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
  66static DEFINE_MUTEX(cgroup_mutex);
  67
  68/*
  69 * Generate an array of cgroup subsystem pointers. At boot time, this is
  70 * populated up to CGROUP_BUILTIN_SUBSYS_COUNT, and modular subsystems are
  71 * registered after that. The mutable section of this array is protected by
  72 * cgroup_mutex.
  73 */
  74#define SUBSYS(_x) &_x ## _subsys,
  75static struct cgroup_subsys *subsys[CGROUP_SUBSYS_COUNT] = {
  76#include <linux/cgroup_subsys.h>
  77};
  78
  79#define MAX_CGROUP_ROOT_NAMELEN 64
  80
  81/*
  82 * A cgroupfs_root represents the root of a cgroup hierarchy,
  83 * and may be associated with a superblock to form an active
  84 * hierarchy
  85 */
  86struct cgroupfs_root {
  87        struct super_block *sb;
  88
  89        /*
  90         * The bitmask of subsystems intended to be attached to this
  91         * hierarchy
  92         */
  93        unsigned long subsys_bits;
  94
  95        /* Unique id for this hierarchy. */
  96        int hierarchy_id;
  97
  98        /* The bitmask of subsystems currently attached to this hierarchy */
  99        unsigned long actual_subsys_bits;
 100
 101        /* A list running through the attached subsystems */
 102        struct list_head subsys_list;
 103
 104        /* The root cgroup for this hierarchy */
 105        struct cgroup top_cgroup;
 106
 107        /* Tracks how many cgroups are currently defined in hierarchy.*/
 108        int number_of_cgroups;
 109
 110        /* A list running through the active hierarchies */
 111        struct list_head root_list;
 112
 113        /* Hierarchy-specific flags */
 114        unsigned long flags;
 115
 116        /* The path to use for release notifications. */
 117        char release_agent_path[PATH_MAX];
 118
 119        /* The name for this hierarchy - may be empty */
 120        char name[MAX_CGROUP_ROOT_NAMELEN];
 121};
 122
 123/*
 124 * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
 125 * subsystems that are otherwise unattached - it never has more than a
 126 * single cgroup, and all tasks are part of that cgroup.
 127 */
 128static struct cgroupfs_root rootnode;
 129
 130/*
 131 * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
 132 * cgroup_subsys->use_id != 0.
 133 */
 134#define CSS_ID_MAX      (65535)
 135struct css_id {
 136        /*
 137         * The css to which this ID points. This pointer is set to valid value
 138         * after cgroup is populated. If cgroup is removed, this will be NULL.
 139         * This pointer is expected to be RCU-safe because destroy()
 140         * is called after synchronize_rcu(). But for safe use, css_is_removed()
 141         * css_tryget() should be used for avoiding race.
 142         */
 143        struct cgroup_subsys_state __rcu *css;
 144        /*
 145         * ID of this css.
 146         */
 147        unsigned short id;
 148        /*
 149         * Depth in hierarchy which this ID belongs to.
 150         */
 151        unsigned short depth;
 152        /*
 153         * ID is freed by RCU. (and lookup routine is RCU safe.)
 154         */
 155        struct rcu_head rcu_head;
 156        /*
 157         * Hierarchy of CSS ID belongs to.
 158         */
 159        unsigned short stack[0]; /* Array of Length (depth+1) */
 160};
 161
 162/*
 163 * cgroup_event represents events which userspace want to receive.
 164 */
 165struct cgroup_event {
 166        /*
 167         * Cgroup which the event belongs to.
 168         */
 169        struct cgroup *cgrp;
 170        /*
 171         * Control file which the event associated.
 172         */
 173        struct cftype *cft;
 174        /*
 175         * eventfd to signal userspace about the event.
 176         */
 177        struct eventfd_ctx *eventfd;
 178        /*
 179         * Each of these stored in a list by the cgroup.
 180         */
 181        struct list_head list;
 182        /*
 183         * All fields below needed to unregister event when
 184         * userspace closes eventfd.
 185         */
 186        poll_table pt;
 187        wait_queue_head_t *wqh;
 188        wait_queue_t wait;
 189        struct work_struct remove;
 190};
 191
 192/* The list of hierarchy roots */
 193
 194static LIST_HEAD(roots);
 195static int root_count;
 196
 197static DEFINE_IDA(hierarchy_ida);
 198static int next_hierarchy_id;
 199static DEFINE_SPINLOCK(hierarchy_id_lock);
 200
 201/* dummytop is a shorthand for the dummy hierarchy's top cgroup */
 202#define dummytop (&rootnode.top_cgroup)
 203
 204/* This flag indicates whether tasks in the fork and exit paths should
 205 * check for fork/exit handlers to call. This avoids us having to do
 206 * extra work in the fork/exit path if none of the subsystems need to
 207 * be called.
 208 */
 209static int need_forkexit_callback __read_mostly;
 210
 211#ifdef CONFIG_PROVE_LOCKING
 212int cgroup_lock_is_held(void)
 213{
 214        return lockdep_is_held(&cgroup_mutex);
 215}
 216#else /* #ifdef CONFIG_PROVE_LOCKING */
 217int cgroup_lock_is_held(void)
 218{
 219        return mutex_is_locked(&cgroup_mutex);
 220}
 221#endif /* #else #ifdef CONFIG_PROVE_LOCKING */
 222
 223EXPORT_SYMBOL_GPL(cgroup_lock_is_held);
 224
 225/* convenient tests for these bits */
 226inline int cgroup_is_removed(const struct cgroup *cgrp)
 227{
 228        return test_bit(CGRP_REMOVED, &cgrp->flags);
 229}
 230
 231/* bits in struct cgroupfs_root flags field */
 232enum {
 233        ROOT_NOPREFIX, /* mounted subsystems have no named prefix */
 234};
 235
 236static int cgroup_is_releasable(const struct cgroup *cgrp)
 237{
 238        const int bits =
 239                (1 << CGRP_RELEASABLE) |
 240                (1 << CGRP_NOTIFY_ON_RELEASE);
 241        return (cgrp->flags & bits) == bits;
 242}
 243
 244static int notify_on_release(const struct cgroup *cgrp)
 245{
 246        return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
 247}
 248
 249static int clone_children(const struct cgroup *cgrp)
 250{
 251        return test_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
 252}
 253
 254/*
 255 * for_each_subsys() allows you to iterate on each subsystem attached to
 256 * an active hierarchy
 257 */
 258#define for_each_subsys(_root, _ss) \
 259list_for_each_entry(_ss, &_root->subsys_list, sibling)
 260
 261/* for_each_active_root() allows you to iterate across the active hierarchies */
 262#define for_each_active_root(_root) \
 263list_for_each_entry(_root, &roots, root_list)
 264
 265/* the list of cgroups eligible for automatic release. Protected by
 266 * release_list_lock */
 267static LIST_HEAD(release_list);
 268static DEFINE_SPINLOCK(release_list_lock);
 269static void cgroup_release_agent(struct work_struct *work);
 270static DECLARE_WORK(release_agent_work, cgroup_release_agent);
 271static void check_for_release(struct cgroup *cgrp);
 272
 273/* Link structure for associating css_set objects with cgroups */
 274struct cg_cgroup_link {
 275        /*
 276         * List running through cg_cgroup_links associated with a
 277         * cgroup, anchored on cgroup->css_sets
 278         */
 279        struct list_head cgrp_link_list;
 280        struct cgroup *cgrp;
 281        /*
 282         * List running through cg_cgroup_links pointing at a
 283         * single css_set object, anchored on css_set->cg_links
 284         */
 285        struct list_head cg_link_list;
 286        struct css_set *cg;
 287};
 288
 289/* The default css_set - used by init and its children prior to any
 290 * hierarchies being mounted. It contains a pointer to the root state
 291 * for each subsystem. Also used to anchor the list of css_sets. Not
 292 * reference-counted, to improve performance when child cgroups
 293 * haven't been created.
 294 */
 295
 296static struct css_set init_css_set;
 297static struct cg_cgroup_link init_css_set_link;
 298
 299static int cgroup_init_idr(struct cgroup_subsys *ss,
 300                           struct cgroup_subsys_state *css);
 301
 302/* css_set_lock protects the list of css_set objects, and the
 303 * chain of tasks off each css_set.  Nests outside task->alloc_lock
 304 * due to cgroup_iter_start() */
 305static DEFINE_RWLOCK(css_set_lock);
 306static int css_set_count;
 307
 308/*
 309 * hash table for cgroup groups. This improves the performance to find
 310 * an existing css_set. This hash doesn't (currently) take into
 311 * account cgroups in empty hierarchies.
 312 */
 313#define CSS_SET_HASH_BITS       7
 314#define CSS_SET_TABLE_SIZE      (1 << CSS_SET_HASH_BITS)
 315static struct hlist_head css_set_table[CSS_SET_TABLE_SIZE];
 316
 317static struct hlist_head *css_set_hash(struct cgroup_subsys_state *css[])
 318{
 319        int i;
 320        int index;
 321        unsigned long tmp = 0UL;
 322
 323        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++)
 324                tmp += (unsigned long)css[i];
 325        tmp = (tmp >> 16) ^ tmp;
 326
 327        index = hash_long(tmp, CSS_SET_HASH_BITS);
 328
 329        return &css_set_table[index];
 330}
 331
 332/* We don't maintain the lists running through each css_set to its
 333 * task until after the first call to cgroup_iter_start(). This
 334 * reduces the fork()/exit() overhead for people who have cgroups
 335 * compiled into their kernel but not actually in use */
 336static int use_task_css_set_links __read_mostly;
 337
 338static void __put_css_set(struct css_set *cg, int taskexit)
 339{
 340        struct cg_cgroup_link *link;
 341        struct cg_cgroup_link *saved_link;
 342        /*
 343         * Ensure that the refcount doesn't hit zero while any readers
 344         * can see it. Similar to atomic_dec_and_lock(), but for an
 345         * rwlock
 346         */
 347        if (atomic_add_unless(&cg->refcount, -1, 1))
 348                return;
 349        write_lock(&css_set_lock);
 350        if (!atomic_dec_and_test(&cg->refcount)) {
 351                write_unlock(&css_set_lock);
 352                return;
 353        }
 354
 355        /* This css_set is dead. unlink it and release cgroup refcounts */
 356        hlist_del(&cg->hlist);
 357        css_set_count--;
 358
 359        list_for_each_entry_safe(link, saved_link, &cg->cg_links,
 360                                 cg_link_list) {
 361                struct cgroup *cgrp = link->cgrp;
 362                list_del(&link->cg_link_list);
 363                list_del(&link->cgrp_link_list);
 364                if (atomic_dec_and_test(&cgrp->count) &&
 365                    notify_on_release(cgrp)) {
 366                        if (taskexit)
 367                                set_bit(CGRP_RELEASABLE, &cgrp->flags);
 368                        check_for_release(cgrp);
 369                }
 370
 371                kfree(link);
 372        }
 373
 374        write_unlock(&css_set_lock);
 375        kfree_rcu(cg, rcu_head);
 376}
 377
 378/*
 379 * refcounted get/put for css_set objects
 380 */
 381static inline void get_css_set(struct css_set *cg)
 382{
 383        atomic_inc(&cg->refcount);
 384}
 385
 386static inline void put_css_set(struct css_set *cg)
 387{
 388        __put_css_set(cg, 0);
 389}
 390
 391static inline void put_css_set_taskexit(struct css_set *cg)
 392{
 393        __put_css_set(cg, 1);
 394}
 395
 396/*
 397 * compare_css_sets - helper function for find_existing_css_set().
 398 * @cg: candidate css_set being tested
 399 * @old_cg: existing css_set for a task
 400 * @new_cgrp: cgroup that's being entered by the task
 401 * @template: desired set of css pointers in css_set (pre-calculated)
 402 *
 403 * Returns true if "cg" matches "old_cg" except for the hierarchy
 404 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
 405 */
 406static bool compare_css_sets(struct css_set *cg,
 407                             struct css_set *old_cg,
 408                             struct cgroup *new_cgrp,
 409                             struct cgroup_subsys_state *template[])
 410{
 411        struct list_head *l1, *l2;
 412
 413        if (memcmp(template, cg->subsys, sizeof(cg->subsys))) {
 414                /* Not all subsystems matched */
 415                return false;
 416        }
 417
 418        /*
 419         * Compare cgroup pointers in order to distinguish between
 420         * different cgroups in heirarchies with no subsystems. We
 421         * could get by with just this check alone (and skip the
 422         * memcmp above) but on most setups the memcmp check will
 423         * avoid the need for this more expensive check on almost all
 424         * candidates.
 425         */
 426
 427        l1 = &cg->cg_links;
 428        l2 = &old_cg->cg_links;
 429        while (1) {
 430                struct cg_cgroup_link *cgl1, *cgl2;
 431                struct cgroup *cg1, *cg2;
 432
 433                l1 = l1->next;
 434                l2 = l2->next;
 435                /* See if we reached the end - both lists are equal length. */
 436                if (l1 == &cg->cg_links) {
 437                        BUG_ON(l2 != &old_cg->cg_links);
 438                        break;
 439                } else {
 440                        BUG_ON(l2 == &old_cg->cg_links);
 441                }
 442                /* Locate the cgroups associated with these links. */
 443                cgl1 = list_entry(l1, struct cg_cgroup_link, cg_link_list);
 444                cgl2 = list_entry(l2, struct cg_cgroup_link, cg_link_list);
 445                cg1 = cgl1->cgrp;
 446                cg2 = cgl2->cgrp;
 447                /* Hierarchies should be linked in the same order. */
 448                BUG_ON(cg1->root != cg2->root);
 449
 450                /*
 451                 * If this hierarchy is the hierarchy of the cgroup
 452                 * that's changing, then we need to check that this
 453                 * css_set points to the new cgroup; if it's any other
 454                 * hierarchy, then this css_set should point to the
 455                 * same cgroup as the old css_set.
 456                 */
 457                if (cg1->root == new_cgrp->root) {
 458                        if (cg1 != new_cgrp)
 459                                return false;
 460                } else {
 461                        if (cg1 != cg2)
 462                                return false;
 463                }
 464        }
 465        return true;
 466}
 467
 468/*
 469 * find_existing_css_set() is a helper for
 470 * find_css_set(), and checks to see whether an existing
 471 * css_set is suitable.
 472 *
 473 * oldcg: the cgroup group that we're using before the cgroup
 474 * transition
 475 *
 476 * cgrp: the cgroup that we're moving into
 477 *
 478 * template: location in which to build the desired set of subsystem
 479 * state objects for the new cgroup group
 480 */
 481static struct css_set *find_existing_css_set(
 482        struct css_set *oldcg,
 483        struct cgroup *cgrp,
 484        struct cgroup_subsys_state *template[])
 485{
 486        int i;
 487        struct cgroupfs_root *root = cgrp->root;
 488        struct hlist_head *hhead;
 489        struct hlist_node *node;
 490        struct css_set *cg;
 491
 492        /*
 493         * Build the set of subsystem state objects that we want to see in the
 494         * new css_set. while subsystems can change globally, the entries here
 495         * won't change, so no need for locking.
 496         */
 497        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 498                if (root->subsys_bits & (1UL << i)) {
 499                        /* Subsystem is in this hierarchy. So we want
 500                         * the subsystem state from the new
 501                         * cgroup */
 502                        template[i] = cgrp->subsys[i];
 503                } else {
 504                        /* Subsystem is not in this hierarchy, so we
 505                         * don't want to change the subsystem state */
 506                        template[i] = oldcg->subsys[i];
 507                }
 508        }
 509
 510        hhead = css_set_hash(template);
 511        hlist_for_each_entry(cg, node, hhead, hlist) {
 512                if (!compare_css_sets(cg, oldcg, cgrp, template))
 513                        continue;
 514
 515                /* This css_set matches what we need */
 516                return cg;
 517        }
 518
 519        /* No existing cgroup group matched */
 520        return NULL;
 521}
 522
 523static void free_cg_links(struct list_head *tmp)
 524{
 525        struct cg_cgroup_link *link;
 526        struct cg_cgroup_link *saved_link;
 527
 528        list_for_each_entry_safe(link, saved_link, tmp, cgrp_link_list) {
 529                list_del(&link->cgrp_link_list);
 530                kfree(link);
 531        }
 532}
 533
 534/*
 535 * allocate_cg_links() allocates "count" cg_cgroup_link structures
 536 * and chains them on tmp through their cgrp_link_list fields. Returns 0 on
 537 * success or a negative error
 538 */
 539static int allocate_cg_links(int count, struct list_head *tmp)
 540{
 541        struct cg_cgroup_link *link;
 542        int i;
 543        INIT_LIST_HEAD(tmp);
 544        for (i = 0; i < count; i++) {
 545                link = kmalloc(sizeof(*link), GFP_KERNEL);
 546                if (!link) {
 547                        free_cg_links(tmp);
 548                        return -ENOMEM;
 549                }
 550                list_add(&link->cgrp_link_list, tmp);
 551        }
 552        return 0;
 553}
 554
 555/**
 556 * link_css_set - a helper function to link a css_set to a cgroup
 557 * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
 558 * @cg: the css_set to be linked
 559 * @cgrp: the destination cgroup
 560 */
 561static void link_css_set(struct list_head *tmp_cg_links,
 562                         struct css_set *cg, struct cgroup *cgrp)
 563{
 564        struct cg_cgroup_link *link;
 565
 566        BUG_ON(list_empty(tmp_cg_links));
 567        link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
 568                                cgrp_link_list);
 569        link->cg = cg;
 570        link->cgrp = cgrp;
 571        atomic_inc(&cgrp->count);
 572        list_move(&link->cgrp_link_list, &cgrp->css_sets);
 573        /*
 574         * Always add links to the tail of the list so that the list
 575         * is sorted by order of hierarchy creation
 576         */
 577        list_add_tail(&link->cg_link_list, &cg->cg_links);
 578}
 579
 580/*
 581 * find_css_set() takes an existing cgroup group and a
 582 * cgroup object, and returns a css_set object that's
 583 * equivalent to the old group, but with the given cgroup
 584 * substituted into the appropriate hierarchy. Must be called with
 585 * cgroup_mutex held
 586 */
 587static struct css_set *find_css_set(
 588        struct css_set *oldcg, struct cgroup *cgrp)
 589{
 590        struct css_set *res;
 591        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
 592
 593        struct list_head tmp_cg_links;
 594
 595        struct hlist_head *hhead;
 596        struct cg_cgroup_link *link;
 597
 598        /* First see if we already have a cgroup group that matches
 599         * the desired set */
 600        read_lock(&css_set_lock);
 601        res = find_existing_css_set(oldcg, cgrp, template);
 602        if (res)
 603                get_css_set(res);
 604        read_unlock(&css_set_lock);
 605
 606        if (res)
 607                return res;
 608
 609        res = kmalloc(sizeof(*res), GFP_KERNEL);
 610        if (!res)
 611                return NULL;
 612
 613        /* Allocate all the cg_cgroup_link objects that we'll need */
 614        if (allocate_cg_links(root_count, &tmp_cg_links) < 0) {
 615                kfree(res);
 616                return NULL;
 617        }
 618
 619        atomic_set(&res->refcount, 1);
 620        INIT_LIST_HEAD(&res->cg_links);
 621        INIT_LIST_HEAD(&res->tasks);
 622        INIT_HLIST_NODE(&res->hlist);
 623
 624        /* Copy the set of subsystem state objects generated in
 625         * find_existing_css_set() */
 626        memcpy(res->subsys, template, sizeof(res->subsys));
 627
 628        write_lock(&css_set_lock);
 629        /* Add reference counts and links from the new css_set. */
 630        list_for_each_entry(link, &oldcg->cg_links, cg_link_list) {
 631                struct cgroup *c = link->cgrp;
 632                if (c->root == cgrp->root)
 633                        c = cgrp;
 634                link_css_set(&tmp_cg_links, res, c);
 635        }
 636
 637        BUG_ON(!list_empty(&tmp_cg_links));
 638
 639        css_set_count++;
 640
 641        /* Add this cgroup group to the hash table */
 642        hhead = css_set_hash(res->subsys);
 643        hlist_add_head(&res->hlist, hhead);
 644
 645        write_unlock(&css_set_lock);
 646
 647        return res;
 648}
 649
 650/*
 651 * Return the cgroup for "task" from the given hierarchy. Must be
 652 * called with cgroup_mutex held.
 653 */
 654static struct cgroup *task_cgroup_from_root(struct task_struct *task,
 655                                            struct cgroupfs_root *root)
 656{
 657        struct css_set *css;
 658        struct cgroup *res = NULL;
 659
 660        BUG_ON(!mutex_is_locked(&cgroup_mutex));
 661        read_lock(&css_set_lock);
 662        /*
 663         * No need to lock the task - since we hold cgroup_mutex the
 664         * task can't change groups, so the only thing that can happen
 665         * is that it exits and its css is set back to init_css_set.
 666         */
 667        css = task->cgroups;
 668        if (css == &init_css_set) {
 669                res = &root->top_cgroup;
 670        } else {
 671                struct cg_cgroup_link *link;
 672                list_for_each_entry(link, &css->cg_links, cg_link_list) {
 673                        struct cgroup *c = link->cgrp;
 674                        if (c->root == root) {
 675                                res = c;
 676                                break;
 677                        }
 678                }
 679        }
 680        read_unlock(&css_set_lock);
 681        BUG_ON(!res);
 682        return res;
 683}
 684
 685/*
 686 * There is one global cgroup mutex. We also require taking
 687 * task_lock() when dereferencing a task's cgroup subsys pointers.
 688 * See "The task_lock() exception", at the end of this comment.
 689 *
 690 * A task must hold cgroup_mutex to modify cgroups.
 691 *
 692 * Any task can increment and decrement the count field without lock.
 693 * So in general, code holding cgroup_mutex can't rely on the count
 694 * field not changing.  However, if the count goes to zero, then only
 695 * cgroup_attach_task() can increment it again.  Because a count of zero
 696 * means that no tasks are currently attached, therefore there is no
 697 * way a task attached to that cgroup can fork (the other way to
 698 * increment the count).  So code holding cgroup_mutex can safely
 699 * assume that if the count is zero, it will stay zero. Similarly, if
 700 * a task holds cgroup_mutex on a cgroup with zero count, it
 701 * knows that the cgroup won't be removed, as cgroup_rmdir()
 702 * needs that mutex.
