2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex
);
87 DECLARE_RWSEM(css_set_rwsem
);
88 EXPORT_SYMBOL_GPL(cgroup_mutex
);
89 EXPORT_SYMBOL_GPL(css_set_rwsem
);
91 static DEFINE_MUTEX(cgroup_mutex
);
92 static DECLARE_RWSEM(css_set_rwsem
);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock
);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock
);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem
;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
111 !lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct
*cgroup_destroy_wq
;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct
*cgroup_pidlist_destroy_wq
;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys
*cgroup_subsys
[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name
[] = {
138 #include <linux/cgroup_subsys.h>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true
*cgroup_subsys_enabled_key
[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true
*cgroup_subsys_on_dfl_key
[] = {
159 #include <linux/cgroup_subsys.h>
164 * The default hierarchy, reserved for the subsystems that are otherwise
165 * unattached - it never has more than a single cgroup, and all tasks are
166 * part of that cgroup.
168 struct cgroup_root cgrp_dfl_root
;
169 EXPORT_SYMBOL_GPL(cgrp_dfl_root
);
172 * The default hierarchy always exists but is hidden until mounted for the
173 * first time. This is for backward compatibility.
175 static bool cgrp_dfl_root_visible
;
178 * Set by the boot param of the same name and makes subsystems with NULL
179 * ->dfl_files to use ->legacy_files on the default hierarchy.
181 static bool cgroup_legacy_files_on_dfl
;
183 /* some controllers are not supported in the default hierarchy */
184 static unsigned long cgrp_dfl_root_inhibit_ss_mask
;
186 /* The list of hierarchy roots */
188 static LIST_HEAD(cgroup_roots
);
189 static int cgroup_root_count
;
191 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
192 static DEFINE_IDR(cgroup_hierarchy_idr
);
195 * Assign a monotonically increasing serial number to csses. It guarantees
196 * cgroups with bigger numbers are newer than those with smaller numbers.
197 * Also, as csses are always appended to the parent's ->children list, it
198 * guarantees that sibling csses are always sorted in the ascending serial
199 * number order on the list. Protected by cgroup_mutex.
201 static u64 css_serial_nr_next
= 1;
204 * These bitmask flags indicate whether tasks in the fork and exit paths have
205 * fork/exit handlers to call. This avoids us having to do extra work in the
206 * fork/exit path to check which subsystems have fork/exit callbacks.
208 static unsigned long have_fork_callback __read_mostly
;
209 static unsigned long have_exit_callback __read_mostly
;
211 /* Ditto for the can_fork callback. */
212 static unsigned long have_canfork_callback __read_mostly
;
214 static struct cftype cgroup_dfl_base_files
[];
215 static struct cftype cgroup_legacy_base_files
[];
217 static int rebind_subsystems(struct cgroup_root
*dst_root
,
218 unsigned long ss_mask
);
219 static int cgroup_destroy_locked(struct cgroup
*cgrp
);
220 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
222 static void css_release(struct percpu_ref
*ref
);
223 static void kill_css(struct cgroup_subsys_state
*css
);
224 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
228 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
229 * @ssid: subsys ID of interest
231 * cgroup_subsys_enabled() can only be used with literal subsys names which
232 * is fine for individual subsystems but unsuitable for cgroup core. This
233 * is slower static_key_enabled() based test indexed by @ssid.
235 static bool cgroup_ssid_enabled(int ssid
)
237 return static_key_enabled(cgroup_subsys_enabled_key
[ssid
]);
241 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
242 * @cgrp: the cgroup of interest
244 * The default hierarchy is the v2 interface of cgroup and this function
245 * can be used to test whether a cgroup is on the default hierarchy for
246 * cases where a subsystem should behave differnetly depending on the
249 * The set of behaviors which change on the default hierarchy are still
250 * being determined and the mount option is prefixed with __DEVEL__.
252 * List of changed behaviors:
254 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
255 * and "name" are disallowed.
257 * - When mounting an existing superblock, mount options should match.
259 * - Remount is disallowed.
261 * - rename(2) is disallowed.
263 * - "tasks" is removed. Everything should be at process granularity. Use
264 * "cgroup.procs" instead.
266 * - "cgroup.procs" is not sorted. pids will be unique unless they got
267 * recycled inbetween reads.
269 * - "release_agent" and "notify_on_release" are removed. Replacement
270 * notification mechanism will be implemented.
272 * - "cgroup.clone_children" is removed.
274 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
275 * and its descendants contain no task; otherwise, 1. The file also
276 * generates kernfs notification which can be monitored through poll and
277 * [di]notify when the value of the file changes.
279 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
280 * take masks of ancestors with non-empty cpus/mems, instead of being
281 * moved to an ancestor.
283 * - cpuset: a task can be moved into an empty cpuset, and again it takes
284 * masks of ancestors.
286 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
289 * - blkcg: blk-throttle becomes properly hierarchical.
291 * - debug: disallowed on the default hierarchy.
293 static bool cgroup_on_dfl(const struct cgroup
*cgrp
)
295 return cgrp
->root
== &cgrp_dfl_root
;
298 /* IDR wrappers which synchronize using cgroup_idr_lock */
299 static int cgroup_idr_alloc(struct idr
*idr
, void *ptr
, int start
, int end
,
304 idr_preload(gfp_mask
);
305 spin_lock_bh(&cgroup_idr_lock
);
306 ret
= idr_alloc(idr
, ptr
, start
, end
, gfp_mask
& ~__GFP_WAIT
);
307 spin_unlock_bh(&cgroup_idr_lock
);
312 static void *cgroup_idr_replace(struct idr
*idr
, void *ptr
, int id
)
316 spin_lock_bh(&cgroup_idr_lock
);
317 ret
= idr_replace(idr
, ptr
, id
);
318 spin_unlock_bh(&cgroup_idr_lock
);
322 static void cgroup_idr_remove(struct idr
*idr
, int id
)
324 spin_lock_bh(&cgroup_idr_lock
);
326 spin_unlock_bh(&cgroup_idr_lock
);
329 static struct cgroup
*cgroup_parent(struct cgroup
*cgrp
)
331 struct cgroup_subsys_state
*parent_css
= cgrp
->self
.parent
;
334 return container_of(parent_css
, struct cgroup
, self
);
339 * cgroup_css - obtain a cgroup's css for the specified subsystem
340 * @cgrp: the cgroup of interest
341 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
343 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
344 * function must be called either under cgroup_mutex or rcu_read_lock() and
345 * the caller is responsible for pinning the returned css if it wants to
346 * keep accessing it outside the said locks. This function may return
347 * %NULL if @cgrp doesn't have @subsys_id enabled.
349 static struct cgroup_subsys_state
*cgroup_css(struct cgroup
*cgrp
,
350 struct cgroup_subsys
*ss
)
353 return rcu_dereference_check(cgrp
->subsys
[ss
->id
],
354 lockdep_is_held(&cgroup_mutex
));
360 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
361 * @cgrp: the cgroup of interest
362 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
364 * Similar to cgroup_css() but returns the effective css, which is defined
365 * as the matching css of the nearest ancestor including self which has @ss
366 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
367 * function is guaranteed to return non-NULL css.
369 static struct cgroup_subsys_state
*cgroup_e_css(struct cgroup
*cgrp
,
370 struct cgroup_subsys
*ss
)
372 lockdep_assert_held(&cgroup_mutex
);
377 if (!(cgrp
->root
->subsys_mask
& (1 << ss
->id
)))
381 * This function is used while updating css associations and thus
382 * can't test the csses directly. Use ->child_subsys_mask.
384 while (cgroup_parent(cgrp
) &&
385 !(cgroup_parent(cgrp
)->child_subsys_mask
& (1 << ss
->id
)))
386 cgrp
= cgroup_parent(cgrp
);
388 return cgroup_css(cgrp
, ss
);
392 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
393 * @cgrp: the cgroup of interest
394 * @ss: the subsystem of interest
396 * Find and get the effective css of @cgrp for @ss. The effective css is
397 * defined as the matching css of the nearest ancestor including self which
398 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
399 * the root css is returned, so this function always returns a valid css.
400 * The returned css must be put using css_put().
402 struct cgroup_subsys_state
*cgroup_get_e_css(struct cgroup
*cgrp
,
403 struct cgroup_subsys
*ss
)
405 struct cgroup_subsys_state
*css
;
410 css
= cgroup_css(cgrp
, ss
);
412 if (css
&& css_tryget_online(css
))
414 cgrp
= cgroup_parent(cgrp
);
417 css
= init_css_set
.subsys
[ss
->id
];
424 /* convenient tests for these bits */
425 static inline bool cgroup_is_dead(const struct cgroup
*cgrp
)
427 return !(cgrp
->self
.flags
& CSS_ONLINE
);
430 struct cgroup_subsys_state
*of_css(struct kernfs_open_file
*of
)
432 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
433 struct cftype
*cft
= of_cft(of
);
436 * This is open and unprotected implementation of cgroup_css().
437 * seq_css() is only called from a kernfs file operation which has
438 * an active reference on the file. Because all the subsystem
439 * files are drained before a css is disassociated with a cgroup,
440 * the matching css from the cgroup's subsys table is guaranteed to
441 * be and stay valid until the enclosing operation is complete.
444 return rcu_dereference_raw(cgrp
->subsys
[cft
->ss
->id
]);
448 EXPORT_SYMBOL_GPL(of_css
);
451 * cgroup_is_descendant - test ancestry
452 * @cgrp: the cgroup to be tested
453 * @ancestor: possible ancestor of @cgrp
455 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
456 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
457 * and @ancestor are accessible.
459 bool cgroup_is_descendant(struct cgroup
*cgrp
, struct cgroup
*ancestor
)
462 if (cgrp
== ancestor
)
464 cgrp
= cgroup_parent(cgrp
);
469 static int notify_on_release(const struct cgroup
*cgrp
)
471 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
475 * for_each_css - iterate all css's of a cgroup
476 * @css: the iteration cursor
477 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
478 * @cgrp: the target cgroup to iterate css's of
480 * Should be called under cgroup_[tree_]mutex.
482 #define for_each_css(css, ssid, cgrp) \
483 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
484 if (!((css) = rcu_dereference_check( \
485 (cgrp)->subsys[(ssid)], \
486 lockdep_is_held(&cgroup_mutex)))) { } \
490 * for_each_e_css - iterate all effective css's of a cgroup
491 * @css: the iteration cursor
492 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
493 * @cgrp: the target cgroup to iterate css's of
495 * Should be called under cgroup_[tree_]mutex.
497 #define for_each_e_css(css, ssid, cgrp) \
498 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
499 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
504 * for_each_subsys - iterate all enabled cgroup subsystems
505 * @ss: the iteration cursor
506 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
508 #define for_each_subsys(ss, ssid) \
509 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
510 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
513 * for_each_subsys_which - filter for_each_subsys with a bitmask
514 * @ss: the iteration cursor
515 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
516 * @ss_maskp: a pointer to the bitmask
518 * The block will only run for cases where the ssid-th bit (1 << ssid) of
521 #define for_each_subsys_which(ss, ssid, ss_maskp) \
522 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
525 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
526 if (((ss) = cgroup_subsys[ssid]) && false) \
530 /* iterate across the hierarchies */
531 #define for_each_root(root) \
532 list_for_each_entry((root), &cgroup_roots, root_list)
534 /* iterate over child cgrps, lock should be held throughout iteration */
535 #define cgroup_for_each_live_child(child, cgrp) \
536 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
537 if (({ lockdep_assert_held(&cgroup_mutex); \
538 cgroup_is_dead(child); })) \
542 static void cgroup_release_agent(struct work_struct
*work
);
543 static void check_for_release(struct cgroup
*cgrp
);
546 * A cgroup can be associated with multiple css_sets as different tasks may
547 * belong to different cgroups on different hierarchies. In the other
548 * direction, a css_set is naturally associated with multiple cgroups.
549 * This M:N relationship is represented by the following link structure
550 * which exists for each association and allows traversing the associations
553 struct cgrp_cset_link
{
554 /* the cgroup and css_set this link associates */
556 struct css_set
*cset
;
558 /* list of cgrp_cset_links anchored at cgrp->cset_links */
559 struct list_head cset_link
;
561 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
562 struct list_head cgrp_link
;
566 * The default css_set - used by init and its children prior to any
567 * hierarchies being mounted. It contains a pointer to the root state
568 * for each subsystem. Also used to anchor the list of css_sets. Not
569 * reference-counted, to improve performance when child cgroups
570 * haven't been created.
572 struct css_set init_css_set
= {
573 .refcount
= ATOMIC_INIT(1),
574 .cgrp_links
= LIST_HEAD_INIT(init_css_set
.cgrp_links
),
575 .tasks
= LIST_HEAD_INIT(init_css_set
.tasks
),
576 .mg_tasks
= LIST_HEAD_INIT(init_css_set
.mg_tasks
),
577 .mg_preload_node
= LIST_HEAD_INIT(init_css_set
.mg_preload_node
),
578 .mg_node
= LIST_HEAD_INIT(init_css_set
.mg_node
),
581 static int css_set_count
= 1; /* 1 for init_css_set */
584 * cgroup_update_populated - updated populated count of a cgroup
585 * @cgrp: the target cgroup
586 * @populated: inc or dec populated count
588 * @cgrp is either getting the first task (css_set) or losing the last.
589 * Update @cgrp->populated_cnt accordingly. The count is propagated
590 * towards root so that a given cgroup's populated_cnt is zero iff the
591 * cgroup and all its descendants are empty.
593 * @cgrp's interface file "cgroup.populated" is zero if
594 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
595 * changes from or to zero, userland is notified that the content of the
596 * interface file has changed. This can be used to detect when @cgrp and
597 * its descendants become populated or empty.
599 static void cgroup_update_populated(struct cgroup
*cgrp
, bool populated
)
601 lockdep_assert_held(&css_set_rwsem
);
607 trigger
= !cgrp
->populated_cnt
++;
609 trigger
= !--cgrp
->populated_cnt
;
615 kernfs_notify(cgrp
->events_kn
);
616 cgrp
= cgroup_parent(cgrp
);
621 * hash table for cgroup groups. This improves the performance to find
622 * an existing css_set. This hash doesn't (currently) take into
623 * account cgroups in empty hierarchies.
625 #define CSS_SET_HASH_BITS 7
626 static DEFINE_HASHTABLE(css_set_table
, CSS_SET_HASH_BITS
);
628 static unsigned long css_set_hash(struct cgroup_subsys_state
*css
[])
630 unsigned long key
= 0UL;
631 struct cgroup_subsys
*ss
;
634 for_each_subsys(ss
, i
)
635 key
+= (unsigned long)css
[i
];
636 key
= (key
>> 16) ^ key
;
641 static void put_css_set_locked(struct css_set
*cset
)
643 struct cgrp_cset_link
*link
, *tmp_link
;
644 struct cgroup_subsys
*ss
;
647 lockdep_assert_held(&css_set_rwsem
);
649 if (!atomic_dec_and_test(&cset
->refcount
))
652 /* This css_set is dead. unlink it and release cgroup refcounts */
653 for_each_subsys(ss
, ssid
)
654 list_del(&cset
->e_cset_node
[ssid
]);
655 hash_del(&cset
->hlist
);
658 list_for_each_entry_safe(link
, tmp_link
, &cset
->cgrp_links
, cgrp_link
) {
659 struct cgroup
*cgrp
= link
->cgrp
;
661 list_del(&link
->cset_link
);
662 list_del(&link
->cgrp_link
);
664 /* @cgrp can't go away while we're holding css_set_rwsem */
665 if (list_empty(&cgrp
->cset_links
)) {
666 cgroup_update_populated(cgrp
, false);
667 check_for_release(cgrp
);
673 kfree_rcu(cset
, rcu_head
);
676 static void put_css_set(struct css_set
*cset
)
679 * Ensure that the refcount doesn't hit zero while any readers
680 * can see it. Similar to atomic_dec_and_lock(), but for an
683 if (atomic_add_unless(&cset
->refcount
, -1, 1))
686 down_write(&css_set_rwsem
);
687 put_css_set_locked(cset
);
688 up_write(&css_set_rwsem
);
692 * refcounted get/put for css_set objects
694 static inline void get_css_set(struct css_set
*cset
)
696 atomic_inc(&cset
->refcount
);
700 * compare_css_sets - helper function for find_existing_css_set().
701 * @cset: candidate css_set being tested
702 * @old_cset: existing css_set for a task
703 * @new_cgrp: cgroup that's being entered by the task
704 * @template: desired set of css pointers in css_set (pre-calculated)
706 * Returns true if "cset" matches "old_cset" except for the hierarchy
707 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
709 static bool compare_css_sets(struct css_set
*cset
,
710 struct css_set
*old_cset
,
711 struct cgroup
*new_cgrp
,
712 struct cgroup_subsys_state
*template[])
714 struct list_head
*l1
, *l2
;
717 * On the default hierarchy, there can be csets which are
718 * associated with the same set of cgroups but different csses.
719 * Let's first ensure that csses match.
721 if (memcmp(template, cset
->subsys
, sizeof(cset
->subsys
)))
725 * Compare cgroup pointers in order to distinguish between
726 * different cgroups in hierarchies. As different cgroups may
727 * share the same effective css, this comparison is always
730 l1
= &cset
->cgrp_links
;
731 l2
= &old_cset
->cgrp_links
;
733 struct cgrp_cset_link
*link1
, *link2
;
734 struct cgroup
*cgrp1
, *cgrp2
;
738 /* See if we reached the end - both lists are equal length. */
739 if (l1
== &cset
->cgrp_links
) {
740 BUG_ON(l2
!= &old_cset
->cgrp_links
);
743 BUG_ON(l2
== &old_cset
->cgrp_links
);
745 /* Locate the cgroups associated with these links. */
746 link1
= list_entry(l1
, struct cgrp_cset_link
, cgrp_link
);
747 link2
= list_entry(l2
, struct cgrp_cset_link
, cgrp_link
);
750 /* Hierarchies should be linked in the same order. */
751 BUG_ON(cgrp1
->root
!= cgrp2
->root
);
754 * If this hierarchy is the hierarchy of the cgroup
755 * that's changing, then we need to check that this
756 * css_set points to the new cgroup; if it's any other
757 * hierarchy, then this css_set should point to the
758 * same cgroup as the old css_set.
760 if (cgrp1
->root
== new_cgrp
->root
) {
761 if (cgrp1
!= new_cgrp
)
772 * find_existing_css_set - init css array and find the matching css_set
773 * @old_cset: the css_set that we're using before the cgroup transition
774 * @cgrp: the cgroup that we're moving into
775 * @template: out param for the new set of csses, should be clear on entry
777 static struct css_set
*find_existing_css_set(struct css_set
*old_cset
,
779 struct cgroup_subsys_state
*template[])
781 struct cgroup_root
*root
= cgrp
->root
;
782 struct cgroup_subsys
*ss
;
783 struct css_set
*cset
;
788 * Build the set of subsystem state objects that we want to see in the
789 * new css_set. while subsystems can change globally, the entries here
790 * won't change, so no need for locking.
792 for_each_subsys(ss
, i
) {
793 if (root
->subsys_mask
& (1UL << i
)) {
795 * @ss is in this hierarchy, so we want the
796 * effective css from @cgrp.
798 template[i
] = cgroup_e_css(cgrp
, ss
);
801 * @ss is not in this hierarchy, so we don't want
804 template[i
] = old_cset
->subsys
[i
];
808 key
= css_set_hash(template);
809 hash_for_each_possible(css_set_table
, cset
, hlist
, key
) {
810 if (!compare_css_sets(cset
, old_cset
, cgrp
, template))
813 /* This css_set matches what we need */
817 /* No existing cgroup group matched */
821 static void free_cgrp_cset_links(struct list_head
*links_to_free
)
823 struct cgrp_cset_link
*link
, *tmp_link
;
825 list_for_each_entry_safe(link
, tmp_link
, links_to_free
, cset_link
) {
826 list_del(&link
->cset_link
);
832 * allocate_cgrp_cset_links - allocate cgrp_cset_links
833 * @count: the number of links to allocate
834 * @tmp_links: list_head the allocated links are put on
836 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
837 * through ->cset_link. Returns 0 on success or -errno.
