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>
48 #include <asm/atomic.h>
50 static DEFINE_MUTEX(cgroup_mutex
);
52 /* Generate an array of cgroup subsystem pointers */
53 #define SUBSYS(_x) &_x ## _subsys,
55 static struct cgroup_subsys
*subsys
[] = {
56 #include <linux/cgroup_subsys.h>
60 * A cgroupfs_root represents the root of a cgroup hierarchy,
61 * and may be associated with a superblock to form an active
64 struct cgroupfs_root
{
65 struct super_block
*sb
;
68 * The bitmask of subsystems intended to be attached to this
71 unsigned long subsys_bits
;
73 /* The bitmask of subsystems currently attached to this hierarchy */
74 unsigned long actual_subsys_bits
;
76 /* A list running through the attached subsystems */
77 struct list_head subsys_list
;
79 /* The root cgroup for this hierarchy */
80 struct cgroup top_cgroup
;
82 /* Tracks how many cgroups are currently defined in hierarchy.*/
83 int number_of_cgroups
;
85 /* A list running through the mounted hierarchies */
86 struct list_head root_list
;
88 /* Hierarchy-specific flags */
91 /* The path to use for release notifications. No locking
92 * between setting and use - so if userspace updates this
93 * while child cgroups exist, you could miss a
94 * notification. We ensure that it's always a valid
95 * NUL-terminated string */
96 char release_agent_path
[PATH_MAX
];
101 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
102 * subsystems that are otherwise unattached - it never has more than a
103 * single cgroup, and all tasks are part of that cgroup.
105 static struct cgroupfs_root rootnode
;
107 /* The list of hierarchy roots */
109 static LIST_HEAD(roots
);
110 static int root_count
;
112 /* dummytop is a shorthand for the dummy hierarchy's top cgroup */
113 #define dummytop (&rootnode.top_cgroup)
115 /* This flag indicates whether tasks in the fork and exit paths should
116 * take callback_mutex and check for fork/exit handlers to call. This
117 * avoids us having to do extra work in the fork/exit path if none of the
118 * subsystems need to be called.
120 static int need_forkexit_callback
;
122 /* bits in struct cgroup flags field */
124 /* Control Group is dead */
126 /* Control Group has previously had a child cgroup or a task,
127 * but no longer (only if CGRP_NOTIFY_ON_RELEASE is set) */
129 /* Control Group requires release notifications to userspace */
130 CGRP_NOTIFY_ON_RELEASE
,
133 /* convenient tests for these bits */
134 inline int cgroup_is_removed(const struct cgroup
*cgrp
)
136 return test_bit(CGRP_REMOVED
, &cgrp
->flags
);
139 /* bits in struct cgroupfs_root flags field */
141 ROOT_NOPREFIX
, /* mounted subsystems have no named prefix */
144 static int cgroup_is_releasable(const struct cgroup
*cgrp
)
147 (1 << CGRP_RELEASABLE
) |
148 (1 << CGRP_NOTIFY_ON_RELEASE
);
149 return (cgrp
->flags
& bits
) == bits
;
152 static int notify_on_release(const struct cgroup
*cgrp
)
154 return test_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
158 * for_each_subsys() allows you to iterate on each subsystem attached to
159 * an active hierarchy
161 #define for_each_subsys(_root, _ss) \
162 list_for_each_entry(_ss, &_root->subsys_list, sibling)
164 /* for_each_root() allows you to iterate across the active hierarchies */
165 #define for_each_root(_root) \
166 list_for_each_entry(_root, &roots, root_list)
168 /* the list of cgroups eligible for automatic release. Protected by
169 * release_list_lock */
170 static LIST_HEAD(release_list
);
171 static DEFINE_SPINLOCK(release_list_lock
);
172 static void cgroup_release_agent(struct work_struct
*work
);
173 static DECLARE_WORK(release_agent_work
, cgroup_release_agent
);
174 static void check_for_release(struct cgroup
*cgrp
);
176 /* Link structure for associating css_set objects with cgroups */
177 struct cg_cgroup_link
{
179 * List running through cg_cgroup_links associated with a
180 * cgroup, anchored on cgroup->css_sets
182 struct list_head cgrp_link_list
;
184 * List running through cg_cgroup_links pointing at a
185 * single css_set object, anchored on css_set->cg_links
187 struct list_head cg_link_list
;
191 /* The default css_set - used by init and its children prior to any
192 * hierarchies being mounted. It contains a pointer to the root state
193 * for each subsystem. Also used to anchor the list of css_sets. Not
194 * reference-counted, to improve performance when child cgroups
195 * haven't been created.
198 static struct css_set init_css_set
;
199 static struct cg_cgroup_link init_css_set_link
;
201 /* css_set_lock protects the list of css_set objects, and the
202 * chain of tasks off each css_set. Nests outside task->alloc_lock
203 * due to cgroup_iter_start() */
204 static DEFINE_RWLOCK(css_set_lock
);
205 static int css_set_count
;
207 /* We don't maintain the lists running through each css_set to its
208 * task until after the first call to cgroup_iter_start(). This
209 * reduces the fork()/exit() overhead for people who have cgroups
210 * compiled into their kernel but not actually in use */
211 static int use_task_css_set_links
;
213 /* When we create or destroy a css_set, the operation simply
214 * takes/releases a reference count on all the cgroups referenced
215 * by subsystems in this css_set. This can end up multiple-counting
216 * some cgroups, but that's OK - the ref-count is just a
217 * busy/not-busy indicator; ensuring that we only count each cgroup
218 * once would require taking a global lock to ensure that no
219 * subsystems moved between hierarchies while we were doing so.
221 * Possible TODO: decide at boot time based on the number of
222 * registered subsystems and the number of CPUs or NUMA nodes whether
223 * it's better for performance to ref-count every subsystem, or to
224 * take a global lock and only add one ref count to each hierarchy.
228 * unlink a css_set from the list and free it
230 static void unlink_css_set(struct css_set
*cg
)
232 write_lock(&css_set_lock
);
235 while (!list_empty(&cg
->cg_links
)) {
236 struct cg_cgroup_link
*link
;
237 link
= list_entry(cg
->cg_links
.next
,
238 struct cg_cgroup_link
, cg_link_list
);
239 list_del(&link
->cg_link_list
);
240 list_del(&link
->cgrp_link_list
);
243 write_unlock(&css_set_lock
);
246 static void __release_css_set(struct kref
*k
, int taskexit
)
249 struct css_set
*cg
= container_of(k
, struct css_set
, ref
);
254 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
255 struct cgroup
*cgrp
= cg
->subsys
[i
]->cgroup
;
256 if (atomic_dec_and_test(&cgrp
->count
) &&
257 notify_on_release(cgrp
)) {
259 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
260 check_for_release(cgrp
);
267 static void release_css_set(struct kref
*k
)
269 __release_css_set(k
, 0);
272 static void release_css_set_taskexit(struct kref
*k
)
274 __release_css_set(k
, 1);
278 * refcounted get/put for css_set objects
280 static inline void get_css_set(struct css_set
*cg
)
285 static inline void put_css_set(struct css_set
*cg
)
287 kref_put(&cg
->ref
, release_css_set
);
290 static inline void put_css_set_taskexit(struct css_set
*cg
)
292 kref_put(&cg
->ref
, release_css_set_taskexit
);
296 * find_existing_css_set() is a helper for
297 * find_css_set(), and checks to see whether an existing
298 * css_set is suitable. This currently walks a linked-list for
299 * simplicity; a later patch will use a hash table for better
302 * oldcg: the cgroup group that we're using before the cgroup
305 * cgrp: the cgroup that we're moving into
307 * template: location in which to build the desired set of subsystem
308 * state objects for the new cgroup group
311 static struct css_set
*find_existing_css_set(
312 struct css_set
*oldcg
,
314 struct cgroup_subsys_state
*template[])
317 struct cgroupfs_root
*root
= cgrp
->root
;
318 struct list_head
*l
= &init_css_set
.list
;
320 /* Built the set of subsystem state objects that we want to
321 * see in the new css_set */
322 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
323 if (root
->subsys_bits
& (1ull << i
)) {
324 /* Subsystem is in this hierarchy. So we want
325 * the subsystem state from the new
327 template[i
] = cgrp
->subsys
[i
];
329 /* Subsystem is not in this hierarchy, so we
330 * don't want to change the subsystem state */
331 template[i
] = oldcg
->subsys
[i
];
335 /* Look through existing cgroup groups to find one to reuse */
338 list_entry(l
, struct css_set
, list
);
340 if (!memcmp(template, cg
->subsys
, sizeof(cg
->subsys
))) {
341 /* All subsystems matched */
344 /* Try the next cgroup group */
346 } while (l
!= &init_css_set
.list
);
348 /* No existing cgroup group matched */
353 * allocate_cg_links() allocates "count" cg_cgroup_link structures
354 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
355 * success or a negative error
358 static int allocate_cg_links(int count
, struct list_head
*tmp
)
360 struct cg_cgroup_link
*link
;
363 for (i
= 0; i
< count
; i
++) {
364 link
= kmalloc(sizeof(*link
), GFP_KERNEL
);
366 while (!list_empty(tmp
)) {
367 link
= list_entry(tmp
->next
,
368 struct cg_cgroup_link
,
370 list_del(&link
->cgrp_link_list
);
375 list_add(&link
->cgrp_link_list
, tmp
);
380 static void free_cg_links(struct list_head
*tmp
)
382 while (!list_empty(tmp
)) {
383 struct cg_cgroup_link
*link
;
384 link
= list_entry(tmp
->next
,
385 struct cg_cgroup_link
,
387 list_del(&link
->cgrp_link_list
);
393 * find_css_set() takes an existing cgroup group and a
394 * cgroup object, and returns a css_set object that's
395 * equivalent to the old group, but with the given cgroup
396 * substituted into the appropriate hierarchy. Must be called with
400 static struct css_set
*find_css_set(
401 struct css_set
*oldcg
, struct cgroup
*cgrp
)
404 struct cgroup_subsys_state
*template[CGROUP_SUBSYS_COUNT
];
407 struct list_head tmp_cg_links
;
408 struct cg_cgroup_link
*link
;
410 /* First see if we already have a cgroup group that matches
412 write_lock(&css_set_lock
);
413 res
= find_existing_css_set(oldcg
, cgrp
, template);
416 write_unlock(&css_set_lock
);
421 res
= kmalloc(sizeof(*res
), GFP_KERNEL
);
425 /* Allocate all the cg_cgroup_link objects that we'll need */
426 if (allocate_cg_links(root_count
, &tmp_cg_links
) < 0) {
431 kref_init(&res
->ref
);
432 INIT_LIST_HEAD(&res
->cg_links
);
433 INIT_LIST_HEAD(&res
->tasks
);
435 /* Copy the set of subsystem state objects generated in
436 * find_existing_css_set() */
437 memcpy(res
->subsys
, template, sizeof(res
->subsys
));
439 write_lock(&css_set_lock
);
440 /* Add reference counts and links from the new css_set. */
441 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
442 struct cgroup
*cgrp
= res
->subsys
[i
]->cgroup
;
443 struct cgroup_subsys
*ss
= subsys
[i
];
444 atomic_inc(&cgrp
->count
);
446 * We want to add a link once per cgroup, so we
447 * only do it for the first subsystem in each
450 if (ss
->root
->subsys_list
.next
== &ss
->sibling
) {
451 BUG_ON(list_empty(&tmp_cg_links
));
452 link
= list_entry(tmp_cg_links
.next
,
453 struct cg_cgroup_link
,
455 list_del(&link
->cgrp_link_list
);
456 list_add(&link
->cgrp_link_list
, &cgrp
->css_sets
);
458 list_add(&link
->cg_link_list
, &res
->cg_links
);
461 if (list_empty(&rootnode
.subsys_list
)) {
462 link
= list_entry(tmp_cg_links
.next
,
463 struct cg_cgroup_link
,
465 list_del(&link
->cgrp_link_list
);
466 list_add(&link
->cgrp_link_list
, &dummytop
->css_sets
);
468 list_add(&link
->cg_link_list
, &res
->cg_links
);
471 BUG_ON(!list_empty(&tmp_cg_links
));
473 /* Link this cgroup group into the list */
474 list_add(&res
->list
, &init_css_set
.list
);
476 INIT_LIST_HEAD(&res
->tasks
);
477 write_unlock(&css_set_lock
);
483 * There is one global cgroup mutex. We also require taking
484 * task_lock() when dereferencing a task's cgroup subsys pointers.
