linux/mm/memcontrol.c
<<
>>
Prefs
   1/* memcontrol.c - Memory Controller
   2 *
   3 * Copyright IBM Corporation, 2007
   4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
   5 *
   6 * Copyright 2007 OpenVZ SWsoft Inc
   7 * Author: Pavel Emelianov <xemul@openvz.org>
   8 *
   9 * Memory thresholds
  10 * Copyright (C) 2009 Nokia Corporation
  11 * Author: Kirill A. Shutemov
  12 *
  13 * Kernel Memory Controller
  14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
  15 * Authors: Glauber Costa and Suleiman Souhlal
  16 *
  17 * This program is free software; you can redistribute it and/or modify
  18 * it under the terms of the GNU General Public License as published by
  19 * the Free Software Foundation; either version 2 of the License, or
  20 * (at your option) any later version.
  21 *
  22 * This program is distributed in the hope that it will be useful,
  23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  25 * GNU General Public License for more details.
  26 */
  27
  28#include <linux/res_counter.h>
  29#include <linux/memcontrol.h>
  30#include <linux/cgroup.h>
  31#include <linux/mm.h>
  32#include <linux/hugetlb.h>
  33#include <linux/pagemap.h>
  34#include <linux/smp.h>
  35#include <linux/page-flags.h>
  36#include <linux/backing-dev.h>
  37#include <linux/bit_spinlock.h>
  38#include <linux/rcupdate.h>
  39#include <linux/limits.h>
  40#include <linux/export.h>
  41#include <linux/mutex.h>
  42#include <linux/rbtree.h>
  43#include <linux/slab.h>
  44#include <linux/swap.h>
  45#include <linux/swapops.h>
  46#include <linux/spinlock.h>
  47#include <linux/eventfd.h>
  48#include <linux/sort.h>
  49#include <linux/fs.h>
  50#include <linux/seq_file.h>
  51#include <linux/vmalloc.h>
  52#include <linux/mm_inline.h>
  53#include <linux/page_cgroup.h>
  54#include <linux/cpu.h>
  55#include <linux/oom.h>
  56#include "internal.h"
  57#include <net/sock.h>
  58#include <net/ip.h>
  59#include <net/tcp_memcontrol.h>
  60
  61#include <asm/uaccess.h>
  62
  63#include <trace/events/vmscan.h>
  64
  65struct cgroup_subsys mem_cgroup_subsys __read_mostly;
  66EXPORT_SYMBOL(mem_cgroup_subsys);
  67
  68#define MEM_CGROUP_RECLAIM_RETRIES      5
  69static struct mem_cgroup *root_mem_cgroup __read_mostly;
  70
  71#ifdef CONFIG_MEMCG_SWAP
  72/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
  73int do_swap_account __read_mostly;
  74
  75/* for remember boot option*/
  76#ifdef CONFIG_MEMCG_SWAP_ENABLED
  77static int really_do_swap_account __initdata = 1;
  78#else
  79static int really_do_swap_account __initdata = 0;
  80#endif
  81
  82#else
  83#define do_swap_account         0
  84#endif
  85
  86
  87/*
  88 * Statistics for memory cgroup.
  89 */
  90enum mem_cgroup_stat_index {
  91        /*
  92         * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  93         */
  94        MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
  95        MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
  96        MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
  97        MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
  98        MEM_CGROUP_STAT_NSTATS,
  99};
 100
 101static const char * const mem_cgroup_stat_names[] = {
 102        "cache",
 103        "rss",
 104        "mapped_file",
 105        "swap",
 106};
 107
 108enum mem_cgroup_events_index {
 109        MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
 110        MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
 111        MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
 112        MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
 113        MEM_CGROUP_EVENTS_NSTATS,
 114};
 115
 116static const char * const mem_cgroup_events_names[] = {
 117        "pgpgin",
 118        "pgpgout",
 119        "pgfault",
 120        "pgmajfault",
 121};
 122
 123static const char * const mem_cgroup_lru_names[] = {
 124        "inactive_anon",
 125        "active_anon",
 126        "inactive_file",
 127        "active_file",
 128        "unevictable",
 129};
 130
 131/*
 132 * Per memcg event counter is incremented at every pagein/pageout. With THP,
 133 * it will be incremated by the number of pages. This counter is used for
 134 * for trigger some periodic events. This is straightforward and better
 135 * than using jiffies etc. to handle periodic memcg event.
 136 */
 137enum mem_cgroup_events_target {
 138        MEM_CGROUP_TARGET_THRESH,
 139        MEM_CGROUP_TARGET_SOFTLIMIT,
 140        MEM_CGROUP_TARGET_NUMAINFO,
 141        MEM_CGROUP_NTARGETS,
 142};
 143#define THRESHOLDS_EVENTS_TARGET 128
 144#define SOFTLIMIT_EVENTS_TARGET 1024
 145#define NUMAINFO_EVENTS_TARGET  1024
 146
 147struct mem_cgroup_stat_cpu {
 148        long count[MEM_CGROUP_STAT_NSTATS];
 149        unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
 150        unsigned long nr_page_events;
 151        unsigned long targets[MEM_CGROUP_NTARGETS];
 152};
 153
 154struct mem_cgroup_reclaim_iter {
 155        /* css_id of the last scanned hierarchy member */
 156        int position;
 157        /* scan generation, increased every round-trip */
 158        unsigned int generation;
 159};
 160
 161/*
 162 * per-zone information in memory controller.
 163 */
 164struct mem_cgroup_per_zone {
 165        struct lruvec           lruvec;
 166        unsigned long           lru_size[NR_LRU_LISTS];
 167
 168        struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
 169
 170        struct rb_node          tree_node;      /* RB tree node */
 171        unsigned long long      usage_in_excess;/* Set to the value by which */
 172                                                /* the soft limit is exceeded*/
 173        bool                    on_tree;
 174        struct mem_cgroup       *memcg;         /* Back pointer, we cannot */
 175                                                /* use container_of        */
 176};
 177
 178struct mem_cgroup_per_node {
 179        struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 180};
 181
 182struct mem_cgroup_lru_info {
 183        struct mem_cgroup_per_node *nodeinfo[0];
 184};
 185
 186/*
 187 * Cgroups above their limits are maintained in a RB-Tree, independent of
 188 * their hierarchy representation
 189 */
 190
 191struct mem_cgroup_tree_per_zone {
 192        struct rb_root rb_root;
 193        spinlock_t lock;
 194};
 195
 196struct mem_cgroup_tree_per_node {
 197        struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
 198};
 199
 200struct mem_cgroup_tree {
 201        struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
 202};
 203
 204static struct mem_cgroup_tree soft_limit_tree __read_mostly;
 205
 206struct mem_cgroup_threshold {
 207        struct eventfd_ctx *eventfd;
 208        u64 threshold;
 209};
 210
 211/* For threshold */
 212struct mem_cgroup_threshold_ary {
 213        /* An array index points to threshold just below or equal to usage. */
 214        int current_threshold;
 215        /* Size of entries[] */
 216        unsigned int size;
 217        /* Array of thresholds */
 218        struct mem_cgroup_threshold entries[0];
 219};
 220
 221struct mem_cgroup_thresholds {
 222        /* Primary thresholds array */
 223        struct mem_cgroup_threshold_ary *primary;
 224        /*
 225         * Spare threshold array.
 226         * This is needed to make mem_cgroup_unregister_event() "never fail".
 227         * It must be able to store at least primary->size - 1 entries.
 228         */
 229        struct mem_cgroup_threshold_ary *spare;
 230};
 231
 232/* for OOM */
 233struct mem_cgroup_eventfd_list {
 234        struct list_head list;
 235        struct eventfd_ctx *eventfd;
 236};
 237
 238static void mem_cgroup_threshold(struct mem_cgroup *memcg);
 239static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
 240
 241/*
 242 * The memory controller data structure. The memory controller controls both
 243 * page cache and RSS per cgroup. We would eventually like to provide
 244 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 245 * to help the administrator determine what knobs to tune.
 246 *
 247 * TODO: Add a water mark for the memory controller. Reclaim will begin when
 248 * we hit the water mark. May be even add a low water mark, such that
 249 * no reclaim occurs from a cgroup at it's low water mark, this is
 250 * a feature that will be implemented much later in the future.
 251 */
 252struct mem_cgroup {
 253        struct cgroup_subsys_state css;
 254        /*
 255         * the counter to account for memory usage
 256         */
 257        struct res_counter res;
 258
 259        union {
 260                /*
 261                 * the counter to account for mem+swap usage.
 262                 */
 263                struct res_counter memsw;
 264
 265                /*
 266                 * rcu_freeing is used only when freeing struct mem_cgroup,
 267                 * so put it into a union to avoid wasting more memory.
 268                 * It must be disjoint from the css field.  It could be
 269                 * in a union with the res field, but res plays a much
 270                 * larger part in mem_cgroup life than memsw, and might
 271                 * be of interest, even at time of free, when debugging.
 272                 * So share rcu_head with the less interesting memsw.
 273                 */
 274                struct rcu_head rcu_freeing;
 275                /*
 276                 * We also need some space for a worker in deferred freeing.
 277                 * By the time we call it, rcu_freeing is no longer in use.
 278                 */
 279                struct work_struct work_freeing;
 280        };
 281
 282        /*
 283         * the counter to account for kernel memory usage.
 284         */
 285        struct res_counter kmem;
 286        /*
 287         * Should the accounting and control be hierarchical, per subtree?
 288         */
 289        bool use_hierarchy;
 290        unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
 291
 292        bool            oom_lock;
 293        atomic_t        under_oom;
 294
 295        atomic_t        refcnt;
 296
 297        int     swappiness;
 298        /* OOM-Killer disable */
 299        int             oom_kill_disable;
 300
 301        /* set when res.limit == memsw.limit */
 302        bool            memsw_is_minimum;
 303
 304        /* protect arrays of thresholds */
 305        struct mutex thresholds_lock;
 306
 307        /* thresholds for memory usage. RCU-protected */
 308        struct mem_cgroup_thresholds thresholds;
 309
 310        /* thresholds for mem+swap usage. RCU-protected */
 311        struct mem_cgroup_thresholds memsw_thresholds;
 312
 313        /* For oom notifier event fd */
 314        struct list_head oom_notify;
 315
 316        /*
 317         * Should we move charges of a task when a task is moved into this
 318         * mem_cgroup ? And what type of charges should we move ?
 319         */
 320        unsigned long   move_charge_at_immigrate;
 321        /*
 322         * set > 0 if pages under this cgroup are moving to other cgroup.
 323         */
 324        atomic_t        moving_account;
 325        /* taken only while moving_account > 0 */
 326        spinlock_t      move_lock;
 327        /*
 328         * percpu counter.
 329         */
 330        struct mem_cgroup_stat_cpu __percpu *stat;
 331        /*
 332         * used when a cpu is offlined or other synchronizations
 333         * See mem_cgroup_read_stat().
 334         */
 335        struct mem_cgroup_stat_cpu nocpu_base;
 336        spinlock_t pcp_counter_lock;
 337
 338#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
 339        struct tcp_memcontrol tcp_mem;
 340#endif
 341#if defined(CONFIG_MEMCG_KMEM)
 342        /* analogous to slab_common's slab_caches list. per-memcg */
 343        struct list_head memcg_slab_caches;
 344        /* Not a spinlock, we can take a lot of time walking the list */
 345        struct mutex slab_caches_mutex;
 346        /* Index in the kmem_cache->memcg_params->memcg_caches array */
 347        int kmemcg_id;
 348#endif
 349
 350        int last_scanned_node;
 351#if MAX_NUMNODES > 1
 352        nodemask_t      scan_nodes;
 353        atomic_t        numainfo_events;
 354        atomic_t        numainfo_updating;
 355#endif
 356        /*
 357         * Per cgroup active and inactive list, similar to the
 358         * per zone LRU lists.
 359         *
 360         * WARNING: This has to be the last element of the struct. Don't
 361         * add new fields after this point.
 362         */
 363        struct mem_cgroup_lru_info info;
 364};
 365
 366static size_t memcg_size(void)
 367{
 368        return sizeof(struct mem_cgroup) +
 369                nr_node_ids * sizeof(struct mem_cgroup_per_node);
 370}
 371
 372/* internal only representation about the status of kmem accounting. */
 373enum {
 374        KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
 375        KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
 376        KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
 377};
 378
 379/* We account when limit is on, but only after call sites are patched */
 380#define KMEM_ACCOUNTED_MASK \
 381                ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
 382
 383#ifdef CONFIG_MEMCG_KMEM
 384static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
 385{
 386        set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 387}
 388
 389static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
 390{
 391        return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 392}
 393
 394static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
 395{
 396        set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 397}
 398
 399static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
 400{
 401        clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 402}
 403
 404static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
 405{
 406        if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
 407                set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
 408}
 409
 410static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
 411{
 412        return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
 413                                  &memcg->kmem_account_flags);
 414}
 415#endif
 416
 417/* Stuffs for move charges at task migration. */
 418/*
 419 * Types of charges to be moved. "move_charge_at_immitgrate" and
 420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
 421 */
 422enum move_type {
 423        MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
 424        MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
 425        NR_MOVE_TYPE,
 426};
 427
 428/* "mc" and its members are protected by cgroup_mutex */
 429static struct move_charge_struct {
 430        spinlock_t        lock; /* for from, to */
 431        struct mem_cgroup *from;
 432        struct mem_cgroup *to;
 433        unsigned long immigrate_flags;
 434        unsigned long precharge;
 435        unsigned long moved_charge;
 436        unsigned long moved_swap;
 437        struct task_struct *moving_task;        /* a task moving charges */
 438        wait_queue_head_t waitq;                /* a waitq for other context */
 439} mc = {
 440        .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
 441        .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
 442};
 443
 444static bool move_anon(void)
 445{
 446        return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
 447}
 448
 449static bool move_file(void)
 450{
 451        return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
 452}
 453
 454/*
 455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
 456 * limit reclaim to prevent infinite loops, if they ever occur.
 457 */
 458#define MEM_CGROUP_MAX_RECLAIM_LOOPS            100
 459#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
 460
 461enum charge_type {
 462        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 463        MEM_CGROUP_CHARGE_TYPE_ANON,
 464        MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
 465        MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
 466        NR_CHARGE_TYPE,
 467};
 468
 469/* for encoding cft->private value on file */
 470enum res_type {
 471        _MEM,
 472        _MEMSWAP,
 473        _OOM_TYPE,
 474        _KMEM,
 475};
 476
 477#define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
 478#define MEMFILE_TYPE(val)       ((val) >> 16 & 0xffff)
 479#define MEMFILE_ATTR(val)       ((val) & 0xffff)
 480/* Used for OOM nofiier */
 481#define OOM_CONTROL             (0)
 482
 483/*
 484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
 485 */
 486#define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
 487#define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
 488#define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
 489#define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
 490
 491/*
 492 * The memcg_create_mutex will be held whenever a new cgroup is created.
 493 * As a consequence, any change that needs to protect against new child cgroups
 494 * appearing has to hold it as well.
 495 */
 496static DEFINE_MUTEX(memcg_create_mutex);
 497
 498static void mem_cgroup_get(struct mem_cgroup *memcg);
 499static void mem_cgroup_put(struct mem_cgroup *memcg);
 500
 501static inline
 502struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
 503{
 504        return container_of(s, struct mem_cgroup, css);
 505}
 506
 507static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
 508{
 509        return (memcg == root_mem_cgroup);
 510}
 511
 512/* Writing them here to avoid exposing memcg's inner layout */
 513#if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
 514
 515void sock_update_memcg(struct sock *sk)
 516{
 517        if (mem_cgroup_sockets_enabled) {
 518                struct mem_cgroup *memcg;
 519                struct cg_proto *cg_proto;
 520
 521                BUG_ON(!sk->sk_prot->proto_cgroup);
 522
 523                /* Socket cloning can throw us here with sk_cgrp already
 524                 * filled. It won't however, necessarily happen from
 525                 * process context. So the test for root memcg given
 526                 * the current task's memcg won't help us in this case.
 527                 *
 528                 * Respecting the original socket's memcg is a better
 529                 * decision in this case.
 530                 */
 531                if (sk->sk_cgrp) {
 532                        BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
 533                        mem_cgroup_get(sk->sk_cgrp->memcg);
 534                        return;
 535                }
 536
 537                rcu_read_lock();
 538                memcg = mem_cgroup_from_task(current);
 539                cg_proto = sk->sk_prot->proto_cgroup(memcg);
 540                if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
 541                        mem_cgroup_get(memcg);
 542                        sk->sk_cgrp = cg_proto;
 543                }
 544                rcu_read_unlock();
 545        }
 546}
 547EXPORT_SYMBOL(sock_update_memcg);
 548
 549void sock_release_memcg(struct sock *sk)
 550{
 551        if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
 552                struct mem_cgroup *memcg;
 553                WARN_ON(!sk->sk_cgrp->memcg);
 554                memcg = sk->sk_cgrp->memcg;
 555                mem_cgroup_put(memcg);
 556        }
 557}
 558
 559struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
 560{
 561        if (!memcg || mem_cgroup_is_root(memcg))
 562                return NULL;
 563
 564        return &memcg->tcp_mem.cg_proto;
 565}
 566EXPORT_SYMBOL(tcp_proto_cgroup);
 567
 568static void disarm_sock_keys(struct mem_cgroup *memcg)
 569{
 570        if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
 571                return;
 572        static_key_slow_dec(&memcg_socket_limit_enabled);
 573}
 574#else
 575static void disarm_sock_keys(struct mem_cgroup *memcg)
 576{
 577}
 578#endif
 579
 580#ifdef CONFIG_MEMCG_KMEM
 581/*
 582 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
 583 * There are two main reasons for not using the css_id for this:
 584 *  1) this works better in sparse environments, where we have a lot of memcgs,
 585 *     but only a few kmem-limited. Or also, if we have, for instance, 200
 586 *     memcgs, and none but the 200th is kmem-limited, we'd have to have a
 587 *     200 entry array for that.
 588 *
 589 *  2) In order not to violate the cgroup API, we would like to do all memory
 590 *     allocation in ->create(). At that point, we haven't yet allocated the
 591 *     css_id. Having a separate index prevents us from messing with the cgroup
 592 *     core for this
 593 *
 594 * The current size of the caches array is stored in
 595 * memcg_limited_groups_array_size.  It will double each time we have to
 596 * increase it.
 597 */
 598static DEFINE_IDA(kmem_limited_groups);
 599int memcg_limited_groups_array_size;
 600
 601/*
 602 * MIN_SIZE is different than 1, because we would like to avoid going through
 603 * the alloc/free process all the time. In a small machine, 4 kmem-limited
 604 * cgroups is a reasonable guess. In the future, it could be a parameter or
 605 * tunable, but that is strictly not necessary.
 606 *
 607 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
 608 * this constant directly from cgroup, but it is understandable that this is
 609 * better kept as an internal representation in cgroup.c. In any case, the
 610 * css_id space is not getting any smaller, and we don't have to necessarily
 611 * increase ours as well if it increases.
 612 */
 613#define MEMCG_CACHES_MIN_SIZE 4
 614#define MEMCG_CACHES_MAX_SIZE 65535
 615
 616/*
 617 * A lot of the calls to the cache allocation functions are expected to be
 618 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
 619 * conditional to this static branch, we'll have to allow modules that does
 620 * kmem_cache_alloc and the such to see this symbol as well
 621 */
 622struct static_key memcg_kmem_enabled_key;
 623EXPORT_SYMBOL(memcg_kmem_enabled_key);
 624
 625static void disarm_kmem_keys(struct mem_cgroup *memcg)
 626{
 627        if (memcg_kmem_is_active(memcg)) {
 628                static_key_slow_dec(&memcg_kmem_enabled_key);
 629                ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
 630        }
 631        /*
 632         * This check can't live in kmem destruction function,
 633         * since the charges will outlive the cgroup
 634         */
 635        WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
 636}
 637#else
 638static void disarm_kmem_keys(struct mem_cgroup *memcg)
 639{
 640}
 641#endif /* CONFIG_MEMCG_KMEM */
 642
 643static void disarm_static_keys(struct mem_cgroup *memcg)
 644{
 645        disarm_sock_keys(memcg);
 646        disarm_kmem_keys(memcg);
 647}
 648
 649static void drain_all_stock_async(struct mem_cgroup *memcg);
 650
 651static struct mem_cgroup_per_zone *
 652mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
 653{
 654        VM_BUG_ON((unsigned)nid >= nr_node_ids);
 655        return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
 656}
 657
 658struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
 659{
 660        return &memcg->css;
 661}
 662
 663static struct mem_cgroup_per_zone *
 664page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
 665{
 666        int nid = page_to_nid(page);
 667        int zid = page_zonenum(page);
 668
 669        return mem_cgroup_zoneinfo(memcg, nid, zid);
 670}
 671
 672static struct mem_cgroup_tree_per_zone *
 673soft_limit_tree_node_zone(int nid, int zid)
 674{
 675        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 676}
 677
 678static struct mem_cgroup_tree_per_zone *
 679soft_limit_tree_from_page(struct page *page)
 680{
 681        int nid = page_to_nid(page);
 682        int zid = page_zonenum(page);
 683
 684        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 685}
 686
 687static void
 688__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
 689                                struct mem_cgroup_per_zone *mz,
 690                                struct mem_cgroup_tree_per_zone *mctz,
 691                                unsigned long long new_usage_in_excess)
 692{
 693        struct rb_node **p = &mctz->rb_root.rb_node;
 694        struct rb_node *parent = NULL;
 695        struct mem_cgroup_per_zone *mz_node;
 696
 697        if (mz->on_tree)
 698                return;
 699
 700        mz->usage_in_excess = new_usage_in_excess;
 701        if (!mz->usage_in_excess)
 702                return;
 703        while (*p) {
 704                parent = *p;
 705                mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
 706                                        tree_node);
 707                if (mz->usage_in_excess < mz_node->usage_in_excess)
 708                        p = &(*p)->rb_left;
 709                /*
 710                 * We can't avoid mem cgroups that are over their soft
 711                 * limit by the same amount
 712                 */
 713                else if (mz->usage_in_excess >= mz_node->usage_in_excess)
 714                        p = &(*p)->rb_right;
 715        }
 716        rb_link_node(&mz->tree_node, parent, p);
 717        rb_insert_color(&mz->tree_node, &mctz->rb_root);
 718        mz->on_tree = true;
 719}
 720
 721static void
 722__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 723                                struct mem_cgroup_per_zone *mz,
 724                                struct mem_cgroup_tree_per_zone *mctz)
 725{
 726        if (!mz->on_tree)
 727                return;
 728        rb_erase(&mz->tree_node, &mctz->rb_root);
 729        mz->on_tree = false;
 730}
 731
 732static void
 733mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 734                                struct mem_cgroup_per_zone *mz,
 735                                struct mem_cgroup_tree_per_zone *mctz)
 736{
 737        spin_lock(&mctz->lock);
 738        __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 739        spin_unlock(&mctz->lock);
 740}
 741
 742
 743static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
 744{
 745        unsigned long long excess;
 746        struct mem_cgroup_per_zone *mz;
 747        struct mem_cgroup_tree_per_zone *mctz;
 748        int nid = page_to_nid(page);
 749        int zid = page_zonenum(page);
 750        mctz = soft_limit_tree_from_page(page);
 751
 752        /*
 753         * Necessary to update all ancestors when hierarchy is used.
