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 * check for fork/exit handlers to call. This avoids us having to do
117 * extra work in the fork/exit path if none of the subsystems need to
120 static int need_forkexit_callback;
122 /* convenient tests for these bits */
123 inline int cgroup_is_removed(const struct cgroup *cgrp)
125 return test_bit(CGRP_REMOVED, &cgrp->flags);
128 /* bits in struct cgroupfs_root flags field */
130 ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
133 static int cgroup_is_releasable(const struct cgroup *cgrp)
136 (1 << CGRP_RELEASABLE) |
137 (1 << CGRP_NOTIFY_ON_RELEASE);
138 return (cgrp->flags & bits) == bits;
141 static int notify_on_release(const struct cgroup *cgrp)
143 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
147 * for_each_subsys() allows you to iterate on each subsystem attached to
148 * an active hierarchy
150 #define for_each_subsys(_root, _ss) \
151 list_for_each_entry(_ss, &_root->subsys_list, sibling)
153 /* for_each_root() allows you to iterate across the active hierarchies */
154 #define for_each_root(_root) \
155 list_for_each_entry(_root, &roots, root_list)
157 /* the list of cgroups eligible for automatic release. Protected by
158 * release_list_lock */
159 static LIST_HEAD(release_list);
160 static DEFINE_SPINLOCK(release_list_lock);
161 static void cgroup_release_agent(struct work_struct *work);
162 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
163 static void check_for_release(struct cgroup *cgrp);
165 /* Link structure for associating css_set objects with cgroups */
166 struct cg_cgroup_link {
168 * List running through cg_cgroup_links associated with a
169 * cgroup, anchored on cgroup->css_sets
171 struct list_head cgrp_link_list;
173 * List running through cg_cgroup_links pointing at a
174 * single css_set object, anchored on css_set->cg_links
176 struct list_head cg_link_list;
180 /* The default css_set - used by init and its children prior to any
181 * hierarchies being mounted. It contains a pointer to the root state
182 * for each subsystem. Also used to anchor the list of css_sets. Not
183 * reference-counted, to improve performance when child cgroups
184 * haven't been created.
187 static struct css_set init_css_set;
188 static struct cg_cgroup_link init_css_set_link;
190 /* css_set_lock protects the list of css_set objects, and the
191 * chain of tasks off each css_set. Nests outside task->alloc_lock
192 * due to cgroup_iter_start() */
193 static DEFINE_RWLOCK(css_set_lock);
194 static int css_set_count;
196 /* We don't maintain the lists running through each css_set to its
197 * task until after the first call to cgroup_iter_start(). This
198 * reduces the fork()/exit() overhead for people who have cgroups
199 * compiled into their kernel but not actually in use */
200 static int use_task_css_set_links;
202 /* When we create or destroy a css_set, the operation simply
203 * takes/releases a reference count on all the cgroups referenced
204 * by subsystems in this css_set. This can end up multiple-counting
205 * some cgroups, but that's OK - the ref-count is just a
206 * busy/not-busy indicator; ensuring that we only count each cgroup
207 * once would require taking a global lock to ensure that no
208 * subsystems moved between hierarchies while we were doing so.
210 * Possible TODO: decide at boot time based on the number of
211 * registered subsystems and the number of CPUs or NUMA nodes whether
212 * it's better for performance to ref-count every subsystem, or to
213 * take a global lock and only add one ref count to each hierarchy.
217 * unlink a css_set from the list and free it
219 static void unlink_css_set(struct css_set *cg)
221 write_lock(&css_set_lock);
224 while (!list_empty(&cg->cg_links)) {
225 struct cg_cgroup_link *link;
226 link = list_entry(cg->cg_links.next,
227 struct cg_cgroup_link, cg_link_list);
228 list_del(&link->cg_link_list);
229 list_del(&link->cgrp_link_list);
232 write_unlock(&css_set_lock);
235 static void __release_css_set(struct kref *k, int taskexit)
238 struct css_set *cg = container_of(k, struct css_set, ref);
243 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
244 struct cgroup *cgrp = cg->subsys[i]->cgroup;
245 if (atomic_dec_and_test(&cgrp->count) &&
246 notify_on_release(cgrp)) {
248 set_bit(CGRP_RELEASABLE, &cgrp->flags);
249 check_for_release(cgrp);
256 static void release_css_set(struct kref *k)
258 __release_css_set(k, 0);
261 static void release_css_set_taskexit(struct kref *k)
263 __release_css_set(k, 1);
267 * refcounted get/put for css_set objects
269 static inline void get_css_set(struct css_set *cg)
274 static inline void put_css_set(struct css_set *cg)
276 kref_put(&cg->ref, release_css_set);
279 static inline void put_css_set_taskexit(struct css_set *cg)
281 kref_put(&cg->ref, release_css_set_taskexit);
285 * find_existing_css_set() is a helper for
286 * find_css_set(), and checks to see whether an existing
287 * css_set is suitable. This currently walks a linked-list for
288 * simplicity; a later patch will use a hash table for better
291 * oldcg: the cgroup group that we're using before the cgroup
294 * cgrp: the cgroup that we're moving into
296 * template: location in which to build the desired set of subsystem
297 * state objects for the new cgroup group
299 static struct css_set *find_existing_css_set(
300 struct css_set *oldcg,
302 struct cgroup_subsys_state *template[])
305 struct cgroupfs_root *root = cgrp->root;
306 struct list_head *l = &init_css_set.list;
308 /* Built the set of subsystem state objects that we want to
309 * see in the new css_set */
310 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
311 if (root->subsys_bits & (1UL << i)) {
312 /* Subsystem is in this hierarchy. So we want
313 * the subsystem state from the new
315 template[i] = cgrp->subsys[i];
317 /* Subsystem is not in this hierarchy, so we
318 * don't want to change the subsystem state */
319 template[i] = oldcg->subsys[i];
323 /* Look through existing cgroup groups to find one to reuse */
326 list_entry(l, struct css_set, list);
328 if (!memcmp(template, cg->subsys, sizeof(cg->subsys))) {
329 /* All subsystems matched */
332 /* Try the next cgroup group */
334 } while (l != &init_css_set.list);
336 /* No existing cgroup group matched */
341 * allocate_cg_links() allocates "count" cg_cgroup_link structures
342 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
343 * success or a negative error
345 static int allocate_cg_links(int count, struct list_head *tmp)
347 struct cg_cgroup_link *link;
350 for (i = 0; i < count; i++) {
351 link = kmalloc(sizeof(*link), GFP_KERNEL);
353 while (!list_empty(tmp)) {
354 link = list_entry(tmp->next,
355 struct cg_cgroup_link,
357 list_del(&link->cgrp_link_list);
362 list_add(&link->cgrp_link_list, tmp);
367 static void free_cg_links(struct list_head *tmp)
369 while (!list_empty(tmp)) {
370 struct cg_cgroup_link *link;
371 link = list_entry(tmp->next,
372 struct cg_cgroup_link,
374 list_del(&link->cgrp_link_list);
380 * find_css_set() takes an existing cgroup group and a
381 * cgroup object, and returns a css_set object that's
382 * equivalent to the old group, but with the given cgroup
383 * substituted into the appropriate hierarchy. Must be called with
386 static struct css_set *find_css_set(
387 struct css_set *oldcg, struct cgroup *cgrp)
390 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
393 struct list_head tmp_cg_links;
394 struct cg_cgroup_link *link;
396 /* First see if we already have a cgroup group that matches
398 write_lock(&css_set_lock);
399 res = find_existing_css_set(oldcg, cgrp, template);
402 write_unlock(&css_set_lock);
407 res = kmalloc(sizeof(*res), GFP_KERNEL);
411 /* Allocate all the cg_cgroup_link objects that we'll need */
412 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
417 kref_init(&res->ref);
418 INIT_LIST_HEAD(&res->cg_links);
419 INIT_LIST_HEAD(&res->tasks);
421 /* Copy the set of subsystem state objects generated in
422 * find_existing_css_set() */
423 memcpy(res->subsys, template, sizeof(res->subsys));
425 write_lock(&css_set_lock);
426 /* Add reference counts and links from the new css_set. */
427 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
428 struct cgroup *cgrp = res->subsys[i]->cgroup;
429 struct cgroup_subsys *ss = subsys[i];
430 atomic_inc(&cgrp->count);
432 * We want to add a link once per cgroup, so we
433 * only do it for the first subsystem in each
436 if (ss->root->subsys_list.next == &ss->sibling) {
437 BUG_ON(list_empty(&tmp_cg_links));
438 link = list_entry(tmp_cg_links.next,
439 struct cg_cgroup_link,
441 list_del(&link->cgrp_link_list);
442 list_add(&link->cgrp_link_list, &cgrp->css_sets);
444 list_add(&link->cg_link_list, &res->cg_links);
447 if (list_empty(&rootnode.subsys_list)) {
448 link = list_entry(tmp_cg_links.next,
449 struct cg_cgroup_link,
451 list_del(&link->cgrp_link_list);
452 list_add(&link->cgrp_link_list, &dummytop->css_sets);
454 list_add(&link->cg_link_list, &res->cg_links);
457 BUG_ON(!list_empty(&tmp_cg_links));
459 /* Link this cgroup group into the list */
460 list_add(&res->list, &init_css_set.list);
462 write_unlock(&css_set_lock);
468 * There is one global cgroup mutex. We also require taking
469 * task_lock() when dereferencing a task's cgroup subsys pointers.
