2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Copyright notices from the original cpuset code:
8 * --------------------------------------------------
9 * Copyright (C) 2003 BULL SA.
10 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
12 * Portions derived from Patrick Mochel's sysfs code.
13 * sysfs is Copyright (c) 2001-3 Patrick Mochel
15 * 2003-10-10 Written by Simon Derr.
16 * 2003-10-22 Updates by Stephen Hemminger.
17 * 2004 May-July Rework by Paul Jackson.
18 * ---------------------------------------------------
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cgroup.h>
26 #include <linux/errno.h>
28 #include <linux/kernel.h>
29 #include <linux/list.h>
31 #include <linux/mutex.h>
32 #include <linux/mount.h>
33 #include <linux/pagemap.h>
34 #include <linux/proc_fs.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched.h>
37 #include <linux/backing-dev.h>
38 #include <linux/seq_file.h>
39 #include <linux/slab.h>
40 #include <linux/magic.h>
41 #include <linux/spinlock.h>
42 #include <linux/string.h>
43 #include <linux/sort.h>
44 #include <linux/kmod.h>
45 #include <linux/delayacct.h>
46 #include <linux/cgroupstats.h>
47 #include <linux/hash.h>
48 #include <linux/namei.h>
50 #include <asm/atomic.h>
52 static DEFINE_MUTEX(cgroup_mutex
);
54 /* Generate an array of cgroup subsystem pointers */
55 #define SUBSYS(_x) &_x ## _subsys,
57 static struct cgroup_subsys
*subsys
[] = {
58 #include <linux/cgroup_subsys.h>
62 * A cgroupfs_root represents the root of a cgroup hierarchy,
63 * and may be associated with a superblock to form an active
66 struct cgroupfs_root
{
67 struct super_block
*sb
;
70 * The bitmask of subsystems intended to be attached to this
73 unsigned long subsys_bits
;
75 /* The bitmask of subsystems currently attached to this hierarchy */
76 unsigned long actual_subsys_bits
;
78 /* A list running through the attached subsystems */
79 struct list_head subsys_list
;
81 /* The root cgroup for this hierarchy */
82 struct cgroup top_cgroup
;
84 /* Tracks how many cgroups are currently defined in hierarchy.*/
85 int number_of_cgroups
;
87 /* A list running through the active hierarchies */
88 struct list_head root_list
;
90 /* Hierarchy-specific flags */
93 /* The path to use for release notifications. */
94 char release_agent_path
[PATH_MAX
];
99 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
100 * subsystems that are otherwise unattached - it never has more than a
101 * single cgroup, and all tasks are part of that cgroup.
103 static struct cgroupfs_root rootnode
;
105 /* The list of hierarchy roots */
107 static LIST_HEAD(roots
);
108 static int root_count
;
110 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
111 #define dummytop (&rootnode.top_cgroup)
113 /* This flag indicates whether tasks in the fork and exit paths should
114 * check for fork/exit handlers to call. This avoids us having to do
115 * extra work in the fork/exit path if none of the subsystems need to
118 static int need_forkexit_callback __read_mostly
;
120 /* convenient tests for these bits */
121 inline int cgroup_is_removed(const struct cgroup
*cgrp
)
123 return test_bit(CGRP_REMOVED
, &cgrp
->flags
);
126 /* bits in struct cgroupfs_root flags field */
128 ROOT_NOPREFIX
, /* mounted subsystems have no named prefix */
131 static int cgroup_is_releasable(const struct cgroup
*cgrp
)
134 (1 << CGRP_RELEASABLE
) |
135 (1 << CGRP_NOTIFY_ON_RELEASE
);
136 return (cgrp
->flags
& bits
) == bits
;
139 static int notify_on_release(const struct cgroup
*cgrp
)
141 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
145 * for_each_subsys() allows you to iterate on each subsystem attached to
146 * an active hierarchy
148 #define for_each_subsys(_root, _ss) \
149 list_for_each_entry(_ss, &_root->subsys_list, sibling)
151 /* for_each_active_root() allows you to iterate across the active hierarchies */
152 #define for_each_active_root(_root) \
153 list_for_each_entry(_root, &roots, root_list)
155 /* the list of cgroups eligible for automatic release. Protected by
156 * release_list_lock */
157 static LIST_HEAD(release_list
);
158 static DEFINE_SPINLOCK(release_list_lock
);
159 static void cgroup_release_agent(struct work_struct
*work
);
160 static DECLARE_WORK(release_agent_work
, cgroup_release_agent
);
161 static void check_for_release(struct cgroup
*cgrp
);
163 /* Link structure for associating css_set objects with cgroups */
164 struct cg_cgroup_link
{
166 * List running through cg_cgroup_links associated with a
167 * cgroup, anchored on cgroup->css_sets
169 struct list_head cgrp_link_list
;
171 * List running through cg_cgroup_links pointing at a
172 * single css_set object, anchored on css_set->cg_links
174 struct list_head cg_link_list
;
178 /* The default css_set - used by init and its children prior to any
179 * hierarchies being mounted. It contains a pointer to the root state
180 * for each subsystem. Also used to anchor the list of css_sets. Not
181 * reference-counted, to improve performance when child cgroups
182 * haven't been created.
185 static struct css_set init_css_set
;
186 static struct cg_cgroup_link init_css_set_link
;
188 /* css_set_lock protects the list of css_set objects, and the
189 * chain of tasks off each css_set. Nests outside task->alloc_lock
190 * due to cgroup_iter_start() */
191 static DEFINE_RWLOCK(css_set_lock
);
192 static int css_set_count
;
194 /* hash table for cgroup groups. This improves the performance to
195 * find an existing css_set */
196 #define CSS_SET_HASH_BITS 7
197 #define CSS_SET_TABLE_SIZE (1 << CSS_SET_HASH_BITS)
198 static struct hlist_head css_set_table
[CSS_SET_TABLE_SIZE
];
200 static struct hlist_head
*css_set_hash(struct cgroup_subsys_state
*css
[])
204 unsigned long tmp
= 0UL;
206 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++)
207 tmp
+= (unsigned long)css
[i
];
208 tmp
= (tmp
>> 16) ^ tmp
;
210 index
= hash_long(tmp
, CSS_SET_HASH_BITS
);
212 return &css_set_table
[index
];
215 /* We don't maintain the lists running through each css_set to its
216 * task until after the first call to cgroup_iter_start(). This
217 * reduces the fork()/exit() overhead for people who have cgroups
218 * compiled into their kernel but not actually in use */
219 static int use_task_css_set_links __read_mostly
;
221 /* When we create or destroy a css_set, the operation simply
222 * takes/releases a reference count on all the cgroups referenced
223 * by subsystems in this css_set. This can end up multiple-counting
224 * some cgroups, but that's OK - the ref-count is just a
225 * busy/not-busy indicator; ensuring that we only count each cgroup
226 * once would require taking a global lock to ensure that no
227 * subsystems moved between hierarchies while we were doing so.
229 * Possible TODO: decide at boot time based on the number of
230 * registered subsystems and the number of CPUs or NUMA nodes whether
231 * it's better for performance to ref-count every subsystem, or to
232 * take a global lock and only add one ref count to each hierarchy.
236 * unlink a css_set from the list and free it
238 static void unlink_css_set(struct css_set
*cg
)
240 struct cg_cgroup_link
*link
;
241 struct cg_cgroup_link
*saved_link
;
243 hlist_del(&cg
->hlist
);
246 list_for_each_entry_safe(link
, saved_link
, &cg
->cg_links
,
248 list_del(&link
->cg_link_list
);
249 list_del(&link
->cgrp_link_list
);
254 static void __put_css_set(struct css_set
*cg
, int taskexit
)
258 * Ensure that the refcount doesn't hit zero while any readers
259 * can see it. Similar to atomic_dec_and_lock(), but for an
262 if (atomic_add_unless(&cg
->refcount
, -1, 1))
264 write_lock(&css_set_lock
);
265 if (!atomic_dec_and_test(&cg
->refcount
)) {
266 write_unlock(&css_set_lock
);
270 write_unlock(&css_set_lock
);
273 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
274 struct cgroup
*cgrp
= cg
->subsys
[i
]->cgroup
;
275 if (atomic_dec_and_test(&cgrp
->count
) &&
276 notify_on_release(cgrp
)) {
278 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
279 check_for_release(cgrp
);
287 * refcounted get/put for css_set objects
289 static inline void get_css_set(struct css_set
*cg
)
291 atomic_inc(&cg
->refcount
);
294 static inline void put_css_set(struct css_set
*cg
)
296 __put_css_set(cg
, 0);
299 static inline void put_css_set_taskexit(struct css_set
*cg
)
301 __put_css_set(cg
, 1);
305 * find_existing_css_set() is a helper for
306 * find_css_set(), and checks to see whether an existing
307 * css_set is suitable.
309 * oldcg: the cgroup group that we're using before the cgroup
312 * cgrp: the cgroup that we're moving into
314 * template: location in which to build the desired set of subsystem
315 * state objects for the new cgroup group
317 static struct css_set
*find_existing_css_set(
318 struct css_set
*oldcg
,
320 struct cgroup_subsys_state
*template[])
323 struct cgroupfs_root
*root
= cgrp
->root
;
324 struct hlist_head
*hhead
;
325 struct hlist_node
*node
;
328 /* Built the set of subsystem state objects that we want to
329 * see in the new css_set */
330 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
331 if (root
->subsys_bits
& (1UL << i
)) {
332 /* Subsystem is in this hierarchy. So we want
333 * the subsystem state from the new
335 template[i
] = cgrp
->subsys
[i
];
337 /* Subsystem is not in this hierarchy, so we
338 * don't want to change the subsystem state */
339 template[i
] = oldcg
->subsys
[i
];
343 hhead
= css_set_hash(template);
344 hlist_for_each_entry(cg
, node
, hhead
, hlist
) {
345 if (!memcmp(template, cg
->subsys
, sizeof(cg
->subsys
))) {
346 /* All subsystems matched */
351 /* No existing cgroup group matched */
355 static void free_cg_links(struct list_head
*tmp
)
357 struct cg_cgroup_link
*link
;
358 struct cg_cgroup_link
*saved_link
;
360 list_for_each_entry_safe(link
, saved_link
, tmp
, cgrp_link_list
) {
361 list_del(&link
->cgrp_link_list
);
367 * allocate_cg_links() allocates "count" cg_cgroup_link structures
368 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
369 * success or a negative error
371 static int allocate_cg_links(int count
, struct list_head
*tmp
)
373 struct cg_cgroup_link
*link
;
376 for (i
= 0; i
< count
; i
++) {
377 link
= kmalloc(sizeof(*link
), GFP_KERNEL
);
382 list_add(&link
->cgrp_link_list
, tmp
);
388 * find_css_set() takes an existing cgroup group and a
389 * cgroup object, and returns a css_set object that's
390 * equivalent to the old group, but with the given cgroup
391 * substituted into the appropriate hierarchy. Must be called with
394 static struct css_set
*find_css_set(
395 struct css_set
*oldcg
, struct cgroup
*cgrp
)
398 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
];
401 struct list_head tmp_cg_links
;
402 struct cg_cgroup_link
*link
;
404 struct hlist_head
*hhead
;
406 /* First see if we already have a cgroup group that matches
408 read_lock(&css_set_lock
);
409 res
= find_existing_css_set(oldcg
, cgrp
, template);
412 read_unlock(&css_set_lock
);
417 res
= kmalloc(sizeof(*res
), GFP_KERNEL
);
421 /* Allocate all the cg_cgroup_link objects that we'll need */
422 if (allocate_cg_links(root_count
, &tmp_cg_links
) < 0) {
427 atomic_set(&res
->refcount
, 1);
428 INIT_LIST_HEAD(&res
->cg_links
);
429 INIT_LIST_HEAD(&res
->tasks
);
430 INIT_HLIST_NODE(&res
->hlist
);
432 /* Copy the set of subsystem state objects generated in
433 * find_existing_css_set() */
434 memcpy(res
->subsys
, template, sizeof(res
->subsys
));
436 write_lock(&css_set_lock
);
437 /* Add reference counts and links from the new css_set. */
438 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
439 struct cgroup
*cgrp
= res
->subsys
[i
]->cgroup
;
440 struct cgroup_subsys
*ss
= subsys
[i
];
441 atomic_inc(&cgrp
->count
);
443 * We want to add a link once per cgroup, so we
444 * only do it for the first subsystem in each
447 if (ss
->root
->subsys_list
.next
== &ss
->sibling
) {
448 BUG_ON(list_empty(&tmp_cg_links
));
449 link
= list_entry(tmp_cg_links
.next
,
450 struct cg_cgroup_link
,
452 list_del(&link
->cgrp_link_list
);
453 list_add(&link
->cgrp_link_list
, &cgrp
->css_sets
);
455 list_add(&link
->cg_link_list
, &res
->cg_links
);
458 if (list_empty(&rootnode
.subsys_list
)) {
459 link
= list_entry(tmp_cg_links
.next
,
460 struct cg_cgroup_link
,
462 list_del(&link
->cgrp_link_list
);
463 list_add(&link
->cgrp_link_list
, &dummytop
->css_sets
);
465 list_add(&link
->cg_link_list
, &res
->cg_links
);
468 BUG_ON(!list_empty(&tmp_cg_links
));
472 /* Add this cgroup group to the hash table */
473 hhead
= css_set_hash(res
->subsys
);
474 hlist_add_head(&res
->hlist
, hhead
);
476 write_unlock(&css_set_lock
);
482 * There is one global cgroup mutex. We also require taking
483 * task_lock() when dereferencing a task's cgroup subsys pointers.
