2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/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>
61 #include <linux/atomic.h>
64 * pidlists linger the following amount before being destroyed. The goal
65 * is avoiding frequent destruction in the middle of consecutive read calls
66 * Expiring in the middle is a performance problem not a correctness one.
67 * 1 sec should be enough.
69 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
71 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
75 * cgroup_mutex is the master lock. Any modification to cgroup or its
76 * hierarchy must be performed while holding it.
78 * css_set_rwsem protects task->cgroups pointer, the list of css_set
79 * objects, and the chain of tasks off each css_set.
81 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
82 * cgroup.h can use them for lockdep annotations.
84 #ifdef CONFIG_PROVE_RCU
85 DEFINE_MUTEX(cgroup_mutex
);
86 DECLARE_RWSEM(css_set_rwsem
);
87 EXPORT_SYMBOL_GPL(cgroup_mutex
);
88 EXPORT_SYMBOL_GPL(css_set_rwsem
);
90 static DEFINE_MUTEX(cgroup_mutex
);
91 static DECLARE_RWSEM(css_set_rwsem
);
95 * Protects cgroup_idr and css_idr so that IDs can be released without
96 * grabbing cgroup_mutex.
98 static DEFINE_SPINLOCK(cgroup_idr_lock
);
101 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
102 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104 static DEFINE_SPINLOCK(release_agent_path_lock
);
106 #define cgroup_assert_mutex_or_rcu_locked() \
107 rcu_lockdep_assert(rcu_read_lock_held() || \
108 lockdep_is_held(&cgroup_mutex), \
109 "cgroup_mutex or RCU read lock required");
112 * cgroup destruction makes heavy use of work items and there can be a lot
113 * of concurrent destructions. Use a separate workqueue so that cgroup
114 * destruction work items don't end up filling up max_active of system_wq
115 * which may lead to deadlock.
117 static struct workqueue_struct
*cgroup_destroy_wq
;
120 * pidlist destructions need to be flushed on cgroup destruction. Use a
121 * separate workqueue as flush domain.
123 static struct workqueue_struct
*cgroup_pidlist_destroy_wq
;
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys
*cgroup_subsys
[] = {
128 #include <linux/cgroup_subsys.h>
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name
[] = {
135 #include <linux/cgroup_subsys.h>
140 * The default hierarchy, reserved for the subsystems that are otherwise
141 * unattached - it never has more than a single cgroup, and all tasks are
142 * part of that cgroup.
144 struct cgroup_root cgrp_dfl_root
;
147 * The default hierarchy always exists but is hidden until mounted for the
148 * first time. This is for backward compatibility.
150 static bool cgrp_dfl_root_visible
;
152 /* some controllers are not supported in the default hierarchy */
153 static const unsigned int cgrp_dfl_root_inhibit_ss_mask
= 0
154 #ifdef CONFIG_CGROUP_DEBUG
155 | (1 << debug_cgrp_id
)
159 /* The list of hierarchy roots */
161 static LIST_HEAD(cgroup_roots
);
162 static int cgroup_root_count
;
164 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
165 static DEFINE_IDR(cgroup_hierarchy_idr
);
168 * Assign a monotonically increasing serial number to csses. It guarantees
169 * cgroups with bigger numbers are newer than those with smaller numbers.
170 * Also, as csses are always appended to the parent's ->children list, it
171 * guarantees that sibling csses are always sorted in the ascending serial
172 * number order on the list. Protected by cgroup_mutex.
174 static u64 css_serial_nr_next
= 1;
176 /* This flag indicates whether tasks in the fork and exit paths should
177 * check for fork/exit handlers to call. This avoids us having to do
178 * extra work in the fork/exit path if none of the subsystems need to
181 static int need_forkexit_callback __read_mostly
;
183 static struct cftype cgroup_base_files
[];
185 static void cgroup_put(struct cgroup
*cgrp
);
186 static int rebind_subsystems(struct cgroup_root
*dst_root
,
187 unsigned int ss_mask
);
188 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
189 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
);
190 static void css_release(struct percpu_ref
*ref
);
191 static void kill_css(struct cgroup_subsys_state
*css
);
192 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
194 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
);
196 /* IDR wrappers which synchronize using cgroup_idr_lock */
197 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
202 idr_preload(gfp_mask
);
203 spin_lock_bh(&cgroup_idr_lock
);
204 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
);
205 spin_unlock_bh(&cgroup_idr_lock
);
210 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
214 spin_lock_bh(&cgroup_idr_lock
);
215 ret
= idr_replace(idr
, ptr
, id
);
216 spin_unlock_bh(&cgroup_idr_lock
);
220 static void cgroup_idr_remove(struct idr
*idr
, int id
)
222 spin_lock_bh(&cgroup_idr_lock
);
224 spin_unlock_bh(&cgroup_idr_lock
);
227 static struct cgroup
*cgroup_parent(struct cgroup
*cgrp
)
229 struct cgroup_subsys_state
*parent_css
= cgrp
->self
.parent
;
232 return container_of(parent_css
, struct cgroup
, self
);
237 * cgroup_css - obtain a cgroup's css for the specified subsystem
238 * @cgrp: the cgroup of interest
239 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
241 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
242 * function must be called either under cgroup_mutex or rcu_read_lock() and
243 * the caller is responsible for pinning the returned css if it wants to
244 * keep accessing it outside the said locks. This function may return
245 * %NULL if @cgrp doesn't have @subsys_id enabled.
247 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
248 struct cgroup_subsys
*ss
)
251 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
252 lockdep_is_held(&cgroup_mutex
));
258 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
259 * @cgrp: the cgroup of interest
260 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
262 * Similar to cgroup_css() but returns the effctive css, which is defined
263 * as the matching css of the nearest ancestor including self which has @ss
264 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
265 * function is guaranteed to return non-NULL css.
267 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
268 struct cgroup_subsys
*ss
)
270 lockdep_assert_held(&cgroup_mutex
);
275 if (!(cgrp
->root
->subsys_mask
& (1 << ss
->id
)))
278 while (cgroup_parent(cgrp
) &&
279 !(cgroup_parent(cgrp
)->child_subsys_mask
& (1 << ss
->id
)))
280 cgrp
= cgroup_parent(cgrp
);
282 return cgroup_css(cgrp
, ss
);
285 /* convenient tests for these bits */
286 static inline bool cgroup_is_dead(const struct cgroup
*cgrp
)
288 return !(cgrp
->self
.flags
& CSS_ONLINE
);
291 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
293 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
294 struct cftype
*cft
= of_cft(of
);
297 * This is open and unprotected implementation of cgroup_css().
298 * seq_css() is only called from a kernfs file operation which has
299 * an active reference on the file. Because all the subsystem
300 * files are drained before a css is disassociated with a cgroup,
301 * the matching css from the cgroup's subsys table is guaranteed to
302 * be and stay valid until the enclosing operation is complete.
305 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
309 EXPORT_SYMBOL_GPL(of_css
);
312 * cgroup_is_descendant - test ancestry
313 * @cgrp: the cgroup to be tested
314 * @ancestor: possible ancestor of @cgrp
316 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
317 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
318 * and @ancestor are accessible.
320 bool cgroup_is_descendant(struct cgroup
*cgrp
, struct cgroup
*ancestor
)
323 if (cgrp
== ancestor
)
325 cgrp
= cgroup_parent(cgrp
);
330 static int cgroup_is_releasable(const struct cgroup
*cgrp
)
333 (1 << CGRP_RELEASABLE
) |
334 (1 << CGRP_NOTIFY_ON_RELEASE
);
335 return (cgrp
->flags
& bits
) == bits
;
338 static int notify_on_release(const struct cgroup
*cgrp
)
340 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
344 * for_each_css - iterate all css's of a cgroup
345 * @css: the iteration cursor
346 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
347 * @cgrp: the target cgroup to iterate css's of
349 * Should be called under cgroup_[tree_]mutex.
351 #define for_each_css(css, ssid, cgrp) \
352 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
353 if (!((css) = rcu_dereference_check( \
354 (cgrp)->subsys[(ssid)], \
355 lockdep_is_held(&cgroup_mutex)))) { } \
359 * for_each_e_css - iterate all effective css's of a cgroup
360 * @css: the iteration cursor
361 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
362 * @cgrp: the target cgroup to iterate css's of
364 * Should be called under cgroup_[tree_]mutex.
366 #define for_each_e_css(css, ssid, cgrp) \
367 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
368 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
373 * for_each_subsys - iterate all enabled cgroup subsystems
374 * @ss: the iteration cursor
375 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
377 #define for_each_subsys(ss, ssid) \
378 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
379 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
381 /* iterate across the hierarchies */
382 #define for_each_root(root) \
383 list_for_each_entry((root), &cgroup_roots, root_list)
385 /* iterate over child cgrps, lock should be held throughout iteration */
386 #define cgroup_for_each_live_child(child, cgrp) \
387 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
388 if (({ lockdep_assert_held(&cgroup_mutex); \
389 cgroup_is_dead(child); })) \
393 /* the list of cgroups eligible for automatic release. Protected by
394 * release_list_lock */
395 static LIST_HEAD(release_list
);
396 static DEFINE_RAW_SPINLOCK(release_list_lock
);
397 static void cgroup_release_agent(struct work_struct
*work
);
398 static DECLARE_WORK(release_agent_work
, cgroup_release_agent
);
399 static void check_for_release(struct cgroup
*cgrp
);
402 * A cgroup can be associated with multiple css_sets as different tasks may
403 * belong to different cgroups on different hierarchies. In the other
404 * direction, a css_set is naturally associated with multiple cgroups.
405 * This M:N relationship is represented by the following link structure
406 * which exists for each association and allows traversing the associations
409 struct cgrp_cset_link
{
410 /* the cgroup and css_set this link associates */
412 struct css_set
*cset
;
414 /* list of cgrp_cset_links anchored at cgrp->cset_links */
415 struct list_head cset_link
;
417 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
418 struct list_head cgrp_link
;
422 * The default css_set - used by init and its children prior to any
423 * hierarchies being mounted. It contains a pointer to the root state
424 * for each subsystem. Also used to anchor the list of css_sets. Not
425 * reference-counted, to improve performance when child cgroups
426 * haven't been created.
428 struct css_set init_css_set
= {
429 .refcount
= ATOMIC_INIT(1),
430 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
431 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
432 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
433 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
434 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
437 static int css_set_count
= 1; /* 1 for init_css_set */
440 * cgroup_update_populated - updated populated count of a cgroup
441 * @cgrp: the target cgroup
442 * @populated: inc or dec populated count
444 * @cgrp is either getting the first task (css_set) or losing the last.
445 * Update @cgrp->populated_cnt accordingly. The count is propagated
446 * towards root so that a given cgroup's populated_cnt is zero iff the
447 * cgroup and all its descendants are empty.
449 * @cgrp's interface file "cgroup.populated" is zero if
450 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
451 * changes from or to zero, userland is notified that the content of the
452 * interface file has changed. This can be used to detect when @cgrp and
453 * its descendants become populated or empty.
455 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
457 lockdep_assert_held(&css_set_rwsem
);
463 trigger
= !cgrp
->populated_cnt
++;
465 trigger
= !--cgrp
->populated_cnt
;
470 if (cgrp
->populated_kn
)
471 kernfs_notify(cgrp
->populated_kn
);
472 cgrp
= cgroup_parent(cgrp
);
477 * hash table for cgroup groups. This improves the performance to find
478 * an existing css_set. This hash doesn't (currently) take into
479 * account cgroups in empty hierarchies.
481 #define CSS_SET_HASH_BITS 7
482 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
484 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
486 unsigned long key
= 0UL;
487 struct cgroup_subsys
*ss
;
490 for_each_subsys(ss
, i
)
491 key
+= (unsigned long)css
[i
];
492 key
= (key
>> 16) ^ key
;
497 static void put_css_set_locked(struct css_set
*cset
, bool taskexit
)
499 struct cgrp_cset_link
*link
, *tmp_link
;
500 struct cgroup_subsys
*ss
;
503 lockdep_assert_held(&css_set_rwsem
);
505 if (!atomic_dec_and_test(&cset
->refcount
))
508 /* This css_set is dead. unlink it and release cgroup refcounts */
509 for_each_subsys(ss
, ssid
)
510 list_del(&cset
->e_cset_node
[ssid
]);
511 hash_del(&cset
->hlist
);
514 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
515 struct cgroup
*cgrp
= link
->cgrp
;
517 list_del(&link
->cset_link
);
518 list_del(&link
->cgrp_link
);
520 /* @cgrp can't go away while we're holding css_set_rwsem */
521 if (list_empty(&cgrp
->cset_links
)) {
522 cgroup_update_populated(cgrp
, false);
523 if (notify_on_release(cgrp
)) {
525 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
526 check_for_release(cgrp
);
533 kfree_rcu(cset
, rcu_head
);
536 static void put_css_set(struct css_set
*cset
, bool taskexit
)
539 * Ensure that the refcount doesn't hit zero while any readers
540 * can see it. Similar to atomic_dec_and_lock(), but for an
543 if (atomic_add_unless(&cset
->refcount
, -1, 1))
546 down_write(&css_set_rwsem
);
547 put_css_set_locked(cset
, taskexit
);
548 up_write(&css_set_rwsem
);
552 * refcounted get/put for css_set objects
554 static inline void get_css_set(struct css_set
*cset
)
556 atomic_inc(&cset
->refcount
);
560 * compare_css_sets - helper function for find_existing_css_set().
561 * @cset: candidate css_set being tested
562 * @old_cset: existing css_set for a task
563 * @new_cgrp: cgroup that's being entered by the task
564 * @template: desired set of css pointers in css_set (pre-calculated)
566 * Returns true if "cset" matches "old_cset" except for the hierarchy
567 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
569 static bool compare_css_sets(struct css_set
*cset
,
570 struct css_set
*old_cset
,
571 struct cgroup
*new_cgrp
,
572 struct cgroup_subsys_state
*template[])
574 struct list_head
*l1
, *l2
;
577 * On the default hierarchy, there can be csets which are
578 * associated with the same set of cgroups but different csses.
579 * Let's first ensure that csses match.
581 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
585 * Compare cgroup pointers in order to distinguish between
586 * different cgroups in hierarchies. As different cgroups may
587 * share the same effective css, this comparison is always
590 l1
= &cset
->cgrp_links
;
591 l2
= &old_cset
->cgrp_links
;
593 struct cgrp_cset_link
*link1
, *link2
;
594 struct cgroup
*cgrp1
, *cgrp2
;
598 /* See if we reached the end - both lists are equal length. */
599 if (l1
== &cset
->cgrp_links
) {
600 BUG_ON(l2
!= &old_cset
->cgrp_links
);
603 BUG_ON(l2
== &old_cset
->cgrp_links
);
605 /* Locate the cgroups associated with these links. */
606 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
607 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
610 /* Hierarchies should be linked in the same order. */
611 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
614 * If this hierarchy is the hierarchy of the cgroup
615 * that's changing, then we need to check that this
616 * css_set points to the new cgroup; if it's any other
617 * hierarchy, then this css_set should point to the
618 * same cgroup as the old css_set.
620 if (cgrp1
->root
== new_cgrp
->root
) {
621 if (cgrp1
!= new_cgrp
)
632 * find_existing_css_set - init css array and find the matching css_set
633 * @old_cset: the css_set that we're using before the cgroup transition
634 * @cgrp: the cgroup that we're moving into
635 * @template: out param for the new set of csses, should be clear on entry
637 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
639 struct cgroup_subsys_state
*template[])
641 struct cgroup_root
*root
= cgrp
->root
;
642 struct cgroup_subsys
*ss
;
643 struct css_set
*cset
;
648 * Build the set of subsystem state objects that we want to see in the
649 * new css_set. while subsystems can change globally, the entries here
650 * won't change, so no need for locking.
652 for_each_subsys(ss
, i
) {
653 if (root
->subsys_mask
& (1UL << i
)) {
655 * @ss is in this hierarchy, so we want the
656 * effective css from @cgrp.
658 template[i
] = cgroup_e_css(cgrp
, ss
);
661 * @ss is not in this hierarchy, so we don't want
664 template[i
] = old_cset
->subsys
[i
];
668 key
= css_set_hash(template);
669 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
670 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
673 /* This css_set matches what we need */
677 /* No existing cgroup group matched */
681 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
683 struct cgrp_cset_link
*link
, *tmp_link
;
685 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
686 list_del(&link
->cset_link
);
692 * allocate_cgrp_cset_links - allocate cgrp_cset_links
693 * @count: the number of links to allocate
694 * @tmp_links: list_head the allocated links are put on
696 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
697 * through ->cset_link. Returns 0 on success or -errno.
699 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
701 struct cgrp_cset_link
*link
;
704 INIT_LIST_HEAD(tmp_links
);
706 for (i
= 0; i
< count
; i
++) {
707 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
709 free_cgrp_cset_links(tmp_links
);
712 list_add(&link
->cset_link
, tmp_links
);
718 * link_css_set - a helper function to link a css_set to a cgroup
719 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
720 * @cset: the css_set to be linked
721 * @cgrp: the destination cgroup
723 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
726 struct cgrp_cset_link
*link
;
728 BUG_ON(list_empty(tmp_links
));
730 if (cgroup_on_dfl(cgrp
))
731 cset
->dfl_cgrp
= cgrp
;
733 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
737 if (list_empty(&cgrp
->cset_links
))
738 cgroup_update_populated(cgrp
, true);
739 list_move(&link
->cset_link
, &cgrp
->cset_links
);
742 * Always add links to the tail of the list so that the list
743 * is sorted by order of hierarchy creation
745 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
749 * find_css_set - return a new css_set with one cgroup updated
750 * @old_cset: the baseline css_set
751 * @cgrp: the cgroup to be updated
753 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
754 * substituted into the appropriate hierarchy.
