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 *  Notifications support
   8 *  Copyright (C) 2009 Nokia Corporation
   9 *  Author: Kirill A. Shutemov
  10 *
  11 *  Copyright notices from the original cpuset code:
  12 *  --------------------------------------------------
  13 *  Copyright (C) 2003 BULL SA.
  14 *  Copyright (C) 2004-2006 Silicon Graphics, Inc.
  15 *
  16 *  Portions derived from Patrick Mochel's sysfs code.
  17 *  sysfs is Copyright (c) 2001-3 Patrick Mochel
  18 *
  19 *  2003-10-10 Written by Simon Derr.
  20 *  2003-10-22 Updates by Stephen Hemminger.
  21 *  2004 May-July Rework by Paul Jackson.
  22 *  ---------------------------------------------------
  23 *
  24 *  This file is subject to the terms and conditions of the GNU General Public
  25 *  License.  See the file COPYING in the main directory of the Linux
  26 *  distribution for more details.
  27 */
  28
  29#include <linux/cgroup.h>
  30#include <linux/cred.h>
  31#include <linux/ctype.h>
  32#include <linux/errno.h>
  33#include <linux/fs.h>
  34#include <linux/init_task.h>
  35#include <linux/kernel.h>
  36#include <linux/list.h>
  37#include <linux/mm.h>
  38#include <linux/mutex.h>
  39#include <linux/mount.h>
  40#include <linux/pagemap.h>
  41#include <linux/proc_fs.h>
  42#include <linux/rcupdate.h>
  43#include <linux/sched.h>
  44#include <linux/backing-dev.h>
  45#include <linux/seq_file.h>
  46#include <linux/slab.h>
  47#include <linux/magic.h>
  48#include <linux/spinlock.h>
  49#include <linux/string.h>
  50#include <linux/sort.h>
  51#include <linux/kmod.h>
  52#include <linux/module.h>
  53#include <linux/delayacct.h>
  54#include <linux/cgroupstats.h>
  55#include <linux/hash.h>
  56#include <linux/namei.h>
  57#include <linux/pid_namespace.h>
  58#include <linux/idr.h>
  59#include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
  60#include <linux/eventfd.h>
  61#include <linux/poll.h>
  62#include <linux/flex_array.h> /* used in cgroup_attach_proc */
  63
  64#include <linux/atomic.h>
  65
  66/*
  67 * cgroup_mutex is the master lock.  Any modification to cgroup or its
  68 * hierarchy must be performed while holding it.
  69 *
  70 * cgroup_root_mutex nests inside cgroup_mutex and should be held to modify
  71 * cgroupfs_root of any cgroup hierarchy - subsys list, flags,
  72 * release_agent_path and so on.  Modifying requires both cgroup_mutex and
  73 * cgroup_root_mutex.  Readers can acquire either of the two.  This is to
  74 * break the following locking order cycle.
  75 *
  76 *  A. cgroup_mutex -> cred_guard_mutex -> s_type->i_mutex_key -> namespace_sem
  77 *  B. namespace_sem -> cgroup_mutex
  78 *
  79 * B happens only through cgroup_show_options() and using cgroup_root_mutex
  80 * breaks it.
  81 */
  82static DEFINE_MUTEX(cgroup_mutex);
  83static DEFINE_MUTEX(cgroup_root_mutex);
  84
  85/*
  86 * Generate an array of cgroup subsystem pointers. At boot time, this is
  87 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
  88 * registered after that. The mutable section of this array is protected by
  89 * cgroup_mutex.
  90 */
  91#define SUBSYS(_x) &_x ## _subsys,
  92static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
  93#include <linux/cgroup_subsys.h>
  94};
  95
  96#define MAX_CGROUP_ROOT_NAMELEN 64
  97
  98/*
  99 * A cgroupfs_root represents the root of a cgroup hierarchy,
 100 * and may be associated with a superblock to form an active
 101 * hierarchy
 102 */
 103struct cgroupfs_root {
 104        struct super_block *sb;
 105
 106        /*
 107         * The bitmask of subsystems intended to be attached to this
 108         * hierarchy
 109         */
 110        unsigned long subsys_bits;
 111
 112        /* Unique id for this hierarchy. */
 113        int hierarchy_id;
 114
 115        /* The bitmask of subsystems currently attached to this hierarchy */
 116        unsigned long actual_subsys_bits;
 117
 118        /* A list running through the attached subsystems */
 119        struct list_head subsys_list;
 120
 121        /* The root cgroup for this hierarchy */
 122        struct cgroup top_cgroup;
 123
 124        /* Tracks how many cgroups are currently defined in hierarchy.*/
 125        int number_of_cgroups;
 126
 127        /* A list running through the active hierarchies */
 128        struct list_head root_list;
 129
 130        /* Hierarchy-specific flags */
 131        unsigned long flags;
 132
 133        /* The path to use for release notifications. */
 134        char release_agent_path[PATH_MAX];
 135
 136        /* The name for this hierarchy - may be empty */
 137        char name[MAX_CGROUP_ROOT_NAMELEN];
 138};
 139
 140/*
 141 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
 142 * subsystems that are otherwise unattached - it never has more than a
 143 * single cgroup, and all tasks are part of that cgroup.
 144 */
 145static struct cgroupfs_root rootnode;
 146
 147/*
 148 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
 149 * cgroup_subsys->use_id != 0.
 150 */
 151#define CSS_ID_MAX      (65535)
 152struct css_id {
 153        /*
 154         * The css to which this ID points. This pointer is set to valid value
 155         * after cgroup is populated. If cgroup is removed, this will be NULL.
 156         * This pointer is expected to be RCU-safe because destroy()
 157         * is called after synchronize_rcu(). But for safe use, css_is_removed()
 158         * css_tryget() should be used for avoiding race.
 159         */
 160        struct cgroup_subsys_state __rcu *css;
 161        /*
 162         * ID of this css.
 163         */
 164        unsigned short id;
 165        /*
 166         * Depth in hierarchy which this ID belongs to.
 167         */
 168        unsigned short depth;
 169        /*
 170         * ID is freed by RCU. (and lookup routine is RCU safe.)
 171         */
 172        struct rcu_head rcu_head;
 173        /*
 174         * Hierarchy of CSS ID belongs to.
 175         */
 176        unsigned short stack[0]; /* Array of Length (depth+1) */
 177};
 178
 179/*
 180 * cgroup_event represents events which userspace want to receive.
 181 */
 182struct cgroup_event {
 183        /*
 184         * Cgroup which the event belongs to.
 185         */
 186        struct cgroup *cgrp;
 187        /*
 188         * Control file which the event associated.
 189         */
 190        struct cftype *cft;
 191        /*
 192         * eventfd to signal userspace about the event.
 193         */
 194        struct eventfd_ctx *eventfd;
 195        /*
 196         * Each of these stored in a list by the cgroup.
 197         */
 198        struct list_head list;
 199        /*
 200         * All fields below needed to unregister event when
 201         * userspace closes eventfd.
 202         */
 203        poll_table pt;
 204        wait_queue_head_t *wqh;
 205        wait_queue_t wait;
 206        struct work_struct remove;
 207};
 208
 209/* The list of hierarchy roots */
 210
 211static LIST_HEAD(roots);
 212static int root_count;
 213
 214static DEFINE_IDA(hierarchy_ida);
 215static int next_hierarchy_id;
 216static DEFINE_SPINLOCK(hierarchy_id_lock);
 217
 218/* dummytop is a shorthand for the dummy hierarchy's top cgroup */
 219#define dummytop (&rootnode.top_cgroup)
 220
 221/* This flag indicates whether tasks in the fork and exit paths should
 222 * check for fork/exit handlers to call. This avoids us having to do
 223 * extra work in the fork/exit path if none of the subsystems need to
 224 * be called.
 225 */
 226static int need_forkexit_callback __read_mostly;
 227
 228#ifdef CONFIG_PROVE_LOCKING
 229int cgroup_lock_is_held(void)
 230{
 231        return lockdep_is_held(&cgroup_mutex);
 232}
 233#else /* #ifdef CONFIG_PROVE_LOCKING */
 234int cgroup_lock_is_held(void)
 235{
 236        return mutex_is_locked(&cgroup_mutex);
 237}
 238#endif /* #else #ifdef CONFIG_PROVE_LOCKING */
 239
 240EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
 241
 242/* convenient tests for these bits */
 243inline int cgroup_is_removed(const struct cgroup *cgrp)
 244{
 245        return test_bit(CGRP_REMOVED, &cgrp->flags);
 246}
 247
 248/* bits in struct cgroupfs_root flags field */
 249enum {
 250        ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
 251};
 252
 253static int cgroup_is_releasable(const struct cgroup *cgrp)
 254{
 255        const int bits =
 256                (1 << CGRP_RELEASABLE) |
 257                (1 << CGRP_NOTIFY_ON_RELEASE);
 258        return (cgrp->flags & bits) == bits;
 259}
 260
 261static int notify_on_release(const struct cgroup *cgrp)
 262{
 263        return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
 264}
 265
 266static int clone_children(const struct cgroup *cgrp)
 267{
 268        return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
 269}
 270
 271/*
 272 * for_each_subsys() allows you to iterate on each subsystem attached to
 273 * an active hierarchy
 274 */
 275#define for_each_subsys(_root, _ss) \
 276list_for_each_entry(_ss, &_root->subsys_list, sibling)
 277
 278/* for_each_active_root() allows you to iterate across the active hierarchies */
 279#define for_each_active_root(_root) \
 280list_for_each_entry(_root, &roots, root_list)
 281
 282/* the list of cgroups eligible for automatic release. Protected by
 283 * release_list_lock */
 284static LIST_HEAD(release_list);
 285static DEFINE_RAW_SPINLOCK(release_list_lock);
 286static void cgroup_release_agent(struct work_struct *work);
 287static DECLARE_WORK(release_agent_work, cgroup_release_agent);
 288static void check_for_release(struct cgroup *cgrp);
 289
 290/* Link structure for associating css_set objects with cgroups */
 291struct cg_cgroup_link {
 292        /*
 293         * List running through cg_cgroup_links associated with a
 294         * cgroup, anchored on cgroup->css_sets
 295         */
 296        struct list_head cgrp_link_list;
 297        struct cgroup *cgrp;
 298        /*
 299         * List running through cg_cgroup_links pointing at a
 300         * single css_set object, anchored on css_set->cg_links
 301         */
 302        struct list_head cg_link_list;
 303        struct css_set *cg;
 304};
 305
 306/* The default css_set - used by init and its children prior to any
 307 * hierarchies being mounted. It contains a pointer to the root state
 308 * for each subsystem. Also used to anchor the list of css_sets. Not
 309 * reference-counted, to improve performance when child cgroups
 310 * haven't been created.
 311 */
 312
 313static struct css_set init_css_set;
 314static struct cg_cgroup_link init_css_set_link;
 315
 316static int cgroup_init_idr(struct cgroup_subsys *ss,
 317                           struct cgroup_subsys_state *css);
 318
 319/* css_set_lock protects the list of css_set objects, and the
 320 * chain of tasks off each css_set.  Nests outside task->alloc_lock
 321 * due to cgroup_iter_start() */
 322static DEFINE_RWLOCK(css_set_lock);
 323static int css_set_count;
 324
 325/*
 326 * hash table for cgroup groups. This improves the performance to find
 327 * an existing css_set. This hash doesn't (currently) take into
 328 * account cgroups in empty hierarchies.
 329 */
 330#define CSS_SET_HASH_BITS       7
 331#define CSS_SET_TABLE_SIZE      (1 << CSS_SET_HASH_BITS)
 332static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
 333
 334static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
 335{
 336        int i;
 337        int index;
 338        unsigned long tmp = 0UL;
 339
 340        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
 341                tmp += (unsigned long)css[i];
 342        tmp = (tmp >> 16) ^ tmp;
 343
 344        index = hash_long(tmp, CSS_SET_HASH_BITS);
 345
 346        return &css_set_table[index];
 347}
 348
 349/* We don't maintain the lists running through each css_set to its
 350 * task until after the first call to cgroup_iter_start(). This
 351 * reduces the fork()/exit() overhead for people who have cgroups
 352 * compiled into their kernel but not actually in use */
 353static int use_task_css_set_links __read_mostly;
 354
 355static void __put_css_set(struct css_set *cg, int taskexit)
 356{
 357        struct cg_cgroup_link *link;
 358        struct cg_cgroup_link *saved_link;
 359        /*
 360         * Ensure that the refcount doesn't hit zero while any readers
 361         * can see it. Similar to atomic_dec_and_lock(), but for an
 362         * rwlock
 363         */
 364        if (atomic_add_unless(&cg->refcount, -1, 1))
 365                return;
 366        write_lock(&css_set_lock);
 367        if (!atomic_dec_and_test(&cg->refcount)) {
 368                write_unlock(&css_set_lock);
 369                return;
 370        }
 371
 372        /* This css_set is dead. unlink it and release cgroup refcounts */
 373        hlist_del(&cg->hlist);
 374        css_set_count--;
 375
 376        list_for_each_entry_safe(link, saved_link, &cg->cg_links,
 377                                 cg_link_list) {
 378                struct cgroup *cgrp = link->cgrp;
 379                list_del(&link->cg_link_list);
 380                list_del(&link->cgrp_link_list);
 381                if (atomic_dec_and_test(&cgrp->count) &&
 382                    notify_on_release(cgrp)) {
 383                        if (taskexit)
 384                                set_bit(CGRP_RELEASABLE, &cgrp->flags);
 385                        check_for_release(cgrp);
 386                }
 387
 388                kfree(link);
 389        }
 390
 391        write_unlock(&css_set_lock);
 392        kfree_rcu(cg, rcu_head);
 393}
 394
 395/*
 396 * refcounted get/put for css_set objects
 397 */
 398static inline void get_css_set(struct css_set *cg)
 399{
 400        atomic_inc(&cg->refcount);
 401}
 402
 403static inline void put_css_set(struct css_set *cg)
 404{
 405        __put_css_set(cg, 0);
 406}
 407
 408static inline void put_css_set_taskexit(struct css_set *cg)
 409{
 410        __put_css_set(cg, 1);
 411}
 412
 413/*
 414 * compare_css_sets - helper function for find_existing_css_set().
 415 * @cg: candidate css_set being tested
 416 * @old_cg: existing css_set for a task
 417 * @new_cgrp: cgroup that's being entered by the task
 418 * @template: desired set of css pointers in css_set (pre-calculated)
 419 *
 420 * Returns true if "cg" matches "old_cg" except for the hierarchy
 421 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
 422 */
 423static bool compare_css_sets(struct css_set *cg,
 424                             struct css_set *old_cg,
 425                             struct cgroup *new_cgrp,
 426                             struct cgroup_subsys_state *template[])
 427{
 428        struct list_head *l1, *l2;
 429
 430        if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
 431                /* Not all subsystems matched */
 432                return false;
 433        }
 434
 435        /*
 436         * Compare cgroup pointers in order to distinguish between
 437         * different cgroups in heirarchies with no subsystems. We
 438         * could get by with just this check alone (and skip the
 439         * memcmp above) but on most setups the memcmp check will
 440         * avoid the need for this more expensive check on almost all
 441         * candidates.
 442         */
 443
 444        l1 = &cg->cg_links;
 445        l2 = &old_cg->cg_links;
 446        while (1) {
 447                struct cg_cgroup_link *cgl1, *cgl2;
 448                struct cgroup *cg1, *cg2;
 449
 450                l1 = l1->next;
 451                l2 = l2->next;
 452                /* See if we reached the end - both lists are equal length. */
 453                if (l1 == &cg->cg_links) {
 454                        BUG_ON(l2 != &old_cg->cg_links);
 455                        break;
 456                } else {
 457                        BUG_ON(l2 == &old_cg->cg_links);
 458                }
 459                /* Locate the cgroups associated with these links. */
 460                cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
 461                cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
 462                cg1 = cgl1->cgrp;
 463                cg2 = cgl2->cgrp;
 464                /* Hierarchies should be linked in the same order. */
 465                BUG_ON(cg1->root != cg2->root);
 466
 467                /*
 468                 * If this hierarchy is the hierarchy of the cgroup
 469                 * that's changing, then we need to check that this
 470                 * css_set points to the new cgroup; if it's any other
 471                 * hierarchy, then this css_set should point to the
 472                 * same cgroup as the old css_set.
 473                 */
 474                if (cg1->root == new_cgrp->root) {
 475                        if (cg1 != new_cgrp)
 476                                return false;
 477                } else {
 478                        if (cg1 != cg2)
 479                                return false;
 480                }
 481        }
 482        return true;
 483}
 484
 485/*
 486 * find_existing_css_set() is a helper for
 487 * find_css_set(), and checks to see whether an existing
 488 * css_set is suitable.
 489 *
 490 * oldcg: the cgroup group that we're using before the cgroup
 491 * transition
 492 *
 493 * cgrp: the cgroup that we're moving into
 494 *
 495 * template: location in which to build the desired set of subsystem
 496 * state objects for the new cgroup group
 497 */
 498static struct css_set *find_existing_css_set(
 499        struct css_set *oldcg,
 500        struct cgroup *cgrp,
 501        struct cgroup_subsys_state *template[])
 502{
 503        int i;
 504        struct cgroupfs_root *root = cgrp->root;
 505        struct hlist_head *hhead;
 506        struct hlist_node *node;
 507        struct css_set *cg;
 508
 509        /*
 510         * Build the set of subsystem state objects that we want to see in the
 511         * new css_set. while subsystems can change globally, the entries here
 512         * won't change, so no need for locking.
