linux/kernel/cgroup.c
<<
>>
Prefs
   1/*
   2 *  Generic process-grouping system.
   3 *
   4 *  Based originally on the cpuset system, extracted by Paul Menage
   5 *  Copyright (C) 2006 Google, Inc
   6 *
   7 *  Copyright notices from the original cpuset code:
   8 *  --------------------------------------------------
   9 *  Copyright (C) 2003 BULL SA.
  10 *  Copyright (C) 2004-2006 Silicon Graphics, Inc.
  11 *
  12 *  Portions derived from Patrick Mochel's sysfs code.
  13 *  sysfs is Copyright (c) 2001-3 Patrick Mochel
  14 *
  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 *  ---------------------------------------------------
  19 *
  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.
  23 */
  24
  25#include <linux/cgroup.h>
  26#include <linux/errno.h>
  27#include <linux/fs.h>
  28#include <linux/kernel.h>
  29#include <linux/list.h>
  30#include <linux/mm.h>
  31#include <linux/mutex.h>
  32#include <linux/mount.h>
  33#include <linux/pagemap.h>
  34#include <linux/proc_fs.h>
  35#include <linux/rcupdate.h>
  36#include <linux/sched.h>
  37#include <linux/backing-dev.h>
  38#include <linux/seq_file.h>
  39#include <linux/slab.h>
  40#include <linux/magic.h>
  41#include <linux/spinlock.h>
  42#include <linux/string.h>
  43#include <linux/sort.h>
  44#include <linux/kmod.h>
  45#include <linux/delayacct.h>
  46#include <linux/cgroupstats.h>
  47#include <linux/hash.h>
  48#include <linux/namei.h>
  49
  50#include <asm/atomic.h>
  51
  52static DEFINE_MUTEX(cgroup_mutex);
  53
  54/* Generate an array of cgroup subsystem pointers */
  55#define SUBSYS(_x) &_x ## _subsys,
  56
  57static struct cgroup_subsys *subsys[] = {
  58#include <linux/cgroup_subsys.h>
  59};
  60
  61/*
  62 * A cgroupfs_root represents the root of a cgroup hierarchy,
  63 * and may be associated with a superblock to form an active
  64 * hierarchy
  65 */
  66struct cgroupfs_root {
  67        struct super_block *sb;
  68
  69        /*
  70         * The bitmask of subsystems intended to be attached to this
  71         * hierarchy
  72         */
  73        unsigned long subsys_bits;
  74
  75        /* The bitmask of subsystems currently attached to this hierarchy */
  76        unsigned long actual_subsys_bits;
  77
  78        /* A list running through the attached subsystems */
  79        struct list_head subsys_list;
  80
  81        /* The root cgroup for this hierarchy */
  82        struct cgroup top_cgroup;
  83
  84        /* Tracks how many cgroups are currently defined in hierarchy.*/
  85        int number_of_cgroups;
  86
  87        /* A list running through the mounted hierarchies */
  88        struct list_head root_list;
  89
  90        /* Hierarchy-specific flags */
  91        unsigned long flags;
  92
  93        /* The path to use for release notifications. */
  94        char release_agent_path[PATH_MAX];
  95};
  96
  97
  98/*
  99 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
 100 * subsystems that are otherwise unattached - it never has more than a
 101 * single cgroup, and all tasks are part of that cgroup.
 102 */
 103static struct cgroupfs_root rootnode;
 104
 105/* The list of hierarchy roots */
 106
 107static LIST_HEAD(roots);
 108static int root_count;
 109
 110/* dummytop is a shorthand for the dummy hierarchy's top cgroup */
 111#define dummytop (&rootnode.top_cgroup)
 112
 113/* This flag indicates whether tasks in the fork and exit paths should
 114 * check for fork/exit handlers to call. This avoids us having to do
 115 * extra work in the fork/exit path if none of the subsystems need to
 116 * be called.
 117 */
 118static int need_forkexit_callback __read_mostly;
 119static int need_mm_owner_callback __read_mostly;
 120
 121/* convenient tests for these bits */
 122inline int cgroup_is_removed(const struct cgroup *cgrp)
 123{
 124        return test_bit(CGRP_REMOVED, &cgrp->flags);
 125}
 126
 127/* bits in struct cgroupfs_root flags field */
 128enum {
 129        ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
 130};
 131
 132static int cgroup_is_releasable(const struct cgroup *cgrp)
 133{
 134        const int bits =
 135                (1 << CGRP_RELEASABLE) |
 136                (1 << CGRP_NOTIFY_ON_RELEASE);
 137        return (cgrp->flags & bits) == bits;
 138}
 139
 140static int notify_on_release(const struct cgroup *cgrp)
 141{
 142        return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
 143}
 144
 145/*
 146 * for_each_subsys() allows you to iterate on each subsystem attached to
 147 * an active hierarchy
 148 */
 149#define for_each_subsys(_root, _ss) \
 150list_for_each_entry(_ss, &_root->subsys_list, sibling)
 151
 152/* for_each_root() allows you to iterate across the active hierarchies */
 153#define for_each_root(_root) \
 154list_for_each_entry(_root, &roots, root_list)
 155
 156/* the list of cgroups eligible for automatic release. Protected by
 157 * release_list_lock */
 158static LIST_HEAD(release_list);
 159static DEFINE_SPINLOCK(release_list_lock);
 160static void cgroup_release_agent(struct work_struct *work);
 161static DECLARE_WORK(release_agent_work, cgroup_release_agent);
 162static void check_for_release(struct cgroup *cgrp);
 163
 164/* Link structure for associating css_set objects with cgroups */
 165struct cg_cgroup_link {
 166        /*
 167         * List running through cg_cgroup_links associated with a
 168         * cgroup, anchored on cgroup->css_sets
 169         */
 170        struct list_head cgrp_link_list;
 171        /*
 172         * List running through cg_cgroup_links pointing at a
 173         * single css_set object, anchored on css_set->cg_links
 174         */
 175        struct list_head cg_link_list;
 176        struct css_set *cg;
 177};
 178
 179/* The default css_set - used by init and its children prior to any
 180 * hierarchies being mounted. It contains a pointer to the root state
 181 * for each subsystem. Also used to anchor the list of css_sets. Not
 182 * reference-counted, to improve performance when child cgroups
 183 * haven't been created.
 184 */
 185
 186static struct css_set init_css_set;
 187static struct cg_cgroup_link init_css_set_link;
 188
 189/* css_set_lock protects the list of css_set objects, and the
 190 * chain of tasks off each css_set.  Nests outside task->alloc_lock
 191 * due to cgroup_iter_start() */
 192static DEFINE_RWLOCK(css_set_lock);
 193static int css_set_count;
 194
 195/* hash table for cgroup groups. This improves the performance to
 196 * find an existing css_set */
 197#define CSS_SET_HASH_BITS       7
 198#define CSS_SET_TABLE_SIZE      (1 << CSS_SET_HASH_BITS)
 199static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
 200
 201static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
 202{
 203        int i;
 204        int index;
 205        unsigned long tmp = 0UL;
 206
 207        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
 208                tmp += (unsigned long)css[i];
 209        tmp = (tmp >> 16) ^ tmp;
 210
 211        index = hash_long(tmp, CSS_SET_HASH_BITS);
 212
 213        return &css_set_table[index];
 214}
 215
 216/* We don't maintain the lists running through each css_set to its
 217 * task until after the first call to cgroup_iter_start(). This
 218 * reduces the fork()/exit() overhead for people who have cgroups
 219 * compiled into their kernel but not actually in use */
 220static int use_task_css_set_links __read_mostly;
 221
 222/* When we create or destroy a css_set, the operation simply
 223 * takes/releases a reference count on all the cgroups referenced
 224 * by subsystems in this css_set. This can end up multiple-counting
 225 * some cgroups, but that's OK - the ref-count is just a
 226 * busy/not-busy indicator; ensuring that we only count each cgroup
 227 * once would require taking a global lock to ensure that no
 228 * subsystems moved between hierarchies while we were doing so.
 229 *
 230 * Possible TODO: decide at boot time based on the number of
 231 * registered subsystems and the number of CPUs or NUMA nodes whether
 232 * it's better for performance to ref-count every subsystem, or to
 233 * take a global lock and only add one ref count to each hierarchy.
 234 */
 235
 236/*
 237 * unlink a css_set from the list and free it
 238 */
 239static void unlink_css_set(struct css_set *cg)
 240{
 241        struct cg_cgroup_link *link;
 242        struct cg_cgroup_link *saved_link;
 243
 244        hlist_del(&cg->hlist);
 245        css_set_count--;
 246
 247        list_for_each_entry_safe(link, saved_link, &cg->cg_links,
 248                                 cg_link_list) {
 249                list_del(&link->cg_link_list);
 250                list_del(&link->cgrp_link_list);
 251                kfree(link);
 252        }
 253}
 254
 255static void __put_css_set(struct css_set *cg, int taskexit)
 256{
 257        int i;
 258        /*
 259         * Ensure that the refcount doesn't hit zero while any readers
 260         * can see it. Similar to atomic_dec_and_lock(), but for an
 261         * rwlock
 262         */
 263        if (atomic_add_unless(&cg->refcount, -1, 1))
 264                return;
 265        write_lock(&css_set_lock);
 266        if (!atomic_dec_and_test(&cg->refcount)) {
 267                write_unlock(&css_set_lock);
 268                return;
 269        }
 270        unlink_css_set(cg);
 271        write_unlock(&css_set_lock);
 272
 273        rcu_read_lock();
 274        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 275                struct cgroup *cgrp = cg->subsys[i]->cgroup;
 276                if (atomic_dec_and_test(&cgrp->count) &&
 277                    notify_on_release(cgrp)) {
 278                        if (taskexit)
 279                                set_bit(CGRP_RELEASABLE, &cgrp->flags);
 280                        check_for_release(cgrp);
 281                }
 282        }
 283        rcu_read_unlock();
 284        kfree(cg);
 285}
 286
 287/*
 288 * refcounted get/put for css_set objects
 289 */
 290static inline void get_css_set(struct css_set *cg)
 291{
 292        atomic_inc(&cg->refcount);
 293}
 294
 295static inline void put_css_set(struct css_set *cg)
 296{
 297        __put_css_set(cg, 0);
 298}
 299
 300static inline void put_css_set_taskexit(struct css_set *cg)
 301{
 302        __put_css_set(cg, 1);
 303}
 304
 305/*
 306 * find_existing_css_set() is a helper for
 307 * find_css_set(), and checks to see whether an existing
 308 * css_set is suitable.
