~__GFP_DIRECT_RECLAIM
);
318 spin_unlock_bh(&cgroup_idr_lock
);
323 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
327 spin_lock_bh(&cgroup_idr_lock
);
328 ret
= idr_replace(idr
, ptr
, id
);
329 spin_unlock_bh(&cgroup_idr_lock
);
333 static void cgroup_idr_remove(struct idr
*idr
, int id
)
335 spin_lock_bh(&cgroup_idr_lock
);
337 spin_unlock_bh(&cgroup_idr_lock
);
340 static struct cgroup
*cgroup_parent(struct cgroup
*cgrp
)
342 struct cgroup_subsys_state
*parent_css
= cgrp
->self
.parent
;
345 return container_of(parent_css
, struct cgroup
, self
);
350 * cgroup_css - obtain a cgroup's css for the specified subsystem
351 * @cgrp: the cgroup of interest
352 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
354 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
355 * function must be called either under cgroup_mutex or rcu_read_lock() and
356 * the caller is responsible for pinning the returned css if it wants to
357 * keep accessing it outside the said locks. This function may return
358 * %NULL if @cgrp doesn't have @subsys_id enabled.
360 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
361 struct cgroup_subsys
*ss
)
364 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
365 lockdep_is_held(&cgroup_mutex
));
371 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
372 * @cgrp: the cgroup of interest
373 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
375 * Similar to cgroup_css() but returns the effective css, which is defined
376 * as the matching css of the nearest ancestor including self which has @ss
377 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
378 * function is guaranteed to return non-NULL css.
380 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
381 struct cgroup_subsys
*ss
)
383 lockdep_assert_held(&cgroup_mutex
);
388 if (!(cgrp
->root
->subsys_mask
& (1 << ss
->id
)))
392 * This function is used while updating css associations and thus
393 * can't test the csses directly. Use ->subtree_ss_mask.
395 while (cgroup_parent(cgrp
) &&
396 !(cgroup_parent(cgrp
)->subtree_ss_mask
& (1 << ss
->id
)))
397 cgrp
= cgroup_parent(cgrp
);
399 return cgroup_css(cgrp
, ss
);
403 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
404 * @cgrp: the cgroup of interest
405 * @ss: the subsystem of interest
407 * Find and get the effective css of @cgrp for @ss. The effective css is
408 * defined as the matching css of the nearest ancestor including self which
409 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
410 * the root css is returned, so this function always returns a valid css.
411 * The returned css must be put using css_put().
413 struct cgroup_subsys_state
*cgroup_get_e_css(struct cgroup
*cgrp
,
414 struct cgroup_subsys
*ss
)
416 struct cgroup_subsys_state
*css
;
421 css
= cgroup_css(cgrp
, ss
);
423 if (css
&& css_tryget_online(css
))
425 cgrp
= cgroup_parent(cgrp
);
428 css
= init_css_set
.subsys
[ss
->id
];
435 /* convenient tests for these bits */
436 static inline bool cgroup_is_dead(const struct cgroup
*cgrp
)
438 return !(cgrp
->self
.flags
& CSS_ONLINE
);
441 static void cgroup_get(struct cgroup
*cgrp
)
443 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
444 css_get(&cgrp
->self
);
447 static bool cgroup_tryget(struct cgroup
*cgrp
)
449 return css_tryget(&cgrp
->self
);
452 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
454 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
455 struct cftype
*cft
= of_cft(of
);
458 * This is open and unprotected implementation of cgroup_css().
459 * seq_css() is only called from a kernfs file operation which has
460 * an active reference on the file. Because all the subsystem
461 * files are drained before a css is disassociated with a cgroup,
462 * the matching css from the cgroup's subsys table is guaranteed to
463 * be and stay valid until the enclosing operation is complete.
466 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
470 EXPORT_SYMBOL_GPL(of_css
);
472 static int notify_on_release(const struct cgroup
*cgrp
)
474 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
478 * for_each_css - iterate all css's of a cgroup
479 * @css: the iteration cursor
480 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
481 * @cgrp: the target cgroup to iterate css's of
483 * Should be called under cgroup_[tree_]mutex.
485 #define for_each_css(css, ssid, cgrp) \
486 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
487 if (!((css) = rcu_dereference_check( \
488 (cgrp)->subsys[(ssid)], \
489 lockdep_is_held(&cgroup_mutex)))) { } \
493 * for_each_e_css - iterate all effective 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_e_css(css, ssid, cgrp) \
501 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
502 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
507 * for_each_subsys - iterate all enabled cgroup subsystems
508 * @ss: the iteration cursor
509 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
511 #define for_each_subsys(ss, ssid) \
512 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
513 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
516 * do_each_subsys_mask - filter for_each_subsys with a bitmask
517 * @ss: the iteration cursor
518 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
519 * @ss_mask: the bitmask
521 * The block will only run for cases where the ssid-th bit (1 << ssid) of
524 #define do_each_subsys_mask(ss, ssid, ss_mask) do { \
525 unsigned long __ss_mask = (ss_mask); \
526 if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
530 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \
531 (ss) = cgroup_subsys[ssid]; \
534 #define while_each_subsys_mask() \
539 /* iterate across the hierarchies */
540 #define for_each_root(root) \
541 list_for_each_entry((root), &cgroup_roots, root_list)
543 /* iterate over child cgrps, lock should be held throughout iteration */
544 #define cgroup_for_each_live_child(child, cgrp) \
545 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
546 if (({ lockdep_assert_held(&cgroup_mutex); \
547 cgroup_is_dead(child); })) \
551 static void cgroup_release_agent(struct work_struct
*work
);
552 static void check_for_release(struct cgroup
*cgrp
);
555 * A cgroup can be associated with multiple css_sets as different tasks may
556 * belong to different cgroups on different hierarchies. In the other
557 * direction, a css_set is naturally associated with multiple cgroups.
558 * This M:N relationship is represented by the following link structure
559 * which exists for each association and allows traversing the associations
562 struct cgrp_cset_link
{
563 /* the cgroup and css_set this link associates */
565 struct css_set
*cset
;
567 /* list of cgrp_cset_links anchored at cgrp->cset_links */
568 struct list_head cset_link
;
570 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
571 struct list_head cgrp_link
;
575 * The default css_set - used by init and its children prior to any
576 * hierarchies being mounted. It contains a pointer to the root state
577 * for each subsystem. Also used to anchor the list of css_sets. Not
578 * reference-counted, to improve performance when child cgroups
579 * haven't been created.
581 struct css_set init_css_set
= {
582 .refcount
= ATOMIC_INIT(1),
583 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
584 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
585 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
586 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
587 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
588 .task_iters
= LIST_HEAD_INIT(init_css_set
.task_iters
),
591 static int css_set_count
= 1; /* 1 for init_css_set */
594 * css_set_populated - does a css_set contain any tasks?
595 * @cset: target css_set
597 static bool css_set_populated(struct css_set
*cset
)
599 lockdep_assert_held(&css_set_lock
);
601 return !list_empty(&cset
->tasks
) || !list_empty(&cset
->mg_tasks
);
605 * cgroup_update_populated - updated populated count of a cgroup
606 * @cgrp: the target cgroup
607 * @populated: inc or dec populated count
609 * One of the css_sets associated with @cgrp is either getting its first
610 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
611 * count is propagated towards root so that a given cgroup's populated_cnt
612 * is zero iff the cgroup and all its descendants don't contain any tasks.
614 * @cgrp's interface file "cgroup.populated" is zero if
615 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
616 * changes from or to zero, userland is notified that the content of the
617 * interface file has changed. This can be used to detect when @cgrp and
618 * its descendants become populated or empty.
620 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
622 lockdep_assert_held(&css_set_lock
);
628 trigger
= !cgrp
->populated_cnt
++;
630 trigger
= !--cgrp
->populated_cnt
;
635 check_for_release(cgrp
);
636 cgroup_file_notify(&cgrp
->events_file
);
638 cgrp
= cgroup_parent(cgrp
);
643 * css_set_update_populated - update populated state of a css_set
644 * @cset: target css_set
645 * @populated: whether @cset is populated or depopulated
647 * @cset is either getting the first task or losing the last. Update the
648 * ->populated_cnt of all associated cgroups accordingly.
650 static void css_set_update_populated(struct css_set
*cset
, bool populated
)
652 struct cgrp_cset_link
*link
;
654 lockdep_assert_held(&css_set_lock
);
656 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
)
657 cgroup_update_populated(link
->cgrp
, populated
);
661 * css_set_move_task - move a task from one css_set to another
662 * @task: task being moved
663 * @from_cset: css_set @task currently belongs to (may be NULL)
664 * @to_cset: new css_set @task is being moved to (may be NULL)
665 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
667 * Move @task from @from_cset to @to_cset. If @task didn't belong to any
668 * css_set, @from_cset can be NULL. If @task is being disassociated
669 * instead of moved, @to_cset can be NULL.
671 * This function automatically handles populated_cnt updates and
672 * css_task_iter adjustments but the caller is responsible for managing
673 * @from_cset and @to_cset's reference counts.
675 static void css_set_move_task(struct task_struct
*task
,
676 struct css_set
*from_cset
, struct css_set
*to_cset
,
679 lockdep_assert_held(&css_set_lock
);
682 struct css_task_iter
*it
, *pos
;
684 WARN_ON_ONCE(list_empty(&task
->cg_list
));
687 * @task is leaving, advance task iterators which are
688 * pointing to it so that they can resume at the next
689 * position. Advancing an iterator might remove it from
690 * the list, use safe walk. See css_task_iter_advance*()
693 list_for_each_entry_safe(it
, pos
, &from_cset
->task_iters
,
695 if (it
->task_pos
== &task
->cg_list
)
696 css_task_iter_advance(it
);
698 list_del_init(&task
->cg_list
);
699 if (!css_set_populated(from_cset
))
700 css_set_update_populated(from_cset
, false);
702 WARN_ON_ONCE(!list_empty(&task
->cg_list
));
707 * We are synchronized through cgroup_threadgroup_rwsem
708 * against PF_EXITING setting such that we can't race
709 * against cgroup_exit() changing the css_set to
710 * init_css_set and dropping the old one.
712 WARN_ON_ONCE(task
->flags
& PF_EXITING
);
714 if (!css_set_populated(to_cset
))
715 css_set_update_populated(to_cset
, true);
716 rcu_assign_pointer(task
->cgroups
, to_cset
);
717 list_add_tail(&task
->cg_list
, use_mg_tasks
? &to_cset
->mg_tasks
:
723 * hash table for cgroup groups. This improves the performance to find
724 * an existing css_set. This hash doesn't (currently) take into
725 * account cgroups in empty hierarchies.
727 #define CSS_SET_HASH_BITS 7
728 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
730 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
732 unsigned long key
= 0UL;
733 struct cgroup_subsys
*ss
;
736 for_each_subsys(ss
, i
)
737 key
+= (unsigned long)css
[i
];
738 key
= (key
>> 16) ^ key
;
743 static void put_css_set_locked(struct css_set
*cset
)
745 struct cgrp_cset_link
*link
, *tmp_link
;
746 struct cgroup_subsys
*ss
;
749 lockdep_assert_held(&css_set_lock
);
751 if (!atomic_dec_and_test(&cset
->refcount
))
754 /* This css_set is dead. unlink it and release cgroup and css refs */
755 for_each_subsys(ss
, ssid
) {
756 list_del(&cset
->e_cset_node
[ssid
]);
757 css_put(cset
->subsys
[ssid
]);
759 hash_del(&cset
->hlist
);
762 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
763 list_del(&link
->cset_link
);
764 list_del(&link
->cgrp_link
);
765 if (cgroup_parent(link
->cgrp
))
766 cgroup_put(link
->cgrp
);
770 kfree_rcu(cset
, rcu_head
);
773 static void put_css_set(struct css_set
*cset
)
776 * Ensure that the refcount doesn't hit zero while any readers
777 * can see it. Similar to atomic_dec_and_lock(), but for an
780 if (atomic_add_unless(&cset
->refcount
, -1, 1))
783 spin_lock_bh(&css_set_lock
);
784 put_css_set_locked(cset
);
785 spin_unlock_bh(&css_set_lock
);
789 * refcounted get/put for css_set objects
791 static inline void get_css_set(struct css_set
*cset
)
793 atomic_inc(&cset
->refcount
);
797 * compare_css_sets - helper function for find_existing_css_set().
798 * @cset: candidate css_set being tested
799 * @old_cset: existing css_set for a task
800 * @new_cgrp: cgroup that's being entered by the task
801 * @template: desired set of css pointers in css_set (pre-calculated)
803 * Returns true if "cset" matches "old_cset" except for the hierarchy
804 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
806 static bool compare_css_sets(struct css_set
*cset
,
807 struct css_set
*old_cset
,
808 struct cgroup
*new_cgrp
,
809 struct cgroup_subsys_state
*template[])
811 struct list_head
*l1
, *l2
;
814 * On the default hierarchy, there can be csets which are
815 * associated with the same set of cgroups but different csses.
816 * Let's first ensure that csses match.
818 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
822 * Compare cgroup pointers in order to distinguish between
823 * different cgroups in hierarchies. As different cgroups may
824 * share the same effective css, this comparison is always
827 l1
= &cset
->cgrp_links
;
828 l2
= &old_cset
->cgrp_links
;
830 struct cgrp_cset_link
*link1
, *link2
;
831 struct cgroup
*cgrp1
, *cgrp2
;
835 /* See if we reached the end - both lists are equal length. */
836 if (l1
== &cset
->cgrp_links
) {
837 BUG_ON(l2
!= &old_cset
->cgrp_links
);
840 BUG_ON(l2
== &old_cset
->cgrp_links
);
842 /* Locate the cgroups associated with these links. */
843 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
844 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
847 /* Hierarchies should be linked in the same order. */
848 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
851 * If this hierarchy is the hierarchy of the cgroup
852 * that's changing, then we need to check that this
853 * css_set points to the new cgroup; if it's any other
854 * hierarchy, then this css_set should point to the
855 * same cgroup as the old css_set.
857 if (cgrp1
->root
== new_cgrp
->root
) {
858 if (cgrp1
!= new_cgrp
)
869 * find_existing_css_set - init css array and find the matching css_set
870 * @old_cset: the css_set that we're using before the cgroup transition
871 * @cgrp: the cgroup that we're moving into
872 * @template: out param for the new set of csses, should be clear on entry
874 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
876 struct cgroup_subsys_state
*template[])
878 struct cgroup_root
*root
= cgrp
->root
;
879 struct cgroup_subsys
*ss
;
880 struct css_set
*cset
;
885 * Build the set of subsystem state objects that we want to see in the
886 * new css_set. while subsystems can change globally, the entries here
887 * won't change, so no need for locking.
889 for_each_subsys(ss
, i
) {
890 if (root
->subsys_mask
& (1UL << i
)) {
892 * @ss is in this hierarchy, so we want the
893 * effective css from @cgrp.
895 template[i
] = cgroup_e_css(cgrp
, ss
);
898 * @ss is not in this hierarchy, so we don't want
901 template[i
] = old_cset
->subsys
[i
];
905 key
= css_set_hash(template);
906 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
907 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
910 /* This css_set matches what we need */
914 /* No existing cgroup group matched */
918 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
920 struct cgrp_cset_link
*link
, *tmp_link
;
922 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
923 list_del(&link
->cset_link
);
929 * allocate_cgrp_cset_links - allocate cgrp_cset_links
930 * @count: the number of links to allocate
931 * @tmp_links: list_head the allocated links are put on
933 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
934 * through ->cset_link. Returns 0 on success or -errno.
936 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
938 struct cgrp_cset_link
*link
;
941 INIT_LIST_HEAD(tmp_links
);
943 for (i
= 0; i
< count
; i
++) {
944 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
946 free_cgrp_cset_links(tmp_links
);
949 list_add(&link
->cset_link
, tmp_links
);
955 * link_css_set - a helper function to link a css_set to a cgroup
956 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
957 * @cset: the css_set to be linked
958 * @cgrp: the destination cgroup
960 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
963 struct cgrp_cset_link
*link
;
965 BUG_ON(list_empty(tmp_links
));
967 if (cgroup_on_dfl(cgrp
))
968 cset
->dfl_cgrp
= cgrp
;
970 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
975 * Always add links to the tail of the lists so that the lists are
976 * in choronological order.
978 list_move_tail(&link
->cset_link
, &cgrp
->cset_links
);
979 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
981 if (cgroup_parent(cgrp
))
986 * find_css_set - return a new css_set with one cgroup updated
987 * @old_cset: the baseline css_set
988 * @cgrp: the cgroup to be updated
990 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
991 * substituted into the appropriate hierarchy.
993 static struct css_set
*find_css_set(struct css_set
*old_cset
,
996 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
997 struct css_set
*cset
;
998 struct list_head tmp_links
;
999 struct cgrp_cset_link
*link
;
1000 struct cgroup_subsys
*ss
;
1004 lockdep_assert_held(&cgroup_mutex
);
1006 /* First see if we already have a cgroup group that matches
1007 * the desired set */
1008 spin_lock_bh(&css_set_lock
);
1009 cset
= find_existing_css_set(old_cset
, cgrp
, template);
1012 spin_unlock_bh(&css_set_lock
);
1017 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
1021 /* Allocate all the cgrp_cset_link objects that we'll need */
1022 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
1027 atomic_set(&cset
->refcount
, 1);
1028 INIT_LIST_HEAD(&cset
->cgrp_links
);
1029 INIT_LIST_HEAD(&cset
->tasks
);
1030 INIT_LIST_HEAD(&cset
->mg_tasks
);
1031 INIT_LIST_HEAD(&cset
->mg_preload_node
);
1032 INIT_LIST_HEAD(&cset
->mg_node
);
1033 INIT_LIST_HEAD(&cset
->task_iters
);
1034 INIT_HLIST_NODE(&cset
->hlist
);
1036 /* Copy the set of subsystem state objects generated in
1037 * find_existing_css_set() */
1038 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
1040 spin_lock_bh(&css_set_lock
);
1041 /* Add reference counts and links from the new css_set. */
1042 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
1043 struct cgroup
*c
= link
->cgrp
;
1045 if (c
->root
== cgrp
->root
)
1047 link_css_set(&tmp_links
, cset
, c
);
1050 BUG_ON(!list_empty(&tmp_links
));
1054 /* Add @cset to the hash table */
1055 key
= css_set_hash(cset
->subsys
);
1056 hash_add(css_set_table
, &cset
->hlist
, key
);
1058 for_each_subsys(ss
, ssid
) {
1059 struct cgroup_subsys_state
*css
= cset
->subsys
[ssid
];
1061 list_add_tail(&cset
->e_cset_node
[ssid
],
1062 &css
->cgroup
->e_csets
[ssid
]);
1066 spin_unlock_bh(&css_set_lock
);
1071 static struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
1073 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
1075 return root_cgrp
->root
;
1078 static int cgroup_init_root_id(struct cgroup_root
*root
)
1082 lockdep_assert_held(&cgroup_mutex
);
1084 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
1088 root
->hierarchy_id
= id
;
1092 static void cgroup_exit_root_id(struct cgroup_root
*root
)
1094 lockdep_assert_held(&cgroup_mutex
);
1096 if (root
->hierarchy_id
) {
1097 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
1098 root
->hierarchy_id
= 0;
1102 static void cgroup_free_root(struct cgroup_root
*root
)
1105 /* hierarchy ID should already have been released */
1106 WARN_ON_ONCE(root
->hierarchy_id
);
1108 idr_destroy(&root
->cgroup_idr
);
1113 static void cgroup_destroy_root(struct cgroup_root
*root
)
1115 struct cgroup
*cgrp
= &root
->cgrp
;
1116 struct cgrp_cset_link
*link
, *tmp_link
;
1118 mutex_lock(&cgroup_mutex
);
1120 BUG_ON(atomic_read(&root
->nr_cgrps
));
1121 BUG_ON(!list_empty(&cgrp
->self
.children
));
1123 /* Rebind all subsystems back to the default hierarchy */
1124 rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
);
1127 * Release all the links from cset_links to this hierarchy's
1130 spin_lock_bh(&css_set_lock
);
1132 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
1133 list_del(&link
->cset_link
);
1134 list_del(&link
->cgrp_link
);
1138 spin_unlock_bh(&css_set_lock
);
1140 if (!list_empty(&root
->root_list
)) {
1141 list_del(&root
->root_list
);
1142 cgroup_root_count
--;
1145 cgroup_exit_root_id(root
);
1147 mutex_unlock(&cgroup_mutex
);
1149 kernfs_destroy_root(root
->kf_root
);
1150 cgroup_free_root(root
);
1153 /* look up cgroup associated with given css_set on the specified hierarchy */
1154 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
1155 struct cgroup_root
*root
)
1157 struct cgroup
*res
= NULL
;
1159 lockdep_assert_held(&cgroup_mutex
);
1160 lockdep_assert_held(&css_set_lock
);
1162 if (cset
== &init_css_set
) {
1165 struct cgrp_cset_link
*link
;
1167 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1168 struct cgroup
*c
= link
->cgrp
;
1170 if (c
->root
== root
) {
1182 * Return the cgroup for "task" from the given hierarchy. Must be
1183 * called with cgroup_mutex and css_set_lock held.