 513         */
 514        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 515                if (root->subsys_bits & (1UL << i)) {
 516                        /* Subsystem is in this hierarchy. So we want
 517                         * the subsystem state from the new
 518                         * cgroup */
 519                        template[i] = cgrp->subsys[i];
 520                } else {
 521                        /* Subsystem is not in this hierarchy, so we
 522                         * don't want to change the subsystem state */
 523                        template[i] = oldcg->subsys[i];
 524                }
 525        }
 526
 527        hhead = css_set_hash(template);
 528        hlist_for_each_entry(cg, node, hhead, hlist) {
 529                if (!compare_css_sets(cg, oldcg, cgrp, template))
 530                        continue;
 531
 532                /* This css_set matches what we need */
 533                return cg;
 534        }
 535
 536        /* No existing cgroup group matched */
 537        return NULL;
 538}
 539
 540static void free_cg_links(struct list_head *tmp)
 541{
 542        struct cg_cgroup_link *link;
 543        struct cg_cgroup_link *saved_link;
 544
 545        list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
 546                list_del(&link->cgrp_link_list);
 547                kfree(link);
 548        }
 549}
 550
 551/*
 552 * allocate_cg_links() allocates "count" cg_cgroup_link structures
 553 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
 554 * success or a negative error
 555 */
 556static int allocate_cg_links(int count, struct list_head *tmp)
 557{
 558        struct cg_cgroup_link *link;
 559        int i;
 560        INIT_LIST_HEAD(tmp);
 561        for (i = 0; i < count; i++) {
 562                link = kmalloc(sizeof(*link), GFP_KERNEL);
 563                if (!link) {
 564                        free_cg_links(tmp);
 565                        return -ENOMEM;
 566                }
 567                list_add(&link->cgrp_link_list, tmp);
 568        }
 569        return 0;
 570}
 571
 572/**
 573 * link_css_set - a helper function to link a css_set to a cgroup
 574 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
 575 * @cg: the css_set to be linked
 576 * @cgrp: the destination cgroup
 577 */
 578static void link_css_set(struct list_head *tmp_cg_links,
 579                         struct css_set *cg, struct cgroup *cgrp)
 580{
 581        struct cg_cgroup_link *link;
 582
 583        BUG_ON(list_empty(tmp_cg_links));
 584        link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
 585                                cgrp_link_list);
 586        link->cg = cg;
 587        link->cgrp = cgrp;
 588        atomic_inc(&cgrp->count);
 589        list_move(&link->cgrp_link_list, &cgrp->css_sets);
 590        /*
 591         * Always add links to the tail of the list so that the list
 592         * is sorted by order of hierarchy creation
 593         */
 594        list_add_tail(&link->cg_link_list, &cg->cg_links);
 595}
 596
 597/*
 598 * find_css_set() takes an existing cgroup group and a
 599 * cgroup object, and returns a css_set object that's
 600 * equivalent to the old group, but with the given cgroup
 601 * substituted into the appropriate hierarchy. Must be called with
 602 * cgroup_mutex held
 603 */
 604static struct css_set *find_css_set(
 605        struct css_set *oldcg, struct cgroup *cgrp)
 606{
 607        struct css_set *res;
 608        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
 609
 610        struct list_head tmp_cg_links;
 611
 612        struct hlist_head *hhead;
 613        struct cg_cgroup_link *link;
 614
 615        /* First see if we already have a cgroup group that matches
 616         * the desired set */
 617        read_lock(&css_set_lock);
 618        res = find_existing_css_set(oldcg, cgrp, template);
 619        if (res)
 620                get_css_set(res);
 621        read_unlock(&css_set_lock);
 622
 623        if (res)
 624                return res;
 625
 626        res = kmalloc(sizeof(*res), GFP_KERNEL);
 627        if (!res)
 628                return NULL;
 629
 630        /* Allocate all the cg_cgroup_link objects that we'll need */
 631        if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
 632                kfree(res);
 633                return NULL;
 634        }
 635
 636        atomic_set(&res->refcount, 1);
 637        INIT_LIST_HEAD(&res->cg_links);
 638        INIT_LIST_HEAD(&res->tasks);
 639        INIT_HLIST_NODE(&res->hlist);
 640
 641        /* Copy the set of subsystem state objects generated in
 642         * find_existing_css_set() */
 643        memcpy(res->subsys, template, sizeof(res->subsys));
 644
 645        write_lock(&css_set_lock);
 646        /* Add reference counts and links from the new css_set. */
 647        list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
 648                struct cgroup *c = link->cgrp;
 649                if (c->root == cgrp->root)
 650                        c = cgrp;
 651                link_css_set(&tmp_cg_links, res, c);
 652        }
 653
 654        BUG_ON(!list_empty(&tmp_cg_links));
 655
 656        css_set_count++;
 657
 658        /* Add this cgroup group to the hash table */
 659        hhead = css_set_hash(res->subsys);
 660        hlist_add_head(&res->hlist, hhead);
 661
 662        write_unlock(&css_set_lock);
 663
 664        return res;
 665}
 666
 667/*
 668 * Return the cgroup for "task" from the given hierarchy. Must be
 669 * called with cgroup_mutex held.
 670 */
 671static struct cgroup *task_cgroup_from_root(struct task_struct *task,
 672                                            struct cgroupfs_root *root)
 673{
 674        struct css_set *css;
 675        struct cgroup *res = NULL;
 676
 677        BUG_ON(!mutex_is_locked(&cgroup_mutex));
 678        read_lock(&css_set_lock);
 679        /*
 680         * No need to lock the task - since we hold cgroup_mutex the
 681         * task can't change groups, so the only thing that can happen
 682         * is that it exits and its css is set back to init_css_set.
 683         */
 684        css = task->cgroups;
 685        if (css == &init_css_set) {
 686                res = &root->top_cgroup;
 687        } else {
 688                struct cg_cgroup_link *link;
 689                list_for_each_entry(link, &css->cg_links, cg_link_list) {
 690                        struct cgroup *c = link->cgrp;
 691                        if (c->root == root) {
 692                                res = c;
 693                                break;
 694                        }
 695                }
 696        }
 697        read_unlock(&css_set_lock);
 698        BUG_ON(!res);
 699        return res;
 700}
 701
 702/*
 703 * There is one global cgroup mutex. We also require taking
 704 * task_lock() when dereferencing a task's cgroup subsys pointers.
 705 * See "The task_lock() exception", at the end of this comment.
 706 *
 707 * A task must hold cgroup_mutex to modify cgroups.
 708 *
 709 * Any task can increment and decrement the count field without lock.
 710 * So in general, code holding cgroup_mutex can't rely on the count
 711 * field not changing.  However, if the count goes to zero, then only
 712 * cgroup_attach_task() can increment it again.  Because a count of zero
 713 * means that no tasks are currently attached, therefore there is no
 714 * way a task attached to that cgroup can fork (the other way to
 715 * increment the count).  So code holding cgroup_mutex can safely
 716 * assume that if the count is zero, it will stay zero. Similarly, if
 717 * a task holds cgroup_mutex on a cgroup with zero count, it
 718 * knows that the cgroup won't be removed, as cgroup_rmdir()
 719 * needs that mutex.
 720 *
 721 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
 722 * (usually) take cgroup_mutex.  These are the two most performance
 723 * critical pieces of code here.  The exception occurs on cgroup_exit(),
 724 * when a task in a notify_on_release cgroup exits.  Then cgroup_mutex
 725 * is taken, and if the cgroup count is zero, a usermode call made
 726 * to the release agent with the name of the cgroup (path relative to
 727 * the root of cgroup file system) as the argument.
 728 *
 729 * A cgroup can only be deleted if both its 'count' of using tasks
 730 * is zero, and its list of 'children' cgroups is empty.  Since all
 731 * tasks in the system use _some_ cgroup, and since there is always at
 732 * least one task in the system (init, pid == 1), therefore, top_cgroup
 733 * always has either children cgroups and/or using tasks.  So we don't
 734 * need a special hack to ensure that top_cgroup cannot be deleted.
 735 *
 736 *      The task_lock() exception
 737 *
 738 * The need for this exception arises from the action of
 739 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
 740 * another.  It does so using cgroup_mutex, however there are
 741 * several performance critical places that need to reference
 742 * task->cgroup without the expense of grabbing a system global
 743 * mutex.  Therefore except as noted below, when dereferencing or, as
 744 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
 745 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
 746 * the task_struct routinely used for such matters.
 747 *
 748 * P.S.  One more locking exception.  RCU is used to guard the
 749 * update of a tasks cgroup pointer by cgroup_attach_task()
 750 */
 751
 752/**
 753 * cgroup_lock - lock out any changes to cgroup structures
 754 *
 755 */
 756void cgroup_lock(void)
 757{
 758        mutex_lock(&cgroup_mutex);
 759}
 760EXPORT_SYMBOL_GPL(cgroup_lock);
 761
 762/**
 763 * cgroup_unlock - release lock on cgroup changes
 764 *
 765 * Undo the lock taken in a previous cgroup_lock() call.
 766 */
 767void cgroup_unlock(void)
 768{
 769        mutex_unlock(&cgroup_mutex);
 770}
 771EXPORT_SYMBOL_GPL(cgroup_unlock);
 772
 773/*
 774 * A couple of forward declarations required, due to cyclic reference loop:
 775 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
 776 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
 777 * -> cgroup_mkdir.
 778 */
 779
 780static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
 781static struct dentry *cgroup_lookup(struct inode *, struct dentry *, struct nameidata *);
 782static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
 783static int cgroup_populate_dir(struct cgroup *cgrp);
 784static const struct inode_operations cgroup_dir_inode_operations;
 785static const struct file_operations proc_cgroupstats_operations;
 786
 787static struct backing_dev_info cgroup_backing_dev_info = {
 788        .name           = "cgroup",
 789        .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 790};
 791
 792static int alloc_css_id(struct cgroup_subsys *ss,
 793                        struct cgroup *parent, struct cgroup *child);
 794
 795static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
 796{
 797        struct inode *inode = new_inode(sb);
 798
 799        if (inode) {
 800                inode->i_ino = get_next_ino();
 801                inode->i_mode = mode;
 802                inode->i_uid = current_fsuid();
 803                inode->i_gid = current_fsgid();
 804                inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 805                inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
 806        }
 807        return inode;
 808}
 809
 810/*
 811 * Call subsys's pre_destroy handler.
 812 * This is called before css refcnt check.
 813 */
 814static int cgroup_call_pre_destroy(struct cgroup *cgrp)
 815{
 816        struct cgroup_subsys *ss;
 817        int ret = 0;
 818
 819        for_each_subsys(cgrp->root, ss)
 820                if (ss->pre_destroy) {
 821                        ret = ss->pre_destroy(ss, cgrp);
 822                        if (ret)
 823                                break;
 824                }
 825
 826        return ret;
 827}
 828
 829static void cgroup_diput(struct dentry *dentry, struct inode *inode)
 830{
 831        /* is dentry a directory ? if so, kfree() associated cgroup */
 832        if (S_ISDIR(inode->i_mode)) {
 833                struct cgroup *cgrp = dentry->d_fsdata;
 834                struct cgroup_subsys *ss;
 835                BUG_ON(!(cgroup_is_removed(cgrp)));
 836                /* It's possible for external users to be holding css
 837                 * reference counts on a cgroup; css_put() needs to
 838                 * be able to access the cgroup after decrementing
 839                 * the reference count in order to know if it needs to
 840                 * queue the cgroup to be handled by the release
 841                 * agent */
 842                synchronize_rcu();
 843
 844                mutex_lock(&cgroup_mutex);
 845                /*
 846                 * Release the subsystem state objects.
 847                 */
 848                for_each_subsys(cgrp->root, ss)
 849                        ss->destroy(ss, cgrp);
 850
 851                cgrp->root->number_of_cgroups--;
 852                mutex_unlock(&cgroup_mutex);
 853
 854                /*
 855                 * Drop the active superblock reference that we took when we
 856                 * created the cgroup
 857                 */
 858                deactivate_super(cgrp->root->sb);
 859
 860                /*
 861                 * if we're getting rid of the cgroup, refcount should ensure
 862                 * that there are no pidlists left.
 863                 */
 864                BUG_ON(!list_empty(&cgrp->pidlists));
 865
 866                kfree_rcu(cgrp, rcu_head);
 867        }
 868        iput(inode);
 869}
 870
 871static int cgroup_delete(const struct dentry *d)
 872{
 873        return 1;
 874}
 875
 876static void remove_dir(struct dentry *d)
 877{
 878        struct dentry *parent = dget(d->d_parent);
 879
 880        d_delete(d);
 881        simple_rmdir(parent->d_inode, d);
 882        dput(parent);
 883}
 884
 885static void cgroup_clear_directory(struct dentry *dentry)
 886{
 887        struct list_head *node;
 888
 889        BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
 890        spin_lock(&dentry->d_lock);
 891        node = dentry->d_subdirs.next;
 892        while (node != &dentry->d_subdirs) {
 893                struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
 894
 895                spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
 896                list_del_init(node);
 897                if (d->d_inode) {
 898                        /* This should never be called on a cgroup
 899                         * directory with child cgroups */
 900                        BUG_ON(d->d_inode->i_mode & S_IFDIR);
 901                        dget_dlock(d);
 902                        spin_unlock(&d->d_lock);
 903                        spin_unlock(&dentry->d_lock);
 904                        d_delete(d);
 905                        simple_unlink(dentry->d_inode, d);
 906                        dput(d);
 907                        spin_lock(&dentry->d_lock);
 908                } else
 909                        spin_unlock(&d->d_lock);
 910                node = dentry->d_subdirs.next;
 911        }
 912        spin_unlock(&dentry->d_lock);
 913}
 914
 915/*
 916 * NOTE : the dentry must have been dget()'ed
 917 */
 918static void cgroup_d_remove_dir(struct dentry *dentry)
 919{
 920        struct dentry *parent;
 921
 922        cgroup_clear_directory(dentry);
 923
 924        parent = dentry->d_parent;
 925        spin_lock(&parent->d_lock);
 926        spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
 927        list_del_init(&dentry->d_u.d_child);
 928        spin_unlock(&dentry->d_lock);
 929        spin_unlock(&parent->d_lock);
 930        remove_dir(dentry);
 931}
 932
 933/*
 934 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
 935 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
 936 * reference to css->refcnt. In general, this refcnt is expected to goes down
 937 * to zero, soon.
 938 *
 939 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
 940 */
 941static DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
 942
 943static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
 944{
 945        if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
 946                wake_up_all(&cgroup_rmdir_waitq);
 947}
 948
 949void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
 950{
 951        css_get(css);
 952}
 953
 954void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
 955{
 956        cgroup_wakeup_rmdir_waiter(css->cgroup);
 957        css_put(css);
 958}
 959
 960/*
 961 * Call with cgroup_mutex held. Drops reference counts on modules, including
 962 * any duplicate ones that parse_cgroupfs_options took. If this function
 963 * returns an error, no reference counts are touched.
 964 */
 965static int rebind_subsystems(struct cgroupfs_root *root,
 966                              unsigned long final_bits)
 967{
 968        unsigned long added_bits, removed_bits;
 969        struct cgroup *cgrp = &root->top_cgroup;
 970        int i;
 971
 972        BUG_ON(!mutex_is_locked(&cgroup_mutex));
 973        BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
 974
 975        removed_bits = root->actual_subsys_bits & ~final_bits;
 976        added_bits = final_bits & ~root->actual_subsys_bits;
 977        /* Check that any added subsystems are currently free */
 978        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 979                unsigned long bit = 1UL << i;
 980                struct cgroup_subsys *ss = subsys[i];
 981                if (!(bit & added_bits))
 982                        continue;
 983                /*
 984                 * Nobody should tell us to do a subsys that doesn't exist:
 985                 * parse_cgroupfs_options should catch that case and refcounts
 986                 * ensure that subsystems won't disappear once selected.
 987                 */
 988                BUG_ON(ss == NULL);
 989                if (ss->root != &rootnode) {
 990                        /* Subsystem isn't free */
 991                        return -EBUSY;
 992                }
 993        }
 994
 995        /* Currently we don't handle adding/removing subsystems when
 996         * any child cgroups exist. This is theoretically supportable
 997         * but involves complex error handling, so it's being left until
 998         * later */
 999        if (root->number_of_cgroups > 1)
1000                return -EBUSY;
1001
1002        /* Process each subsystem */
1003        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1004                struct cgroup_subsys *ss = subsys[i];
1005                unsigned long bit = 1UL << i;
1006                if (bit & added_bits) {
1007                        /* We're binding this subsystem to this hierarchy */
1008                        BUG_ON(ss == NULL);
1009                        BUG_ON(cgrp->subsys[i]);
1010                        BUG_ON(!dummytop->subsys[i]);
1011                        BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
1012                        mutex_lock(&ss->hierarchy_mutex);
1013                        cgrp->subsys[i] = dummytop->subsys[i];
1014                        cgrp->subsys[i]->cgroup = cgrp;
1015                        list_move(&ss->sibling, &root->subsys_list);
1016                        ss->root = root;
1017                        if (ss->bind)
1018                                ss->bind(ss, cgrp);
1019                        mutex_unlock(&ss->hierarchy_mutex);
1020                        /* refcount was already taken, and we're keeping it */
1021                } else if (bit & removed_bits) {
1022                        /* We're removing this subsystem */
1023                        BUG_ON(ss == NULL);
1024                        BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1025                        BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1026                        mutex_lock(&ss->hierarchy_mutex);
1027                        if (ss->bind)
1028                                ss->bind(ss, dummytop);
1029                        dummytop->subsys[i]->cgroup = dummytop;
1030                        cgrp->subsys[i] = NULL;
1031                        subsys[i]->root = &rootnode;
1032                        list_move(&ss->sibling, &rootnode.subsys_list);
1033                        mutex_unlock(&ss->hierarchy_mutex);
1034                        /* subsystem is now free - drop reference on module */
1035                        module_put(ss->module);
1036                } else if (bit & final_bits) {
1037                        /* Subsystem state should already exist */
1038                        BUG_ON(ss == NULL);
1039                        BUG_ON(!cgrp->subsys[i]);
1040                        /*
1041                         * a refcount was taken, but we already had one, so
1042                         * drop the extra reference.
1043                         */
1044                        module_put(ss->module);
1045#ifdef CONFIG_MODULE_UNLOAD
1046                        BUG_ON(ss->module && !module_refcount(ss->module));
1047#endif
1048                } else {
1049                        /* Subsystem state shouldn't exist */
1050                        BUG_ON(cgrp->subsys[i]);
1051                }
1052        }
1053        root->subsys_bits = root->actual_subsys_bits = final_bits;
1054        synchronize_rcu();
1055
1056        return 0;
1057}
1058
1059static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1060{
1061        struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1062        struct cgroup_subsys *ss;
1063
1064        mutex_lock(&cgroup_root_mutex);
1065        for_each_subsys(root, ss)
1066                seq_printf(seq, ",%s", ss->name);
1067        if (test_bit(ROOT_NOPREFIX, &root->flags))
1068                seq_puts(seq, ",noprefix");
1069        if (strlen(root->release_agent_path))
1070                seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1071        if (clone_children(&root->top_cgroup))
1072                seq_puts(seq, ",clone_children");
1073        if (strlen(root->name))
1074                seq_printf(seq, ",name=%s", root->name);
1075        mutex_unlock(&cgroup_root_mutex);
1076        return 0;
1077}
1078
1079struct cgroup_sb_opts {
1080        unsigned long subsys_bits;
1081        unsigned long flags;
1082        char *release_agent;
1083        bool clone_children;
1084        char *name;
1085        /* User explicitly requested empty subsystem */
1086        bool none;
1087
1088        struct cgroupfs_root *new_root;
1089
1090};
1091
1092/*
1093 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1094 * with cgroup_mutex held to protect the subsys[] array. This function takes
1095 * refcounts on subsystems to be used, unless it returns error, in which case
1096 * no refcounts are taken.
1097 */
1098static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1099{
1100        char *token, *o = data;
1101        bool all_ss = false, one_ss = false;
1102        unsigned long mask = (unsigned long)-1;
1103        int i;
1104        bool module_pin_failed = false;
1105
1106        BUG_ON(!mutex_is_locked(&cgroup_mutex));
1107
1108#ifdef CONFIG_CPUSETS
1109        mask = ~(1UL << cpuset_subsys_id);
1110#endif
1111
1112        memset(opts, 0, sizeof(*opts));
1113
1114        while ((token = strsep(&o, ",")) != NULL) {
1115                if (!*token)
1116                        return -EINVAL;
1117                if (!strcmp(token, "none")) {
1118                        /* Explicitly have no subsystems */
1119                        opts->none = true;
1120                        continue;
1121                }
1122                if (!strcmp(token, "all")) {
1123                        /* Mutually exclusive option 'all' + subsystem name */
1124                        if (one_ss)
1125                                return -EINVAL;
1126                        all_ss = true;
1127                        continue;
1128                }
1129                if (!strcmp(token, "noprefix")) {
1130                        set_bit(ROOT_NOPREFIX, &opts->flags);
1131                        continue;
1132                }
1133                if (!strcmp(token, "clone_children")) {
1134                        opts->clone_children = true;
1135                        continue;
1136                }
1137                if (!strncmp(token, "release_agent=", 14)) {
1138                        /* Specifying two release agents is forbidden */
1139                        if (opts->release_agent)
1140                                return -EINVAL;
1141                        opts->release_agent =
1142                                kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1143                        if (!opts->release_agent)
1144                                return -ENOMEM;
1145                        continue;
1146                }
1147                if (!strncmp(token, "name=", 5)) {
1148                        const char *name = token + 5;
1149                        /* Can't specify an empty name */
1150                        if (!strlen(name))
1151                                return -EINVAL;
1152                        /* Must match [\w.-]+ */
1153                        for (i = 0; i < strlen(name); i++) {
1154                                char c = name[i];
1155                                if (isalnum(c))
1156                                        continue;
1157                                if ((c == '.') || (c == '-') || (c == '_'))
1158                                        continue;
1159                                return -EINVAL;
1160                        }
1161                        /* Specifying two names is forbidden */
1162                        if (opts->name)
1163                                return -EINVAL;
1164                        opts->name = kstrndup(name,
1165                                              MAX_CGROUP_ROOT_NAMELEN - 1,
1166                                              GFP_KERNEL);
1167                        if (!opts->name)
1168                                return -ENOMEM;
1169
1170                        continue;
1171                }
1172
1173                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1174                        struct cgroup_subsys *ss = subsys[i];
1175                        if (ss == NULL)
1176                                continue;
1177                        if (strcmp(token, ss->name))
1178                                continue;
1179                        if (ss->disabled)
1180                                continue;
1181
1182                        /* Mutually exclusive option 'all' + subsystem name */
1183                        if (all_ss)
1184                                return -EINVAL;
1185                        set_bit(i, &opts->subsys_bits);
1186                        one_ss = true;
1187
1188                        break;
1189                }
1190                if (i == CGROUP_SUBSYS_COUNT)
1191                        return -ENOENT;
1192        }
1193
1194        /*
1195         * If the 'all' option was specified select all the subsystems,
1196         * otherwise if 'none', 'name=' and a subsystem name options
1197         * were not specified, let's default to 'all'
1198         */
1199        if (all_ss || (!one_ss && !opts->none && !opts->name)) {
1200                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1201                        struct cgroup_subsys *ss = subsys[i];
1202                        if (ss == NULL)
1203                                continue;
1204                        if (ss->disabled)
1205                                continue;
1206                        set_bit(i, &opts->subsys_bits);
1207                }
1208        }
1209
1210        /* Consistency checks */
1211
1212        /*
1213         * Option noprefix was introduced just for backward compatibility
1214         * with the old cpuset, so we allow noprefix only if mounting just
1215         * the cpuset subsystem.