 754         * because their event counter is not touched.
 755         */
 756        for (; memcg; memcg = parent_mem_cgroup(memcg)) {
 757                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 758                excess = res_counter_soft_limit_excess(&memcg->res);
 759                /*
 760                 * We have to update the tree if mz is on RB-tree or
 761                 * mem is over its softlimit.
 762                 */
 763                if (excess || mz->on_tree) {
 764                        spin_lock(&mctz->lock);
 765                        /* if on-tree, remove it */
 766                        if (mz->on_tree)
 767                                __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 768                        /*
 769                         * Insert again. mz->usage_in_excess will be updated.
 770                         * If excess is 0, no tree ops.
 771                         */
 772                        __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
 773                        spin_unlock(&mctz->lock);
 774                }
 775        }
 776}
 777
 778static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
 779{
 780        int node, zone;
 781        struct mem_cgroup_per_zone *mz;
 782        struct mem_cgroup_tree_per_zone *mctz;
 783
 784        for_each_node(node) {
 785                for (zone = 0; zone < MAX_NR_ZONES; zone++) {
 786                        mz = mem_cgroup_zoneinfo(memcg, node, zone);
 787                        mctz = soft_limit_tree_node_zone(node, zone);
 788                        mem_cgroup_remove_exceeded(memcg, mz, mctz);
 789                }
 790        }
 791}
 792
 793static struct mem_cgroup_per_zone *
 794__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 795{
 796        struct rb_node *rightmost = NULL;
 797        struct mem_cgroup_per_zone *mz;
 798
 799retry:
 800        mz = NULL;
 801        rightmost = rb_last(&mctz->rb_root);
 802        if (!rightmost)
 803                goto done;              /* Nothing to reclaim from */
 804
 805        mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
 806        /*
 807         * Remove the node now but someone else can add it back,
 808         * we will to add it back at the end of reclaim to its correct
 809         * position in the tree.
 810         */
 811        __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
 812        if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
 813                !css_tryget(&mz->memcg->css))
 814                goto retry;
 815done:
 816        return mz;
 817}
 818
 819static struct mem_cgroup_per_zone *
 820mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 821{
 822        struct mem_cgroup_per_zone *mz;
 823
 824        spin_lock(&mctz->lock);
 825        mz = __mem_cgroup_largest_soft_limit_node(mctz);
 826        spin_unlock(&mctz->lock);
 827        return mz;
 828}
 829
 830/*
 831 * Implementation Note: reading percpu statistics for memcg.
 832 *
 833 * Both of vmstat[] and percpu_counter has threshold and do periodic
 834 * synchronization to implement "quick" read. There are trade-off between
 835 * reading cost and precision of value. Then, we may have a chance to implement
 836 * a periodic synchronizion of counter in memcg's counter.
 837 *
 838 * But this _read() function is used for user interface now. The user accounts
 839 * memory usage by memory cgroup and he _always_ requires exact value because
 840 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
 841 * have to visit all online cpus and make sum. So, for now, unnecessary
 842 * synchronization is not implemented. (just implemented for cpu hotplug)
 843 *
 844 * If there are kernel internal actions which can make use of some not-exact
 845 * value, and reading all cpu value can be performance bottleneck in some
 846 * common workload, threashold and synchonization as vmstat[] should be
 847 * implemented.
 848 */
 849static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
 850                                 enum mem_cgroup_stat_index idx)
 851{
 852        long val = 0;
 853        int cpu;
 854
 855        get_online_cpus();
 856        for_each_online_cpu(cpu)
 857                val += per_cpu(memcg->stat->count[idx], cpu);
 858#ifdef CONFIG_HOTPLUG_CPU
 859        spin_lock(&memcg->pcp_counter_lock);
 860        val += memcg->nocpu_base.count[idx];
 861        spin_unlock(&memcg->pcp_counter_lock);
 862#endif
 863        put_online_cpus();
 864        return val;
 865}
 866
 867static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
 868                                         bool charge)
 869{
 870        int val = (charge) ? 1 : -1;
 871        this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
 872}
 873
 874static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
 875                                            enum mem_cgroup_events_index idx)
 876{
 877        unsigned long val = 0;
 878        int cpu;
 879
 880        for_each_online_cpu(cpu)
 881                val += per_cpu(memcg->stat->events[idx], cpu);
 882#ifdef CONFIG_HOTPLUG_CPU
 883        spin_lock(&memcg->pcp_counter_lock);
 884        val += memcg->nocpu_base.events[idx];
 885        spin_unlock(&memcg->pcp_counter_lock);
 886#endif
 887        return val;
 888}
 889
 890static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
 891                                         bool anon, int nr_pages)
 892{
 893        preempt_disable();
 894
 895        /*
 896         * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
 897         * counted as CACHE even if it's on ANON LRU.
 898         */
 899        if (anon)
 900                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
 901                                nr_pages);
 902        else
 903                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
 904                                nr_pages);
 905
 906        /* pagein of a big page is an event. So, ignore page size */
 907        if (nr_pages > 0)
 908                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
 909        else {
 910                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
 911                nr_pages = -nr_pages; /* for event */
 912        }
 913
 914        __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
 915
 916        preempt_enable();
 917}
 918
 919unsigned long
 920mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 921{
 922        struct mem_cgroup_per_zone *mz;
 923
 924        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
 925        return mz->lru_size[lru];
 926}
 927
 928static unsigned long
 929mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
 930                        unsigned int lru_mask)
 931{
 932        struct mem_cgroup_per_zone *mz;
 933        enum lru_list lru;
 934        unsigned long ret = 0;
 935
 936        mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 937
 938        for_each_lru(lru) {
 939                if (BIT(lru) & lru_mask)
 940                        ret += mz->lru_size[lru];
 941        }
 942        return ret;
 943}
 944
 945static unsigned long
 946mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
 947                        int nid, unsigned int lru_mask)
 948{
 949        u64 total = 0;
 950        int zid;
 951
 952        for (zid = 0; zid < MAX_NR_ZONES; zid++)
 953                total += mem_cgroup_zone_nr_lru_pages(memcg,
 954                                                nid, zid, lru_mask);
 955
 956        return total;
 957}
 958
 959static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
 960                        unsigned int lru_mask)
 961{
 962        int nid;
 963        u64 total = 0;
 964
 965        for_each_node_state(nid, N_MEMORY)
 966                total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
 967        return total;
 968}
 969
 970static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
 971                                       enum mem_cgroup_events_target target)
 972{
 973        unsigned long val, next;
 974
 975        val = __this_cpu_read(memcg->stat->nr_page_events);
 976        next = __this_cpu_read(memcg->stat->targets[target]);
 977        /* from time_after() in jiffies.h */
 978        if ((long)next - (long)val < 0) {
 979                switch (target) {
 980                case MEM_CGROUP_TARGET_THRESH:
 981                        next = val + THRESHOLDS_EVENTS_TARGET;
 982                        break;
 983                case MEM_CGROUP_TARGET_SOFTLIMIT:
 984                        next = val + SOFTLIMIT_EVENTS_TARGET;
 985                        break;
 986                case MEM_CGROUP_TARGET_NUMAINFO:
 987                        next = val + NUMAINFO_EVENTS_TARGET;
 988                        break;
 989                default:
 990                        break;
 991                }
 992                __this_cpu_write(memcg->stat->targets[target], next);
 993                return true;
 994        }
 995        return false;
 996}
 997
 998/*
 999 * Check events in order.
1000 *
1001 */
1002static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1003{
1004        preempt_disable();
1005        /* threshold event is triggered in finer grain than soft limit */
1006        if (unlikely(mem_cgroup_event_ratelimit(memcg,
1007                                                MEM_CGROUP_TARGET_THRESH))) {
1008                bool do_softlimit;
1009                bool do_numainfo __maybe_unused;
1010
1011                do_softlimit = mem_cgroup_event_ratelimit(memcg,
1012                                                MEM_CGROUP_TARGET_SOFTLIMIT);
1013#if MAX_NUMNODES > 1
1014                do_numainfo = mem_cgroup_event_ratelimit(memcg,
1015                                                MEM_CGROUP_TARGET_NUMAINFO);
1016#endif
1017                preempt_enable();
1018
1019                mem_cgroup_threshold(memcg);
1020                if (unlikely(do_softlimit))
1021                        mem_cgroup_update_tree(memcg, page);
1022#if MAX_NUMNODES > 1
1023                if (unlikely(do_numainfo))
1024                        atomic_inc(&memcg->numainfo_events);
1025#endif
1026        } else
1027                preempt_enable();
1028}
1029
1030struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1031{
1032        return mem_cgroup_from_css(
1033                cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1034}
1035
1036struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1037{
1038        /*
1039         * mm_update_next_owner() may clear mm->owner to NULL
1040         * if it races with swapoff, page migration, etc.
1041         * So this can be called with p == NULL.
1042         */
1043        if (unlikely(!p))
1044                return NULL;
1045
1046        return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1047}
1048
1049struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1050{
1051        struct mem_cgroup *memcg = NULL;
1052
1053        if (!mm)
1054                return NULL;
1055        /*
1056         * Because we have no locks, mm->owner's may be being moved to other
1057         * cgroup. We use css_tryget() here even if this looks
1058         * pessimistic (rather than adding locks here).
1059         */
1060        rcu_read_lock();
1061        do {
1062                memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1063                if (unlikely(!memcg))
1064                        break;
1065        } while (!css_tryget(&memcg->css));
1066        rcu_read_unlock();
1067        return memcg;
1068}
1069
1070/**
1071 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1072 * @root: hierarchy root
1073 * @prev: previously returned memcg, NULL on first invocation
1074 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1075 *
1076 * Returns references to children of the hierarchy below @root, or
1077 * @root itself, or %NULL after a full round-trip.
1078 *
1079 * Caller must pass the return value in @prev on subsequent
1080 * invocations for reference counting, or use mem_cgroup_iter_break()
1081 * to cancel a hierarchy walk before the round-trip is complete.
1082 *
1083 * Reclaimers can specify a zone and a priority level in @reclaim to
1084 * divide up the memcgs in the hierarchy among all concurrent
1085 * reclaimers operating on the same zone and priority.
1086 */
1087struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1088                                   struct mem_cgroup *prev,
1089                                   struct mem_cgroup_reclaim_cookie *reclaim)
1090{
1091        struct mem_cgroup *memcg = NULL;
1092        int id = 0;
1093
1094        if (mem_cgroup_disabled())
1095                return NULL;
1096
1097        if (!root)
1098                root = root_mem_cgroup;
1099
1100        if (prev && !reclaim)
1101                id = css_id(&prev->css);
1102
1103        if (prev && prev != root)
1104                css_put(&prev->css);
1105
1106        if (!root->use_hierarchy && root != root_mem_cgroup) {
1107                if (prev)
1108                        return NULL;
1109                return root;
1110        }
1111
1112        while (!memcg) {
1113                struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1114                struct cgroup_subsys_state *css;
1115
1116                if (reclaim) {
1117                        int nid = zone_to_nid(reclaim->zone);
1118                        int zid = zone_idx(reclaim->zone);
1119                        struct mem_cgroup_per_zone *mz;
1120
1121                        mz = mem_cgroup_zoneinfo(root, nid, zid);
1122                        iter = &mz->reclaim_iter[reclaim->priority];
1123                        if (prev && reclaim->generation != iter->generation)
1124                                return NULL;
1125                        id = iter->position;
1126                }
1127
1128                rcu_read_lock();
1129                css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1130                if (css) {
1131                        if (css == &root->css || css_tryget(css))
1132                                memcg = mem_cgroup_from_css(css);
1133                } else
1134                        id = 0;
1135                rcu_read_unlock();
1136
1137                if (reclaim) {
1138                        iter->position = id;
1139                        if (!css)
1140                                iter->generation++;
1141                        else if (!prev && memcg)
1142                                reclaim->generation = iter->generation;
1143                }
1144
1145                if (prev && !css)
1146                        return NULL;
1147        }
1148        return memcg;
1149}
1150
1151/**
1152 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1153 * @root: hierarchy root
1154 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1155 */
1156void mem_cgroup_iter_break(struct mem_cgroup *root,
1157                           struct mem_cgroup *prev)
1158{
1159        if (!root)
1160                root = root_mem_cgroup;
1161        if (prev && prev != root)
1162                css_put(&prev->css);
1163}
1164
1165/*
1166 * Iteration constructs for visiting all cgroups (under a tree).  If
1167 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1168 * be used for reference counting.
1169 */
1170#define for_each_mem_cgroup_tree(iter, root)            \
1171        for (iter = mem_cgroup_iter(root, NULL, NULL);  \
1172             iter != NULL;                              \
1173             iter = mem_cgroup_iter(root, iter, NULL))
1174
1175#define for_each_mem_cgroup(iter)                       \
1176        for (iter = mem_cgroup_iter(NULL, NULL, NULL);  \
1177             iter != NULL;                              \
1178             iter = mem_cgroup_iter(NULL, iter, NULL))
1179
1180void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1181{
1182        struct mem_cgroup *memcg;
1183
1184        rcu_read_lock();
1185        memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1186        if (unlikely(!memcg))
1187                goto out;
1188
1189        switch (idx) {
1190        case PGFAULT:
1191                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1192                break;
1193        case PGMAJFAULT:
1194                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1195                break;
1196        default:
1197                BUG();
1198        }
1199out:
1200        rcu_read_unlock();
1201}
1202EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1203
1204/**
1205 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1206 * @zone: zone of the wanted lruvec
1207 * @memcg: memcg of the wanted lruvec
1208 *
1209 * Returns the lru list vector holding pages for the given @zone and
1210 * @mem.  This can be the global zone lruvec, if the memory controller
1211 * is disabled.
1212 */
1213struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1214                                      struct mem_cgroup *memcg)
1215{
1216        struct mem_cgroup_per_zone *mz;
1217        struct lruvec *lruvec;
1218
1219        if (mem_cgroup_disabled()) {
1220                lruvec = &zone->lruvec;
1221                goto out;
1222        }
1223
1224        mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1225        lruvec = &mz->lruvec;
1226out:
1227        /*
1228         * Since a node can be onlined after the mem_cgroup was created,
1229         * we have to be prepared to initialize lruvec->zone here;
1230         * and if offlined then reonlined, we need to reinitialize it.
1231         */
1232        if (unlikely(lruvec->zone != zone))
1233                lruvec->zone = zone;
1234        return lruvec;
1235}
1236
1237/*
1238 * Following LRU functions are allowed to be used without PCG_LOCK.
1239 * Operations are called by routine of global LRU independently from memcg.
1240 * What we have to take care of here is validness of pc->mem_cgroup.
1241 *
1242 * Changes to pc->mem_cgroup happens when
1243 * 1. charge
1244 * 2. moving account
1245 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1246 * It is added to LRU before charge.
1247 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1248 * When moving account, the page is not on LRU. It's isolated.
1249 */
1250
1251/**
1252 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1253 * @page: the page
1254 * @zone: zone of the page
1255 */
1256struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1257{
1258        struct mem_cgroup_per_zone *mz;
1259        struct mem_cgroup *memcg;
1260        struct page_cgroup *pc;
1261        struct lruvec *lruvec;
1262
1263        if (mem_cgroup_disabled()) {
1264                lruvec = &zone->lruvec;
1265                goto out;
1266        }
1267
1268        pc = lookup_page_cgroup(page);
1269        memcg = pc->mem_cgroup;
1270
1271        /*
1272         * Surreptitiously switch any uncharged offlist page to root:
1273         * an uncharged page off lru does nothing to secure
1274         * its former mem_cgroup from sudden removal.
1275         *
1276         * Our caller holds lru_lock, and PageCgroupUsed is updated
1277         * under page_cgroup lock: between them, they make all uses
1278         * of pc->mem_cgroup safe.
1279         */
1280        if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1281                pc->mem_cgroup = memcg = root_mem_cgroup;
1282
1283        mz = page_cgroup_zoneinfo(memcg, page);
1284        lruvec = &mz->lruvec;
1285out:
1286        /*
1287         * Since a node can be onlined after the mem_cgroup was created,
1288         * we have to be prepared to initialize lruvec->zone here;
1289         * and if offlined then reonlined, we need to reinitialize it.
1290         */
1291        if (unlikely(lruvec->zone != zone))
1292                lruvec->zone = zone;
1293        return lruvec;
1294}
1295
1296/**
1297 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1298 * @lruvec: mem_cgroup per zone lru vector
1299 * @lru: index of lru list the page is sitting on
1300 * @nr_pages: positive when adding or negative when removing
1301 *
1302 * This function must be called when a page is added to or removed from an
1303 * lru list.
1304 */
1305void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1306                                int nr_pages)
1307{
1308        struct mem_cgroup_per_zone *mz;
1309        unsigned long *lru_size;
1310
1311        if (mem_cgroup_disabled())
1312                return;
1313
1314        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1315        lru_size = mz->lru_size + lru;
1316        *lru_size += nr_pages;
1317        VM_BUG_ON((long)(*lru_size) < 0);
1318}
1319
1320/*
1321 * Checks whether given mem is same or in the root_mem_cgroup's
1322 * hierarchy subtree
1323 */
1324bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1325                                  struct mem_cgroup *memcg)
1326{
1327        if (root_memcg == memcg)
1328                return true;
1329        if (!root_memcg->use_hierarchy || !memcg)
1330                return false;
1331        return css_is_ancestor(&memcg->css, &root_memcg->css);
1332}
1333
1334static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1335                                       struct mem_cgroup *memcg)
1336{
1337        bool ret;
1338
1339        rcu_read_lock();
1340        ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1341        rcu_read_unlock();
1342        return ret;
1343}
1344
1345int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1346{
1347        int ret;
1348        struct mem_cgroup *curr = NULL;
1349        struct task_struct *p;
1350
1351        p = find_lock_task_mm(task);
1352        if (p) {
1353                curr = try_get_mem_cgroup_from_mm(p->mm);
1354                task_unlock(p);
1355        } else {
1356                /*
1357                 * All threads may have already detached their mm's, but the oom
1358                 * killer still needs to detect if they have already been oom
1359                 * killed to prevent needlessly killing additional tasks.
1360                 */
1361                task_lock(task);
1362                curr = mem_cgroup_from_task(task);
1363                if (curr)
1364                        css_get(&curr->css);
1365                task_unlock(task);
1366        }
1367        if (!curr)
1368                return 0;
1369        /*
1370         * We should check use_hierarchy of "memcg" not "curr". Because checking
1371         * use_hierarchy of "curr" here make this function true if hierarchy is
1372         * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1373         * hierarchy(even if use_hierarchy is disabled in "memcg").
1374         */
1375        ret = mem_cgroup_same_or_subtree(memcg, curr);
1376        css_put(&curr->css);
1377        return ret;
1378}
1379
1380int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1381{
1382        unsigned long inactive_ratio;
1383        unsigned long inactive;
1384        unsigned long active;
1385        unsigned long gb;
1386
1387        inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1388        active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1389
1390        gb = (inactive + active) >> (30 - PAGE_SHIFT);
1391        if (gb)
1392                inactive_ratio = int_sqrt(10 * gb);
1393        else
1394                inactive_ratio = 1;
1395
1396        return inactive * inactive_ratio < active;
1397}
1398
1399#define mem_cgroup_from_res_counter(counter, member)    \
1400        container_of(counter, struct mem_cgroup, member)
1401
1402/**
1403 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1404 * @memcg: the memory cgroup
1405 *
1406 * Returns the maximum amount of memory @mem can be charged with, in
1407 * pages.
1408 */
1409static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1410{
1411        unsigned long long margin;
1412
1413        margin = res_counter_margin(&memcg->res);
1414        if (do_swap_account)
1415                margin = min(margin, res_counter_margin(&memcg->memsw));
1416        return margin >> PAGE_SHIFT;
1417}
1418
1419int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1420{
1421        struct cgroup *cgrp = memcg->css.cgroup;
1422
1423        /* root ? */
1424        if (cgrp->parent == NULL)
1425                return vm_swappiness;
1426
1427        return memcg->swappiness;
1428}
1429
1430/*
1431 * memcg->moving_account is used for checking possibility that some thread is
1432 * calling move_account(). When a thread on CPU-A starts moving pages under
1433 * a memcg, other threads should check memcg->moving_account under
1434 * rcu_read_lock(), like this:
1435 *
1436 *         CPU-A                                    CPU-B
1437 *                                              rcu_read_lock()
1438 *         memcg->moving_account+1              if (memcg->mocing_account)
1439 *                                                   take heavy locks.
1440 *         synchronize_rcu()                    update something.
1441 *                                              rcu_read_unlock()
1442 *         start move here.