756 static struct css_set
*find_css_set(struct css_set
*old_cset
,
759 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
760 struct css_set
*cset
;
761 struct list_head tmp_links
;
762 struct cgrp_cset_link
*link
;
763 struct cgroup_subsys
*ss
;
767 lockdep_assert_held(&cgroup_mutex
);
769 /* First see if we already have a cgroup group that matches
771 down_read(&css_set_rwsem
);
772 cset
= find_existing_css_set(old_cset
, cgrp
, template);
775 up_read(&css_set_rwsem
);
780 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
784 /* Allocate all the cgrp_cset_link objects that we'll need */
785 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
790 atomic_set(&cset
->refcount
, 1);
791 INIT_LIST_HEAD(&cset
->cgrp_links
);
792 INIT_LIST_HEAD(&cset
->tasks
);
793 INIT_LIST_HEAD(&cset
->mg_tasks
);
794 INIT_LIST_HEAD(&cset
->mg_preload_node
);
795 INIT_LIST_HEAD(&cset
->mg_node
);
796 INIT_HLIST_NODE(&cset
->hlist
);
798 /* Copy the set of subsystem state objects generated in
799 * find_existing_css_set() */
800 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
802 down_write(&css_set_rwsem
);
803 /* Add reference counts and links from the new css_set. */
804 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
805 struct cgroup
*c
= link
->cgrp
;
807 if (c
->root
== cgrp
->root
)
809 link_css_set(&tmp_links
, cset
, c
);
812 BUG_ON(!list_empty(&tmp_links
));
816 /* Add @cset to the hash table */
817 key
= css_set_hash(cset
->subsys
);
818 hash_add(css_set_table
, &cset
->hlist
, key
);
820 for_each_subsys(ss
, ssid
)
821 list_add_tail(&cset
->e_cset_node
[ssid
],
822 &cset
->subsys
[ssid
]->cgroup
->e_csets
[ssid
]);
824 up_write(&css_set_rwsem
);
829 static struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
831 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
833 return root_cgrp
->root
;
836 static int cgroup_init_root_id(struct cgroup_root
*root
)
840 lockdep_assert_held(&cgroup_mutex
);
842 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
846 root
->hierarchy_id
= id
;
850 static void cgroup_exit_root_id(struct cgroup_root
*root
)
852 lockdep_assert_held(&cgroup_mutex
);
854 if (root
->hierarchy_id
) {
855 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
856 root
->hierarchy_id
= 0;
860 static void cgroup_free_root(struct cgroup_root
*root
)
863 /* hierarhcy ID shoulid already have been released */
864 WARN_ON_ONCE(root
->hierarchy_id
);
866 idr_destroy(&root
->cgroup_idr
);
871 static void cgroup_destroy_root(struct cgroup_root
*root
)
873 struct cgroup
*cgrp
= &root
->cgrp
;
874 struct cgrp_cset_link
*link
, *tmp_link
;
876 mutex_lock(&cgroup_mutex
);
878 BUG_ON(atomic_read(&root
->nr_cgrps
));
879 BUG_ON(!list_empty(&cgrp
->self
.children
));
881 /* Rebind all subsystems back to the default hierarchy */
882 rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
);
885 * Release all the links from cset_links to this hierarchy's
888 down_write(&css_set_rwsem
);
890 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
891 list_del(&link
->cset_link
);
892 list_del(&link
->cgrp_link
);
895 up_write(&css_set_rwsem
);
897 if (!list_empty(&root
->root_list
)) {
898 list_del(&root
->root_list
);
902 cgroup_exit_root_id(root
);
904 mutex_unlock(&cgroup_mutex
);
906 kernfs_destroy_root(root
->kf_root
);
907 cgroup_free_root(root
);
910 /* look up cgroup associated with given css_set on the specified hierarchy */
911 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
912 struct cgroup_root
*root
)
914 struct cgroup
*res
= NULL
;
916 lockdep_assert_held(&cgroup_mutex
);
917 lockdep_assert_held(&css_set_rwsem
);
919 if (cset
== &init_css_set
) {
922 struct cgrp_cset_link
*link
;
924 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
925 struct cgroup
*c
= link
->cgrp
;
927 if (c
->root
== root
) {
939 * Return the cgroup for "task" from the given hierarchy. Must be
940 * called with cgroup_mutex and css_set_rwsem held.
942 static struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
943 struct cgroup_root
*root
)
946 * No need to lock the task - since we hold cgroup_mutex the
947 * task can't change groups, so the only thing that can happen
948 * is that it exits and its css is set back to init_css_set.
950 return cset_cgroup_from_root(task_css_set(task
), root
);
954 * A task must hold cgroup_mutex to modify cgroups.
956 * Any task can increment and decrement the count field without lock.
957 * So in general, code holding cgroup_mutex can't rely on the count
958 * field not changing. However, if the count goes to zero, then only
959 * cgroup_attach_task() can increment it again. Because a count of zero
960 * means that no tasks are currently attached, therefore there is no
961 * way a task attached to that cgroup can fork (the other way to
962 * increment the count). So code holding cgroup_mutex can safely
963 * assume that if the count is zero, it will stay zero. Similarly, if
964 * a task holds cgroup_mutex on a cgroup with zero count, it
965 * knows that the cgroup won't be removed, as cgroup_rmdir()
968 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
969 * (usually) take cgroup_mutex. These are the two most performance
970 * critical pieces of code here. The exception occurs on cgroup_exit(),
971 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
972 * is taken, and if the cgroup count is zero, a usermode call made
973 * to the release agent with the name of the cgroup (path relative to
974 * the root of cgroup file system) as the argument.
976 * A cgroup can only be deleted if both its 'count' of using tasks
977 * is zero, and its list of 'children' cgroups is empty. Since all
978 * tasks in the system use _some_ cgroup, and since there is always at
979 * least one task in the system (init, pid == 1), therefore, root cgroup
980 * always has either children cgroups and/or using tasks. So we don't
981 * need a special hack to ensure that root cgroup cannot be deleted.
983 * P.S. One more locking exception. RCU is used to guard the
984 * update of a tasks cgroup pointer by cgroup_attach_task()
987 static int cgroup_populate_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
);
988 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
989 static const struct file_operations proc_cgroupstats_operations
;
991 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
994 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
995 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
996 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
997 cft
->ss
->name
, cft
->name
);
999 strncpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1004 * cgroup_file_mode - deduce file mode of a control file
1005 * @cft: the control file in question
1007 * returns cft->mode if ->mode is not 0
1008 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1009 * returns S_IRUGO if it has only a read handler
1010 * returns S_IWUSR if it has only a write hander
1012 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1019 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1022 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
)
1028 static void cgroup_get(struct cgroup
*cgrp
)
1030 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
1031 css_get(&cgrp
->self
);
1034 static void cgroup_put(struct cgroup
*cgrp
)
1036 css_put(&cgrp
->self
);
1040 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1041 * @kn: the kernfs_node being serviced
1043 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1044 * the method finishes if locking succeeded. Note that once this function
1045 * returns the cgroup returned by cgroup_kn_lock_live() may become
1046 * inaccessible any time. If the caller intends to continue to access the
1047 * cgroup, it should pin it before invoking this function.
1049 static void cgroup_kn_unlock(struct kernfs_node
*kn
)
1051 struct cgroup
*cgrp
;
1053 if (kernfs_type(kn
) == KERNFS_DIR
)
1056 cgrp
= kn
->parent
->priv
;
1058 mutex_unlock(&cgroup_mutex
);
1060 kernfs_unbreak_active_protection(kn
);
1065 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1066 * @kn: the kernfs_node being serviced
1068 * This helper is to be used by a cgroup kernfs method currently servicing
1069 * @kn. It breaks the active protection, performs cgroup locking and
1070 * verifies that the associated cgroup is alive. Returns the cgroup if
1071 * alive; otherwise, %NULL. A successful return should be undone by a
1072 * matching cgroup_kn_unlock() invocation.
1074 * Any cgroup kernfs method implementation which requires locking the
1075 * associated cgroup should use this helper. It avoids nesting cgroup
1076 * locking under kernfs active protection and allows all kernfs operations
1077 * including self-removal.
1079 static struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
)
1081 struct cgroup
*cgrp
;
1083 if (kernfs_type(kn
) == KERNFS_DIR
)
1086 cgrp
= kn
->parent
->priv
;
1089 * We're gonna grab cgroup_mutex which nests outside kernfs
1090 * active_ref. cgroup liveliness check alone provides enough
1091 * protection against removal. Ensure @cgrp stays accessible and
1092 * break the active_ref protection.
1095 kernfs_break_active_protection(kn
);
1097 mutex_lock(&cgroup_mutex
);
1099 if (!cgroup_is_dead(cgrp
))
1102 cgroup_kn_unlock(kn
);
1106 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1108 char name
[CGROUP_FILE_NAME_MAX
];
1110 lockdep_assert_held(&cgroup_mutex
);
1111 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1115 * cgroup_clear_dir - remove subsys files in a cgroup directory
1116 * @cgrp: target cgroup
1117 * @subsys_mask: mask of the subsystem ids whose files should be removed
1119 static void cgroup_clear_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
)
1121 struct cgroup_subsys
*ss
;
1124 for_each_subsys(ss
, i
) {
1125 struct cftype
*cfts
;
1127 if (!(subsys_mask
& (1 << i
)))
1129 list_for_each_entry(cfts
, &ss
->cfts
, node
)
1130 cgroup_addrm_files(cgrp
, cfts
, false);
1134 static int rebind_subsystems(struct cgroup_root
*dst_root
, unsigned int ss_mask
)
1136 struct cgroup_subsys
*ss
;
1137 unsigned int tmp_ss_mask
;
1140 lockdep_assert_held(&cgroup_mutex
);
1142 for_each_subsys(ss
, ssid
) {
1143 if (!(ss_mask
& (1 << ssid
)))
1146 /* if @ss has non-root csses attached to it, can't move */
1147 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)))
1150 /* can't move between two non-dummy roots either */
1151 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1155 /* skip creating root files on dfl_root for inhibited subsystems */
1156 tmp_ss_mask
= ss_mask
;
1157 if (dst_root
== &cgrp_dfl_root
)
1158 tmp_ss_mask
&= ~cgrp_dfl_root_inhibit_ss_mask
;
1160 ret
= cgroup_populate_dir(&dst_root
->cgrp
, tmp_ss_mask
);
1162 if (dst_root
!= &cgrp_dfl_root
)
1166 * Rebinding back to the default root is not allowed to
1167 * fail. Using both default and non-default roots should
1168 * be rare. Moving subsystems back and forth even more so.
1169 * Just warn about it and continue.
1171 if (cgrp_dfl_root_visible
) {
1172 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1174 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1179 * Nothing can fail from this point on. Remove files for the
1180 * removed subsystems and rebind each subsystem.
1182 for_each_subsys(ss
, ssid
)
1183 if (ss_mask
& (1 << ssid
))
1184 cgroup_clear_dir(&ss
->root
->cgrp
, 1 << ssid
);
1186 for_each_subsys(ss
, ssid
) {
1187 struct cgroup_root
*src_root
;
1188 struct cgroup_subsys_state
*css
;
1189 struct css_set
*cset
;
1191 if (!(ss_mask
& (1 << ssid
)))
1194 src_root
= ss
->root
;
1195 css
= cgroup_css(&src_root
->cgrp
, ss
);
1197 WARN_ON(!css
|| cgroup_css(&dst_root
->cgrp
, ss
));
1199 RCU_INIT_POINTER(src_root
->cgrp
.subsys
[ssid
], NULL
);
1200 rcu_assign_pointer(dst_root
->cgrp
.subsys
[ssid
], css
);
1201 ss
->root
= dst_root
;
1202 css
->cgroup
= &dst_root
->cgrp
;
1204 down_write(&css_set_rwsem
);
1205 hash_for_each(css_set_table
, i
, cset
, hlist
)
1206 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1207 &dst_root
->cgrp
.e_csets
[ss
->id
]);
1208 up_write(&css_set_rwsem
);
1210 src_root
->subsys_mask
&= ~(1 << ssid
);
1211 src_root
->cgrp
.child_subsys_mask
&= ~(1 << ssid
);
1213 /* default hierarchy doesn't enable controllers by default */
1214 dst_root
->subsys_mask
|= 1 << ssid
;
1215 if (dst_root
!= &cgrp_dfl_root
)
1216 dst_root
->cgrp
.child_subsys_mask
|= 1 << ssid
;
1222 kernfs_activate(dst_root
->cgrp
.kn
);
1226 static int cgroup_show_options(struct seq_file
*seq
,
1227 struct kernfs_root
*kf_root
)
1229 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1230 struct cgroup_subsys
*ss
;
1233 for_each_subsys(ss
, ssid
)
1234 if (root
->subsys_mask
& (1 << ssid
))
1235 seq_printf(seq
, ",%s", ss
->name
);
1236 if (root
->flags
& CGRP_ROOT_SANE_BEHAVIOR
)
1237 seq_puts(seq
, ",sane_behavior");
1238 if (root
->flags
& CGRP_ROOT_NOPREFIX
)
1239 seq_puts(seq
, ",noprefix");
1240 if (root
->flags
& CGRP_ROOT_XATTR
)
1241 seq_puts(seq
, ",xattr");
1243 spin_lock(&release_agent_path_lock
);
1244 if (strlen(root
->release_agent_path
))
1245 seq_printf(seq
, ",release_agent=%s", root
->release_agent_path
);
1246 spin_unlock(&release_agent_path_lock
);
1248 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
))
1249 seq_puts(seq
, ",clone_children");
1250 if (strlen(root
->name
))
1251 seq_printf(seq
, ",name=%s", root
->name
);
1255 struct cgroup_sb_opts
{
1256 unsigned int subsys_mask
;
1258 char *release_agent
;
1259 bool cpuset_clone_children
;
1261 /* User explicitly requested empty subsystem */
1265 static int parse_cgroupfs_options(char *data
, struct cgroup_sb_opts
*opts
)
1267 char *token
, *o
= data
;
1268 bool all_ss
= false, one_ss
= false;
1269 unsigned int mask
= -1U;
1270 struct cgroup_subsys
*ss
;
1273 #ifdef CONFIG_CPUSETS
1274 mask
= ~(1U << cpuset_cgrp_id
);
1277 memset(opts
, 0, sizeof(*opts
));
1279 while ((token
= strsep(&o
, ",")) != NULL
) {
1282 if (!strcmp(token
, "none")) {
1283 /* Explicitly have no subsystems */
1287 if (!strcmp(token
, "all")) {
1288 /* Mutually exclusive option 'all' + subsystem name */
1294 if (!strcmp(token
, "__DEVEL__sane_behavior")) {
1295 opts
->flags
|= CGRP_ROOT_SANE_BEHAVIOR
;
1298 if (!strcmp(token
, "noprefix")) {
1299 opts
->flags
|= CGRP_ROOT_NOPREFIX
;
1302 if (!strcmp(token
, "clone_children")) {
1303 opts
->cpuset_clone_children
= true;
1306 if (!strcmp(token
, "xattr")) {
1307 opts
->flags
|= CGRP_ROOT_XATTR
;
1310 if (!strncmp(token
, "release_agent=", 14)) {
1311 /* Specifying two release agents is forbidden */
1312 if (opts
->release_agent
)
1314 opts
->release_agent
=
1315 kstrndup(token
+ 14, PATH_MAX
- 1, GFP_KERNEL
);
1316 if (!opts
->release_agent
)
1320 if (!strncmp(token
, "name=", 5)) {
1321 const char *name
= token
+ 5;
1322 /* Can't specify an empty name */
1325 /* Must match [\w.-]+ */
1326 for (i
= 0; i
< strlen(name
); i
++) {
1330 if ((c
== '.') || (c
== '-') || (c
== '_'))
1334 /* Specifying two names is forbidden */
1337 opts
->name
= kstrndup(name
,
1338 MAX_CGROUP_ROOT_NAMELEN
- 1,
1346 for_each_subsys(ss
, i
) {
1347 if (strcmp(token
, ss
->name
))
1352 /* Mutually exclusive option 'all' + subsystem name */
1355 opts
->subsys_mask
|= (1 << i
);
1360 if (i
== CGROUP_SUBSYS_COUNT
)
1364 /* Consistency checks */
1366 if (opts
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1367 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1369 if ((opts
->flags
& (CGRP_ROOT_NOPREFIX
| CGRP_ROOT_XATTR
)) ||
1370 opts
->cpuset_clone_children
|| opts
->release_agent
||
1372 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1377 * If the 'all' option was specified select all the
1378 * subsystems, otherwise if 'none', 'name=' and a subsystem
1379 * name options were not specified, let's default to 'all'
1381 if (all_ss
|| (!one_ss
&& !opts
->none
&& !opts
->name
))
1382 for_each_subsys(ss
, i
)
1384 opts
->subsys_mask
|= (1 << i
);
1387 * We either have to specify by name or by subsystems. (So
1388 * all empty hierarchies must have a name).
1390 if (!opts
->subsys_mask
&& !opts
->name
)
1395 * Option noprefix was introduced just for backward compatibility
1396 * with the old cpuset, so we allow noprefix only if mounting just
1397 * the cpuset subsystem.
