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 /* 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
310 static struct css_set *find_existing_css_set(
311 struct css_set *oldcg,
313 struct cgroup_subsys_state *template[])
316 struct cgroupfs_root *root = cgrp->root;
317 struct list_head *l = &init_css_set.list;
319 /* Built the set of subsystem state objects that we want to
320 * see in the new css_set */
321 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
322 if (root->subsys_bits & (1UL << i)) {
323 /* Subsystem is in this hierarchy. So we want
324 * the subsystem state from the new
326 template[i] = cgrp->subsys[i];
328 /* Subsystem is not in this hierarchy, so we
329 * don't want to change the subsystem state */
330 template[i] = oldcg->subsys[i];
334 /* Look through existing cgroup groups to find one to reuse */
337 list_entry(l, struct css_set, list);
339 if (!memcmp(template, cg->subsys, sizeof(cg->subsys))) {
340 /* All subsystems matched */
343 /* Try the next cgroup group */
345 } while (l != &init_css_set.list);
347 /* No existing cgroup group matched */
352 * allocate_cg_links() allocates "count" cg_cgroup_link structures
353 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
354 * success or a negative error
356 static int allocate_cg_links(int count, struct list_head *tmp)
358 struct cg_cgroup_link *link;
361 for (i = 0; i < count; i++) {
362 link = kmalloc(sizeof(*link), GFP_KERNEL);
364 while (!list_empty(tmp)) {
365 link = list_entry(tmp->next,
366 struct cg_cgroup_link,
368 list_del(&link->cgrp_link_list);
373 list_add(&link->cgrp_link_list, tmp);
378 static void free_cg_links(struct list_head *tmp)
380 while (!list_empty(tmp)) {
381 struct cg_cgroup_link *link;
382 link = list_entry(tmp->next,
383 struct cg_cgroup_link,
385 list_del(&link->cgrp_link_list);
391 * find_css_set() takes an existing cgroup group and a
392 * cgroup object, and returns a css_set object that's
393 * equivalent to the old group, but with the given cgroup
394 * substituted into the appropriate hierarchy. Must be called with
397 static struct css_set *find_css_set(
398 struct css_set *oldcg, struct cgroup *cgrp)
401 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
404 struct list_head tmp_cg_links;
405 struct cg_cgroup_link *link;
407 /* First see if we already have a cgroup group that matches
409 write_lock(&css_set_lock);
410 res = find_existing_css_set(oldcg, cgrp, template);
413 write_unlock(&css_set_lock);
418 res = kmalloc(sizeof(*res), GFP_KERNEL);
422 /* Allocate all the cg_cgroup_link objects that we'll need */
423 if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
428 kref_init(&res->ref);
429 INIT_LIST_HEAD(&res->cg_links);
430 INIT_LIST_HEAD(&res->tasks);
432 /* Copy the set of subsystem state objects generated in
433 * find_existing_css_set() */
434 memcpy(res->subsys, template, sizeof(res->subsys));
436 write_lock(&css_set_lock);
437 /* Add reference counts and links from the new css_set. */
438 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
439 struct cgroup *cgrp = res->subsys[i]->cgroup;
440 struct cgroup_subsys *ss = subsys[i];
441 atomic_inc(&cgrp->count);
443 * We want to add a link once per cgroup, so we
444 * only do it for the first subsystem in each
447 if (ss->root->subsys_list.next == &ss->sibling) {
448 BUG_ON(list_empty(&tmp_cg_links));
449 link = list_entry(tmp_cg_links.next,
450 struct cg_cgroup_link,
452 list_del(&link->cgrp_link_list);
453 list_add(&link->cgrp_link_list, &cgrp->css_sets);
455 list_add(&link->cg_link_list, &res->cg_links);
458 if (list_empty(&rootnode.subsys_list)) {
459 link = list_entry(tmp_cg_links.next,
460 struct cg_cgroup_link,
462 list_del(&link->cgrp_link_list);
463 list_add(&link->cgrp_link_list, &dummytop->css_sets);
465 list_add(&link->cg_link_list, &res->cg_links);
468 BUG_ON(!list_empty(&tmp_cg_links));
470 /* Link this cgroup group into the list */
471 list_add(&res->list, &init_css_set.list);
473 write_unlock(&css_set_lock);
479 * There is one global cgroup mutex. We also require taking
480 * task_lock() when dereferencing a task's cgroup subsys pointers.
481 * See "The task_lock() exception", at the end of this comment.
483 * A task must hold cgroup_mutex to modify cgroups.
485 * Any task can increment and decrement the count field without lock.
486 * So in general, code holding cgroup_mutex can't rely on the count
487 * field not changing. However, if the count goes to zero, then only
488 * cgroup_attach_task() can increment it again. Because a count of zero
489 * means that no tasks are currently attached, therefore there is no
490 * way a task attached to that cgroup can fork (the other way to
491 * increment the count). So code holding cgroup_mutex can safely
492 * assume that if the count is zero, it will stay zero. Similarly, if
493 * a task holds cgroup_mutex on a cgroup with zero count, it
494 * knows that the cgroup won't be removed, as cgroup_rmdir()
497 * The cgroup_common_file_write handler for operations that modify
498 * the cgroup hierarchy holds cgroup_mutex across the entire operation,
499 * single threading all such cgroup modifications across the system.
501 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
502 * (usually) take cgroup_mutex. These are the two most performance
503 * critical pieces of code here. The exception occurs on cgroup_exit(),
504 * when a task in a notify_on_release cgroup exits. Then cgroup_mutex
505 * is taken, and if the cgroup count is zero, a usermode call made
506 * to the release agent with the name of the cgroup (path relative to
507 * the root of cgroup file system) as the argument.
509 * A cgroup can only be deleted if both its 'count' of using tasks
510 * is zero, and its list of 'children' cgroups is empty. Since all
511 * tasks in the system use _some_ cgroup, and since there is always at
512 * least one task in the system (init, pid == 1), therefore, top_cgroup
513 * always has either children cgroups and/or using tasks. So we don't
514 * need a special hack to ensure that top_cgroup cannot be deleted.
516 * The task_lock() exception
518 * The need for this exception arises from the action of
519 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
520 * another. It does so using cgroup_mutex, however there are
521 * several performance critical places that need to reference
522 * task->cgroup without the expense of grabbing a system global
523 * mutex. Therefore except as noted below, when dereferencing or, as
524 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
525 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
526 * the task_struct routinely used for such matters.
528 * P.S. One more locking exception. RCU is used to guard the
529 * update of a tasks cgroup pointer by cgroup_attach_task()
533 * cgroup_lock - lock out any changes to cgroup structures
536 void cgroup_lock(void)
538 mutex_lock(&cgroup_mutex);
542 * cgroup_unlock - release lock on cgroup changes
544 * Undo the lock taken in a previous cgroup_lock() call.
546 void cgroup_unlock(void)
548 mutex_unlock(&cgroup_mutex);
552 * A couple of forward declarations required, due to cyclic reference loop:
553 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
554 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
558 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
559 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
560 static int cgroup_populate_dir(struct cgroup *cgrp);
561 static struct inode_operations cgroup_dir_inode_operations;
562 static struct file_operations proc_cgroupstats_operations;
564 static struct backing_dev_info cgroup_backing_dev_info = {
565 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
568 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
570 struct inode *inode = new_inode(sb);
573 inode->i_mode = mode;
574 inode->i_uid = current->fsuid;
575 inode->i_gid = current->fsgid;
577 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
578 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
584 * Call subsys's pre_destroy handler.
585 * This is called before css refcnt check.
587 static void cgroup_call_pre_destroy(struct cgroup *cgrp)
589 struct cgroup_subsys *ss;
590 for_each_subsys(cgrp->root, ss)
591 if (ss->pre_destroy && cgrp->subsys[ss->subsys_id])
592 ss->pre_destroy(ss, cgrp);
596 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
598 /* is dentry a directory ? if so, kfree() associated cgroup */
599 if (S_ISDIR(inode->i_mode)) {
600 struct cgroup *cgrp = dentry->d_fsdata;
601 struct cgroup_subsys *ss;
602 BUG_ON(!(cgroup_is_removed(cgrp)));
603 /* It's possible for external users to be holding css
604 * reference counts on a cgroup; css_put() needs to
605 * be able to access the cgroup after decrementing
606 * the reference count in order to know if it needs to
607 * queue the cgroup to be handled by the release
611 mutex_lock(&cgroup_mutex);
613 * Release the subsystem state objects.
