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
 123/*
 124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
 125 * it will be incremated by the number of pages. This counter is used for
 126 * for trigger some periodic events. This is straightforward and better
 127 * than using jiffies etc. to handle periodic memcg event.
 128 */
 129enum mem_cgroup_events_target {
 130        MEM_CGROUP_TARGET_THRESH,
 131        MEM_CGROUP_TARGET_SOFTLIMIT,
 132        MEM_CGROUP_TARGET_NUMAINFO,
 133        MEM_CGROUP_NTARGETS,
 134};
 135#define THRESHOLDS_EVENTS_TARGET 128
 136#define SOFTLIMIT_EVENTS_TARGET 1024
 137#define NUMAINFO_EVENTS_TARGET  1024
 138
 139struct mem_cgroup_stat_cpu {
 140        long count[MEM_CGROUP_STAT_NSTATS];
 141        unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
 142        unsigned long nr_page_events;
 143        unsigned long targets[MEM_CGROUP_NTARGETS];
 144};
 145
 146struct mem_cgroup_reclaim_iter {
 147        /* css_id of the last scanned hierarchy member */
 148        int position;
 149        /* scan generation, increased every round-trip */
 150        unsigned int generation;
 151};
 152
 153/*
 154 * per-zone information in memory controller.
 155 */
 156struct mem_cgroup_per_zone {
 157        struct lruvec           lruvec;
 158        unsigned long           lru_size[NR_LRU_LISTS];
 159
 160        struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
 161
 162        struct rb_node          tree_node;      /* RB tree node */
 163        unsigned long long      usage_in_excess;/* Set to the value by which */
 164                                                /* the soft limit is exceeded*/
 165        bool                    on_tree;
 166        struct mem_cgroup       *memcg;         /* Back pointer, we cannot */
 167                                                /* use container_of        */
 168};
 169
 170struct mem_cgroup_per_node {
 171        struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 172};
 173
 174struct mem_cgroup_lru_info {
 175        struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 176};
 177
 178/*
 179 * Cgroups above their limits are maintained in a RB-Tree, independent of
 180 * their hierarchy representation
 181 */
 182
 183struct mem_cgroup_tree_per_zone {
 184        struct rb_root rb_root;
 185        spinlock_t lock;
 186};
 187
 188struct mem_cgroup_tree_per_node {
 189        struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
 190};
 191
 192struct mem_cgroup_tree {
 193        struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
 194};
 195
 196static struct mem_cgroup_tree soft_limit_tree __read_mostly;
 197
 198struct mem_cgroup_threshold {
 199        struct eventfd_ctx *eventfd;
 200        u64 threshold;
 201};
 202
 203/* For threshold */
 204struct mem_cgroup_threshold_ary {
 205        /* An array index points to threshold just below or equal to usage. */
 206        int current_threshold;
 207        /* Size of entries[] */
 208        unsigned int size;
 209        /* Array of thresholds */
 210        struct mem_cgroup_threshold entries[0];
 211};
 212
 213struct mem_cgroup_thresholds {
 214        /* Primary thresholds array */
 215        struct mem_cgroup_threshold_ary *primary;
 216        /*
 217         * Spare threshold array.
 218         * This is needed to make mem_cgroup_unregister_event() "never fail".
 219         * It must be able to store at least primary->size - 1 entries.
 220         */
 221        struct mem_cgroup_threshold_ary *spare;
 222};
 223
 224/* for OOM */
 225struct mem_cgroup_eventfd_list {
 226        struct list_head list;
 227        struct eventfd_ctx *eventfd;
 228};
 229
 230static void mem_cgroup_threshold(struct mem_cgroup *memcg);
 231static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
 232
 233/*
 234 * The memory controller data structure. The memory controller controls both
 235 * page cache and RSS per cgroup. We would eventually like to provide
 236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 237 * to help the administrator determine what knobs to tune.
 238 *
 239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
 240 * we hit the water mark. May be even add a low water mark, such that
 241 * no reclaim occurs from a cgroup at it's low water mark, this is
 242 * a feature that will be implemented much later in the future.
 243 */
 244struct mem_cgroup {
 245        struct cgroup_subsys_state css;
 246        /*
 247         * the counter to account for memory usage
 248         */
 249        struct res_counter res;
 250
 251        union {
 252                /*
 253                 * the counter to account for mem+swap usage.
 254                 */
 255                struct res_counter memsw;
 256
 257                /*
 258                 * rcu_freeing is used only when freeing struct mem_cgroup,
 259                 * so put it into a union to avoid wasting more memory.
 260                 * It must be disjoint from the css field.  It could be
 261                 * in a union with the res field, but res plays a much
 262                 * larger part in mem_cgroup life than memsw, and might
 263                 * be of interest, even at time of free, when debugging.
 264                 * So share rcu_head with the less interesting memsw.
 265                 */
 266                struct rcu_head rcu_freeing;
 267                /*
 268                 * We also need some space for a worker in deferred freeing.
 269                 * By the time we call it, rcu_freeing is no longer in use.
 270                 */
 271                struct work_struct work_freeing;
 272        };
 273
 274        /*
 275         * the counter to account for kernel memory usage.
 276         */
 277        struct res_counter kmem;
 278        /*
 279         * Per cgroup active and inactive list, similar to the
 280         * per zone LRU lists.
 281         */
 282        struct mem_cgroup_lru_info info;
 283        int last_scanned_node;
 284#if MAX_NUMNODES > 1
 285        nodemask_t      scan_nodes;
 286        atomic_t        numainfo_events;
 287        atomic_t        numainfo_updating;
 288#endif
 289        /*
 290         * Should the accounting and control be hierarchical, per subtree?
 291         */
 292        bool use_hierarchy;
 293        unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
 294
 295        bool            oom_lock;
 296        atomic_t        under_oom;
 297
 298        atomic_t        refcnt;
 299
 300        int     swappiness;
 301        /* OOM-Killer disable */
 302        int             oom_kill_disable;
 303
 304        /* set when res.limit == memsw.limit */
 305        bool            memsw_is_minimum;
 306
 307        /* protect arrays of thresholds */
 308        struct mutex thresholds_lock;
 309
 310        /* thresholds for memory usage. RCU-protected */
 311        struct mem_cgroup_thresholds thresholds;
 312
 313        /* thresholds for mem+swap usage. RCU-protected */
 314        struct mem_cgroup_thresholds memsw_thresholds;
 315
 316        /* For oom notifier event fd */
 317        struct list_head oom_notify;
 318
 319        /*
 320         * Should we move charges of a task when a task is moved into this
 321         * mem_cgroup ? And what type of charges should we move ?
 322         */
 323        unsigned long   move_charge_at_immigrate;
 324        /*
 325         * set > 0 if pages under this cgroup are moving to other cgroup.
 326         */
 327        atomic_t        moving_account;
 328        /* taken only while moving_account > 0 */
 329        spinlock_t      move_lock;
 330        /*
 331         * percpu counter.
 332         */
 333        struct mem_cgroup_stat_cpu __percpu *stat;
 334        /*
 335         * used when a cpu is offlined or other synchronizations
 336         * See mem_cgroup_read_stat().
 337         */
 338        struct mem_cgroup_stat_cpu nocpu_base;
 339        spinlock_t pcp_counter_lock;
 340
 341#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
 342        struct tcp_memcontrol tcp_mem;
 343#endif
 344#if defined(CONFIG_MEMCG_KMEM)
 345        /* analogous to slab_common's slab_caches list. per-memcg */
 346        struct list_head memcg_slab_caches;
 347        /* Not a spinlock, we can take a lot of time walking the list */
 348        struct mutex slab_caches_mutex;
 349        /* Index in the kmem_cache->memcg_params->memcg_caches array */
 350        int kmemcg_id;
 351#endif
 352};
 353
 354/* internal only representation about the status of kmem accounting. */
 355enum {
 356        KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
 357        KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
 358        KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
 359};
 360
 361/* We account when limit is on, but only after call sites are patched */
 362#define KMEM_ACCOUNTED_MASK \
 363                ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
 364
 365#ifdef CONFIG_MEMCG_KMEM
 366static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
 367{
 368        set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 369}
 370
 371static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
 372{
 373        return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 374}
 375
 376static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
 377{
 378        set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 379}
 380
 381static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
 382{
 383        clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 384}
 385
 386static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
 387{
 388        if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
 389                set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
 390}
 391
 392static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
 393{
 394        return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
 395                                  &memcg->kmem_account_flags);
 396}
 397#endif
 398
 399/* Stuffs for move charges at task migration. */
 400/*
 401 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
 402 * left-shifted bitmap of these types.
 403 */
 404enum move_type {
 405        MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
 406        MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
 407        NR_MOVE_TYPE,
 408};
 409
 410/* "mc" and its members are protected by cgroup_mutex */
 411static struct move_charge_struct {
 412        spinlock_t        lock; /* for from, to */
 413        struct mem_cgroup *from;
 414        struct mem_cgroup *to;
 415        unsigned long precharge;
 416        unsigned long moved_charge;
 417        unsigned long moved_swap;
 418        struct task_struct *moving_task;        /* a task moving charges */
 419        wait_queue_head_t waitq;                /* a waitq for other context */
 420} mc = {
 421        .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
 422        .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
 423};
 424
 425static bool move_anon(void)
 426{
 427        return test_bit(MOVE_CHARGE_TYPE_ANON,
 428                                        &mc.to->move_charge_at_immigrate);
 429}
 430
 431static bool move_file(void)
 432{
 433        return test_bit(MOVE_CHARGE_TYPE_FILE,
 434                                        &mc.to->move_charge_at_immigrate);
 435}
 436
 437/*
 438 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
 439 * limit reclaim to prevent infinite loops, if they ever occur.
 440 */
 441#define MEM_CGROUP_MAX_RECLAIM_LOOPS            100
 442#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
 443
 444enum charge_type {
 445        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 446        MEM_CGROUP_CHARGE_TYPE_ANON,
 447        MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
 448        MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
 449        NR_CHARGE_TYPE,
 450};
 451
 452/* for encoding cft->private value on file */
 453enum res_type {
 454        _MEM,
 455        _MEMSWAP,
 456        _OOM_TYPE,
 457        _KMEM,
 458};
 459
 460#define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
 461#define MEMFILE_TYPE(val)       ((val) >> 16 & 0xffff)
 462#define MEMFILE_ATTR(val)       ((val) & 0xffff)
 463/* Used for OOM nofiier */
 464#define OOM_CONTROL             (0)
 465
 466/*
 467 * Reclaim flags for mem_cgroup_hierarchical_reclaim
 468 */
 469#define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
 470#define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
 471#define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
 472#define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
 473
 474static void mem_cgroup_get(struct mem_cgroup *memcg);
 475static void mem_cgroup_put(struct mem_cgroup *memcg);
 476
 477static inline
 478struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
 479{
 480        return container_of(s, struct mem_cgroup, css);
 481}
 482
 483static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
 484{
 485        return (memcg == root_mem_cgroup);
 486}
 487
 488/* Writing them here to avoid exposing memcg's inner layout */
 489#if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
 490
 491void sock_update_memcg(struct sock *sk)
 492{
 493        if (mem_cgroup_sockets_enabled) {
 494                struct mem_cgroup *memcg;
 495                struct cg_proto *cg_proto;
 496
 497                BUG_ON(!sk->sk_prot->proto_cgroup);
 498
 499                /* Socket cloning can throw us here with sk_cgrp already
 500                 * filled. It won't however, necessarily happen from
 501                 * process context. So the test for root memcg given
 502                 * the current task's memcg won't help us in this case.
 503                 *
 504                 * Respecting the original socket's memcg is a better
 505                 * decision in this case.
 506                 */
 507                if (sk->sk_cgrp) {
 508                        BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
 509                        mem_cgroup_get(sk->sk_cgrp->memcg);
 510                        return;
 511                }
 512
 513                rcu_read_lock();
 514                memcg = mem_cgroup_from_task(current);
 515                cg_proto = sk->sk_prot->proto_cgroup(memcg);
 516                if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
 517                        mem_cgroup_get(memcg);
 518                        sk->sk_cgrp = cg_proto;
 519                }
 520                rcu_read_unlock();
 521        }
 522}
 523EXPORT_SYMBOL(sock_update_memcg);
 524
 525void sock_release_memcg(struct sock *sk)
 526{
 527        if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
 528                struct mem_cgroup *memcg;
 529                WARN_ON(!sk->sk_cgrp->memcg);
 530                memcg = sk->sk_cgrp->memcg;
 531                mem_cgroup_put(memcg);
 532        }
 533}
 534
 535struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
 536{
 537        if (!memcg || mem_cgroup_is_root(memcg))
 538                return NULL;
 539
 540        return &memcg->tcp_mem.cg_proto;
 541}
 542EXPORT_SYMBOL(tcp_proto_cgroup);
 543
 544static void disarm_sock_keys(struct mem_cgroup *memcg)
 545{
 546        if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
 547                return;
 548        static_key_slow_dec(&memcg_socket_limit_enabled);
 549}
 550#else
 551static void disarm_sock_keys(struct mem_cgroup *memcg)
 552{
 553}
 554#endif
 555
 556#ifdef CONFIG_MEMCG_KMEM
 557/*
 558 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
 559 * There are two main reasons for not using the css_id for this:
 560 *  1) this works better in sparse environments, where we have a lot of memcgs,
 561 *     but only a few kmem-limited. Or also, if we have, for instance, 200
 562 *     memcgs, and none but the 200th is kmem-limited, we'd have to have a
 563 *     200 entry array for that.
 564 *
 565 *  2) In order not to violate the cgroup API, we would like to do all memory
 566 *     allocation in ->create(). At that point, we haven't yet allocated the
 567 *     css_id. Having a separate index prevents us from messing with the cgroup
 568 *     core for this
 569 *
 570 * The current size of the caches array is stored in
 571 * memcg_limited_groups_array_size.  It will double each time we have to
 572 * increase it.
 573 */
 574static DEFINE_IDA(kmem_limited_groups);
 575int memcg_limited_groups_array_size;
 576
 577/*
 578 * MIN_SIZE is different than 1, because we would like to avoid going through
 579 * the alloc/free process all the time. In a small machine, 4 kmem-limited
 580 * cgroups is a reasonable guess. In the future, it could be a parameter or
 581 * tunable, but that is strictly not necessary.
 582 *
 583 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
 584 * this constant directly from cgroup, but it is understandable that this is
 585 * better kept as an internal representation in cgroup.c. In any case, the
 586 * css_id space is not getting any smaller, and we don't have to necessarily
 587 * increase ours as well if it increases.
 588 */
 589#define MEMCG_CACHES_MIN_SIZE 4
 590#define MEMCG_CACHES_MAX_SIZE 65535
 591
 592/*
 593 * A lot of the calls to the cache allocation functions are expected to be
 594 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
 595 * conditional to this static branch, we'll have to allow modules that does
 596 * kmem_cache_alloc and the such to see this symbol as well
 597 */
 598struct static_key memcg_kmem_enabled_key;
 599EXPORT_SYMBOL(memcg_kmem_enabled_key);
 600
 601static void disarm_kmem_keys(struct mem_cgroup *memcg)
 602{
 603        if (memcg_kmem_is_active(memcg)) {
 604                static_key_slow_dec(&memcg_kmem_enabled_key);
 605                ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
 606        }
 607        /*
 608         * This check can't live in kmem destruction function,
 609         * since the charges will outlive the cgroup
 610         */
 611        WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
 612}
 613#else
 614static void disarm_kmem_keys(struct mem_cgroup *memcg)
 615{
 616}
 617#endif /* CONFIG_MEMCG_KMEM */
 618
 619static void disarm_static_keys(struct mem_cgroup *memcg)
 620{
 621        disarm_sock_keys(memcg);
 622        disarm_kmem_keys(memcg);
 623}
 624
 625static void drain_all_stock_async(struct mem_cgroup *memcg);
 626
 627static struct mem_cgroup_per_zone *
 628mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
 629{
 630        return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
 631}
 632
 633struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
 634{
 635        return &memcg->css;
 636}
 637
 638static struct mem_cgroup_per_zone *
 639page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
 640{
 641        int nid = page_to_nid(page);
 642        int zid = page_zonenum(page);
 643
 644        return mem_cgroup_zoneinfo(memcg, nid, zid);
 645}
 646
 647static struct mem_cgroup_tree_per_zone *
 648soft_limit_tree_node_zone(int nid, int zid)
 649{
 650        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 651}
 652
 653static struct mem_cgroup_tree_per_zone *
 654soft_limit_tree_from_page(struct page *page)
 655{
 656        int nid = page_to_nid(page);
 657        int zid = page_zonenum(page);
 658
 659        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 660}
 661
 662static void
 663__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
 664                                struct mem_cgroup_per_zone *mz,
 665                                struct mem_cgroup_tree_per_zone *mctz,
 666                                unsigned long long new_usage_in_excess)
 667{
 668        struct rb_node **p = &mctz->rb_root.rb_node;
 669        struct rb_node *parent = NULL;
 670        struct mem_cgroup_per_zone *mz_node;
 671
 672        if (mz->on_tree)
 673                return;
 674
 675        mz->usage_in_excess = new_usage_in_excess;
 676        if (!mz->usage_in_excess)
 677                return;
 678        while (*p) {
 679                parent = *p;
 680                mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
 681                                        tree_node);
 682                if (mz->usage_in_excess < mz_node->usage_in_excess)
 683                        p = &(*p)->rb_left;
 684                /*
 685                 * We can't avoid mem cgroups that are over their soft
 686                 * limit by the same amount
 687                 */
 688                else if (mz->usage_in_excess >= mz_node->usage_in_excess)
 689                        p = &(*p)->rb_right;
 690        }
 691        rb_link_node(&mz->tree_node, parent, p);
 692        rb_insert_color(&mz->tree_node, &mctz->rb_root);
 693        mz->on_tree = true;
 694}
 695
 696static void
 697__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 698                                struct mem_cgroup_per_zone *mz,
 699                                struct mem_cgroup_tree_per_zone *mctz)
 700{
 701        if (!mz->on_tree)
 702                return;
 703        rb_erase(&mz->tree_node, &mctz->rb_root);
 704        mz->on_tree = false;
 705}
 706
 707static void
 708mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 709                                struct mem_cgroup_per_zone *mz,
 710                                struct mem_cgroup_tree_per_zone *mctz)
 711{
 712        spin_lock(&mctz->lock);
 713        __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 714        spin_unlock(&mctz->lock);
 715}
 716
 717
 718static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
 719{
 720        unsigned long long excess;
 721        struct mem_cgroup_per_zone *mz;
 722        struct mem_cgroup_tree_per_zone *mctz;
 723        int nid = page_to_nid(page);
 724        int zid = page_zonenum(page);
 725        mctz = soft_limit_tree_from_page(page);
 726
 727        /*
 728         * Necessary to update all ancestors when hierarchy is used.