1399 if ((opts
->flags
& CGRP_ROOT_NOPREFIX
) && (opts
->subsys_mask
& mask
))
1403 /* Can't specify "none" and some subsystems */
1404 if (opts
->subsys_mask
&& opts
->none
)
1410 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1413 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1414 struct cgroup_sb_opts opts
;
1415 unsigned int added_mask
, removed_mask
;
1417 if (root
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1418 pr_err("sane_behavior: remount is not allowed\n");
1422 mutex_lock(&cgroup_mutex
);
1424 /* See what subsystems are wanted */
1425 ret
= parse_cgroupfs_options(data
, &opts
);
1429 if (opts
.subsys_mask
!= root
->subsys_mask
|| opts
.release_agent
)
1430 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1431 task_tgid_nr(current
), current
->comm
);
1433 added_mask
= opts
.subsys_mask
& ~root
->subsys_mask
;
1434 removed_mask
= root
->subsys_mask
& ~opts
.subsys_mask
;
1436 /* Don't allow flags or name to change at remount */
1437 if (((opts
.flags
^ root
->flags
) & CGRP_ROOT_OPTION_MASK
) ||
1438 (opts
.name
&& strcmp(opts
.name
, root
->name
))) {
1439 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1440 opts
.flags
& CGRP_ROOT_OPTION_MASK
, opts
.name
?: "",
1441 root
->flags
& CGRP_ROOT_OPTION_MASK
, root
->name
);
1446 /* remounting is not allowed for populated hierarchies */
1447 if (!list_empty(&root
->cgrp
.self
.children
)) {
1452 ret
= rebind_subsystems(root
, added_mask
);
1456 rebind_subsystems(&cgrp_dfl_root
, removed_mask
);
1458 if (opts
.release_agent
) {
1459 spin_lock(&release_agent_path_lock
);
1460 strcpy(root
->release_agent_path
, opts
.release_agent
);
1461 spin_unlock(&release_agent_path_lock
);
1464 kfree(opts
.release_agent
);
1466 mutex_unlock(&cgroup_mutex
);
1471 * To reduce the fork() overhead for systems that are not actually using
1472 * their cgroups capability, we don't maintain the lists running through
1473 * each css_set to its tasks until we see the list actually used - in other
1474 * words after the first mount.
1476 static bool use_task_css_set_links __read_mostly
;
1478 static void cgroup_enable_task_cg_lists(void)
1480 struct task_struct
*p
, *g
;
1482 down_write(&css_set_rwsem
);
1484 if (use_task_css_set_links
)
1487 use_task_css_set_links
= true;
1490 * We need tasklist_lock because RCU is not safe against
1491 * while_each_thread(). Besides, a forking task that has passed
1492 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1493 * is not guaranteed to have its child immediately visible in the
1494 * tasklist if we walk through it with RCU.
1496 read_lock(&tasklist_lock
);
1497 do_each_thread(g
, p
) {
1498 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1499 task_css_set(p
) != &init_css_set
);
1502 * We should check if the process is exiting, otherwise
1503 * it will race with cgroup_exit() in that the list
1504 * entry won't be deleted though the process has exited.
1505 * Do it while holding siglock so that we don't end up
1506 * racing against cgroup_exit().
1508 spin_lock_irq(&p
->sighand
->siglock
);
1509 if (!(p
->flags
& PF_EXITING
)) {
1510 struct css_set
*cset
= task_css_set(p
);
1512 list_add(&p
->cg_list
, &cset
->tasks
);
1515 spin_unlock_irq(&p
->sighand
->siglock
);
1516 } while_each_thread(g
, p
);
1517 read_unlock(&tasklist_lock
);
1519 up_write(&css_set_rwsem
);
1522 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1524 struct cgroup_subsys
*ss
;
1527 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1528 INIT_LIST_HEAD(&cgrp
->self
.children
);
1529 INIT_LIST_HEAD(&cgrp
->cset_links
);
1530 INIT_LIST_HEAD(&cgrp
->release_list
);
1531 INIT_LIST_HEAD(&cgrp
->pidlists
);
1532 mutex_init(&cgrp
->pidlist_mutex
);
1533 cgrp
->self
.cgroup
= cgrp
;
1534 cgrp
->self
.flags
|= CSS_ONLINE
;
1536 for_each_subsys(ss
, ssid
)
1537 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1539 init_waitqueue_head(&cgrp
->offline_waitq
);
1542 static void init_cgroup_root(struct cgroup_root
*root
,
1543 struct cgroup_sb_opts
*opts
)
1545 struct cgroup
*cgrp
= &root
->cgrp
;
1547 INIT_LIST_HEAD(&root
->root_list
);
1548 atomic_set(&root
->nr_cgrps
, 1);
1550 init_cgroup_housekeeping(cgrp
);
1551 idr_init(&root
->cgroup_idr
);
1553 root
->flags
= opts
->flags
;
1554 if (opts
->release_agent
)
1555 strcpy(root
->release_agent_path
, opts
->release_agent
);
1557 strcpy(root
->name
, opts
->name
);
1558 if (opts
->cpuset_clone_children
)
1559 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1562 static int cgroup_setup_root(struct cgroup_root
*root
, unsigned int ss_mask
)
1564 LIST_HEAD(tmp_links
);
1565 struct cgroup
*root_cgrp
= &root
->cgrp
;
1566 struct css_set
*cset
;
1569 lockdep_assert_held(&cgroup_mutex
);
1571 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_NOWAIT
);
1574 root_cgrp
->id
= ret
;
1576 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
);
1581 * We're accessing css_set_count without locking css_set_rwsem here,
1582 * but that's OK - it can only be increased by someone holding
1583 * cgroup_lock, and that's us. The worst that can happen is that we
1584 * have some link structures left over
1586 ret
= allocate_cgrp_cset_links(css_set_count
, &tmp_links
);
1590 ret
= cgroup_init_root_id(root
);
1594 root
->kf_root
= kernfs_create_root(&cgroup_kf_syscall_ops
,
1595 KERNFS_ROOT_CREATE_DEACTIVATED
,
1597 if (IS_ERR(root
->kf_root
)) {
1598 ret
= PTR_ERR(root
->kf_root
);
1601 root_cgrp
->kn
= root
->kf_root
->kn
;
1603 ret
= cgroup_addrm_files(root_cgrp
, cgroup_base_files
, true);
1607 ret
= rebind_subsystems(root
, ss_mask
);
1612 * There must be no failure case after here, since rebinding takes
1613 * care of subsystems' refcounts, which are explicitly dropped in
1614 * the failure exit path.
1616 list_add(&root
->root_list
, &cgroup_roots
);
1617 cgroup_root_count
++;
1620 * Link the root cgroup in this hierarchy into all the css_set
1623 down_write(&css_set_rwsem
);
1624 hash_for_each(css_set_table
, i
, cset
, hlist
)
1625 link_css_set(&tmp_links
, cset
, root_cgrp
);
1626 up_write(&css_set_rwsem
);
1628 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
1629 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1631 kernfs_activate(root_cgrp
->kn
);
1636 kernfs_destroy_root(root
->kf_root
);
1637 root
->kf_root
= NULL
;
1639 cgroup_exit_root_id(root
);
1641 percpu_ref_cancel_init(&root_cgrp
->self
.refcnt
);
1643 free_cgrp_cset_links(&tmp_links
);
1647 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
1648 int flags
, const char *unused_dev_name
,
1651 struct cgroup_subsys
*ss
;
1652 struct cgroup_root
*root
;
1653 struct cgroup_sb_opts opts
;
1654 struct dentry
*dentry
;
1660 * The first time anyone tries to mount a cgroup, enable the list
1661 * linking each css_set to its tasks and fix up all existing tasks.
1663 if (!use_task_css_set_links
)
1664 cgroup_enable_task_cg_lists();
1666 mutex_lock(&cgroup_mutex
);
1668 /* First find the desired set of subsystems */
1669 ret
= parse_cgroupfs_options(data
, &opts
);
1673 /* look for a matching existing root */
1674 if (!opts
.subsys_mask
&& !opts
.none
&& !opts
.name
) {
1675 cgrp_dfl_root_visible
= true;
1676 root
= &cgrp_dfl_root
;
1677 cgroup_get(&root
->cgrp
);
1683 * Destruction of cgroup root is asynchronous, so subsystems may
1684 * still be dying after the previous unmount. Let's drain the
1685 * dying subsystems. We just need to ensure that the ones
1686 * unmounted previously finish dying and don't care about new ones
1687 * starting. Testing ref liveliness is good enough.
1689 for_each_subsys(ss
, i
) {
1690 if (!(opts
.subsys_mask
& (1 << i
)) ||
1691 ss
->root
== &cgrp_dfl_root
)
1694 if (!percpu_ref_tryget_live(&ss
->root
->cgrp
.self
.refcnt
)) {
1695 mutex_unlock(&cgroup_mutex
);
1697 ret
= restart_syscall();
1700 cgroup_put(&ss
->root
->cgrp
);
1703 for_each_root(root
) {
1704 bool name_match
= false;
1706 if (root
== &cgrp_dfl_root
)
1710 * If we asked for a name then it must match. Also, if
1711 * name matches but sybsys_mask doesn't, we should fail.
1712 * Remember whether name matched.
1715 if (strcmp(opts
.name
, root
->name
))
1721 * If we asked for subsystems (or explicitly for no
1722 * subsystems) then they must match.
1724 if ((opts
.subsys_mask
|| opts
.none
) &&
1725 (opts
.subsys_mask
!= root
->subsys_mask
)) {
1732 if ((root
->flags
^ opts
.flags
) & CGRP_ROOT_OPTION_MASK
) {
1733 if ((root
->flags
| opts
.flags
) & CGRP_ROOT_SANE_BEHAVIOR
) {
1734 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1738 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1743 * A root's lifetime is governed by its root cgroup.
1744 * tryget_live failure indicate that the root is being
1745 * destroyed. Wait for destruction to complete so that the
1746 * subsystems are free. We can use wait_queue for the wait
1747 * but this path is super cold. Let's just sleep for a bit
1750 if (!percpu_ref_tryget_live(&root
->cgrp
.self
.refcnt
)) {
1751 mutex_unlock(&cgroup_mutex
);
1753 ret
= restart_syscall();
1762 * No such thing, create a new one. name= matching without subsys
1763 * specification is allowed for already existing hierarchies but we
1764 * can't create new one without subsys specification.
1766 if (!opts
.subsys_mask
&& !opts
.none
) {
1771 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
1777 init_cgroup_root(root
, &opts
);
1779 ret
= cgroup_setup_root(root
, opts
.subsys_mask
);
1781 cgroup_free_root(root
);
1784 mutex_unlock(&cgroup_mutex
);
1786 kfree(opts
.release_agent
);
1790 return ERR_PTR(ret
);
1792 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
,
1793 CGROUP_SUPER_MAGIC
, &new_sb
);
1794 if (IS_ERR(dentry
) || !new_sb
)
1795 cgroup_put(&root
->cgrp
);
1799 static void cgroup_kill_sb(struct super_block
*sb
)
1801 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
1802 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1805 * If @root doesn't have any mounts or children, start killing it.
1806 * This prevents new mounts by disabling percpu_ref_tryget_live().
1807 * cgroup_mount() may wait for @root's release.
1809 * And don't kill the default root.
1811 if (css_has_online_children(&root
->cgrp
.self
) ||
1812 root
== &cgrp_dfl_root
)
1813 cgroup_put(&root
->cgrp
);
1815 percpu_ref_kill(&root
->cgrp
.self
.refcnt
);
1820 static struct file_system_type cgroup_fs_type
= {
1822 .mount
= cgroup_mount
,
1823 .kill_sb
= cgroup_kill_sb
,
1826 static struct kobject
*cgroup_kobj
;
1829 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1830 * @task: target task
1831 * @buf: the buffer to write the path into
1832 * @buflen: the length of the buffer
1834 * Determine @task's cgroup on the first (the one with the lowest non-zero
1835 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1836 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1837 * cgroup controller callbacks.
1839 * Return value is the same as kernfs_path().
1841 char *task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
1843 struct cgroup_root
*root
;
1844 struct cgroup
*cgrp
;
1845 int hierarchy_id
= 1;
1848 mutex_lock(&cgroup_mutex
);
1849 down_read(&css_set_rwsem
);
1851 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
1854 cgrp
= task_cgroup_from_root(task
, root
);
1855 path
= cgroup_path(cgrp
, buf
, buflen
);
1857 /* if no hierarchy exists, everyone is in "/" */
1858 if (strlcpy(buf
, "/", buflen
) < buflen
)
1862 up_read(&css_set_rwsem
);
1863 mutex_unlock(&cgroup_mutex
);
1866 EXPORT_SYMBOL_GPL(task_cgroup_path
);
1868 /* used to track tasks and other necessary states during migration */
1869 struct cgroup_taskset
{
1870 /* the src and dst cset list running through cset->mg_node */
1871 struct list_head src_csets
;
1872 struct list_head dst_csets
;
1875 * Fields for cgroup_taskset_*() iteration.
1877 * Before migration is committed, the target migration tasks are on
1878 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1879 * the csets on ->dst_csets. ->csets point to either ->src_csets
1880 * or ->dst_csets depending on whether migration is committed.
1882 * ->cur_csets and ->cur_task point to the current task position
1885 struct list_head
*csets
;
1886 struct css_set
*cur_cset
;
1887 struct task_struct
*cur_task
;
1891 * cgroup_taskset_first - reset taskset and return the first task
1892 * @tset: taskset of interest
1894 * @tset iteration is initialized and the first task is returned.
1896 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
)
1898 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
1899 tset
->cur_task
= NULL
;
1901 return cgroup_taskset_next(tset
);
1905 * cgroup_taskset_next - iterate to the next task in taskset
1906 * @tset: taskset of interest
1908 * Return the next task in @tset. Iteration must have been initialized
1909 * with cgroup_taskset_first().
1911 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
)
1913 struct css_set
*cset
= tset
->cur_cset
;
1914 struct task_struct
*task
= tset
->cur_task
;
1916 while (&cset
->mg_node
!= tset
->csets
) {
1918 task
= list_first_entry(&cset
->mg_tasks
,
1919 struct task_struct
, cg_list
);
1921 task
= list_next_entry(task
, cg_list
);
1923 if (&task
->cg_list
!= &cset
->mg_tasks
) {
1924 tset
->cur_cset
= cset
;
1925 tset
->cur_task
= task
;
1929 cset
= list_next_entry(cset
, mg_node
);
1937 * cgroup_task_migrate - move a task from one cgroup to another.
1938 * @old_cgrp: the cgroup @tsk is being migrated from
1939 * @tsk: the task being migrated
1940 * @new_cset: the new css_set @tsk is being attached to
1942 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1944 static void cgroup_task_migrate(struct cgroup
*old_cgrp
,
1945 struct task_struct
*tsk
,
1946 struct css_set
*new_cset
)
1948 struct css_set
*old_cset
;
1950 lockdep_assert_held(&cgroup_mutex
);
1951 lockdep_assert_held(&css_set_rwsem
);
1954 * We are synchronized through threadgroup_lock() against PF_EXITING
1955 * setting such that we can't race against cgroup_exit() changing the
1956 * css_set to init_css_set and dropping the old one.
1958 WARN_ON_ONCE(tsk
->flags
& PF_EXITING
);
1959 old_cset
= task_css_set(tsk
);
1961 get_css_set(new_cset
);
1962 rcu_assign_pointer(tsk
->cgroups
, new_cset
);
1965 * Use move_tail so that cgroup_taskset_first() still returns the
1966 * leader after migration. This works because cgroup_migrate()
1967 * ensures that the dst_cset of the leader is the first on the
1968 * tset's dst_csets list.
1970 list_move_tail(&tsk
->cg_list
, &new_cset
->mg_tasks
);
1973 * We just gained a reference on old_cset by taking it from the
1974 * task. As trading it for new_cset is protected by cgroup_mutex,
1975 * we're safe to drop it here; it will be freed under RCU.
1977 set_bit(CGRP_RELEASABLE
, &old_cgrp
->flags
);
1978 put_css_set_locked(old_cset
, false);
1982 * cgroup_migrate_finish - cleanup after attach
1983 * @preloaded_csets: list of preloaded css_sets
1985 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1986 * those functions for details.
1988 static void cgroup_migrate_finish(struct list_head
*preloaded_csets
)
1990 struct css_set
*cset
, *tmp_cset
;
1992 lockdep_assert_held(&cgroup_mutex
);
1994 down_write(&css_set_rwsem
);
1995 list_for_each_entry_safe(cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
1996 cset
->mg_src_cgrp
= NULL
;
1997 cset
->mg_dst_cset
= NULL
;
1998 list_del_init(&cset
->mg_preload_node
);
1999 put_css_set_locked(cset
, false);
2001 up_write(&css_set_rwsem
);
2005 * cgroup_migrate_add_src - add a migration source css_set
2006 * @src_cset: the source css_set to add
2007 * @dst_cgrp: the destination cgroup
2008 * @preloaded_csets: list of preloaded css_sets
2010 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2011 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2012 * up by cgroup_migrate_finish().
2014 * This function may be called without holding threadgroup_lock even if the
2015 * target is a process. Threads may be created and destroyed but as long
2016 * as cgroup_mutex is not dropped, no new css_set can be put into play and
2017 * the preloaded css_sets are guaranteed to cover all migrations.