615 for_each_subsys(cgrp->root, ss) {
616 if (cgrp->subsys[ss->subsys_id])
617 ss->destroy(ss, cgrp);
620 cgrp->root->number_of_cgroups--;
621 mutex_unlock(&cgroup_mutex);
623 /* Drop the active superblock reference that we took when we
624 * created the cgroup */
625 deactivate_super(cgrp->root->sb);
632 static void remove_dir(struct dentry *d)
634 struct dentry *parent = dget(d->d_parent);
637 simple_rmdir(parent->d_inode, d);
641 static void cgroup_clear_directory(struct dentry *dentry)
643 struct list_head *node;
645 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
646 spin_lock(&dcache_lock);
647 node = dentry->d_subdirs.next;
648 while (node != &dentry->d_subdirs) {
649 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
652 /* This should never be called on a cgroup
653 * directory with child cgroups */
654 BUG_ON(d->d_inode->i_mode & S_IFDIR);
656 spin_unlock(&dcache_lock);
658 simple_unlink(dentry->d_inode, d);
660 spin_lock(&dcache_lock);
662 node = dentry->d_subdirs.next;
664 spin_unlock(&dcache_lock);
668 * NOTE : the dentry must have been dget()'ed
670 static void cgroup_d_remove_dir(struct dentry *dentry)
672 cgroup_clear_directory(dentry);
674 spin_lock(&dcache_lock);
675 list_del_init(&dentry->d_u.d_child);
676 spin_unlock(&dcache_lock);
680 static int rebind_subsystems(struct cgroupfs_root *root,
681 unsigned long final_bits)
683 unsigned long added_bits, removed_bits;
684 struct cgroup *cgrp = &root->top_cgroup;
687 removed_bits = root->actual_subsys_bits & ~final_bits;
688 added_bits = final_bits & ~root->actual_subsys_bits;
689 /* Check that any added subsystems are currently free */
690 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
691 unsigned long bit = 1UL << i;
692 struct cgroup_subsys *ss = subsys[i];
693 if (!(bit & added_bits))
695 if (ss->root != &rootnode) {
696 /* Subsystem isn't free */
701 /* Currently we don't handle adding/removing subsystems when
702 * any child cgroups exist. This is theoretically supportable
703 * but involves complex error handling, so it's being left until
705 if (!list_empty(&cgrp->children))
708 /* Process each subsystem */
709 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
710 struct cgroup_subsys *ss = subsys[i];
711 unsigned long bit = 1UL << i;
712 if (bit & added_bits) {
713 /* We're binding this subsystem to this hierarchy */
714 BUG_ON(cgrp->subsys[i]);
715 BUG_ON(!dummytop->subsys[i]);
716 BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
717 cgrp->subsys[i] = dummytop->subsys[i];
718 cgrp->subsys[i]->cgroup = cgrp;
719 list_add(&ss->sibling, &root->subsys_list);
720 rcu_assign_pointer(ss->root, root);
724 } else if (bit & removed_bits) {
725 /* We're removing this subsystem */
726 BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
727 BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
729 ss->bind(ss, dummytop);
730 dummytop->subsys[i]->cgroup = dummytop;
731 cgrp->subsys[i] = NULL;
732 rcu_assign_pointer(subsys[i]->root, &rootnode);
733 list_del(&ss->sibling);
734 } else if (bit & final_bits) {
735 /* Subsystem state should already exist */
736 BUG_ON(!cgrp->subsys[i]);
738 /* Subsystem state shouldn't exist */
739 BUG_ON(cgrp->subsys[i]);
742 root->subsys_bits = root->actual_subsys_bits = final_bits;
748 static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
750 struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
751 struct cgroup_subsys *ss;
753 mutex_lock(&cgroup_mutex);
754 for_each_subsys(root, ss)
755 seq_printf(seq, ",%s", ss->name);
756 if (test_bit(ROOT_NOPREFIX, &root->flags))
757 seq_puts(seq, ",noprefix");
758 if (strlen(root->release_agent_path))
759 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
760 mutex_unlock(&cgroup_mutex);
764 struct cgroup_sb_opts {
765 unsigned long subsys_bits;
770 /* Convert a hierarchy specifier into a bitmask of subsystems and
772 static int parse_cgroupfs_options(char *data,
773 struct cgroup_sb_opts *opts)
775 char *token, *o = data ?: "all";
777 opts->subsys_bits = 0;
779 opts->release_agent = NULL;
781 while ((token = strsep(&o, ",")) != NULL) {
784 if (!strcmp(token, "all")) {
785 /* Add all non-disabled subsystems */
787 opts->subsys_bits = 0;
788 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
789 struct cgroup_subsys *ss = subsys[i];
791 opts->subsys_bits |= 1ul << i;
793 } else if (!strcmp(token, "noprefix")) {
794 set_bit(ROOT_NOPREFIX, &opts->flags);
795 } else if (!strncmp(token, "release_agent=", 14)) {
796 /* Specifying two release agents is forbidden */
797 if (opts->release_agent)
799 opts->release_agent = kzalloc(PATH_MAX, GFP_KERNEL);
800 if (!opts->release_agent)
802 strncpy(opts->release_agent, token + 14, PATH_MAX - 1);
803 opts->release_agent[PATH_MAX - 1] = 0;
805 struct cgroup_subsys *ss;
807 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
809 if (!strcmp(token, ss->name)) {
811 set_bit(i, &opts->subsys_bits);
815 if (i == CGROUP_SUBSYS_COUNT)
820 /* We can't have an empty hierarchy */
821 if (!opts->subsys_bits)
827 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
830 struct cgroupfs_root *root = sb->s_fs_info;
831 struct cgroup *cgrp = &root->top_cgroup;
832 struct cgroup_sb_opts opts;
834 mutex_lock(&cgrp->dentry->d_inode->i_mutex);
835 mutex_lock(&cgroup_mutex);
837 /* See what subsystems are wanted */
838 ret = parse_cgroupfs_options(data, &opts);
842 /* Don't allow flags to change at remount */
843 if (opts.flags != root->flags) {
848 ret = rebind_subsystems(root, opts.subsys_bits);
850 /* (re)populate subsystem files */
852 cgroup_populate_dir(cgrp);
854 if (opts.release_agent)
855 strcpy(root->release_agent_path, opts.release_agent);
857 if (opts.release_agent)
858 kfree(opts.release_agent);
859 mutex_unlock(&cgroup_mutex);
860 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
864 static struct super_operations cgroup_ops = {
865 .statfs = simple_statfs,
866 .drop_inode = generic_delete_inode,
867 .show_options = cgroup_show_options,
868 .remount_fs = cgroup_remount,
871 static void init_cgroup_root(struct cgroupfs_root *root)
873 struct cgroup *cgrp = &root->top_cgroup;
874 INIT_LIST_HEAD(&root->subsys_list);
875 INIT_LIST_HEAD(&root->root_list);
876 root->number_of_cgroups = 1;
878 cgrp->top_cgroup = cgrp;
879 INIT_LIST_HEAD(&cgrp->sibling);
880 INIT_LIST_HEAD(&cgrp->children);
881 INIT_LIST_HEAD(&cgrp->css_sets);
882 INIT_LIST_HEAD(&cgrp->release_list);
885 static int cgroup_test_super(struct super_block *sb, void *data)
887 struct cgroupfs_root *new = data;
888 struct cgroupfs_root *root = sb->s_fs_info;
890 /* First check subsystems */
891 if (new->subsys_bits != root->subsys_bits)
894 /* Next check flags */
895 if (new->flags != root->flags)
901 static int cgroup_set_super(struct super_block *sb, void *data)
904 struct cgroupfs_root *root = data;
906 ret = set_anon_super(sb, NULL);
910 sb->s_fs_info = root;
913 sb->s_blocksize = PAGE_CACHE_SIZE;
914 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
915 sb->s_magic = CGROUP_SUPER_MAGIC;
916 sb->s_op = &cgroup_ops;
921 static int cgroup_get_rootdir(struct super_block *sb)
923 struct inode *inode =
924 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
925 struct dentry *dentry;
930 inode->i_fop = &simple_dir_operations;
931 inode->i_op = &cgroup_dir_inode_operations;
932 /* directories start off with i_nlink == 2 (for "." entry) */
934 dentry = d_alloc_root(inode);
943 static int cgroup_get_sb(struct file_system_type *fs_type,
944 int flags, const char *unused_dev_name,
945 void *data, struct vfsmount *mnt)
947 struct cgroup_sb_opts opts;
949 struct super_block *sb;
950 struct cgroupfs_root *root;
951 struct list_head tmp_cg_links, *l;
952 INIT_LIST_HEAD(&tmp_cg_links);
954 /* First find the desired set of subsystems */
955 ret = parse_cgroupfs_options(data, &opts);
957 if (opts.release_agent)
958 kfree(opts.release_agent);
962 root = kzalloc(sizeof(*root), GFP_KERNEL);
964 if (opts.release_agent)
