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>
39 #include <linux/mutex.h>
40 #include <linux/mount.h>
41 #include <linux/pagemap.h>
42 #include <linux/proc_fs.h>
43 #include <linux/rcupdate.h>
44 #include <linux/sched.h>
45 #include <linux/slab.h>
46 #include <linux/spinlock.h>
47 #include <linux/rwsem.h>
48 #include <linux/string.h>
49 #include <linux/sort.h>
50 #include <linux/kmod.h>
51 #include <linux/delayacct.h>
52 #include <linux/cgroupstats.h>
53 #include <linux/hashtable.h>
54 #include <linux/pid_namespace.h>
55 #include <linux/idr.h>
56 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
57 #include <linux/kthread.h>
58 #include <linux/delay.h>
60 #include <linux/atomic.h>
63 * pidlists linger the following amount before being destroyed. The goal
64 * is avoiding frequent destruction in the middle of consecutive read calls
65 * Expiring in the middle is a performance problem not a correctness one.
66 * 1 sec should be enough.
68 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
70 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
74 * cgroup_tree_mutex nests above cgroup_mutex and protects cftypes, file
75 * creation/removal and hierarchy changing operations including cgroup
76 * creation, removal, css association and controller rebinding. This outer
77 * lock is needed mainly to resolve the circular dependency between kernfs
78 * active ref and cgroup_mutex. cgroup_tree_mutex nests above both.
80 static DEFINE_MUTEX(cgroup_tree_mutex
);
83 * cgroup_mutex is the master lock. Any modification to cgroup or its
84 * hierarchy must be performed while holding it.
86 * css_set_rwsem protects task->cgroups pointer, the list of css_set
87 * objects, and the chain of tasks off each css_set.
89 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
90 * cgroup.h can use them for lockdep annotations.
92 #ifdef CONFIG_PROVE_RCU
93 DEFINE_MUTEX(cgroup_mutex
);
94 DECLARE_RWSEM(css_set_rwsem
);
95 EXPORT_SYMBOL_GPL(cgroup_mutex
);
96 EXPORT_SYMBOL_GPL(css_set_rwsem
);
98 static DEFINE_MUTEX(cgroup_mutex
);
99 static DECLARE_RWSEM(css_set_rwsem
);
103 * Protects cgroup_idr and css_idr so that IDs can be released without
104 * grabbing cgroup_mutex.
106 static DEFINE_SPINLOCK(cgroup_idr_lock
);
109 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
110 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
112 static DEFINE_SPINLOCK(release_agent_path_lock
);
114 #define cgroup_assert_mutexes_or_rcu_locked() \
115 rcu_lockdep_assert(rcu_read_lock_held() || \
116 lockdep_is_held(&cgroup_tree_mutex) || \
117 lockdep_is_held(&cgroup_mutex), \
118 "cgroup_[tree_]mutex or RCU read lock required");
121 * cgroup destruction makes heavy use of work items and there can be a lot
122 * of concurrent destructions. Use a separate workqueue so that cgroup
123 * destruction work items don't end up filling up max_active of system_wq
124 * which may lead to deadlock.
126 static struct workqueue_struct
*cgroup_destroy_wq
;
129 * pidlist destructions need to be flushed on cgroup destruction. Use a
130 * separate workqueue as flush domain.
132 static struct workqueue_struct
*cgroup_pidlist_destroy_wq
;
134 /* generate an array of cgroup subsystem pointers */
135 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
136 static struct cgroup_subsys
*cgroup_subsys
[] = {
137 #include <linux/cgroup_subsys.h>
141 /* array of cgroup subsystem names */
142 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
143 static const char *cgroup_subsys_name
[] = {
144 #include <linux/cgroup_subsys.h>
149 * The default hierarchy, reserved for the subsystems that are otherwise
150 * unattached - it never has more than a single cgroup, and all tasks are
151 * part of that cgroup.
153 struct cgroup_root cgrp_dfl_root
;
156 * The default hierarchy always exists but is hidden until mounted for the
157 * first time. This is for backward compatibility.
159 static bool cgrp_dfl_root_visible
;
161 /* The list of hierarchy roots */
163 static LIST_HEAD(cgroup_roots
);
164 static int cgroup_root_count
;
166 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
167 static DEFINE_IDR(cgroup_hierarchy_idr
);
170 * Assign a monotonically increasing serial number to cgroups. It
171 * guarantees cgroups with bigger numbers are newer than those with smaller
172 * numbers. Also, as cgroups are always appended to the parent's
173 * ->children list, it guarantees that sibling cgroups are always sorted in
174 * the ascending serial number order on the list. Protected by
177 static u64 cgroup_serial_nr_next
= 1;
179 /* This flag indicates whether tasks in the fork and exit paths should
180 * check for fork/exit handlers to call. This avoids us having to do
181 * extra work in the fork/exit path if none of the subsystems need to
184 static int need_forkexit_callback __read_mostly
;
186 static struct cftype cgroup_base_files
[];
188 static void cgroup_put(struct cgroup
*cgrp
);
189 static int rebind_subsystems(struct cgroup_root
*dst_root
,
190 unsigned int ss_mask
);
191 static void cgroup_destroy_css_killed(struct cgroup
*cgrp
);
192 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
193 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
);
194 static void kill_css(struct cgroup_subsys_state
*css
);
195 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
197 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
);
199 /* IDR wrappers which synchronize using cgroup_idr_lock */
200 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
205 idr_preload(gfp_mask
);
206 spin_lock(&cgroup_idr_lock
);
207 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
);
208 spin_unlock(&cgroup_idr_lock
);
213 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
217 spin_lock(&cgroup_idr_lock
);
218 ret
= idr_replace(idr
, ptr
, id
);
219 spin_unlock(&cgroup_idr_lock
);
223 static void cgroup_idr_remove(struct idr
*idr
, int id
)
225 spin_lock(&cgroup_idr_lock
);
227 spin_unlock(&cgroup_idr_lock
);
231 * cgroup_css - obtain a cgroup's css for the specified subsystem
232 * @cgrp: the cgroup of interest
233 * @ss: the subsystem of interest (%NULL returns the dummy_css)
235 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
236 * function must be called either under cgroup_mutex or rcu_read_lock() and
237 * the caller is responsible for pinning the returned css if it wants to
238 * keep accessing it outside the said locks. This function may return
239 * %NULL if @cgrp doesn't have @subsys_id enabled.
241 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
242 struct cgroup_subsys
*ss
)
245 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
246 lockdep_is_held(&cgroup_tree_mutex
) ||
247 lockdep_is_held(&cgroup_mutex
));
249 return &cgrp
->dummy_css
;
253 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
254 * @cgrp: the cgroup of interest
255 * @ss: the subsystem of interest (%NULL returns the dummy_css)
257 * Similar to cgroup_css() but returns the effctive css, which is defined
258 * as the matching css of the nearest ancestor including self which has @ss
259 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
260 * function is guaranteed to return non-NULL css.
262 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
263 struct cgroup_subsys
*ss
)
265 lockdep_assert_held(&cgroup_mutex
);
268 return &cgrp
->dummy_css
;
270 if (!(cgrp
->root
->subsys_mask
& (1 << ss
->id
)))
273 while (cgrp
->parent
&&
274 !(cgrp
->parent
->child_subsys_mask
& (1 << ss
->id
)))
277 return cgroup_css(cgrp
, ss
);
280 /* convenient tests for these bits */
281 static inline bool cgroup_is_dead(const struct cgroup
*cgrp
)
283 return test_bit(CGRP_DEAD
, &cgrp
->flags
);
286 struct cgroup_subsys_state
*seq_css(struct seq_file
*seq
)
288 struct kernfs_open_file
*of
= seq
->private;
289 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
290 struct cftype
*cft
= seq_cft(seq
);
293 * This is open and unprotected implementation of cgroup_css().
294 * seq_css() is only called from a kernfs file operation which has
295 * an active reference on the file. Because all the subsystem
296 * files are drained before a css is disassociated with a cgroup,
297 * the matching css from the cgroup's subsys table is guaranteed to
298 * be and stay valid until the enclosing operation is complete.
301 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
303 return &cgrp
->dummy_css
;
305 EXPORT_SYMBOL_GPL(seq_css
);
308 * cgroup_is_descendant - test ancestry
309 * @cgrp: the cgroup to be tested
310 * @ancestor: possible ancestor of @cgrp
312 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
313 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
314 * and @ancestor are accessible.
316 bool cgroup_is_descendant(struct cgroup
*cgrp
, struct cgroup
*ancestor
)
319 if (cgrp
== ancestor
)
326 static int cgroup_is_releasable(const struct cgroup
*cgrp
)
329 (1 << CGRP_RELEASABLE
) |
330 (1 << CGRP_NOTIFY_ON_RELEASE
);
331 return (cgrp
->flags
& bits
) == bits
;
334 static int notify_on_release(const struct cgroup
*cgrp
)
336 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
340 * for_each_css - iterate all css's of a cgroup
341 * @css: the iteration cursor
342 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
343 * @cgrp: the target cgroup to iterate css's of
345 * Should be called under cgroup_[tree_]mutex.
347 #define for_each_css(css, ssid, cgrp) \
348 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
349 if (!((css) = rcu_dereference_check( \
350 (cgrp)->subsys[(ssid)], \
351 lockdep_is_held(&cgroup_tree_mutex) || \
352 lockdep_is_held(&cgroup_mutex)))) { } \
356 * for_each_e_css - iterate all effective css's of a cgroup
357 * @css: the iteration cursor
358 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
359 * @cgrp: the target cgroup to iterate css's of
361 * Should be called under cgroup_[tree_]mutex.
363 #define for_each_e_css(css, ssid, cgrp) \
364 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
365 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
370 * for_each_subsys - iterate all enabled cgroup subsystems
371 * @ss: the iteration cursor
372 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
374 #define for_each_subsys(ss, ssid) \
375 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
376 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
378 /* iterate across the hierarchies */
379 #define for_each_root(root) \
380 list_for_each_entry((root), &cgroup_roots, root_list)
382 /* iterate over child cgrps, lock should be held throughout iteration */
383 #define cgroup_for_each_live_child(child, cgrp) \
384 list_for_each_entry((child), &(cgrp)->children, sibling) \
385 if (({ lockdep_assert_held(&cgroup_tree_mutex); \
386 cgroup_is_dead(child); })) \
391 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
392 * @cgrp: the cgroup to be checked for liveness
394 * On success, returns true; the mutex should be later unlocked. On
395 * failure returns false with no lock held.
397 static bool cgroup_lock_live_group(struct cgroup
*cgrp
)
399 mutex_lock(&cgroup_mutex
);
400 if (cgroup_is_dead(cgrp
)) {
401 mutex_unlock(&cgroup_mutex
);
407 /* the list of cgroups eligible for automatic release. Protected by
408 * release_list_lock */
409 static LIST_HEAD(release_list
);
410 static DEFINE_RAW_SPINLOCK(release_list_lock
);
411 static void cgroup_release_agent(struct work_struct
*work
);
412 static DECLARE_WORK(release_agent_work
, cgroup_release_agent
);
413 static void check_for_release(struct cgroup
*cgrp
);
416 * A cgroup can be associated with multiple css_sets as different tasks may
417 * belong to different cgroups on different hierarchies. In the other
418 * direction, a css_set is naturally associated with multiple cgroups.
419 * This M:N relationship is represented by the following link structure
420 * which exists for each association and allows traversing the associations
423 struct cgrp_cset_link
{
424 /* the cgroup and css_set this link associates */
426 struct css_set
*cset
;
428 /* list of cgrp_cset_links anchored at cgrp->cset_links */
429 struct list_head cset_link
;
431 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
432 struct list_head cgrp_link
;
436 * The default css_set - used by init and its children prior to any
437 * hierarchies being mounted. It contains a pointer to the root state
438 * for each subsystem. Also used to anchor the list of css_sets. Not
439 * reference-counted, to improve performance when child cgroups
440 * haven't been created.
442 static struct css_set init_css_set
= {
443 .refcount
= ATOMIC_INIT(1),
444 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
445 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
446 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
447 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
448 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
451 static int css_set_count
= 1; /* 1 for init_css_set */
454 * cgroup_update_populated - updated populated count of a cgroup
455 * @cgrp: the target cgroup
456 * @populated: inc or dec populated count
458 * @cgrp is either getting the first task (css_set) or losing the last.
459 * Update @cgrp->populated_cnt accordingly. The count is propagated
460 * towards root so that a given cgroup's populated_cnt is zero iff the
461 * cgroup and all its descendants are empty.
463 * @cgrp's interface file "cgroup.populated" is zero if
464 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
465 * changes from or to zero, userland is notified that the content of the
466 * interface file has changed. This can be used to detect when @cgrp and
467 * its descendants become populated or empty.
469 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
471 lockdep_assert_held(&css_set_rwsem
);
477 trigger
= !cgrp
->populated_cnt
++;
479 trigger
= !--cgrp
->populated_cnt
;
484 if (cgrp
->populated_kn
)
485 kernfs_notify(cgrp
->populated_kn
);
491 * hash table for cgroup groups. This improves the performance to find
492 * an existing css_set. This hash doesn't (currently) take into
493 * account cgroups in empty hierarchies.
495 #define CSS_SET_HASH_BITS 7
496 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
498 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
500 unsigned long key
= 0UL;
501 struct cgroup_subsys
*ss
;
504 for_each_subsys(ss
, i
)
505 key
+= (unsigned long)css
[i
];
506 key
= (key
>> 16) ^ key
;
511 static void put_css_set_locked(struct css_set
*cset
, bool taskexit
)
513 struct cgrp_cset_link
*link
, *tmp_link
;
514 struct cgroup_subsys
*ss
;
517 lockdep_assert_held(&css_set_rwsem
);
519 if (!atomic_dec_and_test(&cset
->refcount
))
522 /* This css_set is dead. unlink it and release cgroup refcounts */
523 for_each_subsys(ss
, ssid
)
524 list_del(&cset
->e_cset_node
[ssid
]);
525 hash_del(&cset
->hlist
);
528 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
529 struct cgroup
*cgrp
= link
->cgrp
;
531 list_del(&link
->cset_link
);
532 list_del(&link
->cgrp_link
);
534 /* @cgrp can't go away while we're holding css_set_rwsem */
535 if (list_empty(&cgrp
->cset_links
)) {
536 cgroup_update_populated(cgrp
, false);
537 if (notify_on_release(cgrp
)) {
539 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
540 check_for_release(cgrp
);
547 kfree_rcu(cset
, rcu_head
);
550 static void put_css_set(struct css_set
*cset
, bool taskexit
)
553 * Ensure that the refcount doesn't hit zero while any readers
554 * can see it. Similar to atomic_dec_and_lock(), but for an
557 if (atomic_add_unless(&cset
->refcount
, -1, 1))
560 down_write(&css_set_rwsem
);
561 put_css_set_locked(cset
, taskexit
);
562 up_write(&css_set_rwsem
);
566 * refcounted get/put for css_set objects
568 static inline void get_css_set(struct css_set
*cset
)
570 atomic_inc(&cset
->refcount
);
574 * compare_css_sets - helper function for find_existing_css_set().
575 * @cset: candidate css_set being tested
576 * @old_cset: existing css_set for a task
577 * @new_cgrp: cgroup that's being entered by the task
578 * @template: desired set of css pointers in css_set (pre-calculated)
580 * Returns true if "cset" matches "old_cset" except for the hierarchy
581 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
583 static bool compare_css_sets(struct css_set
*cset
,
584 struct css_set
*old_cset
,
585 struct cgroup
*new_cgrp
,
586 struct cgroup_subsys_state
*template[])
588 struct list_head
*l1
, *l2
;
591 * On the default hierarchy, there can be csets which are
592 * associated with the same set of cgroups but different csses.
593 * Let's first ensure that csses match.
595 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
599 * Compare cgroup pointers in order to distinguish between
600 * different cgroups in hierarchies. As different cgroups may
601 * share the same effective css, this comparison is always
604 l1
= &cset
->cgrp_links
;
605 l2
= &old_cset
->cgrp_links
;
607 struct cgrp_cset_link
*link1
, *link2
;
608 struct cgroup
*cgrp1
, *cgrp2
;
612 /* See if we reached the end - both lists are equal length. */
613 if (l1
== &cset
->cgrp_links
) {
614 BUG_ON(l2
!= &old_cset
->cgrp_links
);
617 BUG_ON(l2
== &old_cset
->cgrp_links
);
619 /* Locate the cgroups associated with these links. */
620 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
621 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
624 /* Hierarchies should be linked in the same order. */
625 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
628 * If this hierarchy is the hierarchy of the cgroup
629 * that's changing, then we need to check that this
630 * css_set points to the new cgroup; if it's any other
631 * hierarchy, then this css_set should point to the
632 * same cgroup as the old css_set.
634 if (cgrp1
->root
== new_cgrp
->root
) {
635 if (cgrp1
!= new_cgrp
)
646 * find_existing_css_set - init css array and find the matching css_set
647 * @old_cset: the css_set that we're using before the cgroup transition
648 * @cgrp: the cgroup that we're moving into
649 * @template: out param for the new set of csses, should be clear on entry
651 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
653 struct cgroup_subsys_state
*template[])
655 struct cgroup_root
*root
= cgrp
->root
;
656 struct cgroup_subsys
*ss
;
657 struct css_set
*cset
;
662 * Build the set of subsystem state objects that we want to see in the
663 * new css_set. while subsystems can change globally, the entries here
664 * won't change, so no need for locking.
666 for_each_subsys(ss
, i
) {
667 if (root
->subsys_mask
& (1UL << i
)) {
669 * @ss is in this hierarchy, so we want the
670 * effective css from @cgrp.
672 template[i
] = cgroup_e_css(cgrp
, ss
);
675 * @ss is not in this hierarchy, so we don't want
678 template[i
] = old_cset
->subsys
[i
];
682 key
= css_set_hash(template);
683 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
684 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
687 /* This css_set matches what we need */
691 /* No existing cgroup group matched */
695 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
697 struct cgrp_cset_link
*link
, *tmp_link
;
699 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
700 list_del(&link
->cset_link
);
706 * allocate_cgrp_cset_links - allocate cgrp_cset_links
707 * @count: the number of links to allocate
708 * @tmp_links: list_head the allocated links are put on
710 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
711 * through ->cset_link. Returns 0 on success or -errno.