 309 *
 310 * oldcg: the cgroup group that we're using before the cgroup
 311 * transition
 312 *
 313 * cgrp: the cgroup that we're moving into
 314 *
 315 * template: location in which to build the desired set of subsystem
 316 * state objects for the new cgroup group
 317 */
 318static struct css_set *find_existing_css_set(
 319        struct css_set *oldcg,
 320        struct cgroup *cgrp,
 321        struct cgroup_subsys_state *template[])
 322{
 323        int i;
 324        struct cgroupfs_root *root = cgrp->root;
 325        struct hlist_head *hhead;
 326        struct hlist_node *node;
 327        struct css_set *cg;
 328
 329        /* Built the set of subsystem state objects that we want to
 330         * see in the new css_set */
 331        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 332                if (root->subsys_bits & (1UL << i)) {
 333                        /* Subsystem is in this hierarchy. So we want
 334                         * the subsystem state from the new
 335                         * cgroup */
 336                        template[i] = cgrp->subsys[i];
 337                } else {
 338                        /* Subsystem is not in this hierarchy, so we
 339                         * don't want to change the subsystem state */
 340                        template[i] = oldcg->subsys[i];
 341                }
 342        }
 343
 344        hhead = css_set_hash(template);
 345        hlist_for_each_entry(cg, node, hhead, hlist) {
 346                if (!memcmp(template, cg->subsys, sizeof(cg->subsys))) {
 347                        /* All subsystems matched */
 348                        return cg;
 349                }
 350        }
 351
 352        /* No existing cgroup group matched */
 353        return NULL;
 354}
 355
 356static void free_cg_links(struct list_head *tmp)
 357{
 358        struct cg_cgroup_link *link;
 359        struct cg_cgroup_link *saved_link;
 360
 361        list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
 362                list_del(&link->cgrp_link_list);
 363                kfree(link);
 364        }
 365}
 366
 367/*
 368 * allocate_cg_links() allocates "count" cg_cgroup_link structures
 369 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
 370 * success or a negative error
 371 */
 372static int allocate_cg_links(int count, struct list_head *tmp)
 373{
 374        struct cg_cgroup_link *link;
 375        int i;
 376        INIT_LIST_HEAD(tmp);
 377        for (i = 0; i < count; i++) {
 378                link = kmalloc(sizeof(*link), GFP_KERNEL);
 379                if (!link) {
 380                        free_cg_links(tmp);
 381                        return -ENOMEM;
 382                }
 383                list_add(&link->cgrp_link_list, tmp);
 384        }
 385        return 0;
 386}
 387
 388/*
 389 * find_css_set() takes an existing cgroup group and a
 390 * cgroup object, and returns a css_set object that's
 391 * equivalent to the old group, but with the given cgroup
 392 * substituted into the appropriate hierarchy. Must be called with
 393 * cgroup_mutex held
 394 */
 395static struct css_set *find_css_set(
 396        struct css_set *oldcg, struct cgroup *cgrp)
 397{
 398        struct css_set *res;
 399        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
 400        int i;
 401
 402        struct list_head tmp_cg_links;
 403        struct cg_cgroup_link *link;
 404
 405        struct hlist_head *hhead;
 406
 407        /* First see if we already have a cgroup group that matches
 408         * the desired set */
 409        read_lock(&css_set_lock);
 410        res = find_existing_css_set(oldcg, cgrp, template);
 411        if (res)
 412                get_css_set(res);
 413        read_unlock(&css_set_lock);
 414
 415        if (res)
 416                return res;
 417
 418        res = kmalloc(sizeof(*res), GFP_KERNEL);
 419        if (!res)
 420                return NULL;
 421
 422        /* Allocate all the cg_cgroup_link objects that we'll need */
 423        if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
 424                kfree(res);
 425                return NULL;
 426        }
 427
 428        atomic_set(&res->refcount, 1);
 429        INIT_LIST_HEAD(&res->cg_links);
 430        INIT_LIST_HEAD(&res->tasks);
 431        INIT_HLIST_NODE(&res->hlist);
 432
 433        /* Copy the set of subsystem state objects generated in
 434         * find_existing_css_set() */
 435        memcpy(res->subsys, template, sizeof(res->subsys));
 436
 437        write_lock(&css_set_lock);
 438        /* Add reference counts and links from the new css_set. */
 439        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 440                struct cgroup *cgrp = res->subsys[i]->cgroup;
 441                struct cgroup_subsys *ss = subsys[i];
 442                atomic_inc(&cgrp->count);
 443                /*
 444                 * We want to add a link once per cgroup, so we
 445                 * only do it for the first subsystem in each
 446                 * hierarchy
 447                 */
 448                if (ss->root->subsys_list.next == &ss->sibling) {
 449                        BUG_ON(list_empty(&tmp_cg_links));
 450                        link = list_entry(tmp_cg_links.next,
 451                                          struct cg_cgroup_link,
 452                                          cgrp_link_list);
 453                        list_del(&link->cgrp_link_list);
 454                        list_add(&link->cgrp_link_list, &cgrp->css_sets);
 455                        link->cg = res;
 456                        list_add(&link->cg_link_list, &res->cg_links);
 457                }
 458        }
 459        if (list_empty(&rootnode.subsys_list)) {
 460                link = list_entry(tmp_cg_links.next,
 461                                  struct cg_cgroup_link,
 462                                  cgrp_link_list);
 463                list_del(&link->cgrp_link_list);
 464                list_add(&link->cgrp_link_list, &dummytop->css_sets);
 465                link->cg = res;
 466                list_add(&link->cg_link_list, &res->cg_links);
 467        }
 468
 469        BUG_ON(!list_empty(&tmp_cg_links));
 470
 471        css_set_count++;
 472
 473        /* Add this cgroup group to the hash table */
 474        hhead = css_set_hash(res->subsys);
 475        hlist_add_head(&res->hlist, hhead);
 476
 477        write_unlock(&css_set_lock);
 478
 479        return res;
 480}
 481
 482/*
 483 * There is one global cgroup mutex. We also require taking
 484 * task_lock() when dereferencing a task's cgroup subsys pointers.
 485 * See "The task_lock() exception", at the end of this comment.
 486 *
 487 * A task must hold cgroup_mutex to modify cgroups.
 488 *
 489 * Any task can increment and decrement the count field without lock.
 490 * So in general, code holding cgroup_mutex can't rely on the count
 491 * field not changing.  However, if the count goes to zero, then only
 492 * cgroup_attach_task() can increment it again.  Because a count of zero
 493 * means that no tasks are currently attached, therefore there is no
 494 * way a task attached to that cgroup can fork (the other way to
 495 * increment the count).  So code holding cgroup_mutex can safely
 496 * assume that if the count is zero, it will stay zero. Similarly, if
 497 * a task holds cgroup_mutex on a cgroup with zero count, it
 498 * knows that the cgroup won't be removed, as cgroup_rmdir()
 499 * needs that mutex.
 500 *
 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.
 508 *
 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.
 515 *
 516 *      The task_lock() exception
 517 *
 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.
 527 *
 528 * P.S.  One more locking exception.  RCU is used to guard the
 529 * update of a tasks cgroup pointer by cgroup_attach_task()
 530 */
 531
 532/**
 533 * cgroup_lock - lock out any changes to cgroup structures
 534 *
 535 */
 536void cgroup_lock(void)
 537{
 538        mutex_lock(&cgroup_mutex);
 539}
 540
 541/**
 542 * cgroup_unlock - release lock on cgroup changes
 543 *
 544 * Undo the lock taken in a previous cgroup_lock() call.
 545 */
 546void cgroup_unlock(void)
 547{
 548        mutex_unlock(&cgroup_mutex);
 549}
 550
 551/*
 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
 555 * -> cgroup_mkdir.
 556 */
 557
 558static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
 559static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
 560static int cgroup_populate_dir(struct cgroup *cgrp);
 561static struct inode_operations cgroup_dir_inode_operations;
 562static struct file_operations proc_cgroupstats_operations;
 563
 564static struct backing_dev_info cgroup_backing_dev_info = {
 565        .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 566};
 567
 568static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
 569{
 570        struct inode *inode = new_inode(sb);
 571
 572        if (inode) {
 573                inode->i_mode = mode;
 574                inode->i_uid = current->fsuid;
 575                inode->i_gid = current->fsgid;
 576                inode->i_blocks = 0;
 577                inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 578                inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
 579        }
 580        return inode;
 581}
 582
 583/*
 584 * Call subsys's pre_destroy handler.
 585 * This is called before css refcnt check.
 586 */
 587static void cgroup_call_pre_destroy(struct cgroup *cgrp)
 588{
 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);
 593        return;
 594}
 595
 596static void cgroup_diput(struct dentry *dentry, struct inode *inode)
 597{
 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
 608                 * agent */
 609                synchronize_rcu();
 610
 611                mutex_lock(&cgroup_mutex);
 612                /*
 613                 * Release the subsystem state objects.
 614                 */
 615                for_each_subsys(cgrp->root, ss) {
 616                        if (cgrp->subsys[ss->subsys_id])
 617                                ss->destroy(ss, cgrp);
 618                }
 619
 620                cgrp->root->number_of_cgroups--;
 621                mutex_unlock(&cgroup_mutex);
 622
 623                /* Drop the active superblock reference that we took when we
 624                 * created the cgroup */
 625                deactivate_super(cgrp->root->sb);
 626
 627                kfree(cgrp);
 628        }
 629        iput(inode);
 630}
 631
 632static void remove_dir(struct dentry *d)
 633{
 634        struct dentry *parent = dget(d->d_parent);
 635
 636        d_delete(d);
 637        simple_rmdir(parent->d_inode, d);
 638        dput(parent);
 639}
 640
 641static void cgroup_clear_directory(struct dentry *dentry)
 642{
 643        struct list_head *node;
 644
 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);
 650                list_del_init(node);
 651                if (d->d_inode) {
 652                        /* This should never be called on a cgroup
 653                         * directory with child cgroups */
 654                        BUG_ON(d->d_inode->i_mode & S_IFDIR);
 655                        d = dget_locked(d);
 656                        spin_unlock(&dcache_lock);
 657                        d_delete(d);
 658                        simple_unlink(dentry->d_inode, d);
 659                        dput(d);
 660                        spin_lock(&dcache_lock);
 661                }
 662                node = dentry->d_subdirs.next;
 663        }
 664        spin_unlock(&dcache_lock);
 665}
 666
 667/*
 668 * NOTE : the dentry must have been dget()'ed
 669 */
 670static void cgroup_d_remove_dir(struct dentry *dentry)
 671{
 672        cgroup_clear_directory(dentry);
 673
 674        spin_lock(&dcache_lock);
 675        list_del_init(&dentry->d_u.d_child);
 676        spin_unlock(&dcache_lock);
 677        remove_dir(dentry);
 678}
 679
 680static int rebind_subsystems(struct cgroupfs_root *root,
 681                              unsigned long final_bits)
 682{
 683        unsigned long added_bits, removed_bits;
 684        struct cgroup *cgrp = &root->top_cgroup;
 685        int i;
 686
 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))
 694                        continue;
 695                if (ss->root != &rootnode) {
 696                        /* Subsystem isn't free */
 697                        return -EBUSY;
 698                }
 699        }
 700
 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
 704         * later */
 705        if (root->number_of_cgroups > 1)
 706                return -EBUSY;
 707
 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);
 721                        if (ss->bind)
 722                                ss->bind(ss, cgrp);
 723
 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);
 728                        if (ss->bind)
 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]);
 737                } else {
 738                        /* Subsystem state shouldn't exist */
 739                        BUG_ON(cgrp->subsys[i]);
 740                }
 741        }
 742        root->subsys_bits = root->actual_subsys_bits = final_bits;
 743        synchronize_rcu();
 744
 745        return 0;
 746}
 747
 748static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
 749{
 750        struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
 751        struct cgroup_subsys *ss;
 752
 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);
 761        return 0;
 762}
 763
 764struct cgroup_sb_opts {
 765        unsigned long subsys_bits;
 766        unsigned long flags;
 767        char *release_agent;
 768};
 769
 770/* Convert a hierarchy specifier into a bitmask of subsystems and
 771 * flags. */
 772static int parse_cgroupfs_options(char *data,
 773                                     struct cgroup_sb_opts *opts)
 774{
 775        char *token, *o = data ?: "all";
 776
 777        opts->subsys_bits = 0;
 778        opts->flags = 0;
 779        opts->release_agent = NULL;
 780
 781        while ((token = strsep(&o, ",")) != NULL) {
 782                if (!*token)
 783                        return -EINVAL;
 784                if (!strcmp(token, "all")) {
 785                        /* Add all non-disabled subsystems */
 786                        int i;
 787                        opts->subsys_bits = 0;
 788                        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 789                                struct cgroup_subsys *ss = subsys[i];
 790                                if (!ss->disabled)
 791                                        opts->subsys_bits |= 1ul << i;