965 kfree(opts.release_agent);
969 init_cgroup_root(root);
970 root->subsys_bits = opts.subsys_bits;
971 root->flags = opts.flags;
972 if (opts.release_agent) {
973 strcpy(root->release_agent_path, opts.release_agent);
974 kfree(opts.release_agent);
977 sb = sget(fs_type, cgroup_test_super, cgroup_set_super, root);
984 if (sb->s_fs_info != root) {
985 /* Reusing an existing superblock */
986 BUG_ON(sb->s_root == NULL);
991 struct cgroup *cgrp = &root->top_cgroup;
994 BUG_ON(sb->s_root != NULL);
996 ret = cgroup_get_rootdir(sb);
999 inode = sb->s_root->d_inode;
1001 mutex_lock(&inode->i_mutex);
1002 mutex_lock(&cgroup_mutex);
1005 * We're accessing css_set_count without locking
1006 * css_set_lock here, but that's OK - it can only be
1007 * increased by someone holding cgroup_lock, and
1008 * that's us. The worst that can happen is that we
1009 * have some link structures left over
1011 ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1013 mutex_unlock(&cgroup_mutex);
1014 mutex_unlock(&inode->i_mutex);
1015 goto drop_new_super;
1018 ret = rebind_subsystems(root, root->subsys_bits);
1019 if (ret == -EBUSY) {
1020 mutex_unlock(&cgroup_mutex);
1021 mutex_unlock(&inode->i_mutex);
1022 goto drop_new_super;
1025 /* EBUSY should be the only error here */
1028 list_add(&root->root_list, &roots);
1031 sb->s_root->d_fsdata = &root->top_cgroup;
1032 root->top_cgroup.dentry = sb->s_root;
1034 /* Link the top cgroup in this hierarchy into all
1035 * the css_set objects */
1036 write_lock(&css_set_lock);
1037 l = &init_css_set.list;
1040 struct cg_cgroup_link *link;
1041 cg = list_entry(l, struct css_set, list);
1042 BUG_ON(list_empty(&tmp_cg_links));
1043 link = list_entry(tmp_cg_links.next,
1044 struct cg_cgroup_link,
1046 list_del(&link->cgrp_link_list);
1048 list_add(&link->cgrp_link_list,
1049 &root->top_cgroup.css_sets);
1050 list_add(&link->cg_link_list, &cg->cg_links);
1052 } while (l != &init_css_set.list);
1053 write_unlock(&css_set_lock);
1055 free_cg_links(&tmp_cg_links);
1057 BUG_ON(!list_empty(&cgrp->sibling));
1058 BUG_ON(!list_empty(&cgrp->children));
1059 BUG_ON(root->number_of_cgroups != 1);
1061 cgroup_populate_dir(cgrp);
1062 mutex_unlock(&inode->i_mutex);
1063 mutex_unlock(&cgroup_mutex);
1066 return simple_set_mnt(mnt, sb);
1069 up_write(&sb->s_umount);
1070 deactivate_super(sb);
1071 free_cg_links(&tmp_cg_links);
1075 static void cgroup_kill_sb(struct super_block *sb) {
1076 struct cgroupfs_root *root = sb->s_fs_info;
1077 struct cgroup *cgrp = &root->top_cgroup;
1082 BUG_ON(root->number_of_cgroups != 1);
1083 BUG_ON(!list_empty(&cgrp->children));
1084 BUG_ON(!list_empty(&cgrp->sibling));
1086 mutex_lock(&cgroup_mutex);
1088 /* Rebind all subsystems back to the default hierarchy */
1089 ret = rebind_subsystems(root, 0);
1090 /* Shouldn't be able to fail ... */
1094 * Release all the links from css_sets to this hierarchy's
1097 write_lock(&css_set_lock);
1098 while (!list_empty(&cgrp->css_sets)) {
1099 struct cg_cgroup_link *link;
1100 link = list_entry(cgrp->css_sets.next,
1101 struct cg_cgroup_link, cgrp_link_list);
1102 list_del(&link->cg_link_list);
1103 list_del(&link->cgrp_link_list);
1106 write_unlock(&css_set_lock);
1108 if (!list_empty(&root->root_list)) {
1109 list_del(&root->root_list);
1112 mutex_unlock(&cgroup_mutex);
1115 kill_litter_super(sb);
1118 static struct file_system_type cgroup_fs_type = {
1120 .get_sb = cgroup_get_sb,
1121 .kill_sb = cgroup_kill_sb,
1124 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1126 return dentry->d_fsdata;
1129 static inline struct cftype *__d_cft(struct dentry *dentry)
1131 return dentry->d_fsdata;
1135 * cgroup_path - generate the path of a cgroup
1136 * @cgrp: the cgroup in question
1137 * @buf: the buffer to write the path into
1138 * @buflen: the length of the buffer
1140 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1141 * Returns 0 on success, -errno on error.
1143 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1147 if (cgrp == dummytop) {
1149 * Inactive subsystems have no dentry for their root
1156 start = buf + buflen;
1160 int len = cgrp->dentry->d_name.len;
1161 if ((start -= len) < buf)
1162 return -ENAMETOOLONG;
1163 memcpy(start, cgrp->dentry->d_name.name, len);
1164 cgrp = cgrp->parent;
1170 return -ENAMETOOLONG;
1173 memmove(buf, start, buf + buflen - start);
1178 * Return the first subsystem attached to a cgroup's hierarchy, and
1182 static void get_first_subsys(const struct cgroup *cgrp,
1183 struct cgroup_subsys_state **css, int *subsys_id)
1185 const struct cgroupfs_root *root = cgrp->root;
1186 const struct cgroup_subsys *test_ss;
1187 BUG_ON(list_empty(&root->subsys_list));
1188 test_ss = list_entry(root->subsys_list.next,
1189 struct cgroup_subsys, sibling);
1191 *css = cgrp->subsys[test_ss->subsys_id];
1195 *subsys_id = test_ss->subsys_id;
1199 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1200 * @cgrp: the cgroup the task is attaching to
1201 * @tsk: the task to be attached
1203 * Call holding cgroup_mutex. May take task_lock of
1204 * the task 'tsk' during call.
1206 int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1209 struct cgroup_subsys *ss;
1210 struct cgroup *oldcgrp;
1211 struct css_set *cg = tsk->cgroups;
1212 struct css_set *newcg;
1213 struct cgroupfs_root *root = cgrp->root;
1216 get_first_subsys(cgrp, NULL, &subsys_id);
1218 /* Nothing to do if the task is already in that cgroup */
1219 oldcgrp = task_cgroup(tsk, subsys_id);
1220 if (cgrp == oldcgrp)
1223 for_each_subsys(root, ss) {
1224 if (ss->can_attach) {
1225 retval = ss->can_attach(ss, cgrp, tsk);
1232 * Locate or allocate a new css_set for this task,
1233 * based on its final set of cgroups
1235 newcg = find_css_set(cg, cgrp);
1240 if (tsk->flags & PF_EXITING) {
1245 rcu_assign_pointer(tsk->cgroups, newcg);
1248 /* Update the css_set linked lists if we're using them */
1249 write_lock(&css_set_lock);
1250 if (!list_empty(&tsk->cg_list)) {
1251 list_del(&tsk->cg_list);
1252 list_add(&tsk->cg_list, &newcg->tasks);
1254 write_unlock(&css_set_lock);
1256 for_each_subsys(root, ss) {
1258 ss->attach(ss, cgrp, oldcgrp, tsk);
1260 set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1267 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with
1268 * cgroup_mutex, may take task_lock of task
1270 static int attach_task_by_pid(struct cgroup *cgrp, char *pidbuf)
1273 struct task_struct *tsk;
1276 if (sscanf(pidbuf, "%d", &pid) != 1)
1281 tsk = find_task_by_vpid(pid);
1282 if (!tsk || tsk->flags & PF_EXITING) {
1286 get_task_struct(tsk);
1289 if ((current->euid) && (current->euid != tsk->uid)
1290 && (current->euid != tsk->suid)) {
1291 put_task_struct(tsk);
1296 get_task_struct(tsk);
1299 ret = cgroup_attach_task(cgrp, tsk);
1300 put_task_struct(tsk);
1304 /* The various types of files and directories in a cgroup file system */
1305 enum cgroup_filetype {
1309 FILE_NOTIFY_ON_RELEASE,
1314 static ssize_t cgroup_write_u64(struct cgroup *cgrp, struct cftype *cft,
1316 const char __user *userbuf,
1317 size_t nbytes, loff_t *unused_ppos)
1326 if (nbytes >= sizeof(buffer))
1328 if (copy_from_user(buffer, userbuf, nbytes))
1331 buffer[nbytes] = 0; /* nul-terminate */
1333 val = simple_strtoull(buffer, &end, 0);
1337 /* Pass to subsystem */
1338 retval = cft->write_u64(cgrp, cft, val);
1344 static ssize_t cgroup_common_file_write(struct cgroup *cgrp,
1347 const char __user *userbuf,
1348 size_t nbytes, loff_t *unused_ppos)
1350 enum cgroup_filetype type = cft->private;
1354 if (nbytes >= PATH_MAX)
1357 /* +1 for nul-terminator */
1358 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1362 if (copy_from_user(buffer, userbuf, nbytes)) {
1366 buffer[nbytes] = 0; /* nul-terminate */
1367 strstrip(buffer); /* strip -just- trailing whitespace */
1369 mutex_lock(&cgroup_mutex);
1372 * This was already checked for in cgroup_file_write(), but
1373 * check again now we're holding cgroup_mutex.