713 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
715 struct cgrp_cset_link
*link
;
718 INIT_LIST_HEAD(tmp_links
);
720 for (i
= 0; i
< count
; i
++) {
721 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
723 free_cgrp_cset_links(tmp_links
);
726 list_add(&link
->cset_link
, tmp_links
);
732 * link_css_set - a helper function to link a css_set to a cgroup
733 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
734 * @cset: the css_set to be linked
735 * @cgrp: the destination cgroup
737 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
740 struct cgrp_cset_link
*link
;
742 BUG_ON(list_empty(tmp_links
));
744 if (cgroup_on_dfl(cgrp
))
745 cset
->dfl_cgrp
= cgrp
;
747 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
751 if (list_empty(&cgrp
->cset_links
))
752 cgroup_update_populated(cgrp
, true);
753 list_move(&link
->cset_link
, &cgrp
->cset_links
);
756 * Always add links to the tail of the list so that the list
757 * is sorted by order of hierarchy creation
759 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
763 * find_css_set - return a new css_set with one cgroup updated
764 * @old_cset: the baseline css_set
765 * @cgrp: the cgroup to be updated
767 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
768 * substituted into the appropriate hierarchy.
770 static struct css_set
*find_css_set(struct css_set
*old_cset
,
773 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
774 struct css_set
*cset
;
775 struct list_head tmp_links
;
776 struct cgrp_cset_link
*link
;
777 struct cgroup_subsys
*ss
;
781 lockdep_assert_held(&cgroup_mutex
);
783 /* First see if we already have a cgroup group that matches
785 down_read(&css_set_rwsem
);
786 cset
= find_existing_css_set(old_cset
, cgrp
, template);
789 up_read(&css_set_rwsem
);
794 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
798 /* Allocate all the cgrp_cset_link objects that we'll need */
799 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
804 atomic_set(&cset
->refcount
, 1);
805 INIT_LIST_HEAD(&cset
->cgrp_links
);
806 INIT_LIST_HEAD(&cset
->tasks
);
807 INIT_LIST_HEAD(&cset
->mg_tasks
);
808 INIT_LIST_HEAD(&cset
->mg_preload_node
);
809 INIT_LIST_HEAD(&cset
->mg_node
);
810 INIT_HLIST_NODE(&cset
->hlist
);
812 /* Copy the set of subsystem state objects generated in
813 * find_existing_css_set() */
814 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
816 down_write(&css_set_rwsem
);
817 /* Add reference counts and links from the new css_set. */
818 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
819 struct cgroup
*c
= link
->cgrp
;
821 if (c
->root
== cgrp
->root
)
823 link_css_set(&tmp_links
, cset
, c
);
826 BUG_ON(!list_empty(&tmp_links
));
830 /* Add @cset to the hash table */
831 key
= css_set_hash(cset
->subsys
);
832 hash_add(css_set_table
, &cset
->hlist
, key
);
834 for_each_subsys(ss
, ssid
)
835 list_add_tail(&cset
->e_cset_node
[ssid
],
836 &cset
->subsys
[ssid
]->cgroup
->e_csets
[ssid
]);
838 up_write(&css_set_rwsem
);
843 static struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
845 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
847 return root_cgrp
->root
;
850 static int cgroup_init_root_id(struct cgroup_root
*root
)
854 lockdep_assert_held(&cgroup_mutex
);
856 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
860 root
->hierarchy_id
= id
;
864 static void cgroup_exit_root_id(struct cgroup_root
*root
)
866 lockdep_assert_held(&cgroup_mutex
);
868 if (root
->hierarchy_id
) {
869 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
870 root
->hierarchy_id
= 0;
874 static void cgroup_free_root(struct cgroup_root
*root
)
877 /* hierarhcy ID shoulid already have been released */
878 WARN_ON_ONCE(root
->hierarchy_id
);
880 idr_destroy(&root
->cgroup_idr
);
885 static void cgroup_destroy_root(struct cgroup_root
*root
)
887 struct cgroup
*cgrp
= &root
->cgrp
;
888 struct cgrp_cset_link
*link
, *tmp_link
;
890 mutex_lock(&cgroup_tree_mutex
);
891 mutex_lock(&cgroup_mutex
);
893 BUG_ON(atomic_read(&root
->nr_cgrps
));
894 BUG_ON(!list_empty(&cgrp
->children
));
896 /* Rebind all subsystems back to the default hierarchy */
897 rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
);
900 * Release all the links from cset_links to this hierarchy's
903 down_write(&css_set_rwsem
);
905 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
906 list_del(&link
->cset_link
);
907 list_del(&link
->cgrp_link
);
910 up_write(&css_set_rwsem
);
912 if (!list_empty(&root
->root_list
)) {
913 list_del(&root
->root_list
);
917 cgroup_exit_root_id(root
);
919 mutex_unlock(&cgroup_mutex
);
920 mutex_unlock(&cgroup_tree_mutex
);
922 kernfs_destroy_root(root
->kf_root
);
923 cgroup_free_root(root
);
926 /* look up cgroup associated with given css_set on the specified hierarchy */
927 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
928 struct cgroup_root
*root
)
930 struct cgroup
*res
= NULL
;
932 lockdep_assert_held(&cgroup_mutex
);
933 lockdep_assert_held(&css_set_rwsem
);
935 if (cset
== &init_css_set
) {
938 struct cgrp_cset_link
*link
;
940 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
941 struct cgroup
*c
= link
->cgrp
;
943 if (c
->root
== root
) {
955 * Return the cgroup for "task" from the given hierarchy. Must be
956 * called with cgroup_mutex and css_set_rwsem held.
958 static struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
959 struct cgroup_root
*root
)
962 * No need to lock the task - since we hold cgroup_mutex the
963 * task can't change groups, so the only thing that can happen
964 * is that it exits and its css is set back to init_css_set.
966 return cset_cgroup_from_root(task_css_set(task
), root
);
970 * A task must hold cgroup_mutex to modify cgroups.
972 * Any task can increment and decrement the count field without lock.
973 * So in general, code holding cgroup_mutex can't rely on the count
974 * field not changing. However, if the count goes to zero, then only
975 * cgroup_attach_task() can increment it again. Because a count of zero
976 * means that no tasks are currently attached, therefore there is no
977 * way a task attached to that cgroup can fork (the other way to
978 * increment the count). So code holding cgroup_mutex can safely
979 * assume that if the count is zero, it will stay zero. Similarly, if
980 * a task holds cgroup_mutex on a cgroup with zero count, it
981 * knows that the cgroup won't be removed, as cgroup_rmdir()
984 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
985 * (usually) take cgroup_mutex. These are the two most performance
986 * critical pieces of code here. The exception occurs on cgroup_exit(),
987 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
988 * is taken, and if the cgroup count is zero, a usermode call made
989 * to the release agent with the name of the cgroup (path relative to
990 * the root of cgroup file system) as the argument.
992 * A cgroup can only be deleted if both its 'count' of using tasks
993 * is zero, and its list of 'children' cgroups is empty. Since all
994 * tasks in the system use _some_ cgroup, and since there is always at
995 * least one task in the system (init, pid == 1), therefore, root cgroup
996 * always has either children cgroups and/or using tasks. So we don't
997 * need a special hack to ensure that root cgroup cannot be deleted.
999 * P.S. One more locking exception. RCU is used to guard the
1000 * update of a tasks cgroup pointer by cgroup_attach_task()
1003 static int cgroup_populate_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
);
1004 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1005 static const struct file_operations proc_cgroupstats_operations
;
1007 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1010 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1011 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
1012 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
1013 cft
->ss
->name
, cft
->name
);
1015 strncpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1020 * cgroup_file_mode - deduce file mode of a control file
1021 * @cft: the control file in question
1023 * returns cft->mode if ->mode is not 0
1024 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1025 * returns S_IRUGO if it has only a read handler
1026 * returns S_IWUSR if it has only a write hander
1028 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1035 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1038 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write_string
||
1045 static void cgroup_free_fn(struct work_struct
*work
)
1047 struct cgroup
*cgrp
= container_of(work
, struct cgroup
, destroy_work
);
1049 atomic_dec(&cgrp
->root
->nr_cgrps
);
1050 cgroup_pidlist_destroy_all(cgrp
);
1054 * We get a ref to the parent, and put the ref when this
1055 * cgroup is being freed, so it's guaranteed that the
1056 * parent won't be destroyed before its children.
1058 cgroup_put(cgrp
->parent
);
1059 kernfs_put(cgrp
->kn
);
1063 * This is root cgroup's refcnt reaching zero, which
1064 * indicates that the root should be released.
1066 cgroup_destroy_root(cgrp
->root
);
1070 static void cgroup_free_rcu(struct rcu_head
*head
)
1072 struct cgroup
*cgrp
= container_of(head
, struct cgroup
, rcu_head
);
1074 INIT_WORK(&cgrp
->destroy_work
, cgroup_free_fn
);
1075 queue_work(cgroup_destroy_wq
, &cgrp
->destroy_work
);
1078 static void cgroup_get(struct cgroup
*cgrp
)
1080 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
1081 WARN_ON_ONCE(atomic_read(&cgrp
->refcnt
) <= 0);
1082 atomic_inc(&cgrp
->refcnt
);
1085 static void cgroup_put(struct cgroup
*cgrp
)
1087 if (!atomic_dec_and_test(&cgrp
->refcnt
))
1089 if (WARN_ON_ONCE(cgrp
->parent
&& !cgroup_is_dead(cgrp
)))
1092 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
1095 call_rcu(&cgrp
->rcu_head
, cgroup_free_rcu
);
1098 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1100 char name
[CGROUP_FILE_NAME_MAX
];
1102 lockdep_assert_held(&cgroup_tree_mutex
);
1103 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1107 * cgroup_clear_dir - remove subsys files in a cgroup directory
1108 * @cgrp: target cgroup
1109 * @subsys_mask: mask of the subsystem ids whose files should be removed
1111 static void cgroup_clear_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
)
1113 struct cgroup_subsys
*ss
;
1116 for_each_subsys(ss
, i
) {
1117 struct cftype
*cfts
;
1119 if (!(subsys_mask
& (1 << i
)))
1121 list_for_each_entry(cfts
, &ss
->cfts
, node
)
1122 cgroup_addrm_files(cgrp
, cfts
, false);
1126 static int rebind_subsystems(struct cgroup_root
*dst_root
, unsigned int ss_mask
)
1128 struct cgroup_subsys
*ss
;
1131 lockdep_assert_held(&cgroup_tree_mutex
);
1132 lockdep_assert_held(&cgroup_mutex
);
1134 for_each_subsys(ss
, ssid
) {
1135 if (!(ss_mask
& (1 << ssid
)))
1138 /* if @ss has non-root csses attached to it, can't move */
1139 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)))
1142 /* can't move between two non-dummy roots either */
1143 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1147 ret
= cgroup_populate_dir(&dst_root
->cgrp
, ss_mask
);
1149 if (dst_root
!= &cgrp_dfl_root
)
1153 * Rebinding back to the default root is not allowed to
1154 * fail. Using both default and non-default roots should
1155 * be rare. Moving subsystems back and forth even more so.
1156 * Just warn about it and continue.
1158 if (cgrp_dfl_root_visible
) {
1159 pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1161 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1166 * Nothing can fail from this point on. Remove files for the
1167 * removed subsystems and rebind each subsystem.
1169 mutex_unlock(&cgroup_mutex
);
1170 for_each_subsys(ss
, ssid
)
1171 if (ss_mask
& (1 << ssid
))
1172 cgroup_clear_dir(&ss
->root
->cgrp
, 1 << ssid
);
1173 mutex_lock(&cgroup_mutex
);
1175 for_each_subsys(ss
, ssid
) {
1176 struct cgroup_root
*src_root
;
1177 struct cgroup_subsys_state
*css
;
1178 struct css_set
*cset
;
1180 if (!(ss_mask
& (1 << ssid
)))
1183 src_root
= ss
->root
;
1184 css
= cgroup_css(&src_root
->cgrp
, ss
);
1186 WARN_ON(!css
|| cgroup_css(&dst_root
->cgrp
, ss
));
1188 RCU_INIT_POINTER(src_root
->cgrp
.subsys
[ssid
], NULL
);
1189 rcu_assign_pointer(dst_root
->cgrp
.subsys
[ssid
], css
);
1190 ss
->root
= dst_root
;
1191 css
->cgroup
= &dst_root
->cgrp
;
1193 down_write(&css_set_rwsem
);
1194 hash_for_each(css_set_table
, i
, cset
, hlist
)
1195 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1196 &dst_root
->cgrp
.e_csets
[ss
->id
]);
1197 up_write(&css_set_rwsem
);
1199 src_root
->subsys_mask
&= ~(1 << ssid
);
1200 src_root
->cgrp
.child_subsys_mask
&= ~(1 << ssid
);
1202 /* default hierarchy doesn't enable controllers by default */
1203 dst_root
->subsys_mask
|= 1 << ssid
;
1204 if (dst_root
!= &cgrp_dfl_root
)
1205 dst_root
->cgrp
.child_subsys_mask
|= 1 << ssid
;
1211 kernfs_activate(dst_root
->cgrp
.kn
);
1215 static int cgroup_show_options(struct seq_file
*seq
,
1216 struct kernfs_root
*kf_root
)
1218 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1219 struct cgroup_subsys
*ss
;
1222 for_each_subsys(ss
, ssid
)
1223 if (root
->subsys_mask
& (1 << ssid
))
1224 seq_printf(seq
, ",%s", ss
->name
);
1225 if (root
->flags
& CGRP_ROOT_SANE_BEHAVIOR
)
1226 seq_puts(seq
, ",sane_behavior");
1227 if (root
->flags
& CGRP_ROOT_NOPREFIX
)
1228 seq_puts(seq
, ",noprefix");
1229 if (root
->flags
& CGRP_ROOT_XATTR
)
1230 seq_puts(seq
, ",xattr");
1232 spin_lock(&release_agent_path_lock
);
1233 if (strlen(root
->release_agent_path
))
1234 seq_printf(seq
, ",release_agent=%s", root
->release_agent_path
);
1235 spin_unlock(&release_agent_path_lock
);
1237 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
))
1238 seq_puts(seq
, ",clone_children");
1239 if (strlen(root
->name
))
1240 seq_printf(seq
, ",name=%s", root
->name
);
1244 struct cgroup_sb_opts
{
1245 unsigned int subsys_mask
;
1247 char *release_agent
;
1248 bool cpuset_clone_children
;
1250 /* User explicitly requested empty subsystem */
1254 static int parse_cgroupfs_options(char *data
, struct cgroup_sb_opts
*opts
)
1256 char *token
, *o
= data
;
1257 bool all_ss
= false, one_ss
= false;
1258 unsigned int mask
= -1U;
1259 struct cgroup_subsys
*ss
;
1262 #ifdef CONFIG_CPUSETS
1263 mask
= ~(1U << cpuset_cgrp_id
);
1266 memset(opts
, 0, sizeof(*opts
));
1268 while ((token
= strsep(&o
, ",")) != NULL
) {
1271 if (!strcmp(token
, "none")) {
1272 /* Explicitly have no subsystems */
1276 if (!strcmp(token
, "all")) {
1277 /* Mutually exclusive option 'all' + subsystem name */
1283 if (!strcmp(token
, "__DEVEL__sane_behavior")) {
1284 opts
->flags
|= CGRP_ROOT_SANE_BEHAVIOR
;
1287 if (!strcmp(token
, "noprefix")) {
1288 opts
->flags
|= CGRP_ROOT_NOPREFIX
;
1291 if (!strcmp(token
, "clone_children")) {
1292 opts
->cpuset_clone_children
= true;
1295 if (!strcmp(token
, "xattr")) {
1296 opts
->flags
|= CGRP_ROOT_XATTR
;
1299 if (!strncmp(token
, "release_agent=", 14)) {
1300 /* Specifying two release agents is forbidden */
1301 if (opts
->release_agent
)
1303 opts
->release_agent
=
1304 kstrndup(token
+ 14, PATH_MAX
- 1, GFP_KERNEL
);
1305 if (!opts
->release_agent
)
1309 if (!strncmp(token
, "name=", 5)) {
1310 const char *name
= token
+ 5;
1311 /* Can't specify an empty name */
1314 /* Must match [\w.-]+ */
1315 for (i
= 0; i
< strlen(name
); i
++) {
1319 if ((c
== '.') || (c
== '-') || (c
== '_'))
1323 /* Specifying two names is forbidden */
1326 opts
->name
= kstrndup(name
,
1327 MAX_CGROUP_ROOT_NAMELEN
- 1,
1335 for_each_subsys(ss
, i
) {
1336 if (strcmp(token
, ss
->name
))
1341 /* Mutually exclusive option 'all' + subsystem name */
1344 opts
->subsys_mask
|= (1 << i
);
1349 if (i
== CGROUP_SUBSYS_COUNT
)
1353 /* Consistency checks */
1355 if (opts
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1356 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1358 if ((opts
->flags
& (CGRP_ROOT_NOPREFIX
| CGRP_ROOT_XATTR
)) ||
1359 opts
->cpuset_clone_children
|| opts
->release_agent
||
1361 pr_err("sane_behavior: noprefix, xattr, clone_children, release_agent and name are not allowed\n");
1366 * If the 'all' option was specified select all the
1367 * subsystems, otherwise if 'none', 'name=' and a subsystem
1368 * name options were not specified, let's default to 'all'
1370 if (all_ss
|| (!one_ss
&& !opts
->none
&& !opts
->name
))
1371 for_each_subsys(ss
, i
)
1373 opts
->subsys_mask
|= (1 << i
);
1376 * We either have to specify by name or by subsystems. (So
1377 * all empty hierarchies must have a name).
1379 if (!opts
->subsys_mask
&& !opts
->name
)
1384 * Option noprefix was introduced just for backward compatibility
1385 * with the old cpuset, so we allow noprefix only if mounting just
1386 * the cpuset subsystem.