839 static int allocate_cgrp_cset_links(int count
, struct list_head
*tmp_links
)
841 struct cgrp_cset_link
*link
;
844 INIT_LIST_HEAD(tmp_links
);
846 for (i
= 0; i
< count
; i
++) {
847 link
= kzalloc(sizeof(*link
), GFP_KERNEL
);
849 free_cgrp_cset_links(tmp_links
);
852 list_add(&link
->cset_link
, tmp_links
);
858 * link_css_set - a helper function to link a css_set to a cgroup
859 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
860 * @cset: the css_set to be linked
861 * @cgrp: the destination cgroup
863 static void link_css_set(struct list_head
*tmp_links
, struct css_set
*cset
,
866 struct cgrp_cset_link
*link
;
868 BUG_ON(list_empty(tmp_links
));
870 if (cgroup_on_dfl(cgrp
))
871 cset
->dfl_cgrp
= cgrp
;
873 link
= list_first_entry(tmp_links
, struct cgrp_cset_link
, cset_link
);
877 if (list_empty(&cgrp
->cset_links
))
878 cgroup_update_populated(cgrp
, true);
879 list_move(&link
->cset_link
, &cgrp
->cset_links
);
882 * Always add links to the tail of the list so that the list
883 * is sorted by order of hierarchy creation
885 list_add_tail(&link
->cgrp_link
, &cset
->cgrp_links
);
889 * find_css_set - return a new css_set with one cgroup updated
890 * @old_cset: the baseline css_set
891 * @cgrp: the cgroup to be updated
893 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
894 * substituted into the appropriate hierarchy.
896 static struct css_set
*find_css_set(struct css_set
*old_cset
,
899 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
] = { };
900 struct css_set
*cset
;
901 struct list_head tmp_links
;
902 struct cgrp_cset_link
*link
;
903 struct cgroup_subsys
*ss
;
907 lockdep_assert_held(&cgroup_mutex
);
909 /* First see if we already have a cgroup group that matches
911 down_read(&css_set_rwsem
);
912 cset
= find_existing_css_set(old_cset
, cgrp
, template);
915 up_read(&css_set_rwsem
);
920 cset
= kzalloc(sizeof(*cset
), GFP_KERNEL
);
924 /* Allocate all the cgrp_cset_link objects that we'll need */
925 if (allocate_cgrp_cset_links(cgroup_root_count
, &tmp_links
) < 0) {
930 atomic_set(&cset
->refcount
, 1);
931 INIT_LIST_HEAD(&cset
->cgrp_links
);
932 INIT_LIST_HEAD(&cset
->tasks
);
933 INIT_LIST_HEAD(&cset
->mg_tasks
);
934 INIT_LIST_HEAD(&cset
->mg_preload_node
);
935 INIT_LIST_HEAD(&cset
->mg_node
);
936 INIT_HLIST_NODE(&cset
->hlist
);
938 /* Copy the set of subsystem state objects generated in
939 * find_existing_css_set() */
940 memcpy(cset
->subsys
, template, sizeof(cset
->subsys
));
942 down_write(&css_set_rwsem
);
943 /* Add reference counts and links from the new css_set. */
944 list_for_each_entry(link
, &old_cset
->cgrp_links
, cgrp_link
) {
945 struct cgroup
*c
= link
->cgrp
;
947 if (c
->root
== cgrp
->root
)
949 link_css_set(&tmp_links
, cset
, c
);
952 BUG_ON(!list_empty(&tmp_links
));
956 /* Add @cset to the hash table */
957 key
= css_set_hash(cset
->subsys
);
958 hash_add(css_set_table
, &cset
->hlist
, key
);
960 for_each_subsys(ss
, ssid
)
961 list_add_tail(&cset
->e_cset_node
[ssid
],
962 &cset
->subsys
[ssid
]->cgroup
->e_csets
[ssid
]);
964 up_write(&css_set_rwsem
);
969 static struct cgroup_root
*cgroup_root_from_kf(struct kernfs_root
*kf_root
)
971 struct cgroup
*root_cgrp
= kf_root
->kn
->priv
;
973 return root_cgrp
->root
;
976 static int cgroup_init_root_id(struct cgroup_root
*root
)
980 lockdep_assert_held(&cgroup_mutex
);
982 id
= idr_alloc_cyclic(&cgroup_hierarchy_idr
, root
, 0, 0, GFP_KERNEL
);
986 root
->hierarchy_id
= id
;
990 static void cgroup_exit_root_id(struct cgroup_root
*root
)
992 lockdep_assert_held(&cgroup_mutex
);
994 if (root
->hierarchy_id
) {
995 idr_remove(&cgroup_hierarchy_idr
, root
->hierarchy_id
);
996 root
->hierarchy_id
= 0;
1000 static void cgroup_free_root(struct cgroup_root
*root
)
1003 /* hierarchy ID should already have been released */
1004 WARN_ON_ONCE(root
->hierarchy_id
);
1006 idr_destroy(&root
->cgroup_idr
);
1011 static void cgroup_destroy_root(struct cgroup_root
*root
)
1013 struct cgroup
*cgrp
= &root
->cgrp
;
1014 struct cgrp_cset_link
*link
, *tmp_link
;
1016 mutex_lock(&cgroup_mutex
);
1018 BUG_ON(atomic_read(&root
->nr_cgrps
));
1019 BUG_ON(!list_empty(&cgrp
->self
.children
));
1021 /* Rebind all subsystems back to the default hierarchy */
1022 rebind_subsystems(&cgrp_dfl_root
, root
->subsys_mask
);
1025 * Release all the links from cset_links to this hierarchy's
1028 down_write(&css_set_rwsem
);
1030 list_for_each_entry_safe(link
, tmp_link
, &cgrp
->cset_links
, cset_link
) {
1031 list_del(&link
->cset_link
);
1032 list_del(&link
->cgrp_link
);
1035 up_write(&css_set_rwsem
);
1037 if (!list_empty(&root
->root_list
)) {
1038 list_del(&root
->root_list
);
1039 cgroup_root_count
--;
1042 cgroup_exit_root_id(root
);
1044 mutex_unlock(&cgroup_mutex
);
1046 kernfs_destroy_root(root
->kf_root
);
1047 cgroup_free_root(root
);
1050 /* look up cgroup associated with given css_set on the specified hierarchy */
1051 static struct cgroup
*cset_cgroup_from_root(struct css_set
*cset
,
1052 struct cgroup_root
*root
)
1054 struct cgroup
*res
= NULL
;
1056 lockdep_assert_held(&cgroup_mutex
);
1057 lockdep_assert_held(&css_set_rwsem
);
1059 if (cset
== &init_css_set
) {
1062 struct cgrp_cset_link
*link
;
1064 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
1065 struct cgroup
*c
= link
->cgrp
;
1067 if (c
->root
== root
) {
1079 * Return the cgroup for "task" from the given hierarchy. Must be
1080 * called with cgroup_mutex and css_set_rwsem held.
1082 static struct cgroup
*task_cgroup_from_root(struct task_struct
*task
,
1083 struct cgroup_root
*root
)
1086 * No need to lock the task - since we hold cgroup_mutex the
1087 * task can't change groups, so the only thing that can happen
1088 * is that it exits and its css is set back to init_css_set.
1090 return cset_cgroup_from_root(task_css_set(task
), root
);
1094 * A task must hold cgroup_mutex to modify cgroups.
1096 * Any task can increment and decrement the count field without lock.
1097 * So in general, code holding cgroup_mutex can't rely on the count
1098 * field not changing. However, if the count goes to zero, then only
1099 * cgroup_attach_task() can increment it again. Because a count of zero
1100 * means that no tasks are currently attached, therefore there is no
1101 * way a task attached to that cgroup can fork (the other way to
1102 * increment the count). So code holding cgroup_mutex can safely
1103 * assume that if the count is zero, it will stay zero. Similarly, if
1104 * a task holds cgroup_mutex on a cgroup with zero count, it
1105 * knows that the cgroup won't be removed, as cgroup_rmdir()
1108 * A cgroup can only be deleted if both its 'count' of using tasks
1109 * is zero, and its list of 'children' cgroups is empty. Since all
1110 * tasks in the system use _some_ cgroup, and since there is always at
1111 * least one task in the system (init, pid == 1), therefore, root cgroup
1112 * always has either children cgroups and/or using tasks. So we don't
1113 * need a special hack to ensure that root cgroup cannot be deleted.
1115 * P.S. One more locking exception. RCU is used to guard the
1116 * update of a tasks cgroup pointer by cgroup_attach_task()
1119 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
;
1120 static const struct file_operations proc_cgroupstats_operations
;
1122 static char *cgroup_file_name(struct cgroup
*cgrp
, const struct cftype
*cft
,
1125 struct cgroup_subsys
*ss
= cft
->ss
;
1127 if (cft
->ss
&& !(cft
->flags
& CFTYPE_NO_PREFIX
) &&
1128 !(cgrp
->root
->flags
& CGRP_ROOT_NOPREFIX
))
1129 snprintf(buf
, CGROUP_FILE_NAME_MAX
, "%s.%s",
1130 cgroup_on_dfl(cgrp
) ? ss
->name
: ss
->legacy_name
,
1133 strncpy(buf
, cft
->name
, CGROUP_FILE_NAME_MAX
);
1138 * cgroup_file_mode - deduce file mode of a control file
1139 * @cft: the control file in question
1141 * S_IRUGO for read, S_IWUSR for write.
1143 static umode_t
cgroup_file_mode(const struct cftype
*cft
)
1147 if (cft
->read_u64
|| cft
->read_s64
|| cft
->seq_show
)
1150 if (cft
->write_u64
|| cft
->write_s64
|| cft
->write
) {
1151 if (cft
->flags
& CFTYPE_WORLD_WRITABLE
)
1160 static void cgroup_get(struct cgroup
*cgrp
)
1162 WARN_ON_ONCE(cgroup_is_dead(cgrp
));
1163 css_get(&cgrp
->self
);
1166 static bool cgroup_tryget(struct cgroup
*cgrp
)
1168 return css_tryget(&cgrp
->self
);
1171 static void cgroup_put(struct cgroup
*cgrp
)
1173 css_put(&cgrp
->self
);
1177 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1178 * @cgrp: the target cgroup
1179 * @subtree_control: the new subtree_control mask to consider
1181 * On the default hierarchy, a subsystem may request other subsystems to be
1182 * enabled together through its ->depends_on mask. In such cases, more
1183 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1185 * This function calculates which subsystems need to be enabled if
1186 * @subtree_control is to be applied to @cgrp. The returned mask is always
1187 * a superset of @subtree_control and follows the usual hierarchy rules.
1189 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup
*cgrp
,
1190 unsigned long subtree_control
)
1192 struct cgroup
*parent
= cgroup_parent(cgrp
);
1193 unsigned long cur_ss_mask
= subtree_control
;
1194 struct cgroup_subsys
*ss
;
1197 lockdep_assert_held(&cgroup_mutex
);
1199 if (!cgroup_on_dfl(cgrp
))
1203 unsigned long new_ss_mask
= cur_ss_mask
;
1205 for_each_subsys_which(ss
, ssid
, &cur_ss_mask
)
1206 new_ss_mask
|= ss
->depends_on
;
1209 * Mask out subsystems which aren't available. This can
1210 * happen only if some depended-upon subsystems were bound
1211 * to non-default hierarchies.
1214 new_ss_mask
&= parent
->child_subsys_mask
;
1216 new_ss_mask
&= cgrp
->root
->subsys_mask
;
1218 if (new_ss_mask
== cur_ss_mask
)
1220 cur_ss_mask
= new_ss_mask
;
1227 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1228 * @cgrp: the target cgroup
1230 * Update @cgrp->child_subsys_mask according to the current
1231 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1233 static void cgroup_refresh_child_subsys_mask(struct cgroup
*cgrp
)
1235 cgrp
->child_subsys_mask
=
1236 cgroup_calc_child_subsys_mask(cgrp
, cgrp
->subtree_control
);
1240 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1241 * @kn: the kernfs_node being serviced
1243 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1244 * the method finishes if locking succeeded. Note that once this function
1245 * returns the cgroup returned by cgroup_kn_lock_live() may become
1246 * inaccessible any time. If the caller intends to continue to access the
1247 * cgroup, it should pin it before invoking this function.
1249 static void cgroup_kn_unlock(struct kernfs_node
*kn
)
1251 struct cgroup
*cgrp
;
1253 if (kernfs_type(kn
) == KERNFS_DIR
)
1256 cgrp
= kn
->parent
->priv
;
1258 mutex_unlock(&cgroup_mutex
);
1260 kernfs_unbreak_active_protection(kn
);
1265 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1266 * @kn: the kernfs_node being serviced
1268 * This helper is to be used by a cgroup kernfs method currently servicing
1269 * @kn. It breaks the active protection, performs cgroup locking and
1270 * verifies that the associated cgroup is alive. Returns the cgroup if
1271 * alive; otherwise, %NULL. A successful return should be undone by a
1272 * matching cgroup_kn_unlock() invocation.
1274 * Any cgroup kernfs method implementation which requires locking the
1275 * associated cgroup should use this helper. It avoids nesting cgroup
1276 * locking under kernfs active protection and allows all kernfs operations
1277 * including self-removal.
1279 static struct cgroup
*cgroup_kn_lock_live(struct kernfs_node
*kn
)
1281 struct cgroup
*cgrp
;
1283 if (kernfs_type(kn
) == KERNFS_DIR
)
1286 cgrp
= kn
->parent
->priv
;
1289 * We're gonna grab cgroup_mutex which nests outside kernfs
1290 * active_ref. cgroup liveliness check alone provides enough
1291 * protection against removal. Ensure @cgrp stays accessible and
1292 * break the active_ref protection.
1294 if (!cgroup_tryget(cgrp
))
1296 kernfs_break_active_protection(kn
);
1298 mutex_lock(&cgroup_mutex
);
1300 if (!cgroup_is_dead(cgrp
))
1303 cgroup_kn_unlock(kn
);
1307 static void cgroup_rm_file(struct cgroup
*cgrp
, const struct cftype
*cft
)
1309 char name
[CGROUP_FILE_NAME_MAX
];
1311 lockdep_assert_held(&cgroup_mutex
);
1312 kernfs_remove_by_name(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
));
1316 * cgroup_clear_dir - remove subsys files in a cgroup directory
1317 * @cgrp: target cgroup
1318 * @subsys_mask: mask of the subsystem ids whose files should be removed
1320 static void cgroup_clear_dir(struct cgroup
*cgrp
, unsigned long subsys_mask
)
1322 struct cgroup_subsys
*ss
;
1325 for_each_subsys(ss
, i
) {
1326 struct cftype
*cfts
;
1328 if (!(subsys_mask
& (1 << i
)))
1330 list_for_each_entry(cfts
, &ss
->cfts
, node
)
1331 cgroup_addrm_files(cgrp
, cfts
, false);
1336 * cgroup_populate_dir - create subsys files in a cgroup directory
1337 * @cgrp: target cgroup
1338 * @subsys_mask: mask of the subsystem ids whose files should be added
1340 * On failure, no file is added.
1342 static int cgroup_populate_dir(struct cgroup
*cgrp
, unsigned long subsys_mask
)
1344 struct cgroup_subsys
*ss
;
1347 /* process cftsets of each subsystem */
1348 for_each_subsys(ss
, i
) {
1349 struct cftype
*cfts
;
1351 if (!(subsys_mask
& (1 << i
)))
1354 list_for_each_entry(cfts
, &ss
->cfts
, node
) {
1355 ret
= cgroup_addrm_files(cgrp
, cfts
, true);
1362 cgroup_clear_dir(cgrp
, subsys_mask
);
1366 static int rebind_subsystems(struct cgroup_root
*dst_root
,
1367 unsigned long ss_mask
)
1369 struct cgroup
*dcgrp
= &dst_root
->cgrp
;
1370 struct cgroup_subsys
*ss
;
1371 unsigned long tmp_ss_mask
;
1374 lockdep_assert_held(&cgroup_mutex
);
1376 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
1377 /* if @ss has non-root csses attached to it, can't move */
1378 if (css_next_child(NULL
, cgroup_css(&ss
->root
->cgrp
, ss
)))
1381 /* can't move between two non-dummy roots either */
1382 if (ss
->root
!= &cgrp_dfl_root
&& dst_root
!= &cgrp_dfl_root
)
1386 /* skip creating root files on dfl_root for inhibited subsystems */
1387 tmp_ss_mask
= ss_mask
;
1388 if (dst_root
== &cgrp_dfl_root
)
1389 tmp_ss_mask
&= ~cgrp_dfl_root_inhibit_ss_mask
;
1391 ret
= cgroup_populate_dir(dcgrp
, tmp_ss_mask
);
1393 if (dst_root
!= &cgrp_dfl_root
)
1397 * Rebinding back to the default root is not allowed to
1398 * fail. Using both default and non-default roots should
1399 * be rare. Moving subsystems back and forth even more so.
1400 * Just warn about it and continue.
1402 if (cgrp_dfl_root_visible
) {
1403 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1405 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1410 * Nothing can fail from this point on. Remove files for the
1411 * removed subsystems and rebind each subsystem.
1413 for_each_subsys_which(ss
, ssid
, &ss_mask
)
1414 cgroup_clear_dir(&ss
->root
->cgrp
, 1 << ssid
);
1416 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
1417 struct cgroup_root
*src_root
= ss
->root
;
1418 struct cgroup
*scgrp
= &src_root
->cgrp
;
1419 struct cgroup_subsys_state
*css
= cgroup_css(scgrp
, ss
);
1420 struct css_set
*cset
;
1422 WARN_ON(!css
|| cgroup_css(dcgrp
, ss
));
1424 RCU_INIT_POINTER(scgrp
->subsys
[ssid
], NULL
);
1425 rcu_assign_pointer(dcgrp
->subsys
[ssid
], css
);
1426 ss
->root
= dst_root
;
1427 css
->cgroup
= dcgrp
;
1429 down_write(&css_set_rwsem
);
1430 hash_for_each(css_set_table
, i
, cset
, hlist
)
1431 list_move_tail(&cset
->e_cset_node
[ss
->id
],
1432 &dcgrp
->e_csets
[ss
->id
]);
1433 up_write(&css_set_rwsem
);
1435 src_root
->subsys_mask
&= ~(1 << ssid
);
1436 scgrp
->subtree_control
&= ~(1 << ssid
);
1437 cgroup_refresh_child_subsys_mask(scgrp
);
1439 /* default hierarchy doesn't enable controllers by default */
1440 dst_root
->subsys_mask
|= 1 << ssid
;
1441 if (dst_root
== &cgrp_dfl_root
) {
1442 static_branch_enable(cgroup_subsys_on_dfl_key
[ssid
]);
1444 dcgrp
->subtree_control
|= 1 << ssid
;
1445 cgroup_refresh_child_subsys_mask(dcgrp
);
1446 static_branch_disable(cgroup_subsys_on_dfl_key
[ssid
]);
1453 kernfs_activate(dcgrp
->kn
);
1457 static int cgroup_show_options(struct seq_file
*seq
,
1458 struct kernfs_root
*kf_root
)
1460 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1461 struct cgroup_subsys
*ss
;
1464 if (root
!= &cgrp_dfl_root
)
1465 for_each_subsys(ss
, ssid
)
1466 if (root
->subsys_mask
& (1 << ssid
))
1467 seq_show_option(seq
, ss
->legacy_name
, NULL
);
1468 if (root
->flags
& CGRP_ROOT_NOPREFIX
)
1469 seq_puts(seq
, ",noprefix");
1470 if (root
->flags
& CGRP_ROOT_XATTR
)
1471 seq_puts(seq
, ",xattr");
1473 spin_lock(&release_agent_path_lock
);
1474 if (strlen(root
->release_agent_path
))
1475 seq_show_option(seq
, "release_agent",
1476 root
->release_agent_path
);
1477 spin_unlock(&release_agent_path_lock
);
1479 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
))
1480 seq_puts(seq
, ",clone_children");
1481 if (strlen(root
->name
))
1482 seq_show_option(seq
, "name", root
->name
);
1486 struct cgroup_sb_opts
{
1487 unsigned long subsys_mask
;
1489 char *release_agent
;
1490 bool cpuset_clone_children
;
1492 /* User explicitly requested empty subsystem */
1496 static int parse_cgroupfs_options(char *data
, struct cgroup_sb_opts
*opts
)
1498 char *token
, *o
= data
;
1499 bool all_ss
= false, one_ss
= false;
1500 unsigned long mask
= -1UL;
1501 struct cgroup_subsys
*ss
;
1505 #ifdef CONFIG_CPUSETS
1506 mask
= ~(1U << cpuset_cgrp_id
);
1509 memset(opts
, 0, sizeof(*opts
));
1511 while ((token
= strsep(&o
, ",")) != NULL
) {
1516 if (!strcmp(token
, "none")) {
1517 /* Explicitly have no subsystems */
1521 if (!strcmp(token
, "all")) {
1522 /* Mutually exclusive option 'all' + subsystem name */
1528 if (!strcmp(token
, "__DEVEL__sane_behavior")) {
1529 opts
->flags
|= CGRP_ROOT_SANE_BEHAVIOR
;
1532 if (!strcmp(token
, "noprefix")) {
1533 opts
->flags
|= CGRP_ROOT_NOPREFIX
;
1536 if (!strcmp(token
, "clone_children")) {
1537 opts
->cpuset_clone_children
= true;
1540 if (!strcmp(token
, "xattr")) {
1541 opts
->flags
|= CGRP_ROOT_XATTR
;
1544 if (!strncmp(token
, "release_agent=", 14)) {
1545 /* Specifying two release agents is forbidden */
1546 if (opts
->release_agent
)
1548 opts
->release_agent
=
1549 kstrndup(token
+ 14, PATH_MAX
- 1, GFP_KERNEL
);
1550 if (!opts
->release_agent
)
1554 if (!strncmp(token
, "name=", 5)) {
1555 const char *name
= token
+ 5;
1556 /* Can't specify an empty name */
1559 /* Must match [\w.-]+ */
1560 for (i
= 0; i
< strlen(name
); i
++) {
1564 if ((c
== '.') || (c
== '-') || (c
== '_'))
1568 /* Specifying two names is forbidden */
1571 opts
->name
= kstrndup(name
,
1572 MAX_CGROUP_ROOT_NAMELEN
- 1,
1580 for_each_subsys(ss
, i
) {
1581 if (strcmp(token
, ss
->legacy_name
))
1583 if (!cgroup_ssid_enabled(i
))
1586 /* Mutually exclusive option 'all' + subsystem name */
1589 opts
->subsys_mask
|= (1 << i
);
1594 if (i
== CGROUP_SUBSYS_COUNT
)
1598 if (opts
->flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1599 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1601 pr_err("sane_behavior: no other mount options allowed\n");
1608 * If the 'all' option was specified select all the subsystems,
1609 * otherwise if 'none', 'name=' and a subsystem name options were
1610 * not specified, let's default to 'all'
1612 if (all_ss
|| (!one_ss
&& !opts
->none
&& !opts
->name
))
1613 for_each_subsys(ss
, i
)
1614 if (cgroup_ssid_enabled(i
))
1615 opts
->subsys_mask
|= (1 << i
);
1618 * We either have to specify by name or by subsystems. (So all
1619 * empty hierarchies must have a name).