1375 if (cgroup_is_removed(cgrp)) {
1382 retval = attach_task_by_pid(cgrp, buffer);
1384 case FILE_NOTIFY_ON_RELEASE:
1385 clear_bit(CGRP_RELEASABLE, &cgrp->flags);
1386 if (simple_strtoul(buffer, NULL, 10) != 0)
1387 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
1389 clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
1391 case FILE_RELEASE_AGENT:
1392 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1393 strcpy(cgrp->root->release_agent_path, buffer);
1403 mutex_unlock(&cgroup_mutex);
1409 static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1410 size_t nbytes, loff_t *ppos)
1412 struct cftype *cft = __d_cft(file->f_dentry);
1413 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1415 if (!cft || cgroup_is_removed(cgrp))
1418 return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1420 return cgroup_write_u64(cgrp, cft, file, buf, nbytes, ppos);
1424 static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1426 char __user *buf, size_t nbytes,
1430 u64 val = cft->read_u64(cgrp, cft);
1431 int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1433 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1436 static ssize_t cgroup_common_file_read(struct cgroup *cgrp,
1440 size_t nbytes, loff_t *ppos)
1442 enum cgroup_filetype type = cft->private;
1447 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1453 case FILE_RELEASE_AGENT:
1455 struct cgroupfs_root *root;
1457 mutex_lock(&cgroup_mutex);
1459 n = strnlen(root->release_agent_path,
1460 sizeof(root->release_agent_path));
1461 n = min(n, (size_t) PAGE_SIZE);
1462 strncpy(s, root->release_agent_path, n);
1463 mutex_unlock(&cgroup_mutex);
1473 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1475 free_page((unsigned long)page);
1479 static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1480 size_t nbytes, loff_t *ppos)
1482 struct cftype *cft = __d_cft(file->f_dentry);
1483 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1485 if (!cft || cgroup_is_removed(cgrp))
1489 return cft->read(cgrp, cft, file, buf, nbytes, ppos);
1491 return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
1495 static int cgroup_file_open(struct inode *inode, struct file *file)
1500 err = generic_file_open(inode, file);
1504 cft = __d_cft(file->f_dentry);
1508 err = cft->open(inode, file);
1515 static int cgroup_file_release(struct inode *inode, struct file *file)
1517 struct cftype *cft = __d_cft(file->f_dentry);
1519 return cft->release(inode, file);
1524 * cgroup_rename - Only allow simple rename of directories in place.
1526 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
1527 struct inode *new_dir, struct dentry *new_dentry)
1529 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1531 if (new_dentry->d_inode)
1533 if (old_dir != new_dir)
1535 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1538 static struct file_operations cgroup_file_operations = {
1539 .read = cgroup_file_read,
1540 .write = cgroup_file_write,
1541 .llseek = generic_file_llseek,
1542 .open = cgroup_file_open,
1543 .release = cgroup_file_release,
1546 static struct inode_operations cgroup_dir_inode_operations = {
1547 .lookup = simple_lookup,
1548 .mkdir = cgroup_mkdir,
1549 .rmdir = cgroup_rmdir,
1550 .rename = cgroup_rename,
1553 static int cgroup_create_file(struct dentry *dentry, int mode,
1554 struct super_block *sb)
1556 static struct dentry_operations cgroup_dops = {
1557 .d_iput = cgroup_diput,
1560 struct inode *inode;
1564 if (dentry->d_inode)
1567 inode = cgroup_new_inode(mode, sb);
1571 if (S_ISDIR(mode)) {
1572 inode->i_op = &cgroup_dir_inode_operations;
1573 inode->i_fop = &simple_dir_operations;
1575 /* start off with i_nlink == 2 (for "." entry) */
1578 /* start with the directory inode held, so that we can
1579 * populate it without racing with another mkdir */
1580 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1581 } else if (S_ISREG(mode)) {
1583 inode->i_fop = &cgroup_file_operations;
1585 dentry->d_op = &cgroup_dops;
1586 d_instantiate(dentry, inode);
1587 dget(dentry); /* Extra count - pin the dentry in core */
1592 * cgroup_create_dir - create a directory for an object.
1593 * @cgrp: the cgroup we create the directory for. It must have a valid
1594 * ->parent field. And we are going to fill its ->dentry field.
1595 * @dentry: dentry of the new cgroup
1596 * @mode: mode to set on new directory.
1598 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
1601 struct dentry *parent;
1604 parent = cgrp->parent->dentry;
1605 error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
1607 dentry->d_fsdata = cgrp;
1608 inc_nlink(parent->d_inode);
1609 cgrp->dentry = dentry;
1617 int cgroup_add_file(struct cgroup *cgrp,
1618 struct cgroup_subsys *subsys,
1619 const struct cftype *cft)
1621 struct dentry *dir = cgrp->dentry;
1622 struct dentry *dentry;
1625 char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
1626 if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
1627 strcpy(name, subsys->name);
1630 strcat(name, cft->name);
1631 BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
1632 dentry = lookup_one_len(name, dir, strlen(name));
1633 if (!IS_ERR(dentry)) {
1634 error = cgroup_create_file(dentry, 0644 | S_IFREG,
1637 dentry->d_fsdata = (void *)cft;
1640 error = PTR_ERR(dentry);
1644 int cgroup_add_files(struct cgroup *cgrp,
1645 struct cgroup_subsys *subsys,
1646 const struct cftype cft[],
1650 for (i = 0; i < count; i++) {
1651 err = cgroup_add_file(cgrp, subsys, &cft[i]);
1659 * cgroup_task_count - count the number of tasks in a cgroup.
1660 * @cgrp: the cgroup in question
1662 * Return the number of tasks in the cgroup.
1664 int cgroup_task_count(const struct cgroup *cgrp)
1667 struct list_head *l;
1669 read_lock(&css_set_lock);
1670 l = cgrp->css_sets.next;
1671 while (l != &cgrp->css_sets) {
1672 struct cg_cgroup_link *link =
1673 list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1674 count += atomic_read(&link->cg->ref.refcount);
1677 read_unlock(&css_set_lock);
1682 * Advance a list_head iterator. The iterator should be positioned at
1683 * the start of a css_set
1685 static void cgroup_advance_iter(struct cgroup *cgrp,
1686 struct cgroup_iter *it)
1688 struct list_head *l = it->cg_link;
1689 struct cg_cgroup_link *link;
1692 /* Advance to the next non-empty css_set */
1695 if (l == &cgrp->css_sets) {
1699 link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1701 } while (list_empty(&cg->tasks));
1703 it->task = cg->tasks.next;
1707 * To reduce the fork() overhead for systems that are not actually
1708 * using their cgroups capability, we don't maintain the lists running
1709 * through each css_set to its tasks until we see the list actually
1710 * used - in other words after the first call to cgroup_iter_start().
1712 * The tasklist_lock is not held here, as do_each_thread() and
1713 * while_each_thread() are protected by RCU.
1715 static void cgroup_enable_task_cg_lists(void)
1717 struct task_struct *p, *g;
1718 write_lock(&css_set_lock);
1719 use_task_css_set_links = 1;
1720 do_each_thread(g, p) {
1723 * We should check if the process is exiting, otherwise
1724 * it will race with cgroup_exit() in that the list
1725 * entry won't be deleted though the process has exited.