1216         */
1217        if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1218            (opts->subsys_bits & mask))
1219                return -EINVAL;
1220
1221
1222        /* Can't specify "none" and some subsystems */
1223        if (opts->subsys_bits && opts->none)
1224                return -EINVAL;
1225
1226        /*
1227         * We either have to specify by name or by subsystems. (So all
1228         * empty hierarchies must have a name).
1229         */
1230        if (!opts->subsys_bits && !opts->name)
1231                return -EINVAL;
1232
1233        /*
1234         * Grab references on all the modules we'll need, so the subsystems
1235         * don't dance around before rebind_subsystems attaches them. This may
1236         * take duplicate reference counts on a subsystem that's already used,
1237         * but rebind_subsystems handles this case.
1238         */
1239        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1240                unsigned long bit = 1UL << i;
1241
1242                if (!(bit & opts->subsys_bits))
1243                        continue;
1244                if (!try_module_get(subsys[i]->module)) {
1245                        module_pin_failed = true;
1246                        break;
1247                }
1248        }
1249        if (module_pin_failed) {
1250                /*
1251                 * oops, one of the modules was going away. this means that we
1252                 * raced with a module_delete call, and to the user this is
1253                 * essentially a "subsystem doesn't exist" case.
1254                 */
1255                for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1256                        /* drop refcounts only on the ones we took */
1257                        unsigned long bit = 1UL << i;
1258
1259                        if (!(bit & opts->subsys_bits))
1260                                continue;
1261                        module_put(subsys[i]->module);
1262                }
1263                return -ENOENT;
1264        }
1265
1266        return 0;
1267}
1268
1269static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1270{
1271        int i;
1272        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1273                unsigned long bit = 1UL << i;
1274
1275                if (!(bit & subsys_bits))
1276                        continue;
1277                module_put(subsys[i]->module);
1278        }
1279}
1280
1281static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1282{
1283        int ret = 0;
1284        struct cgroupfs_root *root = sb->s_fs_info;
1285        struct cgroup *cgrp = &root->top_cgroup;
1286        struct cgroup_sb_opts opts;
1287
1288        mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1289        mutex_lock(&cgroup_mutex);
1290        mutex_lock(&cgroup_root_mutex);
1291
1292        /* See what subsystems are wanted */
1293        ret = parse_cgroupfs_options(data, &opts);
1294        if (ret)
1295                goto out_unlock;
1296
1297        /* Don't allow flags or name to change at remount */
1298        if (opts.flags != root->flags ||
1299            (opts.name && strcmp(opts.name, root->name))) {
1300                ret = -EINVAL;
1301                drop_parsed_module_refcounts(opts.subsys_bits);
1302                goto out_unlock;
1303        }
1304
1305        ret = rebind_subsystems(root, opts.subsys_bits);
1306        if (ret) {
1307                drop_parsed_module_refcounts(opts.subsys_bits);
1308                goto out_unlock;
1309        }
1310
1311        /* (re)populate subsystem files */
1312        cgroup_populate_dir(cgrp);
1313
1314        if (opts.release_agent)
1315                strcpy(root->release_agent_path, opts.release_agent);
1316 out_unlock:
1317        kfree(opts.release_agent);
1318        kfree(opts.name);
1319        mutex_unlock(&cgroup_root_mutex);
1320        mutex_unlock(&cgroup_mutex);
1321        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1322        return ret;
1323}
1324
1325static const struct super_operations cgroup_ops = {
1326        .statfs = simple_statfs,
1327        .drop_inode = generic_delete_inode,
1328        .show_options = cgroup_show_options,
1329        .remount_fs = cgroup_remount,
1330};
1331
1332static void init_cgroup_housekeeping(struct cgroup *cgrp)
1333{
1334        INIT_LIST_HEAD(&cgrp->sibling);
1335        INIT_LIST_HEAD(&cgrp->children);
1336        INIT_LIST_HEAD(&cgrp->css_sets);
1337        INIT_LIST_HEAD(&cgrp->release_list);
1338        INIT_LIST_HEAD(&cgrp->pidlists);
1339        mutex_init(&cgrp->pidlist_mutex);
1340        INIT_LIST_HEAD(&cgrp->event_list);
1341        spin_lock_init(&cgrp->event_list_lock);
1342}
1343
1344static void init_cgroup_root(struct cgroupfs_root *root)
1345{
1346        struct cgroup *cgrp = &root->top_cgroup;
1347        INIT_LIST_HEAD(&root->subsys_list);
1348        INIT_LIST_HEAD(&root->root_list);
1349        root->number_of_cgroups = 1;
1350        cgrp->root = root;
1351        cgrp->top_cgroup = cgrp;
1352        init_cgroup_housekeeping(cgrp);
1353}
1354
1355static bool init_root_id(struct cgroupfs_root *root)
1356{
1357        int ret = 0;
1358
1359        do {
1360                if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1361                        return false;
1362                spin_lock(&hierarchy_id_lock);
1363                /* Try to allocate the next unused ID */
1364                ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1365                                        &root->hierarchy_id);
1366                if (ret == -ENOSPC)
1367                        /* Try again starting from 0 */
1368                        ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1369                if (!ret) {
1370                        next_hierarchy_id = root->hierarchy_id + 1;
1371                } else if (ret != -EAGAIN) {
1372                        /* Can only get here if the 31-bit IDR is full ... */
1373                        BUG_ON(ret);
1374                }
1375                spin_unlock(&hierarchy_id_lock);
1376        } while (ret);
1377        return true;
1378}
1379
1380static int cgroup_test_super(struct super_block *sb, void *data)
1381{
1382        struct cgroup_sb_opts *opts = data;
1383        struct cgroupfs_root *root = sb->s_fs_info;
1384
1385        /* If we asked for a name then it must match */
1386        if (opts->name && strcmp(opts->name, root->name))
1387                return 0;
1388
1389        /*
1390         * If we asked for subsystems (or explicitly for no
1391         * subsystems) then they must match
1392         */
1393        if ((opts->subsys_bits || opts->none)
1394            && (opts->subsys_bits != root->subsys_bits))
1395                return 0;
1396
1397        return 1;
1398}
1399
1400static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1401{
1402        struct cgroupfs_root *root;
1403
1404        if (!opts->subsys_bits && !opts->none)
1405                return NULL;
1406
1407        root = kzalloc(sizeof(*root), GFP_KERNEL);
1408        if (!root)
1409                return ERR_PTR(-ENOMEM);
1410
1411        if (!init_root_id(root)) {
1412                kfree(root);
1413                return ERR_PTR(-ENOMEM);
1414        }
1415        init_cgroup_root(root);
1416
1417        root->subsys_bits = opts->subsys_bits;
1418        root->flags = opts->flags;
1419        if (opts->release_agent)
1420                strcpy(root->release_agent_path, opts->release_agent);
1421        if (opts->name)
1422                strcpy(root->name, opts->name);
1423        if (opts->clone_children)
1424                set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1425        return root;
1426}
1427
1428static void cgroup_drop_root(struct cgroupfs_root *root)
1429{
1430        if (!root)
1431                return;
1432
1433        BUG_ON(!root->hierarchy_id);
1434        spin_lock(&hierarchy_id_lock);
1435        ida_remove(&hierarchy_ida, root->hierarchy_id);
1436        spin_unlock(&hierarchy_id_lock);
1437        kfree(root);
1438}
1439
1440static int cgroup_set_super(struct super_block *sb, void *data)
1441{
1442        int ret;
1443        struct cgroup_sb_opts *opts = data;
1444
1445        /* If we don't have a new root, we can't set up a new sb */
1446        if (!opts->new_root)
1447                return -EINVAL;
1448
1449        BUG_ON(!opts->subsys_bits && !opts->none);
1450
1451        ret = set_anon_super(sb, NULL);
1452        if (ret)
1453                return ret;
1454
1455        sb->s_fs_info = opts->new_root;
1456        opts->new_root->sb = sb;
1457
1458        sb->s_blocksize = PAGE_CACHE_SIZE;
1459        sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1460        sb->s_magic = CGROUP_SUPER_MAGIC;
1461        sb->s_op = &cgroup_ops;
1462
1463        return 0;
1464}
1465
1466static int cgroup_get_rootdir(struct super_block *sb)
1467{
1468        static const struct dentry_operations cgroup_dops = {
1469                .d_iput = cgroup_diput,
1470                .d_delete = cgroup_delete,
1471        };
1472
1473        struct inode *inode =
1474                cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1475        struct dentry *dentry;
1476
1477        if (!inode)
1478                return -ENOMEM;
1479
1480        inode->i_fop = &simple_dir_operations;
1481        inode->i_op = &cgroup_dir_inode_operations;
1482        /* directories start off with i_nlink == 2 (for "." entry) */
1483        inc_nlink(inode);
1484        dentry = d_alloc_root(inode);
1485        if (!dentry) {
1486                iput(inode);
1487                return -ENOMEM;
1488        }
1489        sb->s_root = dentry;
1490        /* for everything else we want ->d_op set */
1491        sb->s_d_op = &cgroup_dops;
1492        return 0;
1493}
1494
1495static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1496                         int flags, const char *unused_dev_name,
1497                         void *data)
1498{
1499        struct cgroup_sb_opts opts;
1500        struct cgroupfs_root *root;
1501        int ret = 0;
1502        struct super_block *sb;
1503        struct cgroupfs_root *new_root;
1504        struct inode *inode;
1505
1506        /* First find the desired set of subsystems */
1507        mutex_lock(&cgroup_mutex);
1508        ret = parse_cgroupfs_options(data, &opts);
1509        mutex_unlock(&cgroup_mutex);
1510        if (ret)
1511                goto out_err;
1512
1513        /*
1514         * Allocate a new cgroup root. We may not need it if we're
1515         * reusing an existing hierarchy.
1516         */
1517        new_root = cgroup_root_from_opts(&opts);
1518        if (IS_ERR(new_root)) {
1519                ret = PTR_ERR(new_root);
1520                goto drop_modules;
1521        }
1522        opts.new_root = new_root;
1523
1524        /* Locate an existing or new sb for this hierarchy */
1525        sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1526        if (IS_ERR(sb)) {
1527                ret = PTR_ERR(sb);
1528                cgroup_drop_root(opts.new_root);
1529                goto drop_modules;
1530        }
1531
1532        root = sb->s_fs_info;
1533        BUG_ON(!root);
1534        if (root == opts.new_root) {
1535                /* We used the new root structure, so this is a new hierarchy */
1536                struct list_head tmp_cg_links;
1537                struct cgroup *root_cgrp = &root->top_cgroup;
1538                struct cgroupfs_root *existing_root;
1539                const struct cred *cred;
1540                int i;
1541
1542                BUG_ON(sb->s_root != NULL);
1543
1544                ret = cgroup_get_rootdir(sb);
1545                if (ret)
1546                        goto drop_new_super;
1547                inode = sb->s_root->d_inode;
1548
1549                mutex_lock(&inode->i_mutex);
1550                mutex_lock(&cgroup_mutex);
1551                mutex_lock(&cgroup_root_mutex);
1552
1553                /* Check for name clashes with existing mounts */
1554                ret = -EBUSY;
1555                if (strlen(root->name))
1556                        for_each_active_root(existing_root)
1557                                if (!strcmp(existing_root->name, root->name))
1558                                        goto unlock_drop;
1559
1560                /*
1561                 * We're accessing css_set_count without locking
1562                 * css_set_lock here, but that's OK - it can only be
1563                 * increased by someone holding cgroup_lock, and
1564                 * that's us. The worst that can happen is that we
1565                 * have some link structures left over
1566                 */
1567                ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1568                if (ret)
1569                        goto unlock_drop;
1570
1571                ret = rebind_subsystems(root, root->subsys_bits);
1572                if (ret == -EBUSY) {
1573                        free_cg_links(&tmp_cg_links);
1574                        goto unlock_drop;
1575                }
1576                /*
1577                 * There must be no failure case after here, since rebinding
1578                 * takes care of subsystems' refcounts, which are explicitly
1579                 * dropped in the failure exit path.
1580                 */
1581
1582                /* EBUSY should be the only error here */
1583                BUG_ON(ret);
1584
1585                list_add(&root->root_list, &roots);
1586                root_count++;
1587
1588                sb->s_root->d_fsdata = root_cgrp;
1589                root->top_cgroup.dentry = sb->s_root;
1590
1591                /* Link the top cgroup in this hierarchy into all
1592                 * the css_set objects */
1593                write_lock(&css_set_lock);
1594                for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1595                        struct hlist_head *hhead = &css_set_table[i];
1596                        struct hlist_node *node;
1597                        struct css_set *cg;
1598
1599                        hlist_for_each_entry(cg, node, hhead, hlist)
1600                                link_css_set(&tmp_cg_links, cg, root_cgrp);
1601                }
1602                write_unlock(&css_set_lock);
1603
1604                free_cg_links(&tmp_cg_links);
1605
1606                BUG_ON(!list_empty(&root_cgrp->sibling));
1607                BUG_ON(!list_empty(&root_cgrp->children));
1608                BUG_ON(root->number_of_cgroups != 1);
1609
1610                cred = override_creds(&init_cred);
1611                cgroup_populate_dir(root_cgrp);
1612                revert_creds(cred);
1613                mutex_unlock(&cgroup_root_mutex);
1614                mutex_unlock(&cgroup_mutex);
1615                mutex_unlock(&inode->i_mutex);
1616        } else {
1617                /*
1618                 * We re-used an existing hierarchy - the new root (if
1619                 * any) is not needed
1620                 */
1621                cgroup_drop_root(opts.new_root);
1622                /* no subsys rebinding, so refcounts don't change */
1623                drop_parsed_module_refcounts(opts.subsys_bits);
1624        }
1625
1626        kfree(opts.release_agent);
1627        kfree(opts.name);
1628        return dget(sb->s_root);
1629
1630 unlock_drop:
1631        mutex_unlock(&cgroup_root_mutex);
1632        mutex_unlock(&cgroup_mutex);
1633        mutex_unlock(&inode->i_mutex);
1634 drop_new_super:
1635        deactivate_locked_super(sb);
1636 drop_modules:
1637        drop_parsed_module_refcounts(opts.subsys_bits);
1638 out_err:
1639        kfree(opts.release_agent);
1640        kfree(opts.name);
1641        return ERR_PTR(ret);
1642}
1643
1644static void cgroup_kill_sb(struct super_block *sb) {
1645        struct cgroupfs_root *root = sb->s_fs_info;
1646        struct cgroup *cgrp = &root->top_cgroup;
1647        int ret;
1648        struct cg_cgroup_link *link;
1649        struct cg_cgroup_link *saved_link;
1650
1651        BUG_ON(!root);
1652
1653        BUG_ON(root->number_of_cgroups != 1);
1654        BUG_ON(!list_empty(&cgrp->children));
1655        BUG_ON(!list_empty(&cgrp->sibling));
1656
1657        mutex_lock(&cgroup_mutex);
1658        mutex_lock(&cgroup_root_mutex);
1659
1660        /* Rebind all subsystems back to the default hierarchy */
1661        ret = rebind_subsystems(root, 0);
1662        /* Shouldn't be able to fail ... */
1663        BUG_ON(ret);
1664
1665        /*
1666         * Release all the links from css_sets to this hierarchy's
1667         * root cgroup
1668         */
1669        write_lock(&css_set_lock);
1670
1671        list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1672                                 cgrp_link_list) {
1673                list_del(&link->cg_link_list);
1674                list_del(&link->cgrp_link_list);
1675                kfree(link);
1676        }
1677        write_unlock(&css_set_lock);
1678
1679        if (!list_empty(&root->root_list)) {
1680                list_del(&root->root_list);
1681                root_count--;
1682        }
1683
1684        mutex_unlock(&cgroup_root_mutex);
1685        mutex_unlock(&cgroup_mutex);
1686
1687        kill_litter_super(sb);
1688        cgroup_drop_root(root);
1689}
1690
1691static struct file_system_type cgroup_fs_type = {
1692        .name = "cgroup",
1693        .mount = cgroup_mount,
1694        .kill_sb = cgroup_kill_sb,
1695};
1696
1697static struct kobject *cgroup_kobj;
1698
1699static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1700{
1701        return dentry->d_fsdata;
1702}
1703
1704static inline struct cftype *__d_cft(struct dentry *dentry)
1705{
1706        return dentry->d_fsdata;
1707}
1708
1709/**
1710 * cgroup_path - generate the path of a cgroup
1711 * @cgrp: the cgroup in question
1712 * @buf: the buffer to write the path into
1713 * @buflen: the length of the buffer
1714 *
1715 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1716 * reference.  Writes path of cgroup into buf.  Returns 0 on success,
1717 * -errno on error.
1718 */
1719int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1720{
1721        char *start;
1722        struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1723                                                      cgroup_lock_is_held());
1724
1725        if (!dentry || cgrp == dummytop) {
1726                /*
1727                 * Inactive subsystems have no dentry for their root
1728                 * cgroup
1729                 */
1730                strcpy(buf, "/");
1731                return 0;
1732        }
1733
1734        start = buf + buflen;
1735
1736        *--start = '\0';
1737        for (;;) {
1738                int len = dentry->d_name.len;
1739
1740                if ((start -= len) < buf)
1741                        return -ENAMETOOLONG;
1742                memcpy(start, dentry->d_name.name, len);
1743                cgrp = cgrp->parent;
1744                if (!cgrp)
1745                        break;
1746
1747                dentry = rcu_dereference_check(cgrp->dentry,
1748                                               cgroup_lock_is_held());
1749                if (!cgrp->parent)
1750                        continue;
1751                if (--start < buf)
1752                        return -ENAMETOOLONG;
1753                *start = '/';
1754        }
1755        memmove(buf, start, buf + buflen - start);
1756        return 0;
1757}
1758EXPORT_SYMBOL_GPL(cgroup_path);
1759
1760/*
1761 * Control Group taskset
1762 */
1763struct task_and_cgroup {
1764        struct task_struct      *task;
1765        struct cgroup           *cgrp;
1766};
1767
1768struct cgroup_taskset {
1769        struct task_and_cgroup  single;
1770        struct flex_array       *tc_array;
1771        int                     tc_array_len;
1772        int                     idx;
1773        struct cgroup           *cur_cgrp;
1774};
1775
1776/**
1777 * cgroup_taskset_first - reset taskset and return the first task
1778 * @tset: taskset of interest
1779 *
1780 * @tset iteration is initialized and the first task is returned.
1781 */
1782struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1783{
1784        if (tset->tc_array) {
1785                tset->idx = 0;
1786                return cgroup_taskset_next(tset);
1787        } else {
1788                tset->cur_cgrp = tset->single.cgrp;
1789                return tset->single.task;
1790        }
1791}
1792EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1793
1794/**
1795 * cgroup_taskset_next - iterate to the next task in taskset
1796 * @tset: taskset of interest
1797 *
1798 * Return the next task in @tset.  Iteration must have been initialized
1799 * with cgroup_taskset_first().