 729         * because their event counter is not touched.
 730         */
 731        for (; memcg; memcg = parent_mem_cgroup(memcg)) {
 732                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 733                excess = res_counter_soft_limit_excess(&memcg->res);
 734                /*
 735                 * We have to update the tree if mz is on RB-tree or
 736                 * mem is over its softlimit.
 737                 */
 738                if (excess || mz->on_tree) {
 739                        spin_lock(&mctz->lock);
 740                        /* if on-tree, remove it */
 741                        if (mz->on_tree)
 742                                __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 743                        /*
 744                         * Insert again. mz->usage_in_excess will be updated.
 745                         * If excess is 0, no tree ops.
 746                         */
 747                        __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
 748                        spin_unlock(&mctz->lock);
 749                }
 750        }
 751}
 752
 753static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
 754{
 755        int node, zone;
 756        struct mem_cgroup_per_zone *mz;
 757        struct mem_cgroup_tree_per_zone *mctz;
 758
 759        for_each_node(node) {
 760                for (zone = 0; zone < MAX_NR_ZONES; zone++) {
 761                        mz = mem_cgroup_zoneinfo(memcg, node, zone);
 762                        mctz = soft_limit_tree_node_zone(node, zone);
 763                        mem_cgroup_remove_exceeded(memcg, mz, mctz);
 764                }
 765        }
 766}
 767
 768static struct mem_cgroup_per_zone *
 769__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 770{
 771        struct rb_node *rightmost = NULL;
 772        struct mem_cgroup_per_zone *mz;
 773
 774retry:
 775        mz = NULL;
 776        rightmost = rb_last(&mctz->rb_root);
 777        if (!rightmost)
 778                goto done;              /* Nothing to reclaim from */
 779
 780        mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
 781        /*
 782         * Remove the node now but someone else can add it back,
 783         * we will to add it back at the end of reclaim to its correct
 784         * position in the tree.
 785         */
 786        __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
 787        if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
 788                !css_tryget(&mz->memcg->css))
 789                goto retry;
 790done:
 791        return mz;
 792}
 793
 794static struct mem_cgroup_per_zone *
 795mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 796{
 797        struct mem_cgroup_per_zone *mz;
 798
 799        spin_lock(&mctz->lock);
 800        mz = __mem_cgroup_largest_soft_limit_node(mctz);
 801        spin_unlock(&mctz->lock);
 802        return mz;
 803}
 804
 805/*
 806 * Implementation Note: reading percpu statistics for memcg.
 807 *
 808 * Both of vmstat[] and percpu_counter has threshold and do periodic
 809 * synchronization to implement "quick" read. There are trade-off between
 810 * reading cost and precision of value. Then, we may have a chance to implement
 811 * a periodic synchronizion of counter in memcg's counter.
 812 *
 813 * But this _read() function is used for user interface now. The user accounts
 814 * memory usage by memory cgroup and he _always_ requires exact value because
 815 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
 816 * have to visit all online cpus and make sum. So, for now, unnecessary
 817 * synchronization is not implemented. (just implemented for cpu hotplug)
 818 *
 819 * If there are kernel internal actions which can make use of some not-exact
 820 * value, and reading all cpu value can be performance bottleneck in some
 821 * common workload, threashold and synchonization as vmstat[] should be
 822 * implemented.
 823 */
 824static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
 825                                 enum mem_cgroup_stat_index idx)
 826{
 827        long val = 0;
 828        int cpu;
 829
 830        get_online_cpus();
 831        for_each_online_cpu(cpu)
 832                val += per_cpu(memcg->stat->count[idx], cpu);
 833#ifdef CONFIG_HOTPLUG_CPU
 834        spin_lock(&memcg->pcp_counter_lock);
 835        val += memcg->nocpu_base.count[idx];
 836        spin_unlock(&memcg->pcp_counter_lock);
 837#endif
 838        put_online_cpus();
 839        return val;
 840}
 841
 842static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
 843                                         bool charge)
 844{
 845        int val = (charge) ? 1 : -1;
 846        this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
 847}
 848
 849static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
 850                                            enum mem_cgroup_events_index idx)
 851{
 852        unsigned long val = 0;
 853        int cpu;
 854
 855        for_each_online_cpu(cpu)
 856                val += per_cpu(memcg->stat->events[idx], cpu);
 857#ifdef CONFIG_HOTPLUG_CPU
 858        spin_lock(&memcg->pcp_counter_lock);
 859        val += memcg->nocpu_base.events[idx];
 860        spin_unlock(&memcg->pcp_counter_lock);
 861#endif
 862        return val;
 863}
 864
 865static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
 866                                         bool anon, int nr_pages)
 867{
 868        preempt_disable();
 869
 870        /*
 871         * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
 872         * counted as CACHE even if it's on ANON LRU.
 873         */
 874        if (anon)
 875                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
 876                                nr_pages);
 877        else
 878                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
 879                                nr_pages);
 880
 881        /* pagein of a big page is an event. So, ignore page size */
 882        if (nr_pages > 0)
 883                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
 884        else {
 885                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
 886                nr_pages = -nr_pages; /* for event */
 887        }
 888
 889        __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
 890
 891        preempt_enable();
 892}
 893
 894unsigned long
 895mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 896{
 897        struct mem_cgroup_per_zone *mz;
 898
 899        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
 900        return mz->lru_size[lru];
 901}
 902
 903static unsigned long
 904mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
 905                        unsigned int lru_mask)
 906{
 907        struct mem_cgroup_per_zone *mz;
 908        enum lru_list lru;
 909        unsigned long ret = 0;
 910
 911        mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 912
 913        for_each_lru(lru) {
 914                if (BIT(lru) & lru_mask)
 915                        ret += mz->lru_size[lru];
 916        }
 917        return ret;
 918}
 919
 920static unsigned long
 921mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
 922                        int nid, unsigned int lru_mask)
 923{
 924        u64 total = 0;
 925        int zid;
 926
 927        for (zid = 0; zid < MAX_NR_ZONES; zid++)
 928                total += mem_cgroup_zone_nr_lru_pages(memcg,
 929                                                nid, zid, lru_mask);
 930
 931        return total;
 932}
 933
 934static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
 935                        unsigned int lru_mask)
 936{
 937        int nid;
 938        u64 total = 0;
 939
 940        for_each_node_state(nid, N_MEMORY)
 941                total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
 942        return total;
 943}
 944
 945static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
 946                                       enum mem_cgroup_events_target target)
 947{
 948        unsigned long val, next;
 949
 950        val = __this_cpu_read(memcg->stat->nr_page_events);
 951        next = __this_cpu_read(memcg->stat->targets[target]);
 952        /* from time_after() in jiffies.h */
 953        if ((long)next - (long)val < 0) {
 954                switch (target) {
 955                case MEM_CGROUP_TARGET_THRESH:
 956                        next = val + THRESHOLDS_EVENTS_TARGET;
 957                        break;
 958                case MEM_CGROUP_TARGET_SOFTLIMIT:
 959                        next = val + SOFTLIMIT_EVENTS_TARGET;
 960                        break;
 961                case MEM_CGROUP_TARGET_NUMAINFO:
 962                        next = val + NUMAINFO_EVENTS_TARGET;
 963                        break;
 964                default:
 965                        break;
 966                }
 967                __this_cpu_write(memcg->stat->targets[target], next);
 968                return true;
 969        }
 970        return false;
 971}
 972
 973/*
 974 * Check events in order.
 975 *
 976 */
 977static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
 978{
 979        preempt_disable();
 980        /* threshold event is triggered in finer grain than soft limit */
 981        if (unlikely(mem_cgroup_event_ratelimit(memcg,
 982                                                MEM_CGROUP_TARGET_THRESH))) {
 983                bool do_softlimit;
 984                bool do_numainfo __maybe_unused;
 985
 986                do_softlimit = mem_cgroup_event_ratelimit(memcg,
 987                                                MEM_CGROUP_TARGET_SOFTLIMIT);
 988#if MAX_NUMNODES > 1
 989                do_numainfo = mem_cgroup_event_ratelimit(memcg,
 990                                                MEM_CGROUP_TARGET_NUMAINFO);
 991#endif
 992                preempt_enable();
 993
 994                mem_cgroup_threshold(memcg);
 995                if (unlikely(do_softlimit))
 996                        mem_cgroup_update_tree(memcg, page);
 997#if MAX_NUMNODES > 1
 998                if (unlikely(do_numainfo))
 999                        atomic_inc(&memcg->numainfo_events);
1000#endif
1001        } else
1002                preempt_enable();
1003}
1004
1005struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1006{
1007        return mem_cgroup_from_css(
1008                cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1009}
1010
1011struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1012{
1013        /*
1014         * mm_update_next_owner() may clear mm->owner to NULL
1015         * if it races with swapoff, page migration, etc.
1016         * So this can be called with p == NULL.
1017         */
1018        if (unlikely(!p))
1019                return NULL;
1020
1021        return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1022}
1023
1024struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1025{
1026        struct mem_cgroup *memcg = NULL;
1027
1028        if (!mm)
1029                return NULL;
1030        /*
1031         * Because we have no locks, mm->owner's may be being moved to other
1032         * cgroup. We use css_tryget() here even if this looks
1033         * pessimistic (rather than adding locks here).
1034         */
1035        rcu_read_lock();
1036        do {
1037                memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1038                if (unlikely(!memcg))
1039                        break;
1040        } while (!css_tryget(&memcg->css));
1041        rcu_read_unlock();
1042        return memcg;
1043}
1044
1045/**
1046 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1047 * @root: hierarchy root
1048 * @prev: previously returned memcg, NULL on first invocation
1049 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1050 *
1051 * Returns references to children of the hierarchy below @root, or
1052 * @root itself, or %NULL after a full round-trip.
1053 *
1054 * Caller must pass the return value in @prev on subsequent
1055 * invocations for reference counting, or use mem_cgroup_iter_break()
1056 * to cancel a hierarchy walk before the round-trip is complete.
1057 *
1058 * Reclaimers can specify a zone and a priority level in @reclaim to
1059 * divide up the memcgs in the hierarchy among all concurrent
1060 * reclaimers operating on the same zone and priority.
1061 */
1062struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1063                                   struct mem_cgroup *prev,
1064                                   struct mem_cgroup_reclaim_cookie *reclaim)
1065{
1066        struct mem_cgroup *memcg = NULL;
1067        int id = 0;
1068
1069        if (mem_cgroup_disabled())
1070                return NULL;
1071
1072        if (!root)
1073                root = root_mem_cgroup;
1074
1075        if (prev && !reclaim)
1076                id = css_id(&prev->css);
1077
1078        if (prev && prev != root)
1079                css_put(&prev->css);
1080
1081        if (!root->use_hierarchy && root != root_mem_cgroup) {
1082                if (prev)
1083                        return NULL;
1084                return root;
1085        }
1086
1087        while (!memcg) {
1088                struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1089                struct cgroup_subsys_state *css;
1090
1091                if (reclaim) {
1092                        int nid = zone_to_nid(reclaim->zone);
1093                        int zid = zone_idx(reclaim->zone);
1094                        struct mem_cgroup_per_zone *mz;
1095
1096                        mz = mem_cgroup_zoneinfo(root, nid, zid);
1097                        iter = &mz->reclaim_iter[reclaim->priority];
1098                        if (prev && reclaim->generation != iter->generation)
1099                                return NULL;
1100                        id = iter->position;
1101                }
1102
1103                rcu_read_lock();
1104                css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1105                if (css) {
1106                        if (css == &root->css || css_tryget(css))
1107                                memcg = mem_cgroup_from_css(css);
1108                } else
1109                        id = 0;
1110                rcu_read_unlock();
1111
1112                if (reclaim) {
1113                        iter->position = id;
1114                        if (!css)
1115                                iter->generation++;
1116                        else if (!prev && memcg)
1117                                reclaim->generation = iter->generation;
1118                }
1119
1120                if (prev && !css)
1121                        return NULL;
1122        }
1123        return memcg;
1124}
1125
1126/**
1127 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1128 * @root: hierarchy root
1129 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130 */
1131void mem_cgroup_iter_break(struct mem_cgroup *root,
1132                           struct mem_cgroup *prev)
1133{
1134        if (!root)
1135                root = root_mem_cgroup;
1136        if (prev && prev != root)
1137                css_put(&prev->css);
1138}
1139
1140/*
1141 * Iteration constructs for visiting all cgroups (under a tree).  If
1142 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1143 * be used for reference counting.
1144 */
1145#define for_each_mem_cgroup_tree(iter, root)            \
1146        for (iter = mem_cgroup_iter(root, NULL, NULL);  \
1147             iter != NULL;                              \
1148             iter = mem_cgroup_iter(root, iter, NULL))
1149
1150#define for_each_mem_cgroup(iter)                       \
1151        for (iter = mem_cgroup_iter(NULL, NULL, NULL);  \
1152             iter != NULL;                              \
1153             iter = mem_cgroup_iter(NULL, iter, NULL))
1154
1155void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1156{
1157        struct mem_cgroup *memcg;
1158
1159        rcu_read_lock();
1160        memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1161        if (unlikely(!memcg))
1162                goto out;
1163
1164        switch (idx) {
1165        case PGFAULT:
1166                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1167                break;
1168        case PGMAJFAULT:
1169                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1170                break;
1171        default:
1172                BUG();
1173        }
1174out:
1175        rcu_read_unlock();
1176}
1177EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1178
1179/**
1180 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1181 * @zone: zone of the wanted lruvec
1182 * @memcg: memcg of the wanted lruvec
1183 *
1184 * Returns the lru list vector holding pages for the given @zone and
1185 * @mem.  This can be the global zone lruvec, if the memory controller
1186 * is disabled.
1187 */
1188struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1189                                      struct mem_cgroup *memcg)
1190{
1191        struct mem_cgroup_per_zone *mz;
1192        struct lruvec *lruvec;
1193
1194        if (mem_cgroup_disabled()) {
1195                lruvec = &zone->lruvec;
1196                goto out;
1197        }
1198
1199        mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1200        lruvec = &mz->lruvec;
1201out:
1202        /*
1203         * Since a node can be onlined after the mem_cgroup was created,
1204         * we have to be prepared to initialize lruvec->zone here;
1205         * and if offlined then reonlined, we need to reinitialize it.
1206         */
1207        if (unlikely(lruvec->zone != zone))
1208                lruvec->zone = zone;
1209        return lruvec;
1210}
1211
1212/*
1213 * Following LRU functions are allowed to be used without PCG_LOCK.
1214 * Operations are called by routine of global LRU independently from memcg.
1215 * What we have to take care of here is validness of pc->mem_cgroup.
1216 *
1217 * Changes to pc->mem_cgroup happens when
1218 * 1. charge
1219 * 2. moving account
1220 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1221 * It is added to LRU before charge.