1727 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
1728 list_add(&p->cg_list, &p->cgroups->tasks);
1730 } while_each_thread(g, p);
1731 write_unlock(&css_set_lock);
1734 void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
1737 * The first time anyone tries to iterate across a cgroup,
1738 * we need to enable the list linking each css_set to its
1739 * tasks, and fix up all existing tasks.
1741 if (!use_task_css_set_links)
1742 cgroup_enable_task_cg_lists();
1744 read_lock(&css_set_lock);
1745 it->cg_link = &cgrp->css_sets;
1746 cgroup_advance_iter(cgrp, it);
1749 struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
1750 struct cgroup_iter *it)
1752 struct task_struct *res;
1753 struct list_head *l = it->task;
1755 /* If the iterator cg is NULL, we have no tasks */
1758 res = list_entry(l, struct task_struct, cg_list);
1759 /* Advance iterator to find next entry */
1761 if (l == &res->cgroups->tasks) {
1762 /* We reached the end of this task list - move on to
1763 * the next cg_cgroup_link */
1764 cgroup_advance_iter(cgrp, it);
1771 void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
1773 read_unlock(&css_set_lock);
1776 static inline int started_after_time(struct task_struct *t1,
1777 struct timespec *time,
1778 struct task_struct *t2)
1780 int start_diff = timespec_compare(&t1->start_time, time);
1781 if (start_diff > 0) {
1783 } else if (start_diff < 0) {
1787 * Arbitrarily, if two processes started at the same
1788 * time, we'll say that the lower pointer value
1789 * started first. Note that t2 may have exited by now
1790 * so this may not be a valid pointer any longer, but
1791 * that's fine - it still serves to distinguish
1792 * between two tasks started (effectively) simultaneously.
1799 * This function is a callback from heap_insert() and is used to order
1801 * In this case we order the heap in descending task start time.
1803 static inline int started_after(void *p1, void *p2)
1805 struct task_struct *t1 = p1;
1806 struct task_struct *t2 = p2;
1807 return started_after_time(t1, &t2->start_time, t2);
1811 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
1812 * @scan: struct cgroup_scanner containing arguments for the scan
1814 * Arguments include pointers to callback functions test_task() and
1816 * Iterate through all the tasks in a cgroup, calling test_task() for each,
1817 * and if it returns true, call process_task() for it also.
1818 * The test_task pointer may be NULL, meaning always true (select all tasks).
1819 * Effectively duplicates cgroup_iter_{start,next,end}()
1820 * but does not lock css_set_lock for the call to process_task().
1821 * The struct cgroup_scanner may be embedded in any structure of the caller's
1823 * It is guaranteed that process_task() will act on every task that
1824 * is a member of the cgroup for the duration of this call. This
1825 * function may or may not call process_task() for tasks that exit
1826 * or move to a different cgroup during the call, or are forked or
1827 * move into the cgroup during the call.
1829 * Note that test_task() may be called with locks held, and may in some
1830 * situations be called multiple times for the same task, so it should
1832 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
1833 * pre-allocated and will be used for heap operations (and its "gt" member will
1834 * be overwritten), else a temporary heap will be used (allocation of which
1835 * may cause this function to fail).
1837 int cgroup_scan_tasks(struct cgroup_scanner *scan)
1840 struct cgroup_iter it;
1841 struct task_struct *p, *dropped;
1842 /* Never dereference latest_task, since it's not refcounted */
1843 struct task_struct *latest_task = NULL;
1844 struct ptr_heap tmp_heap;
1845 struct ptr_heap *heap;
1846 struct timespec latest_time = { 0, 0 };
1849 /* The caller supplied our heap and pre-allocated its memory */
1851 heap->gt = &started_after;
1853 /* We need to allocate our own heap memory */
1855 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
1857 /* cannot allocate the heap */
1863 * Scan tasks in the cgroup, using the scanner's "test_task" callback
1864 * to determine which are of interest, and using the scanner's
1865 * "process_task" callback to process any of them that need an update.
1866 * Since we don't want to hold any locks during the task updates,
1867 * gather tasks to be processed in a heap structure.
1868 * The heap is sorted by descending task start time.
1869 * If the statically-sized heap fills up, we overflow tasks that
1870 * started later, and in future iterations only consider tasks that
1871 * started after the latest task in the previous pass. This
1872 * guarantees forward progress and that we don't miss any tasks.
1875 cgroup_iter_start(scan->cg, &it);
1876 while ((p = cgroup_iter_next(scan->cg, &it))) {
1878 * Only affect tasks that qualify per the caller's callback,
1879 * if he provided one
1881 if (scan->test_task && !scan->test_task(p, scan))
1884 * Only process tasks that started after the last task
1887 if (!started_after_time(p, &latest_time, latest_task))
1889 dropped = heap_insert(heap, p);
1890 if (dropped == NULL) {
1892 * The new task was inserted; the heap wasn't
1896 } else if (dropped != p) {
1898 * The new task was inserted, and pushed out a
1902 put_task_struct(dropped);
1905 * Else the new task was newer than anything already in
1906 * the heap and wasn't inserted
1909 cgroup_iter_end(scan->cg, &it);
1912 for (i = 0; i < heap->size; i++) {
1913 struct task_struct *q = heap->ptrs[i];
1915 latest_time = q->start_time;
1918 /* Process the task per the caller's callback */
1919 scan->process_task(q, scan);
1923 * If we had to process any tasks at all, scan again
1924 * in case some of them were in the middle of forking
1925 * children that didn't get processed.
1926 * Not the most efficient way to do it, but it avoids
1927 * having to take callback_mutex in the fork path
1931 if (heap == &tmp_heap)
1932 heap_free(&tmp_heap);
1937 * Stuff for reading the 'tasks' file.
1939 * Reading this file can return large amounts of data if a cgroup has
1940 * *lots* of attached tasks. So it may need several calls to read(),
1941 * but we cannot guarantee that the information we produce is correct
1942 * unless we produce it entirely atomically.
1944 * Upon tasks file open(), a struct ctr_struct is allocated, that
1945 * will have a pointer to an array (also allocated here). The struct
1946 * ctr_struct * is stored in file->private_data. Its resources will
1947 * be freed by release() when the file is closed. The array is used
1948 * to sprintf the PIDs and then used by read().
1956 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
1957 * 'cgrp'. Return actual number of pids loaded. No need to
1958 * task_lock(p) when reading out p->cgroup, since we're in an RCU
1959 * read section, so the css_set can't go away, and is
1960 * immutable after creation.
1962 static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
1965 struct cgroup_iter it;
1966 struct task_struct *tsk;
1967 cgroup_iter_start(cgrp, &it);
1968 while ((tsk = cgroup_iter_next(cgrp, &it))) {
1969 if (unlikely(n == npids))
1971 pidarray[n++] = task_pid_vnr(tsk);
1973 cgroup_iter_end(cgrp, &it);
1978 * cgroupstats_build - build and fill cgroupstats
1979 * @stats: cgroupstats to fill information into
1980 * @dentry: A dentry entry belonging to the cgroup for which stats have
1983 * Build and fill cgroupstats so that taskstats can export it to user
1986 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
1989 struct cgroup *cgrp;
1990 struct cgroup_iter it;
1991 struct task_struct *tsk;
1993 * Validate dentry by checking the superblock operations
1995 if (dentry->d_sb->s_op != &cgroup_ops)
1999 cgrp = dentry->d_fsdata;
2002 cgroup_iter_start(cgrp, &it);
2003 while ((tsk = cgroup_iter_next(cgrp, &it))) {
2004 switch (tsk->state) {
2006 stats->nr_running++;
2008 case TASK_INTERRUPTIBLE:
2009 stats->nr_sleeping++;
2011 case TASK_UNINTERRUPTIBLE:
2012 stats->nr_uninterruptible++;
2015 stats->nr_stopped++;
2018 if (delayacct_is_task_waiting_on_io(tsk))
2019 stats->nr_io_wait++;
2023 cgroup_iter_end(cgrp, &it);
2030 static int cmppid(const void *a, const void *b)
2032 return *(pid_t *)a - *(pid_t *)b;
2036 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
2037 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
2038 * count 'cnt' of how many chars would be written if buf were large enough.
2040 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
2045 for (i = 0; i < npids; i++)
2046 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
2051 * Handle an open on 'tasks' file. Prepare a buffer listing the
2052 * process id's of tasks currently attached to the cgroup being opened.
2054 * Does not require any specific cgroup mutexes, and does not take any.