 792                        }
 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)
 798                                return -EINVAL;
 799                        opts->release_agent = kzalloc(PATH_MAX, GFP_KERNEL);
 800                        if (!opts->release_agent)
 801                                return -ENOMEM;
 802                        strncpy(opts->release_agent, token + 14, PATH_MAX - 1);
 803                        opts->release_agent[PATH_MAX - 1] = 0;
 804                } else {
 805                        struct cgroup_subsys *ss;
 806                        int i;
 807                        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 808                                ss = subsys[i];
 809                                if (!strcmp(token, ss->name)) {
 810                                        if (!ss->disabled)
 811                                                set_bit(i, &opts->subsys_bits);
 812                                        break;
 813                                }
 814                        }
 815                        if (i == CGROUP_SUBSYS_COUNT)
 816                                return -ENOENT;
 817                }
 818        }
 819
 820        /* We can't have an empty hierarchy */
 821        if (!opts->subsys_bits)
 822                return -EINVAL;
 823
 824        return 0;
 825}
 826
 827static int cgroup_remount(struct super_block *sb, int *flags, char *data)
 828{
 829        int ret = 0;
 830        struct cgroupfs_root *root = sb->s_fs_info;
 831        struct cgroup *cgrp = &root->top_cgroup;
 832        struct cgroup_sb_opts opts;
 833
 834        mutex_lock(&cgrp->dentry->d_inode->i_mutex);
 835        mutex_lock(&cgroup_mutex);
 836
 837        /* See what subsystems are wanted */
 838        ret = parse_cgroupfs_options(data, &opts);
 839        if (ret)
 840                goto out_unlock;
 841
 842        /* Don't allow flags to change at remount */
 843        if (opts.flags != root->flags) {
 844                ret = -EINVAL;
 845                goto out_unlock;
 846        }
 847
 848        ret = rebind_subsystems(root, opts.subsys_bits);
 849
 850        /* (re)populate subsystem files */
 851        if (!ret)
 852                cgroup_populate_dir(cgrp);
 853
 854        if (opts.release_agent)
 855                strcpy(root->release_agent_path, opts.release_agent);
 856 out_unlock:
 857        if (opts.release_agent)
 858                kfree(opts.release_agent);
 859        mutex_unlock(&cgroup_mutex);
 860        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
 861        return ret;
 862}
 863
 864static 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,
 869};
 870
 871static void init_cgroup_housekeeping(struct cgroup *cgrp)
 872{
 873        INIT_LIST_HEAD(&cgrp->sibling);
 874        INIT_LIST_HEAD(&cgrp->children);
 875        INIT_LIST_HEAD(&cgrp->css_sets);
 876        INIT_LIST_HEAD(&cgrp->release_list);
 877        init_rwsem(&cgrp->pids_mutex);
 878}
 879static void init_cgroup_root(struct cgroupfs_root *root)
 880{
 881        struct cgroup *cgrp = &root->top_cgroup;
 882        INIT_LIST_HEAD(&root->subsys_list);
 883        INIT_LIST_HEAD(&root->root_list);
 884        root->number_of_cgroups = 1;
 885        cgrp->root = root;
 886        cgrp->top_cgroup = cgrp;
 887        init_cgroup_housekeeping(cgrp);
 888}
 889
 890static int cgroup_test_super(struct super_block *sb, void *data)
 891{
 892        struct cgroupfs_root *new = data;
 893        struct cgroupfs_root *root = sb->s_fs_info;
 894
 895        /* First check subsystems */
 896        if (new->subsys_bits != root->subsys_bits)
 897            return 0;
 898
 899        /* Next check flags */
 900        if (new->flags != root->flags)
 901                return 0;
 902
 903        return 1;
 904}
 905
 906static int cgroup_set_super(struct super_block *sb, void *data)
 907{
 908        int ret;
 909        struct cgroupfs_root *root = data;
 910
 911        ret = set_anon_super(sb, NULL);
 912        if (ret)
 913                return ret;
 914
 915        sb->s_fs_info = root;
 916        root->sb = sb;
 917
 918        sb->s_blocksize = PAGE_CACHE_SIZE;
 919        sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
 920        sb->s_magic = CGROUP_SUPER_MAGIC;
 921        sb->s_op = &cgroup_ops;
 922
 923        return 0;
 924}
 925
 926static int cgroup_get_rootdir(struct super_block *sb)
 927{
 928        struct inode *inode =
 929                cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
 930        struct dentry *dentry;
 931
 932        if (!inode)
 933                return -ENOMEM;
 934
 935        inode->i_fop = &simple_dir_operations;
 936        inode->i_op = &cgroup_dir_inode_operations;
 937        /* directories start off with i_nlink == 2 (for "." entry) */
 938        inc_nlink(inode);
 939        dentry = d_alloc_root(inode);
 940        if (!dentry) {
 941                iput(inode);
 942                return -ENOMEM;
 943        }
 944        sb->s_root = dentry;
 945        return 0;
 946}
 947
 948static int cgroup_get_sb(struct file_system_type *fs_type,
 949                         int flags, const char *unused_dev_name,
 950                         void *data, struct vfsmount *mnt)
 951{
 952        struct cgroup_sb_opts opts;
 953        int ret = 0;
 954        struct super_block *sb;
 955        struct cgroupfs_root *root;
 956        struct list_head tmp_cg_links;
 957
 958        /* First find the desired set of subsystems */
 959        ret = parse_cgroupfs_options(data, &opts);
 960        if (ret) {
 961                if (opts.release_agent)
 962                        kfree(opts.release_agent);
 963                return ret;
 964        }
 965
 966        root = kzalloc(sizeof(*root), GFP_KERNEL);
 967        if (!root) {
 968                if (opts.release_agent)
 969                        kfree(opts.release_agent);
 970                return -ENOMEM;
 971        }
 972
 973        init_cgroup_root(root);
 974        root->subsys_bits = opts.subsys_bits;
 975        root->flags = opts.flags;
 976        if (opts.release_agent) {
 977                strcpy(root->release_agent_path, opts.release_agent);
 978                kfree(opts.release_agent);
 979        }
 980
 981        sb = sget(fs_type, cgroup_test_super, cgroup_set_super, root);
 982
 983        if (IS_ERR(sb)) {
 984                kfree(root);
 985                return PTR_ERR(sb);
 986        }
 987
 988        if (sb->s_fs_info != root) {
 989                /* Reusing an existing superblock */
 990                BUG_ON(sb->s_root == NULL);
 991                kfree(root);
 992                root = NULL;
 993        } else {
 994                /* New superblock */
 995                struct cgroup *cgrp = &root->top_cgroup;
 996                struct inode *inode;
 997                int i;
 998
 999                BUG_ON(sb->s_root != NULL);
1000
1001                ret = cgroup_get_rootdir(sb);
1002                if (ret)
1003                        goto drop_new_super;
1004                inode = sb->s_root->d_inode;
1005
1006                mutex_lock(&inode->i_mutex);
1007                mutex_lock(&cgroup_mutex);
1008
1009                /*
1010                 * We're accessing css_set_count without locking
1011                 * css_set_lock here, but that's OK - it can only be
1012                 * increased by someone holding cgroup_lock, and
1013                 * that's us. The worst that can happen is that we
1014                 * have some link structures left over
1015                 */
1016                ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1017                if (ret) {
1018                        mutex_unlock(&cgroup_mutex);
1019                        mutex_unlock(&inode->i_mutex);
1020                        goto drop_new_super;
1021                }
1022
1023                ret = rebind_subsystems(root, root->subsys_bits);
1024                if (ret == -EBUSY) {
1025                        mutex_unlock(&cgroup_mutex);
1026                        mutex_unlock(&inode->i_mutex);
1027                        goto free_cg_links;
1028                }
1029
1030                /* EBUSY should be the only error here */
1031                BUG_ON(ret);
1032
1033                list_add(&root->root_list, &roots);
1034                root_count++;
1035
1036                sb->s_root->d_fsdata = &root->top_cgroup;
1037                root->top_cgroup.dentry = sb->s_root;
1038
1039                /* Link the top cgroup in this hierarchy into all
1040                 * the css_set objects */
1041                write_lock(&css_set_lock);
1042                for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1043                        struct hlist_head *hhead = &css_set_table[i];
1044                        struct hlist_node *node;
1045                        struct css_set *cg;
1046
1047                        hlist_for_each_entry(cg, node, hhead, hlist) {
1048                                struct cg_cgroup_link *link;
1049
1050                                BUG_ON(list_empty(&tmp_cg_links));
1051                                link = list_entry(tmp_cg_links.next,
1052                                                  struct cg_cgroup_link,
1053                                                  cgrp_link_list);
1054                                list_del(&link->cgrp_link_list);
1055                                link->cg = cg;
1056                                list_add(&link->cgrp_link_list,
1057                                         &root->top_cgroup.css_sets);
1058                                list_add(&link->cg_link_list, &cg->cg_links);
1059                        }
1060                }
1061                write_unlock(&css_set_lock);
1062
1063                free_cg_links(&tmp_cg_links);
1064
1065                BUG_ON(!list_empty(&cgrp->sibling));
1066                BUG_ON(!list_empty(&cgrp->children));
1067                BUG_ON(root->number_of_cgroups != 1);
1068
1069                cgroup_populate_dir(cgrp);
1070                mutex_unlock(&inode->i_mutex);
1071                mutex_unlock(&cgroup_mutex);
1072        }
1073
1074        return simple_set_mnt(mnt, sb);
1075
1076 free_cg_links:
1077        free_cg_links(&tmp_cg_links);
1078 drop_new_super:
1079        up_write(&sb->s_umount);
1080        deactivate_super(sb);
1081        return ret;
1082}
1083
1084static void cgroup_kill_sb(struct super_block *sb) {
1085        struct cgroupfs_root *root = sb->s_fs_info;
1086        struct cgroup *cgrp = &root->top_cgroup;
1087        int ret;
1088        struct cg_cgroup_link *link;
1089        struct cg_cgroup_link *saved_link;
1090
1091        BUG_ON(!root);
1092
1093        BUG_ON(root->number_of_cgroups != 1);
1094        BUG_ON(!list_empty(&cgrp->children));
1095        BUG_ON(!list_empty(&cgrp->sibling));
1096
1097        mutex_lock(&cgroup_mutex);
1098
1099        /* Rebind all subsystems back to the default hierarchy */
1100        ret = rebind_subsystems(root, 0);
1101        /* Shouldn't be able to fail ... */
1102        BUG_ON(ret);
1103
1104        /*
1105         * Release all the links from css_sets to this hierarchy's
1106         * root cgroup
1107         */
1108        write_lock(&css_set_lock);
1109
1110        list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1111                                 cgrp_link_list) {
1112                list_del(&link->cg_link_list);
1113                list_del(&link->cgrp_link_list);
1114                kfree(link);
1115        }
1116        write_unlock(&css_set_lock);
1117
1118        if (!list_empty(&root->root_list)) {
1119                list_del(&root->root_list);
1120                root_count--;
1121        }
1122        mutex_unlock(&cgroup_mutex);
1123
1124        kfree(root);
1125        kill_litter_super(sb);
1126}
1127
1128static struct file_system_type cgroup_fs_type = {
1129        .name = "cgroup",
1130        .get_sb = cgroup_get_sb,
1131        .kill_sb = cgroup_kill_sb,
1132};
1133
1134static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1135{
1136        return dentry->d_fsdata;
1137}
1138
1139static inline struct cftype *__d_cft(struct dentry *dentry)
1140{
1141        return dentry->d_fsdata;
1142}
1143
1144/**
1145 * cgroup_path - generate the path of a cgroup
1146 * @cgrp: the cgroup in question
1147 * @buf: the buffer to write the path into
1148 * @buflen: the length of the buffer
1149 *
1150 * Called with cgroup_mutex held. Writes path of cgroup into buf.
1151 * Returns 0 on success, -errno on error.
1152 */
1153int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1154{
1155        char *start;
1156
1157        if (cgrp == dummytop) {
1158                /*
1159                 * Inactive subsystems have no dentry for their root
1160                 * cgroup
1161                 */
1162                strcpy(buf, "/");
1163                return 0;
1164        }
1165
1166        start = buf + buflen;
1167
1168        *--start = '\0';
1169        for (;;) {
1170                int len = cgrp->dentry->d_name.len;
1171                if ((start -= len) < buf)
1172                        return -ENAMETOOLONG;
1173                memcpy(start, cgrp->dentry->d_name.name, len);
1174                cgrp = cgrp->parent;
1175                if (!cgrp)
1176                        break;
1177                if (!cgrp->parent)
1178                        continue;
1179                if (--start < buf)
1180                        return -ENAMETOOLONG;
1181                *start = '/';
1182        }
1183        memmove(buf, start, buf + buflen - start);
1184        return 0;
1185}
1186
1187/*
1188 * Return the first subsystem attached to a cgroup's hierarchy, and
1189 * its subsystem id.
1190 */
1191
1192static void get_first_subsys(const struct cgroup *cgrp,
1193                        struct cgroup_subsys_state **css, int *subsys_id)
1194{
1195        const struct cgroupfs_root *root = cgrp->root;
1196        const struct cgroup_subsys *test_ss;
1197        BUG_ON(list_empty(&root->subsys_list));
1198        test_ss = list_entry(root->subsys_list.next,
1199                             struct cgroup_subsys, sibling);
1200        if (css) {
1201                *css = cgrp->subsys[test_ss->subsys_id];
1202                BUG_ON(!*css);
1203        }
1204        if (subsys_id)
1205                *subsys_id = test_ss->subsys_id;
1206}
1207
1208/**
1209 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1210 * @cgrp: the cgroup the task is attaching to
1211 * @tsk: the task to be attached
1212 *
1213 * Call holding cgroup_mutex. May take task_lock of
1214 * the task 'tsk' during call.