 703 *
 704 * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
 705 * (usually) take cgroup_mutex.  These are the two most performance
 706 * critical pieces of code here.  The exception occurs on cgroup_exit(),
 707 * when a task in a notify_on_release cgroup exits.  Then cgroup_mutex
 708 * is taken, and if the cgroup count is zero, a usermode call made
 709 * to the release agent with the name of the cgroup (path relative to
 710 * the root of cgroup file system) as the argument.
 711 *
 712 * A cgroup can only be deleted if both its 'count' of using tasks
 713 * is zero, and its list of 'children' cgroups is empty.  Since all
 714 * tasks in the system use _some_ cgroup, and since there is always at
 715 * least one task in the system (init, pid == 1), therefore, top_cgroup
 716 * always has either children cgroups and/or using tasks.  So we don't
 717 * need a special hack to ensure that top_cgroup cannot be deleted.
 718 *
 719 *      The task_lock() exception
 720 *
 721 * The need for this exception arises from the action of
 722 * cgroup_attach_task(), which overwrites one tasks cgroup pointer with
 723 * another.  It does so using cgroup_mutex, however there are
 724 * several performance critical places that need to reference
 725 * task->cgroup without the expense of grabbing a system global
 726 * mutex.  Therefore except as noted below, when dereferencing or, as
 727 * in cgroup_attach_task(), modifying a task'ss cgroup pointer we use
 728 * task_lock(), which acts on a spinlock (task->alloc_lock) already in
 729 * the task_struct routinely used for such matters.
 730 *
 731 * P.S.  One more locking exception.  RCU is used to guard the
 732 * update of a tasks cgroup pointer by cgroup_attach_task()
 733 */
 734
 735/**
 736 * cgroup_lock - lock out any changes to cgroup structures
 737 *
 738 */
 739void cgroup_lock(void)
 740{
 741        mutex_lock(&cgroup_mutex);
 742}
 743EXPORT_SYMBOL_GPL(cgroup_lock);
 744
 745/**
 746 * cgroup_unlock - release lock on cgroup changes
 747 *
 748 * Undo the lock taken in a previous cgroup_lock() call.
 749 */
 750void cgroup_unlock(void)
 751{
 752        mutex_unlock(&cgroup_mutex);
 753}
 754EXPORT_SYMBOL_GPL(cgroup_unlock);
 755
 756/*
 757 * A couple of forward declarations required, due to cyclic reference loop:
 758 * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
 759 * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
 760 * -> cgroup_mkdir.
 761 */
 762
 763static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
 764static struct dentry *cgroup_lookup(struct inode *, struct dentry *, struct nameidata *);
 765static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
 766static int cgroup_populate_dir(struct cgroup *cgrp);
 767static const struct inode_operations cgroup_dir_inode_operations;
 768static const struct file_operations proc_cgroupstats_operations;
 769
 770static struct backing_dev_info cgroup_backing_dev_info = {
 771        .name           = "cgroup",
 772        .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 773};
 774
 775static int alloc_css_id(struct cgroup_subsys *ss,
 776                        struct cgroup *parent, struct cgroup *child);
 777
 778static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
 779{
 780        struct inode *inode = new_inode(sb);
 781
 782        if (inode) {
 783                inode->i_ino = get_next_ino();
 784                inode->i_mode = mode;
 785                inode->i_uid = current_fsuid();
 786                inode->i_gid = current_fsgid();
 787                inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 788                inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
 789        }
 790        return inode;
 791}
 792
 793/*
 794 * Call subsys's pre_destroy handler.
 795 * This is called before css refcnt check.
 796 */
 797static int cgroup_call_pre_destroy(struct cgroup *cgrp)
 798{
 799        struct cgroup_subsys *ss;
 800        int ret = 0;
 801
 802        for_each_subsys(cgrp->root, ss)
 803                if (ss->pre_destroy) {
 804                        ret = ss->pre_destroy(ss, cgrp);
 805                        if (ret)
 806                                break;
 807                }
 808
 809        return ret;
 810}
 811
 812static void cgroup_diput(struct dentry *dentry, struct inode *inode)
 813{
 814        /* is dentry a directory ? if so, kfree() associated cgroup */
 815        if (S_ISDIR(inode->i_mode)) {
 816                struct cgroup *cgrp = dentry->d_fsdata;
 817                struct cgroup_subsys *ss;
 818                BUG_ON(!(cgroup_is_removed(cgrp)));
 819                /* It's possible for external users to be holding css
 820                 * reference counts on a cgroup; css_put() needs to
 821                 * be able to access the cgroup after decrementing
 822                 * the reference count in order to know if it needs to
 823                 * queue the cgroup to be handled by the release
 824                 * agent */
 825                synchronize_rcu();
 826
 827                mutex_lock(&cgroup_mutex);
 828                /*
 829                 * Release the subsystem state objects.
 830                 */
 831                for_each_subsys(cgrp->root, ss)
 832                        ss->destroy(ss, cgrp);
 833
 834                cgrp->root->number_of_cgroups--;
 835                mutex_unlock(&cgroup_mutex);
 836
 837                /*
 838                 * Drop the active superblock reference that we took when we
 839                 * created the cgroup
 840                 */
 841                deactivate_super(cgrp->root->sb);
 842
 843                /*
 844                 * if we're getting rid of the cgroup, refcount should ensure
 845                 * that there are no pidlists left.
 846                 */
 847                BUG_ON(!list_empty(&cgrp->pidlists));
 848
 849                kfree_rcu(cgrp, rcu_head);
 850        }
 851        iput(inode);
 852}
 853
 854static int cgroup_delete(const struct dentry *d)
 855{
 856        return 1;
 857}
 858
 859static void remove_dir(struct dentry *d)
 860{
 861        struct dentry *parent = dget(d->d_parent);
 862
 863        d_delete(d);
 864        simple_rmdir(parent->d_inode, d);
 865        dput(parent);
 866}
 867
 868static void cgroup_clear_directory(struct dentry *dentry)
 869{
 870        struct list_head *node;
 871
 872        BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
 873        spin_lock(&dentry->d_lock);
 874        node = dentry->d_subdirs.next;
 875        while (node != &dentry->d_subdirs) {
 876                struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
 877
 878                spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
 879                list_del_init(node);
 880                if (d->d_inode) {
 881                        /* This should never be called on a cgroup
 882                         * directory with child cgroups */
 883                        BUG_ON(d->d_inode->i_mode & S_IFDIR);
 884                        dget_dlock(d);
 885                        spin_unlock(&d->d_lock);
 886                        spin_unlock(&dentry->d_lock);
 887                        d_delete(d);
 888                        simple_unlink(dentry->d_inode, d);
 889                        dput(d);
 890                        spin_lock(&dentry->d_lock);
 891                } else
 892                        spin_unlock(&d->d_lock);
 893                node = dentry->d_subdirs.next;
 894        }
 895        spin_unlock(&dentry->d_lock);
 896}
 897
 898/*
 899 * NOTE : the dentry must have been dget()'ed
 900 */
 901static void cgroup_d_remove_dir(struct dentry *dentry)
 902{
 903        struct dentry *parent;
 904
 905        cgroup_clear_directory(dentry);
 906
 907        parent = dentry->d_parent;
 908        spin_lock(&parent->d_lock);
 909        spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
 910        list_del_init(&dentry->d_u.d_child);
 911        spin_unlock(&dentry->d_lock);
 912        spin_unlock(&parent->d_lock);
 913        remove_dir(dentry);
 914}
 915
 916/*
 917 * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
 918 * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
 919 * reference to css->refcnt. In general, this refcnt is expected to goes down
 920 * to zero, soon.
 921 *
 922 * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
 923 */
 924DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
 925
 926static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
 927{
 928        if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
 929                wake_up_all(&cgroup_rmdir_waitq);
 930}
 931
 932void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
 933{
 934        css_get(css);
 935}
 936
 937void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
 938{
 939        cgroup_wakeup_rmdir_waiter(css->cgroup);
 940        css_put(css);
 941}
 942
 943/*
 944 * Call with cgroup_mutex held. Drops reference counts on modules, including
 945 * any duplicate ones that parse_cgroupfs_options took. If this function
 946 * returns an error, no reference counts are touched.
 947 */
 948static int rebind_subsystems(struct cgroupfs_root *root,
 949                              unsigned long final_bits)
 950{
 951        unsigned long added_bits, removed_bits;
 952        struct cgroup *cgrp = &root->top_cgroup;
 953        int i;
 954
 955        BUG_ON(!mutex_is_locked(&cgroup_mutex));
 956
 957        removed_bits = root->actual_subsys_bits & ~final_bits;
 958        added_bits = final_bits & ~root->actual_subsys_bits;
 959        /* Check that any added subsystems are currently free */
 960        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 961                unsigned long bit = 1UL << i;
 962                struct cgroup_subsys *ss = subsys[i];
 963                if (!(bit & added_bits))
 964                        continue;
 965                /*
 966                 * Nobody should tell us to do a subsys that doesn't exist:
 967                 * parse_cgroupfs_options should catch that case and refcounts
 968                 * ensure that subsystems won't disappear once selected.
 969                 */
 970                BUG_ON(ss == NULL);
 971                if (ss->root != &rootnode) {
 972                        /* Subsystem isn't free */
 973                        return -EBUSY;
 974                }
 975        }
 976
 977        /* Currently we don't handle adding/removing subsystems when
 978         * any child cgroups exist. This is theoretically supportable
 979         * but involves complex error handling, so it's being left until
 980         * later */
 981        if (root->number_of_cgroups > 1)
 982                return -EBUSY;
 983
 984        /* Process each subsystem */
 985        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
 986                struct cgroup_subsys *ss = subsys[i];
 987                unsigned long bit = 1UL << i;
 988                if (bit & added_bits) {
 989                        /* We're binding this subsystem to this hierarchy */
 990                        BUG_ON(ss == NULL);
 991                        BUG_ON(cgrp->subsys[i]);
 992                        BUG_ON(!dummytop->subsys[i]);
 993                        BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
 994                        mutex_lock(&ss->hierarchy_mutex);
 995                        cgrp->subsys[i] = dummytop->subsys[i];
 996                        cgrp->subsys[i]->cgroup = cgrp;
 997                        list_move(&ss->sibling, &root->subsys_list);
 998                        ss->root = root;
 999                        if (ss->bind)
1000                                ss->bind(ss, cgrp);
1001                        mutex_unlock(&ss->hierarchy_mutex);
1002                        /* refcount was already taken, and we're keeping it */
1003                } else if (bit & removed_bits) {
1004                        /* We're removing this subsystem */
1005                        BUG_ON(ss == NULL);
1006                        BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
1007                        BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
1008                        mutex_lock(&ss->hierarchy_mutex);
1009                        if (ss->bind)
1010                                ss->bind(ss, dummytop);
1011                        dummytop->subsys[i]->cgroup = dummytop;
1012                        cgrp->subsys[i] = NULL;
1013                        subsys[i]->root = &rootnode;
1014                        list_move(&ss->sibling, &rootnode.subsys_list);
1015                        mutex_unlock(&ss->hierarchy_mutex);
1016                        /* subsystem is now free - drop reference on module */
1017                        module_put(ss->module);
1018                } else if (bit & final_bits) {
1019                        /* Subsystem state should already exist */
1020                        BUG_ON(ss == NULL);
1021                        BUG_ON(!cgrp->subsys[i]);
1022                        /*
1023                         * a refcount was taken, but we already had one, so
1024                         * drop the extra reference.
1025                         */
1026                        module_put(ss->module);
1027#ifdef CONFIG_MODULE_UNLOAD
1028                        BUG_ON(ss->module && !module_refcount(ss->module));
1029#endif
1030                } else {
1031                        /* Subsystem state shouldn't exist */
1032                        BUG_ON(cgrp->subsys[i]);
1033                }
1034        }
1035        root->subsys_bits = root->actual_subsys_bits = final_bits;
1036        synchronize_rcu();
1037
1038        return 0;
1039}
1040
1041static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
1042{
1043        struct cgroupfs_root *root = vfs->mnt_sb->s_fs_info;
1044        struct cgroup_subsys *ss;
1045
1046        mutex_lock(&cgroup_mutex);
1047        for_each_subsys(root, ss)
1048                seq_printf(seq, ",%s", ss->name);
1049        if (test_bit(ROOT_NOPREFIX, &root->flags))
1050                seq_puts(seq, ",noprefix");
1051        if (strlen(root->release_agent_path))
1052                seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1053        if (clone_children(&root->top_cgroup))
1054                seq_puts(seq, ",clone_children");
1055        if (strlen(root->name))
1056                seq_printf(seq, ",name=%s", root->name);
1057        mutex_unlock(&cgroup_mutex);
1058        return 0;
1059}
1060
1061struct cgroup_sb_opts {
1062        unsigned long subsys_bits;
1063        unsigned long flags;
1064        char *release_agent;
1065        bool clone_children;
1066        char *name;
1067        /* User explicitly requested empty subsystem */
1068        bool none;
1069
1070        struct cgroupfs_root *new_root;
1071
1072};
1073
1074/*
1075 * Convert a hierarchy specifier into a bitmask of subsystems and flags. Call
1076 * with cgroup_mutex held to protect the subsys[] array. This function takes
1077 * refcounts on subsystems to be used, unless it returns error, in which case
1078 * no refcounts are taken.
1079 */
1080static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1081{
1082        char *token, *o = data;
1083        bool all_ss = false, one_ss = false;
1084        unsigned long mask = (unsigned long)-1;
1085        int i;
1086        bool module_pin_failed = false;
1087
1088        BUG_ON(!mutex_is_locked(&cgroup_mutex));
1089
1090#ifdef CONFIG_CPUSETS
1091        mask = ~(1UL << cpuset_subsys_id);
1092#endif
1093
1094        memset(opts, 0, sizeof(*opts));
1095
1096        while ((token = strsep(&o, ",")) != NULL) {
1097                if (!*token)
1098                        return -EINVAL;
1099                if (!strcmp(token, "none")) {
1100                        /* Explicitly have no subsystems */
1101                        opts->none = true;
1102                        continue;
1103                }
1104                if (!strcmp(token, "all")) {
1105                        /* Mutually exclusive option 'all' + subsystem name */
1106                        if (one_ss)
1107                                return -EINVAL;
1108                        all_ss = true;
1109                        continue;
1110                }
1111                if (!strcmp(token, "noprefix")) {
1112                        set_bit(ROOT_NOPREFIX, &opts->flags);
1113                        continue;
1114                }
1115                if (!strcmp(token, "clone_children")) {
1116                        opts->clone_children = true;
1117                        continue;
1118                }
1119                if (!strncmp(token, "release_agent=", 14)) {
1120                        /* Specifying two release agents is forbidden */
1121                        if (opts->release_agent)
1122                                return -EINVAL;
1123                        opts->release_agent =
1124                                kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1125                        if (!opts->release_agent)
1126                                return -ENOMEM;
1127                        continue;
1128                }
1129                if (!strncmp(token, "name=", 5)) {
1130                        const char *name = token + 5;
1131                        /* Can't specify an empty name */
1132                        if (!strlen(name))
1133                                return -EINVAL;
1134                        /* Must match [\w.-]+ */
1135                        for (i = 0; i < strlen(name); i++) {
1136                                char c = name[i];
1137                                if (isalnum(c))
1138                                        continue;
1139                                if ((c == '.') || (c == '-') || (c == '_'))
1140                                        continue;
1141                                return -EINVAL;
1142                        }
1143                        /* Specifying two names is forbidden */
1144                        if (opts->name)
1145                                return -EINVAL;
1146                        opts->name = kstrndup(name,
1147                                              MAX_CGROUP_ROOT_NAMELEN - 1,
1148                                              GFP_KERNEL);
1149                        if (!opts->name)
1150                                return -ENOMEM;
1151
1152                        continue;
1153                }
1154
1155                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1156                        struct cgroup_subsys *ss = subsys[i];
1157                        if (ss == NULL)
1158                                continue;
1159                        if (strcmp(token, ss->name))
1160                                continue;
1161                        if (ss->disabled)
1162                                continue;
1163
1164                        /* Mutually exclusive option 'all' + subsystem name */
1165                        if (all_ss)
1166                                return -EINVAL;
1167                        set_bit(i, &opts->subsys_bits);
1168                        one_ss = true;
1169
1170                        break;
1171                }
1172                if (i == CGROUP_SUBSYS_COUNT)
1173                        return -ENOENT;
1174        }
1175
1176        /*
1177         * If the 'all' option was specified select all the subsystems,
1178         * otherwise 'all, 'none' and a subsystem name options were not
1179         * specified, let's default to 'all'
1180         */
1181        if (all_ss || (!all_ss && !one_ss && !opts->none)) {
1182                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
1183                        struct cgroup_subsys *ss = subsys[i];
1184                        if (ss == NULL)
1185                                continue;
1186                        if (ss->disabled)
1187                                continue;
1188                        set_bit(i, &opts->subsys_bits);
1189                }
1190        }
1191
1192        /* Consistency checks */
1193
1194        /*
1195         * Option noprefix was introduced just for backward compatibility
1196         * with the old cpuset, so we allow noprefix only if mounting just
1197         * the cpuset subsystem.
1198         */
1199        if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
1200            (opts->subsys_bits & mask))
1201                return -EINVAL;
1202
1203
1204        /* Can't specify "none" and some subsystems */
1205        if (opts->subsys_bits && opts->none)
1206                return -EINVAL;
1207
1208        /*
1209         * We either have to specify by name or by subsystems. (So all
1210         * empty hierarchies must have a name).
1211         */
1212        if (!opts->subsys_bits && !opts->name)
1213                return -EINVAL;
1214
1215        /*
1216         * Grab references on all the modules we'll need, so the subsystems
1217         * don't dance around before rebind_subsystems attaches them. This may
1218         * take duplicate reference counts on a subsystem that's already used,
1219         * but rebind_subsystems handles this case.
1220         */
1221        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1222                unsigned long bit = 1UL << i;
1223
1224                if (!(bit & opts->subsys_bits))
1225                        continue;
1226                if (!try_module_get(subsys[i]->module)) {
1227                        module_pin_failed = true;
1228                        break;
1229                }
1230        }
1231        if (module_pin_failed) {
1232                /*
1233                 * oops, one of the modules was going away. this means that we
1234                 * raced with a module_delete call, and to the user this is
1235                 * essentially a "subsystem doesn't exist" case.