484 * See "The task_lock() exception", at the end of this comment.
486 * A task must hold cgroup_mutex to modify cgroups.
488 * Any task can increment and decrement the count field without lock.
489 * So in general, code holding cgroup_mutex can't rely on the count
490 * field not changing. However, if the count goes to zero, then only
491 * cgroup_attach_task() can increment it again. Because a count of zero
492 * means that no tasks are currently attached, therefore there is no
493 * way a task attached to that cgroup can fork (the other way to
494 * increment the count). So code holding cgroup_mutex can safely
495 * assume that if the count is zero, it will stay zero. Similarly, if
496 * a task holds cgroup_mutex on a cgroup with zero count, it
497 * knows that the cgroup won't be removed, as cgroup_rmdir()
500 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
501 * (usually) take cgroup_mutex. These are the two most performance
502 * critical pieces of code here. The exception occurs on cgroup_exit(),
503 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
504 * is taken, and if the cgroup count is zero, a usermode call made
505 * to the release agent with the name of the cgroup (path relative to
506 * the root of cgroup file system) as the argument.
508 * A cgroup can only be deleted if both its 'count' of using tasks
509 * is zero, and its list of 'children' cgroups is empty. Since all
510 * tasks in the system use _some_ cgroup, and since there is always at
511 * least one task in the system (init, pid == 1), therefore, top_cgroup
512 * always has either children cgroups and/or using tasks. So we don't
513 * need a special hack to ensure that top_cgroup cannot be deleted.
515 * The task_lock() exception
517 * The need for this exception arises from the action of
518 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
519 * another. It does so using cgroup_mutex, however there are
520 * several performance critical places that need to reference
521 * task->cgroup without the expense of grabbing a system global
522 * mutex. Therefore except as noted below, when dereferencing or, as
523 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
524 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
525 * the task_struct routinely used for such matters.
527 * P.S. One more locking exception. RCU is used to guard the
528 * update of a tasks cgroup pointer by cgroup_attach_task()
532 * cgroup_lock - lock out any changes to cgroup structures
535 void cgroup_lock(void)
537 mutex_lock(&cgroup_mutex
);
541 * cgroup_unlock - release lock on cgroup changes
543 * Undo the lock taken in a previous cgroup_lock() call.
545 void cgroup_unlock(void)
547 mutex_unlock(&cgroup_mutex
);
551 * A couple of forward declarations required, due to cyclic reference loop:
552 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
553 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
557 static int cgroup_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
);
558 static int cgroup_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
);
559 static int cgroup_populate_dir(struct cgroup
*cgrp
);
560 static struct inode_operations cgroup_dir_inode_operations
;
561 static struct file_operations proc_cgroupstats_operations
;
563 static struct backing_dev_info cgroup_backing_dev_info
= {
564 .capabilities
= BDI_CAP_NO_ACCT_AND_WRITEBACK
,
567 static struct inode
*cgroup_new_inode(mode_t mode
, struct super_block
*sb
)
569 struct inode
*inode
= new_inode(sb
);
572 inode
->i_mode
= mode
;
573 inode
->i_uid
= current_fsuid();
574 inode
->i_gid
= current_fsgid();
575 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
576 inode
->i_mapping
->backing_dev_info
= &cgroup_backing_dev_info
;
582 * Call subsys's pre_destroy handler.
583 * This is called before css refcnt check.
585 static void cgroup_call_pre_destroy(struct cgroup
*cgrp
)
587 struct cgroup_subsys
*ss
;
588 for_each_subsys(cgrp
->root
, ss
)
590 ss
->pre_destroy(ss
, cgrp
);
594 static void cgroup_diput(struct dentry
*dentry
, struct inode
*inode
)
596 /* is dentry a directory ? if so, kfree() associated cgroup */
597 if (S_ISDIR(inode
->i_mode
)) {
598 struct cgroup
*cgrp
= dentry
->d_fsdata
;
599 struct cgroup_subsys
*ss
;
600 BUG_ON(!(cgroup_is_removed(cgrp
)));
601 /* It's possible for external users to be holding css
602 * reference counts on a cgroup; css_put() needs to
603 * be able to access the cgroup after decrementing
604 * the reference count in order to know if it needs to
605 * queue the cgroup to be handled by the release
609 mutex_lock(&cgroup_mutex
);
611 * Release the subsystem state objects.
613 for_each_subsys(cgrp
->root
, ss
)
614 ss
->destroy(ss
, cgrp
);
616 cgrp
->root
->number_of_cgroups
--;
617 mutex_unlock(&cgroup_mutex
);
619 /* Drop the active superblock reference that we took when we
620 * created the cgroup */
621 deactivate_super(cgrp
->root
->sb
);
628 static void remove_dir(struct dentry
*d
)
630 struct dentry
*parent
= dget(d
->d_parent
);
633 simple_rmdir(parent
->d_inode
, d
);
637 static void cgroup_clear_directory(struct dentry
*dentry
)
639 struct list_head
*node
;
641 BUG_ON(!mutex_is_locked(&dentry
->d_inode
->i_mutex
));
642 spin_lock(&dcache_lock
);
643 node
= dentry
->d_subdirs
.next
;
644 while (node
!= &dentry
->d_subdirs
) {
645 struct dentry
*d
= list_entry(node
, struct dentry
, d_u
.d_child
);
648 /* This should never be called on a cgroup
649 * directory with child cgroups */
650 BUG_ON(d
->d_inode
->i_mode
& S_IFDIR
);
652 spin_unlock(&dcache_lock
);
654 simple_unlink(dentry
->d_inode
, d
);
656 spin_lock(&dcache_lock
);
658 node
= dentry
->d_subdirs
.next
;
660 spin_unlock(&dcache_lock
);
664 * NOTE : the dentry must have been dget()'ed
666 static void cgroup_d_remove_dir(struct dentry
*dentry
)
668 cgroup_clear_directory(dentry
);
670 spin_lock(&dcache_lock
);
671 list_del_init(&dentry
->d_u
.d_child
);
672 spin_unlock(&dcache_lock
);
676 static int rebind_subsystems(struct cgroupfs_root
*root
,
677 unsigned long final_bits
)
679 unsigned long added_bits
, removed_bits
;
680 struct cgroup
*cgrp
= &root
->top_cgroup
;
683 removed_bits
= root
->actual_subsys_bits
& ~final_bits
;
684 added_bits
= final_bits
& ~root
->actual_subsys_bits
;
685 /* Check that any added subsystems are currently free */
686 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
687 unsigned long bit
= 1UL << i
;
688 struct cgroup_subsys
*ss
= subsys
[i
];
689 if (!(bit
& added_bits
))
691 if (ss
->root
!= &rootnode
) {
692 /* Subsystem isn't free */
697 /* Currently we don't handle adding/removing subsystems when
698 * any child cgroups exist. This is theoretically supportable
699 * but involves complex error handling, so it's being left until
701 if (root
->number_of_cgroups
> 1)
704 /* Process each subsystem */
705 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
706 struct cgroup_subsys
*ss
= subsys
[i
];
707 unsigned long bit
= 1UL << i
;
708 if (bit
& added_bits
) {
709 /* We're binding this subsystem to this hierarchy */
710 BUG_ON(cgrp
->subsys
[i
]);
711 BUG_ON(!dummytop
->subsys
[i
]);
712 BUG_ON(dummytop
->subsys
[i
]->cgroup
!= dummytop
);
713 cgrp
->subsys
[i
] = dummytop
->subsys
[i
];
714 cgrp
->subsys
[i
]->cgroup
= cgrp
;
715 list_move(&ss
->sibling
, &root
->subsys_list
);
720 } else if (bit
& removed_bits
) {
721 /* We're removing this subsystem */
722 BUG_ON(cgrp
->subsys
[i
] != dummytop
->subsys
[i
]);
723 BUG_ON(cgrp
->subsys
[i
]->cgroup
!= cgrp
);
725 ss
->bind(ss
, dummytop
);
726 dummytop
->subsys
[i
]->cgroup
= dummytop
;
727 cgrp
->subsys
[i
] = NULL
;
728 subsys
[i
]->root
= &rootnode
;
729 list_move(&ss
->sibling
, &rootnode
.subsys_list
);
730 } else if (bit
& final_bits
) {
731 /* Subsystem state should already exist */
732 BUG_ON(!cgrp
->subsys
[i
]);
734 /* Subsystem state shouldn't exist */
735 BUG_ON(cgrp
->subsys
[i
]);
738 root
->subsys_bits
= root
->actual_subsys_bits
= final_bits
;
744 static int cgroup_show_options(struct seq_file
*seq
, struct vfsmount
*vfs
)
746 struct cgroupfs_root
*root
= vfs
->mnt_sb
->s_fs_info
;
747 struct cgroup_subsys
*ss
;
749 mutex_lock(&cgroup_mutex
);
750 for_each_subsys(root
, ss
)
751 seq_printf(seq
, ",%s", ss
->name
);
752 if (test_bit(ROOT_NOPREFIX
, &root
->flags
))
753 seq_puts(seq
, ",noprefix");
754 if (strlen(root
->release_agent_path
))
755 seq_printf(seq
, ",release_agent=%s", root
->release_agent_path
);
756 mutex_unlock(&cgroup_mutex
);
760 struct cgroup_sb_opts