1185 static struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
1186 struct cgroup_root
*root
)
1189 * No need to lock the task - since we hold cgroup_mutex the
1190 * task can't change groups, so the only thing that can happen
1191 * is that it exits and its css is set back to init_css_set.
1193 return cset_cgroup_from_root(task_css_set(task
), root
);
1197 * A task must hold cgroup_mutex to modify cgroups.
1199 * Any task can increment and decrement the count field without lock.
1200 * So in general, code holding cgroup_mutex can't rely on the count
1201 * field not changing. However, if the count goes to zero, then only
1202 * cgroup_attach_task() can increment it again. Because a count of zero
1203 * means that no tasks are currently attached, therefore there is no
1204 * way a task attached to that cgroup can fork (the other way to
1205 * increment the count). So code holding cgroup_mutex can safely
1206 * assume that if the count is zero, it will stay zero. Similarly, if
1207 * a task holds cgroup_mutex on a cgroup with zero count, it
1208 * knows that the cgroup won't be removed, as cgroup_rmdir()
1211 * A cgroup can only be deleted if both its 'count' of using tasks
1212 * is zero, and its list of 'children' cgroups is empty. Since all
1213 * tasks in the system use _some_ cgroup, and since there is always at
1214 * least one task in the system (init, pid == 1), therefore, root cgroup
1215 * always has either children cgroups and/or using tasks. So we don't
1216 * need a special hack to ensure that root cgroup cannot be deleted.
1218 * P.S. One more locking exception. RCU is used to guard the
1219 * update of a tasks cgroup pointer by cgroup_attach_task()
1222 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1223 static const struct file_operations proc_cgroupstats_operations
;
1225 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1228 struct cgroup_subsys
*ss
= cft
->ss
;
1230 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1231 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
1232 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
1233 cgroup_on_dfl(cgrp
) ? ss
->name
: ss
->legacy_name
,
1236 strncpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1241 * cgroup_file_mode - deduce file mode of a control file
1242 * @cft: the control file in question
1244 * S_IRUGO for read, S_IWUSR for write.
1246 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1250 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1253 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
) {
1254 if (cft
->flags
& CFTYPE_WORLD_WRITABLE
)
1264 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1265 * @cgrp: the target cgroup
1266 * @subtree_control: the new subtree_control mask to consider
1268 * On the default hierarchy, a subsystem may request other subsystems to be
1269 * enabled together through its ->depends_on mask. In such cases, more
1270 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1272 * This function calculates which subsystems need to be enabled if
1273 * @subtree_control is to be applied to @cgrp. The returned mask is always
1274 * a superset of @subtree_control and follows the usual hierarchy rules.
1276 static u16
cgroup_calc_subtree_ss_mask(struct cgroup
*cgrp
, u16 subtree_control
)
1278 struct cgroup
*parent
= cgroup_parent(cgrp
);
1279 u16 cur_ss_mask
= subtree_control
;
1280 struct cgroup_subsys
*ss
;
1283 lockdep_assert_held(&cgroup_mutex
);
1285 if (!cgroup_on_dfl(cgrp
))
1289 u16 new_ss_mask
= cur_ss_mask
;
1291 do_each_subsys_mask(ss
, ssid
, cur_ss_mask
) {
1292 new_ss_mask
|= ss
->depends_on
;
1293 } while_each_subsys_mask();
1296 * Mask out subsystems which aren't available. This can
1297 * happen only if some depended-upon subsystems were bound
1298 * to non-default hierarchies.
1301 new_ss_mask
&= parent
->subtree_ss_mask
;
1303 new_ss_mask
&= cgrp
->root
->subsys_mask
;
1305 if (new_ss_mask
== cur_ss_mask
)
1307 cur_ss_mask
= new_ss_mask
;
1314 * cgroup_refresh_subtree_ss_mask - update subtree_ss_mask
1315 * @cgrp: the target cgroup
1317 * Update @cgrp->subtree_ss_mask according to the current
1318 * @cgrp->subtree_control using cgroup_calc_subtree_ss_mask().
1320 static void cgroup_refresh_subtree_ss_mask(struct cgroup
*cgrp
)
1322 cgrp
->subtree_ss_mask
=
1323 cgroup_calc_subtree_ss_mask(cgrp
, cgrp
->subtree_control
);
1327 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1328 * @kn: the kernfs_node being serviced
1330 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1331 * the method finishes if locking succeeded. Note that once this function
1332 * returns the cgroup returned by cgroup_kn_lock_live() may become
1333 * inaccessible any time. If the caller intends to continue to access the
1334 * cgroup, it should pin it before invoking this function.
1336 static void cgroup_kn_unlock(struct kernfs_node
*kn
)
1338 struct cgroup
*cgrp
;
1340 if (kernfs_type(kn
) == KERNFS_DIR
)
1343 cgrp
= kn
->parent
->priv
;
1345 mutex_unlock(&cgroup_mutex
);
1347 kernfs_unbreak_active_protection(kn
);
1352 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1353 * @kn: the kernfs_node being serviced
1355 * This helper is to be used by a cgroup kernfs method currently servicing
1356 * @kn. It breaks the active protection, performs cgroup locking and
1357 * verifies that the associated cgroup is alive. Returns the cgroup if
1358 * alive; otherwise, %NULL. A successful return should be undone by a
1359 * matching cgroup_kn_unlock() invocation.
1361 * Any cgroup kernfs method implementation which requires locking the
1362 * associated cgroup should use this helper. It avoids nesting cgroup
1363 * locking under kernfs active protection and allows all kernfs operations
1364 * including self-removal.
1366 static struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
)
1368 struct cgroup
*cgrp
;
1370 if (kernfs_type(kn
) == KERNFS_DIR
)
1373 cgrp
= kn
->parent
->priv
;
1376 * We're gonna grab cgroup_mutex which nests outside kernfs
1377 * active_ref. cgroup liveliness check alone provides enough
1378 * protection against removal. Ensure @cgrp stays accessible and
1379 * break the active_ref protection.
1381 if (!cgroup_tryget(cgrp
))
1383 kernfs_break_active_protection(kn
);
1385 mutex_lock(&cgroup_mutex
);
1387 if (!cgroup_is_dead(cgrp
))
1390 cgroup_kn_unlock(kn
);
1394 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1396 char name
[CGROUP_FILE_NAME_MAX
];
1398 lockdep_assert_held(&cgroup_mutex
);
1400 if (cft
->file_offset
) {
1401 struct cgroup_subsys_state
*css
= cgroup_css(cgrp
, cft
->ss
);
1402 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
1404 spin_lock_irq(&cgroup_file_kn_lock
);
1406 spin_unlock_irq(&cgroup_file_kn_lock
);
1409 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1413 * css_clear_dir - remove subsys files in a cgroup directory
1415 * @cgrp_override: specify if target cgroup is different from css->cgroup
1417 static void css_clear_dir(struct cgroup_subsys_state
*css
,
1418 struct cgroup
*cgrp_override
)
1420 struct cgroup
*cgrp
= cgrp_override
?: css
->cgroup
;
1421 struct cftype
*cfts
;
1423 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
)
1424 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1428 * css_populate_dir - create subsys files in a cgroup directory
1430 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1432 * On failure, no file is added.
1434 static int css_populate_dir(struct cgroup_subsys_state
*css
,
1435 struct cgroup
*cgrp_override
)
1437 struct cgroup
*cgrp
= cgrp_override
?: css
->cgroup
;
1438 struct cftype
*cfts
, *failed_cfts
;
1442 if (cgroup_on_dfl(cgrp
))
1443 cfts
= cgroup_dfl_base_files
;
1445 cfts
= cgroup_legacy_base_files
;
1447 return cgroup_addrm_files(&cgrp
->self
, cgrp
, cfts
, true);
1450 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1451 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, true);
1459 list_for_each_entry(cfts
, &css
->ss
->cfts
, node
) {
1460 if (cfts
== failed_cfts
)
1462 cgroup_addrm_files(css
, cgrp
, cfts
, false);
1467 static int rebind_subsystems(struct cgroup_root
*dst_root
, u16 ss_mask
)
1469 struct cgroup
*dcgrp
= &dst_root
->cgrp
;
1470 struct cgroup_subsys
*ss
;
1474 lockdep_assert_held(&cgroup_mutex
);
1476 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1477 /* if @ss has non-root csses attached to it, can't move */
1478 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)))
1481 /* can't move between two non-dummy roots either */
1482 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1484 } while_each_subsys_mask();
1486 /* skip creating root files on dfl_root for inhibited subsystems */
1487 tmp_ss_mask
= ss_mask
;
1488 if (dst_root
== &cgrp_dfl_root
)
1489 tmp_ss_mask
&= ~cgrp_dfl_inhibit_ss_mask
;
1491 do_each_subsys_mask(ss
, ssid
, tmp_ss_mask
) {
1492 struct cgroup
*scgrp
= &ss
->root
->cgrp
;
1495 ret
= css_populate_dir(cgroup_css(scgrp
, ss
), dcgrp
);
1500 * Rebinding back to the default root is not allowed to
1501 * fail. Using both default and non-default roots should
1502 * be rare. Moving subsystems back and forth even more so.
1503 * Just warn about it and continue.
1505 if (dst_root
== &cgrp_dfl_root
) {
1506 if (cgrp_dfl_visible
) {
1507 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1509 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1514 do_each_subsys_mask(ss
, tssid
, tmp_ss_mask
) {
1517 css_clear_dir(cgroup_css(scgrp
, ss
), dcgrp
);
1518 } while_each_subsys_mask();
1520 } while_each_subsys_mask();
1523 * Nothing can fail from this point on. Remove files for the
1524 * removed subsystems and rebind each subsystem.
1526 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
1527 struct cgroup_root
*src_root
= ss
->root
;
1528 struct cgroup
*scgrp
= &src_root
->cgrp
;
1529 struct cgroup_subsys_state
*css
= cgroup_css(scgrp
, ss
);
1530 struct css_set
*cset
;
1532 WARN_ON(!css
|| cgroup_css(dcgrp
, ss
));
1534 css_clear_dir(css
, NULL
);
1536 RCU_INIT_POINTER(scgrp
->subsys
[ssid
], NULL
);
1537 rcu_assign_pointer(dcgrp
->subsys
[ssid
], css
);
1538 ss
->root
= dst_root
;
1539 css
->cgroup
= dcgrp
;
1541 spin_lock_bh(&css_set_lock
);
1542 hash_for_each(css_set_table
, i
, cset
, hlist
)
1543 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1544 &dcgrp
->e_csets
[ss
->id
]);
1545 spin_unlock_bh(&css_set_lock
);
1547 src_root
->subsys_mask
&= ~(1 << ssid
);
1548 scgrp
->subtree_control
&= ~(1 << ssid
);
1549 cgroup_refresh_subtree_ss_mask(scgrp
);
1551 /* default hierarchy doesn't enable controllers by default */
1552 dst_root
->subsys_mask
|= 1 << ssid
;
1553 if (dst_root
== &cgrp_dfl_root
) {
1554 static_branch_enable(cgroup_subsys_on_dfl_key
[ssid
]);
1556 dcgrp
->subtree_control
|= 1 << ssid
;
1557 cgroup_refresh_subtree_ss_mask(dcgrp
);
1558 static_branch_disable(cgroup_subsys_on_dfl_key
[ssid
]);
1563 } while_each_subsys_mask();
1565 kernfs_activate(dcgrp
->kn
);
1569 static int cgroup_show_options(struct seq_file
*seq
,
1570 struct kernfs_root
*kf_root
)
1572 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1573 struct cgroup_subsys
*ss
;
1576 if (root
!= &cgrp_dfl_root
)
1577 for_each_subsys(ss
, ssid
)
1578 if (root
->subsys_mask
& (1 << ssid
))
1579 seq_show_option(seq
, ss
->legacy_name
, NULL
);
1580 if (root
->flags
& CGRP_ROOT_NOPREFIX
)
1581 seq_puts(seq
, ",noprefix");
1582 if (root
->flags
& CGRP_ROOT_XATTR
)
1583 seq_puts(seq
, ",xattr");
1585 spin_lock(&release_agent_path_lock
);
1586 if (strlen(root
->release_agent_path
))
1587 seq_show_option(seq
, "release_agent",
1588 root
->release_agent_path
);
1589 spin_unlock(&release_agent_path_lock
);
1591 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
))
1592 seq_puts(seq
, ",clone_children");
1593 if (strlen(root
->name
))
1594 seq_show_option(seq
, "name", root
->name
);
1598 struct cgroup_sb_opts
{
1601 char *release_agent
;
1602 bool cpuset_clone_children
;
1604 /* User explicitly requested empty subsystem */
1608 static int parse_cgroupfs_options(char *data
, struct cgroup_sb_opts
*opts
)
1610 char *token
, *o
= data
;
1611 bool all_ss
= false, one_ss
= false;
1613 struct cgroup_subsys
*ss
;
1617 #ifdef CONFIG_CPUSETS
1618 mask
= ~((u16
)1 << cpuset_cgrp_id
);
1621 memset(opts
, 0, sizeof(*opts
));
1623 while ((token
= strsep(&o
, ",")) != NULL
) {
1628 if (!strcmp(token
, "none")) {
1629 /* Explicitly have no subsystems */
1633 if (!strcmp(token
, "all")) {
1634 /* Mutually exclusive option 'all' + subsystem name */
1640 if (!strcmp(token
, "noprefix")) {
1641 opts
->flags
|= CGRP_ROOT_NOPREFIX
;
1644 if (!strcmp(token
, "clone_children")) {
1645 opts
->cpuset_clone_children
= true;
1648 if (!strcmp(token
, "xattr")) {
1649 opts
->flags
|= CGRP_ROOT_XATTR
;
1652 if (!strncmp(token
, "release_agent=", 14)) {
1653 /* Specifying two release agents is forbidden */
1654 if (opts
->release_agent
)
1656 opts
->release_agent
=
1657 kstrndup(token
+ 14, PATH_MAX
- 1, GFP_KERNEL
);
1658 if (!opts
->release_agent
)
1662 if (!strncmp(token
, "name=", 5)) {
1663 const char *name
= token
+ 5;
1664 /* Can't specify an empty name */
1667 /* Must match [\w.-]+ */
1668 for (i
= 0; i
< strlen(name
); i
++) {
1672 if ((c
== '.') || (c
== '-') || (c
== '_'))
1676 /* Specifying two names is forbidden */
1679 opts
->name
= kstrndup(name
,
1680 MAX_CGROUP_ROOT_NAMELEN
- 1,
1688 for_each_subsys(ss
, i
) {
1689 if (strcmp(token
, ss
->legacy_name
))
1691 if (!cgroup_ssid_enabled(i
))
1693 if (cgroup_ssid_no_v1(i
))
1696 /* Mutually exclusive option 'all' + subsystem name */
1699 opts
->subsys_mask
|= (1 << i
);
1704 if (i
== CGROUP_SUBSYS_COUNT
)
1709 * If the 'all' option was specified select all the subsystems,
1710 * otherwise if 'none', 'name=' and a subsystem name options were
1711 * not specified, let's default to 'all'
1713 if (all_ss
|| (!one_ss
&& !opts
->none
&& !opts
->name
))
1714 for_each_subsys(ss
, i
)
1715 if (cgroup_ssid_enabled(i
) && !cgroup_ssid_no_v1(i
))
1716 opts
->subsys_mask
|= (1 << i
);
1719 * We either have to specify by name or by subsystems. (So all
1720 * empty hierarchies must have a name).
1722 if (!opts
->subsys_mask
&& !opts
->name
)
1726 * Option noprefix was introduced just for backward compatibility
1727 * with the old cpuset, so we allow noprefix only if mounting just
1728 * the cpuset subsystem.
1730 if ((opts
->flags
& CGRP_ROOT_NOPREFIX
) && (opts
->subsys_mask
& mask
))
1733 /* Can't specify "none" and some subsystems */
1734 if (opts
->subsys_mask
&& opts
->none
)
1740 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1743 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1744 struct cgroup_sb_opts opts
;
1745 u16 added_mask
, removed_mask
;
1747 if (root
== &cgrp_dfl_root
) {
1748 pr_err("remount is not allowed\n");
1752 mutex_lock(&cgroup_mutex
);
1754 /* See what subsystems are wanted */
1755 ret
= parse_cgroupfs_options(data
, &opts
);
1759 if (opts
.subsys_mask
!= root
->subsys_mask
|| opts
.release_agent
)
1760 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1761 task_tgid_nr(current
), current
->comm
);
1763 added_mask
= opts
.subsys_mask
& ~root
->subsys_mask
;
1764 removed_mask
= root
->subsys_mask
& ~opts
.subsys_mask
;
1766 /* Don't allow flags or name to change at remount */
1767 if ((opts
.flags
^ root
->flags
) ||
1768 (opts
.name
&& strcmp(opts
.name
, root
->name
))) {
1769 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1770 opts
.flags
, opts
.name
?: "", root
->flags
, root
->name
);
1775 /* remounting is not allowed for populated hierarchies */
1776 if (!list_empty(&root
->cgrp
.self
.children
)) {
1781 ret
= rebind_subsystems(root
, added_mask
);
1785 rebind_subsystems(&cgrp_dfl_root
, removed_mask
);
1787 if (opts
.release_agent
) {
1788 spin_lock(&release_agent_path_lock
);
1789 strcpy(root
->release_agent_path
, opts
.release_agent
);
1790 spin_unlock(&release_agent_path_lock
);
1793 kfree(opts
.release_agent
);
1795 mutex_unlock(&cgroup_mutex
);
1800 * To reduce the fork() overhead for systems that are not actually using
1801 * their cgroups capability, we don't maintain the lists running through
1802 * each css_set to its tasks until we see the list actually used - in other
1803 * words after the first mount.