1236                 */
1237                for (i--; i >= CGROUP_BUILTIN_SUBSYS_COUNT; i--) {
1238                        /* drop refcounts only on the ones we took */
1239                        unsigned long bit = 1UL << i;
1240
1241                        if (!(bit & opts->subsys_bits))
1242                                continue;
1243                        module_put(subsys[i]->module);
1244                }
1245                return -ENOENT;
1246        }
1247
1248        return 0;
1249}
1250
1251static void drop_parsed_module_refcounts(unsigned long subsys_bits)
1252{
1253        int i;
1254        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
1255                unsigned long bit = 1UL << i;
1256
1257                if (!(bit & subsys_bits))
1258                        continue;
1259                module_put(subsys[i]->module);
1260        }
1261}
1262
1263static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1264{
1265        int ret = 0;
1266        struct cgroupfs_root *root = sb->s_fs_info;
1267        struct cgroup *cgrp = &root->top_cgroup;
1268        struct cgroup_sb_opts opts;
1269
1270        mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1271        mutex_lock(&cgroup_mutex);
1272
1273        /* See what subsystems are wanted */
1274        ret = parse_cgroupfs_options(data, &opts);
1275        if (ret)
1276                goto out_unlock;
1277
1278        /* Don't allow flags or name to change at remount */
1279        if (opts.flags != root->flags ||
1280            (opts.name && strcmp(opts.name, root->name))) {
1281                ret = -EINVAL;
1282                drop_parsed_module_refcounts(opts.subsys_bits);
1283                goto out_unlock;
1284        }
1285
1286        ret = rebind_subsystems(root, opts.subsys_bits);
1287        if (ret) {
1288                drop_parsed_module_refcounts(opts.subsys_bits);
1289                goto out_unlock;
1290        }
1291
1292        /* (re)populate subsystem files */
1293        cgroup_populate_dir(cgrp);
1294
1295        if (opts.release_agent)
1296                strcpy(root->release_agent_path, opts.release_agent);
1297 out_unlock:
1298        kfree(opts.release_agent);
1299        kfree(opts.name);
1300        mutex_unlock(&cgroup_mutex);
1301        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1302        return ret;
1303}
1304
1305static const struct super_operations cgroup_ops = {
1306        .statfs = simple_statfs,
1307        .drop_inode = generic_delete_inode,
1308        .show_options = cgroup_show_options,
1309        .remount_fs = cgroup_remount,
1310};
1311
1312static void init_cgroup_housekeeping(struct cgroup *cgrp)
1313{
1314        INIT_LIST_HEAD(&cgrp->sibling);
1315        INIT_LIST_HEAD(&cgrp->children);
1316        INIT_LIST_HEAD(&cgrp->css_sets);
1317        INIT_LIST_HEAD(&cgrp->release_list);
1318        INIT_LIST_HEAD(&cgrp->pidlists);
1319        mutex_init(&cgrp->pidlist_mutex);
1320        INIT_LIST_HEAD(&cgrp->event_list);
1321        spin_lock_init(&cgrp->event_list_lock);
1322}
1323
1324static void init_cgroup_root(struct cgroupfs_root *root)
1325{
1326        struct cgroup *cgrp = &root->top_cgroup;
1327        INIT_LIST_HEAD(&root->subsys_list);
1328        INIT_LIST_HEAD(&root->root_list);
1329        root->number_of_cgroups = 1;
1330        cgrp->root = root;
1331        cgrp->top_cgroup = cgrp;
1332        init_cgroup_housekeeping(cgrp);
1333}
1334
1335static bool init_root_id(struct cgroupfs_root *root)
1336{
1337        int ret = 0;
1338
1339        do {
1340                if (!ida_pre_get(&hierarchy_ida, GFP_KERNEL))
1341                        return false;
1342                spin_lock(&hierarchy_id_lock);
1343                /* Try to allocate the next unused ID */
1344                ret = ida_get_new_above(&hierarchy_ida, next_hierarchy_id,
1345                                        &root->hierarchy_id);
1346                if (ret == -ENOSPC)
1347                        /* Try again starting from 0 */
1348                        ret = ida_get_new(&hierarchy_ida, &root->hierarchy_id);
1349                if (!ret) {
1350                        next_hierarchy_id = root->hierarchy_id + 1;
1351                } else if (ret != -EAGAIN) {
1352                        /* Can only get here if the 31-bit IDR is full ... */
1353                        BUG_ON(ret);
1354                }
1355                spin_unlock(&hierarchy_id_lock);
1356        } while (ret);
1357        return true;
1358}
1359
1360static int cgroup_test_super(struct super_block *sb, void *data)
1361{
1362        struct cgroup_sb_opts *opts = data;
1363        struct cgroupfs_root *root = sb->s_fs_info;
1364
1365        /* If we asked for a name then it must match */
1366        if (opts->name && strcmp(opts->name, root->name))
1367                return 0;
1368
1369        /*
1370         * If we asked for subsystems (or explicitly for no
1371         * subsystems) then they must match
1372         */
1373        if ((opts->subsys_bits || opts->none)
1374            && (opts->subsys_bits != root->subsys_bits))
1375                return 0;
1376
1377        return 1;
1378}
1379
1380static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1381{
1382        struct cgroupfs_root *root;
1383
1384        if (!opts->subsys_bits && !opts->none)
1385                return NULL;
1386
1387        root = kzalloc(sizeof(*root), GFP_KERNEL);
1388        if (!root)
1389                return ERR_PTR(-ENOMEM);
1390
1391        if (!init_root_id(root)) {
1392                kfree(root);
1393                return ERR_PTR(-ENOMEM);
1394        }
1395        init_cgroup_root(root);
1396
1397        root->subsys_bits = opts->subsys_bits;
1398        root->flags = opts->flags;
1399        if (opts->release_agent)
1400                strcpy(root->release_agent_path, opts->release_agent);
1401        if (opts->name)
1402                strcpy(root->name, opts->name);
1403        if (opts->clone_children)
1404                set_bit(CGRP_CLONE_CHILDREN, &root->top_cgroup.flags);
1405        return root;
1406}
1407
1408static void cgroup_drop_root(struct cgroupfs_root *root)
1409{
1410        if (!root)
1411                return;
1412
1413        BUG_ON(!root->hierarchy_id);
1414        spin_lock(&hierarchy_id_lock);
1415        ida_remove(&hierarchy_ida, root->hierarchy_id);
1416        spin_unlock(&hierarchy_id_lock);
1417        kfree(root);
1418}
1419
1420static int cgroup_set_super(struct super_block *sb, void *data)
1421{
1422        int ret;
1423        struct cgroup_sb_opts *opts = data;
1424
1425        /* If we don't have a new root, we can't set up a new sb */
1426        if (!opts->new_root)
1427                return -EINVAL;
1428
1429        BUG_ON(!opts->subsys_bits && !opts->none);
1430
1431        ret = set_anon_super(sb, NULL);
1432        if (ret)
1433                return ret;
1434
1435        sb->s_fs_info = opts->new_root;
1436        opts->new_root->sb = sb;
1437
1438        sb->s_blocksize = PAGE_CACHE_SIZE;
1439        sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1440        sb->s_magic = CGROUP_SUPER_MAGIC;
1441        sb->s_op = &cgroup_ops;
1442
1443        return 0;
1444}
1445
1446static int cgroup_get_rootdir(struct super_block *sb)
1447{
1448        static const struct dentry_operations cgroup_dops = {
1449                .d_iput = cgroup_diput,
1450                .d_delete = cgroup_delete,
1451        };
1452
1453        struct inode *inode =
1454                cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1455        struct dentry *dentry;
1456
1457        if (!inode)
1458                return -ENOMEM;
1459
1460        inode->i_fop = &simple_dir_operations;
1461        inode->i_op = &cgroup_dir_inode_operations;
1462        /* directories start off with i_nlink == 2 (for "." entry) */
1463        inc_nlink(inode);
1464        dentry = d_alloc_root(inode);
1465        if (!dentry) {
1466                iput(inode);
1467                return -ENOMEM;
1468        }
1469        sb->s_root = dentry;
1470        /* for everything else we want ->d_op set */
1471        sb->s_d_op = &cgroup_dops;
1472        return 0;
1473}
1474
1475static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1476                         int flags, const char *unused_dev_name,
1477                         void *data)
1478{
1479        struct cgroup_sb_opts opts;
1480        struct cgroupfs_root *root;
1481        int ret = 0;
1482        struct super_block *sb;
1483        struct cgroupfs_root *new_root;
1484
1485        /* First find the desired set of subsystems */
1486        mutex_lock(&cgroup_mutex);
1487        ret = parse_cgroupfs_options(data, &opts);
1488        mutex_unlock(&cgroup_mutex);
1489        if (ret)
1490                goto out_err;
1491
1492        /*
1493         * Allocate a new cgroup root. We may not need it if we're
1494         * reusing an existing hierarchy.
1495         */
1496        new_root = cgroup_root_from_opts(&opts);
1497        if (IS_ERR(new_root)) {
1498                ret = PTR_ERR(new_root);
1499                goto drop_modules;
1500        }
1501        opts.new_root = new_root;
1502
1503        /* Locate an existing or new sb for this hierarchy */
1504        sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
1505        if (IS_ERR(sb)) {
1506                ret = PTR_ERR(sb);
1507                cgroup_drop_root(opts.new_root);
1508                goto drop_modules;
1509        }
1510
1511        root = sb->s_fs_info;
1512        BUG_ON(!root);
1513        if (root == opts.new_root) {
1514                /* We used the new root structure, so this is a new hierarchy */
1515                struct list_head tmp_cg_links;
1516                struct cgroup *root_cgrp = &root->top_cgroup;
1517                struct inode *inode;
1518                struct cgroupfs_root *existing_root;
1519                const struct cred *cred;
1520                int i;
1521
1522                BUG_ON(sb->s_root != NULL);
1523
1524                ret = cgroup_get_rootdir(sb);
1525                if (ret)
1526                        goto drop_new_super;
1527                inode = sb->s_root->d_inode;
1528
1529                mutex_lock(&inode->i_mutex);
1530                mutex_lock(&cgroup_mutex);
1531
1532                if (strlen(root->name)) {
1533                        /* Check for name clashes with existing mounts */
1534                        for_each_active_root(existing_root) {
1535                                if (!strcmp(existing_root->name, root->name)) {
1536                                        ret = -EBUSY;
1537                                        mutex_unlock(&cgroup_mutex);
1538                                        mutex_unlock(&inode->i_mutex);
1539                                        goto drop_new_super;
1540                                }
1541                        }
1542                }
1543
1544                /*
1545                 * We're accessing css_set_count without locking
1546                 * css_set_lock here, but that's OK - it can only be
1547                 * increased by someone holding cgroup_lock, and
1548                 * that's us. The worst that can happen is that we
1549                 * have some link structures left over
1550                 */
1551                ret = allocate_cg_links(css_set_count, &tmp_cg_links);
1552                if (ret) {
1553                        mutex_unlock(&cgroup_mutex);
1554                        mutex_unlock(&inode->i_mutex);
1555                        goto drop_new_super;
1556                }
1557
1558                ret = rebind_subsystems(root, root->subsys_bits);
1559                if (ret == -EBUSY) {
1560                        mutex_unlock(&cgroup_mutex);
1561                        mutex_unlock(&inode->i_mutex);
1562                        free_cg_links(&tmp_cg_links);
1563                        goto drop_new_super;
1564                }
1565                /*
1566                 * There must be no failure case after here, since rebinding
1567                 * takes care of subsystems' refcounts, which are explicitly
1568                 * dropped in the failure exit path.
1569                 */
1570
1571                /* EBUSY should be the only error here */
1572                BUG_ON(ret);
1573
1574                list_add(&root->root_list, &roots);
1575                root_count++;
1576
1577                sb->s_root->d_fsdata = root_cgrp;
1578                root->top_cgroup.dentry = sb->s_root;
1579
1580                /* Link the top cgroup in this hierarchy into all
1581                 * the css_set objects */
1582                write_lock(&css_set_lock);
1583                for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
1584                        struct hlist_head *hhead = &css_set_table[i];
1585                        struct hlist_node *node;
1586                        struct css_set *cg;
1587
1588                        hlist_for_each_entry(cg, node, hhead, hlist)
1589                                link_css_set(&tmp_cg_links, cg, root_cgrp);
1590                }
1591                write_unlock(&css_set_lock);
1592
1593                free_cg_links(&tmp_cg_links);
1594
1595                BUG_ON(!list_empty(&root_cgrp->sibling));
1596                BUG_ON(!list_empty(&root_cgrp->children));
1597                BUG_ON(root->number_of_cgroups != 1);
1598
1599                cred = override_creds(&init_cred);
1600                cgroup_populate_dir(root_cgrp);
1601                revert_creds(cred);
1602                mutex_unlock(&cgroup_mutex);
1603                mutex_unlock(&inode->i_mutex);
1604        } else {
1605                /*
1606                 * We re-used an existing hierarchy - the new root (if
1607                 * any) is not needed
1608                 */
1609                cgroup_drop_root(opts.new_root);
1610                /* no subsys rebinding, so refcounts don't change */
1611                drop_parsed_module_refcounts(opts.subsys_bits);
1612        }
1613
1614        kfree(opts.release_agent);
1615        kfree(opts.name);
1616        return dget(sb->s_root);
1617
1618 drop_new_super:
1619        deactivate_locked_super(sb);
1620 drop_modules:
1621        drop_parsed_module_refcounts(opts.subsys_bits);
1622 out_err:
1623        kfree(opts.release_agent);
1624        kfree(opts.name);
1625        return ERR_PTR(ret);
1626}
1627
1628static void cgroup_kill_sb(struct super_block *sb) {
1629        struct cgroupfs_root *root = sb->s_fs_info;
1630        struct cgroup *cgrp = &root->top_cgroup;
1631        int ret;
1632        struct cg_cgroup_link *link;
1633        struct cg_cgroup_link *saved_link;
1634
1635        BUG_ON(!root);
1636
1637        BUG_ON(root->number_of_cgroups != 1);
1638        BUG_ON(!list_empty(&cgrp->children));
1639        BUG_ON(!list_empty(&cgrp->sibling));
1640
1641        mutex_lock(&cgroup_mutex);
1642
1643        /* Rebind all subsystems back to the default hierarchy */
1644        ret = rebind_subsystems(root, 0);
1645        /* Shouldn't be able to fail ... */
1646        BUG_ON(ret);
1647
1648        /*
1649         * Release all the links from css_sets to this hierarchy's
1650         * root cgroup
1651         */
1652        write_lock(&css_set_lock);
1653
1654        list_for_each_entry_safe(link, saved_link, &cgrp->css_sets,
1655                                 cgrp_link_list) {
1656                list_del(&link->cg_link_list);
1657                list_del(&link->cgrp_link_list);
1658                kfree(link);
1659        }
1660        write_unlock(&css_set_lock);
1661
1662        if (!list_empty(&root->root_list)) {
1663                list_del(&root->root_list);
1664                root_count--;
1665        }
1666
1667        mutex_unlock(&cgroup_mutex);
1668
1669        kill_litter_super(sb);
1670        cgroup_drop_root(root);
1671}
1672
1673static struct file_system_type cgroup_fs_type = {
1674        .name = "cgroup",
1675        .mount = cgroup_mount,
1676        .kill_sb = cgroup_kill_sb,
1677};
1678
1679static struct kobject *cgroup_kobj;
1680
1681static inline struct cgroup *__d_cgrp(struct dentry *dentry)
1682{
1683        return dentry->d_fsdata;
1684}
1685
1686static inline struct cftype *__d_cft(struct dentry *dentry)
1687{
1688        return dentry->d_fsdata;
1689}
1690
1691/**
1692 * cgroup_path - generate the path of a cgroup
1693 * @cgrp: the cgroup in question
1694 * @buf: the buffer to write the path into
1695 * @buflen: the length of the buffer
1696 *
1697 * Called with cgroup_mutex held or else with an RCU-protected cgroup
1698 * reference.  Writes path of cgroup into buf.  Returns 0 on success,
1699 * -errno on error.
1700 */
1701int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1702{
1703        char *start;
1704        struct dentry *dentry = rcu_dereference_check(cgrp->dentry,
1705                                                      cgroup_lock_is_held());
1706
1707        if (!dentry || cgrp == dummytop) {
1708                /*
1709                 * Inactive subsystems have no dentry for their root
1710                 * cgroup
1711                 */
1712                strcpy(buf, "/");
1713                return 0;
1714        }
1715
1716        start = buf + buflen;
1717
1718        *--start = '\0';
1719        for (;;) {
1720                int len = dentry->d_name.len;
1721
1722                if ((start -= len) < buf)
1723                        return -ENAMETOOLONG;
1724                memcpy(start, dentry->d_name.name, len);
1725                cgrp = cgrp->parent;
1726                if (!cgrp)
1727                        break;
1728
1729                dentry = rcu_dereference_check(cgrp->dentry,
1730                                               cgroup_lock_is_held());
1731                if (!cgrp->parent)
1732                        continue;
1733                if (--start < buf)
1734                        return -ENAMETOOLONG;
1735                *start = '/';
1736        }
1737        memmove(buf, start, buf + buflen - start);
1738        return 0;
1739}
1740EXPORT_SYMBOL_GPL(cgroup_path);
1741
1742/*
1743 * cgroup_task_migrate - move a task from one cgroup to another.
1744 *
1745 * 'guarantee' is set if the caller promises that a new css_set for the task
1746 * will already exist. If not set, this function might sleep, and can fail with
1747 * -ENOMEM. Otherwise, it can only fail with -ESRCH.
1748 */
1749static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
1750                               struct task_struct *tsk, bool guarantee)
1751{
1752        struct css_set *oldcg;
1753        struct css_set *newcg;
1754
1755        /*
1756         * get old css_set. we need to take task_lock and refcount it, because
1757         * an exiting task can change its css_set to init_css_set and drop its
1758         * old one without taking cgroup_mutex.
1759         */
1760        task_lock(tsk);
1761        oldcg = tsk->cgroups;
1762        get_css_set(oldcg);
1763        task_unlock(tsk);
1764
1765        /* locate or allocate a new css_set for this task. */
1766        if (guarantee) {
1767                /* we know the css_set we want already exists. */
1768                struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
1769                read_lock(&css_set_lock);
1770                newcg = find_existing_css_set(oldcg, cgrp, template);
1771                BUG_ON(!newcg);
1772                get_css_set(newcg);
1773                read_unlock(&css_set_lock);
1774        } else {
1775                might_sleep();
1776                /* find_css_set will give us newcg already referenced. */
1777                newcg = find_css_set(oldcg, cgrp);
1778                if (!newcg) {
1779                        put_css_set(oldcg);
1780                        return -ENOMEM;
1781                }
1782        }
1783        put_css_set(oldcg);
1784
1785        /* if PF_EXITING is set, the tsk->cgroups pointer is no longer safe. */
1786        task_lock(tsk);
1787        if (tsk->flags & PF_EXITING) {
1788                task_unlock(tsk);
1789                put_css_set(newcg);
1790                return -ESRCH;
1791        }
1792        rcu_assign_pointer(tsk->cgroups, newcg);
1793        task_unlock(tsk);
1794
1795        /* Update the css_set linked lists if we're using them */
1796        write_lock(&css_set_lock);
1797        if (!list_empty(&tsk->cg_list))
1798                list_move(&tsk->cg_list, &newcg->tasks);
1799        write_unlock(&css_set_lock);
1800
1801        /*
1802         * We just gained a reference on oldcg by taking it from the task. As
1803         * trading it for newcg is protected by cgroup_mutex, we're safe to drop
1804         * it here; it will be freed under RCU.
1805         */
1806        put_css_set(oldcg);
1807
1808        set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
1809        return 0;
1810}
1811
1812/**
1813 * cgroup_attach_task - attach task 'tsk' to cgroup 'cgrp'
1814 * @cgrp: the cgroup the task is attaching to
1815 * @tsk: the task to be attached
1816 *
1817 * Call holding cgroup_mutex. May take task_lock of
1818 * the task 'tsk' during call.
1819 */
1820int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
1821{
1822        int retval;
1823        struct cgroup_subsys *ss, *failed_ss = NULL;
1824        struct cgroup *oldcgrp;
1825        struct cgroupfs_root *root = cgrp->root;
1826
1827        /* Nothing to do if the task is already in that cgroup */
1828        oldcgrp = task_cgroup_from_root(tsk, root);
1829        if (cgrp == oldcgrp)
1830                return 0;
1831
1832        for_each_subsys(root, ss) {
1833                if (ss->can_attach) {
1834                        retval = ss->can_attach(ss, cgrp, tsk);
1835                        if (retval) {
1836                                /*
1837                                 * Remember on which subsystem the can_attach()
1838                                 * failed, so that we only call cancel_attach()
1839                                 * against the subsystems whose can_attach()
1840                                 * succeeded. (See below)
1841                                 */
1842                                failed_ss = ss;
1843                                goto out;
1844                        }
1845                }
1846                if (ss->can_attach_task) {
1847                        retval = ss->can_attach_task(cgrp, tsk);
1848                        if (retval) {
1849                                failed_ss = ss;
1850                                goto out;
1851                        }
1852                }
1853        }
1854
1855        retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, false);
1856        if (retval)
1857                goto out;
1858
1859        for_each_subsys(root, ss) {
1860                if (ss->pre_attach)
1861                        ss->pre_attach(cgrp);
1862                if (ss->attach_task)
1863                        ss->attach_task(cgrp, tsk);
1864                if (ss->attach)
1865                        ss->attach(ss, cgrp, oldcgrp, tsk);
1866        }
1867
1868        synchronize_rcu();
1869
1870        /*
1871         * wake up rmdir() waiter. the rmdir should fail since the cgroup
1872         * is no longer empty.
1873         */
1874        cgroup_wakeup_rmdir_waiter(cgrp);
1875out:
1876        if (retval) {
1877                for_each_subsys(root, ss) {
1878                        if (ss == failed_ss)
1879                                /*
1880                                 * This subsystem was the one that failed the
1881                                 * can_attach() check earlier, so we don't need
1882                                 * to call cancel_attach() against it or any
1883                                 * remaining subsystems.
1884                                 */
1885                                break;
1886                        if (ss->cancel_attach)
1887                                ss->cancel_attach(ss, cgrp, tsk);
1888                }
1889        }
1890        return retval;
1891}
1892
1893/**
1894 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
1895 * @from: attach to all cgroups of a given task
1896 * @tsk: the task to be attached
1897 */
1898int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
1899{
1900        struct cgroupfs_root *root;
1901        int retval = 0;
1902
1903        cgroup_lock();
1904        for_each_active_root(root) {
1905                struct cgroup *from_cg = task_cgroup_from_root(from, root);
1906
1907                retval = cgroup_attach_task(from_cg, tsk);
1908                if (retval)
1909                        break;
1910        }
1911        cgroup_unlock();
1912
1913        return retval;
1914}
1915EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
1916
1917/*
1918 * cgroup_attach_proc works in two stages, the first of which prefetches all
1919 * new css_sets needed (to make sure we have enough memory before committing
1920 * to the move) and stores them in a list of entries of the following type.
1921 * TODO: possible optimization: use css_set->rcu_head for chaining instead
1922 */
1923struct cg_list_entry {
1924        struct css_set *cg;
1925        struct list_head links;
1926};
1927
1928static bool css_set_check_fetched(struct cgroup *cgrp,
1929                                  struct task_struct *tsk, struct css_set *cg,
1930                                  struct list_head *newcg_list)
1931{
1932        struct css_set *newcg;
1933        struct cg_list_entry *cg_entry;
1934        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
1935
1936        read_lock(&css_set_lock);
1937        newcg = find_existing_css_set(cg, cgrp, template);
1938        if (newcg)
1939                get_css_set(newcg);
1940        read_unlock(&css_set_lock);
1941
1942        /* doesn't exist at all? */
1943        if (!newcg)
1944                return false;
1945        /* see if it's already in the list */
1946        list_for_each_entry(cg_entry, newcg_list, links) {
1947                if (cg_entry->cg == newcg) {
1948                        put_css_set(newcg);
1949                        return true;
1950                }
1951        }
1952
1953        /* not found */
1954        put_css_set(newcg);
1955        return false;
1956}
1957
1958/*
1959 * Find the new css_set and store it in the list in preparation for moving the
1960 * given task to the given cgroup. Returns 0 or -ENOMEM.
1961 */
1962static int css_set_prefetch(struct cgroup *cgrp, struct css_set *cg,
1963                            struct list_head *newcg_list)
1964{
1965        struct css_set *newcg;
1966        struct cg_list_entry *cg_entry;
1967
1968        /* ensure a new css_set will exist for this thread */
1969        newcg = find_css_set(cg, cgrp);
1970        if (!newcg)
1971                return -ENOMEM;
1972        /* add it to the list */
1973        cg_entry = kmalloc(sizeof(struct cg_list_entry), GFP_KERNEL);
1974        if (!cg_entry) {
1975                put_css_set(newcg);
1976                return -ENOMEM;
1977        }
1978        cg_entry->cg = newcg;
1979        list_add(&cg_entry->links, newcg_list);
1980        return 0;
1981}
1982
1983/**
1984 * cgroup_attach_proc - attach all threads in a threadgroup to a cgroup
1985 * @cgrp: the cgroup to attach to
1986 * @leader: the threadgroup leader task_struct of the group to be attached
1987 *
1988 * Call holding cgroup_mutex and the threadgroup_fork_lock of the leader. Will
1989 * take task_lock of each thread in leader's threadgroup individually in turn.
1990 */
1991int cgroup_attach_proc(struct cgroup *cgrp, struct task_struct *leader)
1992{
1993        int retval, i, group_size;
1994        struct cgroup_subsys *ss, *failed_ss = NULL;
1995        bool cancel_failed_ss = false;
1996        /* guaranteed to be initialized later, but the compiler needs this */
1997        struct cgroup *oldcgrp = NULL;
1998        struct css_set *oldcg;
1999        struct cgroupfs_root *root = cgrp->root;
2000        /* threadgroup list cursor and array */
2001        struct task_struct *tsk;
2002        struct flex_array *group;
2003        /*
2004         * we need to make sure we have css_sets for all the tasks we're
2005         * going to move -before- we actually start moving them, so that in
2006         * case we get an ENOMEM we can bail out before making any changes.