470 * See "The task_lock() exception", at the end of this comment.
472 * A task must hold cgroup_mutex to modify cgroups.
474 * Any task can increment and decrement the count field without lock.
475 * So in general, code holding cgroup_mutex can't rely on the count
476 * field not changing. However, if the count goes to zero, then only
477 * cgroup_attach_task() can increment it again. Because a count of zero
478 * means that no tasks are currently attached, therefore there is no
479 * way a task attached to that cgroup can fork (the other way to
480 * increment the count). So code holding cgroup_mutex can safely
481 * assume that if the count is zero, it will stay zero. Similarly, if
482 * a task holds cgroup_mutex on a cgroup with zero count, it
483 * knows that the cgroup won't be removed, as cgroup_rmdir()
486 * The cgroup_common_file_write handler for operations that modify
487 * the cgroup hierarchy holds cgroup_mutex across the entire operation,
488 * single threading all such cgroup modifications across the system.
490 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
491 * (usually) take cgroup_mutex. These are the two most performance
492 * critical pieces of code here. The exception occurs on cgroup_exit(),
493 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
494 * is taken, and if the cgroup count is zero, a usermode call made
495 * to the release agent with the name of the cgroup (path relative to
496 * the root of cgroup file system) as the argument.
498 * A cgroup can only be deleted if both its 'count' of using tasks
499 * is zero, and its list of 'children' cgroups is empty. Since all
500 * tasks in the system use _some_ cgroup, and since there is always at
501 * least one task in the system (init, pid == 1), therefore, top_cgroup
502 * always has either children cgroups and/or using tasks. So we don't
503 * need a special hack to ensure that top_cgroup cannot be deleted.
505 * The task_lock() exception
507 * The need for this exception arises from the action of
508 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
509 * another. It does so using cgroup_mutex, however there are
510 * several performance critical places that need to reference
511 * task->cgroup without the expense of grabbing a system global
512 * mutex. Therefore except as noted below, when dereferencing or, as
513 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
514 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
515 * the task_struct routinely used for such matters.
517 * P.S. One more locking exception. RCU is used to guard the
518 * update of a tasks cgroup pointer by cgroup_attach_task()
522 * cgroup_lock - lock out any changes to cgroup structures
525 void cgroup_lock(void)
527 mutex_lock(&cgroup_mutex);
531 * cgroup_unlock - release lock on cgroup changes
533 * Undo the lock taken in a previous cgroup_lock() call.
535 void cgroup_unlock(void)
537 mutex_unlock(&cgroup_mutex);
541 * A couple of forward declarations required, due to cyclic reference loop:
542 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
543 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
547 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
548 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
549 static int cgroup_populate_dir(struct cgroup *cgrp);
550 static struct inode_operations cgroup_dir_inode_operations;
551 static struct file_operations proc_cgroupstats_operations;
553 static struct backing_dev_info cgroup_backing_dev_info = {
554 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
557 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
559 struct inode *inode = new_inode(sb);
562 inode->i_mode = mode;
563 inode->i_uid = current->fsuid;
564 inode->i_gid = current->fsgid;
566 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
567 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
573 * Call subsys's pre_destroy handler.
574 * This is called before css refcnt check.
576 static void cgroup_call_pre_destroy(struct cgroup *cgrp)
578 struct cgroup_subsys *ss;
579 for_each_subsys(cgrp->root, ss)
580 if (ss->pre_destroy && cgrp->subsys[ss->subsys_id])
581 ss->pre_destroy(ss, cgrp);
585 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
587 /* is dentry a directory ? if so, kfree() associated cgroup */
588 if (S_ISDIR(inode->i_mode)) {
589 struct cgroup *cgrp = dentry->d_fsdata;
590 struct cgroup_subsys *ss;
591 BUG_ON(!(cgroup_is_removed(cgrp)));
592 /* It's possible for external users to be holding css
593 * reference counts on a cgroup; css_put() needs to
594 * be able to access the cgroup after decrementing
595 * the reference count in order to know if it needs to
596 * queue the cgroup to be handled by the release
600 mutex_lock(&cgroup_mutex);
602 * Release the subsystem state objects.
604 for_each_subsys(cgrp->root, ss) {
605 if (cgrp->subsys[ss->subsys_id])
606 ss->destroy(ss, cgrp);
609 cgrp->root->number_of_cgroups--;
610 mutex_unlock(&cgroup_mutex);
612 /* Drop the active superblock reference that we took when we
613 * created the cgroup */
614 deactivate_super(cgrp->root->sb);
621 static void remove_dir(struct dentry *d)
623 struct dentry *parent = dget(d->d_parent);
626 simple_rmdir(parent->d_inode, d);
630 static void cgroup_clear_directory(struct dentry *dentry)
632 struct list_head *node;
634 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
635 spin_lock(&dcache_lock);
636 node = dentry->d_subdirs.next;
637 while (node != &dentry->d_subdirs) {
638 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
641 /* This should never be called on a cgroup
642 * directory with child cgroups */
643 BUG_ON(d->d_inode->i_mode & S_IFDIR);
645 spin_unlock(&dcache_lock);
647 simple_unlink(dentry->d_inode, d);
649 spin_lock(&dcache_lock);
651 node = dentry->d_subdirs.next;
653 spin_unlock(&dcache_lock);
657 * NOTE : the dentry must have been dget()'ed
659 static void cgroup_d_remove_dir(struct dentry *dentry)
661 cgroup_clear_directory(dentry);
663 spin_lock(&dcache_lock);
664 list_del_init(&dentry->d_u.d_child);
665 spin_unlock(&dcache_lock);
669 static int rebind_subsystems(struct cgroupfs_root *root,
670 unsigned long final_bits)
672 unsigned long added_bits, removed_bits;
673 struct cgroup *cgrp = &root->top_cgroup;
676 removed_bits = root->actual_subsys_bits & ~final_bits;
677 added_bits = final_bits & ~root->actual_subsys_bits;
678 /* Check that any added subsystems are currently free */
679 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
680 unsigned long bit = 1UL << i;
681 struct cgroup_subsys *ss = subsys[i];
682 if (!(bit & added_bits))
684 if (ss->root != &rootnode) {
685 /* Subsystem isn't free */
690 /* Currently we don't handle adding/removing subsystems when
691 * any child cgroups exist. This is theoretically supportable
692 * but involves complex error handling, so it's being left until
694 if (!list_empty(&cgrp->children))
697 /* Process each subsystem */
698 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
699 struct cgroup_subsys *ss = subsys[i];
700 unsigned long bit = 1UL << i;
701 if (bit & added_bits) {
702 /* We're binding this subsystem to this hierarchy */
703 BUG_ON(cgrp->subsys[i]);
704 BUG_ON(!dummytop->subsys[i]);
705 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
706 cgrp->subsys[i] = dummytop->subsys[i];
707 cgrp->subsys[i]->cgroup = cgrp;
708 list_add(&ss->sibling, &root->subsys_list);
709 rcu_assign_pointer(ss->root, root);
713 } else if (bit & removed_bits) {
714 /* We're removing this subsystem */
715 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
716 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
718 ss->bind(ss, dummytop);
719 dummytop->subsys[i]->cgroup = dummytop;
720 cgrp->subsys[i] = NULL;
721 rcu_assign_pointer(subsys[i]->root, &rootnode);
722 list_del(&ss->sibling);
723 } else if (bit & final_bits) {
724 /* Subsystem state should already exist */
725 BUG_ON(!cgrp->subsys[i]);
727 /* Subsystem state shouldn't exist */
728 BUG_ON(cgrp->subsys[i]);
731 root->subsys_bits = root->actual_subsys_bits = final_bits;
737 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
739 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
740 struct cgroup_subsys *ss;
742 mutex_lock(&cgroup_mutex);
743 for_each_subsys(root, ss)
744 seq_printf(seq, ",%s", ss->name);