2019 static void cgroup_migrate_add_src(struct css_set
*src_cset
,
2020 struct cgroup
*dst_cgrp
,
2021 struct list_head
*preloaded_csets
)
2023 struct cgroup
*src_cgrp
;
2025 lockdep_assert_held(&cgroup_mutex
);
2026 lockdep_assert_held(&css_set_rwsem
);
2028 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2030 if (!list_empty(&src_cset
->mg_preload_node
))
2033 WARN_ON(src_cset
->mg_src_cgrp
);
2034 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2035 WARN_ON(!list_empty(&src_cset
->mg_node
));
2037 src_cset
->mg_src_cgrp
= src_cgrp
;
2038 get_css_set(src_cset
);
2039 list_add(&src_cset
->mg_preload_node
, preloaded_csets
);
2043 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2044 * @dst_cgrp: the destination cgroup (may be %NULL)
2045 * @preloaded_csets: list of preloaded source css_sets
2047 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2048 * have been preloaded to @preloaded_csets. This function looks up and
2049 * pins all destination css_sets, links each to its source, and append them
2050 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2051 * source css_set is assumed to be its cgroup on the default hierarchy.
2053 * This function must be called after cgroup_migrate_add_src() has been
2054 * called on each migration source css_set. After migration is performed
2055 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2058 static int cgroup_migrate_prepare_dst(struct cgroup
*dst_cgrp
,
2059 struct list_head
*preloaded_csets
)
2062 struct css_set
*src_cset
, *tmp_cset
;
2064 lockdep_assert_held(&cgroup_mutex
);
2067 * Except for the root, child_subsys_mask must be zero for a cgroup
2068 * with tasks so that child cgroups don't compete against tasks.
2070 if (dst_cgrp
&& cgroup_on_dfl(dst_cgrp
) && cgroup_parent(dst_cgrp
) &&
2071 dst_cgrp
->child_subsys_mask
)
2074 /* look up the dst cset for each src cset and link it to src */
2075 list_for_each_entry_safe(src_cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2076 struct css_set
*dst_cset
;
2078 dst_cset
= find_css_set(src_cset
,
2079 dst_cgrp
?: src_cset
->dfl_cgrp
);
2083 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2086 * If src cset equals dst, it's noop. Drop the src.
2087 * cgroup_migrate() will skip the cset too. Note that we
2088 * can't handle src == dst as some nodes are used by both.
2090 if (src_cset
== dst_cset
) {
2091 src_cset
->mg_src_cgrp
= NULL
;
2092 list_del_init(&src_cset
->mg_preload_node
);
2093 put_css_set(src_cset
, false);
2094 put_css_set(dst_cset
, false);
2098 src_cset
->mg_dst_cset
= dst_cset
;
2100 if (list_empty(&dst_cset
->mg_preload_node
))
2101 list_add(&dst_cset
->mg_preload_node
, &csets
);
2103 put_css_set(dst_cset
, false);
2106 list_splice_tail(&csets
, preloaded_csets
);
2109 cgroup_migrate_finish(&csets
);
2114 * cgroup_migrate - migrate a process or task to a cgroup
2115 * @cgrp: the destination cgroup
2116 * @leader: the leader of the process or the task to migrate
2117 * @threadgroup: whether @leader points to the whole process or a single task
2119 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2120 * process, the caller must be holding threadgroup_lock of @leader. The
2121 * caller is also responsible for invoking cgroup_migrate_add_src() and
2122 * cgroup_migrate_prepare_dst() on the targets before invoking this
2123 * function and following up with cgroup_migrate_finish().
2125 * As long as a controller's ->can_attach() doesn't fail, this function is
2126 * guaranteed to succeed. This means that, excluding ->can_attach()
2127 * failure, when migrating multiple targets, the success or failure can be
2128 * decided for all targets by invoking group_migrate_prepare_dst() before
2129 * actually starting migrating.
2131 static int cgroup_migrate(struct cgroup
*cgrp
, struct task_struct
*leader
,
2134 struct cgroup_taskset tset
= {
2135 .src_csets
= LIST_HEAD_INIT(tset
.src_csets
),
2136 .dst_csets
= LIST_HEAD_INIT(tset
.dst_csets
),
2137 .csets
= &tset
.src_csets
,
2139 struct cgroup_subsys_state
*css
, *failed_css
= NULL
;
2140 struct css_set
*cset
, *tmp_cset
;
2141 struct task_struct
*task
, *tmp_task
;
2145 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2146 * already PF_EXITING could be freed from underneath us unless we
2147 * take an rcu_read_lock.
2149 down_write(&css_set_rwsem
);
2153 /* @task either already exited or can't exit until the end */
2154 if (task
->flags
& PF_EXITING
)
2157 /* leave @task alone if post_fork() hasn't linked it yet */
2158 if (list_empty(&task
->cg_list
))
2161 cset
= task_css_set(task
);
2162 if (!cset
->mg_src_cgrp
)
2166 * cgroup_taskset_first() must always return the leader.
2167 * Take care to avoid disturbing the ordering.
2169 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2170 if (list_empty(&cset
->mg_node
))
2171 list_add_tail(&cset
->mg_node
, &tset
.src_csets
);
2172 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2173 list_move_tail(&cset
->mg_dst_cset
->mg_node
,
2178 } while_each_thread(leader
, task
);
2180 up_write(&css_set_rwsem
);
2182 /* methods shouldn't be called if no task is actually migrating */
2183 if (list_empty(&tset
.src_csets
))
2186 /* check that we can legitimately attach to the cgroup */
2187 for_each_e_css(css
, i
, cgrp
) {
2188 if (css
->ss
->can_attach
) {
2189 ret
= css
->ss
->can_attach(css
, &tset
);
2192 goto out_cancel_attach
;
2198 * Now that we're guaranteed success, proceed to move all tasks to
2199 * the new cgroup. There are no failure cases after here, so this
2200 * is the commit point.
2202 down_write(&css_set_rwsem
);
2203 list_for_each_entry(cset
, &tset
.src_csets
, mg_node
) {
2204 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
)
2205 cgroup_task_migrate(cset
->mg_src_cgrp
, task
,
2208 up_write(&css_set_rwsem
);
2211 * Migration is committed, all target tasks are now on dst_csets.
2212 * Nothing is sensitive to fork() after this point. Notify
2213 * controllers that migration is complete.
2215 tset
.csets
= &tset
.dst_csets
;
2217 for_each_e_css(css
, i
, cgrp
)
2218 if (css
->ss
->attach
)
2219 css
->ss
->attach(css
, &tset
);
2222 goto out_release_tset
;
2225 for_each_e_css(css
, i
, cgrp
) {
2226 if (css
== failed_css
)
2228 if (css
->ss
->cancel_attach
)
2229 css
->ss
->cancel_attach(css
, &tset
);
2232 down_write(&css_set_rwsem
);
2233 list_splice_init(&tset
.dst_csets
, &tset
.src_csets
);
2234 list_for_each_entry_safe(cset
, tmp_cset
, &tset
.src_csets
, mg_node
) {
2235 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2236 list_del_init(&cset
->mg_node
);
2238 up_write(&css_set_rwsem
);
2243 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2244 * @dst_cgrp: the cgroup to attach to
2245 * @leader: the task or the leader of the threadgroup to be attached
2246 * @threadgroup: attach the whole threadgroup?
2248 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2250 static int cgroup_attach_task(struct cgroup
*dst_cgrp
,
2251 struct task_struct
*leader
, bool threadgroup
)
2253 LIST_HEAD(preloaded_csets
);
2254 struct task_struct
*task
;
2257 /* look up all src csets */
2258 down_read(&css_set_rwsem
);
2262 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
,
2266 } while_each_thread(leader
, task
);
2268 up_read(&css_set_rwsem
);
2270 /* prepare dst csets and commit */
2271 ret
= cgroup_migrate_prepare_dst(dst_cgrp
, &preloaded_csets
);
2273 ret
= cgroup_migrate(dst_cgrp
, leader
, threadgroup
);
2275 cgroup_migrate_finish(&preloaded_csets
);
2280 * Find the task_struct of the task to attach by vpid and pass it along to the
2281 * function to attach either it or all tasks in its threadgroup. Will lock
2282 * cgroup_mutex and threadgroup.
2284 static ssize_t
__cgroup_procs_write(struct kernfs_open_file
*of
, char *buf
,
2285 size_t nbytes
, loff_t off
, bool threadgroup
)
2287 struct task_struct
*tsk
;
2288 const struct cred
*cred
= current_cred(), *tcred
;
2289 struct cgroup
*cgrp
;
2293 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2296 cgrp
= cgroup_kn_lock_live(of
->kn
);
2303 tsk
= find_task_by_vpid(pid
);
2307 goto out_unlock_cgroup
;
2310 * even if we're attaching all tasks in the thread group, we
2311 * only need to check permissions on one of them.
2313 tcred
= __task_cred(tsk
);
2314 if (!uid_eq(cred
->euid
, GLOBAL_ROOT_UID
) &&
2315 !uid_eq(cred
->euid
, tcred
->uid
) &&
2316 !uid_eq(cred
->euid
, tcred
->suid
)) {
2319 goto out_unlock_cgroup
;
2325 tsk
= tsk
->group_leader
;
2328 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2329 * trapped in a cpuset, or RT worker may be born in a cgroup
2330 * with no rt_runtime allocated. Just say no.
2332 if (tsk
== kthreadd_task
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2335 goto out_unlock_cgroup
;
2338 get_task_struct(tsk
);
2341 threadgroup_lock(tsk
);
2343 if (!thread_group_leader(tsk
)) {
2345 * a race with de_thread from another thread's exec()
2346 * may strip us of our leadership, if this happens,
2347 * there is no choice but to throw this task away and
2348 * try again; this is
2349 * "double-double-toil-and-trouble-check locking".
2351 threadgroup_unlock(tsk
);
2352 put_task_struct(tsk
);
2353 goto retry_find_task
;
2357 ret
= cgroup_attach_task(cgrp
, tsk
, threadgroup
);
2359 threadgroup_unlock(tsk
);
2361 put_task_struct(tsk
);
2363 cgroup_kn_unlock(of
->kn
);
2364 return ret
?: nbytes
;
2368 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2369 * @from: attach to all cgroups of a given task
2370 * @tsk: the task to be attached
2372 int cgroup_attach_task_all(struct task_struct
*from
, struct task_struct
*tsk
)
2374 struct cgroup_root
*root
;
2377 mutex_lock(&cgroup_mutex
);
2378 for_each_root(root
) {
2379 struct cgroup
*from_cgrp
;
2381 if (root
== &cgrp_dfl_root
)
2384 down_read(&css_set_rwsem
);
2385 from_cgrp
= task_cgroup_from_root(from
, root
);
2386 up_read(&css_set_rwsem
);
2388 retval
= cgroup_attach_task(from_cgrp
, tsk
, false);
2392 mutex_unlock(&cgroup_mutex
);
2396 EXPORT_SYMBOL_GPL(cgroup_attach_task_all
);
2398 static ssize_t
cgroup_tasks_write(struct kernfs_open_file
*of
,
2399 char *buf
, size_t nbytes
, loff_t off
)
2401 return __cgroup_procs_write(of
, buf
, nbytes
, off
, false);
2404 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
2405 char *buf
, size_t nbytes
, loff_t off
)
2407 return __cgroup_procs_write(of
, buf
, nbytes
, off
, true);
2410 static ssize_t
cgroup_release_agent_write(struct kernfs_open_file
*of
,
2411 char *buf
, size_t nbytes
, loff_t off
)
2413 struct cgroup
*cgrp
;
2415 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
2417 cgrp
= cgroup_kn_lock_live(of
->kn
);
2420 spin_lock(&release_agent_path_lock
);
2421 strlcpy(cgrp
->root
->release_agent_path
, strstrip(buf
),
2422 sizeof(cgrp
->root
->release_agent_path
));
2423 spin_unlock(&release_agent_path_lock
);
2424 cgroup_kn_unlock(of
->kn
);
2428 static int cgroup_release_agent_show(struct seq_file
*seq
, void *v
)
2430 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2432 spin_lock(&release_agent_path_lock
);
2433 seq_puts(seq
, cgrp
->root
->release_agent_path
);
2434 spin_unlock(&release_agent_path_lock
);
2435 seq_putc(seq
, '\n');
2439 static int cgroup_sane_behavior_show(struct seq_file
*seq
, void *v
)
2441 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2443 seq_printf(seq
, "%d\n", cgroup_sane_behavior(cgrp
));
2447 static void cgroup_print_ss_mask(struct seq_file
*seq
, unsigned int ss_mask
)
2449 struct cgroup_subsys
*ss
;
2450 bool printed
= false;
2453 for_each_subsys(ss
, ssid
) {
2454 if (ss_mask
& (1 << ssid
)) {
2457 seq_printf(seq
, "%s", ss
->name
);
2462 seq_putc(seq
, '\n');
2465 /* show controllers which are currently attached to the default hierarchy */
2466 static int cgroup_root_controllers_show(struct seq_file
*seq
, void *v
)
2468 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2470 cgroup_print_ss_mask(seq
, cgrp
->root
->subsys_mask
&
2471 ~cgrp_dfl_root_inhibit_ss_mask
);
2475 /* show controllers which are enabled from the parent */
2476 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2478 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2480 cgroup_print_ss_mask(seq
, cgroup_parent(cgrp
)->child_subsys_mask
);
2484 /* show controllers which are enabled for a given cgroup's children */
2485 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2487 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2489 cgroup_print_ss_mask(seq
, cgrp
->child_subsys_mask
);
2494 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2495 * @cgrp: root of the subtree to update csses for
2497 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2498 * css associations need to be updated accordingly. This function looks up
2499 * all css_sets which are attached to the subtree, creates the matching
2500 * updated css_sets and migrates the tasks to the new ones.
2502 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2504 LIST_HEAD(preloaded_csets
);
2505 struct cgroup_subsys_state
*css
;
2506 struct css_set
*src_cset
;
2509 lockdep_assert_held(&cgroup_mutex
);
2511 /* look up all csses currently attached to @cgrp's subtree */
2512 down_read(&css_set_rwsem
);
2513 css_for_each_descendant_pre(css
, cgroup_css(cgrp
, NULL
)) {
2514 struct cgrp_cset_link
*link
;
2516 /* self is not affected by child_subsys_mask change */
2517 if (css
->cgroup
== cgrp
)
2520 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
)
2521 cgroup_migrate_add_src(link
->cset
, cgrp
,
2524 up_read(&css_set_rwsem
);
2526 /* NULL dst indicates self on default hierarchy */
2527 ret
= cgroup_migrate_prepare_dst(NULL
, &preloaded_csets
);
2531 list_for_each_entry(src_cset
, &preloaded_csets
, mg_preload_node
) {
2532 struct task_struct
*last_task
= NULL
, *task
;
2534 /* src_csets precede dst_csets, break on the first dst_cset */
2535 if (!src_cset
->mg_src_cgrp
)
2539 * All tasks in src_cset need to be migrated to the
2540 * matching dst_cset. Empty it process by process. We
2541 * walk tasks but migrate processes. The leader might even
2542 * belong to a different cset but such src_cset would also
2543 * be among the target src_csets because the default
2544 * hierarchy enforces per-process membership.
2547 down_read(&css_set_rwsem
);
2548 task
= list_first_entry_or_null(&src_cset
->tasks
,
2549 struct task_struct
, cg_list
);
2551 task
= task
->group_leader
;
2552 WARN_ON_ONCE(!task_css_set(task
)->mg_src_cgrp
);
2553 get_task_struct(task
);
2555 up_read(&css_set_rwsem
);
2560 /* guard against possible infinite loop */
2561 if (WARN(last_task
== task
,
2562 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2566 threadgroup_lock(task
);
2567 /* raced against de_thread() from another thread? */
2568 if (!thread_group_leader(task
)) {
2569 threadgroup_unlock(task
);
2570 put_task_struct(task
);
2574 ret
= cgroup_migrate(src_cset
->dfl_cgrp
, task
, true);
2576 threadgroup_unlock(task
);
2577 put_task_struct(task
);
2579 if (WARN(ret
, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret
))
2585 cgroup_migrate_finish(&preloaded_csets
);
2589 /* change the enabled child controllers for a cgroup in the default hierarchy */
2590 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
2591 char *buf
, size_t nbytes
,
2594 unsigned int enable
= 0, disable
= 0;
2595 struct cgroup
*cgrp
, *child
;
2596 struct cgroup_subsys
*ss
;
2601 * Parse input - space separated list of subsystem names prefixed
2602 * with either + or -.
2604 buf
= strstrip(buf
);
2605 while ((tok
= strsep(&buf
, " "))) {
2608 for_each_subsys(ss
, ssid
) {
2609 if (ss
->disabled
|| strcmp(tok
+ 1, ss
->name
) ||
2610 ((1 << ss
->id
) & cgrp_dfl_root_inhibit_ss_mask
))
2614 enable
|= 1 << ssid
;
2615 disable
&= ~(1 << ssid
);
2616 } else if (*tok
== '-') {
2617 disable
|= 1 << ssid
;
2618 enable
&= ~(1 << ssid
);
2624 if (ssid
== CGROUP_SUBSYS_COUNT
)
2628 cgrp
= cgroup_kn_lock_live(of
->kn
);
2632 for_each_subsys(ss
, ssid
) {
2633 if (enable
& (1 << ssid
)) {
2634 if (cgrp
->child_subsys_mask
& (1 << ssid
)) {
2635 enable
&= ~(1 << ssid
);
2640 * Because css offlining is asynchronous, userland
2641 * might try to re-enable the same controller while
2642 * the previous instance is still around. In such
2643 * cases, wait till it's gone using offline_waitq.