1222 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1223 * When moving account, the page is not on LRU. It's isolated.
1224 */
1225
1226/**
1227 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1228 * @page: the page
1229 * @zone: zone of the page
1230 */
1231struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1232{
1233        struct mem_cgroup_per_zone *mz;
1234        struct mem_cgroup *memcg;
1235        struct page_cgroup *pc;
1236        struct lruvec *lruvec;
1237
1238        if (mem_cgroup_disabled()) {
1239                lruvec = &zone->lruvec;
1240                goto out;
1241        }
1242
1243        pc = lookup_page_cgroup(page);
1244        memcg = pc->mem_cgroup;
1245
1246        /*
1247         * Surreptitiously switch any uncharged offlist page to root:
1248         * an uncharged page off lru does nothing to secure
1249         * its former mem_cgroup from sudden removal.
1250         *
1251         * Our caller holds lru_lock, and PageCgroupUsed is updated
1252         * under page_cgroup lock: between them, they make all uses
1253         * of pc->mem_cgroup safe.
1254         */
1255        if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1256                pc->mem_cgroup = memcg = root_mem_cgroup;
1257
1258        mz = page_cgroup_zoneinfo(memcg, page);
1259        lruvec = &mz->lruvec;
1260out:
1261        /*
1262         * Since a node can be onlined after the mem_cgroup was created,
1263         * we have to be prepared to initialize lruvec->zone here;
1264         * and if offlined then reonlined, we need to reinitialize it.
1265         */
1266        if (unlikely(lruvec->zone != zone))
1267                lruvec->zone = zone;
1268        return lruvec;
1269}
1270
1271/**
1272 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1273 * @lruvec: mem_cgroup per zone lru vector
1274 * @lru: index of lru list the page is sitting on
1275 * @nr_pages: positive when adding or negative when removing
1276 *
1277 * This function must be called when a page is added to or removed from an
1278 * lru list.
1279 */
1280void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1281                                int nr_pages)
1282{
1283        struct mem_cgroup_per_zone *mz;
1284        unsigned long *lru_size;
1285
1286        if (mem_cgroup_disabled())
1287                return;
1288
1289        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1290        lru_size = mz->lru_size + lru;
1291        *lru_size += nr_pages;
1292        VM_BUG_ON((long)(*lru_size) < 0);
1293}
1294
1295/*
1296 * Checks whether given mem is same or in the root_mem_cgroup's
1297 * hierarchy subtree
1298 */
1299bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1300                                  struct mem_cgroup *memcg)
1301{
1302        if (root_memcg == memcg)
1303                return true;
1304        if (!root_memcg->use_hierarchy || !memcg)
1305                return false;
1306        return css_is_ancestor(&memcg->css, &root_memcg->css);
1307}
1308
1309static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1310                                       struct mem_cgroup *memcg)
1311{
1312        bool ret;
1313
1314        rcu_read_lock();
1315        ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1316        rcu_read_unlock();
1317        return ret;
1318}
1319
1320int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1321{
1322        int ret;
1323        struct mem_cgroup *curr = NULL;
1324        struct task_struct *p;
1325
1326        p = find_lock_task_mm(task);
1327        if (p) {
1328                curr = try_get_mem_cgroup_from_mm(p->mm);
1329                task_unlock(p);
1330        } else {
1331                /*
1332                 * All threads may have already detached their mm's, but the oom
1333                 * killer still needs to detect if they have already been oom
1334                 * killed to prevent needlessly killing additional tasks.
1335                 */
1336                task_lock(task);
1337                curr = mem_cgroup_from_task(task);
1338                if (curr)
1339                        css_get(&curr->css);
1340                task_unlock(task);
1341        }
1342        if (!curr)
1343                return 0;
1344        /*
1345         * We should check use_hierarchy of "memcg" not "curr". Because checking
1346         * use_hierarchy of "curr" here make this function true if hierarchy is
1347         * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1348         * hierarchy(even if use_hierarchy is disabled in "memcg").
1349         */
1350        ret = mem_cgroup_same_or_subtree(memcg, curr);
1351        css_put(&curr->css);
1352        return ret;
1353}
1354
1355int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1356{
1357        unsigned long inactive_ratio;
1358        unsigned long inactive;
1359        unsigned long active;
1360        unsigned long gb;
1361
1362        inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1363        active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1364
1365        gb = (inactive + active) >> (30 - PAGE_SHIFT);
1366        if (gb)
1367                inactive_ratio = int_sqrt(10 * gb);
1368        else
1369                inactive_ratio = 1;
1370
1371        return inactive * inactive_ratio < active;
1372}
1373
1374int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1375{
1376        unsigned long active;
1377        unsigned long inactive;
1378
1379        inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1380        active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1381
1382        return (active > inactive);
1383}
1384
1385#define mem_cgroup_from_res_counter(counter, member)    \
1386        container_of(counter, struct mem_cgroup, member)
1387
1388/**
1389 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1390 * @memcg: the memory cgroup
1391 *
1392 * Returns the maximum amount of memory @mem can be charged with, in
1393 * pages.
1394 */
1395static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1396{
1397        unsigned long long margin;
1398
1399        margin = res_counter_margin(&memcg->res);
1400        if (do_swap_account)
1401                margin = min(margin, res_counter_margin(&memcg->memsw));
1402        return margin >> PAGE_SHIFT;
1403}
1404
1405int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1406{
1407        struct cgroup *cgrp = memcg->css.cgroup;
1408
1409        /* root ? */
1410        if (cgrp->parent == NULL)
1411                return vm_swappiness;
1412
1413        return memcg->swappiness;
1414}
1415
1416/*
1417 * memcg->moving_account is used for checking possibility that some thread is
1418 * calling move_account(). When a thread on CPU-A starts moving pages under
1419 * a memcg, other threads should check memcg->moving_account under
1420 * rcu_read_lock(), like this:
1421 *
1422 *         CPU-A                                    CPU-B
1423 *                                              rcu_read_lock()
1424 *         memcg->moving_account+1              if (memcg->mocing_account)
1425 *                                                   take heavy locks.
1426 *         synchronize_rcu()                    update something.
1427 *                                              rcu_read_unlock()
1428 *         start move here.
1429 */
1430
1431/* for quick checking without looking up memcg */
1432atomic_t memcg_moving __read_mostly;
1433
1434static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1435{
1436        atomic_inc(&memcg_moving);
1437        atomic_inc(&memcg->moving_account);
1438        synchronize_rcu();
1439}
1440
1441static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1442{
1443        /*
1444         * Now, mem_cgroup_clear_mc() may call this function with NULL.
1445         * We check NULL in callee rather than caller.
1446         */
1447        if (memcg) {
1448                atomic_dec(&memcg_moving);
1449                atomic_dec(&memcg->moving_account);
1450        }
1451}
1452
1453/*
1454 * 2 routines for checking "mem" is under move_account() or not.
1455 *
1456 * mem_cgroup_stolen() -  checking whether a cgroup is mc.from or not. This
1457 *                        is used for avoiding races in accounting.  If true,
1458 *                        pc->mem_cgroup may be overwritten.
1459 *
1460 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1461 *                        under hierarchy of moving cgroups. This is for
1462 *                        waiting at hith-memory prressure caused by "move".
1463 */
1464
1465static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1466{
1467        VM_BUG_ON(!rcu_read_lock_held());
1468        return atomic_read(&memcg->moving_account) > 0;
1469}
1470
1471static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1472{
1473        struct mem_cgroup *from;
1474        struct mem_cgroup *to;
1475        bool ret = false;
1476        /*
1477         * Unlike task_move routines, we access mc.to, mc.from not under
1478         * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1479         */
1480        spin_lock(&mc.lock);
1481        from = mc.from;
1482        to = mc.to;
1483        if (!from)
1484                goto unlock;
1485
1486        ret = mem_cgroup_same_or_subtree(memcg, from)
1487                || mem_cgroup_same_or_subtree(memcg, to);
1488unlock:
1489        spin_unlock(&mc.lock);
1490        return ret;
1491}
1492
1493static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1494{
1495        if (mc.moving_task && current != mc.moving_task) {
1496                if (mem_cgroup_under_move(memcg)) {
1497                        DEFINE_WAIT(wait);
1498                        prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1499                        /* moving charge context might have finished. */
1500                        if (mc.moving_task)
1501                                schedule();
1502                        finish_wait(&mc.waitq, &wait);
1503                        return true;
1504                }
1505        }
1506        return false;
1507}
1508
1509/*
1510 * Take this lock when
1511 * - a code tries to modify page's memcg while it's USED.
1512 * - a code tries to modify page state accounting in a memcg.
1513 * see mem_cgroup_stolen(), too.
1514 */
1515static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1516                                  unsigned long *flags)
1517{
1518        spin_lock_irqsave(&memcg->move_lock, *flags);
1519}
1520
1521static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1522                                unsigned long *flags)
1523{
1524        spin_unlock_irqrestore(&memcg->move_lock, *flags);
1525}
1526
1527/**
1528 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1529 * @memcg: The memory cgroup that went over limit
1530 * @p: Task that is going to be killed
1531 *
1532 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1533 * enabled
1534 */
1535void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1536{
1537        struct cgroup *task_cgrp;
1538        struct cgroup *mem_cgrp;
1539        /*
1540         * Need a buffer in BSS, can't rely on allocations. The code relies
1541         * on the assumption that OOM is serialized for memory controller.
1542         * If this assumption is broken, revisit this code.
1543         */
1544        static char memcg_name[PATH_MAX];
1545        int ret;
1546
1547        if (!memcg || !p)
1548                return;
1549
1550        rcu_read_lock();
1551
1552        mem_cgrp = memcg->css.cgroup;
1553        task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1554
1555        ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1556        if (ret < 0) {
1557                /*
1558                 * Unfortunately, we are unable to convert to a useful name
1559                 * But we'll still print out the usage information
1560                 */
1561                rcu_read_unlock();
1562                goto done;
1563        }
1564        rcu_read_unlock();
1565
1566        printk(KERN_INFO "Task in %s killed", memcg_name);
1567
1568        rcu_read_lock();
1569        ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1570        if (ret < 0) {
1571                rcu_read_unlock();
1572                goto done;
1573        }
1574        rcu_read_unlock();
1575
1576        /*
1577         * Continues from above, so we don't need an KERN_ level
1578         */
1579        printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1580done:
1581
1582        printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1583                res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1584                res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1585                res_counter_read_u64(&memcg->res, RES_FAILCNT));
1586        printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1587                "failcnt %llu\n",
1588                res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1589                res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1590                res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1591        printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1592                res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1593                res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1594                res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1595}
1596
1597/*
1598 * This function returns the number of memcg under hierarchy tree. Returns
1599 * 1(self count) if no children.
1600 */
1601static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1602{
1603        int num = 0;
1604        struct mem_cgroup *iter;
1605
1606        for_each_mem_cgroup_tree(iter, memcg)
1607                num++;
1608        return num;
1609}
1610
1611/*
1612 * Return the memory (and swap, if configured) limit for a memcg.
1613 */
1614static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1615{
1616        u64 limit;
1617
1618        limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1619
1620        /*
1621         * Do not consider swap space if we cannot swap due to swappiness
1622         */
1623        if (mem_cgroup_swappiness(memcg)) {
1624                u64 memsw;
1625
1626                limit += total_swap_pages << PAGE_SHIFT;
1627                memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1628
1629                /*
1630                 * If memsw is finite and limits the amount of swap space
1631                 * available to this memcg, return that limit.
1632                 */
1633                limit = min(limit, memsw);
1634        }
1635
1636        return limit;
1637}
1638
1639static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1640                                     int order)
1641{
1642        struct mem_cgroup *iter;
1643        unsigned long chosen_points = 0;
1644        unsigned long totalpages;
1645        unsigned int points = 0;
1646        struct task_struct *chosen = NULL;
1647
1648        /*
1649         * If current has a pending SIGKILL, then automatically select it.  The
1650         * goal is to allow it to allocate so that it may quickly exit and free
1651         * its memory.
1652         */
1653        if (fatal_signal_pending(current)) {
1654                set_thread_flag(TIF_MEMDIE);
1655                return;
1656        }
1657
1658        check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1659        totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1660        for_each_mem_cgroup_tree(iter, memcg) {
1661                struct cgroup *cgroup = iter->css.cgroup;
1662                struct cgroup_iter it;
1663                struct task_struct *task;
1664
1665                cgroup_iter_start(cgroup, &it);
1666                while ((task = cgroup_iter_next(cgroup, &it))) {
1667                        switch (oom_scan_process_thread(task, totalpages, NULL,
1668                                                        false)) {
1669                        case OOM_SCAN_SELECT:
1670                                if (chosen)
1671                                        put_task_struct(chosen);
1672                                chosen = task;
1673                                chosen_points = ULONG_MAX;
1674                                get_task_struct(chosen);
1675                                /* fall through */
1676                        case OOM_SCAN_CONTINUE:
1677                                continue;
1678                        case OOM_SCAN_ABORT:
1679                                cgroup_iter_end(cgroup, &it);
1680                                mem_cgroup_iter_break(memcg, iter);
1681                                if (chosen)
1682                                        put_task_struct(chosen);
1683                                return;
1684                        case OOM_SCAN_OK:
1685                                break;
1686                        };
1687                        points = oom_badness(task, memcg, NULL, totalpages);
1688                        if (points > chosen_points) {
1689                                if (chosen)
1690                                        put_task_struct(chosen);
1691                                chosen = task;
1692                                chosen_points = points;
1693                                get_task_struct(chosen);
1694                        }
1695                }
1696                cgroup_iter_end(cgroup, &it);
1697        }
1698
1699        if (!chosen)
1700                return;
1701        points = chosen_points * 1000 / totalpages;
1702        oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1703                         NULL, "Memory cgroup out of memory");
1704}
1705
1706static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1707                                        gfp_t gfp_mask,
1708                                        unsigned long flags)
1709{
1710        unsigned long total = 0;
1711        bool noswap = false;
1712        int loop;
1713
1714        if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1715                noswap = true;
1716        if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1717                noswap = true;
1718
1719        for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1720                if (loop)
1721                        drain_all_stock_async(memcg);
1722                total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1723                /*
1724                 * Allow limit shrinkers, which are triggered directly
1725                 * by userspace, to catch signals and stop reclaim
1726                 * after minimal progress, regardless of the margin.
1727                 */
1728                if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1729                        break;
1730                if (mem_cgroup_margin(memcg))
1731                        break;
1732                /*
1733                 * If nothing was reclaimed after two attempts, there
1734                 * may be no reclaimable pages in this hierarchy.
1735                 */
1736                if (loop && !total)
1737                        break;
1738        }
1739        return total;
1740}
1741
1742/**
1743 * test_mem_cgroup_node_reclaimable
1744 * @memcg: the target memcg
1745 * @nid: the node ID to be checked.
1746 * @noswap : specify true here if the user wants flle only information.
1747 *
1748 * This function returns whether the specified memcg contains any
1749 * reclaimable pages on a node. Returns true if there are any reclaimable
1750 * pages in the node.
1751 */
1752static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1753                int nid, bool noswap)
1754{
1755        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1756                return true;
1757        if (noswap || !total_swap_pages)
1758                return false;
1759        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1760                return true;
1761        return false;
1762
1763}
1764#if MAX_NUMNODES > 1
1765
1766/*
1767 * Always updating the nodemask is not very good - even if we have an empty
1768 * list or the wrong list here, we can start from some node and traverse all
1769 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1770 *
1771 */
1772static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1773{
1774        int nid;
1775        /*
1776         * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1777         * pagein/pageout changes since the last update.
1778         */
1779        if (!atomic_read(&memcg->numainfo_events))
1780                return;
1781        if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1782                return;
1783
1784        /* make a nodemask where this memcg uses memory from */
1785        memcg->scan_nodes = node_states[N_MEMORY];
1786
1787        for_each_node_mask(nid, node_states[N_MEMORY]) {
1788
1789                if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1790                        node_clear(nid, memcg->scan_nodes);
1791        }
1792
1793        atomic_set(&memcg->numainfo_events, 0);
1794        atomic_set(&memcg->numainfo_updating, 0);
1795}
1796
1797/*
1798 * Selecting a node where we start reclaim from. Because what we need is just
1799 * reducing usage counter, start from anywhere is O,K. Considering
1800 * memory reclaim from current node, there are pros. and cons.
1801 *
1802 * Freeing memory from current node means freeing memory from a node which
1803 * we'll use or we've used. So, it may make LRU bad. And if several threads
1804 * hit limits, it will see a contention on a node. But freeing from remote
1805 * node means more costs for memory reclaim because of memory latency.
1806 *
1807 * Now, we use round-robin. Better algorithm is welcomed.