745 if (test_bit(ROOT_NOPREFIX, &root->flags))
746 seq_puts(seq, ",noprefix");
747 if (strlen(root->release_agent_path))
748 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
749 mutex_unlock(&cgroup_mutex);
753 struct cgroup_sb_opts {
754 unsigned long subsys_bits;
759 /* Convert a hierarchy specifier into a bitmask of subsystems and
761 static int parse_cgroupfs_options(char *data,
762 struct cgroup_sb_opts *opts)
764 char *token, *o = data ?: "all";
766 opts->subsys_bits = 0;
768 opts->release_agent = NULL;
770 while ((token = strsep(&o, ",")) != NULL) {
773 if (!strcmp(token, "all")) {
774 /* Add all non-disabled subsystems */
776 opts->subsys_bits = 0;
777 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
778 struct cgroup_subsys *ss = subsys[i];
780 opts->subsys_bits |= 1ul << i;
782 } else if (!strcmp(token, "noprefix")) {
783 set_bit(ROOT_NOPREFIX, &opts->flags);
784 } else if (!strncmp(token, "release_agent=", 14)) {
785 /* Specifying two release agents is forbidden */
786 if (opts->release_agent)
788 opts->release_agent = kzalloc(PATH_MAX, GFP_KERNEL);
789 if (!opts->release_agent)
791 strncpy(opts->release_agent, token + 14, PATH_MAX - 1);
792 opts->release_agent[PATH_MAX - 1] = 0;
794 struct cgroup_subsys *ss;
796 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
798 if (!strcmp(token, ss->name)) {
800 set_bit(i, &opts->subsys_bits);
804 if (i == CGROUP_SUBSYS_COUNT)
809 /* We can't have an empty hierarchy */
810 if (!opts->subsys_bits)
816 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
819 struct cgroupfs_root *root = sb->s_fs_info;
820 struct cgroup *cgrp = &root->top_cgroup;
821 struct cgroup_sb_opts opts;
823 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
824 mutex_lock(&cgroup_mutex);
826 /* See what subsystems are wanted */
827 ret = parse_cgroupfs_options(data, &opts);
831 /* Don't allow flags to change at remount */
832 if (opts.flags != root->flags) {
837 ret = rebind_subsystems(root, opts.subsys_bits);
839 /* (re)populate subsystem files */
841 cgroup_populate_dir(cgrp);
843 if (opts.release_agent)
844 strcpy(root->release_agent_path, opts.release_agent);
846 if (opts.release_agent)
847 kfree(opts.release_agent);
848 mutex_unlock(&cgroup_mutex);
849 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
853 static struct super_operations cgroup_ops = {
854 .statfs = simple_statfs,
855 .drop_inode = generic_delete_inode,
856 .show_options = cgroup_show_options,
857 .remount_fs = cgroup_remount,
860 static void init_cgroup_root(struct cgroupfs_root *root)
862 struct cgroup *cgrp = &root->top_cgroup;
863 INIT_LIST_HEAD(&root->subsys_list);
864 INIT_LIST_HEAD(&root->root_list);
865 root->number_of_cgroups = 1;
867 cgrp->top_cgroup = cgrp;
868 INIT_LIST_HEAD(&cgrp->sibling);
869 INIT_LIST_HEAD(&cgrp->children);
870 INIT_LIST_HEAD(&cgrp->css_sets);
871 INIT_LIST_HEAD(&cgrp->release_list);
874 static int cgroup_test_super(struct super_block *sb, void *data)
876 struct cgroupfs_root *new = data;
877 struct cgroupfs_root *root = sb->s_fs_info;
879 /* First check subsystems */
880 if (new->subsys_bits != root->subsys_bits)
883 /* Next check flags */
884 if (new->flags != root->flags)
890 static int cgroup_set_super(struct super_block *sb, void *data)
893 struct cgroupfs_root *root = data;
895 ret = set_anon_super(sb, NULL);
899 sb->s_fs_info = root;
902 sb->s_blocksize = PAGE_CACHE_SIZE;
903 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
904 sb->s_magic = CGROUP_SUPER_MAGIC;
905 sb->s_op = &cgroup_ops;
910 static int cgroup_get_rootdir(struct super_block *sb)
912 struct inode *inode =
913 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
914 struct dentry *dentry;
919 inode->i_fop = &simple_dir_operations;
920 inode->i_op = &cgroup_dir_inode_operations;
921 /* directories start off with i_nlink == 2 (for "." entry) */
923 dentry = d_alloc_root(inode);
932 static int cgroup_get_sb(struct file_system_type *fs_type,
933 int flags, const char *unused_dev_name,
934 void *data, struct vfsmount *mnt)
936 struct cgroup_sb_opts opts;
938 struct super_block *sb;
939 struct cgroupfs_root *root;
940 struct list_head tmp_cg_links, *l;
941 INIT_LIST_HEAD(&tmp_cg_links);
943 /* First find the desired set of subsystems */
944 ret = parse_cgroupfs_options(data, &opts);
946 if (opts.release_agent)
947 kfree(opts.release_agent);
951 root = kzalloc(sizeof(*root), GFP_KERNEL);
953 if (opts.release_agent)
954 kfree(opts.release_agent);
958 init_cgroup_root(root);
959 root->subsys_bits = opts.subsys_bits;
960 root->flags = opts.flags;
961 if (opts.release_agent) {
962 strcpy(root->release_agent_path, opts.release_agent);
963 kfree(opts.release_agent);
966 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, root);
973 if (sb->s_fs_info != root) {
974 /* Reusing an existing superblock */
975 BUG_ON(sb->s_root == NULL);
980 struct cgroup *cgrp = &root->top_cgroup;
983 BUG_ON(sb->s_root != NULL);
985 ret = cgroup_get_rootdir(sb);
988 inode = sb->s_root->d_inode;
990 mutex_lock(&inode->i_mutex);
991 mutex_lock(&cgroup_mutex);
994 * We're accessing css_set_count without locking
995 * css_set_lock here, but that's OK - it can only be
996 * increased by someone holding cgroup_lock, and
997 * that's us. The worst that can happen is that we
998 * have some link structures left over
1000 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1002 mutex_unlock(&cgroup_mutex);
1003 mutex_unlock(&inode->i_mutex);
1004 goto drop_new_super;
1007 ret = rebind_subsystems(root, root->subsys_bits);
1008 if (ret == -EBUSY) {
1009 mutex_unlock(&cgroup_mutex);
1010 mutex_unlock(&inode->i_mutex);
1011 goto drop_new_super;
1014 /* EBUSY should be the only error here */
1017 list_add(&root->root_list, &roots);
1020 sb->s_root->d_fsdata = &root->top_cgroup;
1021 root->top_cgroup.dentry = sb->s_root;
1023 /* Link the top cgroup in this hierarchy into all
1024 * the css_set objects */
1025 write_lock(&css_set_lock);
1026 l = &init_css_set.list;
1029 struct cg_cgroup_link *link;
1030 cg = list_entry(l, struct css_set, list);
1031 BUG_ON(list_empty(&tmp_cg_links));
1032 link = list_entry(tmp_cg_links.next,
1033 struct cg_cgroup_link,
1035 list_del(&link->cgrp_link_list);
1037 list_add(&link->cgrp_link_list,
1038 &root->top_cgroup.css_sets);
1039 list_add(&link->cg_link_list, &cg->cg_links);
1041 } while (l != &init_css_set.list);
1042 write_unlock(&css_set_lock);
1044 free_cg_links(&tmp_cg_links);
1046 BUG_ON(!list_empty(&cgrp->sibling));
1047 BUG_ON(!list_empty(&cgrp->children));
1048 BUG_ON(root->number_of_cgroups != 1);
1050 cgroup_populate_dir(cgrp);
1051 mutex_unlock(&inode->i_mutex);
1052 mutex_unlock(&cgroup_mutex);
1055 return simple_set_mnt(mnt, sb);
1058 up_write(&sb->s_umount);
1059 deactivate_super(sb);
1060 free_cg_links(&tmp_cg_links);
1064 static void cgroup_kill_sb(struct super_block *sb) {
1065 struct cgroupfs_root *root = sb->s_fs_info;
1066 struct cgroup *cgrp = &root->top_cgroup;
1071 BUG_ON(root->number_of_cgroups != 1);
1072 BUG_ON(!list_empty(&cgrp->children));
1073 BUG_ON(!list_empty(&cgrp->sibling));
1075 mutex_lock(&cgroup_mutex);
1077 /* Rebind all subsystems back to the default hierarchy */
1078 ret = rebind_subsystems(root, 0);
1079 /* Shouldn't be able to fail ... */
1083 * Release all the links from css_sets to this hierarchy's
1086 write_lock(&css_set_lock);
1087 while (!list_empty(&cgrp->css_sets)) {
1088 struct cg_cgroup_link *link;
1089 link = list_entry(cgrp->css_sets.next,
1090 struct cg_cgroup_link, cgrp_link_list);
1091 list_del(&link->cg_link_list);
1092 list_del(&link->cgrp_link_list);
1095 write_unlock(&css_set_lock);
1097 if (!list_empty(&root->root_list)) {
1098 list_del(&root->root_list);
1101 mutex_unlock(&cgroup_mutex);
1104 kill_litter_super(sb);
1107 static struct file_system_type cgroup_fs_type = {
1109 .get_sb = cgroup_get_sb,
1110 .kill_sb = cgroup_kill_sb,
1113 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1115 return dentry->d_fsdata;
1118 static inline struct cftype *__d_cft(struct dentry *dentry)
1120 return dentry->d_fsdata;
1124 * cgroup_path - generate the path of a cgroup
1125 * @cgrp: the cgroup in question
1126 * @buf: the buffer to write the path into
1127 * @buflen: the length of the buffer
1129 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1130 * Returns 0 on success, -errno on error.