2056 static int cgroup_tasks_open(struct inode *unused, struct file *file)
2058 struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2059 struct ctr_struct *ctr;
2064 if (!(file->f_mode & FMODE_READ))
2067 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
2072 * If cgroup gets more users after we read count, we won't have
2073 * enough space - tough. This race is indistinguishable to the
2074 * caller from the case that the additional cgroup users didn't
2075 * show up until sometime later on.
2077 npids = cgroup_task_count(cgrp);
2079 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
2083 npids = pid_array_load(pidarray, npids, cgrp);
2084 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
2086 /* Call pid_array_to_buf() twice, first just to get bufsz */
2087 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
2088 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
2091 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
2098 file->private_data = ctr;
2109 static ssize_t cgroup_tasks_read(struct cgroup *cgrp,
2111 struct file *file, char __user *buf,
2112 size_t nbytes, loff_t *ppos)
2114 struct ctr_struct *ctr = file->private_data;
2116 return simple_read_from_buffer(buf, nbytes, ppos, ctr->buf, ctr->bufsz);
2119 static int cgroup_tasks_release(struct inode *unused_inode,
2122 struct ctr_struct *ctr;
2124 if (file->f_mode & FMODE_READ) {
2125 ctr = file->private_data;
2132 static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2135 return notify_on_release(cgrp);
2138 static u64 cgroup_read_releasable(struct cgroup *cgrp, struct cftype *cft)
2140 return test_bit(CGRP_RELEASABLE, &cgrp->flags);
2144 * for the common functions, 'private' gives the type of file
2146 static struct cftype files[] = {
2149 .open = cgroup_tasks_open,
2150 .read = cgroup_tasks_read,
2151 .write = cgroup_common_file_write,
2152 .release = cgroup_tasks_release,
2153 .private = FILE_TASKLIST,
2157 .name = "notify_on_release",
2158 .read_u64 = cgroup_read_notify_on_release,
2159 .write = cgroup_common_file_write,
2160 .private = FILE_NOTIFY_ON_RELEASE,
2164 .name = "releasable",
2165 .read_u64 = cgroup_read_releasable,
2166 .private = FILE_RELEASABLE,
2170 static struct cftype cft_release_agent = {
2171 .name = "release_agent",
2172 .read = cgroup_common_file_read,
2173 .write = cgroup_common_file_write,
2174 .private = FILE_RELEASE_AGENT,
2177 static int cgroup_populate_dir(struct cgroup *cgrp)
2180 struct cgroup_subsys *ss;
2182 /* First clear out any existing files */
2183 cgroup_clear_directory(cgrp->dentry);
2185 err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
2189 if (cgrp == cgrp->top_cgroup) {
2190 if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
2194 for_each_subsys(cgrp->root, ss) {
2195 if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
2202 static void init_cgroup_css(struct cgroup_subsys_state *css,
2203 struct cgroup_subsys *ss,
2204 struct cgroup *cgrp)
2207 atomic_set(&css->refcnt, 0);
2209 if (cgrp == dummytop)
2210 set_bit(CSS_ROOT, &css->flags);
2211 BUG_ON(cgrp->subsys[ss->subsys_id]);
2212 cgrp->subsys[ss->subsys_id] = css;
2216 * cgroup_create - create a cgroup
2217 * @parent: cgroup that will be parent of the new cgroup
2218 * @dentry: dentry of the new cgroup
2219 * @mode: mode to set on new inode
2221 * Must be called with the mutex on the parent inode held
2223 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
2226 struct cgroup *cgrp;
2227 struct cgroupfs_root *root = parent->root;
2229 struct cgroup_subsys *ss;
2230 struct super_block *sb = root->sb;
2232 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
2236 /* Grab a reference on the superblock so the hierarchy doesn't
2237 * get deleted on unmount if there are child cgroups. This
2238 * can be done outside cgroup_mutex, since the sb can't
2239 * disappear while someone has an open control file on the
2241 atomic_inc(&sb->s_active);
2243 mutex_lock(&cgroup_mutex);
2245 INIT_LIST_HEAD(&cgrp->sibling);
2246 INIT_LIST_HEAD(&cgrp->children);
2247 INIT_LIST_HEAD(&cgrp->css_sets);
2248 INIT_LIST_HEAD(&cgrp->release_list);
2250 cgrp->parent = parent;
2251 cgrp->root = parent->root;
2252 cgrp->top_cgroup = parent->top_cgroup;
2254 if (notify_on_release(parent))
2255 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2257 for_each_subsys(root, ss) {
2258 struct cgroup_subsys_state *css = ss->create(ss, cgrp);
2263 init_cgroup_css(css, ss, cgrp);
2266 list_add(&cgrp->sibling, &cgrp->parent->children);
2267 root->number_of_cgroups++;
2269 err = cgroup_create_dir(cgrp, dentry, mode);
2273 /* The cgroup directory was pre-locked for us */
2274 BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
2276 err = cgroup_populate_dir(cgrp);
2277 /* If err < 0, we have a half-filled directory - oh well ;) */
2279 mutex_unlock(&cgroup_mutex);
2280 mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
2286 list_del(&cgrp->sibling);
2287 root->number_of_cgroups--;
2291 for_each_subsys(root, ss) {
2292 if (cgrp->subsys[ss->subsys_id])
2293 ss->destroy(ss, cgrp);
2296 mutex_unlock(&cgroup_mutex);
2298 /* Release the reference count that we took on the superblock */
2299 deactivate_super(sb);
2305 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
2307 struct cgroup *c_parent = dentry->d_parent->d_fsdata;
2309 /* the vfs holds inode->i_mutex already */
2310 return cgroup_create(c_parent, dentry, mode | S_IFDIR);
2313 static inline int cgroup_has_css_refs(struct cgroup *cgrp)
2315 /* Check the reference count on each subsystem. Since we
2316 * already established that there are no tasks in the
2317 * cgroup, if the css refcount is also 0, then there should
2318 * be no outstanding references, so the subsystem is safe to
2319 * destroy. We scan across all subsystems rather than using
2320 * the per-hierarchy linked list of mounted subsystems since
2321 * we can be called via check_for_release() with no
2322 * synchronization other than RCU, and the subsystem linked
2323 * list isn't RCU-safe */
2325 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2326 struct cgroup_subsys *ss = subsys[i];
2327 struct cgroup_subsys_state *css;
2328 /* Skip subsystems not in this hierarchy */
2329 if (ss->root != cgrp->root)
2331 css = cgrp->subsys[ss->subsys_id];
2332 /* When called from check_for_release() it's possible
2333 * that by this point the cgroup has been removed
2334 * and the css deleted. But a false-positive doesn't
2335 * matter, since it can only happen if the cgroup
2336 * has been deleted and hence no longer needs the
2337 * release agent to be called anyway. */
2338 if (css && atomic_read(&css->refcnt))
2344 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
2346 struct cgroup *cgrp = dentry->d_fsdata;
2348 struct cgroup *parent;
2349 struct super_block *sb;
2350 struct cgroupfs_root *root;
2352 /* the vfs holds both inode->i_mutex already */
2354 mutex_lock(&cgroup_mutex);
2355 if (atomic_read(&cgrp->count) != 0) {
2356 mutex_unlock(&cgroup_mutex);
2359 if (!list_empty(&cgrp->children)) {
2360 mutex_unlock(&cgroup_mutex);
2364 parent = cgrp->parent;
2369 * Call pre_destroy handlers of subsys. Notify subsystems
2370 * that rmdir() request comes.