1388 if ((opts
->flags
& CGRP_ROOT_NOPREFIX
) && (opts
->subsys_mask
& mask
))
1392 /* Can't specify "none" and some subsystems */
1393 if (opts
->subsys_mask
&& opts
->none
)
1399 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1402 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1403 struct cgroup_sb_opts opts
;
1404 unsigned int added_mask
, removed_mask
;
1406 if (root
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1407 pr_err("sane_behavior: remount is not allowed\n");
1411 mutex_lock(&cgroup_tree_mutex
);
1412 mutex_lock(&cgroup_mutex
);
1414 /* See what subsystems are wanted */
1415 ret
= parse_cgroupfs_options(data
, &opts
);
1419 if (opts
.subsys_mask
!= root
->subsys_mask
|| opts
.release_agent
)
1420 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1421 task_tgid_nr(current
), current
->comm
);
1423 added_mask
= opts
.subsys_mask
& ~root
->subsys_mask
;
1424 removed_mask
= root
->subsys_mask
& ~opts
.subsys_mask
;
1426 /* Don't allow flags or name to change at remount */
1427 if (((opts
.flags
^ root
->flags
) & CGRP_ROOT_OPTION_MASK
) ||
1428 (opts
.name
&& strcmp(opts
.name
, root
->name
))) {
1429 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1430 opts
.flags
& CGRP_ROOT_OPTION_MASK
, opts
.name
?: "",
1431 root
->flags
& CGRP_ROOT_OPTION_MASK
, root
->name
);
1436 /* remounting is not allowed for populated hierarchies */
1437 if (!list_empty(&root
->cgrp
.children
)) {
1442 ret
= rebind_subsystems(root
, added_mask
);
1446 rebind_subsystems(&cgrp_dfl_root
, removed_mask
);
1448 if (opts
.release_agent
) {
1449 spin_lock(&release_agent_path_lock
);
1450 strcpy(root
->release_agent_path
, opts
.release_agent
);
1451 spin_unlock(&release_agent_path_lock
);
1454 kfree(opts
.release_agent
);
1456 mutex_unlock(&cgroup_mutex
);
1457 mutex_unlock(&cgroup_tree_mutex
);
1462 * To reduce the fork() overhead for systems that are not actually using
1463 * their cgroups capability, we don't maintain the lists running through
1464 * each css_set to its tasks until we see the list actually used - in other
1465 * words after the first mount.
1467 static bool use_task_css_set_links __read_mostly
;
1469 static void cgroup_enable_task_cg_lists(void)
1471 struct task_struct
*p
, *g
;
1473 down_write(&css_set_rwsem
);
1475 if (use_task_css_set_links
)
1478 use_task_css_set_links
= true;
1481 * We need tasklist_lock because RCU is not safe against
1482 * while_each_thread(). Besides, a forking task that has passed
1483 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1484 * is not guaranteed to have its child immediately visible in the
1485 * tasklist if we walk through it with RCU.
1487 read_lock(&tasklist_lock
);
1488 do_each_thread(g
, p
) {
1489 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1490 task_css_set(p
) != &init_css_set
);
1493 * We should check if the process is exiting, otherwise
1494 * it will race with cgroup_exit() in that the list
1495 * entry won't be deleted though the process has exited.
1496 * Do it while holding siglock so that we don't end up
1497 * racing against cgroup_exit().
1499 spin_lock_irq(&p
->sighand
->siglock
);
1500 if (!(p
->flags
& PF_EXITING
)) {
1501 struct css_set
*cset
= task_css_set(p
);
1503 list_add(&p
->cg_list
, &cset
->tasks
);
1506 spin_unlock_irq(&p
->sighand
->siglock
);
1507 } while_each_thread(g
, p
);
1508 read_unlock(&tasklist_lock
);
1510 up_write(&css_set_rwsem
);
1513 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1515 struct cgroup_subsys
*ss
;
1518 atomic_set(&cgrp
->refcnt
, 1);
1519 INIT_LIST_HEAD(&cgrp
->sibling
);
1520 INIT_LIST_HEAD(&cgrp
->children
);
1521 INIT_LIST_HEAD(&cgrp
->cset_links
);
1522 INIT_LIST_HEAD(&cgrp
->release_list
);
1523 INIT_LIST_HEAD(&cgrp
->pidlists
);
1524 mutex_init(&cgrp
->pidlist_mutex
);
1525 cgrp
->dummy_css
.cgroup
= cgrp
;
1527 for_each_subsys(ss
, ssid
)
1528 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1530 init_waitqueue_head(&cgrp
->offline_waitq
);
1533 static void init_cgroup_root(struct cgroup_root
*root
,
1534 struct cgroup_sb_opts
*opts
)
1536 struct cgroup
*cgrp
= &root
->cgrp
;
1538 INIT_LIST_HEAD(&root
->root_list
);
1539 atomic_set(&root
->nr_cgrps
, 1);
1541 init_cgroup_housekeeping(cgrp
);
1542 idr_init(&root
->cgroup_idr
);
1544 root
->flags
= opts
->flags
;
1545 if (opts
->release_agent
)
1546 strcpy(root
->release_agent_path
, opts
->release_agent
);
1548 strcpy(root
->name
, opts
->name
);
1549 if (opts
->cpuset_clone_children
)
1550 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1553 static int cgroup_setup_root(struct cgroup_root
*root
, unsigned int ss_mask
)
1555 LIST_HEAD(tmp_links
);
1556 struct cgroup
*root_cgrp
= &root
->cgrp
;
1557 struct css_set
*cset
;
1560 lockdep_assert_held(&cgroup_tree_mutex
);
1561 lockdep_assert_held(&cgroup_mutex
);
1563 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_NOWAIT
);
1566 root_cgrp
->id
= ret
;
1569 * We're accessing css_set_count without locking css_set_rwsem here,
1570 * but that's OK - it can only be increased by someone holding
1571 * cgroup_lock, and that's us. The worst that can happen is that we
1572 * have some link structures left over
1574 ret
= allocate_cgrp_cset_links(css_set_count
, &tmp_links
);
1578 ret
= cgroup_init_root_id(root
);
1582 root
->kf_root
= kernfs_create_root(&cgroup_kf_syscall_ops
,
1583 KERNFS_ROOT_CREATE_DEACTIVATED
,
1585 if (IS_ERR(root
->kf_root
)) {
1586 ret
= PTR_ERR(root
->kf_root
);
1589 root_cgrp
->kn
= root
->kf_root
->kn
;
1591 ret
= cgroup_addrm_files(root_cgrp
, cgroup_base_files
, true);
1595 ret
= rebind_subsystems(root
, ss_mask
);
1600 * There must be no failure case after here, since rebinding takes
1601 * care of subsystems' refcounts, which are explicitly dropped in
1602 * the failure exit path.
1604 list_add(&root
->root_list
, &cgroup_roots
);
1605 cgroup_root_count
++;
1608 * Link the root cgroup in this hierarchy into all the css_set
1611 down_write(&css_set_rwsem
);
1612 hash_for_each(css_set_table
, i
, cset
, hlist
)
1613 link_css_set(&tmp_links
, cset
, root_cgrp
);
1614 up_write(&css_set_rwsem
);
1616 BUG_ON(!list_empty(&root_cgrp
->children
));
1617 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1619 kernfs_activate(root_cgrp
->kn
);
1624 kernfs_destroy_root(root
->kf_root
);
1625 root
->kf_root
= NULL
;
1627 cgroup_exit_root_id(root
);
1629 free_cgrp_cset_links(&tmp_links
);
1633 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
1634 int flags
, const char *unused_dev_name
,
1637 struct cgroup_root
*root
;
1638 struct cgroup_sb_opts opts
;
1639 struct dentry
*dentry
;
1644 * The first time anyone tries to mount a cgroup, enable the list
1645 * linking each css_set to its tasks and fix up all existing tasks.
1647 if (!use_task_css_set_links
)
1648 cgroup_enable_task_cg_lists();
1650 mutex_lock(&cgroup_tree_mutex
);
1651 mutex_lock(&cgroup_mutex
);
1653 /* First find the desired set of subsystems */
1654 ret
= parse_cgroupfs_options(data
, &opts
);
1658 /* look for a matching existing root */
1659 if (!opts
.subsys_mask
&& !opts
.none
&& !opts
.name
) {
1660 cgrp_dfl_root_visible
= true;
1661 root
= &cgrp_dfl_root
;
1662 cgroup_get(&root
->cgrp
);
1667 for_each_root(root
) {
1668 bool name_match
= false;
1670 if (root
== &cgrp_dfl_root
)
1674 * If we asked for a name then it must match. Also, if
1675 * name matches but sybsys_mask doesn't, we should fail.
1676 * Remember whether name matched.
1679 if (strcmp(opts
.name
, root
->name
))
1685 * If we asked for subsystems (or explicitly for no
1686 * subsystems) then they must match.
1688 if ((opts
.subsys_mask
|| opts
.none
) &&
1689 (opts
.subsys_mask
!= root
->subsys_mask
)) {
1696 if ((root
->flags
^ opts
.flags
) & CGRP_ROOT_OPTION_MASK
) {
1697 if ((root
->flags
| opts
.flags
) & CGRP_ROOT_SANE_BEHAVIOR
) {
1698 pr_err("sane_behavior: new mount options should match the existing superblock\n");
1702 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1707 * A root's lifetime is governed by its root cgroup. Zero
1708 * ref indicate that the root is being destroyed. Wait for
1709 * destruction to complete so that the subsystems are free.
1710 * We can use wait_queue for the wait but this path is
1711 * super cold. Let's just sleep for a bit and retry.
1713 if (!atomic_inc_not_zero(&root
->cgrp
.refcnt
)) {
1714 mutex_unlock(&cgroup_mutex
);
1715 mutex_unlock(&cgroup_tree_mutex
);
1717 mutex_lock(&cgroup_tree_mutex
);
1718 mutex_lock(&cgroup_mutex
);
1727 * No such thing, create a new one. name= matching without subsys
1728 * specification is allowed for already existing hierarchies but we
1729 * can't create new one without subsys specification.
1731 if (!opts
.subsys_mask
&& !opts
.none
) {
1736 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
1742 init_cgroup_root(root
, &opts
);
1744 ret
= cgroup_setup_root(root
, opts
.subsys_mask
);
1746 cgroup_free_root(root
);
1749 mutex_unlock(&cgroup_mutex
);
1750 mutex_unlock(&cgroup_tree_mutex
);
1752 kfree(opts
.release_agent
);
1756 return ERR_PTR(ret
);
1758 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
, &new_sb
);
1759 if (IS_ERR(dentry
) || !new_sb
)
1760 cgroup_put(&root
->cgrp
);
1764 static void cgroup_kill_sb(struct super_block
*sb
)
1766 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
1767 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1769 cgroup_put(&root
->cgrp
);
1773 static struct file_system_type cgroup_fs_type
= {
1775 .mount
= cgroup_mount
,
1776 .kill_sb
= cgroup_kill_sb
,
1779 static struct kobject
*cgroup_kobj
;
1782 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1783 * @task: target task
1784 * @buf: the buffer to write the path into
1785 * @buflen: the length of the buffer
1787 * Determine @task's cgroup on the first (the one with the lowest non-zero
1788 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1789 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1790 * cgroup controller callbacks.
1792 * Return value is the same as kernfs_path().
1794 char *task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
1796 struct cgroup_root
*root
;
1797 struct cgroup
*cgrp
;
1798 int hierarchy_id
= 1;
1801 mutex_lock(&cgroup_mutex
);
1802 down_read(&css_set_rwsem
);
1804 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
1807 cgrp
= task_cgroup_from_root(task
, root
);
1808 path
= cgroup_path(cgrp
, buf
, buflen
);
1810 /* if no hierarchy exists, everyone is in "/" */
1811 if (strlcpy(buf
, "/", buflen
) < buflen
)
1815 up_read(&css_set_rwsem
);
1816 mutex_unlock(&cgroup_mutex
);
1819 EXPORT_SYMBOL_GPL(task_cgroup_path
);
1821 /* used to track tasks and other necessary states during migration */
1822 struct cgroup_taskset
{
1823 /* the src and dst cset list running through cset->mg_node */
1824 struct list_head src_csets
;
1825 struct list_head dst_csets
;
1828 * Fields for cgroup_taskset_*() iteration.
1830 * Before migration is committed, the target migration tasks are on
1831 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1832 * the csets on ->dst_csets. ->csets point to either ->src_csets
1833 * or ->dst_csets depending on whether migration is committed.
1835 * ->cur_csets and ->cur_task point to the current task position
1838 struct list_head
*csets
;
1839 struct css_set
*cur_cset
;
1840 struct task_struct
*cur_task
;
1844 * cgroup_taskset_first - reset taskset and return the first task
1845 * @tset: taskset of interest
1847 * @tset iteration is initialized and the first task is returned.
1849 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
)
1851 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
1852 tset
->cur_task
= NULL
;
1854 return cgroup_taskset_next(tset
);
1858 * cgroup_taskset_next - iterate to the next task in taskset
1859 * @tset: taskset of interest
1861 * Return the next task in @tset. Iteration must have been initialized
1862 * with cgroup_taskset_first().
1864 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
)
1866 struct css_set
*cset
= tset
->cur_cset
;
1867 struct task_struct
*task
= tset
->cur_task
;
1869 while (&cset
->mg_node
!= tset
->csets
) {
1871 task
= list_first_entry(&cset
->mg_tasks
,
1872 struct task_struct
, cg_list
);
1874 task
= list_next_entry(task
, cg_list
);
1876 if (&task
->cg_list
!= &cset
->mg_tasks
) {
1877 tset
->cur_cset
= cset
;
1878 tset
->cur_task
= task
;
1882 cset
= list_next_entry(cset
, mg_node
);
1890 * cgroup_task_migrate - move a task from one cgroup to another.
1891 * @old_cgrp; the cgroup @tsk is being migrated from
1892 * @tsk: the task being migrated
1893 * @new_cset: the new css_set @tsk is being attached to
1895 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
1897 static void cgroup_task_migrate(struct cgroup
*old_cgrp
,
1898 struct task_struct
*tsk
,
1899 struct css_set
*new_cset
)
1901 struct css_set
*old_cset
;
1903 lockdep_assert_held(&cgroup_mutex
);
1904 lockdep_assert_held(&css_set_rwsem
);
1907 * We are synchronized through threadgroup_lock() against PF_EXITING
1908 * setting such that we can't race against cgroup_exit() changing the
1909 * css_set to init_css_set and dropping the old one.
1911 WARN_ON_ONCE(tsk
->flags
& PF_EXITING
);
1912 old_cset
= task_css_set(tsk
);
1914 get_css_set(new_cset
);
1915 rcu_assign_pointer(tsk
->cgroups
, new_cset
);
1918 * Use move_tail so that cgroup_taskset_first() still returns the
1919 * leader after migration. This works because cgroup_migrate()
1920 * ensures that the dst_cset of the leader is the first on the
1921 * tset's dst_csets list.
1923 list_move_tail(&tsk
->cg_list
, &new_cset
->mg_tasks
);
1926 * We just gained a reference on old_cset by taking it from the
1927 * task. As trading it for new_cset is protected by cgroup_mutex,
1928 * we're safe to drop it here; it will be freed under RCU.
1930 set_bit(CGRP_RELEASABLE
, &old_cgrp
->flags
);
1931 put_css_set_locked(old_cset
, false);
1935 * cgroup_migrate_finish - cleanup after attach
1936 * @preloaded_csets: list of preloaded css_sets
1938 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
1939 * those functions for details.
1941 static void cgroup_migrate_finish(struct list_head
*preloaded_csets
)
1943 struct css_set
*cset
, *tmp_cset
;
1945 lockdep_assert_held(&cgroup_mutex
);
1947 down_write(&css_set_rwsem
);
1948 list_for_each_entry_safe(cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
1949 cset
->mg_src_cgrp
= NULL
;
1950 cset
->mg_dst_cset
= NULL
;
1951 list_del_init(&cset
->mg_preload_node
);
1952 put_css_set_locked(cset
, false);
1954 up_write(&css_set_rwsem
);
1958 * cgroup_migrate_add_src - add a migration source css_set
1959 * @src_cset: the source css_set to add
1960 * @dst_cgrp: the destination cgroup
1961 * @preloaded_csets: list of preloaded css_sets
1963 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
1964 * @src_cset and add it to @preloaded_csets, which should later be cleaned
1965 * up by cgroup_migrate_finish().
1967 * This function may be called without holding threadgroup_lock even if the
1968 * target is a process. Threads may be created and destroyed but as long
1969 * as cgroup_mutex is not dropped, no new css_set can be put into play and
1970 * the preloaded css_sets are guaranteed to cover all migrations.
1972 static void cgroup_migrate_add_src(struct css_set
*src_cset
,
1973 struct cgroup
*dst_cgrp
,
1974 struct list_head
*preloaded_csets
)
1976 struct cgroup
*src_cgrp
;
1978 lockdep_assert_held(&cgroup_mutex
);
1979 lockdep_assert_held(&css_set_rwsem
);
1981 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
1983 if (!list_empty(&src_cset
->mg_preload_node
))
1986 WARN_ON(src_cset
->mg_src_cgrp
);
1987 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
1988 WARN_ON(!list_empty(&src_cset
->mg_node
));
1990 src_cset
->mg_src_cgrp
= src_cgrp
;
1991 get_css_set(src_cset
);
1992 list_add(&src_cset
->mg_preload_node
, preloaded_csets
);
1996 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
1997 * @dst_cgrp: the destination cgroup (may be %NULL)
1998 * @preloaded_csets: list of preloaded source css_sets
2000 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2001 * have been preloaded to @preloaded_csets. This function looks up and
2002 * pins all destination css_sets, links each to its source, and append them
2003 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2004 * source css_set is assumed to be its cgroup on the default hierarchy.
2006 * This function must be called after cgroup_migrate_add_src() has been
2007 * called on each migration source css_set. After migration is performed
2008 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2011 static int cgroup_migrate_prepare_dst(struct cgroup
*dst_cgrp
,
2012 struct list_head
*preloaded_csets
)
2015 struct css_set
*src_cset
, *tmp_cset
;
2017 lockdep_assert_held(&cgroup_mutex
);
2020 * Except for the root, child_subsys_mask must be zero for a cgroup
2021 * with tasks so that child cgroups don't compete against tasks.