1132 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1136 if (cgrp == dummytop) {
1138 * Inactive subsystems have no dentry for their root
1145 start = buf + buflen;
1149 int len = cgrp->dentry->d_name.len;
1150 if ((start -= len) < buf)
1151 return -ENAMETOOLONG;
1152 memcpy(start, cgrp->dentry->d_name.name, len);
1153 cgrp = cgrp->parent;
1159 return -ENAMETOOLONG;
1162 memmove(buf, start, buf + buflen - start);
1167 * Return the first subsystem attached to a cgroup's hierarchy, and
1171 static void get_first_subsys(const struct cgroup *cgrp,
1172 struct cgroup_subsys_state **css, int *subsys_id)
1174 const struct cgroupfs_root *root = cgrp->root;
1175 const struct cgroup_subsys *test_ss;
1176 BUG_ON(list_empty(&root->subsys_list));
1177 test_ss = list_entry(root->subsys_list.next,
1178 struct cgroup_subsys, sibling);
1180 *css = cgrp->subsys[test_ss->subsys_id];
1184 *subsys_id = test_ss->subsys_id;
1188 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1189 * @cgrp: the cgroup the task is attaching to
1190 * @tsk: the task to be attached
1192 * Call holding cgroup_mutex. May take task_lock of
1193 * the task 'tsk' during call.
1195 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1198 struct cgroup_subsys *ss;
1199 struct cgroup *oldcgrp;
1200 struct css_set *cg = tsk->cgroups;
1201 struct css_set *newcg;
1202 struct cgroupfs_root *root = cgrp->root;
1205 get_first_subsys(cgrp, NULL, &subsys_id);
1207 /* Nothing to do if the task is already in that cgroup */
1208 oldcgrp = task_cgroup(tsk, subsys_id);
1209 if (cgrp == oldcgrp)
1212 for_each_subsys(root, ss) {
1213 if (ss->can_attach) {
1214 retval = ss->can_attach(ss, cgrp, tsk);
1221 * Locate or allocate a new css_set for this task,
1222 * based on its final set of cgroups
1224 newcg = find_css_set(cg, cgrp);
1229 if (tsk->flags & PF_EXITING) {
1234 rcu_assign_pointer(tsk->cgroups, newcg);
1237 /* Update the css_set linked lists if we're using them */
1238 write_lock(&css_set_lock);
1239 if (!list_empty(&tsk->cg_list)) {
1240 list_del(&tsk->cg_list);
1241 list_add(&tsk->cg_list, &newcg->tasks);
1243 write_unlock(&css_set_lock);
1245 for_each_subsys(root, ss) {
1247 ss->attach(ss, cgrp, oldcgrp, tsk);
1249 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1256 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with
1257 * cgroup_mutex, may take task_lock of task
1259 static int attach_task_by_pid(struct cgroup *cgrp, char *pidbuf)
1262 struct task_struct *tsk;
1265 if (sscanf(pidbuf, "%d", &pid) != 1)
1270 tsk = find_task_by_vpid(pid);
1271 if (!tsk || tsk->flags & PF_EXITING) {
1275 get_task_struct(tsk);
1278 if ((current->euid) && (current->euid != tsk->uid)
1279 && (current->euid != tsk->suid)) {
1280 put_task_struct(tsk);
1285 get_task_struct(tsk);
1288 ret = cgroup_attach_task(cgrp, tsk);
1289 put_task_struct(tsk);
1293 /* The various types of files and directories in a cgroup file system */
1294 enum cgroup_filetype {
1298 FILE_NOTIFY_ON_RELEASE,
1302 static ssize_t cgroup_write_u64(struct cgroup *cgrp, struct cftype *cft,
1304 const char __user *userbuf,
1305 size_t nbytes, loff_t *unused_ppos)
1314 if (nbytes >= sizeof(buffer))
1316 if (copy_from_user(buffer, userbuf, nbytes))
1319 buffer[nbytes] = 0; /* nul-terminate */
1321 val = simple_strtoull(buffer, &end, 0);
1325 /* Pass to subsystem */
1326 retval = cft->write_u64(cgrp, cft, val);
1332 static ssize_t cgroup_common_file_write(struct cgroup *cgrp,
1335 const char __user *userbuf,
1336 size_t nbytes, loff_t *unused_ppos)
1338 enum cgroup_filetype type = cft->private;
1342 if (nbytes >= PATH_MAX)
1345 /* +1 for nul-terminator */
1346 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1350 if (copy_from_user(buffer, userbuf, nbytes)) {
1354 buffer[nbytes] = 0; /* nul-terminate */
1355 strstrip(buffer); /* strip -just- trailing whitespace */
1357 mutex_lock(&cgroup_mutex);
1360 * This was already checked for in cgroup_file_write(), but
1361 * check again now we're holding cgroup_mutex.
1363 if (cgroup_is_removed(cgrp)) {
1370 retval = attach_task_by_pid(cgrp, buffer);
1372 case FILE_NOTIFY_ON_RELEASE:
1373 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
1374 if (simple_strtoul(buffer, NULL, 10) != 0)
1375 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
1377 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
1379 case FILE_RELEASE_AGENT:
1380 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1381 strcpy(cgrp->root->release_agent_path, buffer);
1391 mutex_unlock(&cgroup_mutex);
1397 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1398 size_t nbytes, loff_t *ppos)
1400 struct cftype *cft = __d_cft(file->f_dentry);
1401 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1403 if (!cft || cgroup_is_removed(cgrp))
1406 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1408 return cgroup_write_u64(cgrp, cft, file, buf, nbytes, ppos);
1412 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1414 char __user *buf, size_t nbytes,
1418 u64 val = cft->read_u64(cgrp, cft);
1419 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1421 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1424 static ssize_t cgroup_common_file_read(struct cgroup *cgrp,
1428 size_t nbytes, loff_t *ppos)
1430 enum cgroup_filetype type = cft->private;
1435 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1441 case FILE_RELEASE_AGENT:
1443 struct cgroupfs_root *root;
1445 mutex_lock(&cgroup_mutex);
1447 n = strnlen(root->release_agent_path,
1448 sizeof(root->release_agent_path));
1449 n = min(n, (size_t) PAGE_SIZE);
1450 strncpy(s, root->release_agent_path, n);
1451 mutex_unlock(&cgroup_mutex);
1461 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1463 free_page((unsigned long)page);
1467 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1468 size_t nbytes, loff_t *ppos)
1470 struct cftype *cft = __d_cft(file->f_dentry);
1471 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1473 if (!cft || cgroup_is_removed(cgrp))
1477 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
1479 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
1484 * seqfile ops/methods for returning structured data. Currently just
1485 * supports string->u64 maps, but can be extended in future.
1488 struct cgroup_seqfile_state {
1490 struct cgroup *cgroup;
1493 static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
1495 struct seq_file *sf = cb->state;
1496 return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
1499 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
1501 struct cgroup_seqfile_state *state = m->private;
1502 struct cftype *cft = state->cft;
1503 struct cgroup_map_cb cb = {
1504 .fill = cgroup_map_add,
1507 return cft->read_map(state->cgroup, cft, &cb);
1510 int cgroup_seqfile_release(struct inode *inode, struct file *file)
1512 struct seq_file *seq = file->private_data;
1513 kfree(seq->private);
1514 return single_release(inode, file);
1517 static struct file_operations cgroup_seqfile_operations = {
1519 .llseek = seq_lseek,
1520 .release = cgroup_seqfile_release,
1523 static int cgroup_file_open(struct inode *inode, struct file *file)
1528 err = generic_file_open(inode, file);
1532 cft = __d_cft(file->f_dentry);
1535 if (cft->read_map) {
1536 struct cgroup_seqfile_state *state =
1537 kzalloc(sizeof(*state), GFP_USER);
1541 state->cgroup = __d_cgrp(file->f_dentry->d_parent);
1542 file->f_op = &cgroup_seqfile_operations;
1543 err = single_open(file, cgroup_seqfile_show, state);
1546 } else if (cft->open)
1547 err = cft->open(inode, file);
1554 static int cgroup_file_release(struct inode *inode, struct file *file)
1556 struct cftype *cft = __d_cft(file->f_dentry);
1558 return cft->release(inode, file);
1563 * cgroup_rename - Only allow simple rename of directories in place.
1565 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
1566 struct inode *new_dir, struct dentry *new_dentry)
1568 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1570 if (new_dentry->d_inode)
1572 if (old_dir != new_dir)
1574 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1577 static struct file_operations cgroup_file_operations = {
1578 .read = cgroup_file_read,
1579 .write = cgroup_file_write,
1580 .llseek = generic_file_llseek,
1581 .open = cgroup_file_open,
1582 .release = cgroup_file_release,
1585 static struct inode_operations cgroup_dir_inode_operations = {
1586 .lookup = simple_lookup,
1587 .mkdir = cgroup_mkdir,
1588 .rmdir = cgroup_rmdir,
1589 .rename = cgroup_rename,
1592 static int cgroup_create_file(struct dentry *dentry, int mode,
1593 struct super_block *sb)
1595 static struct dentry_operations cgroup_dops = {
1596 .d_iput = cgroup_diput,
1599 struct inode *inode;
1603 if (dentry->d_inode)
1606 inode = cgroup_new_inode(mode, sb);
1610 if (S_ISDIR(mode)) {
1611 inode->i_op = &cgroup_dir_inode_operations;
1612 inode->i_fop = &simple_dir_operations;
1614 /* start off with i_nlink == 2 (for "." entry) */
1617 /* start with the directory inode held, so that we can
1618 * populate it without racing with another mkdir */
1619 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1620 } else if (S_ISREG(mode)) {
1622 inode->i_fop = &cgroup_file_operations;
1624 dentry->d_op = &cgroup_dops;
1625 d_instantiate(dentry, inode);
1626 dget(dentry); /* Extra count - pin the dentry in core */
1631 * cgroup_create_dir - create a directory for an object.