1800 */
1801struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1802{
1803        struct task_and_cgroup *tc;
1804
1805        if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1806                return NULL;
1807
1808        tc = flex_array_get(tset->tc_array, tset->idx++);
1809        tset->cur_cgrp = tc->cgrp;
1810        return tc->task;
1811}
1812EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1813
1814/**
1815 * cgroup_taskset_cur_cgroup - return the matching cgroup for the current task
1816 * @tset: taskset of interest
1817 *
1818 * Return the cgroup for the current (last returned) task of @tset.  This
1819 * function must be preceded by either cgroup_taskset_first() or
1820 * cgroup_taskset_next().
1821 */
1822struct cgroup *cgroup_taskset_cur_cgroup(struct cgroup_taskset *tset)
1823{
1824        return tset->cur_cgrp;
1825}
1826EXPORT_SYMBOL_GPL(cgroup_taskset_cur_cgroup);
1827
1828/**
1829 * cgroup_taskset_size - return the number of tasks in taskset
1830 * @tset: taskset of interest
1831 */
1832int cgroup_taskset_size(struct cgroup_taskset *tset)
1833{
1834        return tset->tc_array ? tset->tc_array_len : 1;
1835}
1836EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1837
1838
1839/*
1840 * cgroup_task_migrate - move a task from one cgroup to another.
1841 *
1842 * 'guarantee' is set if the caller promises that a new css_set for the task
1843 * will already exist. If not set, this function might sleep, and can fail with
1844 * -ENOMEM. Must be called with cgroup_mutex and threadgroup locked.
1845 */
1846static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
1847                               struct task_struct *tsk, bool guarantee)
1848{
1849        struct css_set *oldcg;
1850        struct css_set *newcg;
1851
1852        /*
1853         * We are synchronized through threadgroup_lock() against PF_EXITING
1854         * setting such that we can't race against cgroup_exit() changing the
1855         * css_set to init_css_set and dropping the old one.
1856         */
1857        WARN_ON_ONCE(tsk->flags & PF_EXITING);
1858        oldcg = tsk->cgroups;
1859
1860        /* locate or allocate a new css_set for this task. */
1861        if (guarantee) {
1862                /* we know the css_set we want already exists. */
1863                struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
1864                read_lock(&css_set_lock);
1865                newcg = find_existing_css_set(oldcg, cgrp, template);
1866                BUG_ON(!newcg);
1867                get_css_set(newcg);
1868                read_unlock(&css_set_lock);
1869        } else {
1870                might_sleep();
1871                /* find_css_set will give us newcg already referenced. */
1872                newcg = find_css_set(oldcg, cgrp);
1873                if (!newcg)
1874                        return -ENOMEM;
1875        }
1876
1877        task_lock(tsk);
1878        rcu_assign_pointer(tsk->cgroups, newcg);
1879        task_unlock(tsk);
1880
1881        /* Update the css_set linked lists if we're using them */
1882        write_lock(&css_set_lock);
1883        if (!list_empty(&tsk->cg_list))
1884                list_move(&tsk->cg_list, &newcg->tasks);
1885        write_unlock(&css_set_lock);
1886
1887        /*
1888         * We just gained a reference on oldcg by taking it from the task. As
1889         * trading it for newcg is protected by cgroup_mutex, we're safe to drop
1890         * it here; it will be freed under RCU.
1891         */
1892        put_css_set(oldcg);
1893
1894        set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1895        return 0;
1896}
1897
1898/**
1899 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1900 * @cgrp: the cgroup the task is attaching to
1901 * @tsk: the task to be attached
1902 *
1903 * Call with cgroup_mutex and threadgroup locked. May take task_lock of
1904 * @tsk during call.
1905 */
1906int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1907{
1908        int retval;
1909        struct cgroup_subsys *ss, *failed_ss = NULL;
1910        struct cgroup *oldcgrp;
1911        struct cgroupfs_root *root = cgrp->root;
1912        struct cgroup_taskset tset = { };
1913
1914        /* @tsk either already exited or can't exit until the end */
1915        if (tsk->flags & PF_EXITING)
1916                return -ESRCH;
1917
1918        /* Nothing to do if the task is already in that cgroup */
1919        oldcgrp = task_cgroup_from_root(tsk, root);
1920        if (cgrp == oldcgrp)
1921                return 0;
1922
1923        tset.single.task = tsk;
1924        tset.single.cgrp = oldcgrp;
1925
1926        for_each_subsys(root, ss) {
1927                if (ss->can_attach) {
1928                        retval = ss->can_attach(ss, cgrp, &tset);
1929                        if (retval) {
1930                                /*
1931                                 * Remember on which subsystem the can_attach()
1932                                 * failed, so that we only call cancel_attach()
1933                                 * against the subsystems whose can_attach()
1934                                 * succeeded. (See below)
1935                                 */
1936                                failed_ss = ss;
1937                                goto out;
1938                        }
1939                }
1940        }
1941
1942        retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, false);
1943        if (retval)
1944                goto out;
1945
1946        for_each_subsys(root, ss) {
1947                if (ss->attach)
1948                        ss->attach(ss, cgrp, &tset);
1949        }
1950
1951        synchronize_rcu();
1952
1953        /*
1954         * wake up rmdir() waiter. the rmdir should fail since the cgroup
1955         * is no longer empty.
1956         */
1957        cgroup_wakeup_rmdir_waiter(cgrp);
1958out:
1959        if (retval) {
1960                for_each_subsys(root, ss) {
1961                        if (ss == failed_ss)
1962                                /*
1963                                 * This subsystem was the one that failed the
1964                                 * can_attach() check earlier, so we don't need
1965                                 * to call cancel_attach() against it or any
1966                                 * remaining subsystems.
1967                                 */
1968                                break;
1969                        if (ss->cancel_attach)
1970                                ss->cancel_attach(ss, cgrp, &tset);
1971                }
1972        }
1973        return retval;
1974}
1975
1976/**
1977 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1978 * @from: attach to all cgroups of a given task
1979 * @tsk: the task to be attached
1980 */
1981int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1982{
1983        struct cgroupfs_root *root;
1984        int retval = 0;
1985
1986        cgroup_lock();
1987        for_each_active_root(root) {
1988                struct cgroup *from_cg = task_cgroup_from_root(from, root);
1989
1990                retval = cgroup_attach_task(from_cg, tsk);
1991                if (retval)
1992                        break;
1993        }
1994        cgroup_unlock();
1995
1996        return retval;
1997}
1998EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1999
2000/*
2001 * cgroup_attach_proc works in two stages, the first of which prefetches all
2002 * new css_sets needed (to make sure we have enough memory before committing
2003 * to the move) and stores them in a list of entries of the following type.
2004 * TODO: possible optimization: use css_set->rcu_head for chaining instead
2005 */
2006struct cg_list_entry {
2007        struct css_set *cg;
2008        struct list_head links;
2009};
2010
2011static bool css_set_check_fetched(struct cgroup *cgrp,
2012                                  struct task_struct *tsk, struct css_set *cg,
2013                                  struct list_head *newcg_list)
2014{
2015        struct css_set *newcg;
2016        struct cg_list_entry *cg_entry;
2017        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
2018
2019        read_lock(&css_set_lock);
2020        newcg = find_existing_css_set(cg, cgrp, template);
2021        read_unlock(&css_set_lock);
2022
2023        /* doesn't exist at all? */
2024        if (!newcg)
2025                return false;
2026        /* see if it's already in the list */
2027        list_for_each_entry(cg_entry, newcg_list, links)
2028                if (cg_entry->cg == newcg)
2029                        return true;
2030
2031        /* not found */
2032        return false;
2033}
2034
2035/*
2036 * Find the new css_set and store it in the list in preparation for moving the
2037 * given task to the given cgroup. Returns 0 or -ENOMEM.
2038 */
2039static int css_set_prefetch(struct cgroup *cgrp, struct css_set *cg,
2040                            struct list_head *newcg_list)
2041{
2042        struct css_set *newcg;
2043        struct cg_list_entry *cg_entry;
2044
2045        /* ensure a new css_set will exist for this thread */
2046        newcg = find_css_set(cg, cgrp);
2047        if (!newcg)
2048                return -ENOMEM;
2049        /* add it to the list */
2050        cg_entry = kmalloc(sizeof(struct cg_list_entry), GFP_KERNEL);
2051        if (!cg_entry) {
2052                put_css_set(newcg);
2053                return -ENOMEM;
2054        }
2055        cg_entry->cg = newcg;
2056        list_add(&cg_entry->links, newcg_list);
2057        return 0;
2058}
2059
2060/**
2061 * cgroup_attach_proc - attach all threads in a threadgroup to a cgroup
2062 * @cgrp: the cgroup to attach to
2063 * @leader: the threadgroup leader task_struct of the group to be attached
2064 *
2065 * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
2066 * task_lock of each thread in leader's threadgroup individually in turn.
2067 */
2068static int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
2069{
2070        int retval, i, group_size;
2071        struct cgroup_subsys *ss, *failed_ss = NULL;
2072        /* guaranteed to be initialized later, but the compiler needs this */
2073        struct css_set *oldcg;
2074        struct cgroupfs_root *root = cgrp->root;
2075        /* threadgroup list cursor and array */
2076        struct task_struct *tsk;
2077        struct task_and_cgroup *tc;
2078        struct flex_array *group;
2079        struct cgroup_taskset tset = { };
2080        /*
2081         * we need to make sure we have css_sets for all the tasks we're
2082         * going to move -before- we actually start moving them, so that in
2083         * case we get an ENOMEM we can bail out before making any changes.
2084         */
2085        struct list_head newcg_list;
2086        struct cg_list_entry *cg_entry, *temp_nobe;
2087
2088        /*
2089         * step 0: in order to do expensive, possibly blocking operations for
2090         * every thread, we cannot iterate the thread group list, since it needs
2091         * rcu or tasklist locked. instead, build an array of all threads in the
2092         * group - group_rwsem prevents new threads from appearing, and if
2093         * threads exit, this will just be an over-estimate.
2094         */
2095        group_size = get_nr_threads(leader);
2096        /* flex_array supports very large thread-groups better than kmalloc. */
2097        group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
2098        if (!group)
2099                return -ENOMEM;
2100        /* pre-allocate to guarantee space while iterating in rcu read-side. */
2101        retval = flex_array_prealloc(group, 0, group_size - 1, GFP_KERNEL);
2102        if (retval)
2103                goto out_free_group_list;
2104
2105        /* prevent changes to the threadgroup list while we take a snapshot. */
2106        read_lock(&tasklist_lock);
2107        if (!thread_group_leader(leader)) {
2108                /*
2109                 * a race with de_thread from another thread's exec() may strip
2110                 * us of our leadership, making while_each_thread unsafe to use
2111                 * on this task. if this happens, there is no choice but to
2112                 * throw this task away and try again (from cgroup_procs_write);
2113                 * this is "double-double-toil-and-trouble-check locking".
2114                 */
2115                read_unlock(&tasklist_lock);
2116                retval = -EAGAIN;
2117                goto out_free_group_list;
2118        }
2119
2120        tsk = leader;
2121        i = 0;
2122        do {
2123                struct task_and_cgroup ent;
2124
2125                /* @tsk either already exited or can't exit until the end */
2126                if (tsk->flags & PF_EXITING)
2127                        continue;
2128
2129                /* as per above, nr_threads may decrease, but not increase. */
2130                BUG_ON(i >= group_size);
2131                /*
2132                 * saying GFP_ATOMIC has no effect here because we did prealloc
2133                 * earlier, but it's good form to communicate our expectations.
2134                 */
2135                ent.task = tsk;
2136                ent.cgrp = task_cgroup_from_root(tsk, root);
2137                /* nothing to do if this task is already in the cgroup */
2138                if (ent.cgrp == cgrp)
2139                        continue;
2140                retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2141                BUG_ON(retval != 0);
2142                i++;
2143        } while_each_thread(leader, tsk);
2144        /* remember the number of threads in the array for later. */
2145        group_size = i;
2146        tset.tc_array = group;
2147        tset.tc_array_len = group_size;
2148        read_unlock(&tasklist_lock);
2149
2150        /* methods shouldn't be called if no task is actually migrating */
2151        retval = 0;
2152        if (!group_size)
2153                goto out_free_group_list;
2154
2155        /*
2156         * step 1: check that we can legitimately attach to the cgroup.
2157         */
2158        for_each_subsys(root, ss) {
2159                if (ss->can_attach) {
2160                        retval = ss->can_attach(ss, cgrp, &tset);
2161                        if (retval) {
2162                                failed_ss = ss;
2163                                goto out_cancel_attach;
2164                        }
2165                }
2166        }
2167
2168        /*
2169         * step 2: make sure css_sets exist for all threads to be migrated.
2170         * we use find_css_set, which allocates a new one if necessary.
2171         */
2172        INIT_LIST_HEAD(&newcg_list);
2173        for (i = 0; i < group_size; i++) {
2174                tc = flex_array_get(group, i);
2175                oldcg = tc->task->cgroups;
2176
2177                /* if we don't already have it in the list get a new one */
2178                if (!css_set_check_fetched(cgrp, tc->task, oldcg,
2179                                           &newcg_list)) {
2180                        retval = css_set_prefetch(cgrp, oldcg, &newcg_list);
2181                        if (retval)
2182                                goto out_list_teardown;
2183                }
2184        }
2185
2186        /*
2187         * step 3: now that we're guaranteed success wrt the css_sets,
2188         * proceed to move all tasks to the new cgroup.  There are no
2189         * failure cases after here, so this is the commit point.
2190         */
2191        for (i = 0; i < group_size; i++) {
2192                tc = flex_array_get(group, i);
2193                retval = cgroup_task_migrate(cgrp, tc->cgrp, tc->task, true);
2194                BUG_ON(retval);
2195        }
2196        /* nothing is sensitive to fork() after this point. */
2197
2198        /*
2199         * step 4: do subsystem attach callbacks.
2200         */
2201        for_each_subsys(root, ss) {
2202                if (ss->attach)
2203                        ss->attach(ss, cgrp, &tset);
2204        }
2205
2206        /*
2207         * step 5: success! and cleanup
2208         */
2209        synchronize_rcu();
2210        cgroup_wakeup_rmdir_waiter(cgrp);
2211        retval = 0;
2212out_list_teardown:
2213        /* clean up the list of prefetched css_sets. */
2214        list_for_each_entry_safe(cg_entry, temp_nobe, &newcg_list, links) {
2215                list_del(&cg_entry->links);
2216                put_css_set(cg_entry->cg);
2217                kfree(cg_entry);
2218        }
2219out_cancel_attach:
2220        /* same deal as in cgroup_attach_task */
2221        if (retval) {
2222                for_each_subsys(root, ss) {
2223                        if (ss == failed_ss)
2224                                break;
2225                        if (ss->cancel_attach)
2226                                ss->cancel_attach(ss, cgrp, &tset);
2227                }
2228        }
2229out_free_group_list:
2230        flex_array_free(group);
2231        return retval;
2232}
2233
2234/*
2235 * Find the task_struct of the task to attach by vpid and pass it along to the
2236 * function to attach either it or all tasks in its threadgroup. Will lock
2237 * cgroup_mutex and threadgroup; may take task_lock of task.
2238 */
2239static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2240{
2241        struct task_struct *tsk;
2242        const struct cred *cred = current_cred(), *tcred;
2243        int ret;
2244
2245        if (!cgroup_lock_live_group(cgrp))
2246                return -ENODEV;
2247
2248        if (pid) {
2249                rcu_read_lock();
2250                tsk = find_task_by_vpid(pid);
2251                if (!tsk) {
2252                        rcu_read_unlock();
2253                        cgroup_unlock();
2254                        return -ESRCH;
2255                }
2256                if (threadgroup) {
2257                        /*
2258                         * RCU protects this access, since tsk was found in the
2259                         * tid map. a race with de_thread may cause group_leader
2260                         * to stop being the leader, but cgroup_attach_proc will
2261                         * detect it later.
2262                         */
2263                        tsk = tsk->group_leader;
2264                }
2265                /*
2266                 * even if we're attaching all tasks in the thread group, we
2267                 * only need to check permissions on one of them.
2268                 */
2269                tcred = __task_cred(tsk);
2270                if (cred->euid &&
2271                    cred->euid != tcred->uid &&
2272                    cred->euid != tcred->suid) {
2273                        rcu_read_unlock();
2274                        cgroup_unlock();
2275                        return -EACCES;
2276                }
2277                get_task_struct(tsk);
2278                rcu_read_unlock();
2279        } else {
2280                if (threadgroup)
2281                        tsk = current->group_leader;
2282                else
2283                        tsk = current;
2284                get_task_struct(tsk);
2285        }
2286
2287        threadgroup_lock(tsk);
2288
2289        if (threadgroup)
2290                ret = cgroup_attach_proc(cgrp, tsk);
2291        else
2292                ret = cgroup_attach_task(cgrp, tsk);
2293
2294        threadgroup_unlock(tsk);
2295
2296        put_task_struct(tsk);
2297        cgroup_unlock();
2298        return ret;
2299}
2300
2301static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
2302{
2303        return attach_task_by_pid(cgrp, pid, false);
2304}
2305
2306static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
2307{
2308        int ret;
2309        do {
2310                /*
2311                 * attach_proc fails with -EAGAIN if threadgroup leadership
2312                 * changes in the middle of the operation, in which case we need
2313                 * to find the task_struct for the new leader and start over.
2314                 */
2315                ret = attach_task_by_pid(cgrp, tgid, true);
2316        } while (ret == -EAGAIN);
2317        return ret;
2318}
2319
2320/**
2321 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
2322 * @cgrp: the cgroup to be checked for liveness
2323 *
2324 * On success, returns true; the lock should be later released with
2325 * cgroup_unlock(). On failure returns false with no lock held.