2645 cgroup_for_each_live_child(child
, cgrp
) {
2648 if (!cgroup_css(child
, ss
))
2652 prepare_to_wait(&child
->offline_waitq
, &wait
,
2653 TASK_UNINTERRUPTIBLE
);
2654 cgroup_kn_unlock(of
->kn
);
2656 finish_wait(&child
->offline_waitq
, &wait
);
2659 return restart_syscall();
2662 /* unavailable or not enabled on the parent? */
2663 if (!(cgrp_dfl_root
.subsys_mask
& (1 << ssid
)) ||
2664 (cgroup_parent(cgrp
) &&
2665 !(cgroup_parent(cgrp
)->child_subsys_mask
& (1 << ssid
)))) {
2669 } else if (disable
& (1 << ssid
)) {
2670 if (!(cgrp
->child_subsys_mask
& (1 << ssid
))) {
2671 disable
&= ~(1 << ssid
);
2675 /* a child has it enabled? */
2676 cgroup_for_each_live_child(child
, cgrp
) {
2677 if (child
->child_subsys_mask
& (1 << ssid
)) {
2685 if (!enable
&& !disable
) {
2691 * Except for the root, child_subsys_mask must be zero for a cgroup
2692 * with tasks so that child cgroups don't compete against tasks.
2694 if (enable
&& cgroup_parent(cgrp
) && !list_empty(&cgrp
->cset_links
)) {
2700 * Create csses for enables and update child_subsys_mask. This
2701 * changes cgroup_e_css() results which in turn makes the
2702 * subsequent cgroup_update_dfl_csses() associate all tasks in the
2703 * subtree to the updated csses.
2705 for_each_subsys(ss
, ssid
) {
2706 if (!(enable
& (1 << ssid
)))
2709 cgroup_for_each_live_child(child
, cgrp
) {
2710 ret
= create_css(child
, ss
);
2716 cgrp
->child_subsys_mask
|= enable
;
2717 cgrp
->child_subsys_mask
&= ~disable
;
2719 ret
= cgroup_update_dfl_csses(cgrp
);
2723 /* all tasks are now migrated away from the old csses, kill them */
2724 for_each_subsys(ss
, ssid
) {
2725 if (!(disable
& (1 << ssid
)))
2728 cgroup_for_each_live_child(child
, cgrp
)
2729 kill_css(cgroup_css(child
, ss
));
2732 kernfs_activate(cgrp
->kn
);
2735 cgroup_kn_unlock(of
->kn
);
2736 return ret
?: nbytes
;
2739 cgrp
->child_subsys_mask
&= ~enable
;
2740 cgrp
->child_subsys_mask
|= disable
;
2742 for_each_subsys(ss
, ssid
) {
2743 if (!(enable
& (1 << ssid
)))
2746 cgroup_for_each_live_child(child
, cgrp
) {
2747 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
2755 static int cgroup_populated_show(struct seq_file
*seq
, void *v
)
2757 seq_printf(seq
, "%d\n", (bool)seq_css(seq
)->cgroup
->populated_cnt
);
2761 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
2762 size_t nbytes
, loff_t off
)
2764 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
2765 struct cftype
*cft
= of
->kn
->priv
;
2766 struct cgroup_subsys_state
*css
;
2770 return cft
->write(of
, buf
, nbytes
, off
);
2773 * kernfs guarantees that a file isn't deleted with operations in
2774 * flight, which means that the matching css is and stays alive and
2775 * doesn't need to be pinned. The RCU locking is not necessary
2776 * either. It's just for the convenience of using cgroup_css().
2779 css
= cgroup_css(cgrp
, cft
->ss
);
2782 if (cft
->write_u64
) {
2783 unsigned long long v
;
2784 ret
= kstrtoull(buf
, 0, &v
);
2786 ret
= cft
->write_u64(css
, cft
, v
);
2787 } else if (cft
->write_s64
) {
2789 ret
= kstrtoll(buf
, 0, &v
);
2791 ret
= cft
->write_s64(css
, cft
, v
);
2796 return ret
?: nbytes
;
2799 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
2801 return seq_cft(seq
)->seq_start(seq
, ppos
);
2804 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
2806 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
2809 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
2811 seq_cft(seq
)->seq_stop(seq
, v
);
2814 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
2816 struct cftype
*cft
= seq_cft(m
);
2817 struct cgroup_subsys_state
*css
= seq_css(m
);
2820 return cft
->seq_show(m
, arg
);
2823 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
2824 else if (cft
->read_s64
)
2825 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
2831 static struct kernfs_ops cgroup_kf_single_ops
= {
2832 .atomic_write_len
= PAGE_SIZE
,
2833 .write
= cgroup_file_write
,
2834 .seq_show
= cgroup_seqfile_show
,
2837 static struct kernfs_ops cgroup_kf_ops
= {
2838 .atomic_write_len
= PAGE_SIZE
,
2839 .write
= cgroup_file_write
,
2840 .seq_start
= cgroup_seqfile_start
,
2841 .seq_next
= cgroup_seqfile_next
,
2842 .seq_stop
= cgroup_seqfile_stop
,
2843 .seq_show
= cgroup_seqfile_show
,
2847 * cgroup_rename - Only allow simple rename of directories in place.
2849 static int cgroup_rename(struct kernfs_node
*kn
, struct kernfs_node
*new_parent
,
2850 const char *new_name_str
)
2852 struct cgroup
*cgrp
= kn
->priv
;
2855 if (kernfs_type(kn
) != KERNFS_DIR
)
2857 if (kn
->parent
!= new_parent
)
2861 * This isn't a proper migration and its usefulness is very
2862 * limited. Disallow if sane_behavior.
2864 if (cgroup_sane_behavior(cgrp
))
2868 * We're gonna grab cgroup_mutex which nests outside kernfs
2869 * active_ref. kernfs_rename() doesn't require active_ref
2870 * protection. Break them before grabbing cgroup_mutex.
2872 kernfs_break_active_protection(new_parent
);
2873 kernfs_break_active_protection(kn
);
2875 mutex_lock(&cgroup_mutex
);
2877 ret
= kernfs_rename(kn
, new_parent
, new_name_str
);
2879 mutex_unlock(&cgroup_mutex
);
2881 kernfs_unbreak_active_protection(kn
);
2882 kernfs_unbreak_active_protection(new_parent
);
2886 /* set uid and gid of cgroup dirs and files to that of the creator */
2887 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
2889 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
2890 .ia_uid
= current_fsuid(),
2891 .ia_gid
= current_fsgid(), };
2893 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
2894 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
2897 return kernfs_setattr(kn
, &iattr
);
2900 static int cgroup_add_file(struct cgroup
*cgrp
, struct cftype
*cft
)
2902 char name
[CGROUP_FILE_NAME_MAX
];
2903 struct kernfs_node
*kn
;
2904 struct lock_class_key
*key
= NULL
;
2907 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2908 key
= &cft
->lockdep_key
;
2910 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
2911 cgroup_file_mode(cft
), 0, cft
->kf_ops
, cft
,
2916 ret
= cgroup_kn_set_ugid(kn
);
2922 if (cft
->seq_show
== cgroup_populated_show
)
2923 cgrp
->populated_kn
= kn
;
2928 * cgroup_addrm_files - add or remove files to a cgroup directory
2929 * @cgrp: the target cgroup
2930 * @cfts: array of cftypes to be added
2931 * @is_add: whether to add or remove
2933 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2934 * For removals, this function never fails. If addition fails, this
2935 * function doesn't remove files already added. The caller is responsible
2938 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
2944 lockdep_assert_held(&cgroup_mutex
);
2946 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
2947 /* does cft->flags tell us to skip this file on @cgrp? */
2948 if ((cft
->flags
& CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
2950 if ((cft
->flags
& CFTYPE_INSANE
) && cgroup_sane_behavior(cgrp
))
2952 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
2954 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
2958 ret
= cgroup_add_file(cgrp
, cft
);
2960 pr_warn("%s: failed to add %s, err=%d\n",
2961 __func__
, cft
->name
, ret
);
2965 cgroup_rm_file(cgrp
, cft
);
2971 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
2974 struct cgroup_subsys
*ss
= cfts
[0].ss
;
2975 struct cgroup
*root
= &ss
->root
->cgrp
;
2976 struct cgroup_subsys_state
*css
;
2979 lockdep_assert_held(&cgroup_mutex
);
2981 /* add/rm files for all cgroups created before */
2982 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
2983 struct cgroup
*cgrp
= css
->cgroup
;
2985 if (cgroup_is_dead(cgrp
))
2988 ret
= cgroup_addrm_files(cgrp
, cfts
, is_add
);
2994 kernfs_activate(root
->kn
);
2998 static void cgroup_exit_cftypes(struct cftype
*cfts
)
3002 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3003 /* free copy for custom atomic_write_len, see init_cftypes() */
3004 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
3011 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3015 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3016 struct kernfs_ops
*kf_ops
;
3018 WARN_ON(cft
->ss
|| cft
->kf_ops
);
3021 kf_ops
= &cgroup_kf_ops
;
3023 kf_ops
= &cgroup_kf_single_ops
;
3026 * Ugh... if @cft wants a custom max_write_len, we need to
3027 * make a copy of kf_ops to set its atomic_write_len.
3029 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
3030 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
3032 cgroup_exit_cftypes(cfts
);
3035 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3038 cft
->kf_ops
= kf_ops
;
3045 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3047 lockdep_assert_held(&cgroup_mutex
);
3049 if (!cfts
|| !cfts
[0].ss
)
3052 list_del(&cfts
->node
);
3053 cgroup_apply_cftypes(cfts
, false);
3054 cgroup_exit_cftypes(cfts
);
3059 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3060 * @cfts: zero-length name terminated array of cftypes
3062 * Unregister @cfts. Files described by @cfts are removed from all
3063 * existing cgroups and all future cgroups won't have them either. This
3064 * function can be called anytime whether @cfts' subsys is attached or not.
3066 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3069 int cgroup_rm_cftypes(struct cftype
*cfts
)
3073 mutex_lock(&cgroup_mutex
);
3074 ret
= cgroup_rm_cftypes_locked(cfts
);
3075 mutex_unlock(&cgroup_mutex
);
3080 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3081 * @ss: target cgroup subsystem
3082 * @cfts: zero-length name terminated array of cftypes
3084 * Register @cfts to @ss. Files described by @cfts are created for all
3085 * existing cgroups to which @ss is attached and all future cgroups will
3086 * have them too. This function can be called anytime whether @ss is
3089 * Returns 0 on successful registration, -errno on failure. Note that this
3090 * function currently returns 0 as long as @cfts registration is successful
3091 * even if some file creation attempts on existing cgroups fail.
3093 int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3100 if (!cfts
|| cfts
[0].name
[0] == '\0')
3103 ret
= cgroup_init_cftypes(ss
, cfts
);
3107 mutex_lock(&cgroup_mutex
);
3109 list_add_tail(&cfts
->node
, &ss
->cfts
);
3110 ret
= cgroup_apply_cftypes(cfts
, true);
3112 cgroup_rm_cftypes_locked(cfts
);
3114 mutex_unlock(&cgroup_mutex
);
3119 * cgroup_task_count - count the number of tasks in a cgroup.
3120 * @cgrp: the cgroup in question
3122 * Return the number of tasks in the cgroup.
3124 static int cgroup_task_count(const struct cgroup
*cgrp
)
3127 struct cgrp_cset_link
*link
;
3129 down_read(&css_set_rwsem
);
3130 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
3131 count
+= atomic_read(&link
->cset
->refcount
);
3132 up_read(&css_set_rwsem
);
3137 * css_next_child - find the next child of a given css
3138 * @pos: the current position (%NULL to initiate traversal)
3139 * @parent: css whose children to walk
3141 * This function returns the next child of @parent and should be called
3142 * under either cgroup_mutex or RCU read lock. The only requirement is
3143 * that @parent and @pos are accessible. The next sibling is guaranteed to
3144 * be returned regardless of their states.
3146 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3147 * css which finished ->css_online() is guaranteed to be visible in the
3148 * future iterations and will stay visible until the last reference is put.
3149 * A css which hasn't finished ->css_online() or already finished
3150 * ->css_offline() may show up during traversal. It's each subsystem's
3151 * responsibility to synchronize against on/offlining.
3153 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
3154 struct cgroup_subsys_state
*parent
)
3156 struct cgroup_subsys_state
*next
;
3158 cgroup_assert_mutex_or_rcu_locked();
3161 * @pos could already have been unlinked from the sibling list.
3162 * Once a cgroup is removed, its ->sibling.next is no longer
3163 * updated when its next sibling changes. CSS_RELEASED is set when
3164 * @pos is taken off list, at which time its next pointer is valid,
3165 * and, as releases are serialized, the one pointed to by the next
3166 * pointer is guaranteed to not have started release yet. This
3167 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3168 * critical section, the one pointed to by its next pointer is
3169 * guaranteed to not have finished its RCU grace period even if we
3170 * have dropped rcu_read_lock() inbetween iterations.
3172 * If @pos has CSS_RELEASED set, its next pointer can't be
3173 * dereferenced; however, as each css is given a monotonically
3174 * increasing unique serial number and always appended to the
3175 * sibling list, the next one can be found by walking the parent's
3176 * children until the first css with higher serial number than
3177 * @pos's. While this path can be slower, it happens iff iteration
3178 * races against release and the race window is very small.
3181 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
3182 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
3183 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
3185 list_for_each_entry_rcu(next
, &parent
->children
, sibling
)
3186 if (next
->serial_nr
> pos
->serial_nr
)
3191 * @next, if not pointing to the head, can be dereferenced and is
3194 if (&next
->sibling
!= &parent
->children
)
3200 * css_next_descendant_pre - find the next descendant for pre-order walk
3201 * @pos: the current position (%NULL to initiate traversal)
3202 * @root: css whose descendants to walk
3204 * To be used by css_for_each_descendant_pre(). Find the next descendant
3205 * to visit for pre-order traversal of @root's descendants. @root is
3206 * included in the iteration and the first node to be visited.
3208 * While this function requires cgroup_mutex or RCU read locking, it
3209 * doesn't require the whole traversal to be contained in a single critical
3210 * section. This function will return the correct next descendant as long
3211 * as both @pos and @root are accessible and @pos is a descendant of @root.
3213 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3214 * css which finished ->css_online() is guaranteed to be visible in the
3215 * future iterations and will stay visible until the last reference is put.
3216 * A css which hasn't finished ->css_online() or already finished
3217 * ->css_offline() may show up during traversal. It's each subsystem's
3218 * responsibility to synchronize against on/offlining.
3220 struct cgroup_subsys_state
*
3221 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3222 struct cgroup_subsys_state
*root
)
3224 struct cgroup_subsys_state
*next
;
3226 cgroup_assert_mutex_or_rcu_locked();
3228 /* if first iteration, visit @root */
3232 /* visit the first child if exists */
3233 next
= css_next_child(NULL
, pos
);
3237 /* no child, visit my or the closest ancestor's next sibling */
3238 while (pos
!= root
) {
3239 next
= css_next_child(pos
, pos
->parent
);
3249 * css_rightmost_descendant - return the rightmost descendant of a css
3250 * @pos: css of interest
3252 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3253 * is returned. This can be used during pre-order traversal to skip
3256 * While this function requires cgroup_mutex or RCU read locking, it
3257 * doesn't require the whole traversal to be contained in a single critical
3258 * section. This function will return the correct rightmost descendant as
3259 * long as @pos is accessible.
3261 struct cgroup_subsys_state
*
3262 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3264 struct cgroup_subsys_state
*last
, *tmp
;
3266 cgroup_assert_mutex_or_rcu_locked();
3270 /* ->prev isn't RCU safe, walk ->next till the end */
3272 css_for_each_child(tmp
, last
)
3279 static struct cgroup_subsys_state
*
3280 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3282 struct cgroup_subsys_state
*last
;
3286 pos
= css_next_child(NULL
, pos
);
3293 * css_next_descendant_post - find the next descendant for post-order walk
3294 * @pos: the current position (%NULL to initiate traversal)
3295 * @root: css whose descendants to walk
3297 * To be used by css_for_each_descendant_post(). Find the next descendant
3298 * to visit for post-order traversal of @root's descendants. @root is
3299 * included in the iteration and the last node to be visited.
3301 * While this function requires cgroup_mutex or RCU read locking, it
3302 * doesn't require the whole traversal to be contained in a single critical
3303 * section. This function will return the correct next descendant as long
3304 * as both @pos and @cgroup are accessible and @pos is a descendant of
3307 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3308 * css which finished ->css_online() is guaranteed to be visible in the
3309 * future iterations and will stay visible until the last reference is put.
3310 * A css which hasn't finished ->css_online() or already finished
3311 * ->css_offline() may show up during traversal. It's each subsystem's
3312 * responsibility to synchronize against on/offlining.
3314 struct cgroup_subsys_state
*
3315 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
3316 struct cgroup_subsys_state
*root
)
3318 struct cgroup_subsys_state
*next
;
3320 cgroup_assert_mutex_or_rcu_locked();
3322 /* if first iteration, visit leftmost descendant which may be @root */
3324 return css_leftmost_descendant(root
);
3326 /* if we visited @root, we're done */
3330 /* if there's an unvisited sibling, visit its leftmost descendant */
3331 next
= css_next_child(pos
, pos
->parent
);
3333 return css_leftmost_descendant(next
);
3335 /* no sibling left, visit parent */
3340 * css_has_online_children - does a css have online children
3341 * @css: the target css
3343 * Returns %true if @css has any online children; otherwise, %false. This
3344 * function can be called from any context but the caller is responsible
3345 * for synchronizing against on/offlining as necessary.
3347 bool css_has_online_children(struct cgroup_subsys_state
*css
)
3349 struct cgroup_subsys_state
*child
;
3353 css_for_each_child(child
, css
) {
3354 if (child
->flags
& CSS_ONLINE
) {
3364 * css_advance_task_iter - advance a task itererator to the next css_set
3365 * @it: the iterator to advance
3367 * Advance @it to the next css_set to walk.