2023 if (dst_cgrp
&& cgroup_on_dfl(dst_cgrp
) && dst_cgrp
->parent
&&
2024 dst_cgrp
->child_subsys_mask
)
2027 /* look up the dst cset for each src cset and link it to src */
2028 list_for_each_entry_safe(src_cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2029 struct css_set
*dst_cset
;
2031 dst_cset
= find_css_set(src_cset
,
2032 dst_cgrp
?: src_cset
->dfl_cgrp
);
2036 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2039 * If src cset equals dst, it's noop. Drop the src.
2040 * cgroup_migrate() will skip the cset too. Note that we
2041 * can't handle src == dst as some nodes are used by both.
2043 if (src_cset
== dst_cset
) {
2044 src_cset
->mg_src_cgrp
= NULL
;
2045 list_del_init(&src_cset
->mg_preload_node
);
2046 put_css_set(src_cset
, false);
2047 put_css_set(dst_cset
, false);
2051 src_cset
->mg_dst_cset
= dst_cset
;
2053 if (list_empty(&dst_cset
->mg_preload_node
))
2054 list_add(&dst_cset
->mg_preload_node
, &csets
);
2056 put_css_set(dst_cset
, false);
2059 list_splice_tail(&csets
, preloaded_csets
);
2062 cgroup_migrate_finish(&csets
);
2067 * cgroup_migrate - migrate a process or task to a cgroup
2068 * @cgrp: the destination cgroup
2069 * @leader: the leader of the process or the task to migrate
2070 * @threadgroup: whether @leader points to the whole process or a single task
2072 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2073 * process, the caller must be holding threadgroup_lock of @leader. The
2074 * caller is also responsible for invoking cgroup_migrate_add_src() and
2075 * cgroup_migrate_prepare_dst() on the targets before invoking this
2076 * function and following up with cgroup_migrate_finish().
2078 * As long as a controller's ->can_attach() doesn't fail, this function is
2079 * guaranteed to succeed. This means that, excluding ->can_attach()
2080 * failure, when migrating multiple targets, the success or failure can be
2081 * decided for all targets by invoking group_migrate_prepare_dst() before
2082 * actually starting migrating.
2084 static int cgroup_migrate(struct cgroup
*cgrp
, struct task_struct
*leader
,
2087 struct cgroup_taskset tset
= {
2088 .src_csets
= LIST_HEAD_INIT(tset
.src_csets
),
2089 .dst_csets
= LIST_HEAD_INIT(tset
.dst_csets
),
2090 .csets
= &tset
.src_csets
,
2092 struct cgroup_subsys_state
*css
, *failed_css
= NULL
;
2093 struct css_set
*cset
, *tmp_cset
;
2094 struct task_struct
*task
, *tmp_task
;
2098 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2099 * already PF_EXITING could be freed from underneath us unless we
2100 * take an rcu_read_lock.
2102 down_write(&css_set_rwsem
);
2106 /* @task either already exited or can't exit until the end */
2107 if (task
->flags
& PF_EXITING
)
2110 /* leave @task alone if post_fork() hasn't linked it yet */
2111 if (list_empty(&task
->cg_list
))
2114 cset
= task_css_set(task
);
2115 if (!cset
->mg_src_cgrp
)
2119 * cgroup_taskset_first() must always return the leader.
2120 * Take care to avoid disturbing the ordering.
2122 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2123 if (list_empty(&cset
->mg_node
))
2124 list_add_tail(&cset
->mg_node
, &tset
.src_csets
);
2125 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2126 list_move_tail(&cset
->mg_dst_cset
->mg_node
,
2131 } while_each_thread(leader
, task
);
2133 up_write(&css_set_rwsem
);
2135 /* methods shouldn't be called if no task is actually migrating */
2136 if (list_empty(&tset
.src_csets
))
2139 /* check that we can legitimately attach to the cgroup */
2140 for_each_e_css(css
, i
, cgrp
) {
2141 if (css
->ss
->can_attach
) {
2142 ret
= css
->ss
->can_attach(css
, &tset
);
2145 goto out_cancel_attach
;
2151 * Now that we're guaranteed success, proceed to move all tasks to
2152 * the new cgroup. There are no failure cases after here, so this
2153 * is the commit point.
2155 down_write(&css_set_rwsem
);
2156 list_for_each_entry(cset
, &tset
.src_csets
, mg_node
) {
2157 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
)
2158 cgroup_task_migrate(cset
->mg_src_cgrp
, task
,
2161 up_write(&css_set_rwsem
);
2164 * Migration is committed, all target tasks are now on dst_csets.
2165 * Nothing is sensitive to fork() after this point. Notify
2166 * controllers that migration is complete.
2168 tset
.csets
= &tset
.dst_csets
;
2170 for_each_e_css(css
, i
, cgrp
)
2171 if (css
->ss
->attach
)
2172 css
->ss
->attach(css
, &tset
);
2175 goto out_release_tset
;
2178 for_each_e_css(css
, i
, cgrp
) {
2179 if (css
== failed_css
)
2181 if (css
->ss
->cancel_attach
)
2182 css
->ss
->cancel_attach(css
, &tset
);
2185 down_write(&css_set_rwsem
);
2186 list_splice_init(&tset
.dst_csets
, &tset
.src_csets
);
2187 list_for_each_entry_safe(cset
, tmp_cset
, &tset
.src_csets
, mg_node
) {
2188 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2189 list_del_init(&cset
->mg_node
);
2191 up_write(&css_set_rwsem
);
2196 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2197 * @dst_cgrp: the cgroup to attach to
2198 * @leader: the task or the leader of the threadgroup to be attached
2199 * @threadgroup: attach the whole threadgroup?
2201 * Call holding cgroup_mutex and threadgroup_lock of @leader.
2203 static int cgroup_attach_task(struct cgroup
*dst_cgrp
,
2204 struct task_struct
*leader
, bool threadgroup
)
2206 LIST_HEAD(preloaded_csets
);
2207 struct task_struct
*task
;
2210 /* look up all src csets */
2211 down_read(&css_set_rwsem
);
2215 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
,
2219 } while_each_thread(leader
, task
);
2221 up_read(&css_set_rwsem
);
2223 /* prepare dst csets and commit */
2224 ret
= cgroup_migrate_prepare_dst(dst_cgrp
, &preloaded_csets
);
2226 ret
= cgroup_migrate(dst_cgrp
, leader
, threadgroup
);
2228 cgroup_migrate_finish(&preloaded_csets
);
2233 * Find the task_struct of the task to attach by vpid and pass it along to the
2234 * function to attach either it or all tasks in its threadgroup. Will lock
2235 * cgroup_mutex and threadgroup.
2237 static int attach_task_by_pid(struct cgroup
*cgrp
, u64 pid
, bool threadgroup
)
2239 struct task_struct
*tsk
;
2240 const struct cred
*cred
= current_cred(), *tcred
;
2243 if (!cgroup_lock_live_group(cgrp
))
2249 tsk
= find_task_by_vpid(pid
);
2253 goto out_unlock_cgroup
;
2256 * even if we're attaching all tasks in the thread group, we
2257 * only need to check permissions on one of them.
2259 tcred
= __task_cred(tsk
);
2260 if (!uid_eq(cred
->euid
, GLOBAL_ROOT_UID
) &&
2261 !uid_eq(cred
->euid
, tcred
->uid
) &&
2262 !uid_eq(cred
->euid
, tcred
->suid
)) {
2265 goto out_unlock_cgroup
;
2271 tsk
= tsk
->group_leader
;
2274 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2275 * trapped in a cpuset, or RT worker may be born in a cgroup
2276 * with no rt_runtime allocated. Just say no.
2278 if (tsk
== kthreadd_task
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2281 goto out_unlock_cgroup
;
2284 get_task_struct(tsk
);
2287 threadgroup_lock(tsk
);
2289 if (!thread_group_leader(tsk
)) {
2291 * a race with de_thread from another thread's exec()
2292 * may strip us of our leadership, if this happens,
2293 * there is no choice but to throw this task away and
2294 * try again; this is
2295 * "double-double-toil-and-trouble-check locking".
2297 threadgroup_unlock(tsk
);
2298 put_task_struct(tsk
);
2299 goto retry_find_task
;
2303 ret
= cgroup_attach_task(cgrp
, tsk
, threadgroup
);
2305 threadgroup_unlock(tsk
);
2307 put_task_struct(tsk
);
2309 mutex_unlock(&cgroup_mutex
);
2314 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2315 * @from: attach to all cgroups of a given task
2316 * @tsk: the task to be attached
2318 int cgroup_attach_task_all(struct task_struct
*from
, struct task_struct
*tsk
)
2320 struct cgroup_root
*root
;
2323 mutex_lock(&cgroup_mutex
);
2324 for_each_root(root
) {
2325 struct cgroup
*from_cgrp
;
2327 if (root
== &cgrp_dfl_root
)
2330 down_read(&css_set_rwsem
);
2331 from_cgrp
= task_cgroup_from_root(from
, root
);
2332 up_read(&css_set_rwsem
);
2334 retval
= cgroup_attach_task(from_cgrp
, tsk
, false);
2338 mutex_unlock(&cgroup_mutex
);
2342 EXPORT_SYMBOL_GPL(cgroup_attach_task_all
);
2344 static int cgroup_tasks_write(struct cgroup_subsys_state
*css
,
2345 struct cftype
*cft
, u64 pid
)
2347 return attach_task_by_pid(css
->cgroup
, pid
, false);
2350 static int cgroup_procs_write(struct cgroup_subsys_state
*css
,
2351 struct cftype
*cft
, u64 tgid
)
2353 return attach_task_by_pid(css
->cgroup
, tgid
, true);
2356 static int cgroup_release_agent_write(struct cgroup_subsys_state
*css
,
2357 struct cftype
*cft
, char *buffer
)
2359 struct cgroup_root
*root
= css
->cgroup
->root
;
2361 BUILD_BUG_ON(sizeof(root
->release_agent_path
) < PATH_MAX
);
2362 if (!cgroup_lock_live_group(css
->cgroup
))
2364 spin_lock(&release_agent_path_lock
);
2365 strlcpy(root
->release_agent_path
, buffer
,
2366 sizeof(root
->release_agent_path
));
2367 spin_unlock(&release_agent_path_lock
);
2368 mutex_unlock(&cgroup_mutex
);
2372 static int cgroup_release_agent_show(struct seq_file
*seq
, void *v
)
2374 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2376 if (!cgroup_lock_live_group(cgrp
))
2378 seq_puts(seq
, cgrp
->root
->release_agent_path
);
2379 seq_putc(seq
, '\n');
2380 mutex_unlock(&cgroup_mutex
);
2384 static int cgroup_sane_behavior_show(struct seq_file
*seq
, void *v
)
2386 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2388 seq_printf(seq
, "%d\n", cgroup_sane_behavior(cgrp
));
2392 static void cgroup_print_ss_mask(struct seq_file
*seq
, unsigned int ss_mask
)
2394 struct cgroup_subsys
*ss
;
2395 bool printed
= false;
2398 for_each_subsys(ss
, ssid
) {
2399 if (ss_mask
& (1 << ssid
)) {
2402 seq_printf(seq
, "%s", ss
->name
);
2407 seq_putc(seq
, '\n');
2410 /* show controllers which are currently attached to the default hierarchy */
2411 static int cgroup_root_controllers_show(struct seq_file
*seq
, void *v
)
2413 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2415 cgroup_print_ss_mask(seq
, cgrp
->root
->subsys_mask
);
2419 /* show controllers which are enabled from the parent */
2420 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2422 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2424 cgroup_print_ss_mask(seq
, cgrp
->parent
->child_subsys_mask
);
2428 /* show controllers which are enabled for a given cgroup's children */
2429 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2431 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2433 cgroup_print_ss_mask(seq
, cgrp
->child_subsys_mask
);
2438 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2439 * @cgrp: root of the subtree to update csses for
2441 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2442 * css associations need to be updated accordingly. This function looks up
2443 * all css_sets which are attached to the subtree, creates the matching
2444 * updated css_sets and migrates the tasks to the new ones.
2446 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2448 LIST_HEAD(preloaded_csets
);
2449 struct cgroup_subsys_state
*css
;
2450 struct css_set
*src_cset
;
2453 lockdep_assert_held(&cgroup_tree_mutex
);
2454 lockdep_assert_held(&cgroup_mutex
);
2456 /* look up all csses currently attached to @cgrp's subtree */
2457 down_read(&css_set_rwsem
);
2458 css_for_each_descendant_pre(css
, cgroup_css(cgrp
, NULL
)) {
2459 struct cgrp_cset_link
*link
;
2461 /* self is not affected by child_subsys_mask change */
2462 if (css
->cgroup
== cgrp
)
2465 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
)
2466 cgroup_migrate_add_src(link
->cset
, cgrp
,
2469 up_read(&css_set_rwsem
);
2471 /* NULL dst indicates self on default hierarchy */
2472 ret
= cgroup_migrate_prepare_dst(NULL
, &preloaded_csets
);
2476 list_for_each_entry(src_cset
, &preloaded_csets
, mg_preload_node
) {
2477 struct task_struct
*last_task
= NULL
, *task
;
2479 /* src_csets precede dst_csets, break on the first dst_cset */
2480 if (!src_cset
->mg_src_cgrp
)
2484 * All tasks in src_cset need to be migrated to the
2485 * matching dst_cset. Empty it process by process. We
2486 * walk tasks but migrate processes. The leader might even
2487 * belong to a different cset but such src_cset would also
2488 * be among the target src_csets because the default
2489 * hierarchy enforces per-process membership.
2492 down_read(&css_set_rwsem
);
2493 task
= list_first_entry_or_null(&src_cset
->tasks
,
2494 struct task_struct
, cg_list
);
2496 task
= task
->group_leader
;
2497 WARN_ON_ONCE(!task_css_set(task
)->mg_src_cgrp
);
2498 get_task_struct(task
);
2500 up_read(&css_set_rwsem
);
2505 /* guard against possible infinite loop */
2506 if (WARN(last_task
== task
,
2507 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2511 threadgroup_lock(task
);
2512 /* raced against de_thread() from another thread? */
2513 if (!thread_group_leader(task
)) {
2514 threadgroup_unlock(task
);
2515 put_task_struct(task
);
2519 ret
= cgroup_migrate(src_cset
->dfl_cgrp
, task
, true);
2521 threadgroup_unlock(task
);
2522 put_task_struct(task
);
2524 if (WARN(ret
, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret
))
2530 cgroup_migrate_finish(&preloaded_csets
);
2534 /* change the enabled child controllers for a cgroup in the default hierarchy */
2535 static int cgroup_subtree_control_write(struct cgroup_subsys_state
*dummy_css
,
2536 struct cftype
*cft
, char *buffer
)
2538 unsigned int enable_req
= 0, disable_req
= 0, enable
, disable
;
2539 struct cgroup
*cgrp
= dummy_css
->cgroup
, *child
;
2540 struct cgroup_subsys
*ss
;
2545 * Parse input - white space separated list of subsystem names
2546 * prefixed with either + or -.
2549 while ((tok
= strsep(&p
, " \t\n"))) {
2550 for_each_subsys(ss
, ssid
) {
2551 if (ss
->disabled
|| strcmp(tok
+ 1, ss
->name
))
2555 enable_req
|= 1 << ssid
;
2556 disable_req
&= ~(1 << ssid
);
2557 } else if (*tok
== '-') {
2558 disable_req
|= 1 << ssid
;
2559 enable_req
&= ~(1 << ssid
);
2565 if (ssid
== CGROUP_SUBSYS_COUNT
)
2570 * We're gonna grab cgroup_tree_mutex which nests outside kernfs
2571 * active_ref. cgroup_lock_live_group() already provides enough
2572 * protection. Ensure @cgrp stays accessible and break the
2573 * active_ref protection.
2576 kernfs_break_active_protection(cgrp
->control_kn
);
2578 enable
= enable_req
;
2579 disable
= disable_req
;
2581 mutex_lock(&cgroup_tree_mutex
);
2583 for_each_subsys(ss
, ssid
) {
2584 if (enable
& (1 << ssid
)) {
2585 if (cgrp
->child_subsys_mask
& (1 << ssid
)) {
2586 enable
&= ~(1 << ssid
);
2591 * Because css offlining is asynchronous, userland
2592 * might try to re-enable the same controller while
2593 * the previous instance is still around. In such
2594 * cases, wait till it's gone using offline_waitq.
2596 cgroup_for_each_live_child(child
, cgrp
) {
2599 if (!cgroup_css(child
, ss
))
2602 prepare_to_wait(&child
->offline_waitq
, &wait
,
2603 TASK_UNINTERRUPTIBLE
);
2604 mutex_unlock(&cgroup_tree_mutex
);
2606 finish_wait(&child
->offline_waitq
, &wait
);
2610 /* unavailable or not enabled on the parent? */
2611 if (!(cgrp_dfl_root
.subsys_mask
& (1 << ssid
)) ||
2613 !(cgrp
->parent
->child_subsys_mask
& (1 << ssid
)))) {
2615 goto out_unlock_tree
;
2617 } else if (disable
& (1 << ssid
)) {
2618 if (!(cgrp
->child_subsys_mask
& (1 << ssid
))) {
2619 disable
&= ~(1 << ssid
);
2623 /* a child has it enabled? */
2624 cgroup_for_each_live_child(child
, cgrp
) {
2625 if (child
->child_subsys_mask
& (1 << ssid
)) {
2627 goto out_unlock_tree
;
2633 if (!enable
&& !disable
) {
2635 goto out_unlock_tree
;
2638 if (!cgroup_lock_live_group(cgrp
)) {
2640 goto out_unlock_tree
;
2644 * Except for the root, child_subsys_mask must be zero for a cgroup
2645 * with tasks so that child cgroups don't compete against tasks.
2647 if (enable
&& cgrp
->parent
&& !list_empty(&cgrp
->cset_links
)) {
2653 * Create csses for enables and update child_subsys_mask. This
2654 * changes cgroup_e_css() results which in turn makes the
2655 * subsequent cgroup_update_dfl_csses() associate all tasks in the
2656 * subtree to the updated csses.