485 * See "The task_lock() exception", at the end of this comment.
487 * A task must hold cgroup_mutex to modify cgroups.
489 * Any task can increment and decrement the count field without lock.
490 * So in general, code holding cgroup_mutex can't rely on the count
491 * field not changing. However, if the count goes to zero, then only
492 * attach_task() can increment it again. Because a count of zero
493 * means that no tasks are currently attached, therefore there is no
494 * way a task attached to that cgroup can fork (the other way to
495 * increment the count). So code holding cgroup_mutex can safely
496 * assume that if the count is zero, it will stay zero. Similarly, if
497 * a task holds cgroup_mutex on a cgroup with zero count, it
498 * knows that the cgroup won't be removed, as cgroup_rmdir()
501 * The cgroup_common_file_write handler for operations that modify
502 * the cgroup hierarchy holds cgroup_mutex across the entire operation,
503 * single threading all such cgroup modifications across the system.
505 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
506 * (usually) take cgroup_mutex. These are the two most performance
507 * critical pieces of code here. The exception occurs on cgroup_exit(),
508 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
509 * is taken, and if the cgroup count is zero, a usermode call made
510 * to /sbin/cgroup_release_agent with the name of the cgroup (path
511 * relative to the root of cgroup file system) as the argument.
513 * A cgroup can only be deleted if both its 'count' of using tasks
514 * is zero, and its list of 'children' cgroups is empty. Since all
515 * tasks in the system use _some_ cgroup, and since there is always at
516 * least one task in the system (init, pid == 1), therefore, top_cgroup
517 * always has either children cgroups and/or using tasks. So we don't
518 * need a special hack to ensure that top_cgroup cannot be deleted.
520 * The task_lock() exception
522 * The need for this exception arises from the action of
523 * attach_task(), which overwrites one tasks cgroup pointer with
524 * another. It does so using cgroup_mutexe, however there are
525 * several performance critical places that need to reference
526 * task->cgroup without the expense of grabbing a system global
527 * mutex. Therefore except as noted below, when dereferencing or, as
528 * in attach_task(), modifying a task'ss cgroup pointer we use
529 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
530 * the task_struct routinely used for such matters.
532 * P.S. One more locking exception. RCU is used to guard the
533 * update of a tasks cgroup pointer by attach_task()
537 * cgroup_lock - lock out any changes to cgroup structures
541 void cgroup_lock(void)
543 mutex_lock(&cgroup_mutex
);
547 * cgroup_unlock - release lock on cgroup changes
549 * Undo the lock taken in a previous cgroup_lock() call.
552 void cgroup_unlock(void)
554 mutex_unlock(&cgroup_mutex
);
558 * A couple of forward declarations required, due to cyclic reference loop:
559 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
560 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
564 static int cgroup_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
);
565 static int cgroup_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
);
566 static int cgroup_populate_dir(struct cgroup
*cgrp
);
567 static struct inode_operations cgroup_dir_inode_operations
;
568 static struct file_operations proc_cgroupstats_operations
;
570 static struct backing_dev_info cgroup_backing_dev_info
= {
571 .capabilities
= BDI_CAP_NO_ACCT_DIRTY
| BDI_CAP_NO_WRITEBACK
,
574 static struct inode
*cgroup_new_inode(mode_t mode
, struct super_block
*sb
)
576 struct inode
*inode
= new_inode(sb
);
579 inode
->i_mode
= mode
;
580 inode
->i_uid
= current
->fsuid
;
581 inode
->i_gid
= current
->fsgid
;
583 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
584 inode
->i_mapping
->backing_dev_info
= &cgroup_backing_dev_info
;
589 static void cgroup_diput(struct dentry
*dentry
, struct inode
*inode
)
591 /* is dentry a directory ? if so, kfree() associated cgroup */
592 if (S_ISDIR(inode
->i_mode
)) {
593 struct cgroup
*cgrp
= dentry
->d_fsdata
;
594 struct cgroup_subsys
*ss
;
595 BUG_ON(!(cgroup_is_removed(cgrp
)));
596 /* It's possible for external users to be holding css
597 * reference counts on a cgroup; css_put() needs to
598 * be able to access the cgroup after decrementing
599 * the reference count in order to know if it needs to
600 * queue the cgroup to be handled by the release
604 mutex_lock(&cgroup_mutex
);
606 * Release the subsystem state objects.
608 for_each_subsys(cgrp
->root
, ss
) {
609 if (cgrp
->subsys
[ss
->subsys_id
])
610 ss
->destroy(ss
, cgrp
);
613 cgrp
->root
->number_of_cgroups
--;
614 mutex_unlock(&cgroup_mutex
);
616 /* Drop the active superblock reference that we took when we
617 * created the cgroup */
618 deactivate_super(cgrp
->root
->sb
);
625 static void remove_dir(struct dentry
*d
)
627 struct dentry
*parent
= dget(d
->d_parent
);
630 simple_rmdir(parent
->d_inode
, d
);
634 static void cgroup_clear_directory(struct dentry
*dentry
)
636 struct list_head
*node
;
638 BUG_ON(!mutex_is_locked(&dentry
->d_inode
->i_mutex
));
639 spin_lock(&dcache_lock
);
640 node
= dentry
->d_subdirs
.next
;
641 while (node
!= &dentry
->d_subdirs
) {
642 struct dentry
*d
= list_entry(node
, struct dentry
, d_u
.d_child
);
645 /* This should never be called on a cgroup
646 * directory with child cgroups */
647 BUG_ON(d
->d_inode
->i_mode
& S_IFDIR
);
649 spin_unlock(&dcache_lock
);
651 simple_unlink(dentry
->d_inode
, d
);
653 spin_lock(&dcache_lock
);
655 node
= dentry
->d_subdirs
.next
;
657 spin_unlock(&dcache_lock
);
661 * NOTE : the dentry must have been dget()'ed
663 static void cgroup_d_remove_dir(struct dentry
*dentry
)
665 cgroup_clear_directory(dentry
);
667 spin_lock(&dcache_lock
);
668 list_del_init(&dentry
->d_u
.d_child
);
669 spin_unlock(&dcache_lock
);
673 static int rebind_subsystems(struct cgroupfs_root
*root
,
674 unsigned long final_bits
)
676 unsigned long added_bits
, removed_bits
;
677 struct cgroup
*cgrp
= &root
->top_cgroup
;
680 removed_bits
= root
->actual_subsys_bits
& ~final_bits
;
681 added_bits
= final_bits
& ~root
->actual_subsys_bits
;
682 /* Check that any added subsystems are currently free */
683 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
684 unsigned long long bit
= 1ull << i
;
685 struct cgroup_subsys
*ss
= subsys
[i
];
686 if (!(bit
& added_bits
))
688 if (ss
->root
!= &rootnode
) {
689 /* Subsystem isn't free */
694 /* Currently we don't handle adding/removing subsystems when
695 * any child cgroups exist. This is theoretically supportable
696 * but involves complex error handling, so it's being left until
698 if (!list_empty(&cgrp
->children
))
701 /* Process each subsystem */
702 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
703 struct cgroup_subsys
*ss
= subsys
[i
];
704 unsigned long bit
= 1UL << i
;
705 if (bit
& added_bits
) {
706 /* We're binding this subsystem to this hierarchy */
707 BUG_ON(cgrp
->subsys
[i
]);
708 BUG_ON(!dummytop
->subsys
[i
]);
709 BUG_ON(dummytop
->subsys
[i
]->cgroup
!= dummytop
);
710 cgrp
->subsys
[i
] = dummytop
->subsys
[i
];
711 cgrp
->subsys
[i
]->cgroup
= cgrp
;
712 list_add(&ss
->sibling
, &root
->subsys_list
);
713 rcu_assign_pointer(ss
->root
, root
);
717 } else if (bit
& removed_bits
) {
718 /* We're removing this subsystem */
719 BUG_ON(cgrp
->subsys
[i
] != dummytop
->subsys
[i
]);
720 BUG_ON(cgrp
->subsys
[i
]->cgroup
!= cgrp
);
722 ss
->bind(ss
, dummytop
);
723 dummytop
->subsys
[i
]->cgroup
= dummytop
;
724 cgrp
->subsys
[i
] = NULL
;
725 rcu_assign_pointer(subsys
[i
]->root
, &rootnode
);
726 list_del(&ss
->sibling
);
727 } else if (bit
& final_bits
) {
728 /* Subsystem state should already exist */
729 BUG_ON(!cgrp
->subsys
[i
]);
731 /* Subsystem state shouldn't exist */
732 BUG_ON(cgrp
->subsys
[i
]);
735 root
->subsys_bits
= root
->actual_subsys_bits
= final_bits
;
741 static int cgroup_show_options(struct seq_file
*seq
, struct vfsmount
*vfs
)
743 struct cgroupfs_root
*root
= vfs
->mnt_sb
->s_fs_info
;
744 struct cgroup_subsys
*ss
;
746 mutex_lock(&cgroup_mutex
);
747 for_each_subsys(root
, ss
)
748 seq_printf(seq