1443 */
1444
1445/* for quick checking without looking up memcg */
1446atomic_t memcg_moving __read_mostly;
1447
1448static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1449{
1450        atomic_inc(&memcg_moving);
1451        atomic_inc(&memcg->moving_account);
1452        synchronize_rcu();
1453}
1454
1455static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1456{
1457        /*
1458         * Now, mem_cgroup_clear_mc() may call this function with NULL.
1459         * We check NULL in callee rather than caller.
1460         */
1461        if (memcg) {
1462                atomic_dec(&memcg_moving);
1463                atomic_dec(&memcg->moving_account);
1464        }
1465}
1466
1467/*
1468 * 2 routines for checking "mem" is under move_account() or not.
1469 *
1470 * mem_cgroup_stolen() -  checking whether a cgroup is mc.from or not. This
1471 *                        is used for avoiding races in accounting.  If true,
1472 *                        pc->mem_cgroup may be overwritten.
1473 *
1474 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1475 *                        under hierarchy of moving cgroups. This is for
1476 *                        waiting at hith-memory prressure caused by "move".
1477 */
1478
1479static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1480{
1481        VM_BUG_ON(!rcu_read_lock_held());
1482        return atomic_read(&memcg->moving_account) > 0;
1483}
1484
1485static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1486{
1487        struct mem_cgroup *from;
1488        struct mem_cgroup *to;
1489        bool ret = false;
1490        /*
1491         * Unlike task_move routines, we access mc.to, mc.from not under
1492         * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1493         */
1494        spin_lock(&mc.lock);
1495        from = mc.from;
1496        to = mc.to;
1497        if (!from)
1498                goto unlock;
1499
1500        ret = mem_cgroup_same_or_subtree(memcg, from)
1501                || mem_cgroup_same_or_subtree(memcg, to);
1502unlock:
1503        spin_unlock(&mc.lock);
1504        return ret;
1505}
1506
1507static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1508{
1509        if (mc.moving_task && current != mc.moving_task) {
1510                if (mem_cgroup_under_move(memcg)) {
1511                        DEFINE_WAIT(wait);
1512                        prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1513                        /* moving charge context might have finished. */
1514                        if (mc.moving_task)
1515                                schedule();
1516                        finish_wait(&mc.waitq, &wait);
1517                        return true;
1518                }
1519        }
1520        return false;
1521}
1522
1523/*
1524 * Take this lock when
1525 * - a code tries to modify page's memcg while it's USED.
1526 * - a code tries to modify page state accounting in a memcg.
1527 * see mem_cgroup_stolen(), too.
1528 */
1529static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1530                                  unsigned long *flags)
1531{
1532        spin_lock_irqsave(&memcg->move_lock, *flags);
1533}
1534
1535static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1536                                unsigned long *flags)
1537{
1538        spin_unlock_irqrestore(&memcg->move_lock, *flags);
1539}
1540
1541#define K(x) ((x) << (PAGE_SHIFT-10))
1542/**
1543 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1544 * @memcg: The memory cgroup that went over limit
1545 * @p: Task that is going to be killed
1546 *
1547 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1548 * enabled
1549 */
1550void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1551{
1552        struct cgroup *task_cgrp;
1553        struct cgroup *mem_cgrp;
1554        /*
1555         * Need a buffer in BSS, can't rely on allocations. The code relies
1556         * on the assumption that OOM is serialized for memory controller.
1557         * If this assumption is broken, revisit this code.
1558         */
1559        static char memcg_name[PATH_MAX];
1560        int ret;
1561        struct mem_cgroup *iter;
1562        unsigned int i;
1563
1564        if (!p)
1565                return;
1566
1567        rcu_read_lock();
1568
1569        mem_cgrp = memcg->css.cgroup;
1570        task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1571
1572        ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1573        if (ret < 0) {
1574                /*
1575                 * Unfortunately, we are unable to convert to a useful name
1576                 * But we'll still print out the usage information
1577                 */
1578                rcu_read_unlock();
1579                goto done;
1580        }
1581        rcu_read_unlock();
1582
1583        pr_info("Task in %s killed", memcg_name);
1584
1585        rcu_read_lock();
1586        ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1587        if (ret < 0) {
1588                rcu_read_unlock();
1589                goto done;
1590        }
1591        rcu_read_unlock();
1592
1593        /*
1594         * Continues from above, so we don't need an KERN_ level
1595         */
1596        pr_cont(" as a result of limit of %s\n", memcg_name);
1597done:
1598
1599        pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1600                res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1601                res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1602                res_counter_read_u64(&memcg->res, RES_FAILCNT));
1603        pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1604                res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1605                res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1606                res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1607        pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1608                res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1609                res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1610                res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1611
1612        for_each_mem_cgroup_tree(iter, memcg) {
1613                pr_info("Memory cgroup stats");
1614
1615                rcu_read_lock();
1616                ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1617                if (!ret)
1618                        pr_cont(" for %s", memcg_name);
1619                rcu_read_unlock();
1620                pr_cont(":");
1621
1622                for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1623                        if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1624                                continue;
1625                        pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1626                                K(mem_cgroup_read_stat(iter, i)));
1627                }
1628
1629                for (i = 0; i < NR_LRU_LISTS; i++)
1630                        pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1631                                K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1632
1633                pr_cont("\n");
1634        }
1635}
1636
1637/*
1638 * This function returns the number of memcg under hierarchy tree. Returns
1639 * 1(self count) if no children.
1640 */
1641static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1642{
1643        int num = 0;
1644        struct mem_cgroup *iter;
1645
1646        for_each_mem_cgroup_tree(iter, memcg)
1647                num++;
1648        return num;
1649}
1650
1651/*
1652 * Return the memory (and swap, if configured) limit for a memcg.
1653 */
1654static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1655{
1656        u64 limit;
1657
1658        limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1659
1660        /*
1661         * Do not consider swap space if we cannot swap due to swappiness
1662         */
1663        if (mem_cgroup_swappiness(memcg)) {
1664                u64 memsw;
1665
1666                limit += total_swap_pages << PAGE_SHIFT;
1667                memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1668
1669                /*
1670                 * If memsw is finite and limits the amount of swap space
1671                 * available to this memcg, return that limit.
1672                 */
1673                limit = min(limit, memsw);
1674        }
1675
1676        return limit;
1677}
1678
1679static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1680                                     int order)
1681{
1682        struct mem_cgroup *iter;
1683        unsigned long chosen_points = 0;
1684        unsigned long totalpages;
1685        unsigned int points = 0;
1686        struct task_struct *chosen = NULL;
1687
1688        /*
1689         * If current has a pending SIGKILL, then automatically select it.  The
1690         * goal is to allow it to allocate so that it may quickly exit and free
1691         * its memory.
1692         */
1693        if (fatal_signal_pending(current)) {
1694                set_thread_flag(TIF_MEMDIE);
1695                return;
1696        }
1697
1698        check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1699        totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1700        for_each_mem_cgroup_tree(iter, memcg) {
1701                struct cgroup *cgroup = iter->css.cgroup;
1702                struct cgroup_iter it;
1703                struct task_struct *task;
1704
1705                cgroup_iter_start(cgroup, &it);
1706                while ((task = cgroup_iter_next(cgroup, &it))) {
1707                        switch (oom_scan_process_thread(task, totalpages, NULL,
1708                                                        false)) {
1709                        case OOM_SCAN_SELECT:
1710                                if (chosen)
1711                                        put_task_struct(chosen);
1712                                chosen = task;
1713                                chosen_points = ULONG_MAX;
1714                                get_task_struct(chosen);
1715                                /* fall through */
1716                        case OOM_SCAN_CONTINUE:
1717                                continue;
1718                        case OOM_SCAN_ABORT:
1719                                cgroup_iter_end(cgroup, &it);
1720                                mem_cgroup_iter_break(memcg, iter);
1721                                if (chosen)
1722                                        put_task_struct(chosen);
1723                                return;
1724                        case OOM_SCAN_OK:
1725                                break;
1726                        };
1727                        points = oom_badness(task, memcg, NULL, totalpages);
1728                        if (points > chosen_points) {
1729                                if (chosen)
1730                                        put_task_struct(chosen);
1731                                chosen = task;
1732                                chosen_points = points;
1733                                get_task_struct(chosen);
1734                        }
1735                }
1736                cgroup_iter_end(cgroup, &it);
1737        }
1738
1739        if (!chosen)
1740                return;
1741        points = chosen_points * 1000 / totalpages;
1742        oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1743                         NULL, "Memory cgroup out of memory");
1744}
1745
1746static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1747                                        gfp_t gfp_mask,
1748                                        unsigned long flags)
1749{
1750        unsigned long total = 0;
1751        bool noswap = false;
1752        int loop;
1753
1754        if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1755                noswap = true;
1756        if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1757                noswap = true;
1758
1759        for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1760                if (loop)
1761                        drain_all_stock_async(memcg);
1762                total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1763                /*
1764                 * Allow limit shrinkers, which are triggered directly
1765                 * by userspace, to catch signals and stop reclaim
1766                 * after minimal progress, regardless of the margin.
1767                 */
1768                if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1769                        break;
1770                if (mem_cgroup_margin(memcg))
1771                        break;
1772                /*
1773                 * If nothing was reclaimed after two attempts, there
1774                 * may be no reclaimable pages in this hierarchy.
1775                 */
1776                if (loop && !total)
1777                        break;
1778        }
1779        return total;
1780}
1781
1782/**
1783 * test_mem_cgroup_node_reclaimable
1784 * @memcg: the target memcg
1785 * @nid: the node ID to be checked.
1786 * @noswap : specify true here if the user wants flle only information.
1787 *
1788 * This function returns whether the specified memcg contains any
1789 * reclaimable pages on a node. Returns true if there are any reclaimable
1790 * pages in the node.
1791 */
1792static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1793                int nid, bool noswap)
1794{
1795        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1796                return true;
1797        if (noswap || !total_swap_pages)
1798                return false;
1799        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1800                return true;
1801        return false;
1802
1803}
1804#if MAX_NUMNODES > 1
1805
1806/*
1807 * Always updating the nodemask is not very good - even if we have an empty
1808 * list or the wrong list here, we can start from some node and traverse all
1809 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1810 *
1811 */
1812static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1813{
1814        int nid;
1815        /*
1816         * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1817         * pagein/pageout changes since the last update.
1818         */
1819        if (!atomic_read(&memcg->numainfo_events))
1820                return;
1821        if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1822                return;
1823
1824        /* make a nodemask where this memcg uses memory from */
1825        memcg->scan_nodes = node_states[N_MEMORY];
1826
1827        for_each_node_mask(nid, node_states[N_MEMORY]) {
1828
1829                if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1830                        node_clear(nid, memcg->scan_nodes);
1831        }
1832
1833        atomic_set(&memcg->numainfo_events, 0);
1834        atomic_set(&memcg->numainfo_updating, 0);
1835}
1836
1837/*
1838 * Selecting a node where we start reclaim from. Because what we need is just
1839 * reducing usage counter, start from anywhere is O,K. Considering
1840 * memory reclaim from current node, there are pros. and cons.
1841 *
1842 * Freeing memory from current node means freeing memory from a node which
1843 * we'll use or we've used. So, it may make LRU bad. And if several threads
1844 * hit limits, it will see a contention on a node. But freeing from remote
1845 * node means more costs for memory reclaim because of memory latency.
1846 *
1847 * Now, we use round-robin. Better algorithm is welcomed.
1848 */
1849int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1850{
1851        int node;
1852
1853        mem_cgroup_may_update_nodemask(memcg);
1854        node = memcg->last_scanned_node;
1855
1856        node = next_node(node, memcg->scan_nodes);
1857        if (node == MAX_NUMNODES)
1858                node = first_node(memcg->scan_nodes);
1859        /*
1860         * We call this when we hit limit, not when pages are added to LRU.
1861         * No LRU may hold pages because all pages are UNEVICTABLE or
1862         * memcg is too small and all pages are not on LRU. In that case,
1863         * we use curret node.
1864         */
1865        if (unlikely(node == MAX_NUMNODES))
1866                node = numa_node_id();
1867
1868        memcg->last_scanned_node = node;
1869        return node;
1870}
1871
1872/*
1873 * Check all nodes whether it contains reclaimable pages or not.
1874 * For quick scan, we make use of scan_nodes. This will allow us to skip
1875 * unused nodes. But scan_nodes is lazily updated and may not cotain
1876 * enough new information. We need to do double check.
1877 */
1878static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1879{
1880        int nid;
1881
1882        /*
1883         * quick check...making use of scan_node.
1884         * We can skip unused nodes.
1885         */
1886        if (!nodes_empty(memcg->scan_nodes)) {
1887                for (nid = first_node(memcg->scan_nodes);
1888                     nid < MAX_NUMNODES;
1889                     nid = next_node(nid, memcg->scan_nodes)) {
1890
1891                        if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1892                                return true;
1893                }
1894        }
1895        /*
1896         * Check rest of nodes.
1897         */
1898        for_each_node_state(nid, N_MEMORY) {
1899                if (node_isset(nid, memcg->scan_nodes))
1900                        continue;
1901                if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1902                        return true;
1903        }
1904        return false;
1905}
1906
1907#else
1908int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1909{
1910        return 0;
1911}
1912
1913static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1914{
1915        return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1916}
1917#endif
1918
1919static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1920                                   struct zone *zone,
1921                                   gfp_t gfp_mask,
1922                                   unsigned long *total_scanned)
1923{
1924        struct mem_cgroup *victim = NULL;
1925        int total = 0;
1926        int loop = 0;
1927        unsigned long excess;
1928        unsigned long nr_scanned;
1929        struct mem_cgroup_reclaim_cookie reclaim = {
1930                .zone = zone,
1931                .priority = 0,
1932        };
1933
1934        excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1935
1936        while (1) {
1937                victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1938                if (!victim) {
1939                        loop++;
1940                        if (loop >= 2) {
1941                                /*
1942                                 * If we have not been able to reclaim
1943                                 * anything, it might because there are
1944                                 * no reclaimable pages under this hierarchy
1945                                 */
1946                                if (!total)
1947                                        break;
1948                                /*
1949                                 * We want to do more targeted reclaim.
1950                                 * excess >> 2 is not to excessive so as to
1951                                 * reclaim too much, nor too less that we keep
1952                                 * coming back to reclaim from this cgroup
1953                                 */
1954                                if (total >= (excess >> 2) ||
1955                                        (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1956                                        break;
1957                        }
1958                        continue;
1959                }
1960                if (!mem_cgroup_reclaimable(victim, false))
1961                        continue;
1962                total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1963                                                     zone, &nr_scanned);
1964                *total_scanned += nr_scanned;
1965                if (!res_counter_soft_limit_excess(&root_memcg->res))
1966                        break;
1967        }
1968        mem_cgroup_iter_break(root_memcg, victim);
1969        return total;
1970}
1971
1972/*
1973 * Check OOM-Killer is already running under our hierarchy.
1974 * If someone is running, return false.
1975 * Has to be called with memcg_oom_lock
1976 */
1977static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1978{
1979        struct mem_cgroup *iter, *failed = NULL;
1980
1981        for_each_mem_cgroup_tree(iter, memcg) {
1982                if (iter->oom_lock) {
1983                        /*
1984                         * this subtree of our hierarchy is already locked
1985                         * so we cannot give a lock.
1986                         */
1987                        failed = iter;
1988                        mem_cgroup_iter_break(memcg, iter);
1989                        break;
1990                } else
1991                        iter->oom_lock = true;
1992        }
1993
1994        if (!failed)
1995                return true;
1996
1997        /*
1998         * OK, we failed to lock the whole subtree so we have to clean up
1999         * what we set up to the failing subtree
2000         */
2001        for_each_mem_cgroup_tree(iter, memcg) {
2002                if (iter == failed) {
2003                        mem_cgroup_iter_break(memcg, iter);
2004                        break;
2005                }
2006                iter->oom_lock = false;
2007        }
2008        return false;
2009}
2010
2011/*
2012 * Has to be called with memcg_oom_lock
2013 */
2014static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2015{
2016        struct mem_cgroup *iter;
2017
2018        for_each_mem_cgroup_tree(iter, memcg)
2019                iter->oom_lock = false;
2020        return 0;
2021}
2022
2023static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2024{
2025        struct mem_cgroup *iter;
2026
2027        for_each_mem_cgroup_tree(iter, memcg)
2028                atomic_inc(&iter->under_oom);
2029}
2030
2031static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2032{
2033        struct mem_cgroup *iter;
2034
2035        /*
2036         * When a new child is created while the hierarchy is under oom,
2037         * mem_cgroup_oom_lock() may not be called. We have to use
2038         * atomic_add_unless() here.
2039         */
2040        for_each_mem_cgroup_tree(iter, memcg)
2041                atomic_add_unless(&iter->under_oom, -1, 0);
2042}
2043
2044static DEFINE_SPINLOCK(memcg_oom_lock);
2045static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2046
2047struct oom_wait_info {
2048        struct mem_cgroup *memcg;
2049        wait_queue_t    wait;
2050};
2051
2052static int memcg_oom_wake_function(wait_queue_t *wait,
2053        unsigned mode, int sync, void *arg)
2054{
2055        struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2056        struct mem_cgroup *oom_wait_memcg;
2057        struct oom_wait_info *oom_wait_info;
2058
2059        oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2060        oom_wait_memcg = oom_wait_info->memcg;
2061
2062        /*
2063         * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2064         * Then we can use css_is_ancestor without taking care of RCU.
2065         */
2066        if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2067                && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2068                return 0;
2069        return autoremove_wake_function(wait, mode, sync, arg);
2070}
2071
2072static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2073{
2074        /* for filtering, pass "memcg" as argument. */
2075        __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2076}
2077
2078static void memcg_oom_recover(struct mem_cgroup *memcg)
2079{
2080        if (memcg && atomic_read(&memcg->under_oom))
2081                memcg_wakeup_oom(memcg);
2082}
2083
2084/*
2085 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2086 */
2087static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2088                                  int order)
2089{
2090        struct oom_wait_info owait;
2091        bool locked, need_to_kill;
2092
2093        owait.memcg = memcg;
2094        owait.wait.flags = 0;
2095        owait.wait.func = memcg_oom_wake_function;
2096        owait.wait.private = current;
2097        INIT_LIST_HEAD(&owait.wait.task_list);
2098        need_to_kill = true;
2099        mem_cgroup_mark_under_oom(memcg);
2100
2101        /* At first, try to OOM lock hierarchy under memcg.*/
2102        spin_lock(&memcg_oom_lock);
2103        locked = mem_cgroup_oom_lock(memcg);
2104        /*
2105         * Even if signal_pending(), we can't quit charge() loop without
2106         * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2107         * under OOM is always welcomed, use TASK_KILLABLE here.
2108         */
2109        prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2110        if (!locked || memcg->oom_kill_disable)
2111                need_to_kill = false;
2112        if (locked)
2113                mem_cgroup_oom_notify(memcg);
2114        spin_unlock(&memcg_oom_lock);
2115
2116        if (need_to_kill) {
2117                finish_wait(&memcg_oom_waitq, &owait.wait);
2118                mem_cgroup_out_of_memory(memcg, mask, order);
2119        } else {
2120                schedule();
2121                finish_wait(&memcg_oom_waitq, &owait.wait);
2122        }
2123        spin_lock(&memcg_oom_lock);
2124        if (locked)
2125                mem_cgroup_oom_unlock(memcg);
2126        memcg_wakeup_oom(memcg);
2127        spin_unlock(&memcg_oom_lock);
2128
2129        mem_cgroup_unmark_under_oom(memcg);
2130
2131        if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2132                return false;
2133        /* Give chance to dying process */
2134        schedule_timeout_uninterruptible(1);
2135        return true;
2136}
2137
2138/*
2139 * Currently used to update mapped file statistics, but the routine can be
2140 * generalized to update other statistics as well.
2141 *
2142 * Notes: Race condition
2143 *
2144 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2145 * it tends to be costly. But considering some conditions, we doesn't need
2146 * to do so _always_.
2147 *
2148 * Considering "charge", lock_page_cgroup() is not required because all
2149 * file-stat operations happen after a page is attached to radix-tree. There
2150 * are no race with "charge".
2151 *
2152 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2153 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2154 * if there are race with "uncharge". Statistics itself is properly handled
2155 * by flags.
2156 *
2157 * Considering "move", this is an only case we see a race. To make the race
2158 * small, we check mm->moving_account and detect there are possibility of race
2159 * If there is, we take a lock.
2160 */
2161
2162void __mem_cgroup_begin_update_page_stat(struct page *page,
2163                                bool *locked, unsigned long *flags)
2164{
2165        struct mem_cgroup *memcg;
2166        struct page_cgroup *pc;
2167
2168        pc = lookup_page_cgroup(page);
2169again:
2170        memcg = pc->mem_cgroup;
2171        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2172                return;
2173        /*
2174         * If this memory cgroup is not under account moving, we don't
2175         * need to take move_lock_mem_cgroup(). Because we already hold
2176         * rcu_read_lock(), any calls to move_account will be delayed until
2177         * rcu_read_unlock() if mem_cgroup_stolen() == true.
2178         */
2179        if (!mem_cgroup_stolen(memcg))
2180                return;
2181
2182        move_lock_mem_cgroup(memcg, flags);
2183        if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2184                move_unlock_mem_cgroup(memcg, flags);
2185                goto again;
2186        }
2187        *locked = true;
2188}
2189
2190void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2191{
2192        struct page_cgroup *pc = lookup_page_cgroup(page);
2193
2194        /*
2195         * It's guaranteed that pc->mem_cgroup never changes while
2196         * lock is held because a routine modifies pc->mem_cgroup
2197         * should take move_lock_mem_cgroup().