1215 */
1216int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1217{
1218        int retval = 0;
1219        struct cgroup_subsys *ss;
1220        struct cgroup *oldcgrp;
1221        struct css_set *cg = tsk->cgroups;
1222        struct css_set *newcg;
1223        struct cgroupfs_root *root = cgrp->root;
1224        int subsys_id;
1225
1226        get_first_subsys(cgrp, NULL, &subsys_id);
1227
1228        /* Nothing to do if the task is already in that cgroup */
1229        oldcgrp = task_cgroup(tsk, subsys_id);
1230        if (cgrp == oldcgrp)
1231                return 0;
1232
1233        for_each_subsys(root, ss) {
1234                if (ss->can_attach) {
1235                        retval = ss->can_attach(ss, cgrp, tsk);
1236                        if (retval)
1237                                return retval;
1238                }
1239        }
1240
1241        /*
1242         * Locate or allocate a new css_set for this task,
1243         * based on its final set of cgroups
1244         */
1245        newcg = find_css_set(cg, cgrp);
1246        if (!newcg)
1247                return -ENOMEM;
1248
1249        task_lock(tsk);
1250        if (tsk->flags & PF_EXITING) {
1251                task_unlock(tsk);
1252                put_css_set(newcg);
1253                return -ESRCH;
1254        }
1255        rcu_assign_pointer(tsk->cgroups, newcg);
1256        task_unlock(tsk);
1257
1258        /* Update the css_set linked lists if we're using them */
1259        write_lock(&css_set_lock);
1260        if (!list_empty(&tsk->cg_list)) {
1261                list_del(&tsk->cg_list);
1262                list_add(&tsk->cg_list, &newcg->tasks);
1263        }
1264        write_unlock(&css_set_lock);
1265
1266        for_each_subsys(root, ss) {
1267                if (ss->attach)
1268                        ss->attach(ss, cgrp, oldcgrp, tsk);
1269        }
1270        set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1271        synchronize_rcu();
1272        put_css_set(cg);
1273        return 0;
1274}
1275
1276/*
1277 * Attach task with pid 'pid' to cgroup 'cgrp'. Call with cgroup_mutex
1278 * held. May take task_lock of task
1279 */
1280static int attach_task_by_pid(struct cgroup *cgrp, u64 pid)
1281{
1282        struct task_struct *tsk;
1283        int ret;
1284
1285        if (pid) {
1286                rcu_read_lock();
1287                tsk = find_task_by_vpid(pid);
1288                if (!tsk || tsk->flags & PF_EXITING) {
1289                        rcu_read_unlock();
1290                        return -ESRCH;
1291                }
1292                get_task_struct(tsk);
1293                rcu_read_unlock();
1294
1295                if ((current->euid) && (current->euid != tsk->uid)
1296                    && (current->euid != tsk->suid)) {
1297                        put_task_struct(tsk);
1298                        return -EACCES;
1299                }
1300        } else {
1301                tsk = current;
1302                get_task_struct(tsk);
1303        }
1304
1305        ret = cgroup_attach_task(cgrp, tsk);
1306        put_task_struct(tsk);
1307        return ret;
1308}
1309
1310static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
1311{
1312        int ret;
1313        if (!cgroup_lock_live_group(cgrp))
1314                return -ENODEV;
1315        ret = attach_task_by_pid(cgrp, pid);
1316        cgroup_unlock();
1317        return ret;
1318}
1319
1320/* The various types of files and directories in a cgroup file system */
1321enum cgroup_filetype {
1322        FILE_ROOT,
1323        FILE_DIR,
1324        FILE_TASKLIST,
1325        FILE_NOTIFY_ON_RELEASE,
1326        FILE_RELEASE_AGENT,
1327};
1328
1329/**
1330 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
1331 * @cgrp: the cgroup to be checked for liveness
1332 *
1333 * On success, returns true; the lock should be later released with
1334 * cgroup_unlock(). On failure returns false with no lock held.
1335 */
1336bool cgroup_lock_live_group(struct cgroup *cgrp)
1337{
1338        mutex_lock(&cgroup_mutex);
1339        if (cgroup_is_removed(cgrp)) {
1340                mutex_unlock(&cgroup_mutex);
1341                return false;
1342        }
1343        return true;
1344}
1345
1346static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
1347                                      const char *buffer)
1348{
1349        BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
1350        if (!cgroup_lock_live_group(cgrp))
1351                return -ENODEV;
1352        strcpy(cgrp->root->release_agent_path, buffer);
1353        cgroup_unlock();
1354        return 0;
1355}
1356
1357static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
1358                                     struct seq_file *seq)
1359{
1360        if (!cgroup_lock_live_group(cgrp))
1361                return -ENODEV;
1362        seq_puts(seq, cgrp->root->release_agent_path);
1363        seq_putc(seq, '\n');
1364        cgroup_unlock();
1365        return 0;
1366}
1367
1368/* A buffer size big enough for numbers or short strings */
1369#define CGROUP_LOCAL_BUFFER_SIZE 64
1370
1371static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
1372                                struct file *file,
1373                                const char __user *userbuf,
1374                                size_t nbytes, loff_t *unused_ppos)
1375{
1376        char buffer[CGROUP_LOCAL_BUFFER_SIZE];
1377        int retval = 0;
1378        char *end;
1379
1380        if (!nbytes)
1381                return -EINVAL;
1382        if (nbytes >= sizeof(buffer))
1383                return -E2BIG;
1384        if (copy_from_user(buffer, userbuf, nbytes))
1385                return -EFAULT;
1386
1387        buffer[nbytes] = 0;     /* nul-terminate */
1388        strstrip(buffer);
1389        if (cft->write_u64) {
1390                u64 val = simple_strtoull(buffer, &end, 0);
1391                if (*end)
1392                        return -EINVAL;
1393                retval = cft->write_u64(cgrp, cft, val);
1394        } else {
1395                s64 val = simple_strtoll(buffer, &end, 0);
1396                if (*end)
1397                        return -EINVAL;
1398                retval = cft->write_s64(cgrp, cft, val);
1399        }
1400        if (!retval)
1401                retval = nbytes;
1402        return retval;
1403}
1404
1405static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
1406                                   struct file *file,
1407                                   const char __user *userbuf,
1408                                   size_t nbytes, loff_t *unused_ppos)
1409{
1410        char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
1411        int retval = 0;
1412        size_t max_bytes = cft->max_write_len;
1413        char *buffer = local_buffer;
1414
1415        if (!max_bytes)
1416                max_bytes = sizeof(local_buffer) - 1;
1417        if (nbytes >= max_bytes)
1418                return -E2BIG;
1419        /* Allocate a dynamic buffer if we need one */
1420        if (nbytes >= sizeof(local_buffer)) {
1421                buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1422                if (buffer == NULL)
1423                        return -ENOMEM;
1424        }
1425        if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
1426                retval = -EFAULT;
1427                goto out;
1428        }
1429
1430        buffer[nbytes] = 0;     /* nul-terminate */
1431        strstrip(buffer);
1432        retval = cft->write_string(cgrp, cft, buffer);
1433        if (!retval)
1434                retval = nbytes;
1435out:
1436        if (buffer != local_buffer)
1437                kfree(buffer);
1438        return retval;
1439}
1440
1441static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
1442                                                size_t nbytes, loff_t *ppos)
1443{
1444        struct cftype *cft = __d_cft(file->f_dentry);
1445        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1446
1447        if (!cft || cgroup_is_removed(cgrp))
1448                return -ENODEV;
1449        if (cft->write)
1450                return cft->write(cgrp, cft, file, buf, nbytes, ppos);
1451        if (cft->write_u64 || cft->write_s64)
1452                return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
1453        if (cft->write_string)
1454                return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
1455        if (cft->trigger) {
1456                int ret = cft->trigger(cgrp, (unsigned int)cft->private);
1457                return ret ? ret : nbytes;
1458        }
1459        return -EINVAL;
1460}
1461
1462static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
1463                               struct file *file,
1464                               char __user *buf, size_t nbytes,
1465                               loff_t *ppos)
1466{
1467        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1468        u64 val = cft->read_u64(cgrp, cft);
1469        int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
1470
1471        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1472}
1473
1474static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
1475                               struct file *file,
1476                               char __user *buf, size_t nbytes,
1477                               loff_t *ppos)
1478{
1479        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
1480        s64 val = cft->read_s64(cgrp, cft);
1481        int len = sprintf(tmp, "%lld\n", (long long) val);
1482
1483        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
1484}
1485
1486static ssize_t cgroup_file_read(struct file *file, char __user *buf,
1487                                   size_t nbytes, loff_t *ppos)
1488{
1489        struct cftype *cft = __d_cft(file->f_dentry);
1490        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
1491
1492        if (!cft || cgroup_is_removed(cgrp))
1493                return -ENODEV;
1494
1495        if (cft->read)
1496                return cft->read(cgrp, cft, file, buf, nbytes, ppos);
1497        if (cft->read_u64)
1498                return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
1499        if (cft->read_s64)
1500                return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
1501        return -EINVAL;
1502}
1503
1504/*
1505 * seqfile ops/methods for returning structured data. Currently just
1506 * supports string->u64 maps, but can be extended in future.
1507 */
1508
1509struct cgroup_seqfile_state {
1510        struct cftype *cft;
1511        struct cgroup *cgroup;
1512};
1513
1514static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
1515{
1516        struct seq_file *sf = cb->state;
1517        return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
1518}
1519
1520static int cgroup_seqfile_show(struct seq_file *m, void *arg)
1521{
1522        struct cgroup_seqfile_state *state = m->private;
1523        struct cftype *cft = state->cft;
1524        if (cft->read_map) {
1525                struct cgroup_map_cb cb = {
1526                        .fill = cgroup_map_add,
1527                        .state = m,
1528                };
1529                return cft->read_map(state->cgroup, cft, &cb);
1530        }
1531        return cft->read_seq_string(state->cgroup, cft, m);
1532}
1533
1534static int cgroup_seqfile_release(struct inode *inode, struct file *file)
1535{
1536        struct seq_file *seq = file->private_data;
1537        kfree(seq->private);
1538        return single_release(inode, file);
1539}
1540
1541static struct file_operations cgroup_seqfile_operations = {
1542        .read = seq_read,
1543        .write = cgroup_file_write,
1544        .llseek = seq_lseek,
1545        .release = cgroup_seqfile_release,
1546};
1547
1548static int cgroup_file_open(struct inode *inode, struct file *file)
1549{
1550        int err;
1551        struct cftype *cft;
1552
1553        err = generic_file_open(inode, file);
1554        if (err)
1555                return err;
1556
1557        cft = __d_cft(file->f_dentry);
1558        if (!cft)
1559                return -ENODEV;
1560        if (cft->read_map || cft->read_seq_string) {
1561                struct cgroup_seqfile_state *state =
1562                        kzalloc(sizeof(*state), GFP_USER);
1563                if (!state)
1564                        return -ENOMEM;
1565                state->cft = cft;
1566                state->cgroup = __d_cgrp(file->f_dentry->d_parent);
1567                file->f_op = &cgroup_seqfile_operations;
1568                err = single_open(file, cgroup_seqfile_show, state);
1569                if (err < 0)
1570                        kfree(state);
1571        } else if (cft->open)
1572                err = cft->open(inode, file);
1573        else
1574                err = 0;
1575
1576        return err;
1577}
1578
1579static int cgroup_file_release(struct inode *inode, struct file *file)
1580{
1581        struct cftype *cft = __d_cft(file->f_dentry);
1582        if (cft->release)
1583                return cft->release(inode, file);
1584        return 0;
1585}
1586
1587/*
1588 * cgroup_rename - Only allow simple rename of directories in place.
1589 */
1590static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
1591                            struct inode *new_dir, struct dentry *new_dentry)
1592{
1593        if (!S_ISDIR(old_dentry->d_inode->i_mode))
1594                return -ENOTDIR;
1595        if (new_dentry->d_inode)
1596                return -EEXIST;
1597        if (old_dir != new_dir)
1598                return -EIO;
1599        return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1600}
1601
1602static struct file_operations cgroup_file_operations = {
1603        .read = cgroup_file_read,
1604        .write = cgroup_file_write,
1605        .llseek = generic_file_llseek,
1606        .open = cgroup_file_open,
1607        .release = cgroup_file_release,
1608};
1609
1610static struct inode_operations cgroup_dir_inode_operations = {
1611        .lookup = simple_lookup,
1612        .mkdir = cgroup_mkdir,
1613        .rmdir = cgroup_rmdir,
1614        .rename = cgroup_rename,
1615};
1616
1617static int cgroup_create_file(struct dentry *dentry, int mode,
1618                                struct super_block *sb)
1619{
1620        static struct dentry_operations cgroup_dops = {
1621                .d_iput = cgroup_diput,
1622        };
1623
1624        struct inode *inode;
1625
1626        if (!dentry)
1627                return -ENOENT;
1628        if (dentry->d_inode)
1629                return -EEXIST;
1630
1631        inode = cgroup_new_inode(mode, sb);
1632        if (!inode)
1633                return -ENOMEM;
1634
1635        if (S_ISDIR(mode)) {
1636                inode->i_op = &cgroup_dir_inode_operations;
1637                inode->i_fop = &simple_dir_operations;
1638
1639                /* start off with i_nlink == 2 (for "." entry) */
1640                inc_nlink(inode);
1641
1642                /* start with the directory inode held, so that we can
1643                 * populate it without racing with another mkdir */
1644                mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1645        } else if (S_ISREG(mode)) {
1646                inode->i_size = 0;
1647                inode->i_fop = &cgroup_file_operations;
1648        }
1649        dentry->d_op = &cgroup_dops;
1650        d_instantiate(dentry, inode);
1651        dget(dentry);   /* Extra count - pin the dentry in core */
1652        return 0;
1653}
1654
1655/*
1656 * cgroup_create_dir - create a directory for an object.
1657 * @cgrp: the cgroup we create the directory for. It must have a valid
1658 *        ->parent field. And we are going to fill its ->dentry field.
1659 * @dentry: dentry of the new cgroup
1660 * @mode: mode to set on new directory.