1632 * @cgrp: the cgroup we create the directory for. It must have a valid
1633 * ->parent field. And we are going to fill its ->dentry field.
1634 * @dentry: dentry of the new cgroup
1635 * @mode: mode to set on new directory.
1637 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
1640 struct dentry *parent;
1643 parent = cgrp->parent->dentry;
1644 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
1646 dentry->d_fsdata = cgrp;
1647 inc_nlink(parent->d_inode);
1648 cgrp->dentry = dentry;
1656 int cgroup_add_file(struct cgroup *cgrp,
1657 struct cgroup_subsys *subsys,
1658 const struct cftype *cft)
1660 struct dentry *dir = cgrp->dentry;
1661 struct dentry *dentry;
1664 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
1665 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
1666 strcpy(name, subsys->name);
1669 strcat(name, cft->name);
1670 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
1671 dentry = lookup_one_len(name, dir, strlen(name));
1672 if (!IS_ERR(dentry)) {
1673 error = cgroup_create_file(dentry, 0644 | S_IFREG,
1676 dentry->d_fsdata = (void *)cft;
1679 error = PTR_ERR(dentry);
1683 int cgroup_add_files(struct cgroup *cgrp,
1684 struct cgroup_subsys *subsys,
1685 const struct cftype cft[],
1689 for (i = 0; i < count; i++) {
1690 err = cgroup_add_file(cgrp, subsys, &cft[i]);
1698 * cgroup_task_count - count the number of tasks in a cgroup.
1699 * @cgrp: the cgroup in question
1701 * Return the number of tasks in the cgroup.
1703 int cgroup_task_count(const struct cgroup *cgrp)
1706 struct list_head *l;
1708 read_lock(&css_set_lock);
1709 l = cgrp->css_sets.next;
1710 while (l != &cgrp->css_sets) {
1711 struct cg_cgroup_link *link =
1712 list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1713 count += atomic_read(&link->cg->ref.refcount);
1716 read_unlock(&css_set_lock);
1721 * Advance a list_head iterator. The iterator should be positioned at
1722 * the start of a css_set
1724 static void cgroup_advance_iter(struct cgroup *cgrp,
1725 struct cgroup_iter *it)
1727 struct list_head *l = it->cg_link;
1728 struct cg_cgroup_link *link;
1731 /* Advance to the next non-empty css_set */
1734 if (l == &cgrp->css_sets) {
1738 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1740 } while (list_empty(&cg->tasks));
1742 it->task = cg->tasks.next;
1746 * To reduce the fork() overhead for systems that are not actually
1747 * using their cgroups capability, we don't maintain the lists running
1748 * through each css_set to its tasks until we see the list actually
1749 * used - in other words after the first call to cgroup_iter_start().
1751 * The tasklist_lock is not held here, as do_each_thread() and
1752 * while_each_thread() are protected by RCU.
1754 static void cgroup_enable_task_cg_lists(void)
1756 struct task_struct *p, *g;
1757 write_lock(&css_set_lock);
1758 use_task_css_set_links = 1;
1759 do_each_thread(g, p) {
1762 * We should check if the process is exiting, otherwise
1763 * it will race with cgroup_exit() in that the list
1764 * entry won't be deleted though the process has exited.
1766 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
1767 list_add(&p->cg_list, &p->cgroups->tasks);
1769 } while_each_thread(g, p);
1770 write_unlock(&css_set_lock);
1773 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
1776 * The first time anyone tries to iterate across a cgroup,
1777 * we need to enable the list linking each css_set to its
1778 * tasks, and fix up all existing tasks.
1780 if (!use_task_css_set_links)
1781 cgroup_enable_task_cg_lists();
1783 read_lock(&css_set_lock);
1784 it->cg_link = &cgrp->css_sets;
1785 cgroup_advance_iter(cgrp, it);
1788 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
1789 struct cgroup_iter *it)
1791 struct task_struct *res;
1792 struct list_head *l = it->task;
1794 /* If the iterator cg is NULL, we have no tasks */
1797 res = list_entry(l, struct task_struct, cg_list);
1798 /* Advance iterator to find next entry */
1800 if (l == &res->cgroups->tasks) {
1801 /* We reached the end of this task list - move on to
1802 * the next cg_cgroup_link */
1803 cgroup_advance_iter(cgrp, it);
1810 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
1812 read_unlock(&css_set_lock);
1815 static inline int started_after_time(struct task_struct *t1,
1816 struct timespec *time,
1817 struct task_struct *t2)
1819 int start_diff = timespec_compare(&t1->start_time, time);
1820 if (start_diff > 0) {
1822 } else if (start_diff < 0) {
1826 * Arbitrarily, if two processes started at the same
1827 * time, we'll say that the lower pointer value
1828 * started first. Note that t2 may have exited by now
1829 * so this may not be a valid pointer any longer, but
1830 * that's fine - it still serves to distinguish
1831 * between two tasks started (effectively) simultaneously.
1838 * This function is a callback from heap_insert() and is used to order
1840 * In this case we order the heap in descending task start time.
1842 static inline int started_after(void *p1, void *p2)
1844 struct task_struct *t1 = p1;
1845 struct task_struct *t2 = p2;
1846 return started_after_time(t1, &t2->start_time, t2);
1850 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
1851 * @scan: struct cgroup_scanner containing arguments for the scan
1853 * Arguments include pointers to callback functions test_task() and
1855 * Iterate through all the tasks in a cgroup, calling test_task() for each,
1856 * and if it returns true, call process_task() for it also.
1857 * The test_task pointer may be NULL, meaning always true (select all tasks).
1858 * Effectively duplicates cgroup_iter_{start,next,end}()
1859 * but does not lock css_set_lock for the call to process_task().
1860 * The struct cgroup_scanner may be embedded in any structure of the caller's
1862 * It is guaranteed that process_task() will act on every task that
1863 * is a member of the cgroup for the duration of this call. This
1864 * function may or may not call process_task() for tasks that exit
1865 * or move to a different cgroup during the call, or are forked or
1866 * move into the cgroup during the call.
1868 * Note that test_task() may be called with locks held, and may in some
1869 * situations be called multiple times for the same task, so it should
1871 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
1872 * pre-allocated and will be used for heap operations (and its "gt" member will
1873 * be overwritten), else a temporary heap will be used (allocation of which
1874 * may cause this function to fail).
1876 int cgroup_scan_tasks(struct cgroup_scanner *scan)
1879 struct cgroup_iter it;
1880 struct task_struct *p, *dropped;
1881 /* Never dereference latest_task, since it's not refcounted */
1882 struct task_struct *latest_task = NULL;
1883 struct ptr_heap tmp_heap;
1884 struct ptr_heap *heap;
1885 struct timespec latest_time = { 0, 0 };
1888 /* The caller supplied our heap and pre-allocated its memory */
1890 heap->gt = &started_after;
1892 /* We need to allocate our own heap memory */
1894 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
1896 /* cannot allocate the heap */
1902 * Scan tasks in the cgroup, using the scanner's "test_task" callback
1903 * to determine which are of interest, and using the scanner's
1904 * "process_task" callback to process any of them that need an update.
1905 * Since we don't want to hold any locks during the task updates,
1906 * gather tasks to be processed in a heap structure.
1907 * The heap is sorted by descending task start time.
1908 * If the statically-sized heap fills up, we overflow tasks that
1909 * started later, and in future iterations only consider tasks that
1910 * started after the latest task in the previous pass. This
1911 * guarantees forward progress and that we don't miss any tasks.
1914 cgroup_iter_start(scan->cg, &it);
1915 while ((p = cgroup_iter_next(scan->cg, &it))) {
1917 * Only affect tasks that qualify per the caller's callback,
1918 * if he provided one
1920 if (scan->test_task && !scan->test_task(p, scan))
1923 * Only process tasks that started after the last task
1926 if (!started_after_time(p, &latest_time, latest_task))
1928 dropped = heap_insert(heap, p);
1929 if (dropped == NULL) {
1931 * The new task was inserted; the heap wasn't
1935 } else if (dropped != p) {
1937 * The new task was inserted, and pushed out a
1941 put_task_struct(dropped);
1944 * Else the new task was newer than anything already in
1945 * the heap and wasn't inserted
1948 cgroup_iter_end(scan->cg, &it);
1951 for (i = 0; i < heap->size; i++) {
1952 struct task_struct *q = heap->ptrs[i];
1954 latest_time = q->start_time;
1957 /* Process the task per the caller's callback */
1958 scan->process_task(q, scan);
1962 * If we had to process any tasks at all, scan again
1963 * in case some of them were in the middle of forking
1964 * children that didn't get processed.
1965 * Not the most efficient way to do it, but it avoids
1966 * having to take callback_mutex in the fork path
1970 if (heap == &tmp_heap)
1971 heap_free(&tmp_heap);
1976 * Stuff for reading the 'tasks' file.