{
761 unsigned long subsys_bits
;
766 /* Convert a hierarchy specifier into a bitmask of subsystems and
768 static int parse_cgroupfs_options(char *data
,
769 struct cgroup_sb_opts
*opts
)
771 char *token
, *o
= data
?: "all";
773 opts
->subsys_bits
= 0;
775 opts
->release_agent
= NULL
;
777 while ((token
= strsep(&o
, ",")) != NULL
) {
780 if (!strcmp(token
, "all")) {
781 /* Add all non-disabled subsystems */
783 opts
->subsys_bits
= 0;
784 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
785 struct cgroup_subsys
*ss
= subsys
[i
];
787 opts
->subsys_bits
|= 1ul << i
;
789 } else if (!strcmp(token
, "noprefix")) {
790 set_bit(ROOT_NOPREFIX
, &opts
->flags
);
791 } else if (!strncmp(token
, "release_agent=", 14)) {
792 /* Specifying two release agents is forbidden */
793 if (opts
->release_agent
)
795 opts
->release_agent
= kzalloc(PATH_MAX
, GFP_KERNEL
);
796 if (!opts
->release_agent
)
798 strncpy(opts
->release_agent
, token
+ 14, PATH_MAX
- 1);
799 opts
->release_agent
[PATH_MAX
- 1] = 0;
801 struct cgroup_subsys
*ss
;
803 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
805 if (!strcmp(token
, ss
->name
)) {
807 set_bit(i
, &opts
->subsys_bits
);
811 if (i
== CGROUP_SUBSYS_COUNT
)
816 /* We can't have an empty hierarchy */
817 if (!opts
->subsys_bits
)
823 static int cgroup_remount(struct super_block
*sb
, int *flags
, char *data
)
826 struct cgroupfs_root
*root
= sb
->s_fs_info
;
827 struct cgroup
*cgrp
= &root
->top_cgroup
;
828 struct cgroup_sb_opts opts
;
830 mutex_lock(&cgrp
->dentry
->d_inode
->i_mutex
);
831 mutex_lock(&cgroup_mutex
);
833 /* See what subsystems are wanted */
834 ret
= parse_cgroupfs_options(data
, &opts
);
838 /* Don't allow flags to change at remount */
839 if (opts
.flags
!= root
->flags
) {
844 ret
= rebind_subsystems(root
, opts
.subsys_bits
);
846 /* (re)populate subsystem files */
848 cgroup_populate_dir(cgrp
);
850 if (opts
.release_agent
)
851 strcpy(root
->release_agent_path
, opts
.release_agent
);
853 if (opts
.release_agent
)
854 kfree(opts
.release_agent
);
855 mutex_unlock(&cgroup_mutex
);
856 mutex_unlock(&cgrp
->dentry
->d_inode
->i_mutex
);
860 static struct super_operations cgroup_ops
= {
861 .statfs
= simple_statfs
,
862 .drop_inode
= generic_delete_inode
,
863 .show_options
= cgroup_show_options
,
864 .remount_fs
= cgroup_remount
,
867 static void init_cgroup_housekeeping(struct cgroup
*cgrp
)
869 INIT_LIST_HEAD(&cgrp
->sibling
);
870 INIT_LIST_HEAD(&cgrp
->children
);
871 INIT_LIST_HEAD(&cgrp
->css_sets
);
872 INIT_LIST_HEAD(&cgrp
->release_list
);
873 init_rwsem(&cgrp
->pids_mutex
);
875 static void init_cgroup_root(struct cgroupfs_root
*root
)
877 struct cgroup
*cgrp
= &root
->top_cgroup
;
878 INIT_LIST_HEAD(&root
->subsys_list
);
879 INIT_LIST_HEAD(&root
->root_list
);
880 root
->number_of_cgroups
= 1;
882 cgrp
->top_cgroup
= cgrp
;
883 init_cgroup_housekeeping(cgrp
);
886 static int cgroup_test_super(struct super_block
*sb
, void *data
)
888 struct cgroupfs_root
*new = data
;
889 struct cgroupfs_root
*root
= sb
->s_fs_info
;
891 /* First check subsystems */
892 if (new->subsys_bits
!= root
->subsys_bits
)
895 /* Next check flags */
896 if (new->flags
!= root
->flags
)
902 static int cgroup_set_super(struct super_block
*sb
, void *data
)
905 struct cgroupfs_root
*root
= data
;
907 ret
= set_anon_super(sb
, NULL
);
911 sb
->s_fs_info
= root
;
914 sb
->s_blocksize
= PAGE_CACHE_SIZE
;
915 sb
->s_blocksize_bits
= PAGE_CACHE_SHIFT
;
916 sb
->s_magic
= CGROUP_SUPER_MAGIC
;
917 sb
->s_op
= &cgroup_ops
;
922 static int cgroup_get_rootdir(struct super_block
*sb
)
924 struct inode
*inode
=
925 cgroup_new_inode(S_IFDIR
| S_IRUGO
| S_IXUGO
| S_IWUSR
, sb
);
926 struct dentry
*dentry
;
931 inode
->i_fop
= &simple_dir_operations
;
932 inode
->i_op
= &cgroup_dir_inode_operations
;
933 /* directories start off with i_nlink == 2 (for "." entry) */
935 dentry
= d_alloc_root(inode
);
944 static int cgroup_get_sb(struct file_system_type
*fs_type
,
945 int flags
, const char *unused_dev_name
,
946 void *data
, struct vfsmount
*mnt
)
948 struct cgroup_sb_opts opts
;
950 struct super_block
*sb
;
951 struct cgroupfs_root
*root
;
952 struct list_head tmp_cg_links
;
954 /* First find the desired set of subsystems */
955 ret
= parse_cgroupfs_options(data
, &opts
);
957 if (opts
.release_agent
)
958 kfree(opts
.release_agent
);
962 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
964 if (opts
.release_agent
)
965 kfree(opts
.release_agent
);
969 init_cgroup_root(root
);
970 root
->subsys_bits
= opts
.subsys_bits
;
971 root
->flags
= opts
.flags
;
972 if (opts
.release_agent
) {
973 strcpy(root
->release_agent_path
, opts
.release_agent
);
974 kfree(opts
.release_agent
);
977 sb
= sget(fs_type
, cgroup_test_super
, cgroup_set_super
, root
);
984 if (sb
->s_fs_info
!= root
) {
985 /* Reusing an existing superblock */
986 BUG_ON(sb
->s_root
== NULL
);
991 struct cgroup
*cgrp
= &root
->top_cgroup
;
995 BUG_ON(sb
->s_root
!= NULL
);
997 ret
= cgroup_get_rootdir(sb
);
1000 inode
= sb
->s_root
->d_inode
;
1002 mutex_lock(&inode
->i_mutex
);
1003 mutex_lock(&cgroup_mutex
);
1006 * We're accessing css_set_count without locking
1007 * css_set_lock here, but that's OK - it can only be
1008 * increased by someone holding cgroup_lock, and
1009 * that's us. The worst that can happen is that we
1010 * have some link structures left over
1012 ret
= allocate_cg_links(css_set_count
, &tmp_cg_links
);
1014 mutex_unlock(&cgroup_mutex
);
1015 mutex_unlock(&inode
->i_mutex
);
1016 goto drop_new_super
;
1019 ret
= rebind_subsystems(root
, root
->subsys_bits
);
1020 if (ret
== -EBUSY
) {
1021 mutex_unlock(&cgroup_mutex
);
1022 mutex_unlock(&inode
->i_mutex
);
1026 /* EBUSY should be the only error here */
1029 list_add(&root
->root_list
, &roots
);
1032 sb
->s_root
->d_fsdata
= &root
->top_cgroup
;
1033 root
->top_cgroup
.dentry
= sb
->s_root
;
1035 /* Link the top cgroup in this hierarchy into all
1036 * the css_set objects */
1037 write_lock(&css_set_lock
);
1038 for (i
= 0; i
< CSS_SET_TABLE_SIZE
; i
++) {
1039 struct hlist_head
*hhead
= &css_set_table
[i
];
1040 struct hlist_node
*node
;
1043 hlist_for_each_entry(cg
, node
, hhead
, hlist
) {
1044 struct cg_cgroup_link
*link
;
1046 BUG_ON(list_empty(&tmp_cg_links
));
1047 link
= list_entry(tmp_cg_links
.next
,
1048 struct cg_cgroup_link
,
1050 list_del(&link
->cgrp_link_list
);
1052 list_add(&link
->cgrp_link_list
,
1053 &root
->top_cgroup
.css_sets
);
1054 list_add(&link
->cg_link_list
, &cg
->cg_links
);
1057 write_unlock(&css_set_lock
);
1059 free_cg_links(&tmp_cg_links
);
1061 BUG_ON(!list_empty(&cgrp
->sibling
));
1062 BUG_ON(!list_empty(&cgrp
->children
));
1063 BUG_ON(root
->number_of_cgroups
!= 1);
1065 cgroup_populate_dir(cgrp
);
1066 mutex_unlock(&inode
->i_mutex
);
1067 mutex_unlock(&cgroup_mutex
);
1070 return simple_set_mnt(mnt
, sb
);
1073 free_cg_links(&tmp_cg_links
);
1075 up_write(&sb
->s_umount
);
1076 deactivate_super(sb
);
1080 static void cgroup_kill_sb(struct super_block
*sb
) {
1081 struct cgroupfs_root
*root
= sb
->s_fs_info
;
1082 struct cgroup
*cgrp
= &root
->top_cgroup
;
1084 struct cg_cgroup_link
*link
;
1085 struct cg_cgroup_link
*saved_link
;
1089 BUG_ON(root
->number_of_cgroups
!= 1);
1090 BUG_ON(!list_empty(&cgrp
->children
));
1091 BUG_ON(!list_empty(&cgrp
->sibling
));
1093 mutex_lock(&cgroup_mutex
);
1095 /* Rebind all subsystems back to the default hierarchy */
1096 ret
= rebind_subsystems(root
, 0);
1097 /* Shouldn't be able to fail ... */
1101 * Release all the links from css_sets to this hierarchy's
1104 write_lock(&css_set_lock
);
1106 list_for_each_entry_safe(link
, saved_link
, &cgrp
->css_sets
,
1108 list_del(&link
->cg_link_list
);
1109 list_del(&link
->cgrp_link_list
);
1112 write_unlock(&css_set_lock
);
1114 list_del(&root
->root_list
);
1117 mutex_unlock(&cgroup_mutex
);
1120 kill_litter_super(sb
);
1123 static struct file_system_type cgroup_fs_type
= {
1125 .get_sb
= cgroup_get_sb
,
1126 .kill_sb
= cgroup_kill_sb
,
1129 static inline struct cgroup
*__d_cgrp(struct dentry
*dentry
)
1131 return dentry
->d_fsdata
;
1134 static inline struct cftype
*__d_cft(struct dentry
*dentry
)
1136 return dentry
->d_fsdata
;
1140 * cgroup_path - generate the path of a cgroup
1141 * @cgrp: the cgroup in question
1142 * @buf: the buffer to write the path into
1143 * @buflen: the length of the buffer
1145 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1146 * Returns 0 on success, -errno on error.