1621 if (!opts
->subsys_mask
&& !opts
->name
)
1625 * Option noprefix was introduced just for backward compatibility
1626 * with the old cpuset, so we allow noprefix only if mounting just
1627 * the cpuset subsystem.
1629 if ((opts
->flags
& CGRP_ROOT_NOPREFIX
) && (opts
->subsys_mask
& mask
))
1632 /* Can't specify "none" and some subsystems */
1633 if (opts
->subsys_mask
&& opts
->none
)
1639 static int cgroup_remount(struct kernfs_root
*kf_root
, int *flags
, char *data
)
1642 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
1643 struct cgroup_sb_opts opts
;
1644 unsigned long added_mask
, removed_mask
;
1646 if (root
== &cgrp_dfl_root
) {
1647 pr_err("remount is not allowed\n");
1651 mutex_lock(&cgroup_mutex
);
1653 /* See what subsystems are wanted */
1654 ret
= parse_cgroupfs_options(data
, &opts
);
1658 if (opts
.subsys_mask
!= root
->subsys_mask
|| opts
.release_agent
)
1659 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1660 task_tgid_nr(current
), current
->comm
);
1662 added_mask
= opts
.subsys_mask
& ~root
->subsys_mask
;
1663 removed_mask
= root
->subsys_mask
& ~opts
.subsys_mask
;
1665 /* Don't allow flags or name to change at remount */
1666 if ((opts
.flags
^ root
->flags
) ||
1667 (opts
.name
&& strcmp(opts
.name
, root
->name
))) {
1668 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1669 opts
.flags
, opts
.name
?: "", root
->flags
, root
->name
);
1674 /* remounting is not allowed for populated hierarchies */
1675 if (!list_empty(&root
->cgrp
.self
.children
)) {
1680 ret
= rebind_subsystems(root
, added_mask
);
1684 rebind_subsystems(&cgrp_dfl_root
, removed_mask
);
1686 if (opts
.release_agent
) {
1687 spin_lock(&release_agent_path_lock
);
1688 strcpy(root
->release_agent_path
, opts
.release_agent
);
1689 spin_unlock(&release_agent_path_lock
);
1692 kfree(opts
.release_agent
);
1694 mutex_unlock(&cgroup_mutex
);
1699 * To reduce the fork() overhead for systems that are not actually using
1700 * their cgroups capability, we don't maintain the lists running through
1701 * each css_set to its tasks until we see the list actually used - in other
1702 * words after the first mount.
1704 static bool use_task_css_set_links __read_mostly
;
1706 static void cgroup_enable_task_cg_lists(void)
1708 struct task_struct
*p
, *g
;
1710 down_write(&css_set_rwsem
);
1712 if (use_task_css_set_links
)
1715 use_task_css_set_links
= true;
1718 * We need tasklist_lock because RCU is not safe against
1719 * while_each_thread(). Besides, a forking task that has passed
1720 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1721 * is not guaranteed to have its child immediately visible in the
1722 * tasklist if we walk through it with RCU.
1724 read_lock(&tasklist_lock
);
1725 do_each_thread(g
, p
) {
1726 WARN_ON_ONCE(!list_empty(&p
->cg_list
) ||
1727 task_css_set(p
) != &init_css_set
);
1730 * We should check if the process is exiting, otherwise
1731 * it will race with cgroup_exit() in that the list
1732 * entry won't be deleted though the process has exited.
1733 * Do it while holding siglock so that we don't end up
1734 * racing against cgroup_exit().
1736 spin_lock_irq(&p
->sighand
->siglock
);
1737 if (!(p
->flags
& PF_EXITING
)) {
1738 struct css_set
*cset
= task_css_set(p
);
1740 list_add(&p
->cg_list
, &cset
->tasks
);
1743 spin_unlock_irq(&p
->sighand
->siglock
);
1744 } while_each_thread(g
, p
);
1745 read_unlock(&tasklist_lock
);
1747 up_write(&css_set_rwsem
);
1750 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
1752 struct cgroup_subsys
*ss
;
1755 INIT_LIST_HEAD(&cgrp
->self
.sibling
);
1756 INIT_LIST_HEAD(&cgrp
->self
.children
);
1757 INIT_LIST_HEAD(&cgrp
->cset_links
);
1758 INIT_LIST_HEAD(&cgrp
->pidlists
);
1759 mutex_init(&cgrp
->pidlist_mutex
);
1760 cgrp
->self
.cgroup
= cgrp
;
1761 cgrp
->self
.flags
|= CSS_ONLINE
;
1763 for_each_subsys(ss
, ssid
)
1764 INIT_LIST_HEAD(&cgrp
->e_csets
[ssid
]);
1766 init_waitqueue_head(&cgrp
->offline_waitq
);
1767 INIT_WORK(&cgrp
->release_agent_work
, cgroup_release_agent
);
1770 static void init_cgroup_root(struct cgroup_root
*root
,
1771 struct cgroup_sb_opts
*opts
)
1773 struct cgroup
*cgrp
= &root
->cgrp
;
1775 INIT_LIST_HEAD(&root
->root_list
);
1776 atomic_set(&root
->nr_cgrps
, 1);
1778 init_cgroup_housekeeping(cgrp
);
1779 idr_init(&root
->cgroup_idr
);
1781 root
->flags
= opts
->flags
;
1782 if (opts
->release_agent
)
1783 strcpy(root
->release_agent_path
, opts
->release_agent
);
1785 strcpy(root
->name
, opts
->name
);
1786 if (opts
->cpuset_clone_children
)
1787 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &root
->cgrp
.flags
);
1790 static int cgroup_setup_root(struct cgroup_root
*root
, unsigned long ss_mask
)
1792 LIST_HEAD(tmp_links
);
1793 struct cgroup
*root_cgrp
= &root
->cgrp
;
1794 struct cftype
*base_files
;
1795 struct css_set
*cset
;
1798 lockdep_assert_held(&cgroup_mutex
);
1800 ret
= cgroup_idr_alloc(&root
->cgroup_idr
, root_cgrp
, 1, 2, GFP_KERNEL
);
1803 root_cgrp
->id
= ret
;
1805 ret
= percpu_ref_init(&root_cgrp
->self
.refcnt
, css_release
, 0,
1811 * We're accessing css_set_count without locking css_set_rwsem here,
1812 * but that's OK - it can only be increased by someone holding
1813 * cgroup_lock, and that's us. The worst that can happen is that we
1814 * have some link structures left over
1816 ret
= allocate_cgrp_cset_links(css_set_count
, &tmp_links
);
1820 ret
= cgroup_init_root_id(root
);
1824 root
->kf_root
= kernfs_create_root(&cgroup_kf_syscall_ops
,
1825 KERNFS_ROOT_CREATE_DEACTIVATED
,
1827 if (IS_ERR(root
->kf_root
)) {
1828 ret
= PTR_ERR(root
->kf_root
);
1831 root_cgrp
->kn
= root
->kf_root
->kn
;
1833 if (root
== &cgrp_dfl_root
)
1834 base_files
= cgroup_dfl_base_files
;
1836 base_files
= cgroup_legacy_base_files
;
1838 ret
= cgroup_addrm_files(root_cgrp
, base_files
, true);
1842 ret
= rebind_subsystems(root
, ss_mask
);
1847 * There must be no failure case after here, since rebinding takes
1848 * care of subsystems' refcounts, which are explicitly dropped in
1849 * the failure exit path.
1851 list_add(&root
->root_list
, &cgroup_roots
);
1852 cgroup_root_count
++;
1855 * Link the root cgroup in this hierarchy into all the css_set
1858 down_write(&css_set_rwsem
);
1859 hash_for_each(css_set_table
, i
, cset
, hlist
)
1860 link_css_set(&tmp_links
, cset
, root_cgrp
);
1861 up_write(&css_set_rwsem
);
1863 BUG_ON(!list_empty(&root_cgrp
->self
.children
));
1864 BUG_ON(atomic_read(&root
->nr_cgrps
) != 1);
1866 kernfs_activate(root_cgrp
->kn
);
1871 kernfs_destroy_root(root
->kf_root
);
1872 root
->kf_root
= NULL
;
1874 cgroup_exit_root_id(root
);
1876 percpu_ref_exit(&root_cgrp
->self
.refcnt
);
1878 free_cgrp_cset_links(&tmp_links
);
1882 static struct dentry
*cgroup_mount(struct file_system_type
*fs_type
,
1883 int flags
, const char *unused_dev_name
,
1886 struct super_block
*pinned_sb
= NULL
;
1887 struct cgroup_subsys
*ss
;
1888 struct cgroup_root
*root
;
1889 struct cgroup_sb_opts opts
;
1890 struct dentry
*dentry
;
1896 * The first time anyone tries to mount a cgroup, enable the list
1897 * linking each css_set to its tasks and fix up all existing tasks.
1899 if (!use_task_css_set_links
)
1900 cgroup_enable_task_cg_lists();
1902 mutex_lock(&cgroup_mutex
);
1904 /* First find the desired set of subsystems */
1905 ret
= parse_cgroupfs_options(data
, &opts
);
1909 /* look for a matching existing root */
1910 if (opts
.flags
& CGRP_ROOT_SANE_BEHAVIOR
) {
1911 cgrp_dfl_root_visible
= true;
1912 root
= &cgrp_dfl_root
;
1913 cgroup_get(&root
->cgrp
);
1919 * Destruction of cgroup root is asynchronous, so subsystems may
1920 * still be dying after the previous unmount. Let's drain the
1921 * dying subsystems. We just need to ensure that the ones
1922 * unmounted previously finish dying and don't care about new ones
1923 * starting. Testing ref liveliness is good enough.
1925 for_each_subsys(ss
, i
) {
1926 if (!(opts
.subsys_mask
& (1 << i
)) ||
1927 ss
->root
== &cgrp_dfl_root
)
1930 if (!percpu_ref_tryget_live(&ss
->root
->cgrp
.self
.refcnt
)) {
1931 mutex_unlock(&cgroup_mutex
);
1933 ret
= restart_syscall();
1936 cgroup_put(&ss
->root
->cgrp
);
1939 for_each_root(root
) {
1940 bool name_match
= false;
1942 if (root
== &cgrp_dfl_root
)
1946 * If we asked for a name then it must match. Also, if
1947 * name matches but sybsys_mask doesn't, we should fail.
1948 * Remember whether name matched.
1951 if (strcmp(opts
.name
, root
->name
))
1957 * If we asked for subsystems (or explicitly for no
1958 * subsystems) then they must match.
1960 if ((opts
.subsys_mask
|| opts
.none
) &&
1961 (opts
.subsys_mask
!= root
->subsys_mask
)) {
1968 if (root
->flags
^ opts
.flags
)
1969 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1972 * We want to reuse @root whose lifetime is governed by its
1973 * ->cgrp. Let's check whether @root is alive and keep it
1974 * that way. As cgroup_kill_sb() can happen anytime, we
1975 * want to block it by pinning the sb so that @root doesn't
1976 * get killed before mount is complete.
1978 * With the sb pinned, tryget_live can reliably indicate
1979 * whether @root can be reused. If it's being killed,
1980 * drain it. We can use wait_queue for the wait but this
1981 * path is super cold. Let's just sleep a bit and retry.
1983 pinned_sb
= kernfs_pin_sb(root
->kf_root
, NULL
);
1984 if (IS_ERR(pinned_sb
) ||
1985 !percpu_ref_tryget_live(&root
->cgrp
.self
.refcnt
)) {
1986 mutex_unlock(&cgroup_mutex
);
1987 if (!IS_ERR_OR_NULL(pinned_sb
))
1988 deactivate_super(pinned_sb
);
1990 ret
= restart_syscall();
1999 * No such thing, create a new one. name= matching without subsys
2000 * specification is allowed for already existing hierarchies but we
2001 * can't create new one without subsys specification.
2003 if (!opts
.subsys_mask
&& !opts
.none
) {
2008 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
2014 init_cgroup_root(root
, &opts
);
2016 ret
= cgroup_setup_root(root
, opts
.subsys_mask
);
2018 cgroup_free_root(root
);
2021 mutex_unlock(&cgroup_mutex
);
2023 kfree(opts
.release_agent
);
2027 return ERR_PTR(ret
);
2029 dentry
= kernfs_mount(fs_type
, flags
, root
->kf_root
,
2030 CGROUP_SUPER_MAGIC
, &new_sb
);
2031 if (IS_ERR(dentry
) || !new_sb
)
2032 cgroup_put(&root
->cgrp
);
2035 * If @pinned_sb, we're reusing an existing root and holding an
2036 * extra ref on its sb. Mount is complete. Put the extra ref.
2040 deactivate_super(pinned_sb
);
2046 static void cgroup_kill_sb(struct super_block
*sb
)
2048 struct kernfs_root
*kf_root
= kernfs_root_from_sb(sb
);
2049 struct cgroup_root
*root
= cgroup_root_from_kf(kf_root
);
2052 * If @root doesn't have any mounts or children, start killing it.
2053 * This prevents new mounts by disabling percpu_ref_tryget_live().
2054 * cgroup_mount() may wait for @root's release.
2056 * And don't kill the default root.
2058 if (!list_empty(&root
->cgrp
.self
.children
) ||
2059 root
== &cgrp_dfl_root
)
2060 cgroup_put(&root
->cgrp
);
2062 percpu_ref_kill(&root
->cgrp
.self
.refcnt
);
2067 static struct file_system_type cgroup_fs_type
= {
2069 .mount
= cgroup_mount
,
2070 .kill_sb
= cgroup_kill_sb
,
2074 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2075 * @task: target task
2076 * @buf: the buffer to write the path into
2077 * @buflen: the length of the buffer
2079 * Determine @task's cgroup on the first (the one with the lowest non-zero
2080 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2081 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2082 * cgroup controller callbacks.
2084 * Return value is the same as kernfs_path().
2086 char *task_cgroup_path(struct task_struct
*task
, char *buf
, size_t buflen
)
2088 struct cgroup_root
*root
;
2089 struct cgroup
*cgrp
;
2090 int hierarchy_id
= 1;
2093 mutex_lock(&cgroup_mutex
);
2094 down_read(&css_set_rwsem
);
2096 root
= idr_get_next(&cgroup_hierarchy_idr
, &hierarchy_id
);
2099 cgrp
= task_cgroup_from_root(task
, root
);
2100 path
= cgroup_path(cgrp
, buf
, buflen
);
2102 /* if no hierarchy exists, everyone is in "/" */
2103 if (strlcpy(buf
, "/", buflen
) < buflen
)
2107 up_read(&css_set_rwsem
);
2108 mutex_unlock(&cgroup_mutex
);
2111 EXPORT_SYMBOL_GPL(task_cgroup_path
);
2113 /* used to track tasks and other necessary states during migration */
2114 struct cgroup_taskset
{
2115 /* the src and dst cset list running through cset->mg_node */
2116 struct list_head src_csets
;
2117 struct list_head dst_csets
;
2120 * Fields for cgroup_taskset_*() iteration.
2122 * Before migration is committed, the target migration tasks are on
2123 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2124 * the csets on ->dst_csets. ->csets point to either ->src_csets
2125 * or ->dst_csets depending on whether migration is committed.
2127 * ->cur_csets and ->cur_task point to the current task position
2130 struct list_head
*csets
;
2131 struct css_set
*cur_cset
;
2132 struct task_struct
*cur_task
;
2136 * cgroup_taskset_first - reset taskset and return the first task
2137 * @tset: taskset of interest
2139 * @tset iteration is initialized and the first task is returned.
2141 struct task_struct
*cgroup_taskset_first(struct cgroup_taskset
*tset
)
2143 tset
->cur_cset
= list_first_entry(tset
->csets
, struct css_set
, mg_node
);
2144 tset
->cur_task
= NULL
;
2146 return cgroup_taskset_next(tset
);
2150 * cgroup_taskset_next - iterate to the next task in taskset
2151 * @tset: taskset of interest
2153 * Return the next task in @tset. Iteration must have been initialized
2154 * with cgroup_taskset_first().
2156 struct task_struct
*cgroup_taskset_next(struct cgroup_taskset
*tset
)
2158 struct css_set
*cset
= tset
->cur_cset
;
2159 struct task_struct
*task
= tset
->cur_task
;
2161 while (&cset
->mg_node
!= tset
->csets
) {
2163 task
= list_first_entry(&cset
->mg_tasks
,
2164 struct task_struct
, cg_list
);
2166 task
= list_next_entry(task
, cg_list
);
2168 if (&task
->cg_list
!= &cset
->mg_tasks
) {
2169 tset
->cur_cset
= cset
;
2170 tset
->cur_task
= task
;
2174 cset
= list_next_entry(cset
, mg_node
);
2182 * cgroup_task_migrate - move a task from one cgroup to another.
2183 * @old_cgrp: the cgroup @tsk is being migrated from
2184 * @tsk: the task being migrated
2185 * @new_cset: the new css_set @tsk is being attached to
2187 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2189 static void cgroup_task_migrate(struct cgroup
*old_cgrp
,
2190 struct task_struct
*tsk
,
2191 struct css_set
*new_cset
)
2193 struct css_set
*old_cset
;
2195 lockdep_assert_held(&cgroup_mutex
);
2196 lockdep_assert_held(&css_set_rwsem
);
2199 * We are synchronized through cgroup_threadgroup_rwsem against
2200 * PF_EXITING setting such that we can't race against cgroup_exit()
2201 * changing the css_set to init_css_set and dropping the old one.
2203 WARN_ON_ONCE(tsk
->flags
& PF_EXITING
);
2204 old_cset
= task_css_set(tsk
);
2206 get_css_set(new_cset
);
2207 rcu_assign_pointer(tsk
->cgroups
, new_cset
);
2210 * Use move_tail so that cgroup_taskset_first() still returns the
2211 * leader after migration. This works because cgroup_migrate()
2212 * ensures that the dst_cset of the leader is the first on the
2213 * tset's dst_csets list.
2215 list_move_tail(&tsk
->cg_list
, &new_cset
->mg_tasks
);
2218 * We just gained a reference on old_cset by taking it from the
2219 * task. As trading it for new_cset is protected by cgroup_mutex,
2220 * we're safe to drop it here; it will be freed under RCU.
2222 put_css_set_locked(old_cset
);
2226 * cgroup_migrate_finish - cleanup after attach
2227 * @preloaded_csets: list of preloaded css_sets
2229 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2230 * those functions for details.
2232 static void cgroup_migrate_finish(struct list_head
*preloaded_csets
)
2234 struct css_set
*cset
, *tmp_cset
;
2236 lockdep_assert_held(&cgroup_mutex
);
2238 down_write(&css_set_rwsem
);
2239 list_for_each_entry_safe(cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2240 cset
->mg_src_cgrp
= NULL
;
2241 cset
->mg_dst_cset
= NULL
;
2242 list_del_init(&cset
->mg_preload_node
);
2243 put_css_set_locked(cset
);
2245 up_write(&css_set_rwsem
);
2249 * cgroup_migrate_add_src - add a migration source css_set
2250 * @src_cset: the source css_set to add
2251 * @dst_cgrp: the destination cgroup
2252 * @preloaded_csets: list of preloaded css_sets
2254 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2255 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2256 * up by cgroup_migrate_finish().