2007         */
2008        struct list_head newcg_list;
2009        struct cg_list_entry *cg_entry, *temp_nobe;
2010
2011        /*
2012         * step 0: in order to do expensive, possibly blocking operations for
2013         * every thread, we cannot iterate the thread group list, since it needs
2014         * rcu or tasklist locked. instead, build an array of all threads in the
2015         * group - threadgroup_fork_lock prevents new threads from appearing,
2016         * and if threads exit, this will just be an over-estimate.
2017         */
2018        group_size = get_nr_threads(leader);
2019        /* flex_array supports very large thread-groups better than kmalloc. */
2020        group = flex_array_alloc(sizeof(struct task_struct *), group_size,
2021                                 GFP_KERNEL);
2022        if (!group)
2023                return -ENOMEM;
2024        /* pre-allocate to guarantee space while iterating in rcu read-side. */
2025        retval = flex_array_prealloc(group, 0, group_size - 1, GFP_KERNEL);
2026        if (retval)
2027                goto out_free_group_list;
2028
2029        /* prevent changes to the threadgroup list while we take a snapshot. */
2030        rcu_read_lock();
2031        if (!thread_group_leader(leader)) {
2032                /*
2033                 * a race with de_thread from another thread's exec() may strip
2034                 * us of our leadership, making while_each_thread unsafe to use
2035                 * on this task. if this happens, there is no choice but to
2036                 * throw this task away and try again (from cgroup_procs_write);
2037                 * this is "double-double-toil-and-trouble-check locking".
2038                 */
2039                rcu_read_unlock();
2040                retval = -EAGAIN;
2041                goto out_free_group_list;
2042        }
2043        /* take a reference on each task in the group to go in the array. */
2044        tsk = leader;
2045        i = 0;
2046        do {
2047                /* as per above, nr_threads may decrease, but not increase. */
2048                BUG_ON(i >= group_size);
2049                get_task_struct(tsk);
2050                /*
2051                 * saying GFP_ATOMIC has no effect here because we did prealloc
2052                 * earlier, but it's good form to communicate our expectations.
2053                 */
2054                retval = flex_array_put_ptr(group, i, tsk, GFP_ATOMIC);
2055                BUG_ON(retval != 0);
2056                i++;
2057        } while_each_thread(leader, tsk);
2058        /* remember the number of threads in the array for later. */
2059        group_size = i;
2060        rcu_read_unlock();
2061
2062        /*
2063         * step 1: check that we can legitimately attach to the cgroup.
2064         */
2065        for_each_subsys(root, ss) {
2066                if (ss->can_attach) {
2067                        retval = ss->can_attach(ss, cgrp, leader);
2068                        if (retval) {
2069                                failed_ss = ss;
2070                                goto out_cancel_attach;
2071                        }
2072                }
2073                /* a callback to be run on every thread in the threadgroup. */
2074                if (ss->can_attach_task) {
2075                        /* run on each task in the threadgroup. */
2076                        for (i = 0; i < group_size; i++) {
2077                                tsk = flex_array_get_ptr(group, i);
2078                                retval = ss->can_attach_task(cgrp, tsk);
2079                                if (retval) {
2080                                        failed_ss = ss;
2081                                        cancel_failed_ss = true;
2082                                        goto out_cancel_attach;
2083                                }
2084                        }
2085                }
2086        }
2087
2088        /*
2089         * step 2: make sure css_sets exist for all threads to be migrated.
2090         * we use find_css_set, which allocates a new one if necessary.
2091         */
2092        INIT_LIST_HEAD(&newcg_list);
2093        for (i = 0; i < group_size; i++) {
2094                tsk = flex_array_get_ptr(group, i);
2095                /* nothing to do if this task is already in the cgroup */
2096                oldcgrp = task_cgroup_from_root(tsk, root);
2097                if (cgrp == oldcgrp)
2098                        continue;
2099                /* get old css_set pointer */
2100                task_lock(tsk);
2101                if (tsk->flags & PF_EXITING) {
2102                        /* ignore this task if it's going away */
2103                        task_unlock(tsk);
2104                        continue;
2105                }
2106                oldcg = tsk->cgroups;
2107                get_css_set(oldcg);
2108                task_unlock(tsk);
2109                /* see if the new one for us is already in the list? */
2110                if (css_set_check_fetched(cgrp, tsk, oldcg, &newcg_list)) {
2111                        /* was already there, nothing to do. */
2112                        put_css_set(oldcg);
2113                } else {
2114                        /* we don't already have it. get new one. */
2115                        retval = css_set_prefetch(cgrp, oldcg, &newcg_list);
2116                        put_css_set(oldcg);
2117                        if (retval)
2118                                goto out_list_teardown;
2119                }
2120        }
2121
2122        /*
2123         * step 3: now that we're guaranteed success wrt the css_sets, proceed
2124         * to move all tasks to the new cgroup, calling ss->attach_task for each
2125         * one along the way. there are no failure cases after here, so this is
2126         * the commit point.
2127         */
2128        for_each_subsys(root, ss) {
2129                if (ss->pre_attach)
2130                        ss->pre_attach(cgrp);
2131        }
2132        for (i = 0; i < group_size; i++) {
2133                tsk = flex_array_get_ptr(group, i);
2134                /* leave current thread as it is if it's already there */
2135                oldcgrp = task_cgroup_from_root(tsk, root);
2136                if (cgrp == oldcgrp)
2137                        continue;
2138                /* attach each task to each subsystem */
2139                for_each_subsys(root, ss) {
2140                        if (ss->attach_task)
2141                                ss->attach_task(cgrp, tsk);
2142                }
2143                /* if the thread is PF_EXITING, it can just get skipped. */
2144                retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, true);
2145                BUG_ON(retval != 0 && retval != -ESRCH);
2146        }
2147        /* nothing is sensitive to fork() after this point. */
2148
2149        /*
2150         * step 4: do expensive, non-thread-specific subsystem callbacks.
2151         * TODO: if ever a subsystem needs to know the oldcgrp for each task
2152         * being moved, this call will need to be reworked to communicate that.
2153         */
2154        for_each_subsys(root, ss) {
2155                if (ss->attach)
2156                        ss->attach(ss, cgrp, oldcgrp, leader);
2157        }
2158
2159        /*
2160         * step 5: success! and cleanup
2161         */
2162        synchronize_rcu();
2163        cgroup_wakeup_rmdir_waiter(cgrp);
2164        retval = 0;
2165out_list_teardown:
2166        /* clean up the list of prefetched css_sets. */
2167        list_for_each_entry_safe(cg_entry, temp_nobe, &newcg_list, links) {
2168                list_del(&cg_entry->links);
2169                put_css_set(cg_entry->cg);
2170                kfree(cg_entry);
2171        }
2172out_cancel_attach:
2173        /* same deal as in cgroup_attach_task */
2174        if (retval) {
2175                for_each_subsys(root, ss) {
2176                        if (ss == failed_ss) {
2177                                if (cancel_failed_ss && ss->cancel_attach)
2178                                        ss->cancel_attach(ss, cgrp, leader);
2179                                break;
2180                        }
2181                        if (ss->cancel_attach)
2182                                ss->cancel_attach(ss, cgrp, leader);
2183                }
2184        }
2185        /* clean up the array of referenced threads in the group. */
2186        for (i = 0; i < group_size; i++) {
2187                tsk = flex_array_get_ptr(group, i);
2188                put_task_struct(tsk);
2189        }
2190out_free_group_list:
2191        flex_array_free(group);
2192        return retval;
2193}
2194
2195/*
2196 * Find the task_struct of the task to attach by vpid and pass it along to the
2197 * function to attach either it or all tasks in its threadgroup. Will take
2198 * cgroup_mutex; may take task_lock of task.
2199 */
2200static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2201{
2202        struct task_struct *tsk;
2203        const struct cred *cred = current_cred(), *tcred;
2204        int ret;
2205
2206        if (!cgroup_lock_live_group(cgrp))
2207                return -ENODEV;
2208
2209        if (pid) {
2210                rcu_read_lock();
2211                tsk = find_task_by_vpid(pid);
2212                if (!tsk) {
2213                        rcu_read_unlock();
2214                        cgroup_unlock();
2215                        return -ESRCH;
2216                }
2217                if (threadgroup) {
2218                        /*
2219                         * RCU protects this access, since tsk was found in the
2220                         * tid map. a race with de_thread may cause group_leader
2221                         * to stop being the leader, but cgroup_attach_proc will
2222                         * detect it later.
2223                         */
2224                        tsk = tsk->group_leader;
2225                } else if (tsk->flags & PF_EXITING) {
2226                        /* optimization for the single-task-only case */
2227                        rcu_read_unlock();
2228                        cgroup_unlock();
2229                        return -ESRCH;
2230                }
2231
2232                /*
2233                 * even if we're attaching all tasks in the thread group, we
2234                 * only need to check permissions on one of them.
2235                 */
2236                tcred = __task_cred(tsk);
2237                if (cred->euid &&
2238                    cred->euid != tcred->uid &&
2239                    cred->euid != tcred->suid) {
2240                        rcu_read_unlock();
2241                        cgroup_unlock();
2242                        return -EACCES;
2243                }
2244                get_task_struct(tsk);
2245                rcu_read_unlock();
2246        } else {
2247                if (threadgroup)
2248                        tsk = current->group_leader;
2249                else
2250                        tsk = current;
2251                get_task_struct(tsk);
2252        }
2253
2254        if (threadgroup) {
2255                threadgroup_fork_write_lock(tsk);
2256                ret = cgroup_attach_proc(cgrp, tsk);
2257                threadgroup_fork_write_unlock(tsk);
2258        } else {
2259                ret = cgroup_attach_task(cgrp, tsk);
2260        }
2261        put_task_struct(tsk);
2262        cgroup_unlock();
2263        return ret;
2264}
2265
2266static int cgroup_tasks_write(struct cgroup *cgrp, struct cftype *cft, u64 pid)
2267{
2268        return attach_task_by_pid(cgrp, pid, false);
2269}
2270
2271static int cgroup_procs_write(struct cgroup *cgrp, struct cftype *cft, u64 tgid)
2272{
2273        int ret;
2274        do {
2275                /*
2276                 * attach_proc fails with -EAGAIN if threadgroup leadership
2277                 * changes in the middle of the operation, in which case we need
2278                 * to find the task_struct for the new leader and start over.
2279                 */
2280                ret = attach_task_by_pid(cgrp, tgid, true);
2281        } while (ret == -EAGAIN);
2282        return ret;
2283}
2284
2285/**
2286 * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
2287 * @cgrp: the cgroup to be checked for liveness
2288 *
2289 * On success, returns true; the lock should be later released with
2290 * cgroup_unlock(). On failure returns false with no lock held.
2291 */
2292bool cgroup_lock_live_group(struct cgroup *cgrp)
2293{
2294        mutex_lock(&cgroup_mutex);
2295        if (cgroup_is_removed(cgrp)) {
2296                mutex_unlock(&cgroup_mutex);
2297                return false;
2298        }
2299        return true;
2300}
2301EXPORT_SYMBOL_GPL(cgroup_lock_live_group);
2302
2303static int cgroup_release_agent_write(struct cgroup *cgrp, struct cftype *cft,
2304                                      const char *buffer)
2305{
2306        BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2307        if (strlen(buffer) >= PATH_MAX)
2308                return -EINVAL;
2309        if (!cgroup_lock_live_group(cgrp))
2310                return -ENODEV;
2311        strcpy(cgrp->root->release_agent_path, buffer);
2312        cgroup_unlock();
2313        return 0;
2314}
2315
2316static int cgroup_release_agent_show(struct cgroup *cgrp, struct cftype *cft,
2317                                     struct seq_file *seq)
2318{
2319        if (!cgroup_lock_live_group(cgrp))
2320                return -ENODEV;
2321        seq_puts(seq, cgrp->root->release_agent_path);
2322        seq_putc(seq, '\n');
2323        cgroup_unlock();
2324        return 0;
2325}
2326
2327/* A buffer size big enough for numbers or short strings */
2328#define CGROUP_LOCAL_BUFFER_SIZE 64
2329
2330static ssize_t cgroup_write_X64(struct cgroup *cgrp, struct cftype *cft,
2331                                struct file *file,
2332                                const char __user *userbuf,
2333                                size_t nbytes, loff_t *unused_ppos)
2334{
2335        char buffer[CGROUP_LOCAL_BUFFER_SIZE];
2336        int retval = 0;
2337        char *end;
2338
2339        if (!nbytes)
2340                return -EINVAL;
2341        if (nbytes >= sizeof(buffer))
2342                return -E2BIG;
2343        if (copy_from_user(buffer, userbuf, nbytes))
2344                return -EFAULT;
2345
2346        buffer[nbytes] = 0;     /* nul-terminate */
2347        if (cft->write_u64) {
2348                u64 val = simple_strtoull(strstrip(buffer), &end, 0);
2349                if (*end)
2350                        return -EINVAL;
2351                retval = cft->write_u64(cgrp, cft, val);
2352        } else {
2353                s64 val = simple_strtoll(strstrip(buffer), &end, 0);
2354                if (*end)
2355                        return -EINVAL;
2356                retval = cft->write_s64(cgrp, cft, val);
2357        }
2358        if (!retval)
2359                retval = nbytes;
2360        return retval;
2361}
2362
2363static ssize_t cgroup_write_string(struct cgroup *cgrp, struct cftype *cft,
2364                                   struct file *file,
2365                                   const char __user *userbuf,
2366                                   size_t nbytes, loff_t *unused_ppos)
2367{
2368        char local_buffer[CGROUP_LOCAL_BUFFER_SIZE];
2369        int retval = 0;
2370        size_t max_bytes = cft->max_write_len;
2371        char *buffer = local_buffer;
2372
2373        if (!max_bytes)
2374                max_bytes = sizeof(local_buffer) - 1;
2375        if (nbytes >= max_bytes)
2376                return -E2BIG;
2377        /* Allocate a dynamic buffer if we need one */
2378        if (nbytes >= sizeof(local_buffer)) {
2379                buffer = kmalloc(nbytes + 1, GFP_KERNEL);
2380                if (buffer == NULL)
2381                        return -ENOMEM;
2382        }
2383        if (nbytes && copy_from_user(buffer, userbuf, nbytes)) {
2384                retval = -EFAULT;
2385                goto out;
2386        }
2387
2388        buffer[nbytes] = 0;     /* nul-terminate */
2389        retval = cft->write_string(cgrp, cft, strstrip(buffer));
2390        if (!retval)
2391                retval = nbytes;
2392out:
2393        if (buffer != local_buffer)
2394                kfree(buffer);
2395        return retval;
2396}
2397
2398static ssize_t cgroup_file_write(struct file *file, const char __user *buf,
2399                                                size_t nbytes, loff_t *ppos)
2400{
2401        struct cftype *cft = __d_cft(file->f_dentry);
2402        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2403
2404        if (cgroup_is_removed(cgrp))
2405                return -ENODEV;
2406        if (cft->write)
2407                return cft->write(cgrp, cft, file, buf, nbytes, ppos);
2408        if (cft->write_u64 || cft->write_s64)
2409                return cgroup_write_X64(cgrp, cft, file, buf, nbytes, ppos);
2410        if (cft->write_string)
2411                return cgroup_write_string(cgrp, cft, file, buf, nbytes, ppos);
2412        if (cft->trigger) {
2413                int ret = cft->trigger(cgrp, (unsigned int)cft->private);
2414                return ret ? ret : nbytes;
2415        }
2416        return -EINVAL;
2417}
2418
2419static ssize_t cgroup_read_u64(struct cgroup *cgrp, struct cftype *cft,
2420                               struct file *file,
2421                               char __user *buf, size_t nbytes,
2422                               loff_t *ppos)
2423{
2424        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2425        u64 val = cft->read_u64(cgrp, cft);
2426        int len = sprintf(tmp, "%llu\n", (unsigned long long) val);
2427
2428        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2429}
2430
2431static ssize_t cgroup_read_s64(struct cgroup *cgrp, struct cftype *cft,
2432                               struct file *file,
2433                               char __user *buf, size_t nbytes,
2434                               loff_t *ppos)
2435{
2436        char tmp[CGROUP_LOCAL_BUFFER_SIZE];
2437        s64 val = cft->read_s64(cgrp, cft);
2438        int len = sprintf(tmp, "%lld\n", (long long) val);
2439
2440        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
2441}
2442
2443static ssize_t cgroup_file_read(struct file *file, char __user *buf,
2444                                   size_t nbytes, loff_t *ppos)
2445{
2446        struct cftype *cft = __d_cft(file->f_dentry);
2447        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
2448
2449        if (cgroup_is_removed(cgrp))
2450                return -ENODEV;
2451
2452        if (cft->read)
2453                return cft->read(cgrp, cft, file, buf, nbytes, ppos);
2454        if (cft->read_u64)
2455                return cgroup_read_u64(cgrp, cft, file, buf, nbytes, ppos);
2456        if (cft->read_s64)
2457                return cgroup_read_s64(cgrp, cft, file, buf, nbytes, ppos);
2458        return -EINVAL;
2459}
2460
2461/*
2462 * seqfile ops/methods for returning structured data. Currently just
2463 * supports string->u64 maps, but can be extended in future.
2464 */
2465
2466struct cgroup_seqfile_state {
2467        struct cftype *cft;
2468        struct cgroup *cgroup;
2469};
2470
2471static int cgroup_map_add(struct cgroup_map_cb *cb, const char *key, u64 value)
2472{
2473        struct seq_file *sf = cb->state;
2474        return seq_printf(sf, "%s %llu\n", key, (unsigned long long)value);
2475}
2476
2477static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2478{
2479        struct cgroup_seqfile_state *state = m->private;
2480        struct cftype *cft = state->cft;
2481        if (cft->read_map) {
2482                struct cgroup_map_cb cb = {
2483                        .fill = cgroup_map_add,
2484                        .state = m,
2485                };
2486                return cft->read_map(state->cgroup, cft, &cb);
2487        }
2488        return cft->read_seq_string(state->cgroup, cft, m);
2489}
2490
2491static int cgroup_seqfile_release(struct inode *inode, struct file *file)
2492{
2493        struct seq_file *seq = file->private_data;
2494        kfree(seq->private);
2495        return single_release(inode, file);
2496}
2497
2498static const struct file_operations cgroup_seqfile_operations = {
2499        .read = seq_read,
2500        .write = cgroup_file_write,
2501        .llseek = seq_lseek,
2502        .release = cgroup_seqfile_release,
2503};
2504
2505static int cgroup_file_open(struct inode *inode, struct file *file)
2506{
2507        int err;
2508        struct cftype *cft;
2509
2510        err = generic_file_open(inode, file);
2511        if (err)
2512                return err;
2513        cft = __d_cft(file->f_dentry);
2514
2515        if (cft->read_map || cft->read_seq_string) {
2516                struct cgroup_seqfile_state *state =
2517                        kzalloc(sizeof(*state), GFP_USER);
2518                if (!state)
2519                        return -ENOMEM;
2520                state->cft = cft;
2521                state->cgroup = __d_cgrp(file->f_dentry->d_parent);
2522                file->f_op = &cgroup_seqfile_operations;
2523                err = single_open(file, cgroup_seqfile_show, state);
2524                if (err < 0)
2525                        kfree(state);
2526        } else if (cft->open)
2527                err = cft->open(inode, file);
2528        else
2529                err = 0;
2530
2531        return err;
2532}
2533
2534static int cgroup_file_release(struct inode *inode, struct file *file)
2535{
2536        struct cftype *cft = __d_cft(file->f_dentry);
2537        if (cft->release)
2538                return cft->release(inode, file);
2539        return 0;
2540}
2541
2542/*
2543 * cgroup_rename - Only allow simple rename of directories in place.
2544 */
2545static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2546                            struct inode *new_dir, struct dentry *new_dentry)
2547{
2548        if (!S_ISDIR(old_dentry->d_inode->i_mode))
2549                return -ENOTDIR;
2550        if (new_dentry->d_inode)
2551                return -EEXIST;
2552        if (old_dir != new_dir)
2553                return -EIO;
2554        return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2555}
2556
2557static const struct file_operations cgroup_file_operations = {
2558        .read = cgroup_file_read,
2559        .write = cgroup_file_write,
2560        .llseek = generic_file_llseek,
2561        .open = cgroup_file_open,
2562        .release = cgroup_file_release,
2563};
2564
2565static const struct inode_operations cgroup_dir_inode_operations = {
2566        .lookup = cgroup_lookup,
2567        .mkdir = cgroup_mkdir,
2568        .rmdir = cgroup_rmdir,
2569        .rename = cgroup_rename,
2570};
2571
2572static struct dentry *cgroup_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
2573{
2574        if (dentry->d_name.len > NAME_MAX)
2575                return ERR_PTR(-ENAMETOOLONG);
2576        d_add(dentry, NULL);
2577        return NULL;
2578}
2579
2580/*
2581 * Check if a file is a control file
2582 */
2583static inline struct cftype *__file_cft(struct file *file)
2584{
2585        if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
2586                return ERR_PTR(-EINVAL);
2587        return __d_cft(file->f_dentry);
2588}
2589
2590static int cgroup_create_file(struct dentry *dentry, mode_t mode,
2591                                struct super_block *sb)
2592{
2593        struct inode *inode;
2594
2595        if (!dentry)
2596                return -ENOENT;
2597        if (dentry->d_inode)
2598                return -EEXIST;
2599
2600        inode = cgroup_new_inode(mode, sb);
2601        if (!inode)
2602                return -ENOMEM;
2603
2604        if (S_ISDIR(mode)) {
2605                inode->i_op = &cgroup_dir_inode_operations;
2606                inode->i_fop = &simple_dir_operations;
2607
2608                /* start off with i_nlink == 2 (for "." entry) */
2609                inc_nlink(inode);
2610
2611                /* start with the directory inode held, so that we can
2612                 * populate it without racing with another mkdir */
2613                mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2614        } else if (S_ISREG(mode)) {
2615                inode->i_size = 0;
2616                inode->i_fop = &cgroup_file_operations;
2617        }
2618        d_instantiate(dentry, inode);
2619        dget(dentry);   /* Extra count - pin the dentry in core */
2620        return 0;
2621}
2622
2623/*
2624 * cgroup_create_dir - create a directory for an object.