, ",%s", ss
->name
);
749 if (test_bit(ROOT_NOPREFIX
, &root
->flags
))
750 seq_puts(seq
, ",noprefix");
751 if (strlen(root
->release_agent_path
))
752 seq_printf(seq
, ",release_agent=%s", root
->release_agent_path
);
753 mutex_unlock(&cgroup_mutex
);
757 struct cgroup_sb_opts
{
758 unsigned long subsys_bits
;
763 /* Convert a hierarchy specifier into a bitmask of subsystems and
765 static int parse_cgroupfs_options(char *data
,
766 struct cgroup_sb_opts
*opts
)
768 char *token
, *o
= data
?: "all";
770 opts
->subsys_bits
= 0;
772 opts
->release_agent
= NULL
;
774 while ((token
= strsep(&o
, ",")) != NULL
) {
777 if (!strcmp(token
, "all")) {
778 opts
->subsys_bits
= (1 << CGROUP_SUBSYS_COUNT
) - 1;
779 } else if (!strcmp(token
, "noprefix")) {
780 set_bit(ROOT_NOPREFIX
, &opts
->flags
);
781 } else if (!strncmp(token
, "release_agent=", 14)) {
782 /* Specifying two release agents is forbidden */
783 if (opts
->release_agent
)
785 opts
->release_agent
= kzalloc(PATH_MAX
, GFP_KERNEL
);
786 if (!opts
->release_agent
)
788 strncpy(opts
->release_agent
, token
+ 14, PATH_MAX
- 1);
789 opts
->release_agent
[PATH_MAX
- 1] = 0;
791 struct cgroup_subsys
*ss
;
793 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
795 if (!strcmp(token
, ss
->name
)) {
796 set_bit(i
, &opts
->subsys_bits
);
800 if (i
== CGROUP_SUBSYS_COUNT
)
805 /* We can't have an empty hierarchy */
806 if (!opts
->subsys_bits
)
812 static int cgroup_remount(struct super_block
*sb
, int *flags
, char *data
)
815 struct cgroupfs_root
*root
= sb
->s_fs_info
;
816 struct cgroup
*cgrp
= &root
->top_cgroup
;
817 struct cgroup_sb_opts opts
;
819 mutex_lock(&cgrp
->dentry
->d_inode
->i_mutex
);
820 mutex_lock(&cgroup_mutex
);
822 /* See what subsystems are wanted */
823 ret
= parse_cgroupfs_options(data
, &opts
);
827 /* Don't allow flags to change at remount */
828 if (opts
.flags
!= root
->flags
) {
833 ret
= rebind_subsystems(root
, opts
.subsys_bits
);
835 /* (re)populate subsystem files */
837 cgroup_populate_dir(cgrp
);
839 if (opts
.release_agent
)
840 strcpy(root
->release_agent_path
, opts
.release_agent
);
842 if (opts
.release_agent
)
843 kfree(opts
.release_agent
);
844 mutex_unlock(&cgroup_mutex
);
845 mutex_unlock(&cgrp
->dentry
->d_inode
->i_mutex
);
849 static struct super_operations cgroup_ops
= {
850 .statfs
= simple_statfs
,
851 .drop_inode
= generic_delete_inode
,
852 .show_options
= cgroup_show_options
,
853 .remount_fs
= cgroup_remount
,
856 static void init_cgroup_root(struct cgroupfs_root
*root
)
858 struct cgroup
*cgrp
= &root
->top_cgroup
;
859 INIT_LIST_HEAD(&root
->subsys_list
);
860 INIT_LIST_HEAD(&root
->root_list
);
861 root
->number_of_cgroups
= 1;
863 cgrp
->top_cgroup
= cgrp
;
864 INIT_LIST_HEAD(&cgrp
->sibling
);
865 INIT_LIST_HEAD(&cgrp
->children
);
866 INIT_LIST_HEAD(&cgrp
->css_sets
);
867 INIT_LIST_HEAD(&cgrp
->release_list
);
870 static int cgroup_test_super(struct super_block
*sb
, void *data
)
872 struct cgroupfs_root
*new = data
;
873 struct cgroupfs_root
*root
= sb
->s_fs_info
;
875 /* First check subsystems */
876 if (new->subsys_bits
!= root
->subsys_bits
)
879 /* Next check flags */
880 if (new->flags
!= root
->flags
)
886 static int cgroup_set_super(struct super_block
*sb
, void *data
)
889 struct cgroupfs_root
*root
= data
;
891 ret
= set_anon_super(sb
, NULL
);
895 sb
->s_fs_info
= root
;
898 sb
->s_blocksize
= PAGE_CACHE_SIZE
;
899 sb
->s_blocksize_bits
= PAGE_CACHE_SHIFT
;
900 sb
->s_magic
= CGROUP_SUPER_MAGIC
;
901 sb
->s_op
= &cgroup_ops
;
906 static int cgroup_get_rootdir(struct super_block
*sb
)
908 struct inode
*inode
=
909 cgroup_new_inode(S_IFDIR
| S_IRUGO
| S_IXUGO
| S_IWUSR
, sb
);
910 struct dentry
*dentry
;
915 inode
->i_op
= &simple_dir_inode_operations
;
916 inode
->i_fop
= &simple_dir_operations
;
917 inode
->i_op
= &cgroup_dir_inode_operations
;
918 /* directories start off with i_nlink == 2 (for "." entry) */
920 dentry
= d_alloc_root(inode
);
929 static int cgroup_get_sb(struct file_system_type
*fs_type
,
930 int flags
, const char *unused_dev_name
,
931 void *data
, struct vfsmount
*mnt
)
933 struct cgroup_sb_opts opts
;
935 struct super_block
*sb
;
936 struct cgroupfs_root
*root
;
937 struct list_head tmp_cg_links
, *l
;
938 INIT_LIST_HEAD(&tmp_cg_links
);
940 /* First find the desired set of subsystems */
941 ret
= parse_cgroupfs_options(data
, &opts
);
943 if (opts
.release_agent
)
944 kfree(opts
.release_agent
);
948 root
= kzalloc(sizeof(*root
), GFP_KERNEL
);
952 init_cgroup_root(root
);
953 root
->subsys_bits
= opts
.subsys_bits
;
954 root
->flags
= opts
.flags
;
955 if (opts
.release_agent
) {
956 strcpy(root
->release_agent_path
, opts
.release_agent
);
957 kfree(opts
.release_agent
);
960 sb
= sget(fs_type
, cgroup_test_super
, cgroup_set_super
, root
);
967 if (sb
->s_fs_info
!= root
) {
968 /* Reusing an existing superblock */
969 BUG_ON(sb
->s_root
== NULL
);
974 struct cgroup
*cgrp
= &root
->top_cgroup
;
977 BUG_ON(sb
->s_root
!= NULL
);
979 ret
= cgroup_get_rootdir(sb
);
982 inode
= sb
->s_root
->d_inode
;
984 mutex_lock(&inode
->i_mutex
);
985 mutex_lock(&cgroup_mutex
);
988 * We're accessing css_set_count without locking
989 * css_set_lock here, but that's OK - it can only be
990 * increased by someone holding cgroup_lock, and
991 * that's us. The worst that can happen is that we
992 * have some link structures left over
994 ret
= allocate_cg_links(css_set_count
, &tmp_cg_links
);
996 mutex_unlock(&cgroup_mutex
);
997 mutex_unlock(&inode
->i_mutex
);
1001 ret
= rebind_subsystems(root
, root
->subsys_bits
);
1002 if (ret
== -EBUSY
) {
1003 mutex_unlock(&cgroup_mutex
);
1004 mutex_unlock(&inode
->i_mutex
);
1005 goto drop_new_super
;
1008 /* EBUSY should be the only error here */
1011 list_add(&root
->root_list
, &roots
);
1014 sb
->s_root
->d_fsdata
= &root
->top_cgroup
;
1015 root
->top_cgroup
.dentry
= sb
->s_root
;
1017 /* Link the top cgroup in this hierarchy into all
1018 * the css_set objects */
1019 write_lock(&css_set_lock
);
1020 l
= &init_css_set
.list
;
1023 struct cg_cgroup_link
*link
;
1024 cg
= list_entry(l
, struct css_set
, list
);
1025 BUG_ON(list_empty(&tmp_cg_links
));
1026 link
= list_entry(tmp_cg_links
.next
,
1027 struct cg_cgroup_link
,
1029 list_del(&link
->cgrp_link_list
);
1031 list_add(&link
->cgrp_link_list
,
1032 &root
->top_cgroup
.css_sets
);
1033 list_add(&link
->cg_link_list
, &cg
->cg_links
);
1035 } while (l
!= &init_css_set
.list
);
1036 write_unlock(&css_set_lock
);
1038 free_cg_links(&tmp_cg_links
);
1040 BUG_ON(!list_empty(&cgrp
->sibling
));
1041 BUG_ON(!list_empty(&cgrp
->children
));
1042 BUG_ON(root
->number_of_cgroups
!= 1);
1044 cgroup_populate_dir(cgrp
);
1045 mutex_unlock(&inode
->i_mutex
);
1046 mutex_unlock(&cgroup_mutex
);
1049 return simple_set_mnt(mnt
, sb
);
1052 up_write(&sb
->s_umount
);
1053 deactivate_super(sb
);
1054 free_cg_links(&tmp_cg_links
);
1058 static void cgroup_kill_sb(struct super_block
*sb
) {
1059 struct cgroupfs_root
*root
= sb
->s_fs_info
;
1060 struct cgroup
*cgrp
= &root
->top_cgroup
;
1065 BUG_ON(root
->number_of_cgroups
!= 1);
1066 BUG_ON(!list_empty(&cgrp
->children
));
1067 BUG_ON(!list_empty(&cgrp
->sibling
));
1069 mutex_lock(&cgroup_mutex
);
1071 /* Rebind all subsystems back to the default hierarchy */
1072 ret
= rebind_subsystems(root
, 0);
1073 /* Shouldn't be able to fail ... */
1077 * Release all the links from css_sets to this hierarchy's
1080 write_lock(&css_set_lock
);
1081 while (!list_empty(&cgrp
->css_sets
)) {
1082 struct cg_cgroup_link
*link
;
1083 link
= list_entry(cgrp
->css_sets
.next
,
1084 struct cg_cgroup_link
, cgrp_link_list
);
1085 list_del(&link
->cg_link_list
);
1086 list_del(&link
->cgrp_link_list
);
1089 write_unlock(&css_set_lock
);
1091 if (!list_empty(&root
->root_list
)) {
1092 list_del(&root
->root_list
);
1095 mutex_unlock(&cgroup_mutex
);
1098 kill_litter_super(sb
);
1101 static struct file_system_type cgroup_fs_type
= {
1103 .get_sb
= cgroup_get_sb
,
1104 .kill_sb
= cgroup_kill_sb
,
1107 static inline struct cgroup
*__d_cgrp(struct dentry
*dentry
)
1109 return dentry
->d_fsdata
;
1112 static inline struct cftype
*__d_cft(struct dentry
*dentry
)
1114 return dentry
->d_fsdata
;
1118 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1119 * Returns 0 on success, -errno on error.