1805 static bool use_task_css_set_links __read_mostly
;
1807 static void cgroup_enable_task_cg_lists(void)
1809 struct task_struct
*p
, *g
;
1811 spin_lock_bh(&css_set_lock
);
1813 if (use_task_css_set_links
)
1816 use_task_css_set_links
= true;
1819 * We need tasklist_lock because RCU is not safe against
1820 * while_each_thread(). Besides, a forking task that has passed
1821 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1822 * is not guaranteed to have its child immediately visible in the
1823 * tasklist if we walk through it with RCU.
1825 read_lock(&tasklist_lock
);
1826 do_each_thread(g
, p
) {
1827 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1828 task_css_set(p
) != &init_css_set
);
1831 * We should check if the process is exiting, otherwise
1832 * it will race with cgroup_exit() in that the list
1833 * entry won't be deleted though the process has exited.
1834 * Do it while holding siglock so that we don't end up
1835 * racing against cgroup_exit().
1837 spin_lock_irq(&p
->sighand
->siglock
);
1838 if (!(p
->flags
& PF_EXITING
)) {
1839 struct css_set
*cset
= task_css_set(p
);
1841 if (!css_set_populated(cset
))
1842 css_set_update_populated(cset
, true);
1843 list_add_tail(&p
->cg_list
, &cset
->tasks
);
1846 spin_unlock_irq(&p
->sighand
->siglock
);
1847 } while_each_thread(g
, p
);
1848 read_unlock(&tasklist_lock
);
1850 spin_unlock_bh(&css_set_lock
);
1853 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1855 struct cgroup_subsys
*ss
;
1858 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1859 INIT_LIST_HEAD(&cgrp
->self
.children
);
1860 INIT_LIST_HEAD(&cgrp
->cset_links
);
1861 INIT_LIST_HEAD(&cgrp
->pidlists
);
1862 mutex_init(&cgrp
->pidlist_mutex
);
1863 cgrp
->self
.cgroup
= cgrp
;
1864 cgrp
->self
.flags
|= CSS_ONLINE
;
1866 for_each_subsys(ss
, ssid
)
1867 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1869 init_waitqueue_head(&cgrp
->offline_waitq
);
1870 INIT_WORK(&cgrp
->release_agent_work
, cgroup_release_agent
);
1873 static void init_cgroup_root(struct cgroup_root
*root
,
1874 struct cgroup_sb_opts
*opts
)
1876 struct cgroup
*cgrp
= &root
->cgrp
;
1878 INIT_LIST_HEAD(&root
->root_list
);
1879 atomic_set(&root
->nr_cgrps
, 1);
1881 init_cgroup_housekeeping(cgrp
);
1882 idr_init(&root
->cgroup_idr
);
1884 root
->flags
= opts
->flags
;
1885 if (opts
->release_agent
)
1886 strcpy(root
->release_agent_path
, opts
->release_agent
);
1888 strcpy(root
->name
, opts
->name
);
1889 if (opts
->cpuset_clone_children
)
1890 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1893 static int cgroup_setup_root(struct cgroup_root
*root
, u16 ss_mask
)
1895 LIST_HEAD(tmp_links
);
1896 struct cgroup
*root_cgrp
= &root
->cgrp
;
1897 struct css_set
*cset
;
1900 lockdep_assert_held(&cgroup_mutex
);
1902 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_KERNEL
);
1905 root_cgrp
->id
= ret
;
1906 root_cgrp
->ancestor_ids
[0] = ret
;
1908 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
, 0,
1914 * We're accessing css_set_count without locking css_set_lock here,
1915 * but that's OK - it can only be increased by someone holding
1916 * cgroup_lock, and that's us. The worst that can happen is that we
1917 * have some link structures left over
1919 ret
= allocate_cgrp_cset_links(css_set_count
, &tmp_links
);
1923 ret
= cgroup_init_root_id(root
);
1927 root
->kf_root
= kernfs_create_root(&cgroup_kf_syscall_ops
,
1928 KERNFS_ROOT_CREATE_DEACTIVATED
,
1930 if (IS_ERR(root
->kf_root
)) {
1931 ret
= PTR_ERR(root
->kf_root
);
1934 root_cgrp
->kn
= root
->kf_root
->kn
;
1936 ret
= css_populate_dir(&root_cgrp
->self
, NULL
);
1940 ret
= rebind_subsystems(root
, ss_mask
);
1945 * There must be no failure case after here, since rebinding takes
1946 * care of subsystems' refcounts, which are explicitly dropped in
1947 * the failure exit path.
1949 list_add(&root
->root_list
, &cgroup_roots
);
1950 cgroup_root_count
++;
1953 * Link the root cgroup in this hierarchy into all the css_set
1956 spin_lock_bh(&css_set_lock
);
1957 hash_for_each(css_set_table
, i
, cset
, hlist
) {
1958 link_css_set(&tmp_links
, cset
, root_cgrp
);
1959 if (css_set_populated(cset
))
1960 cgroup_update_populated(root_cgrp
, true);
1962 spin_unlock_bh(&css_set_lock
);
1964 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
1965 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1967 kernfs_activate(root_cgrp
->kn
);
1972 kernfs_destroy_root(root
->kf_root
);
1973 root
->kf_root
= NULL
;
1975 cgroup_exit_root_id(root
);
1977 percpu_ref_exit(&root_cgrp
->self
.refcnt
);
1979 free_cgrp_cset_links(&tmp_links
);
1983 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
1984 int flags
, const char *unused_dev_name
,
1987 bool is_v2
= fs_type
== &cgroup2_fs_type
;
1988 struct super_block
*pinned_sb
= NULL
;
1989 struct cgroup_subsys
*ss
;
1990 struct cgroup_root
*root
;
1991 struct cgroup_sb_opts opts
;
1992 struct dentry
*dentry
;
1998 * The first time anyone tries to mount a cgroup, enable the list
1999 * linking each css_set to its tasks and fix up all existing tasks.
2001 if (!use_task_css_set_links
)
2002 cgroup_enable_task_cg_lists();
2006 pr_err("cgroup2: unknown option \"%s\"\n", (char *)data
);
2007 return ERR_PTR(-EINVAL
);
2009 cgrp_dfl_visible
= true;
2010 root
= &cgrp_dfl_root
;
2011 cgroup_get(&root
->cgrp
);
2015 mutex_lock(&cgroup_mutex
);
2017 /* First find the desired set of subsystems */
2018 ret
= parse_cgroupfs_options(data
, &opts
);
2023 * Destruction of cgroup root is asynchronous, so subsystems may
2024 * still be dying after the previous unmount. Let's drain the
2025 * dying subsystems. We just need to ensure that the ones
2026 * unmounted previously finish dying and don't care about new ones
2027 * starting. Testing ref liveliness is good enough.
2029 for_each_subsys(ss
, i
) {
2030 if (!(opts
.subsys_mask
& (1 << i
)) ||
2031 ss
->root
== &cgrp_dfl_root
)
2034 if (!percpu_ref_tryget_live(&ss
->root
->cgrp
.self
.refcnt
)) {
2035 mutex_unlock(&cgroup_mutex
);
2037 ret
= restart_syscall();
2040 cgroup_put(&ss
->root
->cgrp
);
2043 for_each_root(root
) {
2044 bool name_match
= false;
2046 if (root
== &cgrp_dfl_root
)
2050 * If we asked for a name then it must match. Also, if
2051 * name matches but sybsys_mask doesn't, we should fail.
2052 * Remember whether name matched.
2055 if (strcmp(opts
.name
, root
->name
))
2061 * If we asked for subsystems (or explicitly for no
2062 * subsystems) then they must match.
2064 if ((opts
.subsys_mask
|| opts
.none
) &&
2065 (opts
.subsys_mask
!= root
->subsys_mask
)) {
2072 if (root
->flags
^ opts
.flags
)
2073 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2076 * We want to reuse @root whose lifetime is governed by its
2077 * ->cgrp. Let's check whether @root is alive and keep it
2078 * that way. As cgroup_kill_sb() can happen anytime, we
2079 * want to block it by pinning the sb so that @root doesn't
2080 * get killed before mount is complete.
2082 * With the sb pinned, tryget_live can reliably indicate
2083 * whether @root can be reused. If it's being killed,
2084 * drain it. We can use wait_queue for the wait but this
2085 * path is super cold. Let's just sleep a bit and retry.
2087 pinned_sb
= kernfs_pin_sb(root
->kf_root
, NULL
);
2088 if (IS_ERR(pinned_sb
) ||
2089 !percpu_ref_tryget_live(&root
->cgrp
.self
.refcnt
)) {
2090 mutex_unlock(&cgroup_mutex
);
2091 if (!IS_ERR_OR_NULL(pinned_sb
))
2092 deactivate_super(pinned_sb
);
2094 ret
= restart_syscall();
2103 * No such thing, create a new one. name= matching without subsys
2104 * specification is allowed for already existing hierarchies but we
2105 * can't create new one without subsys specification.
2107 if (!opts
.subsys_mask
&& !opts
.none
) {
2112 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
2118 init_cgroup_root(root
, &opts
);
2120 ret
= cgroup_setup_root(root
, opts
.subsys_mask
);
2122 cgroup_free_root(root
);
2125 mutex_unlock(&cgroup_mutex
);
2127 kfree(opts
.release_agent
);
2131 return ERR_PTR(ret
);
2133 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
,
2134 is_v2
? CGROUP2_SUPER_MAGIC
: CGROUP_SUPER_MAGIC
,
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
,
2178 static struct file_system_type cgroup2_fs_type
= {
2180 .mount
= cgroup_mount
,
2181 .kill_sb
= cgroup_kill_sb
,
2185 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2186 * @task: target task
2187 * @buf: the buffer to write the path into
2188 * @buflen: the length of the buffer
2190 * Determine @task's cgroup on the first (the one with the lowest non-zero
2191 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2192 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2193 * cgroup controller callbacks.
2195 * Return value is the same as kernfs_path().
2197 char *task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
2199 struct cgroup_root
*root
;
2200 struct cgroup
*cgrp
;
2201 int hierarchy_id
= 1;
2204 mutex_lock(&cgroup_mutex
);
2205 spin_lock_bh(&css_set_lock
);
2207 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
2210 cgrp
= task_cgroup_from_root(task
, root
);
2211 path
= cgroup_path(cgrp
, buf
, buflen
);
2213 /* if no hierarchy exists, everyone is in "/" */
2214 if (strlcpy(buf
, "/", buflen
) < buflen
)
2218 spin_unlock_bh(&css_set_lock
);
2219 mutex_unlock(&cgroup_mutex
);
2222 EXPORT_SYMBOL_GPL(task_cgroup_path
);
2224 /* used to track tasks and other necessary states during migration */
2225 struct cgroup_taskset
{
2226 /* the src and dst cset list running through cset->mg_node */
2227 struct list_head src_csets
;
2228 struct list_head dst_csets
;
2230 /* the subsys currently being processed */
2234 * Fields for cgroup_taskset_*() iteration.
2236 * Before migration is committed, the target migration tasks are on
2237 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2238 * the csets on ->dst_csets. ->csets point to either ->src_csets
2239 * or ->dst_csets depending on whether migration is committed.
2241 * ->cur_csets and ->cur_task point to the current task position
2244 struct list_head
*csets
;
2245 struct css_set
*cur_cset
;
2246 struct task_struct
*cur_task
;
2249 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2250 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2251 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2252 .csets = &tset.src_csets, \
2256 * cgroup_taskset_add - try to add a migration target task to a taskset
2257 * @task: target task
2258 * @tset: target taskset
2260 * Add @task, which is a migration target, to @tset. This function becomes
2261 * noop if @task doesn't need to be migrated. @task's css_set should have
2262 * been added as a migration source and @task->cg_list will be moved from
2263 * the css_set's tasks list to mg_tasks one.
2265 static void cgroup_taskset_add(struct task_struct
*task
,
2266 struct cgroup_taskset
*tset
)
2268 struct css_set
*cset
;
2270 lockdep_assert_held(&css_set_lock
);
2272 /* @task either already exited or can't exit until the end */
2273 if (task
->flags
& PF_EXITING
)
2276 /* leave @task alone if post_fork() hasn't linked it yet */
2277 if (list_empty(&task
->cg_list
))
2280 cset
= task_css_set(task
);
2281 if (!cset
->mg_src_cgrp
)
2284 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2285 if (list_empty(&cset
->mg_node
))
2286 list_add_tail(&cset
->mg_node
, &tset
->src_csets
);
2287 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2288 list_move_tail(&cset
->mg_dst_cset
->mg_node
,
2293 * cgroup_taskset_first - reset taskset and return the first task
2294 * @tset: taskset of interest
2295 * @dst_cssp: output variable for the destination css
2297 * @tset iteration is initialized and the first task is returned.
2299 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
,
2300 struct cgroup_subsys_state
**dst_cssp
)
2302 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
2303 tset
->cur_task
= NULL
;
2305 return cgroup_taskset_next(tset
, dst_cssp
);
2309 * cgroup_taskset_next - iterate to the next task in taskset
2310 * @tset: taskset of interest
2311 * @dst_cssp: output variable for the destination css
2313 * Return the next task in @tset. Iteration must have been initialized
2314 * with cgroup_taskset_first().
2316 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
,
2317 struct cgroup_subsys_state
**dst_cssp
)
2319 struct css_set
*cset
= tset
->cur_cset
;
2320 struct task_struct
*task
= tset
->cur_task
;
2322 while (&cset
->mg_node
!= tset
->csets
) {
2324 task
= list_first_entry(&cset
->mg_tasks
,
2325 struct task_struct
, cg_list
);
2327 task
= list_next_entry(task
, cg_list
);
2329 if (&task
->cg_list
!= &cset
->mg_tasks
) {
2330 tset
->cur_cset
= cset
;
2331 tset
->cur_task
= task
;
2334 * This function may be called both before and
2335 * after cgroup_taskset_migrate(). The two cases
2336 * can be distinguished by looking at whether @cset
2337 * has its ->mg_dst_cset set.
2339 if (cset
->mg_dst_cset
)
2340 *dst_cssp
= cset
->mg_dst_cset
->subsys
[tset
->ssid
];
2342 *dst_cssp
= cset
->subsys
[tset
->ssid
];
2347 cset
= list_next_entry(cset
, mg_node
);
2355 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2356 * @tset: taget taskset
2357 * @dst_cgrp: destination cgroup
2359 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2360 * ->can_attach callbacks fails and guarantees that either all or none of
2361 * the tasks in @tset are migrated. @tset is consumed regardless of
2364 static int cgroup_taskset_migrate(struct cgroup_taskset
*tset
,
2365 struct cgroup
*dst_cgrp
)
2367 struct cgroup_subsys_state
*css
, *failed_css
= NULL
;
2368 struct task_struct
*task
, *tmp_task
;
2369 struct css_set
*cset
, *tmp_cset
;
2372 /* methods shouldn't be called if no task is actually migrating */
2373 if (list_empty(&tset
->src_csets
))
2376 /* check that we can legitimately attach to the cgroup */
2377 for_each_e_css(css
, i
, dst_cgrp
) {
2378 if (css
->ss
->can_attach
) {
2380 ret
= css
->ss
->can_attach(tset
);
2383 goto out_cancel_attach
;
2389 * Now that we're guaranteed success, proceed to move all tasks to
2390 * the new cgroup. There are no failure cases after here, so this
2391 * is the commit point.
2393 spin_lock_bh(&css_set_lock
);
2394 list_for_each_entry(cset
, &tset
->src_csets
, mg_node
) {
2395 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
) {
2396 struct css_set
*from_cset
= task_css_set(task
);
2397 struct css_set
*to_cset
= cset
->mg_dst_cset
;
2399 get_css_set(to_cset
);
2400 css_set_move_task(task
, from_cset
, to_cset
, true);
2401 put_css_set_locked(from_cset
);
2404 spin_unlock_bh(&css_set_lock
);
2407 * Migration is committed, all target tasks are now on dst_csets.
2408 * Nothing is sensitive to fork() after this point. Notify
2409 * controllers that migration is complete.
2411 tset
->csets
= &tset
->dst_csets
;
2413 for_each_e_css(css
, i
, dst_cgrp
) {
2414 if (css
->ss
->attach
) {
2416 css
->ss
->attach(tset
);
2421 goto out_release_tset
;
2424 for_each_e_css(css
, i
, dst_cgrp
) {
2425 if (css
== failed_css
)
2427 if (css
->ss
->cancel_attach
) {
2429 css
->ss
->cancel_attach(tset
);
2433 spin_lock_bh(&css_set_lock
);
2434 list_splice_init(&tset
->dst_csets
, &tset
->src_csets
);
2435 list_for_each_entry_safe(cset
, tmp_cset
, &tset
->src_csets
, mg_node
) {
2436 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2437 list_del_init(&cset
->mg_node
);
2439 spin_unlock_bh(&css_set_lock
);
2444 * cgroup_migrate_finish - cleanup after attach
2445 * @preloaded_csets: list of preloaded css_sets
2447 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2448 * those functions for details.
2450 static void cgroup_migrate_finish(struct list_head
*preloaded_csets
)
2452 struct css_set
*cset
, *tmp_cset
;
2454 lockdep_assert_held(&cgroup_mutex
);
2456 spin_lock_bh(&css_set_lock
);
2457 list_for_each_entry_safe(cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2458 cset
->mg_src_cgrp
= NULL
;
2459 cset
->mg_dst_cset
= NULL
;
2460 list_del_init(&cset
->mg_preload_node
);
2461 put_css_set_locked(cset
);
2463 spin_unlock_bh(&css_set_lock
);
2467 * cgroup_migrate_add_src - add a migration source css_set
2468 * @src_cset: the source css_set to add
2469 * @dst_cgrp: the destination cgroup
2470 * @preloaded_csets: list of preloaded css_sets
2472 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2473 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2474 * up by cgroup_migrate_finish().