2625 * @cgrp: the cgroup we create the directory for. It must have a valid
2626 *        ->parent field. And we are going to fill its ->dentry field.
2627 * @dentry: dentry of the new cgroup
2628 * @mode: mode to set on new directory.
2629 */
2630static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
2631                                mode_t mode)
2632{
2633        struct dentry *parent;
2634        int error = 0;
2635
2636        parent = cgrp->parent->dentry;
2637        error = cgroup_create_file(dentry, S_IFDIR | mode, cgrp->root->sb);
2638        if (!error) {
2639                dentry->d_fsdata = cgrp;
2640                inc_nlink(parent->d_inode);
2641                rcu_assign_pointer(cgrp->dentry, dentry);
2642                dget(dentry);
2643        }
2644        dput(dentry);
2645
2646        return error;
2647}
2648
2649/**
2650 * cgroup_file_mode - deduce file mode of a control file
2651 * @cft: the control file in question
2652 *
2653 * returns cft->mode if ->mode is not 0
2654 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2655 * returns S_IRUGO if it has only a read handler
2656 * returns S_IWUSR if it has only a write hander
2657 */
2658static mode_t cgroup_file_mode(const struct cftype *cft)
2659{
2660        mode_t mode = 0;
2661
2662        if (cft->mode)
2663                return cft->mode;
2664
2665        if (cft->read || cft->read_u64 || cft->read_s64 ||
2666            cft->read_map || cft->read_seq_string)
2667                mode |= S_IRUGO;
2668
2669        if (cft->write || cft->write_u64 || cft->write_s64 ||
2670            cft->write_string || cft->trigger)
2671                mode |= S_IWUSR;
2672
2673        return mode;
2674}
2675
2676int cgroup_add_file(struct cgroup *cgrp,
2677                       struct cgroup_subsys *subsys,
2678                       const struct cftype *cft)
2679{
2680        struct dentry *dir = cgrp->dentry;
2681        struct dentry *dentry;
2682        int error;
2683        mode_t mode;
2684
2685        char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2686        if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
2687                strcpy(name, subsys->name);
2688                strcat(name, ".");
2689        }
2690        strcat(name, cft->name);
2691        BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2692        dentry = lookup_one_len(name, dir, strlen(name));
2693        if (!IS_ERR(dentry)) {
2694                mode = cgroup_file_mode(cft);
2695                error = cgroup_create_file(dentry, mode | S_IFREG,
2696                                                cgrp->root->sb);
2697                if (!error)
2698                        dentry->d_fsdata = (void *)cft;
2699                dput(dentry);
2700        } else
2701                error = PTR_ERR(dentry);
2702        return error;
2703}
2704EXPORT_SYMBOL_GPL(cgroup_add_file);
2705
2706int cgroup_add_files(struct cgroup *cgrp,
2707                        struct cgroup_subsys *subsys,
2708                        const struct cftype cft[],
2709                        int count)
2710{
2711        int i, err;
2712        for (i = 0; i < count; i++) {
2713                err = cgroup_add_file(cgrp, subsys, &cft[i]);
2714                if (err)
2715                        return err;
2716        }
2717        return 0;
2718}
2719EXPORT_SYMBOL_GPL(cgroup_add_files);
2720
2721/**
2722 * cgroup_task_count - count the number of tasks in a cgroup.
2723 * @cgrp: the cgroup in question
2724 *
2725 * Return the number of tasks in the cgroup.
2726 */
2727int cgroup_task_count(const struct cgroup *cgrp)
2728{
2729        int count = 0;
2730        struct cg_cgroup_link *link;
2731
2732        read_lock(&css_set_lock);
2733        list_for_each_entry(link, &cgrp->css_sets, cgrp_link_list) {
2734                count += atomic_read(&link->cg->refcount);
2735        }
2736        read_unlock(&css_set_lock);
2737        return count;
2738}
2739
2740/*
2741 * Advance a list_head iterator.  The iterator should be positioned at
2742 * the start of a css_set
2743 */
2744static void cgroup_advance_iter(struct cgroup *cgrp,
2745                                struct cgroup_iter *it)
2746{
2747        struct list_head *l = it->cg_link;
2748        struct cg_cgroup_link *link;
2749        struct css_set *cg;
2750
2751        /* Advance to the next non-empty css_set */
2752        do {
2753                l = l->next;
2754                if (l == &cgrp->css_sets) {
2755                        it->cg_link = NULL;
2756                        return;
2757                }
2758                link = list_entry(l, struct cg_cgroup_link, cgrp_link_list);
2759                cg = link->cg;
2760        } while (list_empty(&cg->tasks));
2761        it->cg_link = l;
2762        it->task = cg->tasks.next;
2763}
2764
2765/*
2766 * To reduce the fork() overhead for systems that are not actually
2767 * using their cgroups capability, we don't maintain the lists running
2768 * through each css_set to its tasks until we see the list actually
2769 * used - in other words after the first call to cgroup_iter_start().
2770 *
2771 * The tasklist_lock is not held here, as do_each_thread() and
2772 * while_each_thread() are protected by RCU.
2773 */
2774static void cgroup_enable_task_cg_lists(void)
2775{
2776        struct task_struct *p, *g;
2777        write_lock(&css_set_lock);
2778        use_task_css_set_links = 1;
2779        do_each_thread(g, p) {
2780                task_lock(p);
2781                /*
2782                 * We should check if the process is exiting, otherwise
2783                 * it will race with cgroup_exit() in that the list
2784                 * entry won't be deleted though the process has exited.
2785                 */
2786                if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2787                        list_add(&p->cg_list, &p->cgroups->tasks);
2788                task_unlock(p);
2789        } while_each_thread(g, p);
2790        write_unlock(&css_set_lock);
2791}
2792
2793void cgroup_iter_start(struct cgroup *cgrp, struct cgroup_iter *it)
2794{
2795        /*
2796         * The first time anyone tries to iterate across a cgroup,
2797         * we need to enable the list linking each css_set to its
2798         * tasks, and fix up all existing tasks.
2799         */
2800        if (!use_task_css_set_links)
2801                cgroup_enable_task_cg_lists();
2802
2803        read_lock(&css_set_lock);
2804        it->cg_link = &cgrp->css_sets;
2805        cgroup_advance_iter(cgrp, it);
2806}
2807
2808struct task_struct *cgroup_iter_next(struct cgroup *cgrp,
2809                                        struct cgroup_iter *it)
2810{
2811        struct task_struct *res;
2812        struct list_head *l = it->task;
2813        struct cg_cgroup_link *link;
2814
2815        /* If the iterator cg is NULL, we have no tasks */
2816        if (!it->cg_link)
2817                return NULL;
2818        res = list_entry(l, struct task_struct, cg_list);
2819        /* Advance iterator to find next entry */
2820        l = l->next;
2821        link = list_entry(it->cg_link, struct cg_cgroup_link, cgrp_link_list);
2822        if (l == &link->cg->tasks) {
2823                /* We reached the end of this task list - move on to
2824                 * the next cg_cgroup_link */
2825                cgroup_advance_iter(cgrp, it);
2826        } else {
2827                it->task = l;
2828        }
2829        return res;
2830}
2831
2832void cgroup_iter_end(struct cgroup *cgrp, struct cgroup_iter *it)
2833{
2834        read_unlock(&css_set_lock);
2835}
2836
2837static inline int started_after_time(struct task_struct *t1,
2838                                     struct timespec *time,
2839                                     struct task_struct *t2)
2840{
2841        int start_diff = timespec_compare(&t1->start_time, time);
2842        if (start_diff > 0) {
2843                return 1;
2844        } else if (start_diff < 0) {
2845                return 0;
2846        } else {
2847                /*
2848                 * Arbitrarily, if two processes started at the same
2849                 * time, we'll say that the lower pointer value
2850                 * started first. Note that t2 may have exited by now
2851                 * so this may not be a valid pointer any longer, but
2852                 * that's fine - it still serves to distinguish
2853                 * between two tasks started (effectively) simultaneously.
2854                 */
2855                return t1 > t2;
2856        }
2857}
2858
2859/*
2860 * This function is a callback from heap_insert() and is used to order
2861 * the heap.
2862 * In this case we order the heap in descending task start time.
2863 */
2864static inline int started_after(void *p1, void *p2)
2865{
2866        struct task_struct *t1 = p1;
2867        struct task_struct *t2 = p2;
2868        return started_after_time(t1, &t2->start_time, t2);
2869}
2870
2871/**
2872 * cgroup_scan_tasks - iterate though all the tasks in a cgroup
2873 * @scan: struct cgroup_scanner containing arguments for the scan
2874 *
2875 * Arguments include pointers to callback functions test_task() and
2876 * process_task().
2877 * Iterate through all the tasks in a cgroup, calling test_task() for each,
2878 * and if it returns true, call process_task() for it also.
2879 * The test_task pointer may be NULL, meaning always true (select all tasks).
2880 * Effectively duplicates cgroup_iter_{start,next,end}()
2881 * but does not lock css_set_lock for the call to process_task().
2882 * The struct cgroup_scanner may be embedded in any structure of the caller's
2883 * creation.
2884 * It is guaranteed that process_task() will act on every task that
2885 * is a member of the cgroup for the duration of this call. This
2886 * function may or may not call process_task() for tasks that exit
2887 * or move to a different cgroup during the call, or are forked or
2888 * move into the cgroup during the call.
2889 *
2890 * Note that test_task() may be called with locks held, and may in some
2891 * situations be called multiple times for the same task, so it should
2892 * be cheap.
2893 * If the heap pointer in the struct cgroup_scanner is non-NULL, a heap has been
2894 * pre-allocated and will be used for heap operations (and its "gt" member will
2895 * be overwritten), else a temporary heap will be used (allocation of which
2896 * may cause this function to fail).
2897 */
2898int cgroup_scan_tasks(struct cgroup_scanner *scan)
2899{
2900        int retval, i;
2901        struct cgroup_iter it;
2902        struct task_struct *p, *dropped;
2903        /* Never dereference latest_task, since it's not refcounted */
2904        struct task_struct *latest_task = NULL;
2905        struct ptr_heap tmp_heap;
2906        struct ptr_heap *heap;
2907        struct timespec latest_time = { 0, 0 };
2908
2909        if (scan->heap) {
2910                /* The caller supplied our heap and pre-allocated its memory */
2911                heap = scan->heap;
2912                heap->gt = &started_after;
2913        } else {
2914                /* We need to allocate our own heap memory */
2915                heap = &tmp_heap;
2916                retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
2917                if (retval)
2918                        /* cannot allocate the heap */
2919                        return retval;
2920        }
2921
2922 again:
2923        /*
2924         * Scan tasks in the cgroup, using the scanner's "test_task" callback
2925         * to determine which are of interest, and using the scanner's
2926         * "process_task" callback to process any of them that need an update.
2927         * Since we don't want to hold any locks during the task updates,
2928         * gather tasks to be processed in a heap structure.
2929         * The heap is sorted by descending task start time.
2930         * If the statically-sized heap fills up, we overflow tasks that
2931         * started later, and in future iterations only consider tasks that
2932         * started after the latest task in the previous pass. This
2933         * guarantees forward progress and that we don't miss any tasks.
2934         */
2935        heap->size = 0;
2936        cgroup_iter_start(scan->cg, &it);
2937        while ((p = cgroup_iter_next(scan->cg, &it))) {
2938                /*
2939                 * Only affect tasks that qualify per the caller's callback,
2940                 * if he provided one
2941                 */
2942                if (scan->test_task && !scan->test_task(p, scan))
2943                        continue;
2944                /*
2945                 * Only process tasks that started after the last task
2946                 * we processed
2947                 */
2948                if (!started_after_time(p, &latest_time, latest_task))
2949                        continue;
2950                dropped = heap_insert(heap, p);
2951                if (dropped == NULL) {
2952                        /*
2953                         * The new task was inserted; the heap wasn't
2954                         * previously full
2955                         */
2956                        get_task_struct(p);
2957                } else if (dropped != p) {
2958                        /*
2959                         * The new task was inserted, and pushed out a
2960                         * different task
2961                         */
2962                        get_task_struct(p);
2963                        put_task_struct(dropped);
2964                }
2965                /*
2966                 * Else the new task was newer than anything already in
2967                 * the heap and wasn't inserted
2968                 */
2969        }
2970        cgroup_iter_end(scan->cg, &it);
2971
2972        if (heap->size) {
2973                for (i = 0; i < heap->size; i++) {
2974                        struct task_struct *q = heap->ptrs[i];
2975                        if (i == 0) {
2976                                latest_time = q->start_time;
2977                                latest_task = q;
2978                        }
2979                        /* Process the task per the caller's callback */
2980                        scan->process_task(q, scan);
2981                        put_task_struct(q);
2982                }
2983                /*
2984                 * If we had to process any tasks at all, scan again
2985                 * in case some of them were in the middle of forking
2986                 * children that didn't get processed.
2987                 * Not the most efficient way to do it, but it avoids
2988                 * having to take callback_mutex in the fork path
2989                 */
2990                goto again;
2991        }
2992        if (heap == &tmp_heap)
2993                heap_free(&tmp_heap);
2994        return 0;
2995}
2996
2997/*
2998 * Stuff for reading the 'tasks'/'procs' files.
2999 *
3000 * Reading this file can return large amounts of data if a cgroup has
3001 * *lots* of attached tasks. So it may need several calls to read(),
3002 * but we cannot guarantee that the information we produce is correct
3003 * unless we produce it entirely atomically.
3004 *
3005 */
3006
3007/*
3008 * The following two functions "fix" the issue where there are more pids
3009 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3010 * TODO: replace with a kernel-wide solution to this problem
3011 */
3012#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3013static void *pidlist_allocate(int count)
3014{
3015        if (PIDLIST_TOO_LARGE(count))
3016                return vmalloc(count * sizeof(pid_t));
3017        else
3018                return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3019}
3020static void pidlist_free(void *p)
3021{
3022        if (is_vmalloc_addr(p))
3023                vfree(p);
3024        else
3025                kfree(p);
3026}
3027static void *pidlist_resize(void *p, int newcount)
3028{
3029        void *newlist;
3030        /* note: if new alloc fails, old p will still be valid either way */
3031        if (is_vmalloc_addr(p)) {
3032                newlist = vmalloc(newcount * sizeof(pid_t));
3033                if (!newlist)
3034                        return NULL;
3035                memcpy(newlist, p, newcount * sizeof(pid_t));
3036                vfree(p);
3037        } else {
3038                newlist = krealloc(p, newcount * sizeof(pid_t), GFP_KERNEL);
3039        }
3040        return newlist;
3041}
3042
3043/*
3044 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3045 * If the new stripped list is sufficiently smaller and there's enough memory
3046 * to allocate a new buffer, will let go of the unneeded memory. Returns the
3047 * number of unique elements.
3048 */
3049/* is the size difference enough that we should re-allocate the array? */
3050#define PIDLIST_REALLOC_DIFFERENCE(old, new) ((old) - PAGE_SIZE >= (new))
3051static int pidlist_uniq(pid_t **p, int length)
3052{
3053        int src, dest = 1;
3054        pid_t *list = *p;
3055        pid_t *newlist;
3056
3057        /*
3058         * we presume the 0th element is unique, so i starts at 1. trivial
3059         * edge cases first; no work needs to be done for either
3060         */
3061        if (length == 0 || length == 1)
3062                return length;
3063        /* src and dest walk down the list; dest counts unique elements */
3064        for (src = 1; src < length; src++) {
3065                /* find next unique element */
3066                while (list[src] == list[src-1]) {
3067                        src++;
3068                        if (src == length)
3069                                goto after;
3070                }
3071                /* dest always points to where the next unique element goes */
3072                list[dest] = list[src];
3073                dest++;
3074        }
3075after:
3076        /*
3077         * if the length difference is large enough, we want to allocate a
3078         * smaller buffer to save memory. if this fails due to out of memory,
3079         * we'll just stay with what we've got.
3080         */
3081        if (PIDLIST_REALLOC_DIFFERENCE(length, dest)) {
3082                newlist = pidlist_resize(list, dest);
3083                if (newlist)
3084                        *p = newlist;
3085        }
3086        return dest;
3087}
3088
3089static int cmppid(const void *a, const void *b)
3090{
3091        return *(pid_t *)a - *(pid_t *)b;
3092}
3093
3094/*
3095 * find the appropriate pidlist for our purpose (given procs vs tasks)
3096 * returns with the lock on that pidlist already held, and takes care
3097 * of the use count, or returns NULL with no locks held if we're out of
3098 * memory.
3099 */
3100static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3101                                                  enum cgroup_filetype type)
3102{
3103        struct cgroup_pidlist *l;
3104        /* don't need task_nsproxy() if we're looking at ourself */
3105        struct pid_namespace *ns = current->nsproxy->pid_ns;
3106
3107        /*
3108         * We can't drop the pidlist_mutex before taking the l->mutex in case
3109         * the last ref-holder is trying to remove l from the list at the same
3110         * time. Holding the pidlist_mutex precludes somebody taking whichever
3111         * list we find out from under us - compare release_pid_array().
3112         */
3113        mutex_lock(&cgrp->pidlist_mutex);
3114        list_for_each_entry(l, &cgrp->pidlists, links) {
3115                if (l->key.type == type && l->key.ns == ns) {
3116                        /* make sure l doesn't vanish out from under us */
3117                        down_write(&l->mutex);
3118                        mutex_unlock(&cgrp->pidlist_mutex);
3119                        return l;
3120                }
3121        }
3122        /* entry not found; create a new one */
3123        l = kmalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3124        if (!l) {
3125                mutex_unlock(&cgrp->pidlist_mutex);
3126                return l;
3127        }
3128        init_rwsem(&l->mutex);
3129        down_write(&l->mutex);
3130        l->key.type = type;
3131        l->key.ns = get_pid_ns(ns);
3132        l->use_count = 0; /* don't increment here */
3133        l->list = NULL;
3134        l->owner = cgrp;
3135        list_add(&l->links, &cgrp->pidlists);
3136        mutex_unlock(&cgrp->pidlist_mutex);
3137        return l;
3138}
3139
3140/*
3141 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3142 */
3143static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3144                              struct cgroup_pidlist **lp)
3145{
3146        pid_t *array;
3147        int length;
3148        int pid, n = 0; /* used for populating the array */
3149        struct cgroup_iter it;
3150        struct task_struct *tsk;
3151        struct cgroup_pidlist *l;
3152
3153        /*
3154         * If cgroup gets more users after we read count, we won't have
3155         * enough space - tough.  This race is indistinguishable to the
3156         * caller from the case that the additional cgroup users didn't
3157         * show up until sometime later on.
3158         */
3159        length = cgroup_task_count(cgrp);
3160        array = pidlist_allocate(length);
3161        if (!array)
3162                return -ENOMEM;
3163        /* now, populate the array */
3164        cgroup_iter_start(cgrp, &it);
3165        while ((tsk = cgroup_iter_next(cgrp, &it))) {
3166                if (unlikely(n == length))
3167                        break;
3168                /* get tgid or pid for procs or tasks file respectively */
3169                if (type == CGROUP_FILE_PROCS)
3170                        pid = task_tgid_vnr(tsk);
3171                else
3172                        pid = task_pid_vnr(tsk);
3173                if (pid > 0) /* make sure to only use valid results */
3174                        array[n++] = pid;
3175        }
3176        cgroup_iter_end(cgrp, &it);
3177        length = n;
3178        /* now sort & (if procs) strip out duplicates */
3179        sort(array, length, sizeof(pid_t), cmppid, NULL);
3180        if (type == CGROUP_FILE_PROCS)
3181                length = pidlist_uniq(&array, length);
3182        l = cgroup_pidlist_find(cgrp, type);
3183        if (!l) {
3184                pidlist_free(array);
3185                return -ENOMEM;
3186        }
3187        /* store array, freeing old if necessary - lock already held */
3188        pidlist_free(l->list);
3189        l->list = array;
3190        l->length = length;
3191        l->use_count++;
3192        up_write(&l->mutex);
3193        *lp = l;
3194        return 0;
3195}
3196
3197/**
3198 * cgroupstats_build - build and fill cgroupstats
3199 * @stats: cgroupstats to fill information into
3200 * @dentry: A dentry entry belonging to the cgroup for which stats have
3201 * been requested.
3202 *
3203 * Build and fill cgroupstats so that taskstats can export it to user
3204 * space.
3205 */
3206int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3207{
3208        int ret = -EINVAL;
3209        struct cgroup *cgrp;
3210        struct cgroup_iter it;
3211        struct task_struct *tsk;
3212
3213        /*
3214         * Validate dentry by checking the superblock operations,
3215         * and make sure it's a directory.
3216         */
3217        if (dentry->d_sb->s_op != &cgroup_ops ||
3218            !S_ISDIR(dentry->d_inode->i_mode))
3219                 goto err;
3220
3221        ret = 0;
3222        cgrp = dentry->d_fsdata;
3223
3224        cgroup_iter_start(cgrp, &it);
3225        while ((tsk = cgroup_iter_next(cgrp, &it))) {
3226                switch (tsk->state) {
3227                case TASK_RUNNING:
3228                        stats->nr_running++;
3229                        break;
3230                case TASK_INTERRUPTIBLE:
3231                        stats->nr_sleeping++;
3232                        break;
3233                case TASK_UNINTERRUPTIBLE:
3234                        stats->nr_uninterruptible++;
3235                        break;
3236                case TASK_STOPPED:
3237                        stats->nr_stopped++;
3238                        break;
3239                default:
3240                        if (delayacct_is_task_waiting_on_io(tsk))
3241                                stats->nr_io_wait++;
3242                        break;
3243                }
3244        }
3245        cgroup_iter_end(cgrp, &it);
3246
3247err:
3248        return ret;
3249}
3250
3251
3252/*
3253 * seq_file methods for the tasks/procs files. The seq_file position is the
3254 * next pid to display; the seq_file iterator is a pointer to the pid
3255 * in the cgroup->l->list array.