2326 */
2327bool cgroup_lock_live_group(struct cgroup *cgrp)
2328{
2329        mutex_lock(&cgroup_mutex);
2330        if (cgroup_is_removed(cgrp)) {
2331                mutex_unlock(&cgroup_mutex);
2332                return false;
2333        }
2334        return true;
2335}
2336EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
2337
2338static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
2339                                      const char *buffer)
2340{
2341        BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2342        if (strlen(buffer) >= PATH_MAX)
2343                return -EINVAL;
2344        if (!cgroup_lock_live_group(cgrp))
2345                return -ENODEV;
2346        mutex_lock(&cgroup_root_mutex);
2347        strcpy(cgrp->root->release_agent_path, buffer);
2348        mutex_unlock(&cgroup_root_mutex);
2349        cgroup_unlock();
2350        return 0;
2351}
2352
2353static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
2354                                     struct seq_file *seq)
2355{
2356        if (!cgroup_lock_live_group(cgrp))
2357                return -ENODEV;
2358        seq_puts(seq, cgrp->root->release_agent_path);
2359        seq_putc(seq, '\n');
2360        cgroup_unlock();
2361        return 0;
2362}
2363
2364/* A buffer size big enough for numbers or short strings */
2365#define CGROUP_LOCAL_BUFFER_SIZE 64
2366
2367static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
2368                                struct file *file,
2369                                const char __user *userbuf,
2370                                size_t nbytes, loff_t *unused_ppos)
2371{
2372        char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2373        int retval = 0;
2374        char *end;
2375
2376        if (!nbytes)
2377                return -EINVAL;
2378        if (nbytes >= sizeof(buffer))
2379                return -E2BIG;
2380        if (copy_from_user(buffer, userbuf, nbytes))
2381                return -EFAULT;
2382
2383        buffer[nbytes] = 0;     /* nul-terminate */
2384        if (cft->write_u64) {
2385                u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2386                if (*end)
2387                        return -EINVAL;
2388                retval = cft->write_u64(cgrp, cft, val);
2389        } else {
2390                s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2391                if (*end)
2392                        return -EINVAL;
2393                retval = cft->write_s64(cgrp, cft, val);
2394        }
2395        if (!retval)
2396                retval = nbytes;
2397        return retval;
2398}
2399
2400static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2401                                   struct file *file,
2402                                   const char __user *userbuf,
2403                                   size_t nbytes, loff_t *unused_ppos)
2404{
2405        char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2406        int retval = 0;
2407        size_t max_bytes = cft->max_write_len;
2408        char *buffer = local_buffer;
2409
2410        if (!max_bytes)
2411                max_bytes = sizeof(local_buffer) - 1;
2412        if (nbytes >= max_bytes)
2413                return -E2BIG;
2414        /* Allocate a dynamic buffer if we need one */
2415        if (nbytes >= sizeof(local_buffer)) {
2416                buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2417                if (buffer == NULL)
2418                        return -ENOMEM;
2419        }
2420        if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2421                retval = -EFAULT;
2422                goto out;
2423        }
2424
2425        buffer[nbytes] = 0;     /* nul-terminate */
2426        retval = cft->write_string(cgrp, cft, strstrip(buffer));
2427        if (!retval)
2428                retval = nbytes;
2429out:
2430        if (buffer != local_buffer)
2431                kfree(buffer);
2432        return retval;
2433}
2434
2435static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2436                                                size_t nbytes, loff_t *ppos)
2437{
2438        struct cftype *cft = __d_cft(file->f_dentry);
2439        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2440
2441        if (cgroup_is_removed(cgrp))
2442                return -ENODEV;
2443        if (cft->write)
2444                return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2445        if (cft->write_u64 || cft->write_s64)
2446                return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2447        if (cft->write_string)
2448                return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2449        if (cft->trigger) {
2450                int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2451                return ret ? ret : nbytes;
2452        }
2453        return -EINVAL;
2454}
2455
2456static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2457                               struct file *file,
2458                               char __user *buf, size_t nbytes,
2459                               loff_t *ppos)
2460{
2461        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2462        u64 val = cft->read_u64(cgrp, cft);
2463        int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2464
2465        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2466}
2467
2468static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2469                               struct file *file,
2470                               char __user *buf, size_t nbytes,
2471                               loff_t *ppos)
2472{
2473        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2474        s64 val = cft->read_s64(cgrp, cft);
2475        int len = sprintf(tmp, "%lld\n", (long long) val);
2476
2477        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2478}
2479
2480static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2481                                   size_t nbytes, loff_t *ppos)
2482{
2483        struct cftype *cft = __d_cft(file->f_dentry);
2484        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2485
2486        if (cgroup_is_removed(cgrp))
2487                return -ENODEV;
2488
2489        if (cft->read)
2490                return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2491        if (cft->read_u64)
2492                return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2493        if (cft->read_s64)
2494                return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2495        return -EINVAL;
2496}
2497
2498/*
2499 * seqfile ops/methods for returning structured data. Currently just
2500 * supports string->u64 maps, but can be extended in future.
2501 */
2502
2503struct cgroup_seqfile_state {
2504        struct cftype *cft;
2505        struct cgroup *cgroup;
2506};
2507
2508static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2509{
2510        struct seq_file *sf = cb->state;
2511        return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2512}
2513
2514static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2515{
2516        struct cgroup_seqfile_state *state = m->private;
2517        struct cftype *cft = state->cft;
2518        if (cft->read_map) {
2519                struct cgroup_map_cb cb = {
2520                        .fill = cgroup_map_add,
2521                        .state = m,
2522                };
2523                return cft->read_map(state->cgroup, cft, &cb);
2524        }
2525        return cft->read_seq_string(state->cgroup, cft, m);
2526}
2527
2528static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2529{
2530        struct seq_file *seq = file->private_data;
2531        kfree(seq->private);
2532        return single_release(inode, file);
2533}
2534
2535static const struct file_operations cgroup_seqfile_operations = {
2536        .read = seq_read,
2537        .write = cgroup_file_write,
2538        .llseek = seq_lseek,
2539        .release = cgroup_seqfile_release,
2540};
2541
2542static int cgroup_file_open(struct inode *inode, struct file *file)
2543{
2544        int err;
2545        struct cftype *cft;
2546
2547        err = generic_file_open(inode, file);
2548        if (err)
2549                return err;
2550        cft = __d_cft(file->f_dentry);
2551
2552        if (cft->read_map || cft->read_seq_string) {
2553                struct cgroup_seqfile_state *state =
2554                        kzalloc(sizeof(*state), GFP_USER);
2555                if (!state)
2556                        return -ENOMEM;
2557                state->cft = cft;
2558                state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2559                file->f_op = &cgroup_seqfile_operations;
2560                err = single_open(file, cgroup_seqfile_show, state);
2561                if (err < 0)
2562                        kfree(state);
2563        } else if (cft->open)
2564                err = cft->open(inode, file);
2565        else
2566                err = 0;
2567
2568        return err;
2569}
2570
2571static int cgroup_file_release(struct inode *inode, struct file *file)
2572{
2573        struct cftype *cft = __d_cft(file->f_dentry);
2574        if (cft->release)
2575                return cft->release(inode, file);
2576        return 0;
2577}
2578
2579/*
2580 * cgroup_rename - Only allow simple rename of directories in place.
2581 */
2582static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2583                            struct inode *new_dir, struct dentry *new_dentry)
2584{
2585        if (!S_ISDIR(old_dentry->d_inode->i_mode))
2586                return -ENOTDIR;
2587        if (new_dentry->d_inode)
2588                return -EEXIST;
2589        if (old_dir != new_dir)
2590                return -EIO;
2591        return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2592}
2593
2594static const struct file_operations cgroup_file_operations = {
2595        .read = cgroup_file_read,
2596        .write = cgroup_file_write,
2597        .llseek = generic_file_llseek,
2598        .open = cgroup_file_open,
2599        .release = cgroup_file_release,
2600};
2601
2602static const struct inode_operations cgroup_dir_inode_operations = {
2603        .lookup = cgroup_lookup,
2604        .mkdir = cgroup_mkdir,
2605        .rmdir = cgroup_rmdir,
2606        .rename = cgroup_rename,
2607};
2608
2609static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
2610{
2611        if (dentry->d_name.len > NAME_MAX)
2612                return ERR_PTR(-ENAMETOOLONG);
2613        d_add(dentry, NULL);
2614        return NULL;
2615}
2616
2617/*
2618 * Check if a file is a control file
2619 */
2620static inline struct cftype *__file_cft(struct file *file)
2621{
2622        if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2623                return ERR_PTR(-EINVAL);
2624        return __d_cft(file->f_dentry);
2625}
2626
2627static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2628                                struct super_block *sb)
2629{
2630        struct inode *inode;
2631
2632        if (!dentry)
2633                return -ENOENT;
2634        if (dentry->d_inode)
2635                return -EEXIST;
2636
2637        inode = cgroup_new_inode(mode, sb);
2638        if (!inode)
2639                return -ENOMEM;
2640
2641        if (S_ISDIR(mode)) {
2642                inode->i_op = &cgroup_dir_inode_operations;
2643                inode->i_fop = &simple_dir_operations;
2644
2645                /* start off with i_nlink == 2 (for "." entry) */
2646                inc_nlink(inode);
2647
2648                /* start with the directory inode held, so that we can
2649                 * populate it without racing with another mkdir */
2650                mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2651        } else if (S_ISREG(mode)) {
2652                inode->i_size = 0;
2653                inode->i_fop = &cgroup_file_operations;
2654        }
2655        d_instantiate(dentry, inode);
2656        dget(dentry);   /* Extra count - pin the dentry in core */
2657        return 0;
2658}
2659
2660/*
2661 * cgroup_create_dir - create a directory for an object.
2662 * @cgrp: the cgroup we create the directory for. It must have a valid
2663 *        ->parent field. And we are going to fill its ->dentry field.
2664 * @dentry: dentry of the new cgroup
2665 * @mode: mode to set on new directory.
2666 */
2667static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2668                                umode_t mode)
2669{
2670        struct dentry *parent;
2671        int error = 0;
2672
2673        parent = cgrp->parent->dentry;
2674        error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2675        if (!error) {
2676                dentry->d_fsdata = cgrp;
2677                inc_nlink(parent->d_inode);
2678                rcu_assign_pointer(cgrp->dentry, dentry);
2679                dget(dentry);
2680        }
2681        dput(dentry);
2682
2683        return error;
2684}
2685
2686/**
2687 * cgroup_file_mode - deduce file mode of a control file
2688 * @cft: the control file in question
2689 *
2690 * returns cft->mode if ->mode is not 0
2691 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2692 * returns S_IRUGO if it has only a read handler
2693 * returns S_IWUSR if it has only a write hander
2694 */
2695static umode_t cgroup_file_mode(const struct cftype *cft)
2696{
2697        umode_t mode = 0;
2698
2699        if (cft->mode)
2700                return cft->mode;
2701
2702        if (cft->read || cft->read_u64 || cft->read_s64 ||
2703            cft->read_map || cft->read_seq_string)
2704                mode |= S_IRUGO;
2705
2706        if (cft->write || cft->write_u64 || cft->write_s64 ||
2707            cft->write_string || cft->trigger)
2708                mode |= S_IWUSR;
2709
2710        return mode;
2711}
2712
2713int cgroup_add_file(struct cgroup *cgrp,
2714                       struct cgroup_subsys *subsys,
2715                       const struct cftype *cft)
2716{
2717        struct dentry *dir = cgrp->dentry;
2718        struct dentry *dentry;
2719        int error;
2720        umode_t mode;
2721
2722        char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2723        if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2724                strcpy(name, subsys->name);
2725                strcat(name, ".");
2726        }
2727        strcat(name, cft->name);
2728        BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2729        dentry = lookup_one_len(name, dir, strlen(name));
2730        if (!IS_ERR(dentry)) {
2731                mode = cgroup_file_mode(cft);
2732                error = cgroup_create_file(dentry, mode | S_IFREG,
2733                                                cgrp->root->sb);
2734                if (!error)
2735                        dentry->d_fsdata = (void *)cft;
2736                dput(dentry);
2737        } else
2738                error = PTR_ERR(dentry);
2739        return error;
2740}
2741EXPORT_SYMBOL_GPL(cgroup_add_file);
2742
2743int cgroup_add_files(struct cgroup *cgrp,
2744                        struct cgroup_subsys *subsys,
2745                        const struct cftype cft[],
2746                        int count)
2747{
2748        int i, err;
2749        for (i = 0; i < count; i++) {
2750                err = cgroup_add_file(cgrp, subsys, &cft[i]);
2751                if (err)
2752                        return err;
2753        }
2754        return 0;
2755}
2756EXPORT_SYMBOL_GPL(cgroup_add_files);
2757
2758/**
2759 * cgroup_task_count - count the number of tasks in a cgroup.
2760 * @cgrp: the cgroup in question
2761 *
2762 * Return the number of tasks in the cgroup.
2763 */
2764int cgroup_task_count(const struct cgroup *cgrp)
2765{
2766        int count = 0;
2767        struct cg_cgroup_link *link;
2768
2769        read_lock(&css_set_lock);
2770        list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2771                count += atomic_read(&link->cg->refcount);
2772        }
2773        read_unlock(&css_set_lock);
2774        return count;
2775}
2776
2777/*
2778 * Advance a list_head iterator.  The iterator should be positioned at
2779 * the start of a css_set
2780 */
2781static void cgroup_advance_iter(struct cgroup *cgrp,
2782                                struct cgroup_iter *it)
2783{
2784        struct list_head *l = it->cg_link;
2785        struct cg_cgroup_link *link;
2786        struct css_set *cg;
2787
2788        /* Advance to the next non-empty css_set */
2789        do {
2790                l = l->next;
2791                if (l == &cgrp->css_sets) {
2792                        it->cg_link = NULL;
2793                        return;
2794                }
2795                link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2796                cg = link->cg;
2797        } while (list_empty(&cg->tasks));
2798        it->cg_link = l;
2799        it->task = cg->tasks.next;
2800}
2801
2802/*
2803 * To reduce the fork() overhead for systems that are not actually
2804 * using their cgroups capability, we don't maintain the lists running
2805 * through each css_set to its tasks until we see the list actually
2806 * used - in other words after the first call to cgroup_iter_start().
2807 *
2808 * The tasklist_lock is not held here, as do_each_thread() and
2809 * while_each_thread() are protected by RCU.
2810 */
2811static void cgroup_enable_task_cg_lists(void)
2812{
2813        struct task_struct *p, *g;
2814        write_lock(&css_set_lock);
2815        use_task_css_set_links = 1;
2816        do_each_thread(g, p) {
2817                task_lock(p);
2818                /*
2819                 * We should check if the process is exiting, otherwise
2820                 * it will race with cgroup_exit() in that the list
2821                 * entry won't be deleted though the process has exited.
2822                 */
2823                if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2824                        list_add(&p->cg_list, &p->cgroups->tasks);
2825                task_unlock(p);
2826        } while_each_thread(g, p);
2827        write_unlock(&css_set_lock);
2828}
2829
2830void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2831        __acquires(css_set_lock)
2832{
2833        /*
2834         * The first time anyone tries to iterate across a cgroup,
2835         * we need to enable the list linking each css_set to its
2836         * tasks, and fix up all existing tasks.
2837         */
2838        if (!use_task_css_set_links)
2839                cgroup_enable_task_cg_lists();
2840
2841        read_lock(&css_set_lock);
2842        it->cg_link = &cgrp->css_sets;
2843        cgroup_advance_iter(cgrp, it);
2844}
2845
2846struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2847                                        struct cgroup_iter *it)
2848{
2849        struct task_struct *res;
2850        struct list_head *l = it->task;
2851        struct cg_cgroup_link *link;
2852
2853        /* If the iterator cg is NULL, we have no tasks */
2854        if (!it->cg_link)
2855                return NULL;
2856        res = list_entry(l, struct task_struct, cg_list);
2857        /* Advance iterator to find next entry */
2858        l = l->next;
2859        link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2860        if (l == &link->cg->tasks) {
2861                /* We reached the end of this task list - move on to
2862                 * the next cg_cgroup_link */
2863                cgroup_advance_iter(cgrp, it);
2864        } else {
2865                it->task = l;
2866        }
2867        return res;
2868}
2869
2870void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2871        __releases(css_set_lock)
2872{
2873        read_unlock(&css_set_lock);
2874}
2875
2876static inline int started_after_time(struct task_struct *t1,
2877                                     struct timespec *time,
2878                                     struct task_struct *t2)
2879{
2880        int start_diff = timespec_compare(&t1->start_time, time);
2881        if (start_diff > 0) {
2882                return 1;
2883        } else if (start_diff < 0) {
2884                return 0;
2885        } else {
2886                /*
2887                 * Arbitrarily, if two processes started at the same
2888                 * time, we'll say that the lower pointer value
2889                 * started first. Note that t2 may have exited by now
2890                 * so this may not be a valid pointer any longer, but
2891                 * that's fine - it still serves to distinguish
2892                 * between two tasks started (effectively) simultaneously.
2893                 */
2894                return t1 > t2;
2895        }
2896}
2897
2898/*
2899 * This function is a callback from heap_insert() and is used to order
2900 * the heap.
2901 * In this case we order the heap in descending task start time.
2902 */
2903static inline int started_after(void *p1, void *p2)
2904{
2905        struct task_struct *t1 = p1;
2906        struct task_struct *t2 = p2;
2907        return started_after_time(t1, &t2->start_time, t2);
2908}
2909
2910/**
2911 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2912 * @scan: struct cgroup_scanner containing arguments for the scan
2913 *
2914 * Arguments include pointers to callback functions test_task() and
2915 * process_task().
2916 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2917 * and if it returns true, call process_task() for it also.
2918 * The test_task pointer may be NULL, meaning always true (select all tasks).
2919 * Effectively duplicates cgroup_iter_{start,next,end}()
2920 * but does not lock css_set_lock for the call to process_task().
2921 * The struct cgroup_scanner may be embedded in any structure of the caller's
2922 * creation.
2923 * It is guaranteed that process_task() will act on every task that
2924 * is a member of the cgroup for the duration of this call. This
2925 * function may or may not call process_task() for tasks that exit
2926 * or move to a different cgroup during the call, or are forked or
2927 * move into the cgroup during the call.
2928 *
2929 * Note that test_task() may be called with locks held, and may in some
2930 * situations be called multiple times for the same task, so it should
2931 * be cheap.
2932 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2933 * pre-allocated and will be used for heap operations (and its "gt" member will
2934 * be overwritten), else a temporary heap will be used (allocation of which
2935 * may cause this function to fail).
2936 */
2937int cgroup_scan_tasks(struct cgroup_scanner *scan)
2938{
2939        int retval, i;
2940        struct cgroup_iter it;
2941        struct task_struct *p, *dropped;
2942        /* Never dereference latest_task, since it's not refcounted */
2943        struct task_struct *latest_task = NULL;
2944        struct ptr_heap tmp_heap;
2945        struct ptr_heap *heap;
2946        struct timespec latest_time = { 0, 0 };
2947
2948        if (scan->heap) {
2949                /* The caller supplied our heap and pre-allocated its memory */
2950                heap = scan->heap;
2951                heap->gt = &started_after;
2952        } else {
2953                /* We need to allocate our own heap memory */
2954                heap = &tmp_heap;
2955                retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2956                if (retval)
2957                        /* cannot allocate the heap */
2958                        return retval;
2959        }
2960
2961 again:
2962        /*
2963         * Scan tasks in the cgroup, using the scanner's "test_task" callback
2964         * to determine which are of interest, and using the scanner's
2965         * "process_task" callback to process any of them that need an update.
2966         * Since we don't want to hold any locks during the task updates,
2967         * gather tasks to be processed in a heap structure.
2968         * The heap is sorted by descending task start time.
2969         * If the statically-sized heap fills up, we overflow tasks that
2970         * started later, and in future iterations only consider tasks that
2971         * started after the latest task in the previous pass. This
2972         * guarantees forward progress and that we don't miss any tasks.
2973         */
2974        heap->size = 0;
2975        cgroup_iter_start(scan->cg, &it);
2976        while ((p = cgroup_iter_next(scan->cg, &it))) {
2977                /*
2978                 * Only affect tasks that qualify per the caller's callback,
2979                 * if he provided one
2980                 */
2981                if (scan->test_task && !scan->test_task(p, scan))
2982                        continue;
2983                /*
2984                 * Only process tasks that started after the last task
2985                 * we processed
2986                 */
2987                if (!started_after_time(p, &latest_time, latest_task))
2988                        continue;
2989                dropped = heap_insert(heap, p);
2990                if (dropped == NULL) {
2991                        /*
2992                         * The new task was inserted; the heap wasn't
2993                         * previously full
2994                         */
2995                        get_task_struct(p);
2996                } else if (dropped != p) {
2997                        /*
2998                         * The new task was inserted, and pushed out a
2999                         * different task
3000                         */
3001                        get_task_struct(p);
3002                        put_task_struct(dropped);
3003                }
3004                /*
3005                 * Else the new task was newer than anything already in
3006                 * the heap and wasn't inserted
3007                 */
3008        }
3009        cgroup_iter_end(scan->cg, &it);
3010
3011        if (heap->size) {
3012                for (i = 0; i < heap->size; i++) {
3013                        struct task_struct *q = heap->ptrs[i];
3014                        if (i == 0) {
3015                                latest_time = q->start_time;
3016                                latest_task = q;
3017                        }
3018                        /* Process the task per the caller's callback */
3019                        scan->process_task(q, scan);
3020                        put_task_struct(q);
3021                }
3022                /*
3023                 * If we had to process any tasks at all, scan again
3024                 * in case some of them were in the middle of forking
3025                 * children that didn't get processed.
3026                 * Not the most efficient way to do it, but it avoids
3027                 * having to take callback_mutex in the fork path
3028                 */
3029                goto again;
3030        }
3031        if (heap == &tmp_heap)
3032                heap_free(&tmp_heap);
3033        return 0;
3034}
3035
3036/*
3037 * Stuff for reading the 'tasks'/'procs' files.