1808 */
1809int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1810{
1811        int node;
1812
1813        mem_cgroup_may_update_nodemask(memcg);
1814        node = memcg->last_scanned_node;
1815
1816        node = next_node(node, memcg->scan_nodes);
1817        if (node == MAX_NUMNODES)
1818                node = first_node(memcg->scan_nodes);
1819        /*
1820         * We call this when we hit limit, not when pages are added to LRU.
1821         * No LRU may hold pages because all pages are UNEVICTABLE or
1822         * memcg is too small and all pages are not on LRU. In that case,
1823         * we use curret node.
1824         */
1825        if (unlikely(node == MAX_NUMNODES))
1826                node = numa_node_id();
1827
1828        memcg->last_scanned_node = node;
1829        return node;
1830}
1831
1832/*
1833 * Check all nodes whether it contains reclaimable pages or not.
1834 * For quick scan, we make use of scan_nodes. This will allow us to skip
1835 * unused nodes. But scan_nodes is lazily updated and may not cotain
1836 * enough new information. We need to do double check.
1837 */
1838static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1839{
1840        int nid;
1841
1842        /*
1843         * quick check...making use of scan_node.
1844         * We can skip unused nodes.
1845         */
1846        if (!nodes_empty(memcg->scan_nodes)) {
1847                for (nid = first_node(memcg->scan_nodes);
1848                     nid < MAX_NUMNODES;
1849                     nid = next_node(nid, memcg->scan_nodes)) {
1850
1851                        if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1852                                return true;
1853                }
1854        }
1855        /*
1856         * Check rest of nodes.
1857         */
1858        for_each_node_state(nid, N_MEMORY) {
1859                if (node_isset(nid, memcg->scan_nodes))
1860                        continue;
1861                if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1862                        return true;
1863        }
1864        return false;
1865}
1866
1867#else
1868int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1869{
1870        return 0;
1871}
1872
1873static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1874{
1875        return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1876}
1877#endif
1878
1879static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1880                                   struct zone *zone,
1881                                   gfp_t gfp_mask,
1882                                   unsigned long *total_scanned)
1883{
1884        struct mem_cgroup *victim = NULL;
1885        int total = 0;
1886        int loop = 0;
1887        unsigned long excess;
1888        unsigned long nr_scanned;
1889        struct mem_cgroup_reclaim_cookie reclaim = {
1890                .zone = zone,
1891                .priority = 0,
1892        };
1893
1894        excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1895
1896        while (1) {
1897                victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1898                if (!victim) {
1899                        loop++;
1900                        if (loop >= 2) {
1901                                /*
1902                                 * If we have not been able to reclaim
1903                                 * anything, it might because there are
1904                                 * no reclaimable pages under this hierarchy
1905                                 */
1906                                if (!total)
1907                                        break;
1908                                /*
1909                                 * We want to do more targeted reclaim.
1910                                 * excess >> 2 is not to excessive so as to
1911                                 * reclaim too much, nor too less that we keep
1912                                 * coming back to reclaim from this cgroup
1913                                 */
1914                                if (total >= (excess >> 2) ||
1915                                        (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1916                                        break;
1917                        }
1918                        continue;
1919                }
1920                if (!mem_cgroup_reclaimable(victim, false))
1921                        continue;
1922                total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1923                                                     zone, &nr_scanned);
1924                *total_scanned += nr_scanned;
1925                if (!res_counter_soft_limit_excess(&root_memcg->res))
1926                        break;
1927        }
1928        mem_cgroup_iter_break(root_memcg, victim);
1929        return total;
1930}
1931
1932/*
1933 * Check OOM-Killer is already running under our hierarchy.
1934 * If someone is running, return false.
1935 * Has to be called with memcg_oom_lock
1936 */
1937static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1938{
1939        struct mem_cgroup *iter, *failed = NULL;
1940
1941        for_each_mem_cgroup_tree(iter, memcg) {
1942                if (iter->oom_lock) {
1943                        /*
1944                         * this subtree of our hierarchy is already locked
1945                         * so we cannot give a lock.
1946                         */
1947                        failed = iter;
1948                        mem_cgroup_iter_break(memcg, iter);
1949                        break;
1950                } else
1951                        iter->oom_lock = true;
1952        }
1953
1954        if (!failed)
1955                return true;
1956
1957        /*
1958         * OK, we failed to lock the whole subtree so we have to clean up
1959         * what we set up to the failing subtree
1960         */
1961        for_each_mem_cgroup_tree(iter, memcg) {
1962                if (iter == failed) {
1963                        mem_cgroup_iter_break(memcg, iter);
1964                        break;
1965                }
1966                iter->oom_lock = false;
1967        }
1968        return false;
1969}
1970
1971/*
1972 * Has to be called with memcg_oom_lock
1973 */
1974static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1975{
1976        struct mem_cgroup *iter;
1977
1978        for_each_mem_cgroup_tree(iter, memcg)
1979                iter->oom_lock = false;
1980        return 0;
1981}
1982
1983static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1984{
1985        struct mem_cgroup *iter;
1986
1987        for_each_mem_cgroup_tree(iter, memcg)
1988                atomic_inc(&iter->under_oom);
1989}
1990
1991static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1992{
1993        struct mem_cgroup *iter;
1994
1995        /*
1996         * When a new child is created while the hierarchy is under oom,
1997         * mem_cgroup_oom_lock() may not be called. We have to use
1998         * atomic_add_unless() here.
1999         */
2000        for_each_mem_cgroup_tree(iter, memcg)
2001                atomic_add_unless(&iter->under_oom, -1, 0);
2002}
2003
2004static DEFINE_SPINLOCK(memcg_oom_lock);
2005static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2006
2007struct oom_wait_info {
2008        struct mem_cgroup *memcg;
2009        wait_queue_t    wait;
2010};
2011
2012static int memcg_oom_wake_function(wait_queue_t *wait,
2013        unsigned mode, int sync, void *arg)
2014{
2015        struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2016        struct mem_cgroup *oom_wait_memcg;
2017        struct oom_wait_info *oom_wait_info;
2018
2019        oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2020        oom_wait_memcg = oom_wait_info->memcg;
2021
2022        /*
2023         * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2024         * Then we can use css_is_ancestor without taking care of RCU.
2025         */
2026        if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2027                && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2028                return 0;
2029        return autoremove_wake_function(wait, mode, sync, arg);
2030}
2031
2032static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2033{
2034        /* for filtering, pass "memcg" as argument. */
2035        __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2036}
2037
2038static void memcg_oom_recover(struct mem_cgroup *memcg)
2039{
2040        if (memcg && atomic_read(&memcg->under_oom))
2041                memcg_wakeup_oom(memcg);
2042}
2043
2044/*
2045 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2046 */
2047static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2048                                  int order)
2049{
2050        struct oom_wait_info owait;
2051        bool locked, need_to_kill;
2052
2053        owait.memcg = memcg;
2054        owait.wait.flags = 0;
2055        owait.wait.func = memcg_oom_wake_function;
2056        owait.wait.private = current;
2057        INIT_LIST_HEAD(&owait.wait.task_list);
2058        need_to_kill = true;
2059        mem_cgroup_mark_under_oom(memcg);
2060
2061        /* At first, try to OOM lock hierarchy under memcg.*/
2062        spin_lock(&memcg_oom_lock);
2063        locked = mem_cgroup_oom_lock(memcg);
2064        /*
2065         * Even if signal_pending(), we can't quit charge() loop without
2066         * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2067         * under OOM is always welcomed, use TASK_KILLABLE here.
2068         */
2069        prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2070        if (!locked || memcg->oom_kill_disable)
2071                need_to_kill = false;
2072        if (locked)
2073                mem_cgroup_oom_notify(memcg);
2074        spin_unlock(&memcg_oom_lock);
2075
2076        if (need_to_kill) {
2077                finish_wait(&memcg_oom_waitq, &owait.wait);
2078                mem_cgroup_out_of_memory(memcg, mask, order);
2079        } else {
2080                schedule();
2081                finish_wait(&memcg_oom_waitq, &owait.wait);
2082        }
2083        spin_lock(&memcg_oom_lock);
2084        if (locked)
2085                mem_cgroup_oom_unlock(memcg);
2086        memcg_wakeup_oom(memcg);
2087        spin_unlock(&memcg_oom_lock);
2088
2089        mem_cgroup_unmark_under_oom(memcg);
2090
2091        if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2092                return false;
2093        /* Give chance to dying process */
2094        schedule_timeout_uninterruptible(1);
2095        return true;
2096}
2097
2098/*
2099 * Currently used to update mapped file statistics, but the routine can be
2100 * generalized to update other statistics as well.
2101 *
2102 * Notes: Race condition
2103 *
2104 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2105 * it tends to be costly. But considering some conditions, we doesn't need
2106 * to do so _always_.
2107 *
2108 * Considering "charge", lock_page_cgroup() is not required because all
2109 * file-stat operations happen after a page is attached to radix-tree. There
2110 * are no race with "charge".
2111 *
2112 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2113 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2114 * if there are race with "uncharge". Statistics itself is properly handled
2115 * by flags.
2116 *
2117 * Considering "move", this is an only case we see a race. To make the race
2118 * small, we check mm->moving_account and detect there are possibility of race
2119 * If there is, we take a lock.
2120 */
2121
2122void __mem_cgroup_begin_update_page_stat(struct page *page,
2123                                bool *locked, unsigned long *flags)
2124{
2125        struct mem_cgroup *memcg;
2126        struct page_cgroup *pc;
2127
2128        pc = lookup_page_cgroup(page);
2129again:
2130        memcg = pc->mem_cgroup;
2131        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2132                return;
2133        /*
2134         * If this memory cgroup is not under account moving, we don't
2135         * need to take move_lock_mem_cgroup(). Because we already hold
2136         * rcu_read_lock(), any calls to move_account will be delayed until
2137         * rcu_read_unlock() if mem_cgroup_stolen() == true.
2138         */
2139        if (!mem_cgroup_stolen(memcg))
2140                return;
2141
2142        move_lock_mem_cgroup(memcg, flags);
2143        if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2144                move_unlock_mem_cgroup(memcg, flags);
2145                goto again;
2146        }
2147        *locked = true;
2148}
2149
2150void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2151{
2152        struct page_cgroup *pc = lookup_page_cgroup(page);
2153
2154        /*
2155         * It's guaranteed that pc->mem_cgroup never changes while
2156         * lock is held because a routine modifies pc->mem_cgroup
2157         * should take move_lock_mem_cgroup().
2158         */
2159        move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2160}
2161
2162void mem_cgroup_update_page_stat(struct page *page,
2163                                 enum mem_cgroup_page_stat_item idx, int val)
2164{
2165        struct mem_cgroup *memcg;
2166        struct page_cgroup *pc = lookup_page_cgroup(page);
2167        unsigned long uninitialized_var(flags);
2168
2169        if (mem_cgroup_disabled())
2170                return;
2171
2172        memcg = pc->mem_cgroup;
2173        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2174                return;
2175
2176        switch (idx) {
2177        case MEMCG_NR_FILE_MAPPED:
2178                idx = MEM_CGROUP_STAT_FILE_MAPPED;
2179                break;
2180        default:
2181                BUG();
2182        }
2183
2184        this_cpu_add(memcg->stat->count[idx], val);
2185}
2186
2187/*
2188 * size of first charge trial. "32" comes from vmscan.c's magic value.
2189 * TODO: maybe necessary to use big numbers in big irons.
2190 */
2191#define CHARGE_BATCH    32U
2192struct memcg_stock_pcp {
2193        struct mem_cgroup *cached; /* this never be root cgroup */
2194        unsigned int nr_pages;
2195        struct work_struct work;
2196        unsigned long flags;
2197#define FLUSHING_CACHED_CHARGE  0
2198};
2199static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2200static DEFINE_MUTEX(percpu_charge_mutex);
2201
2202/**
2203 * consume_stock: Try to consume stocked charge on this cpu.
2204 * @memcg: memcg to consume from.
2205 * @nr_pages: how many pages to charge.
2206 *
2207 * The charges will only happen if @memcg matches the current cpu's memcg
2208 * stock, and at least @nr_pages are available in that stock.  Failure to
2209 * service an allocation will refill the stock.
2210 *
2211 * returns true if successful, false otherwise.
2212 */
2213static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2214{
2215        struct memcg_stock_pcp *stock;
2216        bool ret = true;
2217
2218        if (nr_pages > CHARGE_BATCH)
2219                return false;
2220
2221        stock = &get_cpu_var(memcg_stock);
2222        if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2223                stock->nr_pages -= nr_pages;
2224        else /* need to call res_counter_charge */
2225                ret = false;
2226        put_cpu_var(memcg_stock);
2227        return ret;
2228}
2229
2230/*
2231 * Returns stocks cached in percpu to res_counter and reset cached information.
2232 */
2233static void drain_stock(struct memcg_stock_pcp *stock)
2234{
2235        struct mem_cgroup *old = stock->cached;
2236
2237        if (stock->nr_pages) {
2238                unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2239
2240                res_counter_uncharge(&old->res, bytes);
2241                if (do_swap_account)
2242                        res_counter_uncharge(&old->memsw, bytes);
2243                stock->nr_pages = 0;
2244        }
2245        stock->cached = NULL;
2246}
2247
2248/*
2249 * This must be called under preempt disabled or must be called by
2250 * a thread which is pinned to local cpu.
2251 */
2252static void drain_local_stock(struct work_struct *dummy)
2253{
2254        struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2255        drain_stock(stock);
2256        clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2257}
2258
2259/*
2260 * Cache charges(val) which is from res_counter, to local per_cpu area.
2261 * This will be consumed by consume_stock() function, later.
2262 */
2263static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2264{
2265        struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2266
2267        if (stock->cached != memcg) { /* reset if necessary */
2268                drain_stock(stock);
2269                stock->cached = memcg;
2270        }
2271        stock->nr_pages += nr_pages;
2272        put_cpu_var(memcg_stock);
2273}
2274
2275/*
2276 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2277 * of the hierarchy under it. sync flag says whether we should block
2278 * until the work is done.
2279 */
2280static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2281{
2282        int cpu, curcpu;
2283
2284        /* Notify other cpus that system-wide "drain" is running */
2285        get_online_cpus();
2286        curcpu = get_cpu();
2287        for_each_online_cpu(cpu) {
2288                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2289                struct mem_cgroup *memcg;
2290
2291                memcg = stock->cached;
2292                if (!memcg || !stock->nr_pages)
2293                        continue;
2294                if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2295                        continue;
2296                if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2297                        if (cpu == curcpu)
2298                                drain_local_stock(&stock->work);
2299                        else
2300                                schedule_work_on(cpu, &stock->work);
2301                }
2302        }
2303        put_cpu();
2304
2305        if (!sync)
2306                goto out;
2307
2308        for_each_online_cpu(cpu) {
2309                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2310                if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2311                        flush_work(&stock->work);
2312        }
2313out:
2314        put_online_cpus();
2315}
2316
2317/*
2318 * Tries to drain stocked charges in other cpus. This function is asynchronous
2319 * and just put a work per cpu for draining localy on each cpu. Caller can
2320 * expects some charges will be back to res_counter later but cannot wait for
2321 * it.
2322 */
2323static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2324{
2325        /*
2326         * If someone calls draining, avoid adding more kworker runs.
2327         */
2328        if (!mutex_trylock(&percpu_charge_mutex))
2329                return;
2330        drain_all_stock(root_memcg, false);
2331        mutex_unlock(&percpu_charge_mutex);
2332}
2333
2334/* This is a synchronous drain interface. */
2335static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2336{
2337        /* called when force_empty is called */
2338        mutex_lock(&percpu_charge_mutex);
2339        drain_all_stock(root_memcg, true);
2340        mutex_unlock(&percpu_charge_mutex);
2341}
2342
2343/*
2344 * This function drains percpu counter value from DEAD cpu and
2345 * move it to local cpu. Note that this function can be preempted.