2372 cgroup_call_pre_destroy(cgrp);
2374 if (cgroup_has_css_refs(cgrp)) {
2375 mutex_unlock(&cgroup_mutex);
2379 spin_lock(&release_list_lock);
2380 set_bit(CGRP_REMOVED, &cgrp->flags);
2381 if (!list_empty(&cgrp->release_list))
2382 list_del(&cgrp->release_list);
2383 spin_unlock(&release_list_lock);
2384 /* delete my sibling from parent->children */
2385 list_del(&cgrp->sibling);
2386 spin_lock(&cgrp->dentry->d_lock);
2387 d = dget(cgrp->dentry);
2388 cgrp->dentry = NULL;
2389 spin_unlock(&d->d_lock);
2391 cgroup_d_remove_dir(d);
2394 set_bit(CGRP_RELEASABLE, &parent->flags);
2395 check_for_release(parent);
2397 mutex_unlock(&cgroup_mutex);
2401 static void cgroup_init_subsys(struct cgroup_subsys *ss)
2403 struct cgroup_subsys_state *css;
2404 struct list_head *l;
2406 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
2408 /* Create the top cgroup state for this subsystem */
2409 ss->root = &rootnode;
2410 css = ss->create(ss, dummytop);
2411 /* We don't handle early failures gracefully */
2412 BUG_ON(IS_ERR(css));
2413 init_cgroup_css(css, ss, dummytop);
2415 /* Update all cgroup groups to contain a subsys
2416 * pointer to this state - since the subsystem is
2417 * newly registered, all tasks and hence all cgroup
2418 * groups are in the subsystem's top cgroup. */
2419 write_lock(&css_set_lock);
2420 l = &init_css_set.list;
2422 struct css_set *cg =
2423 list_entry(l, struct css_set, list);
2424 cg->subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
2426 } while (l != &init_css_set.list);
2427 write_unlock(&css_set_lock);
2429 /* If this subsystem requested that it be notified with fork
2430 * events, we should send it one now for every process in the
2433 struct task_struct *g, *p;
2435 read_lock(&tasklist_lock);
2436 do_each_thread(g, p) {
2438 } while_each_thread(g, p);
2439 read_unlock(&tasklist_lock);
2442 need_forkexit_callback |= ss->fork || ss->exit;
2448 * cgroup_init_early - cgroup initialization at system boot
2450 * Initialize cgroups at system boot, and initialize any
2451 * subsystems that request early init.
2453 int __init cgroup_init_early(void)
2456 kref_init(&init_css_set.ref);
2457 kref_get(&init_css_set.ref);
2458 INIT_LIST_HEAD(&init_css_set.list);
2459 INIT_LIST_HEAD(&init_css_set.cg_links);
2460 INIT_LIST_HEAD(&init_css_set.tasks);
2462 init_cgroup_root(&rootnode);
2463 list_add(&rootnode.root_list, &roots);
2465 init_task.cgroups = &init_css_set;
2467 init_css_set_link.cg = &init_css_set;
2468 list_add(&init_css_set_link.cgrp_link_list,
2469 &rootnode.top_cgroup.css_sets);
2470 list_add(&init_css_set_link.cg_link_list,
2471 &init_css_set.cg_links);
2473 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2474 struct cgroup_subsys *ss = subsys[i];
2477 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
2478 BUG_ON(!ss->create);
2479 BUG_ON(!ss->destroy);
2480 if (ss->subsys_id != i) {
2481 printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
2482 ss->name, ss->subsys_id);
2487 cgroup_init_subsys(ss);
2493 * cgroup_init - cgroup initialization
2495 * Register cgroup filesystem and /proc file, and initialize
2496 * any subsystems that didn't request early init.
2498 int __init cgroup_init(void)
2502 struct proc_dir_entry *entry;
2504 err = bdi_init(&cgroup_backing_dev_info);
2508 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2509 struct cgroup_subsys *ss = subsys[i];
2510 if (!ss->early_init)
2511 cgroup_init_subsys(ss);
2514 err = register_filesystem(&cgroup_fs_type);
2518 entry = create_proc_entry("cgroups", 0, NULL);
2520 entry->proc_fops = &proc_cgroupstats_operations;
2524 bdi_destroy(&cgroup_backing_dev_info);
2530 * proc_cgroup_show()
2531 * - Print task's cgroup paths into seq_file, one line for each hierarchy
2532 * - Used for /proc/<pid>/cgroup.
2533 * - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2534 * doesn't really matter if tsk->cgroup changes after we read it,
2535 * and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
2536 * anyway. No need to check that tsk->cgroup != NULL, thanks to
2537 * the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2538 * cgroup to top_cgroup.
2541 /* TODO: Use a proper seq_file iterator */
2542 static int proc_cgroup_show(struct seq_file *m, void *v)
2545 struct task_struct *tsk;
2548 struct cgroupfs_root *root;
2551 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2557 tsk = get_pid_task(pid, PIDTYPE_PID);
2563 mutex_lock(&cgroup_mutex);
2565 for_each_root(root) {
2566 struct cgroup_subsys *ss;
2567 struct cgroup *cgrp;
2571 /* Skip this hierarchy if it has no active subsystems */
2572 if (!root->actual_subsys_bits)
2574 seq_printf(m, "%lu:", root->subsys_bits);
2575 for_each_subsys(root, ss)
2576 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
2578 get_first_subsys(&root->top_cgroup, NULL, &subsys_id);
2579 cgrp = task_cgroup(tsk, subsys_id);
2580 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
2588 mutex_unlock(&cgroup_mutex);
2589 put_task_struct(tsk);
2596 static int cgroup_open(struct inode *inode, struct file *file)
2598 struct pid *pid = PROC_I(inode)->pid;
2599 return single_open(file, proc_cgroup_show, pid);
2602 struct file_operations proc_cgroup_operations = {
2603 .open = cgroup_open,
2605 .llseek = seq_lseek,
2606 .release = single_release,
2609 /* Display information about each subsystem and each hierarchy */
2610 static int proc_cgroupstats_show(struct seq_file *m, void *v)
2614 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
2615 mutex_lock(&cgroup_mutex);
2616 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2617 struct cgroup_subsys *ss = subsys[i];
2618 seq_printf(m, "%s\t%lu\t%d\t%d\n",
2619 ss->name, ss->root->subsys_bits,
2620 ss->root->number_of_cgroups, !ss->disabled);
2622 mutex_unlock(&cgroup_mutex);
2626 static int cgroupstats_open(struct inode *inode, struct file *file)
2628 return single_open(file, proc_cgroupstats_show, NULL);
2631 static struct file_operations proc_cgroupstats_operations = {
2632 .open = cgroupstats_open,
2634 .llseek = seq_lseek,
2635 .release = single_release,
2639 * cgroup_fork - attach newly forked task to its parents cgroup.
2640 * @child: pointer to task_struct of forking parent process.
2642 * Description: A task inherits its parent's cgroup at fork().
2644 * A pointer to the shared css_set was automatically copied in
2645 * fork.c by dup_task_struct(). However, we ignore that copy, since
2646 * it was not made under the protection of RCU or cgroup_mutex, so
2647 * might no longer be a valid cgroup pointer. cgroup_attach_task() might
2648 * have already changed current->cgroups, allowing the previously
2649 * referenced cgroup group to be removed and freed.
2651 * At the point that cgroup_fork() is called, 'current' is the parent
2652 * task, and the passed argument 'child' points to the child task.
2654 void cgroup_fork(struct task_struct *child)
2657 child->cgroups = current->cgroups;
2658 get_css_set(child->cgroups);
2659 task_unlock(current);
2660 INIT_LIST_HEAD(&child->cg_list);
2664 * cgroup_fork_callbacks - run fork callbacks
2665 * @child: the new task
2667 * Called on a new task very soon before adding it to the
2668 * tasklist. No need to take any locks since no-one can
2669 * be operating on this task.
2671 void cgroup_fork_callbacks(struct task_struct *child)
2673 if (need_forkexit_callback) {
2675 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2676 struct cgroup_subsys *ss = subsys[i];
2678 ss->fork(ss, child);
2684 * cgroup_post_fork - called on a new task after adding it to the task list
2685 * @child: the task in question
2687 * Adds the task to the list running through its css_set if necessary.
2688 * Has to be after the task is visible on the task list in case we race
2689 * with the first call to cgroup_iter_start() - to guarantee that the
2690 * new task ends up on its list.
2692 void cgroup_post_fork(struct task_struct *child)
2694 if (use_task_css_set_links) {
2695 write_lock(&css_set_lock);
2696 if (list_empty(&child->cg_list))
2697 list_add(&child->cg_list, &child->cgroups->tasks);
2698 write_unlock(&css_set_lock);
2702 * cgroup_exit - detach cgroup from exiting task
2703 * @tsk: pointer to task_struct of exiting process
2704 * @run_callback: run exit callbacks?
2706 * Description: Detach cgroup from @tsk and release it.
2708 * Note that cgroups marked notify_on_release force every task in
2709 * them to take the global cgroup_mutex mutex when exiting.
2710 * This could impact scaling on very large systems. Be reluctant to
2711 * use notify_on_release cgroups where very high task exit scaling
2712 * is required on large systems.
2714 * the_top_cgroup_hack:
2716 * Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2718 * We call cgroup_exit() while the task is still competent to
2719 * handle notify_on_release(), then leave the task attached to the
2720 * root cgroup in each hierarchy for the remainder of its exit.
2722 * To do this properly, we would increment the reference count on
2723 * top_cgroup, and near the very end of the kernel/exit.c do_exit()
2724 * code we would add a second cgroup function call, to drop that
2725 * reference. This would just create an unnecessary hot spot on
2726 * the top_cgroup reference count, to no avail.