2476 * This function may be called without holding cgroup_threadgroup_rwsem
2477 * even if the target is a process. Threads may be created and destroyed
2478 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2479 * into play and the preloaded css_sets are guaranteed to cover all
2482 static void cgroup_migrate_add_src(struct css_set
*src_cset
,
2483 struct cgroup
*dst_cgrp
,
2484 struct list_head
*preloaded_csets
)
2486 struct cgroup
*src_cgrp
;
2488 lockdep_assert_held(&cgroup_mutex
);
2489 lockdep_assert_held(&css_set_lock
);
2491 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2493 if (!list_empty(&src_cset
->mg_preload_node
))
2496 WARN_ON(src_cset
->mg_src_cgrp
);
2497 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2498 WARN_ON(!list_empty(&src_cset
->mg_node
));
2500 src_cset
->mg_src_cgrp
= src_cgrp
;
2501 get_css_set(src_cset
);
2502 list_add(&src_cset
->mg_preload_node
, preloaded_csets
);
2506 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2507 * @dst_cgrp: the destination cgroup (may be %NULL)
2508 * @preloaded_csets: list of preloaded source css_sets
2510 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2511 * have been preloaded to @preloaded_csets. This function looks up and
2512 * pins all destination css_sets, links each to its source, and append them
2513 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2514 * source css_set is assumed to be its cgroup on the default hierarchy.
2516 * This function must be called after cgroup_migrate_add_src() has been
2517 * called on each migration source css_set. After migration is performed
2518 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2521 static int cgroup_migrate_prepare_dst(struct cgroup
*dst_cgrp
,
2522 struct list_head
*preloaded_csets
)
2525 struct css_set
*src_cset
, *tmp_cset
;
2527 lockdep_assert_held(&cgroup_mutex
);
2530 * Except for the root, subtree_control must be zero for a cgroup
2531 * with tasks so that child cgroups don't compete against tasks.
2533 if (dst_cgrp
&& cgroup_on_dfl(dst_cgrp
) && cgroup_parent(dst_cgrp
) &&
2534 dst_cgrp
->subtree_control
)
2537 /* look up the dst cset for each src cset and link it to src */
2538 list_for_each_entry_safe(src_cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2539 struct css_set
*dst_cset
;
2541 dst_cset
= find_css_set(src_cset
,
2542 dst_cgrp
?: src_cset
->dfl_cgrp
);
2546 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2549 * If src cset equals dst, it's noop. Drop the src.
2550 * cgroup_migrate() will skip the cset too. Note that we
2551 * can't handle src == dst as some nodes are used by both.
2553 if (src_cset
== dst_cset
) {
2554 src_cset
->mg_src_cgrp
= NULL
;
2555 list_del_init(&src_cset
->mg_preload_node
);
2556 put_css_set(src_cset
);
2557 put_css_set(dst_cset
);
2561 src_cset
->mg_dst_cset
= dst_cset
;
2563 if (list_empty(&dst_cset
->mg_preload_node
))
2564 list_add(&dst_cset
->mg_preload_node
, &csets
);
2566 put_css_set(dst_cset
);
2569 list_splice_tail(&csets
, preloaded_csets
);
2572 cgroup_migrate_finish(&csets
);
2577 * cgroup_migrate - migrate a process or task to a cgroup
2578 * @leader: the leader of the process or the task to migrate
2579 * @threadgroup: whether @leader points to the whole process or a single task
2580 * @cgrp: the destination cgroup
2582 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2583 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2584 * caller is also responsible for invoking cgroup_migrate_add_src() and
2585 * cgroup_migrate_prepare_dst() on the targets before invoking this
2586 * function and following up with cgroup_migrate_finish().
2588 * As long as a controller's ->can_attach() doesn't fail, this function is
2589 * guaranteed to succeed. This means that, excluding ->can_attach()
2590 * failure, when migrating multiple targets, the success or failure can be
2591 * decided for all targets by invoking group_migrate_prepare_dst() before
2592 * actually starting migrating.
2594 static int cgroup_migrate(struct task_struct
*leader
, bool threadgroup
,
2595 struct cgroup
*cgrp
)
2597 struct cgroup_taskset tset
= CGROUP_TASKSET_INIT(tset
);
2598 struct task_struct
*task
;
2601 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2602 * already PF_EXITING could be freed from underneath us unless we
2603 * take an rcu_read_lock.
2605 spin_lock_bh(&css_set_lock
);
2609 cgroup_taskset_add(task
, &tset
);
2612 } while_each_thread(leader
, task
);
2614 spin_unlock_bh(&css_set_lock
);
2616 return cgroup_taskset_migrate(&tset
, cgrp
);
2620 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2621 * @dst_cgrp: the cgroup to attach to
2622 * @leader: the task or the leader of the threadgroup to be attached
2623 * @threadgroup: attach the whole threadgroup?
2625 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2627 static int cgroup_attach_task(struct cgroup
*dst_cgrp
,
2628 struct task_struct
*leader
, bool threadgroup
)
2630 LIST_HEAD(preloaded_csets
);
2631 struct task_struct
*task
;
2634 /* look up all src csets */
2635 spin_lock_bh(&css_set_lock
);
2639 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
,
2643 } while_each_thread(leader
, task
);
2645 spin_unlock_bh(&css_set_lock
);
2647 /* prepare dst csets and commit */
2648 ret
= cgroup_migrate_prepare_dst(dst_cgrp
, &preloaded_csets
);
2650 ret
= cgroup_migrate(leader
, threadgroup
, dst_cgrp
);
2652 cgroup_migrate_finish(&preloaded_csets
);
2656 static int cgroup_procs_write_permission(struct task_struct
*task
,
2657 struct cgroup
*dst_cgrp
,
2658 struct kernfs_open_file
*of
)
2660 const struct cred
*cred
= current_cred();
2661 const struct cred
*tcred
= get_task_cred(task
);
2665 * even if we're attaching all tasks in the thread group, we only
2666 * need to check permissions on one of them.
2668 if (!uid_eq(cred
->euid
, GLOBAL_ROOT_UID
) &&
2669 !uid_eq(cred
->euid
, tcred
->uid
) &&
2670 !uid_eq(cred
->euid
, tcred
->suid
))
2673 if (!ret
&& cgroup_on_dfl(dst_cgrp
)) {
2674 struct super_block
*sb
= of
->file
->f_path
.dentry
->d_sb
;
2675 struct cgroup
*cgrp
;
2676 struct inode
*inode
;
2678 spin_lock_bh(&css_set_lock
);
2679 cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
2680 spin_unlock_bh(&css_set_lock
);
2682 while (!cgroup_is_descendant(dst_cgrp
, cgrp
))
2683 cgrp
= cgroup_parent(cgrp
);
2686 inode
= kernfs_get_inode(sb
, cgrp
->procs_file
.kn
);
2688 ret
= inode_permission(inode
, MAY_WRITE
);
2698 * Find the task_struct of the task to attach by vpid and pass it along to the
2699 * function to attach either it or all tasks in its threadgroup. Will lock
2700 * cgroup_mutex and threadgroup.
2702 static ssize_t
__cgroup_procs_write(struct kernfs_open_file
*of
, char *buf
,
2703 size_t nbytes
, loff_t off
, bool threadgroup
)
2705 struct task_struct
*tsk
;
2706 struct cgroup
*cgrp
;
2710 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2713 cgrp
= cgroup_kn_lock_live(of
->kn
);
2717 percpu_down_write(&cgroup_threadgroup_rwsem
);
2720 tsk
= find_task_by_vpid(pid
);
2723 goto out_unlock_rcu
;
2730 tsk
= tsk
->group_leader
;
2733 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2734 * trapped in a cpuset, or RT worker may be born in a cgroup
2735 * with no rt_runtime allocated. Just say no.
2737 if (tsk
== kthreadd_task
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2739 goto out_unlock_rcu
;
2742 get_task_struct(tsk
);
2745 ret
= cgroup_procs_write_permission(tsk
, cgrp
, of
);
2747 ret
= cgroup_attach_task(cgrp
, tsk
, threadgroup
);
2749 put_task_struct(tsk
);
2750 goto out_unlock_threadgroup
;
2754 out_unlock_threadgroup
:
2755 percpu_up_write(&cgroup_threadgroup_rwsem
);
2756 cgroup_kn_unlock(of
->kn
);
2757 cpuset_post_attach_flush();
2758 return ret
?: nbytes
;
2762 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2763 * @from: attach to all cgroups of a given task
2764 * @tsk: the task to be attached
2766 int cgroup_attach_task_all(struct task_struct
*from
, struct task_struct
*tsk
)
2768 struct cgroup_root
*root
;
2771 mutex_lock(&cgroup_mutex
);
2772 for_each_root(root
) {
2773 struct cgroup
*from_cgrp
;
2775 if (root
== &cgrp_dfl_root
)
2778 spin_lock_bh(&css_set_lock
);
2779 from_cgrp
= task_cgroup_from_root(from
, root
);
2780 spin_unlock_bh(&css_set_lock
);
2782 retval
= cgroup_attach_task(from_cgrp
, tsk
, false);
2786 mutex_unlock(&cgroup_mutex
);
2790 EXPORT_SYMBOL_GPL(cgroup_attach_task_all
);
2792 static ssize_t
cgroup_tasks_write(struct kernfs_open_file
*of
,
2793 char *buf
, size_t nbytes
, loff_t off
)
2795 return __cgroup_procs_write(of
, buf
, nbytes
, off
, false);
2798 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
2799 char *buf
, size_t nbytes
, loff_t off
)
2801 return __cgroup_procs_write(of
, buf
, nbytes
, off
, true);
2804 static ssize_t
cgroup_release_agent_write(struct kernfs_open_file
*of
,
2805 char *buf
, size_t nbytes
, loff_t off
)
2807 struct cgroup
*cgrp
;
2809 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
2811 cgrp
= cgroup_kn_lock_live(of
->kn
);
2814 spin_lock(&release_agent_path_lock
);
2815 strlcpy(cgrp
->root
->release_agent_path
, strstrip(buf
),
2816 sizeof(cgrp
->root
->release_agent_path
));
2817 spin_unlock(&release_agent_path_lock
);
2818 cgroup_kn_unlock(of
->kn
);
2822 static int cgroup_release_agent_show(struct seq_file
*seq
, void *v
)
2824 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2826 spin_lock(&release_agent_path_lock
);
2827 seq_puts(seq
, cgrp
->root
->release_agent_path
);
2828 spin_unlock(&release_agent_path_lock
);
2829 seq_putc(seq
, '\n');
2833 static int cgroup_sane_behavior_show(struct seq_file
*seq
, void *v
)
2835 seq_puts(seq
, "0\n");
2839 static void cgroup_print_ss_mask(struct seq_file
*seq
, u16 ss_mask
)
2841 struct cgroup_subsys
*ss
;
2842 bool printed
= false;
2845 do_each_subsys_mask(ss
, ssid
, ss_mask
) {
2848 seq_printf(seq
, "%s", ss
->name
);
2850 } while_each_subsys_mask();
2852 seq_putc(seq
, '\n');
2855 /* show controllers which are currently attached to the default hierarchy */
2856 static int cgroup_root_controllers_show(struct seq_file
*seq
, void *v
)
2858 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2860 cgroup_print_ss_mask(seq
, cgrp
->root
->subsys_mask
&
2861 ~cgrp_dfl_inhibit_ss_mask
);
2865 /* show controllers which are enabled from the parent */
2866 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2868 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2870 cgroup_print_ss_mask(seq
, cgroup_parent(cgrp
)->subtree_control
);
2874 /* show controllers which are enabled for a given cgroup's children */
2875 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2877 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2879 cgroup_print_ss_mask(seq
, cgrp
->subtree_control
);
2884 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2885 * @cgrp: root of the subtree to update csses for
2887 * @cgrp's subtree_ss_mask has changed and its subtree's (self excluded)
2888 * css associations need to be updated accordingly. This function looks up
2889 * all css_sets which are attached to the subtree, creates the matching
2890 * updated css_sets and migrates the tasks to the new ones.
2892 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2894 LIST_HEAD(preloaded_csets
);
2895 struct cgroup_taskset tset
= CGROUP_TASKSET_INIT(tset
);
2896 struct cgroup_subsys_state
*css
;
2897 struct css_set
*src_cset
;
2900 lockdep_assert_held(&cgroup_mutex
);
2902 percpu_down_write(&cgroup_threadgroup_rwsem
);
2904 /* look up all csses currently attached to @cgrp's subtree */
2905 spin_lock_bh(&css_set_lock
);
2906 css_for_each_descendant_pre(css
, cgroup_css(cgrp
, NULL
)) {
2907 struct cgrp_cset_link
*link
;
2909 /* self is not affected by subtree_ss_mask change */
2910 if (css
->cgroup
== cgrp
)
2913 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
)
2914 cgroup_migrate_add_src(link
->cset
, cgrp
,
2917 spin_unlock_bh(&css_set_lock
);
2919 /* NULL dst indicates self on default hierarchy */
2920 ret
= cgroup_migrate_prepare_dst(NULL
, &preloaded_csets
);
2924 spin_lock_bh(&css_set_lock
);
2925 list_for_each_entry(src_cset
, &preloaded_csets
, mg_preload_node
) {
2926 struct task_struct
*task
, *ntask
;
2928 /* src_csets precede dst_csets, break on the first dst_cset */
2929 if (!src_cset
->mg_src_cgrp
)
2932 /* all tasks in src_csets need to be migrated */
2933 list_for_each_entry_safe(task
, ntask
, &src_cset
->tasks
, cg_list
)
2934 cgroup_taskset_add(task
, &tset
);
2936 spin_unlock_bh(&css_set_lock
);
2938 ret
= cgroup_taskset_migrate(&tset
, cgrp
);
2940 cgroup_migrate_finish(&preloaded_csets
);
2941 percpu_up_write(&cgroup_threadgroup_rwsem
);
2945 /* change the enabled child controllers for a cgroup in the default hierarchy */
2946 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
2947 char *buf
, size_t nbytes
,
2950 u16 enable
= 0, disable
= 0;
2951 u16 css_enable
, css_disable
, old_sc
, new_sc
, old_ss
, new_ss
;
2952 struct cgroup
*cgrp
, *child
;
2953 struct cgroup_subsys
*ss
;
2958 * Parse input - space separated list of subsystem names prefixed
2959 * with either + or -.
2961 buf
= strstrip(buf
);
2962 while ((tok
= strsep(&buf
, " "))) {
2965 do_each_subsys_mask(ss
, ssid
, ~cgrp_dfl_inhibit_ss_mask
) {
2966 if (!cgroup_ssid_enabled(ssid
) ||
2967 strcmp(tok
+ 1, ss
->name
))
2971 enable
|= 1 << ssid
;
2972 disable
&= ~(1 << ssid
);
2973 } else if (*tok
== '-') {
2974 disable
|= 1 << ssid
;
2975 enable
&= ~(1 << ssid
);
2980 } while_each_subsys_mask();
2981 if (ssid
== CGROUP_SUBSYS_COUNT
)
2985 cgrp
= cgroup_kn_lock_live(of
->kn
);
2989 for_each_subsys(ss
, ssid
) {
2990 if (enable
& (1 << ssid
)) {
2991 if (cgrp
->subtree_control
& (1 << ssid
)) {
2992 enable
&= ~(1 << ssid
);
2996 /* unavailable or not enabled on the parent? */
2997 if (!(cgrp_dfl_root
.subsys_mask
& (1 << ssid
)) ||
2998 (cgroup_parent(cgrp
) &&
2999 !(cgroup_parent(cgrp
)->subtree_control
& (1 << ssid
)))) {
3003 } else if (disable
& (1 << ssid
)) {
3004 if (!(cgrp
->subtree_control
& (1 << ssid
))) {
3005 disable
&= ~(1 << ssid
);
3009 /* a child has it enabled? */
3010 cgroup_for_each_live_child(child
, cgrp
) {
3011 if (child
->subtree_control
& (1 << ssid
)) {
3019 if (!enable
&& !disable
) {
3025 * Except for the root, subtree_control must be zero for a cgroup
3026 * with tasks so that child cgroups don't compete against tasks.
3028 if (enable
&& cgroup_parent(cgrp
) && !list_empty(&cgrp
->cset_links
)) {
3034 * Update subsys masks and calculate what needs to be done. More
3035 * subsystems than specified may need to be enabled or disabled
3036 * depending on subsystem dependencies.
3038 old_sc
= cgrp
->subtree_control
;
3039 old_ss
= cgrp
->subtree_ss_mask
;
3040 new_sc
= (old_sc
| enable
) & ~disable
;
3041 new_ss
= cgroup_calc_subtree_ss_mask(cgrp
, new_sc
);
3043 css_enable
= ~old_ss
& new_ss
;
3044 css_disable
= old_ss
& ~new_ss
;
3045 enable
|= css_enable
;
3046 disable
|= css_disable
;
3049 * Because css offlining is asynchronous, userland might try to
3050 * re-enable the same controller while the previous instance is
3051 * still around. In such cases, wait till it's gone using
3054 do_each_subsys_mask(ss
, ssid
, css_enable
) {
3055 cgroup_for_each_live_child(child
, cgrp
) {
3058 if (!cgroup_css(child
, ss
))
3062 prepare_to_wait(&child
->offline_waitq
, &wait
,
3063 TASK_UNINTERRUPTIBLE
);
3064 cgroup_kn_unlock(of
->kn
);
3066 finish_wait(&child
->offline_waitq
, &wait
);
3069 return restart_syscall();
3071 } while_each_subsys_mask();
3073 cgrp
->subtree_control
= new_sc
;
3074 cgrp
->subtree_ss_mask
= new_ss
;
3077 * Create new csses or make the existing ones visible. A css is
3078 * created invisible if it's being implicitly enabled through
3079 * dependency. An invisible css is made visible when the userland
3080 * explicitly enables it.
3082 do_each_subsys_mask(ss
, ssid
, enable
) {
3083 cgroup_for_each_live_child(child
, cgrp
) {
3084 if (css_enable
& (1 << ssid
))
3085 ret
= create_css(child
, ss
,
3086 cgrp
->subtree_control
& (1 << ssid
));
3088 ret
= css_populate_dir(cgroup_css(child
, ss
),
3093 } while_each_subsys_mask();
3096 * At this point, cgroup_e_css() results reflect the new csses
3097 * making the following cgroup_update_dfl_csses() properly update
3098 * css associations of all tasks in the subtree.
3100 ret
= cgroup_update_dfl_csses(cgrp
);
3105 * All tasks are migrated out of disabled csses. Kill or hide
3106 * them. A css is hidden when the userland requests it to be
3107 * disabled while other subsystems are still depending on it. The
3108 * css must not actively control resources and be in the vanilla
3109 * state if it's made visible again later. Controllers which may
3110 * be depended upon should provide ->css_reset() for this purpose.
3112 do_each_subsys_mask(ss
, ssid
, disable
) {
3113 cgroup_for_each_live_child(child
, cgrp
) {
3114 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3116 if (css_disable
& (1 << ssid
)) {
3119 css_clear_dir(css
, NULL
);
3124 } while_each_subsys_mask();
3126 kernfs_activate(cgrp
->kn
);
3129 cgroup_kn_unlock(of
->kn
);
3130 return ret
?: nbytes
;
3133 cgrp
->subtree_control
= old_sc
;
3134 cgrp
->subtree_ss_mask
= old_ss
;
3136 do_each_subsys_mask(ss
, ssid
, enable
) {
3137 cgroup_for_each_live_child(child
, cgrp
) {
3138 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3143 if (css_enable
& (1 << ssid
))
3146 css_clear_dir(css
, NULL
);
3148 } while_each_subsys_mask();
3152 static int cgroup_events_show(struct seq_file
*seq
, void *v
)
3154 seq_printf(seq
, "populated %d\n",
3155 cgroup_is_populated(seq_css(seq
)->cgroup
));
3159 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
3160 size_t nbytes
, loff_t off
)
3162 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
3163 struct cftype
*cft
= of
->kn
->priv
;
3164 struct cgroup_subsys_state
*css
;
3168 return cft
->write(of
, buf
, nbytes
, off
);
3171 * kernfs guarantees that a file isn't deleted with operations in
3172 * flight, which means that the matching css is and stays alive and
3173 * doesn't need to be pinned. The RCU locking is not necessary
3174 * either. It's just for the convenience of using cgroup_css().