2658 for_each_subsys(ss
, ssid
) {
2659 if (!(enable
& (1 << ssid
)))
2662 cgroup_for_each_live_child(child
, cgrp
) {
2663 ret
= create_css(child
, ss
);
2669 cgrp
->child_subsys_mask
|= enable
;
2670 cgrp
->child_subsys_mask
&= ~disable
;
2672 ret
= cgroup_update_dfl_csses(cgrp
);
2676 /* all tasks are now migrated away from the old csses, kill them */
2677 for_each_subsys(ss
, ssid
) {
2678 if (!(disable
& (1 << ssid
)))
2681 cgroup_for_each_live_child(child
, cgrp
)
2682 kill_css(cgroup_css(child
, ss
));
2685 kernfs_activate(cgrp
->kn
);
2688 mutex_unlock(&cgroup_mutex
);
2690 mutex_unlock(&cgroup_tree_mutex
);
2691 kernfs_unbreak_active_protection(cgrp
->control_kn
);
2696 cgrp
->child_subsys_mask
&= ~enable
;
2697 cgrp
->child_subsys_mask
|= disable
;
2699 for_each_subsys(ss
, ssid
) {
2700 if (!(enable
& (1 << ssid
)))
2703 cgroup_for_each_live_child(child
, cgrp
) {
2704 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
2712 static int cgroup_populated_show(struct seq_file
*seq
, void *v
)
2714 seq_printf(seq
, "%d\n", (bool)seq_css(seq
)->cgroup
->populated_cnt
);
2718 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
2719 size_t nbytes
, loff_t off
)
2721 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
2722 struct cftype
*cft
= of
->kn
->priv
;
2723 struct cgroup_subsys_state
*css
;
2727 * kernfs guarantees that a file isn't deleted with operations in
2728 * flight, which means that the matching css is and stays alive and
2729 * doesn't need to be pinned. The RCU locking is not necessary
2730 * either. It's just for the convenience of using cgroup_css().
2733 css
= cgroup_css(cgrp
, cft
->ss
);
2736 if (cft
->write_string
) {
2737 ret
= cft
->write_string(css
, cft
, strstrip(buf
));
2738 } else if (cft
->write_u64
) {
2739 unsigned long long v
;
2740 ret
= kstrtoull(buf
, 0, &v
);
2742 ret
= cft
->write_u64(css
, cft
, v
);
2743 } else if (cft
->write_s64
) {
2745 ret
= kstrtoll(buf
, 0, &v
);
2747 ret
= cft
->write_s64(css
, cft
, v
);
2748 } else if (cft
->trigger
) {
2749 ret
= cft
->trigger(css
, (unsigned int)cft
->private);
2754 return ret
?: nbytes
;
2757 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
2759 return seq_cft(seq
)->seq_start(seq
, ppos
);
2762 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
2764 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
2767 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
2769 seq_cft(seq
)->seq_stop(seq
, v
);
2772 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
2774 struct cftype
*cft
= seq_cft(m
);
2775 struct cgroup_subsys_state
*css
= seq_css(m
);
2778 return cft
->seq_show(m
, arg
);
2781 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
2782 else if (cft
->read_s64
)
2783 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
2789 static struct kernfs_ops cgroup_kf_single_ops
= {
2790 .atomic_write_len
= PAGE_SIZE
,
2791 .write
= cgroup_file_write
,
2792 .seq_show
= cgroup_seqfile_show
,
2795 static struct kernfs_ops cgroup_kf_ops
= {
2796 .atomic_write_len
= PAGE_SIZE
,
2797 .write
= cgroup_file_write
,
2798 .seq_start
= cgroup_seqfile_start
,
2799 .seq_next
= cgroup_seqfile_next
,
2800 .seq_stop
= cgroup_seqfile_stop
,
2801 .seq_show
= cgroup_seqfile_show
,
2805 * cgroup_rename - Only allow simple rename of directories in place.
2807 static int cgroup_rename(struct kernfs_node
*kn
, struct kernfs_node
*new_parent
,
2808 const char *new_name_str
)
2810 struct cgroup
*cgrp
= kn
->priv
;
2813 if (kernfs_type(kn
) != KERNFS_DIR
)
2815 if (kn
->parent
!= new_parent
)
2819 * This isn't a proper migration and its usefulness is very
2820 * limited. Disallow if sane_behavior.
2822 if (cgroup_sane_behavior(cgrp
))
2826 * We're gonna grab cgroup_tree_mutex which nests outside kernfs
2827 * active_ref. kernfs_rename() doesn't require active_ref
2828 * protection. Break them before grabbing cgroup_tree_mutex.
2830 kernfs_break_active_protection(new_parent
);
2831 kernfs_break_active_protection(kn
);
2833 mutex_lock(&cgroup_tree_mutex
);
2834 mutex_lock(&cgroup_mutex
);
2836 ret
= kernfs_rename(kn
, new_parent
, new_name_str
);
2838 mutex_unlock(&cgroup_mutex
);
2839 mutex_unlock(&cgroup_tree_mutex
);
2841 kernfs_unbreak_active_protection(kn
);
2842 kernfs_unbreak_active_protection(new_parent
);
2846 /* set uid and gid of cgroup dirs and files to that of the creator */
2847 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
2849 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
2850 .ia_uid
= current_fsuid(),
2851 .ia_gid
= current_fsgid(), };
2853 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
2854 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
2857 return kernfs_setattr(kn
, &iattr
);
2860 static int cgroup_add_file(struct cgroup
*cgrp
, struct cftype
*cft
)
2862 char name
[CGROUP_FILE_NAME_MAX
];
2863 struct kernfs_node
*kn
;
2864 struct lock_class_key
*key
= NULL
;
2867 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2868 key
= &cft
->lockdep_key
;
2870 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
2871 cgroup_file_mode(cft
), 0, cft
->kf_ops
, cft
,
2876 ret
= cgroup_kn_set_ugid(kn
);
2882 if (cft
->seq_show
== cgroup_subtree_control_show
)
2883 cgrp
->control_kn
= kn
;
2884 else if (cft
->seq_show
== cgroup_populated_show
)
2885 cgrp
->populated_kn
= kn
;
2890 * cgroup_addrm_files - add or remove files to a cgroup directory
2891 * @cgrp: the target cgroup
2892 * @cfts: array of cftypes to be added
2893 * @is_add: whether to add or remove
2895 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2896 * For removals, this function never fails. If addition fails, this
2897 * function doesn't remove files already added. The caller is responsible
2900 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
2906 lockdep_assert_held(&cgroup_tree_mutex
);
2908 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
2909 /* does cft->flags tell us to skip this file on @cgrp? */
2910 if ((cft
->flags
& CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
2912 if ((cft
->flags
& CFTYPE_INSANE
) && cgroup_sane_behavior(cgrp
))
2914 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgrp
->parent
)
2916 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgrp
->parent
)
2920 ret
= cgroup_add_file(cgrp
, cft
);
2922 pr_warn("%s: failed to add %s, err=%d\n",
2923 __func__
, cft
->name
, ret
);
2927 cgroup_rm_file(cgrp
, cft
);
2933 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
2936 struct cgroup_subsys
*ss
= cfts
[0].ss
;
2937 struct cgroup
*root
= &ss
->root
->cgrp
;
2938 struct cgroup_subsys_state
*css
;
2941 lockdep_assert_held(&cgroup_tree_mutex
);
2943 /* add/rm files for all cgroups created before */
2944 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
2945 struct cgroup
*cgrp
= css
->cgroup
;
2947 if (cgroup_is_dead(cgrp
))
2950 ret
= cgroup_addrm_files(cgrp
, cfts
, is_add
);
2956 kernfs_activate(root
->kn
);
2960 static void cgroup_exit_cftypes(struct cftype
*cfts
)
2964 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
2965 /* free copy for custom atomic_write_len, see init_cftypes() */
2966 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
2973 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
2977 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
2978 struct kernfs_ops
*kf_ops
;
2980 WARN_ON(cft
->ss
|| cft
->kf_ops
);
2983 kf_ops
= &cgroup_kf_ops
;
2985 kf_ops
= &cgroup_kf_single_ops
;
2988 * Ugh... if @cft wants a custom max_write_len, we need to
2989 * make a copy of kf_ops to set its atomic_write_len.
2991 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
2992 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
2994 cgroup_exit_cftypes(cfts
);
2997 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3000 cft
->kf_ops
= kf_ops
;
3007 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3009 lockdep_assert_held(&cgroup_tree_mutex
);
3011 if (!cfts
|| !cfts
[0].ss
)
3014 list_del(&cfts
->node
);
3015 cgroup_apply_cftypes(cfts
, false);
3016 cgroup_exit_cftypes(cfts
);
3021 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3022 * @cfts: zero-length name terminated array of cftypes
3024 * Unregister @cfts. Files described by @cfts are removed from all
3025 * existing cgroups and all future cgroups won't have them either. This
3026 * function can be called anytime whether @cfts' subsys is attached or not.
3028 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3031 int cgroup_rm_cftypes(struct cftype
*cfts
)
3035 mutex_lock(&cgroup_tree_mutex
);
3036 ret
= cgroup_rm_cftypes_locked(cfts
);
3037 mutex_unlock(&cgroup_tree_mutex
);
3042 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3043 * @ss: target cgroup subsystem
3044 * @cfts: zero-length name terminated array of cftypes
3046 * Register @cfts to @ss. Files described by @cfts are created for all
3047 * existing cgroups to which @ss is attached and all future cgroups will
3048 * have them too. This function can be called anytime whether @ss is
3051 * Returns 0 on successful registration, -errno on failure. Note that this
3052 * function currently returns 0 as long as @cfts registration is successful
3053 * even if some file creation attempts on existing cgroups fail.
3055 int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3059 if (!cfts
|| cfts
[0].name
[0] == '\0')
3062 ret
= cgroup_init_cftypes(ss
, cfts
);
3066 mutex_lock(&cgroup_tree_mutex
);
3068 list_add_tail(&cfts
->node
, &ss
->cfts
);
3069 ret
= cgroup_apply_cftypes(cfts
, true);
3071 cgroup_rm_cftypes_locked(cfts
);
3073 mutex_unlock(&cgroup_tree_mutex
);
3078 * cgroup_task_count - count the number of tasks in a cgroup.
3079 * @cgrp: the cgroup in question
3081 * Return the number of tasks in the cgroup.
3083 static int cgroup_task_count(const struct cgroup
*cgrp
)
3086 struct cgrp_cset_link
*link
;
3088 down_read(&css_set_rwsem
);
3089 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
3090 count
+= atomic_read(&link
->cset
->refcount
);
3091 up_read(&css_set_rwsem
);
3096 * css_next_child - find the next child of a given css
3097 * @pos_css: the current position (%NULL to initiate traversal)
3098 * @parent_css: css whose children to walk
3100 * This function returns the next child of @parent_css and should be called
3101 * under either cgroup_mutex or RCU read lock. The only requirement is
3102 * that @parent_css and @pos_css are accessible. The next sibling is
3103 * guaranteed to be returned regardless of their states.
3105 struct cgroup_subsys_state
*
3106 css_next_child(struct cgroup_subsys_state
*pos_css
,
3107 struct cgroup_subsys_state
*parent_css
)
3109 struct cgroup
*pos
= pos_css
? pos_css
->cgroup
: NULL
;
3110 struct cgroup
*cgrp
= parent_css
->cgroup
;
3111 struct cgroup
*next
;
3113 cgroup_assert_mutexes_or_rcu_locked();
3116 * @pos could already have been removed. Once a cgroup is removed,
3117 * its ->sibling.next is no longer updated when its next sibling
3118 * changes. As CGRP_DEAD assertion is serialized and happens
3119 * before the cgroup is taken off the ->sibling list, if we see it
3120 * unasserted, it's guaranteed that the next sibling hasn't
3121 * finished its grace period even if it's already removed, and thus
3122 * safe to dereference from this RCU critical section. If
3123 * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
3124 * to be visible as %true here.
3126 * If @pos is dead, its next pointer can't be dereferenced;
3127 * however, as each cgroup is given a monotonically increasing
3128 * unique serial number and always appended to the sibling list,
3129 * the next one can be found by walking the parent's children until
3130 * we see a cgroup with higher serial number than @pos's. While
3131 * this path can be slower, it's taken only when either the current
3132 * cgroup is removed or iteration and removal race.
3135 next
= list_entry_rcu(cgrp
->children
.next
, struct cgroup
, sibling
);
3136 } else if (likely(!cgroup_is_dead(pos
))) {
3137 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup
, sibling
);
3139 list_for_each_entry_rcu(next
, &cgrp
->children
, sibling
)
3140 if (next
->serial_nr
> pos
->serial_nr
)
3145 * @next, if not pointing to the head, can be dereferenced and is
3146 * the next sibling; however, it might have @ss disabled. If so,
3147 * fast-forward to the next enabled one.
3149 while (&next
->sibling
!= &cgrp
->children
) {
3150 struct cgroup_subsys_state
*next_css
= cgroup_css(next
, parent_css
->ss
);
3154 next
= list_entry_rcu(next
->sibling
.next
, struct cgroup
, sibling
);
3160 * css_next_descendant_pre - find the next descendant for pre-order walk
3161 * @pos: the current position (%NULL to initiate traversal)
3162 * @root: css whose descendants to walk
3164 * To be used by css_for_each_descendant_pre(). Find the next descendant
3165 * to visit for pre-order traversal of @root's descendants. @root is
3166 * included in the iteration and the first node to be visited.
3168 * While this function requires cgroup_mutex or RCU read locking, it
3169 * doesn't require the whole traversal to be contained in a single critical
3170 * section. This function will return the correct next descendant as long
3171 * as both @pos and @root are accessible and @pos is a descendant of @root.
3173 struct cgroup_subsys_state
*
3174 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3175 struct cgroup_subsys_state
*root
)
3177 struct cgroup_subsys_state
*next
;
3179 cgroup_assert_mutexes_or_rcu_locked();
3181 /* if first iteration, visit @root */
3185 /* visit the first child if exists */
3186 next
= css_next_child(NULL
, pos
);
3190 /* no child, visit my or the closest ancestor's next sibling */
3191 while (pos
!= root
) {
3192 next
= css_next_child(pos
, css_parent(pos
));
3195 pos
= css_parent(pos
);
3202 * css_rightmost_descendant - return the rightmost descendant of a css
3203 * @pos: css of interest
3205 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3206 * is returned. This can be used during pre-order traversal to skip
3209 * While this function requires cgroup_mutex or RCU read locking, it
3210 * doesn't require the whole traversal to be contained in a single critical
3211 * section. This function will return the correct rightmost descendant as
3212 * long as @pos is accessible.
3214 struct cgroup_subsys_state
*
3215 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3217 struct cgroup_subsys_state
*last
, *tmp
;
3219 cgroup_assert_mutexes_or_rcu_locked();
3223 /* ->prev isn't RCU safe, walk ->next till the end */
3225 css_for_each_child(tmp
, last
)
3232 static struct cgroup_subsys_state
*
3233 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3235 struct cgroup_subsys_state
*last
;
3239 pos
= css_next_child(NULL
, pos
);
3246 * css_next_descendant_post - find the next descendant for post-order walk
3247 * @pos: the current position (%NULL to initiate traversal)
3248 * @root: css whose descendants to walk
3250 * To be used by css_for_each_descendant_post(). Find the next descendant
3251 * to visit for post-order traversal of @root's descendants. @root is
3252 * included in the iteration and the last node to be visited.
3254 * While this function requires cgroup_mutex or RCU read locking, it
3255 * doesn't require the whole traversal to be contained in a single critical
3256 * section. This function will return the correct next descendant as long
3257 * as both @pos and @cgroup are accessible and @pos is a descendant of
3260 struct cgroup_subsys_state
*
3261 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
3262 struct cgroup_subsys_state
*root
)
3264 struct cgroup_subsys_state
*next
;
3266 cgroup_assert_mutexes_or_rcu_locked();
3268 /* if first iteration, visit leftmost descendant which may be @root */
3270 return css_leftmost_descendant(root
);
3272 /* if we visited @root, we're done */
3276 /* if there's an unvisited sibling, visit its leftmost descendant */
3277 next
= css_next_child(pos
, css_parent(pos
));
3279 return css_leftmost_descendant(next
);
3281 /* no sibling left, visit parent */
3282 return css_parent(pos
);
3286 * css_advance_task_iter - advance a task itererator to the next css_set
3287 * @it: the iterator to advance
3289 * Advance @it to the next css_set to walk.
3291 static void css_advance_task_iter(struct css_task_iter
*it
)
3293 struct list_head
*l
= it
->cset_pos
;
3294 struct cgrp_cset_link
*link
;
3295 struct css_set
*cset
;
3297 /* Advance to the next non-empty css_set */
3300 if (l
== it
->cset_head
) {
3301 it
->cset_pos
= NULL
;
3306 cset
= container_of(l
, struct css_set
,
3307 e_cset_node
[it
->ss
->id
]);
3309 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
3312 } while (list_empty(&cset
->tasks
) && list_empty(&cset
->mg_tasks
));
3316 if (!list_empty(&cset
->tasks
))
3317 it
->task_pos
= cset
->tasks
.next
;
3319 it
->task_pos
= cset
->mg_tasks
.next
;
3321 it
->tasks_head
= &cset
->tasks
;
3322 it
->mg_tasks_head
= &cset
->mg_tasks
;
3326 * css_task_iter_start - initiate task iteration
3327 * @css: the css to walk tasks of
3328 * @it: the task iterator to use
3330 * Initiate iteration through the tasks of @css. The caller can call
3331 * css_task_iter_next() to walk through the tasks until the function
3332 * returns NULL. On completion of iteration, css_task_iter_end() must be
3335 * Note that this function acquires a lock which is released when the
3336 * iteration finishes. The caller can't sleep while iteration is in
3339 void css_task_iter_start(struct cgroup_subsys_state
*css
,
3340 struct css_task_iter
*it
)
3341 __acquires(css_set_rwsem
)
3343 /* no one should try to iterate before mounting cgroups */
3344 WARN_ON_ONCE(!use_task_css_set_links
);
3346 down_read(&css_set_rwsem
);
3351 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
3353 it
->cset_pos
= &css
->cgroup
->cset_links
;
3355 it
->cset_head
= it
->cset_pos
;
3357 css_advance_task_iter(it
);
3361 * css_task_iter_next - return the next task for the iterator
3362 * @it: the task iterator being iterated
3364 * The "next" function for task iteration. @it should have been
3365 * initialized via css_task_iter_start(). Returns NULL when the iteration
3368 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
3370 struct task_struct
*res
;
3371 struct list_head
*l
= it
->task_pos
;
3373 /* If the iterator cg is NULL, we have no tasks */
3376 res
= list_entry(l
, struct task_struct
, cg_list
);
3379 * Advance iterator to find next entry. cset->tasks is consumed
3380 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3385 if (l
== it
->tasks_head
)
3386 l
= it
->mg_tasks_head
->next
;
3388 if (l
== it
->mg_tasks_head
)
3389 css_advance_task_iter(it
);
3397 * css_task_iter_end - finish task iteration
3398 * @it: the task iterator to finish
3400 * Finish task iteration started by css_task_iter_start().