3369 static void css_advance_task_iter(struct css_task_iter
*it
)
3371 struct list_head
*l
= it
->cset_pos
;
3372 struct cgrp_cset_link
*link
;
3373 struct css_set
*cset
;
3375 /* Advance to the next non-empty css_set */
3378 if (l
== it
->cset_head
) {
3379 it
->cset_pos
= NULL
;
3384 cset
= container_of(l
, struct css_set
,
3385 e_cset_node
[it
->ss
->id
]);
3387 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
3390 } while (list_empty(&cset
->tasks
) && list_empty(&cset
->mg_tasks
));
3394 if (!list_empty(&cset
->tasks
))
3395 it
->task_pos
= cset
->tasks
.next
;
3397 it
->task_pos
= cset
->mg_tasks
.next
;
3399 it
->tasks_head
= &cset
->tasks
;
3400 it
->mg_tasks_head
= &cset
->mg_tasks
;
3404 * css_task_iter_start - initiate task iteration
3405 * @css: the css to walk tasks of
3406 * @it: the task iterator to use
3408 * Initiate iteration through the tasks of @css. The caller can call
3409 * css_task_iter_next() to walk through the tasks until the function
3410 * returns NULL. On completion of iteration, css_task_iter_end() must be
3413 * Note that this function acquires a lock which is released when the
3414 * iteration finishes. The caller can't sleep while iteration is in
3417 void css_task_iter_start(struct cgroup_subsys_state
*css
,
3418 struct css_task_iter
*it
)
3419 __acquires(css_set_rwsem
)
3421 /* no one should try to iterate before mounting cgroups */
3422 WARN_ON_ONCE(!use_task_css_set_links
);
3424 down_read(&css_set_rwsem
);
3429 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
3431 it
->cset_pos
= &css
->cgroup
->cset_links
;
3433 it
->cset_head
= it
->cset_pos
;
3435 css_advance_task_iter(it
);
3439 * css_task_iter_next - return the next task for the iterator
3440 * @it: the task iterator being iterated
3442 * The "next" function for task iteration. @it should have been
3443 * initialized via css_task_iter_start(). Returns NULL when the iteration
3446 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
3448 struct task_struct
*res
;
3449 struct list_head
*l
= it
->task_pos
;
3451 /* If the iterator cg is NULL, we have no tasks */
3454 res
= list_entry(l
, struct task_struct
, cg_list
);
3457 * Advance iterator to find next entry. cset->tasks is consumed
3458 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3463 if (l
== it
->tasks_head
)
3464 l
= it
->mg_tasks_head
->next
;
3466 if (l
== it
->mg_tasks_head
)
3467 css_advance_task_iter(it
);
3475 * css_task_iter_end - finish task iteration
3476 * @it: the task iterator to finish
3478 * Finish task iteration started by css_task_iter_start().
3480 void css_task_iter_end(struct css_task_iter
*it
)
3481 __releases(css_set_rwsem
)
3483 up_read(&css_set_rwsem
);
3487 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3488 * @to: cgroup to which the tasks will be moved
3489 * @from: cgroup in which the tasks currently reside
3491 * Locking rules between cgroup_post_fork() and the migration path
3492 * guarantee that, if a task is forking while being migrated, the new child
3493 * is guaranteed to be either visible in the source cgroup after the
3494 * parent's migration is complete or put into the target cgroup. No task
3495 * can slip out of migration through forking.
3497 int cgroup_transfer_tasks(struct cgroup
*to
, struct cgroup
*from
)
3499 LIST_HEAD(preloaded_csets
);
3500 struct cgrp_cset_link
*link
;
3501 struct css_task_iter it
;
3502 struct task_struct
*task
;
3505 mutex_lock(&cgroup_mutex
);
3507 /* all tasks in @from are being moved, all csets are source */
3508 down_read(&css_set_rwsem
);
3509 list_for_each_entry(link
, &from
->cset_links
, cset_link
)
3510 cgroup_migrate_add_src(link
->cset
, to
, &preloaded_csets
);
3511 up_read(&css_set_rwsem
);
3513 ret
= cgroup_migrate_prepare_dst(to
, &preloaded_csets
);
3518 * Migrate tasks one-by-one until @form is empty. This fails iff
3519 * ->can_attach() fails.
3522 css_task_iter_start(&from
->self
, &it
);
3523 task
= css_task_iter_next(&it
);
3525 get_task_struct(task
);
3526 css_task_iter_end(&it
);
3529 ret
= cgroup_migrate(to
, task
, false);
3530 put_task_struct(task
);
3532 } while (task
&& !ret
);
3534 cgroup_migrate_finish(&preloaded_csets
);
3535 mutex_unlock(&cgroup_mutex
);
3540 * Stuff for reading the 'tasks'/'procs' files.
3542 * Reading this file can return large amounts of data if a cgroup has
3543 * *lots* of attached tasks. So it may need several calls to read(),
3544 * but we cannot guarantee that the information we produce is correct
3545 * unless we produce it entirely atomically.
3549 /* which pidlist file are we talking about? */
3550 enum cgroup_filetype
{
3556 * A pidlist is a list of pids that virtually represents the contents of one
3557 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3558 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3561 struct cgroup_pidlist
{
3563 * used to find which pidlist is wanted. doesn't change as long as
3564 * this particular list stays in the list.
3566 struct { enum cgroup_filetype type
; struct pid_namespace
*ns
; } key
;
3569 /* how many elements the above list has */
3571 /* each of these stored in a list by its cgroup */
3572 struct list_head links
;
3573 /* pointer to the cgroup we belong to, for list removal purposes */
3574 struct cgroup
*owner
;
3575 /* for delayed destruction */
3576 struct delayed_work destroy_dwork
;
3580 * The following two functions "fix" the issue where there are more pids
3581 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3582 * TODO: replace with a kernel-wide solution to this problem
3584 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3585 static void *pidlist_allocate(int count
)
3587 if (PIDLIST_TOO_LARGE(count
))
3588 return vmalloc(count
* sizeof(pid_t
));
3590 return kmalloc(count
* sizeof(pid_t
), GFP_KERNEL
);
3593 static void pidlist_free(void *p
)
3595 if (is_vmalloc_addr(p
))
3602 * Used to destroy all pidlists lingering waiting for destroy timer. None
3603 * should be left afterwards.
3605 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
)
3607 struct cgroup_pidlist
*l
, *tmp_l
;
3609 mutex_lock(&cgrp
->pidlist_mutex
);
3610 list_for_each_entry_safe(l
, tmp_l
, &cgrp
->pidlists
, links
)
3611 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
, 0);
3612 mutex_unlock(&cgrp
->pidlist_mutex
);
3614 flush_workqueue(cgroup_pidlist_destroy_wq
);
3615 BUG_ON(!list_empty(&cgrp
->pidlists
));
3618 static void cgroup_pidlist_destroy_work_fn(struct work_struct
*work
)
3620 struct delayed_work
*dwork
= to_delayed_work(work
);
3621 struct cgroup_pidlist
*l
= container_of(dwork
, struct cgroup_pidlist
,
3623 struct cgroup_pidlist
*tofree
= NULL
;
3625 mutex_lock(&l
->owner
->pidlist_mutex
);
3628 * Destroy iff we didn't get queued again. The state won't change
3629 * as destroy_dwork can only be queued while locked.
3631 if (!delayed_work_pending(dwork
)) {
3632 list_del(&l
->links
);
3633 pidlist_free(l
->list
);
3634 put_pid_ns(l
->key
.ns
);
3638 mutex_unlock(&l
->owner
->pidlist_mutex
);
3643 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3644 * Returns the number of unique elements.
3646 static int pidlist_uniq(pid_t
*list
, int length
)
3651 * we presume the 0th element is unique, so i starts at 1. trivial
3652 * edge cases first; no work needs to be done for either
3654 if (length
== 0 || length
== 1)
3656 /* src and dest walk down the list; dest counts unique elements */
3657 for (src
= 1; src
< length
; src
++) {
3658 /* find next unique element */
3659 while (list
[src
] == list
[src
-1]) {
3664 /* dest always points to where the next unique element goes */
3665 list
[dest
] = list
[src
];
3673 * The two pid files - task and cgroup.procs - guaranteed that the result
3674 * is sorted, which forced this whole pidlist fiasco. As pid order is
3675 * different per namespace, each namespace needs differently sorted list,
3676 * making it impossible to use, for example, single rbtree of member tasks
3677 * sorted by task pointer. As pidlists can be fairly large, allocating one
3678 * per open file is dangerous, so cgroup had to implement shared pool of
3679 * pidlists keyed by cgroup and namespace.
3681 * All this extra complexity was caused by the original implementation
3682 * committing to an entirely unnecessary property. In the long term, we
3683 * want to do away with it. Explicitly scramble sort order if
3684 * sane_behavior so that no such expectation exists in the new interface.
3686 * Scrambling is done by swapping every two consecutive bits, which is
3687 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3689 static pid_t
pid_fry(pid_t pid
)
3691 unsigned a
= pid
& 0x55555555;
3692 unsigned b
= pid
& 0xAAAAAAAA;
3694 return (a
<< 1) | (b
>> 1);
3697 static pid_t
cgroup_pid_fry(struct cgroup
*cgrp
, pid_t pid
)
3699 if (cgroup_sane_behavior(cgrp
))
3700 return pid_fry(pid
);
3705 static int cmppid(const void *a
, const void *b
)
3707 return *(pid_t
*)a
- *(pid_t
*)b
;
3710 static int fried_cmppid(const void *a
, const void *b
)
3712 return pid_fry(*(pid_t
*)a
) - pid_fry(*(pid_t
*)b
);
3715 static struct cgroup_pidlist
*cgroup_pidlist_find(struct cgroup
*cgrp
,
3716 enum cgroup_filetype type
)
3718 struct cgroup_pidlist
*l
;
3719 /* don't need task_nsproxy() if we're looking at ourself */
3720 struct pid_namespace
*ns
= task_active_pid_ns(current
);
3722 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3724 list_for_each_entry(l
, &cgrp
->pidlists
, links
)
3725 if (l
->key
.type
== type
&& l
->key
.ns
== ns
)
3731 * find the appropriate pidlist for our purpose (given procs vs tasks)
3732 * returns with the lock on that pidlist already held, and takes care
3733 * of the use count, or returns NULL with no locks held if we're out of
3736 static struct cgroup_pidlist
*cgroup_pidlist_find_create(struct cgroup
*cgrp
,
3737 enum cgroup_filetype type
)
3739 struct cgroup_pidlist
*l
;
3741 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3743 l
= cgroup_pidlist_find(cgrp
, type
);
3747 /* entry not found; create a new one */
3748 l
= kzalloc(sizeof(struct cgroup_pidlist
), GFP_KERNEL
);
3752 INIT_DELAYED_WORK(&l
->destroy_dwork
, cgroup_pidlist_destroy_work_fn
);
3754 /* don't need task_nsproxy() if we're looking at ourself */
3755 l
->key
.ns
= get_pid_ns(task_active_pid_ns(current
));
3757 list_add(&l
->links
, &cgrp
->pidlists
);
3762 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3764 static int pidlist_array_load(struct cgroup
*cgrp
, enum cgroup_filetype type
,
3765 struct cgroup_pidlist
**lp
)
3769 int pid
, n
= 0; /* used for populating the array */
3770 struct css_task_iter it
;
3771 struct task_struct
*tsk
;
3772 struct cgroup_pidlist
*l
;
3774 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3777 * If cgroup gets more users after we read count, we won't have
3778 * enough space - tough. This race is indistinguishable to the
3779 * caller from the case that the additional cgroup users didn't
3780 * show up until sometime later on.
3782 length
= cgroup_task_count(cgrp
);
3783 array
= pidlist_allocate(length
);
3786 /* now, populate the array */
3787 css_task_iter_start(&cgrp
->self
, &it
);
3788 while ((tsk
= css_task_iter_next(&it
))) {
3789 if (unlikely(n
== length
))
3791 /* get tgid or pid for procs or tasks file respectively */
3792 if (type
== CGROUP_FILE_PROCS
)
3793 pid
= task_tgid_vnr(tsk
);
3795 pid
= task_pid_vnr(tsk
);
3796 if (pid
> 0) /* make sure to only use valid results */
3799 css_task_iter_end(&it
);
3801 /* now sort & (if procs) strip out duplicates */
3802 if (cgroup_sane_behavior(cgrp
))
3803 sort(array
, length
, sizeof(pid_t
), fried_cmppid
, NULL
);
3805 sort(array
, length
, sizeof(pid_t
), cmppid
, NULL
);
3806 if (type
== CGROUP_FILE_PROCS
)
3807 length
= pidlist_uniq(array
, length
);
3809 l
= cgroup_pidlist_find_create(cgrp
, type
);
3811 mutex_unlock(&cgrp
->pidlist_mutex
);
3812 pidlist_free(array
);
3816 /* store array, freeing old if necessary */
3817 pidlist_free(l
->list
);
3825 * cgroupstats_build - build and fill cgroupstats
3826 * @stats: cgroupstats to fill information into
3827 * @dentry: A dentry entry belonging to the cgroup for which stats have
3830 * Build and fill cgroupstats so that taskstats can export it to user
3833 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
3835 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
3836 struct cgroup
*cgrp
;
3837 struct css_task_iter it
;
3838 struct task_struct
*tsk
;
3840 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3841 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
3842 kernfs_type(kn
) != KERNFS_DIR
)
3845 mutex_lock(&cgroup_mutex
);
3848 * We aren't being called from kernfs and there's no guarantee on
3849 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
3850 * @kn->priv is RCU safe. Let's do the RCU dancing.
3853 cgrp
= rcu_dereference(kn
->priv
);
3854 if (!cgrp
|| cgroup_is_dead(cgrp
)) {
3856 mutex_unlock(&cgroup_mutex
);
3861 css_task_iter_start(&cgrp
->self
, &it
);
3862 while ((tsk
= css_task_iter_next(&it
))) {
3863 switch (tsk
->state
) {
3865 stats
->nr_running
++;
3867 case TASK_INTERRUPTIBLE
:
3868 stats
->nr_sleeping
++;
3870 case TASK_UNINTERRUPTIBLE
:
3871 stats
->nr_uninterruptible
++;
3874 stats
->nr_stopped
++;
3877 if (delayacct_is_task_waiting_on_io(tsk
))
3878 stats
->nr_io_wait
++;
3882 css_task_iter_end(&it
);
3884 mutex_unlock(&cgroup_mutex
);
3890 * seq_file methods for the tasks/procs files. The seq_file position is the
3891 * next pid to display; the seq_file iterator is a pointer to the pid
3892 * in the cgroup->l->list array.
3895 static void *cgroup_pidlist_start(struct seq_file
*s
, loff_t
*pos
)
3898 * Initially we receive a position value that corresponds to
3899 * one more than the last pid shown (or 0 on the first call or
3900 * after a seek to the start). Use a binary-search to find the
3901 * next pid to display, if any
3903 struct kernfs_open_file
*of
= s
->private;
3904 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
3905 struct cgroup_pidlist
*l
;
3906 enum cgroup_filetype type
= seq_cft(s
)->private;
3907 int index
= 0, pid
= *pos
;
3910 mutex_lock(&cgrp
->pidlist_mutex
);
3913 * !NULL @of->priv indicates that this isn't the first start()
3914 * after open. If the matching pidlist is around, we can use that.
3915 * Look for it. Note that @of->priv can't be used directly. It
3916 * could already have been destroyed.
3919 of
->priv
= cgroup_pidlist_find(cgrp
, type
);
3922 * Either this is the first start() after open or the matching
3923 * pidlist has been destroyed inbetween. Create a new one.