1661 */
1662static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
1663                                int mode)
1664{
1665        struct dentry *parent;
1666        int error = 0;
1667
1668        parent = cgrp->parent->dentry;
1669        error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
1670        if (!error) {
1671                dentry->d_fsdata = cgrp;
1672                inc_nlink(parent->d_inode);
1673                cgrp->dentry = dentry;
1674                dget(dentry);
1675        }
1676        dput(dentry);
1677
1678        return error;
1679}
1680
1681int cgroup_add_file(struct cgroup *cgrp,
1682                       struct cgroup_subsys *subsys,
1683                       const struct cftype *cft)
1684{
1685        struct dentry *dir = cgrp->dentry;
1686        struct dentry *dentry;
1687        int error;
1688
1689        char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
1690        if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
1691                strcpy(name, subsys->name);
1692                strcat(name, ".");
1693        }
1694        strcat(name, cft->name);
1695        BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
1696        dentry = lookup_one_len(name, dir, strlen(name));
1697        if (!IS_ERR(dentry)) {
1698                error = cgroup_create_file(dentry, 0644 | S_IFREG,
1699                                                cgrp->root->sb);
1700                if (!error)
1701                        dentry->d_fsdata = (void *)cft;
1702                dput(dentry);
1703        } else
1704                error = PTR_ERR(dentry);
1705        return error;
1706}
1707
1708int cgroup_add_files(struct cgroup *cgrp,
1709                        struct cgroup_subsys *subsys,
1710                        const struct cftype cft[],
1711                        int count)
1712{
1713        int i, err;
1714        for (i = 0; i < count; i++) {
1715                err = cgroup_add_file(cgrp, subsys, &cft[i]);
1716                if (err)
1717                        return err;
1718        }
1719        return 0;
1720}
1721
1722/**
1723 * cgroup_task_count - count the number of tasks in a cgroup.
1724 * @cgrp: the cgroup in question
1725 *
1726 * Return the number of tasks in the cgroup.
1727 */
1728int cgroup_task_count(const struct cgroup *cgrp)
1729{
1730        int count = 0;
1731        struct cg_cgroup_link *link;
1732
1733        read_lock(&css_set_lock);
1734        list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
1735                count += atomic_read(&link->cg->refcount);
1736        }
1737        read_unlock(&css_set_lock);
1738        return count;
1739}
1740
1741/*
1742 * Advance a list_head iterator.  The iterator should be positioned at
1743 * the start of a css_set
1744 */
1745static void cgroup_advance_iter(struct cgroup *cgrp,
1746                                          struct cgroup_iter *it)
1747{
1748        struct list_head *l = it->cg_link;
1749        struct cg_cgroup_link *link;
1750        struct css_set *cg;
1751
1752        /* Advance to the next non-empty css_set */
1753        do {
1754                l = l->next;
1755                if (l == &cgrp->css_sets) {
1756                        it->cg_link = NULL;
1757                        return;
1758                }
1759                link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
1760                cg = link->cg;
1761        } while (list_empty(&cg->tasks));
1762        it->cg_link = l;
1763        it->task = cg->tasks.next;
1764}
1765
1766/*
1767 * To reduce the fork() overhead for systems that are not actually
1768 * using their cgroups capability, we don't maintain the lists running
1769 * through each css_set to its tasks until we see the list actually
1770 * used - in other words after the first call to cgroup_iter_start().
1771 *
1772 * The tasklist_lock is not held here, as do_each_thread() and
1773 * while_each_thread() are protected by RCU.
1774 */
1775static void cgroup_enable_task_cg_lists(void)
1776{
1777        struct task_struct *p, *g;
1778        write_lock(&css_set_lock);
1779        use_task_css_set_links = 1;
1780        do_each_thread(g, p) {
1781                task_lock(p);
1782                /*
1783                 * We should check if the process is exiting, otherwise
1784                 * it will race with cgroup_exit() in that the list
1785                 * entry won't be deleted though the process has exited.
1786                 */
1787                if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
1788                        list_add(&p->cg_list, &p->cgroups->tasks);
1789                task_unlock(p);
1790        } while_each_thread(g, p);
1791        write_unlock(&css_set_lock);
1792}
1793
1794void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
1795{
1796        /*
1797         * The first time anyone tries to iterate across a cgroup,
1798         * we need to enable the list linking each css_set to its
1799         * tasks, and fix up all existing tasks.
1800         */
1801        if (!use_task_css_set_links)
1802                cgroup_enable_task_cg_lists();
1803
1804        read_lock(&css_set_lock);
1805        it->cg_link = &cgrp->css_sets;
1806        cgroup_advance_iter(cgrp, it);
1807}
1808
1809struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
1810                                        struct cgroup_iter *it)
1811{
1812        struct task_struct *res;
1813        struct list_head *l = it->task;
1814
1815        /* If the iterator cg is NULL, we have no tasks */
1816        if (!it->cg_link)
1817                return NULL;
1818        res = list_entry(l, struct task_struct, cg_list);
1819        /* Advance iterator to find next entry */
1820        l = l->next;
1821        if (l == &res->cgroups->tasks) {
1822                /* We reached the end of this task list - move on to
1823                 * the next cg_cgroup_link */
1824                cgroup_advance_iter(cgrp, it);
1825        } else {
1826                it->task = l;
1827        }
1828        return res;
1829}
1830
1831void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
1832{
1833        read_unlock(&css_set_lock);
1834}
1835
1836static inline int started_after_time(struct task_struct *t1,
1837                                     struct timespec *time,
1838                                     struct task_struct *t2)
1839{
1840        int start_diff = timespec_compare(&t1->start_time, time);
1841        if (start_diff > 0) {
1842                return 1;
1843        } else if (start_diff < 0) {
1844                return 0;
1845        } else {
1846                /*
1847                 * Arbitrarily, if two processes started at the same
1848                 * time, we'll say that the lower pointer value
1849                 * started first. Note that t2 may have exited by now
1850                 * so this may not be a valid pointer any longer, but
1851                 * that's fine - it still serves to distinguish
1852                 * between two tasks started (effectively) simultaneously.
1853                 */
1854                return t1 > t2;
1855        }
1856}
1857
1858/*
1859 * This function is a callback from heap_insert() and is used to order
1860 * the heap.
1861 * In this case we order the heap in descending task start time.
1862 */
1863static inline int started_after(void *p1, void *p2)
1864{
1865        struct task_struct *t1 = p1;
1866        struct task_struct *t2 = p2;
1867        return started_after_time(t1, &t2->start_time, t2);
1868}
1869
1870/**
1871 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
1872 * @scan: struct cgroup_scanner containing arguments for the scan
1873 *
1874 * Arguments include pointers to callback functions test_task() and
1875 * process_task().
1876 * Iterate through all the tasks in a cgroup, calling test_task() for each,
1877 * and if it returns true, call process_task() for it also.
1878 * The test_task pointer may be NULL, meaning always true (select all tasks).
1879 * Effectively duplicates cgroup_iter_{start,next,end}()
1880 * but does not lock css_set_lock for the call to process_task().
1881 * The struct cgroup_scanner may be embedded in any structure of the caller's
1882 * creation.
1883 * It is guaranteed that process_task() will act on every task that
1884 * is a member of the cgroup for the duration of this call. This
1885 * function may or may not call process_task() for tasks that exit
1886 * or move to a different cgroup during the call, or are forked or
1887 * move into the cgroup during the call.
1888 *
1889 * Note that test_task() may be called with locks held, and may in some
1890 * situations be called multiple times for the same task, so it should
1891 * be cheap.
1892 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
1893 * pre-allocated and will be used for heap operations (and its "gt" member will
1894 * be overwritten), else a temporary heap will be used (allocation of which
1895 * may cause this function to fail).
1896 */
1897int cgroup_scan_tasks(struct cgroup_scanner *scan)
1898{
1899        int retval, i;
1900        struct cgroup_iter it;
1901        struct task_struct *p, *dropped;
1902        /* Never dereference latest_task, since it's not refcounted */
1903        struct task_struct *latest_task = NULL;
1904        struct ptr_heap tmp_heap;
1905        struct ptr_heap *heap;
1906        struct timespec latest_time = { 0, 0 };
1907
1908        if (scan->heap) {
1909                /* The caller supplied our heap and pre-allocated its memory */
1910                heap = scan->heap;
1911                heap->gt = &started_after;
1912        } else {
1913                /* We need to allocate our own heap memory */
1914                heap = &tmp_heap;
1915                retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
1916                if (retval)
1917                        /* cannot allocate the heap */
1918                        return retval;
1919        }
1920
1921 again:
1922        /*
1923         * Scan tasks in the cgroup, using the scanner's "test_task" callback
1924         * to determine which are of interest, and using the scanner's
1925         * "process_task" callback to process any of them that need an update.
1926         * Since we don't want to hold any locks during the task updates,
1927         * gather tasks to be processed in a heap structure.
1928         * The heap is sorted by descending task start time.
1929         * If the statically-sized heap fills up, we overflow tasks that
1930         * started later, and in future iterations only consider tasks that
1931         * started after the latest task in the previous pass. This
1932         * guarantees forward progress and that we don't miss any tasks.
1933         */
1934        heap->size = 0;
1935        cgroup_iter_start(scan->cg, &it);
1936        while ((p = cgroup_iter_next(scan->cg, &it))) {
1937                /*
1938                 * Only affect tasks that qualify per the caller's callback,
1939                 * if he provided one
1940                 */
1941                if (scan->test_task && !scan->test_task(p, scan))
1942                        continue;
1943                /*
1944                 * Only process tasks that started after the last task
1945                 * we processed
1946                 */
1947                if (!started_after_time(p, &latest_time, latest_task))
1948                        continue;
1949                dropped = heap_insert(heap, p);
1950                if (dropped == NULL) {
1951                        /*
1952                         * The new task was inserted; the heap wasn't
1953                         * previously full
1954                         */
1955                        get_task_struct(p);
1956                } else if (dropped != p) {
1957                        /*
1958                         * The new task was inserted, and pushed out a
1959                         * different task
1960                         */
1961                        get_task_struct(p);
1962                        put_task_struct(dropped);
1963                }
1964                /*
1965                 * Else the new task was newer than anything already in
1966                 * the heap and wasn't inserted
1967                 */
1968        }
1969        cgroup_iter_end(scan->cg, &it);
1970
1971        if (heap->size) {
1972                for (i = 0; i < heap->size; i++) {
1973                        struct task_struct *q = heap->ptrs[i];
1974                        if (i == 0) {
1975                                latest_time = q->start_time;
1976                                latest_task = q;
1977                        }
1978                        /* Process the task per the caller's callback */
1979                        scan->process_task(q, scan);
1980                        put_task_struct(q);
1981                }
1982                /*
1983                 * If we had to process any tasks at all, scan again
1984                 * in case some of them were in the middle of forking
1985                 * children that didn't get processed.
1986                 * Not the most efficient way to do it, but it avoids
1987                 * having to take callback_mutex in the fork path
1988                 */
1989                goto again;
1990        }
1991        if (heap == &tmp_heap)
1992                heap_free(&tmp_heap);
1993        return 0;
1994}
1995
1996/*
1997 * Stuff for reading the 'tasks' file.
1998 *
1999 * Reading this file can return large amounts of data if a cgroup has
2000 * *lots* of attached tasks. So it may need several calls to read(),
2001 * but we cannot guarantee that the information we produce is correct
2002 * unless we produce it entirely atomically.
2003 *
2004 */
2005
2006/*
2007 * Load into 'pidarray' up to 'npids' of the tasks using cgroup
2008 * 'cgrp'.  Return actual number of pids loaded.  No need to
2009 * task_lock(p) when reading out p->cgroup, since we're in an RCU
2010 * read section, so the css_set can't go away, and is
2011 * immutable after creation.
2012 */
2013static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
2014{
2015        int n = 0;
2016        struct cgroup_iter it;
2017        struct task_struct *tsk;
2018        cgroup_iter_start(cgrp, &it);
2019        while ((tsk = cgroup_iter_next(cgrp, &it))) {
2020                if (unlikely(n == npids))
2021                        break;
2022                pidarray[n++] = task_pid_vnr(tsk);
2023        }
2024        cgroup_iter_end(cgrp, &it);
2025        return n;
2026}
2027
2028/**
2029 * cgroupstats_build - build and fill cgroupstats
2030 * @stats: cgroupstats to fill information into
2031 * @dentry: A dentry entry belonging to the cgroup for which stats have
2032 * been requested.
2033 *
2034 * Build and fill cgroupstats so that taskstats can export it to user
2035 * space.
2036 */
2037int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
2038{
2039        int ret = -EINVAL;
2040        struct cgroup *cgrp;
2041        struct cgroup_iter it;
2042        struct task_struct *tsk;
2043
2044        /*
2045         * Validate dentry by checking the superblock operations,
2046         * and make sure it's a directory.