3402 void css_task_iter_end(struct css_task_iter
*it
)
3403 __releases(css_set_rwsem
)
3405 up_read(&css_set_rwsem
);
3409 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3410 * @to: cgroup to which the tasks will be moved
3411 * @from: cgroup in which the tasks currently reside
3413 * Locking rules between cgroup_post_fork() and the migration path
3414 * guarantee that, if a task is forking while being migrated, the new child
3415 * is guaranteed to be either visible in the source cgroup after the
3416 * parent's migration is complete or put into the target cgroup. No task
3417 * can slip out of migration through forking.
3419 int cgroup_transfer_tasks(struct cgroup
*to
, struct cgroup
*from
)
3421 LIST_HEAD(preloaded_csets
);
3422 struct cgrp_cset_link
*link
;
3423 struct css_task_iter it
;
3424 struct task_struct
*task
;
3427 mutex_lock(&cgroup_mutex
);
3429 /* all tasks in @from are being moved, all csets are source */
3430 down_read(&css_set_rwsem
);
3431 list_for_each_entry(link
, &from
->cset_links
, cset_link
)
3432 cgroup_migrate_add_src(link
->cset
, to
, &preloaded_csets
);
3433 up_read(&css_set_rwsem
);
3435 ret
= cgroup_migrate_prepare_dst(to
, &preloaded_csets
);
3440 * Migrate tasks one-by-one until @form is empty. This fails iff
3441 * ->can_attach() fails.
3444 css_task_iter_start(&from
->dummy_css
, &it
);
3445 task
= css_task_iter_next(&it
);
3447 get_task_struct(task
);
3448 css_task_iter_end(&it
);
3451 ret
= cgroup_migrate(to
, task
, false);
3452 put_task_struct(task
);
3454 } while (task
&& !ret
);
3456 cgroup_migrate_finish(&preloaded_csets
);
3457 mutex_unlock(&cgroup_mutex
);
3462 * Stuff for reading the 'tasks'/'procs' files.
3464 * Reading this file can return large amounts of data if a cgroup has
3465 * *lots* of attached tasks. So it may need several calls to read(),
3466 * but we cannot guarantee that the information we produce is correct
3467 * unless we produce it entirely atomically.
3471 /* which pidlist file are we talking about? */
3472 enum cgroup_filetype
{
3478 * A pidlist is a list of pids that virtually represents the contents of one
3479 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3480 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3483 struct cgroup_pidlist
{
3485 * used to find which pidlist is wanted. doesn't change as long as
3486 * this particular list stays in the list.
3488 struct { enum cgroup_filetype type
; struct pid_namespace
*ns
; } key
;
3491 /* how many elements the above list has */
3493 /* each of these stored in a list by its cgroup */
3494 struct list_head links
;
3495 /* pointer to the cgroup we belong to, for list removal purposes */
3496 struct cgroup
*owner
;
3497 /* for delayed destruction */
3498 struct delayed_work destroy_dwork
;
3502 * The following two functions "fix" the issue where there are more pids
3503 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3504 * TODO: replace with a kernel-wide solution to this problem
3506 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3507 static void *pidlist_allocate(int count
)
3509 if (PIDLIST_TOO_LARGE(count
))
3510 return vmalloc(count
* sizeof(pid_t
));
3512 return kmalloc(count
* sizeof(pid_t
), GFP_KERNEL
);
3515 static void pidlist_free(void *p
)
3517 if (is_vmalloc_addr(p
))
3524 * Used to destroy all pidlists lingering waiting for destroy timer. None
3525 * should be left afterwards.
3527 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
)
3529 struct cgroup_pidlist
*l
, *tmp_l
;
3531 mutex_lock(&cgrp
->pidlist_mutex
);
3532 list_for_each_entry_safe(l
, tmp_l
, &cgrp
->pidlists
, links
)
3533 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
, 0);
3534 mutex_unlock(&cgrp
->pidlist_mutex
);
3536 flush_workqueue(cgroup_pidlist_destroy_wq
);
3537 BUG_ON(!list_empty(&cgrp
->pidlists
));
3540 static void cgroup_pidlist_destroy_work_fn(struct work_struct
*work
)
3542 struct delayed_work
*dwork
= to_delayed_work(work
);
3543 struct cgroup_pidlist
*l
= container_of(dwork
, struct cgroup_pidlist
,
3545 struct cgroup_pidlist
*tofree
= NULL
;
3547 mutex_lock(&l
->owner
->pidlist_mutex
);
3550 * Destroy iff we didn't get queued again. The state won't change
3551 * as destroy_dwork can only be queued while locked.
3553 if (!delayed_work_pending(dwork
)) {
3554 list_del(&l
->links
);
3555 pidlist_free(l
->list
);
3556 put_pid_ns(l
->key
.ns
);
3560 mutex_unlock(&l
->owner
->pidlist_mutex
);
3565 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3566 * Returns the number of unique elements.
3568 static int pidlist_uniq(pid_t
*list
, int length
)
3573 * we presume the 0th element is unique, so i starts at 1. trivial
3574 * edge cases first; no work needs to be done for either
3576 if (length
== 0 || length
== 1)
3578 /* src and dest walk down the list; dest counts unique elements */
3579 for (src
= 1; src
< length
; src
++) {
3580 /* find next unique element */
3581 while (list
[src
] == list
[src
-1]) {
3586 /* dest always points to where the next unique element goes */
3587 list
[dest
] = list
[src
];
3595 * The two pid files - task and cgroup.procs - guaranteed that the result
3596 * is sorted, which forced this whole pidlist fiasco. As pid order is
3597 * different per namespace, each namespace needs differently sorted list,
3598 * making it impossible to use, for example, single rbtree of member tasks
3599 * sorted by task pointer. As pidlists can be fairly large, allocating one
3600 * per open file is dangerous, so cgroup had to implement shared pool of
3601 * pidlists keyed by cgroup and namespace.
3603 * All this extra complexity was caused by the original implementation
3604 * committing to an entirely unnecessary property. In the long term, we
3605 * want to do away with it. Explicitly scramble sort order if
3606 * sane_behavior so that no such expectation exists in the new interface.
3608 * Scrambling is done by swapping every two consecutive bits, which is
3609 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3611 static pid_t
pid_fry(pid_t pid
)
3613 unsigned a
= pid
& 0x55555555;
3614 unsigned b
= pid
& 0xAAAAAAAA;
3616 return (a
<< 1) | (b
>> 1);
3619 static pid_t
cgroup_pid_fry(struct cgroup
*cgrp
, pid_t pid
)
3621 if (cgroup_sane_behavior(cgrp
))
3622 return pid_fry(pid
);
3627 static int cmppid(const void *a
, const void *b
)
3629 return *(pid_t
*)a
- *(pid_t
*)b
;
3632 static int fried_cmppid(const void *a
, const void *b
)
3634 return pid_fry(*(pid_t
*)a
) - pid_fry(*(pid_t
*)b
);
3637 static struct cgroup_pidlist
*cgroup_pidlist_find(struct cgroup
*cgrp
,
3638 enum cgroup_filetype type
)
3640 struct cgroup_pidlist
*l
;
3641 /* don't need task_nsproxy() if we're looking at ourself */
3642 struct pid_namespace
*ns
= task_active_pid_ns(current
);
3644 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3646 list_for_each_entry(l
, &cgrp
->pidlists
, links
)
3647 if (l
->key
.type
== type
&& l
->key
.ns
== ns
)
3653 * find the appropriate pidlist for our purpose (given procs vs tasks)
3654 * returns with the lock on that pidlist already held, and takes care
3655 * of the use count, or returns NULL with no locks held if we're out of
3658 static struct cgroup_pidlist
*cgroup_pidlist_find_create(struct cgroup
*cgrp
,
3659 enum cgroup_filetype type
)
3661 struct cgroup_pidlist
*l
;
3663 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3665 l
= cgroup_pidlist_find(cgrp
, type
);
3669 /* entry not found; create a new one */
3670 l
= kzalloc(sizeof(struct cgroup_pidlist
), GFP_KERNEL
);
3674 INIT_DELAYED_WORK(&l
->destroy_dwork
, cgroup_pidlist_destroy_work_fn
);
3676 /* don't need task_nsproxy() if we're looking at ourself */
3677 l
->key
.ns
= get_pid_ns(task_active_pid_ns(current
));
3679 list_add(&l
->links
, &cgrp
->pidlists
);
3684 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3686 static int pidlist_array_load(struct cgroup
*cgrp
, enum cgroup_filetype type
,
3687 struct cgroup_pidlist
**lp
)
3691 int pid
, n
= 0; /* used for populating the array */
3692 struct css_task_iter it
;
3693 struct task_struct
*tsk
;
3694 struct cgroup_pidlist
*l
;
3696 lockdep_assert_held(&cgrp
->pidlist_mutex
);
3699 * If cgroup gets more users after we read count, we won't have
3700 * enough space - tough. This race is indistinguishable to the
3701 * caller from the case that the additional cgroup users didn't
3702 * show up until sometime later on.
3704 length
= cgroup_task_count(cgrp
);
3705 array
= pidlist_allocate(length
);
3708 /* now, populate the array */
3709 css_task_iter_start(&cgrp
->dummy_css
, &it
);
3710 while ((tsk
= css_task_iter_next(&it
))) {
3711 if (unlikely(n
== length
))
3713 /* get tgid or pid for procs or tasks file respectively */
3714 if (type
== CGROUP_FILE_PROCS
)
3715 pid
= task_tgid_vnr(tsk
);
3717 pid
= task_pid_vnr(tsk
);
3718 if (pid
> 0) /* make sure to only use valid results */
3721 css_task_iter_end(&it
);
3723 /* now sort & (if procs) strip out duplicates */
3724 if (cgroup_sane_behavior(cgrp
))
3725 sort(array
, length
, sizeof(pid_t
), fried_cmppid
, NULL
);
3727 sort(array
, length
, sizeof(pid_t
), cmppid
, NULL
);
3728 if (type
== CGROUP_FILE_PROCS
)
3729 length
= pidlist_uniq(array
, length
);
3731 l
= cgroup_pidlist_find_create(cgrp
, type
);
3733 mutex_unlock(&cgrp
->pidlist_mutex
);
3734 pidlist_free(array
);
3738 /* store array, freeing old if necessary */
3739 pidlist_free(l
->list
);
3747 * cgroupstats_build - build and fill cgroupstats
3748 * @stats: cgroupstats to fill information into
3749 * @dentry: A dentry entry belonging to the cgroup for which stats have
3752 * Build and fill cgroupstats so that taskstats can export it to user
3755 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
3757 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
3758 struct cgroup
*cgrp
;
3759 struct css_task_iter it
;
3760 struct task_struct
*tsk
;
3762 /* it should be kernfs_node belonging to cgroupfs and is a directory */
3763 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
3764 kernfs_type(kn
) != KERNFS_DIR
)
3767 mutex_lock(&cgroup_mutex
);
3770 * We aren't being called from kernfs and there's no guarantee on
3771 * @kn->priv's validity. For this and css_tryget_from_dir(),
3772 * @kn->priv is RCU safe. Let's do the RCU dancing.
3775 cgrp
= rcu_dereference(kn
->priv
);
3776 if (!cgrp
|| cgroup_is_dead(cgrp
)) {
3778 mutex_unlock(&cgroup_mutex
);
3783 css_task_iter_start(&cgrp
->dummy_css
, &it
);
3784 while ((tsk
= css_task_iter_next(&it
))) {
3785 switch (tsk
->state
) {
3787 stats
->nr_running
++;
3789 case TASK_INTERRUPTIBLE
:
3790 stats
->nr_sleeping
++;
3792 case TASK_UNINTERRUPTIBLE
:
3793 stats
->nr_uninterruptible
++;
3796 stats
->nr_stopped
++;
3799 if (delayacct_is_task_waiting_on_io(tsk
))
3800 stats
->nr_io_wait
++;
3804 css_task_iter_end(&it
);
3806 mutex_unlock(&cgroup_mutex
);
3812 * seq_file methods for the tasks/procs files. The seq_file position is the
3813 * next pid to display; the seq_file iterator is a pointer to the pid
3814 * in the cgroup->l->list array.
3817 static void *cgroup_pidlist_start(struct seq_file
*s
, loff_t
*pos
)
3820 * Initially we receive a position value that corresponds to
3821 * one more than the last pid shown (or 0 on the first call or
3822 * after a seek to the start). Use a binary-search to find the
3823 * next pid to display, if any
3825 struct kernfs_open_file
*of
= s
->private;
3826 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
3827 struct cgroup_pidlist
*l
;
3828 enum cgroup_filetype type
= seq_cft(s
)->private;
3829 int index
= 0, pid
= *pos
;
3832 mutex_lock(&cgrp
->pidlist_mutex
);
3835 * !NULL @of->priv indicates that this isn't the first start()
3836 * after open. If the matching pidlist is around, we can use that.
3837 * Look for it. Note that @of->priv can't be used directly. It
3838 * could already have been destroyed.
3841 of
->priv
= cgroup_pidlist_find(cgrp
, type
);
3844 * Either this is the first start() after open or the matching
3845 * pidlist has been destroyed inbetween. Create a new one.
3848 ret
= pidlist_array_load(cgrp
, type
,
3849 (struct cgroup_pidlist
**)&of
->priv
);
3851 return ERR_PTR(ret
);
3856 int end
= l
->length
;
3858 while (index
< end
) {
3859 int mid
= (index
+ end
) / 2;
3860 if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) == pid
) {
3863 } else if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) <= pid
)
3869 /* If we're off the end of the array, we're done */
3870 if (index
>= l
->length
)
3872 /* Update the abstract position to be the actual pid that we found */
3873 iter
= l
->list
+ index
;
3874 *pos
= cgroup_pid_fry(cgrp
, *iter
);
3878 static void cgroup_pidlist_stop(struct seq_file
*s
, void *v
)
3880 struct kernfs_open_file
*of
= s
->private;
3881 struct cgroup_pidlist
*l
= of
->priv
;
3884 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
,
3885 CGROUP_PIDLIST_DESTROY_DELAY
);
3886 mutex_unlock(&seq_css(s
)->cgroup
->pidlist_mutex
);
3889 static void *cgroup_pidlist_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
3891 struct kernfs_open_file
*of
= s
->private;
3892 struct cgroup_pidlist
*l
= of
->priv
;
3894 pid_t
*end
= l
->list
+ l
->length
;
3896 * Advance to the next pid in the array. If this goes off the
3903 *pos
= cgroup_pid_fry(seq_css(s
)->cgroup
, *p
);
3908 static int cgroup_pidlist_show(struct seq_file
*s
, void *v
)
3910 return seq_printf(s
, "%d\n", *(int *)v
);
3913 static u64
cgroup_read_notify_on_release(struct cgroup_subsys_state
*css
,
3916 return notify_on_release(css
->cgroup
);
3919 static int cgroup_write_notify_on_release(struct cgroup_subsys_state
*css
,
3920 struct cftype
*cft
, u64 val
)
3922 clear_bit(CGRP_RELEASABLE
, &css
->cgroup
->flags
);
3924 set_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
3926 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
3930 static u64
cgroup_clone_children_read(struct cgroup_subsys_state
*css
,
3933 return test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
3936 static int cgroup_clone_children_write(struct cgroup_subsys_state
*css
,
3937 struct cftype
*cft
, u64 val
)
3940 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
3942 clear_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
3946 static struct cftype cgroup_base_files
[] = {
3948 .name
= "cgroup.procs",
3949 .seq_start
= cgroup_pidlist_start
,
3950 .seq_next
= cgroup_pidlist_next
,
3951 .seq_stop
= cgroup_pidlist_stop
,
3952 .seq_show
= cgroup_pidlist_show
,
3953 .private = CGROUP_FILE_PROCS
,
3954 .write_u64
= cgroup_procs_write
,
3955 .mode
= S_IRUGO
| S_IWUSR
,
3958 .name
= "cgroup.clone_children",
3959 .flags
= CFTYPE_INSANE
,
3960 .read_u64
= cgroup_clone_children_read
,
3961 .write_u64
= cgroup_clone_children_write
,
3964 .name
= "cgroup.sane_behavior",
3965 .flags
= CFTYPE_ONLY_ON_ROOT
,
3966 .seq_show
= cgroup_sane_behavior_show
,
3969 .name
= "cgroup.controllers",
3970 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_ONLY_ON_ROOT
,
3971 .seq_show
= cgroup_root_controllers_show
,
3974 .name
= "cgroup.controllers",
3975 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_NOT_ON_ROOT
,
3976 .seq_show
= cgroup_controllers_show
,
3979 .name
= "cgroup.subtree_control",
3980 .flags
= CFTYPE_ONLY_ON_DFL
,
3981 .seq_show
= cgroup_subtree_control_show
,
3982 .write_string
= cgroup_subtree_control_write
,
3985 .name
= "cgroup.populated",
3986 .flags
= CFTYPE_ONLY_ON_DFL
| CFTYPE_NOT_ON_ROOT
,
3987 .seq_show
= cgroup_populated_show
,
3991 * Historical crazy stuff. These don't have "cgroup." prefix and
3992 * don't exist if sane_behavior. If you're depending on these, be
3993 * prepared to be burned.