3177 css
= cgroup_css(cgrp
, cft
->ss
);
3180 if (cft
->write_u64
) {
3181 unsigned long long v
;
3182 ret
= kstrtoull(buf
, 0, &v
);
3184 ret
= cft
->write_u64(css
, cft
, v
);
3185 } else if (cft
->write_s64
) {
3187 ret
= kstrtoll(buf
, 0, &v
);
3189 ret
= cft
->write_s64(css
, cft
, v
);
3194 return ret
?: nbytes
;
3197 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
3199 return seq_cft(seq
)->seq_start(seq
, ppos
);
3202 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
3204 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
3207 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
3209 seq_cft(seq
)->seq_stop(seq
, v
);
3212 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
3214 struct cftype
*cft
= seq_cft(m
);
3215 struct cgroup_subsys_state
*css
= seq_css(m
);
3218 return cft
->seq_show(m
, arg
);
3221 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
3222 else if (cft
->read_s64
)
3223 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
3229 static struct kernfs_ops cgroup_kf_single_ops
= {
3230 .atomic_write_len
= PAGE_SIZE
,
3231 .write
= cgroup_file_write
,
3232 .seq_show
= cgroup_seqfile_show
,
3235 static struct kernfs_ops cgroup_kf_ops
= {
3236 .atomic_write_len
= PAGE_SIZE
,
3237 .write
= cgroup_file_write
,
3238 .seq_start
= cgroup_seqfile_start
,
3239 .seq_next
= cgroup_seqfile_next
,
3240 .seq_stop
= cgroup_seqfile_stop
,
3241 .seq_show
= cgroup_seqfile_show
,
3245 * cgroup_rename - Only allow simple rename of directories in place.
3247 static int cgroup_rename(struct kernfs_node
*kn
, struct kernfs_node
*new_parent
,
3248 const char *new_name_str
)
3250 struct cgroup
*cgrp
= kn
->priv
;
3253 if (kernfs_type(kn
) != KERNFS_DIR
)
3255 if (kn
->parent
!= new_parent
)
3259 * This isn't a proper migration and its usefulness is very
3260 * limited. Disallow on the default hierarchy.
3262 if (cgroup_on_dfl(cgrp
))
3266 * We're gonna grab cgroup_mutex which nests outside kernfs
3267 * active_ref. kernfs_rename() doesn't require active_ref
3268 * protection. Break them before grabbing cgroup_mutex.
3270 kernfs_break_active_protection(new_parent
);
3271 kernfs_break_active_protection(kn
);
3273 mutex_lock(&cgroup_mutex
);
3275 ret
= kernfs_rename(kn
, new_parent
, new_name_str
);
3277 mutex_unlock(&cgroup_mutex
);
3279 kernfs_unbreak_active_protection(kn
);
3280 kernfs_unbreak_active_protection(new_parent
);
3284 /* set uid and gid of cgroup dirs and files to that of the creator */
3285 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
3287 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
3288 .ia_uid
= current_fsuid(),
3289 .ia_gid
= current_fsgid(), };
3291 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
3292 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
3295 return kernfs_setattr(kn
, &iattr
);
3298 static int cgroup_add_file(struct cgroup_subsys_state
*css
, struct cgroup
*cgrp
,
3301 char name
[CGROUP_FILE_NAME_MAX
];
3302 struct kernfs_node
*kn
;
3303 struct lock_class_key
*key
= NULL
;
3306 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3307 key
= &cft
->lockdep_key
;
3309 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
3310 cgroup_file_mode(cft
), 0, cft
->kf_ops
, cft
,
3315 ret
= cgroup_kn_set_ugid(kn
);
3321 if (cft
->file_offset
) {
3322 struct cgroup_file
*cfile
= (void *)css
+ cft
->file_offset
;
3324 spin_lock_irq(&cgroup_file_kn_lock
);
3326 spin_unlock_irq(&cgroup_file_kn_lock
);
3333 * cgroup_addrm_files - add or remove files to a cgroup directory
3334 * @css: the target css
3335 * @cgrp: the target cgroup (usually css->cgroup)
3336 * @cfts: array of cftypes to be added
3337 * @is_add: whether to add or remove
3339 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3340 * For removals, this function never fails.
3342 static int cgroup_addrm_files(struct cgroup_subsys_state
*css
,
3343 struct cgroup
*cgrp
, struct cftype cfts
[],
3346 struct cftype
*cft
, *cft_end
= NULL
;
3349 lockdep_assert_held(&cgroup_mutex
);
3352 for (cft
= cfts
; cft
!= cft_end
&& cft
->name
[0] != '\0'; cft
++) {
3353 /* does cft->flags tell us to skip this file on @cgrp? */
3354 if ((cft
->flags
& __CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
3356 if ((cft
->flags
& __CFTYPE_NOT_ON_DFL
) && cgroup_on_dfl(cgrp
))
3358 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
3360 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
3364 ret
= cgroup_add_file(css
, cgrp
, cft
);
3366 pr_warn("%s: failed to add %s, err=%d\n",
3367 __func__
, cft
->name
, ret
);
3373 cgroup_rm_file(cgrp
, cft
);
3379 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
3382 struct cgroup_subsys
*ss
= cfts
[0].ss
;
3383 struct cgroup
*root
= &ss
->root
->cgrp
;
3384 struct cgroup_subsys_state
*css
;
3387 lockdep_assert_held(&cgroup_mutex
);
3389 /* add/rm files for all cgroups created before */
3390 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
3391 struct cgroup
*cgrp
= css
->cgroup
;
3393 if (cgroup_is_dead(cgrp
))
3396 ret
= cgroup_addrm_files(css
, cgrp
, cfts
, is_add
);
3402 kernfs_activate(root
->kn
);
3406 static void cgroup_exit_cftypes(struct cftype
*cfts
)
3410 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3411 /* free copy for custom atomic_write_len, see init_cftypes() */
3412 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
3417 /* revert flags set by cgroup core while adding @cfts */
3418 cft
->flags
&= ~(__CFTYPE_ONLY_ON_DFL
| __CFTYPE_NOT_ON_DFL
);
3422 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3426 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3427 struct kernfs_ops
*kf_ops
;
3429 WARN_ON(cft
->ss
|| cft
->kf_ops
);
3432 kf_ops
= &cgroup_kf_ops
;
3434 kf_ops
= &cgroup_kf_single_ops
;
3437 * Ugh... if @cft wants a custom max_write_len, we need to
3438 * make a copy of kf_ops to set its atomic_write_len.
3440 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
3441 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
3443 cgroup_exit_cftypes(cfts
);
3446 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3449 cft
->kf_ops
= kf_ops
;
3456 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3458 lockdep_assert_held(&cgroup_mutex
);
3460 if (!cfts
|| !cfts
[0].ss
)
3463 list_del(&cfts
->node
);
3464 cgroup_apply_cftypes(cfts
, false);
3465 cgroup_exit_cftypes(cfts
);
3470 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3471 * @cfts: zero-length name terminated array of cftypes
3473 * Unregister @cfts. Files described by @cfts are removed from all
3474 * existing cgroups and all future cgroups won't have them either. This
3475 * function can be called anytime whether @cfts' subsys is attached or not.
3477 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3480 int cgroup_rm_cftypes(struct cftype
*cfts
)
3484 mutex_lock(&cgroup_mutex
);
3485 ret
= cgroup_rm_cftypes_locked(cfts
);
3486 mutex_unlock(&cgroup_mutex
);
3491 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3492 * @ss: target cgroup subsystem
3493 * @cfts: zero-length name terminated array of cftypes
3495 * Register @cfts to @ss. Files described by @cfts are created for all
3496 * existing cgroups to which @ss is attached and all future cgroups will
3497 * have them too. This function can be called anytime whether @ss is
3500 * Returns 0 on successful registration, -errno on failure. Note that this
3501 * function currently returns 0 as long as @cfts registration is successful
3502 * even if some file creation attempts on existing cgroups fail.
3504 static int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3508 if (!cgroup_ssid_enabled(ss
->id
))
3511 if (!cfts
|| cfts
[0].name
[0] == '\0')
3514 ret
= cgroup_init_cftypes(ss
, cfts
);
3518 mutex_lock(&cgroup_mutex
);
3520 list_add_tail(&cfts
->node
, &ss
->cfts
);
3521 ret
= cgroup_apply_cftypes(cfts
, true);
3523 cgroup_rm_cftypes_locked(cfts
);
3525 mutex_unlock(&cgroup_mutex
);
3530 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3531 * @ss: target cgroup subsystem
3532 * @cfts: zero-length name terminated array of cftypes
3534 * Similar to cgroup_add_cftypes() but the added files are only used for
3535 * the default hierarchy.
3537 int cgroup_add_dfl_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3541 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3542 cft
->flags
|= __CFTYPE_ONLY_ON_DFL
;
3543 return cgroup_add_cftypes(ss
, cfts
);
3547 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3548 * @ss: target cgroup subsystem
3549 * @cfts: zero-length name terminated array of cftypes
3551 * Similar to cgroup_add_cftypes() but the added files are only used for
3552 * the legacy hierarchies.
3554 int cgroup_add_legacy_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3558 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3559 cft
->flags
|= __CFTYPE_NOT_ON_DFL
;
3560 return cgroup_add_cftypes(ss
, cfts
);
3564 * cgroup_file_notify - generate a file modified event for a cgroup_file
3565 * @cfile: target cgroup_file
3567 * @cfile must have been obtained by setting cftype->file_offset.
3569 void cgroup_file_notify(struct cgroup_file
*cfile
)
3571 unsigned long flags
;
3573 spin_lock_irqsave(&cgroup_file_kn_lock
, flags
);
3575 kernfs_notify(cfile
->kn
);
3576 spin_unlock_irqrestore(&cgroup_file_kn_lock
, flags
);
3580 * cgroup_task_count - count the number of tasks in a cgroup.
3581 * @cgrp: the cgroup in question
3583 * Return the number of tasks in the cgroup.
3585 static int cgroup_task_count(const struct cgroup
*cgrp
)
3588 struct cgrp_cset_link
*link
;
3590 spin_lock_bh(&css_set_lock
);
3591 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
3592 count
+= atomic_read(&link
->cset
->refcount
);
3593 spin_unlock_bh(&css_set_lock
);
3598 * css_next_child - find the next child of a given css
3599 * @pos: the current position (%NULL to initiate traversal)
3600 * @parent: css whose children to walk
3602 * This function returns the next child of @parent and should be called
3603 * under either cgroup_mutex or RCU read lock. The only requirement is
3604 * that @parent and @pos are accessible. The next sibling is guaranteed to
3605 * be returned regardless of their states.
3607 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3608 * css which finished ->css_online() is guaranteed to be visible in the
3609 * future iterations and will stay visible until the last reference is put.
3610 * A css which hasn't finished ->css_online() or already finished
3611 * ->css_offline() may show up during traversal. It's each subsystem's
3612 * responsibility to synchronize against on/offlining.
3614 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
3615 struct cgroup_subsys_state
*parent
)
3617 struct cgroup_subsys_state
*next
;
3619 cgroup_assert_mutex_or_rcu_locked();
3622 * @pos could already have been unlinked from the sibling list.
3623 * Once a cgroup is removed, its ->sibling.next is no longer
3624 * updated when its next sibling changes. CSS_RELEASED is set when
3625 * @pos is taken off list, at which time its next pointer is valid,
3626 * and, as releases are serialized, the one pointed to by the next
3627 * pointer is guaranteed to not have started release yet. This
3628 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3629 * critical section, the one pointed to by its next pointer is
3630 * guaranteed to not have finished its RCU grace period even if we
3631 * have dropped rcu_read_lock() inbetween iterations.
3633 * If @pos has CSS_RELEASED set, its next pointer can't be
3634 * dereferenced; however, as each css is given a monotonically
3635 * increasing unique serial number and always appended to the
3636 * sibling list, the next one can be found by walking the parent's
3637 * children until the first css with higher serial number than
3638 * @pos's. While this path can be slower, it happens iff iteration
3639 * races against release and the race window is very small.
3642 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
3643 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
3644 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
3646 list_for_each_entry_rcu(next
, &parent
->children
, sibling
)
3647 if (next
->serial_nr
> pos
->serial_nr
)
3652 * @next, if not pointing to the head, can be dereferenced and is
3655 if (&next
->sibling
!= &parent
->children
)
3661 * css_next_descendant_pre - find the next descendant for pre-order walk
3662 * @pos: the current position (%NULL to initiate traversal)
3663 * @root: css whose descendants to walk
3665 * To be used by css_for_each_descendant_pre(). Find the next descendant
3666 * to visit for pre-order traversal of @root's descendants. @root is
3667 * included in the iteration and the first node to be visited.
3669 * While this function requires cgroup_mutex or RCU read locking, it
3670 * doesn't require the whole traversal to be contained in a single critical
3671 * section. This function will return the correct next descendant as long
3672 * as both @pos and @root are accessible and @pos is a descendant of @root.
3674 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3675 * css which finished ->css_online() is guaranteed to be visible in the
3676 * future iterations and will stay visible until the last reference is put.
3677 * A css which hasn't finished ->css_online() or already finished
3678 * ->css_offline() may show up during traversal. It's each subsystem's
3679 * responsibility to synchronize against on/offlining.
3681 struct cgroup_subsys_state
*
3682 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3683 struct cgroup_subsys_state
*root
)
3685 struct cgroup_subsys_state
*next
;
3687 cgroup_assert_mutex_or_rcu_locked();
3689 /* if first iteration, visit @root */
3693 /* visit the first child if exists */
3694 next
= css_next_child(NULL
, pos
);
3698 /* no child, visit my or the closest ancestor's next sibling */
3699 while (pos
!= root
) {
3700 next
= css_next_child(pos
, pos
->parent
);
3710 * css_rightmost_descendant - return the rightmost descendant of a css
3711 * @pos: css of interest
3713 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3714 * is returned. This can be used during pre-order traversal to skip
3717 * While this function requires cgroup_mutex or RCU read locking, it
3718 * doesn't require the whole traversal to be contained in a single critical
3719 * section. This function will return the correct rightmost descendant as
3720 * long as @pos is accessible.
3722 struct cgroup_subsys_state
*
3723 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3725 struct cgroup_subsys_state
*last
, *tmp
;
3727 cgroup_assert_mutex_or_rcu_locked();
3731 /* ->prev isn't RCU safe, walk ->next till the end */
3733 css_for_each_child(tmp
, last
)
3740 static struct cgroup_subsys_state
*
3741 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3743 struct cgroup_subsys_state
*last
;
3747 pos
= css_next_child(NULL
, pos
);
3754 * css_next_descendant_post - find the next descendant for post-order walk
3755 * @pos: the current position (%NULL to initiate traversal)
3756 * @root: css whose descendants to walk
3758 * To be used by css_for_each_descendant_post(). Find the next descendant
3759 * to visit for post-order traversal of @root's descendants. @root is
3760 * included in the iteration and the last node to be visited.
3762 * While this function requires cgroup_mutex or RCU read locking, it
3763 * doesn't require the whole traversal to be contained in a single critical
3764 * section. This function will return the correct next descendant as long
3765 * as both @pos and @cgroup are accessible and @pos is a descendant of
3768 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3769 * css which finished ->css_online() is guaranteed to be visible in the
3770 * future iterations and will stay visible until the last reference is put.
3771 * A css which hasn't finished ->css_online() or already finished
3772 * ->css_offline() may show up during traversal. It's each subsystem's
3773 * responsibility to synchronize against on/offlining.
3775 struct cgroup_subsys_state
*
3776 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
3777 struct cgroup_subsys_state
*root
)
3779 struct cgroup_subsys_state
*next
;
3781 cgroup_assert_mutex_or_rcu_locked();
3783 /* if first iteration, visit leftmost descendant which may be @root */
3785 return css_leftmost_descendant(root
);
3787 /* if we visited @root, we're done */
3791 /* if there's an unvisited sibling, visit its leftmost descendant */
3792 next
= css_next_child(pos
, pos
->parent
);
3794 return css_leftmost_descendant(next
);
3796 /* no sibling left, visit parent */
3801 * css_has_online_children - does a css have online children
3802 * @css: the target css
3804 * Returns %true if @css has any online children; otherwise, %false. This
3805 * function can be called from any context but the caller is responsible
3806 * for synchronizing against on/offlining as necessary.
3808 bool css_has_online_children(struct cgroup_subsys_state
*css
)
3810 struct cgroup_subsys_state
*child
;
3814 css_for_each_child(child
, css
) {
3815 if (child
->flags
& CSS_ONLINE
) {
3825 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
3826 * @it: the iterator to advance
3828 * Advance @it to the next css_set to walk.
3830 static void css_task_iter_advance_css_set(struct css_task_iter
*it
)
3832 struct list_head
*l
= it
->cset_pos
;
3833 struct cgrp_cset_link
*link
;
3834 struct css_set
*cset
;
3836 lockdep_assert_held(&css_set_lock
);
3838 /* Advance to the next non-empty css_set */
3841 if (l
== it
->cset_head
) {
3842 it
->cset_pos
= NULL
;
3843 it
->task_pos
= NULL
;
3848 cset
= container_of(l
, struct css_set
,
3849 e_cset_node
[it
->ss
->id
]);
3851 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
3854 } while (!css_set_populated(cset
));
3858 if (!list_empty(&cset
->tasks
))
3859 it
->task_pos
= cset
->tasks
.next
;
3861 it
->task_pos
= cset
->mg_tasks
.next
;
3863 it
->tasks_head
= &cset
->tasks
;
3864 it
->mg_tasks_head
= &cset
->mg_tasks
;
3867 * We don't keep css_sets locked across iteration steps and thus
3868 * need to take steps to ensure that iteration can be resumed after
3869 * the lock is re-acquired. Iteration is performed at two levels -
3870 * css_sets and tasks in them.
3872 * Once created, a css_set never leaves its cgroup lists, so a
3873 * pinned css_set is guaranteed to stay put and we can resume
3874 * iteration afterwards.
3876 * Tasks may leave @cset across iteration steps. This is resolved
3877 * by registering each iterator with the css_set currently being
3878 * walked and making css_set_move_task() advance iterators whose
3879 * next task is leaving.