2258 * This function may be called without holding cgroup_threadgroup_rwsem
2259 * even if the target is a process. Threads may be created and destroyed
2260 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2261 * into play and the preloaded css_sets are guaranteed to cover all
2264 static void cgroup_migrate_add_src(struct css_set
*src_cset
,
2265 struct cgroup
*dst_cgrp
,
2266 struct list_head
*preloaded_csets
)
2268 struct cgroup
*src_cgrp
;
2270 lockdep_assert_held(&cgroup_mutex
);
2271 lockdep_assert_held(&css_set_rwsem
);
2273 src_cgrp
= cset_cgroup_from_root(src_cset
, dst_cgrp
->root
);
2275 if (!list_empty(&src_cset
->mg_preload_node
))
2278 WARN_ON(src_cset
->mg_src_cgrp
);
2279 WARN_ON(!list_empty(&src_cset
->mg_tasks
));
2280 WARN_ON(!list_empty(&src_cset
->mg_node
));
2282 src_cset
->mg_src_cgrp
= src_cgrp
;
2283 get_css_set(src_cset
);
2284 list_add(&src_cset
->mg_preload_node
, preloaded_csets
);
2288 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2289 * @dst_cgrp: the destination cgroup (may be %NULL)
2290 * @preloaded_csets: list of preloaded source css_sets
2292 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2293 * have been preloaded to @preloaded_csets. This function looks up and
2294 * pins all destination css_sets, links each to its source, and append them
2295 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2296 * source css_set is assumed to be its cgroup on the default hierarchy.
2298 * This function must be called after cgroup_migrate_add_src() has been
2299 * called on each migration source css_set. After migration is performed
2300 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2303 static int cgroup_migrate_prepare_dst(struct cgroup
*dst_cgrp
,
2304 struct list_head
*preloaded_csets
)
2307 struct css_set
*src_cset
, *tmp_cset
;
2309 lockdep_assert_held(&cgroup_mutex
);
2312 * Except for the root, child_subsys_mask must be zero for a cgroup
2313 * with tasks so that child cgroups don't compete against tasks.
2315 if (dst_cgrp
&& cgroup_on_dfl(dst_cgrp
) && cgroup_parent(dst_cgrp
) &&
2316 dst_cgrp
->child_subsys_mask
)
2319 /* look up the dst cset for each src cset and link it to src */
2320 list_for_each_entry_safe(src_cset
, tmp_cset
, preloaded_csets
, mg_preload_node
) {
2321 struct css_set
*dst_cset
;
2323 dst_cset
= find_css_set(src_cset
,
2324 dst_cgrp
?: src_cset
->dfl_cgrp
);
2328 WARN_ON_ONCE(src_cset
->mg_dst_cset
|| dst_cset
->mg_dst_cset
);
2331 * If src cset equals dst, it's noop. Drop the src.
2332 * cgroup_migrate() will skip the cset too. Note that we
2333 * can't handle src == dst as some nodes are used by both.
2335 if (src_cset
== dst_cset
) {
2336 src_cset
->mg_src_cgrp
= NULL
;
2337 list_del_init(&src_cset
->mg_preload_node
);
2338 put_css_set(src_cset
);
2339 put_css_set(dst_cset
);
2343 src_cset
->mg_dst_cset
= dst_cset
;
2345 if (list_empty(&dst_cset
->mg_preload_node
))
2346 list_add(&dst_cset
->mg_preload_node
, &csets
);
2348 put_css_set(dst_cset
);
2351 list_splice_tail(&csets
, preloaded_csets
);
2354 cgroup_migrate_finish(&csets
);
2359 * cgroup_migrate - migrate a process or task to a cgroup
2360 * @cgrp: the destination cgroup
2361 * @leader: the leader of the process or the task to migrate
2362 * @threadgroup: whether @leader points to the whole process or a single task
2364 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2365 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2366 * caller is also responsible for invoking cgroup_migrate_add_src() and
2367 * cgroup_migrate_prepare_dst() on the targets before invoking this
2368 * function and following up with cgroup_migrate_finish().
2370 * As long as a controller's ->can_attach() doesn't fail, this function is
2371 * guaranteed to succeed. This means that, excluding ->can_attach()
2372 * failure, when migrating multiple targets, the success or failure can be
2373 * decided for all targets by invoking group_migrate_prepare_dst() before
2374 * actually starting migrating.
2376 static int cgroup_migrate(struct cgroup
*cgrp
, struct task_struct
*leader
,
2379 struct cgroup_taskset tset
= {
2380 .src_csets
= LIST_HEAD_INIT(tset
.src_csets
),
2381 .dst_csets
= LIST_HEAD_INIT(tset
.dst_csets
),
2382 .csets
= &tset
.src_csets
,
2384 struct cgroup_subsys_state
*css
, *failed_css
= NULL
;
2385 struct css_set
*cset
, *tmp_cset
;
2386 struct task_struct
*task
, *tmp_task
;
2390 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2391 * already PF_EXITING could be freed from underneath us unless we
2392 * take an rcu_read_lock.
2394 down_write(&css_set_rwsem
);
2398 /* @task either already exited or can't exit until the end */
2399 if (task
->flags
& PF_EXITING
)
2402 /* leave @task alone if post_fork() hasn't linked it yet */
2403 if (list_empty(&task
->cg_list
))
2406 cset
= task_css_set(task
);
2407 if (!cset
->mg_src_cgrp
)
2411 * cgroup_taskset_first() must always return the leader.
2412 * Take care to avoid disturbing the ordering.
2414 list_move_tail(&task
->cg_list
, &cset
->mg_tasks
);
2415 if (list_empty(&cset
->mg_node
))
2416 list_add_tail(&cset
->mg_node
, &tset
.src_csets
);
2417 if (list_empty(&cset
->mg_dst_cset
->mg_node
))
2418 list_move_tail(&cset
->mg_dst_cset
->mg_node
,
2423 } while_each_thread(leader
, task
);
2425 up_write(&css_set_rwsem
);
2427 /* methods shouldn't be called if no task is actually migrating */
2428 if (list_empty(&tset
.src_csets
))
2431 /* check that we can legitimately attach to the cgroup */
2432 for_each_e_css(css
, i
, cgrp
) {
2433 if (css
->ss
->can_attach
) {
2434 ret
= css
->ss
->can_attach(css
, &tset
);
2437 goto out_cancel_attach
;
2443 * Now that we're guaranteed success, proceed to move all tasks to
2444 * the new cgroup. There are no failure cases after here, so this
2445 * is the commit point.
2447 down_write(&css_set_rwsem
);
2448 list_for_each_entry(cset
, &tset
.src_csets
, mg_node
) {
2449 list_for_each_entry_safe(task
, tmp_task
, &cset
->mg_tasks
, cg_list
)
2450 cgroup_task_migrate(cset
->mg_src_cgrp
, task
,
2453 up_write(&css_set_rwsem
);
2456 * Migration is committed, all target tasks are now on dst_csets.
2457 * Nothing is sensitive to fork() after this point. Notify
2458 * controllers that migration is complete.
2460 tset
.csets
= &tset
.dst_csets
;
2462 for_each_e_css(css
, i
, cgrp
)
2463 if (css
->ss
->attach
)
2464 css
->ss
->attach(css
, &tset
);
2467 goto out_release_tset
;
2470 for_each_e_css(css
, i
, cgrp
) {
2471 if (css
== failed_css
)
2473 if (css
->ss
->cancel_attach
)
2474 css
->ss
->cancel_attach(css
, &tset
);
2477 down_write(&css_set_rwsem
);
2478 list_splice_init(&tset
.dst_csets
, &tset
.src_csets
);
2479 list_for_each_entry_safe(cset
, tmp_cset
, &tset
.src_csets
, mg_node
) {
2480 list_splice_tail_init(&cset
->mg_tasks
, &cset
->tasks
);
2481 list_del_init(&cset
->mg_node
);
2483 up_write(&css_set_rwsem
);
2488 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2489 * @dst_cgrp: the cgroup to attach to
2490 * @leader: the task or the leader of the threadgroup to be attached
2491 * @threadgroup: attach the whole threadgroup?
2493 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2495 static int cgroup_attach_task(struct cgroup
*dst_cgrp
,
2496 struct task_struct
*leader
, bool threadgroup
)
2498 LIST_HEAD(preloaded_csets
);
2499 struct task_struct
*task
;
2502 /* look up all src csets */
2503 down_read(&css_set_rwsem
);
2507 cgroup_migrate_add_src(task_css_set(task
), dst_cgrp
,
2511 } while_each_thread(leader
, task
);
2513 up_read(&css_set_rwsem
);
2515 /* prepare dst csets and commit */
2516 ret
= cgroup_migrate_prepare_dst(dst_cgrp
, &preloaded_csets
);
2518 ret
= cgroup_migrate(dst_cgrp
, leader
, threadgroup
);
2520 cgroup_migrate_finish(&preloaded_csets
);
2524 static int cgroup_procs_write_permission(struct task_struct
*task
,
2525 struct cgroup
*dst_cgrp
,
2526 struct kernfs_open_file
*of
)
2528 const struct cred
*cred
= current_cred();
2529 const struct cred
*tcred
= get_task_cred(task
);
2533 * even if we're attaching all tasks in the thread group, we only
2534 * need to check permissions on one of them.
2536 if (!uid_eq(cred
->euid
, GLOBAL_ROOT_UID
) &&
2537 !uid_eq(cred
->euid
, tcred
->uid
) &&
2538 !uid_eq(cred
->euid
, tcred
->suid
))
2541 if (!ret
&& cgroup_on_dfl(dst_cgrp
)) {
2542 struct super_block
*sb
= of
->file
->f_path
.dentry
->d_sb
;
2543 struct cgroup
*cgrp
;
2544 struct inode
*inode
;
2546 down_read(&css_set_rwsem
);
2547 cgrp
= task_cgroup_from_root(task
, &cgrp_dfl_root
);
2548 up_read(&css_set_rwsem
);
2550 while (!cgroup_is_descendant(dst_cgrp
, cgrp
))
2551 cgrp
= cgroup_parent(cgrp
);
2554 inode
= kernfs_get_inode(sb
, cgrp
->procs_kn
);
2556 ret
= inode_permission(inode
, MAY_WRITE
);
2566 * Find the task_struct of the task to attach by vpid and pass it along to the
2567 * function to attach either it or all tasks in its threadgroup. Will lock
2568 * cgroup_mutex and threadgroup.
2570 static ssize_t
__cgroup_procs_write(struct kernfs_open_file
*of
, char *buf
,
2571 size_t nbytes
, loff_t off
, bool threadgroup
)
2573 struct task_struct
*tsk
;
2574 struct cgroup
*cgrp
;
2578 if (kstrtoint(strstrip(buf
), 0, &pid
) || pid
< 0)
2581 cgrp
= cgroup_kn_lock_live(of
->kn
);
2585 percpu_down_write(&cgroup_threadgroup_rwsem
);
2588 tsk
= find_task_by_vpid(pid
);
2591 goto out_unlock_rcu
;
2598 tsk
= tsk
->group_leader
;
2601 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2602 * trapped in a cpuset, or RT worker may be born in a cgroup
2603 * with no rt_runtime allocated. Just say no.
2605 if (tsk
== kthreadd_task
|| (tsk
->flags
& PF_NO_SETAFFINITY
)) {
2607 goto out_unlock_rcu
;
2610 get_task_struct(tsk
);
2613 ret
= cgroup_procs_write_permission(tsk
, cgrp
, of
);
2615 ret
= cgroup_attach_task(cgrp
, tsk
, threadgroup
);
2617 put_task_struct(tsk
);
2618 goto out_unlock_threadgroup
;
2622 out_unlock_threadgroup
:
2623 percpu_up_write(&cgroup_threadgroup_rwsem
);
2624 cgroup_kn_unlock(of
->kn
);
2625 return ret
?: nbytes
;
2629 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2630 * @from: attach to all cgroups of a given task
2631 * @tsk: the task to be attached
2633 int cgroup_attach_task_all(struct task_struct
*from
, struct task_struct
*tsk
)
2635 struct cgroup_root
*root
;
2638 mutex_lock(&cgroup_mutex
);
2639 for_each_root(root
) {
2640 struct cgroup
*from_cgrp
;
2642 if (root
== &cgrp_dfl_root
)
2645 down_read(&css_set_rwsem
);
2646 from_cgrp
= task_cgroup_from_root(from
, root
);
2647 up_read(&css_set_rwsem
);
2649 retval
= cgroup_attach_task(from_cgrp
, tsk
, false);
2653 mutex_unlock(&cgroup_mutex
);
2657 EXPORT_SYMBOL_GPL(cgroup_attach_task_all
);
2659 static ssize_t
cgroup_tasks_write(struct kernfs_open_file
*of
,
2660 char *buf
, size_t nbytes
, loff_t off
)
2662 return __cgroup_procs_write(of
, buf
, nbytes
, off
, false);
2665 static ssize_t
cgroup_procs_write(struct kernfs_open_file
*of
,
2666 char *buf
, size_t nbytes
, loff_t off
)
2668 return __cgroup_procs_write(of
, buf
, nbytes
, off
, true);
2671 static ssize_t
cgroup_release_agent_write(struct kernfs_open_file
*of
,
2672 char *buf
, size_t nbytes
, loff_t off
)
2674 struct cgroup
*cgrp
;
2676 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
2678 cgrp
= cgroup_kn_lock_live(of
->kn
);
2681 spin_lock(&release_agent_path_lock
);
2682 strlcpy(cgrp
->root
->release_agent_path
, strstrip(buf
),
2683 sizeof(cgrp
->root
->release_agent_path
));
2684 spin_unlock(&release_agent_path_lock
);
2685 cgroup_kn_unlock(of
->kn
);
2689 static int cgroup_release_agent_show(struct seq_file
*seq
, void *v
)
2691 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2693 spin_lock(&release_agent_path_lock
);
2694 seq_puts(seq
, cgrp
->root
->release_agent_path
);
2695 spin_unlock(&release_agent_path_lock
);
2696 seq_putc(seq
, '\n');
2700 static int cgroup_sane_behavior_show(struct seq_file
*seq
, void *v
)
2702 seq_puts(seq
, "0\n");
2706 static void cgroup_print_ss_mask(struct seq_file
*seq
, unsigned long ss_mask
)
2708 struct cgroup_subsys
*ss
;
2709 bool printed
= false;
2712 for_each_subsys_which(ss
, ssid
, &ss_mask
) {
2715 seq_printf(seq
, "%s", ss
->name
);
2719 seq_putc(seq
, '\n');
2722 /* show controllers which are currently attached to the default hierarchy */
2723 static int cgroup_root_controllers_show(struct seq_file
*seq
, void *v
)
2725 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2727 cgroup_print_ss_mask(seq
, cgrp
->root
->subsys_mask
&
2728 ~cgrp_dfl_root_inhibit_ss_mask
);
2732 /* show controllers which are enabled from the parent */
2733 static int cgroup_controllers_show(struct seq_file
*seq
, void *v
)
2735 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2737 cgroup_print_ss_mask(seq
, cgroup_parent(cgrp
)->subtree_control
);
2741 /* show controllers which are enabled for a given cgroup's children */
2742 static int cgroup_subtree_control_show(struct seq_file
*seq
, void *v
)
2744 struct cgroup
*cgrp
= seq_css(seq
)->cgroup
;
2746 cgroup_print_ss_mask(seq
, cgrp
->subtree_control
);
2751 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2752 * @cgrp: root of the subtree to update csses for
2754 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2755 * css associations need to be updated accordingly. This function looks up
2756 * all css_sets which are attached to the subtree, creates the matching
2757 * updated css_sets and migrates the tasks to the new ones.
2759 static int cgroup_update_dfl_csses(struct cgroup
*cgrp
)
2761 LIST_HEAD(preloaded_csets
);
2762 struct cgroup_subsys_state
*css
;
2763 struct css_set
*src_cset
;
2766 lockdep_assert_held(&cgroup_mutex
);
2768 percpu_down_write(&cgroup_threadgroup_rwsem
);
2770 /* look up all csses currently attached to @cgrp's subtree */
2771 down_read(&css_set_rwsem
);
2772 css_for_each_descendant_pre(css
, cgroup_css(cgrp
, NULL
)) {
2773 struct cgrp_cset_link
*link
;
2775 /* self is not affected by child_subsys_mask change */
2776 if (css
->cgroup
== cgrp
)
2779 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
)
2780 cgroup_migrate_add_src(link
->cset
, cgrp
,
2783 up_read(&css_set_rwsem
);
2785 /* NULL dst indicates self on default hierarchy */
2786 ret
= cgroup_migrate_prepare_dst(NULL
, &preloaded_csets
);
2790 list_for_each_entry(src_cset
, &preloaded_csets
, mg_preload_node
) {
2791 struct task_struct
*last_task
= NULL
, *task
;
2793 /* src_csets precede dst_csets, break on the first dst_cset */
2794 if (!src_cset
->mg_src_cgrp
)
2798 * All tasks in src_cset need to be migrated to the
2799 * matching dst_cset. Empty it process by process. We
2800 * walk tasks but migrate processes. The leader might even
2801 * belong to a different cset but such src_cset would also
2802 * be among the target src_csets because the default
2803 * hierarchy enforces per-process membership.
2806 down_read(&css_set_rwsem
);
2807 task
= list_first_entry_or_null(&src_cset
->tasks
,
2808 struct task_struct
, cg_list
);
2810 task
= task
->group_leader
;
2811 WARN_ON_ONCE(!task_css_set(task
)->mg_src_cgrp
);
2812 get_task_struct(task
);
2814 up_read(&css_set_rwsem
);
2819 /* guard against possible infinite loop */
2820 if (WARN(last_task
== task
,
2821 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2825 ret
= cgroup_migrate(src_cset
->dfl_cgrp
, task
, true);
2827 put_task_struct(task
);
2829 if (WARN(ret
, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret
))
2835 cgroup_migrate_finish(&preloaded_csets
);
2836 percpu_up_write(&cgroup_threadgroup_rwsem
);
2840 /* change the enabled child controllers for a cgroup in the default hierarchy */
2841 static ssize_t
cgroup_subtree_control_write(struct kernfs_open_file
*of
,
2842 char *buf
, size_t nbytes
,
2845 unsigned long enable
= 0, disable
= 0;
2846 unsigned long css_enable
, css_disable
, old_sc
, new_sc
, old_ss
, new_ss
;
2847 struct cgroup
*cgrp
, *child
;
2848 struct cgroup_subsys
*ss
;
2853 * Parse input - space separated list of subsystem names prefixed
2854 * with either + or -.
2856 buf
= strstrip(buf
);
2857 while ((tok
= strsep(&buf
, " "))) {
2858 unsigned long tmp_ss_mask
= ~cgrp_dfl_root_inhibit_ss_mask
;
2862 for_each_subsys_which(ss
, ssid
, &tmp_ss_mask
) {
2863 if (!cgroup_ssid_enabled(ssid
) ||
2864 strcmp(tok
+ 1, ss
->name
))
2868 enable
|= 1 << ssid
;
2869 disable
&= ~(1 << ssid
);
2870 } else if (*tok
== '-') {
2871 disable
|= 1 << ssid
;
2872 enable
&= ~(1 << ssid
);
2878 if (ssid
== CGROUP_SUBSYS_COUNT
)
2882 cgrp
= cgroup_kn_lock_live(of
->kn
);
2886 for_each_subsys(ss
, ssid
) {
2887 if (enable
& (1 << ssid
)) {
2888 if (cgrp
->subtree_control
& (1 << ssid
)) {
2889 enable
&= ~(1 << ssid
);
2893 /* unavailable or not enabled on the parent? */
2894 if (!(cgrp_dfl_root
.subsys_mask
& (1 << ssid
)) ||
2895 (cgroup_parent(cgrp
) &&
2896 !(cgroup_parent(cgrp
)->subtree_control
& (1 << ssid
)))) {
2900 } else if (disable
& (1 << ssid
)) {
2901 if (!(cgrp
->subtree_control
& (1 << ssid
))) {
2902 disable
&= ~(1 << ssid
);
2906 /* a child has it enabled? */
2907 cgroup_for_each_live_child(child
, cgrp
) {
2908 if (child
->subtree_control
& (1 << ssid
)) {
2916 if (!enable
&& !disable
) {
2922 * Except for the root, subtree_control must be zero for a cgroup
2923 * with tasks so that child cgroups don't compete against tasks.
2925 if (enable
&& cgroup_parent(cgrp
) && !list_empty(&cgrp
->cset_links
)) {
2931 * Update subsys masks and calculate what needs to be done. More
2932 * subsystems than specified may need to be enabled or disabled
2933 * depending on subsystem dependencies.