1148 int cgroup_path(const struct cgroup
*cgrp
, char *buf
, int buflen
)
1152 if (cgrp
== dummytop
) {
1154 * Inactive subsystems have no dentry for their root
1161 start
= buf
+ buflen
;
1165 int len
= cgrp
->dentry
->d_name
.len
;
1166 if ((start
-= len
) < buf
)
1167 return -ENAMETOOLONG
;
1168 memcpy(start
, cgrp
->dentry
->d_name
.name
, len
);
1169 cgrp
= cgrp
->parent
;
1175 return -ENAMETOOLONG
;
1178 memmove(buf
, start
, buf
+ buflen
- start
);
1183 * Return the first subsystem attached to a cgroup's hierarchy, and
1187 static void get_first_subsys(const struct cgroup
*cgrp
,
1188 struct cgroup_subsys_state
**css
, int *subsys_id
)
1190 const struct cgroupfs_root
*root
= cgrp
->root
;
1191 const struct cgroup_subsys
*test_ss
;
1192 BUG_ON(list_empty(&root
->subsys_list
));
1193 test_ss
= list_entry(root
->subsys_list
.next
,
1194 struct cgroup_subsys
, sibling
);
1196 *css
= cgrp
->subsys
[test_ss
->subsys_id
];
1200 *subsys_id
= test_ss
->subsys_id
;
1204 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1205 * @cgrp: the cgroup the task is attaching to
1206 * @tsk: the task to be attached
1208 * Call holding cgroup_mutex. May take task_lock of
1209 * the task 'tsk' during call.
1211 int cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
1214 struct cgroup_subsys
*ss
;
1215 struct cgroup
*oldcgrp
;
1217 struct css_set
*newcg
;
1218 struct cgroupfs_root
*root
= cgrp
->root
;
1221 get_first_subsys(cgrp
, NULL
, &subsys_id
);
1223 /* Nothing to do if the task is already in that cgroup */
1224 oldcgrp
= task_cgroup(tsk
, subsys_id
);
1225 if (cgrp
== oldcgrp
)
1228 for_each_subsys(root
, ss
) {
1229 if (ss
->can_attach
) {
1230 retval
= ss
->can_attach(ss
, cgrp
, tsk
);
1241 * Locate or allocate a new css_set for this task,
1242 * based on its final set of cgroups
1244 newcg
= find_css_set(cg
, cgrp
);
1250 if (tsk
->flags
& PF_EXITING
) {
1255 rcu_assign_pointer(tsk
->cgroups
, newcg
);
1258 /* Update the css_set linked lists if we're using them */
1259 write_lock(&css_set_lock
);
1260 if (!list_empty(&tsk
->cg_list
)) {
1261 list_del(&tsk
->cg_list
);
1262 list_add(&tsk
->cg_list
, &newcg
->tasks
);
1264 write_unlock(&css_set_lock
);
1266 for_each_subsys(root
, ss
) {
1268 ss
->attach(ss
, cgrp
, oldcgrp
, tsk
);
1270 set_bit(CGRP_RELEASABLE
, &oldcgrp
->flags
);
1277 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1278 * held. May take task_lock of task
1280 static int attach_task_by_pid(struct cgroup
*cgrp
, u64 pid
)
1282 struct task_struct
*tsk
;
1283 const struct cred
*cred
= current_cred(), *tcred
;
1288 tsk
= find_task_by_vpid(pid
);
1289 if (!tsk
|| tsk
->flags
& PF_EXITING
) {
1294 tcred
= __task_cred(tsk
);
1296 cred
->euid
!= tcred
->uid
&&
1297 cred
->euid
!= tcred
->suid
) {
1301 get_task_struct(tsk
);
1305 get_task_struct(tsk
);
1308 ret
= cgroup_attach_task(cgrp
, tsk
);
1309 put_task_struct(tsk
);
1313 static int cgroup_tasks_write(struct cgroup
*cgrp
, struct cftype
*cft
, u64 pid
)
1316 if (!cgroup_lock_live_group(cgrp
))
1318 ret
= attach_task_by_pid(cgrp
, pid
);
1323 /* The various types of files and directories in a cgroup file system */
1324 enum cgroup_filetype
{
1328 FILE_NOTIFY_ON_RELEASE
,
1333 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1334 * @cgrp: the cgroup to be checked for liveness
1336 * On success, returns true; the lock should be later released with
1337 * cgroup_unlock(). On failure returns false with no lock held.
1339 bool cgroup_lock_live_group(struct cgroup
*cgrp
)
1341 mutex_lock(&cgroup_mutex
);
1342 if (cgroup_is_removed(cgrp
)) {
1343 mutex_unlock(&cgroup_mutex
);
1349 static int cgroup_release_agent_write(struct cgroup
*cgrp
, struct cftype
*cft
,
1352 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
1353 if (!cgroup_lock_live_group(cgrp
))
1355 strcpy(cgrp
->root
->release_agent_path
, buffer
);
1360 static int cgroup_release_agent_show(struct cgroup
*cgrp
, struct cftype
*cft
,
1361 struct seq_file
*seq
)
1363 if (!cgroup_lock_live_group(cgrp
))
1365 seq_puts(seq
, cgrp
->root
->release_agent_path
);
1366 seq_putc(seq
, '\n');
1371 /* A buffer size big enough for numbers or short strings */
1372 #define CGROUP_LOCAL_BUFFER_SIZE 64
1374 static ssize_t
cgroup_write_X64(struct cgroup
*cgrp
, struct cftype
*cft
,
1376 const char __user
*userbuf
,
1377 size_t nbytes
, loff_t
*unused_ppos
)
1379 char buffer
[CGROUP_LOCAL_BUFFER_SIZE
];
1385 if (nbytes
>= sizeof(buffer
))
1387 if (copy_from_user(buffer
, userbuf
, nbytes
))
1390 buffer
[nbytes
] = 0; /* nul-terminate */
1392 if (cft
->write_u64
) {
1393 u64 val
= simple_strtoull(buffer
, &end
, 0);
1396 retval
= cft
->write_u64(cgrp
, cft
, val
);
1398 s64 val
= simple_strtoll(buffer
, &end
, 0);
1401 retval
= cft
->write_s64(cgrp
, cft
, val
);
1408 static ssize_t
cgroup_write_string(struct cgroup
*cgrp
, struct cftype
*cft
,
1410 const char __user
*userbuf
,
1411 size_t nbytes
, loff_t
*unused_ppos
)
1413 char local_buffer
[CGROUP_LOCAL_BUFFER_SIZE
];
1415 size_t max_bytes
= cft
->max_write_len
;
1416 char *buffer
= local_buffer
;
1419 max_bytes
= sizeof(local_buffer
) - 1;
1420 if (nbytes
>= max_bytes
)
1422 /* Allocate a dynamic buffer if we need one */
1423 if (nbytes
>= sizeof(local_buffer
)) {
1424 buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
);
1428 if (nbytes
&& copy_from_user(buffer
, userbuf
, nbytes
)) {
1433 buffer
[nbytes
] = 0; /* nul-terminate */
1435 retval
= cft
->write_string(cgrp
, cft
, buffer
);
1439 if (buffer
!= local_buffer
)
1444 static ssize_t
cgroup_file_write(struct file
*file
, const char __user
*buf
,
1445 size_t nbytes
, loff_t
*ppos
)
1447 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1448 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
1450 if (cgroup_is_removed(cgrp
))
1453 return cft
->write(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1454 if (cft
->write_u64
|| cft
->write_s64
)
1455 return cgroup_write_X64(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1456 if (cft
->write_string
)
1457 return cgroup_write_string(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1459 int ret
= cft
->trigger(cgrp
, (unsigned int)cft
->private);
1460 return ret
? ret
: nbytes
;
1465 static ssize_t
cgroup_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
,
1467 char __user
*buf
, size_t nbytes
,
1470 char tmp
[CGROUP_LOCAL_BUFFER_SIZE
];
1471 u64 val
= cft
->read_u64(cgrp
, cft
);
1472 int len
= sprintf(tmp
, "%llu\n", (unsigned long long) val
);
1474 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
1477 static ssize_t
cgroup_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
,
1479 char __user
*buf
, size_t nbytes
,
1482 char tmp
[CGROUP_LOCAL_BUFFER_SIZE
];
1483 s64 val
= cft
->read_s64(cgrp
, cft
);
1484 int len
= sprintf(tmp
, "%lld\n", (long long) val
);
1486 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
1489 static ssize_t
cgroup_file_read(struct file
*file
, char __user
*buf
,
1490 size_t nbytes
, loff_t
*ppos
)
1492 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1493 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
1495 if (cgroup_is_removed(cgrp
))
1499 return cft
->read(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1501 return cgroup_read_u64(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1503 return cgroup_read_s64(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1508 * seqfile ops/methods for returning structured data. Currently just
1509 * supports string->u64 maps, but can be extended in future.
1512 struct cgroup_seqfile_state
{
1514 struct cgroup
*cgroup
;
1517 static int cgroup_map_add(struct cgroup_map_cb
*cb
, const char *key
, u64 value
)
1519 struct seq_file
*sf
= cb
->state
;
1520 return seq_printf(sf
, "%s %llu\n", key
, (unsigned long long)value
);
1523 static int cgroup_seqfile_show(struct seq_file
*m
, void *arg
)
1525 struct cgroup_seqfile_state
*state
= m
->private;
1526 struct cftype
*cft
= state
->cft
;
1527 if (cft
->read_map
) {
1528 struct cgroup_map_cb cb
= {
1529 .fill
= cgroup_map_add
,
1532 return cft
->read_map(state
->cgroup
, cft
, &cb
);
1534 return cft
->read_seq_string(state
->cgroup
, cft
, m
);
1537 static int cgroup_seqfile_release(struct inode
*inode
, struct file
*file
)
1539 struct seq_file
*seq
= file
->private_data
;
1540 kfree(seq
->private);
1541 return single_release(inode
, file
);
1544 static struct file_operations cgroup_seqfile_operations
= {
1546 .write
= cgroup_file_write
,
1547 .llseek
= seq_lseek
,
1548 .release
= cgroup_seqfile_release
,
1551 static int cgroup_file_open(struct inode
*inode
, struct file
*file
)
1556 err
= generic_file_open(inode
, file
);
1559 cft
= __d_cft(file
->f_dentry
);
1561 if (cft
->read_map
|| cft
->read_seq_string
) {
1562 struct cgroup_seqfile_state
*state
=
1563 kzalloc(sizeof(*state
), GFP_USER
);
1567 state
->cgroup
= __d_cgrp(file
->f_dentry
->d_parent
);
1568 file
->f_op
= &cgroup_seqfile_operations
;
1569 err
= single_open(file
, cgroup_seqfile_show
, state
);
1572 } else if (cft
->open
)
1573 err
= cft
->open(inode
, file
);
1580 static int cgroup_file_release(struct inode
*inode
, struct file
*file
)
1582 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1584 return cft
->release(inode
, file
);
1589 * cgroup_rename - Only allow simple rename of directories in place.