2198         */
2199        move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2200}
2201
2202void mem_cgroup_update_page_stat(struct page *page,
2203                                 enum mem_cgroup_page_stat_item idx, int val)
2204{
2205        struct mem_cgroup *memcg;
2206        struct page_cgroup *pc = lookup_page_cgroup(page);
2207        unsigned long uninitialized_var(flags);
2208
2209        if (mem_cgroup_disabled())
2210                return;
2211
2212        memcg = pc->mem_cgroup;
2213        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2214                return;
2215
2216        switch (idx) {
2217        case MEMCG_NR_FILE_MAPPED:
2218                idx = MEM_CGROUP_STAT_FILE_MAPPED;
2219                break;
2220        default:
2221                BUG();
2222        }
2223
2224        this_cpu_add(memcg->stat->count[idx], val);
2225}
2226
2227/*
2228 * size of first charge trial. "32" comes from vmscan.c's magic value.
2229 * TODO: maybe necessary to use big numbers in big irons.
2230 */
2231#define CHARGE_BATCH    32U
2232struct memcg_stock_pcp {
2233        struct mem_cgroup *cached; /* this never be root cgroup */
2234        unsigned int nr_pages;
2235        struct work_struct work;
2236        unsigned long flags;
2237#define FLUSHING_CACHED_CHARGE  0
2238};
2239static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2240static DEFINE_MUTEX(percpu_charge_mutex);
2241
2242/**
2243 * consume_stock: Try to consume stocked charge on this cpu.
2244 * @memcg: memcg to consume from.
2245 * @nr_pages: how many pages to charge.
2246 *
2247 * The charges will only happen if @memcg matches the current cpu's memcg
2248 * stock, and at least @nr_pages are available in that stock.  Failure to
2249 * service an allocation will refill the stock.
2250 *
2251 * returns true if successful, false otherwise.
2252 */
2253static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2254{
2255        struct memcg_stock_pcp *stock;
2256        bool ret = true;
2257
2258        if (nr_pages > CHARGE_BATCH)
2259                return false;
2260
2261        stock = &get_cpu_var(memcg_stock);
2262        if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2263                stock->nr_pages -= nr_pages;
2264        else /* need to call res_counter_charge */
2265                ret = false;
2266        put_cpu_var(memcg_stock);
2267        return ret;
2268}
2269
2270/*
2271 * Returns stocks cached in percpu to res_counter and reset cached information.
2272 */
2273static void drain_stock(struct memcg_stock_pcp *stock)
2274{
2275        struct mem_cgroup *old = stock->cached;
2276
2277        if (stock->nr_pages) {
2278                unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2279
2280                res_counter_uncharge(&old->res, bytes);
2281                if (do_swap_account)
2282                        res_counter_uncharge(&old->memsw, bytes);
2283                stock->nr_pages = 0;
2284        }
2285        stock->cached = NULL;
2286}
2287
2288/*
2289 * This must be called under preempt disabled or must be called by
2290 * a thread which is pinned to local cpu.
2291 */
2292static void drain_local_stock(struct work_struct *dummy)
2293{
2294        struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2295        drain_stock(stock);
2296        clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2297}
2298
2299static void __init memcg_stock_init(void)
2300{
2301        int cpu;
2302
2303        for_each_possible_cpu(cpu) {
2304                struct memcg_stock_pcp *stock =
2305                                        &per_cpu(memcg_stock, cpu);
2306                INIT_WORK(&stock->work, drain_local_stock);
2307        }
2308}
2309
2310/*
2311 * Cache charges(val) which is from res_counter, to local per_cpu area.
2312 * This will be consumed by consume_stock() function, later.
2313 */
2314static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2315{
2316        struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2317
2318        if (stock->cached != memcg) { /* reset if necessary */
2319                drain_stock(stock);
2320                stock->cached = memcg;
2321        }
2322        stock->nr_pages += nr_pages;
2323        put_cpu_var(memcg_stock);
2324}
2325
2326/*
2327 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2328 * of the hierarchy under it. sync flag says whether we should block
2329 * until the work is done.
2330 */
2331static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2332{
2333        int cpu, curcpu;
2334
2335        /* Notify other cpus that system-wide "drain" is running */
2336        get_online_cpus();
2337        curcpu = get_cpu();
2338        for_each_online_cpu(cpu) {
2339                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2340                struct mem_cgroup *memcg;
2341
2342                memcg = stock->cached;
2343                if (!memcg || !stock->nr_pages)
2344                        continue;
2345                if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2346                        continue;
2347                if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2348                        if (cpu == curcpu)
2349                                drain_local_stock(&stock->work);
2350                        else
2351                                schedule_work_on(cpu, &stock->work);
2352                }
2353        }
2354        put_cpu();
2355
2356        if (!sync)
2357                goto out;
2358
2359        for_each_online_cpu(cpu) {
2360                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2361                if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2362                        flush_work(&stock->work);
2363        }
2364out:
2365        put_online_cpus();
2366}
2367
2368/*
2369 * Tries to drain stocked charges in other cpus. This function is asynchronous
2370 * and just put a work per cpu for draining localy on each cpu. Caller can
2371 * expects some charges will be back to res_counter later but cannot wait for
2372 * it.
2373 */
2374static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2375{
2376        /*
2377         * If someone calls draining, avoid adding more kworker runs.
2378         */
2379        if (!mutex_trylock(&percpu_charge_mutex))
2380                return;
2381        drain_all_stock(root_memcg, false);
2382        mutex_unlock(&percpu_charge_mutex);
2383}
2384
2385/* This is a synchronous drain interface. */
2386static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2387{
2388        /* called when force_empty is called */
2389        mutex_lock(&percpu_charge_mutex);
2390        drain_all_stock(root_memcg, true);
2391        mutex_unlock(&percpu_charge_mutex);
2392}
2393
2394/*
2395 * This function drains percpu counter value from DEAD cpu and
2396 * move it to local cpu. Note that this function can be preempted.
2397 */
2398static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2399{
2400        int i;
2401
2402        spin_lock(&memcg->pcp_counter_lock);
2403        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2404                long x = per_cpu(memcg->stat->count[i], cpu);
2405
2406                per_cpu(memcg->stat->count[i], cpu) = 0;
2407                memcg->nocpu_base.count[i] += x;
2408        }
2409        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2410                unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2411
2412                per_cpu(memcg->stat->events[i], cpu) = 0;
2413                memcg->nocpu_base.events[i] += x;
2414        }
2415        spin_unlock(&memcg->pcp_counter_lock);
2416}
2417
2418static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2419                                        unsigned long action,
2420                                        void *hcpu)
2421{
2422        int cpu = (unsigned long)hcpu;
2423        struct memcg_stock_pcp *stock;
2424        struct mem_cgroup *iter;
2425
2426        if (action == CPU_ONLINE)
2427                return NOTIFY_OK;
2428
2429        if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2430                return NOTIFY_OK;
2431
2432        for_each_mem_cgroup(iter)
2433                mem_cgroup_drain_pcp_counter(iter, cpu);
2434
2435        stock = &per_cpu(memcg_stock, cpu);
2436        drain_stock(stock);
2437        return NOTIFY_OK;
2438}
2439
2440
2441/* See __mem_cgroup_try_charge() for details */
2442enum {
2443        CHARGE_OK,              /* success */
2444        CHARGE_RETRY,           /* need to retry but retry is not bad */
2445        CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2446        CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2447        CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2448};
2449
2450static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2451                                unsigned int nr_pages, unsigned int min_pages,
2452                                bool oom_check)
2453{
2454        unsigned long csize = nr_pages * PAGE_SIZE;
2455        struct mem_cgroup *mem_over_limit;
2456        struct res_counter *fail_res;
2457        unsigned long flags = 0;
2458        int ret;
2459
2460        ret = res_counter_charge(&memcg->res, csize, &fail_res);
2461
2462        if (likely(!ret)) {
2463                if (!do_swap_account)
2464                        return CHARGE_OK;
2465                ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2466                if (likely(!ret))
2467                        return CHARGE_OK;
2468
2469                res_counter_uncharge(&memcg->res, csize);
2470                mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2471                flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2472        } else
2473                mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2474        /*
2475         * Never reclaim on behalf of optional batching, retry with a
2476         * single page instead.
2477         */
2478        if (nr_pages > min_pages)
2479                return CHARGE_RETRY;
2480
2481        if (!(gfp_mask & __GFP_WAIT))
2482                return CHARGE_WOULDBLOCK;
2483
2484        if (gfp_mask & __GFP_NORETRY)
2485                return CHARGE_NOMEM;
2486
2487        ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2488        if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2489                return CHARGE_RETRY;
2490        /*
2491         * Even though the limit is exceeded at this point, reclaim
2492         * may have been able to free some pages.  Retry the charge
2493         * before killing the task.
2494         *
2495         * Only for regular pages, though: huge pages are rather
2496         * unlikely to succeed so close to the limit, and we fall back
2497         * to regular pages anyway in case of failure.
2498         */
2499        if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2500                return CHARGE_RETRY;
2501
2502        /*
2503         * At task move, charge accounts can be doubly counted. So, it's
2504         * better to wait until the end of task_move if something is going on.
2505         */
2506        if (mem_cgroup_wait_acct_move(mem_over_limit))
2507                return CHARGE_RETRY;
2508
2509        /* If we don't need to call oom-killer at el, return immediately */
2510        if (!oom_check)
2511                return CHARGE_NOMEM;
2512        /* check OOM */
2513        if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2514                return CHARGE_OOM_DIE;
2515
2516        return CHARGE_RETRY;
2517}
2518
2519/*
2520 * __mem_cgroup_try_charge() does
2521 * 1. detect memcg to be charged against from passed *mm and *ptr,
2522 * 2. update res_counter
2523 * 3. call memory reclaim if necessary.
2524 *
2525 * In some special case, if the task is fatal, fatal_signal_pending() or
2526 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2527 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2528 * as possible without any hazards. 2: all pages should have a valid
2529 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2530 * pointer, that is treated as a charge to root_mem_cgroup.
2531 *
2532 * So __mem_cgroup_try_charge() will return
2533 *  0       ...  on success, filling *ptr with a valid memcg pointer.
2534 *  -ENOMEM ...  charge failure because of resource limits.
2535 *  -EINTR  ...  if thread is fatal. *ptr is filled with root_mem_cgroup.
2536 *
2537 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2538 * the oom-killer can be invoked.
2539 */
2540static int __mem_cgroup_try_charge(struct mm_struct *mm,
2541                                   gfp_t gfp_mask,
2542                                   unsigned int nr_pages,
2543                                   struct mem_cgroup **ptr,
2544                                   bool oom)
2545{
2546        unsigned int batch = max(CHARGE_BATCH, nr_pages);
2547        int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2548        struct mem_cgroup *memcg = NULL;
2549        int ret;
2550
2551        /*
2552         * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2553         * in system level. So, allow to go ahead dying process in addition to
2554         * MEMDIE process.
2555         */
2556        if (unlikely(test_thread_flag(TIF_MEMDIE)
2557                     || fatal_signal_pending(current)))
2558                goto bypass;
2559
2560        /*
2561         * We always charge the cgroup the mm_struct belongs to.
2562         * The mm_struct's mem_cgroup changes on task migration if the
2563         * thread group leader migrates. It's possible that mm is not
2564         * set, if so charge the root memcg (happens for pagecache usage).
2565         */
2566        if (!*ptr && !mm)
2567                *ptr = root_mem_cgroup;
2568again:
2569        if (*ptr) { /* css should be a valid one */
2570                memcg = *ptr;
2571                if (mem_cgroup_is_root(memcg))
2572                        goto done;
2573                if (consume_stock(memcg, nr_pages))
2574                        goto done;
2575                css_get(&memcg->css);
2576        } else {
2577                struct task_struct *p;
2578
2579                rcu_read_lock();
2580                p = rcu_dereference(mm->owner);
2581                /*
2582                 * Because we don't have task_lock(), "p" can exit.
2583                 * In that case, "memcg" can point to root or p can be NULL with
2584                 * race with swapoff. Then, we have small risk of mis-accouning.
2585                 * But such kind of mis-account by race always happens because
2586                 * we don't have cgroup_mutex(). It's overkill and we allo that
2587                 * small race, here.
2588                 * (*) swapoff at el will charge against mm-struct not against
2589                 * task-struct. So, mm->owner can be NULL.
2590                 */
2591                memcg = mem_cgroup_from_task(p);
2592                if (!memcg)
2593                        memcg = root_mem_cgroup;
2594                if (mem_cgroup_is_root(memcg)) {
2595                        rcu_read_unlock();
2596                        goto done;
2597                }
2598                if (consume_stock(memcg, nr_pages)) {
2599                        /*
2600                         * It seems dagerous to access memcg without css_get().
2601                         * But considering how consume_stok works, it's not
2602                         * necessary. If consume_stock success, some charges
2603                         * from this memcg are cached on this cpu. So, we
2604                         * don't need to call css_get()/css_tryget() before
2605                         * calling consume_stock().
2606                         */
2607                        rcu_read_unlock();
2608                        goto done;
2609                }
2610                /* after here, we may be blocked. we need to get refcnt */
2611                if (!css_tryget(&memcg->css)) {
2612                        rcu_read_unlock();
2613                        goto again;
2614                }
2615                rcu_read_unlock();
2616        }
2617
2618        do {
2619                bool oom_check;
2620
2621                /* If killed, bypass charge */
2622                if (fatal_signal_pending(current)) {
2623                        css_put(&memcg->css);
2624                        goto bypass;
2625                }
2626
2627                oom_check = false;
2628                if (oom && !nr_oom_retries) {
2629                        oom_check = true;
2630                        nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2631                }
2632
2633                ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2634                    oom_check);
2635                switch (ret) {
2636                case CHARGE_OK:
2637                        break;
2638                case CHARGE_RETRY: /* not in OOM situation but retry */
2639                        batch = nr_pages;
2640                        css_put(&memcg->css);
2641                        memcg = NULL;
2642                        goto again;
2643                case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2644                        css_put(&memcg->css);
2645                        goto nomem;
2646                case CHARGE_NOMEM: /* OOM routine works */
2647                        if (!oom) {
2648                                css_put(&memcg->css);
2649                                goto nomem;
2650                        }
2651                        /* If oom, we never return -ENOMEM */
2652                        nr_oom_retries--;
2653                        break;
2654                case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2655                        css_put(&memcg->css);
2656                        goto bypass;
2657                }
2658        } while (ret != CHARGE_OK);
2659
2660        if (batch > nr_pages)
2661                refill_stock(memcg, batch - nr_pages);
2662        css_put(&memcg->css);
2663done:
2664        *ptr = memcg;
2665        return 0;
2666nomem:
2667        *ptr = NULL;
2668        return -ENOMEM;
2669bypass:
2670        *ptr = root_mem_cgroup;
2671        return -EINTR;
2672}
2673
2674/*
2675 * Somemtimes we have to undo a charge we got by try_charge().
2676 * This function is for that and do uncharge, put css's refcnt.
2677 * gotten by try_charge().
2678 */
2679static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2680                                       unsigned int nr_pages)
2681{
2682        if (!mem_cgroup_is_root(memcg)) {
2683                unsigned long bytes = nr_pages * PAGE_SIZE;
2684
2685                res_counter_uncharge(&memcg->res, bytes);
2686                if (do_swap_account)
2687                        res_counter_uncharge(&memcg->memsw, bytes);
2688        }
2689}
2690
2691/*
2692 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2693 * This is useful when moving usage to parent cgroup.
2694 */
2695static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2696                                        unsigned int nr_pages)
2697{
2698        unsigned long bytes = nr_pages * PAGE_SIZE;
2699
2700        if (mem_cgroup_is_root(memcg))
2701                return;
2702
2703        res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2704        if (do_swap_account)
2705                res_counter_uncharge_until(&memcg->memsw,
2706                                                memcg->memsw.parent, bytes);
2707}
2708
2709/*
2710 * A helper function to get mem_cgroup from ID. must be called under
2711 * rcu_read_lock().  The caller is responsible for calling css_tryget if
2712 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2713 * called against removed memcg.)
2714 */
2715static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2716{
2717        struct cgroup_subsys_state *css;
2718
2719        /* ID 0 is unused ID */
2720        if (!id)
2721                return NULL;
2722        css = css_lookup(&mem_cgroup_subsys, id);
2723        if (!css)
2724                return NULL;
2725        return mem_cgroup_from_css(css);
2726}
2727
2728struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2729{
2730        struct mem_cgroup *memcg = NULL;
2731        struct page_cgroup *pc;
2732        unsigned short id;
2733        swp_entry_t ent;
2734
2735        VM_BUG_ON(!PageLocked(page));
2736
2737        pc = lookup_page_cgroup(page);
2738        lock_page_cgroup(pc);
2739        if (PageCgroupUsed(pc)) {
2740                memcg = pc->mem_cgroup;
2741                if (memcg && !css_tryget(&memcg->css))
2742                        memcg = NULL;
2743        } else if (PageSwapCache(page)) {
2744                ent.val = page_private(page);
2745                id = lookup_swap_cgroup_id(ent);
2746                rcu_read_lock();
2747                memcg = mem_cgroup_lookup(id);
2748                if (memcg && !css_tryget(&memcg->css))
2749                        memcg = NULL;
2750                rcu_read_unlock();
2751        }
2752        unlock_page_cgroup(pc);
2753        return memcg;
2754}
2755
2756static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2757                                       struct page *page,
2758                                       unsigned int nr_pages,
2759                                       enum charge_type ctype,
2760                                       bool lrucare)
2761{
2762        struct page_cgroup *pc = lookup_page_cgroup(page);
2763        struct zone *uninitialized_var(zone);
2764        struct lruvec *lruvec;
2765        bool was_on_lru = false;
2766        bool anon;
2767
2768        lock_page_cgroup(pc);
2769        VM_BUG_ON(PageCgroupUsed(pc));
2770        /*
2771         * we don't need page_cgroup_lock about tail pages, becase they are not
2772         * accessed by any other context at this point.
2773         */
2774
2775        /*
2776         * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2777         * may already be on some other mem_cgroup's LRU.  Take care of it.
2778         */
2779        if (lrucare) {
2780                zone = page_zone(page);
2781                spin_lock_irq(&zone->lru_lock);
2782                if (PageLRU(page)) {
2783                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2784                        ClearPageLRU(page);
2785                        del_page_from_lru_list(page, lruvec, page_lru(page));
2786                        was_on_lru = true;
2787                }
2788        }
2789
2790        pc->mem_cgroup = memcg;
2791        /*
2792         * We access a page_cgroup asynchronously without lock_page_cgroup().
2793         * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2794         * is accessed after testing USED bit. To make pc->mem_cgroup visible
2795         * before USED bit, we need memory barrier here.
2796         * See mem_cgroup_add_lru_list(), etc.
2797         */
2798        smp_wmb();
2799        SetPageCgroupUsed(pc);
2800
2801        if (lrucare) {
2802                if (was_on_lru) {
2803                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2804                        VM_BUG_ON(PageLRU(page));
2805                        SetPageLRU(page);
2806                        add_page_to_lru_list(page, lruvec, page_lru(page));
2807                }
2808                spin_unlock_irq(&zone->lru_lock);
2809        }
2810
2811        if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2812                anon = true;
2813        else
2814                anon = false;
2815
2816        mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2817        unlock_page_cgroup(pc);
2818
2819        /*
2820         * "charge_statistics" updated event counter. Then, check it.
2821         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2822         * if they exceeds softlimit.
2823         */
2824        memcg_check_events(memcg, page);
2825}
2826
2827static DEFINE_MUTEX(set_limit_mutex);
2828
2829#ifdef CONFIG_MEMCG_KMEM
2830static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2831{
2832        return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2833                (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2834}
2835
2836/*
2837 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2838 * in the memcg_cache_params struct.
2839 */
2840static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2841{
2842        struct kmem_cache *cachep;
2843
2844        VM_BUG_ON(p->is_root_cache);
2845        cachep = p->root_cache;
2846        return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2847}
2848
2849#ifdef CONFIG_SLABINFO
2850static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2851                                        struct seq_file *m)
2852{
2853        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2854        struct memcg_cache_params *params;
2855
2856        if (!memcg_can_account_kmem(memcg))
2857                return -EIO;
2858
2859        print_slabinfo_header(m);
2860
2861        mutex_lock(&memcg->slab_caches_mutex);
2862        list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2863                cache_show(memcg_params_to_cache(params), m);
2864        mutex_unlock(&memcg->slab_caches_mutex);
2865
2866        return 0;
2867}
2868#endif
2869
2870static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2871{
2872        struct res_counter *fail_res;
2873        struct mem_cgroup *_memcg;
2874        int ret = 0;
2875        bool may_oom;
2876
2877        ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2878        if (ret)
2879                return ret;
2880
2881        /*
2882         * Conditions under which we can wait for the oom_killer. Those are
2883         * the same conditions tested by the core page allocator
2884         */
2885        may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2886
2887        _memcg = memcg;
2888        ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2889                                      &_memcg, may_oom);
2890
2891        if (ret == -EINTR)  {
2892                /*
2893                 * __mem_cgroup_try_charge() chosed to bypass to root due to
2894                 * OOM kill or fatal signal.  Since our only options are to
2895                 * either fail the allocation or charge it to this cgroup, do
2896                 * it as a temporary condition. But we can't fail. From a
2897                 * kmem/slab perspective, the cache has already been selected,
2898                 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2899                 * our minds.
2900                 *
2901                 * This condition will only trigger if the task entered
2902                 * memcg_charge_kmem in a sane state, but was OOM-killed during
2903                 * __mem_cgroup_try_charge() above. Tasks that were already
2904                 * dying when the allocation triggers should have been already
2905                 * directed to the root cgroup in memcontrol.h
2906                 */
2907                res_counter_charge_nofail(&memcg->res, size, &fail_res);
2908                if (do_swap_account)
2909                        res_counter_charge_nofail(&memcg->memsw, size,
2910                                                  &fail_res);
2911                ret = 0;
2912        } else if (ret)
2913                res_counter_uncharge(&memcg->kmem, size);
2914
2915        return ret;
2916}
2917
2918static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2919{
2920        res_counter_uncharge(&memcg->res, size);
2921        if (do_swap_account)
2922                res_counter_uncharge(&memcg->memsw, size);
2923
2924        /* Not down to 0 */
2925        if (res_counter_uncharge(&memcg->kmem, size))
2926                return;
2927
2928        if (memcg_kmem_test_and_clear_dead(memcg))
2929                mem_cgroup_put(memcg);
2930}
2931
2932void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2933{
2934        if (!memcg)
2935                return;
2936
2937        mutex_lock(&memcg->slab_caches_mutex);
2938        list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2939        mutex_unlock(&memcg->slab_caches_mutex);
2940}
2941
2942/*
2943 * helper for acessing a memcg's index. It will be used as an index in the
2944 * child cache array in kmem_cache, and also to derive its name. This function
2945 * will return -1 when this is not a kmem-limited memcg.