1121 int cgroup_path(const struct cgroup
*cgrp
, char *buf
, int buflen
)
1125 if (cgrp
== dummytop
) {
1127 * Inactive subsystems have no dentry for their root
1134 start
= buf
+ buflen
;
1138 int len
= cgrp
->dentry
->d_name
.len
;
1139 if ((start
-= len
) < buf
)
1140 return -ENAMETOOLONG
;
1141 memcpy(start
, cgrp
->dentry
->d_name
.name
, len
);
1142 cgrp
= cgrp
->parent
;
1148 return -ENAMETOOLONG
;
1151 memmove(buf
, start
, buf
+ buflen
- start
);
1156 * Return the first subsystem attached to a cgroup's hierarchy, and
1160 static void get_first_subsys(const struct cgroup
*cgrp
,
1161 struct cgroup_subsys_state
**css
, int *subsys_id
)
1163 const struct cgroupfs_root
*root
= cgrp
->root
;
1164 const struct cgroup_subsys
*test_ss
;
1165 BUG_ON(list_empty(&root
->subsys_list
));
1166 test_ss
= list_entry(root
->subsys_list
.next
,
1167 struct cgroup_subsys
, sibling
);
1169 *css
= cgrp
->subsys
[test_ss
->subsys_id
];
1173 *subsys_id
= test_ss
->subsys_id
;
1177 * Attach task 'tsk' to cgroup 'cgrp'
1179 * Call holding cgroup_mutex. May take task_lock of
1180 * the task 'pid' during call.
1182 static int attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
1185 struct cgroup_subsys
*ss
;
1186 struct cgroup
*oldcgrp
;
1187 struct css_set
*cg
= tsk
->cgroups
;
1188 struct css_set
*newcg
;
1189 struct cgroupfs_root
*root
= cgrp
->root
;
1192 get_first_subsys(cgrp
, NULL
, &subsys_id
);
1194 /* Nothing to do if the task is already in that cgroup */
1195 oldcgrp
= task_cgroup(tsk
, subsys_id
);
1196 if (cgrp
== oldcgrp
)
1199 for_each_subsys(root
, ss
) {
1200 if (ss
->can_attach
) {
1201 retval
= ss
->can_attach(ss
, cgrp
, tsk
);
1208 * Locate or allocate a new css_set for this task,
1209 * based on its final set of cgroups
1211 newcg
= find_css_set(cg
, cgrp
);
1216 if (tsk
->flags
& PF_EXITING
) {
1221 rcu_assign_pointer(tsk
->cgroups
, newcg
);
1224 /* Update the css_set linked lists if we're using them */
1225 write_lock(&css_set_lock
);
1226 if (!list_empty(&tsk
->cg_list
)) {
1227 list_del(&tsk
->cg_list
);
1228 list_add(&tsk
->cg_list
, &newcg
->tasks
);
1230 write_unlock(&css_set_lock
);
1232 for_each_subsys(root
, ss
) {
1234 ss
->attach(ss
, cgrp
, oldcgrp
, tsk
);
1236 set_bit(CGRP_RELEASABLE
, &oldcgrp
->flags
);
1243 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with
1244 * cgroup_mutex, may take task_lock of task
1246 static int attach_task_by_pid(struct cgroup
*cgrp
, char *pidbuf
)
1249 struct task_struct
*tsk
;
1252 if (sscanf(pidbuf
, "%d", &pid
) != 1)
1257 tsk
= find_task_by_pid(pid
);
1258 if (!tsk
|| tsk
->flags
& PF_EXITING
) {
1262 get_task_struct(tsk
);
1265 if ((current
->euid
) && (current
->euid
!= tsk
->uid
)
1266 && (current
->euid
!= tsk
->suid
)) {
1267 put_task_struct(tsk
);
1272 get_task_struct(tsk
);
1275 ret
= attach_task(cgrp
, tsk
);
1276 put_task_struct(tsk
);
1280 /* The various types of files and directories in a cgroup file system */
1282 enum cgroup_filetype
{
1286 FILE_NOTIFY_ON_RELEASE
,
1291 static ssize_t
cgroup_write_uint(struct cgroup
*cgrp
, struct cftype
*cft
,
1293 const char __user
*userbuf
,
1294 size_t nbytes
, loff_t
*unused_ppos
)
1303 if (nbytes
>= sizeof(buffer
))
1305 if (copy_from_user(buffer
, userbuf
, nbytes
))
1308 buffer
[nbytes
] = 0; /* nul-terminate */
1310 /* strip newline if necessary */
1311 if (nbytes
&& (buffer
[nbytes
-1] == '\n'))
1312 buffer
[nbytes
-1] = 0;
1313 val
= simple_strtoull(buffer
, &end
, 0);
1317 /* Pass to subsystem */
1318 retval
= cft
->write_uint(cgrp
, cft
, val
);
1324 static ssize_t
cgroup_common_file_write(struct cgroup
*cgrp
,
1327 const char __user
*userbuf
,
1328 size_t nbytes
, loff_t
*unused_ppos
)
1330 enum cgroup_filetype type
= cft
->private;
1334 if (nbytes
>= PATH_MAX
)
1337 /* +1 for nul-terminator */
1338 buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
);
1342 if (copy_from_user(buffer
, userbuf
, nbytes
)) {
1346 buffer
[nbytes
] = 0; /* nul-terminate */
1347 strstrip(buffer
); /* strip -just- trailing whitespace */
1349 mutex_lock(&cgroup_mutex
);
1352 * This was already checked for in cgroup_file_write(), but
1353 * check again now we're holding cgroup_mutex.
1355 if (cgroup_is_removed(cgrp
)) {
1362 retval
= attach_task_by_pid(cgrp
, buffer
);
1364 case FILE_NOTIFY_ON_RELEASE
:
1365 clear_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
1366 if (simple_strtoul(buffer
, NULL
, 10) != 0)
1367 set_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
1369 clear_bit(CGRP_NOTIFY_ON_RELEASE
, &cgrp
->flags
);
1371 case FILE_RELEASE_AGENT
:
1372 BUILD_BUG_ON(sizeof(cgrp
->root
->release_agent_path
) < PATH_MAX
);
1373 strcpy(cgrp
->root
->release_agent_path
, buffer
);
1383 mutex_unlock(&cgroup_mutex
);
1389 static ssize_t
cgroup_file_write(struct file
*file
, const char __user
*buf
,
1390 size_t nbytes
, loff_t
*ppos
)
1392 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1393 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
1395 if (!cft
|| cgroup_is_removed(cgrp
))
1398 return cft
->write(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1399 if (cft
->write_uint
)
1400 return cgroup_write_uint(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1404 static ssize_t
cgroup_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
,
1406 char __user
*buf
, size_t nbytes
,
1410 u64 val
= cft
->read_uint(cgrp
, cft
);
1411 int len
= sprintf(tmp
, "%llu\n", (unsigned long long) val
);
1413 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
1416 static ssize_t
cgroup_common_file_read(struct cgroup
*cgrp
,
1420 size_t nbytes
, loff_t
*ppos
)
1422 enum cgroup_filetype type
= cft
->private;
1427 if (!(page
= (char *)__get_free_page(GFP_KERNEL
)))
1433 case FILE_RELEASE_AGENT
:
1435 struct cgroupfs_root
*root
;
1437 mutex_lock(&cgroup_mutex
);
1439 n
= strnlen(root
->release_agent_path
,
1440 sizeof(root
->release_agent_path
));
1441 n
= min(n
, (size_t) PAGE_SIZE
);
1442 strncpy(s
, root
->release_agent_path
, n
);
1443 mutex_unlock(&cgroup_mutex
);
1453 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1455 free_page((unsigned long)page
);
1459 static ssize_t
cgroup_file_read(struct file
*file
, char __user
*buf
,
1460 size_t nbytes
, loff_t
*ppos
)
1462 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1463 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
1465 if (!cft
|| cgroup_is_removed(cgrp
))
1469 return cft
->read(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1471 return cgroup_read_uint(cgrp
, cft
, file
, buf
, nbytes
, ppos
);
1475 static int cgroup_file_open(struct inode
*inode
, struct file
*file
)
1480 err
= generic_file_open(inode
, file
);
1484 cft
= __d_cft(file
->f_dentry
);
1488 err
= cft
->open(inode
, file
);
1495 static int cgroup_file_release(struct inode
*inode
, struct file
*file
)
1497 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1499 return cft
->release(inode
, file
);
1504 * cgroup_rename - Only allow simple rename of directories in place.