1591 static int cgroup_rename(struct inode
*old_dir
, struct dentry
*old_dentry
,
1592 struct inode
*new_dir
, struct dentry
*new_dentry
)
1594 if (!S_ISDIR(old_dentry
->d_inode
->i_mode
))
1596 if (new_dentry
->d_inode
)
1598 if (old_dir
!= new_dir
)
1600 return simple_rename(old_dir
, old_dentry
, new_dir
, new_dentry
);
1603 static struct file_operations cgroup_file_operations
= {
1604 .read
= cgroup_file_read
,
1605 .write
= cgroup_file_write
,
1606 .llseek
= generic_file_llseek
,
1607 .open
= cgroup_file_open
,
1608 .release
= cgroup_file_release
,
1611 static struct inode_operations cgroup_dir_inode_operations
= {
1612 .lookup
= simple_lookup
,
1613 .mkdir
= cgroup_mkdir
,
1614 .rmdir
= cgroup_rmdir
,
1615 .rename
= cgroup_rename
,
1618 static int cgroup_create_file(struct dentry
*dentry
, int mode
,
1619 struct super_block
*sb
)
1621 static struct dentry_operations cgroup_dops
= {
1622 .d_iput
= cgroup_diput
,
1625 struct inode
*inode
;
1629 if (dentry
->d_inode
)
1632 inode
= cgroup_new_inode(mode
, sb
);
1636 if (S_ISDIR(mode
)) {
1637 inode
->i_op
= &cgroup_dir_inode_operations
;
1638 inode
->i_fop
= &simple_dir_operations
;
1640 /* start off with i_nlink == 2 (for "." entry) */
1643 /* start with the directory inode held, so that we can
1644 * populate it without racing with another mkdir */
1645 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_CHILD
);
1646 } else if (S_ISREG(mode
)) {
1648 inode
->i_fop
= &cgroup_file_operations
;
1650 dentry
->d_op
= &cgroup_dops
;
1651 d_instantiate(dentry
, inode
);
1652 dget(dentry
); /* Extra count - pin the dentry in core */
1657 * cgroup_create_dir - create a directory for an object.
1658 * @cgrp: the cgroup we create the directory for. It must have a valid
1659 * ->parent field. And we are going to fill its ->dentry field.
1660 * @dentry: dentry of the new cgroup
1661 * @mode: mode to set on new directory.
1663 static int cgroup_create_dir(struct cgroup
*cgrp
, struct dentry
*dentry
,
1666 struct dentry
*parent
;
1669 parent
= cgrp
->parent
->dentry
;
1670 error
= cgroup_create_file(dentry
, S_IFDIR
| mode
, cgrp
->root
->sb
);
1672 dentry
->d_fsdata
= cgrp
;
1673 inc_nlink(parent
->d_inode
);
1674 cgrp
->dentry
= dentry
;
1682 int cgroup_add_file(struct cgroup
*cgrp
,
1683 struct cgroup_subsys
*subsys
,
1684 const struct cftype
*cft
)
1686 struct dentry
*dir
= cgrp
->dentry
;
1687 struct dentry
*dentry
;
1690 char name
[MAX_CGROUP_TYPE_NAMELEN
+ MAX_CFTYPE_NAME
+ 2] = { 0 };
1691 if (subsys
&& !test_bit(ROOT_NOPREFIX
, &cgrp
->root
->flags
)) {
1692 strcpy(name
, subsys
->name
);
1695 strcat(name
, cft
->name
);
1696 BUG_ON(!mutex_is_locked(&dir
->d_inode
->i_mutex
));
1697 dentry
= lookup_one_len(name
, dir
, strlen(name
));
1698 if (!IS_ERR(dentry
)) {
1699 error
= cgroup_create_file(dentry
, 0644 | S_IFREG
,
1702 dentry
->d_fsdata
= (void *)cft
;
1705 error
= PTR_ERR(dentry
);
1709 int cgroup_add_files(struct cgroup
*cgrp
,
1710 struct cgroup_subsys
*subsys
,
1711 const struct cftype cft
[],
1715 for (i
= 0; i
< count
; i
++) {
1716 err
= cgroup_add_file(cgrp
, subsys
, &cft
[i
]);
1724 * cgroup_task_count - count the number of tasks in a cgroup.
1725 * @cgrp: the cgroup in question
1727 * Return the number of tasks in the cgroup.
1729 int cgroup_task_count(const struct cgroup
*cgrp
)
1732 struct cg_cgroup_link
*link
;
1734 read_lock(&css_set_lock
);
1735 list_for_each_entry(link
, &cgrp
->css_sets
, cgrp_link_list
) {
1736 count
+= atomic_read(&link
->cg
->refcount
);
1738 read_unlock(&css_set_lock
);
1743 * Advance a list_head iterator. The iterator should be positioned at
1744 * the start of a css_set
1746 static void cgroup_advance_iter(struct cgroup
*cgrp
,
1747 struct cgroup_iter
*it
)
1749 struct list_head
*l
= it
->cg_link
;
1750 struct cg_cgroup_link
*link
;
1753 /* Advance to the next non-empty css_set */
1756 if (l
== &cgrp
->css_sets
) {
1760 link
= list_entry(l
, struct cg_cgroup_link
, cgrp_link_list
);
1762 } while (list_empty(&cg
->tasks
));
1764 it
->task
= cg
->tasks
.next
;
1768 * To reduce the fork() overhead for systems that are not actually
1769 * using their cgroups capability, we don't maintain the lists running
1770 * through each css_set to its tasks until we see the list actually
1771 * used - in other words after the first call to cgroup_iter_start().
1773 * The tasklist_lock is not held here, as do_each_thread() and
1774 * while_each_thread() are protected by RCU.
1776 static void cgroup_enable_task_cg_lists(void)
1778 struct task_struct
*p
, *g
;
1779 write_lock(&css_set_lock
);
1780 use_task_css_set_links
= 1;
1781 do_each_thread(g
, p
) {
1784 * We should check if the process is exiting, otherwise
1785 * it will race with cgroup_exit() in that the list
1786 * entry won't be deleted though the process has exited.
1788 if (!(p
->flags
& PF_EXITING
) && list_empty(&p
->cg_list
))
1789 list_add(&p
->cg_list
, &p
->cgroups
->tasks
);
1791 } while_each_thread(g
, p
);
1792 write_unlock(&css_set_lock
);
1795 void cgroup_iter_start(struct cgroup
*cgrp
, struct cgroup_iter
*it
)
1798 * The first time anyone tries to iterate across a cgroup,
1799 * we need to enable the list linking each css_set to its
1800 * tasks, and fix up all existing tasks.
1802 if (!use_task_css_set_links
)
1803 cgroup_enable_task_cg_lists();
1805 read_lock(&css_set_lock
);
1806 it
->cg_link
= &cgrp
->css_sets
;
1807 cgroup_advance_iter(cgrp
, it
);
1810 struct task_struct
*cgroup_iter_next(struct cgroup
*cgrp
,
1811 struct cgroup_iter
*it
)
1813 struct task_struct
*res
;
1814 struct list_head
*l
= it
->task
;
1815 struct cg_cgroup_link
*link
;
1817 /* If the iterator cg is NULL, we have no tasks */
1820 res
= list_entry(l
, struct task_struct
, cg_list
);
1821 /* Advance iterator to find next entry */
1823 link
= list_entry(it
->cg_link
, struct cg_cgroup_link
, cgrp_link_list
);
1824 if (l
== &link
->cg
->tasks
) {
1825 /* We reached the end of this task list - move on to
1826 * the next cg_cgroup_link */
1827 cgroup_advance_iter(cgrp
, it
);
1834 void cgroup_iter_end(struct cgroup
*cgrp
, struct cgroup_iter
*it
)
1836 read_unlock(&css_set_lock
);
1839 static inline int started_after_time(struct task_struct
*t1
,
1840 struct timespec
*time
,
1841 struct task_struct
*t2
)
1843 int start_diff
= timespec_compare(&t1
->start_time
, time
);
1844 if (start_diff
> 0) {
1846 } else if (start_diff
< 0) {
1850 * Arbitrarily, if two processes started at the same
1851 * time, we'll say that the lower pointer value
1852 * started first. Note that t2 may have exited by now
1853 * so this may not be a valid pointer any longer, but
1854 * that's fine - it still serves to distinguish
1855 * between two tasks started (effectively) simultaneously.
1862 * This function is a callback from heap_insert() and is used to order
1864 * In this case we order the heap in descending task start time.
1866 static inline int started_after(void *p1
, void *p2
)
1868 struct task_struct
*t1
= p1
;
1869 struct task_struct
*t2
= p2
;
1870 return started_after_time(t1
, &t2
->start_time
, t2
);
1874 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
1875 * @scan: struct cgroup_scanner containing arguments for the scan
1877 * Arguments include pointers to callback functions test_task() and
1879 * Iterate through all the tasks in a cgroup, calling test_task() for each,
1880 * and if it returns true, call process_task() for it also.
1881 * The test_task pointer may be NULL, meaning always true (select all tasks).
1882 * Effectively duplicates cgroup_iter_{start,next,end}()
1883 * but does not lock css_set_lock for the call to process_task().
1884 * The struct cgroup_scanner may be embedded in any structure of the caller's
1886 * It is guaranteed that process_task() will act on every task that
1887 * is a member of the cgroup for the duration of this call. This
1888 * function may or may not call process_task() for tasks that exit
1889 * or move to a different cgroup during the call, or are forked or
1890 * move into the cgroup during the call.
1892 * Note that test_task() may be called with locks held, and may in some
1893 * situations be called multiple times for the same task, so it should
1895 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
1896 * pre-allocated and will be used for heap operations (and its "gt" member will
1897 * be overwritten), else a temporary heap will be used (allocation of which
1898 * may cause this function to fail).
1900 int cgroup_scan_tasks(struct cgroup_scanner
*scan
)
1903 struct cgroup_iter it
;
1904 struct task_struct
*p
, *dropped
;
1905 /* Never dereference latest_task, since it's not refcounted */
1906 struct task_struct
*latest_task
= NULL
;
1907 struct ptr_heap tmp_heap
;
1908 struct ptr_heap
*heap
;
1909 struct timespec latest_time
= { 0, 0 };
1912 /* The caller supplied our heap and pre-allocated its memory */
1914 heap
->gt
= &started_after
;
1916 /* We need to allocate our own heap memory */
1918 retval
= heap_init(heap
, PAGE_SIZE
, GFP_KERNEL
, &started_after
);
1920 /* cannot allocate the heap */
1926 * Scan tasks in the cgroup, using the scanner's "test_task" callback
1927 * to determine which are of interest, and using the scanner's
1928 * "process_task" callback to process any of them that need an update.
1929 * Since we don't want to hold any locks during the task updates,
1930 * gather tasks to be processed in a heap structure.
1931 * The heap is sorted by descending task start time.
1932 * If the statically-sized heap fills up, we overflow tasks that
1933 * started later, and in future iterations only consider tasks that
1934 * started after the latest task in the previous pass. This
1935 * guarantees forward progress and that we don't miss any tasks.
1938 cgroup_iter_start(scan
->cg
, &it
);
1939 while ((p
= cgroup_iter_next(scan
->cg
, &it
))) {
1941 * Only affect tasks that qualify per the caller's callback,
1942 * if he provided one
1944 if (scan
->test_task
&& !scan
->test_task(p
, scan
))
1947 * Only process tasks that started after the last task
1950 if (!started_after_time(p
, &latest_time
, latest_task
))
1952 dropped
= heap_insert(heap
, p
);
1953 if (dropped
== NULL
) {
1955 * The new task was inserted; the heap wasn't
1959 } else if (dropped
!= p
) {
1961 * The new task was inserted, and pushed out a
1965 put_task_struct(dropped
);
1968 * Else the new task was newer than anything already in
1969 * the heap and wasn't inserted
1972 cgroup_iter_end(scan
->cg
, &it
);
1975 for (i
= 0; i
< heap
->size
; i
++) {
1976 struct task_struct
*q
= heap
->ptrs
[i
];
1978 latest_time
= q
->start_time
;
1981 /* Process the task per the caller's callback */
1982 scan
->process_task(q
, scan
);
1986 * If we had to process any tasks at all, scan again
1987 * in case some of them were in the middle of forking
1988 * children that didn't get processed.