2946 */
2947int memcg_cache_id(struct mem_cgroup *memcg)
2948{
2949        return memcg ? memcg->kmemcg_id : -1;
2950}
2951
2952/*
2953 * This ends up being protected by the set_limit mutex, during normal
2954 * operation, because that is its main call site.
2955 *
2956 * But when we create a new cache, we can call this as well if its parent
2957 * is kmem-limited. That will have to hold set_limit_mutex as well.
2958 */
2959int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2960{
2961        int num, ret;
2962
2963        num = ida_simple_get(&kmem_limited_groups,
2964                                0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2965        if (num < 0)
2966                return num;
2967        /*
2968         * After this point, kmem_accounted (that we test atomically in
2969         * the beginning of this conditional), is no longer 0. This
2970         * guarantees only one process will set the following boolean
2971         * to true. We don't need test_and_set because we're protected
2972         * by the set_limit_mutex anyway.
2973         */
2974        memcg_kmem_set_activated(memcg);
2975
2976        ret = memcg_update_all_caches(num+1);
2977        if (ret) {
2978                ida_simple_remove(&kmem_limited_groups, num);
2979                memcg_kmem_clear_activated(memcg);
2980                return ret;
2981        }
2982
2983        memcg->kmemcg_id = num;
2984        INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2985        mutex_init(&memcg->slab_caches_mutex);
2986        return 0;
2987}
2988
2989static size_t memcg_caches_array_size(int num_groups)
2990{
2991        ssize_t size;
2992        if (num_groups <= 0)
2993                return 0;
2994
2995        size = 2 * num_groups;
2996        if (size < MEMCG_CACHES_MIN_SIZE)
2997                size = MEMCG_CACHES_MIN_SIZE;
2998        else if (size > MEMCG_CACHES_MAX_SIZE)
2999                size = MEMCG_CACHES_MAX_SIZE;
3000
3001        return size;
3002}
3003
3004/*
3005 * We should update the current array size iff all caches updates succeed. This
3006 * can only be done from the slab side. The slab mutex needs to be held when
3007 * calling this.
3008 */
3009void memcg_update_array_size(int num)
3010{
3011        if (num > memcg_limited_groups_array_size)
3012                memcg_limited_groups_array_size = memcg_caches_array_size(num);
3013}
3014
3015static void kmem_cache_destroy_work_func(struct work_struct *w);
3016
3017int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3018{
3019        struct memcg_cache_params *cur_params = s->memcg_params;
3020
3021        VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3022
3023        if (num_groups > memcg_limited_groups_array_size) {
3024                int i;
3025                ssize_t size = memcg_caches_array_size(num_groups);
3026
3027                size *= sizeof(void *);
3028                size += sizeof(struct memcg_cache_params);
3029
3030                s->memcg_params = kzalloc(size, GFP_KERNEL);
3031                if (!s->memcg_params) {
3032                        s->memcg_params = cur_params;
3033                        return -ENOMEM;
3034                }
3035
3036                INIT_WORK(&s->memcg_params->destroy,
3037                                kmem_cache_destroy_work_func);
3038                s->memcg_params->is_root_cache = true;
3039
3040                /*
3041                 * There is the chance it will be bigger than
3042                 * memcg_limited_groups_array_size, if we failed an allocation
3043                 * in a cache, in which case all caches updated before it, will
3044                 * have a bigger array.
3045                 *
3046                 * But if that is the case, the data after
3047                 * memcg_limited_groups_array_size is certainly unused
3048                 */
3049                for (i = 0; i < memcg_limited_groups_array_size; i++) {
3050                        if (!cur_params->memcg_caches[i])
3051                                continue;
3052                        s->memcg_params->memcg_caches[i] =
3053                                                cur_params->memcg_caches[i];
3054                }
3055
3056                /*
3057                 * Ideally, we would wait until all caches succeed, and only
3058                 * then free the old one. But this is not worth the extra
3059                 * pointer per-cache we'd have to have for this.
3060                 *
3061                 * It is not a big deal if some caches are left with a size
3062                 * bigger than the others. And all updates will reset this
3063                 * anyway.
3064                 */
3065                kfree(cur_params);
3066        }
3067        return 0;
3068}
3069
3070int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3071                         struct kmem_cache *root_cache)
3072{
3073        size_t size = sizeof(struct memcg_cache_params);
3074
3075        if (!memcg_kmem_enabled())
3076                return 0;
3077
3078        if (!memcg)
3079                size += memcg_limited_groups_array_size * sizeof(void *);
3080
3081        s->memcg_params = kzalloc(size, GFP_KERNEL);
3082        if (!s->memcg_params)
3083                return -ENOMEM;
3084
3085        INIT_WORK(&s->memcg_params->destroy,
3086                        kmem_cache_destroy_work_func);
3087        if (memcg) {
3088                s->memcg_params->memcg = memcg;
3089                s->memcg_params->root_cache = root_cache;
3090        } else
3091                s->memcg_params->is_root_cache = true;
3092
3093        return 0;
3094}
3095
3096void memcg_release_cache(struct kmem_cache *s)
3097{
3098        struct kmem_cache *root;
3099        struct mem_cgroup *memcg;
3100        int id;
3101
3102        /*
3103         * This happens, for instance, when a root cache goes away before we
3104         * add any memcg.
3105         */
3106        if (!s->memcg_params)
3107                return;
3108
3109        if (s->memcg_params->is_root_cache)
3110                goto out;
3111
3112        memcg = s->memcg_params->memcg;
3113        id  = memcg_cache_id(memcg);
3114
3115        root = s->memcg_params->root_cache;
3116        root->memcg_params->memcg_caches[id] = NULL;
3117        mem_cgroup_put(memcg);
3118
3119        mutex_lock(&memcg->slab_caches_mutex);
3120        list_del(&s->memcg_params->list);
3121        mutex_unlock(&memcg->slab_caches_mutex);
3122
3123out:
3124        kfree(s->memcg_params);
3125}
3126
3127/*
3128 * During the creation a new cache, we need to disable our accounting mechanism
3129 * altogether. This is true even if we are not creating, but rather just
3130 * enqueing new caches to be created.
3131 *
3132 * This is because that process will trigger allocations; some visible, like
3133 * explicit kmallocs to auxiliary data structures, name strings and internal
3134 * cache structures; some well concealed, like INIT_WORK() that can allocate
3135 * objects during debug.
3136 *
3137 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3138 * to it. This may not be a bounded recursion: since the first cache creation
3139 * failed to complete (waiting on the allocation), we'll just try to create the
3140 * cache again, failing at the same point.
3141 *
3142 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3143 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3144 * inside the following two functions.
3145 */
3146static inline void memcg_stop_kmem_account(void)
3147{
3148        VM_BUG_ON(!current->mm);
3149        current->memcg_kmem_skip_account++;
3150}
3151
3152static inline void memcg_resume_kmem_account(void)
3153{
3154        VM_BUG_ON(!current->mm);
3155        current->memcg_kmem_skip_account--;
3156}
3157
3158static void kmem_cache_destroy_work_func(struct work_struct *w)
3159{
3160        struct kmem_cache *cachep;
3161        struct memcg_cache_params *p;
3162
3163        p = container_of(w, struct memcg_cache_params, destroy);
3164
3165        cachep = memcg_params_to_cache(p);
3166
3167        /*
3168         * If we get down to 0 after shrink, we could delete right away.
3169         * However, memcg_release_pages() already puts us back in the workqueue
3170         * in that case. If we proceed deleting, we'll get a dangling
3171         * reference, and removing the object from the workqueue in that case
3172         * is unnecessary complication. We are not a fast path.
3173         *
3174         * Note that this case is fundamentally different from racing with
3175         * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3176         * kmem_cache_shrink, not only we would be reinserting a dead cache
3177         * into the queue, but doing so from inside the worker racing to
3178         * destroy it.
3179         *
3180         * So if we aren't down to zero, we'll just schedule a worker and try
3181         * again
3182         */
3183        if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3184                kmem_cache_shrink(cachep);
3185                if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3186                        return;
3187        } else
3188                kmem_cache_destroy(cachep);
3189}
3190
3191void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3192{
3193        if (!cachep->memcg_params->dead)
3194                return;
3195
3196        /*
3197         * There are many ways in which we can get here.
3198         *
3199         * We can get to a memory-pressure situation while the delayed work is
3200         * still pending to run. The vmscan shrinkers can then release all
3201         * cache memory and get us to destruction. If this is the case, we'll
3202         * be executed twice, which is a bug (the second time will execute over
3203         * bogus data). In this case, cancelling the work should be fine.
3204         *
3205         * But we can also get here from the worker itself, if
3206         * kmem_cache_shrink is enough to shake all the remaining objects and
3207         * get the page count to 0. In this case, we'll deadlock if we try to
3208         * cancel the work (the worker runs with an internal lock held, which
3209         * is the same lock we would hold for cancel_work_sync().)
3210         *
3211         * Since we can't possibly know who got us here, just refrain from
3212         * running if there is already work pending
3213         */
3214        if (work_pending(&cachep->memcg_params->destroy))
3215                return;
3216        /*
3217         * We have to defer the actual destroying to a workqueue, because
3218         * we might currently be in a context that cannot sleep.
3219         */
3220        schedule_work(&cachep->memcg_params->destroy);
3221}
3222
3223static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3224{
3225        char *name;
3226        struct dentry *dentry;
3227
3228        rcu_read_lock();
3229        dentry = rcu_dereference(memcg->css.cgroup->dentry);
3230        rcu_read_unlock();
3231
3232        BUG_ON(dentry == NULL);
3233
3234        name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3235                         memcg_cache_id(memcg), dentry->d_name.name);
3236
3237        return name;
3238}
3239
3240static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3241                                         struct kmem_cache *s)
3242{
3243        char *name;
3244        struct kmem_cache *new;
3245
3246        name = memcg_cache_name(memcg, s);
3247        if (!name)
3248                return NULL;
3249
3250        new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3251                                      (s->flags & ~SLAB_PANIC), s->ctor, s);
3252
3253        if (new)
3254                new->allocflags |= __GFP_KMEMCG;
3255
3256        kfree(name);
3257        return new;
3258}
3259
3260/*
3261 * This lock protects updaters, not readers. We want readers to be as fast as
3262 * they can, and they will either see NULL or a valid cache value. Our model
3263 * allow them to see NULL, in which case the root memcg will be selected.
3264 *
3265 * We need this lock because multiple allocations to the same cache from a non
3266 * will span more than one worker. Only one of them can create the cache.
3267 */
3268static DEFINE_MUTEX(memcg_cache_mutex);
3269static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3270                                                  struct kmem_cache *cachep)
3271{
3272        struct kmem_cache *new_cachep;
3273        int idx;
3274
3275        BUG_ON(!memcg_can_account_kmem(memcg));
3276
3277        idx = memcg_cache_id(memcg);
3278
3279        mutex_lock(&memcg_cache_mutex);
3280        new_cachep = cachep->memcg_params->memcg_caches[idx];
3281        if (new_cachep)
3282                goto out;
3283
3284        new_cachep = kmem_cache_dup(memcg, cachep);
3285        if (new_cachep == NULL) {
3286                new_cachep = cachep;
3287                goto out;
3288        }
3289
3290        mem_cgroup_get(memcg);
3291        atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3292
3293        cachep->memcg_params->memcg_caches[idx] = new_cachep;
3294        /*
3295         * the readers won't lock, make sure everybody sees the updated value,
3296         * so they won't put stuff in the queue again for no reason
3297         */
3298        wmb();
3299out:
3300        mutex_unlock(&memcg_cache_mutex);
3301        return new_cachep;
3302}
3303
3304void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3305{
3306        struct kmem_cache *c;
3307        int i;
3308
3309        if (!s->memcg_params)
3310                return;
3311        if (!s->memcg_params->is_root_cache)
3312                return;
3313
3314        /*
3315         * If the cache is being destroyed, we trust that there is no one else
3316         * requesting objects from it. Even if there are, the sanity checks in
3317         * kmem_cache_destroy should caught this ill-case.
3318         *
3319         * Still, we don't want anyone else freeing memcg_caches under our
3320         * noses, which can happen if a new memcg comes to life. As usual,
3321         * we'll take the set_limit_mutex to protect ourselves against this.
3322         */
3323        mutex_lock(&set_limit_mutex);
3324        for (i = 0; i < memcg_limited_groups_array_size; i++) {
3325                c = s->memcg_params->memcg_caches[i];
3326                if (!c)
3327                        continue;
3328
3329                /*
3330                 * We will now manually delete the caches, so to avoid races
3331                 * we need to cancel all pending destruction workers and
3332                 * proceed with destruction ourselves.
3333                 *
3334                 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3335                 * and that could spawn the workers again: it is likely that
3336                 * the cache still have active pages until this very moment.
3337                 * This would lead us back to mem_cgroup_destroy_cache.
3338                 *
3339                 * But that will not execute at all if the "dead" flag is not
3340                 * set, so flip it down to guarantee we are in control.
3341                 */
3342                c->memcg_params->dead = false;
3343                cancel_work_sync(&c->memcg_params->destroy);
3344                kmem_cache_destroy(c);
3345        }
3346        mutex_unlock(&set_limit_mutex);
3347}
3348
3349struct create_work {
3350        struct mem_cgroup *memcg;
3351        struct kmem_cache *cachep;
3352        struct work_struct work;
3353};
3354
3355static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3356{
3357        struct kmem_cache *cachep;
3358        struct memcg_cache_params *params;
3359
3360        if (!memcg_kmem_is_active(memcg))
3361                return;
3362
3363        mutex_lock(&memcg->slab_caches_mutex);
3364        list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3365                cachep = memcg_params_to_cache(params);
3366                cachep->memcg_params->dead = true;
3367                schedule_work(&cachep->memcg_params->destroy);
3368        }
3369        mutex_unlock(&memcg->slab_caches_mutex);
3370}
3371
3372static void memcg_create_cache_work_func(struct work_struct *w)
3373{
3374        struct create_work *cw;
3375
3376        cw = container_of(w, struct create_work, work);
3377        memcg_create_kmem_cache(cw->memcg, cw->cachep);
3378        /* Drop the reference gotten when we enqueued. */
3379        css_put(&cw->memcg->css);
3380        kfree(cw);
3381}
3382
3383/*
3384 * Enqueue the creation of a per-memcg kmem_cache.
3385 * Called with rcu_read_lock.
3386 */
3387static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3388                                         struct kmem_cache *cachep)
3389{
3390        struct create_work *cw;
3391
3392        cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3393        if (cw == NULL)
3394                return;
3395
3396        /* The corresponding put will be done in the workqueue. */
3397        if (!css_tryget(&memcg->css)) {
3398                kfree(cw);
3399                return;
3400        }
3401
3402        cw->memcg = memcg;
3403        cw->cachep = cachep;
3404
3405        INIT_WORK(&cw->work, memcg_create_cache_work_func);
3406        schedule_work(&cw->work);
3407}
3408
3409static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3410                                       struct kmem_cache *cachep)
3411{
3412        /*
3413         * We need to stop accounting when we kmalloc, because if the
3414         * corresponding kmalloc cache is not yet created, the first allocation
3415         * in __memcg_create_cache_enqueue will recurse.
3416         *
3417         * However, it is better to enclose the whole function. Depending on
3418         * the debugging options enabled, INIT_WORK(), for instance, can
3419         * trigger an allocation. This too, will make us recurse. Because at
3420         * this point we can't allow ourselves back into memcg_kmem_get_cache,
3421         * the safest choice is to do it like this, wrapping the whole function.
3422         */
3423        memcg_stop_kmem_account();
3424        __memcg_create_cache_enqueue(memcg, cachep);
3425        memcg_resume_kmem_account();
3426}
3427/*
3428 * Return the kmem_cache we're supposed to use for a slab allocation.
3429 * We try to use the current memcg's version of the cache.
3430 *
3431 * If the cache does not exist yet, if we are the first user of it,
3432 * we either create it immediately, if possible, or create it asynchronously
3433 * in a workqueue.
3434 * In the latter case, we will let the current allocation go through with
3435 * the original cache.
3436 *
3437 * Can't be called in interrupt context or from kernel threads.
3438 * This function needs to be called with rcu_read_lock() held.
3439 */
3440struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3441                                          gfp_t gfp)
3442{
3443        struct mem_cgroup *memcg;
3444        int idx;
3445
3446        VM_BUG_ON(!cachep->memcg_params);
3447        VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3448
3449        if (!current->mm || current->memcg_kmem_skip_account)
3450                return cachep;
3451
3452        rcu_read_lock();
3453        memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3454        rcu_read_unlock();
3455
3456        if (!memcg_can_account_kmem(memcg))
3457                return cachep;
3458
3459        idx = memcg_cache_id(memcg);
3460
3461        /*
3462         * barrier to mare sure we're always seeing the up to date value.  The
3463         * code updating memcg_caches will issue a write barrier to match this.
3464         */
3465        read_barrier_depends();
3466        if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3467                /*
3468                 * If we are in a safe context (can wait, and not in interrupt
3469                 * context), we could be be predictable and return right away.
3470                 * This would guarantee that the allocation being performed
3471                 * already belongs in the new cache.
3472                 *
3473                 * However, there are some clashes that can arrive from locking.
3474                 * For instance, because we acquire the slab_mutex while doing
3475                 * kmem_cache_dup, this means no further allocation could happen
3476                 * with the slab_mutex held.
3477                 *
3478                 * Also, because cache creation issue get_online_cpus(), this
3479                 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3480                 * that ends up reversed during cpu hotplug. (cpuset allocates
3481                 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3482                 * better to defer everything.
3483                 */
3484                memcg_create_cache_enqueue(memcg, cachep);
3485                return cachep;
3486        }
3487
3488        return cachep->memcg_params->memcg_caches[idx];
3489}
3490EXPORT_SYMBOL(__memcg_kmem_get_cache);
3491
3492/*
3493 * We need to verify if the allocation against current->mm->owner's memcg is
3494 * possible for the given order. But the page is not allocated yet, so we'll
3495 * need a further commit step to do the final arrangements.
3496 *
3497 * It is possible for the task to switch cgroups in this mean time, so at
3498 * commit time, we can't rely on task conversion any longer.  We'll then use
3499 * the handle argument to return to the caller which cgroup we should commit
3500 * against. We could also return the memcg directly and avoid the pointer
3501 * passing, but a boolean return value gives better semantics considering
3502 * the compiled-out case as well.
3503 *
3504 * Returning true means the allocation is possible.
3505 */
3506bool
3507__memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3508{
3509        struct mem_cgroup *memcg;
3510        int ret;
3511
3512        *_memcg = NULL;
3513        memcg = try_get_mem_cgroup_from_mm(current->mm);
3514
3515        /*
3516         * very rare case described in mem_cgroup_from_task. Unfortunately there
3517         * isn't much we can do without complicating this too much, and it would
3518         * be gfp-dependent anyway. Just let it go
3519         */
3520        if (unlikely(!memcg))
3521                return true;
3522
3523        if (!memcg_can_account_kmem(memcg)) {
3524                css_put(&memcg->css);
3525                return true;
3526        }
3527
3528        ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3529        if (!ret)
3530                *_memcg = memcg;
3531
3532        css_put(&memcg->css);
3533        return (ret == 0);
3534}
3535
3536void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3537                              int order)
3538{
3539        struct page_cgroup *pc;
3540
3541        VM_BUG_ON(mem_cgroup_is_root(memcg));
3542
3543        /* The page allocation failed. Revert */
3544        if (!page) {
3545                memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3546                return;
3547        }
3548
3549        pc = lookup_page_cgroup(page);
3550        lock_page_cgroup(pc);
3551        pc->mem_cgroup = memcg;
3552        SetPageCgroupUsed(pc);
3553        unlock_page_cgroup(pc);
3554}
3555
3556void __memcg_kmem_uncharge_pages(struct page *page, int order)
3557{
3558        struct mem_cgroup *memcg = NULL;
3559        struct page_cgroup *pc;
3560
3561
3562        pc = lookup_page_cgroup(page);
3563        /*
3564         * Fast unlocked return. Theoretically might have changed, have to
3565         * check again after locking.
3566         */
3567        if (!PageCgroupUsed(pc))
3568                return;
3569
3570        lock_page_cgroup(pc);
3571        if (PageCgroupUsed(pc)) {
3572                memcg = pc->mem_cgroup;
3573                ClearPageCgroupUsed(pc);
3574        }
3575        unlock_page_cgroup(pc);
3576
3577        /*
3578         * We trust that only if there is a memcg associated with the page, it
3579         * is a valid allocation
3580         */
3581        if (!memcg)
3582                return;
3583
3584        VM_BUG_ON(mem_cgroup_is_root(memcg));
3585        memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3586}
3587#else
3588static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3589{
3590}
3591#endif /* CONFIG_MEMCG_KMEM */
3592
3593#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3594
3595#define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3596/*
3597 * Because tail pages are not marked as "used", set it. We're under
3598 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3599 * charge/uncharge will be never happen and move_account() is done under
3600 * compound_lock(), so we don't have to take care of races.