3038 *
3039 * Reading this file can return large amounts of data if a cgroup has
3040 * *lots* of attached tasks. So it may need several calls to read(),
3041 * but we cannot guarantee that the information we produce is correct
3042 * unless we produce it entirely atomically.
3043 *
3044 */
3045
3046/*
3047 * The following two functions "fix" the issue where there are more pids
3048 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3049 * TODO: replace with a kernel-wide solution to this problem
3050 */
3051#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3052static void *pidlist_allocate(int count)
3053{
3054        if (PIDLIST_TOO_LARGE(count))
3055                return vmalloc(count * sizeof(pid_t));
3056        else
3057                return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3058}
3059static void pidlist_free(void *p)
3060{
3061        if (is_vmalloc_addr(p))
3062                vfree(p);
3063        else
3064                kfree(p);
3065}
3066static void *pidlist_resize(void *p, int newcount)
3067{
3068        void *newlist;
3069        /* note: if new alloc fails, old p will still be valid either way */
3070        if (is_vmalloc_addr(p)) {
3071                newlist = vmalloc(newcount * sizeof(pid_t));
3072                if (!newlist)
3073                        return NULL;
3074                memcpy(newlist, p, newcount * sizeof(pid_t));
3075                vfree(p);
3076        } else {
3077                newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
3078        }
3079        return newlist;
3080}
3081
3082/*
3083 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3084 * If the new stripped list is sufficiently smaller and there's enough memory
3085 * to allocate a new buffer, will let go of the unneeded memory. Returns the
3086 * number of unique elements.
3087 */
3088/* is the size difference enough that we should re-allocate the array? */
3089#define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
3090static int pidlist_uniq(pid_t **p, int length)
3091{
3092        int src, dest = 1;
3093        pid_t *list = *p;
3094        pid_t *newlist;
3095
3096        /*
3097         * we presume the 0th element is unique, so i starts at 1. trivial
3098         * edge cases first; no work needs to be done for either
3099         */
3100        if (length == 0 || length == 1)
3101                return length;
3102        /* src and dest walk down the list; dest counts unique elements */
3103        for (src = 1; src < length; src++) {
3104                /* find next unique element */
3105                while (list[src] == list[src-1]) {
3106                        src++;
3107                        if (src == length)
3108                                goto after;
3109                }
3110                /* dest always points to where the next unique element goes */
3111                list[dest] = list[src];
3112                dest++;
3113        }
3114after:
3115        /*
3116         * if the length difference is large enough, we want to allocate a
3117         * smaller buffer to save memory. if this fails due to out of memory,
3118         * we'll just stay with what we've got.
3119         */
3120        if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
3121                newlist = pidlist_resize(list, dest);
3122                if (newlist)
3123                        *p = newlist;
3124        }
3125        return dest;
3126}
3127
3128static int cmppid(const void *a, const void *b)
3129{
3130        return *(pid_t *)a - *(pid_t *)b;
3131}
3132
3133/*
3134 * find the appropriate pidlist for our purpose (given procs vs tasks)
3135 * returns with the lock on that pidlist already held, and takes care
3136 * of the use count, or returns NULL with no locks held if we're out of
3137 * memory.
3138 */
3139static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3140                                                  enum cgroup_filetype type)
3141{
3142        struct cgroup_pidlist *l;
3143        /* don't need task_nsproxy() if we're looking at ourself */
3144        struct pid_namespace *ns = current->nsproxy->pid_ns;
3145
3146        /*
3147         * We can't drop the pidlist_mutex before taking the l->mutex in case
3148         * the last ref-holder is trying to remove l from the list at the same
3149         * time. Holding the pidlist_mutex precludes somebody taking whichever
3150         * list we find out from under us - compare release_pid_array().
3151         */
3152        mutex_lock(&cgrp->pidlist_mutex);
3153        list_for_each_entry(l, &cgrp->pidlists, links) {
3154                if (l->key.type == type && l->key.ns == ns) {
3155                        /* make sure l doesn't vanish out from under us */
3156                        down_write(&l->mutex);
3157                        mutex_unlock(&cgrp->pidlist_mutex);
3158                        return l;
3159                }
3160        }
3161        /* entry not found; create a new one */
3162        l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3163        if (!l) {
3164                mutex_unlock(&cgrp->pidlist_mutex);
3165                return l;
3166        }
3167        init_rwsem(&l->mutex);
3168        down_write(&l->mutex);
3169        l->key.type = type;
3170        l->key.ns = get_pid_ns(ns);
3171        l->use_count = 0; /* don't increment here */
3172        l->list = NULL;
3173        l->owner = cgrp;
3174        list_add(&l->links, &cgrp->pidlists);
3175        mutex_unlock(&cgrp->pidlist_mutex);
3176        return l;
3177}
3178
3179/*
3180 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3181 */
3182static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3183                              struct cgroup_pidlist **lp)
3184{
3185        pid_t *array;
3186        int length;
3187        int pid, n = 0; /* used for populating the array */
3188        struct cgroup_iter it;
3189        struct task_struct *tsk;
3190        struct cgroup_pidlist *l;
3191
3192        /*
3193         * If cgroup gets more users after we read count, we won't have
3194         * enough space - tough.  This race is indistinguishable to the
3195         * caller from the case that the additional cgroup users didn't
3196         * show up until sometime later on.
3197         */
3198        length = cgroup_task_count(cgrp);
3199        array = pidlist_allocate(length);
3200        if (!array)
3201                return -ENOMEM;
3202        /* now, populate the array */
3203        cgroup_iter_start(cgrp, &it);
3204        while ((tsk = cgroup_iter_next(cgrp, &it))) {
3205                if (unlikely(n == length))
3206                        break;
3207                /* get tgid or pid for procs or tasks file respectively */
3208                if (type == CGROUP_FILE_PROCS)
3209                        pid = task_tgid_vnr(tsk);
3210                else
3211                        pid = task_pid_vnr(tsk);
3212                if (pid > 0) /* make sure to only use valid results */
3213                        array[n++] = pid;
3214        }
3215        cgroup_iter_end(cgrp, &it);
3216        length = n;
3217        /* now sort & (if procs) strip out duplicates */
3218        sort(array, length, sizeof(pid_t), cmppid, NULL);
3219        if (type == CGROUP_FILE_PROCS)
3220                length = pidlist_uniq(&array, length);
3221        l = cgroup_pidlist_find(cgrp, type);
3222        if (!l) {
3223                pidlist_free(array);
3224                return -ENOMEM;
3225        }
3226        /* store array, freeing old if necessary - lock already held */
3227        pidlist_free(l->list);
3228        l->list = array;
3229        l->length = length;
3230        l->use_count++;
3231        up_write(&l->mutex);
3232        *lp = l;
3233        return 0;
3234}
3235
3236/**
3237 * cgroupstats_build - build and fill cgroupstats
3238 * @stats: cgroupstats to fill information into
3239 * @dentry: A dentry entry belonging to the cgroup for which stats have
3240 * been requested.
3241 *
3242 * Build and fill cgroupstats so that taskstats can export it to user
3243 * space.
3244 */
3245int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3246{
3247        int ret = -EINVAL;
3248        struct cgroup *cgrp;
3249        struct cgroup_iter it;
3250        struct task_struct *tsk;
3251
3252        /*
3253         * Validate dentry by checking the superblock operations,
3254         * and make sure it's a directory.
3255         */
3256        if (dentry->d_sb->s_op != &cgroup_ops ||
3257            !S_ISDIR(dentry->d_inode->i_mode))
3258                 goto err;
3259
3260        ret = 0;
3261        cgrp = dentry->d_fsdata;
3262
3263        cgroup_iter_start(cgrp, &it);
3264        while ((tsk = cgroup_iter_next(cgrp, &it))) {
3265                switch (tsk->state) {
3266                case TASK_RUNNING:
3267                        stats->nr_running++;
3268                        break;
3269                case TASK_INTERRUPTIBLE:
3270                        stats->nr_sleeping++;
3271                        break;
3272                case TASK_UNINTERRUPTIBLE:
3273                        stats->nr_uninterruptible++;
3274                        break;
3275                case TASK_STOPPED:
3276                        stats->nr_stopped++;
3277                        break;
3278                default:
3279                        if (delayacct_is_task_waiting_on_io(tsk))
3280                                stats->nr_io_wait++;
3281                        break;
3282                }
3283        }
3284        cgroup_iter_end(cgrp, &it);
3285
3286err:
3287        return ret;
3288}
3289
3290
3291/*
3292 * seq_file methods for the tasks/procs files. The seq_file position is the
3293 * next pid to display; the seq_file iterator is a pointer to the pid
3294 * in the cgroup->l->list array.
3295 */
3296
3297static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3298{
3299        /*
3300         * Initially we receive a position value that corresponds to
3301         * one more than the last pid shown (or 0 on the first call or
3302         * after a seek to the start). Use a binary-search to find the
3303         * next pid to display, if any
3304         */
3305        struct cgroup_pidlist *l = s->private;
3306        int index = 0, pid = *pos;
3307        int *iter;
3308
3309        down_read(&l->mutex);
3310        if (pid) {
3311                int end = l->length;
3312
3313                while (index < end) {
3314                        int mid = (index + end) / 2;
3315                        if (l->list[mid] == pid) {
3316                                index = mid;
3317                                break;
3318                        } else if (l->list[mid] <= pid)
3319                                index = mid + 1;
3320                        else
3321                                end = mid;
3322                }
3323        }
3324        /* If we're off the end of the array, we're done */
3325        if (index >= l->length)
3326                return NULL;
3327        /* Update the abstract position to be the actual pid that we found */
3328        iter = l->list + index;
3329        *pos = *iter;
3330        return iter;
3331}
3332
3333static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3334{
3335        struct cgroup_pidlist *l = s->private;
3336        up_read(&l->mutex);
3337}
3338
3339static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3340{
3341        struct cgroup_pidlist *l = s->private;
3342        pid_t *p = v;
3343        pid_t *end = l->list + l->length;
3344        /*
3345         * Advance to the next pid in the array. If this goes off the
3346         * end, we're done
3347         */
3348        p++;
3349        if (p >= end) {
3350                return NULL;
3351        } else {
3352                *pos = *p;
3353                return p;
3354        }
3355}
3356
3357static int cgroup_pidlist_show(struct seq_file *s, void *v)
3358{
3359        return seq_printf(s, "%d\n", *(int *)v);
3360}
3361
3362/*
3363 * seq_operations functions for iterating on pidlists through seq_file -
3364 * independent of whether it's tasks or procs
3365 */
3366static const struct seq_operations cgroup_pidlist_seq_operations = {
3367        .start = cgroup_pidlist_start,
3368        .stop = cgroup_pidlist_stop,
3369        .next = cgroup_pidlist_next,
3370        .show = cgroup_pidlist_show,
3371};
3372
3373static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3374{
3375        /*
3376         * the case where we're the last user of this particular pidlist will
3377         * have us remove it from the cgroup's list, which entails taking the
3378         * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3379         * pidlist_mutex, we have to take pidlist_mutex first.
3380         */
3381        mutex_lock(&l->owner->pidlist_mutex);
3382        down_write(&l->mutex);
3383        BUG_ON(!l->use_count);
3384        if (!--l->use_count) {
3385                /* we're the last user if refcount is 0; remove and free */
3386                list_del(&l->links);
3387                mutex_unlock(&l->owner->pidlist_mutex);
3388                pidlist_free(l->list);
3389                put_pid_ns(l->key.ns);
3390                up_write(&l->mutex);
3391                kfree(l);
3392                return;
3393        }
3394        mutex_unlock(&l->owner->pidlist_mutex);
3395        up_write(&l->mutex);
3396}
3397
3398static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3399{
3400        struct cgroup_pidlist *l;
3401        if (!(file->f_mode & FMODE_READ))
3402                return 0;
3403        /*
3404         * the seq_file will only be initialized if the file was opened for
3405         * reading; hence we check if it's not null only in that case.
3406         */
3407        l = ((struct seq_file *)file->private_data)->private;
3408        cgroup_release_pid_array(l);
3409        return seq_release(inode, file);
3410}
3411
3412static const struct file_operations cgroup_pidlist_operations = {
3413        .read = seq_read,
3414        .llseek = seq_lseek,
3415        .write = cgroup_file_write,
3416        .release = cgroup_pidlist_release,
3417};
3418
3419/*
3420 * The following functions handle opens on a file that displays a pidlist
3421 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3422 * in the cgroup.
3423 */
3424/* helper function for the two below it */
3425static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3426{
3427        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3428        struct cgroup_pidlist *l;
3429        int retval;
3430
3431        /* Nothing to do for write-only files */
3432        if (!(file->f_mode & FMODE_READ))
3433                return 0;
3434
3435        /* have the array populated */
3436        retval = pidlist_array_load(cgrp, type, &l);
3437        if (retval)
3438                return retval;
3439        /* configure file information */
3440        file->f_op = &cgroup_pidlist_operations;
3441
3442        retval = seq_open(file, &cgroup_pidlist_seq_operations);
3443        if (retval) {
3444                cgroup_release_pid_array(l);
3445                return retval;
3446        }
3447        ((struct seq_file *)file->private_data)->private = l;
3448        return 0;
3449}
3450static int cgroup_tasks_open(struct inode *unused, struct file *file)
3451{
3452        return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3453}
3454static int cgroup_procs_open(struct inode *unused, struct file *file)
3455{
3456        return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3457}
3458
3459static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3460                                            struct cftype *cft)
3461{
3462        return notify_on_release(cgrp);
3463}
3464
3465static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3466                                          struct cftype *cft,
3467                                          u64 val)
3468{
3469        clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3470        if (val)
3471                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3472        else
3473                clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3474        return 0;
3475}
3476
3477/*
3478 * Unregister event and free resources.
3479 *
3480 * Gets called from workqueue.
3481 */
3482static void cgroup_event_remove(struct work_struct *work)
3483{
3484        struct cgroup_event *event = container_of(work, struct cgroup_event,
3485                        remove);
3486        struct cgroup *cgrp = event->cgrp;
3487
3488        event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3489
3490        eventfd_ctx_put(event->eventfd);
3491        kfree(event);
3492        dput(cgrp->dentry);
3493}
3494
3495/*
3496 * Gets called on POLLHUP on eventfd when user closes it.
3497 *
3498 * Called with wqh->lock held and interrupts disabled.
3499 */
3500static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3501                int sync, void *key)
3502{
3503        struct cgroup_event *event = container_of(wait,
3504                        struct cgroup_event, wait);
3505        struct cgroup *cgrp = event->cgrp;
3506        unsigned long flags = (unsigned long)key;
3507
3508        if (flags & POLLHUP) {
3509                __remove_wait_queue(event->wqh, &event->wait);
3510                spin_lock(&cgrp->event_list_lock);
3511                list_del(&event->list);
3512                spin_unlock(&cgrp->event_list_lock);
3513                /*
3514                 * We are in atomic context, but cgroup_event_remove() may
3515                 * sleep, so we have to call it in workqueue.
3516                 */
3517                schedule_work(&event->remove);
3518        }
3519
3520        return 0;
3521}
3522
3523static void cgroup_event_ptable_queue_proc(struct file *file,
3524                wait_queue_head_t *wqh, poll_table *pt)
3525{
3526        struct cgroup_event *event = container_of(pt,
3527                        struct cgroup_event, pt);
3528
3529        event->wqh = wqh;
3530        add_wait_queue(wqh, &event->wait);
3531}
3532
3533/*
3534 * Parse input and register new cgroup event handler.
3535 *
3536 * Input must be in format '<event_fd> <control_fd> <args>'.
3537 * Interpretation of args is defined by control file implementation.
3538 */
3539static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3540                                      const char *buffer)
3541{
3542        struct cgroup_event *event = NULL;
3543        unsigned int efd, cfd;
3544        struct file *efile = NULL;
3545        struct file *cfile = NULL;
3546        char *endp;
3547        int ret;
3548
3549        efd = simple_strtoul(buffer, &endp, 10);
3550        if (*endp != ' ')
3551                return -EINVAL;
3552        buffer = endp + 1;
3553
3554        cfd = simple_strtoul(buffer, &endp, 10);
3555        if ((*endp != ' ') && (*endp != '\0'))
3556                return -EINVAL;
3557        buffer = endp + 1;
3558
3559        event = kzalloc(sizeof(*event), GFP_KERNEL);
3560        if (!event)
3561                return -ENOMEM;
3562        event->cgrp = cgrp;
3563        INIT_LIST_HEAD(&event->list);
3564        init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3565        init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3566        INIT_WORK(&event->remove, cgroup_event_remove);
3567
3568        efile = eventfd_fget(efd);
3569        if (IS_ERR(efile)) {
3570                ret = PTR_ERR(efile);
3571                goto fail;
3572        }
3573
3574        event->eventfd = eventfd_ctx_fileget(efile);
3575        if (IS_ERR(event->eventfd)) {
3576                ret = PTR_ERR(event->eventfd);
3577                goto fail;
3578        }
3579
3580        cfile = fget(cfd);
3581        if (!cfile) {
3582                ret = -EBADF;
3583                goto fail;
3584        }
3585
3586        /* the process need read permission on control file */
3587        /* AV: shouldn't we check that it's been opened for read instead? */
3588        ret = inode_permission(cfile->f_path.dentry->d_inode, MAY_READ);
3589        if (ret < 0)
3590                goto fail;
3591
3592        event->cft = __file_cft(cfile);
3593        if (IS_ERR(event->cft)) {
3594                ret = PTR_ERR(event->cft);
3595                goto fail;
3596        }
3597
3598        if (!event->cft->register_event || !event->cft->unregister_event) {
3599                ret = -EINVAL;
3600                goto fail;
3601        }
3602
3603        ret = event->cft->register_event(cgrp, event->cft,
3604                        event->eventfd, buffer);
3605        if (ret)
3606                goto fail;
3607
3608        if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3609                event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3610                ret = 0;
3611                goto fail;
3612        }
3613
3614        /*
3615         * Events should be removed after rmdir of cgroup directory, but before
3616         * destroying subsystem state objects. Let's take reference to cgroup
3617         * directory dentry to do that.
3618         */
3619        dget(cgrp->dentry);
3620
3621        spin_lock(&cgrp->event_list_lock);
3622        list_add(&event->list, &cgrp->event_list);
3623        spin_unlock(&cgrp->event_list_lock);
3624
3625        fput(cfile);
3626        fput(efile);
3627
3628        return 0;
3629
3630fail:
3631        if (cfile)
3632                fput(cfile);
3633
3634        if (event && event->eventfd && !IS_ERR(event->eventfd))
3635                eventfd_ctx_put(event->eventfd);
3636
3637        if (!IS_ERR_OR_NULL(efile))
3638                fput(efile);
3639
3640        kfree(event);
3641
3642        return ret;
3643}
3644
3645static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3646                                    struct cftype *cft)
3647{
3648        return clone_children(cgrp);
3649}
3650
3651static int cgroup_clone_children_write(struct cgroup *cgrp,
3652                                     struct cftype *cft,
3653                                     u64 val)
3654{
3655        if (val)
3656                set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3657        else
3658                clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3659        return 0;
3660}
3661
3662/*
3663 * for the common functions, 'private' gives the type of file
3664 */
3665/* for hysterical raisins, we can't put this on the older files */
3666#define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3667static struct cftype files[] = {
3668        {
3669                .name = "tasks",
3670                .open = cgroup_tasks_open,
3671                .write_u64 = cgroup_tasks_write,
3672                .release = cgroup_pidlist_release,
3673                .mode = S_IRUGO | S_IWUSR,
3674        },
3675        {
3676                .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3677                .open = cgroup_procs_open,
3678                .write_u64 = cgroup_procs_write,
3679                .release = cgroup_pidlist_release,
3680                .mode = S_IRUGO | S_IWUSR,
3681        },
3682        {
3683                .name = "notify_on_release",
3684                .read_u64 = cgroup_read_notify_on_release,
3685                .write_u64 = cgroup_write_notify_on_release,
3686        },
3687        {
3688                .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3689                .write_string = cgroup_write_event_control,
3690                .mode = S_IWUGO,
3691        },
3692        {
3693                .name = "cgroup.clone_children",
3694                .read_u64 = cgroup_clone_children_read,
3695                .write_u64 = cgroup_clone_children_write,
3696        },
3697};
3698
3699static struct cftype cft_release_agent = {
3700        .name = "release_agent",
3701        .read_seq_string = cgroup_release_agent_show,
3702        .write_string = cgroup_release_agent_write,
3703        .max_write_len = PATH_MAX,
3704};
3705
3706static int cgroup_populate_dir(struct cgroup *cgrp)
3707{
3708        int err;
3709        struct cgroup_subsys *ss;
3710
3711        /* First clear out any existing files */
3712        cgroup_clear_directory(cgrp->dentry);
3713
3714        err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3715        if (err < 0)
3716                return err;
3717
3718        if (cgrp == cgrp->top_cgroup) {
3719                if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3720                        return err;
3721        }
3722
3723        for_each_subsys(cgrp->root, ss) {
3724                if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3725                        return err;
3726        }
3727        /* This cgroup is ready now */
3728        for_each_subsys(cgrp->root, ss) {
3729                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3730                /*
3731                 * Update id->css pointer and make this css visible from
3732                 * CSS ID functions. This pointer will be dereferened
3733                 * from RCU-read-side without locks.