2935 old_sc
= cgrp
->subtree_control
;
2936 old_ss
= cgrp
->child_subsys_mask
;
2937 new_sc
= (old_sc
| enable
) & ~disable
;
2938 new_ss
= cgroup_calc_child_subsys_mask(cgrp
, new_sc
);
2940 css_enable
= ~old_ss
& new_ss
;
2941 css_disable
= old_ss
& ~new_ss
;
2942 enable
|= css_enable
;
2943 disable
|= css_disable
;
2946 * Because css offlining is asynchronous, userland might try to
2947 * re-enable the same controller while the previous instance is
2948 * still around. In such cases, wait till it's gone using
2951 for_each_subsys_which(ss
, ssid
, &css_enable
) {
2952 cgroup_for_each_live_child(child
, cgrp
) {
2955 if (!cgroup_css(child
, ss
))
2959 prepare_to_wait(&child
->offline_waitq
, &wait
,
2960 TASK_UNINTERRUPTIBLE
);
2961 cgroup_kn_unlock(of
->kn
);
2963 finish_wait(&child
->offline_waitq
, &wait
);
2966 return restart_syscall();
2970 cgrp
->subtree_control
= new_sc
;
2971 cgrp
->child_subsys_mask
= new_ss
;
2974 * Create new csses or make the existing ones visible. A css is
2975 * created invisible if it's being implicitly enabled through
2976 * dependency. An invisible css is made visible when the userland
2977 * explicitly enables it.
2979 for_each_subsys(ss
, ssid
) {
2980 if (!(enable
& (1 << ssid
)))
2983 cgroup_for_each_live_child(child
, cgrp
) {
2984 if (css_enable
& (1 << ssid
))
2985 ret
= create_css(child
, ss
,
2986 cgrp
->subtree_control
& (1 << ssid
));
2988 ret
= cgroup_populate_dir(child
, 1 << ssid
);
2995 * At this point, cgroup_e_css() results reflect the new csses
2996 * making the following cgroup_update_dfl_csses() properly update
2997 * css associations of all tasks in the subtree.
2999 ret
= cgroup_update_dfl_csses(cgrp
);
3004 * All tasks are migrated out of disabled csses. Kill or hide
3005 * them. A css is hidden when the userland requests it to be
3006 * disabled while other subsystems are still depending on it. The
3007 * css must not actively control resources and be in the vanilla
3008 * state if it's made visible again later. Controllers which may
3009 * be depended upon should provide ->css_reset() for this purpose.
3011 for_each_subsys(ss
, ssid
) {
3012 if (!(disable
& (1 << ssid
)))
3015 cgroup_for_each_live_child(child
, cgrp
) {
3016 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3018 if (css_disable
& (1 << ssid
)) {
3021 cgroup_clear_dir(child
, 1 << ssid
);
3029 * The effective csses of all the descendants (excluding @cgrp) may
3030 * have changed. Subsystems can optionally subscribe to this event
3031 * by implementing ->css_e_css_changed() which is invoked if any of
3032 * the effective csses seen from the css's cgroup may have changed.
3034 for_each_subsys(ss
, ssid
) {
3035 struct cgroup_subsys_state
*this_css
= cgroup_css(cgrp
, ss
);
3036 struct cgroup_subsys_state
*css
;
3038 if (!ss
->css_e_css_changed
|| !this_css
)
3041 css_for_each_descendant_pre(css
, this_css
)
3042 if (css
!= this_css
)
3043 ss
->css_e_css_changed(css
);
3046 kernfs_activate(cgrp
->kn
);
3049 cgroup_kn_unlock(of
->kn
);
3050 return ret
?: nbytes
;
3053 cgrp
->subtree_control
= old_sc
;
3054 cgrp
->child_subsys_mask
= old_ss
;
3056 for_each_subsys(ss
, ssid
) {
3057 if (!(enable
& (1 << ssid
)))
3060 cgroup_for_each_live_child(child
, cgrp
) {
3061 struct cgroup_subsys_state
*css
= cgroup_css(child
, ss
);
3066 if (css_enable
& (1 << ssid
))
3069 cgroup_clear_dir(child
, 1 << ssid
);
3075 static int cgroup_events_show(struct seq_file
*seq
, void *v
)
3077 seq_printf(seq
, "populated %d\n",
3078 (bool)seq_css(seq
)->cgroup
->populated_cnt
);
3082 static ssize_t
cgroup_file_write(struct kernfs_open_file
*of
, char *buf
,
3083 size_t nbytes
, loff_t off
)
3085 struct cgroup
*cgrp
= of
->kn
->parent
->priv
;
3086 struct cftype
*cft
= of
->kn
->priv
;
3087 struct cgroup_subsys_state
*css
;
3091 return cft
->write(of
, buf
, nbytes
, off
);
3094 * kernfs guarantees that a file isn't deleted with operations in
3095 * flight, which means that the matching css is and stays alive and
3096 * doesn't need to be pinned. The RCU locking is not necessary
3097 * either. It's just for the convenience of using cgroup_css().
3100 css
= cgroup_css(cgrp
, cft
->ss
);
3103 if (cft
->write_u64
) {
3104 unsigned long long v
;
3105 ret
= kstrtoull(buf
, 0, &v
);
3107 ret
= cft
->write_u64(css
, cft
, v
);
3108 } else if (cft
->write_s64
) {
3110 ret
= kstrtoll(buf
, 0, &v
);
3112 ret
= cft
->write_s64(css
, cft
, v
);
3117 return ret
?: nbytes
;
3120 static void *cgroup_seqfile_start(struct seq_file
*seq
, loff_t
*ppos
)
3122 return seq_cft(seq
)->seq_start(seq
, ppos
);
3125 static void *cgroup_seqfile_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
3127 return seq_cft(seq
)->seq_next(seq
, v
, ppos
);
3130 static void cgroup_seqfile_stop(struct seq_file
*seq
, void *v
)
3132 seq_cft(seq
)->seq_stop(seq
, v
);
3135 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
3137 struct cftype
*cft
= seq_cft(m
);
3138 struct cgroup_subsys_state
*css
= seq_css(m
);
3141 return cft
->seq_show(m
, arg
);
3144 seq_printf(m
, "%llu\n", cft
->read_u64(css
, cft
));
3145 else if (cft
->read_s64
)
3146 seq_printf(m
, "%lld\n", cft
->read_s64(css
, cft
));
3152 static struct kernfs_ops cgroup_kf_single_ops
= {
3153 .atomic_write_len
= PAGE_SIZE
,
3154 .write
= cgroup_file_write
,
3155 .seq_show
= cgroup_seqfile_show
,
3158 static struct kernfs_ops cgroup_kf_ops
= {
3159 .atomic_write_len
= PAGE_SIZE
,
3160 .write
= cgroup_file_write
,
3161 .seq_start
= cgroup_seqfile_start
,
3162 .seq_next
= cgroup_seqfile_next
,
3163 .seq_stop
= cgroup_seqfile_stop
,
3164 .seq_show
= cgroup_seqfile_show
,
3168 * cgroup_rename - Only allow simple rename of directories in place.
3170 static int cgroup_rename(struct kernfs_node
*kn
, struct kernfs_node
*new_parent
,
3171 const char *new_name_str
)
3173 struct cgroup
*cgrp
= kn
->priv
;
3176 if (kernfs_type(kn
) != KERNFS_DIR
)
3178 if (kn
->parent
!= new_parent
)
3182 * This isn't a proper migration and its usefulness is very
3183 * limited. Disallow on the default hierarchy.
3185 if (cgroup_on_dfl(cgrp
))
3189 * We're gonna grab cgroup_mutex which nests outside kernfs
3190 * active_ref. kernfs_rename() doesn't require active_ref
3191 * protection. Break them before grabbing cgroup_mutex.
3193 kernfs_break_active_protection(new_parent
);
3194 kernfs_break_active_protection(kn
);
3196 mutex_lock(&cgroup_mutex
);
3198 ret
= kernfs_rename(kn
, new_parent
, new_name_str
);
3200 mutex_unlock(&cgroup_mutex
);
3202 kernfs_unbreak_active_protection(kn
);
3203 kernfs_unbreak_active_protection(new_parent
);
3207 /* set uid and gid of cgroup dirs and files to that of the creator */
3208 static int cgroup_kn_set_ugid(struct kernfs_node
*kn
)
3210 struct iattr iattr
= { .ia_valid
= ATTR_UID
| ATTR_GID
,
3211 .ia_uid
= current_fsuid(),
3212 .ia_gid
= current_fsgid(), };
3214 if (uid_eq(iattr
.ia_uid
, GLOBAL_ROOT_UID
) &&
3215 gid_eq(iattr
.ia_gid
, GLOBAL_ROOT_GID
))
3218 return kernfs_setattr(kn
, &iattr
);
3221 static int cgroup_add_file(struct cgroup
*cgrp
, struct cftype
*cft
)
3223 char name
[CGROUP_FILE_NAME_MAX
];
3224 struct kernfs_node
*kn
;
3225 struct lock_class_key
*key
= NULL
;
3228 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3229 key
= &cft
->lockdep_key
;
3231 kn
= __kernfs_create_file(cgrp
->kn
, cgroup_file_name(cgrp
, cft
, name
),
3232 cgroup_file_mode(cft
), 0, cft
->kf_ops
, cft
,
3237 ret
= cgroup_kn_set_ugid(kn
);
3243 if (cft
->write
== cgroup_procs_write
)
3244 cgrp
->procs_kn
= kn
;
3245 else if (cft
->seq_show
== cgroup_events_show
)
3246 cgrp
->events_kn
= kn
;
3251 * cgroup_addrm_files - add or remove files to a cgroup directory
3252 * @cgrp: the target cgroup
3253 * @cfts: array of cftypes to be added
3254 * @is_add: whether to add or remove
3256 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3257 * For removals, this function never fails.
3259 static int cgroup_addrm_files(struct cgroup
*cgrp
, struct cftype cfts
[],
3262 struct cftype
*cft
, *cft_end
= NULL
;
3265 lockdep_assert_held(&cgroup_mutex
);
3268 for (cft
= cfts
; cft
!= cft_end
&& cft
->name
[0] != '\0'; cft
++) {
3269 /* does cft->flags tell us to skip this file on @cgrp? */
3270 if ((cft
->flags
& __CFTYPE_ONLY_ON_DFL
) && !cgroup_on_dfl(cgrp
))
3272 if ((cft
->flags
& __CFTYPE_NOT_ON_DFL
) && cgroup_on_dfl(cgrp
))
3274 if ((cft
->flags
& CFTYPE_NOT_ON_ROOT
) && !cgroup_parent(cgrp
))
3276 if ((cft
->flags
& CFTYPE_ONLY_ON_ROOT
) && cgroup_parent(cgrp
))
3280 ret
= cgroup_add_file(cgrp
, cft
);
3282 pr_warn("%s: failed to add %s, err=%d\n",
3283 __func__
, cft
->name
, ret
);
3289 cgroup_rm_file(cgrp
, cft
);
3295 static int cgroup_apply_cftypes(struct cftype
*cfts
, bool is_add
)
3298 struct cgroup_subsys
*ss
= cfts
[0].ss
;
3299 struct cgroup
*root
= &ss
->root
->cgrp
;
3300 struct cgroup_subsys_state
*css
;
3303 lockdep_assert_held(&cgroup_mutex
);
3305 /* add/rm files for all cgroups created before */
3306 css_for_each_descendant_pre(css
, cgroup_css(root
, ss
)) {
3307 struct cgroup
*cgrp
= css
->cgroup
;
3309 if (cgroup_is_dead(cgrp
))
3312 ret
= cgroup_addrm_files(cgrp
, cfts
, is_add
);
3318 kernfs_activate(root
->kn
);
3322 static void cgroup_exit_cftypes(struct cftype
*cfts
)
3326 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3327 /* free copy for custom atomic_write_len, see init_cftypes() */
3328 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
)
3333 /* revert flags set by cgroup core while adding @cfts */
3334 cft
->flags
&= ~(__CFTYPE_ONLY_ON_DFL
| __CFTYPE_NOT_ON_DFL
);
3338 static int cgroup_init_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3342 for (cft
= cfts
; cft
->name
[0] != '\0'; cft
++) {
3343 struct kernfs_ops
*kf_ops
;
3345 WARN_ON(cft
->ss
|| cft
->kf_ops
);
3348 kf_ops
= &cgroup_kf_ops
;
3350 kf_ops
= &cgroup_kf_single_ops
;
3353 * Ugh... if @cft wants a custom max_write_len, we need to
3354 * make a copy of kf_ops to set its atomic_write_len.
3356 if (cft
->max_write_len
&& cft
->max_write_len
!= PAGE_SIZE
) {
3357 kf_ops
= kmemdup(kf_ops
, sizeof(*kf_ops
), GFP_KERNEL
);
3359 cgroup_exit_cftypes(cfts
);
3362 kf_ops
->atomic_write_len
= cft
->max_write_len
;
3365 cft
->kf_ops
= kf_ops
;
3372 static int cgroup_rm_cftypes_locked(struct cftype
*cfts
)
3374 lockdep_assert_held(&cgroup_mutex
);
3376 if (!cfts
|| !cfts
[0].ss
)
3379 list_del(&cfts
->node
);
3380 cgroup_apply_cftypes(cfts
, false);
3381 cgroup_exit_cftypes(cfts
);
3386 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3387 * @cfts: zero-length name terminated array of cftypes
3389 * Unregister @cfts. Files described by @cfts are removed from all
3390 * existing cgroups and all future cgroups won't have them either. This
3391 * function can be called anytime whether @cfts' subsys is attached or not.
3393 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3396 int cgroup_rm_cftypes(struct cftype
*cfts
)
3400 mutex_lock(&cgroup_mutex
);
3401 ret
= cgroup_rm_cftypes_locked(cfts
);
3402 mutex_unlock(&cgroup_mutex
);
3407 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3408 * @ss: target cgroup subsystem
3409 * @cfts: zero-length name terminated array of cftypes
3411 * Register @cfts to @ss. Files described by @cfts are created for all
3412 * existing cgroups to which @ss is attached and all future cgroups will
3413 * have them too. This function can be called anytime whether @ss is
3416 * Returns 0 on successful registration, -errno on failure. Note that this
3417 * function currently returns 0 as long as @cfts registration is successful
3418 * even if some file creation attempts on existing cgroups fail.
3420 static int cgroup_add_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3424 if (!cgroup_ssid_enabled(ss
->id
))
3427 if (!cfts
|| cfts
[0].name
[0] == '\0')
3430 ret
= cgroup_init_cftypes(ss
, cfts
);
3434 mutex_lock(&cgroup_mutex
);
3436 list_add_tail(&cfts
->node
, &ss
->cfts
);
3437 ret
= cgroup_apply_cftypes(cfts
, true);
3439 cgroup_rm_cftypes_locked(cfts
);
3441 mutex_unlock(&cgroup_mutex
);
3446 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3447 * @ss: target cgroup subsystem
3448 * @cfts: zero-length name terminated array of cftypes
3450 * Similar to cgroup_add_cftypes() but the added files are only used for
3451 * the default hierarchy.
3453 int cgroup_add_dfl_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3457 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3458 cft
->flags
|= __CFTYPE_ONLY_ON_DFL
;
3459 return cgroup_add_cftypes(ss
, cfts
);
3463 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3464 * @ss: target cgroup subsystem
3465 * @cfts: zero-length name terminated array of cftypes
3467 * Similar to cgroup_add_cftypes() but the added files are only used for
3468 * the legacy hierarchies.
3470 int cgroup_add_legacy_cftypes(struct cgroup_subsys
*ss
, struct cftype
*cfts
)
3475 * If legacy_flies_on_dfl, we want to show the legacy files on the
3476 * dfl hierarchy but iff the target subsystem hasn't been updated
3477 * for the dfl hierarchy yet.
3479 if (!cgroup_legacy_files_on_dfl
||
3480 ss
->dfl_cftypes
!= ss
->legacy_cftypes
) {
3481 for (cft
= cfts
; cft
&& cft
->name
[0] != '\0'; cft
++)
3482 cft
->flags
|= __CFTYPE_NOT_ON_DFL
;
3485 return cgroup_add_cftypes(ss
, cfts
);
3489 * cgroup_task_count - count the number of tasks in a cgroup.
3490 * @cgrp: the cgroup in question
3492 * Return the number of tasks in the cgroup.
3494 static int cgroup_task_count(const struct cgroup
*cgrp
)
3497 struct cgrp_cset_link
*link
;
3499 down_read(&css_set_rwsem
);
3500 list_for_each_entry(link
, &cgrp
->cset_links
, cset_link
)
3501 count
+= atomic_read(&link
->cset
->refcount
);
3502 up_read(&css_set_rwsem
);
3507 * css_next_child - find the next child of a given css
3508 * @pos: the current position (%NULL to initiate traversal)
3509 * @parent: css whose children to walk
3511 * This function returns the next child of @parent and should be called
3512 * under either cgroup_mutex or RCU read lock. The only requirement is
3513 * that @parent and @pos are accessible. The next sibling is guaranteed to
3514 * be returned regardless of their states.
3516 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3517 * css which finished ->css_online() is guaranteed to be visible in the
3518 * future iterations and will stay visible until the last reference is put.
3519 * A css which hasn't finished ->css_online() or already finished
3520 * ->css_offline() may show up during traversal. It's each subsystem's
3521 * responsibility to synchronize against on/offlining.
3523 struct cgroup_subsys_state
*css_next_child(struct cgroup_subsys_state
*pos
,
3524 struct cgroup_subsys_state
*parent
)
3526 struct cgroup_subsys_state
*next
;
3528 cgroup_assert_mutex_or_rcu_locked();
3531 * @pos could already have been unlinked from the sibling list.
3532 * Once a cgroup is removed, its ->sibling.next is no longer
3533 * updated when its next sibling changes. CSS_RELEASED is set when
3534 * @pos is taken off list, at which time its next pointer is valid,
3535 * and, as releases are serialized, the one pointed to by the next
3536 * pointer is guaranteed to not have started release yet. This
3537 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3538 * critical section, the one pointed to by its next pointer is
3539 * guaranteed to not have finished its RCU grace period even if we
3540 * have dropped rcu_read_lock() inbetween iterations.
3542 * If @pos has CSS_RELEASED set, its next pointer can't be
3543 * dereferenced; however, as each css is given a monotonically
3544 * increasing unique serial number and always appended to the
3545 * sibling list, the next one can be found by walking the parent's
3546 * children until the first css with higher serial number than
3547 * @pos's. While this path can be slower, it happens iff iteration
3548 * races against release and the race window is very small.
3551 next
= list_entry_rcu(parent
->children
.next
, struct cgroup_subsys_state
, sibling
);
3552 } else if (likely(!(pos
->flags
& CSS_RELEASED
))) {
3553 next
= list_entry_rcu(pos
->sibling
.next
, struct cgroup_subsys_state
, sibling
);
3555 list_for_each_entry_rcu(next
, &parent
->children
, sibling
)
3556 if (next
->serial_nr
> pos
->serial_nr
)
3561 * @next, if not pointing to the head, can be dereferenced and is
3564 if (&next
->sibling
!= &parent
->children
)
3570 * css_next_descendant_pre - find the next descendant for pre-order walk
3571 * @pos: the current position (%NULL to initiate traversal)
3572 * @root: css whose descendants to walk
3574 * To be used by css_for_each_descendant_pre(). Find the next descendant
3575 * to visit for pre-order traversal of @root's descendants. @root is
3576 * included in the iteration and the first node to be visited.
3578 * While this function requires cgroup_mutex or RCU read locking, it
3579 * doesn't require the whole traversal to be contained in a single critical
3580 * section. This function will return the correct next descendant as long
3581 * as both @pos and @root are accessible and @pos is a descendant of @root.
3583 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3584 * css which finished ->css_online() is guaranteed to be visible in the
3585 * future iterations and will stay visible until the last reference is put.
3586 * A css which hasn't finished ->css_online() or already finished
3587 * ->css_offline() may show up during traversal. It's each subsystem's
3588 * responsibility to synchronize against on/offlining.
3590 struct cgroup_subsys_state
*
3591 css_next_descendant_pre(struct cgroup_subsys_state
*pos
,
3592 struct cgroup_subsys_state
*root
)
3594 struct cgroup_subsys_state
*next
;
3596 cgroup_assert_mutex_or_rcu_locked();
3598 /* if first iteration, visit @root */
3602 /* visit the first child if exists */
3603 next
= css_next_child(NULL
, pos
);
3607 /* no child, visit my or the closest ancestor's next sibling */
3608 while (pos
!= root
) {
3609 next
= css_next_child(pos
, pos
->parent
);
3619 * css_rightmost_descendant - return the rightmost descendant of a css
3620 * @pos: css of interest
3622 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3623 * is returned. This can be used during pre-order traversal to skip
3626 * While this function requires cgroup_mutex or RCU read locking, it
3627 * doesn't require the whole traversal to be contained in a single critical
3628 * section. This function will return the correct rightmost descendant as
3629 * long as @pos is accessible.