3926 ret
= pidlist_array_load(cgrp
, type
,
3927 (struct cgroup_pidlist
**)&of
->priv
);
3929 return ERR_PTR(ret
);
3934 int end
= l
->length
;
3936 while (index
< end
) {
3937 int mid
= (index
+ end
) / 2;
3938 if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) == pid
) {
3941 } else if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) <= pid
)
3947 /* If we're off the end of the array, we're done */
3948 if (index
>= l
->length
)
3950 /* Update the abstract position to be the actual pid that we found */
3951 iter
= l
->list
+ index
;
3952 *pos
= cgroup_pid_fry(cgrp
, *iter
);
3956 static void cgroup_pidlist_stop(struct seq_file
*s
, void *v
)
3958 struct kernfs_open_file
*of
= s
->private;
3959 struct cgroup_pidlist
*l
= of
->priv
;
3962 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
,
3963 CGROUP_PIDLIST_DESTROY_DELAY
);
3964 mutex_unlock(&seq_css(s
)->cgroup
->pidlist_mutex
);
3967 static void *cgroup_pidlist_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
3969 struct kernfs_open_file
*of
= s
->private;
3970 struct cgroup_pidlist
*l
= of
->priv
;
3972 pid_t
*end
= l
->list
+ l
->length
;
3974 * Advance to the next pid in the array. If this goes off the
3981 *pos
= cgroup_pid_fry(seq_css(s
)->cgroup
, *p
);
3986 static int cgroup_pidlist_show(struct seq_file
*s
, void *v
)
3988 return seq_printf(s
, "%d\n", *(int *)v
);
3991 static u64
cgroup_read_notify_on_release(struct cgroup_subsys_state
*css
,
3994 return notify_on_release(css
->cgroup
);
3997 static int cgroup_write_notify_on_release(struct cgroup_subsys_state
*css
,
3998 struct cftype
*cft
, u64 val
)
4000 clear_bit(CGRP_RELEASABLE
, &css
->cgroup
->flags
);
4002 set_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4004 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4008 static u64
cgroup_clone_children_read(struct cgroup_subsys_state
*css
,
4011 return test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4014 static int cgroup_clone_children_write(struct cgroup_subsys_state
*css
,
4015 struct cftype
*cft
, u64 val
)
4018 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4020 clear_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4024 static struct cftype cgroup_base_files
[] = {
4026 .name
= "cgroup.procs",
4027 .seq_start
= cgroup_pidlist_start
,
4028 .seq_next
= cgroup_pidlist_next
,
4029 .seq_stop
= cgroup_pidlist_stop
,
4030 .seq_show
= cgroup_pidlist_show
,
4031 .private = CGROUP_FILE_PROCS
,
4032 .write
= cgroup_procs_write
,
4033 .mode
= S_IRUGO
| S_IWUSR
,
4036 .name
= "cgroup.clone_children",
4037 .flags
= CFTYPE_INSANE
,
4038 .read_u64
= cgroup_clone_children_read
,
4039 .write_u64
= cgroup_clone_children_write
,
4042 .name
= "cgroup.sane_behavior",
4043 .flags
= CFTYPE_ONLY_ON_ROOT
,
4044 .seq_show
= cgroup_sane_behavior_show
,
4047 .name
= "cgroup.controllers",
4048 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_ONLY_ON_ROOT
,
4049 .seq_show
= cgroup_root_controllers_show
,
4052 .name
= "cgroup.controllers",
4053 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_NOT_ON_ROOT
,
4054 .seq_show
= cgroup_controllers_show
,
4057 .name
= "cgroup.subtree_control",
4058 .flags
= CFTYPE_ONLY_ON_DFL
,
4059 .seq_show
= cgroup_subtree_control_show
,
4060 .write
= cgroup_subtree_control_write
,
4063 .name
= "cgroup.populated",
4064 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_NOT_ON_ROOT
,
4065 .seq_show
= cgroup_populated_show
,
4069 * Historical crazy stuff. These don't have "cgroup." prefix and
4070 * don't exist if sane_behavior. If you're depending on these, be
4071 * prepared to be burned.
4075 .flags
= CFTYPE_INSANE
, /* use "procs" instead */
4076 .seq_start
= cgroup_pidlist_start
,
4077 .seq_next
= cgroup_pidlist_next
,
4078 .seq_stop
= cgroup_pidlist_stop
,
4079 .seq_show
= cgroup_pidlist_show
,
4080 .private = CGROUP_FILE_TASKS
,
4081 .write
= cgroup_tasks_write
,
4082 .mode
= S_IRUGO
| S_IWUSR
,
4085 .name
= "notify_on_release",
4086 .flags
= CFTYPE_INSANE
,
4087 .read_u64
= cgroup_read_notify_on_release
,
4088 .write_u64
= cgroup_write_notify_on_release
,
4091 .name
= "release_agent",
4092 .flags
= CFTYPE_INSANE
| CFTYPE_ONLY_ON_ROOT
,
4093 .seq_show
= cgroup_release_agent_show
,
4094 .write
= cgroup_release_agent_write
,
4095 .max_write_len
= PATH_MAX
- 1,
4101 * cgroup_populate_dir - create subsys files in a cgroup directory
4102 * @cgrp: target cgroup
4103 * @subsys_mask: mask of the subsystem ids whose files should be added
4105 * On failure, no file is added.
4107 static int cgroup_populate_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
)
4109 struct cgroup_subsys
*ss
;
4112 /* process cftsets of each subsystem */
4113 for_each_subsys(ss
, i
) {
4114 struct cftype
*cfts
;
4116 if (!(subsys_mask
& (1 << i
)))
4119 list_for_each_entry(cfts
, &ss
->cfts
, node
) {
4120 ret
= cgroup_addrm_files(cgrp
, cfts
, true);
4127 cgroup_clear_dir(cgrp
, subsys_mask
);
4132 * css destruction is four-stage process.
4134 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4135 * Implemented in kill_css().
4137 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4138 * and thus css_tryget_online() is guaranteed to fail, the css can be
4139 * offlined by invoking offline_css(). After offlining, the base ref is
4140 * put. Implemented in css_killed_work_fn().
4142 * 3. When the percpu_ref reaches zero, the only possible remaining
4143 * accessors are inside RCU read sections. css_release() schedules the
4146 * 4. After the grace period, the css can be freed. Implemented in
4147 * css_free_work_fn().
4149 * It is actually hairier because both step 2 and 4 require process context
4150 * and thus involve punting to css->destroy_work adding two additional
4151 * steps to the already complex sequence.
4153 static void css_free_work_fn(struct work_struct
*work
)
4155 struct cgroup_subsys_state
*css
=
4156 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4157 struct cgroup
*cgrp
= css
->cgroup
;
4162 css_put(css
->parent
);
4164 css
->ss
->css_free(css
);
4167 /* cgroup free path */
4168 atomic_dec(&cgrp
->root
->nr_cgrps
);
4169 cgroup_pidlist_destroy_all(cgrp
);
4171 if (cgroup_parent(cgrp
)) {
4173 * We get a ref to the parent, and put the ref when
4174 * this cgroup is being freed, so it's guaranteed
4175 * that the parent won't be destroyed before its
4178 cgroup_put(cgroup_parent(cgrp
));
4179 kernfs_put(cgrp
->kn
);
4183 * This is root cgroup's refcnt reaching zero,
4184 * which indicates that the root should be
4187 cgroup_destroy_root(cgrp
->root
);
4192 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4194 struct cgroup_subsys_state
*css
=
4195 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4197 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4198 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4201 static void css_release_work_fn(struct work_struct
*work
)
4203 struct cgroup_subsys_state
*css
=
4204 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4205 struct cgroup_subsys
*ss
= css
->ss
;
4206 struct cgroup
*cgrp
= css
->cgroup
;
4208 mutex_lock(&cgroup_mutex
);
4210 css
->flags
|= CSS_RELEASED
;
4211 list_del_rcu(&css
->sibling
);
4214 /* css release path */
4215 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4217 /* cgroup release path */
4218 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
4222 mutex_unlock(&cgroup_mutex
);
4224 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4227 static void css_release(struct percpu_ref
*ref
)
4229 struct cgroup_subsys_state
*css
=
4230 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4232 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
4233 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4236 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4237 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4239 lockdep_assert_held(&cgroup_mutex
);
4243 memset(css
, 0, sizeof(*css
));
4246 INIT_LIST_HEAD(&css
->sibling
);
4247 INIT_LIST_HEAD(&css
->children
);
4248 css
->serial_nr
= css_serial_nr_next
++;
4250 if (cgroup_parent(cgrp
)) {
4251 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
4252 css_get(css
->parent
);
4255 BUG_ON(cgroup_css(cgrp
, ss
));
4258 /* invoke ->css_online() on a new CSS and mark it online if successful */
4259 static int online_css(struct cgroup_subsys_state
*css
)
4261 struct cgroup_subsys
*ss
= css
->ss
;
4264 lockdep_assert_held(&cgroup_mutex
);
4267 ret
= ss
->css_online(css
);
4269 css
->flags
|= CSS_ONLINE
;
4270 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4275 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4276 static void offline_css(struct cgroup_subsys_state
*css
)
4278 struct cgroup_subsys
*ss
= css
->ss
;
4280 lockdep_assert_held(&cgroup_mutex
);
4282 if (!(css
->flags
& CSS_ONLINE
))
4285 if (ss
->css_offline
)
4286 ss
->css_offline(css
);
4288 css
->flags
&= ~CSS_ONLINE
;
4289 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4291 wake_up_all(&css
->cgroup
->offline_waitq
);
4295 * create_css - create a cgroup_subsys_state
4296 * @cgrp: the cgroup new css will be associated with
4297 * @ss: the subsys of new css
4299 * Create a new css associated with @cgrp - @ss pair. On success, the new
4300 * css is online and installed in @cgrp with all interface files created.
4301 * Returns 0 on success, -errno on failure.
4303 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
)
4305 struct cgroup
*parent
= cgroup_parent(cgrp
);
4306 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
4307 struct cgroup_subsys_state
*css
;
4310 lockdep_assert_held(&cgroup_mutex
);
4312 css
= ss
->css_alloc(parent_css
);
4314 return PTR_ERR(css
);
4316 init_and_link_css(css
, ss
, cgrp
);
4318 err
= percpu_ref_init(&css
->refcnt
, css_release
);
4322 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_NOWAIT
);
4324 goto err_free_percpu_ref
;
4327 err
= cgroup_populate_dir(cgrp
, 1 << ss
->id
);
4331 /* @css is ready to be brought online now, make it visible */
4332 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
4333 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4335 err
= online_css(css
);
4339 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4340 cgroup_parent(parent
)) {
4341 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4342 current
->comm
, current
->pid
, ss
->name
);
4343 if (!strcmp(ss
->name
, "memory"))
4344 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4345 ss
->warned_broken_hierarchy
= true;
4351 list_del_rcu(&css
->sibling
);
4352 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4354 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4355 err_free_percpu_ref
:
4356 percpu_ref_cancel_init(&css
->refcnt
);
4358 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4362 static int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
,
4365 struct cgroup
*parent
, *cgrp
;
4366 struct cgroup_root
*root
;
4367 struct cgroup_subsys
*ss
;
4368 struct kernfs_node
*kn
;
4371 parent
= cgroup_kn_lock_live(parent_kn
);
4374 root
= parent
->root
;
4376 /* allocate the cgroup and its ID, 0 is reserved for the root */
4377 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
4383 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
);
4388 * Temporarily set the pointer to NULL, so idr_find() won't return
4389 * a half-baked cgroup.
4391 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_NOWAIT
);
4394 goto out_cancel_ref
;
4397 init_cgroup_housekeeping(cgrp
);
4399 cgrp
->self
.parent
= &parent
->self
;
4402 if (notify_on_release(parent
))
4403 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4405 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4406 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4408 /* create the directory */
4409 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4417 * This extra ref will be put in cgroup_free_fn() and guarantees
4418 * that @cgrp->kn is always accessible.
4422 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
4424 /* allocation complete, commit to creation */
4425 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
4426 atomic_inc(&root
->nr_cgrps
);
4430 * @cgrp is now fully operational. If something fails after this
4431 * point, it'll be released via the normal destruction path.
4433 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4435 ret
= cgroup_kn_set_ugid(kn
);
4439 ret
= cgroup_addrm_files(cgrp
, cgroup_base_files
, true);
4443 /* let's create and online css's */
4444 for_each_subsys(ss
, ssid
) {
4445 if (parent
->child_subsys_mask
& (1 << ssid
)) {
4446 ret
= create_css(cgrp
, ss
);
4453 * On the default hierarchy, a child doesn't automatically inherit
4454 * child_subsys_mask from the parent. Each is configured manually.
4456 if (!cgroup_on_dfl(cgrp
))
4457 cgrp
->child_subsys_mask
= parent
->child_subsys_mask
;
4459 kernfs_activate(kn
);
4465 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4467 percpu_ref_cancel_init(&cgrp
->self
.refcnt
);
4471 cgroup_kn_unlock(parent_kn
);
4475 cgroup_destroy_locked(cgrp
);
4480 * This is called when the refcnt of a css is confirmed to be killed.
4481 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4482 * initate destruction and put the css ref from kill_css().
4484 static void css_killed_work_fn(struct work_struct
*work
)
4486 struct cgroup_subsys_state
*css
=
4487 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4489 mutex_lock(&cgroup_mutex
);
4491 mutex_unlock(&cgroup_mutex
);
4496 /* css kill confirmation processing requires process context, bounce */
4497 static void css_killed_ref_fn(struct percpu_ref
*ref
)
4499 struct cgroup_subsys_state
*css
=
4500 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4502 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
4503 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4507 * kill_css - destroy a css
4508 * @css: css to destroy
4510 * This function initiates destruction of @css by removing cgroup interface
4511 * files and putting its base reference. ->css_offline() will be invoked
4512 * asynchronously once css_tryget_online() is guaranteed to fail and when
4513 * the reference count reaches zero, @css will be released.
4515 static void kill_css(struct cgroup_subsys_state
*css
)
4517 lockdep_assert_held(&cgroup_mutex
);
4520 * This must happen before css is disassociated with its cgroup.
4521 * See seq_css() for details.
4523 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4526 * Killing would put the base ref, but we need to keep it alive
4527 * until after ->css_offline().
4532 * cgroup core guarantees that, by the time ->css_offline() is
4533 * invoked, no new css reference will be given out via
4534 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4535 * proceed to offlining css's because percpu_ref_kill() doesn't
4536 * guarantee that the ref is seen as killed on all CPUs on return.
4538 * Use percpu_ref_kill_and_confirm() to get notifications as each
4539 * css is confirmed to be seen as killed on all CPUs.
4541 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
4545 * cgroup_destroy_locked - the first stage of cgroup destruction
4546 * @cgrp: cgroup to be destroyed
4548 * css's make use of percpu refcnts whose killing latency shouldn't be
4549 * exposed to userland and are RCU protected. Also, cgroup core needs to
4550 * guarantee that css_tryget_online() won't succeed by the time
4551 * ->css_offline() is invoked. To satisfy all the requirements,
4552 * destruction is implemented in the following two steps.
4554 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4555 * userland visible parts and start killing the percpu refcnts of
4556 * css's. Set up so that the next stage will be kicked off once all
4557 * the percpu refcnts are confirmed to be killed.
4559 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4560 * rest of destruction. Once all cgroup references are gone, the
4561 * cgroup is RCU-freed.
4563 * This function implements s1. After this step, @cgrp is gone as far as
4564 * the userland is concerned and a new cgroup with the same name may be
4565 * created. As cgroup doesn't care about the names internally, this
4566 * doesn't cause any problem.
4568 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
4569 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
4571 struct cgroup_subsys_state
*css
;
4575 lockdep_assert_held(&cgroup_mutex
);
4578 * css_set_rwsem synchronizes access to ->cset_links and prevents
4579 * @cgrp from being removed while put_css_set() is in progress.
4581 down_read(&css_set_rwsem
);
4582 empty
= list_empty(&cgrp
->cset_links
);
4583 up_read(&css_set_rwsem
);
4588 * Make sure there's no live children. We can't test emptiness of
4589 * ->self.children as dead children linger on it while being
4590 * drained; otherwise, "rmdir parent/child parent" may fail.
4592 if (css_has_online_children(&cgrp
->self
))
4596 * Mark @cgrp dead. This prevents further task migration and child
4597 * creation by disabling cgroup_lock_live_group().
4599 cgrp
->self
.flags
&= ~CSS_ONLINE
;
4601 /* initiate massacre of all css's */
4602 for_each_css(css
, ssid
, cgrp
)
4605 /* CSS_ONLINE is clear, remove from ->release_list for the last time */
4606 raw_spin_lock(&release_list_lock
);
4607 if (!list_empty(&cgrp
->release_list
))
4608 list_del_init(&cgrp
->release_list
);
4609 raw_spin_unlock(&release_list_lock
);
4612 * Remove @cgrp directory along with the base files. @cgrp has an
4613 * extra ref on its kn.
4615 kernfs_remove(cgrp
->kn
);
4617 set_bit(CGRP_RELEASABLE
, &cgroup_parent(cgrp
)->flags
);
4618 check_for_release(cgroup_parent(cgrp
));
4620 /* put the base reference */
4621 percpu_ref_kill(&cgrp
->self
.refcnt
);
4626 static int cgroup_rmdir(struct kernfs_node
*kn
)
4628 struct cgroup
*cgrp
;
4631 cgrp
= cgroup_kn_lock_live(kn
);
4634 cgroup_get(cgrp
); /* for @kn->priv clearing */
4636 ret
= cgroup_destroy_locked(cgrp
);
4638 cgroup_kn_unlock(kn
);
4641 * There are two control paths which try to determine cgroup from
4642 * dentry without going through kernfs - cgroupstats_build() and
4643 * css_tryget_online_from_dir(). Those are supported by RCU
4644 * protecting clearing of cgrp->kn->priv backpointer, which should
4645 * happen after all files under it have been removed.
4648 RCU_INIT_POINTER(*(void __rcu __force
**)&kn
->priv
, NULL
);
4654 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
4655 .remount_fs
= cgroup_remount
,
4656 .show_options
= cgroup_show_options
,
4657 .mkdir
= cgroup_mkdir
,
4658 .rmdir
= cgroup_rmdir
,
4659 .rename
= cgroup_rename
,
4662 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
4664 struct cgroup_subsys_state
*css
;
4666 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
4668 mutex_lock(&cgroup_mutex
);
4670 idr_init(&ss
->css_idr
);
4671 INIT_LIST_HEAD(&ss
->cfts
);
4673 /* Create the root cgroup state for this subsystem */
4674 ss
->root
= &cgrp_dfl_root
;
4675 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
4676 /* We don't handle early failures gracefully */
4677 BUG_ON(IS_ERR(css
));
4678 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
4681 * Root csses are never destroyed and we can't initialize
4682 * percpu_ref during early init. Disable refcnting.
4684 css
->flags
|= CSS_NO_REF
;
4687 /* allocation can't be done safely during early init */
4690 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
4691 BUG_ON(css
->id
< 0);
4694 /* Update the init_css_set to contain a subsys
4695 * pointer to this state - since the subsystem is
4696 * newly registered, all tasks and hence the
4697 * init_css_set is in the subsystem's root cgroup. */
4698 init_css_set
.subsys
[ss
->id
] = css
;
4700 need_forkexit_callback
|= ss
->fork
|| ss
->exit
;
4702 /* At system boot, before all subsystems have been
4703 * registered, no tasks have been forked, so we don't
4704 * need to invoke fork callbacks here. */
4705 BUG_ON(!list_empty(&init_task
.tasks
));
4707 BUG_ON(online_css(css
));
4709 mutex_unlock(&cgroup_mutex
);
4713 * cgroup_init_early - cgroup initialization at system boot
4715 * Initialize cgroups at system boot, and initialize any
4716 * subsystems that request early init.