2346 */
2347static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2348{
2349        int i;
2350
2351        spin_lock(&memcg->pcp_counter_lock);
2352        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2353                long x = per_cpu(memcg->stat->count[i], cpu);
2354
2355                per_cpu(memcg->stat->count[i], cpu) = 0;
2356                memcg->nocpu_base.count[i] += x;
2357        }
2358        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2359                unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2360
2361                per_cpu(memcg->stat->events[i], cpu) = 0;
2362                memcg->nocpu_base.events[i] += x;
2363        }
2364        spin_unlock(&memcg->pcp_counter_lock);
2365}
2366
2367static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2368                                        unsigned long action,
2369                                        void *hcpu)
2370{
2371        int cpu = (unsigned long)hcpu;
2372        struct memcg_stock_pcp *stock;
2373        struct mem_cgroup *iter;
2374
2375        if (action == CPU_ONLINE)
2376                return NOTIFY_OK;
2377
2378        if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2379                return NOTIFY_OK;
2380
2381        for_each_mem_cgroup(iter)
2382                mem_cgroup_drain_pcp_counter(iter, cpu);
2383
2384        stock = &per_cpu(memcg_stock, cpu);
2385        drain_stock(stock);
2386        return NOTIFY_OK;
2387}
2388
2389
2390/* See __mem_cgroup_try_charge() for details */
2391enum {
2392        CHARGE_OK,              /* success */
2393        CHARGE_RETRY,           /* need to retry but retry is not bad */
2394        CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2395        CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2396        CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2397};
2398
2399static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2400                                unsigned int nr_pages, unsigned int min_pages,
2401                                bool oom_check)
2402{
2403        unsigned long csize = nr_pages * PAGE_SIZE;
2404        struct mem_cgroup *mem_over_limit;
2405        struct res_counter *fail_res;
2406        unsigned long flags = 0;
2407        int ret;
2408
2409        ret = res_counter_charge(&memcg->res, csize, &fail_res);
2410
2411        if (likely(!ret)) {
2412                if (!do_swap_account)
2413                        return CHARGE_OK;
2414                ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2415                if (likely(!ret))
2416                        return CHARGE_OK;
2417
2418                res_counter_uncharge(&memcg->res, csize);
2419                mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2420                flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2421        } else
2422                mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2423        /*
2424         * Never reclaim on behalf of optional batching, retry with a
2425         * single page instead.
2426         */
2427        if (nr_pages > min_pages)
2428                return CHARGE_RETRY;
2429
2430        if (!(gfp_mask & __GFP_WAIT))
2431                return CHARGE_WOULDBLOCK;
2432
2433        if (gfp_mask & __GFP_NORETRY)
2434                return CHARGE_NOMEM;
2435
2436        ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2437        if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2438                return CHARGE_RETRY;
2439        /*
2440         * Even though the limit is exceeded at this point, reclaim
2441         * may have been able to free some pages.  Retry the charge
2442         * before killing the task.
2443         *
2444         * Only for regular pages, though: huge pages are rather
2445         * unlikely to succeed so close to the limit, and we fall back
2446         * to regular pages anyway in case of failure.
2447         */
2448        if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2449                return CHARGE_RETRY;
2450
2451        /*
2452         * At task move, charge accounts can be doubly counted. So, it's
2453         * better to wait until the end of task_move if something is going on.
2454         */
2455        if (mem_cgroup_wait_acct_move(mem_over_limit))
2456                return CHARGE_RETRY;
2457
2458        /* If we don't need to call oom-killer at el, return immediately */
2459        if (!oom_check)
2460                return CHARGE_NOMEM;
2461        /* check OOM */
2462        if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2463                return CHARGE_OOM_DIE;
2464
2465        return CHARGE_RETRY;
2466}
2467
2468/*
2469 * __mem_cgroup_try_charge() does
2470 * 1. detect memcg to be charged against from passed *mm and *ptr,
2471 * 2. update res_counter
2472 * 3. call memory reclaim if necessary.
2473 *
2474 * In some special case, if the task is fatal, fatal_signal_pending() or
2475 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2476 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2477 * as possible without any hazards. 2: all pages should have a valid
2478 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2479 * pointer, that is treated as a charge to root_mem_cgroup.
2480 *
2481 * So __mem_cgroup_try_charge() will return
2482 *  0       ...  on success, filling *ptr with a valid memcg pointer.
2483 *  -ENOMEM ...  charge failure because of resource limits.
2484 *  -EINTR  ...  if thread is fatal. *ptr is filled with root_mem_cgroup.
2485 *
2486 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2487 * the oom-killer can be invoked.
2488 */
2489static int __mem_cgroup_try_charge(struct mm_struct *mm,
2490                                   gfp_t gfp_mask,
2491                                   unsigned int nr_pages,
2492                                   struct mem_cgroup **ptr,
2493                                   bool oom)
2494{
2495        unsigned int batch = max(CHARGE_BATCH, nr_pages);
2496        int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2497        struct mem_cgroup *memcg = NULL;
2498        int ret;
2499
2500        /*
2501         * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2502         * in system level. So, allow to go ahead dying process in addition to
2503         * MEMDIE process.
2504         */
2505        if (unlikely(test_thread_flag(TIF_MEMDIE)
2506                     || fatal_signal_pending(current)))
2507                goto bypass;
2508
2509        /*
2510         * We always charge the cgroup the mm_struct belongs to.
2511         * The mm_struct's mem_cgroup changes on task migration if the
2512         * thread group leader migrates. It's possible that mm is not
2513         * set, if so charge the root memcg (happens for pagecache usage).
2514         */
2515        if (!*ptr && !mm)
2516                *ptr = root_mem_cgroup;
2517again:
2518        if (*ptr) { /* css should be a valid one */
2519                memcg = *ptr;
2520                if (mem_cgroup_is_root(memcg))
2521                        goto done;
2522                if (consume_stock(memcg, nr_pages))
2523                        goto done;
2524                css_get(&memcg->css);
2525        } else {
2526                struct task_struct *p;
2527
2528                rcu_read_lock();
2529                p = rcu_dereference(mm->owner);
2530                /*
2531                 * Because we don't have task_lock(), "p" can exit.
2532                 * In that case, "memcg" can point to root or p can be NULL with
2533                 * race with swapoff. Then, we have small risk of mis-accouning.
2534                 * But such kind of mis-account by race always happens because
2535                 * we don't have cgroup_mutex(). It's overkill and we allo that
2536                 * small race, here.
2537                 * (*) swapoff at el will charge against mm-struct not against
2538                 * task-struct. So, mm->owner can be NULL.
2539                 */
2540                memcg = mem_cgroup_from_task(p);
2541                if (!memcg)
2542                        memcg = root_mem_cgroup;
2543                if (mem_cgroup_is_root(memcg)) {
2544                        rcu_read_unlock();
2545                        goto done;
2546                }
2547                if (consume_stock(memcg, nr_pages)) {
2548                        /*
2549                         * It seems dagerous to access memcg without css_get().
2550                         * But considering how consume_stok works, it's not
2551                         * necessary. If consume_stock success, some charges
2552                         * from this memcg are cached on this cpu. So, we
2553                         * don't need to call css_get()/css_tryget() before
2554                         * calling consume_stock().
2555                         */
2556                        rcu_read_unlock();
2557                        goto done;
2558                }
2559                /* after here, we may be blocked. we need to get refcnt */
2560                if (!css_tryget(&memcg->css)) {
2561                        rcu_read_unlock();
2562                        goto again;
2563                }
2564                rcu_read_unlock();
2565        }
2566
2567        do {
2568                bool oom_check;
2569
2570                /* If killed, bypass charge */
2571                if (fatal_signal_pending(current)) {
2572                        css_put(&memcg->css);
2573                        goto bypass;
2574                }
2575
2576                oom_check = false;
2577                if (oom && !nr_oom_retries) {
2578                        oom_check = true;
2579                        nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2580                }
2581
2582                ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2583                    oom_check);
2584                switch (ret) {
2585                case CHARGE_OK:
2586                        break;
2587                case CHARGE_RETRY: /* not in OOM situation but retry */
2588                        batch = nr_pages;
2589                        css_put(&memcg->css);
2590                        memcg = NULL;
2591                        goto again;
2592                case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2593                        css_put(&memcg->css);
2594                        goto nomem;
2595                case CHARGE_NOMEM: /* OOM routine works */
2596                        if (!oom) {
2597                                css_put(&memcg->css);
2598                                goto nomem;
2599                        }
2600                        /* If oom, we never return -ENOMEM */
2601                        nr_oom_retries--;
2602                        break;
2603                case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2604                        css_put(&memcg->css);
2605                        goto bypass;
2606                }
2607        } while (ret != CHARGE_OK);
2608
2609        if (batch > nr_pages)
2610                refill_stock(memcg, batch - nr_pages);
2611        css_put(&memcg->css);
2612done:
2613        *ptr = memcg;
2614        return 0;
2615nomem:
2616        *ptr = NULL;
2617        return -ENOMEM;
2618bypass:
2619        *ptr = root_mem_cgroup;
2620        return -EINTR;
2621}
2622
2623/*
2624 * Somemtimes we have to undo a charge we got by try_charge().
2625 * This function is for that and do uncharge, put css's refcnt.
2626 * gotten by try_charge().
2627 */
2628static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2629                                       unsigned int nr_pages)
2630{
2631        if (!mem_cgroup_is_root(memcg)) {
2632                unsigned long bytes = nr_pages * PAGE_SIZE;
2633
2634                res_counter_uncharge(&memcg->res, bytes);
2635                if (do_swap_account)
2636                        res_counter_uncharge(&memcg->memsw, bytes);
2637        }
2638}
2639
2640/*
2641 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2642 * This is useful when moving usage to parent cgroup.
2643 */
2644static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2645                                        unsigned int nr_pages)
2646{
2647        unsigned long bytes = nr_pages * PAGE_SIZE;
2648
2649        if (mem_cgroup_is_root(memcg))
2650                return;
2651
2652        res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2653        if (do_swap_account)
2654                res_counter_uncharge_until(&memcg->memsw,
2655                                                memcg->memsw.parent, bytes);
2656}
2657
2658/*
2659 * A helper function to get mem_cgroup from ID. must be called under
2660 * rcu_read_lock().  The caller is responsible for calling css_tryget if
2661 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2662 * called against removed memcg.)
2663 */
2664static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2665{
2666        struct cgroup_subsys_state *css;
2667
2668        /* ID 0 is unused ID */
2669        if (!id)
2670                return NULL;
2671        css = css_lookup(&mem_cgroup_subsys, id);
2672        if (!css)
2673                return NULL;
2674        return mem_cgroup_from_css(css);
2675}
2676
2677struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2678{
2679        struct mem_cgroup *memcg = NULL;
2680        struct page_cgroup *pc;
2681        unsigned short id;
2682        swp_entry_t ent;
2683
2684        VM_BUG_ON(!PageLocked(page));
2685
2686        pc = lookup_page_cgroup(page);
2687        lock_page_cgroup(pc);
2688        if (PageCgroupUsed(pc)) {
2689                memcg = pc->mem_cgroup;
2690                if (memcg && !css_tryget(&memcg->css))
2691                        memcg = NULL;
2692        } else if (PageSwapCache(page)) {
2693                ent.val = page_private(page);
2694                id = lookup_swap_cgroup_id(ent);
2695                rcu_read_lock();
2696                memcg = mem_cgroup_lookup(id);
2697                if (memcg && !css_tryget(&memcg->css))
2698                        memcg = NULL;
2699                rcu_read_unlock();
2700        }
2701        unlock_page_cgroup(pc);
2702        return memcg;
2703}
2704
2705static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2706                                       struct page *page,
2707                                       unsigned int nr_pages,
2708                                       enum charge_type ctype,
2709                                       bool lrucare)
2710{
2711        struct page_cgroup *pc = lookup_page_cgroup(page);
2712        struct zone *uninitialized_var(zone);
2713        struct lruvec *lruvec;
2714        bool was_on_lru = false;
2715        bool anon;
2716
2717        lock_page_cgroup(pc);
2718        VM_BUG_ON(PageCgroupUsed(pc));
2719        /*
2720         * we don't need page_cgroup_lock about tail pages, becase they are not
2721         * accessed by any other context at this point.
2722         */
2723
2724        /*
2725         * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2726         * may already be on some other mem_cgroup's LRU.  Take care of it.
2727         */
2728        if (lrucare) {
2729                zone = page_zone(page);
2730                spin_lock_irq(&zone->lru_lock);
2731                if (PageLRU(page)) {
2732                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2733                        ClearPageLRU(page);
2734                        del_page_from_lru_list(page, lruvec, page_lru(page));
2735                        was_on_lru = true;
2736                }
2737        }
2738
2739        pc->mem_cgroup = memcg;
2740        /*
2741         * We access a page_cgroup asynchronously without lock_page_cgroup().
2742         * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2743         * is accessed after testing USED bit. To make pc->mem_cgroup visible
2744         * before USED bit, we need memory barrier here.
2745         * See mem_cgroup_add_lru_list(), etc.
2746         */
2747        smp_wmb();
2748        SetPageCgroupUsed(pc);
2749
2750        if (lrucare) {
2751                if (was_on_lru) {
2752                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2753                        VM_BUG_ON(PageLRU(page));
2754                        SetPageLRU(page);
2755                        add_page_to_lru_list(page, lruvec, page_lru(page));
2756                }
2757                spin_unlock_irq(&zone->lru_lock);
2758        }
2759
2760        if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2761                anon = true;
2762        else
2763                anon = false;
2764
2765        mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2766        unlock_page_cgroup(pc);
2767
2768        /*
2769         * "charge_statistics" updated event counter. Then, check it.
2770         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2771         * if they exceeds softlimit.
2772         */
2773        memcg_check_events(memcg, page);
2774}
2775
2776static DEFINE_MUTEX(set_limit_mutex);
2777
2778#ifdef CONFIG_MEMCG_KMEM
2779static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2780{
2781        return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2782                (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2783}
2784
2785/*
2786 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2787 * in the memcg_cache_params struct.
2788 */
2789static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2790{
2791        struct kmem_cache *cachep;
2792
2793        VM_BUG_ON(p->is_root_cache);
2794        cachep = p->root_cache;
2795        return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2796}
2797
2798#ifdef CONFIG_SLABINFO
2799static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2800                                        struct seq_file *m)
2801{
2802        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2803        struct memcg_cache_params *params;
2804
2805        if (!memcg_can_account_kmem(memcg))
2806                return -EIO;
2807
2808        print_slabinfo_header(m);
2809
2810        mutex_lock(&memcg->slab_caches_mutex);
2811        list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2812                cache_show(memcg_params_to_cache(params), m);
2813        mutex_unlock(&memcg->slab_caches_mutex);
2814
2815        return 0;
2816}
2817#endif
2818
2819static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2820{
2821        struct res_counter *fail_res;
2822        struct mem_cgroup *_memcg;
2823        int ret = 0;
2824        bool may_oom;
2825
2826        ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2827        if (ret)
2828                return ret;
2829
2830        /*
2831         * Conditions under which we can wait for the oom_killer. Those are
2832         * the same conditions tested by the core page allocator
2833         */
2834        may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2835
2836        _memcg = memcg;
2837        ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2838                                      &_memcg, may_oom);
2839
2840        if (ret == -EINTR)  {
2841                /*
2842                 * __mem_cgroup_try_charge() chosed to bypass to root due to
2843                 * OOM kill or fatal signal.  Since our only options are to
2844                 * either fail the allocation or charge it to this cgroup, do
2845                 * it as a temporary condition. But we can't fail. From a
2846                 * kmem/slab perspective, the cache has already been selected,
2847                 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2848                 * our minds.
2849                 *
2850                 * This condition will only trigger if the task entered
2851                 * memcg_charge_kmem in a sane state, but was OOM-killed during
2852                 * __mem_cgroup_try_charge() above. Tasks that were already
2853                 * dying when the allocation triggers should have been already
2854                 * directed to the root cgroup in memcontrol.h
2855                 */
2856                res_counter_charge_nofail(&memcg->res, size, &fail_res);
2857                if (do_swap_account)
2858                        res_counter_charge_nofail(&memcg->memsw, size,
2859                                                  &fail_res);
2860                ret = 0;
2861        } else if (ret)
2862                res_counter_uncharge(&memcg->kmem, size);
2863
2864        return ret;
2865}
2866
2867static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2868{
2869        res_counter_uncharge(&memcg->res, size);
2870        if (do_swap_account)
2871                res_counter_uncharge(&memcg->memsw, size);
2872
2873        /* Not down to 0 */
2874        if (res_counter_uncharge(&memcg->kmem, size))
2875                return;
2876
2877        if (memcg_kmem_test_and_clear_dead(memcg))
2878                mem_cgroup_put(memcg);
2879}
2880
2881void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2882{
2883        if (!memcg)
2884                return;
2885
2886        mutex_lock(&memcg->slab_caches_mutex);
2887        list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2888        mutex_unlock(&memcg->slab_caches_mutex);
2889}
2890
2891/*
2892 * helper for acessing a memcg's index. It will be used as an index in the
2893 * child cache array in kmem_cache, and also to derive its name. This function
2894 * will return -1 when this is not a kmem-limited memcg.
2895 */
2896int memcg_cache_id(struct mem_cgroup *memcg)
2897{
2898        return memcg ? memcg->kmemcg_id : -1;
2899}
2900
2901/*
2902 * This ends up being protected by the set_limit mutex, during normal
2903 * operation, because that is its main call site.
2904 *
2905 * But when we create a new cache, we can call this as well if its parent
2906 * is kmem-limited. That will have to hold set_limit_mutex as well.