3631 struct cgroup_subsys_state
*
3632 css_rightmost_descendant(struct cgroup_subsys_state
*pos
)
3634 struct cgroup_subsys_state
*last
, *tmp
;
3636 cgroup_assert_mutex_or_rcu_locked();
3640 /* ->prev isn't RCU safe, walk ->next till the end */
3642 css_for_each_child(tmp
, last
)
3649 static struct cgroup_subsys_state
*
3650 css_leftmost_descendant(struct cgroup_subsys_state
*pos
)
3652 struct cgroup_subsys_state
*last
;
3656 pos
= css_next_child(NULL
, pos
);
3663 * css_next_descendant_post - find the next descendant for post-order walk
3664 * @pos: the current position (%NULL to initiate traversal)
3665 * @root: css whose descendants to walk
3667 * To be used by css_for_each_descendant_post(). Find the next descendant
3668 * to visit for post-order traversal of @root's descendants. @root is
3669 * included in the iteration and the last node to be visited.
3671 * While this function requires cgroup_mutex or RCU read locking, it
3672 * doesn't require the whole traversal to be contained in a single critical
3673 * section. This function will return the correct next descendant as long
3674 * as both @pos and @cgroup are accessible and @pos is a descendant of
3677 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3678 * css which finished ->css_online() is guaranteed to be visible in the
3679 * future iterations and will stay visible until the last reference is put.
3680 * A css which hasn't finished ->css_online() or already finished
3681 * ->css_offline() may show up during traversal. It's each subsystem's
3682 * responsibility to synchronize against on/offlining.
3684 struct cgroup_subsys_state
*
3685 css_next_descendant_post(struct cgroup_subsys_state
*pos
,
3686 struct cgroup_subsys_state
*root
)
3688 struct cgroup_subsys_state
*next
;
3690 cgroup_assert_mutex_or_rcu_locked();
3692 /* if first iteration, visit leftmost descendant which may be @root */
3694 return css_leftmost_descendant(root
);
3696 /* if we visited @root, we're done */
3700 /* if there's an unvisited sibling, visit its leftmost descendant */
3701 next
= css_next_child(pos
, pos
->parent
);
3703 return css_leftmost_descendant(next
);
3705 /* no sibling left, visit parent */
3710 * css_has_online_children - does a css have online children
3711 * @css: the target css
3713 * Returns %true if @css has any online children; otherwise, %false. This
3714 * function can be called from any context but the caller is responsible
3715 * for synchronizing against on/offlining as necessary.
3717 bool css_has_online_children(struct cgroup_subsys_state
*css
)
3719 struct cgroup_subsys_state
*child
;
3723 css_for_each_child(child
, css
) {
3724 if (child
->flags
& CSS_ONLINE
) {
3734 * css_advance_task_iter - advance a task itererator to the next css_set
3735 * @it: the iterator to advance
3737 * Advance @it to the next css_set to walk.
3739 static void css_advance_task_iter(struct css_task_iter
*it
)
3741 struct list_head
*l
= it
->cset_pos
;
3742 struct cgrp_cset_link
*link
;
3743 struct css_set
*cset
;
3745 /* Advance to the next non-empty css_set */
3748 if (l
== it
->cset_head
) {
3749 it
->cset_pos
= NULL
;
3754 cset
= container_of(l
, struct css_set
,
3755 e_cset_node
[it
->ss
->id
]);
3757 link
= list_entry(l
, struct cgrp_cset_link
, cset_link
);
3760 } while (list_empty(&cset
->tasks
) && list_empty(&cset
->mg_tasks
));
3764 if (!list_empty(&cset
->tasks
))
3765 it
->task_pos
= cset
->tasks
.next
;
3767 it
->task_pos
= cset
->mg_tasks
.next
;
3769 it
->tasks_head
= &cset
->tasks
;
3770 it
->mg_tasks_head
= &cset
->mg_tasks
;
3774 * css_task_iter_start - initiate task iteration
3775 * @css: the css to walk tasks of
3776 * @it: the task iterator to use
3778 * Initiate iteration through the tasks of @css. The caller can call
3779 * css_task_iter_next() to walk through the tasks until the function
3780 * returns NULL. On completion of iteration, css_task_iter_end() must be
3783 * Note that this function acquires a lock which is released when the
3784 * iteration finishes. The caller can't sleep while iteration is in
3787 void css_task_iter_start(struct cgroup_subsys_state
*css
,
3788 struct css_task_iter
*it
)
3789 __acquires(css_set_rwsem
)
3791 /* no one should try to iterate before mounting cgroups */
3792 WARN_ON_ONCE(!use_task_css_set_links
);
3794 down_read(&css_set_rwsem
);
3799 it
->cset_pos
= &css
->cgroup
->e_csets
[css
->ss
->id
];
3801 it
->cset_pos
= &css
->cgroup
->cset_links
;
3803 it
->cset_head
= it
->cset_pos
;
3805 css_advance_task_iter(it
);
3809 * css_task_iter_next - return the next task for the iterator
3810 * @it: the task iterator being iterated
3812 * The "next" function for task iteration. @it should have been
3813 * initialized via css_task_iter_start(). Returns NULL when the iteration
3816 struct task_struct
*css_task_iter_next(struct css_task_iter
*it
)
3818 struct task_struct
*res
;
3819 struct list_head
*l
= it
->task_pos
;
3821 /* If the iterator cg is NULL, we have no tasks */
3824 res
= list_entry(l
, struct task_struct
, cg_list
);
3827 * Advance iterator to find next entry. cset->tasks is consumed
3828 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3833 if (l
== it
->tasks_head
)
3834 l
= it
->mg_tasks_head
->next
;
3836 if (l
== it
->mg_tasks_head
)
3837 css_advance_task_iter(it
);
3845 * css_task_iter_end - finish task iteration
3846 * @it: the task iterator to finish
3848 * Finish task iteration started by css_task_iter_start().
3850 void css_task_iter_end(struct css_task_iter
*it
)
3851 __releases(css_set_rwsem
)
3853 up_read(&css_set_rwsem
);
3857 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3858 * @to: cgroup to which the tasks will be moved
3859 * @from: cgroup in which the tasks currently reside
3861 * Locking rules between cgroup_post_fork() and the migration path
3862 * guarantee that, if a task is forking while being migrated, the new child
3863 * is guaranteed to be either visible in the source cgroup after the
3864 * parent's migration is complete or put into the target cgroup. No task
3865 * can slip out of migration through forking.
3867 int cgroup_transfer_tasks(struct cgroup
*to
, struct cgroup
*from
)
3869 LIST_HEAD(preloaded_csets
);
3870 struct cgrp_cset_link
*link
;
3871 struct css_task_iter it
;
3872 struct task_struct
*task
;
3875 mutex_lock(&cgroup_mutex
);
3877 /* all tasks in @from are being moved, all csets are source */
3878 down_read(&css_set_rwsem
);
3879 list_for_each_entry(link
, &from
->cset_links
, cset_link
)
3880 cgroup_migrate_add_src(link
->cset
, to
, &preloaded_csets
);
3881 up_read(&css_set_rwsem
);
3883 ret
= cgroup_migrate_prepare_dst(to
, &preloaded_csets
);
3888 * Migrate tasks one-by-one until @form is empty. This fails iff
3889 * ->can_attach() fails.
3892 css_task_iter_start(&from
->self
, &it
);
3893 task
= css_task_iter_next(&it
);
3895 get_task_struct(task
);
3896 css_task_iter_end(&it
);
3899 ret
= cgroup_migrate(to
, task
, false);
3900 put_task_struct(task
);
3902 } while (task
&& !ret
);
3904 cgroup_migrate_finish(&preloaded_csets
);
3905 mutex_unlock(&cgroup_mutex
);
3910 * Stuff for reading the 'tasks'/'procs' files.
3912 * Reading this file can return large amounts of data if a cgroup has
3913 * *lots* of attached tasks. So it may need several calls to read(),
3914 * but we cannot guarantee that the information we produce is correct
3915 * unless we produce it entirely atomically.
3919 /* which pidlist file are we talking about? */
3920 enum cgroup_filetype
{
3926 * A pidlist is a list of pids that virtually represents the contents of one
3927 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3928 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3931 struct cgroup_pidlist
{
3933 * used to find which pidlist is wanted. doesn't change as long as
3934 * this particular list stays in the list.
3936 struct { enum cgroup_filetype type
; struct pid_namespace
*ns
; } key
;
3939 /* how many elements the above list has */
3941 /* each of these stored in a list by its cgroup */
3942 struct list_head links
;
3943 /* pointer to the cgroup we belong to, for list removal purposes */
3944 struct cgroup
*owner
;
3945 /* for delayed destruction */
3946 struct delayed_work destroy_dwork
;
3950 * The following two functions "fix" the issue where there are more pids
3951 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3952 * TODO: replace with a kernel-wide solution to this problem
3954 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3955 static void *pidlist_allocate(int count
)
3957 if (PIDLIST_TOO_LARGE(count
))
3958 return vmalloc(count
* sizeof(pid_t
));
3960 return kmalloc(count
* sizeof(pid_t
), GFP_KERNEL
);
3963 static void pidlist_free(void *p
)
3969 * Used to destroy all pidlists lingering waiting for destroy timer. None
3970 * should be left afterwards.
3972 static void cgroup_pidlist_destroy_all(struct cgroup
*cgrp
)
3974 struct cgroup_pidlist
*l
, *tmp_l
;
3976 mutex_lock(&cgrp
->pidlist_mutex
);
3977 list_for_each_entry_safe(l
, tmp_l
, &cgrp
->pidlists
, links
)
3978 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
, 0);
3979 mutex_unlock(&cgrp
->pidlist_mutex
);
3981 flush_workqueue(cgroup_pidlist_destroy_wq
);
3982 BUG_ON(!list_empty(&cgrp
->pidlists
));
3985 static void cgroup_pidlist_destroy_work_fn(struct work_struct
*work
)
3987 struct delayed_work
*dwork
= to_delayed_work(work
);
3988 struct cgroup_pidlist
*l
= container_of(dwork
, struct cgroup_pidlist
,
3990 struct cgroup_pidlist
*tofree
= NULL
;
3992 mutex_lock(&l
->owner
->pidlist_mutex
);
3995 * Destroy iff we didn't get queued again. The state won't change
3996 * as destroy_dwork can only be queued while locked.
3998 if (!delayed_work_pending(dwork
)) {
3999 list_del(&l
->links
);
4000 pidlist_free(l
->list
);
4001 put_pid_ns(l
->key
.ns
);
4005 mutex_unlock(&l
->owner
->pidlist_mutex
);
4010 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4011 * Returns the number of unique elements.
4013 static int pidlist_uniq(pid_t
*list
, int length
)
4018 * we presume the 0th element is unique, so i starts at 1. trivial
4019 * edge cases first; no work needs to be done for either
4021 if (length
== 0 || length
== 1)
4023 /* src and dest walk down the list; dest counts unique elements */
4024 for (src
= 1; src
< length
; src
++) {
4025 /* find next unique element */
4026 while (list
[src
] == list
[src
-1]) {
4031 /* dest always points to where the next unique element goes */
4032 list
[dest
] = list
[src
];
4040 * The two pid files - task and cgroup.procs - guaranteed that the result
4041 * is sorted, which forced this whole pidlist fiasco. As pid order is
4042 * different per namespace, each namespace needs differently sorted list,
4043 * making it impossible to use, for example, single rbtree of member tasks
4044 * sorted by task pointer. As pidlists can be fairly large, allocating one
4045 * per open file is dangerous, so cgroup had to implement shared pool of
4046 * pidlists keyed by cgroup and namespace.
4048 * All this extra complexity was caused by the original implementation
4049 * committing to an entirely unnecessary property. In the long term, we
4050 * want to do away with it. Explicitly scramble sort order if on the
4051 * default hierarchy so that no such expectation exists in the new
4054 * Scrambling is done by swapping every two consecutive bits, which is
4055 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4057 static pid_t
pid_fry(pid_t pid
)
4059 unsigned a
= pid
& 0x55555555;
4060 unsigned b
= pid
& 0xAAAAAAAA;
4062 return (a
<< 1) | (b
>> 1);
4065 static pid_t
cgroup_pid_fry(struct cgroup
*cgrp
, pid_t pid
)
4067 if (cgroup_on_dfl(cgrp
))
4068 return pid_fry(pid
);
4073 static int cmppid(const void *a
, const void *b
)
4075 return *(pid_t
*)a
- *(pid_t
*)b
;
4078 static int fried_cmppid(const void *a
, const void *b
)
4080 return pid_fry(*(pid_t
*)a
) - pid_fry(*(pid_t
*)b
);
4083 static struct cgroup_pidlist
*cgroup_pidlist_find(struct cgroup
*cgrp
,
4084 enum cgroup_filetype type
)
4086 struct cgroup_pidlist
*l
;
4087 /* don't need task_nsproxy() if we're looking at ourself */
4088 struct pid_namespace
*ns
= task_active_pid_ns(current
);
4090 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4092 list_for_each_entry(l
, &cgrp
->pidlists
, links
)
4093 if (l
->key
.type
== type
&& l
->key
.ns
== ns
)
4099 * find the appropriate pidlist for our purpose (given procs vs tasks)
4100 * returns with the lock on that pidlist already held, and takes care
4101 * of the use count, or returns NULL with no locks held if we're out of
4104 static struct cgroup_pidlist
*cgroup_pidlist_find_create(struct cgroup
*cgrp
,
4105 enum cgroup_filetype type
)
4107 struct cgroup_pidlist
*l
;
4109 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4111 l
= cgroup_pidlist_find(cgrp
, type
);
4115 /* entry not found; create a new one */
4116 l
= kzalloc(sizeof(struct cgroup_pidlist
), GFP_KERNEL
);
4120 INIT_DELAYED_WORK(&l
->destroy_dwork
, cgroup_pidlist_destroy_work_fn
);
4122 /* don't need task_nsproxy() if we're looking at ourself */
4123 l
->key
.ns
= get_pid_ns(task_active_pid_ns(current
));
4125 list_add(&l
->links
, &cgrp
->pidlists
);
4130 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4132 static int pidlist_array_load(struct cgroup
*cgrp
, enum cgroup_filetype type
,
4133 struct cgroup_pidlist
**lp
)
4137 int pid
, n
= 0; /* used for populating the array */
4138 struct css_task_iter it
;
4139 struct task_struct
*tsk
;
4140 struct cgroup_pidlist
*l
;
4142 lockdep_assert_held(&cgrp
->pidlist_mutex
);
4145 * If cgroup gets more users after we read count, we won't have
4146 * enough space - tough. This race is indistinguishable to the
4147 * caller from the case that the additional cgroup users didn't
4148 * show up until sometime later on.
4150 length
= cgroup_task_count(cgrp
);
4151 array
= pidlist_allocate(length
);
4154 /* now, populate the array */
4155 css_task_iter_start(&cgrp
->self
, &it
);
4156 while ((tsk
= css_task_iter_next(&it
))) {
4157 if (unlikely(n
== length
))
4159 /* get tgid or pid for procs or tasks file respectively */
4160 if (type
== CGROUP_FILE_PROCS
)
4161 pid
= task_tgid_vnr(tsk
);
4163 pid
= task_pid_vnr(tsk
);
4164 if (pid
> 0) /* make sure to only use valid results */
4167 css_task_iter_end(&it
);
4169 /* now sort & (if procs) strip out duplicates */
4170 if (cgroup_on_dfl(cgrp
))
4171 sort(array
, length
, sizeof(pid_t
), fried_cmppid
, NULL
);
4173 sort(array
, length
, sizeof(pid_t
), cmppid
, NULL
);
4174 if (type
== CGROUP_FILE_PROCS
)
4175 length
= pidlist_uniq(array
, length
);
4177 l
= cgroup_pidlist_find_create(cgrp
, type
);
4179 pidlist_free(array
);
4183 /* store array, freeing old if necessary */
4184 pidlist_free(l
->list
);
4192 * cgroupstats_build - build and fill cgroupstats
4193 * @stats: cgroupstats to fill information into
4194 * @dentry: A dentry entry belonging to the cgroup for which stats have
4197 * Build and fill cgroupstats so that taskstats can export it to user
4200 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
4202 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
4203 struct cgroup
*cgrp
;
4204 struct css_task_iter it
;
4205 struct task_struct
*tsk
;
4207 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4208 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
4209 kernfs_type(kn
) != KERNFS_DIR
)
4212 mutex_lock(&cgroup_mutex
);
4215 * We aren't being called from kernfs and there's no guarantee on
4216 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4217 * @kn->priv is RCU safe. Let's do the RCU dancing.
4220 cgrp
= rcu_dereference(kn
->priv
);
4221 if (!cgrp
|| cgroup_is_dead(cgrp
)) {
4223 mutex_unlock(&cgroup_mutex
);
4228 css_task_iter_start(&cgrp
->self
, &it
);
4229 while ((tsk
= css_task_iter_next(&it
))) {
4230 switch (tsk
->state
) {
4232 stats
->nr_running
++;
4234 case TASK_INTERRUPTIBLE
:
4235 stats
->nr_sleeping
++;
4237 case TASK_UNINTERRUPTIBLE
:
4238 stats
->nr_uninterruptible
++;
4241 stats
->nr_stopped
++;
4244 if (delayacct_is_task_waiting_on_io(tsk
))
4245 stats
->nr_io_wait
++;
4249 css_task_iter_end(&it
);
4251 mutex_unlock(&cgroup_mutex
);
4257 * seq_file methods for the tasks/procs files. The seq_file position is the
4258 * next pid to display; the seq_file iterator is a pointer to the pid
4259 * in the cgroup->l->list array.
4262 static void *cgroup_pidlist_start(struct seq_file
*s
, loff_t
*pos
)
4265 * Initially we receive a position value that corresponds to
4266 * one more than the last pid shown (or 0 on the first call or
4267 * after a seek to the start). Use a binary-search to find the
4268 * next pid to display, if any
4270 struct kernfs_open_file
*of
= s
->private;
4271 struct cgroup
*cgrp
= seq_css(s
)->cgroup
;
4272 struct cgroup_pidlist
*l
;
4273 enum cgroup_filetype type
= seq_cft(s
)->private;
4274 int index
= 0, pid
= *pos
;
4277 mutex_lock(&cgrp
->pidlist_mutex
);
4280 * !NULL @of->priv indicates that this isn't the first start()
4281 * after open. If the matching pidlist is around, we can use that.
4282 * Look for it. Note that @of->priv can't be used directly. It
4283 * could already have been destroyed.
4286 of
->priv
= cgroup_pidlist_find(cgrp
, type
);
4289 * Either this is the first start() after open or the matching
4290 * pidlist has been destroyed inbetween. Create a new one.