3882 list_del(&it
->iters_node
);
3883 put_css_set_locked(it
->cur_cset
);
3886 it
->cur_cset
= cset
;
3887 list_add(&it
->iters_node
, &cset
->task_iters
);
3890 static void css_task_iter_advance(struct css_task_iter
*it
)
3892 struct list_head
*l
= it
->task_pos
;
3894 lockdep_assert_held(&css_set_lock
);
3898 * Advance iterator to find next entry. cset->tasks is consumed
3899 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3904 if (l
== it
->tasks_head
)
3905 l
= it
->mg_tasks_head
->next
;
3907 if (l
== it
->mg_tasks_head
)
3908 css_task_iter_advance_css_set(it
);
3914 * css_task_iter_start - initiate task iteration
3915 * @css: the css to walk tasks of
3916 * @it: the task iterator to use
3918 * Initiate iteration through the tasks of @css. The caller can call
3919 * css_task_iter_next() to walk through the tasks until the function
3920 * returns NULL. On completion of iteration, css_task_iter_end() must be
3923 void css_task_iter_start(struct cgroup_subsys_state
*css
,
3924 struct css_task_iter
*it
)
3926 /* no one should try to iterate before mounting cgroups */
3927 WARN_ON_ONCE(!use_task_css_set_links
);
3929 memset(it
, 0, sizeof(*it
));
3931 spin_lock_bh(&css_set_lock
);
3936 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
3938 it
->cset_pos
= &css
->cgroup
->cset_links
;
3940 it
->cset_head
= it
->cset_pos
;
3942 css_task_iter_advance_css_set(it
);
3944 spin_unlock_bh(&css_set_lock
);
3948 * css_task_iter_next - return the next task for the iterator
3949 * @it: the task iterator being iterated
3951 * The "next" function for task iteration. @it should have been
3952 * initialized via css_task_iter_start(). Returns NULL when the iteration
3955 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
3958 put_task_struct(it
->cur_task
);
3959 it
->cur_task
= NULL
;
3962 spin_lock_bh(&css_set_lock
);
3965 it
->cur_task
= list_entry(it
->task_pos
, struct task_struct
,
3967 get_task_struct(it
->cur_task
);
3968 css_task_iter_advance(it
);
3971 spin_unlock_bh(&css_set_lock
);
3973 return it
->cur_task
;
3977 * css_task_iter_end - finish task iteration
3978 * @it: the task iterator to finish
3980 * Finish task iteration started by css_task_iter_start().
3982 void css_task_iter_end(struct css_task_iter
*it
)
3985 spin_lock_bh(&css_set_lock
);
3986 list_del(&it
->iters_node
);
3987 put_css_set_locked(it
->cur_cset
);
3988 spin_unlock_bh(&css_set_lock
);
3992 put_task_struct(it
->cur_task
);
3996 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3997 * @to: cgroup to which the tasks will be moved
3998 * @from: cgroup in which the tasks currently reside
4000 * Locking rules between cgroup_post_fork() and the migration path
4001 * guarantee that, if a task is forking while being migrated, the new child
4002 * is guaranteed to be either visible in the source cgroup after the
4003 * parent's migration is complete or put into the target cgroup. No task
4004 * can slip out of migration through forking.
4006 int cgroup_transfer_tasks(struct cgroup
*to
, struct cgroup
*from
)
4008 LIST_HEAD(preloaded_csets
);
4009 struct cgrp_cset_link
*link
;
4010 struct css_task_iter it
;
4011 struct task_struct
*task
;
4014 mutex_lock(&cgroup_mutex
);
4016 /* all tasks in @from are being moved, all csets are source */
4017 spin_lock_bh(&css_set_lock
);
4018 list_for_each_entry(link
, &from
->cset_links
, cset_link
)
4019 cgroup_migrate_add_src(link
->cset
, to
, &preloaded_csets
);
4020 spin_unlock_bh(&css_set_lock
);
4022 ret
= cgroup_migrate_prepare_dst(to
, &preloaded_csets
);
4027 * Migrate tasks one-by-one until @from is empty. This fails iff
4028 * ->can_attach() fails.
4031 css_task_iter_start(&from
->self
, &it
);
4032 task
= css_task_iter_next(&it
);
4034 get_task_struct(task
);
4035 css_task_iter_end(&it
);
4038 ret
= cgroup_migrate(task
, false, to
);
4039 put_task_struct(task
);
4041 } while (task
&& !ret
);
4043 cgroup_migrate_finish(&preloaded_csets
);
4044 mutex_unlock(&cgroup_mutex
);
4049 * Stuff for reading the 'tasks'/'procs' files.
4051 * Reading this file can return large amounts of data if a cgroup has
4052 * *lots* of attached tasks. So it may need several calls to read(),
4053 * but we cannot guarantee that the information we produce is correct
4054 * unless we produce it entirely atomically.
4058 /* which pidlist file are we talking about? */
4059 enum cgroup_filetype
{
4065 * A pidlist is a list of pids that virtually represents the contents of one
4066 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
4067 * a pair (one each for procs, tasks) for each pid namespace that's relevant
4070 struct cgroup_pidlist
{
4072 * used to find which pidlist is wanted. doesn't change as long as
4073 * this particular list stays in the list.
4075 struct { enum cgroup_filetype type
; struct pid_namespace
*ns
; } key
;
4078 /* how many elements the above list has */
4080 /* each of these stored in a list by its cgroup */
4081 struct list_head links
;
4082 /* pointer to the cgroup we belong to, for list removal purposes */
4083 struct cgroup
*owner
;
4084 /* for delayed destruction */
4085 struct delayed_work destroy_dwork
;
4089 * The following two functions "fix" the issue where there are more pids
4090 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
4091 * TODO: replace with a kernel-wide solution to this problem
4093 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4094 static void *pidlist_allocate(int count
)
4096 if (PIDLIST_TOO_LARGE(count
))
4097 return vmalloc(count
* sizeof(pid_t
));
4099 return kmalloc(count
* sizeof(pid_t
), GFP_KERNEL
);
4102 static void pidlist_free(void *p
)
4108 * Used to destroy all pidlists lingering waiting for destroy timer. None
4109 * should be left afterwards.
4111 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
)
4113 struct cgroup_pidlist
*l
, *tmp_l
;
4115 mutex_lock(&cgrp
->pidlist_mutex
);
4116 list_for_each_entry_safe(l
, tmp_l
, &cgrp
->pidlists
, links
)
4117 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
, 0);
4118 mutex_unlock(&cgrp
->pidlist_mutex
);
4120 flush_workqueue(cgroup_pidlist_destroy_wq
);
4121 BUG_ON(!list_empty(&cgrp
->pidlists
));
4124 static void cgroup_pidlist_destroy_work_fn(struct work_struct
*work
)
4126 struct delayed_work
*dwork
= to_delayed_work(work
);
4127 struct cgroup_pidlist
*l
= container_of(dwork
, struct cgroup_pidlist
,
4129 struct cgroup_pidlist
*tofree
= NULL
;
4131 mutex_lock(&l
->owner
->pidlist_mutex
);
4134 * Destroy iff we didn't get queued again. The state won't change
4135 * as destroy_dwork can only be queued while locked.
4137 if (!delayed_work_pending(dwork
)) {
4138 list_del(&l
->links
);
4139 pidlist_free(l
->list
);
4140 put_pid_ns(l
->key
.ns
);
4144 mutex_unlock(&l
->owner
->pidlist_mutex
);
4149 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4150 * Returns the number of unique elements.
4152 static int pidlist_uniq(pid_t
*list
, int length
)
4157 * we presume the 0th element is unique, so i starts at 1. trivial
4158 * edge cases first; no work needs to be done for either
4160 if (length
== 0 || length
== 1)
4162 /* src and dest walk down the list; dest counts unique elements */
4163 for (src
= 1; src
< length
; src
++) {
4164 /* find next unique element */
4165 while (list
[src
] == list
[src
-1]) {
4170 /* dest always points to where the next unique element goes */
4171 list
[dest
] = list
[src
];
4179 * The two pid files - task and cgroup.procs - guaranteed that the result
4180 * is sorted, which forced this whole pidlist fiasco. As pid order is
4181 * different per namespace, each namespace needs differently sorted list,
4182 * making it impossible to use, for example, single rbtree of member tasks
4183 * sorted by task pointer. As pidlists can be fairly large, allocating one
4184 * per open file is dangerous, so cgroup had to implement shared pool of
4185 * pidlists keyed by cgroup and namespace.
4187 * All this extra complexity was caused by the original implementation
4188 * committing to an entirely unnecessary property. In the long term, we
4189 * want to do away with it. Explicitly scramble sort order if on the
4190 * default hierarchy so that no such expectation exists in the new
4193 * Scrambling is done by swapping every two consecutive bits, which is
4194 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4196 static pid_t
pid_fry(pid_t pid
)
4198 unsigned a
= pid
& 0x55555555;
4199 unsigned b
= pid
& 0xAAAAAAAA;
4201 return (a
<< 1) | (b
>> 1);
4204 static pid_t
cgroup_pid_fry(struct cgroup
*cgrp
, pid_t pid
)
4206 if (cgroup_on_dfl(cgrp
))
4207 return pid_fry(pid
);
4212 static int cmppid(const void *a
, const void *b
)
4214 return *(pid_t
*)a
- *(pid_t
*)b
;
4217 static int fried_cmppid(const void *a
, const void *b
)
4219 return pid_fry(*(pid_t
*)a
) - pid_fry(*(pid_t
*)b
);
4222 static struct cgroup_pidlist
*cgroup_pidlist_find(struct cgroup
*cgrp
,
4223 enum cgroup_filetype type
)
4225 struct cgroup_pidlist
*l
;
4226 /* don't need task_nsproxy() if we're looking at ourself */
4227 struct pid_namespace
*ns
= task_active_pid_ns(current
);
4229 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4231 list_for_each_entry(l
, &cgrp
->pidlists
, links
)
4232 if (l
->key
.type
== type
&& l
->key
.ns
== ns
)
4238 * find the appropriate pidlist for our purpose (given procs vs tasks)
4239 * returns with the lock on that pidlist already held, and takes care
4240 * of the use count, or returns NULL with no locks held if we're out of
4243 static struct cgroup_pidlist
*cgroup_pidlist_find_create(struct cgroup
*cgrp
,
4244 enum cgroup_filetype type
)
4246 struct cgroup_pidlist
*l
;
4248 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4250 l
= cgroup_pidlist_find(cgrp
, type
);
4254 /* entry not found; create a new one */
4255 l
= kzalloc(sizeof(struct cgroup_pidlist
), GFP_KERNEL
);
4259 INIT_DELAYED_WORK(&l
->destroy_dwork
, cgroup_pidlist_destroy_work_fn
);
4261 /* don't need task_nsproxy() if we're looking at ourself */
4262 l
->key
.ns
= get_pid_ns(task_active_pid_ns(current
));
4264 list_add(&l
->links
, &cgrp
->pidlists
);
4269 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4271 static int pidlist_array_load(struct cgroup
*cgrp
, enum cgroup_filetype type
,
4272 struct cgroup_pidlist
**lp
)
4276 int pid
, n
= 0; /* used for populating the array */
4277 struct css_task_iter it
;
4278 struct task_struct
*tsk
;
4279 struct cgroup_pidlist
*l
;
4281 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4284 * If cgroup gets more users after we read count, we won't have
4285 * enough space - tough. This race is indistinguishable to the
4286 * caller from the case that the additional cgroup users didn't
4287 * show up until sometime later on.
4289 length
= cgroup_task_count(cgrp
);
4290 array
= pidlist_allocate(length
);
4293 /* now, populate the array */
4294 css_task_iter_start(&cgrp
->self
, &it
);
4295 while ((tsk
= css_task_iter_next(&it
))) {
4296 if (unlikely(n
== length
))
4298 /* get tgid or pid for procs or tasks file respectively */
4299 if (type
== CGROUP_FILE_PROCS
)
4300 pid
= task_tgid_vnr(tsk
);
4302 pid
= task_pid_vnr(tsk
);
4303 if (pid
> 0) /* make sure to only use valid results */
4306 css_task_iter_end(&it
);
4308 /* now sort & (if procs) strip out duplicates */
4309 if (cgroup_on_dfl(cgrp
))
4310 sort(array
, length
, sizeof(pid_t
), fried_cmppid
, NULL
);
4312 sort(array
, length
, sizeof(pid_t
), cmppid
, NULL
);
4313 if (type
== CGROUP_FILE_PROCS
)
4314 length
= pidlist_uniq(array
, length
);
4316 l
= cgroup_pidlist_find_create(cgrp
, type
);
4318 pidlist_free(array
);
4322 /* store array, freeing old if necessary */
4323 pidlist_free(l
->list
);
4331 * cgroupstats_build - build and fill cgroupstats
4332 * @stats: cgroupstats to fill information into
4333 * @dentry: A dentry entry belonging to the cgroup for which stats have
4336 * Build and fill cgroupstats so that taskstats can export it to user
4339 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
4341 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
4342 struct cgroup
*cgrp
;
4343 struct css_task_iter it
;
4344 struct task_struct
*tsk
;
4346 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4347 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
4348 kernfs_type(kn
) != KERNFS_DIR
)
4351 mutex_lock(&cgroup_mutex
);
4354 * We aren't being called from kernfs and there's no guarantee on
4355 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4356 * @kn->priv is RCU safe. Let's do the RCU dancing.
4359 cgrp
= rcu_dereference(kn
->priv
);
4360 if (!cgrp
|| cgroup_is_dead(cgrp
)) {
4362 mutex_unlock(&cgroup_mutex
);
4367 css_task_iter_start(&cgrp
->self
, &it
);
4368 while ((tsk
= css_task_iter_next(&it
))) {
4369 switch (tsk
->state
) {
4371 stats
->nr_running
++;
4373 case TASK_INTERRUPTIBLE
:
4374 stats
->nr_sleeping
++;
4376 case TASK_UNINTERRUPTIBLE
:
4377 stats
->nr_uninterruptible
++;
4380 stats
->nr_stopped
++;
4383 if (delayacct_is_task_waiting_on_io(tsk
))
4384 stats
->nr_io_wait
++;
4388 css_task_iter_end(&it
);
4390 mutex_unlock(&cgroup_mutex
);
4396 * seq_file methods for the tasks/procs files. The seq_file position is the
4397 * next pid to display; the seq_file iterator is a pointer to the pid
4398 * in the cgroup->l->list array.
4401 static void *cgroup_pidlist_start(struct seq_file
*s
, loff_t
*pos
)
4404 * Initially we receive a position value that corresponds to
4405 * one more than the last pid shown (or 0 on the first call or
4406 * after a seek to the start). Use a binary-search to find the
4407 * next pid to display, if any
4409 struct kernfs_open_file
*of
= s
->private;
4410 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4411 struct cgroup_pidlist
*l
;
4412 enum cgroup_filetype type
= seq_cft(s
)->private;
4413 int index
= 0, pid
= *pos
;
4416 mutex_lock(&cgrp
->pidlist_mutex
);
4419 * !NULL @of->priv indicates that this isn't the first start()
4420 * after open. If the matching pidlist is around, we can use that.
4421 * Look for it. Note that @of->priv can't be used directly. It
4422 * could already have been destroyed.
4425 of
->priv
= cgroup_pidlist_find(cgrp
, type
);
4428 * Either this is the first start() after open or the matching
4429 * pidlist has been destroyed inbetween. Create a new one.
4432 ret
= pidlist_array_load(cgrp
, type
,
4433 (struct cgroup_pidlist
**)&of
->priv
);
4435 return ERR_PTR(ret
);
4440 int end
= l
->length
;
4442 while (index
< end
) {
4443 int mid
= (index
+ end
) / 2;
4444 if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) == pid
) {
4447 } else if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) <= pid
)
4453 /* If we're off the end of the array, we're done */
4454 if (index
>= l
->length
)
4456 /* Update the abstract position to be the actual pid that we found */
4457 iter
= l
->list
+ index
;
4458 *pos
= cgroup_pid_fry(cgrp
, *iter
);
4462 static void cgroup_pidlist_stop(struct seq_file
*s
, void *v
)
4464 struct kernfs_open_file
*of
= s
->private;
4465 struct cgroup_pidlist
*l
= of
->priv
;
4468 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
,
4469 CGROUP_PIDLIST_DESTROY_DELAY
);
4470 mutex_unlock(&seq_css(s
)->cgroup
->pidlist_mutex
);
4473 static void *cgroup_pidlist_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
4475 struct kernfs_open_file
*of
= s
->private;
4476 struct cgroup_pidlist
*l
= of
->priv
;
4478 pid_t
*end
= l
->list
+ l
->length
;
4480 * Advance to the next pid in the array. If this goes off the
4487 *pos
= cgroup_pid_fry(seq_css(s
)->cgroup
, *p
);
4492 static int cgroup_pidlist_show(struct seq_file
*s
, void *v
)
4494 seq_printf(s
, "%d\n", *(int *)v
);
4499 static u64
cgroup_read_notify_on_release(struct cgroup_subsys_state
*css
,
4502 return notify_on_release(css
->cgroup
);
4505 static int cgroup_write_notify_on_release(struct cgroup_subsys_state
*css
,
4506 struct cftype
*cft
, u64 val
)
4509 set_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4511 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4515 static u64
cgroup_clone_children_read(struct cgroup_subsys_state
*css
,
4518 return test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4521 static int cgroup_clone_children_write(struct cgroup_subsys_state
*css
,
4522 struct cftype
*cft
, u64 val
)
4525 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4527 clear_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4531 /* cgroup core interface files for the default hierarchy */
4532 static struct cftype cgroup_dfl_base_files
[] = {
4534 .name
= "cgroup.procs",
4535 .file_offset
= offsetof(struct cgroup
, procs_file
),
4536 .seq_start
= cgroup_pidlist_start
,
4537 .seq_next
= cgroup_pidlist_next
,
4538 .seq_stop
= cgroup_pidlist_stop
,
4539 .seq_show
= cgroup_pidlist_show
,
4540 .private = CGROUP_FILE_PROCS
,
4541 .write
= cgroup_procs_write
,
4544 .name
= "cgroup.controllers",
4545 .flags
= CFTYPE_ONLY_ON_ROOT
,
4546 .seq_show
= cgroup_root_controllers_show
,
4549 .name
= "cgroup.controllers",
4550 .flags
= CFTYPE_NOT_ON_ROOT
,
4551 .seq_show
= cgroup_controllers_show
,
4554 .name
= "cgroup.subtree_control",
4555 .seq_show
= cgroup_subtree_control_show
,
4556 .write
= cgroup_subtree_control_write
,
4559 .name
= "cgroup.events",
4560 .flags
= CFTYPE_NOT_ON_ROOT
,
4561 .file_offset
= offsetof(struct cgroup
, events_file
),
4562 .seq_show
= cgroup_events_show
,
4567 /* cgroup core interface files for the legacy hierarchies */
4568 static struct cftype cgroup_legacy_base_files
[] = {
4570 .name
= "cgroup.procs",
4571 .seq_start
= cgroup_pidlist_start
,
4572 .seq_next
= cgroup_pidlist_next
,
4573 .seq_stop
= cgroup_pidlist_stop
,
4574 .seq_show
= cgroup_pidlist_show
,
4575 .private = CGROUP_FILE_PROCS
,
4576 .write
= cgroup_procs_write
,
4579 .name
= "cgroup.clone_children",
4580 .read_u64
= cgroup_clone_children_read
,
4581 .write_u64
= cgroup_clone_children_write
,
4584 .name
= "cgroup.sane_behavior",
4585 .flags
= CFTYPE_ONLY_ON_ROOT
,
4586 .seq_show
= cgroup_sane_behavior_show
,
4590 .seq_start
= cgroup_pidlist_start
,
4591 .seq_next
= cgroup_pidlist_next
,
4592 .seq_stop
= cgroup_pidlist_stop
,
4593 .seq_show
= cgroup_pidlist_show
,
4594 .private = CGROUP_FILE_TASKS
,
4595 .write
= cgroup_tasks_write
,
4598 .name
= "notify_on_release",
4599 .read_u64
= cgroup_read_notify_on_release
,
4600 .write_u64
= cgroup_write_notify_on_release
,
4603 .name
= "release_agent",
4604 .flags
= CFTYPE_ONLY_ON_ROOT
,
4605 .seq_show
= cgroup_release_agent_show
,
4606 .write
= cgroup_release_agent_write
,
4607 .max_write_len
= PATH_MAX
- 1,
4613 * css destruction is four-stage process.