3601 */
3602void mem_cgroup_split_huge_fixup(struct page *head)
3603{
3604        struct page_cgroup *head_pc = lookup_page_cgroup(head);
3605        struct page_cgroup *pc;
3606        int i;
3607
3608        if (mem_cgroup_disabled())
3609                return;
3610        for (i = 1; i < HPAGE_PMD_NR; i++) {
3611                pc = head_pc + i;
3612                pc->mem_cgroup = head_pc->mem_cgroup;
3613                smp_wmb();/* see __commit_charge() */
3614                pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3615        }
3616}
3617#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3618
3619/**
3620 * mem_cgroup_move_account - move account of the page
3621 * @page: the page
3622 * @nr_pages: number of regular pages (>1 for huge pages)
3623 * @pc: page_cgroup of the page.
3624 * @from: mem_cgroup which the page is moved from.
3625 * @to: mem_cgroup which the page is moved to. @from != @to.
3626 *
3627 * The caller must confirm following.
3628 * - page is not on LRU (isolate_page() is useful.)
3629 * - compound_lock is held when nr_pages > 1
3630 *
3631 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3632 * from old cgroup.
3633 */
3634static int mem_cgroup_move_account(struct page *page,
3635                                   unsigned int nr_pages,
3636                                   struct page_cgroup *pc,
3637                                   struct mem_cgroup *from,
3638                                   struct mem_cgroup *to)
3639{
3640        unsigned long flags;
3641        int ret;
3642        bool anon = PageAnon(page);
3643
3644        VM_BUG_ON(from == to);
3645        VM_BUG_ON(PageLRU(page));
3646        /*
3647         * The page is isolated from LRU. So, collapse function
3648         * will not handle this page. But page splitting can happen.
3649         * Do this check under compound_page_lock(). The caller should
3650         * hold it.
3651         */
3652        ret = -EBUSY;
3653        if (nr_pages > 1 && !PageTransHuge(page))
3654                goto out;
3655
3656        lock_page_cgroup(pc);
3657
3658        ret = -EINVAL;
3659        if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3660                goto unlock;
3661
3662        move_lock_mem_cgroup(from, &flags);
3663
3664        if (!anon && page_mapped(page)) {
3665                /* Update mapped_file data for mem_cgroup */
3666                preempt_disable();
3667                __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3668                __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3669                preempt_enable();
3670        }
3671        mem_cgroup_charge_statistics(from, anon, -nr_pages);
3672
3673        /* caller should have done css_get */
3674        pc->mem_cgroup = to;
3675        mem_cgroup_charge_statistics(to, anon, nr_pages);
3676        move_unlock_mem_cgroup(from, &flags);
3677        ret = 0;
3678unlock:
3679        unlock_page_cgroup(pc);
3680        /*
3681         * check events
3682         */
3683        memcg_check_events(to, page);
3684        memcg_check_events(from, page);
3685out:
3686        return ret;
3687}
3688
3689/**
3690 * mem_cgroup_move_parent - moves page to the parent group
3691 * @page: the page to move
3692 * @pc: page_cgroup of the page
3693 * @child: page's cgroup
3694 *
3695 * move charges to its parent or the root cgroup if the group has no
3696 * parent (aka use_hierarchy==0).
3697 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3698 * mem_cgroup_move_account fails) the failure is always temporary and
3699 * it signals a race with a page removal/uncharge or migration. In the
3700 * first case the page is on the way out and it will vanish from the LRU
3701 * on the next attempt and the call should be retried later.
3702 * Isolation from the LRU fails only if page has been isolated from
3703 * the LRU since we looked at it and that usually means either global
3704 * reclaim or migration going on. The page will either get back to the
3705 * LRU or vanish.
3706 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3707 * (!PageCgroupUsed) or moved to a different group. The page will
3708 * disappear in the next attempt.
3709 */
3710static int mem_cgroup_move_parent(struct page *page,
3711                                  struct page_cgroup *pc,
3712                                  struct mem_cgroup *child)
3713{
3714        struct mem_cgroup *parent;
3715        unsigned int nr_pages;
3716        unsigned long uninitialized_var(flags);
3717        int ret;
3718
3719        VM_BUG_ON(mem_cgroup_is_root(child));
3720
3721        ret = -EBUSY;
3722        if (!get_page_unless_zero(page))
3723                goto out;
3724        if (isolate_lru_page(page))
3725                goto put;
3726
3727        nr_pages = hpage_nr_pages(page);
3728
3729        parent = parent_mem_cgroup(child);
3730        /*
3731         * If no parent, move charges to root cgroup.
3732         */
3733        if (!parent)
3734                parent = root_mem_cgroup;
3735
3736        if (nr_pages > 1) {
3737                VM_BUG_ON(!PageTransHuge(page));
3738                flags = compound_lock_irqsave(page);
3739        }
3740
3741        ret = mem_cgroup_move_account(page, nr_pages,
3742                                pc, child, parent);
3743        if (!ret)
3744                __mem_cgroup_cancel_local_charge(child, nr_pages);
3745
3746        if (nr_pages > 1)
3747                compound_unlock_irqrestore(page, flags);
3748        putback_lru_page(page);
3749put:
3750        put_page(page);
3751out:
3752        return ret;
3753}
3754
3755/*
3756 * Charge the memory controller for page usage.
3757 * Return
3758 * 0 if the charge was successful
3759 * < 0 if the cgroup is over its limit
3760 */
3761static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3762                                gfp_t gfp_mask, enum charge_type ctype)
3763{
3764        struct mem_cgroup *memcg = NULL;
3765        unsigned int nr_pages = 1;
3766        bool oom = true;
3767        int ret;
3768
3769        if (PageTransHuge(page)) {
3770                nr_pages <<= compound_order(page);
3771                VM_BUG_ON(!PageTransHuge(page));
3772                /*
3773                 * Never OOM-kill a process for a huge page.  The
3774                 * fault handler will fall back to regular pages.
3775                 */
3776                oom = false;
3777        }
3778
3779        ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3780        if (ret == -ENOMEM)
3781                return ret;
3782        __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3783        return 0;
3784}
3785
3786int mem_cgroup_newpage_charge(struct page *page,
3787                              struct mm_struct *mm, gfp_t gfp_mask)
3788{
3789        if (mem_cgroup_disabled())
3790                return 0;
3791        VM_BUG_ON(page_mapped(page));
3792        VM_BUG_ON(page->mapping && !PageAnon(page));
3793        VM_BUG_ON(!mm);
3794        return mem_cgroup_charge_common(page, mm, gfp_mask,
3795                                        MEM_CGROUP_CHARGE_TYPE_ANON);
3796}
3797
3798/*
3799 * While swap-in, try_charge -> commit or cancel, the page is locked.
3800 * And when try_charge() successfully returns, one refcnt to memcg without
3801 * struct page_cgroup is acquired. This refcnt will be consumed by
3802 * "commit()" or removed by "cancel()"
3803 */
3804static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3805                                          struct page *page,
3806                                          gfp_t mask,
3807                                          struct mem_cgroup **memcgp)
3808{
3809        struct mem_cgroup *memcg;
3810        struct page_cgroup *pc;
3811        int ret;
3812
3813        pc = lookup_page_cgroup(page);
3814        /*
3815         * Every swap fault against a single page tries to charge the
3816         * page, bail as early as possible.  shmem_unuse() encounters
3817         * already charged pages, too.  The USED bit is protected by
3818         * the page lock, which serializes swap cache removal, which
3819         * in turn serializes uncharging.
3820         */
3821        if (PageCgroupUsed(pc))
3822                return 0;
3823        if (!do_swap_account)
3824                goto charge_cur_mm;
3825        memcg = try_get_mem_cgroup_from_page(page);
3826        if (!memcg)
3827                goto charge_cur_mm;
3828        *memcgp = memcg;
3829        ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3830        css_put(&memcg->css);
3831        if (ret == -EINTR)
3832                ret = 0;
3833        return ret;
3834charge_cur_mm:
3835        ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3836        if (ret == -EINTR)
3837                ret = 0;
3838        return ret;
3839}
3840
3841int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3842                                 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3843{
3844        *memcgp = NULL;
3845        if (mem_cgroup_disabled())
3846                return 0;
3847        /*
3848         * A racing thread's fault, or swapoff, may have already
3849         * updated the pte, and even removed page from swap cache: in
3850         * those cases unuse_pte()'s pte_same() test will fail; but
3851         * there's also a KSM case which does need to charge the page.
3852         */
3853        if (!PageSwapCache(page)) {
3854                int ret;
3855
3856                ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3857                if (ret == -EINTR)
3858                        ret = 0;
3859                return ret;
3860        }
3861        return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3862}
3863
3864void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3865{
3866        if (mem_cgroup_disabled())
3867                return;
3868        if (!memcg)
3869                return;
3870        __mem_cgroup_cancel_charge(memcg, 1);
3871}
3872
3873static void
3874__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3875                                        enum charge_type ctype)
3876{
3877        if (mem_cgroup_disabled())
3878                return;
3879        if (!memcg)
3880                return;
3881
3882        __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3883        /*
3884         * Now swap is on-memory. This means this page may be
3885         * counted both as mem and swap....double count.
3886         * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3887         * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3888         * may call delete_from_swap_cache() before reach here.
3889         */
3890        if (do_swap_account && PageSwapCache(page)) {
3891                swp_entry_t ent = {.val = page_private(page)};
3892                mem_cgroup_uncharge_swap(ent);
3893        }
3894}
3895
3896void mem_cgroup_commit_charge_swapin(struct page *page,
3897                                     struct mem_cgroup *memcg)
3898{
3899        __mem_cgroup_commit_charge_swapin(page, memcg,
3900                                          MEM_CGROUP_CHARGE_TYPE_ANON);
3901}
3902
3903int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3904                                gfp_t gfp_mask)
3905{
3906        struct mem_cgroup *memcg = NULL;
3907        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3908        int ret;
3909
3910        if (mem_cgroup_disabled())
3911                return 0;
3912        if (PageCompound(page))
3913                return 0;
3914
3915        if (!PageSwapCache(page))
3916                ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3917        else { /* page is swapcache/shmem */
3918                ret = __mem_cgroup_try_charge_swapin(mm, page,
3919                                                     gfp_mask, &memcg);
3920                if (!ret)
3921                        __mem_cgroup_commit_charge_swapin(page, memcg, type);
3922        }
3923        return ret;
3924}
3925
3926static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3927                                   unsigned int nr_pages,
3928                                   const enum charge_type ctype)
3929{
3930        struct memcg_batch_info *batch = NULL;
3931        bool uncharge_memsw = true;
3932
3933        /* If swapout, usage of swap doesn't decrease */
3934        if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3935                uncharge_memsw = false;
3936
3937        batch = &current->memcg_batch;
3938        /*
3939         * In usual, we do css_get() when we remember memcg pointer.
3940         * But in this case, we keep res->usage until end of a series of
3941         * uncharges. Then, it's ok to ignore memcg's refcnt.
3942         */
3943        if (!batch->memcg)
3944                batch->memcg = memcg;
3945        /*
3946         * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3947         * In those cases, all pages freed continuously can be expected to be in
3948         * the same cgroup and we have chance to coalesce uncharges.
3949         * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3950         * because we want to do uncharge as soon as possible.
3951         */
3952
3953        if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3954                goto direct_uncharge;
3955
3956        if (nr_pages > 1)
3957                goto direct_uncharge;
3958
3959        /*
3960         * In typical case, batch->memcg == mem. This means we can
3961         * merge a series of uncharges to an uncharge of res_counter.
3962         * If not, we uncharge res_counter ony by one.
3963         */
3964        if (batch->memcg != memcg)
3965                goto direct_uncharge;
3966        /* remember freed charge and uncharge it later */
3967        batch->nr_pages++;
3968        if (uncharge_memsw)
3969                batch->memsw_nr_pages++;
3970        return;
3971direct_uncharge:
3972        res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3973        if (uncharge_memsw)
3974                res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3975        if (unlikely(batch->memcg != memcg))
3976                memcg_oom_recover(memcg);
3977}
3978
3979/*
3980 * uncharge if !page_mapped(page)
3981 */
3982static struct mem_cgroup *
3983__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3984                             bool end_migration)
3985{
3986        struct mem_cgroup *memcg = NULL;
3987        unsigned int nr_pages = 1;
3988        struct page_cgroup *pc;
3989        bool anon;
3990
3991        if (mem_cgroup_disabled())
3992                return NULL;
3993
3994        if (PageTransHuge(page)) {
3995                nr_pages <<= compound_order(page);
3996                VM_BUG_ON(!PageTransHuge(page));
3997        }
3998        /*
3999         * Check if our page_cgroup is valid
4000         */
4001        pc = lookup_page_cgroup(page);
4002        if (unlikely(!PageCgroupUsed(pc)))
4003                return NULL;
4004
4005        lock_page_cgroup(pc);
4006
4007        memcg = pc->mem_cgroup;
4008
4009        if (!PageCgroupUsed(pc))
4010                goto unlock_out;
4011
4012        anon = PageAnon(page);
4013
4014        switch (ctype) {
4015        case MEM_CGROUP_CHARGE_TYPE_ANON:
4016                /*
4017                 * Generally PageAnon tells if it's the anon statistics to be
4018                 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4019                 * used before page reached the stage of being marked PageAnon.
4020                 */
4021                anon = true;
4022                /* fallthrough */
4023        case MEM_CGROUP_CHARGE_TYPE_DROP:
4024                /* See mem_cgroup_prepare_migration() */
4025                if (page_mapped(page))
4026                        goto unlock_out;
4027                /*
4028                 * Pages under migration may not be uncharged.  But
4029                 * end_migration() /must/ be the one uncharging the
4030                 * unused post-migration page and so it has to call
4031                 * here with the migration bit still set.  See the
4032                 * res_counter handling below.
4033                 */
4034                if (!end_migration && PageCgroupMigration(pc))
4035                        goto unlock_out;
4036                break;
4037        case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4038                if (!PageAnon(page)) {  /* Shared memory */
4039                        if (page->mapping && !page_is_file_cache(page))
4040                                goto unlock_out;
4041                } else if (page_mapped(page)) /* Anon */
4042                                goto unlock_out;
4043                break;
4044        default:
4045                break;
4046        }
4047
4048        mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
4049
4050        ClearPageCgroupUsed(pc);
4051        /*
4052         * pc->mem_cgroup is not cleared here. It will be accessed when it's
4053         * freed from LRU. This is safe because uncharged page is expected not
4054         * to be reused (freed soon). Exception is SwapCache, it's handled by
4055         * special functions.
4056         */
4057
4058        unlock_page_cgroup(pc);
4059        /*
4060         * even after unlock, we have memcg->res.usage here and this memcg
4061         * will never be freed.
4062         */
4063        memcg_check_events(memcg, page);
4064        if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4065                mem_cgroup_swap_statistics(memcg, true);
4066                mem_cgroup_get(memcg);
4067        }
4068        /*
4069         * Migration does not charge the res_counter for the
4070         * replacement page, so leave it alone when phasing out the
4071         * page that is unused after the migration.
4072         */
4073        if (!end_migration && !mem_cgroup_is_root(memcg))
4074                mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4075
4076        return memcg;
4077
4078unlock_out:
4079        unlock_page_cgroup(pc);
4080        return NULL;
4081}
4082
4083void mem_cgroup_uncharge_page(struct page *page)
4084{
4085        /* early check. */
4086        if (page_mapped(page))
4087                return;
4088        VM_BUG_ON(page->mapping && !PageAnon(page));
4089        /*
4090         * If the page is in swap cache, uncharge should be deferred
4091         * to the swap path, which also properly accounts swap usage
4092         * and handles memcg lifetime.
4093         *
4094         * Note that this check is not stable and reclaim may add the
4095         * page to swap cache at any time after this.  However, if the
4096         * page is not in swap cache by the time page->mapcount hits
4097         * 0, there won't be any page table references to the swap
4098         * slot, and reclaim will free it and not actually write the
4099         * page to disk.
4100         */
4101        if (PageSwapCache(page))
4102                return;
4103        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4104}
4105
4106void mem_cgroup_uncharge_cache_page(struct page *page)
4107{
4108        VM_BUG_ON(page_mapped(page));
4109        VM_BUG_ON(page->mapping);
4110        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4111}
4112
4113/*
4114 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4115 * In that cases, pages are freed continuously and we can expect pages
4116 * are in the same memcg. All these calls itself limits the number of
4117 * pages freed at once, then uncharge_start/end() is called properly.
4118 * This may be called prural(2) times in a context,
4119 */
4120
4121void mem_cgroup_uncharge_start(void)
4122{
4123        current->memcg_batch.do_batch++;
4124        /* We can do nest. */
4125        if (current->memcg_batch.do_batch == 1) {
4126                current->memcg_batch.memcg = NULL;
4127                current->memcg_batch.nr_pages = 0;
4128                current->memcg_batch.memsw_nr_pages = 0;
4129        }
4130}
4131
4132void mem_cgroup_uncharge_end(void)
4133{
4134        struct memcg_batch_info *batch = &current->memcg_batch;
4135
4136        if (!batch->do_batch)
4137                return;
4138
4139        batch->do_batch--;
4140        if (batch->do_batch) /* If stacked, do nothing. */
4141                return;
4142
4143        if (!batch->memcg)
4144                return;
4145        /*
4146         * This "batch->memcg" is valid without any css_get/put etc...
4147         * bacause we hide charges behind us.
4148         */
4149        if (batch->nr_pages)
4150                res_counter_uncharge(&batch->memcg->res,
4151                                     batch->nr_pages * PAGE_SIZE);
4152        if (batch->memsw_nr_pages)
4153                res_counter_uncharge(&batch->memcg->memsw,
4154                                     batch->memsw_nr_pages * PAGE_SIZE);
4155        memcg_oom_recover(batch->memcg);
4156        /* forget this pointer (for sanity check) */
4157        batch->memcg = NULL;
4158}
4159
4160#ifdef CONFIG_SWAP
4161/*
4162 * called after __delete_from_swap_cache() and drop "page" account.
4163 * memcg information is recorded to swap_cgroup of "ent"
4164 */
4165void
4166mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4167{
4168        struct mem_cgroup *memcg;
4169        int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4170
4171        if (!swapout) /* this was a swap cache but the swap is unused ! */
4172                ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4173
4174        memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4175
4176        /*
4177         * record memcg information,  if swapout && memcg != NULL,
4178         * mem_cgroup_get() was called in uncharge().
4179         */
4180        if (do_swap_account && swapout && memcg)
4181                swap_cgroup_record(ent, css_id(&memcg->css));
4182}
4183#endif
4184
4185#ifdef CONFIG_MEMCG_SWAP
4186/*
4187 * called from swap_entry_free(). remove record in swap_cgroup and
4188 * uncharge "memsw" account.
4189 */
4190void mem_cgroup_uncharge_swap(swp_entry_t ent)
4191{
4192        struct mem_cgroup *memcg;
4193        unsigned short id;
4194
4195        if (!do_swap_account)
4196                return;
4197
4198        id = swap_cgroup_record(ent, 0);
4199        rcu_read_lock();
4200        memcg = mem_cgroup_lookup(id);
4201        if (memcg) {
4202                /*
4203                 * We uncharge this because swap is freed.
4204                 * This memcg can be obsolete one. We avoid calling css_tryget
4205                 */
4206                if (!mem_cgroup_is_root(memcg))
4207                        res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4208                mem_cgroup_swap_statistics(memcg, false);
4209                mem_cgroup_put(memcg);
4210        }
4211        rcu_read_unlock();
4212}
4213
4214/**
4215 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4216 * @entry: swap entry to be moved
4217 * @from:  mem_cgroup which the entry is moved from
4218 * @to:  mem_cgroup which the entry is moved to
4219 *
4220 * It succeeds only when the swap_cgroup's record for this entry is the same
4221 * as the mem_cgroup's id of @from.
4222 *
4223 * Returns 0 on success, -EINVAL on failure.
4224 *
4225 * The caller must have charged to @to, IOW, called res_counter_charge() about
4226 * both res and memsw, and called css_get().
4227 */
4228static int mem_cgroup_move_swap_account(swp_entry_t entry,
4229                                struct mem_cgroup *from, struct mem_cgroup *to)
4230{
4231        unsigned short old_id, new_id;
4232
4233        old_id = css_id(&from->css);
4234        new_id = css_id(&to->css);
4235
4236        if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4237                mem_cgroup_swap_statistics(from, false);
4238                mem_cgroup_swap_statistics(to, true);
4239                /*
4240                 * This function is only called from task migration context now.
4241                 * It postpones res_counter and refcount handling till the end
4242                 * of task migration(mem_cgroup_clear_mc()) for performance
4243                 * improvement. But we cannot postpone mem_cgroup_get(to)
4244                 * because if the process that has been moved to @to does
4245                 * swap-in, the refcount of @to might be decreased to 0.
4246                 */
4247                mem_cgroup_get(to);
4248                return 0;
4249        }
4250        return -EINVAL;
4251}
4252#else
4253static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4254                                struct mem_cgroup *from, struct mem_cgroup *to)
4255{
4256        return -EINVAL;
4257}
4258#endif
4259
4260/*
4261 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4262 * page belongs to.
4263 */
4264void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4265                                  struct mem_cgroup **memcgp)
4266{
4267        struct mem_cgroup *memcg = NULL;
4268        unsigned int nr_pages = 1;
4269        struct page_cgroup *pc;
4270        enum charge_type ctype;
4271
4272        *memcgp = NULL;
4273
4274        if (mem_cgroup_disabled())
4275                return;
4276
4277        if (PageTransHuge(page))
4278                nr_pages <<= compound_order(page);
4279
4280        pc = lookup_page_cgroup(page);
4281        lock_page_cgroup(pc);
4282        if (PageCgroupUsed(pc)) {
4283                memcg = pc->mem_cgroup;
4284                css_get(&memcg->css);
4285                /*
4286                 * At migrating an anonymous page, its mapcount goes down
4287                 * to 0 and uncharge() will be called. But, even if it's fully
4288                 * unmapped, migration may fail and this page has to be
4289                 * charged again. We set MIGRATION flag here and delay uncharge
4290                 * until end_migration() is called
4291                 *
4292                 * Corner Case Thinking
4293                 * A)
4294                 * When the old page was mapped as Anon and it's unmap-and-freed
4295                 * while migration was ongoing.