3997 .flags
= CFTYPE_INSANE
, /* use "procs" instead */
3998 .seq_start
= cgroup_pidlist_start
,
3999 .seq_next
= cgroup_pidlist_next
,
4000 .seq_stop
= cgroup_pidlist_stop
,
4001 .seq_show
= cgroup_pidlist_show
,
4002 .private = CGROUP_FILE_TASKS
,
4003 .write_u64
= cgroup_tasks_write
,
4004 .mode
= S_IRUGO
| S_IWUSR
,
4007 .name
= "notify_on_release",
4008 .flags
= CFTYPE_INSANE
,
4009 .read_u64
= cgroup_read_notify_on_release
,
4010 .write_u64
= cgroup_write_notify_on_release
,
4013 .name
= "release_agent",
4014 .flags
= CFTYPE_INSANE
| CFTYPE_ONLY_ON_ROOT
,
4015 .seq_show
= cgroup_release_agent_show
,
4016 .write_string
= cgroup_release_agent_write
,
4017 .max_write_len
= PATH_MAX
- 1,
4023 * cgroup_populate_dir - create subsys files in a cgroup directory
4024 * @cgrp: target cgroup
4025 * @subsys_mask: mask of the subsystem ids whose files should be added
4027 * On failure, no file is added.
4029 static int cgroup_populate_dir(struct cgroup
*cgrp
, unsigned int subsys_mask
)
4031 struct cgroup_subsys
*ss
;
4034 /* process cftsets of each subsystem */
4035 for_each_subsys(ss
, i
) {
4036 struct cftype
*cfts
;
4038 if (!(subsys_mask
& (1 << i
)))
4041 list_for_each_entry(cfts
, &ss
->cfts
, node
) {
4042 ret
= cgroup_addrm_files(cgrp
, cfts
, true);
4049 cgroup_clear_dir(cgrp
, subsys_mask
);
4054 * css destruction is four-stage process.
4056 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4057 * Implemented in kill_css().
4059 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4060 * and thus css_tryget() is guaranteed to fail, the css can be offlined
4061 * by invoking offline_css(). After offlining, the base ref is put.
4062 * Implemented in css_killed_work_fn().
4064 * 3. When the percpu_ref reaches zero, the only possible remaining
4065 * accessors are inside RCU read sections. css_release() schedules the
4068 * 4. After the grace period, the css can be freed. Implemented in
4069 * css_free_work_fn().
4071 * It is actually hairier because both step 2 and 4 require process context
4072 * and thus involve punting to css->destroy_work adding two additional
4073 * steps to the already complex sequence.
4075 static void css_free_work_fn(struct work_struct
*work
)
4077 struct cgroup_subsys_state
*css
=
4078 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4079 struct cgroup
*cgrp
= css
->cgroup
;
4082 css_put(css
->parent
);
4084 css
->ss
->css_free(css
);
4088 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4090 struct cgroup_subsys_state
*css
=
4091 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4093 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4094 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4097 static void css_release(struct percpu_ref
*ref
)
4099 struct cgroup_subsys_state
*css
=
4100 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4101 struct cgroup_subsys
*ss
= css
->ss
;
4103 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4104 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4106 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4109 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4110 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4119 css
->parent
= cgroup_css(cgrp
->parent
, ss
);
4120 css_get(css
->parent
);
4122 css
->flags
|= CSS_ROOT
;
4125 BUG_ON(cgroup_css(cgrp
, ss
));
4128 /* invoke ->css_online() on a new CSS and mark it online if successful */
4129 static int online_css(struct cgroup_subsys_state
*css
)
4131 struct cgroup_subsys
*ss
= css
->ss
;
4134 lockdep_assert_held(&cgroup_tree_mutex
);
4135 lockdep_assert_held(&cgroup_mutex
);
4138 ret
= ss
->css_online(css
);
4140 css
->flags
|= CSS_ONLINE
;
4141 css
->cgroup
->nr_css
++;
4142 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4147 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4148 static void offline_css(struct cgroup_subsys_state
*css
)
4150 struct cgroup_subsys
*ss
= css
->ss
;
4152 lockdep_assert_held(&cgroup_tree_mutex
);
4153 lockdep_assert_held(&cgroup_mutex
);
4155 if (!(css
->flags
& CSS_ONLINE
))
4158 if (ss
->css_offline
)
4159 ss
->css_offline(css
);
4161 css
->flags
&= ~CSS_ONLINE
;
4162 css
->cgroup
->nr_css
--;
4163 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4165 wake_up_all(&css
->cgroup
->offline_waitq
);
4169 * create_css - create a cgroup_subsys_state
4170 * @cgrp: the cgroup new css will be associated with
4171 * @ss: the subsys of new css
4173 * Create a new css associated with @cgrp - @ss pair. On success, the new
4174 * css is online and installed in @cgrp with all interface files created.
4175 * Returns 0 on success, -errno on failure.
4177 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
)
4179 struct cgroup
*parent
= cgrp
->parent
;
4180 struct cgroup_subsys_state
*css
;
4183 lockdep_assert_held(&cgroup_mutex
);
4185 css
= ss
->css_alloc(cgroup_css(parent
, ss
));
4187 return PTR_ERR(css
);
4189 init_and_link_css(css
, ss
, cgrp
);
4191 err
= percpu_ref_init(&css
->refcnt
, css_release
);
4195 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_NOWAIT
);
4197 goto err_free_percpu_ref
;
4200 err
= cgroup_populate_dir(cgrp
, 1 << ss
->id
);
4204 /* @css is ready to be brought online now, make it visible */
4205 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4207 err
= online_css(css
);
4211 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4213 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4214 current
->comm
, current
->pid
, ss
->name
);
4215 if (!strcmp(ss
->name
, "memory"))
4216 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4217 ss
->warned_broken_hierarchy
= true;
4223 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4225 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4226 err_free_percpu_ref
:
4227 percpu_ref_cancel_init(&css
->refcnt
);
4229 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4234 * cgroup_create - create a cgroup
4235 * @parent: cgroup that will be parent of the new cgroup
4236 * @name: name of the new cgroup
4237 * @mode: mode to set on new cgroup
4239 static long cgroup_create(struct cgroup
*parent
, const char *name
,
4242 struct cgroup
*cgrp
;
4243 struct cgroup_root
*root
= parent
->root
;
4245 struct cgroup_subsys
*ss
;
4246 struct kernfs_node
*kn
;
4248 /* allocate the cgroup and its ID, 0 is reserved for the root */
4249 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
4253 mutex_lock(&cgroup_tree_mutex
);
4256 * Only live parents can have children. Note that the liveliness
4257 * check isn't strictly necessary because cgroup_mkdir() and
4258 * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4259 * anyway so that locking is contained inside cgroup proper and we
4260 * don't get nasty surprises if we ever grow another caller.
4262 if (!cgroup_lock_live_group(parent
)) {
4264 goto err_unlock_tree
;
4268 * Temporarily set the pointer to NULL, so idr_find() won't return
4269 * a half-baked cgroup.
4271 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_NOWAIT
);
4277 init_cgroup_housekeeping(cgrp
);
4279 cgrp
->parent
= parent
;
4280 cgrp
->dummy_css
.parent
= &parent
->dummy_css
;
4281 cgrp
->root
= parent
->root
;
4283 if (notify_on_release(parent
))
4284 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4286 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4287 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4289 /* create the directory */
4290 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4298 * This extra ref will be put in cgroup_free_fn() and guarantees
4299 * that @cgrp->kn is always accessible.
4303 cgrp
->serial_nr
= cgroup_serial_nr_next
++;
4305 /* allocation complete, commit to creation */
4306 list_add_tail_rcu(&cgrp
->sibling
, &cgrp
->parent
->children
);
4307 atomic_inc(&root
->nr_cgrps
);
4311 * @cgrp is now fully operational. If something fails after this
4312 * point, it'll be released via the normal destruction path.
4314 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4316 err
= cgroup_kn_set_ugid(kn
);
4320 err
= cgroup_addrm_files(cgrp
, cgroup_base_files
, true);
4324 /* let's create and online css's */
4325 for_each_subsys(ss
, ssid
) {
4326 if (parent
->child_subsys_mask
& (1 << ssid
)) {
4327 err
= create_css(cgrp
, ss
);
4334 * On the default hierarchy, a child doesn't automatically inherit
4335 * child_subsys_mask from the parent. Each is configured manually.
4337 if (!cgroup_on_dfl(cgrp
))
4338 cgrp
->child_subsys_mask
= parent
->child_subsys_mask
;
4340 kernfs_activate(kn
);
4342 mutex_unlock(&cgroup_mutex
);
4343 mutex_unlock(&cgroup_tree_mutex
);
4348 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4350 mutex_unlock(&cgroup_mutex
);
4352 mutex_unlock(&cgroup_tree_mutex
);
4357 cgroup_destroy_locked(cgrp
);
4358 mutex_unlock(&cgroup_mutex
);
4359 mutex_unlock(&cgroup_tree_mutex
);
4363 static int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
,
4366 struct cgroup
*parent
= parent_kn
->priv
;
4370 * cgroup_create() grabs cgroup_tree_mutex which nests outside
4371 * kernfs active_ref and cgroup_create() already synchronizes
4372 * properly against removal through cgroup_lock_live_group().
4373 * Break it before calling cgroup_create().
4376 kernfs_break_active_protection(parent_kn
);
4378 ret
= cgroup_create(parent
, name
, mode
);
4380 kernfs_unbreak_active_protection(parent_kn
);
4386 * This is called when the refcnt of a css is confirmed to be killed.
4387 * css_tryget() is now guaranteed to fail.
4389 static void css_killed_work_fn(struct work_struct
*work
)
4391 struct cgroup_subsys_state
*css
=
4392 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4393 struct cgroup
*cgrp
= css
->cgroup
;
4395 mutex_lock(&cgroup_tree_mutex
);
4396 mutex_lock(&cgroup_mutex
);
4399 * css_tryget() is guaranteed to fail now. Tell subsystems to
4400 * initate destruction.
4405 * If @cgrp is marked dead, it's waiting for refs of all css's to
4406 * be disabled before proceeding to the second phase of cgroup
4407 * destruction. If we are the last one, kick it off.
4409 if (!cgrp
->nr_css
&& cgroup_is_dead(cgrp
))
4410 cgroup_destroy_css_killed(cgrp
);
4412 mutex_unlock(&cgroup_mutex
);
4413 mutex_unlock(&cgroup_tree_mutex
);
4416 * Put the css refs from kill_css(). Each css holds an extra
4417 * reference to the cgroup's dentry and cgroup removal proceeds
4418 * regardless of css refs. On the last put of each css, whenever
4419 * that may be, the extra dentry ref is put so that dentry
4420 * destruction happens only after all css's are released.
4425 /* css kill confirmation processing requires process context, bounce */
4426 static void css_killed_ref_fn(struct percpu_ref
*ref
)
4428 struct cgroup_subsys_state
*css
=
4429 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4431 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
4432 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4436 * kill_css - destroy a css
4437 * @css: css to destroy
4439 * This function initiates destruction of @css by removing cgroup interface
4440 * files and putting its base reference. ->css_offline() will be invoked
4441 * asynchronously once css_tryget() is guaranteed to fail and when the
4442 * reference count reaches zero, @css will be released.
4444 static void kill_css(struct cgroup_subsys_state
*css
)
4446 lockdep_assert_held(&cgroup_tree_mutex
);
4449 * This must happen before css is disassociated with its cgroup.
4450 * See seq_css() for details.
4452 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4455 * Killing would put the base ref, but we need to keep it alive
4456 * until after ->css_offline().
4461 * cgroup core guarantees that, by the time ->css_offline() is
4462 * invoked, no new css reference will be given out via
4463 * css_tryget(). We can't simply call percpu_ref_kill() and
4464 * proceed to offlining css's because percpu_ref_kill() doesn't
4465 * guarantee that the ref is seen as killed on all CPUs on return.
4467 * Use percpu_ref_kill_and_confirm() to get notifications as each
4468 * css is confirmed to be seen as killed on all CPUs.
4470 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
4474 * cgroup_destroy_locked - the first stage of cgroup destruction
4475 * @cgrp: cgroup to be destroyed
4477 * css's make use of percpu refcnts whose killing latency shouldn't be
4478 * exposed to userland and are RCU protected. Also, cgroup core needs to
4479 * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4480 * invoked. To satisfy all the requirements, destruction is implemented in
4481 * the following two steps.
4483 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4484 * userland visible parts and start killing the percpu refcnts of
4485 * css's. Set up so that the next stage will be kicked off once all
4486 * the percpu refcnts are confirmed to be killed.
4488 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4489 * rest of destruction. Once all cgroup references are gone, the
4490 * cgroup is RCU-freed.
4492 * This function implements s1. After this step, @cgrp is gone as far as
4493 * the userland is concerned and a new cgroup with the same name may be
4494 * created. As cgroup doesn't care about the names internally, this
4495 * doesn't cause any problem.
4497 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
4498 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
4500 struct cgroup
*child
;
4501 struct cgroup_subsys_state
*css
;
4505 lockdep_assert_held(&cgroup_tree_mutex
);
4506 lockdep_assert_held(&cgroup_mutex
);
4509 * css_set_rwsem synchronizes access to ->cset_links and prevents
4510 * @cgrp from being removed while put_css_set() is in progress.
4512 down_read(&css_set_rwsem
);
4513 empty
= list_empty(&cgrp
->cset_links
);
4514 up_read(&css_set_rwsem
);
4519 * Make sure there's no live children. We can't test ->children
4520 * emptiness as dead children linger on it while being destroyed;
4521 * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4525 list_for_each_entry_rcu(child
, &cgrp
->children
, sibling
) {
4526 empty
= cgroup_is_dead(child
);
4535 * Mark @cgrp dead. This prevents further task migration and child
4536 * creation by disabling cgroup_lock_live_group(). Note that
4537 * CGRP_DEAD assertion is depended upon by css_next_child() to
4538 * resume iteration after dropping RCU read lock. See
4539 * css_next_child() for details.
4541 set_bit(CGRP_DEAD
, &cgrp
->flags
);
4544 * Initiate massacre of all css's. cgroup_destroy_css_killed()
4545 * will be invoked to perform the rest of destruction once the
4546 * percpu refs of all css's are confirmed to be killed. This
4547 * involves removing the subsystem's files, drop cgroup_mutex.
4549 mutex_unlock(&cgroup_mutex
);
4550 for_each_css(css
, ssid
, cgrp
)
4552 mutex_lock(&cgroup_mutex
);
4554 /* CGRP_DEAD is set, remove from ->release_list for the last time */
4555 raw_spin_lock(&release_list_lock
);
4556 if (!list_empty(&cgrp
->release_list
))
4557 list_del_init(&cgrp
->release_list
);
4558 raw_spin_unlock(&release_list_lock
);
4561 * If @cgrp has css's attached, the second stage of cgroup
4562 * destruction is kicked off from css_killed_work_fn() after the
4563 * refs of all attached css's are killed. If @cgrp doesn't have
4564 * any css, we kick it off here.
4567 cgroup_destroy_css_killed(cgrp
);
4569 /* remove @cgrp directory along with the base files */
4570 mutex_unlock(&cgroup_mutex
);
4573 * There are two control paths which try to determine cgroup from
4574 * dentry without going through kernfs - cgroupstats_build() and
4575 * css_tryget_from_dir(). Those are supported by RCU protecting
4576 * clearing of cgrp->kn->priv backpointer, which should happen
4577 * after all files under it have been removed.
4579 kernfs_remove(cgrp
->kn
); /* @cgrp has an extra ref on its kn */
4580 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
, NULL
);
4582 mutex_lock(&cgroup_mutex
);
4588 * cgroup_destroy_css_killed - the second step of cgroup destruction
4589 * @work: cgroup->destroy_free_work
4591 * This function is invoked from a work item for a cgroup which is being
4592 * destroyed after all css's are offlined and performs the rest of
4593 * destruction. This is the second step of destruction described in the
4594 * comment above cgroup_destroy_locked().
4596 static void cgroup_destroy_css_killed(struct cgroup
*cgrp
)
4598 struct cgroup
*parent
= cgrp
->parent
;
4600 lockdep_assert_held(&cgroup_tree_mutex
);
4601 lockdep_assert_held(&cgroup_mutex
);
4603 /* delete this cgroup from parent->children */
4604 list_del_rcu(&cgrp
->sibling
);
4608 set_bit(CGRP_RELEASABLE
, &parent
->flags
);
4609 check_for_release(parent
);
4612 static int cgroup_rmdir(struct kernfs_node
*kn
)
4614 struct cgroup
*cgrp
= kn
->priv
;
4618 * This is self-destruction but @kn can't be removed while this
4619 * callback is in progress. Let's break active protection. Once
4620 * the protection is broken, @cgrp can be destroyed at any point.
4621 * Pin it so that it stays accessible.
4624 kernfs_break_active_protection(kn
);
4626 mutex_lock(&cgroup_tree_mutex
);
4627 mutex_lock(&cgroup_mutex
);
4630 * @cgrp might already have been destroyed while we're trying to
4633 if (!cgroup_is_dead(cgrp
))
4634 ret
= cgroup_destroy_locked(cgrp
);
4636 mutex_unlock(&cgroup_mutex
);
4637 mutex_unlock(&cgroup_tree_mutex
);
4639 kernfs_unbreak_active_protection(kn
);
4644 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
4645 .remount_fs
= cgroup_remount
,
4646 .show_options
= cgroup_show_options
,
4647 .mkdir
= cgroup_mkdir
,
4648 .rmdir
= cgroup_rmdir
,
4649 .rename
= cgroup_rename
,
4652 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
4654 struct cgroup_subsys_state
*css
;
4656 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
4658 mutex_lock(&cgroup_tree_mutex
);
4659 mutex_lock(&cgroup_mutex
);
4661 idr_init(&ss
->css_idr
);
4662 INIT_LIST_HEAD(&ss
->cfts
);
4664 /* Create the root cgroup state for this subsystem */
4665 ss
->root
= &cgrp_dfl_root
;
4666 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
4667 /* We don't handle early failures gracefully */
4668 BUG_ON(IS_ERR(css
));
4669 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
4671 /* idr_alloc() can't be called safely during early init */
4674 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
4675 BUG_ON(css
->id
< 0);
4678 /* Update the init_css_set to contain a subsys
4679 * pointer to this state - since the subsystem is
4680 * newly registered, all tasks and hence the
4681 * init_css_set is in the subsystem's root cgroup. */
4682 init_css_set
.subsys
[ss
->id
] = css
;
4684 need_forkexit_callback
|= ss
->fork
|| ss
->exit
;
4686 /* At system boot, before all subsystems have been
4687 * registered, no tasks have been forked, so we don't
4688 * need to invoke fork callbacks here. */
4689 BUG_ON(!list_empty(&init_task
.tasks
));
4691 BUG_ON(online_css(css
));
4693 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
4695 mutex_unlock(&cgroup_mutex
);
4696 mutex_unlock(&cgroup_tree_mutex
);
4700 * cgroup_init_early - cgroup initialization at system boot
4702 * Initialize cgroups at system boot, and initialize any
4703 * subsystems that request early init.