4615 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4616 * Implemented in kill_css().
4618 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4619 * and thus css_tryget_online() is guaranteed to fail, the css can be
4620 * offlined by invoking offline_css(). After offlining, the base ref is
4621 * put. Implemented in css_killed_work_fn().
4623 * 3. When the percpu_ref reaches zero, the only possible remaining
4624 * accessors are inside RCU read sections. css_release() schedules the
4627 * 4. After the grace period, the css can be freed. Implemented in
4628 * css_free_work_fn().
4630 * It is actually hairier because both step 2 and 4 require process context
4631 * and thus involve punting to css->destroy_work adding two additional
4632 * steps to the already complex sequence.
4634 static void css_free_work_fn(struct work_struct
*work
)
4636 struct cgroup_subsys_state
*css
=
4637 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4638 struct cgroup_subsys
*ss
= css
->ss
;
4639 struct cgroup
*cgrp
= css
->cgroup
;
4641 percpu_ref_exit(&css
->refcnt
);
4645 struct cgroup_subsys_state
*parent
= css
->parent
;
4649 cgroup_idr_remove(&ss
->css_idr
, id
);
4655 /* cgroup free path */
4656 atomic_dec(&cgrp
->root
->nr_cgrps
);
4657 cgroup_pidlist_destroy_all(cgrp
);
4658 cancel_work_sync(&cgrp
->release_agent_work
);
4660 if (cgroup_parent(cgrp
)) {
4662 * We get a ref to the parent, and put the ref when
4663 * this cgroup is being freed, so it's guaranteed
4664 * that the parent won't be destroyed before its
4667 cgroup_put(cgroup_parent(cgrp
));
4668 kernfs_put(cgrp
->kn
);
4672 * This is root cgroup's refcnt reaching zero,
4673 * which indicates that the root should be
4676 cgroup_destroy_root(cgrp
->root
);
4681 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4683 struct cgroup_subsys_state
*css
=
4684 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4686 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4687 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4690 static void css_release_work_fn(struct work_struct
*work
)
4692 struct cgroup_subsys_state
*css
=
4693 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4694 struct cgroup_subsys
*ss
= css
->ss
;
4695 struct cgroup
*cgrp
= css
->cgroup
;
4697 mutex_lock(&cgroup_mutex
);
4699 css
->flags
|= CSS_RELEASED
;
4700 list_del_rcu(&css
->sibling
);
4703 /* css release path */
4704 cgroup_idr_replace(&ss
->css_idr
, NULL
, css
->id
);
4705 if (ss
->css_released
)
4706 ss
->css_released(css
);
4708 /* cgroup release path */
4709 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
4713 * There are two control paths which try to determine
4714 * cgroup from dentry without going through kernfs -
4715 * cgroupstats_build() and css_tryget_online_from_dir().
4716 * Those are supported by RCU protecting clearing of
4717 * cgrp->kn->priv backpointer.
4719 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
, NULL
);
4722 mutex_unlock(&cgroup_mutex
);
4724 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4727 static void css_release(struct percpu_ref
*ref
)
4729 struct cgroup_subsys_state
*css
=
4730 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4732 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
4733 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4736 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4737 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4739 lockdep_assert_held(&cgroup_mutex
);
4743 memset(css
, 0, sizeof(*css
));
4746 INIT_LIST_HEAD(&css
->sibling
);
4747 INIT_LIST_HEAD(&css
->children
);
4748 css
->serial_nr
= css_serial_nr_next
++;
4749 atomic_set(&css
->online_cnt
, 0);
4751 if (cgroup_parent(cgrp
)) {
4752 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
4753 css_get(css
->parent
);
4756 BUG_ON(cgroup_css(cgrp
, ss
));
4759 /* invoke ->css_online() on a new CSS and mark it online if successful */
4760 static int online_css(struct cgroup_subsys_state
*css
)
4762 struct cgroup_subsys
*ss
= css
->ss
;
4765 lockdep_assert_held(&cgroup_mutex
);
4768 ret
= ss
->css_online(css
);
4770 css
->flags
|= CSS_ONLINE
;
4771 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4773 atomic_inc(&css
->online_cnt
);
4775 atomic_inc(&css
->parent
->online_cnt
);
4780 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4781 static void offline_css(struct cgroup_subsys_state
*css
)
4783 struct cgroup_subsys
*ss
= css
->ss
;
4785 lockdep_assert_held(&cgroup_mutex
);
4787 if (!(css
->flags
& CSS_ONLINE
))
4790 if (ss
->css_offline
)
4791 ss
->css_offline(css
);
4793 css
->flags
&= ~CSS_ONLINE
;
4794 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4796 wake_up_all(&css
->cgroup
->offline_waitq
);
4800 * create_css - create a cgroup_subsys_state
4801 * @cgrp: the cgroup new css will be associated with
4802 * @ss: the subsys of new css
4803 * @visible: whether to create control knobs for the new css or not
4805 * Create a new css associated with @cgrp - @ss pair. On success, the new
4806 * css is online and installed in @cgrp with all interface files created if
4807 * @visible. Returns 0 on success, -errno on failure.
4809 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
4812 struct cgroup
*parent
= cgroup_parent(cgrp
);
4813 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
4814 struct cgroup_subsys_state
*css
;
4817 lockdep_assert_held(&cgroup_mutex
);
4819 css
= ss
->css_alloc(parent_css
);
4821 return PTR_ERR(css
);
4823 init_and_link_css(css
, ss
, cgrp
);
4825 err
= percpu_ref_init(&css
->refcnt
, css_release
, 0, GFP_KERNEL
);
4829 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_KERNEL
);
4831 goto err_free_percpu_ref
;
4835 err
= css_populate_dir(css
, NULL
);
4840 /* @css is ready to be brought online now, make it visible */
4841 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
4842 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4844 err
= online_css(css
);
4848 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4849 cgroup_parent(parent
)) {
4850 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4851 current
->comm
, current
->pid
, ss
->name
);
4852 if (!strcmp(ss
->name
, "memory"))
4853 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4854 ss
->warned_broken_hierarchy
= true;
4860 list_del_rcu(&css
->sibling
);
4861 css_clear_dir(css
, NULL
);
4863 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4864 err_free_percpu_ref
:
4865 percpu_ref_exit(&css
->refcnt
);
4867 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4871 static int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
,
4874 struct cgroup
*parent
, *cgrp
, *tcgrp
;
4875 struct cgroup_root
*root
;
4876 struct cgroup_subsys
*ss
;
4877 struct kernfs_node
*kn
;
4878 int level
, ssid
, ret
;
4880 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4882 if (strchr(name
, '\n'))
4885 parent
= cgroup_kn_lock_live(parent_kn
);
4888 root
= parent
->root
;
4889 level
= parent
->level
+ 1;
4891 /* allocate the cgroup and its ID, 0 is reserved for the root */
4892 cgrp
= kzalloc(sizeof(*cgrp
) +
4893 sizeof(cgrp
->ancestor_ids
[0]) * (level
+ 1), GFP_KERNEL
);
4899 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
, 0, GFP_KERNEL
);
4904 * Temporarily set the pointer to NULL, so idr_find() won't return
4905 * a half-baked cgroup.
4907 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_KERNEL
);
4910 goto out_cancel_ref
;
4913 init_cgroup_housekeeping(cgrp
);
4915 cgrp
->self
.parent
= &parent
->self
;
4917 cgrp
->level
= level
;
4919 for (tcgrp
= cgrp
; tcgrp
; tcgrp
= cgroup_parent(tcgrp
))
4920 cgrp
->ancestor_ids
[tcgrp
->level
] = tcgrp
->id
;
4922 if (notify_on_release(parent
))
4923 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4925 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4926 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4928 /* create the directory */
4929 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4937 * This extra ref will be put in cgroup_free_fn() and guarantees
4938 * that @cgrp->kn is always accessible.
4942 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
4944 /* allocation complete, commit to creation */
4945 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
4946 atomic_inc(&root
->nr_cgrps
);
4950 * @cgrp is now fully operational. If something fails after this
4951 * point, it'll be released via the normal destruction path.
4953 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4955 ret
= cgroup_kn_set_ugid(kn
);
4959 ret
= css_populate_dir(&cgrp
->self
, NULL
);
4963 /* let's create and online css's */
4964 do_each_subsys_mask(ss
, ssid
, parent
->subtree_ss_mask
) {
4965 ret
= create_css(cgrp
, ss
,
4966 parent
->subtree_control
& (1 << ssid
));
4969 } while_each_subsys_mask();
4972 * On the default hierarchy, a child doesn't automatically inherit
4973 * subtree_control from the parent. Each is configured manually.
4975 if (!cgroup_on_dfl(cgrp
)) {
4976 cgrp
->subtree_control
= parent
->subtree_control
;
4977 cgroup_refresh_subtree_ss_mask(cgrp
);
4980 kernfs_activate(kn
);
4986 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4988 percpu_ref_exit(&cgrp
->self
.refcnt
);
4992 cgroup_kn_unlock(parent_kn
);
4996 cgroup_destroy_locked(cgrp
);
5001 * This is called when the refcnt of a css is confirmed to be killed.
5002 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
5003 * initate destruction and put the css ref from kill_css().
5005 static void css_killed_work_fn(struct work_struct
*work
)
5007 struct cgroup_subsys_state
*css
=
5008 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
5010 mutex_lock(&cgroup_mutex
);
5015 /* @css can't go away while we're holding cgroup_mutex */
5017 } while (css
&& atomic_dec_and_test(&css
->online_cnt
));
5019 mutex_unlock(&cgroup_mutex
);
5022 /* css kill confirmation processing requires process context, bounce */
5023 static void css_killed_ref_fn(struct percpu_ref
*ref
)
5025 struct cgroup_subsys_state
*css
=
5026 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
5028 if (atomic_dec_and_test(&css
->online_cnt
)) {
5029 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
5030 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
5035 * kill_css - destroy a css
5036 * @css: css to destroy
5038 * This function initiates destruction of @css by removing cgroup interface
5039 * files and putting its base reference. ->css_offline() will be invoked
5040 * asynchronously once css_tryget_online() is guaranteed to fail and when
5041 * the reference count reaches zero, @css will be released.
5043 static void kill_css(struct cgroup_subsys_state
*css
)
5045 lockdep_assert_held(&cgroup_mutex
);
5048 * This must happen before css is disassociated with its cgroup.
5049 * See seq_css() for details.
5051 css_clear_dir(css
, NULL
);
5054 * Killing would put the base ref, but we need to keep it alive
5055 * until after ->css_offline().
5060 * cgroup core guarantees that, by the time ->css_offline() is
5061 * invoked, no new css reference will be given out via
5062 * css_tryget_online(). We can't simply call percpu_ref_kill() and
5063 * proceed to offlining css's because percpu_ref_kill() doesn't
5064 * guarantee that the ref is seen as killed on all CPUs on return.
5066 * Use percpu_ref_kill_and_confirm() to get notifications as each
5067 * css is confirmed to be seen as killed on all CPUs.
5069 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
5073 * cgroup_destroy_locked - the first stage of cgroup destruction
5074 * @cgrp: cgroup to be destroyed
5076 * css's make use of percpu refcnts whose killing latency shouldn't be
5077 * exposed to userland and are RCU protected. Also, cgroup core needs to
5078 * guarantee that css_tryget_online() won't succeed by the time
5079 * ->css_offline() is invoked. To satisfy all the requirements,
5080 * destruction is implemented in the following two steps.
5082 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
5083 * userland visible parts and start killing the percpu refcnts of
5084 * css's. Set up so that the next stage will be kicked off once all
5085 * the percpu refcnts are confirmed to be killed.
5087 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
5088 * rest of destruction. Once all cgroup references are gone, the
5089 * cgroup is RCU-freed.
5091 * This function implements s1. After this step, @cgrp is gone as far as
5092 * the userland is concerned and a new cgroup with the same name may be
5093 * created. As cgroup doesn't care about the names internally, this
5094 * doesn't cause any problem.
5096 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
5097 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
5099 struct cgroup_subsys_state
*css
;
5102 lockdep_assert_held(&cgroup_mutex
);
5105 * Only migration can raise populated from zero and we're already
5106 * holding cgroup_mutex.
5108 if (cgroup_is_populated(cgrp
))
5112 * Make sure there's no live children. We can't test emptiness of
5113 * ->self.children as dead children linger on it while being
5114 * drained; otherwise, "rmdir parent/child parent" may fail.
5116 if (css_has_online_children(&cgrp
->self
))
5120 * Mark @cgrp dead. This prevents further task migration and child
5121 * creation by disabling cgroup_lock_live_group().
5123 cgrp
->self
.flags
&= ~CSS_ONLINE
;
5125 /* initiate massacre of all css's */
5126 for_each_css(css
, ssid
, cgrp
)
5130 * Remove @cgrp directory along with the base files. @cgrp has an
5131 * extra ref on its kn.
5133 kernfs_remove(cgrp
->kn
);
5135 check_for_release(cgroup_parent(cgrp
));
5137 /* put the base reference */
5138 percpu_ref_kill(&cgrp
->self
.refcnt
);
5143 static int cgroup_rmdir(struct kernfs_node
*kn
)
5145 struct cgroup
*cgrp
;
5148 cgrp
= cgroup_kn_lock_live(kn
);
5152 ret
= cgroup_destroy_locked(cgrp
);
5154 cgroup_kn_unlock(kn
);
5158 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
5159 .remount_fs
= cgroup_remount
,
5160 .show_options
= cgroup_show_options
,
5161 .mkdir
= cgroup_mkdir
,
5162 .rmdir
= cgroup_rmdir
,
5163 .rename
= cgroup_rename
,
5166 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
5168 struct cgroup_subsys_state
*css
;
5170 pr_debug("Initializing cgroup subsys %s\n", ss
->name
);
5172 mutex_lock(&cgroup_mutex
);
5174 idr_init(&ss
->css_idr
);
5175 INIT_LIST_HEAD(&ss
->cfts
);
5177 /* Create the root cgroup state for this subsystem */
5178 ss
->root
= &cgrp_dfl_root
;
5179 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
5180 /* We don't handle early failures gracefully */
5181 BUG_ON(IS_ERR(css
));
5182 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
5185 * Root csses are never destroyed and we can't initialize
5186 * percpu_ref during early init. Disable refcnting.
5188 css
->flags
|= CSS_NO_REF
;
5191 /* allocation can't be done safely during early init */
5194 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
5195 BUG_ON(css
->id
< 0);
5198 /* Update the init_css_set to contain a subsys
5199 * pointer to this state - since the subsystem is
5200 * newly registered, all tasks and hence the
5201 * init_css_set is in the subsystem's root cgroup. */
5202 init_css_set
.subsys
[ss
->id
] = css
;
5204 have_fork_callback
|= (bool)ss
->fork
<< ss
->id
;
5205 have_exit_callback
|= (bool)ss
->exit
<< ss
->id
;
5206 have_free_callback
|= (bool)ss
->free
<< ss
->id
;
5207 have_canfork_callback
|= (bool)ss
->can_fork
<< ss
->id
;
5209 /* At system boot, before all subsystems have been
5210 * registered, no tasks have been forked, so we don't
5211 * need to invoke fork callbacks here. */
5212 BUG_ON(!list_empty(&init_task
.tasks
));
5214 BUG_ON(online_css(css
));
5216 mutex_unlock(&cgroup_mutex
);
5220 * cgroup_init_early - cgroup initialization at system boot
5222 * Initialize cgroups at system boot, and initialize any
5223 * subsystems that request early init.
5225 int __init
cgroup_init_early(void)
5227 static struct cgroup_sb_opts __initdata opts
;
5228 struct cgroup_subsys
*ss
;
5231 init_cgroup_root(&cgrp_dfl_root
, &opts
);
5232 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
5234 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
5236 for_each_subsys(ss
, i
) {
5237 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
5238 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5239 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
5241 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
5242 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
5245 ss
->name
= cgroup_subsys_name
[i
];
5246 if (!ss
->legacy_name
)
5247 ss
->legacy_name
= cgroup_subsys_name
[i
];
5250 cgroup_init_subsys(ss
, true);
5255 static u16 cgroup_disable_mask __initdata
;
5258 * cgroup_init - cgroup initialization
5260 * Register cgroup filesystem and /proc file, and initialize
5261 * any subsystems that didn't request early init.
5263 int __init
cgroup_init(void)
5265 struct cgroup_subsys
*ss
;
5269 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT
> 16);
5270 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem
));
5271 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_dfl_base_files
));
5272 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_legacy_base_files
));
5274 mutex_lock(&cgroup_mutex
);
5276 /* Add init_css_set to the hash table */
5277 key
= css_set_hash(init_css_set
.subsys
);
5278 hash_add(css_set_table
, &init_css_set
.hlist
, key
);
5280 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
5282 mutex_unlock(&cgroup_mutex
);
5284 for_each_subsys(ss
, ssid
) {
5285 if (ss
->early_init
) {
5286 struct cgroup_subsys_state
*css
=
5287 init_css_set
.subsys
[ss
->id
];
5289 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
5291 BUG_ON(css
->id
< 0);
5293 cgroup_init_subsys(ss
, false);
5296 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
5297 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
5300 * Setting dfl_root subsys_mask needs to consider the
5301 * disabled flag and cftype registration needs kmalloc,
5302 * both of which aren't available during early_init.
5304 if (cgroup_disable_mask
& (1 << ssid
)) {
5305 static_branch_disable(cgroup_subsys_enabled_key
[ssid
]);
5306 printk(KERN_INFO
"Disabling %s control group subsystem\n",
5311 if (cgroup_ssid_no_v1(ssid
))
5312 printk(KERN_INFO
"Disabling %s control group subsystem in v1 mounts\n",
5315 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
5317 if (!ss
->dfl_cftypes
)
5318 cgrp_dfl_inhibit_ss_mask
|= 1 << ss
->id
;
5320 if (ss
->dfl_cftypes
== ss
->legacy_cftypes
) {
5321 WARN_ON(cgroup_add_cftypes(ss
, ss
->dfl_cftypes
));
5323 WARN_ON(cgroup_add_dfl_cftypes(ss
, ss
->dfl_cftypes
));
5324 WARN_ON(cgroup_add_legacy_cftypes(ss
, ss
->legacy_cftypes
));
5328 ss
->bind(init_css_set
.subsys
[ssid
]);
5331 WARN_ON(sysfs_create_mount_point(fs_kobj
, "cgroup"));
5332 WARN_ON(register_filesystem(&cgroup_fs_type
));
5333 WARN_ON(register_filesystem(&cgroup2_fs_type
));
5334 WARN_ON(!proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
));
5339 static int __init
cgroup_wq_init(void)
5342 * There isn't much point in executing destruction path in
5343 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5344 * Use 1 for @max_active.