2907 */
2908int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2909{
2910        int num, ret;
2911
2912        num = ida_simple_get(&kmem_limited_groups,
2913                                0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2914        if (num < 0)
2915                return num;
2916        /*
2917         * After this point, kmem_accounted (that we test atomically in
2918         * the beginning of this conditional), is no longer 0. This
2919         * guarantees only one process will set the following boolean
2920         * to true. We don't need test_and_set because we're protected
2921         * by the set_limit_mutex anyway.
2922         */
2923        memcg_kmem_set_activated(memcg);
2924
2925        ret = memcg_update_all_caches(num+1);
2926        if (ret) {
2927                ida_simple_remove(&kmem_limited_groups, num);
2928                memcg_kmem_clear_activated(memcg);
2929                return ret;
2930        }
2931
2932        memcg->kmemcg_id = num;
2933        INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2934        mutex_init(&memcg->slab_caches_mutex);
2935        return 0;
2936}
2937
2938static size_t memcg_caches_array_size(int num_groups)
2939{
2940        ssize_t size;
2941        if (num_groups <= 0)
2942                return 0;
2943
2944        size = 2 * num_groups;
2945        if (size < MEMCG_CACHES_MIN_SIZE)
2946                size = MEMCG_CACHES_MIN_SIZE;
2947        else if (size > MEMCG_CACHES_MAX_SIZE)
2948                size = MEMCG_CACHES_MAX_SIZE;
2949
2950        return size;
2951}
2952
2953/*
2954 * We should update the current array size iff all caches updates succeed. This
2955 * can only be done from the slab side. The slab mutex needs to be held when
2956 * calling this.
2957 */
2958void memcg_update_array_size(int num)
2959{
2960        if (num > memcg_limited_groups_array_size)
2961                memcg_limited_groups_array_size = memcg_caches_array_size(num);
2962}
2963
2964int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2965{
2966        struct memcg_cache_params *cur_params = s->memcg_params;
2967
2968        VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2969
2970        if (num_groups > memcg_limited_groups_array_size) {
2971                int i;
2972                ssize_t size = memcg_caches_array_size(num_groups);
2973
2974                size *= sizeof(void *);
2975                size += sizeof(struct memcg_cache_params);
2976
2977                s->memcg_params = kzalloc(size, GFP_KERNEL);
2978                if (!s->memcg_params) {
2979                        s->memcg_params = cur_params;
2980                        return -ENOMEM;
2981                }
2982
2983                s->memcg_params->is_root_cache = true;
2984
2985                /*
2986                 * There is the chance it will be bigger than
2987                 * memcg_limited_groups_array_size, if we failed an allocation
2988                 * in a cache, in which case all caches updated before it, will
2989                 * have a bigger array.
2990                 *
2991                 * But if that is the case, the data after
2992                 * memcg_limited_groups_array_size is certainly unused
2993                 */
2994                for (i = 0; i < memcg_limited_groups_array_size; i++) {
2995                        if (!cur_params->memcg_caches[i])
2996                                continue;
2997                        s->memcg_params->memcg_caches[i] =
2998                                                cur_params->memcg_caches[i];
2999                }
3000
3001                /*
3002                 * Ideally, we would wait until all caches succeed, and only
3003                 * then free the old one. But this is not worth the extra
3004                 * pointer per-cache we'd have to have for this.
3005                 *
3006                 * It is not a big deal if some caches are left with a size
3007                 * bigger than the others. And all updates will reset this
3008                 * anyway.
3009                 */
3010                kfree(cur_params);
3011        }
3012        return 0;
3013}
3014
3015int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3016                         struct kmem_cache *root_cache)
3017{
3018        size_t size = sizeof(struct memcg_cache_params);
3019
3020        if (!memcg_kmem_enabled())
3021                return 0;
3022
3023        if (!memcg)
3024                size += memcg_limited_groups_array_size * sizeof(void *);
3025
3026        s->memcg_params = kzalloc(size, GFP_KERNEL);
3027        if (!s->memcg_params)
3028                return -ENOMEM;
3029
3030        if (memcg) {
3031                s->memcg_params->memcg = memcg;
3032                s->memcg_params->root_cache = root_cache;
3033        } else
3034                s->memcg_params->is_root_cache = true;
3035
3036        return 0;
3037}
3038
3039void memcg_release_cache(struct kmem_cache *s)
3040{
3041        struct kmem_cache *root;
3042        struct mem_cgroup *memcg;
3043        int id;
3044
3045        /*
3046         * This happens, for instance, when a root cache goes away before we
3047         * add any memcg.
3048         */
3049        if (!s->memcg_params)
3050                return;
3051
3052        if (s->memcg_params->is_root_cache)
3053                goto out;
3054
3055        memcg = s->memcg_params->memcg;
3056        id  = memcg_cache_id(memcg);
3057
3058        root = s->memcg_params->root_cache;
3059        root->memcg_params->memcg_caches[id] = NULL;
3060        mem_cgroup_put(memcg);
3061
3062        mutex_lock(&memcg->slab_caches_mutex);
3063        list_del(&s->memcg_params->list);
3064        mutex_unlock(&memcg->slab_caches_mutex);
3065
3066out:
3067        kfree(s->memcg_params);
3068}
3069
3070/*
3071 * During the creation a new cache, we need to disable our accounting mechanism
3072 * altogether. This is true even if we are not creating, but rather just
3073 * enqueing new caches to be created.
3074 *
3075 * This is because that process will trigger allocations; some visible, like
3076 * explicit kmallocs to auxiliary data structures, name strings and internal
3077 * cache structures; some well concealed, like INIT_WORK() that can allocate
3078 * objects during debug.
3079 *
3080 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3081 * to it. This may not be a bounded recursion: since the first cache creation
3082 * failed to complete (waiting on the allocation), we'll just try to create the
3083 * cache again, failing at the same point.
3084 *
3085 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3086 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3087 * inside the following two functions.
3088 */
3089static inline void memcg_stop_kmem_account(void)
3090{
3091        VM_BUG_ON(!current->mm);
3092        current->memcg_kmem_skip_account++;
3093}
3094
3095static inline void memcg_resume_kmem_account(void)
3096{
3097        VM_BUG_ON(!current->mm);
3098        current->memcg_kmem_skip_account--;
3099}
3100
3101static void kmem_cache_destroy_work_func(struct work_struct *w)
3102{
3103        struct kmem_cache *cachep;
3104        struct memcg_cache_params *p;
3105
3106        p = container_of(w, struct memcg_cache_params, destroy);
3107
3108        cachep = memcg_params_to_cache(p);
3109
3110        /*
3111         * If we get down to 0 after shrink, we could delete right away.
3112         * However, memcg_release_pages() already puts us back in the workqueue
3113         * in that case. If we proceed deleting, we'll get a dangling
3114         * reference, and removing the object from the workqueue in that case
3115         * is unnecessary complication. We are not a fast path.
3116         *
3117         * Note that this case is fundamentally different from racing with
3118         * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3119         * kmem_cache_shrink, not only we would be reinserting a dead cache
3120         * into the queue, but doing so from inside the worker racing to
3121         * destroy it.
3122         *
3123         * So if we aren't down to zero, we'll just schedule a worker and try
3124         * again
3125         */
3126        if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3127                kmem_cache_shrink(cachep);
3128                if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3129                        return;
3130        } else
3131                kmem_cache_destroy(cachep);
3132}
3133
3134void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3135{
3136        if (!cachep->memcg_params->dead)
3137                return;
3138
3139        /*
3140         * There are many ways in which we can get here.
3141         *
3142         * We can get to a memory-pressure situation while the delayed work is
3143         * still pending to run. The vmscan shrinkers can then release all
3144         * cache memory and get us to destruction. If this is the case, we'll
3145         * be executed twice, which is a bug (the second time will execute over
3146         * bogus data). In this case, cancelling the work should be fine.
3147         *
3148         * But we can also get here from the worker itself, if
3149         * kmem_cache_shrink is enough to shake all the remaining objects and
3150         * get the page count to 0. In this case, we'll deadlock if we try to
3151         * cancel the work (the worker runs with an internal lock held, which
3152         * is the same lock we would hold for cancel_work_sync().)
3153         *
3154         * Since we can't possibly know who got us here, just refrain from
3155         * running if there is already work pending
3156         */
3157        if (work_pending(&cachep->memcg_params->destroy))
3158                return;
3159        /*
3160         * We have to defer the actual destroying to a workqueue, because
3161         * we might currently be in a context that cannot sleep.
3162         */
3163        schedule_work(&cachep->memcg_params->destroy);
3164}
3165
3166static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3167{
3168        char *name;
3169        struct dentry *dentry;
3170
3171        rcu_read_lock();
3172        dentry = rcu_dereference(memcg->css.cgroup->dentry);
3173        rcu_read_unlock();
3174
3175        BUG_ON(dentry == NULL);
3176
3177        name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3178                         memcg_cache_id(memcg), dentry->d_name.name);
3179
3180        return name;
3181}
3182
3183static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3184                                         struct kmem_cache *s)
3185{
3186        char *name;
3187        struct kmem_cache *new;
3188
3189        name = memcg_cache_name(memcg, s);
3190        if (!name)
3191                return NULL;
3192
3193        new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3194                                      (s->flags & ~SLAB_PANIC), s->ctor, s);
3195
3196        if (new)
3197                new->allocflags |= __GFP_KMEMCG;
3198
3199        kfree(name);
3200        return new;
3201}
3202
3203/*
3204 * This lock protects updaters, not readers. We want readers to be as fast as
3205 * they can, and they will either see NULL or a valid cache value. Our model
3206 * allow them to see NULL, in which case the root memcg will be selected.
3207 *
3208 * We need this lock because multiple allocations to the same cache from a non
3209 * will span more than one worker. Only one of them can create the cache.
3210 */
3211static DEFINE_MUTEX(memcg_cache_mutex);
3212static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3213                                                  struct kmem_cache *cachep)
3214{
3215        struct kmem_cache *new_cachep;
3216        int idx;
3217
3218        BUG_ON(!memcg_can_account_kmem(memcg));
3219
3220        idx = memcg_cache_id(memcg);
3221
3222        mutex_lock(&memcg_cache_mutex);
3223        new_cachep = cachep->memcg_params->memcg_caches[idx];
3224        if (new_cachep)
3225                goto out;
3226
3227        new_cachep = kmem_cache_dup(memcg, cachep);
3228        if (new_cachep == NULL) {
3229                new_cachep = cachep;
3230                goto out;
3231        }
3232
3233        mem_cgroup_get(memcg);
3234        atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3235
3236        cachep->memcg_params->memcg_caches[idx] = new_cachep;
3237        /*
3238         * the readers won't lock, make sure everybody sees the updated value,
3239         * so they won't put stuff in the queue again for no reason
3240         */
3241        wmb();
3242out:
3243        mutex_unlock(&memcg_cache_mutex);
3244        return new_cachep;
3245}
3246
3247void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3248{
3249        struct kmem_cache *c;
3250        int i;
3251
3252        if (!s->memcg_params)
3253                return;
3254        if (!s->memcg_params->is_root_cache)
3255                return;
3256
3257        /*
3258         * If the cache is being destroyed, we trust that there is no one else
3259         * requesting objects from it. Even if there are, the sanity checks in
3260         * kmem_cache_destroy should caught this ill-case.
3261         *
3262         * Still, we don't want anyone else freeing memcg_caches under our
3263         * noses, which can happen if a new memcg comes to life. As usual,
3264         * we'll take the set_limit_mutex to protect ourselves against this.
3265         */
3266        mutex_lock(&set_limit_mutex);
3267        for (i = 0; i < memcg_limited_groups_array_size; i++) {
3268                c = s->memcg_params->memcg_caches[i];
3269                if (!c)
3270                        continue;
3271
3272                /*
3273                 * We will now manually delete the caches, so to avoid races
3274                 * we need to cancel all pending destruction workers and
3275                 * proceed with destruction ourselves.
3276                 *
3277                 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3278                 * and that could spawn the workers again: it is likely that
3279                 * the cache still have active pages until this very moment.
3280                 * This would lead us back to mem_cgroup_destroy_cache.
3281                 *
3282                 * But that will not execute at all if the "dead" flag is not
3283                 * set, so flip it down to guarantee we are in control.
3284                 */
3285                c->memcg_params->dead = false;
3286                cancel_work_sync(&c->memcg_params->destroy);
3287                kmem_cache_destroy(c);
3288        }
3289        mutex_unlock(&set_limit_mutex);
3290}
3291
3292struct create_work {
3293        struct mem_cgroup *memcg;
3294        struct kmem_cache *cachep;
3295        struct work_struct work;
3296};
3297
3298static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3299{
3300        struct kmem_cache *cachep;
3301        struct memcg_cache_params *params;
3302
3303        if (!memcg_kmem_is_active(memcg))
3304                return;
3305
3306        mutex_lock(&memcg->slab_caches_mutex);
3307        list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3308                cachep = memcg_params_to_cache(params);
3309                cachep->memcg_params->dead = true;
3310                INIT_WORK(&cachep->memcg_params->destroy,
3311                                  kmem_cache_destroy_work_func);
3312                schedule_work(&cachep->memcg_params->destroy);
3313        }
3314        mutex_unlock(&memcg->slab_caches_mutex);
3315}
3316
3317static void memcg_create_cache_work_func(struct work_struct *w)
3318{
3319        struct create_work *cw;
3320
3321        cw = container_of(w, struct create_work, work);
3322        memcg_create_kmem_cache(cw->memcg, cw->cachep);
3323        /* Drop the reference gotten when we enqueued. */
3324        css_put(&cw->memcg->css);
3325        kfree(cw);
3326}
3327
3328/*
3329 * Enqueue the creation of a per-memcg kmem_cache.
3330 * Called with rcu_read_lock.
3331 */
3332static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3333                                         struct kmem_cache *cachep)
3334{
3335        struct create_work *cw;
3336
3337        cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3338        if (cw == NULL)
3339                return;
3340
3341        /* The corresponding put will be done in the workqueue. */
3342        if (!css_tryget(&memcg->css)) {
3343                kfree(cw);
3344                return;
3345        }
3346
3347        cw->memcg = memcg;
3348        cw->cachep = cachep;
3349
3350        INIT_WORK(&cw->work, memcg_create_cache_work_func);
3351        schedule_work(&cw->work);
3352}
3353
3354static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3355                                       struct kmem_cache *cachep)
3356{
3357        /*
3358         * We need to stop accounting when we kmalloc, because if the
3359         * corresponding kmalloc cache is not yet created, the first allocation
3360         * in __memcg_create_cache_enqueue will recurse.
3361         *
3362         * However, it is better to enclose the whole function. Depending on
3363         * the debugging options enabled, INIT_WORK(), for instance, can
3364         * trigger an allocation. This too, will make us recurse. Because at
3365         * this point we can't allow ourselves back into memcg_kmem_get_cache,
3366         * the safest choice is to do it like this, wrapping the whole function.
3367         */
3368        memcg_stop_kmem_account();
3369        __memcg_create_cache_enqueue(memcg, cachep);
3370        memcg_resume_kmem_account();
3371}
3372/*
3373 * Return the kmem_cache we're supposed to use for a slab allocation.
3374 * We try to use the current memcg's version of the cache.
3375 *
3376 * If the cache does not exist yet, if we are the first user of it,
3377 * we either create it immediately, if possible, or create it asynchronously
3378 * in a workqueue.
3379 * In the latter case, we will let the current allocation go through with
3380 * the original cache.
3381 *
3382 * Can't be called in interrupt context or from kernel threads.
3383 * This function needs to be called with rcu_read_lock() held.
3384 */
3385struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3386                                          gfp_t gfp)
3387{
3388        struct mem_cgroup *memcg;
3389        int idx;
3390
3391        VM_BUG_ON(!cachep->memcg_params);
3392        VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3393
3394        if (!current->mm || current->memcg_kmem_skip_account)
3395                return cachep;
3396
3397        rcu_read_lock();
3398        memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3399        rcu_read_unlock();
3400
3401        if (!memcg_can_account_kmem(memcg))
3402                return cachep;
3403
3404        idx = memcg_cache_id(memcg);
3405
3406        /*
3407         * barrier to mare sure we're always seeing the up to date value.  The
3408         * code updating memcg_caches will issue a write barrier to match this.
3409         */
3410        read_barrier_depends();
3411        if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3412                /*
3413                 * If we are in a safe context (can wait, and not in interrupt
3414                 * context), we could be be predictable and return right away.
3415                 * This would guarantee that the allocation being performed
3416                 * already belongs in the new cache.
3417                 *
3418                 * However, there are some clashes that can arrive from locking.
3419                 * For instance, because we acquire the slab_mutex while doing
3420                 * kmem_cache_dup, this means no further allocation could happen
3421                 * with the slab_mutex held.