3256 */
3257
3258static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3259{
3260        /*
3261         * Initially we receive a position value that corresponds to
3262         * one more than the last pid shown (or 0 on the first call or
3263         * after a seek to the start). Use a binary-search to find the
3264         * next pid to display, if any
3265         */
3266        struct cgroup_pidlist *l = s->private;
3267        int index = 0, pid = *pos;
3268        int *iter;
3269
3270        down_read(&l->mutex);
3271        if (pid) {
3272                int end = l->length;
3273
3274                while (index < end) {
3275                        int mid = (index + end) / 2;
3276                        if (l->list[mid] == pid) {
3277                                index = mid;
3278                                break;
3279                        } else if (l->list[mid] <= pid)
3280                                index = mid + 1;
3281                        else
3282                                end = mid;
3283                }
3284        }
3285        /* If we're off the end of the array, we're done */
3286        if (index >= l->length)
3287                return NULL;
3288        /* Update the abstract position to be the actual pid that we found */
3289        iter = l->list + index;
3290        *pos = *iter;
3291        return iter;
3292}
3293
3294static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3295{
3296        struct cgroup_pidlist *l = s->private;
3297        up_read(&l->mutex);
3298}
3299
3300static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3301{
3302        struct cgroup_pidlist *l = s->private;
3303        pid_t *p = v;
3304        pid_t *end = l->list + l->length;
3305        /*
3306         * Advance to the next pid in the array. If this goes off the
3307         * end, we're done
3308         */
3309        p++;
3310        if (p >= end) {
3311                return NULL;
3312        } else {
3313                *pos = *p;
3314                return p;
3315        }
3316}
3317
3318static int cgroup_pidlist_show(struct seq_file *s, void *v)
3319{
3320        return seq_printf(s, "%d\n", *(int *)v);
3321}
3322
3323/*
3324 * seq_operations functions for iterating on pidlists through seq_file -
3325 * independent of whether it's tasks or procs
3326 */
3327static const struct seq_operations cgroup_pidlist_seq_operations = {
3328        .start = cgroup_pidlist_start,
3329        .stop = cgroup_pidlist_stop,
3330        .next = cgroup_pidlist_next,
3331        .show = cgroup_pidlist_show,
3332};
3333
3334static void cgroup_release_pid_array(struct cgroup_pidlist *l)
3335{
3336        /*
3337         * the case where we're the last user of this particular pidlist will
3338         * have us remove it from the cgroup's list, which entails taking the
3339         * mutex. since in pidlist_find the pidlist->lock depends on cgroup->
3340         * pidlist_mutex, we have to take pidlist_mutex first.
3341         */
3342        mutex_lock(&l->owner->pidlist_mutex);
3343        down_write(&l->mutex);
3344        BUG_ON(!l->use_count);
3345        if (!--l->use_count) {
3346                /* we're the last user if refcount is 0; remove and free */
3347                list_del(&l->links);
3348                mutex_unlock(&l->owner->pidlist_mutex);
3349                pidlist_free(l->list);
3350                put_pid_ns(l->key.ns);
3351                up_write(&l->mutex);
3352                kfree(l);
3353                return;
3354        }
3355        mutex_unlock(&l->owner->pidlist_mutex);
3356        up_write(&l->mutex);
3357}
3358
3359static int cgroup_pidlist_release(struct inode *inode, struct file *file)
3360{
3361        struct cgroup_pidlist *l;
3362        if (!(file->f_mode & FMODE_READ))
3363                return 0;
3364        /*
3365         * the seq_file will only be initialized if the file was opened for
3366         * reading; hence we check if it's not null only in that case.
3367         */
3368        l = ((struct seq_file *)file->private_data)->private;
3369        cgroup_release_pid_array(l);
3370        return seq_release(inode, file);
3371}
3372
3373static const struct file_operations cgroup_pidlist_operations = {
3374        .read = seq_read,
3375        .llseek = seq_lseek,
3376        .write = cgroup_file_write,
3377        .release = cgroup_pidlist_release,
3378};
3379
3380/*
3381 * The following functions handle opens on a file that displays a pidlist
3382 * (tasks or procs). Prepare an array of the process/thread IDs of whoever's
3383 * in the cgroup.
3384 */
3385/* helper function for the two below it */
3386static int cgroup_pidlist_open(struct file *file, enum cgroup_filetype type)
3387{
3388        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
3389        struct cgroup_pidlist *l;
3390        int retval;
3391
3392        /* Nothing to do for write-only files */
3393        if (!(file->f_mode & FMODE_READ))
3394                return 0;
3395
3396        /* have the array populated */
3397        retval = pidlist_array_load(cgrp, type, &l);
3398        if (retval)
3399                return retval;
3400        /* configure file information */
3401        file->f_op = &cgroup_pidlist_operations;
3402
3403        retval = seq_open(file, &cgroup_pidlist_seq_operations);
3404        if (retval) {
3405                cgroup_release_pid_array(l);
3406                return retval;
3407        }
3408        ((struct seq_file *)file->private_data)->private = l;
3409        return 0;
3410}
3411static int cgroup_tasks_open(struct inode *unused, struct file *file)
3412{
3413        return cgroup_pidlist_open(file, CGROUP_FILE_TASKS);
3414}
3415static int cgroup_procs_open(struct inode *unused, struct file *file)
3416{
3417        return cgroup_pidlist_open(file, CGROUP_FILE_PROCS);
3418}
3419
3420static u64 cgroup_read_notify_on_release(struct cgroup *cgrp,
3421                                            struct cftype *cft)
3422{
3423        return notify_on_release(cgrp);
3424}
3425
3426static int cgroup_write_notify_on_release(struct cgroup *cgrp,
3427                                          struct cftype *cft,
3428                                          u64 val)
3429{
3430        clear_bit(CGRP_RELEASABLE, &cgrp->flags);
3431        if (val)
3432                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3433        else
3434                clear_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3435        return 0;
3436}
3437
3438/*
3439 * Unregister event and free resources.
3440 *
3441 * Gets called from workqueue.
3442 */
3443static void cgroup_event_remove(struct work_struct *work)
3444{
3445        struct cgroup_event *event = container_of(work, struct cgroup_event,
3446                        remove);
3447        struct cgroup *cgrp = event->cgrp;
3448
3449        event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3450
3451        eventfd_ctx_put(event->eventfd);
3452        kfree(event);
3453        dput(cgrp->dentry);
3454}
3455
3456/*
3457 * Gets called on POLLHUP on eventfd when user closes it.
3458 *
3459 * Called with wqh->lock held and interrupts disabled.
3460 */
3461static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
3462                int sync, void *key)
3463{
3464        struct cgroup_event *event = container_of(wait,
3465                        struct cgroup_event, wait);
3466        struct cgroup *cgrp = event->cgrp;
3467        unsigned long flags = (unsigned long)key;
3468
3469        if (flags & POLLHUP) {
3470                __remove_wait_queue(event->wqh, &event->wait);
3471                spin_lock(&cgrp->event_list_lock);
3472                list_del(&event->list);
3473                spin_unlock(&cgrp->event_list_lock);
3474                /*
3475                 * We are in atomic context, but cgroup_event_remove() may
3476                 * sleep, so we have to call it in workqueue.
3477                 */
3478                schedule_work(&event->remove);
3479        }
3480
3481        return 0;
3482}
3483
3484static void cgroup_event_ptable_queue_proc(struct file *file,
3485                wait_queue_head_t *wqh, poll_table *pt)
3486{
3487        struct cgroup_event *event = container_of(pt,
3488                        struct cgroup_event, pt);
3489
3490        event->wqh = wqh;
3491        add_wait_queue(wqh, &event->wait);
3492}
3493
3494/*
3495 * Parse input and register new cgroup event handler.
3496 *
3497 * Input must be in format '<event_fd> <control_fd> <args>'.
3498 * Interpretation of args is defined by control file implementation.
3499 */
3500static int cgroup_write_event_control(struct cgroup *cgrp, struct cftype *cft,
3501                                      const char *buffer)
3502{
3503        struct cgroup_event *event = NULL;
3504        unsigned int efd, cfd;
3505        struct file *efile = NULL;
3506        struct file *cfile = NULL;
3507        char *endp;
3508        int ret;
3509
3510        efd = simple_strtoul(buffer, &endp, 10);
3511        if (*endp != ' ')
3512                return -EINVAL;
3513        buffer = endp + 1;
3514
3515        cfd = simple_strtoul(buffer, &endp, 10);
3516        if ((*endp != ' ') && (*endp != '\0'))
3517                return -EINVAL;
3518        buffer = endp + 1;
3519
3520        event = kzalloc(sizeof(*event), GFP_KERNEL);
3521        if (!event)
3522                return -ENOMEM;
3523        event->cgrp = cgrp;
3524        INIT_LIST_HEAD(&event->list);
3525        init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
3526        init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
3527        INIT_WORK(&event->remove, cgroup_event_remove);
3528
3529        efile = eventfd_fget(efd);
3530        if (IS_ERR(efile)) {
3531                ret = PTR_ERR(efile);
3532                goto fail;
3533        }
3534
3535        event->eventfd = eventfd_ctx_fileget(efile);
3536        if (IS_ERR(event->eventfd)) {
3537                ret = PTR_ERR(event->eventfd);
3538                goto fail;
3539        }
3540
3541        cfile = fget(cfd);
3542        if (!cfile) {
3543                ret = -EBADF;
3544                goto fail;
3545        }
3546
3547        /* the process need read permission on control file */
3548        /* AV: shouldn't we check that it's been opened for read instead? */
3549        ret = inode_permission(cfile->f_path.dentry->d_inode, MAY_READ);
3550        if (ret < 0)
3551                goto fail;
3552
3553        event->cft = __file_cft(cfile);
3554        if (IS_ERR(event->cft)) {
3555                ret = PTR_ERR(event->cft);
3556                goto fail;
3557        }
3558
3559        if (!event->cft->register_event || !event->cft->unregister_event) {
3560                ret = -EINVAL;
3561                goto fail;
3562        }
3563
3564        ret = event->cft->register_event(cgrp, event->cft,
3565                        event->eventfd, buffer);
3566        if (ret)
3567                goto fail;
3568
3569        if (efile->f_op->poll(efile, &event->pt) & POLLHUP) {
3570                event->cft->unregister_event(cgrp, event->cft, event->eventfd);
3571                ret = 0;
3572                goto fail;
3573        }
3574
3575        /*
3576         * Events should be removed after rmdir of cgroup directory, but before
3577         * destroying subsystem state objects. Let's take reference to cgroup
3578         * directory dentry to do that.
3579         */
3580        dget(cgrp->dentry);
3581
3582        spin_lock(&cgrp->event_list_lock);
3583        list_add(&event->list, &cgrp->event_list);
3584        spin_unlock(&cgrp->event_list_lock);
3585
3586        fput(cfile);
3587        fput(efile);
3588
3589        return 0;
3590
3591fail:
3592        if (cfile)
3593                fput(cfile);
3594
3595        if (event && event->eventfd && !IS_ERR(event->eventfd))
3596                eventfd_ctx_put(event->eventfd);
3597
3598        if (!IS_ERR_OR_NULL(efile))
3599                fput(efile);
3600
3601        kfree(event);
3602
3603        return ret;
3604}
3605
3606static u64 cgroup_clone_children_read(struct cgroup *cgrp,
3607                                    struct cftype *cft)
3608{
3609        return clone_children(cgrp);
3610}
3611
3612static int cgroup_clone_children_write(struct cgroup *cgrp,
3613                                     struct cftype *cft,
3614                                     u64 val)
3615{
3616        if (val)
3617                set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3618        else
3619                clear_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3620        return 0;
3621}
3622
3623/*
3624 * for the common functions, 'private' gives the type of file
3625 */
3626/* for hysterical raisins, we can't put this on the older files */
3627#define CGROUP_FILE_GENERIC_PREFIX "cgroup."
3628static struct cftype files[] = {
3629        {
3630                .name = "tasks",
3631                .open = cgroup_tasks_open,
3632                .write_u64 = cgroup_tasks_write,
3633                .release = cgroup_pidlist_release,
3634                .mode = S_IRUGO | S_IWUSR,
3635        },
3636        {
3637                .name = CGROUP_FILE_GENERIC_PREFIX "procs",
3638                .open = cgroup_procs_open,
3639                .write_u64 = cgroup_procs_write,
3640                .release = cgroup_pidlist_release,
3641                .mode = S_IRUGO | S_IWUSR,
3642        },
3643        {
3644                .name = "notify_on_release",
3645                .read_u64 = cgroup_read_notify_on_release,
3646                .write_u64 = cgroup_write_notify_on_release,
3647        },
3648        {
3649                .name = CGROUP_FILE_GENERIC_PREFIX "event_control",
3650                .write_string = cgroup_write_event_control,
3651                .mode = S_IWUGO,
3652        },
3653        {
3654                .name = "cgroup.clone_children",
3655                .read_u64 = cgroup_clone_children_read,
3656                .write_u64 = cgroup_clone_children_write,
3657        },
3658};
3659
3660static struct cftype cft_release_agent = {
3661        .name = "release_agent",
3662        .read_seq_string = cgroup_release_agent_show,
3663        .write_string = cgroup_release_agent_write,
3664        .max_write_len = PATH_MAX,
3665};
3666
3667static int cgroup_populate_dir(struct cgroup *cgrp)
3668{
3669        int err;
3670        struct cgroup_subsys *ss;
3671
3672        /* First clear out any existing files */
3673        cgroup_clear_directory(cgrp->dentry);
3674
3675        err = cgroup_add_files(cgrp, NULL, files, ARRAY_SIZE(files));
3676        if (err < 0)
3677                return err;
3678
3679        if (cgrp == cgrp->top_cgroup) {
3680                if ((err = cgroup_add_file(cgrp, NULL, &cft_release_agent)) < 0)
3681                        return err;
3682        }
3683
3684        for_each_subsys(cgrp->root, ss) {
3685                if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
3686                        return err;
3687        }
3688        /* This cgroup is ready now */
3689        for_each_subsys(cgrp->root, ss) {
3690                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3691                /*
3692                 * Update id->css pointer and make this css visible from
3693                 * CSS ID functions. This pointer will be dereferened
3694                 * from RCU-read-side without locks.
3695                 */
3696                if (css->id)
3697                        rcu_assign_pointer(css->id->css, css);
3698        }
3699
3700        return 0;
3701}
3702
3703static void init_cgroup_css(struct cgroup_subsys_state *css,
3704                               struct cgroup_subsys *ss,
3705                               struct cgroup *cgrp)
3706{
3707        css->cgroup = cgrp;
3708        atomic_set(&css->refcnt, 1);
3709        css->flags = 0;
3710        css->id = NULL;
3711        if (cgrp == dummytop)
3712                set_bit(CSS_ROOT, &css->flags);
3713        BUG_ON(cgrp->subsys[ss->subsys_id]);
3714        cgrp->subsys[ss->subsys_id] = css;
3715}
3716
3717static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
3718{
3719        /* We need to take each hierarchy_mutex in a consistent order */
3720        int i;
3721
3722        /*
3723         * No worry about a race with rebind_subsystems that might mess up the
3724         * locking order, since both parties are under cgroup_mutex.
3725         */
3726        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3727                struct cgroup_subsys *ss = subsys[i];
3728                if (ss == NULL)
3729                        continue;
3730                if (ss->root == root)
3731                        mutex_lock(&ss->hierarchy_mutex);
3732        }
3733}
3734
3735static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
3736{
3737        int i;
3738
3739        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3740                struct cgroup_subsys *ss = subsys[i];
3741                if (ss == NULL)
3742                        continue;
3743                if (ss->root == root)
3744                        mutex_unlock(&ss->hierarchy_mutex);
3745        }
3746}
3747
3748/*
3749 * cgroup_create - create a cgroup
3750 * @parent: cgroup that will be parent of the new cgroup
3751 * @dentry: dentry of the new cgroup
3752 * @mode: mode to set on new inode
3753 *
3754 * Must be called with the mutex on the parent inode held
3755 */
3756static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
3757                             mode_t mode)
3758{
3759        struct cgroup *cgrp;
3760        struct cgroupfs_root *root = parent->root;
3761        int err = 0;
3762        struct cgroup_subsys *ss;
3763        struct super_block *sb = root->sb;
3764
3765        cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
3766        if (!cgrp)
3767                return -ENOMEM;
3768
3769        /* Grab a reference on the superblock so the hierarchy doesn't
3770         * get deleted on unmount if there are child cgroups.  This
3771         * can be done outside cgroup_mutex, since the sb can't
3772         * disappear while someone has an open control file on the
3773         * fs */
3774        atomic_inc(&sb->s_active);
3775
3776        mutex_lock(&cgroup_mutex);
3777
3778        init_cgroup_housekeeping(cgrp);
3779
3780        cgrp->parent = parent;
3781        cgrp->root = parent->root;
3782        cgrp->top_cgroup = parent->top_cgroup;
3783
3784        if (notify_on_release(parent))
3785                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
3786
3787        if (clone_children(parent))
3788                set_bit(CGRP_CLONE_CHILDREN, &cgrp->flags);
3789
3790        for_each_subsys(root, ss) {
3791                struct cgroup_subsys_state *css = ss->create(ss, cgrp);
3792
3793                if (IS_ERR(css)) {
3794                        err = PTR_ERR(css);
3795                        goto err_destroy;
3796                }
3797                init_cgroup_css(css, ss, cgrp);
3798                if (ss->use_id) {
3799                        err = alloc_css_id(ss, parent, cgrp);
3800                        if (err)
3801                                goto err_destroy;
3802                }
3803                /* At error, ->destroy() callback has to free assigned ID. */
3804                if (clone_children(parent) && ss->post_clone)
3805                        ss->post_clone(ss, cgrp);
3806        }
3807
3808        cgroup_lock_hierarchy(root);
3809        list_add(&cgrp->sibling, &cgrp->parent->children);
3810        cgroup_unlock_hierarchy(root);
3811        root->number_of_cgroups++;
3812
3813        err = cgroup_create_dir(cgrp, dentry, mode);
3814        if (err < 0)
3815                goto err_remove;
3816
3817        /* The cgroup directory was pre-locked for us */
3818        BUG_ON(!mutex_is_locked(&cgrp->dentry->d_inode->i_mutex));
3819
3820        err = cgroup_populate_dir(cgrp);
3821        /* If err < 0, we have a half-filled directory - oh well ;) */
3822
3823        mutex_unlock(&cgroup_mutex);
3824        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
3825
3826        return 0;
3827
3828 err_remove:
3829
3830        cgroup_lock_hierarchy(root);
3831        list_del(&cgrp->sibling);
3832        cgroup_unlock_hierarchy(root);
3833        root->number_of_cgroups--;
3834
3835 err_destroy:
3836
3837        for_each_subsys(root, ss) {
3838                if (cgrp->subsys[ss->subsys_id])
3839                        ss->destroy(ss, cgrp);
3840        }
3841
3842        mutex_unlock(&cgroup_mutex);
3843
3844        /* Release the reference count that we took on the superblock */
3845        deactivate_super(sb);
3846
3847        kfree(cgrp);
3848        return err;
3849}
3850
3851static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode)
3852{
3853        struct cgroup *c_parent = dentry->d_parent->d_fsdata;
3854
3855        /* the vfs holds inode->i_mutex already */
3856        return cgroup_create(c_parent, dentry, mode | S_IFDIR);
3857}
3858
3859static int cgroup_has_css_refs(struct cgroup *cgrp)
3860{
3861        /* Check the reference count on each subsystem. Since we
3862         * already established that there are no tasks in the
3863         * cgroup, if the css refcount is also 1, then there should
3864         * be no outstanding references, so the subsystem is safe to
3865         * destroy. We scan across all subsystems rather than using
3866         * the per-hierarchy linked list of mounted subsystems since
3867         * we can be called via check_for_release() with no
3868         * synchronization other than RCU, and the subsystem linked
3869         * list isn't RCU-safe */
3870        int i;
3871        /*
3872         * We won't need to lock the subsys array, because the subsystems
3873         * we're concerned about aren't going anywhere since our cgroup root
3874         * has a reference on them.
3875         */
3876        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
3877                struct cgroup_subsys *ss = subsys[i];
3878                struct cgroup_subsys_state *css;
3879                /* Skip subsystems not present or not in this hierarchy */
3880                if (ss == NULL || ss->root != cgrp->root)
3881                        continue;
3882                css = cgrp->subsys[ss->subsys_id];
3883                /* When called from check_for_release() it's possible
3884                 * that by this point the cgroup has been removed
3885                 * and the css deleted. But a false-positive doesn't
3886                 * matter, since it can only happen if the cgroup
3887                 * has been deleted and hence no longer needs the
3888                 * release agent to be called anyway. */
3889                if (css && (atomic_read(&css->refcnt) > 1))
3890                        return 1;
3891        }
3892        return 0;
3893}
3894
3895/*
3896 * Atomically mark all (or else none) of the cgroup's CSS objects as
3897 * CSS_REMOVED. Return true on success, or false if the cgroup has
3898 * busy subsystems. Call with cgroup_mutex held
3899 */
3900
3901static int cgroup_clear_css_refs(struct cgroup *cgrp)
3902{
3903        struct cgroup_subsys *ss;
3904        unsigned long flags;
3905        bool failed = false;
3906        local_irq_save(flags);
3907        for_each_subsys(cgrp->root, ss) {
3908                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3909                int refcnt;
3910                while (1) {
3911                        /* We can only remove a CSS with a refcnt==1 */
3912                        refcnt = atomic_read(&css->refcnt);
3913                        if (refcnt > 1) {
3914                                failed = true;
3915                                goto done;
3916                        }
3917                        BUG_ON(!refcnt);
3918                        /*
3919                         * Drop the refcnt to 0 while we check other
3920                         * subsystems. This will cause any racing
3921                         * css_tryget() to spin until we set the
3922                         * CSS_REMOVED bits or abort
3923                         */
3924                        if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
3925                                break;
3926                        cpu_relax();
3927                }
3928        }
3929 done:
3930        for_each_subsys(cgrp->root, ss) {
3931                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
3932                if (failed) {
3933                        /*
3934                         * Restore old refcnt if we previously managed
3935                         * to clear it from 1 to 0
3936                         */
3937                        if (!atomic_read(&css->refcnt))
3938                                atomic_set(&css->refcnt, 1);
3939                } else {
3940                        /* Commit the fact that the CSS is removed */
3941                        set_bit(CSS_REMOVED, &css->flags);
3942                }
3943        }
3944        local_irq_restore(flags);
3945        return !failed;
3946}
3947
3948static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
3949{
3950        struct cgroup *cgrp = dentry->d_fsdata;
3951        struct dentry *d;
3952        struct cgroup *parent;
3953        DEFINE_WAIT(wait);
3954        struct cgroup_event *event, *tmp;
3955        int ret;
3956
3957        /* the vfs holds both inode->i_mutex already */
3958again:
3959        mutex_lock(&cgroup_mutex);
3960        if (atomic_read(&cgrp->count) != 0) {
3961                mutex_unlock(&cgroup_mutex);
3962                return -EBUSY;
3963        }
3964        if (!list_empty(&cgrp->children)) {
3965                mutex_unlock(&cgroup_mutex);
3966                return -EBUSY;
3967        }
3968        mutex_unlock(&cgroup_mutex);
3969
3970        /*
3971         * In general, subsystem has no css->refcnt after pre_destroy(). But
3972         * in racy cases, subsystem may have to get css->refcnt after
3973         * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
3974         * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
3975         * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
3976         * and subsystem's reference count handling. Please see css_get/put
3977         * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
3978         */
3979        set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3980
3981        /*
3982         * Call pre_destroy handlers of subsys. Notify subsystems
3983         * that rmdir() request comes.