5346 * We would prefer to do this in cgroup_init() above, but that
5347 * is called before init_workqueues(): so leave this until after.
5349 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
5350 BUG_ON(!cgroup_destroy_wq
);
5353 * Used to destroy pidlists and separate to serve as flush domain.
5354 * Cap @max_active to 1 too.
5356 cgroup_pidlist_destroy_wq
= alloc_workqueue("cgroup_pidlist_destroy",
5358 BUG_ON(!cgroup_pidlist_destroy_wq
);
5362 core_initcall(cgroup_wq_init
);
5365 * proc_cgroup_show()
5366 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5367 * - Used for /proc/<pid>/cgroup.
5369 int proc_cgroup_show(struct seq_file
*m
, struct pid_namespace
*ns
,
5370 struct pid
*pid
, struct task_struct
*tsk
)
5374 struct cgroup_root
*root
;
5377 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5381 mutex_lock(&cgroup_mutex
);
5382 spin_lock_bh(&css_set_lock
);
5384 for_each_root(root
) {
5385 struct cgroup_subsys
*ss
;
5386 struct cgroup
*cgrp
;
5387 int ssid
, count
= 0;
5389 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_visible
)
5392 seq_printf(m
, "%d:", root
->hierarchy_id
);
5393 if (root
!= &cgrp_dfl_root
)
5394 for_each_subsys(ss
, ssid
)
5395 if (root
->subsys_mask
& (1 << ssid
))
5396 seq_printf(m
, "%s%s", count
++ ? "," : "",
5398 if (strlen(root
->name
))
5399 seq_printf(m
, "%sname=%s", count
? "," : "",
5403 cgrp
= task_cgroup_from_root(tsk
, root
);
5406 * On traditional hierarchies, all zombie tasks show up as
5407 * belonging to the root cgroup. On the default hierarchy,
5408 * while a zombie doesn't show up in "cgroup.procs" and
5409 * thus can't be migrated, its /proc/PID/cgroup keeps
5410 * reporting the cgroup it belonged to before exiting. If
5411 * the cgroup is removed before the zombie is reaped,
5412 * " (deleted)" is appended to the cgroup path.
5414 if (cgroup_on_dfl(cgrp
) || !(tsk
->flags
& PF_EXITING
)) {
5415 path
= cgroup_path(cgrp
, buf
, PATH_MAX
);
5417 retval
= -ENAMETOOLONG
;
5426 if (cgroup_on_dfl(cgrp
) && cgroup_is_dead(cgrp
))
5427 seq_puts(m
, " (deleted)\n");
5434 spin_unlock_bh(&css_set_lock
);
5435 mutex_unlock(&cgroup_mutex
);
5441 /* Display information about each subsystem and each hierarchy */
5442 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
5444 struct cgroup_subsys
*ss
;
5447 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5449 * ideally we don't want subsystems moving around while we do this.
5450 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5451 * subsys/hierarchy state.
5453 mutex_lock(&cgroup_mutex
);
5455 for_each_subsys(ss
, i
)
5456 seq_printf(m
, "%s\t%d\t%d\t%d\n",
5457 ss
->legacy_name
, ss
->root
->hierarchy_id
,
5458 atomic_read(&ss
->root
->nr_cgrps
),
5459 cgroup_ssid_enabled(i
));
5461 mutex_unlock(&cgroup_mutex
);
5465 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
5467 return single_open(file
, proc_cgroupstats_show
, NULL
);
5470 static const struct file_operations proc_cgroupstats_operations
= {
5471 .open
= cgroupstats_open
,
5473 .llseek
= seq_lseek
,
5474 .release
= single_release
,
5478 * cgroup_fork - initialize cgroup related fields during copy_process()
5479 * @child: pointer to task_struct of forking parent process.
5481 * A task is associated with the init_css_set until cgroup_post_fork()
5482 * attaches it to the parent's css_set. Empty cg_list indicates that
5483 * @child isn't holding reference to its css_set.
5485 void cgroup_fork(struct task_struct
*child
)
5487 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
5488 INIT_LIST_HEAD(&child
->cg_list
);
5492 * cgroup_can_fork - called on a new task before the process is exposed
5493 * @child: the task in question.
5495 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5496 * returns an error, the fork aborts with that error code. This allows for
5497 * a cgroup subsystem to conditionally allow or deny new forks.
5499 int cgroup_can_fork(struct task_struct
*child
)
5501 struct cgroup_subsys
*ss
;
5504 do_each_subsys_mask(ss
, i
, have_canfork_callback
) {
5505 ret
= ss
->can_fork(child
);
5508 } while_each_subsys_mask();
5513 for_each_subsys(ss
, j
) {
5516 if (ss
->cancel_fork
)
5517 ss
->cancel_fork(child
);
5524 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5525 * @child: the task in question
5527 * This calls the cancel_fork() callbacks if a fork failed *after*
5528 * cgroup_can_fork() succeded.
5530 void cgroup_cancel_fork(struct task_struct
*child
)
5532 struct cgroup_subsys
*ss
;
5535 for_each_subsys(ss
, i
)
5536 if (ss
->cancel_fork
)
5537 ss
->cancel_fork(child
);
5541 * cgroup_post_fork - called on a new task after adding it to the task list
5542 * @child: the task in question
5544 * Adds the task to the list running through its css_set if necessary and
5545 * call the subsystem fork() callbacks. Has to be after the task is
5546 * visible on the task list in case we race with the first call to
5547 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5550 void cgroup_post_fork(struct task_struct
*child
)
5552 struct cgroup_subsys
*ss
;
5556 * This may race against cgroup_enable_task_cg_lists(). As that
5557 * function sets use_task_css_set_links before grabbing
5558 * tasklist_lock and we just went through tasklist_lock to add
5559 * @child, it's guaranteed that either we see the set
5560 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5561 * @child during its iteration.
5563 * If we won the race, @child is associated with %current's
5564 * css_set. Grabbing css_set_lock guarantees both that the
5565 * association is stable, and, on completion of the parent's
5566 * migration, @child is visible in the source of migration or
5567 * already in the destination cgroup. This guarantee is necessary
5568 * when implementing operations which need to migrate all tasks of
5569 * a cgroup to another.
5571 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5572 * will remain in init_css_set. This is safe because all tasks are
5573 * in the init_css_set before cg_links is enabled and there's no
5574 * operation which transfers all tasks out of init_css_set.
5576 if (use_task_css_set_links
) {
5577 struct css_set
*cset
;
5579 spin_lock_bh(&css_set_lock
);
5580 cset
= task_css_set(current
);
5581 if (list_empty(&child
->cg_list
)) {
5583 css_set_move_task(child
, NULL
, cset
, false);
5585 spin_unlock_bh(&css_set_lock
);
5589 * Call ss->fork(). This must happen after @child is linked on
5590 * css_set; otherwise, @child might change state between ->fork()
5591 * and addition to css_set.
5593 do_each_subsys_mask(ss
, i
, have_fork_callback
) {
5595 } while_each_subsys_mask();
5599 * cgroup_exit - detach cgroup from exiting task
5600 * @tsk: pointer to task_struct of exiting process
5602 * Description: Detach cgroup from @tsk and release it.
5604 * Note that cgroups marked notify_on_release force every task in
5605 * them to take the global cgroup_mutex mutex when exiting.
5606 * This could impact scaling on very large systems. Be reluctant to
5607 * use notify_on_release cgroups where very high task exit scaling
5608 * is required on large systems.
5610 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5611 * call cgroup_exit() while the task is still competent to handle
5612 * notify_on_release(), then leave the task attached to the root cgroup in
5613 * each hierarchy for the remainder of its exit. No need to bother with
5614 * init_css_set refcnting. init_css_set never goes away and we can't race
5615 * with migration path - PF_EXITING is visible to migration path.
5617 void cgroup_exit(struct task_struct
*tsk
)
5619 struct cgroup_subsys
*ss
;
5620 struct css_set
*cset
;
5624 * Unlink from @tsk from its css_set. As migration path can't race
5625 * with us, we can check css_set and cg_list without synchronization.
5627 cset
= task_css_set(tsk
);
5629 if (!list_empty(&tsk
->cg_list
)) {
5630 spin_lock_bh(&css_set_lock
);
5631 css_set_move_task(tsk
, cset
, NULL
, false);
5632 spin_unlock_bh(&css_set_lock
);
5637 /* see cgroup_post_fork() for details */
5638 do_each_subsys_mask(ss
, i
, have_exit_callback
) {
5640 } while_each_subsys_mask();
5643 void cgroup_free(struct task_struct
*task
)
5645 struct css_set
*cset
= task_css_set(task
);
5646 struct cgroup_subsys
*ss
;
5649 do_each_subsys_mask(ss
, ssid
, have_free_callback
) {
5651 } while_each_subsys_mask();
5656 static void check_for_release(struct cgroup
*cgrp
)
5658 if (notify_on_release(cgrp
) && !cgroup_is_populated(cgrp
) &&
5659 !css_has_online_children(&cgrp
->self
) && !cgroup_is_dead(cgrp
))
5660 schedule_work(&cgrp
->release_agent_work
);
5664 * Notify userspace when a cgroup is released, by running the
5665 * configured release agent with the name of the cgroup (path
5666 * relative to the root of cgroup file system) as the argument.
5668 * Most likely, this user command will try to rmdir this cgroup.
5670 * This races with the possibility that some other task will be
5671 * attached to this cgroup before it is removed, or that some other
5672 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5673 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5674 * unused, and this cgroup will be reprieved from its death sentence,
5675 * to continue to serve a useful existence. Next time it's released,
5676 * we will get notified again, if it still has 'notify_on_release' set.
5678 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5679 * means only wait until the task is successfully execve()'d. The
5680 * separate release agent task is forked by call_usermodehelper(),
5681 * then control in this thread returns here, without waiting for the
5682 * release agent task. We don't bother to wait because the caller of
5683 * this routine has no use for the exit status of the release agent
5684 * task, so no sense holding our caller up for that.
5686 static void cgroup_release_agent(struct work_struct
*work
)
5688 struct cgroup
*cgrp
=
5689 container_of(work
, struct cgroup
, release_agent_work
);
5690 char *pathbuf
= NULL
, *agentbuf
= NULL
, *path
;
5691 char *argv
[3], *envp
[3];
5693 mutex_lock(&cgroup_mutex
);
5695 pathbuf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5696 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
5697 if (!pathbuf
|| !agentbuf
)
5700 path
= cgroup_path(cgrp
, pathbuf
, PATH_MAX
);
5708 /* minimal command environment */
5710 envp
[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5713 mutex_unlock(&cgroup_mutex
);
5714 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
5717 mutex_unlock(&cgroup_mutex
);
5723 static int __init
cgroup_disable(char *str
)
5725 struct cgroup_subsys
*ss
;
5729 while ((token
= strsep(&str
, ",")) != NULL
) {
5733 for_each_subsys(ss
, i
) {
5734 if (strcmp(token
, ss
->name
) &&
5735 strcmp(token
, ss
->legacy_name
))
5737 cgroup_disable_mask
|= 1 << i
;
5742 __setup("cgroup_disable=", cgroup_disable
);
5744 static int __init
cgroup_no_v1(char *str
)
5746 struct cgroup_subsys
*ss
;
5750 while ((token
= strsep(&str
, ",")) != NULL
) {
5754 if (!strcmp(token
, "all")) {
5755 cgroup_no_v1_mask
= U16_MAX
;
5759 for_each_subsys(ss
, i
) {
5760 if (strcmp(token
, ss
->name
) &&
5761 strcmp(token
, ss
->legacy_name
))
5764 cgroup_no_v1_mask
|= 1 << i
;
5769 __setup("cgroup_no_v1=", cgroup_no_v1
);
5772 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5773 * @dentry: directory dentry of interest
5774 * @ss: subsystem of interest
5776 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5777 * to get the corresponding css and return it. If such css doesn't exist
5778 * or can't be pinned, an ERR_PTR value is returned.
5780 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
5781 struct cgroup_subsys
*ss
)
5783 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5784 struct file_system_type
*s_type
= dentry
->d_sb
->s_type
;
5785 struct cgroup_subsys_state
*css
= NULL
;
5786 struct cgroup
*cgrp
;
5788 /* is @dentry a cgroup dir? */
5789 if ((s_type
!= &cgroup_fs_type
&& s_type
!= &cgroup2_fs_type
) ||
5790 !kn
|| kernfs_type(kn
) != KERNFS_DIR
)
5791 return ERR_PTR(-EBADF
);
5796 * This path doesn't originate from kernfs and @kn could already
5797 * have been or be removed at any point. @kn->priv is RCU
5798 * protected for this access. See css_release_work_fn() for details.
5800 cgrp
= rcu_dereference(kn
->priv
);
5802 css
= cgroup_css(cgrp
, ss
);
5804 if (!css
|| !css_tryget_online(css
))
5805 css
= ERR_PTR(-ENOENT
);
5812 * css_from_id - lookup css by id
5813 * @id: the cgroup id
5814 * @ss: cgroup subsys to be looked into
5816 * Returns the css if there's valid one with @id, otherwise returns NULL.
5817 * Should be called under rcu_read_lock().
5819 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5821 WARN_ON_ONCE(!rcu_read_lock_held());
5822 return id
> 0 ? idr_find(&ss
->css_idr
, id
) : NULL
;
5826 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5827 * @path: path on the default hierarchy
5829 * Find the cgroup at @path on the default hierarchy, increment its
5830 * reference count and return it. Returns pointer to the found cgroup on
5831 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5832 * if @path points to a non-directory.
5834 struct cgroup
*cgroup_get_from_path(const char *path
)
5836 struct kernfs_node
*kn
;
5837 struct cgroup
*cgrp
;
5839 mutex_lock(&cgroup_mutex
);
5841 kn
= kernfs_walk_and_get(cgrp_dfl_root
.cgrp
.kn
, path
);
5843 if (kernfs_type(kn
) == KERNFS_DIR
) {
5847 cgrp
= ERR_PTR(-ENOTDIR
);
5851 cgrp
= ERR_PTR(-ENOENT
);
5854 mutex_unlock(&cgroup_mutex
);
5857 EXPORT_SYMBOL_GPL(cgroup_get_from_path
);
5860 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data
5861 * definition in cgroup-defs.h.
5863 #ifdef CONFIG_SOCK_CGROUP_DATA
5865 #if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5867 DEFINE_SPINLOCK(cgroup_sk_update_lock
);
5868 static bool cgroup_sk_alloc_disabled __read_mostly
;
5870 void cgroup_sk_alloc_disable(void)
5872 if (cgroup_sk_alloc_disabled
)
5874 pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5875 cgroup_sk_alloc_disabled
= true;
5880 #define cgroup_sk_alloc_disabled false
5884 void cgroup_sk_alloc(struct sock_cgroup_data
*skcd
)
5886 if (cgroup_sk_alloc_disabled
)
5892 struct css_set
*cset
;
5894 cset
= task_css_set(current
);
5895 if (likely(cgroup_tryget(cset
->dfl_cgrp
))) {
5896 skcd
->val
= (unsigned long)cset
->dfl_cgrp
;
5905 void cgroup_sk_free(struct sock_cgroup_data
*skcd
)
5907 cgroup_put(sock_cgroup_ptr(skcd
));
5910 #endif /* CONFIG_SOCK_CGROUP_DATA */
5912 #ifdef CONFIG_CGROUP_DEBUG
5913 static struct cgroup_subsys_state
*
5914 debug_css_alloc(struct cgroup_subsys_state
*parent_css
)
5916 struct cgroup_subsys_state
*css
= kzalloc(sizeof(*css
), GFP_KERNEL
);
5919 return ERR_PTR(-ENOMEM
);
5924 static void debug_css_free(struct cgroup_subsys_state
*css
)
5929 static u64
debug_taskcount_read(struct cgroup_subsys_state
*css
,
5932 return cgroup_task_count(css
->cgroup
);
5935 static u64
current_css_set_read(struct cgroup_subsys_state
*css
,
5938 return (u64
)(unsigned long)current
->cgroups
;
5941 static u64
current_css_set_refcount_read(struct cgroup_subsys_state
*css
,
5947 count
= atomic_read(&task_css_set(current
)->refcount
);
5952 static int current_css_set_cg_links_read(struct seq_file
*seq
, void *v
)
5954 struct cgrp_cset_link
*link
;
5955 struct css_set
*cset
;
5958 name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
5962 spin_lock_bh(&css_set_lock
);
5964 cset
= rcu_dereference(current
->cgroups
);
5965 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
5966 struct cgroup
*c
= link
->cgrp
;
5968 cgroup_name(c
, name_buf
, NAME_MAX
+ 1);
5969 seq_printf(seq
, "Root %d group %s\n",
5970 c
->root
->hierarchy_id
, name_buf
);
5973 spin_unlock_bh(&css_set_lock
);
5978 #define MAX_TASKS_SHOWN_PER_CSS 25
5979 static int cgroup_css_links_read(struct seq_file
*seq
, void *v
)
5981 struct cgroup_subsys_state
*css
= seq_css(seq
);
5982 struct cgrp_cset_link
*link
;
5984 spin_lock_bh(&css_set_lock
);
5985 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
) {
5986 struct css_set
*cset
= link
->cset
;
5987 struct task_struct
*task
;
5990 seq_printf(seq
, "css_set %p\n", cset
);
5992 list_for_each_entry(task
, &cset
->tasks
, cg_list
) {
5993 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5995 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5998 list_for_each_entry(task
, &cset
->mg_tasks
, cg_list
) {
5999 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
6001 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
6005 seq_puts(seq
, " ...\n");
6007 spin_unlock_bh(&css_set_lock
);
6011 static u64
releasable_read(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
6013 return (!cgroup_is_populated(css
->cgroup
) &&
6014 !css_has_online_children(&css
->cgroup
->self
));
6017 static struct cftype debug_files
[] = {
6019 .name
= "taskcount",
6020 .read_u64
= debug_taskcount_read
,
6024 .name
= "current_css_set",
6025 .read_u64
= current_css_set_read
,
6029 .name
= "current_css_set_refcount",
6030 .read_u64
= current_css_set_refcount_read
,
6034 .name
= "current_css_set_cg_links",
6035 .seq_show
= current_css_set_cg_links_read
,
6039 .name
= "cgroup_css_links",
6040 .seq_show
= cgroup_css_links_read
,
6044 .name
= "releasable",
6045 .read_u64
= releasable_read
,
6051 struct cgroup_subsys debug_cgrp_subsys
= {
6052 .css_alloc
= debug_css_alloc
,
6053 .css_free
= debug_css_free
,
6054 .legacy_cftypes
= debug_files
,
6056 #endif /* CONFIG_CGROUP_DEBUG */