2047         */
2048        if (dentry->d_sb->s_op != &cgroup_ops ||
2049            !S_ISDIR(dentry->d_inode->i_mode))
2050                 goto err;
2051
2052        ret = 0;
2053        cgrp = dentry->d_fsdata;
2054        rcu_read_lock();
2055
2056        cgroup_iter_start(cgrp, &it);
2057        while ((tsk = cgroup_iter_next(cgrp, &it))) {
2058                switch (tsk->state) {
2059                case TASK_RUNNING:
2060                        stats->nr_running++;
2061                        break;
2062                case TASK_INTERRUPTIBLE:
2063                        stats->nr_sleeping++;
2064                        break;
2065                case TASK_UNINTERRUPTIBLE:
2066                        stats->nr_uninterruptible++;
2067                        break;
2068                case TASK_STOPPED:
2069                        stats->nr_stopped++;
2070                        break;
2071                default:
2072                        if (delayacct_is_task_waiting_on_io(tsk))
2073                                stats->nr_io_wait++;
2074                        break;
2075                }
2076        }
2077        cgroup_iter_end(cgrp, &it);
2078
2079        rcu_read_unlock();
2080err:
2081        return ret;
2082}
2083
2084static int cmppid(const void *a, const void *b)
2085{
2086        return *(pid_t *)a - *(pid_t *)b;
2087}
2088
2089
2090/*
2091 * seq_file methods for the "tasks" file. The seq_file position is the
2092 * next pid to display; the seq_file iterator is a pointer to the pid
2093 * in the cgroup->tasks_pids array.
2094 */
2095
2096static void *cgroup_tasks_start(struct seq_file *s, loff_t *pos)
2097{
2098        /*
2099         * Initially we receive a position value that corresponds to
2100         * one more than the last pid shown (or 0 on the first call or
2101         * after a seek to the start). Use a binary-search to find the
2102         * next pid to display, if any
2103         */
2104        struct cgroup *cgrp = s->private;
2105        int index = 0, pid = *pos;
2106        int *iter;
2107
2108        down_read(&cgrp->pids_mutex);
2109        if (pid) {
2110                int end = cgrp->pids_length;
2111
2112                while (index < end) {
2113                        int mid = (index + end) / 2;
2114                        if (cgrp->tasks_pids[mid] == pid) {
2115                                index = mid;
2116                                break;
2117                        } else if (cgrp->tasks_pids[mid] <= pid)
2118                                index = mid + 1;
2119                        else
2120                                end = mid;
2121                }
2122        }
2123        /* If we're off the end of the array, we're done */
2124        if (index >= cgrp->pids_length)
2125                return NULL;
2126        /* Update the abstract position to be the actual pid that we found */
2127        iter = cgrp->tasks_pids + index;
2128        *pos = *iter;
2129        return iter;
2130}
2131
2132static void cgroup_tasks_stop(struct seq_file *s, void *v)
2133{
2134        struct cgroup *cgrp = s->private;
2135        up_read(&cgrp->pids_mutex);
2136}
2137
2138static void *cgroup_tasks_next(struct seq_file *s, void *v, loff_t *pos)
2139{
2140        struct cgroup *cgrp = s->private;
2141        int *p = v;
2142        int *end = cgrp->tasks_pids + cgrp->pids_length;
2143
2144        /*
2145         * Advance to the next pid in the array. If this goes off the
2146         * end, we're done
2147         */
2148        p++;
2149        if (p >= end) {
2150                return NULL;
2151        } else {
2152                *pos = *p;
2153                return p;
2154        }
2155}
2156
2157static int cgroup_tasks_show(struct seq_file *s, void *v)
2158{
2159        return seq_printf(s, "%d\n", *(int *)v);
2160}
2161
2162static struct seq_operations cgroup_tasks_seq_operations = {
2163        .start = cgroup_tasks_start,
2164        .stop = cgroup_tasks_stop,
2165        .next = cgroup_tasks_next,
2166        .show = cgroup_tasks_show,
2167};
2168
2169static void release_cgroup_pid_array(struct cgroup *cgrp)
2170{
2171        down_write(&cgrp->pids_mutex);
2172        BUG_ON(!cgrp->pids_use_count);
2173        if (!--cgrp->pids_use_count) {
2174                kfree(cgrp->tasks_pids);
2175                cgrp->tasks_pids = NULL;
2176                cgrp->pids_length = 0;
2177        }
2178        up_write(&cgrp->pids_mutex);
2179}
2180
2181static int cgroup_tasks_release(struct inode *inode, struct file *file)
2182{
2183        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2184
2185        if (!(file->f_mode & FMODE_READ))
2186                return 0;
2187
2188        release_cgroup_pid_array(cgrp);
2189        return seq_release(inode, file);
2190}
2191
2192static struct file_operations cgroup_tasks_operations = {
2193        .read = seq_read,
2194        .llseek = seq_lseek,
2195        .write = cgroup_file_write,
2196        .release = cgroup_tasks_release,
2197};
2198
2199/*
2200 * Handle an open on 'tasks' file.  Prepare an array containing the
2201 * process id's of tasks currently attached to the cgroup being opened.
2202 */
2203
2204static int cgroup_tasks_open(struct inode *unused, struct file *file)
2205{
2206        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2207        pid_t *pidarray;
2208        int npids;
2209        int retval;
2210
2211        /* Nothing to do for write-only files */
2212        if (!(file->f_mode & FMODE_READ))
2213                return 0;
2214
2215        /*
2216         * If cgroup gets more users after we read count, we won't have
2217         * enough space - tough.  This race is indistinguishable to the
2218         * caller from the case that the additional cgroup users didn't
2219         * show up until sometime later on.
2220         */
2221        npids = cgroup_task_count(cgrp);
2222        pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
2223        if (!pidarray)
2224                return -ENOMEM;
2225        npids = pid_array_load(pidarray, npids, cgrp);
2226        sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
2227
2228        /*
2229         * Store the array in the cgroup, freeing the old
2230         * array if necessary
2231         */
2232        down_write(&cgrp->pids_mutex);
2233        kfree(cgrp->tasks_pids);
2234        cgrp->tasks_pids = pidarray;
2235        cgrp->pids_length = npids;
2236        cgrp->pids_use_count++;
2237        up_write(&cgrp->pids_mutex);
2238
2239        file->f_op = &cgroup_tasks_operations;
2240
2241        retval = seq_open(file, &cgroup_tasks_seq_operations);
2242        if (retval) {
2243                release_cgroup_pid_array(cgrp);
2244                return retval;
2245        }
2246        ((struct seq_file *)file->private_data)->private = cgrp;
2247        return 0;
2248}
2249
2250static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
2251                                            struct cftype *cft)
2252{
2253        return notify_on_release(cgrp);
2254}
2255
2256static int cgroup_write_notify_on_release(struct cgroup *cgrp,
2257                                          struct cftype *cft,
2258                                          u64 val)
2259{
2260        clear_bit(CGRP_RELEASABLE, &cgrp->flags);
2261        if (val)
2262                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2263        else
2264                clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2265        return 0;
2266}
2267
2268/*
2269 * for the common functions, 'private' gives the type of file
2270 */
2271static struct cftype files[] = {
2272        {
2273                .name = "tasks",
2274                .open = cgroup_tasks_open,
2275                .write_u64 = cgroup_tasks_write,
2276                .release = cgroup_tasks_release,
2277                .private = FILE_TASKLIST,
2278        },
2279
2280        {
2281                .name = "notify_on_release",
2282                .read_u64 = cgroup_read_notify_on_release,
2283                .write_u64 = cgroup_write_notify_on_release,
2284                .private = FILE_NOTIFY_ON_RELEASE,
2285        },
2286};
2287
2288static struct cftype cft_release_agent = {
2289        .name = "release_agent",
2290        .read_seq_string = cgroup_release_agent_show,
2291        .write_string = cgroup_release_agent_write,
2292        .max_write_len = PATH_MAX,
2293        .private = FILE_RELEASE_AGENT,
2294};
2295
2296static int cgroup_populate_dir(struct cgroup *cgrp)
2297{
2298        int err;
2299        struct cgroup_subsys *ss;
2300
2301        /* First clear out any existing files */
2302        cgroup_clear_directory(cgrp->dentry);
2303
2304        err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
2305        if (err < 0)
2306                return err;
2307
2308        if (cgrp == cgrp->top_cgroup) {
2309                if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
2310                        return err;
2311        }
2312
2313        for_each_subsys(cgrp->root, ss) {
2314                if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
2315                        return err;
2316        }
2317
2318        return 0;
2319}
2320
2321static void init_cgroup_css(struct cgroup_subsys_state *css,
2322                               struct cgroup_subsys *ss,
2323                               struct cgroup *cgrp)
2324{
2325        css->cgroup = cgrp;
2326        atomic_set(&css->refcnt, 0);
2327        css->flags = 0;
2328        if (cgrp == dummytop)
2329                set_bit(CSS_ROOT, &css->flags);
2330        BUG_ON(cgrp->subsys[ss->subsys_id]);
2331        cgrp->subsys[ss->subsys_id] = css;
2332}
2333
2334/*
2335 * cgroup_create - create a cgroup
2336 * @parent: cgroup that will be parent of the new cgroup
2337 * @dentry: dentry of the new cgroup
2338 * @mode: mode to set on new inode
2339 *
2340 * Must be called with the mutex on the parent inode held
2341 */
2342static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
2343                             int mode)
2344{
2345        struct cgroup *cgrp;
2346        struct cgroupfs_root *root = parent->root;
2347        int err = 0;
2348        struct cgroup_subsys *ss;
2349        struct super_block *sb = root->sb;
2350
2351        cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
2352        if (!cgrp)
2353                return -ENOMEM;
2354
2355        /* Grab a reference on the superblock so the hierarchy doesn't
2356         * get deleted on unmount if there are child cgroups.  This
2357         * can be done outside cgroup_mutex, since the sb can't
2358         * disappear while someone has an open control file on the
2359         * fs */
2360        atomic_inc(&sb->s_active);
2361
2362        mutex_lock(&cgroup_mutex);
2363
2364        init_cgroup_housekeeping(cgrp);
2365
2366        cgrp->parent = parent;
2367        cgrp->root = parent->root;
2368        cgrp->top_cgroup = parent->top_cgroup;
2369
2370        if (notify_on_release(parent))
2371                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
2372
2373        for_each_subsys(root, ss) {
2374                struct cgroup_subsys_state *css = ss->create(ss, cgrp);
2375                if (IS_ERR(css)) {
2376                        err = PTR_ERR(css);
2377                        goto err_destroy;
2378                }
2379                init_cgroup_css(css, ss, cgrp);
2380        }
2381
2382        list_add(&cgrp->sibling, &cgrp->parent->children);
2383        root->number_of_cgroups++;
2384
2385        err = cgroup_create_dir(cgrp, dentry, mode);
2386        if (err < 0)
2387                goto err_remove;
2388
2389        /* The cgroup directory was pre-locked for us */
2390        BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
2391
2392        err = cgroup_populate_dir(cgrp);
2393        /* If err < 0, we have a half-filled directory - oh well ;) */
2394
2395        mutex_unlock(&cgroup_mutex);
2396        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
2397
2398        return 0;
2399
2400 err_remove:
2401
2402        list_del(&cgrp->sibling);
2403        root->number_of_cgroups--;
2404
2405 err_destroy:
2406
2407        for_each_subsys(root, ss) {
2408                if (cgrp->subsys[ss->subsys_id])
2409                        ss->destroy(ss, cgrp);
2410        }
2411
2412        mutex_unlock(&cgroup_mutex);
2413
2414        /* Release the reference count that we took on the superblock */
2415        deactivate_super(sb);
2416
2417        kfree(cgrp);
2418        return err;
2419}
2420
2421static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
2422{
2423        struct cgroup *c_parent = dentry->d_parent->d_fsdata;
2424
2425        /* the vfs holds inode->i_mutex already */
2426        return cgroup_create(c_parent, dentry, mode | S_IFDIR);
2427}
2428
2429static int cgroup_has_css_refs(struct cgroup *cgrp)
2430{
2431        /* Check the reference count on each subsystem. Since we
2432         * already established that there are no tasks in the
2433         * cgroup, if the css refcount is also 0, then there should
2434         * be no outstanding references, so the subsystem is safe to
2435         * destroy. We scan across all subsystems rather than using
2436         * the per-hierarchy linked list of mounted subsystems since
2437         * we can be called via check_for_release() with no
2438         * synchronization other than RCU, and the subsystem linked
2439         * list isn't RCU-safe */
2440        int i;
2441        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2442                struct cgroup_subsys *ss = subsys[i];
2443                struct cgroup_subsys_state *css;
2444                /* Skip subsystems not in this hierarchy */
2445                if (ss->root != cgrp->root)
2446                        continue;
2447                css = cgrp->subsys[ss->subsys_id];
2448                /* When called from check_for_release() it's possible
2449                 * that by this point the cgroup has been removed
2450                 * and the css deleted. But a false-positive doesn't
2451                 * matter, since it can only happen if the cgroup
2452                 * has been deleted and hence no longer needs the
2453                 * release agent to be called anyway. */
2454                if (css && atomic_read(&css->refcnt))
2455                        return 1;
2456        }
2457        return 0;
2458}
2459
2460static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
2461{
2462        struct cgroup *cgrp = dentry->d_fsdata;
2463        struct dentry *d;
2464        struct cgroup *parent;
2465        struct super_block *sb;
2466        struct cgroupfs_root *root;
2467
2468        /* the vfs holds both inode->i_mutex already */
2469
2470        mutex_lock(&cgroup_mutex);
2471        if (atomic_read(&cgrp->count) != 0) {
2472                mutex_unlock(&cgroup_mutex);
2473                return -EBUSY;
2474        }
2475        if (!list_empty(&cgrp->children)) {
2476                mutex_unlock(&cgroup_mutex);
2477                return -EBUSY;
2478        }
2479        mutex_unlock(&cgroup_mutex);
2480
2481        /*
2482         * Call pre_destroy handlers of subsys. Notify subsystems
2483         * that rmdir() request comes.