1506 static int cgroup_rename(struct inode
*old_dir
, struct dentry
*old_dentry
,
1507 struct inode
*new_dir
, struct dentry
*new_dentry
)
1509 if (!S_ISDIR(old_dentry
->d_inode
->i_mode
))
1511 if (new_dentry
->d_inode
)
1513 if (old_dir
!= new_dir
)
1515 return simple_rename(old_dir
, old_dentry
, new_dir
, new_dentry
);
1518 static struct file_operations cgroup_file_operations
= {
1519 .read
= cgroup_file_read
,
1520 .write
= cgroup_file_write
,
1521 .llseek
= generic_file_llseek
,
1522 .open
= cgroup_file_open
,
1523 .release
= cgroup_file_release
,
1526 static struct inode_operations cgroup_dir_inode_operations
= {
1527 .lookup
= simple_lookup
,
1528 .mkdir
= cgroup_mkdir
,
1529 .rmdir
= cgroup_rmdir
,
1530 .rename
= cgroup_rename
,
1533 static int cgroup_create_file(struct dentry
*dentry
, int mode
,
1534 struct super_block
*sb
)
1536 static struct dentry_operations cgroup_dops
= {
1537 .d_iput
= cgroup_diput
,
1540 struct inode
*inode
;
1544 if (dentry
->d_inode
)
1547 inode
= cgroup_new_inode(mode
, sb
);
1551 if (S_ISDIR(mode
)) {
1552 inode
->i_op
= &cgroup_dir_inode_operations
;
1553 inode
->i_fop
= &simple_dir_operations
;
1555 /* start off with i_nlink == 2 (for "." entry) */
1558 /* start with the directory inode held, so that we can
1559 * populate it without racing with another mkdir */
1560 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_CHILD
);
1561 } else if (S_ISREG(mode
)) {
1563 inode
->i_fop
= &cgroup_file_operations
;
1565 dentry
->d_op
= &cgroup_dops
;
1566 d_instantiate(dentry
, inode
);
1567 dget(dentry
); /* Extra count - pin the dentry in core */
1572 * cgroup_create_dir - create a directory for an object.
1573 * cgrp: the cgroup we create the directory for.
1574 * It must have a valid ->parent field
1575 * And we are going to fill its ->dentry field.
1576 * dentry: dentry of the new cgroup
1577 * mode: mode to set on new directory.
1579 static int cgroup_create_dir(struct cgroup
*cgrp
, struct dentry
*dentry
,
1582 struct dentry
*parent
;
1585 parent
= cgrp
->parent
->dentry
;
1586 error
= cgroup_create_file(dentry
, S_IFDIR
| mode
, cgrp
->root
->sb
);
1588 dentry
->d_fsdata
= cgrp
;
1589 inc_nlink(parent
->d_inode
);
1590 cgrp
->dentry
= dentry
;
1598 int cgroup_add_file(struct cgroup
*cgrp
,
1599 struct cgroup_subsys
*subsys
,
1600 const struct cftype
*cft
)
1602 struct dentry
*dir
= cgrp
->dentry
;
1603 struct dentry
*dentry
;
1606 char name
[MAX_CGROUP_TYPE_NAMELEN
+ MAX_CFTYPE_NAME
+ 2] = { 0 };
1607 if (subsys
&& !test_bit(ROOT_NOPREFIX
, &cgrp
->root
->flags
)) {
1608 strcpy(name
, subsys
->name
);
1611 strcat(name
, cft
->name
);
1612 BUG_ON(!mutex_is_locked(&dir
->d_inode
->i_mutex
));
1613 dentry
= lookup_one_len(name
, dir
, strlen(name
));
1614 if (!IS_ERR(dentry
)) {
1615 error
= cgroup_create_file(dentry
, 0644 | S_IFREG
,
1618 dentry
->d_fsdata
= (void *)cft
;
1621 error
= PTR_ERR(dentry
);
1625 int cgroup_add_files(struct cgroup
*cgrp
,
1626 struct cgroup_subsys
*subsys
,
1627 const struct cftype cft
[],
1631 for (i
= 0; i
< count
; i
++) {
1632 err
= cgroup_add_file(cgrp
, subsys
, &cft
[i
]);
1639 /* Count the number of tasks in a cgroup. */
1641 int cgroup_task_count(const struct cgroup
*cgrp
)
1644 struct list_head
*l
;
1646 read_lock(&css_set_lock
);
1647 l
= cgrp
->css_sets
.next
;
1648 while (l
!= &cgrp
->css_sets
) {
1649 struct cg_cgroup_link
*link
=
1650 list_entry(l
, struct cg_cgroup_link
, cgrp_link_list
);
1651 count
+= atomic_read(&link
->cg
->ref
.refcount
);
1654 read_unlock(&css_set_lock
);
1659 * Advance a list_head iterator. The iterator should be positioned at
1660 * the start of a css_set
1662 static void cgroup_advance_iter(struct cgroup
*cgrp
,
1663 struct cgroup_iter
*it
)
1665 struct list_head
*l
= it
->cg_link
;
1666 struct cg_cgroup_link
*link
;
1669 /* Advance to the next non-empty css_set */
1672 if (l
== &cgrp
->css_sets
) {
1676 link
= list_entry(l
, struct cg_cgroup_link
, cgrp_link_list
);
1678 } while (list_empty(&cg
->tasks
));
1680 it
->task
= cg
->tasks
.next
;
1683 void cgroup_iter_start(struct cgroup
*cgrp
, struct cgroup_iter
*it
)
1686 * The first time anyone tries to iterate across a cgroup,
1687 * we need to enable the list linking each css_set to its
1688 * tasks, and fix up all existing tasks.
1690 if (!use_task_css_set_links
) {
1691 struct task_struct
*p
, *g
;
1692 write_lock(&css_set_lock
);
1693 use_task_css_set_links
= 1;
1694 do_each_thread(g
, p
) {
1696 if (list_empty(&p
->cg_list
))
1697 list_add(&p
->cg_list
, &p
->cgroups
->tasks
);
1699 } while_each_thread(g
, p
);
1700 write_unlock(&css_set_lock
);
1702 read_lock(&css_set_lock
);
1703 it
->cg_link
= &cgrp
->css_sets
;
1704 cgroup_advance_iter(cgrp
, it
);
1707 struct task_struct
*cgroup_iter_next(struct cgroup
*cgrp
,
1708 struct cgroup_iter
*it
)
1710 struct task_struct
*res
;
1711 struct list_head
*l
= it
->task
;
1713 /* If the iterator cg is NULL, we have no tasks */
1716 res
= list_entry(l
, struct task_struct
, cg_list
);
1717 /* Advance iterator to find next entry */
1719 if (l
== &res
->cgroups
->tasks
) {
1720 /* We reached the end of this task list - move on to
1721 * the next cg_cgroup_link */
1722 cgroup_advance_iter(cgrp
, it
);
1729 void cgroup_iter_end(struct cgroup
*cgrp
, struct cgroup_iter
*it
)
1731 read_unlock(&css_set_lock
);
1735 * Stuff for reading the 'tasks' file.
1737 * Reading this file can return large amounts of data if a cgroup has
1738 * *lots* of attached tasks. So it may need several calls to read(),
1739 * but we cannot guarantee that the information we produce is correct
1740 * unless we produce it entirely atomically.
1742 * Upon tasks file open(), a struct ctr_struct is allocated, that
1743 * will have a pointer to an array (also allocated here). The struct
1744 * ctr_struct * is stored in file->private_data. Its resources will
1745 * be freed by release() when the file is closed. The array is used
1746 * to sprintf the PIDs and then used by read().
1754 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
1755 * 'cgrp'. Return actual number of pids loaded. No need to
1756 * task_lock(p) when reading out p->cgroup, since we're in an RCU
1757 * read section, so the css_set can't go away, and is
1758 * immutable after creation.
1760 static int pid_array_load(pid_t
*pidarray
, int npids
, struct cgroup
*cgrp
)
1763 struct cgroup_iter it
;
1764 struct task_struct
*tsk
;
1765 cgroup_iter_start(cgrp
, &it
);
1766 while ((tsk
= cgroup_iter_next(cgrp
, &it
))) {
1767 if (unlikely(n
== npids
))
1769 pidarray
[n
++] = task_pid_nr(tsk
);
1771 cgroup_iter_end(cgrp
, &it
);
1776 * Build and fill cgroupstats so that taskstats can export it to user
1779 * @stats: cgroupstats to fill information into
1780 * @dentry: A dentry entry belonging to the cgroup for which stats have
1783 int cgroupstats_build(struct cgroupstats
*stats
, struct dentry
*dentry
)
1786 struct cgroup
*cgrp
;
1787 struct cgroup_iter it
;
1788 struct task_struct
*tsk
;
1790 * Validate dentry by checking the superblock operations
1792 if (dentry
->d_sb
->s_op
!= &cgroup_ops
)
1796 cgrp
= dentry
->d_fsdata
;
1799 cgroup_iter_start(cgrp
, &it
);
1800 while ((tsk
= cgroup_iter_next(cgrp
, &it
))) {
1801 switch (tsk
->state
) {
1803 stats
->nr_running
++;
1805 case TASK_INTERRUPTIBLE
:
1806 stats
->nr_sleeping
++;
1808 case TASK_UNINTERRUPTIBLE
:
1809 stats
->nr_uninterruptible
++;
1812 stats
->nr_stopped
++;
1815 if (delayacct_is_task_waiting_on_io(tsk
))
1816 stats
->nr_io_wait
++;
1820 cgroup_iter_end(cgrp
, &it
);
1827 static int cmppid(const void *a
, const void *b
)
1829 return *(pid_t
*)a
- *(pid_t
*)b
;
1833 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1834 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1835 * count 'cnt' of how many chars would be written if buf were large enough.