1978 * Reading this file can return large amounts of data if a cgroup has
1979 * *lots* of attached tasks. So it may need several calls to read(),
1980 * but we cannot guarantee that the information we produce is correct
1981 * unless we produce it entirely atomically.
1983 * Upon tasks file open(), a struct ctr_struct is allocated, that
1984 * will have a pointer to an array (also allocated here). The struct
1985 * ctr_struct * is stored in file->private_data. Its resources will
1986 * be freed by release() when the file is closed. The array is used
1987 * to sprintf the PIDs and then used by read().
1995 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
1996 * 'cgrp'. Return actual number of pids loaded. No need to
1997 * task_lock(p) when reading out p->cgroup, since we're in an RCU
1998 * read section, so the css_set can't go away, and is
1999 * immutable after creation.
2001 static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
2004 struct cgroup_iter it;
2005 struct task_struct *tsk;
2006 cgroup_iter_start(cgrp, &it);
2007 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2008 if (unlikely(n == npids))
2010 pidarray[n++] = task_pid_vnr(tsk);
2012 cgroup_iter_end(cgrp, &it);
2017 * cgroupstats_build - build and fill cgroupstats
2018 * @stats: cgroupstats to fill information into
2019 * @dentry: A dentry entry belonging to the cgroup for which stats have
2022 * Build and fill cgroupstats so that taskstats can export it to user
2025 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2028 struct cgroup *cgrp;
2029 struct cgroup_iter it;
2030 struct task_struct *tsk;
2032 * Validate dentry by checking the superblock operations
2034 if (dentry->d_sb->s_op != &cgroup_ops)
2038 cgrp = dentry->d_fsdata;
2041 cgroup_iter_start(cgrp, &it);
2042 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2043 switch (tsk->state) {
2045 stats->nr_running++;
2047 case TASK_INTERRUPTIBLE:
2048 stats->nr_sleeping++;
2050 case TASK_UNINTERRUPTIBLE:
2051 stats->nr_uninterruptible++;
2054 stats->nr_stopped++;
2057 if (delayacct_is_task_waiting_on_io(tsk))
2058 stats->nr_io_wait++;
2062 cgroup_iter_end(cgrp, &it);
2069 static int cmppid(const void *a, const void *b)
2071 return *(pid_t *)a - *(pid_t *)b;
2075 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
2076 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
2077 * count 'cnt' of how many chars would be written if buf were large enough.
2079 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
2084 for (i = 0; i < npids; i++)
2085 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
2090 * Handle an open on 'tasks' file. Prepare a buffer listing the
2091 * process id's of tasks currently attached to the cgroup being opened.
2093 * Does not require any specific cgroup mutexes, and does not take any.
2095 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2097 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2098 struct ctr_struct *ctr;
2103 if (!(file->f_mode & FMODE_READ))
2106 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
2111 * If cgroup gets more users after we read count, we won't have
2112 * enough space - tough. This race is indistinguishable to the
2113 * caller from the case that the additional cgroup users didn't
2114 * show up until sometime later on.
2116 npids = cgroup_task_count(cgrp);
2118 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
2122 npids = pid_array_load(pidarray, npids, cgrp);
2123 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
2125 /* Call pid_array_to_buf() twice, first just to get bufsz */
2126 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
2127 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
2130 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
2137 file->private_data = ctr;
2148 static ssize_t cgroup_tasks_read(struct cgroup *cgrp,
2150 struct file *file, char __user *buf,
2151 size_t nbytes, loff_t *ppos)
2153 struct ctr_struct *ctr = file->private_data;
2155 return simple_read_from_buffer(buf, nbytes, ppos, ctr->buf, ctr->bufsz);
2158 static int cgroup_tasks_release(struct inode *unused_inode,
2161 struct ctr_struct *ctr;
2163 if (file->f_mode & FMODE_READ) {
2164 ctr = file->private_data;
2171 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2174 return notify_on_release(cgrp);
2178 * for the common functions, 'private' gives the type of file
2180 static struct cftype files[] = {
2183 .open = cgroup_tasks_open,
2184 .read = cgroup_tasks_read,
2185 .write = cgroup_common_file_write,
2186 .release = cgroup_tasks_release,
2187 .private = FILE_TASKLIST,
2191 .name = "notify_on_release",
2192 .read_u64 = cgroup_read_notify_on_release,
2193 .write = cgroup_common_file_write,
2194 .private = FILE_NOTIFY_ON_RELEASE,
2198 static struct cftype cft_release_agent = {
2199 .name = "release_agent",
2200 .read = cgroup_common_file_read,
2201 .write = cgroup_common_file_write,
2202 .private = FILE_RELEASE_AGENT,
2205 static int cgroup_populate_dir(struct cgroup *cgrp)
2208 struct cgroup_subsys *ss;
2210 /* First clear out any existing files */
2211 cgroup_clear_directory(cgrp->dentry);
2213 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
2217 if (cgrp == cgrp->top_cgroup) {
2218 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
2222 for_each_subsys(cgrp->root, ss) {
2223 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
2230 static void init_cgroup_css(struct cgroup_subsys_state *css,
2231 struct cgroup_subsys *ss,
2232 struct cgroup *cgrp)
2235 atomic_set(&css->refcnt, 0);
2237 if (cgrp == dummytop)
2238 set_bit(CSS_ROOT, &css->flags);
2239 BUG_ON(cgrp->subsys[ss->subsys_id]);
2240 cgrp->subsys[ss->subsys_id] = css;
2244 * cgroup_create - create a cgroup
2245 * @parent: cgroup that will be parent of the new cgroup
2246 * @dentry: dentry of the new cgroup
2247 * @mode: mode to set on new inode
2249 * Must be called with the mutex on the parent inode held
2251 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
2254 struct cgroup *cgrp;
2255 struct cgroupfs_root *root = parent->root;
2257 struct cgroup_subsys *ss;
2258 struct super_block *sb = root->sb;
2260 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
2264 /* Grab a reference on the superblock so the hierarchy doesn't
2265 * get deleted on unmount if there are child cgroups. This
2266 * can be done outside cgroup_mutex, since the sb can't
2267 * disappear while someone has an open control file on the
2269 atomic_inc(&sb->s_active);
2271 mutex_lock(&cgroup_mutex);
2273 INIT_LIST_HEAD(&cgrp->sibling);
2274 INIT_LIST_HEAD(&cgrp->children);
2275 INIT_LIST_HEAD(&cgrp->css_sets);
2276 INIT_LIST_HEAD(&cgrp->release_list);
2278 cgrp->parent = parent;
2279 cgrp->root = parent->root;
2280 cgrp->top_cgroup = parent->top_cgroup;
2282 if (notify_on_release(parent))
2283 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2285 for_each_subsys(root, ss) {
2286 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
2291 init_cgroup_css(css, ss, cgrp);
2294 list_add(&cgrp->sibling, &cgrp->parent->children);
2295 root->number_of_cgroups++;
2297 err = cgroup_create_dir(cgrp, dentry, mode);
2301 /* The cgroup directory was pre-locked for us */
2302 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
2304 err = cgroup_populate_dir(cgrp);
2305 /* If err < 0, we have a half-filled directory - oh well ;) */
2307 mutex_unlock(&cgroup_mutex);
2308 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
2314 list_del(&cgrp->sibling);
2315 root->number_of_cgroups--;
2319 for_each_subsys(root, ss) {
2320 if (cgrp->subsys[ss->subsys_id])
2321 ss->destroy(ss, cgrp);
2324 mutex_unlock(&cgroup_mutex);
2326 /* Release the reference count that we took on the superblock */
2327 deactivate_super(sb);
2333 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
2335 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
2337 /* the vfs holds inode->i_mutex already */
2338 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
2341 static inline int cgroup_has_css_refs(struct cgroup *cgrp)
2343 /* Check the reference count on each subsystem. Since we
2344 * already established that there are no tasks in the
2345 * cgroup, if the css refcount is also 0, then there should
2346 * be no outstanding references, so the subsystem is safe to
2347 * destroy. We scan across all subsystems rather than using
2348 * the per-hierarchy linked list of mounted subsystems since
2349 * we can be called via check_for_release() with no
2350 * synchronization other than RCU, and the subsystem linked
2351 * list isn't RCU-safe */
2353 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2354 struct cgroup_subsys *ss = subsys[i];
2355 struct cgroup_subsys_state *css;
2356 /* Skip subsystems not in this hierarchy */
2357 if (ss->root != cgrp->root)
2359 css = cgrp->subsys[ss->subsys_id];
2360 /* When called from check_for_release() it's possible
2361 * that by this point the cgroup has been removed
2362 * and the css deleted. But a false-positive doesn't
2363 * matter, since it can only happen if the cgroup
2364 * has been deleted and hence no longer needs the
2365 * release agent to be called anyway. */
2366 if (css && atomic_read(&css->refcnt))
2372 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
2374 struct cgroup *cgrp = dentry->d_fsdata;
2376 struct cgroup *parent;
2377 struct super_block *sb;
2378 struct cgroupfs_root *root;
2380 /* the vfs holds both inode->i_mutex already */
2382 mutex_lock(&cgroup_mutex);
2383 if (atomic_read(&cgrp->count) != 0) {
2384 mutex_unlock(&cgroup_mutex);
2387 if (!list_empty(&cgrp->children)) {
2388 mutex_unlock(&cgroup_mutex);
2392 parent = cgrp->parent;
2397 * Call pre_destroy handlers of subsys. Notify subsystems
2398 * that rmdir() request comes.