2728 * Normally, holding a reference to a cgroup without bumping its
2729 * count is unsafe. The cgroup could go away, or someone could
2730 * attach us to a different cgroup, decrementing the count on
2731 * the first cgroup that we never incremented. But in this case,
2732 * top_cgroup isn't going away, and either task has PF_EXITING set,
2733 * which wards off any cgroup_attach_task() attempts, or task is a failed
2734 * fork, never visible to cgroup_attach_task.
2736 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
2741 if (run_callbacks && need_forkexit_callback) {
2742 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2743 struct cgroup_subsys *ss = subsys[i];
2750 * Unlink from the css_set task list if necessary.
2751 * Optimistically check cg_list before taking
2754 if (!list_empty(&tsk->cg_list)) {
2755 write_lock(&css_set_lock);
2756 if (!list_empty(&tsk->cg_list))
2757 list_del(&tsk->cg_list);
2758 write_unlock(&css_set_lock);
2761 /* Reassign the task to the init_css_set. */
2764 tsk->cgroups = &init_css_set;
2767 put_css_set_taskexit(cg);
2771 * cgroup_clone - clone the cgroup the given subsystem is attached to
2772 * @tsk: the task to be moved
2773 * @subsys: the given subsystem
2775 * Duplicate the current cgroup in the hierarchy that the given
2776 * subsystem is attached to, and move this task into the new
2779 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys)
2781 struct dentry *dentry;
2783 char nodename[MAX_CGROUP_TYPE_NAMELEN];
2784 struct cgroup *parent, *child;
2785 struct inode *inode;
2787 struct cgroupfs_root *root;
2788 struct cgroup_subsys *ss;
2790 /* We shouldn't be called by an unregistered subsystem */
2791 BUG_ON(!subsys->active);
2793 /* First figure out what hierarchy and cgroup we're dealing
2794 * with, and pin them so we can drop cgroup_mutex */
2795 mutex_lock(&cgroup_mutex);
2797 root = subsys->root;
2798 if (root == &rootnode) {
2800 "Not cloning cgroup for unused subsystem %s\n",
2802 mutex_unlock(&cgroup_mutex);
2806 parent = task_cgroup(tsk, subsys->subsys_id);
2808 snprintf(nodename, MAX_CGROUP_TYPE_NAMELEN, "node_%d", tsk->pid);
2810 /* Pin the hierarchy */
2811 atomic_inc(&parent->root->sb->s_active);
2813 /* Keep the cgroup alive */
2815 mutex_unlock(&cgroup_mutex);
2817 /* Now do the VFS work to create a cgroup */
2818 inode = parent->dentry->d_inode;
2820 /* Hold the parent directory mutex across this operation to
2821 * stop anyone else deleting the new cgroup */
2822 mutex_lock(&inode->i_mutex);
2823 dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
2824 if (IS_ERR(dentry)) {
2826 "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
2828 ret = PTR_ERR(dentry);
2832 /* Create the cgroup directory, which also creates the cgroup */
2833 ret = vfs_mkdir(inode, dentry, S_IFDIR | 0755);
2834 child = __d_cgrp(dentry);
2838 "Failed to create cgroup %s: %d\n", nodename,
2845 "Couldn't find new cgroup %s\n", nodename);
2850 /* The cgroup now exists. Retake cgroup_mutex and check
2851 * that we're still in the same state that we thought we
2853 mutex_lock(&cgroup_mutex);
2854 if ((root != subsys->root) ||
2855 (parent != task_cgroup(tsk, subsys->subsys_id))) {
2856 /* Aargh, we raced ... */
2857 mutex_unlock(&inode->i_mutex);
2860 deactivate_super(parent->root->sb);
2861 /* The cgroup is still accessible in the VFS, but
2862 * we're not going to try to rmdir() it at this
2865 "Race in cgroup_clone() - leaking cgroup %s\n",
2870 /* do any required auto-setup */
2871 for_each_subsys(root, ss) {
2873 ss->post_clone(ss, child);
2876 /* All seems fine. Finish by moving the task into the new cgroup */
2877 ret = cgroup_attach_task(child, tsk);
2878 mutex_unlock(&cgroup_mutex);
2881 mutex_unlock(&inode->i_mutex);
2883 mutex_lock(&cgroup_mutex);
2885 mutex_unlock(&cgroup_mutex);
2886 deactivate_super(parent->root->sb);
2891 * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
2892 * @cgrp: the cgroup in question
2894 * See if @cgrp is a descendant of the current task's cgroup in
2895 * the appropriate hierarchy.
2897 * If we are sending in dummytop, then presumably we are creating
2898 * the top cgroup in the subsystem.
2900 * Called only by the ns (nsproxy) cgroup.
2902 int cgroup_is_descendant(const struct cgroup *cgrp)
2905 struct cgroup *target;
2908 if (cgrp == dummytop)
2911 get_first_subsys(cgrp, NULL, &subsys_id);
2912 target = task_cgroup(current, subsys_id);
2913 while (cgrp != target && cgrp!= cgrp->top_cgroup)
2914 cgrp = cgrp->parent;
2915 ret = (cgrp == target);
2919 static void check_for_release(struct cgroup *cgrp)
2921 /* All of these checks rely on RCU to keep the cgroup
2922 * structure alive */
2923 if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
2924 && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
2925 /* Control Group is currently removeable. If it's not
2926 * already queued for a userspace notification, queue
2928 int need_schedule_work = 0;
2929 spin_lock(&release_list_lock);
2930 if (!cgroup_is_removed(cgrp) &&
2931 list_empty(&cgrp->release_list)) {
2932 list_add(&cgrp->release_list, &release_list);
2933 need_schedule_work = 1;
2935 spin_unlock(&release_list_lock);
2936 if (need_schedule_work)
2937 schedule_work(&release_agent_work);
2941 void __css_put(struct cgroup_subsys_state *css)
2943 struct cgroup *cgrp = css->cgroup;
2945 if (atomic_dec_and_test(&css->refcnt) && notify_on_release(cgrp)) {
2946 set_bit(CGRP_RELEASABLE, &cgrp->flags);
2947 check_for_release(cgrp);
2953 * Notify userspace when a cgroup is released, by running the
2954 * configured release agent with the name of the cgroup (path
2955 * relative to the root of cgroup file system) as the argument.
2957 * Most likely, this user command will try to rmdir this cgroup.
2959 * This races with the possibility that some other task will be
2960 * attached to this cgroup before it is removed, or that some other
2961 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
2962 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
2963 * unused, and this cgroup will be reprieved from its death sentence,
2964 * to continue to serve a useful existence. Next time it's released,
2965 * we will get notified again, if it still has 'notify_on_release' set.
2967 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
2968 * means only wait until the task is successfully execve()'d. The
2969 * separate release agent task is forked by call_usermodehelper(),
2970 * then control in this thread returns here, without waiting for the
2971 * release agent task. We don't bother to wait because the caller of
2972 * this routine has no use for the exit status of the release agent
2973 * task, so no sense holding our caller up for that.
2975 static void cgroup_release_agent(struct work_struct *work)
2977 BUG_ON(work != &release_agent_work);
2978 mutex_lock(&cgroup_mutex);
2979 spin_lock(&release_list_lock);
2980 while (!list_empty(&release_list)) {
2981 char *argv[3], *envp[3];
2984 struct cgroup *cgrp = list_entry(release_list.next,
2987 list_del_init(&cgrp->release_list);
2988 spin_unlock(&release_list_lock);
2989 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2991 spin_lock(&release_list_lock);
2995 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0) {
2997 spin_lock(&release_list_lock);
3002 argv[i++] = cgrp->root->release_agent_path;
3003 argv[i++] = (char *)pathbuf;
3007 /* minimal command environment */
3008 envp[i++] = "HOME=/";
3009 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3012 /* Drop the lock while we invoke the usermode helper,
3013 * since the exec could involve hitting disk and hence
3014 * be a slow process */
3015 mutex_unlock(&cgroup_mutex);
3016 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
3018 mutex_lock(&cgroup_mutex);
3019 spin_lock(&release_list_lock);
3021 spin_unlock(&release_list_lock);
3022 mutex_unlock(&cgroup_mutex);
3025 static int __init cgroup_disable(char *str)
3030 while ((token = strsep(&str, ",")) != NULL) {
3034 for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3035 struct cgroup_subsys *ss = subsys[i];
3037 if (!strcmp(token, ss->name)) {
3039 printk(KERN_INFO "Disabling %s control group"
3040 " subsystem\n", ss->name);
3047 __setup("cgroup_disable=", cgroup_disable);