1837 static int pid_array_to_buf(char *buf
, int sz
, pid_t
*a
, int npids
)
1842 for (i
= 0; i
< npids
; i
++)
1843 cnt
+= snprintf(buf
+ cnt
, max(sz
- cnt
, 0), "%d\n", a
[i
]);
1848 * Handle an open on 'tasks' file. Prepare a buffer listing the
1849 * process id's of tasks currently attached to the cgroup being opened.
1851 * Does not require any specific cgroup mutexes, and does not take any.
1853 static int cgroup_tasks_open(struct inode
*unused
, struct file
*file
)
1855 struct cgroup
*cgrp
= __d_cgrp(file
->f_dentry
->d_parent
);
1856 struct ctr_struct
*ctr
;
1861 if (!(file
->f_mode
& FMODE_READ
))
1864 ctr
= kmalloc(sizeof(*ctr
), GFP_KERNEL
);
1869 * If cgroup gets more users after we read count, we won't have
1870 * enough space - tough. This race is indistinguishable to the
1871 * caller from the case that the additional cgroup users didn't
1872 * show up until sometime later on.
1874 npids
= cgroup_task_count(cgrp
);
1876 pidarray
= kmalloc(npids
* sizeof(pid_t
), GFP_KERNEL
);
1880 npids
= pid_array_load(pidarray
, npids
, cgrp
);
1881 sort(pidarray
, npids
, sizeof(pid_t
), cmppid
, NULL
);
1883 /* Call pid_array_to_buf() twice, first just to get bufsz */
1884 ctr
->bufsz
= pid_array_to_buf(&c
, sizeof(c
), pidarray
, npids
) + 1;
1885 ctr
->buf
= kmalloc(ctr
->bufsz
, GFP_KERNEL
);
1888 ctr
->bufsz
= pid_array_to_buf(ctr
->buf
, ctr
->bufsz
, pidarray
, npids
);
1895 file
->private_data
= ctr
;
1906 static ssize_t
cgroup_tasks_read(struct cgroup
*cgrp
,
1908 struct file
*file
, char __user
*buf
,
1909 size_t nbytes
, loff_t
*ppos
)
1911 struct ctr_struct
*ctr
= file
->private_data
;
1913 return simple_read_from_buffer(buf
, nbytes
, ppos
, ctr
->buf
, ctr
->bufsz
);
1916 static int cgroup_tasks_release(struct inode
*unused_inode
,
1919 struct ctr_struct
*ctr
;
1921 if (file
->f_mode
& FMODE_READ
) {
1922 ctr
= file
->private_data
;
1929 static u64
cgroup_read_notify_on_release(struct cgroup
*cgrp
,
1932 return notify_on_release(cgrp
);
1935 static u64
cgroup_read_releasable(struct cgroup
*cgrp
, struct cftype
*cft
)
1937 return test_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
1941 * for the common functions, 'private' gives the type of file
1943 static struct cftype files
[] = {
1946 .open
= cgroup_tasks_open
,
1947 .read
= cgroup_tasks_read
,
1948 .write
= cgroup_common_file_write
,
1949 .release
= cgroup_tasks_release
,
1950 .private = FILE_TASKLIST
,
1954 .name
= "notify_on_release",
1955 .read_uint
= cgroup_read_notify_on_release
,
1956 .write
= cgroup_common_file_write
,
1957 .private = FILE_NOTIFY_ON_RELEASE
,
1961 .name
= "releasable",
1962 .read_uint
= cgroup_read_releasable
,
1963 .private = FILE_RELEASABLE
,
1967 static struct cftype cft_release_agent
= {
1968 .name
= "release_agent",
1969 .read
= cgroup_common_file_read
,
1970 .write
= cgroup_common_file_write
,
1971 .private = FILE_RELEASE_AGENT
,
1974 static int cgroup_populate_dir(struct cgroup
*cgrp
)
1977 struct cgroup_subsys
*ss
;
1979 /* First clear out any existing files */
1980 cgroup_clear_directory(cgrp
->dentry
);
1982 err
= cgroup_add_files(cgrp
, NULL
, files
, ARRAY_SIZE(files
));
1986 if (cgrp
== cgrp
->top_cgroup
) {
1987 if ((err
= cgroup_add_file(cgrp
, NULL
, &cft_release_agent
)) < 0)
1991 for_each_subsys(cgrp
->root
, ss
) {
1992 if (ss
->populate
&& (err
= ss
->populate(ss
, cgrp
)) < 0)
1999 static void init_cgroup_css(struct cgroup_subsys_state
*css
,
2000 struct cgroup_subsys
*ss
,
2001 struct cgroup
*cgrp
)
2004 atomic_set(&css
->refcnt
, 0);
2006 if (cgrp
== dummytop
)
2007 set_bit(CSS_ROOT
, &css
->flags
);
2008 BUG_ON(cgrp
->subsys
[ss
->subsys_id
]);
2009 cgrp
->subsys
[ss
->subsys_id
] = css
;
2013 * cgroup_create - create a cgroup
2014 * parent: cgroup that will be parent of the new cgroup.
2015 * name: name of the new cgroup. Will be strcpy'ed.
2016 * mode: mode to set on new inode
2018 * Must be called with the mutex on the parent inode held
2021 static long cgroup_create(struct cgroup
*parent
, struct dentry
*dentry
,
2024 struct cgroup
*cgrp
;
2025 struct cgroupfs_root
*root
= parent
->root
;
2027 struct cgroup_subsys
*ss
;
2028 struct super_block
*sb
= root
->sb
;
2030 cgrp
= kzalloc(sizeof(*cgrp
), GFP_KERNEL
);
2034 /* Grab a reference on the superblock so the hierarchy doesn't
2035 * get deleted on unmount if there are child cgroups. This
2036 * can be done outside cgroup_mutex, since the sb can't
2037 * disappear while someone has an open control file on the
2039 atomic_inc(&sb
->s_active
);
2041 mutex_lock(&cgroup_mutex
);
2044 INIT_LIST_HEAD(&cgrp
->sibling
);
2045 INIT_LIST_HEAD(&cgrp
->children
);
2046 INIT_LIST_HEAD(&cgrp
->css_sets
);
2047 INIT_LIST_HEAD(&cgrp
->release_list
);
2049 cgrp
->parent
= parent
;
2050 cgrp
->root
= parent
->root
;
2051 cgrp
->top_cgroup
= parent
->top_cgroup
;
2053 for_each_subsys(root
, ss
) {
2054 struct cgroup_subsys_state
*css
= ss
->create(ss
, cgrp
);
2059 init_cgroup_css(css
, ss
, cgrp
);
2062 list_add(&cgrp
->sibling
, &cgrp
->parent
->children
);
2063 root
->number_of_cgroups
++;
2065 err
= cgroup_create_dir(cgrp
, dentry
, mode
);
2069 /* The cgroup directory was pre-locked for us */
2070 BUG_ON(!mutex_is_locked(&cgrp
->dentry
->d_inode
->i_mutex
));
2072 err
= cgroup_populate_dir(cgrp
);
2073 /* If err < 0, we have a half-filled directory - oh well ;) */
2075 mutex_unlock(&cgroup_mutex
);
2076 mutex_unlock(&cgrp
->dentry
->d_inode
->i_mutex
);
2082 list_del(&cgrp
->sibling
);
2083 root
->number_of_cgroups
--;
2087 for_each_subsys(root
, ss
) {
2088 if (cgrp
->subsys
[ss
->subsys_id
])
2089 ss
->destroy(ss
, cgrp
);
2092 mutex_unlock(&cgroup_mutex
);
2094 /* Release the reference count that we took on the superblock */
2095 deactivate_super(sb
);
2101 static int cgroup_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
)
2103 struct cgroup
*c_parent
= dentry
->d_parent
->d_fsdata
;
2105 /* the vfs holds inode->i_mutex already */
2106 return cgroup_create(c_parent
, dentry
, mode
| S_IFDIR
);
2109 static inline int cgroup_has_css_refs(struct cgroup
*cgrp
)
2111 /* Check the reference count on each subsystem. Since we
2112 * already established that there are no tasks in the
2113 * cgroup, if the css refcount is also 0, then there should
2114 * be no outstanding references, so the subsystem is safe to
2115 * destroy. We scan across all subsystems rather than using
2116 * the per-hierarchy linked list of mounted subsystems since
2117 * we can be called via check_for_release() with no
2118 * synchronization other than RCU, and the subsystem linked
2119 * list isn't RCU-safe */
2121 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2122 struct cgroup_subsys
*ss
= subsys
[i
];
2123 struct cgroup_subsys_state
*css
;
2124 /* Skip subsystems not in this hierarchy */
2125 if (ss
->root
!= cgrp
->root
)
2127 css
= cgrp
->subsys
[ss
->subsys_id
];
2128 /* When called from check_for_release() it's possible
2129 * that by this point the cgroup has been removed
2130 * and the css deleted. But a false-positive doesn't
2131 * matter, since it can only happen if the cgroup
2132 * has been deleted and hence no longer needs the
2133 * release agent to be called anyway. */
2134 if (css
&& atomic_read(&css
->refcnt
))
2140 static int cgroup_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
)
2142 struct cgroup
*cgrp
= dentry
->d_fsdata
;
2144 struct cgroup
*parent
;
2145 struct super_block
*sb
;
2146 struct cgroupfs_root
*root
;
2148 /* the vfs holds both inode->i_mutex already */
2150 mutex_lock(&cgroup_mutex
);
2151 if (atomic_read(&cgrp
->count
) != 0) {
2152 mutex_unlock(&cgroup_mutex
);
2155 if (!list_empty(&cgrp
->children
)) {
2156 mutex_unlock(&cgroup_mutex
);
2160 parent
= cgrp
->parent
;
2164 if (cgroup_has_css_refs(cgrp
)) {
2165 mutex_unlock(&cgroup_mutex
);
2169 spin_lock(&release_list_lock
);
2170 set_bit(CGRP_REMOVED
, &cgrp
->flags
);
2171 if (!list_empty(&cgrp
->release_list
))
2172 list_del(&cgrp
->release_list
);