4718 int __init
cgroup_init_early(void)
4720 static struct cgroup_sb_opts __initdata opts
=
4721 { .flags
= CGRP_ROOT_SANE_BEHAVIOR
};
4722 struct cgroup_subsys
*ss
;
4725 init_cgroup_root(&cgrp_dfl_root
, &opts
);
4726 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
4728 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
4730 for_each_subsys(ss
, i
) {
4731 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
4732 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4733 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
4735 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
4736 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
4739 ss
->name
= cgroup_subsys_name
[i
];
4742 cgroup_init_subsys(ss
, true);
4748 * cgroup_init - cgroup initialization
4750 * Register cgroup filesystem and /proc file, and initialize
4751 * any subsystems that didn't request early init.
4753 int __init
cgroup_init(void)
4755 struct cgroup_subsys
*ss
;
4759 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_base_files
));
4761 mutex_lock(&cgroup_mutex
);
4763 /* Add init_css_set to the hash table */
4764 key
= css_set_hash(init_css_set
.subsys
);
4765 hash_add(css_set_table
, &init_css_set
.hlist
, key
);
4767 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
4769 mutex_unlock(&cgroup_mutex
);
4771 for_each_subsys(ss
, ssid
) {
4772 if (ss
->early_init
) {
4773 struct cgroup_subsys_state
*css
=
4774 init_css_set
.subsys
[ss
->id
];
4776 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
4778 BUG_ON(css
->id
< 0);
4780 cgroup_init_subsys(ss
, false);
4783 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
4784 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
4787 * Setting dfl_root subsys_mask needs to consider the
4788 * disabled flag and cftype registration needs kmalloc,
4789 * both of which aren't available during early_init.
4791 if (!ss
->disabled
) {
4792 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
4793 WARN_ON(cgroup_add_cftypes(ss
, ss
->base_cftypes
));
4797 cgroup_kobj
= kobject_create_and_add("cgroup", fs_kobj
);
4801 err
= register_filesystem(&cgroup_fs_type
);
4803 kobject_put(cgroup_kobj
);
4807 proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
);
4811 static int __init
cgroup_wq_init(void)
4814 * There isn't much point in executing destruction path in
4815 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4816 * Use 1 for @max_active.
4818 * We would prefer to do this in cgroup_init() above, but that
4819 * is called before init_workqueues(): so leave this until after.
4821 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
4822 BUG_ON(!cgroup_destroy_wq
);
4825 * Used to destroy pidlists and separate to serve as flush domain.
4826 * Cap @max_active to 1 too.
4828 cgroup_pidlist_destroy_wq
= alloc_workqueue("cgroup_pidlist_destroy",
4830 BUG_ON(!cgroup_pidlist_destroy_wq
);
4834 core_initcall(cgroup_wq_init
);
4837 * proc_cgroup_show()
4838 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4839 * - Used for /proc/<pid>/cgroup.
4842 /* TODO: Use a proper seq_file iterator */
4843 int proc_cgroup_show(struct seq_file
*m
, void *v
)
4846 struct task_struct
*tsk
;
4849 struct cgroup_root
*root
;
4852 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
4858 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
4864 mutex_lock(&cgroup_mutex
);
4865 down_read(&css_set_rwsem
);
4867 for_each_root(root
) {
4868 struct cgroup_subsys
*ss
;
4869 struct cgroup
*cgrp
;
4870 int ssid
, count
= 0;
4872 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_root_visible
)
4875 seq_printf(m
, "%d:", root
->hierarchy_id
);
4876 for_each_subsys(ss
, ssid
)
4877 if (root
->subsys_mask
& (1 << ssid
))
4878 seq_printf(m
, "%s%s", count
++ ? "," : "", ss
->name
);
4879 if (strlen(root
->name
))
4880 seq_printf(m
, "%sname=%s", count
? "," : "",
4883 cgrp
= task_cgroup_from_root(tsk
, root
);
4884 path
= cgroup_path(cgrp
, buf
, PATH_MAX
);
4886 retval
= -ENAMETOOLONG
;
4894 up_read(&css_set_rwsem
);
4895 mutex_unlock(&cgroup_mutex
);
4896 put_task_struct(tsk
);
4903 /* Display information about each subsystem and each hierarchy */
4904 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
4906 struct cgroup_subsys
*ss
;
4909 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4911 * ideally we don't want subsystems moving around while we do this.
4912 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4913 * subsys/hierarchy state.
4915 mutex_lock(&cgroup_mutex
);
4917 for_each_subsys(ss
, i
)
4918 seq_printf(m
, "%s\t%d\t%d\t%d\n",
4919 ss
->name
, ss
->root
->hierarchy_id
,
4920 atomic_read(&ss
->root
->nr_cgrps
), !ss
->disabled
);
4922 mutex_unlock(&cgroup_mutex
);
4926 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
4928 return single_open(file
, proc_cgroupstats_show
, NULL
);
4931 static const struct file_operations proc_cgroupstats_operations
= {
4932 .open
= cgroupstats_open
,
4934 .llseek
= seq_lseek
,
4935 .release
= single_release
,
4939 * cgroup_fork - initialize cgroup related fields during copy_process()
4940 * @child: pointer to task_struct of forking parent process.
4942 * A task is associated with the init_css_set until cgroup_post_fork()
4943 * attaches it to the parent's css_set. Empty cg_list indicates that
4944 * @child isn't holding reference to its css_set.
4946 void cgroup_fork(struct task_struct
*child
)
4948 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
4949 INIT_LIST_HEAD(&child
->cg_list
);
4953 * cgroup_post_fork - called on a new task after adding it to the task list
4954 * @child: the task in question
4956 * Adds the task to the list running through its css_set if necessary and
4957 * call the subsystem fork() callbacks. Has to be after the task is
4958 * visible on the task list in case we race with the first call to
4959 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4962 void cgroup_post_fork(struct task_struct
*child
)
4964 struct cgroup_subsys
*ss
;
4968 * This may race against cgroup_enable_task_cg_links(). As that
4969 * function sets use_task_css_set_links before grabbing
4970 * tasklist_lock and we just went through tasklist_lock to add
4971 * @child, it's guaranteed that either we see the set
4972 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4973 * @child during its iteration.
4975 * If we won the race, @child is associated with %current's
4976 * css_set. Grabbing css_set_rwsem guarantees both that the
4977 * association is stable, and, on completion of the parent's
4978 * migration, @child is visible in the source of migration or
4979 * already in the destination cgroup. This guarantee is necessary
4980 * when implementing operations which need to migrate all tasks of
4981 * a cgroup to another.
4983 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4984 * will remain in init_css_set. This is safe because all tasks are
4985 * in the init_css_set before cg_links is enabled and there's no
4986 * operation which transfers all tasks out of init_css_set.
4988 if (use_task_css_set_links
) {
4989 struct css_set
*cset
;
4991 down_write(&css_set_rwsem
);
4992 cset
= task_css_set(current
);
4993 if (list_empty(&child
->cg_list
)) {
4994 rcu_assign_pointer(child
->cgroups
, cset
);
4995 list_add(&child
->cg_list
, &cset
->tasks
);
4998 up_write(&css_set_rwsem
);
5002 * Call ss->fork(). This must happen after @child is linked on
5003 * css_set; otherwise, @child might change state between ->fork()
5004 * and addition to css_set.
5006 if (need_forkexit_callback
) {
5007 for_each_subsys(ss
, i
)
5014 * cgroup_exit - detach cgroup from exiting task
5015 * @tsk: pointer to task_struct of exiting process
5017 * Description: Detach cgroup from @tsk and release it.
5019 * Note that cgroups marked notify_on_release force every task in
5020 * them to take the global cgroup_mutex mutex when exiting.
5021 * This could impact scaling on very large systems. Be reluctant to
5022 * use notify_on_release cgroups where very high task exit scaling
5023 * is required on large systems.
5025 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5026 * call cgroup_exit() while the task is still competent to handle
5027 * notify_on_release(), then leave the task attached to the root cgroup in
5028 * each hierarchy for the remainder of its exit. No need to bother with
5029 * init_css_set refcnting. init_css_set never goes away and we can't race
5030 * with migration path - PF_EXITING is visible to migration path.
5032 void cgroup_exit(struct task_struct
*tsk
)
5034 struct cgroup_subsys
*ss
;
5035 struct css_set
*cset
;
5036 bool put_cset
= false;
5040 * Unlink from @tsk from its css_set. As migration path can't race
5041 * with us, we can check cg_list without grabbing css_set_rwsem.
5043 if (!list_empty(&tsk
->cg_list
)) {
5044 down_write(&css_set_rwsem
);
5045 list_del_init(&tsk
->cg_list
);
5046 up_write(&css_set_rwsem
);
5050 /* Reassign the task to the init_css_set. */
5051 cset
= task_css_set(tsk
);
5052 RCU_INIT_POINTER(tsk
->cgroups
, &init_css_set
);
5054 if (need_forkexit_callback
) {
5055 /* see cgroup_post_fork() for details */
5056 for_each_subsys(ss
, i
) {
5058 struct cgroup_subsys_state
*old_css
= cset
->subsys
[i
];
5059 struct cgroup_subsys_state
*css
= task_css(tsk
, i
);
5061 ss
->exit(css
, old_css
, tsk
);
5067 put_css_set(cset
, true);
5070 static void check_for_release(struct cgroup
*cgrp
)
5072 if (cgroup_is_releasable(cgrp
) && list_empty(&cgrp
->cset_links
) &&
5073 !css_has_online_children(&cgrp
->self
)) {
5075 * Control Group is currently removeable. If it's not
5076 * already queued for a userspace notification, queue
5079 int need_schedule_work
= 0;
5081 raw_spin_lock(&release_list_lock
);
5082 if (!cgroup_is_dead(cgrp
) &&
5083 list_empty(&cgrp
->release_list
)) {
5084 list_add(&cgrp
->release_list
, &release_list
);
5085 need_schedule_work
= 1;
5087 raw_spin_unlock(&release_list_lock
);
5088 if (need_schedule_work
)
5089 schedule_work(&release_agent_work
);
5094 * Notify userspace when a cgroup is released, by running the
5095 * configured release agent with the name of the cgroup (path
5096 * relative to the root of cgroup file system) as the argument.
5098 * Most likely, this user command will try to rmdir this cgroup.
5100 * This races with the possibility that some other task will be
5101 * attached to this cgroup before it is removed, or that some other
5102 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5103 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5104 * unused, and this cgroup will be reprieved from its death sentence,
5105 * to continue to serve a useful existence. Next time it's released,
5106 * we will get notified again, if it still has 'notify_on_release' set.
5108 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5109 * means only wait until the task is successfully execve()'d. The
5110 * separate release agent task is forked by call_usermodehelper(),
5111 * then control in this thread returns here, without waiting for the
5112 * release agent task. We don't bother to wait because the caller of
5113 * this routine has no use for the exit status of the release agent
5114 * task, so no sense holding our caller up for that.
5116 static void cgroup_release_agent(struct work_struct
*work
)
5118 BUG_ON(work
!= &release_agent_work
);
5119 mutex_lock(&cgroup_mutex
);
5120 raw_spin_lock(&release_list_lock
);
5121 while (!list_empty(&release_list
)) {
5122 char *argv
[3], *envp
[3];
5124 char *pathbuf
= NULL
, *agentbuf
= NULL
, *path
;
5125 struct cgroup
*cgrp
= list_entry(release_list
.next
,
5128 list_del_init(&cgrp
->release_list
);
5129 raw_spin_unlock(&release_list_lock
);
5130 pathbuf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5133 path
= cgroup_path(cgrp
, pathbuf
, PATH_MAX
);
5136 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
5141 argv
[i
++] = agentbuf
;
5146 /* minimal command environment */
5147 envp
[i
++] = "HOME=/";
5148 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5151 /* Drop the lock while we invoke the usermode helper,
5152 * since the exec could involve hitting disk and hence
5153 * be a slow process */
5154 mutex_unlock(&cgroup_mutex
);
5155 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
5156 mutex_lock(&cgroup_mutex
);
5160 raw_spin_lock(&release_list_lock
);
5162 raw_spin_unlock(&release_list_lock
);
5163 mutex_unlock(&cgroup_mutex
);
5166 static int __init
cgroup_disable(char *str
)
5168 struct cgroup_subsys
*ss
;
5172 while ((token
= strsep(&str
, ",")) != NULL
) {
5176 for_each_subsys(ss
, i
) {
5177 if (!strcmp(token
, ss
->name
)) {
5179 printk(KERN_INFO
"Disabling %s control group"
5180 " subsystem\n", ss
->name
);
5187 __setup("cgroup_disable=", cgroup_disable
);
5190 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5191 * @dentry: directory dentry of interest
5192 * @ss: subsystem of interest
5194 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5195 * to get the corresponding css and return it. If such css doesn't exist
5196 * or can't be pinned, an ERR_PTR value is returned.
5198 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
5199 struct cgroup_subsys
*ss
)
5201 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5202 struct cgroup_subsys_state
*css
= NULL
;
5203 struct cgroup
*cgrp
;
5205 /* is @dentry a cgroup dir? */
5206 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
5207 kernfs_type(kn
) != KERNFS_DIR
)
5208 return ERR_PTR(-EBADF
);
5213 * This path doesn't originate from kernfs and @kn could already
5214 * have been or be removed at any point. @kn->priv is RCU
5215 * protected for this access. See cgroup_rmdir() for details.
5217 cgrp
= rcu_dereference(kn
->priv
);
5219 css
= cgroup_css(cgrp
, ss
);
5221 if (!css
|| !css_tryget_online(css
))
5222 css
= ERR_PTR(-ENOENT
);
5229 * css_from_id - lookup css by id
5230 * @id: the cgroup id
5231 * @ss: cgroup subsys to be looked into
5233 * Returns the css if there's valid one with @id, otherwise returns NULL.
5234 * Should be called under rcu_read_lock().
5236 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5238 WARN_ON_ONCE(!rcu_read_lock_held());
5239 return idr_find(&ss
->css_idr
, id
);
5242 #ifdef CONFIG_CGROUP_DEBUG
5243 static struct cgroup_subsys_state
*
5244 debug_css_alloc(struct cgroup_subsys_state
*parent_css
)
5246 struct cgroup_subsys_state
*css
= kzalloc(sizeof(*css
), GFP_KERNEL
);
5249 return ERR_PTR(-ENOMEM
);
5254 static void debug_css_free(struct cgroup_subsys_state
*css
)
5259 static u64
debug_taskcount_read(struct cgroup_subsys_state
*css
,
5262 return cgroup_task_count(css
->cgroup
);
5265 static u64
current_css_set_read(struct cgroup_subsys_state
*css
,
5268 return (u64
)(unsigned long)current
->cgroups
;
5271 static u64
current_css_set_refcount_read(struct cgroup_subsys_state
*css
,
5277 count
= atomic_read(&task_css_set(current
)->refcount
);
5282 static int current_css_set_cg_links_read(struct seq_file
*seq
, void *v
)
5284 struct cgrp_cset_link
*link
;
5285 struct css_set
*cset
;
5288 name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
5292 down_read(&css_set_rwsem
);
5294 cset
= rcu_dereference(current
->cgroups
);
5295 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
5296 struct cgroup
*c
= link
->cgrp
;
5298 cgroup_name(c
, name_buf
, NAME_MAX
+ 1);
5299 seq_printf(seq
, "Root %d group %s\n",
5300 c
->root
->hierarchy_id
, name_buf
);
5303 up_read(&css_set_rwsem
);
5308 #define MAX_TASKS_SHOWN_PER_CSS 25
5309 static int cgroup_css_links_read(struct seq_file
*seq
, void *v
)
5311 struct cgroup_subsys_state
*css
= seq_css(seq
);
5312 struct cgrp_cset_link
*link
;
5314 down_read(&css_set_rwsem
);
5315 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
) {
5316 struct css_set
*cset
= link
->cset
;
5317 struct task_struct
*task
;
5320 seq_printf(seq
, "css_set %p\n", cset
);
5322 list_for_each_entry(task
, &cset
->tasks
, cg_list
) {
5323 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5325 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5328 list_for_each_entry(task
, &cset
->mg_tasks
, cg_list
) {
5329 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5331 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5335 seq_puts(seq
, " ...\n");
5337 up_read(&css_set_rwsem
);
5341 static u64
releasable_read(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
5343 return test_bit(CGRP_RELEASABLE
, &css
->cgroup
->flags
);
5346 static struct cftype debug_files
[] = {
5348 .name
= "taskcount",
5349 .read_u64
= debug_taskcount_read
,
5353 .name
= "current_css_set",
5354 .read_u64
= current_css_set_read
,
5358 .name
= "current_css_set_refcount",
5359 .read_u64
= current_css_set_refcount_read
,
5363 .name
= "current_css_set_cg_links",
5364 .seq_show
= current_css_set_cg_links_read
,
5368 .name
= "cgroup_css_links",
5369 .seq_show
= cgroup_css_links_read
,
5373 .name
= "releasable",
5374 .read_u64
= releasable_read
,
5380 struct cgroup_subsys debug_cgrp_subsys
= {
5381 .css_alloc
= debug_css_alloc
,
5382 .css_free
= debug_css_free
,
5383 .base_cftypes
= debug_files
,
5385 #endif /* CONFIG_CGROUP_DEBUG */