2400 cgroup_call_pre_destroy(cgrp);
2402 if (cgroup_has_css_refs(cgrp)) {
2403 mutex_unlock(&cgroup_mutex);
2407 spin_lock(&release_list_lock);
2408 set_bit(CGRP_REMOVED, &cgrp->flags);
2409 if (!list_empty(&cgrp->release_list))
2410 list_del(&cgrp->release_list);
2411 spin_unlock(&release_list_lock);
2412 /* delete my sibling from parent->children */
2413 list_del(&cgrp->sibling);
2414 spin_lock(&cgrp->dentry->d_lock);
2415 d = dget(cgrp->dentry);
2416 cgrp->dentry = NULL;
2417 spin_unlock(&d->d_lock);
2419 cgroup_d_remove_dir(d);
2422 set_bit(CGRP_RELEASABLE, &parent->flags);
2423 check_for_release(parent);
2425 mutex_unlock(&cgroup_mutex);
2429 static void cgroup_init_subsys(struct cgroup_subsys *ss)
2431 struct cgroup_subsys_state *css;
2432 struct list_head *l;
2434 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
2436 /* Create the top cgroup state for this subsystem */
2437 ss->root = &rootnode;
2438 css = ss->create(ss, dummytop);
2439 /* We don't handle early failures gracefully */
2440 BUG_ON(IS_ERR(css));
2441 init_cgroup_css(css, ss, dummytop);
2443 /* Update all cgroup groups to contain a subsys
2444 * pointer to this state - since the subsystem is
2445 * newly registered, all tasks and hence all cgroup
2446 * groups are in the subsystem's top cgroup. */
2447 write_lock(&css_set_lock);
2448 l = &init_css_set.list;
2450 struct css_set *cg =
2451 list_entry(l, struct css_set, list);
2452 cg->subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
2454 } while (l != &init_css_set.list);
2455 write_unlock(&css_set_lock);
2457 /* If this subsystem requested that it be notified with fork
2458 * events, we should send it one now for every process in the
2461 struct task_struct *g, *p;
2463 read_lock(&tasklist_lock);
2464 do_each_thread(g, p) {
2466 } while_each_thread(g, p);
2467 read_unlock(&tasklist_lock);
2470 need_forkexit_callback |= ss->fork || ss->exit;
2476 * cgroup_init_early - cgroup initialization at system boot
2478 * Initialize cgroups at system boot, and initialize any
2479 * subsystems that request early init.
2481 int __init cgroup_init_early(void)
2484 kref_init(&init_css_set.ref);
2485 kref_get(&init_css_set.ref);
2486 INIT_LIST_HEAD(&init_css_set.list);
2487 INIT_LIST_HEAD(&init_css_set.cg_links);
2488 INIT_LIST_HEAD(&init_css_set.tasks);
2490 init_cgroup_root(&rootnode);
2491 list_add(&rootnode.root_list, &roots);
2493 init_task.cgroups = &init_css_set;
2495 init_css_set_link.cg = &init_css_set;
2496 list_add(&init_css_set_link.cgrp_link_list,
2497 &rootnode.top_cgroup.css_sets);
2498 list_add(&init_css_set_link.cg_link_list,
2499 &init_css_set.cg_links);
2501 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2502 struct cgroup_subsys *ss = subsys[i];
2505 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
2506 BUG_ON(!ss->create);
2507 BUG_ON(!ss->destroy);
2508 if (ss->subsys_id != i) {
2509 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
2510 ss->name, ss->subsys_id);
2515 cgroup_init_subsys(ss);
2521 * cgroup_init - cgroup initialization
2523 * Register cgroup filesystem and /proc file, and initialize
2524 * any subsystems that didn't request early init.
2526 int __init cgroup_init(void)
2530 struct proc_dir_entry *entry;
2532 err = bdi_init(&cgroup_backing_dev_info);
2536 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2537 struct cgroup_subsys *ss = subsys[i];
2538 if (!ss->early_init)
2539 cgroup_init_subsys(ss);
2542 err = register_filesystem(&cgroup_fs_type);
2546 entry = create_proc_entry("cgroups", 0, NULL);
2548 entry->proc_fops = &proc_cgroupstats_operations;
2552 bdi_destroy(&cgroup_backing_dev_info);
2558 * proc_cgroup_show()
2559 * - Print task's cgroup paths into seq_file, one line for each hierarchy
2560 * - Used for /proc/<pid>/cgroup.
2561 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2562 * doesn't really matter if tsk->cgroup changes after we read it,
2563 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
2564 * anyway. No need to check that tsk->cgroup != NULL, thanks to
2565 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2566 * cgroup to top_cgroup.
2569 /* TODO: Use a proper seq_file iterator */
2570 static int proc_cgroup_show(struct seq_file *m, void *v)
2573 struct task_struct *tsk;
2576 struct cgroupfs_root *root;
2579 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2585 tsk = get_pid_task(pid, PIDTYPE_PID);
2591 mutex_lock(&cgroup_mutex);
2593 for_each_root(root) {
2594 struct cgroup_subsys *ss;
2595 struct cgroup *cgrp;
2599 /* Skip this hierarchy if it has no active subsystems */
2600 if (!root->actual_subsys_bits)
2602 seq_printf(m, "%lu:", root->subsys_bits);
2603 for_each_subsys(root, ss)
2604 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
2606 get_first_subsys(&root->top_cgroup, NULL, &subsys_id);
2607 cgrp = task_cgroup(tsk, subsys_id);
2608 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
2616 mutex_unlock(&cgroup_mutex);
2617 put_task_struct(tsk);
2624 static int cgroup_open(struct inode *inode, struct file *file)
2626 struct pid *pid = PROC_I(inode)->pid;
2627 return single_open(file, proc_cgroup_show, pid);
2630 struct file_operations proc_cgroup_operations = {
2631 .open = cgroup_open,
2633 .llseek = seq_lseek,
2634 .release = single_release,
2637 /* Display information about each subsystem and each hierarchy */
2638 static int proc_cgroupstats_show(struct seq_file *m, void *v)
2642 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
2643 mutex_lock(&cgroup_mutex);
2644 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2645 struct cgroup_subsys *ss = subsys[i];
2646 seq_printf(m, "%s\t%lu\t%d\t%d\n",
2647 ss->name, ss->root->subsys_bits,
2648 ss->root->number_of_cgroups, !ss->disabled);
2650 mutex_unlock(&cgroup_mutex);
2654 static int cgroupstats_open(struct inode *inode, struct file *file)
2656 return single_open(file, proc_cgroupstats_show, NULL);
2659 static struct file_operations proc_cgroupstats_operations = {
2660 .open = cgroupstats_open,
2662 .llseek = seq_lseek,
2663 .release = single_release,
2667 * cgroup_fork - attach newly forked task to its parents cgroup.
2668 * @child: pointer to task_struct of forking parent process.
2670 * Description: A task inherits its parent's cgroup at fork().
2672 * A pointer to the shared css_set was automatically copied in
2673 * fork.c by dup_task_struct(). However, we ignore that copy, since
2674 * it was not made under the protection of RCU or cgroup_mutex, so
2675 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
2676 * have already changed current->cgroups, allowing the previously
2677 * referenced cgroup group to be removed and freed.
2679 * At the point that cgroup_fork() is called, 'current' is the parent
2680 * task, and the passed argument 'child' points to the child task.
2682 void cgroup_fork(struct task_struct *child)
2685 child->cgroups = current->cgroups;
2686 get_css_set(child->cgroups);
2687 task_unlock(current);
2688 INIT_LIST_HEAD(&child->cg_list);
2692 * cgroup_fork_callbacks - run fork callbacks
2693 * @child: the new task
2695 * Called on a new task very soon before adding it to the
2696 * tasklist. No need to take any locks since no-one can
2697 * be operating on this task.
2699 void cgroup_fork_callbacks(struct task_struct *child)
2701 if (need_forkexit_callback) {
2703 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2704 struct cgroup_subsys *ss = subsys[i];
2706 ss->fork(ss, child);
2712 * cgroup_post_fork - called on a new task after adding it to the task list
2713 * @child: the task in question
2715 * Adds the task to the list running through its css_set if necessary.
2716 * Has to be after the task is visible on the task list in case we race
2717 * with the first call to cgroup_iter_start() - to guarantee that the
2718 * new task ends up on its list.
2720 void cgroup_post_fork(struct task_struct *child)
2722 if (use_task_css_set_links) {
2723 write_lock(&css_set_lock);
2724 if (list_empty(&child->cg_list))
2725 list_add(&child->cg_list, &child->cgroups->tasks);
2726 write_unlock(&css_set_lock);
2730 * cgroup_exit - detach cgroup from exiting task
2731 * @tsk: pointer to task_struct of exiting process
2732 * @run_callback: run exit callbacks?
2734 * Description: Detach cgroup from @tsk and release it.
2736 * Note that cgroups marked notify_on_release force every task in
2737 * them to take the global cgroup_mutex mutex when exiting.