3734                 */
3735                if (css->id)
3736                        rcu_assign_pointer(css->id->css, css);
3737        }
3738
3739        return 0;
3740}
3741
3742static void init_cgroup_css(struct cgroup_subsys_state *css,
3743                               struct cgroup_subsys *ss,
3744                               struct cgroup *cgrp)
3745{
3746        css->cgroup = cgrp;
3747        atomic_set(&css->refcnt, 1);
3748        css->flags = 0;
3749        css->id = NULL;
3750        if (cgrp == dummytop)
3751                set_bit(CSS_ROOT, &css->flags);
3752        BUG_ON(cgrp->subsys[ss->subsys_id]);
3753        cgrp->subsys[ss->subsys_id] = css;
3754}
3755
3756static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3757{
3758        /* We need to take each hierarchy_mutex in a consistent order */
3759        int i;
3760
3761        /*
3762         * No worry about a race with rebind_subsystems that might mess up the
3763         * locking order, since both parties are under cgroup_mutex.
3764         */
3765        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3766                struct cgroup_subsys *ss = subsys[i];
3767                if (ss == NULL)
3768                        continue;
3769                if (ss->root == root)
3770                        mutex_lock(&ss->hierarchy_mutex);
3771        }
3772}
3773
3774static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3775{
3776        int i;
3777
3778        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3779                struct cgroup_subsys *ss = subsys[i];
3780                if (ss == NULL)
3781                        continue;
3782                if (ss->root == root)
3783                        mutex_unlock(&ss->hierarchy_mutex);
3784        }
3785}
3786
3787/*
3788 * cgroup_create - create a cgroup
3789 * @parent: cgroup that will be parent of the new cgroup
3790 * @dentry: dentry of the new cgroup
3791 * @mode: mode to set on new inode
3792 *
3793 * Must be called with the mutex on the parent inode held
3794 */
3795static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3796                             umode_t mode)
3797{
3798        struct cgroup *cgrp;
3799        struct cgroupfs_root *root = parent->root;
3800        int err = 0;
3801        struct cgroup_subsys *ss;
3802        struct super_block *sb = root->sb;
3803
3804        cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3805        if (!cgrp)
3806                return -ENOMEM;
3807
3808        /* Grab a reference on the superblock so the hierarchy doesn't
3809         * get deleted on unmount if there are child cgroups.  This
3810         * can be done outside cgroup_mutex, since the sb can't
3811         * disappear while someone has an open control file on the
3812         * fs */
3813        atomic_inc(&sb->s_active);
3814
3815        mutex_lock(&cgroup_mutex);
3816
3817        init_cgroup_housekeeping(cgrp);
3818
3819        cgrp->parent = parent;
3820        cgrp->root = parent->root;
3821        cgrp->top_cgroup = parent->top_cgroup;
3822
3823        if (notify_on_release(parent))
3824                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3825
3826        if (clone_children(parent))
3827                set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3828
3829        for_each_subsys(root, ss) {
3830                struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3831
3832                if (IS_ERR(css)) {
3833                        err = PTR_ERR(css);
3834                        goto err_destroy;
3835                }
3836                init_cgroup_css(css, ss, cgrp);
3837                if (ss->use_id) {
3838                        err = alloc_css_id(ss, parent, cgrp);
3839                        if (err)
3840                                goto err_destroy;
3841                }
3842                /* At error, ->destroy() callback has to free assigned ID. */
3843                if (clone_children(parent) && ss->post_clone)
3844                        ss->post_clone(ss, cgrp);
3845        }
3846
3847        cgroup_lock_hierarchy(root);
3848        list_add(&cgrp->sibling, &cgrp->parent->children);
3849        cgroup_unlock_hierarchy(root);
3850        root->number_of_cgroups++;
3851
3852        err = cgroup_create_dir(cgrp, dentry, mode);
3853        if (err < 0)
3854                goto err_remove;
3855
3856        /* The cgroup directory was pre-locked for us */
3857        BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3858
3859        err = cgroup_populate_dir(cgrp);
3860        /* If err < 0, we have a half-filled directory - oh well ;) */
3861
3862        mutex_unlock(&cgroup_mutex);
3863        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3864
3865        return 0;
3866
3867 err_remove:
3868
3869        cgroup_lock_hierarchy(root);
3870        list_del(&cgrp->sibling);
3871        cgroup_unlock_hierarchy(root);
3872        root->number_of_cgroups--;
3873
3874 err_destroy:
3875
3876        for_each_subsys(root, ss) {
3877                if (cgrp->subsys[ss->subsys_id])
3878                        ss->destroy(ss, cgrp);
3879        }
3880
3881        mutex_unlock(&cgroup_mutex);
3882
3883        /* Release the reference count that we took on the superblock */
3884        deactivate_super(sb);
3885
3886        kfree(cgrp);
3887        return err;
3888}
3889
3890static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3891{
3892        struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3893
3894        /* the vfs holds inode->i_mutex already */
3895        return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3896}
3897
3898static int cgroup_has_css_refs(struct cgroup *cgrp)
3899{
3900        /* Check the reference count on each subsystem. Since we
3901         * already established that there are no tasks in the
3902         * cgroup, if the css refcount is also 1, then there should
3903         * be no outstanding references, so the subsystem is safe to
3904         * destroy. We scan across all subsystems rather than using
3905         * the per-hierarchy linked list of mounted subsystems since
3906         * we can be called via check_for_release() with no
3907         * synchronization other than RCU, and the subsystem linked
3908         * list isn't RCU-safe */
3909        int i;
3910        /*
3911         * We won't need to lock the subsys array, because the subsystems
3912         * we're concerned about aren't going anywhere since our cgroup root
3913         * has a reference on them.
3914         */
3915        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3916                struct cgroup_subsys *ss = subsys[i];
3917                struct cgroup_subsys_state *css;
3918                /* Skip subsystems not present or not in this hierarchy */
3919                if (ss == NULL || ss->root != cgrp->root)
3920                        continue;
3921                css = cgrp->subsys[ss->subsys_id];
3922                /* When called from check_for_release() it's possible
3923                 * that by this point the cgroup has been removed
3924                 * and the css deleted. But a false-positive doesn't
3925                 * matter, since it can only happen if the cgroup
3926                 * has been deleted and hence no longer needs the
3927                 * release agent to be called anyway. */
3928                if (css && (atomic_read(&css->refcnt) > 1))
3929                        return 1;
3930        }
3931        return 0;
3932}
3933
3934/*
3935 * Atomically mark all (or else none) of the cgroup's CSS objects as
3936 * CSS_REMOVED. Return true on success, or false if the cgroup has
3937 * busy subsystems. Call with cgroup_mutex held
3938 */
3939
3940static int cgroup_clear_css_refs(struct cgroup *cgrp)
3941{
3942        struct cgroup_subsys *ss;
3943        unsigned long flags;
3944        bool failed = false;
3945        local_irq_save(flags);
3946        for_each_subsys(cgrp->root, ss) {
3947                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3948                int refcnt;
3949                while (1) {
3950                        /* We can only remove a CSS with a refcnt==1 */
3951                        refcnt = atomic_read(&css->refcnt);
3952                        if (refcnt > 1) {
3953                                failed = true;
3954                                goto done;
3955                        }
3956                        BUG_ON(!refcnt);
3957                        /*
3958                         * Drop the refcnt to 0 while we check other
3959                         * subsystems. This will cause any racing
3960                         * css_tryget() to spin until we set the
3961                         * CSS_REMOVED bits or abort
3962                         */
3963                        if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3964                                break;
3965                        cpu_relax();
3966                }
3967        }
3968 done:
3969        for_each_subsys(cgrp->root, ss) {
3970                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3971                if (failed) {
3972                        /*
3973                         * Restore old refcnt if we previously managed
3974                         * to clear it from 1 to 0
3975                         */
3976                        if (!atomic_read(&css->refcnt))
3977                                atomic_set(&css->refcnt, 1);
3978                } else {
3979                        /* Commit the fact that the CSS is removed */
3980                        set_bit(CSS_REMOVED, &css->flags);
3981                }
3982        }
3983        local_irq_restore(flags);
3984        return !failed;
3985}
3986
3987static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3988{
3989        struct cgroup *cgrp = dentry->d_fsdata;
3990        struct dentry *d;
3991        struct cgroup *parent;
3992        DEFINE_WAIT(wait);
3993        struct cgroup_event *event, *tmp;
3994        int ret;
3995
3996        /* the vfs holds both inode->i_mutex already */
3997again:
3998        mutex_lock(&cgroup_mutex);
3999        if (atomic_read(&cgrp->count) != 0) {
4000                mutex_unlock(&cgroup_mutex);
4001                return -EBUSY;
4002        }
4003        if (!list_empty(&cgrp->children)) {
4004                mutex_unlock(&cgroup_mutex);
4005                return -EBUSY;
4006        }
4007        mutex_unlock(&cgroup_mutex);
4008
4009        /*
4010         * In general, subsystem has no css->refcnt after pre_destroy(). But
4011         * in racy cases, subsystem may have to get css->refcnt after
4012         * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
4013         * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
4014         * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
4015         * and subsystem's reference count handling. Please see css_get/put
4016         * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
4017         */
4018        set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4019
4020        /*
4021         * Call pre_destroy handlers of subsys. Notify subsystems
4022         * that rmdir() request comes.
4023         */
4024        ret = cgroup_call_pre_destroy(cgrp);
4025        if (ret) {
4026                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4027                return ret;
4028        }
4029
4030        mutex_lock(&cgroup_mutex);
4031        parent = cgrp->parent;
4032        if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
4033                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4034                mutex_unlock(&cgroup_mutex);
4035                return -EBUSY;
4036        }
4037        prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
4038        if (!cgroup_clear_css_refs(cgrp)) {
4039                mutex_unlock(&cgroup_mutex);
4040                /*
4041                 * Because someone may call cgroup_wakeup_rmdir_waiter() before
4042                 * prepare_to_wait(), we need to check this flag.
4043                 */
4044                if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
4045                        schedule();
4046                finish_wait(&cgroup_rmdir_waitq, &wait);
4047                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4048                if (signal_pending(current))
4049                        return -EINTR;
4050                goto again;
4051        }
4052        /* NO css_tryget() can success after here. */
4053        finish_wait(&cgroup_rmdir_waitq, &wait);
4054        clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4055
4056        raw_spin_lock(&release_list_lock);
4057        set_bit(CGRP_REMOVED, &cgrp->flags);
4058        if (!list_empty(&cgrp->release_list))
4059                list_del_init(&cgrp->release_list);
4060        raw_spin_unlock(&release_list_lock);
4061
4062        cgroup_lock_hierarchy(cgrp->root);
4063        /* delete this cgroup from parent->children */
4064        list_del_init(&cgrp->sibling);
4065        cgroup_unlock_hierarchy(cgrp->root);
4066
4067        d = dget(cgrp->dentry);
4068
4069        cgroup_d_remove_dir(d);
4070        dput(d);
4071
4072        set_bit(CGRP_RELEASABLE, &parent->flags);
4073        check_for_release(parent);
4074
4075        /*
4076         * Unregister events and notify userspace.
4077         * Notify userspace about cgroup removing only after rmdir of cgroup
4078         * directory to avoid race between userspace and kernelspace
4079         */
4080        spin_lock(&cgrp->event_list_lock);
4081        list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4082                list_del(&event->list);
4083                remove_wait_queue(event->wqh, &event->wait);
4084                eventfd_signal(event->eventfd, 1);
4085                schedule_work(&event->remove);
4086        }
4087        spin_unlock(&cgrp->event_list_lock);
4088
4089        mutex_unlock(&cgroup_mutex);
4090        return 0;
4091}
4092
4093static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4094{
4095        struct cgroup_subsys_state *css;
4096
4097        printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4098
4099        /* Create the top cgroup state for this subsystem */
4100        list_add(&ss->sibling, &rootnode.subsys_list);
4101        ss->root = &rootnode;
4102        css = ss->create(ss, dummytop);
4103        /* We don't handle early failures gracefully */
4104        BUG_ON(IS_ERR(css));
4105        init_cgroup_css(css, ss, dummytop);
4106
4107        /* Update the init_css_set to contain a subsys
4108         * pointer to this state - since the subsystem is
4109         * newly registered, all tasks and hence the
4110         * init_css_set is in the subsystem's top cgroup. */
4111        init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
4112
4113        need_forkexit_callback |= ss->fork || ss->exit;
4114
4115        /* At system boot, before all subsystems have been
4116         * registered, no tasks have been forked, so we don't
4117         * need to invoke fork callbacks here. */
4118        BUG_ON(!list_empty(&init_task.tasks));
4119
4120        mutex_init(&ss->hierarchy_mutex);
4121        lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4122        ss->active = 1;
4123
4124        /* this function shouldn't be used with modular subsystems, since they
4125         * need to register a subsys_id, among other things */
4126        BUG_ON(ss->module);
4127}
4128
4129/**
4130 * cgroup_load_subsys: load and register a modular subsystem at runtime
4131 * @ss: the subsystem to load
4132 *
4133 * This function should be called in a modular subsystem's initcall. If the
4134 * subsystem is built as a module, it will be assigned a new subsys_id and set
4135 * up for use. If the subsystem is built-in anyway, work is delegated to the
4136 * simpler cgroup_init_subsys.
4137 */
4138int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4139{
4140        int i;
4141        struct cgroup_subsys_state *css;
4142
4143        /* check name and function validity */
4144        if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4145            ss->create == NULL || ss->destroy == NULL)
4146                return -EINVAL;
4147
4148        /*
4149         * we don't support callbacks in modular subsystems. this check is
4150         * before the ss->module check for consistency; a subsystem that could
4151         * be a module should still have no callbacks even if the user isn't
4152         * compiling it as one.
4153         */
4154        if (ss->fork || ss->exit)
4155                return -EINVAL;
4156
4157        /*
4158         * an optionally modular subsystem is built-in: we want to do nothing,
4159         * since cgroup_init_subsys will have already taken care of it.
4160         */
4161        if (ss->module == NULL) {
4162                /* a few sanity checks */
4163                BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
4164                BUG_ON(subsys[ss->subsys_id] != ss);
4165                return 0;
4166        }
4167
4168        /*
4169         * need to register a subsys id before anything else - for example,
4170         * init_cgroup_css needs it.
4171         */
4172        mutex_lock(&cgroup_mutex);
4173        /* find the first empty slot in the array */
4174        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
4175                if (subsys[i] == NULL)
4176                        break;
4177        }
4178        if (i == CGROUP_SUBSYS_COUNT) {
4179                /* maximum number of subsystems already registered! */
4180                mutex_unlock(&cgroup_mutex);
4181                return -EBUSY;
4182        }
4183        /* assign ourselves the subsys_id */
4184        ss->subsys_id = i;
4185        subsys[i] = ss;
4186
4187        /*
4188         * no ss->create seems to need anything important in the ss struct, so
4189         * this can happen first (i.e. before the rootnode attachment).
4190         */
4191        css = ss->create(ss, dummytop);
4192        if (IS_ERR(css)) {
4193                /* failure case - need to deassign the subsys[] slot. */
4194                subsys[i] = NULL;
4195                mutex_unlock(&cgroup_mutex);
4196                return PTR_ERR(css);
4197        }
4198
4199        list_add(&ss->sibling, &rootnode.subsys_list);
4200        ss->root = &rootnode;
4201
4202        /* our new subsystem will be attached to the dummy hierarchy. */
4203        init_cgroup_css(css, ss, dummytop);
4204        /* init_idr must be after init_cgroup_css because it sets css->id. */
4205        if (ss->use_id) {
4206                int ret = cgroup_init_idr(ss, css);
4207                if (ret) {
4208                        dummytop->subsys[ss->subsys_id] = NULL;
4209                        ss->destroy(ss, dummytop);
4210                        subsys[i] = NULL;
4211                        mutex_unlock(&cgroup_mutex);
4212                        return ret;
4213                }
4214        }
4215
4216        /*
4217         * Now we need to entangle the css into the existing css_sets. unlike
4218         * in cgroup_init_subsys, there are now multiple css_sets, so each one
4219         * will need a new pointer to it; done by iterating the css_set_table.
4220         * furthermore, modifying the existing css_sets will corrupt the hash
4221         * table state, so each changed css_set will need its hash recomputed.
4222         * this is all done under the css_set_lock.
4223         */
4224        write_lock(&css_set_lock);
4225        for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
4226                struct css_set *cg;
4227                struct hlist_node *node, *tmp;
4228                struct hlist_head *bucket = &css_set_table[i], *new_bucket;
4229
4230                hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
4231                        /* skip entries that we already rehashed */
4232                        if (cg->subsys[ss->subsys_id])
4233                                continue;
4234                        /* remove existing entry */
4235                        hlist_del(&cg->hlist);
4236                        /* set new value */
4237                        cg->subsys[ss->subsys_id] = css;
4238                        /* recompute hash and restore entry */
4239                        new_bucket = css_set_hash(cg->subsys);
4240                        hlist_add_head(&cg->hlist, new_bucket);
4241                }
4242        }
4243        write_unlock(&css_set_lock);
4244
4245        mutex_init(&ss->hierarchy_mutex);
4246        lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4247        ss->active = 1;
4248
4249        /* success! */
4250        mutex_unlock(&cgroup_mutex);
4251        return 0;
4252}
4253EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4254
4255/**
4256 * cgroup_unload_subsys: unload a modular subsystem
4257 * @ss: the subsystem to unload
4258 *
4259 * This function should be called in a modular subsystem's exitcall. When this
4260 * function is invoked, the refcount on the subsystem's module will be 0, so
4261 * the subsystem will not be attached to any hierarchy.
4262 */
4263void cgroup_unload_subsys(struct cgroup_subsys *ss)
4264{
4265        struct cg_cgroup_link *link;
4266        struct hlist_head *hhead;
4267
4268        BUG_ON(ss->module == NULL);
4269
4270        /*
4271         * we shouldn't be called if the subsystem is in use, and the use of
4272         * try_module_get in parse_cgroupfs_options should ensure that it
4273         * doesn't start being used while we're killing it off.
4274         */
4275        BUG_ON(ss->root != &rootnode);
4276
4277        mutex_lock(&cgroup_mutex);
4278        /* deassign the subsys_id */
4279        BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
4280        subsys[ss->subsys_id] = NULL;
4281
4282        /* remove subsystem from rootnode's list of subsystems */
4283        list_del_init(&ss->sibling);
4284
4285        /*
4286         * disentangle the css from all css_sets attached to the dummytop. as
4287         * in loading, we need to pay our respects to the hashtable gods.