1989 * Not the most efficient way to do it, but it avoids
1990 * having to take callback_mutex in the fork path
1994 if (heap
== &tmp_heap
)
1995 heap_free(&tmp_heap
);
2000 * Stuff for reading the 'tasks' file.
2002 * Reading this file can return large amounts of data if a cgroup has
2003 * *lots* of attached tasks. So it may need several calls to read(),
2004 * but we cannot guarantee that the information we produce is correct
2005 * unless we produce it entirely atomically.
2010 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
2011 * 'cgrp'. Return actual number of pids loaded. No need to
2012 * task_lock(p) when reading out p->cgroup, since we're in an RCU
2013 * read section, so the css_set can't go away, and is
2014 * immutable after creation.
2016 static int pid_array_load(pid_t
*pidarray
, int npids
, struct cgroup
*cgrp
)
2019 struct cgroup_iter it
;
2020 struct task_struct
*tsk
;
2021 cgroup_iter_start(cgrp
, &it
);
2022 while ((tsk
= cgroup_iter_next(cgrp
, &it
))) {
2023 if (unlikely(n
== npids
))
2025 pidarray
[n
++] = task_pid_vnr(tsk
);
2027 cgroup_iter_end(cgrp
, &it
);
2032 * cgroupstats_build - build and fill cgroupstats
2033 * @stats: cgroupstats to fill information into
2034 * @dentry: A dentry entry belonging to the cgroup for which stats have
2037 * Build and fill cgroupstats so that taskstats can export it to user
2040 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
2043 struct cgroup
*cgrp
;
2044 struct cgroup_iter it
;
2045 struct task_struct
*tsk
;
2048 * Validate dentry by checking the superblock operations,
2049 * and make sure it's a directory.
2051 if (dentry
->d_sb
->s_op
!= &cgroup_ops
||
2052 !S_ISDIR(dentry
->d_inode
->i_mode
))
2056 cgrp
= dentry
->d_fsdata
;
2058 cgroup_iter_start(cgrp
, &it
);
2059 while ((tsk
= cgroup_iter_next(cgrp
, &it
))) {
2060 switch (tsk
->state
) {
2062 stats
->nr_running
++;
2064 case TASK_INTERRUPTIBLE
:
2065 stats
->nr_sleeping
++;
2067 case TASK_UNINTERRUPTIBLE
:
2068 stats
->nr_uninterruptible
++;
2071 stats
->nr_stopped
++;
2074 if (delayacct_is_task_waiting_on_io(tsk
))
2075 stats
->nr_io_wait
++;
2079 cgroup_iter_end(cgrp
, &it
);
2085 static int cmppid(const void *a
, const void *b
)
2087 return *(pid_t
*)a
- *(pid_t
*)b
;
2092 * seq_file methods for the "tasks" file. The seq_file position is the
2093 * next pid to display; the seq_file iterator is a pointer to the pid
2094 * in the cgroup->tasks_pids array.
2097 static void *cgroup_tasks_start(struct seq_file
*s
, loff_t
*pos
)
2100 * Initially we receive a position value that corresponds to
2101 * one more than the last pid shown (or 0 on the first call or
2102 * after a seek to the start). Use a binary-search to find the
2103 * next pid to display, if any
2105 struct cgroup
*cgrp
= s
->private;
2106 int index
= 0, pid
= *pos
;
2109 down_read(&cgrp
->pids_mutex
);
2111 int end
= cgrp
->pids_length
;
2113 while (index
< end
) {
2114 int mid
= (index
+ end
) / 2;
2115 if (cgrp
->tasks_pids
[mid
] == pid
) {
2118 } else if (cgrp
->tasks_pids
[mid
] <= pid
)
2124 /* If we're off the end of the array, we're done */
2125 if (index
>= cgrp
->pids_length
)
2127 /* Update the abstract position to be the actual pid that we found */
2128 iter
= cgrp
->tasks_pids
+ index
;
2133 static void cgroup_tasks_stop(struct seq_file
*s
, void *v
)
2135 struct cgroup
*cgrp
= s
->private;
2136 up_read(&cgrp
->pids_mutex
);
2139 static void *cgroup_tasks_next(struct seq_file
*s
, void *v
, loff_t
*pos
)
2141 struct cgroup
*cgrp
= s
->private;
2143 int *end
= cgrp
->tasks_pids
+ cgrp
->pids_length
;
2146 * Advance to the next pid in the array. If this goes off the
2158 static int cgroup_tasks_show(struct seq_file
*s
, void *v
)
2160 return seq_printf(s
, "%d\n", *(int *)v
);
2163 static struct seq_operations cgroup_tasks_seq_operations
= {
2164 .start
= cgroup_tasks_start
,
2165 .stop
= cgroup_tasks_stop
,
2166 .next
= cgroup_tasks_next
,
2167 .show
= cgroup_tasks_show
,
2170 static void release_cgroup_pid_array(struct cgroup
*cgrp
)
2172 down_write(&cgrp
->pids_mutex
);
2173 BUG_ON(!cgrp
->pids_use_count
);
2174 if (!--cgrp
->pids_use_count
) {
2175 kfree(cgrp
->tasks_pids
);
2176 cgrp
->tasks_pids
= NULL
;
2177 cgrp
->pids_length
= 0;
2179 up_write(&cgrp
->pids_mutex
);
2182 static int cgroup_tasks_release(struct inode
*inode
, struct file
*file
)
2184 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
2186 if (!(file
->f_mode
& FMODE_READ
))
2189 release_cgroup_pid_array(cgrp
);
2190 return seq_release(inode
, file
);
2193 static struct file_operations cgroup_tasks_operations
= {
2195 .llseek
= seq_lseek
,
2196 .write
= cgroup_file_write
,
2197 .release
= cgroup_tasks_release
,
2201 * Handle an open on 'tasks' file. Prepare an array containing the
2202 * process id's of tasks currently attached to the cgroup being opened.
2205 static int cgroup_tasks_open(struct inode
*unused
, struct file
*file
)
2207 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
2212 /* Nothing to do for write-only files */
2213 if (!(file
->f_mode
& FMODE_READ
))
2217 * If cgroup gets more users after we read count, we won't have
2218 * enough space - tough. This race is indistinguishable to the
2219 * caller from the case that the additional cgroup users didn't
2220 * show up until sometime later on.
2222 npids
= cgroup_task_count(cgrp
);
2223 pidarray
= kmalloc(npids
* sizeof(pid_t
), GFP_KERNEL
);
2226 npids
= pid_array_load(pidarray
, npids
, cgrp
);
2227 sort(pidarray
, npids
, sizeof(pid_t
), cmppid
, NULL
);
2230 * Store the array in the cgroup, freeing the old
2231 * array if necessary
2233 down_write(&cgrp
->pids_mutex
);
2234 kfree(cgrp
->tasks_pids
);
2235 cgrp
->tasks_pids
= pidarray
;
2236 cgrp
->pids_length
= npids
;
2237 cgrp
->pids_use_count
++;
2238 up_write(&cgrp
->pids_mutex
);
2240 file
->f_op
= &cgroup_tasks_operations
;
2242 retval
= seq_open(file
, &cgroup_tasks_seq_operations
);
2244 release_cgroup_pid_array(cgrp
);
2247 ((struct seq_file
*)file
->private_data
)->private = cgrp
;
2251 static u64
cgroup_read_notify_on_release(struct cgroup
*cgrp
,
2254 return notify_on_release(cgrp
);
2257 static int cgroup_write_notify_on_release(struct cgroup
*cgrp
,
2261 clear_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
2263 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
2265 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
2270 * for the common functions, 'private' gives the type of file
2272 static struct cftype files
[] = {
2275 .open
= cgroup_tasks_open
,
2276 .write_u64
= cgroup_tasks_write
,
2277 .release
= cgroup_tasks_release
,
2278 .private = FILE_TASKLIST
,
2282 .name
= "notify_on_release",
2283 .read_u64
= cgroup_read_notify_on_release
,
2284 .write_u64
= cgroup_write_notify_on_release
,
2285 .private = FILE_NOTIFY_ON_RELEASE
,
2289 static struct cftype cft_release_agent
= {
2290 .name
= "release_agent",
2291 .read_seq_string
= cgroup_release_agent_show
,
2292 .write_string
= cgroup_release_agent_write
,
2293 .max_write_len
= PATH_MAX
,
2294 .private = FILE_RELEASE_AGENT
,
2297 static int cgroup_populate_dir(struct cgroup
*cgrp
)
2300 struct cgroup_subsys
*ss
;
2302 /* First clear out any existing files */
2303 cgroup_clear_directory(cgrp
->dentry
);
2305 err
= cgroup_add_files(cgrp
, NULL
, files
, ARRAY_SIZE(files
));
2309 if (cgrp
== cgrp
->top_cgroup
) {
2310 if ((err
= cgroup_add_file(cgrp
, NULL
, &cft_release_agent
)) < 0)
2314 for_each_subsys(cgrp
->root
, ss
) {
2315 if (ss
->populate
&& (err
= ss
->populate(ss
, cgrp
)) < 0)
2322 static void init_cgroup_css(struct cgroup_subsys_state
*css
,
2323 struct cgroup_subsys
*ss
,
2324 struct cgroup
*cgrp
)
2327 atomic_set(&css
->refcnt
, 0);
2329 if (cgrp
== dummytop
)
2330 set_bit(CSS_ROOT
, &css
->flags
);
2331 BUG_ON(cgrp
->subsys
[ss
->subsys_id
]);
2332 cgrp
->subsys
[ss
->subsys_id
] = css
;
2336 * cgroup_create - create a cgroup
2337 * @parent: cgroup that will be parent of the new cgroup
2338 * @dentry: dentry of the new cgroup
2339 * @mode: mode to set on new inode
2341 * Must be called with the mutex on the parent inode held
2343 static long cgroup_create(struct cgroup
*parent
, struct dentry
*dentry
,
2346 struct cgroup
*cgrp
;
2347 struct cgroupfs_root
*root
= parent
->root
;
2349 struct cgroup_subsys
*ss
;
2350 struct super_block
*sb
= root
->sb
;
2352 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
2356 /* Grab a reference on the superblock so the hierarchy doesn't
2357 * get deleted on unmount if there are child cgroups. This
2358 * can be done outside cgroup_mutex, since the sb can't
2359 * disappear while someone has an open control file on the
2361 atomic_inc(&sb
->s_active
);
2363 mutex_lock(&cgroup_mutex
);
2365 init_cgroup_housekeeping(cgrp
);
2367 cgrp
->parent
= parent
;
2368 cgrp
->root
= parent
->root
;
2369 cgrp
->top_cgroup
= parent
->top_cgroup
;
2371 if (notify_on_release(parent
))
2372 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
2374 for_each_subsys(root
, ss
) {
2375 struct cgroup_subsys_state
*css
= ss
->create(ss
, cgrp
);
2380 init_cgroup_css(css
, ss
, cgrp
);
2383 list_add(&cgrp
->sibling
, &cgrp
->parent
->children
);
2384 root
->number_of_cgroups
++;
2386 err
= cgroup_create_dir(cgrp
, dentry
, mode
);
2390 /* The cgroup directory was pre-locked for us */
2391 BUG_ON(!mutex_is_locked(&cgrp
->dentry
->d_inode
->i_mutex
));
2393 err
= cgroup_populate_dir(cgrp
);
2394 /* If err < 0, we have a half-filled directory - oh well ;) */
2396 mutex_unlock(&cgroup_mutex
);
2397 mutex_unlock(&cgrp
->dentry
->d_inode
->i_mutex
);
2403 list_del(&cgrp
->sibling
);
2404 root
->number_of_cgroups
--;
2408 for_each_subsys(root
, ss
) {
2409 if (cgrp
->subsys
[ss
->subsys_id
])
2410 ss
->destroy(ss
, cgrp
);
2413 mutex_unlock(&cgroup_mutex
);
2415 /* Release the reference count that we took on the superblock */
2416 deactivate_super(sb
);
2422 static int cgroup_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
)
2424 struct cgroup
*c_parent
= dentry
->d_parent
->d_fsdata
;
2426 /* the vfs holds inode->i_mutex already */
2427 return cgroup_create(c_parent
, dentry
, mode
| S_IFDIR
);
2430 static int cgroup_has_css_refs(struct cgroup
*cgrp
)
2432 /* Check the reference count on each subsystem. Since we
2433 * already established that there are no tasks in the
2434 * cgroup, if the css refcount is also 0, then there should
2435 * be no outstanding references, so the subsystem is safe to
2436 * destroy. We scan across all subsystems rather than using
2437 * the per-hierarchy linked list of mounted subsystems since
2438 * we can be called via check_for_release() with no
2439 * synchronization other than RCU, and the subsystem linked
2440 * list isn't RCU-safe */
2442 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2443 struct cgroup_subsys
*ss
= subsys
[i
];
2444 struct cgroup_subsys_state
*css
;
2445 /* Skip subsystems not in this hierarchy */
2446 if (ss
->root
!= cgrp
->root
)
2448 css
= cgrp
->subsys
[ss
->subsys_id
];
2449 /* When called from check_for_release() it's possible
2450 * that by this point the cgroup has been removed
2451 * and the css deleted. But a false-positive doesn't
2452 * matter, since it can only happen if the cgroup
2453 * has been deleted and hence no longer needs the
2454 * release agent to be called anyway. */
2455 if (css
&& atomic_read(&css
->refcnt
))
2461 static int cgroup_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
)
2463 struct cgroup
*cgrp
= dentry
->d_fsdata
;
2465 struct cgroup
*parent
;
2467 /* the vfs holds both inode->i_mutex already */
2469 mutex_lock(&cgroup_mutex
);
2470 if (atomic_read(&cgrp
->count
) != 0) {
2471 mutex_unlock(&cgroup_mutex
);
2474 if (!list_empty(&cgrp
->children
)) {
2475 mutex_unlock(&cgroup_mutex
);
2478 mutex_unlock(&cgroup_mutex
);
2481 * Call pre_destroy handlers of subsys. Notify subsystems
2482 * that rmdir() request comes.