4296                 * If unmap finds the old page, uncharge() of it will be delayed
4297                 * until end_migration(). If unmap finds a new page, it's
4298                 * uncharged when it make mapcount to be 1->0. If unmap code
4299                 * finds swap_migration_entry, the new page will not be mapped
4300                 * and end_migration() will find it(mapcount==0).
4301                 *
4302                 * B)
4303                 * When the old page was mapped but migraion fails, the kernel
4304                 * remaps it. A charge for it is kept by MIGRATION flag even
4305                 * if mapcount goes down to 0. We can do remap successfully
4306                 * without charging it again.
4307                 *
4308                 * C)
4309                 * The "old" page is under lock_page() until the end of
4310                 * migration, so, the old page itself will not be swapped-out.
4311                 * If the new page is swapped out before end_migraton, our
4312                 * hook to usual swap-out path will catch the event.
4313                 */
4314                if (PageAnon(page))
4315                        SetPageCgroupMigration(pc);
4316        }
4317        unlock_page_cgroup(pc);
4318        /*
4319         * If the page is not charged at this point,
4320         * we return here.
4321         */
4322        if (!memcg)
4323                return;
4324
4325        *memcgp = memcg;
4326        /*
4327         * We charge new page before it's used/mapped. So, even if unlock_page()
4328         * is called before end_migration, we can catch all events on this new
4329         * page. In the case new page is migrated but not remapped, new page's
4330         * mapcount will be finally 0 and we call uncharge in end_migration().
4331         */
4332        if (PageAnon(page))
4333                ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4334        else
4335                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4336        /*
4337         * The page is committed to the memcg, but it's not actually
4338         * charged to the res_counter since we plan on replacing the
4339         * old one and only one page is going to be left afterwards.
4340         */
4341        __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4342}
4343
4344/* remove redundant charge if migration failed*/
4345void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4346        struct page *oldpage, struct page *newpage, bool migration_ok)
4347{
4348        struct page *used, *unused;
4349        struct page_cgroup *pc;
4350        bool anon;
4351
4352        if (!memcg)
4353                return;
4354
4355        if (!migration_ok) {
4356                used = oldpage;
4357                unused = newpage;
4358        } else {
4359                used = newpage;
4360                unused = oldpage;
4361        }
4362        anon = PageAnon(used);
4363        __mem_cgroup_uncharge_common(unused,
4364                                     anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4365                                     : MEM_CGROUP_CHARGE_TYPE_CACHE,
4366                                     true);
4367        css_put(&memcg->css);
4368        /*
4369         * We disallowed uncharge of pages under migration because mapcount
4370         * of the page goes down to zero, temporarly.
4371         * Clear the flag and check the page should be charged.
4372         */
4373        pc = lookup_page_cgroup(oldpage);
4374        lock_page_cgroup(pc);
4375        ClearPageCgroupMigration(pc);
4376        unlock_page_cgroup(pc);
4377
4378        /*
4379         * If a page is a file cache, radix-tree replacement is very atomic
4380         * and we can skip this check. When it was an Anon page, its mapcount
4381         * goes down to 0. But because we added MIGRATION flage, it's not
4382         * uncharged yet. There are several case but page->mapcount check
4383         * and USED bit check in mem_cgroup_uncharge_page() will do enough
4384         * check. (see prepare_charge() also)
4385         */
4386        if (anon)
4387                mem_cgroup_uncharge_page(used);
4388}
4389
4390/*
4391 * At replace page cache, newpage is not under any memcg but it's on
4392 * LRU. So, this function doesn't touch res_counter but handles LRU
4393 * in correct way. Both pages are locked so we cannot race with uncharge.
4394 */
4395void mem_cgroup_replace_page_cache(struct page *oldpage,
4396                                  struct page *newpage)
4397{
4398        struct mem_cgroup *memcg = NULL;
4399        struct page_cgroup *pc;
4400        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4401
4402        if (mem_cgroup_disabled())
4403                return;
4404
4405        pc = lookup_page_cgroup(oldpage);
4406        /* fix accounting on old pages */
4407        lock_page_cgroup(pc);
4408        if (PageCgroupUsed(pc)) {
4409                memcg = pc->mem_cgroup;
4410                mem_cgroup_charge_statistics(memcg, false, -1);
4411                ClearPageCgroupUsed(pc);
4412        }
4413        unlock_page_cgroup(pc);
4414
4415        /*
4416         * When called from shmem_replace_page(), in some cases the
4417         * oldpage has already been charged, and in some cases not.
4418         */
4419        if (!memcg)
4420                return;
4421        /*
4422         * Even if newpage->mapping was NULL before starting replacement,
4423         * the newpage may be on LRU(or pagevec for LRU) already. We lock
4424         * LRU while we overwrite pc->mem_cgroup.
4425         */
4426        __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4427}
4428
4429#ifdef CONFIG_DEBUG_VM
4430static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4431{
4432        struct page_cgroup *pc;
4433
4434        pc = lookup_page_cgroup(page);
4435        /*
4436         * Can be NULL while feeding pages into the page allocator for
4437         * the first time, i.e. during boot or memory hotplug;
4438         * or when mem_cgroup_disabled().
4439         */
4440        if (likely(pc) && PageCgroupUsed(pc))
4441                return pc;
4442        return NULL;
4443}
4444
4445bool mem_cgroup_bad_page_check(struct page *page)
4446{
4447        if (mem_cgroup_disabled())
4448                return false;
4449
4450        return lookup_page_cgroup_used(page) != NULL;
4451}
4452
4453void mem_cgroup_print_bad_page(struct page *page)
4454{
4455        struct page_cgroup *pc;
4456
4457        pc = lookup_page_cgroup_used(page);
4458        if (pc) {
4459                pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4460                         pc, pc->flags, pc->mem_cgroup);
4461        }
4462}
4463#endif
4464
4465static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4466                                unsigned long long val)
4467{
4468        int retry_count;
4469        u64 memswlimit, memlimit;
4470        int ret = 0;
4471        int children = mem_cgroup_count_children(memcg);
4472        u64 curusage, oldusage;
4473        int enlarge;
4474
4475        /*
4476         * For keeping hierarchical_reclaim simple, how long we should retry
4477         * is depends on callers. We set our retry-count to be function
4478         * of # of children which we should visit in this loop.
4479         */
4480        retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4481
4482        oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4483
4484        enlarge = 0;
4485        while (retry_count) {
4486                if (signal_pending(current)) {
4487                        ret = -EINTR;
4488                        break;
4489                }
4490                /*
4491                 * Rather than hide all in some function, I do this in
4492                 * open coded manner. You see what this really does.
4493                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4494                 */
4495                mutex_lock(&set_limit_mutex);
4496                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4497                if (memswlimit < val) {
4498                        ret = -EINVAL;
4499                        mutex_unlock(&set_limit_mutex);
4500                        break;
4501                }
4502
4503                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4504                if (memlimit < val)
4505                        enlarge = 1;
4506
4507                ret = res_counter_set_limit(&memcg->res, val);
4508                if (!ret) {
4509                        if (memswlimit == val)
4510                                memcg->memsw_is_minimum = true;
4511                        else
4512                                memcg->memsw_is_minimum = false;
4513                }
4514                mutex_unlock(&set_limit_mutex);
4515
4516                if (!ret)
4517                        break;
4518
4519                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4520                                   MEM_CGROUP_RECLAIM_SHRINK);
4521                curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4522                /* Usage is reduced ? */
4523                if (curusage >= oldusage)
4524                        retry_count--;
4525                else
4526                        oldusage = curusage;
4527        }
4528        if (!ret && enlarge)
4529                memcg_oom_recover(memcg);
4530
4531        return ret;
4532}
4533
4534static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4535                                        unsigned long long val)
4536{
4537        int retry_count;
4538        u64 memlimit, memswlimit, oldusage, curusage;
4539        int children = mem_cgroup_count_children(memcg);
4540        int ret = -EBUSY;
4541        int enlarge = 0;
4542
4543        /* see mem_cgroup_resize_res_limit */
4544        retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4545        oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4546        while (retry_count) {
4547                if (signal_pending(current)) {
4548                        ret = -EINTR;
4549                        break;
4550                }
4551                /*
4552                 * Rather than hide all in some function, I do this in
4553                 * open coded manner. You see what this really does.
4554                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4555                 */
4556                mutex_lock(&set_limit_mutex);
4557                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4558                if (memlimit > val) {
4559                        ret = -EINVAL;
4560                        mutex_unlock(&set_limit_mutex);
4561                        break;
4562                }
4563                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4564                if (memswlimit < val)
4565                        enlarge = 1;
4566                ret = res_counter_set_limit(&memcg->memsw, val);
4567                if (!ret) {
4568                        if (memlimit == val)
4569                                memcg->memsw_is_minimum = true;
4570                        else
4571                                memcg->memsw_is_minimum = false;
4572                }
4573                mutex_unlock(&set_limit_mutex);
4574
4575                if (!ret)
4576                        break;
4577
4578                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4579                                   MEM_CGROUP_RECLAIM_NOSWAP |
4580                                   MEM_CGROUP_RECLAIM_SHRINK);
4581                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4582                /* Usage is reduced ? */
4583                if (curusage >= oldusage)
4584                        retry_count--;
4585                else
4586                        oldusage = curusage;
4587        }
4588        if (!ret && enlarge)
4589                memcg_oom_recover(memcg);
4590        return ret;
4591}
4592
4593unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4594                                            gfp_t gfp_mask,
4595                                            unsigned long *total_scanned)
4596{
4597        unsigned long nr_reclaimed = 0;
4598        struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4599        unsigned long reclaimed;
4600        int loop = 0;
4601        struct mem_cgroup_tree_per_zone *mctz;
4602        unsigned long long excess;
4603        unsigned long nr_scanned;
4604
4605        if (order > 0)
4606                return 0;
4607
4608        mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4609        /*
4610         * This loop can run a while, specially if mem_cgroup's continuously
4611         * keep exceeding their soft limit and putting the system under
4612         * pressure
4613         */
4614        do {
4615                if (next_mz)
4616                        mz = next_mz;
4617                else
4618                        mz = mem_cgroup_largest_soft_limit_node(mctz);
4619                if (!mz)
4620                        break;
4621
4622                nr_scanned = 0;
4623                reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4624                                                    gfp_mask, &nr_scanned);
4625                nr_reclaimed += reclaimed;
4626                *total_scanned += nr_scanned;
4627                spin_lock(&mctz->lock);
4628
4629                /*
4630                 * If we failed to reclaim anything from this memory cgroup
4631                 * it is time to move on to the next cgroup
4632                 */
4633                next_mz = NULL;
4634                if (!reclaimed) {
4635                        do {
4636                                /*
4637                                 * Loop until we find yet another one.
4638                                 *
4639                                 * By the time we get the soft_limit lock
4640                                 * again, someone might have aded the
4641                                 * group back on the RB tree. Iterate to
4642                                 * make sure we get a different mem.
4643                                 * mem_cgroup_largest_soft_limit_node returns
4644                                 * NULL if no other cgroup is present on
4645                                 * the tree
4646                                 */
4647                                next_mz =
4648                                __mem_cgroup_largest_soft_limit_node(mctz);
4649                                if (next_mz == mz)
4650                                        css_put(&next_mz->memcg->css);
4651                                else /* next_mz == NULL or other memcg */
4652                                        break;
4653                        } while (1);
4654                }
4655                __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4656                excess = res_counter_soft_limit_excess(&mz->memcg->res);
4657                /*
4658                 * One school of thought says that we should not add
4659                 * back the node to the tree if reclaim returns 0.
4660                 * But our reclaim could return 0, simply because due
4661                 * to priority we are exposing a smaller subset of
4662                 * memory to reclaim from. Consider this as a longer
4663                 * term TODO.
4664                 */
4665                /* If excess == 0, no tree ops */
4666                __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4667                spin_unlock(&mctz->lock);
4668                css_put(&mz->memcg->css);
4669                loop++;
4670                /*
4671                 * Could not reclaim anything and there are no more
4672                 * mem cgroups to try or we seem to be looping without
4673                 * reclaiming anything.
4674                 */
4675                if (!nr_reclaimed &&
4676                        (next_mz == NULL ||
4677                        loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4678                        break;
4679        } while (!nr_reclaimed);
4680        if (next_mz)
4681                css_put(&next_mz->memcg->css);
4682        return nr_reclaimed;
4683}
4684
4685/**
4686 * mem_cgroup_force_empty_list - clears LRU of a group
4687 * @memcg: group to clear
4688 * @node: NUMA node
4689 * @zid: zone id
4690 * @lru: lru to to clear
4691 *
4692 * Traverse a specified page_cgroup list and try to drop them all.  This doesn't
4693 * reclaim the pages page themselves - pages are moved to the parent (or root)
4694 * group.
4695 */
4696static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4697                                int node, int zid, enum lru_list lru)
4698{
4699        struct lruvec *lruvec;
4700        unsigned long flags;
4701        struct list_head *list;
4702        struct page *busy;
4703        struct zone *zone;
4704
4705        zone = &NODE_DATA(node)->node_zones[zid];
4706        lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4707        list = &lruvec->lists[lru];
4708
4709        busy = NULL;
4710        do {
4711                struct page_cgroup *pc;
4712                struct page *page;
4713
4714                spin_lock_irqsave(&zone->lru_lock, flags);
4715                if (list_empty(list)) {
4716                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4717                        break;
4718                }
4719                page = list_entry(list->prev, struct page, lru);
4720                if (busy == page) {
4721                        list_move(&page->lru, list);
4722                        busy = NULL;
4723                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4724                        continue;
4725                }
4726                spin_unlock_irqrestore(&zone->lru_lock, flags);
4727
4728                pc = lookup_page_cgroup(page);
4729
4730                if (mem_cgroup_move_parent(page, pc, memcg)) {
4731                        /* found lock contention or "pc" is obsolete. */
4732                        busy = page;
4733                        cond_resched();
4734                } else
4735                        busy = NULL;
4736        } while (!list_empty(list));
4737}
4738
4739/*
4740 * make mem_cgroup's charge to be 0 if there is no task by moving
4741 * all the charges and pages to the parent.
4742 * This enables deleting this mem_cgroup.
4743 *
4744 * Caller is responsible for holding css reference on the memcg.
4745 */
4746static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4747{
4748        int node, zid;
4749        u64 usage;
4750
4751        do {
4752                /* This is for making all *used* pages to be on LRU. */
4753                lru_add_drain_all();
4754                drain_all_stock_sync(memcg);
4755                mem_cgroup_start_move(memcg);
4756                for_each_node_state(node, N_MEMORY) {
4757                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4758                                enum lru_list lru;
4759                                for_each_lru(lru) {
4760                                        mem_cgroup_force_empty_list(memcg,
4761                                                        node, zid, lru);
4762                                }
4763                        }
4764                }
4765                mem_cgroup_end_move(memcg);
4766                memcg_oom_recover(memcg);
4767                cond_resched();
4768
4769                /*
4770                 * Kernel memory may not necessarily be trackable to a specific
4771                 * process. So they are not migrated, and therefore we can't
4772                 * expect their value to drop to 0 here.
4773                 * Having res filled up with kmem only is enough.
4774                 *
4775                 * This is a safety check because mem_cgroup_force_empty_list
4776                 * could have raced with mem_cgroup_replace_page_cache callers
4777                 * so the lru seemed empty but the page could have been added
4778                 * right after the check. RES_USAGE should be safe as we always
4779                 * charge before adding to the LRU.
4780                 */
4781                usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4782                        res_counter_read_u64(&memcg->kmem, RES_USAGE);
4783        } while (usage > 0);
4784}
4785
4786/*
4787 * This mainly exists for tests during the setting of set of use_hierarchy.
4788 * Since this is the very setting we are changing, the current hierarchy value
4789 * is meaningless
4790 */
4791static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4792{
4793        struct cgroup *pos;
4794
4795        /* bounce at first found */
4796        cgroup_for_each_child(pos, memcg->css.cgroup)
4797                return true;
4798        return false;
4799}
4800
4801/*
4802 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4803 * to be already dead (as in mem_cgroup_force_empty, for instance).  This is
4804 * from mem_cgroup_count_children(), in the sense that we don't really care how
4805 * many children we have; we only need to know if we have any.  It also counts
4806 * any memcg without hierarchy as infertile.
4807 */
4808static inline bool memcg_has_children(struct mem_cgroup *memcg)
4809{
4810        return memcg->use_hierarchy && __memcg_has_children(memcg);
4811}
4812
4813/*
4814 * Reclaims as many pages from the given memcg as possible and moves
4815 * the rest to the parent.
4816 *
4817 * Caller is responsible for holding css reference for memcg.
4818 */
4819static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4820{
4821        int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4822        struct cgroup *cgrp = memcg->css.cgroup;
4823
4824        /* returns EBUSY if there is a task or if we come here twice. */
4825        if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4826                return -EBUSY;
4827
4828        /* we call try-to-free pages for make this cgroup empty */
4829        lru_add_drain_all();
4830        /* try to free all pages in this cgroup */
4831        while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4832                int progress;
4833
4834                if (signal_pending(current))
4835                        return -EINTR;
4836
4837                progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4838                                                false);
4839                if (!progress) {
4840                        nr_retries--;
4841                        /* maybe some writeback is necessary */
4842                        congestion_wait(BLK_RW_ASYNC, HZ/10);
4843                }
4844
4845        }
4846        lru_add_drain();
4847        mem_cgroup_reparent_charges(memcg);
4848
4849        return 0;
4850}
4851
4852static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4853{
4854        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4855        int ret;
4856
4857        if (mem_cgroup_is_root(memcg))
4858                return -EINVAL;
4859        css_get(&memcg->css);
4860        ret = mem_cgroup_force_empty(memcg);
4861        css_put(&memcg->css);
4862
4863        return ret;
4864}
4865
4866
4867static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4868{
4869        return mem_cgroup_from_cont(cont)->use_hierarchy;
4870}
4871
4872static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4873                                        u64 val)
4874{
4875        int retval = 0;
4876        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4877        struct cgroup *parent = cont->parent;
4878        struct mem_cgroup *parent_memcg = NULL;
4879
4880        if (parent)
4881                parent_memcg = mem_cgroup_from_cont(parent);
4882
4883        mutex_lock(&memcg_create_mutex);
4884
4885        if (memcg->use_hierarchy == val)
4886                goto out;
4887
4888        /*
4889         * If parent's use_hierarchy is set, we can't make any modifications
4890         * in the child subtrees. If it is unset, then the change can
4891         * occur, provided the current cgroup has no children.
4892         *
4893         * For the root cgroup, parent_mem is NULL, we allow value to be
4894         * set if there are no children.
4895         */
4896        if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4897                                (val == 1 || val == 0)) {
4898                if (!__memcg_has_children(memcg))
4899                        memcg->use_hierarchy = val;
4900                else
4901                        retval = -EBUSY;
4902        } else
4903                retval = -EINVAL;
4904
4905out:
4906        mutex_unlock(&memcg_create_mutex);
4907
4908        return retval;
4909}
4910
4911
4912static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4913                                               enum mem_cgroup_stat_index idx)
4914{
4915        struct mem_cgroup *iter;
4916        long val = 0;
4917
4918        /* Per-cpu values can be negative, use a signed accumulator */
4919        for_each_mem_cgroup_tree(iter, memcg)
4920                val += mem_cgroup_read_stat(iter, idx);
4921
4922        if (val < 0) /* race ? */
4923                val = 0;
4924        return val;
4925}
4926
4927static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4928{
4929        u64 val;
4930
4931        if (!mem_cgroup_is_root(memcg)) {
4932                if (!swap)
4933                        return res_counter_read_u64(&memcg->res, RES_USAGE);
4934                else
4935                        return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4936        }
4937
4938        val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4939        val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4940
4941        if (swap)
4942                val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4943
4944        return val << PAGE_SHIFT;
4945}
4946
4947static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4948                               struct file *file, char __user *buf,
4949                               size_t nbytes, loff_t *ppos)
4950{
4951        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4952        char str[64];
4953        u64 val;
4954        int name, len;
4955        enum res_type type;
4956
4957        type = MEMFILE_TYPE(cft->private);
4958        name = MEMFILE_ATTR(cft->private);
4959
4960        if (!do_swap_account && type == _MEMSWAP)
4961                return -EOPNOTSUPP;
4962
4963        switch (type) {
4964        case _MEM:
4965                if (name == RES_USAGE)
4966                        val = mem_cgroup_usage(memcg, false);
4967                else
4968                        val = res_counter_read_u64(&memcg->res, name);
4969                break;
4970        case _MEMSWAP:
4971                if (name == RES_USAGE)
4972                        val = mem_cgroup_usage(memcg, true);
4973                else
4974                        val = res_counter_read_u64(&memcg->memsw, name);
4975                break;
4976        case _KMEM:
4977                val = res_counter_read_u64(&memcg->kmem, name);
4978                break;
4979        default:
4980                BUG();
4981        }
4982
4983        len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4984        return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4985}
4986
4987static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4988{
4989        int ret = -EINVAL;
4990#ifdef CONFIG_MEMCG_KMEM
4991        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4992        /*
4993         * For simplicity, we won't allow this to be disabled.  It also can't
4994         * be changed if the cgroup has children already, or if tasks had
4995         * already joined.
4996         *
4997         * If tasks join before we set the limit, a person looking at
4998         * kmem.usage_in_bytes will have no way to determine when it took
4999         * place, which makes the value quite meaningless.
5000         *
5001         * After it first became limited, changes in the value of the limit are
5002         * of course permitted.
5003         */
5004        mutex_lock(&memcg_create_mutex);
5005        mutex_lock(&set_limit_mutex);
5006        if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5007                if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5008                        ret = -EBUSY;
5009                        goto out;
5010                }
5011                ret = res_counter_set_limit(&memcg->kmem, val);
5012                VM_BUG_ON(ret);
5013
5014                ret = memcg_update_cache_sizes(memcg);
5015                if (ret) {
5016                        res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5017                        goto out;
5018                }
5019                static_key_slow_inc(&memcg_kmem_enabled_key);
5020                /*
5021                 * setting the active bit after the inc will guarantee no one
5022                 * starts accounting before all call sites are patched
5023                 */
5024                memcg_kmem_set_active(memcg);
5025
5026                /*
5027                 * kmem charges can outlive the cgroup. In the case of slab
5028                 * pages, for instance, a page contain objects from various
5029                 * processes, so it is unfeasible to migrate them away. We
5030                 * need to reference count the memcg because of that.