3422                 *
3423                 * Also, because cache creation issue get_online_cpus(), this
3424                 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3425                 * that ends up reversed during cpu hotplug. (cpuset allocates
3426                 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3427                 * better to defer everything.
3428                 */
3429                memcg_create_cache_enqueue(memcg, cachep);
3430                return cachep;
3431        }
3432
3433        return cachep->memcg_params->memcg_caches[idx];
3434}
3435EXPORT_SYMBOL(__memcg_kmem_get_cache);
3436
3437/*
3438 * We need to verify if the allocation against current->mm->owner's memcg is
3439 * possible for the given order. But the page is not allocated yet, so we'll
3440 * need a further commit step to do the final arrangements.
3441 *
3442 * It is possible for the task to switch cgroups in this mean time, so at
3443 * commit time, we can't rely on task conversion any longer.  We'll then use
3444 * the handle argument to return to the caller which cgroup we should commit
3445 * against. We could also return the memcg directly and avoid the pointer
3446 * passing, but a boolean return value gives better semantics considering
3447 * the compiled-out case as well.
3448 *
3449 * Returning true means the allocation is possible.
3450 */
3451bool
3452__memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3453{
3454        struct mem_cgroup *memcg;
3455        int ret;
3456
3457        *_memcg = NULL;
3458        memcg = try_get_mem_cgroup_from_mm(current->mm);
3459
3460        /*
3461         * very rare case described in mem_cgroup_from_task. Unfortunately there
3462         * isn't much we can do without complicating this too much, and it would
3463         * be gfp-dependent anyway. Just let it go
3464         */
3465        if (unlikely(!memcg))
3466                return true;
3467
3468        if (!memcg_can_account_kmem(memcg)) {
3469                css_put(&memcg->css);
3470                return true;
3471        }
3472
3473        ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3474        if (!ret)
3475                *_memcg = memcg;
3476
3477        css_put(&memcg->css);
3478        return (ret == 0);
3479}
3480
3481void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3482                              int order)
3483{
3484        struct page_cgroup *pc;
3485
3486        VM_BUG_ON(mem_cgroup_is_root(memcg));
3487
3488        /* The page allocation failed. Revert */
3489        if (!page) {
3490                memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3491                return;
3492        }
3493
3494        pc = lookup_page_cgroup(page);
3495        lock_page_cgroup(pc);
3496        pc->mem_cgroup = memcg;
3497        SetPageCgroupUsed(pc);
3498        unlock_page_cgroup(pc);
3499}
3500
3501void __memcg_kmem_uncharge_pages(struct page *page, int order)
3502{
3503        struct mem_cgroup *memcg = NULL;
3504        struct page_cgroup *pc;
3505
3506
3507        pc = lookup_page_cgroup(page);
3508        /*
3509         * Fast unlocked return. Theoretically might have changed, have to
3510         * check again after locking.
3511         */
3512        if (!PageCgroupUsed(pc))
3513                return;
3514
3515        lock_page_cgroup(pc);
3516        if (PageCgroupUsed(pc)) {
3517                memcg = pc->mem_cgroup;
3518                ClearPageCgroupUsed(pc);
3519        }
3520        unlock_page_cgroup(pc);
3521
3522        /*
3523         * We trust that only if there is a memcg associated with the page, it
3524         * is a valid allocation
3525         */
3526        if (!memcg)
3527                return;
3528
3529        VM_BUG_ON(mem_cgroup_is_root(memcg));
3530        memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3531}
3532#else
3533static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3534{
3535}
3536#endif /* CONFIG_MEMCG_KMEM */
3537
3538#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3539
3540#define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3541/*
3542 * Because tail pages are not marked as "used", set it. We're under
3543 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3544 * charge/uncharge will be never happen and move_account() is done under
3545 * compound_lock(), so we don't have to take care of races.
3546 */
3547void mem_cgroup_split_huge_fixup(struct page *head)
3548{
3549        struct page_cgroup *head_pc = lookup_page_cgroup(head);
3550        struct page_cgroup *pc;
3551        int i;
3552
3553        if (mem_cgroup_disabled())
3554                return;
3555        for (i = 1; i < HPAGE_PMD_NR; i++) {
3556                pc = head_pc + i;
3557                pc->mem_cgroup = head_pc->mem_cgroup;
3558                smp_wmb();/* see __commit_charge() */
3559                pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3560        }
3561}
3562#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3563
3564/**
3565 * mem_cgroup_move_account - move account of the page
3566 * @page: the page
3567 * @nr_pages: number of regular pages (>1 for huge pages)
3568 * @pc: page_cgroup of the page.
3569 * @from: mem_cgroup which the page is moved from.
3570 * @to: mem_cgroup which the page is moved to. @from != @to.
3571 *
3572 * The caller must confirm following.
3573 * - page is not on LRU (isolate_page() is useful.)
3574 * - compound_lock is held when nr_pages > 1
3575 *
3576 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3577 * from old cgroup.
3578 */
3579static int mem_cgroup_move_account(struct page *page,
3580                                   unsigned int nr_pages,
3581                                   struct page_cgroup *pc,
3582                                   struct mem_cgroup *from,
3583                                   struct mem_cgroup *to)
3584{
3585        unsigned long flags;
3586        int ret;
3587        bool anon = PageAnon(page);
3588
3589        VM_BUG_ON(from == to);
3590        VM_BUG_ON(PageLRU(page));
3591        /*
3592         * The page is isolated from LRU. So, collapse function
3593         * will not handle this page. But page splitting can happen.
3594         * Do this check under compound_page_lock(). The caller should
3595         * hold it.
3596         */
3597        ret = -EBUSY;
3598        if (nr_pages > 1 && !PageTransHuge(page))
3599                goto out;
3600
3601        lock_page_cgroup(pc);
3602
3603        ret = -EINVAL;
3604        if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3605                goto unlock;
3606
3607        move_lock_mem_cgroup(from, &flags);
3608
3609        if (!anon && page_mapped(page)) {
3610                /* Update mapped_file data for mem_cgroup */
3611                preempt_disable();
3612                __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3613                __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3614                preempt_enable();
3615        }
3616        mem_cgroup_charge_statistics(from, anon, -nr_pages);
3617
3618        /* caller should have done css_get */
3619        pc->mem_cgroup = to;
3620        mem_cgroup_charge_statistics(to, anon, nr_pages);
3621        move_unlock_mem_cgroup(from, &flags);
3622        ret = 0;
3623unlock:
3624        unlock_page_cgroup(pc);
3625        /*
3626         * check events
3627         */
3628        memcg_check_events(to, page);
3629        memcg_check_events(from, page);
3630out:
3631        return ret;
3632}
3633
3634/**
3635 * mem_cgroup_move_parent - moves page to the parent group
3636 * @page: the page to move
3637 * @pc: page_cgroup of the page
3638 * @child: page's cgroup
3639 *
3640 * move charges to its parent or the root cgroup if the group has no
3641 * parent (aka use_hierarchy==0).
3642 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3643 * mem_cgroup_move_account fails) the failure is always temporary and
3644 * it signals a race with a page removal/uncharge or migration. In the
3645 * first case the page is on the way out and it will vanish from the LRU
3646 * on the next attempt and the call should be retried later.
3647 * Isolation from the LRU fails only if page has been isolated from
3648 * the LRU since we looked at it and that usually means either global
3649 * reclaim or migration going on. The page will either get back to the
3650 * LRU or vanish.
3651 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3652 * (!PageCgroupUsed) or moved to a different group. The page will
3653 * disappear in the next attempt.
3654 */
3655static int mem_cgroup_move_parent(struct page *page,
3656                                  struct page_cgroup *pc,
3657                                  struct mem_cgroup *child)
3658{
3659        struct mem_cgroup *parent;
3660        unsigned int nr_pages;
3661        unsigned long uninitialized_var(flags);
3662        int ret;
3663
3664        VM_BUG_ON(mem_cgroup_is_root(child));
3665
3666        ret = -EBUSY;
3667        if (!get_page_unless_zero(page))
3668                goto out;
3669        if (isolate_lru_page(page))
3670                goto put;
3671
3672        nr_pages = hpage_nr_pages(page);
3673
3674        parent = parent_mem_cgroup(child);
3675        /*
3676         * If no parent, move charges to root cgroup.
3677         */
3678        if (!parent)
3679                parent = root_mem_cgroup;
3680
3681        if (nr_pages > 1) {
3682                VM_BUG_ON(!PageTransHuge(page));
3683                flags = compound_lock_irqsave(page);
3684        }
3685
3686        ret = mem_cgroup_move_account(page, nr_pages,
3687                                pc, child, parent);
3688        if (!ret)
3689                __mem_cgroup_cancel_local_charge(child, nr_pages);
3690
3691        if (nr_pages > 1)
3692                compound_unlock_irqrestore(page, flags);
3693        putback_lru_page(page);
3694put:
3695        put_page(page);
3696out:
3697        return ret;
3698}
3699
3700/*
3701 * Charge the memory controller for page usage.
3702 * Return
3703 * 0 if the charge was successful
3704 * < 0 if the cgroup is over its limit
3705 */
3706static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3707                                gfp_t gfp_mask, enum charge_type ctype)
3708{
3709        struct mem_cgroup *memcg = NULL;
3710        unsigned int nr_pages = 1;
3711        bool oom = true;
3712        int ret;
3713
3714        if (PageTransHuge(page)) {
3715                nr_pages <<= compound_order(page);
3716                VM_BUG_ON(!PageTransHuge(page));
3717                /*
3718                 * Never OOM-kill a process for a huge page.  The
3719                 * fault handler will fall back to regular pages.
3720                 */
3721                oom = false;
3722        }
3723
3724        ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3725        if (ret == -ENOMEM)
3726                return ret;
3727        __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3728        return 0;
3729}
3730
3731int mem_cgroup_newpage_charge(struct page *page,
3732                              struct mm_struct *mm, gfp_t gfp_mask)
3733{
3734        if (mem_cgroup_disabled())
3735                return 0;
3736        VM_BUG_ON(page_mapped(page));
3737        VM_BUG_ON(page->mapping && !PageAnon(page));
3738        VM_BUG_ON(!mm);
3739        return mem_cgroup_charge_common(page, mm, gfp_mask,
3740                                        MEM_CGROUP_CHARGE_TYPE_ANON);
3741}
3742
3743/*
3744 * While swap-in, try_charge -> commit or cancel, the page is locked.
3745 * And when try_charge() successfully returns, one refcnt to memcg without
3746 * struct page_cgroup is acquired. This refcnt will be consumed by
3747 * "commit()" or removed by "cancel()"
3748 */
3749static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3750                                          struct page *page,
3751                                          gfp_t mask,
3752                                          struct mem_cgroup **memcgp)
3753{
3754        struct mem_cgroup *memcg;
3755        struct page_cgroup *pc;
3756        int ret;
3757
3758        pc = lookup_page_cgroup(page);
3759        /*
3760         * Every swap fault against a single page tries to charge the
3761         * page, bail as early as possible.  shmem_unuse() encounters
3762         * already charged pages, too.  The USED bit is protected by
3763         * the page lock, which serializes swap cache removal, which
3764         * in turn serializes uncharging.
3765         */
3766        if (PageCgroupUsed(pc))
3767                return 0;
3768        if (!do_swap_account)
3769                goto charge_cur_mm;
3770        memcg = try_get_mem_cgroup_from_page(page);
3771        if (!memcg)
3772                goto charge_cur_mm;
3773        *memcgp = memcg;
3774        ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3775        css_put(&memcg->css);
3776        if (ret == -EINTR)
3777                ret = 0;
3778        return ret;
3779charge_cur_mm:
3780        ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3781        if (ret == -EINTR)
3782                ret = 0;
3783        return ret;
3784}
3785
3786int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3787                                 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3788{
3789        *memcgp = NULL;
3790        if (mem_cgroup_disabled())
3791                return 0;
3792        /*
3793         * A racing thread's fault, or swapoff, may have already
3794         * updated the pte, and even removed page from swap cache: in
3795         * those cases unuse_pte()'s pte_same() test will fail; but
3796         * there's also a KSM case which does need to charge the page.
3797         */
3798        if (!PageSwapCache(page)) {
3799                int ret;
3800
3801                ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3802                if (ret == -EINTR)
3803                        ret = 0;
3804                return ret;
3805        }
3806        return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3807}
3808
3809void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3810{
3811        if (mem_cgroup_disabled())
3812                return;
3813        if (!memcg)
3814                return;
3815        __mem_cgroup_cancel_charge(memcg, 1);
3816}
3817
3818static void
3819__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3820                                        enum charge_type ctype)
3821{
3822        if (mem_cgroup_disabled())
3823                return;
3824        if (!memcg)
3825                return;
3826
3827        __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3828        /*
3829         * Now swap is on-memory. This means this page may be
3830         * counted both as mem and swap....double count.
3831         * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3832         * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3833         * may call delete_from_swap_cache() before reach here.
3834         */
3835        if (do_swap_account && PageSwapCache(page)) {
3836                swp_entry_t ent = {.val = page_private(page)};
3837                mem_cgroup_uncharge_swap(ent);
3838        }
3839}
3840
3841void mem_cgroup_commit_charge_swapin(struct page *page,
3842                                     struct mem_cgroup *memcg)
3843{
3844        __mem_cgroup_commit_charge_swapin(page, memcg,
3845                                          MEM_CGROUP_CHARGE_TYPE_ANON);
3846}
3847
3848int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3849                                gfp_t gfp_mask)
3850{
3851        struct mem_cgroup *memcg = NULL;
3852        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3853        int ret;
3854
3855        if (mem_cgroup_disabled())
3856                return 0;
3857        if (PageCompound(page))
3858                return 0;
3859
3860        if (!PageSwapCache(page))
3861                ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3862        else { /* page is swapcache/shmem */
3863                ret = __mem_cgroup_try_charge_swapin(mm, page,
3864                                                     gfp_mask, &memcg);
3865                if (!ret)
3866                        __mem_cgroup_commit_charge_swapin(page, memcg, type);
3867        }
3868        return ret;
3869}
3870
3871static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3872                                   unsigned int nr_pages,
3873                                   const enum charge_type ctype)
3874{
3875        struct memcg_batch_info *batch = NULL;
3876        bool uncharge_memsw = true;
3877
3878        /* If swapout, usage of swap doesn't decrease */
3879        if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3880                uncharge_memsw = false;
3881
3882        batch = &current->memcg_batch;
3883        /*
3884         * In usual, we do css_get() when we remember memcg pointer.
3885         * But in this case, we keep res->usage until end of a series of
3886         * uncharges. Then, it's ok to ignore memcg's refcnt.
3887         */
3888        if (!batch->memcg)
3889                batch->memcg = memcg;
3890        /*
3891         * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3892         * In those cases, all pages freed continuously can be expected to be in
3893         * the same cgroup and we have chance to coalesce uncharges.
3894         * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3895         * because we want to do uncharge as soon as possible.
3896         */
3897
3898        if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3899                goto direct_uncharge;
3900
3901        if (nr_pages > 1)
3902                goto direct_uncharge;
3903
3904        /*
3905         * In typical case, batch->memcg == mem. This means we can
3906         * merge a series of uncharges to an uncharge of res_counter.
3907         * If not, we uncharge res_counter ony by one.
3908         */
3909        if (batch->memcg != memcg)
3910                goto direct_uncharge;
3911        /* remember freed charge and uncharge it later */
3912        batch->nr_pages++;
3913        if (uncharge_memsw)
3914                batch->memsw_nr_pages++;
3915        return;
3916direct_uncharge:
3917        res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3918        if (uncharge_memsw)
3919                res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3920        if (unlikely(batch->memcg != memcg))
3921                memcg_oom_recover(memcg);
3922}
3923
3924/*
3925 * uncharge if !page_mapped(page)
3926 */
3927static struct mem_cgroup *
3928__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3929                             bool end_migration)
3930{
3931        struct mem_cgroup *memcg = NULL;
3932        unsigned int nr_pages = 1;
3933        struct page_cgroup *pc;
3934        bool anon;
3935
3936        if (mem_cgroup_disabled())
3937                return NULL;
3938
3939        VM_BUG_ON(PageSwapCache(page));
3940
3941        if (PageTransHuge(page)) {
3942                nr_pages <<= compound_order(page);
3943                VM_BUG_ON(!PageTransHuge(page));
3944        }
3945        /*
3946         * Check if our page_cgroup is valid
3947         */
3948        pc = lookup_page_cgroup(page);
3949        if (unlikely(!PageCgroupUsed(pc)))
3950                return NULL;
3951
3952        lock_page_cgroup(pc);
3953
3954        memcg = pc->mem_cgroup;
3955
3956        if (!PageCgroupUsed(pc))
3957                goto unlock_out;
3958
3959        anon = PageAnon(page);
3960
3961        switch (ctype) {
3962        case MEM_CGROUP_CHARGE_TYPE_ANON:
3963                /*
3964                 * Generally PageAnon tells if it's the anon statistics to be
3965                 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3966                 * used before page reached the stage of being marked PageAnon.