4293 ret
= pidlist_array_load(cgrp
, type
,
4294 (struct cgroup_pidlist
**)&of
->priv
);
4296 return ERR_PTR(ret
);
4301 int end
= l
->length
;
4303 while (index
< end
) {
4304 int mid
= (index
+ end
) / 2;
4305 if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) == pid
) {
4308 } else if (cgroup_pid_fry(cgrp
, l
->list
[mid
]) <= pid
)
4314 /* If we're off the end of the array, we're done */
4315 if (index
>= l
->length
)
4317 /* Update the abstract position to be the actual pid that we found */
4318 iter
= l
->list
+ index
;
4319 *pos
= cgroup_pid_fry(cgrp
, *iter
);
4323 static void cgroup_pidlist_stop(struct seq_file
*s
, void *v
)
4325 struct kernfs_open_file
*of
= s
->private;
4326 struct cgroup_pidlist
*l
= of
->priv
;
4329 mod_delayed_work(cgroup_pidlist_destroy_wq
, &l
->destroy_dwork
,
4330 CGROUP_PIDLIST_DESTROY_DELAY
);
4331 mutex_unlock(&seq_css(s
)->cgroup
->pidlist_mutex
);
4334 static void *cgroup_pidlist_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
4336 struct kernfs_open_file
*of
= s
->private;
4337 struct cgroup_pidlist
*l
= of
->priv
;
4339 pid_t
*end
= l
->list
+ l
->length
;
4341 * Advance to the next pid in the array. If this goes off the
4348 *pos
= cgroup_pid_fry(seq_css(s
)->cgroup
, *p
);
4353 static int cgroup_pidlist_show(struct seq_file
*s
, void *v
)
4355 seq_printf(s
, "%d\n", *(int *)v
);
4360 static u64
cgroup_read_notify_on_release(struct cgroup_subsys_state
*css
,
4363 return notify_on_release(css
->cgroup
);
4366 static int cgroup_write_notify_on_release(struct cgroup_subsys_state
*css
,
4367 struct cftype
*cft
, u64 val
)
4370 set_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4372 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &css
->cgroup
->flags
);
4376 static u64
cgroup_clone_children_read(struct cgroup_subsys_state
*css
,
4379 return test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4382 static int cgroup_clone_children_write(struct cgroup_subsys_state
*css
,
4383 struct cftype
*cft
, u64 val
)
4386 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4388 clear_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
);
4392 /* cgroup core interface files for the default hierarchy */
4393 static struct cftype cgroup_dfl_base_files
[] = {
4395 .name
= "cgroup.procs",
4396 .seq_start
= cgroup_pidlist_start
,
4397 .seq_next
= cgroup_pidlist_next
,
4398 .seq_stop
= cgroup_pidlist_stop
,
4399 .seq_show
= cgroup_pidlist_show
,
4400 .private = CGROUP_FILE_PROCS
,
4401 .write
= cgroup_procs_write
,
4404 .name
= "cgroup.controllers",
4405 .flags
= CFTYPE_ONLY_ON_ROOT
,
4406 .seq_show
= cgroup_root_controllers_show
,
4409 .name
= "cgroup.controllers",
4410 .flags
= CFTYPE_NOT_ON_ROOT
,
4411 .seq_show
= cgroup_controllers_show
,
4414 .name
= "cgroup.subtree_control",
4415 .seq_show
= cgroup_subtree_control_show
,
4416 .write
= cgroup_subtree_control_write
,
4419 .name
= "cgroup.events",
4420 .flags
= CFTYPE_NOT_ON_ROOT
,
4421 .seq_show
= cgroup_events_show
,
4426 /* cgroup core interface files for the legacy hierarchies */
4427 static struct cftype cgroup_legacy_base_files
[] = {
4429 .name
= "cgroup.procs",
4430 .seq_start
= cgroup_pidlist_start
,
4431 .seq_next
= cgroup_pidlist_next
,
4432 .seq_stop
= cgroup_pidlist_stop
,
4433 .seq_show
= cgroup_pidlist_show
,
4434 .private = CGROUP_FILE_PROCS
,
4435 .write
= cgroup_procs_write
,
4438 .name
= "cgroup.clone_children",
4439 .read_u64
= cgroup_clone_children_read
,
4440 .write_u64
= cgroup_clone_children_write
,
4443 .name
= "cgroup.sane_behavior",
4444 .flags
= CFTYPE_ONLY_ON_ROOT
,
4445 .seq_show
= cgroup_sane_behavior_show
,
4449 .seq_start
= cgroup_pidlist_start
,
4450 .seq_next
= cgroup_pidlist_next
,
4451 .seq_stop
= cgroup_pidlist_stop
,
4452 .seq_show
= cgroup_pidlist_show
,
4453 .private = CGROUP_FILE_TASKS
,
4454 .write
= cgroup_tasks_write
,
4457 .name
= "notify_on_release",
4458 .read_u64
= cgroup_read_notify_on_release
,
4459 .write_u64
= cgroup_write_notify_on_release
,
4462 .name
= "release_agent",
4463 .flags
= CFTYPE_ONLY_ON_ROOT
,
4464 .seq_show
= cgroup_release_agent_show
,
4465 .write
= cgroup_release_agent_write
,
4466 .max_write_len
= PATH_MAX
- 1,
4472 * css destruction is four-stage process.
4474 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4475 * Implemented in kill_css().
4477 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4478 * and thus css_tryget_online() is guaranteed to fail, the css can be
4479 * offlined by invoking offline_css(). After offlining, the base ref is
4480 * put. Implemented in css_killed_work_fn().
4482 * 3. When the percpu_ref reaches zero, the only possible remaining
4483 * accessors are inside RCU read sections. css_release() schedules the
4486 * 4. After the grace period, the css can be freed. Implemented in
4487 * css_free_work_fn().
4489 * It is actually hairier because both step 2 and 4 require process context
4490 * and thus involve punting to css->destroy_work adding two additional
4491 * steps to the already complex sequence.
4493 static void css_free_work_fn(struct work_struct
*work
)
4495 struct cgroup_subsys_state
*css
=
4496 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4497 struct cgroup_subsys
*ss
= css
->ss
;
4498 struct cgroup
*cgrp
= css
->cgroup
;
4500 percpu_ref_exit(&css
->refcnt
);
4507 css_put(css
->parent
);
4510 cgroup_idr_remove(&ss
->css_idr
, id
);
4513 /* cgroup free path */
4514 atomic_dec(&cgrp
->root
->nr_cgrps
);
4515 cgroup_pidlist_destroy_all(cgrp
);
4516 cancel_work_sync(&cgrp
->release_agent_work
);
4518 if (cgroup_parent(cgrp
)) {
4520 * We get a ref to the parent, and put the ref when
4521 * this cgroup is being freed, so it's guaranteed
4522 * that the parent won't be destroyed before its
4525 cgroup_put(cgroup_parent(cgrp
));
4526 kernfs_put(cgrp
->kn
);
4530 * This is root cgroup's refcnt reaching zero,
4531 * which indicates that the root should be
4534 cgroup_destroy_root(cgrp
->root
);
4539 static void css_free_rcu_fn(struct rcu_head
*rcu_head
)
4541 struct cgroup_subsys_state
*css
=
4542 container_of(rcu_head
, struct cgroup_subsys_state
, rcu_head
);
4544 INIT_WORK(&css
->destroy_work
, css_free_work_fn
);
4545 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4548 static void css_release_work_fn(struct work_struct
*work
)
4550 struct cgroup_subsys_state
*css
=
4551 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4552 struct cgroup_subsys
*ss
= css
->ss
;
4553 struct cgroup
*cgrp
= css
->cgroup
;
4555 mutex_lock(&cgroup_mutex
);
4557 css
->flags
|= CSS_RELEASED
;
4558 list_del_rcu(&css
->sibling
);
4561 /* css release path */
4562 cgroup_idr_replace(&ss
->css_idr
, NULL
, css
->id
);
4563 if (ss
->css_released
)
4564 ss
->css_released(css
);
4566 /* cgroup release path */
4567 cgroup_idr_remove(&cgrp
->root
->cgroup_idr
, cgrp
->id
);
4571 * There are two control paths which try to determine
4572 * cgroup from dentry without going through kernfs -
4573 * cgroupstats_build() and css_tryget_online_from_dir().
4574 * Those are supported by RCU protecting clearing of
4575 * cgrp->kn->priv backpointer.
4577 RCU_INIT_POINTER(*(void __rcu __force
**)&cgrp
->kn
->priv
, NULL
);
4580 mutex_unlock(&cgroup_mutex
);
4582 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4585 static void css_release(struct percpu_ref
*ref
)
4587 struct cgroup_subsys_state
*css
=
4588 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4590 INIT_WORK(&css
->destroy_work
, css_release_work_fn
);
4591 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4594 static void init_and_link_css(struct cgroup_subsys_state
*css
,
4595 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
4597 lockdep_assert_held(&cgroup_mutex
);
4601 memset(css
, 0, sizeof(*css
));
4604 INIT_LIST_HEAD(&css
->sibling
);
4605 INIT_LIST_HEAD(&css
->children
);
4606 css
->serial_nr
= css_serial_nr_next
++;
4608 if (cgroup_parent(cgrp
)) {
4609 css
->parent
= cgroup_css(cgroup_parent(cgrp
), ss
);
4610 css_get(css
->parent
);
4613 BUG_ON(cgroup_css(cgrp
, ss
));
4616 /* invoke ->css_online() on a new CSS and mark it online if successful */
4617 static int online_css(struct cgroup_subsys_state
*css
)
4619 struct cgroup_subsys
*ss
= css
->ss
;
4622 lockdep_assert_held(&cgroup_mutex
);
4625 ret
= ss
->css_online(css
);
4627 css
->flags
|= CSS_ONLINE
;
4628 rcu_assign_pointer(css
->cgroup
->subsys
[ss
->id
], css
);
4633 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4634 static void offline_css(struct cgroup_subsys_state
*css
)
4636 struct cgroup_subsys
*ss
= css
->ss
;
4638 lockdep_assert_held(&cgroup_mutex
);
4640 if (!(css
->flags
& CSS_ONLINE
))
4643 if (ss
->css_offline
)
4644 ss
->css_offline(css
);
4646 css
->flags
&= ~CSS_ONLINE
;
4647 RCU_INIT_POINTER(css
->cgroup
->subsys
[ss
->id
], NULL
);
4649 wake_up_all(&css
->cgroup
->offline_waitq
);
4653 * create_css - create a cgroup_subsys_state
4654 * @cgrp: the cgroup new css will be associated with
4655 * @ss: the subsys of new css
4656 * @visible: whether to create control knobs for the new css or not
4658 * Create a new css associated with @cgrp - @ss pair. On success, the new
4659 * css is online and installed in @cgrp with all interface files created if
4660 * @visible. Returns 0 on success, -errno on failure.
4662 static int create_css(struct cgroup
*cgrp
, struct cgroup_subsys
*ss
,
4665 struct cgroup
*parent
= cgroup_parent(cgrp
);
4666 struct cgroup_subsys_state
*parent_css
= cgroup_css(parent
, ss
);
4667 struct cgroup_subsys_state
*css
;
4670 lockdep_assert_held(&cgroup_mutex
);
4672 css
= ss
->css_alloc(parent_css
);
4674 return PTR_ERR(css
);
4676 init_and_link_css(css
, ss
, cgrp
);
4678 err
= percpu_ref_init(&css
->refcnt
, css_release
, 0, GFP_KERNEL
);
4682 err
= cgroup_idr_alloc(&ss
->css_idr
, NULL
, 2, 0, GFP_KERNEL
);
4684 goto err_free_percpu_ref
;
4688 err
= cgroup_populate_dir(cgrp
, 1 << ss
->id
);
4693 /* @css is ready to be brought online now, make it visible */
4694 list_add_tail_rcu(&css
->sibling
, &parent_css
->children
);
4695 cgroup_idr_replace(&ss
->css_idr
, css
, css
->id
);
4697 err
= online_css(css
);
4701 if (ss
->broken_hierarchy
&& !ss
->warned_broken_hierarchy
&&
4702 cgroup_parent(parent
)) {
4703 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4704 current
->comm
, current
->pid
, ss
->name
);
4705 if (!strcmp(ss
->name
, "memory"))
4706 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4707 ss
->warned_broken_hierarchy
= true;
4713 list_del_rcu(&css
->sibling
);
4714 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4716 cgroup_idr_remove(&ss
->css_idr
, css
->id
);
4717 err_free_percpu_ref
:
4718 percpu_ref_exit(&css
->refcnt
);
4720 call_rcu(&css
->rcu_head
, css_free_rcu_fn
);
4724 static int cgroup_mkdir(struct kernfs_node
*parent_kn
, const char *name
,
4727 struct cgroup
*parent
, *cgrp
;
4728 struct cgroup_root
*root
;
4729 struct cgroup_subsys
*ss
;
4730 struct kernfs_node
*kn
;
4731 struct cftype
*base_files
;
4734 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4736 if (strchr(name
, '\n'))
4739 parent
= cgroup_kn_lock_live(parent_kn
);
4742 root
= parent
->root
;
4744 /* allocate the cgroup and its ID, 0 is reserved for the root */
4745 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
4751 ret
= percpu_ref_init(&cgrp
->self
.refcnt
, css_release
, 0, GFP_KERNEL
);
4756 * Temporarily set the pointer to NULL, so idr_find() won't return
4757 * a half-baked cgroup.
4759 cgrp
->id
= cgroup_idr_alloc(&root
->cgroup_idr
, NULL
, 2, 0, GFP_KERNEL
);
4762 goto out_cancel_ref
;
4765 init_cgroup_housekeeping(cgrp
);
4767 cgrp
->self
.parent
= &parent
->self
;
4770 if (notify_on_release(parent
))
4771 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
4773 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &parent
->flags
))
4774 set_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
);
4776 /* create the directory */
4777 kn
= kernfs_create_dir(parent
->kn
, name
, mode
, cgrp
);
4785 * This extra ref will be put in cgroup_free_fn() and guarantees
4786 * that @cgrp->kn is always accessible.
4790 cgrp
->self
.serial_nr
= css_serial_nr_next
++;
4792 /* allocation complete, commit to creation */
4793 list_add_tail_rcu(&cgrp
->self
.sibling
, &cgroup_parent(cgrp
)->self
.children
);
4794 atomic_inc(&root
->nr_cgrps
);
4798 * @cgrp is now fully operational. If something fails after this
4799 * point, it'll be released via the normal destruction path.
4801 cgroup_idr_replace(&root
->cgroup_idr
, cgrp
, cgrp
->id
);
4803 ret
= cgroup_kn_set_ugid(kn
);
4807 if (cgroup_on_dfl(cgrp
))
4808 base_files
= cgroup_dfl_base_files
;
4810 base_files
= cgroup_legacy_base_files
;
4812 ret
= cgroup_addrm_files(cgrp
, base_files
, true);
4816 /* let's create and online css's */
4817 for_each_subsys(ss
, ssid
) {
4818 if (parent
->child_subsys_mask
& (1 << ssid
)) {
4819 ret
= create_css(cgrp
, ss
,
4820 parent
->subtree_control
& (1 << ssid
));
4827 * On the default hierarchy, a child doesn't automatically inherit
4828 * subtree_control from the parent. Each is configured manually.
4830 if (!cgroup_on_dfl(cgrp
)) {
4831 cgrp
->subtree_control
= parent
->subtree_control
;
4832 cgroup_refresh_child_subsys_mask(cgrp
);
4835 kernfs_activate(kn
);
4841 cgroup_idr_remove(&root
->cgroup_idr
, cgrp
->id
);
4843 percpu_ref_exit(&cgrp
->self
.refcnt
);
4847 cgroup_kn_unlock(parent_kn
);
4851 cgroup_destroy_locked(cgrp
);
4856 * This is called when the refcnt of a css is confirmed to be killed.
4857 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4858 * initate destruction and put the css ref from kill_css().
4860 static void css_killed_work_fn(struct work_struct
*work
)
4862 struct cgroup_subsys_state
*css
=
4863 container_of(work
, struct cgroup_subsys_state
, destroy_work
);
4865 mutex_lock(&cgroup_mutex
);
4867 mutex_unlock(&cgroup_mutex
);
4872 /* css kill confirmation processing requires process context, bounce */
4873 static void css_killed_ref_fn(struct percpu_ref
*ref
)
4875 struct cgroup_subsys_state
*css
=
4876 container_of(ref
, struct cgroup_subsys_state
, refcnt
);
4878 INIT_WORK(&css
->destroy_work
, css_killed_work_fn
);
4879 queue_work(cgroup_destroy_wq
, &css
->destroy_work
);
4883 * kill_css - destroy a css
4884 * @css: css to destroy
4886 * This function initiates destruction of @css by removing cgroup interface
4887 * files and putting its base reference. ->css_offline() will be invoked
4888 * asynchronously once css_tryget_online() is guaranteed to fail and when
4889 * the reference count reaches zero, @css will be released.
4891 static void kill_css(struct cgroup_subsys_state
*css
)
4893 lockdep_assert_held(&cgroup_mutex
);
4896 * This must happen before css is disassociated with its cgroup.
4897 * See seq_css() for details.
4899 cgroup_clear_dir(css
->cgroup
, 1 << css
->ss
->id
);
4902 * Killing would put the base ref, but we need to keep it alive
4903 * until after ->css_offline().
4908 * cgroup core guarantees that, by the time ->css_offline() is
4909 * invoked, no new css reference will be given out via
4910 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4911 * proceed to offlining css's because percpu_ref_kill() doesn't
4912 * guarantee that the ref is seen as killed on all CPUs on return.
4914 * Use percpu_ref_kill_and_confirm() to get notifications as each
4915 * css is confirmed to be seen as killed on all CPUs.
4917 percpu_ref_kill_and_confirm(&css
->refcnt
, css_killed_ref_fn
);
4921 * cgroup_destroy_locked - the first stage of cgroup destruction
4922 * @cgrp: cgroup to be destroyed
4924 * css's make use of percpu refcnts whose killing latency shouldn't be
4925 * exposed to userland and are RCU protected. Also, cgroup core needs to
4926 * guarantee that css_tryget_online() won't succeed by the time
4927 * ->css_offline() is invoked. To satisfy all the requirements,
4928 * destruction is implemented in the following two steps.
4930 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4931 * userland visible parts and start killing the percpu refcnts of
4932 * css's. Set up so that the next stage will be kicked off once all
4933 * the percpu refcnts are confirmed to be killed.
4935 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4936 * rest of destruction. Once all cgroup references are gone, the
4937 * cgroup is RCU-freed.
4939 * This function implements s1. After this step, @cgrp is gone as far as
4940 * the userland is concerned and a new cgroup with the same name may be
4941 * created. As cgroup doesn't care about the names internally, this
4942 * doesn't cause any problem.
4944 static int cgroup_destroy_locked(struct cgroup
*cgrp
)
4945 __releases(&cgroup_mutex
) __acquires(&cgroup_mutex
)
4947 struct cgroup_subsys_state
*css
;
4951 lockdep_assert_held(&cgroup_mutex
);
4954 * css_set_rwsem synchronizes access to ->cset_links and prevents
4955 * @cgrp from being removed while put_css_set() is in progress.
4957 down_read(&css_set_rwsem
);
4958 empty
= list_empty(&cgrp
->cset_links
);
4959 up_read(&css_set_rwsem
);
4964 * Make sure there's no live children. We can't test emptiness of
4965 * ->self.children as dead children linger on it while being
4966 * drained; otherwise, "rmdir parent/child parent" may fail.
4968 if (css_has_online_children(&cgrp
->self
))
4972 * Mark @cgrp dead. This prevents further task migration and child
4973 * creation by disabling cgroup_lock_live_group().
4975 cgrp
->self
.flags
&= ~CSS_ONLINE
;
4977 /* initiate massacre of all css's */
4978 for_each_css(css
, ssid
, cgrp
)
4982 * Remove @cgrp directory along with the base files. @cgrp has an
4983 * extra ref on its kn.
4985 kernfs_remove(cgrp
->kn
);
4987 check_for_release(cgroup_parent(cgrp
));
4989 /* put the base reference */
4990 percpu_ref_kill(&cgrp
->self
.refcnt
);
4995 static int cgroup_rmdir(struct kernfs_node
*kn
)
4997 struct cgroup
*cgrp
;
5000 cgrp
= cgroup_kn_lock_live(kn
);
5004 ret
= cgroup_destroy_locked(cgrp
);
5006 cgroup_kn_unlock(kn
);
5010 static struct kernfs_syscall_ops cgroup_kf_syscall_ops
= {
5011 .remount_fs
= cgroup_remount
,
5012 .show_options
= cgroup_show_options
,
5013 .mkdir
= cgroup_mkdir
,
5014 .rmdir
= cgroup_rmdir
,
5015 .rename
= cgroup_rename
,
5018 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
, bool early
)
5020 struct cgroup_subsys_state
*css
;
5022 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
5024 mutex_lock(&cgroup_mutex
);
5026 idr_init(&ss
->css_idr
);
5027 INIT_LIST_HEAD(&ss
->cfts
);
5029 /* Create the root cgroup state for this subsystem */
5030 ss
->root
= &cgrp_dfl_root
;
5031 css
= ss
->css_alloc(cgroup_css(&cgrp_dfl_root
.cgrp
, ss
));
5032 /* We don't handle early failures gracefully */
5033 BUG_ON(IS_ERR(css
));
5034 init_and_link_css(css
, ss
, &cgrp_dfl_root
.cgrp
);
5037 * Root csses are never destroyed and we can't initialize
5038 * percpu_ref during early init. Disable refcnting.
5040 css
->flags
|= CSS_NO_REF
;
5043 /* allocation can't be done safely during early init */
5046 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2, GFP_KERNEL
);
5047 BUG_ON(css
->id
< 0);
5050 /* Update the init_css_set to contain a subsys
5051 * pointer to this state - since the subsystem is
5052 * newly registered, all tasks and hence the
5053 * init_css_set is in the subsystem's root cgroup. */
5054 init_css_set
.subsys
[ss
->id
] = css
;
5056 have_fork_callback
|= (bool)ss
->fork
<< ss
->id
;
5057 have_exit_callback
|= (bool)ss
->exit
<< ss
->id
;
5058 have_canfork_callback
|= (bool)ss
->can_fork
<< ss
->id
;
5060 /* At system boot, before all subsystems have been
5061 * registered, no tasks have been forked, so we don't
5062 * need to invoke fork callbacks here. */
5063 BUG_ON(!list_empty(&init_task
.tasks
));
5065 BUG_ON(online_css(css
));
5067 mutex_unlock(&cgroup_mutex
);
5071 * cgroup_init_early - cgroup initialization at system boot
5073 * Initialize cgroups at system boot, and initialize any
5074 * subsystems that request early init.