3984         */
3985        ret = cgroup_call_pre_destroy(cgrp);
3986        if (ret) {
3987                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3988                return ret;
3989        }
3990
3991        mutex_lock(&cgroup_mutex);
3992        parent = cgrp->parent;
3993        if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
3994                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
3995                mutex_unlock(&cgroup_mutex);
3996                return -EBUSY;
3997        }
3998        prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
3999        if (!cgroup_clear_css_refs(cgrp)) {
4000                mutex_unlock(&cgroup_mutex);
4001                /*
4002                 * Because someone may call cgroup_wakeup_rmdir_waiter() before
4003                 * prepare_to_wait(), we need to check this flag.
4004                 */
4005                if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
4006                        schedule();
4007                finish_wait(&cgroup_rmdir_waitq, &wait);
4008                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4009                if (signal_pending(current))
4010                        return -EINTR;
4011                goto again;
4012        }
4013        /* NO css_tryget() can success after here. */
4014        finish_wait(&cgroup_rmdir_waitq, &wait);
4015        clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
4016
4017        spin_lock(&release_list_lock);
4018        set_bit(CGRP_REMOVED, &cgrp->flags);
4019        if (!list_empty(&cgrp->release_list))
4020                list_del_init(&cgrp->release_list);
4021        spin_unlock(&release_list_lock);
4022
4023        cgroup_lock_hierarchy(cgrp->root);
4024        /* delete this cgroup from parent->children */
4025        list_del_init(&cgrp->sibling);
4026        cgroup_unlock_hierarchy(cgrp->root);
4027
4028        d = dget(cgrp->dentry);
4029
4030        cgroup_d_remove_dir(d);
4031        dput(d);
4032
4033        set_bit(CGRP_RELEASABLE, &parent->flags);
4034        check_for_release(parent);
4035
4036        /*
4037         * Unregister events and notify userspace.
4038         * Notify userspace about cgroup removing only after rmdir of cgroup
4039         * directory to avoid race between userspace and kernelspace
4040         */
4041        spin_lock(&cgrp->event_list_lock);
4042        list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
4043                list_del(&event->list);
4044                remove_wait_queue(event->wqh, &event->wait);
4045                eventfd_signal(event->eventfd, 1);
4046                schedule_work(&event->remove);
4047        }
4048        spin_unlock(&cgrp->event_list_lock);
4049
4050        mutex_unlock(&cgroup_mutex);
4051        return 0;
4052}
4053
4054static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4055{
4056        struct cgroup_subsys_state *css;
4057
4058        printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4059
4060        /* Create the top cgroup state for this subsystem */
4061        list_add(&ss->sibling, &rootnode.subsys_list);
4062        ss->root = &rootnode;
4063        css = ss->create(ss, dummytop);
4064        /* We don't handle early failures gracefully */
4065        BUG_ON(IS_ERR(css));
4066        init_cgroup_css(css, ss, dummytop);
4067
4068        /* Update the init_css_set to contain a subsys
4069         * pointer to this state - since the subsystem is
4070         * newly registered, all tasks and hence the
4071         * init_css_set is in the subsystem's top cgroup. */
4072        init_css_set.subsys[ss->subsys_id] = dummytop->subsys[ss->subsys_id];
4073
4074        need_forkexit_callback |= ss->fork || ss->exit;
4075
4076        /* At system boot, before all subsystems have been
4077         * registered, no tasks have been forked, so we don't
4078         * need to invoke fork callbacks here. */
4079        BUG_ON(!list_empty(&init_task.tasks));
4080
4081        mutex_init(&ss->hierarchy_mutex);
4082        lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4083        ss->active = 1;
4084
4085        /* this function shouldn't be used with modular subsystems, since they
4086         * need to register a subsys_id, among other things */
4087        BUG_ON(ss->module);
4088}
4089
4090/**
4091 * cgroup_load_subsys: load and register a modular subsystem at runtime
4092 * @ss: the subsystem to load
4093 *
4094 * This function should be called in a modular subsystem's initcall. If the
4095 * subsystem is built as a module, it will be assigned a new subsys_id and set
4096 * up for use. If the subsystem is built-in anyway, work is delegated to the
4097 * simpler cgroup_init_subsys.
4098 */
4099int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4100{
4101        int i;
4102        struct cgroup_subsys_state *css;
4103
4104        /* check name and function validity */
4105        if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4106            ss->create == NULL || ss->destroy == NULL)
4107                return -EINVAL;
4108
4109        /*
4110         * we don't support callbacks in modular subsystems. this check is
4111         * before the ss->module check for consistency; a subsystem that could
4112         * be a module should still have no callbacks even if the user isn't
4113         * compiling it as one.
4114         */
4115        if (ss->fork || ss->exit)
4116                return -EINVAL;
4117
4118        /*
4119         * an optionally modular subsystem is built-in: we want to do nothing,
4120         * since cgroup_init_subsys will have already taken care of it.
4121         */
4122        if (ss->module == NULL) {
4123                /* a few sanity checks */
4124                BUG_ON(ss->subsys_id >= CGROUP_BUILTIN_SUBSYS_COUNT);
4125                BUG_ON(subsys[ss->subsys_id] != ss);
4126                return 0;
4127        }
4128
4129        /*
4130         * need to register a subsys id before anything else - for example,
4131         * init_cgroup_css needs it.
4132         */
4133        mutex_lock(&cgroup_mutex);
4134        /* find the first empty slot in the array */
4135        for (i = CGROUP_BUILTIN_SUBSYS_COUNT; i < CGROUP_SUBSYS_COUNT; i++) {
4136                if (subsys[i] == NULL)
4137                        break;
4138        }
4139        if (i == CGROUP_SUBSYS_COUNT) {
4140                /* maximum number of subsystems already registered! */
4141                mutex_unlock(&cgroup_mutex);
4142                return -EBUSY;
4143        }
4144        /* assign ourselves the subsys_id */
4145        ss->subsys_id = i;
4146        subsys[i] = ss;
4147
4148        /*
4149         * no ss->create seems to need anything important in the ss struct, so
4150         * this can happen first (i.e. before the rootnode attachment).
4151         */
4152        css = ss->create(ss, dummytop);
4153        if (IS_ERR(css)) {
4154                /* failure case - need to deassign the subsys[] slot. */
4155                subsys[i] = NULL;
4156                mutex_unlock(&cgroup_mutex);
4157                return PTR_ERR(css);
4158        }
4159
4160        list_add(&ss->sibling, &rootnode.subsys_list);
4161        ss->root = &rootnode;
4162
4163        /* our new subsystem will be attached to the dummy hierarchy. */
4164        init_cgroup_css(css, ss, dummytop);
4165        /* init_idr must be after init_cgroup_css because it sets css->id. */
4166        if (ss->use_id) {
4167                int ret = cgroup_init_idr(ss, css);
4168                if (ret) {
4169                        dummytop->subsys[ss->subsys_id] = NULL;
4170                        ss->destroy(ss, dummytop);
4171                        subsys[i] = NULL;
4172                        mutex_unlock(&cgroup_mutex);
4173                        return ret;
4174                }
4175        }
4176
4177        /*
4178         * Now we need to entangle the css into the existing css_sets. unlike
4179         * in cgroup_init_subsys, there are now multiple css_sets, so each one
4180         * will need a new pointer to it; done by iterating the css_set_table.
4181         * furthermore, modifying the existing css_sets will corrupt the hash
4182         * table state, so each changed css_set will need its hash recomputed.
4183         * this is all done under the css_set_lock.
4184         */
4185        write_lock(&css_set_lock);
4186        for (i = 0; i < CSS_SET_TABLE_SIZE; i++) {
4187                struct css_set *cg;
4188                struct hlist_node *node, *tmp;
4189                struct hlist_head *bucket = &css_set_table[i], *new_bucket;
4190
4191                hlist_for_each_entry_safe(cg, node, tmp, bucket, hlist) {
4192                        /* skip entries that we already rehashed */
4193                        if (cg->subsys[ss->subsys_id])
4194                                continue;
4195                        /* remove existing entry */
4196                        hlist_del(&cg->hlist);
4197                        /* set new value */
4198                        cg->subsys[ss->subsys_id] = css;
4199                        /* recompute hash and restore entry */
4200                        new_bucket = css_set_hash(cg->subsys);
4201                        hlist_add_head(&cg->hlist, new_bucket);
4202                }
4203        }
4204        write_unlock(&css_set_lock);
4205
4206        mutex_init(&ss->hierarchy_mutex);
4207        lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
4208        ss->active = 1;
4209
4210        /* success! */
4211        mutex_unlock(&cgroup_mutex);
4212        return 0;
4213}
4214EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4215
4216/**
4217 * cgroup_unload_subsys: unload a modular subsystem
4218 * @ss: the subsystem to unload
4219 *
4220 * This function should be called in a modular subsystem's exitcall. When this
4221 * function is invoked, the refcount on the subsystem's module will be 0, so
4222 * the subsystem will not be attached to any hierarchy.
4223 */
4224void cgroup_unload_subsys(struct cgroup_subsys *ss)
4225{
4226        struct cg_cgroup_link *link;
4227        struct hlist_head *hhead;
4228
4229        BUG_ON(ss->module == NULL);
4230
4231        /*
4232         * we shouldn't be called if the subsystem is in use, and the use of
4233         * try_module_get in parse_cgroupfs_options should ensure that it
4234         * doesn't start being used while we're killing it off.
4235         */
4236        BUG_ON(ss->root != &rootnode);
4237
4238        mutex_lock(&cgroup_mutex);
4239        /* deassign the subsys_id */
4240        BUG_ON(ss->subsys_id < CGROUP_BUILTIN_SUBSYS_COUNT);
4241        subsys[ss->subsys_id] = NULL;
4242
4243        /* remove subsystem from rootnode's list of subsystems */
4244        list_del_init(&ss->sibling);
4245
4246        /*
4247         * disentangle the css from all css_sets attached to the dummytop. as
4248         * in loading, we need to pay our respects to the hashtable gods.
4249         */
4250        write_lock(&css_set_lock);
4251        list_for_each_entry(link, &dummytop->css_sets, cgrp_link_list) {
4252                struct css_set *cg = link->cg;
4253
4254                hlist_del(&cg->hlist);
4255                BUG_ON(!cg->subsys[ss->subsys_id]);
4256                cg->subsys[ss->subsys_id] = NULL;
4257                hhead = css_set_hash(cg->subsys);
4258                hlist_add_head(&cg->hlist, hhead);
4259        }
4260        write_unlock(&css_set_lock);
4261
4262        /*
4263         * remove subsystem's css from the dummytop and free it - need to free
4264         * before marking as null because ss->destroy needs the cgrp->subsys
4265         * pointer to find their state. note that this also takes care of
4266         * freeing the css_id.
4267         */
4268        ss->destroy(ss, dummytop);
4269        dummytop->subsys[ss->subsys_id] = NULL;
4270
4271        mutex_unlock(&cgroup_mutex);
4272}
4273EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4274
4275/**
4276 * cgroup_init_early - cgroup initialization at system boot
4277 *
4278 * Initialize cgroups at system boot, and initialize any
4279 * subsystems that request early init.
4280 */
4281int __init cgroup_init_early(void)
4282{
4283        int i;
4284        atomic_set(&init_css_set.refcount, 1);
4285        INIT_LIST_HEAD(&init_css_set.cg_links);
4286        INIT_LIST_HEAD(&init_css_set.tasks);
4287        INIT_HLIST_NODE(&init_css_set.hlist);
4288        css_set_count = 1;
4289        init_cgroup_root(&rootnode);
4290        root_count = 1;
4291        init_task.cgroups = &init_css_set;
4292
4293        init_css_set_link.cg = &init_css_set;
4294        init_css_set_link.cgrp = dummytop;
4295        list_add(&init_css_set_link.cgrp_link_list,
4296                 &rootnode.top_cgroup.css_sets);
4297        list_add(&init_css_set_link.cg_link_list,
4298                 &init_css_set.cg_links);
4299
4300        for (i = 0; i < CSS_SET_TABLE_SIZE; i++)
4301                INIT_HLIST_HEAD(&css_set_table[i]);
4302
4303        /* at bootup time, we don't worry about modular subsystems */
4304        for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4305                struct cgroup_subsys *ss = subsys[i];
4306
4307                BUG_ON(!ss->name);
4308                BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4309                BUG_ON(!ss->create);
4310                BUG_ON(!ss->destroy);
4311                if (ss->subsys_id != i) {
4312                        printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4313                               ss->name, ss->subsys_id);
4314                        BUG();
4315                }
4316
4317                if (ss->early_init)
4318                        cgroup_init_subsys(ss);
4319        }
4320        return 0;
4321}
4322
4323/**
4324 * cgroup_init - cgroup initialization
4325 *
4326 * Register cgroup filesystem and /proc file, and initialize
4327 * any subsystems that didn't request early init.
4328 */
4329int __init cgroup_init(void)
4330{
4331        int err;
4332        int i;
4333        struct hlist_head *hhead;
4334
4335        err = bdi_init(&cgroup_backing_dev_info);
4336        if (err)
4337                return err;
4338
4339        /* at bootup time, we don't worry about modular subsystems */
4340        for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4341                struct cgroup_subsys *ss = subsys[i];
4342                if (!ss->early_init)
4343                        cgroup_init_subsys(ss);
4344                if (ss->use_id)
4345                        cgroup_init_idr(ss, init_css_set.subsys[ss->subsys_id]);
4346        }
4347
4348        /* Add init_css_set to the hash table */
4349        hhead = css_set_hash(init_css_set.subsys);
4350        hlist_add_head(&init_css_set.hlist, hhead);
4351        BUG_ON(!init_root_id(&rootnode));
4352
4353        cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4354        if (!cgroup_kobj) {
4355                err = -ENOMEM;
4356                goto out;
4357        }
4358
4359        err = register_filesystem(&cgroup_fs_type);
4360        if (err < 0) {
4361                kobject_put(cgroup_kobj);
4362                goto out;
4363        }
4364
4365        proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4366
4367out:
4368        if (err)
4369                bdi_destroy(&cgroup_backing_dev_info);
4370
4371        return err;
4372}
4373
4374/*
4375 * proc_cgroup_show()
4376 *  - Print task's cgroup paths into seq_file, one line for each hierarchy
4377 *  - Used for /proc/<pid>/cgroup.
4378 *  - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4379 *    doesn't really matter if tsk->cgroup changes after we read it,
4380 *    and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4381 *    anyway.  No need to check that tsk->cgroup != NULL, thanks to
4382 *    the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4383 *    cgroup to top_cgroup.
4384 */
4385
4386/* TODO: Use a proper seq_file iterator */
4387static int proc_cgroup_show(struct seq_file *m, void *v)
4388{
4389        struct pid *pid;
4390        struct task_struct *tsk;
4391        char *buf;
4392        int retval;
4393        struct cgroupfs_root *root;
4394
4395        retval = -ENOMEM;
4396        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4397        if (!buf)
4398                goto out;
4399
4400        retval = -ESRCH;
4401        pid = m->private;
4402        tsk = get_pid_task(pid, PIDTYPE_PID);
4403        if (!tsk)
4404                goto out_free;
4405
4406        retval = 0;
4407
4408        mutex_lock(&cgroup_mutex);
4409
4410        for_each_active_root(root) {
4411                struct cgroup_subsys *ss;
4412                struct cgroup *cgrp;
4413                int count = 0;
4414
4415                seq_printf(m, "%d:", root->hierarchy_id);
4416                for_each_subsys(root, ss)
4417                        seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4418                if (strlen(root->name))
4419                        seq_printf(m, "%sname=%s", count ? "," : "",
4420                                   root->name);
4421                seq_putc(m, ':');
4422                cgrp = task_cgroup_from_root(tsk, root);
4423                retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4424                if (retval < 0)
4425                        goto out_unlock;
4426                seq_puts(m, buf);
4427                seq_putc(m, '\n');
4428        }
4429
4430out_unlock:
4431        mutex_unlock(&cgroup_mutex);
4432        put_task_struct(tsk);
4433out_free:
4434        kfree(buf);
4435out:
4436        return retval;
4437}
4438
4439static int cgroup_open(struct inode *inode, struct file *file)
4440{
4441        struct pid *pid = PROC_I(inode)->pid;
4442        return single_open(file, proc_cgroup_show, pid);
4443}
4444
4445const struct file_operations proc_cgroup_operations = {
4446        .open           = cgroup_open,
4447        .read           = seq_read,
4448        .llseek         = seq_lseek,
4449        .release        = single_release,
4450};
4451
4452/* Display information about each subsystem and each hierarchy */
4453static int proc_cgroupstats_show(struct seq_file *m, void *v)
4454{
4455        int i;
4456
4457        seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4458        /*
4459         * ideally we don't want subsystems moving around while we do this.
4460         * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4461         * subsys/hierarchy state.
4462         */
4463        mutex_lock(&cgroup_mutex);
4464        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
4465                struct cgroup_subsys *ss = subsys[i];
4466                if (ss == NULL)
4467                        continue;
4468                seq_printf(m, "%s\t%d\t%d\t%d\n",
4469                           ss->name, ss->root->hierarchy_id,
4470                           ss->root->number_of_cgroups, !ss->disabled);
4471        }
4472        mutex_unlock(&cgroup_mutex);
4473        return 0;
4474}
4475
4476static int cgroupstats_open(struct inode *inode, struct file *file)
4477{
4478        return single_open(file, proc_cgroupstats_show, NULL);
4479}
4480
4481static const struct file_operations proc_cgroupstats_operations = {
4482        .open = cgroupstats_open,
4483        .read = seq_read,
4484        .llseek = seq_lseek,
4485        .release = single_release,
4486};
4487
4488/**
4489 * cgroup_fork - attach newly forked task to its parents cgroup.
4490 * @child: pointer to task_struct of forking parent process.
4491 *
4492 * Description: A task inherits its parent's cgroup at fork().
4493 *
4494 * A pointer to the shared css_set was automatically copied in
4495 * fork.c by dup_task_struct().  However, we ignore that copy, since
4496 * it was not made under the protection of RCU or cgroup_mutex, so
4497 * might no longer be a valid cgroup pointer.  cgroup_attach_task() might
4498 * have already changed current->cgroups, allowing the previously
4499 * referenced cgroup group to be removed and freed.
4500 *
4501 * At the point that cgroup_fork() is called, 'current' is the parent
4502 * task, and the passed argument 'child' points to the child task.
4503 */
4504void cgroup_fork(struct task_struct *child)
4505{
4506        task_lock(current);
4507        child->cgroups = current->cgroups;
4508        get_css_set(child->cgroups);
4509        task_unlock(current);
4510        INIT_LIST_HEAD(&child->cg_list);
4511}
4512
4513/**
4514 * cgroup_fork_callbacks - run fork callbacks
4515 * @child: the new task
4516 *
4517 * Called on a new task very soon before adding it to the
4518 * tasklist. No need to take any locks since no-one can
4519 * be operating on this task.
4520 */
4521void cgroup_fork_callbacks(struct task_struct *child)
4522{
4523        if (need_forkexit_callback) {
4524                int i;
4525                /*
4526                 * forkexit callbacks are only supported for builtin
4527                 * subsystems, and the builtin section of the subsys array is
4528                 * immutable, so we don't need to lock the subsys array here.
4529                 */
4530                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4531                        struct cgroup_subsys *ss = subsys[i];
4532                        if (ss->fork)
4533                                ss->fork(ss, child);
4534                }
4535        }
4536}
4537
4538/**
4539 * cgroup_post_fork - called on a new task after adding it to the task list
4540 * @child: the task in question
4541 *
4542 * Adds the task to the list running through its css_set if necessary.
4543 * Has to be after the task is visible on the task list in case we race
4544 * with the first call to cgroup_iter_start() - to guarantee that the
4545 * new task ends up on its list.
4546 */
4547void cgroup_post_fork(struct task_struct *child)
4548{
4549        if (use_task_css_set_links) {
4550                write_lock(&css_set_lock);
4551                task_lock(child);
4552                if (list_empty(&child->cg_list))
4553                        list_add(&child->cg_list, &child->cgroups->tasks);
4554                task_unlock(child);
4555                write_unlock(&css_set_lock);
4556        }
4557}
4558/**
4559 * cgroup_exit - detach cgroup from exiting task
4560 * @tsk: pointer to task_struct of exiting process
4561 * @run_callback: run exit callbacks?
4562 *
4563 * Description: Detach cgroup from @tsk and release it.
4564 *
4565 * Note that cgroups marked notify_on_release force every task in
4566 * them to take the global cgroup_mutex mutex when exiting.
4567 * This could impact scaling on very large systems.  Be reluctant to
4568 * use notify_on_release cgroups where very high task exit scaling
4569 * is required on large systems.