2484         */
2485        cgroup_call_pre_destroy(cgrp);
2486
2487        mutex_lock(&cgroup_mutex);
2488        parent = cgrp->parent;
2489        root = cgrp->root;
2490        sb = root->sb;
2491
2492        if (atomic_read(&cgrp->count)
2493            || !list_empty(&cgrp->children)
2494            || cgroup_has_css_refs(cgrp)) {
2495                mutex_unlock(&cgroup_mutex);
2496                return -EBUSY;
2497        }
2498
2499        spin_lock(&release_list_lock);
2500        set_bit(CGRP_REMOVED, &cgrp->flags);
2501        if (!list_empty(&cgrp->release_list))
2502                list_del(&cgrp->release_list);
2503        spin_unlock(&release_list_lock);
2504        /* delete my sibling from parent->children */
2505        list_del(&cgrp->sibling);
2506        spin_lock(&cgrp->dentry->d_lock);
2507        d = dget(cgrp->dentry);
2508        spin_unlock(&d->d_lock);
2509
2510        cgroup_d_remove_dir(d);
2511        dput(d);
2512
2513        set_bit(CGRP_RELEASABLE, &parent->flags);
2514        check_for_release(parent);
2515
2516        mutex_unlock(&cgroup_mutex);
2517        return 0;
2518}
2519
2520static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
2521{
2522        struct cgroup_subsys_state *css;
2523
2524        printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
2525
2526        /* Create the top cgroup state for this subsystem */
2527        ss->root = &rootnode;
2528        css = ss->create(ss, dummytop);
2529        /* We don't handle early failures gracefully */
2530        BUG_ON(IS_ERR(css));
2531        init_cgroup_css(css, ss, dummytop);
2532
2533        /* Update the init_css_set to contain a subsys
2534         * pointer to this state - since the subsystem is
2535         * newly registered, all tasks and hence the
2536         * init_css_set is in the subsystem's top cgroup. */
2537        init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
2538
2539        need_forkexit_callback |= ss->fork || ss->exit;
2540        need_mm_owner_callback |= !!ss->mm_owner_changed;
2541
2542        /* At system boot, before all subsystems have been
2543         * registered, no tasks have been forked, so we don't
2544         * need to invoke fork callbacks here. */
2545        BUG_ON(!list_empty(&init_task.tasks));
2546
2547        ss->active = 1;
2548}
2549
2550/**
2551 * cgroup_init_early - cgroup initialization at system boot
2552 *
2553 * Initialize cgroups at system boot, and initialize any
2554 * subsystems that request early init.
2555 */
2556int __init cgroup_init_early(void)
2557{
2558        int i;
2559        atomic_set(&init_css_set.refcount, 1);
2560        INIT_LIST_HEAD(&init_css_set.cg_links);
2561        INIT_LIST_HEAD(&init_css_set.tasks);
2562        INIT_HLIST_NODE(&init_css_set.hlist);
2563        css_set_count = 1;
2564        init_cgroup_root(&rootnode);
2565        list_add(&rootnode.root_list, &roots);
2566        root_count = 1;
2567        init_task.cgroups = &init_css_set;
2568
2569        init_css_set_link.cg = &init_css_set;
2570        list_add(&init_css_set_link.cgrp_link_list,
2571                 &rootnode.top_cgroup.css_sets);
2572        list_add(&init_css_set_link.cg_link_list,
2573                 &init_css_set.cg_links);
2574
2575        for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
2576                INIT_HLIST_HEAD(&css_set_table[i]);
2577
2578        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2579                struct cgroup_subsys *ss = subsys[i];
2580
2581                BUG_ON(!ss->name);
2582                BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
2583                BUG_ON(!ss->create);
2584                BUG_ON(!ss->destroy);
2585                if (ss->subsys_id != i) {
2586                        printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
2587                               ss->name, ss->subsys_id);
2588                        BUG();
2589                }
2590
2591                if (ss->early_init)
2592                        cgroup_init_subsys(ss);
2593        }
2594        return 0;
2595}
2596
2597/**
2598 * cgroup_init - cgroup initialization
2599 *
2600 * Register cgroup filesystem and /proc file, and initialize
2601 * any subsystems that didn't request early init.
2602 */
2603int __init cgroup_init(void)
2604{
2605        int err;
2606        int i;
2607        struct hlist_head *hhead;
2608
2609        err = bdi_init(&cgroup_backing_dev_info);
2610        if (err)
2611                return err;
2612
2613        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2614                struct cgroup_subsys *ss = subsys[i];
2615                if (!ss->early_init)
2616                        cgroup_init_subsys(ss);
2617        }
2618
2619        /* Add init_css_set to the hash table */
2620        hhead = css_set_hash(init_css_set.subsys);
2621        hlist_add_head(&init_css_set.hlist, hhead);
2622
2623        err = register_filesystem(&cgroup_fs_type);
2624        if (err < 0)
2625                goto out;
2626
2627        proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
2628
2629out:
2630        if (err)
2631                bdi_destroy(&cgroup_backing_dev_info);
2632
2633        return err;
2634}
2635
2636/*
2637 * proc_cgroup_show()
2638 *  - Print task's cgroup paths into seq_file, one line for each hierarchy
2639 *  - Used for /proc/<pid>/cgroup.
2640 *  - No need to task_lock(tsk) on this tsk->cgroup reference, as it
2641 *    doesn't really matter if tsk->cgroup changes after we read it,
2642 *    and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
2643 *    anyway.  No need to check that tsk->cgroup != NULL, thanks to
2644 *    the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
2645 *    cgroup to top_cgroup.
2646 */
2647
2648/* TODO: Use a proper seq_file iterator */
2649static int proc_cgroup_show(struct seq_file *m, void *v)
2650{
2651        struct pid *pid;
2652        struct task_struct *tsk;
2653        char *buf;
2654        int retval;
2655        struct cgroupfs_root *root;
2656
2657        retval = -ENOMEM;
2658        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2659        if (!buf)
2660                goto out;
2661
2662        retval = -ESRCH;
2663        pid = m->private;
2664        tsk = get_pid_task(pid, PIDTYPE_PID);
2665        if (!tsk)
2666                goto out_free;
2667
2668        retval = 0;
2669
2670        mutex_lock(&cgroup_mutex);
2671
2672        for_each_root(root) {
2673                struct cgroup_subsys *ss;
2674                struct cgroup *cgrp;
2675                int subsys_id;
2676                int count = 0;
2677
2678                /* Skip this hierarchy if it has no active subsystems */
2679                if (!root->actual_subsys_bits)
2680                        continue;
2681                seq_printf(m, "%lu:", root->subsys_bits);
2682                for_each_subsys(root, ss)
2683                        seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
2684                seq_putc(m, ':');
2685                get_first_subsys(&root->top_cgroup, NULL, &subsys_id);
2686                cgrp = task_cgroup(tsk, subsys_id);
2687                retval = cgroup_path(cgrp, buf, PAGE_SIZE);
2688                if (retval < 0)
2689                        goto out_unlock;
2690                seq_puts(m, buf);
2691                seq_putc(m, '\n');
2692        }
2693
2694out_unlock:
2695        mutex_unlock(&cgroup_mutex);
2696        put_task_struct(tsk);
2697out_free:
2698        kfree(buf);
2699out:
2700        return retval;
2701}
2702
2703static int cgroup_open(struct inode *inode, struct file *file)
2704{
2705        struct pid *pid = PROC_I(inode)->pid;
2706        return single_open(file, proc_cgroup_show, pid);
2707}
2708
2709struct file_operations proc_cgroup_operations = {
2710        .open           = cgroup_open,
2711        .read           = seq_read,
2712        .llseek         = seq_lseek,
2713        .release        = single_release,
2714};
2715
2716/* Display information about each subsystem and each hierarchy */
2717static int proc_cgroupstats_show(struct seq_file *m, void *v)
2718{
2719        int i;
2720
2721        seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
2722        mutex_lock(&cgroup_mutex);
2723        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2724                struct cgroup_subsys *ss = subsys[i];
2725                seq_printf(m, "%s\t%lu\t%d\t%d\n",
2726                           ss->name, ss->root->subsys_bits,
2727                           ss->root->number_of_cgroups, !ss->disabled);
2728        }
2729        mutex_unlock(&cgroup_mutex);
2730        return 0;
2731}
2732
2733static int cgroupstats_open(struct inode *inode, struct file *file)
2734{
2735        return single_open(file, proc_cgroupstats_show, NULL);
2736}
2737
2738static struct file_operations proc_cgroupstats_operations = {
2739        .open = cgroupstats_open,
2740        .read = seq_read,
2741        .llseek = seq_lseek,
2742        .release = single_release,
2743};
2744
2745/**
2746 * cgroup_fork - attach newly forked task to its parents cgroup.
2747 * @child: pointer to task_struct of forking parent process.
2748 *
2749 * Description: A task inherits its parent's cgroup at fork().
2750 *
2751 * A pointer to the shared css_set was automatically copied in
2752 * fork.c by dup_task_struct().  However, we ignore that copy, since
2753 * it was not made under the protection of RCU or cgroup_mutex, so
2754 * might no longer be a valid cgroup pointer.  cgroup_attach_task() might
2755 * have already changed current->cgroups, allowing the previously
2756 * referenced cgroup group to be removed and freed.
2757 *
2758 * At the point that cgroup_fork() is called, 'current' is the parent
2759 * task, and the passed argument 'child' points to the child task.
2760 */
2761void cgroup_fork(struct task_struct *child)
2762{
2763        task_lock(current);
2764        child->cgroups = current->cgroups;
2765        get_css_set(child->cgroups);
2766        task_unlock(current);
2767        INIT_LIST_HEAD(&child->cg_list);
2768}
2769
2770/**
2771 * cgroup_fork_callbacks - run fork callbacks
2772 * @child: the new task
2773 *
2774 * Called on a new task very soon before adding it to the
2775 * tasklist. No need to take any locks since no-one can
2776 * be operating on this task.
2777 */
2778void cgroup_fork_callbacks(struct task_struct *child)
2779{
2780        if (need_forkexit_callback) {
2781                int i;
2782                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2783                        struct cgroup_subsys *ss = subsys[i];
2784                        if (ss->fork)
2785                                ss->fork(ss, child);
2786                }
2787        }
2788}
2789
2790#ifdef CONFIG_MM_OWNER
2791/**
2792 * cgroup_mm_owner_callbacks - run callbacks when the mm->owner changes
2793 * @p: the new owner
2794 *
2795 * Called on every change to mm->owner. mm_init_owner() does not
2796 * invoke this routine, since it assigns the mm->owner the first time
2797 * and does not change it.
2798 *
2799 * The callbacks are invoked with mmap_sem held in read mode.
2800 */
2801void cgroup_mm_owner_callbacks(struct task_struct *old, struct task_struct *new)
2802{
2803        struct cgroup *oldcgrp, *newcgrp = NULL;
2804
2805        if (need_mm_owner_callback) {
2806                int i;
2807                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2808                        struct cgroup_subsys *ss = subsys[i];
2809                        oldcgrp = task_cgroup(old, ss->subsys_id);
2810                        if (new)
2811                                newcgrp = task_cgroup(new, ss->subsys_id);
2812                        if (oldcgrp == newcgrp)
2813                                continue;
2814                        if (ss->mm_owner_changed)
2815                                ss->mm_owner_changed(ss, oldcgrp, newcgrp, new);
2816                }
2817        }
2818}
2819#endif /* CONFIG_MM_OWNER */
2820
2821/**
2822 * cgroup_post_fork - called on a new task after adding it to the task list
2823 * @child: the task in question
2824 *
2825 * Adds the task to the list running through its css_set if necessary.
2826 * Has to be after the task is visible on the task list in case we race
2827 * with the first call to cgroup_iter_start() - to guarantee that the
2828 * new task ends up on its list.