4288         */
4289        write_lock(&css_set_lock);
4290        list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
4291                struct css_set *cg = link->cg;
4292
4293                hlist_del(&cg->hlist);
4294                BUG_ON(!cg->subsys[ss->subsys_id]);
4295                cg->subsys[ss->subsys_id] = NULL;
4296                hhead = css_set_hash(cg->subsys);
4297                hlist_add_head(&cg->hlist, hhead);
4298        }
4299        write_unlock(&css_set_lock);
4300
4301        /*
4302         * remove subsystem's css from the dummytop and free it - need to free
4303         * before marking as null because ss->destroy needs the cgrp->subsys
4304         * pointer to find their state. note that this also takes care of
4305         * freeing the css_id.
4306         */
4307        ss->destroy(ss, dummytop);
4308        dummytop->subsys[ss->subsys_id] = NULL;
4309
4310        mutex_unlock(&cgroup_mutex);
4311}
4312EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4313
4314/**
4315 * cgroup_init_early - cgroup initialization at system boot
4316 *
4317 * Initialize cgroups at system boot, and initialize any
4318 * subsystems that request early init.
4319 */
4320int __init cgroup_init_early(void)
4321{
4322        int i;
4323        atomic_set(&init_css_set.refcount, 1);
4324        INIT_LIST_HEAD(&init_css_set.cg_links);
4325        INIT_LIST_HEAD(&init_css_set.tasks);
4326        INIT_HLIST_NODE(&init_css_set.hlist);
4327        css_set_count = 1;
4328        init_cgroup_root(&rootnode);
4329        root_count = 1;
4330        init_task.cgroups = &init_css_set;
4331
4332        init_css_set_link.cg = &init_css_set;
4333        init_css_set_link.cgrp = dummytop;
4334        list_add(&init_css_set_link.cgrp_link_list,
4335                 &rootnode.top_cgroup.css_sets);
4336        list_add(&init_css_set_link.cg_link_list,
4337                 &init_css_set.cg_links);
4338
4339        for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
4340                INIT_HLIST_HEAD(&css_set_table[i]);
4341
4342        /* at bootup time, we don't worry about modular subsystems */
4343        for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4344                struct cgroup_subsys *ss = subsys[i];
4345
4346                BUG_ON(!ss->name);
4347                BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4348                BUG_ON(!ss->create);
4349                BUG_ON(!ss->destroy);
4350                if (ss->subsys_id != i) {
4351                        printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4352                               ss->name, ss->subsys_id);
4353                        BUG();
4354                }
4355
4356                if (ss->early_init)
4357                        cgroup_init_subsys(ss);
4358        }
4359        return 0;
4360}
4361
4362/**
4363 * cgroup_init - cgroup initialization
4364 *
4365 * Register cgroup filesystem and /proc file, and initialize
4366 * any subsystems that didn't request early init.
4367 */
4368int __init cgroup_init(void)
4369{
4370        int err;
4371        int i;
4372        struct hlist_head *hhead;
4373
4374        err = bdi_init(&cgroup_backing_dev_info);
4375        if (err)
4376                return err;
4377
4378        /* at bootup time, we don't worry about modular subsystems */
4379        for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4380                struct cgroup_subsys *ss = subsys[i];
4381                if (!ss->early_init)
4382                        cgroup_init_subsys(ss);
4383                if (ss->use_id)
4384                        cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
4385        }
4386
4387        /* Add init_css_set to the hash table */
4388        hhead = css_set_hash(init_css_set.subsys);
4389        hlist_add_head(&init_css_set.hlist, hhead);
4390        BUG_ON(!init_root_id(&rootnode));
4391
4392        cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4393        if (!cgroup_kobj) {
4394                err = -ENOMEM;
4395                goto out;
4396        }
4397
4398        err = register_filesystem(&cgroup_fs_type);
4399        if (err < 0) {
4400                kobject_put(cgroup_kobj);
4401                goto out;
4402        }
4403
4404        proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4405
4406out:
4407        if (err)
4408                bdi_destroy(&cgroup_backing_dev_info);
4409
4410        return err;
4411}
4412
4413/*
4414 * proc_cgroup_show()
4415 *  - Print task's cgroup paths into seq_file, one line for each hierarchy
4416 *  - Used for /proc/<pid>/cgroup.
4417 *  - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4418 *    doesn't really matter if tsk->cgroup changes after we read it,
4419 *    and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4420 *    anyway.  No need to check that tsk->cgroup != NULL, thanks to
4421 *    the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4422 *    cgroup to top_cgroup.
4423 */
4424
4425/* TODO: Use a proper seq_file iterator */
4426static int proc_cgroup_show(struct seq_file *m, void *v)
4427{
4428        struct pid *pid;
4429        struct task_struct *tsk;
4430        char *buf;
4431        int retval;
4432        struct cgroupfs_root *root;
4433
4434        retval = -ENOMEM;
4435        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4436        if (!buf)
4437                goto out;
4438
4439        retval = -ESRCH;
4440        pid = m->private;
4441        tsk = get_pid_task(pid, PIDTYPE_PID);
4442        if (!tsk)
4443                goto out_free;
4444
4445        retval = 0;
4446
4447        mutex_lock(&cgroup_mutex);
4448
4449        for_each_active_root(root) {
4450                struct cgroup_subsys *ss;
4451                struct cgroup *cgrp;
4452                int count = 0;
4453
4454                seq_printf(m, "%d:", root->hierarchy_id);
4455                for_each_subsys(root, ss)
4456                        seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4457                if (strlen(root->name))
4458                        seq_printf(m, "%sname=%s", count ? "," : "",
4459                                   root->name);
4460                seq_putc(m, ':');
4461                cgrp = task_cgroup_from_root(tsk, root);
4462                retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4463                if (retval < 0)
4464                        goto out_unlock;
4465                seq_puts(m, buf);
4466                seq_putc(m, '\n');
4467        }
4468
4469out_unlock:
4470        mutex_unlock(&cgroup_mutex);
4471        put_task_struct(tsk);
4472out_free:
4473        kfree(buf);
4474out:
4475        return retval;
4476}
4477
4478static int cgroup_open(struct inode *inode, struct file *file)
4479{
4480        struct pid *pid = PROC_I(inode)->pid;
4481        return single_open(file, proc_cgroup_show, pid);
4482}
4483
4484const struct file_operations proc_cgroup_operations = {
4485        .open           = cgroup_open,
4486        .read           = seq_read,
4487        .llseek         = seq_lseek,
4488        .release        = single_release,
4489};
4490
4491/* Display information about each subsystem and each hierarchy */
4492static int proc_cgroupstats_show(struct seq_file *m, void *v)
4493{
4494        int i;
4495
4496        seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4497        /*
4498         * ideally we don't want subsystems moving around while we do this.
4499         * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4500         * subsys/hierarchy state.
4501         */
4502        mutex_lock(&cgroup_mutex);
4503        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4504                struct cgroup_subsys *ss = subsys[i];
4505                if (ss == NULL)
4506                        continue;
4507                seq_printf(m, "%s\t%d\t%d\t%d\n",
4508                           ss->name, ss->root->hierarchy_id,
4509                           ss->root->number_of_cgroups, !ss->disabled);
4510        }
4511        mutex_unlock(&cgroup_mutex);
4512        return 0;
4513}
4514
4515static int cgroupstats_open(struct inode *inode, struct file *file)
4516{
4517        return single_open(file, proc_cgroupstats_show, NULL);
4518}
4519
4520static const struct file_operations proc_cgroupstats_operations = {
4521        .open = cgroupstats_open,
4522        .read = seq_read,
4523        .llseek = seq_lseek,
4524        .release = single_release,
4525};
4526
4527/**
4528 * cgroup_fork - attach newly forked task to its parents cgroup.
4529 * @child: pointer to task_struct of forking parent process.
4530 *
4531 * Description: A task inherits its parent's cgroup at fork().
4532 *
4533 * A pointer to the shared css_set was automatically copied in
4534 * fork.c by dup_task_struct().  However, we ignore that copy, since
4535 * it was not made under the protection of RCU, cgroup_mutex or
4536 * threadgroup_change_begin(), so it might no longer be a valid
4537 * cgroup pointer.  cgroup_attach_task() might have already changed
4538 * current->cgroups, allowing the previously referenced cgroup
4539 * group to be removed and freed.
4540 *
4541 * Outside the pointer validity we also need to process the css_set
4542 * inheritance between threadgoup_change_begin() and
4543 * threadgoup_change_end(), this way there is no leak in any process
4544 * wide migration performed by cgroup_attach_proc() that could otherwise
4545 * miss a thread because it is too early or too late in the fork stage.
4546 *
4547 * At the point that cgroup_fork() is called, 'current' is the parent
4548 * task, and the passed argument 'child' points to the child task.
4549 */
4550void cgroup_fork(struct task_struct *child)
4551{
4552        /*
4553         * We don't need to task_lock() current because current->cgroups
4554         * can't be changed concurrently here. The parent obviously hasn't
4555         * exited and called cgroup_exit(), and we are synchronized against
4556         * cgroup migration through threadgroup_change_begin().
4557         */
4558        child->cgroups = current->cgroups;
4559        get_css_set(child->cgroups);
4560        INIT_LIST_HEAD(&child->cg_list);
4561}
4562
4563/**
4564 * cgroup_fork_callbacks - run fork callbacks
4565 * @child: the new task
4566 *
4567 * Called on a new task very soon before adding it to the
4568 * tasklist. No need to take any locks since no-one can
4569 * be operating on this task.
4570 */
4571void cgroup_fork_callbacks(struct task_struct *child)
4572{
4573        if (need_forkexit_callback) {
4574                int i;
4575                /*
4576                 * forkexit callbacks are only supported for builtin
4577                 * subsystems, and the builtin section of the subsys array is
4578                 * immutable, so we don't need to lock the subsys array here.
4579                 */
4580                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4581                        struct cgroup_subsys *ss = subsys[i];
4582                        if (ss->fork)
4583                                ss->fork(ss, child);
4584                }
4585        }
4586}
4587
4588/**
4589 * cgroup_post_fork - called on a new task after adding it to the task list
4590 * @child: the task in question
4591 *
4592 * Adds the task to the list running through its css_set if necessary.
4593 * Has to be after the task is visible on the task list in case we race
4594 * with the first call to cgroup_iter_start() - to guarantee that the
4595 * new task ends up on its list.
4596 */
4597void cgroup_post_fork(struct task_struct *child)
4598{
4599        if (use_task_css_set_links) {
4600                write_lock(&css_set_lock);
4601                if (list_empty(&child->cg_list)) {
4602                        /*
4603                         * It's safe to use child->cgroups without task_lock()
4604                         * here because we are protected through
4605                         * threadgroup_change_begin() against concurrent
4606                         * css_set change in cgroup_task_migrate(). Also
4607                         * the task can't exit at that point until
4608                         * wake_up_new_task() is called, so we are protected
4609                         * against cgroup_exit() setting child->cgroup to
4610                         * init_css_set.
4611                         */
4612                        list_add(&child->cg_list, &child->cgroups->tasks);
4613                }
4614                write_unlock(&css_set_lock);
4615        }
4616}
4617/**
4618 * cgroup_exit - detach cgroup from exiting task
4619 * @tsk: pointer to task_struct of exiting process
4620 * @run_callback: run exit callbacks?
4621 *
4622 * Description: Detach cgroup from @tsk and release it.
4623 *
4624 * Note that cgroups marked notify_on_release force every task in
4625 * them to take the global cgroup_mutex mutex when exiting.
4626 * This could impact scaling on very large systems.  Be reluctant to
4627 * use notify_on_release cgroups where very high task exit scaling
4628 * is required on large systems.
4629 *
4630 * the_top_cgroup_hack:
4631 *
4632 *    Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4633 *
4634 *    We call cgroup_exit() while the task is still competent to
4635 *    handle notify_on_release(), then leave the task attached to the
4636 *    root cgroup in each hierarchy for the remainder of its exit.
4637 *
4638 *    To do this properly, we would increment the reference count on
4639 *    top_cgroup, and near the very end of the kernel/exit.c do_exit()
4640 *    code we would add a second cgroup function call, to drop that
4641 *    reference.  This would just create an unnecessary hot spot on
4642 *    the top_cgroup reference count, to no avail.
4643 *
4644 *    Normally, holding a reference to a cgroup without bumping its
4645 *    count is unsafe.   The cgroup could go away, or someone could
4646 *    attach us to a different cgroup, decrementing the count on
4647 *    the first cgroup that we never incremented.  But in this case,
4648 *    top_cgroup isn't going away, and either task has PF_EXITING set,
4649 *    which wards off any cgroup_attach_task() attempts, or task is a failed
4650 *    fork, never visible to cgroup_attach_task.
4651 */
4652void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4653{
4654        struct css_set *cg;
4655        int i;
4656
4657        /*
4658         * Unlink from the css_set task list if necessary.
4659         * Optimistically check cg_list before taking
4660         * css_set_lock
4661         */
4662        if (!list_empty(&tsk->cg_list)) {
4663                write_lock(&css_set_lock);
4664                if (!list_empty(&tsk->cg_list))
4665                        list_del_init(&tsk->cg_list);
4666                write_unlock(&css_set_lock);
4667        }
4668
4669        /* Reassign the task to the init_css_set. */
4670        task_lock(tsk);
4671        cg = tsk->cgroups;
4672        tsk->cgroups = &init_css_set;
4673
4674        if (run_callbacks && need_forkexit_callback) {
4675                /*
4676                 * modular subsystems can't use callbacks, so no need to lock
4677                 * the subsys array
4678                 */
4679                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4680                        struct cgroup_subsys *ss = subsys[i];
4681                        if (ss->exit) {
4682                                struct cgroup *old_cgrp =
4683                                        rcu_dereference_raw(cg->subsys[i])->cgroup;
4684                                struct cgroup *cgrp = task_cgroup(tsk, i);
4685                                ss->exit(ss, cgrp, old_cgrp, tsk);
4686                        }
4687                }
4688        }
4689        task_unlock(tsk);
4690
4691        if (cg)
4692                put_css_set_taskexit(cg);
4693}
4694
4695/**
4696 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4697 * @cgrp: the cgroup in question
4698 * @task: the task in question
4699 *
4700 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4701 * hierarchy.
4702 *
4703 * If we are sending in dummytop, then presumably we are creating
4704 * the top cgroup in the subsystem.
4705 *
4706 * Called only by the ns (nsproxy) cgroup.
4707 */
4708int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4709{
4710        int ret;
4711        struct cgroup *target;
4712
4713        if (cgrp == dummytop)
4714                return 1;
4715
4716        target = task_cgroup_from_root(task, cgrp->root);
4717        while (cgrp != target && cgrp!= cgrp->top_cgroup)
4718                cgrp = cgrp->parent;
4719        ret = (cgrp == target);
4720        return ret;
4721}
4722
4723static void check_for_release(struct cgroup *cgrp)
4724{
4725        /* All of these checks rely on RCU to keep the cgroup
4726         * structure alive */
4727        if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4728            && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4729                /* Control Group is currently removeable. If it's not
4730                 * already queued for a userspace notification, queue
4731                 * it now */
4732                int need_schedule_work = 0;
4733                raw_spin_lock(&release_list_lock);
4734                if (!cgroup_is_removed(cgrp) &&
4735                    list_empty(&cgrp->release_list)) {
4736                        list_add(&cgrp->release_list, &release_list);
4737                        need_schedule_work = 1;
4738                }
4739                raw_spin_unlock(&release_list_lock);
4740                if (need_schedule_work)
4741                        schedule_work(&release_agent_work);
4742        }
4743}
4744
4745/* Caller must verify that the css is not for root cgroup */
4746void __css_put(struct cgroup_subsys_state *css, int count)
4747{
4748        struct cgroup *cgrp = css->cgroup;
4749        int val;
4750        rcu_read_lock();
4751        val = atomic_sub_return(count, &css->refcnt);
4752        if (val == 1) {
4753                if (notify_on_release(cgrp)) {
4754                        set_bit(CGRP_RELEASABLE, &cgrp->flags);
4755                        check_for_release(cgrp);
4756                }
4757                cgroup_wakeup_rmdir_waiter(cgrp);
4758        }
4759        rcu_read_unlock();
4760        WARN_ON_ONCE(val < 1);
4761}
4762EXPORT_SYMBOL_GPL(__css_put);
4763
4764/*
4765 * Notify userspace when a cgroup is released, by running the
4766 * configured release agent with the name of the cgroup (path
4767 * relative to the root of cgroup file system) as the argument.
4768 *
4769 * Most likely, this user command will try to rmdir this cgroup.
4770 *
4771 * This races with the possibility that some other task will be
4772 * attached to this cgroup before it is removed, or that some other
4773 * user task will 'mkdir' a child cgroup of this cgroup.  That's ok.
4774 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4775 * unused, and this cgroup will be reprieved from its death sentence,
4776 * to continue to serve a useful existence.  Next time it's released,
4777 * we will get notified again, if it still has 'notify_on_release' set.
4778 *
4779 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4780 * means only wait until the task is successfully execve()'d.  The
4781 * separate release agent task is forked by call_usermodehelper(),
4782 * then control in this thread returns here, without waiting for the
4783 * release agent task.  We don't bother to wait because the caller of
4784 * this routine has no use for the exit status of the release agent
4785 * task, so no sense holding our caller up for that.
4786 */
4787static void cgroup_release_agent(struct work_struct *work)
4788{
4789        BUG_ON(work != &release_agent_work);
4790        mutex_lock(&cgroup_mutex);
4791        raw_spin_lock(&release_list_lock);
4792        while (!list_empty(&release_list)) {
4793                char *argv[3], *envp[3];
4794                int i;
4795                char *pathbuf = NULL, *agentbuf = NULL;
4796                struct cgroup *cgrp = list_entry(release_list.next,
4797                                                    struct cgroup,
4798                                                    release_list);
4799                list_del_init(&cgrp->release_list);
4800                raw_spin_unlock(&release_list_lock);
4801                pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4802                if (!pathbuf)
4803                        goto continue_free;
4804                if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4805                        goto continue_free;
4806                agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4807                if (!agentbuf)
4808                        goto continue_free;
4809
4810                i = 0;
4811                argv[i++] = agentbuf;
4812                argv[i++] = pathbuf;
4813                argv[i] = NULL;
4814
4815                i = 0;
4816                /* minimal command environment */
4817                envp[i++] = "HOME=/";
4818                envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4819                envp[i] = NULL;
4820
4821                /* Drop the lock while we invoke the usermode helper,
4822                 * since the exec could involve hitting disk and hence
4823                 * be a slow process */
4824                mutex_unlock(&cgroup_mutex);
4825                call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4826                mutex_lock(&cgroup_mutex);
4827 continue_free:
4828                kfree(pathbuf);
4829                kfree(agentbuf);
4830                raw_spin_lock(&release_list_lock);
4831        }
4832        raw_spin_unlock(&release_list_lock);
4833        mutex_unlock(&cgroup_mutex);
4834}
4835
4836static int __init cgroup_disable(char *str)
4837{
4838        int i;
4839        char *token;
4840
4841        while ((token = strsep(&str, ",")) != NULL) {
4842                if (!*token)
4843                        continue;
4844                /*
4845                 * cgroup_disable, being at boot time, can't know about module
4846                 * subsystems, so we don't worry about them.
4847                 */
4848                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4849                        struct cgroup_subsys *ss = subsys[i];
4850
4851                        if (!strcmp(token, ss->name)) {
4852                                ss->disabled = 1;
4853                                printk(KERN_INFO "Disabling %s control group"
4854                                        " subsystem\n", ss->name);
4855                                break;
4856                        }
4857                }
4858        }
4859        return 1;
4860}
4861__setup("cgroup_disable=", cgroup_disable);
4862
4863/*
4864 * Functons for CSS ID.
4865 */
4866
4867/*
4868 *To get ID other than 0, this should be called when !cgroup_is_removed().
4869 */
4870unsigned short css_id(struct cgroup_subsys_state *css)
4871{
4872        struct css_id *cssid;
4873
4874        /*
4875         * This css_id() can return correct value when somone has refcnt
4876         * on this or this is under rcu_read_lock(). Once css->id is allocated,
4877         * it's unchanged until freed.