4570 *
4571 * the_top_cgroup_hack:
4572 *
4573 *    Set the exiting tasks cgroup to the root cgroup (top_cgroup).
4574 *
4575 *    We call cgroup_exit() while the task is still competent to
4576 *    handle notify_on_release(), then leave the task attached to the
4577 *    root cgroup in each hierarchy for the remainder of its exit.
4578 *
4579 *    To do this properly, we would increment the reference count on
4580 *    top_cgroup, and near the very end of the kernel/exit.c do_exit()
4581 *    code we would add a second cgroup function call, to drop that
4582 *    reference.  This would just create an unnecessary hot spot on
4583 *    the top_cgroup reference count, to no avail.
4584 *
4585 *    Normally, holding a reference to a cgroup without bumping its
4586 *    count is unsafe.   The cgroup could go away, or someone could
4587 *    attach us to a different cgroup, decrementing the count on
4588 *    the first cgroup that we never incremented.  But in this case,
4589 *    top_cgroup isn't going away, and either task has PF_EXITING set,
4590 *    which wards off any cgroup_attach_task() attempts, or task is a failed
4591 *    fork, never visible to cgroup_attach_task.
4592 */
4593void cgroup_exit(struct task_struct *tsk, int run_callbacks)
4594{
4595        struct css_set *cg;
4596        int i;
4597
4598        /*
4599         * Unlink from the css_set task list if necessary.
4600         * Optimistically check cg_list before taking
4601         * css_set_lock
4602         */
4603        if (!list_empty(&tsk->cg_list)) {
4604                write_lock(&css_set_lock);
4605                if (!list_empty(&tsk->cg_list))
4606                        list_del_init(&tsk->cg_list);
4607                write_unlock(&css_set_lock);
4608        }
4609
4610        /* Reassign the task to the init_css_set. */
4611        task_lock(tsk);
4612        cg = tsk->cgroups;
4613        tsk->cgroups = &init_css_set;
4614
4615        if (run_callbacks && need_forkexit_callback) {
4616                /*
4617                 * modular subsystems can't use callbacks, so no need to lock
4618                 * the subsys array
4619                 */
4620                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4621                        struct cgroup_subsys *ss = subsys[i];
4622                        if (ss->exit) {
4623                                struct cgroup *old_cgrp =
4624                                        rcu_dereference_raw(cg->subsys[i])->cgroup;
4625                                struct cgroup *cgrp = task_cgroup(tsk, i);
4626                                ss->exit(ss, cgrp, old_cgrp, tsk);
4627                        }
4628                }
4629        }
4630        task_unlock(tsk);
4631
4632        if (cg)
4633                put_css_set_taskexit(cg);
4634}
4635
4636/**
4637 * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
4638 * @cgrp: the cgroup in question
4639 * @task: the task in question
4640 *
4641 * See if @cgrp is a descendant of @task's cgroup in the appropriate
4642 * hierarchy.
4643 *
4644 * If we are sending in dummytop, then presumably we are creating
4645 * the top cgroup in the subsystem.
4646 *
4647 * Called only by the ns (nsproxy) cgroup.
4648 */
4649int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
4650{
4651        int ret;
4652        struct cgroup *target;
4653
4654        if (cgrp == dummytop)
4655                return 1;
4656
4657        target = task_cgroup_from_root(task, cgrp->root);
4658        while (cgrp != target && cgrp!= cgrp->top_cgroup)
4659                cgrp = cgrp->parent;
4660        ret = (cgrp == target);
4661        return ret;
4662}
4663
4664static void check_for_release(struct cgroup *cgrp)
4665{
4666        /* All of these checks rely on RCU to keep the cgroup
4667         * structure alive */
4668        if (cgroup_is_releasable(cgrp) && !atomic_read(&cgrp->count)
4669            && list_empty(&cgrp->children) && !cgroup_has_css_refs(cgrp)) {
4670                /* Control Group is currently removeable. If it's not
4671                 * already queued for a userspace notification, queue
4672                 * it now */
4673                int need_schedule_work = 0;
4674                spin_lock(&release_list_lock);
4675                if (!cgroup_is_removed(cgrp) &&
4676                    list_empty(&cgrp->release_list)) {
4677                        list_add(&cgrp->release_list, &release_list);
4678                        need_schedule_work = 1;
4679                }
4680                spin_unlock(&release_list_lock);
4681                if (need_schedule_work)
4682                        schedule_work(&release_agent_work);
4683        }
4684}
4685
4686/* Caller must verify that the css is not for root cgroup */
4687void __css_put(struct cgroup_subsys_state *css, int count)
4688{
4689        struct cgroup *cgrp = css->cgroup;
4690        int val;
4691        rcu_read_lock();
4692        val = atomic_sub_return(count, &css->refcnt);
4693        if (val == 1) {
4694                if (notify_on_release(cgrp)) {
4695                        set_bit(CGRP_RELEASABLE, &cgrp->flags);
4696                        check_for_release(cgrp);
4697                }
4698                cgroup_wakeup_rmdir_waiter(cgrp);
4699        }
4700        rcu_read_unlock();
4701        WARN_ON_ONCE(val < 1);
4702}
4703EXPORT_SYMBOL_GPL(__css_put);
4704
4705/*
4706 * Notify userspace when a cgroup is released, by running the
4707 * configured release agent with the name of the cgroup (path
4708 * relative to the root of cgroup file system) as the argument.
4709 *
4710 * Most likely, this user command will try to rmdir this cgroup.
4711 *
4712 * This races with the possibility that some other task will be
4713 * attached to this cgroup before it is removed, or that some other
4714 * user task will 'mkdir' a child cgroup of this cgroup.  That's ok.
4715 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
4716 * unused, and this cgroup will be reprieved from its death sentence,
4717 * to continue to serve a useful existence.  Next time it's released,
4718 * we will get notified again, if it still has 'notify_on_release' set.
4719 *
4720 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
4721 * means only wait until the task is successfully execve()'d.  The
4722 * separate release agent task is forked by call_usermodehelper(),
4723 * then control in this thread returns here, without waiting for the
4724 * release agent task.  We don't bother to wait because the caller of
4725 * this routine has no use for the exit status of the release agent
4726 * task, so no sense holding our caller up for that.
4727 */
4728static void cgroup_release_agent(struct work_struct *work)
4729{
4730        BUG_ON(work != &release_agent_work);
4731        mutex_lock(&cgroup_mutex);
4732        spin_lock(&release_list_lock);
4733        while (!list_empty(&release_list)) {
4734                char *argv[3], *envp[3];
4735                int i;
4736                char *pathbuf = NULL, *agentbuf = NULL;
4737                struct cgroup *cgrp = list_entry(release_list.next,
4738                                                    struct cgroup,
4739                                                    release_list);
4740                list_del_init(&cgrp->release_list);
4741                spin_unlock(&release_list_lock);
4742                pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4743                if (!pathbuf)
4744                        goto continue_free;
4745                if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
4746                        goto continue_free;
4747                agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
4748                if (!agentbuf)
4749                        goto continue_free;
4750
4751                i = 0;
4752                argv[i++] = agentbuf;
4753                argv[i++] = pathbuf;
4754                argv[i] = NULL;
4755
4756                i = 0;
4757                /* minimal command environment */
4758                envp[i++] = "HOME=/";
4759                envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
4760                envp[i] = NULL;
4761
4762                /* Drop the lock while we invoke the usermode helper,
4763                 * since the exec could involve hitting disk and hence
4764                 * be a slow process */
4765                mutex_unlock(&cgroup_mutex);
4766                call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
4767                mutex_lock(&cgroup_mutex);
4768 continue_free:
4769                kfree(pathbuf);
4770                kfree(agentbuf);
4771                spin_lock(&release_list_lock);
4772        }
4773        spin_unlock(&release_list_lock);
4774        mutex_unlock(&cgroup_mutex);
4775}
4776
4777static int __init cgroup_disable(char *str)
4778{
4779        int i;
4780        char *token;
4781
4782        while ((token = strsep(&str, ",")) != NULL) {
4783                if (!*token)
4784                        continue;
4785                /*
4786                 * cgroup_disable, being at boot time, can't know about module
4787                 * subsystems, so we don't worry about them.
4788                 */
4789                for (i = 0; i < CGROUP_BUILTIN_SUBSYS_COUNT; i++) {
4790                        struct cgroup_subsys *ss = subsys[i];
4791
4792                        if (!strcmp(token, ss->name)) {
4793                                ss->disabled = 1;
4794                                printk(KERN_INFO "Disabling %s control group"
4795                                        " subsystem\n", ss->name);
4796                                break;
4797                        }
4798                }
4799        }
4800        return 1;
4801}
4802__setup("cgroup_disable=", cgroup_disable);
4803
4804/*
4805 * Functons for CSS ID.
4806 */
4807
4808/*
4809 *To get ID other than 0, this should be called when !cgroup_is_removed().
4810 */
4811unsigned short css_id(struct cgroup_subsys_state *css)
4812{
4813        struct css_id *cssid;
4814
4815        /*
4816         * This css_id() can return correct value when somone has refcnt
4817         * on this or this is under rcu_read_lock(). Once css->id is allocated,
4818         * it's unchanged until freed.
4819         */
4820        cssid = rcu_dereference_check(css->id, atomic_read(&css->refcnt));
4821
4822        if (cssid)
4823                return cssid->id;
4824        return 0;
4825}
4826EXPORT_SYMBOL_GPL(css_id);
4827
4828unsigned short css_depth(struct cgroup_subsys_state *css)
4829{
4830        struct css_id *cssid;
4831
4832        cssid = rcu_dereference_check(css->id, atomic_read(&css->refcnt));
4833
4834        if (cssid)
4835                return cssid->depth;
4836        return 0;
4837}
4838EXPORT_SYMBOL_GPL(css_depth);
4839
4840/**
4841 *  css_is_ancestor - test "root" css is an ancestor of "child"
4842 * @child: the css to be tested.
4843 * @root: the css supporsed to be an ancestor of the child.
4844 *
4845 * Returns true if "root" is an ancestor of "child" in its hierarchy. Because
4846 * this function reads css->id, this use rcu_dereference() and rcu_read_lock().
4847 * But, considering usual usage, the csses should be valid objects after test.
4848 * Assuming that the caller will do some action to the child if this returns
4849 * returns true, the caller must take "child";s reference count.
4850 * If "child" is valid object and this returns true, "root" is valid, too.
4851 */
4852
4853bool css_is_ancestor(struct cgroup_subsys_state *child,
4854                    const struct cgroup_subsys_state *root)
4855{
4856        struct css_id *child_id;
4857        struct css_id *root_id;
4858        bool ret = true;
4859
4860        rcu_read_lock();
4861        child_id  = rcu_dereference(child->id);
4862        root_id = rcu_dereference(root->id);
4863        if (!child_id
4864            || !root_id
4865            || (child_id->depth < root_id->depth)
4866            || (child_id->stack[root_id->depth] != root_id->id))
4867                ret = false;
4868        rcu_read_unlock();
4869        return ret;
4870}
4871
4872void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
4873{
4874        struct css_id *id = css->id;
4875        /* When this is called before css_id initialization, id can be NULL */
4876        if (!id)
4877                return;
4878
4879        BUG_ON(!ss->use_id);
4880
4881        rcu_assign_pointer(id->css, NULL);
4882        rcu_assign_pointer(css->id, NULL);
4883        spin_lock(&ss->id_lock);
4884        idr_remove(&ss->idr, id->id);
4885        spin_unlock(&ss->id_lock);
4886        kfree_rcu(id, rcu_head);
4887}
4888EXPORT_SYMBOL_GPL(free_css_id);
4889
4890/*
4891 * This is called by init or create(). Then, calls to this function are
4892 * always serialized (By cgroup_mutex() at create()).
4893 */
4894
4895static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
4896{
4897        struct css_id *newid;
4898        int myid, error, size;
4899
4900        BUG_ON(!ss->use_id);
4901
4902        size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
4903        newid = kzalloc(size, GFP_KERNEL);
4904        if (!newid)
4905                return ERR_PTR(-ENOMEM);
4906        /* get id */
4907        if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
4908                error = -ENOMEM;
4909                goto err_out;
4910        }
4911        spin_lock(&ss->id_lock);
4912        /* Don't use 0. allocates an ID of 1-65535 */
4913        error = idr_get_new_above(&ss->idr, newid, 1, &myid);
4914        spin_unlock(&ss->id_lock);
4915
4916        /* Returns error when there are no free spaces for new ID.*/
4917        if (error) {
4918                error = -ENOSPC;
4919                goto err_out;
4920        }
4921        if (myid > CSS_ID_MAX)
4922                goto remove_idr;
4923
4924        newid->id = myid;
4925        newid->depth = depth;
4926        return newid;
4927remove_idr:
4928        error = -ENOSPC;
4929        spin_lock(&ss->id_lock);
4930        idr_remove(&ss->idr, myid);
4931        spin_unlock(&ss->id_lock);
4932err_out:
4933        kfree(newid);
4934        return ERR_PTR(error);
4935
4936}
4937
4938static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
4939                                            struct cgroup_subsys_state *rootcss)
4940{
4941        struct css_id *newid;
4942
4943        spin_lock_init(&ss->id_lock);
4944        idr_init(&ss->idr);
4945
4946        newid = get_new_cssid(ss, 0);
4947        if (IS_ERR(newid))
4948                return PTR_ERR(newid);
4949
4950        newid->stack[0] = newid->id;
4951        newid->css = rootcss;
4952        rootcss->id = newid;
4953        return 0;
4954}
4955
4956static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
4957                        struct cgroup *child)
4958{
4959        int subsys_id, i, depth = 0;
4960        struct cgroup_subsys_state *parent_css, *child_css;
4961        struct css_id *child_id, *parent_id;
4962
4963        subsys_id = ss->subsys_id;
4964        parent_css = parent->subsys[subsys_id];
4965        child_css = child->subsys[subsys_id];
4966        parent_id = parent_css->id;
4967        depth = parent_id->depth + 1;
4968
4969        child_id = get_new_cssid(ss, depth);
4970        if (IS_ERR(child_id))
4971                return PTR_ERR(child_id);
4972
4973        for (i = 0; i < depth; i++)
4974                child_id->stack[i] = parent_id->stack[i];
4975        child_id->stack[depth] = child_id->id;
4976        /*
4977         * child_id->css pointer will be set after this cgroup is available
4978         * see cgroup_populate_dir()
4979         */
4980        rcu_assign_pointer(child_css->id, child_id);
4981
4982        return 0;
4983}
4984
4985/**
4986 * css_lookup - lookup css by id
4987 * @ss: cgroup subsys to be looked into.
4988 * @id: the id
4989 *
4990 * Returns pointer to cgroup_subsys_state if there is valid one with id.
4991 * NULL if not. Should be called under rcu_read_lock()
4992 */
4993struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
4994{
4995        struct css_id *cssid = NULL;
4996
4997        BUG_ON(!ss->use_id);
4998        cssid = idr_find(&ss->idr, id);
4999
5000        if (unlikely(!cssid))
5001                return NULL;
5002
5003        return rcu_dereference(cssid->css);
5004}
5005EXPORT_SYMBOL_GPL(css_lookup);
5006
5007/**
5008 * css_get_next - lookup next cgroup under specified hierarchy.
5009 * @ss: pointer to subsystem
5010 * @id: current position of iteration.
5011 * @root: pointer to css. search tree under this.
5012 * @foundid: position of found object.
5013 *
5014 * Search next css under the specified hierarchy of rootid. Calling under
5015 * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
5016 */
5017struct cgroup_subsys_state *
5018css_get_next(struct cgroup_subsys *ss, int id,
5019             struct cgroup_subsys_state *root, int *foundid)
5020{
5021        struct cgroup_subsys_state *ret = NULL;
5022        struct css_id *tmp;
5023        int tmpid;
5024        int rootid = css_id(root);
5025        int depth = css_depth(root);
5026
5027        if (!rootid)
5028                return NULL;
5029
5030        BUG_ON(!ss->use_id);
5031        /* fill start point for scan */
5032        tmpid = id;
5033        while (1) {
5034                /*
5035                 * scan next entry from bitmap(tree), tmpid is updated after
5036                 * idr_get_next().
5037                 */
5038                spin_lock(&ss->id_lock);
5039                tmp = idr_get_next(&ss->idr, &tmpid);
5040                spin_unlock(&ss->id_lock);
5041
5042                if (!tmp)
5043                        break;
5044                if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
5045                        ret = rcu_dereference(tmp->css);
5046                        if (ret) {
5047                                *foundid = tmpid;
5048                                break;
5049                        }
5050                }
5051                /* continue to scan from next id */
5052                tmpid = tmpid + 1;
5053        }
5054        return ret;
5055}
5056
5057/*
5058 * get corresponding css from file open on cgroupfs directory
5059 */
5060struct cgroup_subsys_state *cgroup_css_from_dir(struct file *f, int id)
5061{
5062        struct cgroup *cgrp;
5063        struct inode *inode;
5064        struct cgroup_subsys_state *css;
5065
5066        inode = f->f_dentry->d_inode;
5067        /* check in cgroup filesystem dir */
5068        if (inode->i_op != &cgroup_dir_inode_operations)
5069                return ERR_PTR(-EBADF);
5070
5071        if (id < 0 || id >= CGROUP_SUBSYS_COUNT)
5072                return ERR_PTR(-EINVAL);
5073
5074        /* get cgroup */
5075        cgrp = __d_cgrp(f->f_dentry);
5076        css = cgrp->subsys[id];
5077        return css ? css : ERR_PTR(-ENOENT);
5078}
5079
5080#ifdef CONFIG_CGROUP_DEBUG
5081static struct cgroup_subsys_state *debug_create(struct cgroup_subsys *ss,
5082                                                   struct cgroup *cont)
5083{
5084        struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5085
5086        if (!css)
5087                return ERR_PTR(-ENOMEM);
5088
5089        return css;
5090}
5091
5092static void debug_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
5093{
5094        kfree(cont->subsys[debug_subsys_id]);
5095}
5096
5097static u64 cgroup_refcount_read(struct cgroup *cont, struct cftype *cft)
5098{
5099        return atomic_read(&cont->count);
5100}
5101
5102static u64 debug_taskcount_read(struct cgroup *cont, struct cftype *cft)
5103{
5104        return cgroup_task_count(cont);
5105}
5106
5107static u64 current_css_set_read(struct cgroup *cont, struct cftype *cft)
5108{
5109        return (u64)(unsigned long)current->cgroups;
5110}
5111
5112static u64 current_css_set_refcount_read(struct cgroup *cont,
5113                                           struct cftype *cft)
5114{
5115        u64 count;
5116
5117        rcu_read_lock();
5118        count = atomic_read(&current->cgroups->refcount);
5119        rcu_read_unlock();
5120        return count;
5121}
5122
5123static int current_css_set_cg_links_read(struct cgroup *cont,
5124                                         struct cftype *cft,
5125                                         struct seq_file *seq)
5126{
5127        struct cg_cgroup_link *link;
5128        struct css_set *cg;
5129
5130        read_lock(&css_set_lock);
5131        rcu_read_lock();
5132        cg = rcu_dereference(current->cgroups);
5133        list_for_each_entry(link, &cg->cg_links, cg_link_list) {
5134                struct cgroup *c = link->cgrp;
5135                const char *name;
5136
5137                if (c->dentry)
5138                        name = c->dentry->d_name.name;
5139                else
5140                        name = "?";
5141                seq_printf(seq, "Root %d group %s\n",
5142                           c->root->hierarchy_id, name);
5143        }
5144        rcu_read_unlock();
5145        read_unlock(&css_set_lock);
5146        return 0;
5147}
5148
5149#define MAX_TASKS_SHOWN_PER_CSS 25
5150static int cgroup_css_links_read(struct cgroup *cont,
5151                                 struct cftype *cft,
5152                                 struct seq_file *seq)
5153{
5154        struct cg_cgroup_link *link;
5155
5156        read_lock(&css_set_lock);
5157        list_for_each_entry(link, &cont->css_sets, cgrp_link_list) {
5158                struct css_set *cg = link->cg;
5159                struct task_struct *task;
5160                int count = 0;
5161                seq_printf(seq, "css_set %p\n", cg);
5162                list_for_each_entry(task, &cg->tasks, cg_list) {
5163                        if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5164                                seq_puts(seq, "  ...\n");
5165                                break;
5166                        } else {
5167                                seq_printf(seq, "  task %d\n",
5168                                           task_pid_vnr(task));
5169                        }
5170                }
5171        }
5172        read_unlock(&css_set_lock);
5173        return 0;
5174}
5175
5176static u64 releasable_read(struct cgroup *cgrp, struct cftype *cft)
5177{
5178        return test_bit(CGRP_RELEASABLE, &cgrp->flags);
5179}
5180
5181static struct cftype debug_files[] =  {
5182        {
5183                .name = "cgroup_refcount",
5184                .read_u64 = cgroup_refcount_read,
5185        },
5186        {
5187                .name = "taskcount",
5188                .read_u64 = debug_taskcount_read,
5189        },
5190
5191        {
5192                .name = "current_css_set",
5193                .read_u64 = current_css_set_read,
5194        },
5195
5196        {
5197                .name = "current_css_set_refcount",
5198                .read_u64 = current_css_set_refcount_read,
5199        },
5200
5201        {
5202                .name = "current_css_set_cg_links",
5203                .read_seq_string = current_css_set_cg_links_read,
5204        },
5205
5206        {
5207                .name = "cgroup_css_links",
5208                .read_seq_string = cgroup_css_links_read,
5209        },
5210
5211        {
5212                .name = "releasable",
5213                .read_u64 = releasable_read,
5214        },
5215};
5216
5217static int debug_populate(struct cgroup_subsys *ss, struct cgroup *cont)
5218{
5219        return cgroup_add_files(cont, ss, debug_files,
5220                                ARRAY_SIZE(debug_files));
5221}
5222
5223struct cgroup_subsys debug_subsys = {
5224        .name = "debug",
5225        .create = debug_create,
5226        .destroy = debug_destroy,
5227        .populate = debug_populate,
5228        .subsys_id = debug_subsys_id,
5229};
5230#endif /* CONFIG_CGROUP_DEBUG */
5231
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.