4705 int __init
cgroup_init_early(void)
4707 static struct cgroup_sb_opts __initdata opts
=
4708 { .flags
= CGRP_ROOT_SANE_BEHAVIOR
};
4709 struct cgroup_subsys
*ss
;
4712 init_cgroup_root(&cgrp_dfl_root
, &opts
);
4713 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
4715 for_each_subsys(ss
, i
) {
4716 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
4717 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4718 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
4720 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
4721 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
4724 ss
->name
= cgroup_subsys_name
[i
];
4727 cgroup_init_subsys(ss
, true);
4733 * cgroup_init - cgroup initialization
4735 * Register cgroup filesystem and /proc file, and initialize
4736 * any subsystems that didn't request early init.
4738 int __init
cgroup_init(void)
4740 struct cgroup_subsys
*ss
;
4744 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_base_files
));
4746 mutex_lock(&cgroup_tree_mutex
);
4747 mutex_lock(&cgroup_mutex
);
4749 /* Add init_css_set to the hash table */
4750 key
= css_set_hash(init_css_set
.subsys
);
4751 hash_add(css_set_table
, &init_css_set
.hlist
, key
);
4753 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
4755 mutex_unlock(&cgroup_mutex
);
4756 mutex_unlock(&cgroup_tree_mutex
);
4758 for_each_subsys(ss
, ssid
) {
4759 if (ss
->early_init
) {
4760 struct cgroup_subsys_state
*css
=
4761 init_css_set
.subsys
[ss
->id
];
4763 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
4765 BUG_ON(css
->id
< 0);
4767 cgroup_init_subsys(ss
, false);
4770 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
4771 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
4774 * cftype registration needs kmalloc and can't be done
4775 * during early_init. Register base cftypes separately.
4777 if (ss
->base_cftypes
)
4778 WARN_ON(cgroup_add_cftypes(ss
, ss
->base_cftypes
));
4781 cgroup_kobj
= kobject_create_and_add("cgroup", fs_kobj
);
4785 err
= register_filesystem(&cgroup_fs_type
);
4787 kobject_put(cgroup_kobj
);
4791 proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
);
4795 static int __init
cgroup_wq_init(void)
4798 * There isn't much point in executing destruction path in
4799 * parallel. Good chunk is serialized with cgroup_mutex anyway.
4800 * Use 1 for @max_active.
4802 * We would prefer to do this in cgroup_init() above, but that
4803 * is called before init_workqueues(): so leave this until after.
4805 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
4806 BUG_ON(!cgroup_destroy_wq
);
4809 * Used to destroy pidlists and separate to serve as flush domain.
4810 * Cap @max_active to 1 too.
4812 cgroup_pidlist_destroy_wq
= alloc_workqueue("cgroup_pidlist_destroy",
4814 BUG_ON(!cgroup_pidlist_destroy_wq
);
4818 core_initcall(cgroup_wq_init
);
4821 * proc_cgroup_show()
4822 * - Print task's cgroup paths into seq_file, one line for each hierarchy
4823 * - Used for /proc/<pid>/cgroup.
4826 /* TODO: Use a proper seq_file iterator */
4827 int proc_cgroup_show(struct seq_file
*m
, void *v
)
4830 struct task_struct
*tsk
;
4833 struct cgroup_root
*root
;
4836 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
4842 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
4848 mutex_lock(&cgroup_mutex
);
4849 down_read(&css_set_rwsem
);
4851 for_each_root(root
) {
4852 struct cgroup_subsys
*ss
;
4853 struct cgroup
*cgrp
;
4854 int ssid
, count
= 0;
4856 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_root_visible
)
4859 seq_printf(m
, "%d:", root
->hierarchy_id
);
4860 for_each_subsys(ss
, ssid
)
4861 if (root
->subsys_mask
& (1 << ssid
))
4862 seq_printf(m
, "%s%s", count
++ ? "," : "", ss
->name
);
4863 if (strlen(root
->name
))
4864 seq_printf(m
, "%sname=%s", count
? "," : "",
4867 cgrp
= task_cgroup_from_root(tsk
, root
);
4868 path
= cgroup_path(cgrp
, buf
, PATH_MAX
);
4870 retval
= -ENAMETOOLONG
;
4878 up_read(&css_set_rwsem
);
4879 mutex_unlock(&cgroup_mutex
);
4880 put_task_struct(tsk
);
4887 /* Display information about each subsystem and each hierarchy */
4888 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
4890 struct cgroup_subsys
*ss
;
4893 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4895 * ideally we don't want subsystems moving around while we do this.
4896 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4897 * subsys/hierarchy state.
4899 mutex_lock(&cgroup_mutex
);
4901 for_each_subsys(ss
, i
)
4902 seq_printf(m
, "%s\t%d\t%d\t%d\n",
4903 ss
->name
, ss
->root
->hierarchy_id
,
4904 atomic_read(&ss
->root
->nr_cgrps
), !ss
->disabled
);
4906 mutex_unlock(&cgroup_mutex
);
4910 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
4912 return single_open(file
, proc_cgroupstats_show
, NULL
);
4915 static const struct file_operations proc_cgroupstats_operations
= {
4916 .open
= cgroupstats_open
,
4918 .llseek
= seq_lseek
,
4919 .release
= single_release
,
4923 * cgroup_fork - initialize cgroup related fields during copy_process()
4924 * @child: pointer to task_struct of forking parent process.
4926 * A task is associated with the init_css_set until cgroup_post_fork()
4927 * attaches it to the parent's css_set. Empty cg_list indicates that
4928 * @child isn't holding reference to its css_set.
4930 void cgroup_fork(struct task_struct
*child
)
4932 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
4933 INIT_LIST_HEAD(&child
->cg_list
);
4937 * cgroup_post_fork - called on a new task after adding it to the task list
4938 * @child: the task in question
4940 * Adds the task to the list running through its css_set if necessary and
4941 * call the subsystem fork() callbacks. Has to be after the task is
4942 * visible on the task list in case we race with the first call to
4943 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
4946 void cgroup_post_fork(struct task_struct
*child
)
4948 struct cgroup_subsys
*ss
;
4952 * This may race against cgroup_enable_task_cg_links(). As that
4953 * function sets use_task_css_set_links before grabbing
4954 * tasklist_lock and we just went through tasklist_lock to add
4955 * @child, it's guaranteed that either we see the set
4956 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
4957 * @child during its iteration.
4959 * If we won the race, @child is associated with %current's
4960 * css_set. Grabbing css_set_rwsem guarantees both that the
4961 * association is stable, and, on completion of the parent's
4962 * migration, @child is visible in the source of migration or
4963 * already in the destination cgroup. This guarantee is necessary
4964 * when implementing operations which need to migrate all tasks of
4965 * a cgroup to another.
4967 * Note that if we lose to cgroup_enable_task_cg_links(), @child
4968 * will remain in init_css_set. This is safe because all tasks are
4969 * in the init_css_set before cg_links is enabled and there's no
4970 * operation which transfers all tasks out of init_css_set.
4972 if (use_task_css_set_links
) {
4973 struct css_set
*cset
;
4975 down_write(&css_set_rwsem
);
4976 cset
= task_css_set(current
);
4977 if (list_empty(&child
->cg_list
)) {
4978 rcu_assign_pointer(child
->cgroups
, cset
);
4979 list_add(&child
->cg_list
, &cset
->tasks
);
4982 up_write(&css_set_rwsem
);
4986 * Call ss->fork(). This must happen after @child is linked on
4987 * css_set; otherwise, @child might change state between ->fork()
4988 * and addition to css_set.
4990 if (need_forkexit_callback
) {
4991 for_each_subsys(ss
, i
)
4998 * cgroup_exit - detach cgroup from exiting task
4999 * @tsk: pointer to task_struct of exiting process
5001 * Description: Detach cgroup from @tsk and release it.
5003 * Note that cgroups marked notify_on_release force every task in
5004 * them to take the global cgroup_mutex mutex when exiting.
5005 * This could impact scaling on very large systems. Be reluctant to
5006 * use notify_on_release cgroups where very high task exit scaling
5007 * is required on large systems.
5009 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5010 * call cgroup_exit() while the task is still competent to handle
5011 * notify_on_release(), then leave the task attached to the root cgroup in
5012 * each hierarchy for the remainder of its exit. No need to bother with
5013 * init_css_set refcnting. init_css_set never goes away and we can't race
5014 * with migration path - PF_EXITING is visible to migration path.
5016 void cgroup_exit(struct task_struct
*tsk
)
5018 struct cgroup_subsys
*ss
;
5019 struct css_set
*cset
;
5020 bool put_cset
= false;
5024 * Unlink from @tsk from its css_set. As migration path can't race
5025 * with us, we can check cg_list without grabbing css_set_rwsem.
5027 if (!list_empty(&tsk
->cg_list
)) {
5028 down_write(&css_set_rwsem
);
5029 list_del_init(&tsk
->cg_list
);
5030 up_write(&css_set_rwsem
);
5034 /* Reassign the task to the init_css_set. */
5035 cset
= task_css_set(tsk
);
5036 RCU_INIT_POINTER(tsk
->cgroups
, &init_css_set
);
5038 if (need_forkexit_callback
) {
5039 /* see cgroup_post_fork() for details */
5040 for_each_subsys(ss
, i
) {
5042 struct cgroup_subsys_state
*old_css
= cset
->subsys
[i
];
5043 struct cgroup_subsys_state
*css
= task_css(tsk
, i
);
5045 ss
->exit(css
, old_css
, tsk
);
5051 put_css_set(cset
, true);
5054 static void check_for_release(struct cgroup
*cgrp
)
5056 if (cgroup_is_releasable(cgrp
) &&
5057 list_empty(&cgrp
->cset_links
) && list_empty(&cgrp
->children
)) {
5059 * Control Group is currently removeable. If it's not
5060 * already queued for a userspace notification, queue
5063 int need_schedule_work
= 0;
5065 raw_spin_lock(&release_list_lock
);
5066 if (!cgroup_is_dead(cgrp
) &&
5067 list_empty(&cgrp
->release_list
)) {
5068 list_add(&cgrp
->release_list
, &release_list
);
5069 need_schedule_work
= 1;
5071 raw_spin_unlock(&release_list_lock
);
5072 if (need_schedule_work
)
5073 schedule_work(&release_agent_work
);
5078 * Notify userspace when a cgroup is released, by running the
5079 * configured release agent with the name of the cgroup (path
5080 * relative to the root of cgroup file system) as the argument.
5082 * Most likely, this user command will try to rmdir this cgroup.
5084 * This races with the possibility that some other task will be
5085 * attached to this cgroup before it is removed, or that some other
5086 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5087 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5088 * unused, and this cgroup will be reprieved from its death sentence,
5089 * to continue to serve a useful existence. Next time it's released,
5090 * we will get notified again, if it still has 'notify_on_release' set.
5092 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5093 * means only wait until the task is successfully execve()'d. The
5094 * separate release agent task is forked by call_usermodehelper(),
5095 * then control in this thread returns here, without waiting for the
5096 * release agent task. We don't bother to wait because the caller of
5097 * this routine has no use for the exit status of the release agent
5098 * task, so no sense holding our caller up for that.
5100 static void cgroup_release_agent(struct work_struct
*work
)
5102 BUG_ON(work
!= &release_agent_work
);
5103 mutex_lock(&cgroup_mutex
);
5104 raw_spin_lock(&release_list_lock
);
5105 while (!list_empty(&release_list
)) {
5106 char *argv
[3], *envp
[3];
5108 char *pathbuf
= NULL
, *agentbuf
= NULL
, *path
;
5109 struct cgroup
*cgrp
= list_entry(release_list
.next
,
5112 list_del_init(&cgrp
->release_list
);
5113 raw_spin_unlock(&release_list_lock
);
5114 pathbuf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5117 path
= cgroup_path(cgrp
, pathbuf
, PATH_MAX
);
5120 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
5125 argv
[i
++] = agentbuf
;
5130 /* minimal command environment */
5131 envp
[i
++] = "HOME=/";
5132 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5135 /* Drop the lock while we invoke the usermode helper,
5136 * since the exec could involve hitting disk and hence
5137 * be a slow process */
5138 mutex_unlock(&cgroup_mutex
);
5139 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
5140 mutex_lock(&cgroup_mutex
);
5144 raw_spin_lock(&release_list_lock
);
5146 raw_spin_unlock(&release_list_lock
);
5147 mutex_unlock(&cgroup_mutex
);
5150 static int __init
cgroup_disable(char *str
)
5152 struct cgroup_subsys
*ss
;
5156 while ((token
= strsep(&str
, ",")) != NULL
) {
5160 for_each_subsys(ss
, i
) {
5161 if (!strcmp(token
, ss
->name
)) {
5163 printk(KERN_INFO
"Disabling %s control group"
5164 " subsystem\n", ss
->name
);
5171 __setup("cgroup_disable=", cgroup_disable
);
5174 * css_tryget_from_dir - get corresponding css from the dentry of a cgroup dir
5175 * @dentry: directory dentry of interest
5176 * @ss: subsystem of interest
5178 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5179 * to get the corresponding css and return it. If such css doesn't exist
5180 * or can't be pinned, an ERR_PTR value is returned.
5182 struct cgroup_subsys_state
*css_tryget_from_dir(struct dentry
*dentry
,
5183 struct cgroup_subsys
*ss
)
5185 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5186 struct cgroup_subsys_state
*css
= NULL
;
5187 struct cgroup
*cgrp
;
5189 /* is @dentry a cgroup dir? */
5190 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
5191 kernfs_type(kn
) != KERNFS_DIR
)
5192 return ERR_PTR(-EBADF
);
5197 * This path doesn't originate from kernfs and @kn could already
5198 * have been or be removed at any point. @kn->priv is RCU
5199 * protected for this access. See destroy_locked() for details.
5201 cgrp
= rcu_dereference(kn
->priv
);
5203 css
= cgroup_css(cgrp
, ss
);
5205 if (!css
|| !css_tryget(css
))
5206 css
= ERR_PTR(-ENOENT
);
5213 * css_from_id - lookup css by id
5214 * @id: the cgroup id
5215 * @ss: cgroup subsys to be looked into
5217 * Returns the css if there's valid one with @id, otherwise returns NULL.
5218 * Should be called under rcu_read_lock().
5220 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5222 WARN_ON_ONCE(!rcu_read_lock_held());
5223 return idr_find(&ss
->css_idr
, id
);
5226 #ifdef CONFIG_CGROUP_DEBUG
5227 static struct cgroup_subsys_state
*
5228 debug_css_alloc(struct cgroup_subsys_state
*parent_css
)
5230 struct cgroup_subsys_state
*css
= kzalloc(sizeof(*css
), GFP_KERNEL
);
5233 return ERR_PTR(-ENOMEM
);
5238 static void debug_css_free(struct cgroup_subsys_state
*css
)
5243 static u64
debug_taskcount_read(struct cgroup_subsys_state
*css
,
5246 return cgroup_task_count(css
->cgroup
);
5249 static u64
current_css_set_read(struct cgroup_subsys_state
*css
,
5252 return (u64
)(unsigned long)current
->cgroups
;
5255 static u64
current_css_set_refcount_read(struct cgroup_subsys_state
*css
,
5261 count
= atomic_read(&task_css_set(current
)->refcount
);
5266 static int current_css_set_cg_links_read(struct seq_file
*seq
, void *v
)
5268 struct cgrp_cset_link
*link
;
5269 struct css_set
*cset
;
5272 name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
5276 down_read(&css_set_rwsem
);
5278 cset
= rcu_dereference(current
->cgroups
);
5279 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
5280 struct cgroup
*c
= link
->cgrp
;
5282 cgroup_name(c
, name_buf
, NAME_MAX
+ 1);
5283 seq_printf(seq
, "Root %d group %s\n",
5284 c
->root
->hierarchy_id
, name_buf
);
5287 up_read(&css_set_rwsem
);
5292 #define MAX_TASKS_SHOWN_PER_CSS 25
5293 static int cgroup_css_links_read(struct seq_file
*seq
, void *v
)
5295 struct cgroup_subsys_state
*css
= seq_css(seq
);
5296 struct cgrp_cset_link
*link
;
5298 down_read(&css_set_rwsem
);
5299 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
) {
5300 struct css_set
*cset
= link
->cset
;
5301 struct task_struct
*task
;
5304 seq_printf(seq
, "css_set %p\n", cset
);
5306 list_for_each_entry(task
, &cset
->tasks
, cg_list
) {
5307 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5309 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5312 list_for_each_entry(task
, &cset
->mg_tasks
, cg_list
) {
5313 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5315 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5319 seq_puts(seq
, " ...\n");
5321 up_read(&css_set_rwsem
);
5325 static u64
releasable_read(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
5327 return test_bit(CGRP_RELEASABLE
, &css
->cgroup
->flags
);
5330 static struct cftype debug_files
[] = {
5332 .name
= "taskcount",
5333 .read_u64
= debug_taskcount_read
,
5337 .name
= "current_css_set",
5338 .read_u64
= current_css_set_read
,
5342 .name
= "current_css_set_refcount",
5343 .read_u64
= current_css_set_refcount_read
,
5347 .name
= "current_css_set_cg_links",
5348 .seq_show
= current_css_set_cg_links_read
,
5352 .name
= "cgroup_css_links",
5353 .seq_show
= cgroup_css_links_read
,
5357 .name
= "releasable",
5358 .read_u64
= releasable_read
,
5364 struct cgroup_subsys debug_cgrp_subsys
= {
5365 .css_alloc
= debug_css_alloc
,
5366 .css_free
= debug_css_free
,
5367 .base_cftypes
= debug_files
,
5369 #endif /* CONFIG_CGROUP_DEBUG */