2829 */
2830void cgroup_post_fork(struct task_struct *child)
2831{
2832        if (use_task_css_set_links) {
2833                write_lock(&css_set_lock);
2834                if (list_empty(&child->cg_list))
2835                        list_add(&child->cg_list, &child->cgroups->tasks);
2836                write_unlock(&css_set_lock);
2837        }
2838}
2839/**
2840 * cgroup_exit - detach cgroup from exiting task
2841 * @tsk: pointer to task_struct of exiting process
2842 * @run_callback: run exit callbacks?
2843 *
2844 * Description: Detach cgroup from @tsk and release it.
2845 *
2846 * Note that cgroups marked notify_on_release force every task in
2847 * them to take the global cgroup_mutex mutex when exiting.
2848 * This could impact scaling on very large systems.  Be reluctant to
2849 * use notify_on_release cgroups where very high task exit scaling
2850 * is required on large systems.
2851 *
2852 * the_top_cgroup_hack:
2853 *
2854 *    Set the exiting tasks cgroup to the root cgroup (top_cgroup).
2855 *
2856 *    We call cgroup_exit() while the task is still competent to
2857 *    handle notify_on_release(), then leave the task attached to the
2858 *    root cgroup in each hierarchy for the remainder of its exit.
2859 *
2860 *    To do this properly, we would increment the reference count on
2861 *    top_cgroup, and near the very end of the kernel/exit.c do_exit()
2862 *    code we would add a second cgroup function call, to drop that
2863 *    reference.  This would just create an unnecessary hot spot on
2864 *    the top_cgroup reference count, to no avail.
2865 *
2866 *    Normally, holding a reference to a cgroup without bumping its
2867 *    count is unsafe.   The cgroup could go away, or someone could
2868 *    attach us to a different cgroup, decrementing the count on
2869 *    the first cgroup that we never incremented.  But in this case,
2870 *    top_cgroup isn't going away, and either task has PF_EXITING set,
2871 *    which wards off any cgroup_attach_task() attempts, or task is a failed
2872 *    fork, never visible to cgroup_attach_task.
2873 */
2874void cgroup_exit(struct task_struct *tsk, int run_callbacks)
2875{
2876        int i;
2877        struct css_set *cg;
2878
2879        if (run_callbacks && need_forkexit_callback) {
2880                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
2881                        struct cgroup_subsys *ss = subsys[i];
2882                        if (ss->exit)
2883                                ss->exit(ss, tsk);
2884                }
2885        }
2886
2887        /*
2888         * Unlink from the css_set task list if necessary.
2889         * Optimistically check cg_list before taking
2890         * css_set_lock
2891         */
2892        if (!list_empty(&tsk->cg_list)) {
2893                write_lock(&css_set_lock);
2894                if (!list_empty(&tsk->cg_list))
2895                        list_del(&tsk->cg_list);
2896                write_unlock(&css_set_lock);
2897        }
2898
2899        /* Reassign the task to the init_css_set. */
2900        task_lock(tsk);
2901        cg = tsk->cgroups;
2902        tsk->cgroups = &init_css_set;
2903        task_unlock(tsk);
2904        if (cg)
2905                put_css_set_taskexit(cg);
2906}
2907
2908/**
2909 * cgroup_clone - clone the cgroup the given subsystem is attached to
2910 * @tsk: the task to be moved
2911 * @subsys: the given subsystem
2912 * @nodename: the name for the new cgroup
2913 *
2914 * Duplicate the current cgroup in the hierarchy that the given
2915 * subsystem is attached to, and move this task into the new
2916 * child.
2917 */
2918int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
2919                                                        char *nodename)
2920{
2921        struct dentry *dentry;
2922        int ret = 0;
2923        struct cgroup *parent, *child;
2924        struct inode *inode;
2925        struct css_set *cg;
2926        struct cgroupfs_root *root;
2927        struct cgroup_subsys *ss;
2928
2929        /* We shouldn't be called by an unregistered subsystem */
2930        BUG_ON(!subsys->active);
2931
2932        /* First figure out what hierarchy and cgroup we're dealing
2933         * with, and pin them so we can drop cgroup_mutex */
2934        mutex_lock(&cgroup_mutex);
2935 again:
2936        root = subsys->root;
2937        if (root == &rootnode) {
2938                mutex_unlock(&cgroup_mutex);
2939                return 0;
2940        }
2941        cg = tsk->cgroups;
2942        parent = task_cgroup(tsk, subsys->subsys_id);
2943
2944        /* Pin the hierarchy */
2945        if (!atomic_inc_not_zero(&parent->root->sb->s_active)) {
2946                /* We race with the final deactivate_super() */
2947                mutex_unlock(&cgroup_mutex);
2948                return 0;
2949        }
2950
2951        /* Keep the cgroup alive */
2952        get_css_set(cg);
2953        mutex_unlock(&cgroup_mutex);
2954
2955        /* Now do the VFS work to create a cgroup */
2956        inode = parent->dentry->d_inode;
2957
2958        /* Hold the parent directory mutex across this operation to
2959         * stop anyone else deleting the new cgroup */
2960        mutex_lock(&inode->i_mutex);
2961        dentry = lookup_one_len(nodename, parent->dentry, strlen(nodename));
2962        if (IS_ERR(dentry)) {
2963                printk(KERN_INFO
2964                       "cgroup: Couldn't allocate dentry for %s: %ld\n", nodename,
2965                       PTR_ERR(dentry));
2966                ret = PTR_ERR(dentry);
2967                goto out_release;
2968        }
2969
2970        /* Create the cgroup directory, which also creates the cgroup */
2971        ret = vfs_mkdir(inode, dentry, S_IFDIR | 0755);
2972        child = __d_cgrp(dentry);
2973        dput(dentry);
2974        if (ret) {
2975                printk(KERN_INFO
2976                       "Failed to create cgroup %s: %d\n", nodename,
2977                       ret);
2978                goto out_release;
2979        }
2980
2981        if (!child) {
2982                printk(KERN_INFO
2983                       "Couldn't find new cgroup %s\n", nodename);
2984                ret = -ENOMEM;
2985                goto out_release;
2986        }
2987
2988        /* The cgroup now exists. Retake cgroup_mutex and check
2989         * that we're still in the same state that we thought we
2990         * were. */
2991        mutex_lock(&cgroup_mutex);
2992        if ((root != subsys->root) ||
2993            (parent != task_cgroup(tsk, subsys->subsys_id))) {
2994                /* Aargh, we raced ... */
2995                mutex_unlock(&inode->i_mutex);
2996                put_css_set(cg);
2997
2998                deactivate_super(parent->root->sb);
2999                /* The cgroup is still accessible in the VFS, but
3000                 * we're not going to try to rmdir() it at this
3001                 * point. */
3002                printk(KERN_INFO
3003                       "Race in cgroup_clone() - leaking cgroup %s\n",
3004                       nodename);
3005                goto again;
3006        }
3007
3008        /* do any required auto-setup */
3009        for_each_subsys(root, ss) {
3010                if (ss->post_clone)
3011                        ss->post_clone(ss, child);
3012        }
3013
3014        /* All seems fine. Finish by moving the task into the new cgroup */
3015        ret = cgroup_attach_task(child, tsk);
3016        mutex_unlock(&cgroup_mutex);
3017
3018 out_release:
3019        mutex_unlock(&inode->i_mutex);
3020
3021        mutex_lock(&cgroup_mutex);
3022        put_css_set(cg);
3023        mutex_unlock(&cgroup_mutex);
3024        deactivate_super(parent->root->sb);
3025        return ret;
3026}
3027
3028/**
3029 * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
3030 * @cgrp: the cgroup in question
3031 *
3032 * See if @cgrp is a descendant of the current task's cgroup in
3033 * the appropriate hierarchy.
3034 *
3035 * If we are sending in dummytop, then presumably we are creating
3036 * the top cgroup in the subsystem.
3037 *
3038 * Called only by the ns (nsproxy) cgroup.
3039 */
3040int cgroup_is_descendant(const struct cgroup *cgrp)
3041{
3042        int ret;
3043        struct cgroup *target;
3044        int subsys_id;
3045
3046        if (cgrp == dummytop)
3047                return 1;
3048
3049        get_first_subsys(cgrp, NULL, &subsys_id);
3050        target = task_cgroup(current, subsys_id);
3051        while (cgrp != target && cgrp!= cgrp->top_cgroup)
3052                cgrp = cgrp->parent;
3053        ret = (cgrp == target);
3054        return ret;
3055}
3056
3057static void check_for_release(struct cgroup *cgrp)
3058{
3059        /* All of these checks rely on RCU to keep the cgroup
3060         * structure alive */
3061        if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
3062            && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
3063                /* Control Group is currently removeable. If it's not
3064                 * already queued for a userspace notification, queue
3065                 * it now */
3066                int need_schedule_work = 0;
3067                spin_lock(&release_list_lock);
3068                if (!cgroup_is_removed(cgrp) &&
3069                    list_empty(&cgrp->release_list)) {
3070                        list_add(&cgrp->release_list, &release_list);
3071                        need_schedule_work = 1;
3072                }
3073                spin_unlock(&release_list_lock);
3074                if (need_schedule_work)
3075                        schedule_work(&release_agent_work);
3076        }
3077}
3078
3079void __css_put(struct cgroup_subsys_state *css)
3080{
3081        struct cgroup *cgrp = css->cgroup;
3082        rcu_read_lock();
3083        if (atomic_dec_and_test(&css->refcnt) && notify_on_release(cgrp)) {
3084                set_bit(CGRP_RELEASABLE, &cgrp->flags);
3085                check_for_release(cgrp);
3086        }
3087        rcu_read_unlock();
3088}
3089
3090/*
3091 * Notify userspace when a cgroup is released, by running the
3092 * configured release agent with the name of the cgroup (path
3093 * relative to the root of cgroup file system) as the argument.
3094 *
3095 * Most likely, this user command will try to rmdir this cgroup.
3096 *
3097 * This races with the possibility that some other task will be
3098 * attached to this cgroup before it is removed, or that some other
3099 * user task will 'mkdir' a child cgroup of this cgroup.  That's ok.
3100 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
3101 * unused, and this cgroup will be reprieved from its death sentence,
3102 * to continue to serve a useful existence.  Next time it's released,
3103 * we will get notified again, if it still has 'notify_on_release' set.
3104 *
3105 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
3106 * means only wait until the task is successfully execve()'d.  The
3107 * separate release agent task is forked by call_usermodehelper(),
3108 * then control in this thread returns here, without waiting for the
3109 * release agent task.  We don't bother to wait because the caller of
3110 * this routine has no use for the exit status of the release agent
3111 * task, so no sense holding our caller up for that.
3112 */
3113static void cgroup_release_agent(struct work_struct *work)
3114{
3115        BUG_ON(work != &release_agent_work);
3116        mutex_lock(&cgroup_mutex);
3117        spin_lock(&release_list_lock);
3118        while (!list_empty(&release_list)) {
3119                char *argv[3], *envp[3];
3120                int i;
3121                char *pathbuf = NULL, *agentbuf = NULL;
3122                struct cgroup *cgrp = list_entry(release_list.next,
3123                                                    struct cgroup,
3124                                                    release_list);
3125                list_del_init(&cgrp->release_list);
3126                spin_unlock(&release_list_lock);
3127                pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
3128                if (!pathbuf)
3129                        goto continue_free;
3130                if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
3131                        goto continue_free;
3132                agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
3133                if (!agentbuf)
3134                        goto continue_free;
3135
3136                i = 0;
3137                argv[i++] = agentbuf;
3138                argv[i++] = pathbuf;
3139                argv[i] = NULL;
3140
3141                i = 0;
3142                /* minimal command environment */
3143                envp[i++] = "HOME=/";
3144                envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
3145                envp[i] = NULL;
3146
3147                /* Drop the lock while we invoke the usermode helper,
3148                 * since the exec could involve hitting disk and hence
3149                 * be a slow process */
3150                mutex_unlock(&cgroup_mutex);
3151                call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
3152                mutex_lock(&cgroup_mutex);
3153 continue_free:
3154                kfree(pathbuf);
3155                kfree(agentbuf);
3156                spin_lock(&release_list_lock);
3157        }
3158        spin_unlock(&release_list_lock);
3159        mutex_unlock(&cgroup_mutex);
3160}
3161
3162static int __init cgroup_disable(char *str)
3163{
3164        int i;
3165        char *token;
3166
3167        while ((token = strsep(&str, ",")) != NULL) {
3168                if (!*token)
3169                        continue;
3170
3171                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3172                        struct cgroup_subsys *ss = subsys[i];
3173
3174                        if (!strcmp(token, ss->name)) {
3175                                ss->disabled = 1;
3176                                printk(KERN_INFO "Disabling %s control group"
3177                                        " subsystem\n", ss->name);
3178                                break;
3179                        }
3180                }
3181        }
3182        return 1;
3183}
3184__setup("cgroup_disable=", cgroup_disable);
3185