2173 spin_unlock(&release_list_lock
);
2174 /* delete my sibling from parent->children */
2175 list_del(&cgrp
->sibling
);
2176 spin_lock(&cgrp
->dentry
->d_lock
);
2177 d
= dget(cgrp
->dentry
);
2178 cgrp
->dentry
= NULL
;
2179 spin_unlock(&d
->d_lock
);
2181 cgroup_d_remove_dir(d
);
2184 set_bit(CGRP_RELEASABLE
, &parent
->flags
);
2185 check_for_release(parent
);
2187 mutex_unlock(&cgroup_mutex
);
2191 static void cgroup_init_subsys(struct cgroup_subsys
*ss
)
2193 struct cgroup_subsys_state
*css
;
2194 struct list_head
*l
;
2196 printk(KERN_INFO
"Initializing cgroup subsys %s\n", ss
->name
);
2198 /* Create the top cgroup state for this subsystem */
2199 ss
->root
= &rootnode
;
2200 css
= ss
->create(ss
, dummytop
);
2201 /* We don't handle early failures gracefully */
2202 BUG_ON(IS_ERR(css
));
2203 init_cgroup_css(css
, ss
, dummytop
);
2205 /* Update all cgroup groups to contain a subsys
2206 * pointer to this state - since the subsystem is
2207 * newly registered, all tasks and hence all cgroup
2208 * groups are in the subsystem's top cgroup. */
2209 write_lock(&css_set_lock
);
2210 l
= &init_css_set
.list
;
2212 struct css_set
*cg
=
2213 list_entry(l
, struct css_set
, list
);
2214 cg
->subsys
[ss
->subsys_id
] = dummytop
->subsys
[ss
->subsys_id
];
2216 } while (l
!= &init_css_set
.list
);
2217 write_unlock(&css_set_lock
);
2219 /* If this subsystem requested that it be notified with fork
2220 * events, we should send it one now for every process in the
2223 struct task_struct
*g
, *p
;
2225 read_lock(&tasklist_lock
);
2226 do_each_thread(g
, p
) {
2228 } while_each_thread(g
, p
);
2229 read_unlock(&tasklist_lock
);
2232 need_forkexit_callback
|= ss
->fork
|| ss
->exit
;
2238 * cgroup_init_early - initialize cgroups at system boot, and
2239 * initialize any subsystems that request early init.
2241 int __init
cgroup_init_early(void)
2244 kref_init(&init_css_set
.ref
);
2245 kref_get(&init_css_set
.ref
);
2246 INIT_LIST_HEAD(&init_css_set
.list
);
2247 INIT_LIST_HEAD(&init_css_set
.cg_links
);
2248 INIT_LIST_HEAD(&init_css_set
.tasks
);
2250 init_cgroup_root(&rootnode
);
2251 list_add(&rootnode
.root_list
, &roots
);
2253 init_task
.cgroups
= &init_css_set
;
2255 init_css_set_link
.cg
= &init_css_set
;
2256 list_add(&init_css_set_link
.cgrp_link_list
,
2257 &rootnode
.top_cgroup
.css_sets
);
2258 list_add(&init_css_set_link
.cg_link_list
,
2259 &init_css_set
.cg_links
);
2261 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2262 struct cgroup_subsys
*ss
= subsys
[i
];
2265 BUG_ON(strlen(ss
->name
) > MAX_CGROUP_TYPE_NAMELEN
);
2266 BUG_ON(!ss
->create
);
2267 BUG_ON(!ss
->destroy
);
2268 if (ss
->subsys_id
!= i
) {
2269 printk(KERN_ERR
"cgroup: Subsys %s id == %d\n",
2270 ss
->name
, ss
->subsys_id
);
2275 cgroup_init_subsys(ss
);
2281 * cgroup_init - register cgroup filesystem and /proc file, and
2282 * initialize any subsystems that didn't request early init.
2284 int __init
cgroup_init(void)
2288 struct proc_dir_entry
*entry
;
2290 err
= bdi_init(&cgroup_backing_dev_info
);
2294 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2295 struct cgroup_subsys
*ss
= subsys
[i
];
2296 if (!ss
->early_init
)
2297 cgroup_init_subsys(ss
);
2300 err
= register_filesystem(&cgroup_fs_type
);
2304 entry
= create_proc_entry("cgroups", 0, NULL
);
2306 entry
->proc_fops
= &proc_cgroupstats_operations
;
2310 bdi_destroy(&cgroup_backing_dev_info
);
2316 * proc_cgroup_show()
2317 * - Print task's cgroup paths into seq_file, one line for each hierarchy
2318 * - Used for /proc/<pid>/cgroup.
2319 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2320 * doesn't really matter if tsk->cgroup changes after we read it,
2321 * and we take cgroup_mutex, keeping attach_task() from changing it
2322 * anyway. No need to check that tsk->cgroup != NULL, thanks to
2323 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2324 * cgroup to top_cgroup.
2327 /* TODO: Use a proper seq_file iterator */
2328 static int proc_cgroup_show(struct seq_file
*m
, void *v
)
2331 struct task_struct
*tsk
;
2334 struct cgroupfs_root
*root
;
2337 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2343 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2349 mutex_lock(&cgroup_mutex
);
2351 for_each_root(root
) {
2352 struct cgroup_subsys
*ss
;
2353 struct cgroup
*cgrp
;
2357 /* Skip this hierarchy if it has no active subsystems */
2358 if (!root
->actual_subsys_bits
)
2360 for_each_subsys(root
, ss
)
2361 seq_printf(m
, "%s%s", count
++ ? "," : "", ss
->name
);
2363 get_first_subsys(&root
->top_cgroup
, NULL
, &subsys_id
);
2364 cgrp
= task_cgroup(tsk
, subsys_id
);
2365 retval
= cgroup_path(cgrp
, buf
, PAGE_SIZE
);
2373 mutex_unlock(&cgroup_mutex
);
2374 put_task_struct(tsk
);
2381 static int cgroup_open(struct inode
*inode
, struct file
*file
)
2383 struct pid
*pid
= PROC_I(inode
)->pid
;
2384 return single_open(file
, proc_cgroup_show
, pid
);
2387 struct file_operations proc_cgroup_operations
= {
2388 .open
= cgroup_open
,
2390 .llseek
= seq_lseek
,
2391 .release
= single_release
,
2394 /* Display information about each subsystem and each hierarchy */
2395 static int proc_cgroupstats_show(struct seq_file
*m
, void *v
)
2399 seq_puts(m
, "#subsys_name\thierarchy\tnum_cgroups\n");
2400 mutex_lock(&cgroup_mutex
);
2401 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2402 struct cgroup_subsys
*ss
= subsys
[i
];
2403 seq_printf(m
, "%s\t%lu\t%d\n",
2404 ss
->name
, ss
->root
->subsys_bits
,
2405 ss
->root
->number_of_cgroups
);
2407 mutex_unlock(&cgroup_mutex
);
2411 static int cgroupstats_open(struct inode
*inode
, struct file
*file
)
2413 return single_open(file
, proc_cgroupstats_show
, 0);
2416 static struct file_operations proc_cgroupstats_operations
= {
2417 .open
= cgroupstats_open
,
2419 .llseek
= seq_lseek
,
2420 .release
= single_release
,
2424 * cgroup_fork - attach newly forked task to its parents cgroup.
2425 * @tsk: pointer to task_struct of forking parent process.
2427 * Description: A task inherits its parent's cgroup at fork().
2429 * A pointer to the shared css_set was automatically copied in
2430 * fork.c by dup_task_struct(). However, we ignore that copy, since
2431 * it was not made under the protection of RCU or cgroup_mutex, so
2432 * might no longer be a valid cgroup pointer. attach_task() might
2433 * have already changed current->cgroups, allowing the previously
2434 * referenced cgroup group to be removed and freed.
2436 * At the point that cgroup_fork() is called, 'current' is the parent
2437 * task, and the passed argument 'child' points to the child task.
2439 void cgroup_fork(struct task_struct
*child
)
2442 child
->cgroups
= current
->cgroups
;
2443 get_css_set(child
->cgroups
);
2444 task_unlock(current
);
2445 INIT_LIST_HEAD(&child
->cg_list
);
2449 * cgroup_fork_callbacks - called on a new task very soon before
2450 * adding it to the tasklist. No need to take any locks since no-one
2451 * can be operating on this task
2453 void cgroup_fork_callbacks(struct task_struct
*child
)
2455 if (need_forkexit_callback
) {
2457 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2458 struct cgroup_subsys
*ss
= subsys
[i
];
2460 ss
->fork(ss
, child
);
2466 * cgroup_post_fork - called on a new task after adding it to the
2467 * task list. Adds the task to the list running through its css_set
2468 * if necessary. Has to be after the task is visible on the task list
2469 * in case we race with the first call to cgroup_iter_start() - to
2470 * guarantee that the new task ends up on its list. */
2471 void cgroup_post_fork(struct task_struct
*child
)
2473 if (use_task_css_set_links
) {
2474 write_lock(&css_set_lock
);
2475 if (list_empty(&child
->cg_list
))
2476 list_add(&child
->cg_list
, &child
->cgroups
->tasks
);
2477 write_unlock(&css_set_lock
);
2481 * cgroup_exit - detach cgroup from exiting task
2482 * @tsk: pointer to task_struct of exiting process
2484 * Description: Detach cgroup from @tsk and release it.