2484 cgroup_call_pre_destroy(cgrp
);
2486 mutex_lock(&cgroup_mutex
);
2487 parent
= cgrp
->parent
;
2489 if (atomic_read(&cgrp
->count
)
2490 || !list_empty(&cgrp
->children
)
2491 || cgroup_has_css_refs(cgrp
)) {
2492 mutex_unlock(&cgroup_mutex
);
2496 spin_lock(&release_list_lock
);
2497 set_bit(CGRP_REMOVED
, &cgrp
->flags
);
2498 if (!list_empty(&cgrp
->release_list
))
2499 list_del(&cgrp
->release_list
);
2500 spin_unlock(&release_list_lock
);
2501 /* delete my sibling from parent->children */
2502 list_del(&cgrp
->sibling
);
2503 spin_lock(&cgrp
->dentry
->d_lock
);
2504 d
= dget(cgrp
->dentry
);
2505 spin_unlock(&d
->d_lock
);
2507 cgroup_d_remove_dir(d
);
2510 set_bit(CGRP_RELEASABLE
, &parent
->flags
);
2511 check_for_release(parent
);
2513 mutex_unlock(&cgroup_mutex
);
2517 static void __init
cgroup_init_subsys(struct cgroup_subsys
*ss
)
2519 struct cgroup_subsys_state
*css
;
2521 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
2523 /* Create the top cgroup state for this subsystem */
2524 list_add(&ss
->sibling
, &rootnode
.subsys_list
);
2525 ss
->root
= &rootnode
;
2526 css
= ss
->create(ss
, dummytop
);
2527 /* We don't handle early failures gracefully */
2528 BUG_ON(IS_ERR(css
));
2529 init_cgroup_css(css
, ss
, dummytop
);
2531 /* Update the init_css_set to contain a subsys
2532 * pointer to this state - since the subsystem is
2533 * newly registered, all tasks and hence the
2534 * init_css_set is in the subsystem's top cgroup. */
2535 init_css_set
.subsys
[ss
->subsys_id
] = dummytop
->subsys
[ss
->subsys_id
];
2537 need_forkexit_callback
|= ss
->fork
|| ss
->exit
;
2539 /* At system boot, before all subsystems have been
2540 * registered, no tasks have been forked, so we don't
2541 * need to invoke fork callbacks here. */
2542 BUG_ON(!list_empty(&init_task
.tasks
));
2548 * cgroup_init_early - cgroup initialization at system boot
2550 * Initialize cgroups at system boot, and initialize any
2551 * subsystems that request early init.
2553 int __init
cgroup_init_early(void)
2556 atomic_set(&init_css_set
.refcount
, 1);
2557 INIT_LIST_HEAD(&init_css_set
.cg_links
);
2558 INIT_LIST_HEAD(&init_css_set
.tasks
);
2559 INIT_HLIST_NODE(&init_css_set
.hlist
);
2561 init_cgroup_root(&rootnode
);
2563 init_task
.cgroups
= &init_css_set
;
2565 init_css_set_link
.cg
= &init_css_set
;
2566 list_add(&init_css_set_link
.cgrp_link_list
,
2567 &rootnode
.top_cgroup
.css_sets
);
2568 list_add(&init_css_set_link
.cg_link_list
,
2569 &init_css_set
.cg_links
);
2571 for (i
= 0; i
< CSS_SET_TABLE_SIZE
; i
++)
2572 INIT_HLIST_HEAD(&css_set_table
[i
]);
2574 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2575 struct cgroup_subsys
*ss
= subsys
[i
];
2578 BUG_ON(strlen(ss
->name
) > MAX_CGROUP_TYPE_NAMELEN
);
2579 BUG_ON(!ss
->create
);
2580 BUG_ON(!ss
->destroy
);
2581 if (ss
->subsys_id
!= i
) {
2582 printk(KERN_ERR
"cgroup: Subsys %s id == %d\n",
2583 ss
->name
, ss
->subsys_id
);
2588 cgroup_init_subsys(ss
);
2594 * cgroup_init - cgroup initialization
2596 * Register cgroup filesystem and /proc file, and initialize
2597 * any subsystems that didn't request early init.
2599 int __init
cgroup_init(void)
2603 struct hlist_head
*hhead
;
2605 err
= bdi_init(&cgroup_backing_dev_info
);
2609 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2610 struct cgroup_subsys
*ss
= subsys
[i
];
2611 if (!ss
->early_init
)
2612 cgroup_init_subsys(ss
);
2615 /* Add init_css_set to the hash table */
2616 hhead
= css_set_hash(init_css_set
.subsys
);
2617 hlist_add_head(&init_css_set
.hlist
, hhead
);
2619 err
= register_filesystem(&cgroup_fs_type
);
2623 proc_create("cgroups", 0, NULL
, &proc_cgroupstats_operations
);
2627 bdi_destroy(&cgroup_backing_dev_info
);
2633 * proc_cgroup_show()
2634 * - Print task's cgroup paths into seq_file, one line for each hierarchy
2635 * - Used for /proc/<pid>/cgroup.
2636 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2637 * doesn't really matter if tsk->cgroup changes after we read it,
2638 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
2639 * anyway. No need to check that tsk->cgroup != NULL, thanks to
2640 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2641 * cgroup to top_cgroup.
2644 /* TODO: Use a proper seq_file iterator */
2645 static int proc_cgroup_show(struct seq_file
*m
, void *v
)
2648 struct task_struct
*tsk
;
2651 struct cgroupfs_root
*root
;
2654 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2660 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2666 mutex_lock(&cgroup_mutex
);
2668 for_each_active_root(root
) {
2669 struct cgroup_subsys
*ss
;
2670 struct cgroup
*cgrp
;
2674 seq_printf(m
, "%lu:", root
->subsys_bits
);
2675 for_each_subsys(root
, ss
)
2676 seq_printf(m
, "%s%s", count
++ ? "," : "", ss
->name
);
2678 get_first_subsys(&root
->top_cgroup
, NULL
, &subsys_id
);
2679 cgrp
= task_cgroup(tsk
, subsys_id
);
2680 retval
= cgroup_path(cgrp
, buf
, PAGE_SIZE
);
2688 mutex_unlock(&cgroup_mutex
);
2689 put_task_struct(tsk
);
2696 static int cgroup_open(struct inode
*inode
, struct file
*file
)
2698 struct pid
*pid
= PROC_I(inode
)->pid
;
2699 return single_open(file
, proc_cgroup_show
, pid
);
2702 struct file_operations proc_cgroup_operations
= {
2703 .open
= cgroup_open
,
2705 .llseek
= seq_lseek
,
2706 .release
= single_release
,
2709 /* Display information about each subsystem and each hierarchy */
2710 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
2714 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
2715 mutex_lock(&cgroup_mutex
);
2716 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2717 struct cgroup_subsys
*ss
= subsys
[i
];
2718 seq_printf(m
, "%s\t%lu\t%d\t%d\n",
2719 ss
->name
, ss
->root
->subsys_bits
,
2720 ss
->root
->number_of_cgroups
, !ss
->disabled
);
2722 mutex_unlock(&cgroup_mutex
);
2726 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
2728 return single_open(file
, proc_cgroupstats_show
, NULL
);
2731 static struct file_operations proc_cgroupstats_operations
= {
2732 .open
= cgroupstats_open
,
2734 .llseek
= seq_lseek
,
2735 .release
= single_release
,
2739 * cgroup_fork - attach newly forked task to its parents cgroup.
2740 * @child: pointer to task_struct of forking parent process.
2742 * Description: A task inherits its parent's cgroup at fork().
2744 * A pointer to the shared css_set was automatically copied in
2745 * fork.c by dup_task_struct(). However, we ignore that copy, since
2746 * it was not made under the protection of RCU or cgroup_mutex, so
2747 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
2748 * have already changed current->cgroups, allowing the previously
2749 * referenced cgroup group to be removed and freed.
2751 * At the point that cgroup_fork() is called, 'current' is the parent
2752 * task, and the passed argument 'child' points to the child task.
2754 void cgroup_fork(struct task_struct
*child
)
2757 child
->cgroups
= current
->cgroups
;
2758 get_css_set(child
->cgroups
);
2759 task_unlock(current
);
2760 INIT_LIST_HEAD(&child
->cg_list
);
2764 * cgroup_fork_callbacks - run fork callbacks
2765 * @child: the new task
2767 * Called on a new task very soon before adding it to the
2768 * tasklist. No need to take any locks since no-one can
2769 * be operating on this task.