5076 int __init
cgroup_init_early(void)
5078 static struct cgroup_sb_opts __initdata opts
;
5079 struct cgroup_subsys
*ss
;
5082 init_cgroup_root(&cgrp_dfl_root
, &opts
);
5083 cgrp_dfl_root
.cgrp
.self
.flags
|= CSS_NO_REF
;
5085 RCU_INIT_POINTER(init_task
.cgroups
, &init_css_set
);
5087 for_each_subsys(ss
, i
) {
5088 WARN(!ss
->css_alloc
|| !ss
->css_free
|| ss
->name
|| ss
->id
,
5089 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5090 i
, cgroup_subsys_name
[i
], ss
->css_alloc
, ss
->css_free
,
5092 WARN(strlen(cgroup_subsys_name
[i
]) > MAX_CGROUP_TYPE_NAMELEN
,
5093 "cgroup_subsys_name %s too long\n", cgroup_subsys_name
[i
]);
5096 ss
->name
= cgroup_subsys_name
[i
];
5097 if (!ss
->legacy_name
)
5098 ss
->legacy_name
= cgroup_subsys_name
[i
];
5101 cgroup_init_subsys(ss
, true);
5107 * cgroup_init - cgroup initialization
5109 * Register cgroup filesystem and /proc file, and initialize
5110 * any subsystems that didn't request early init.
5112 int __init
cgroup_init(void)
5114 struct cgroup_subsys
*ss
;
5118 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem
));
5119 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_dfl_base_files
));
5120 BUG_ON(cgroup_init_cftypes(NULL
, cgroup_legacy_base_files
));
5122 mutex_lock(&cgroup_mutex
);
5124 /* Add init_css_set to the hash table */
5125 key
= css_set_hash(init_css_set
.subsys
);
5126 hash_add(css_set_table
, &init_css_set
.hlist
, key
);
5128 BUG_ON(cgroup_setup_root(&cgrp_dfl_root
, 0));
5130 mutex_unlock(&cgroup_mutex
);
5132 for_each_subsys(ss
, ssid
) {
5133 if (ss
->early_init
) {
5134 struct cgroup_subsys_state
*css
=
5135 init_css_set
.subsys
[ss
->id
];
5137 css
->id
= cgroup_idr_alloc(&ss
->css_idr
, css
, 1, 2,
5139 BUG_ON(css
->id
< 0);
5141 cgroup_init_subsys(ss
, false);
5144 list_add_tail(&init_css_set
.e_cset_node
[ssid
],
5145 &cgrp_dfl_root
.cgrp
.e_csets
[ssid
]);
5148 * Setting dfl_root subsys_mask needs to consider the
5149 * disabled flag and cftype registration needs kmalloc,
5150 * both of which aren't available during early_init.
5152 if (!cgroup_ssid_enabled(ssid
))
5155 cgrp_dfl_root
.subsys_mask
|= 1 << ss
->id
;
5157 if (cgroup_legacy_files_on_dfl
&& !ss
->dfl_cftypes
)
5158 ss
->dfl_cftypes
= ss
->legacy_cftypes
;
5160 if (!ss
->dfl_cftypes
)
5161 cgrp_dfl_root_inhibit_ss_mask
|= 1 << ss
->id
;
5163 if (ss
->dfl_cftypes
== ss
->legacy_cftypes
) {
5164 WARN_ON(cgroup_add_cftypes(ss
, ss
->dfl_cftypes
));
5166 WARN_ON(cgroup_add_dfl_cftypes(ss
, ss
->dfl_cftypes
));
5167 WARN_ON(cgroup_add_legacy_cftypes(ss
, ss
->legacy_cftypes
));
5171 ss
->bind(init_css_set
.subsys
[ssid
]);
5174 err
= sysfs_create_mount_point(fs_kobj
, "cgroup");
5178 err
= register_filesystem(&cgroup_fs_type
);
5180 sysfs_remove_mount_point(fs_kobj
, "cgroup");
5184 proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
);
5188 static int __init
cgroup_wq_init(void)
5191 * There isn't much point in executing destruction path in
5192 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5193 * Use 1 for @max_active.
5195 * We would prefer to do this in cgroup_init() above, but that
5196 * is called before init_workqueues(): so leave this until after.
5198 cgroup_destroy_wq
= alloc_workqueue("cgroup_destroy", 0, 1);
5199 BUG_ON(!cgroup_destroy_wq
);
5202 * Used to destroy pidlists and separate to serve as flush domain.
5203 * Cap @max_active to 1 too.
5205 cgroup_pidlist_destroy_wq
= alloc_workqueue("cgroup_pidlist_destroy",
5207 BUG_ON(!cgroup_pidlist_destroy_wq
);
5211 core_initcall(cgroup_wq_init
);
5214 * proc_cgroup_show()
5215 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5216 * - Used for /proc/<pid>/cgroup.
5218 int proc_cgroup_show(struct seq_file
*m
, struct pid_namespace
*ns
,
5219 struct pid
*pid
, struct task_struct
*tsk
)
5223 struct cgroup_root
*root
;
5226 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5230 mutex_lock(&cgroup_mutex
);
5231 down_read(&css_set_rwsem
);
5233 for_each_root(root
) {
5234 struct cgroup_subsys
*ss
;
5235 struct cgroup
*cgrp
;
5236 int ssid
, count
= 0;
5238 if (root
== &cgrp_dfl_root
&& !cgrp_dfl_root_visible
)
5241 seq_printf(m
, "%d:", root
->hierarchy_id
);
5242 if (root
!= &cgrp_dfl_root
)
5243 for_each_subsys(ss
, ssid
)
5244 if (root
->subsys_mask
& (1 << ssid
))
5245 seq_printf(m
, "%s%s", count
++ ? "," : "",
5247 if (strlen(root
->name
))
5248 seq_printf(m
, "%sname=%s", count
? "," : "",
5251 cgrp
= task_cgroup_from_root(tsk
, root
);
5252 path
= cgroup_path(cgrp
, buf
, PATH_MAX
);
5254 retval
= -ENAMETOOLONG
;
5263 up_read(&css_set_rwsem
);
5264 mutex_unlock(&cgroup_mutex
);
5270 /* Display information about each subsystem and each hierarchy */
5271 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
5273 struct cgroup_subsys
*ss
;
5276 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5278 * ideally we don't want subsystems moving around while we do this.
5279 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5280 * subsys/hierarchy state.
5282 mutex_lock(&cgroup_mutex
);
5284 for_each_subsys(ss
, i
)
5285 seq_printf(m
, "%s\t%d\t%d\t%d\n",
5286 ss
->legacy_name
, ss
->root
->hierarchy_id
,
5287 atomic_read(&ss
->root
->nr_cgrps
),
5288 cgroup_ssid_enabled(i
));
5290 mutex_unlock(&cgroup_mutex
);
5294 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
5296 return single_open(file
, proc_cgroupstats_show
, NULL
);
5299 static const struct file_operations proc_cgroupstats_operations
= {
5300 .open
= cgroupstats_open
,
5302 .llseek
= seq_lseek
,
5303 .release
= single_release
,
5306 static void **subsys_canfork_priv_p(void *ss_priv
[CGROUP_CANFORK_COUNT
], int i
)
5308 if (CGROUP_CANFORK_START
<= i
&& i
< CGROUP_CANFORK_END
)
5309 return &ss_priv
[i
- CGROUP_CANFORK_START
];
5313 static void *subsys_canfork_priv(void *ss_priv
[CGROUP_CANFORK_COUNT
], int i
)
5315 void **private = subsys_canfork_priv_p(ss_priv
, i
);
5316 return private ? *private : NULL
;
5320 * cgroup_fork - initialize cgroup related fields during copy_process()
5321 * @child: pointer to task_struct of forking parent process.
5323 * A task is associated with the init_css_set until cgroup_post_fork()
5324 * attaches it to the parent's css_set. Empty cg_list indicates that
5325 * @child isn't holding reference to its css_set.
5327 void cgroup_fork(struct task_struct
*child
)
5329 RCU_INIT_POINTER(child
->cgroups
, &init_css_set
);
5330 INIT_LIST_HEAD(&child
->cg_list
);
5334 * cgroup_can_fork - called on a new task before the process is exposed
5335 * @child: the task in question.
5337 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5338 * returns an error, the fork aborts with that error code. This allows for
5339 * a cgroup subsystem to conditionally allow or deny new forks.
5341 int cgroup_can_fork(struct task_struct
*child
,
5342 void *ss_priv
[CGROUP_CANFORK_COUNT
])
5344 struct cgroup_subsys
*ss
;
5347 for_each_subsys_which(ss
, i
, &have_canfork_callback
) {
5348 ret
= ss
->can_fork(child
, subsys_canfork_priv_p(ss_priv
, i
));
5356 for_each_subsys(ss
, j
) {
5359 if (ss
->cancel_fork
)
5360 ss
->cancel_fork(child
, subsys_canfork_priv(ss_priv
, j
));
5367 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5368 * @child: the task in question
5370 * This calls the cancel_fork() callbacks if a fork failed *after*
5371 * cgroup_can_fork() succeded.
5373 void cgroup_cancel_fork(struct task_struct
*child
,
5374 void *ss_priv
[CGROUP_CANFORK_COUNT
])
5376 struct cgroup_subsys
*ss
;
5379 for_each_subsys(ss
, i
)
5380 if (ss
->cancel_fork
)
5381 ss
->cancel_fork(child
, subsys_canfork_priv(ss_priv
, i
));
5385 * cgroup_post_fork - called on a new task after adding it to the task list
5386 * @child: the task in question
5388 * Adds the task to the list running through its css_set if necessary and
5389 * call the subsystem fork() callbacks. Has to be after the task is
5390 * visible on the task list in case we race with the first call to
5391 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5394 void cgroup_post_fork(struct task_struct
*child
,
5395 void *old_ss_priv
[CGROUP_CANFORK_COUNT
])
5397 struct cgroup_subsys
*ss
;
5401 * This may race against cgroup_enable_task_cg_lists(). As that
5402 * function sets use_task_css_set_links before grabbing
5403 * tasklist_lock and we just went through tasklist_lock to add
5404 * @child, it's guaranteed that either we see the set
5405 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5406 * @child during its iteration.
5408 * If we won the race, @child is associated with %current's
5409 * css_set. Grabbing css_set_rwsem guarantees both that the
5410 * association is stable, and, on completion of the parent's
5411 * migration, @child is visible in the source of migration or
5412 * already in the destination cgroup. This guarantee is necessary
5413 * when implementing operations which need to migrate all tasks of
5414 * a cgroup to another.
5416 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5417 * will remain in init_css_set. This is safe because all tasks are
5418 * in the init_css_set before cg_links is enabled and there's no
5419 * operation which transfers all tasks out of init_css_set.
5421 if (use_task_css_set_links
) {
5422 struct css_set
*cset
;
5424 down_write(&css_set_rwsem
);
5425 cset
= task_css_set(current
);
5426 if (list_empty(&child
->cg_list
)) {
5427 rcu_assign_pointer(child
->cgroups
, cset
);
5428 list_add(&child
->cg_list
, &cset
->tasks
);
5431 up_write(&css_set_rwsem
);
5435 * Call ss->fork(). This must happen after @child is linked on
5436 * css_set; otherwise, @child might change state between ->fork()
5437 * and addition to css_set.
5439 for_each_subsys_which(ss
, i
, &have_fork_callback
)
5440 ss
->fork(child
, subsys_canfork_priv(old_ss_priv
, i
));
5444 * cgroup_exit - detach cgroup from exiting task
5445 * @tsk: pointer to task_struct of exiting process
5447 * Description: Detach cgroup from @tsk and release it.
5449 * Note that cgroups marked notify_on_release force every task in
5450 * them to take the global cgroup_mutex mutex when exiting.
5451 * This could impact scaling on very large systems. Be reluctant to
5452 * use notify_on_release cgroups where very high task exit scaling
5453 * is required on large systems.
5455 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5456 * call cgroup_exit() while the task is still competent to handle
5457 * notify_on_release(), then leave the task attached to the root cgroup in
5458 * each hierarchy for the remainder of its exit. No need to bother with
5459 * init_css_set refcnting. init_css_set never goes away and we can't race
5460 * with migration path - PF_EXITING is visible to migration path.
5462 void cgroup_exit(struct task_struct
*tsk
)
5464 struct cgroup_subsys
*ss
;
5465 struct css_set
*cset
;
5466 bool put_cset
= false;
5470 * Unlink from @tsk from its css_set. As migration path can't race
5471 * with us, we can check cg_list without grabbing css_set_rwsem.
5473 if (!list_empty(&tsk
->cg_list
)) {
5474 down_write(&css_set_rwsem
);
5475 list_del_init(&tsk
->cg_list
);
5476 up_write(&css_set_rwsem
);
5480 /* Reassign the task to the init_css_set. */
5481 cset
= task_css_set(tsk
);
5482 RCU_INIT_POINTER(tsk
->cgroups
, &init_css_set
);
5484 /* see cgroup_post_fork() for details */
5485 for_each_subsys_which(ss
, i
, &have_exit_callback
) {
5486 struct cgroup_subsys_state
*old_css
= cset
->subsys
[i
];
5487 struct cgroup_subsys_state
*css
= task_css(tsk
, i
);
5489 ss
->exit(css
, old_css
, tsk
);
5496 static void check_for_release(struct cgroup
*cgrp
)
5498 if (notify_on_release(cgrp
) && !cgroup_has_tasks(cgrp
) &&
5499 !css_has_online_children(&cgrp
->self
) && !cgroup_is_dead(cgrp
))
5500 schedule_work(&cgrp
->release_agent_work
);
5504 * Notify userspace when a cgroup is released, by running the
5505 * configured release agent with the name of the cgroup (path
5506 * relative to the root of cgroup file system) as the argument.
5508 * Most likely, this user command will try to rmdir this cgroup.
5510 * This races with the possibility that some other task will be
5511 * attached to this cgroup before it is removed, or that some other
5512 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5513 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5514 * unused, and this cgroup will be reprieved from its death sentence,
5515 * to continue to serve a useful existence. Next time it's released,
5516 * we will get notified again, if it still has 'notify_on_release' set.
5518 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5519 * means only wait until the task is successfully execve()'d. The
5520 * separate release agent task is forked by call_usermodehelper(),
5521 * then control in this thread returns here, without waiting for the
5522 * release agent task. We don't bother to wait because the caller of
5523 * this routine has no use for the exit status of the release agent
5524 * task, so no sense holding our caller up for that.
5526 static void cgroup_release_agent(struct work_struct
*work
)
5528 struct cgroup
*cgrp
=
5529 container_of(work
, struct cgroup
, release_agent_work
);
5530 char *pathbuf
= NULL
, *agentbuf
= NULL
, *path
;
5531 char *argv
[3], *envp
[3];
5533 mutex_lock(&cgroup_mutex
);
5535 pathbuf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
5536 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
5537 if (!pathbuf
|| !agentbuf
)
5540 path
= cgroup_path(cgrp
, pathbuf
, PATH_MAX
);
5548 /* minimal command environment */
5550 envp
[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5553 mutex_unlock(&cgroup_mutex
);
5554 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
5557 mutex_unlock(&cgroup_mutex
);
5563 static int __init
cgroup_disable(char *str
)
5565 struct cgroup_subsys
*ss
;
5569 while ((token
= strsep(&str
, ",")) != NULL
) {
5573 for_each_subsys(ss
, i
) {
5574 if (strcmp(token
, ss
->name
) &&
5575 strcmp(token
, ss
->legacy_name
))
5578 static_branch_disable(cgroup_subsys_enabled_key
[i
]);
5579 printk(KERN_INFO
"Disabling %s control group subsystem\n",
5586 __setup("cgroup_disable=", cgroup_disable
);
5588 static int __init
cgroup_set_legacy_files_on_dfl(char *str
)
5590 printk("cgroup: using legacy files on the default hierarchy\n");
5591 cgroup_legacy_files_on_dfl
= true;
5594 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl
);
5597 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5598 * @dentry: directory dentry of interest
5599 * @ss: subsystem of interest
5601 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5602 * to get the corresponding css and return it. If such css doesn't exist
5603 * or can't be pinned, an ERR_PTR value is returned.
5605 struct cgroup_subsys_state
*css_tryget_online_from_dir(struct dentry
*dentry
,
5606 struct cgroup_subsys
*ss
)
5608 struct kernfs_node
*kn
= kernfs_node_from_dentry(dentry
);
5609 struct cgroup_subsys_state
*css
= NULL
;
5610 struct cgroup
*cgrp
;
5612 /* is @dentry a cgroup dir? */
5613 if (dentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !kn
||
5614 kernfs_type(kn
) != KERNFS_DIR
)
5615 return ERR_PTR(-EBADF
);
5620 * This path doesn't originate from kernfs and @kn could already
5621 * have been or be removed at any point. @kn->priv is RCU
5622 * protected for this access. See css_release_work_fn() for details.
5624 cgrp
= rcu_dereference(kn
->priv
);
5626 css
= cgroup_css(cgrp
, ss
);
5628 if (!css
|| !css_tryget_online(css
))
5629 css
= ERR_PTR(-ENOENT
);
5636 * css_from_id - lookup css by id
5637 * @id: the cgroup id
5638 * @ss: cgroup subsys to be looked into
5640 * Returns the css if there's valid one with @id, otherwise returns NULL.
5641 * Should be called under rcu_read_lock().
5643 struct cgroup_subsys_state
*css_from_id(int id
, struct cgroup_subsys
*ss
)
5645 WARN_ON_ONCE(!rcu_read_lock_held());
5646 return id
> 0 ? idr_find(&ss
->css_idr
, id
) : NULL
;
5649 #ifdef CONFIG_CGROUP_DEBUG
5650 static struct cgroup_subsys_state
*
5651 debug_css_alloc(struct cgroup_subsys_state
*parent_css
)
5653 struct cgroup_subsys_state
*css
= kzalloc(sizeof(*css
), GFP_KERNEL
);
5656 return ERR_PTR(-ENOMEM
);
5661 static void debug_css_free(struct cgroup_subsys_state
*css
)
5666 static u64
debug_taskcount_read(struct cgroup_subsys_state
*css
,
5669 return cgroup_task_count(css
->cgroup
);
5672 static u64
current_css_set_read(struct cgroup_subsys_state
*css
,
5675 return (u64
)(unsigned long)current
->cgroups
;
5678 static u64
current_css_set_refcount_read(struct cgroup_subsys_state
*css
,
5684 count
= atomic_read(&task_css_set(current
)->refcount
);
5689 static int current_css_set_cg_links_read(struct seq_file
*seq
, void *v
)
5691 struct cgrp_cset_link
*link
;
5692 struct css_set
*cset
;
5695 name_buf
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
5699 down_read(&css_set_rwsem
);
5701 cset
= rcu_dereference(current
->cgroups
);
5702 list_for_each_entry(link
, &cset
->cgrp_links
, cgrp_link
) {
5703 struct cgroup
*c
= link
->cgrp
;
5705 cgroup_name(c
, name_buf
, NAME_MAX
+ 1);
5706 seq_printf(seq
, "Root %d group %s\n",
5707 c
->root
->hierarchy_id
, name_buf
);
5710 up_read(&css_set_rwsem
);
5715 #define MAX_TASKS_SHOWN_PER_CSS 25
5716 static int cgroup_css_links_read(struct seq_file
*seq
, void *v
)
5718 struct cgroup_subsys_state
*css
= seq_css(seq
);
5719 struct cgrp_cset_link
*link
;
5721 down_read(&css_set_rwsem
);
5722 list_for_each_entry(link
, &css
->cgroup
->cset_links
, cset_link
) {
5723 struct css_set
*cset
= link
->cset
;
5724 struct task_struct
*task
;
5727 seq_printf(seq
, "css_set %p\n", cset
);
5729 list_for_each_entry(task
, &cset
->tasks
, cg_list
) {
5730 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5732 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5735 list_for_each_entry(task
, &cset
->mg_tasks
, cg_list
) {
5736 if (count
++ > MAX_TASKS_SHOWN_PER_CSS
)
5738 seq_printf(seq
, " task %d\n", task_pid_vnr(task
));
5742 seq_puts(seq
, " ...\n");
5744 up_read(&css_set_rwsem
);
5748 static u64
releasable_read(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
5750 return (!cgroup_has_tasks(css
->cgroup
) &&
5751 !css_has_online_children(&css
->cgroup
->self
));
5754 static struct cftype debug_files
[] = {
5756 .name
= "taskcount",
5757 .read_u64
= debug_taskcount_read
,
5761 .name
= "current_css_set",
5762 .read_u64
= current_css_set_read
,
5766 .name
= "current_css_set_refcount",
5767 .read_u64
= current_css_set_refcount_read
,
5771 .name
= "current_css_set_cg_links",
5772 .seq_show
= current_css_set_cg_links_read
,
5776 .name
= "cgroup_css_links",
5777 .seq_show
= cgroup_css_links_read
,
5781 .name
= "releasable",
5782 .read_u64
= releasable_read
,
5788 struct cgroup_subsys debug_cgrp_subsys
= {
5789 .css_alloc
= debug_css_alloc
,
5790 .css_free
= debug_css_free
,
5791 .legacy_cftypes
= debug_files
,
5793 #endif /* CONFIG_CGROUP_DEBUG */