2738 * This could impact scaling on very large systems. Be reluctant to
2739 * use notify_on_release cgroups where very high task exit scaling
2740 * is required on large systems.
2742 * the_top_cgroup_hack:
2744 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2746 * We call cgroup_exit() while the task is still competent to
2747 * handle notify_on_release(), then leave the task attached to the
2748 * root cgroup in each hierarchy for the remainder of its exit.
2750 * To do this properly, we would increment the reference count on
2751 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
2752 * code we would add a second cgroup function call, to drop that
2753 * reference. This would just create an unnecessary hot spot on
2754 * the top_cgroup reference count, to no avail.
2756 * Normally, holding a reference to a cgroup without bumping its
2757 * count is unsafe. The cgroup could go away, or someone could
2758 * attach us to a different cgroup, decrementing the count on
2759 * the first cgroup that we never incremented. But in this case,
2760 * top_cgroup isn't going away, and either task has PF_EXITING set,
2761 * which wards off any cgroup_attach_task() attempts, or task is a failed
2762 * fork, never visible to cgroup_attach_task.
2764 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
2769 if (run_callbacks && need_forkexit_callback) {
2770 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2771 struct cgroup_subsys *ss = subsys[i];
2778 * Unlink from the css_set task list if necessary.
2779 * Optimistically check cg_list before taking
2782 if (!list_empty(&tsk->cg_list)) {
2783 write_lock(&css_set_lock);
2784 if (!list_empty(&tsk->cg_list))
2785 list_del(&tsk->cg_list);
2786 write_unlock(&css_set_lock);
2789 /* Reassign the task to the init_css_set. */
2792 tsk->cgroups = &init_css_set;
2795 put_css_set_taskexit(cg);
2799 * cgroup_clone - clone the cgroup the given subsystem is attached to
2800 * @tsk: the task to be moved
2801 * @subsys: the given subsystem
2803 * Duplicate the current cgroup in the hierarchy that the given
2804 * subsystem is attached to, and move this task into the new
2807 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys)
2809 struct dentry *dentry;
2811 char nodename[MAX_CGROUP_TYPE_NAMELEN];
2812 struct cgroup *parent, *child;
2813 struct inode *inode;
2815 struct cgroupfs_root *root;
2816 struct cgroup_subsys *ss;
2818 /* We shouldn't be called by an unregistered subsystem */
2819 BUG_ON(!subsys->active);
2821 /* First figure out what hierarchy and cgroup we're dealing
2822 * with, and pin them so we can drop cgroup_mutex */
2823 mutex_lock(&cgroup_mutex);
2825 root = subsys->root;
2826 if (root == &rootnode) {
2828 "Not cloning cgroup for unused subsystem %s\n",
2830 mutex_unlock(&cgroup_mutex);
2834 parent = task_cgroup(tsk, subsys->subsys_id);
2836 snprintf(nodename, MAX_CGROUP_TYPE_NAMELEN, "node_%d", tsk->pid);
2838 /* Pin the hierarchy */
2839 atomic_inc(&parent->root->sb->s_active);
2841 /* Keep the cgroup alive */
2843 mutex_unlock(&cgroup_mutex);
2845 /* Now do the VFS work to create a cgroup */
2846 inode = parent->dentry->d_inode;
2848 /* Hold the parent directory mutex across this operation to
2849 * stop anyone else deleting the new cgroup */
2850 mutex_lock(&inode->i_mutex);
2851 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
2852 if (IS_ERR(dentry)) {
2854 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
2856 ret = PTR_ERR(dentry);
2860 /* Create the cgroup directory, which also creates the cgroup */
2861 ret = vfs_mkdir(inode, dentry, S_IFDIR | 0755);
2862 child = __d_cgrp(dentry);
2866 "Failed to create cgroup %s: %d\n", nodename,
2873 "Couldn't find new cgroup %s\n", nodename);
2878 /* The cgroup now exists. Retake cgroup_mutex and check
2879 * that we're still in the same state that we thought we
2881 mutex_lock(&cgroup_mutex);
2882 if ((root != subsys->root) ||
2883 (parent != task_cgroup(tsk, subsys->subsys_id))) {
2884 /* Aargh, we raced ... */
2885 mutex_unlock(&inode->i_mutex);
2888 deactivate_super(parent->root->sb);
2889 /* The cgroup is still accessible in the VFS, but
2890 * we're not going to try to rmdir() it at this
2893 "Race in cgroup_clone() - leaking cgroup %s\n",
2898 /* do any required auto-setup */
2899 for_each_subsys(root, ss) {
2901 ss->post_clone(ss, child);
2904 /* All seems fine. Finish by moving the task into the new cgroup */
2905 ret = cgroup_attach_task(child, tsk);
2906 mutex_unlock(&cgroup_mutex);
2909 mutex_unlock(&inode->i_mutex);
2911 mutex_lock(&cgroup_mutex);
2913 mutex_unlock(&cgroup_mutex);
2914 deactivate_super(parent->root->sb);
2919 * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
2920 * @cgrp: the cgroup in question
2922 * See if @cgrp is a descendant of the current task's cgroup in
2923 * the appropriate hierarchy.
2925 * If we are sending in dummytop, then presumably we are creating
2926 * the top cgroup in the subsystem.
2928 * Called only by the ns (nsproxy) cgroup.
2930 int cgroup_is_descendant(const struct cgroup *cgrp)
2933 struct cgroup *target;
2936 if (cgrp == dummytop)
2939 get_first_subsys(cgrp, NULL, &subsys_id);
2940 target = task_cgroup(current, subsys_id);
2941 while (cgrp != target && cgrp!= cgrp->top_cgroup)
2942 cgrp = cgrp->parent;
2943 ret = (cgrp == target);
2947 static void check_for_release(struct cgroup *cgrp)
2949 /* All of these checks rely on RCU to keep the cgroup
2950 * structure alive */
2951 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
2952 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
2953 /* Control Group is currently removeable. If it's not
2954 * already queued for a userspace notification, queue
2956 int need_schedule_work = 0;
2957 spin_lock(&release_list_lock);
2958 if (!cgroup_is_removed(cgrp) &&
2959 list_empty(&cgrp->release_list)) {
2960 list_add(&cgrp->release_list, &release_list);
2961 need_schedule_work = 1;
2963 spin_unlock(&release_list_lock);
2964 if (need_schedule_work)
2965 schedule_work(&release_agent_work);
2969 void __css_put(struct cgroup_subsys_state *css)
2971 struct cgroup *cgrp = css->cgroup;
2973 if (atomic_dec_and_test(&css->refcnt) && notify_on_release(cgrp)) {
2974 set_bit(CGRP_RELEASABLE, &cgrp->flags);
2975 check_for_release(cgrp);
2981 * Notify userspace when a cgroup is released, by running the
2982 * configured release agent with the name of the cgroup (path
2983 * relative to the root of cgroup file system) as the argument.
2985 * Most likely, this user command will try to rmdir this cgroup.
2987 * This races with the possibility that some other task will be
2988 * attached to this cgroup before it is removed, or that some other
2989 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
2990 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
2991 * unused, and this cgroup will be reprieved from its death sentence,
2992 * to continue to serve a useful existence. Next time it's released,
2993 * we will get notified again, if it still has 'notify_on_release' set.
2995 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
2996 * means only wait until the task is successfully execve()'d. The
2997 * separate release agent task is forked by call_usermodehelper(),
2998 * then control in this thread returns here, without waiting for the
2999 * release agent task. We don't bother to wait because the caller of
3000 * this routine has no use for the exit status of the release agent
3001 * task, so no sense holding our caller up for that.
3003 static void cgroup_release_agent(struct work_struct *work)
3005 BUG_ON(work != &release_agent_work);
3006 mutex_lock(&cgroup_mutex);
3007 spin_lock(&release_list_lock);
3008 while (!list_empty(&release_list)) {
3009 char *argv[3], *envp[3];
3012 struct cgroup *cgrp = list_entry(release_list.next,
3015 list_del_init(&cgrp->release_list);
3016 spin_unlock(&release_list_lock);
3017 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3019 spin_lock(&release_list_lock);
3023 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0) {
3025 spin_lock(&release_list_lock);
3030 argv[i++] = cgrp->root->release_agent_path;
3031 argv[i++] = (char *)pathbuf;
3035 /* minimal command environment */
3036 envp[i++] = "HOME=/";
3037 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3040 /* Drop the lock while we invoke the usermode helper,
3041 * since the exec could involve hitting disk and hence
3042 * be a slow process */
3043 mutex_unlock(&cgroup_mutex);
3044 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
3046 mutex_lock(&cgroup_mutex);
3047 spin_lock(&release_list_lock);
3049 spin_unlock(&release_list_lock);
3050 mutex_unlock(&cgroup_mutex);
3053 static int __init cgroup_disable(char *str)
3058 while ((token = strsep(&str, ",")) != NULL) {
3062 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3063 struct cgroup_subsys *ss = subsys[i];
3065 if (!strcmp(token, ss->name)) {
3067 printk(KERN_INFO "Disabling %s control group"
3068 " subsystem\n", ss->name);
3075 __setup("cgroup_disable=", cgroup_disable);