5031                 */
5032                mem_cgroup_get(memcg);
5033        } else
5034                ret = res_counter_set_limit(&memcg->kmem, val);
5035out:
5036        mutex_unlock(&set_limit_mutex);
5037        mutex_unlock(&memcg_create_mutex);
5038#endif
5039        return ret;
5040}
5041
5042#ifdef CONFIG_MEMCG_KMEM
5043static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5044{
5045        int ret = 0;
5046        struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5047        if (!parent)
5048                goto out;
5049
5050        memcg->kmem_account_flags = parent->kmem_account_flags;
5051        /*
5052         * When that happen, we need to disable the static branch only on those
5053         * memcgs that enabled it. To achieve this, we would be forced to
5054         * complicate the code by keeping track of which memcgs were the ones
5055         * that actually enabled limits, and which ones got it from its
5056         * parents.
5057         *
5058         * It is a lot simpler just to do static_key_slow_inc() on every child
5059         * that is accounted.
5060         */
5061        if (!memcg_kmem_is_active(memcg))
5062                goto out;
5063
5064        /*
5065         * destroy(), called if we fail, will issue static_key_slow_inc() and
5066         * mem_cgroup_put() if kmem is enabled. We have to either call them
5067         * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5068         * this more consistent, since it always leads to the same destroy path
5069         */
5070        mem_cgroup_get(memcg);
5071        static_key_slow_inc(&memcg_kmem_enabled_key);
5072
5073        mutex_lock(&set_limit_mutex);
5074        ret = memcg_update_cache_sizes(memcg);
5075        mutex_unlock(&set_limit_mutex);
5076out:
5077        return ret;
5078}
5079#endif /* CONFIG_MEMCG_KMEM */
5080
5081/*
5082 * The user of this function is...
5083 * RES_LIMIT.
5084 */
5085static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5086                            const char *buffer)
5087{
5088        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5089        enum res_type type;
5090        int name;
5091        unsigned long long val;
5092        int ret;
5093
5094        type = MEMFILE_TYPE(cft->private);
5095        name = MEMFILE_ATTR(cft->private);
5096
5097        if (!do_swap_account && type == _MEMSWAP)
5098                return -EOPNOTSUPP;
5099
5100        switch (name) {
5101        case RES_LIMIT:
5102                if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5103                        ret = -EINVAL;
5104                        break;
5105                }
5106                /* This function does all necessary parse...reuse it */
5107                ret = res_counter_memparse_write_strategy(buffer, &val);
5108                if (ret)
5109                        break;
5110                if (type == _MEM)
5111                        ret = mem_cgroup_resize_limit(memcg, val);
5112                else if (type == _MEMSWAP)
5113                        ret = mem_cgroup_resize_memsw_limit(memcg, val);
5114                else if (type == _KMEM)
5115                        ret = memcg_update_kmem_limit(cont, val);
5116                else
5117                        return -EINVAL;
5118                break;
5119        case RES_SOFT_LIMIT:
5120                ret = res_counter_memparse_write_strategy(buffer, &val);
5121                if (ret)
5122                        break;
5123                /*
5124                 * For memsw, soft limits are hard to implement in terms
5125                 * of semantics, for now, we support soft limits for
5126                 * control without swap
5127                 */
5128                if (type == _MEM)
5129                        ret = res_counter_set_soft_limit(&memcg->res, val);
5130                else
5131                        ret = -EINVAL;
5132                break;
5133        default:
5134                ret = -EINVAL; /* should be BUG() ? */
5135                break;
5136        }
5137        return ret;
5138}
5139
5140static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5141                unsigned long long *mem_limit, unsigned long long *memsw_limit)
5142{
5143        struct cgroup *cgroup;
5144        unsigned long long min_limit, min_memsw_limit, tmp;
5145
5146        min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5147        min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5148        cgroup = memcg->css.cgroup;
5149        if (!memcg->use_hierarchy)
5150                goto out;
5151
5152        while (cgroup->parent) {
5153                cgroup = cgroup->parent;
5154                memcg = mem_cgroup_from_cont(cgroup);
5155                if (!memcg->use_hierarchy)
5156                        break;
5157                tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5158                min_limit = min(min_limit, tmp);
5159                tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5160                min_memsw_limit = min(min_memsw_limit, tmp);
5161        }
5162out:
5163        *mem_limit = min_limit;
5164        *memsw_limit = min_memsw_limit;
5165}
5166
5167static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5168{
5169        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5170        int name;
5171        enum res_type type;
5172
5173        type = MEMFILE_TYPE(event);
5174        name = MEMFILE_ATTR(event);
5175
5176        if (!do_swap_account && type == _MEMSWAP)
5177                return -EOPNOTSUPP;
5178
5179        switch (name) {
5180        case RES_MAX_USAGE:
5181                if (type == _MEM)
5182                        res_counter_reset_max(&memcg->res);
5183                else if (type == _MEMSWAP)
5184                        res_counter_reset_max(&memcg->memsw);
5185                else if (type == _KMEM)
5186                        res_counter_reset_max(&memcg->kmem);
5187                else
5188                        return -EINVAL;
5189                break;
5190        case RES_FAILCNT:
5191                if (type == _MEM)
5192                        res_counter_reset_failcnt(&memcg->res);
5193                else if (type == _MEMSWAP)
5194                        res_counter_reset_failcnt(&memcg->memsw);
5195                else if (type == _KMEM)
5196                        res_counter_reset_failcnt(&memcg->kmem);
5197                else
5198                        return -EINVAL;
5199                break;
5200        }
5201
5202        return 0;
5203}
5204
5205static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5206                                        struct cftype *cft)
5207{
5208        return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5209}
5210
5211#ifdef CONFIG_MMU
5212static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5213                                        struct cftype *cft, u64 val)
5214{
5215        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5216
5217        if (val >= (1 << NR_MOVE_TYPE))
5218                return -EINVAL;
5219
5220        /*
5221         * No kind of locking is needed in here, because ->can_attach() will
5222         * check this value once in the beginning of the process, and then carry
5223         * on with stale data. This means that changes to this value will only
5224         * affect task migrations starting after the change.
5225         */
5226        memcg->move_charge_at_immigrate = val;
5227        return 0;
5228}
5229#else
5230static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5231                                        struct cftype *cft, u64 val)
5232{
5233        return -ENOSYS;
5234}
5235#endif
5236
5237#ifdef CONFIG_NUMA
5238static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5239                                      struct seq_file *m)
5240{
5241        int nid;
5242        unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5243        unsigned long node_nr;
5244        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5245
5246        total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5247        seq_printf(m, "total=%lu", total_nr);
5248        for_each_node_state(nid, N_MEMORY) {
5249                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5250                seq_printf(m, " N%d=%lu", nid, node_nr);
5251        }
5252        seq_putc(m, '\n');
5253
5254        file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5255        seq_printf(m, "file=%lu", file_nr);
5256        for_each_node_state(nid, N_MEMORY) {
5257                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5258                                LRU_ALL_FILE);
5259                seq_printf(m, " N%d=%lu", nid, node_nr);
5260        }
5261        seq_putc(m, '\n');
5262
5263        anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5264        seq_printf(m, "anon=%lu", anon_nr);
5265        for_each_node_state(nid, N_MEMORY) {
5266                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5267                                LRU_ALL_ANON);
5268                seq_printf(m, " N%d=%lu", nid, node_nr);
5269        }
5270        seq_putc(m, '\n');
5271
5272        unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5273        seq_printf(m, "unevictable=%lu", unevictable_nr);
5274        for_each_node_state(nid, N_MEMORY) {
5275                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5276                                BIT(LRU_UNEVICTABLE));
5277                seq_printf(m, " N%d=%lu", nid, node_nr);
5278        }
5279        seq_putc(m, '\n');
5280        return 0;
5281}
5282#endif /* CONFIG_NUMA */
5283
5284static inline void mem_cgroup_lru_names_not_uptodate(void)
5285{
5286        BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5287}
5288
5289static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5290                                 struct seq_file *m)
5291{
5292        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5293        struct mem_cgroup *mi;
5294        unsigned int i;
5295
5296        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5297                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5298                        continue;
5299                seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5300                           mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5301        }
5302
5303        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5304                seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5305                           mem_cgroup_read_events(memcg, i));
5306
5307        for (i = 0; i < NR_LRU_LISTS; i++)
5308                seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5309                           mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5310
5311        /* Hierarchical information */
5312        {
5313                unsigned long long limit, memsw_limit;
5314                memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5315                seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5316                if (do_swap_account)
5317                        seq_printf(m, "hierarchical_memsw_limit %llu\n",
5318                                   memsw_limit);
5319        }
5320
5321        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5322                long long val = 0;
5323
5324                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5325                        continue;
5326                for_each_mem_cgroup_tree(mi, memcg)
5327                        val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5328                seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5329        }
5330
5331        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5332                unsigned long long val = 0;
5333
5334                for_each_mem_cgroup_tree(mi, memcg)
5335                        val += mem_cgroup_read_events(mi, i);
5336                seq_printf(m, "total_%s %llu\n",
5337                           mem_cgroup_events_names[i], val);
5338        }
5339
5340        for (i = 0; i < NR_LRU_LISTS; i++) {
5341                unsigned long long val = 0;
5342
5343                for_each_mem_cgroup_tree(mi, memcg)
5344                        val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5345                seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5346        }
5347
5348#ifdef CONFIG_DEBUG_VM
5349        {
5350                int nid, zid;
5351                struct mem_cgroup_per_zone *mz;
5352                struct zone_reclaim_stat *rstat;
5353                unsigned long recent_rotated[2] = {0, 0};
5354                unsigned long recent_scanned[2] = {0, 0};
5355
5356                for_each_online_node(nid)
5357                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5358                                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5359                                rstat = &mz->lruvec.reclaim_stat;
5360
5361                                recent_rotated[0] += rstat->recent_rotated[0];
5362                                recent_rotated[1] += rstat->recent_rotated[1];
5363                                recent_scanned[0] += rstat->recent_scanned[0];
5364                                recent_scanned[1] += rstat->recent_scanned[1];
5365                        }
5366                seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5367                seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5368                seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5369                seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5370        }
5371#endif
5372
5373        return 0;
5374}
5375
5376static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5377{
5378        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5379
5380        return mem_cgroup_swappiness(memcg);
5381}
5382
5383static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5384                                       u64 val)
5385{
5386        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5387        struct mem_cgroup *parent;
5388
5389        if (val > 100)
5390                return -EINVAL;
5391
5392        if (cgrp->parent == NULL)
5393                return -EINVAL;
5394
5395        parent = mem_cgroup_from_cont(cgrp->parent);
5396
5397        mutex_lock(&memcg_create_mutex);
5398
5399        /* If under hierarchy, only empty-root can set this value */
5400        if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5401                mutex_unlock(&memcg_create_mutex);
5402                return -EINVAL;
5403        }
5404
5405        memcg->swappiness = val;
5406
5407        mutex_unlock(&memcg_create_mutex);
5408
5409        return 0;
5410}
5411
5412static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5413{
5414        struct mem_cgroup_threshold_ary *t;
5415        u64 usage;
5416        int i;
5417
5418        rcu_read_lock();
5419        if (!swap)
5420                t = rcu_dereference(memcg->thresholds.primary);
5421        else
5422                t = rcu_dereference(memcg->memsw_thresholds.primary);
5423
5424        if (!t)
5425                goto unlock;
5426
5427        usage = mem_cgroup_usage(memcg, swap);
5428
5429        /*
5430         * current_threshold points to threshold just below or equal to usage.
5431         * If it's not true, a threshold was crossed after last
5432         * call of __mem_cgroup_threshold().
5433         */
5434        i = t->current_threshold;
5435
5436        /*
5437         * Iterate backward over array of thresholds starting from
5438         * current_threshold and check if a threshold is crossed.
5439         * If none of thresholds below usage is crossed, we read
5440         * only one element of the array here.
5441         */
5442        for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5443                eventfd_signal(t->entries[i].eventfd, 1);
5444
5445        /* i = current_threshold + 1 */
5446        i++;
5447
5448        /*
5449         * Iterate forward over array of thresholds starting from
5450         * current_threshold+1 and check if a threshold is crossed.
5451         * If none of thresholds above usage is crossed, we read
5452         * only one element of the array here.
5453         */
5454        for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5455                eventfd_signal(t->entries[i].eventfd, 1);
5456
5457        /* Update current_threshold */
5458        t->current_threshold = i - 1;
5459unlock:
5460        rcu_read_unlock();
5461}
5462
5463static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5464{
5465        while (memcg) {
5466                __mem_cgroup_threshold(memcg, false);
5467                if (do_swap_account)
5468                        __mem_cgroup_threshold(memcg, true);
5469
5470                memcg = parent_mem_cgroup(memcg);
5471        }
5472}
5473
5474static int compare_thresholds(const void *a, const void *b)
5475{
5476        const struct mem_cgroup_threshold *_a = a;
5477        const struct mem_cgroup_threshold *_b = b;
5478
5479        return _a->threshold - _b->threshold;
5480}
5481
5482static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5483{
5484        struct mem_cgroup_eventfd_list *ev;
5485
5486        list_for_each_entry(ev, &memcg->oom_notify, list)
5487                eventfd_signal(ev->eventfd, 1);
5488        return 0;
5489}
5490
5491static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5492{
5493        struct mem_cgroup *iter;
5494
5495        for_each_mem_cgroup_tree(iter, memcg)
5496                mem_cgroup_oom_notify_cb(iter);
5497}
5498
5499static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5500        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5501{
5502        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5503        struct mem_cgroup_thresholds *thresholds;
5504        struct mem_cgroup_threshold_ary *new;
5505        enum res_type type = MEMFILE_TYPE(cft->private);
5506        u64 threshold, usage;
5507        int i, size, ret;
5508
5509        ret = res_counter_memparse_write_strategy(args, &threshold);
5510        if (ret)
5511                return ret;
5512
5513        mutex_lock(&memcg->thresholds_lock);
5514
5515        if (type == _MEM)
5516                thresholds = &memcg->thresholds;
5517        else if (type == _MEMSWAP)
5518                thresholds = &memcg->memsw_thresholds;
5519        else
5520                BUG();
5521
5522        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5523
5524        /* Check if a threshold crossed before adding a new one */
5525        if (thresholds->primary)
5526                __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5527
5528        size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5529
5530        /* Allocate memory for new array of thresholds */
5531        new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5532                        GFP_KERNEL);
5533        if (!new) {
5534                ret = -ENOMEM;
5535                goto unlock;
5536        }
5537        new->size = size;
5538
5539        /* Copy thresholds (if any) to new array */
5540        if (thresholds->primary) {
5541                memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5542                                sizeof(struct mem_cgroup_threshold));
5543        }
5544
5545        /* Add new threshold */
5546        new->entries[size - 1].eventfd = eventfd;
5547        new->entries[size - 1].threshold = threshold;
5548
5549        /* Sort thresholds. Registering of new threshold isn't time-critical */
5550        sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5551                        compare_thresholds, NULL);
5552
5553        /* Find current threshold */
5554        new->current_threshold = -1;
5555        for (i = 0; i < size; i++) {
5556                if (new->entries[i].threshold <= usage) {
5557                        /*
5558                         * new->current_threshold will not be used until
5559                         * rcu_assign_pointer(), so it's safe to increment
5560                         * it here.
5561                         */
5562                        ++new->current_threshold;
5563                } else
5564                        break;
5565        }
5566
5567        /* Free old spare buffer and save old primary buffer as spare */
5568        kfree(thresholds->spare);
5569        thresholds->spare = thresholds->primary;
5570
5571        rcu_assign_pointer(thresholds->primary, new);
5572
5573        /* To be sure that nobody uses thresholds */
5574        synchronize_rcu();
5575
5576unlock:
5577        mutex_unlock(&memcg->thresholds_lock);
5578
5579        return ret;
5580}
5581
5582static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5583        struct cftype *cft, struct eventfd_ctx *eventfd)
5584{
5585        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5586        struct mem_cgroup_thresholds *thresholds;
5587        struct mem_cgroup_threshold_ary *new;
5588        enum res_type type = MEMFILE_TYPE(cft->private);
5589        u64 usage;
5590        int i, j, size;
5591
5592        mutex_lock(&memcg->thresholds_lock);
5593        if (type == _MEM)
5594                thresholds = &memcg->thresholds;
5595        else if (type == _MEMSWAP)
5596                thresholds = &memcg->memsw_thresholds;
5597        else
5598                BUG();
5599
5600        if (!thresholds->primary)
5601                goto unlock;
5602
5603        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5604
5605        /* Check if a threshold crossed before removing */
5606        __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5607
5608        /* Calculate new number of threshold */
5609        size = 0;
5610        for (i = 0; i < thresholds->primary->size; i++) {
5611                if (thresholds->primary->entries[i].eventfd != eventfd)
5612                        size++;
5613        }
5614
5615        new = thresholds->spare;
5616
5617        /* Set thresholds array to NULL if we don't have thresholds */
5618        if (!size) {
5619                kfree(new);
5620                new = NULL;
5621                goto swap_buffers;
5622        }
5623
5624        new->size = size;
5625
5626        /* Copy thresholds and find current threshold */
5627        new->current_threshold = -1;
5628        for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5629                if (thresholds->primary->entries[i].eventfd == eventfd)
5630                        continue;
5631
5632                new->entries[j] = thresholds->primary->entries[i];
5633                if (new->entries[j].threshold <= usage) {
5634                        /*
5635                         * new->current_threshold will not be used
5636                         * until rcu_assign_pointer(), so it's safe to increment
5637                         * it here.
5638                         */
5639                        ++new->current_threshold;
5640                }
5641                j++;
5642        }
5643
5644swap_buffers:
5645        /* Swap primary and spare array */
5646        thresholds->spare = thresholds->primary;
5647        /* If all events are unregistered, free the spare array */
5648        if (!new) {
5649                kfree(thresholds->spare);
5650                thresholds->spare = NULL;
5651        }
5652
5653        rcu_assign_pointer(thresholds->primary, new);
5654
5655        /* To be sure that nobody uses thresholds */
5656        synchronize_rcu();
5657unlock:
5658        mutex_unlock(&memcg->thresholds_lock);
5659}
5660
5661static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5662        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5663{
5664        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5665        struct mem_cgroup_eventfd_list *event;
5666        enum res_type type = MEMFILE_TYPE(cft->private);
5667
5668        BUG_ON(type != _OOM_TYPE);
5669        event = kmalloc(sizeof(*event), GFP_KERNEL);
5670        if (!event)
5671                return -ENOMEM;
5672
5673        spin_lock(&memcg_oom_lock);
5674
5675        event->eventfd = eventfd;
5676        list_add(&event->list, &memcg->oom_notify);
5677
5678        /* already in OOM ? */
5679        if (atomic_read(&memcg->under_oom))
5680                eventfd_signal(eventfd, 1);
5681        spin_unlock(&memcg_oom_lock);
5682
5683        return 0;
5684}
5685
5686static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5687        struct cftype *cft, struct eventfd_ctx *eventfd)
5688{
5689        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5690        struct mem_cgroup_eventfd_list *ev, *tmp;
5691        enum res_type type = MEMFILE_TYPE(cft->private);
5692
5693        BUG_ON(type != _OOM_TYPE);
5694
5695        spin_lock(&memcg_oom_lock);
5696
5697        list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5698                if (ev->eventfd == eventfd) {
5699                        list_del(&ev->list);
5700                        kfree(ev);
5701                }
5702        }
5703
5704        spin_unlock(&memcg_oom_lock);
5705}
5706
5707static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5708        struct cftype *cft,  struct cgroup_map_cb *cb)
5709{
5710        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5711
5712        cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5713
5714        if (atomic_read(&memcg->under_oom))
5715                cb->fill(cb, "under_oom", 1);
5716        else
5717                cb->fill(cb, "under_oom", 0);
5718        return 0;
5719}
5720
5721static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5722        struct cftype *cft, u64 val)
5723{
5724        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5725        struct mem_cgroup *parent;
5726
5727        /* cannot set to root cgroup and only 0 and 1 are allowed */
5728        if (!cgrp->parent || !((val == 0) || (val == 1)))
5729                return -EINVAL;
5730
5731        parent = mem_cgroup_from_cont(cgrp->parent);
5732
5733        mutex_lock(&memcg_create_mutex);
5734        /* oom-kill-disable is a flag for subhierarchy. */
5735        if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5736                mutex_unlock(&memcg_create_mutex);
5737                return -EINVAL;
5738        }
5739        memcg->oom_kill_disable = val;
5740        if (!val)
5741                memcg_oom_recover(memcg);
5742        mutex_unlock(&memcg_create_mutex);
5743        return 0;
5744}
5745
5746#ifdef CONFIG_MEMCG_KMEM
5747static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5748{
5749        int ret;
5750
5751        memcg->kmemcg_id = -1;
5752        ret = memcg_propagate_kmem(memcg);
5753        if (ret)
5754                return ret;
5755
5756        return mem_cgroup_sockets_init(memcg, ss);
5757};
5758
5759static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5760{
5761        mem_cgroup_sockets_destroy(memcg);
5762
5763        memcg_kmem_mark_dead(memcg);
5764
5765        if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5766                return;
5767
5768        /*
5769         * Charges already down to 0, undo mem_cgroup_get() done in the charge
5770         * path here, being careful not to race with memcg_uncharge_kmem: it is
5771         * possible that the charges went down to 0 between mark_dead and the
5772         * res_counter read, so in that case, we don't need the put
5773         */
5774        if (memcg_kmem_test_and_clear_dead(memcg))