2486 * Note that cgroups marked notify_on_release force every task in
2487 * them to take the global cgroup_mutex mutex when exiting.
2488 * This could impact scaling on very large systems. Be reluctant to
2489 * use notify_on_release cgroups where very high task exit scaling
2490 * is required on large systems.
2492 * the_top_cgroup_hack:
2494 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2496 * We call cgroup_exit() while the task is still competent to
2497 * handle notify_on_release(), then leave the task attached to the
2498 * root cgroup in each hierarchy for the remainder of its exit.
2500 * To do this properly, we would increment the reference count on
2501 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
2502 * code we would add a second cgroup function call, to drop that
2503 * reference. This would just create an unnecessary hot spot on
2504 * the top_cgroup reference count, to no avail.
2506 * Normally, holding a reference to a cgroup without bumping its
2507 * count is unsafe. The cgroup could go away, or someone could
2508 * attach us to a different cgroup, decrementing the count on
2509 * the first cgroup that we never incremented. But in this case,
2510 * top_cgroup isn't going away, and either task has PF_EXITING set,
2511 * which wards off any attach_task() attempts, or task is a failed
2512 * fork, never visible to attach_task.
2515 void cgroup_exit(struct task_struct
*tsk
, int run_callbacks
)
2520 if (run_callbacks
&& need_forkexit_callback
) {
2521 for (i
= 0; i
< CGROUP_SUBSYS_COUNT
; i
++) {
2522 struct cgroup_subsys
*ss
= subsys
[i
];
2529 * Unlink from the css_set task list if necessary.
2530 * Optimistically check cg_list before taking
2533 if (!list_empty(&tsk
->cg_list
)) {
2534 write_lock(&css_set_lock
);
2535 if (!list_empty(&tsk
->cg_list
))
2536 list_del(&tsk
->cg_list
);
2537 write_unlock(&css_set_lock
);
2540 /* Reassign the task to the init_css_set. */
2543 tsk
->cgroups
= &init_css_set
;
2546 put_css_set_taskexit(cg
);
2550 * cgroup_clone - duplicate the current cgroup in the hierarchy
2551 * that the given subsystem is attached to, and move this task into
2554 int cgroup_clone(struct task_struct
*tsk
, struct cgroup_subsys
*subsys
)
2556 struct dentry
*dentry
;
2558 char nodename
[MAX_CGROUP_TYPE_NAMELEN
];
2559 struct cgroup
*parent
, *child
;
2560 struct inode
*inode
;
2562 struct cgroupfs_root
*root
;
2563 struct cgroup_subsys
*ss
;
2565 /* We shouldn't be called by an unregistered subsystem */
2566 BUG_ON(!subsys
->active
);
2568 /* First figure out what hierarchy and cgroup we're dealing
2569 * with, and pin them so we can drop cgroup_mutex */
2570 mutex_lock(&cgroup_mutex
);
2572 root
= subsys
->root
;
2573 if (root
== &rootnode
) {
2575 "Not cloning cgroup for unused subsystem %s\n",
2577 mutex_unlock(&cgroup_mutex
);
2581 parent
= task_cgroup(tsk
, subsys
->subsys_id
);
2583 snprintf(nodename
, MAX_CGROUP_TYPE_NAMELEN
, "node_%d", tsk
->pid
);
2585 /* Pin the hierarchy */
2586 atomic_inc(&parent
->root
->sb
->s_active
);
2588 /* Keep the cgroup alive */
2590 mutex_unlock(&cgroup_mutex
);
2592 /* Now do the VFS work to create a cgroup */
2593 inode
= parent
->dentry
->d_inode
;
2595 /* Hold the parent directory mutex across this operation to
2596 * stop anyone else deleting the new cgroup */
2597 mutex_lock(&inode
->i_mutex
);
2598 dentry
= lookup_one_len(nodename
, parent
->dentry
, strlen(nodename
));
2599 if (IS_ERR(dentry
)) {
2601 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename
,
2603 ret
= PTR_ERR(dentry
);
2607 /* Create the cgroup directory, which also creates the cgroup */
2608 ret
= vfs_mkdir(inode
, dentry
, S_IFDIR
| 0755);
2609 child
= __d_cgrp(dentry
);
2613 "Failed to create cgroup %s: %d\n", nodename
,
2620 "Couldn't find new cgroup %s\n", nodename
);
2625 /* The cgroup now exists. Retake cgroup_mutex and check
2626 * that we're still in the same state that we thought we
2628 mutex_lock(&cgroup_mutex
);
2629 if ((root
!= subsys
->root
) ||
2630 (parent
!= task_cgroup(tsk
, subsys
->subsys_id
))) {
2631 /* Aargh, we raced ... */
2632 mutex_unlock(&inode
->i_mutex
);
2635 deactivate_super(parent
->root
->sb
);
2636 /* The cgroup is still accessible in the VFS, but
2637 * we're not going to try to rmdir() it at this
2640 "Race in cgroup_clone() - leaking cgroup %s\n",
2645 /* do any required auto-setup */
2646 for_each_subsys(root
, ss
) {
2648 ss
->post_clone(ss
, child
);
2651 /* All seems fine. Finish by moving the task into the new cgroup */
2652 ret
= attach_task(child
, tsk
);
2653 mutex_unlock(&cgroup_mutex
);
2656 mutex_unlock(&inode
->i_mutex
);
2658 mutex_lock(&cgroup_mutex
);
2660 mutex_unlock(&cgroup_mutex
);
2661 deactivate_super(parent
->root
->sb
);
2666 * See if "cgrp" is a descendant of the current task's cgroup in
2667 * the appropriate hierarchy
2669 * If we are sending in dummytop, then presumably we are creating
2670 * the top cgroup in the subsystem.
2672 * Called only by the ns (nsproxy) cgroup.
2674 int cgroup_is_descendant(const struct cgroup
*cgrp
)
2677 struct cgroup
*target
;
2680 if (cgrp
== dummytop
)
2683 get_first_subsys(cgrp
, NULL
, &subsys_id
);
2684 target
= task_cgroup(current
, subsys_id
);
2685 while (cgrp
!= target
&& cgrp
!= cgrp
->top_cgroup
)
2686 cgrp
= cgrp
->parent
;
2687 ret
= (cgrp
== target
);
2691 static void check_for_release(struct cgroup
*cgrp
)
2693 /* All of these checks rely on RCU to keep the cgroup
2694 * structure alive */
2695 if (cgroup_is_releasable(cgrp
) && !atomic_read(&cgrp
->count
)
2696 && list_empty(&cgrp
->children
) && !cgroup_has_css_refs(cgrp
)) {
2697 /* Control Group is currently removeable. If it's not
2698 * already queued for a userspace notification, queue
2700 int need_schedule_work
= 0;
2701 spin_lock(&release_list_lock
);
2702 if (!cgroup_is_removed(cgrp
) &&
2703 list_empty(&cgrp
->release_list
)) {
2704 list_add(&cgrp
->release_list
, &release_list
);
2705 need_schedule_work
= 1;
2707 spin_unlock(&release_list_lock
);
2708 if (need_schedule_work
)
2709 schedule_work(&release_agent_work
);
2713 void __css_put(struct cgroup_subsys_state
*css
)
2715 struct cgroup
*cgrp
= css
->cgroup
;
2717 if (atomic_dec_and_test(&css
->refcnt
) && notify_on_release(cgrp
)) {
2718 set_bit(CGRP_RELEASABLE
, &cgrp
->flags
);
2719 check_for_release(cgrp
);
2725 * Notify userspace when a cgroup is released, by running the
2726 * configured release agent with the name of the cgroup (path
2727 * relative to the root of cgroup file system) as the argument.
2729 * Most likely, this user command will try to rmdir this cgroup.
2731 * This races with the possibility that some other task will be
2732 * attached to this cgroup before it is removed, or that some other
2733 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
2734 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
2735 * unused, and this cgroup will be reprieved from its death sentence,
2736 * to continue to serve a useful existence. Next time it's released,
2737 * we will get notified again, if it still has 'notify_on_release' set.
2739 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
2740 * means only wait until the task is successfully execve()'d. The
2741 * separate release agent task is forked by call_usermodehelper(),
2742 * then control in this thread returns here, without waiting for the
2743 * release agent task. We don't bother to wait because the caller of
2744 * this routine has no use for the exit status of the release agent
2745 * task, so no sense holding our caller up for that.
2749 static void cgroup_release_agent(struct work_struct
*work
)
2751 BUG_ON(work
!= &release_agent_work
);
2752 mutex_lock(&cgroup_mutex
);
2753 spin_lock(&release_list_lock
);
2754 while (!list_empty(&release_list
)) {
2755 char *argv
[3], *envp
[3];
2758 struct cgroup
*cgrp
= list_entry(release_list
.next
,
2761 list_del_init(&cgrp
->release_list
);
2762 spin_unlock(&release_list_lock
);
2763 pathbuf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2765 spin_lock(&release_list_lock
);
2769 if (cgroup_path(cgrp
, pathbuf
, PAGE_SIZE
) < 0) {
2771 spin_lock(&release_list_lock
);
2776 argv
[i
++] = cgrp
->root
->release_agent_path
;
2777 argv
[i
++] = (char *)pathbuf
;
2781 /* minimal command environment */
2782 envp
[i
++] = "HOME=/";
2783 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
2786 /* Drop the lock while we invoke the usermode helper,
2787 * since the exec could involve hitting disk and hence
2788 * be a slow process */
2789 mutex_unlock(&cgroup_mutex
);
2790 call_usermodehelper(argv
[0], argv
, envp
, UMH_WAIT_EXEC
);
2792 mutex_lock(&cgroup_mutex
);
2793 spin_lock(&release_list_lock
);
2795 spin_unlock(&release_list_lock
);
2796 mutex_unlock(&cgroup_mutex
);