2771 void cgroup_fork_callbacks(struct task_struct
*child
)
2773 if (need_forkexit_callback
) {
2775 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2776 struct cgroup_subsys
*ss
= subsys
[i
];
2778 ss
->fork(ss
, child
);
2784 * cgroup_post_fork - called on a new task after adding it to the task list
2785 * @child: the task in question
2787 * Adds the task to the list running through its css_set if necessary.
2788 * Has to be after the task is visible on the task list in case we race
2789 * with the first call to cgroup_iter_start() - to guarantee that the
2790 * new task ends up on its list.
2792 void cgroup_post_fork(struct task_struct
*child
)
2794 if (use_task_css_set_links
) {
2795 write_lock(&css_set_lock
);
2797 if (list_empty(&child
->cg_list
))
2798 list_add(&child
->cg_list
, &child
->cgroups
->tasks
);
2800 write_unlock(&css_set_lock
);
2804 * cgroup_exit - detach cgroup from exiting task
2805 * @tsk: pointer to task_struct of exiting process
2806 * @run_callback: run exit callbacks?
2808 * Description: Detach cgroup from @tsk and release it.
2810 * Note that cgroups marked notify_on_release force every task in
2811 * them to take the global cgroup_mutex mutex when exiting.
2812 * This could impact scaling on very large systems. Be reluctant to
2813 * use notify_on_release cgroups where very high task exit scaling
2814 * is required on large systems.
2816 * the_top_cgroup_hack:
2818 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2820 * We call cgroup_exit() while the task is still competent to
2821 * handle notify_on_release(), then leave the task attached to the
2822 * root cgroup in each hierarchy for the remainder of its exit.
2824 * To do this properly, we would increment the reference count on
2825 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
2826 * code we would add a second cgroup function call, to drop that
2827 * reference. This would just create an unnecessary hot spot on
2828 * the top_cgroup reference count, to no avail.
2830 * Normally, holding a reference to a cgroup without bumping its
2831 * count is unsafe. The cgroup could go away, or someone could
2832 * attach us to a different cgroup, decrementing the count on
2833 * the first cgroup that we never incremented. But in this case,
2834 * top_cgroup isn't going away, and either task has PF_EXITING set,
2835 * which wards off any cgroup_attach_task() attempts, or task is a failed
2836 * fork, never visible to cgroup_attach_task.
2838 void cgroup_exit(struct task_struct
*tsk
, int run_callbacks
)
2843 if (run_callbacks
&& need_forkexit_callback
) {
2844 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2845 struct cgroup_subsys
*ss
= subsys
[i
];
2852 * Unlink from the css_set task list if necessary.
2853 * Optimistically check cg_list before taking
2856 if (!list_empty(&tsk
->cg_list
)) {
2857 write_lock(&css_set_lock
);
2858 if (!list_empty(&tsk
->cg_list
))
2859 list_del(&tsk
->cg_list
);
2860 write_unlock(&css_set_lock
);
2863 /* Reassign the task to the init_css_set. */
2866 tsk
->cgroups
= &init_css_set
;
2869 put_css_set_taskexit(cg
);
2873 * cgroup_clone - clone the cgroup the given subsystem is attached to
2874 * @tsk: the task to be moved
2875 * @subsys: the given subsystem
2876 * @nodename: the name for the new cgroup
2878 * Duplicate the current cgroup in the hierarchy that the given
2879 * subsystem is attached to, and move this task into the new
2882 int cgroup_clone(struct task_struct
*tsk
, struct cgroup_subsys
*subsys
,
2885 struct dentry
*dentry
;
2887 struct cgroup
*parent
, *child
;
2888 struct inode
*inode
;
2890 struct cgroupfs_root
*root
;
2891 struct cgroup_subsys
*ss
;
2893 /* We shouldn't be called by an unregistered subsystem */
2894 BUG_ON(!subsys
->active
);
2896 /* First figure out what hierarchy and cgroup we're dealing
2897 * with, and pin them so we can drop cgroup_mutex */
2898 mutex_lock(&cgroup_mutex
);
2900 root
= subsys
->root
;
2901 if (root
== &rootnode
) {
2902 mutex_unlock(&cgroup_mutex
);
2907 parent
= task_cgroup(tsk
, subsys
->subsys_id
);
2909 /* Pin the hierarchy */
2910 if (!atomic_inc_not_zero(&parent
->root
->sb
->s_active
)) {
2911 /* We race with the final deactivate_super() */
2912 mutex_unlock(&cgroup_mutex
);
2916 /* Keep the cgroup alive */
2919 mutex_unlock(&cgroup_mutex
);
2921 /* Now do the VFS work to create a cgroup */
2922 inode
= parent
->dentry
->d_inode
;
2924 /* Hold the parent directory mutex across this operation to
2925 * stop anyone else deleting the new cgroup */
2926 mutex_lock(&inode
->i_mutex
);
2927 dentry
= lookup_one_len(nodename
, parent
->dentry
, strlen(nodename
));
2928 if (IS_ERR(dentry
)) {
2930 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename
,
2932 ret
= PTR_ERR(dentry
);
2936 /* Create the cgroup directory, which also creates the cgroup */
2937 ret
= vfs_mkdir(inode
, dentry
, 0755);
2938 child
= __d_cgrp(dentry
);
2942 "Failed to create cgroup %s: %d\n", nodename
,
2947 /* The cgroup now exists. Retake cgroup_mutex and check
2948 * that we're still in the same state that we thought we
2950 mutex_lock(&cgroup_mutex
);
2951 if ((root
!= subsys
->root
) ||
2952 (parent
!= task_cgroup(tsk
, subsys
->subsys_id
))) {
2953 /* Aargh, we raced ... */
2954 mutex_unlock(&inode
->i_mutex
);
2957 deactivate_super(parent
->root
->sb
);
2958 /* The cgroup is still accessible in the VFS, but
2959 * we're not going to try to rmdir() it at this
2962 "Race in cgroup_clone() - leaking cgroup %s\n",
2967 /* do any required auto-setup */
2968 for_each_subsys(root
, ss
) {
2970 ss
->post_clone(ss
, child
);
2973 /* All seems fine. Finish by moving the task into the new cgroup */
2974 ret
= cgroup_attach_task(child
, tsk
);
2975 mutex_unlock(&cgroup_mutex
);
2978 mutex_unlock(&inode
->i_mutex
);
2980 mutex_lock(&cgroup_mutex
);
2982 mutex_unlock(&cgroup_mutex
);
2983 deactivate_super(parent
->root
->sb
);
2988 * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
2989 * @cgrp: the cgroup in question
2991 * See if @cgrp is a descendant of the current task's cgroup in
2992 * the appropriate hierarchy.
2994 * If we are sending in dummytop, then presumably we are creating
2995 * the top cgroup in the subsystem.
2997 * Called only by the ns (nsproxy) cgroup.
2999 int cgroup_is_descendant(const struct cgroup
*cgrp
)
3002 struct cgroup
*target
;
3005 if (cgrp
== dummytop
)
3008 get_first_subsys(cgrp
, NULL
, &subsys_id
);
3009 target
= task_cgroup(current
, subsys_id
);
3010 while (cgrp
!= target
&& cgrp
!= cgrp
->top_cgroup
)
3011 cgrp
= cgrp
->parent
;
3012 ret
= (cgrp
== target
);
3016 static void check_for_release(struct cgroup
*cgrp
)
3018 /* All of these checks rely on RCU to keep the cgroup
3019 * structure alive */
3020 if (cgroup_is_releasable(cgrp
) && !atomic_read(&cgrp
->count
)
3021 && list_empty(&cgrp
->children
) && !cgroup_has_css_refs(cgrp
)) {
3022 /* Control Group is currently removeable. If it's not
3023 * already queued for a userspace notification, queue
3025 int need_schedule_work
= 0;
3026 spin_lock(&release_list_lock
);
3027 if (!cgroup_is_removed(cgrp
) &&
3028 list_empty(&cgrp
->release_list
)) {
3029 list_add(&cgrp
->release_list
, &release_list
);
3030 need_schedule_work
= 1;
3032 spin_unlock(&release_list_lock
);
3033 if (need_schedule_work
)
3034 schedule_work(&release_agent_work
);
3038 void __css_put(struct cgroup_subsys_state
*css
)
3040 struct cgroup
*cgrp
= css
->cgroup
;
3042 if (atomic_dec_and_test(&css
->refcnt
) && notify_on_release(cgrp
)) {
3043 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
3044 check_for_release(cgrp
);
3050 * Notify userspace when a cgroup is released, by running the
3051 * configured release agent with the name of the cgroup (path
3052 * relative to the root of cgroup file system) as the argument.
3054 * Most likely, this user command will try to rmdir this cgroup.
3056 * This races with the possibility that some other task will be
3057 * attached to this cgroup before it is removed, or that some other
3058 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
3059 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
3060 * unused, and this cgroup will be reprieved from its death sentence,
3061 * to continue to serve a useful existence. Next time it's released,
3062 * we will get notified again, if it still has 'notify_on_release' set.
3064 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
3065 * means only wait until the task is successfully execve()'d. The
3066 * separate release agent task is forked by call_usermodehelper(),
3067 * then control in this thread returns here, without waiting for the
3068 * release agent task. We don't bother to wait because the caller of
3069 * this routine has no use for the exit status of the release agent
3070 * task, so no sense holding our caller up for that.
3072 static void cgroup_release_agent(struct work_struct
*work
)
3074 BUG_ON(work
!= &release_agent_work
);
3075 mutex_lock(&cgroup_mutex
);
3076 spin_lock(&release_list_lock
);
3077 while (!list_empty(&release_list
)) {
3078 char *argv
[3], *envp
[3];
3080 char *pathbuf
= NULL
, *agentbuf
= NULL
;
3081 struct cgroup
*cgrp
= list_entry(release_list
.next
,
3084 list_del_init(&cgrp
->release_list
);
3085 spin_unlock(&release_list_lock
);
3086 pathbuf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
3089 if (cgroup_path(cgrp
, pathbuf
, PAGE_SIZE
) < 0)
3091 agentbuf
= kstrdup(cgrp
->root
->release_agent_path
, GFP_KERNEL
);
3096 argv
[i
++] = agentbuf
;
3097 argv
[i
++] = pathbuf
;
3101 /* minimal command environment */
3102 envp
[i
++] = "HOME=/";
3103 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3106 /* Drop the lock while we invoke the usermode helper,
3107 * since the exec could involve hitting disk and hence
3108 * be a slow process */
3109 mutex_unlock(&cgroup_mutex
);
3110 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
3111 mutex_lock(&cgroup_mutex
);
3115 spin_lock(&release_list_lock
);
3117 spin_unlock(&release_list_lock
);
3118 mutex_unlock(&cgroup_mutex
);
3121 static int __init
cgroup_disable(char *str
)
3126 while ((token
= strsep(&str
, ",")) != NULL
) {
3130 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
3131 struct cgroup_subsys
*ss
= subsys
[i
];
3133 if (!strcmp(token
, ss
->name
)) {
3135 printk(KERN_INFO
"Disabling %s control group"
3136 " subsystem\n", ss
->name
);
3143 __setup("cgroup_disable=", cgroup_disable
);