3967                 */
3968                anon = true;
3969                /* fallthrough */
3970        case MEM_CGROUP_CHARGE_TYPE_DROP:
3971                /* See mem_cgroup_prepare_migration() */
3972                if (page_mapped(page))
3973                        goto unlock_out;
3974                /*
3975                 * Pages under migration may not be uncharged.  But
3976                 * end_migration() /must/ be the one uncharging the
3977                 * unused post-migration page and so it has to call
3978                 * here with the migration bit still set.  See the
3979                 * res_counter handling below.
3980                 */
3981                if (!end_migration && PageCgroupMigration(pc))
3982                        goto unlock_out;
3983                break;
3984        case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3985                if (!PageAnon(page)) {  /* Shared memory */
3986                        if (page->mapping && !page_is_file_cache(page))
3987                                goto unlock_out;
3988                } else if (page_mapped(page)) /* Anon */
3989                                goto unlock_out;
3990                break;
3991        default:
3992                break;
3993        }
3994
3995        mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3996
3997        ClearPageCgroupUsed(pc);
3998        /*
3999         * pc->mem_cgroup is not cleared here. It will be accessed when it's
4000         * freed from LRU. This is safe because uncharged page is expected not
4001         * to be reused (freed soon). Exception is SwapCache, it's handled by
4002         * special functions.
4003         */
4004
4005        unlock_page_cgroup(pc);
4006        /*
4007         * even after unlock, we have memcg->res.usage here and this memcg
4008         * will never be freed.
4009         */
4010        memcg_check_events(memcg, page);
4011        if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4012                mem_cgroup_swap_statistics(memcg, true);
4013                mem_cgroup_get(memcg);
4014        }
4015        /*
4016         * Migration does not charge the res_counter for the
4017         * replacement page, so leave it alone when phasing out the
4018         * page that is unused after the migration.
4019         */
4020        if (!end_migration && !mem_cgroup_is_root(memcg))
4021                mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4022
4023        return memcg;
4024
4025unlock_out:
4026        unlock_page_cgroup(pc);
4027        return NULL;
4028}
4029
4030void mem_cgroup_uncharge_page(struct page *page)
4031{
4032        /* early check. */
4033        if (page_mapped(page))
4034                return;
4035        VM_BUG_ON(page->mapping && !PageAnon(page));
4036        if (PageSwapCache(page))
4037                return;
4038        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4039}
4040
4041void mem_cgroup_uncharge_cache_page(struct page *page)
4042{
4043        VM_BUG_ON(page_mapped(page));
4044        VM_BUG_ON(page->mapping);
4045        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4046}
4047
4048/*
4049 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4050 * In that cases, pages are freed continuously and we can expect pages
4051 * are in the same memcg. All these calls itself limits the number of
4052 * pages freed at once, then uncharge_start/end() is called properly.
4053 * This may be called prural(2) times in a context,
4054 */
4055
4056void mem_cgroup_uncharge_start(void)
4057{
4058        current->memcg_batch.do_batch++;
4059        /* We can do nest. */
4060        if (current->memcg_batch.do_batch == 1) {
4061                current->memcg_batch.memcg = NULL;
4062                current->memcg_batch.nr_pages = 0;
4063                current->memcg_batch.memsw_nr_pages = 0;
4064        }
4065}
4066
4067void mem_cgroup_uncharge_end(void)
4068{
4069        struct memcg_batch_info *batch = &current->memcg_batch;
4070
4071        if (!batch->do_batch)
4072                return;
4073
4074        batch->do_batch--;
4075        if (batch->do_batch) /* If stacked, do nothing. */
4076                return;
4077
4078        if (!batch->memcg)
4079                return;
4080        /*
4081         * This "batch->memcg" is valid without any css_get/put etc...
4082         * bacause we hide charges behind us.
4083         */
4084        if (batch->nr_pages)
4085                res_counter_uncharge(&batch->memcg->res,
4086                                     batch->nr_pages * PAGE_SIZE);
4087        if (batch->memsw_nr_pages)
4088                res_counter_uncharge(&batch->memcg->memsw,
4089                                     batch->memsw_nr_pages * PAGE_SIZE);
4090        memcg_oom_recover(batch->memcg);
4091        /* forget this pointer (for sanity check) */
4092        batch->memcg = NULL;
4093}
4094
4095#ifdef CONFIG_SWAP
4096/*
4097 * called after __delete_from_swap_cache() and drop "page" account.
4098 * memcg information is recorded to swap_cgroup of "ent"
4099 */
4100void
4101mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4102{
4103        struct mem_cgroup *memcg;
4104        int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4105
4106        if (!swapout) /* this was a swap cache but the swap is unused ! */
4107                ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4108
4109        memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4110
4111        /*
4112         * record memcg information,  if swapout && memcg != NULL,
4113         * mem_cgroup_get() was called in uncharge().
4114         */
4115        if (do_swap_account && swapout && memcg)
4116                swap_cgroup_record(ent, css_id(&memcg->css));
4117}
4118#endif
4119
4120#ifdef CONFIG_MEMCG_SWAP
4121/*
4122 * called from swap_entry_free(). remove record in swap_cgroup and
4123 * uncharge "memsw" account.
4124 */
4125void mem_cgroup_uncharge_swap(swp_entry_t ent)
4126{
4127        struct mem_cgroup *memcg;
4128        unsigned short id;
4129
4130        if (!do_swap_account)
4131                return;
4132
4133        id = swap_cgroup_record(ent, 0);
4134        rcu_read_lock();
4135        memcg = mem_cgroup_lookup(id);
4136        if (memcg) {
4137                /*
4138                 * We uncharge this because swap is freed.
4139                 * This memcg can be obsolete one. We avoid calling css_tryget
4140                 */
4141                if (!mem_cgroup_is_root(memcg))
4142                        res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4143                mem_cgroup_swap_statistics(memcg, false);
4144                mem_cgroup_put(memcg);
4145        }
4146        rcu_read_unlock();
4147}
4148
4149/**
4150 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4151 * @entry: swap entry to be moved
4152 * @from:  mem_cgroup which the entry is moved from
4153 * @to:  mem_cgroup which the entry is moved to
4154 *
4155 * It succeeds only when the swap_cgroup's record for this entry is the same
4156 * as the mem_cgroup's id of @from.
4157 *
4158 * Returns 0 on success, -EINVAL on failure.
4159 *
4160 * The caller must have charged to @to, IOW, called res_counter_charge() about
4161 * both res and memsw, and called css_get().
4162 */
4163static int mem_cgroup_move_swap_account(swp_entry_t entry,
4164                                struct mem_cgroup *from, struct mem_cgroup *to)
4165{
4166        unsigned short old_id, new_id;
4167
4168        old_id = css_id(&from->css);
4169        new_id = css_id(&to->css);
4170
4171        if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4172                mem_cgroup_swap_statistics(from, false);
4173                mem_cgroup_swap_statistics(to, true);
4174                /*
4175                 * This function is only called from task migration context now.
4176                 * It postpones res_counter and refcount handling till the end
4177                 * of task migration(mem_cgroup_clear_mc()) for performance
4178                 * improvement. But we cannot postpone mem_cgroup_get(to)
4179                 * because if the process that has been moved to @to does
4180                 * swap-in, the refcount of @to might be decreased to 0.
4181                 */
4182                mem_cgroup_get(to);
4183                return 0;
4184        }
4185        return -EINVAL;
4186}
4187#else
4188static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4189                                struct mem_cgroup *from, struct mem_cgroup *to)
4190{
4191        return -EINVAL;
4192}
4193#endif
4194
4195/*
4196 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4197 * page belongs to.
4198 */
4199void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4200                                  struct mem_cgroup **memcgp)
4201{
4202        struct mem_cgroup *memcg = NULL;
4203        unsigned int nr_pages = 1;
4204        struct page_cgroup *pc;
4205        enum charge_type ctype;
4206
4207        *memcgp = NULL;
4208
4209        if (mem_cgroup_disabled())
4210                return;
4211
4212        if (PageTransHuge(page))
4213                nr_pages <<= compound_order(page);
4214
4215        pc = lookup_page_cgroup(page);
4216        lock_page_cgroup(pc);
4217        if (PageCgroupUsed(pc)) {
4218                memcg = pc->mem_cgroup;
4219                css_get(&memcg->css);
4220                /*
4221                 * At migrating an anonymous page, its mapcount goes down
4222                 * to 0 and uncharge() will be called. But, even if it's fully
4223                 * unmapped, migration may fail and this page has to be
4224                 * charged again. We set MIGRATION flag here and delay uncharge
4225                 * until end_migration() is called
4226                 *
4227                 * Corner Case Thinking
4228                 * A)
4229                 * When the old page was mapped as Anon and it's unmap-and-freed
4230                 * while migration was ongoing.
4231                 * If unmap finds the old page, uncharge() of it will be delayed
4232                 * until end_migration(). If unmap finds a new page, it's
4233                 * uncharged when it make mapcount to be 1->0. If unmap code
4234                 * finds swap_migration_entry, the new page will not be mapped
4235                 * and end_migration() will find it(mapcount==0).
4236                 *
4237                 * B)
4238                 * When the old page was mapped but migraion fails, the kernel
4239                 * remaps it. A charge for it is kept by MIGRATION flag even
4240                 * if mapcount goes down to 0. We can do remap successfully
4241                 * without charging it again.
4242                 *
4243                 * C)
4244                 * The "old" page is under lock_page() until the end of
4245                 * migration, so, the old page itself will not be swapped-out.
4246                 * If the new page is swapped out before end_migraton, our
4247                 * hook to usual swap-out path will catch the event.
4248                 */
4249                if (PageAnon(page))
4250                        SetPageCgroupMigration(pc);
4251        }
4252        unlock_page_cgroup(pc);
4253        /*
4254         * If the page is not charged at this point,
4255         * we return here.
4256         */
4257        if (!memcg)
4258                return;
4259
4260        *memcgp = memcg;
4261        /*
4262         * We charge new page before it's used/mapped. So, even if unlock_page()
4263         * is called before end_migration, we can catch all events on this new
4264         * page. In the case new page is migrated but not remapped, new page's
4265         * mapcount will be finally 0 and we call uncharge in end_migration().
4266         */
4267        if (PageAnon(page))
4268                ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4269        else
4270                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4271        /*
4272         * The page is committed to the memcg, but it's not actually
4273         * charged to the res_counter since we plan on replacing the
4274         * old one and only one page is going to be left afterwards.
4275         */
4276        __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4277}
4278
4279/* remove redundant charge if migration failed*/
4280void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4281        struct page *oldpage, struct page *newpage, bool migration_ok)
4282{
4283        struct page *used, *unused;
4284        struct page_cgroup *pc;
4285        bool anon;
4286
4287        if (!memcg)
4288                return;
4289
4290        if (!migration_ok) {
4291                used = oldpage;
4292                unused = newpage;
4293        } else {
4294                used = newpage;
4295                unused = oldpage;
4296        }
4297        anon = PageAnon(used);
4298        __mem_cgroup_uncharge_common(unused,
4299                                     anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4300                                     : MEM_CGROUP_CHARGE_TYPE_CACHE,
4301                                     true);
4302        css_put(&memcg->css);
4303        /*
4304         * We disallowed uncharge of pages under migration because mapcount
4305         * of the page goes down to zero, temporarly.
4306         * Clear the flag and check the page should be charged.
4307         */
4308        pc = lookup_page_cgroup(oldpage);
4309        lock_page_cgroup(pc);
4310        ClearPageCgroupMigration(pc);
4311        unlock_page_cgroup(pc);
4312
4313        /*
4314         * If a page is a file cache, radix-tree replacement is very atomic
4315         * and we can skip this check. When it was an Anon page, its mapcount
4316         * goes down to 0. But because we added MIGRATION flage, it's not
4317         * uncharged yet. There are several case but page->mapcount check
4318         * and USED bit check in mem_cgroup_uncharge_page() will do enough
4319         * check. (see prepare_charge() also)
4320         */
4321        if (anon)
4322                mem_cgroup_uncharge_page(used);
4323}
4324
4325/*
4326 * At replace page cache, newpage is not under any memcg but it's on
4327 * LRU. So, this function doesn't touch res_counter but handles LRU
4328 * in correct way. Both pages are locked so we cannot race with uncharge.
4329 */
4330void mem_cgroup_replace_page_cache(struct page *oldpage,
4331                                  struct page *newpage)
4332{
4333        struct mem_cgroup *memcg = NULL;
4334        struct page_cgroup *pc;
4335        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4336
4337        if (mem_cgroup_disabled())
4338                return;
4339
4340        pc = lookup_page_cgroup(oldpage);
4341        /* fix accounting on old pages */
4342        lock_page_cgroup(pc);
4343        if (PageCgroupUsed(pc)) {
4344                memcg = pc->mem_cgroup;
4345                mem_cgroup_charge_statistics(memcg, false, -1);
4346                ClearPageCgroupUsed(pc);
4347        }
4348        unlock_page_cgroup(pc);
4349
4350        /*
4351         * When called from shmem_replace_page(), in some cases the
4352         * oldpage has already been charged, and in some cases not.
4353         */
4354        if (!memcg)
4355                return;
4356        /*
4357         * Even if newpage->mapping was NULL before starting replacement,
4358         * the newpage may be on LRU(or pagevec for LRU) already. We lock
4359         * LRU while we overwrite pc->mem_cgroup.
4360         */
4361        __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4362}
4363
4364#ifdef CONFIG_DEBUG_VM
4365static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4366{
4367        struct page_cgroup *pc;
4368
4369        pc = lookup_page_cgroup(page);
4370        /*
4371         * Can be NULL while feeding pages into the page allocator for
4372         * the first time, i.e. during boot or memory hotplug;
4373         * or when mem_cgroup_disabled().
4374         */
4375        if (likely(pc) && PageCgroupUsed(pc))
4376                return pc;
4377        return NULL;
4378}
4379
4380bool mem_cgroup_bad_page_check(struct page *page)
4381{
4382        if (mem_cgroup_disabled())
4383                return false;
4384
4385        return lookup_page_cgroup_used(page) != NULL;
4386}
4387
4388void mem_cgroup_print_bad_page(struct page *page)
4389{
4390        struct page_cgroup *pc;
4391
4392        pc = lookup_page_cgroup_used(page);
4393        if (pc) {
4394                printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4395                       pc, pc->flags, pc->mem_cgroup);
4396        }
4397}
4398#endif
4399
4400static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4401                                unsigned long long val)
4402{
4403        int retry_count;
4404        u64 memswlimit, memlimit;
4405        int ret = 0;
4406        int children = mem_cgroup_count_children(memcg);
4407        u64 curusage, oldusage;
4408        int enlarge;
4409
4410        /*
4411         * For keeping hierarchical_reclaim simple, how long we should retry
4412         * is depends on callers. We set our retry-count to be function
4413         * of # of children which we should visit in this loop.
4414         */
4415        retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4416
4417        oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4418
4419        enlarge = 0;
4420        while (retry_count) {
4421                if (signal_pending(current)) {
4422                        ret = -EINTR;
4423                        break;
4424                }
4425                /*
4426                 * Rather than hide all in some function, I do this in
4427                 * open coded manner. You see what this really does.
4428                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4429                 */
4430                mutex_lock(&set_limit_mutex);
4431                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4432                if (memswlimit < val) {
4433                        ret = -EINVAL;
4434                        mutex_unlock(&set_limit_mutex);
4435                        break;
4436                }
4437
4438                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4439                if (memlimit < val)
4440                        enlarge = 1;
4441
4442                ret = res_counter_set_limit(&memcg->res, val);
4443                if (!ret) {
4444                        if (memswlimit == val)
4445                                memcg->memsw_is_minimum = true;
4446                        else
4447                                memcg->memsw_is_minimum = false;
4448                }
4449                mutex_unlock(&set_limit_mutex);
4450
4451                if (!ret)
4452                        break;
4453
4454                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4455                                   MEM_CGROUP_RECLAIM_SHRINK);
4456                curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4457                /* Usage is reduced ? */
4458                if (curusage >= oldusage)
4459                        retry_count--;
4460                else
4461                        oldusage = curusage;
4462        }
4463        if (!ret && enlarge)
4464                memcg_oom_recover(memcg);
4465
4466        return ret;
4467}
4468
4469static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4470                                        unsigned long long val)
4471{
4472        int retry_count;
4473        u64 memlimit, memswlimit, oldusage, curusage;
4474        int children = mem_cgroup_count_children(memcg);
4475        int ret = -EBUSY;
4476        int enlarge = 0;
4477
4478        /* see mem_cgroup_resize_res_limit */
4479        retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4480        oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4481        while (retry_count) {
4482                if (signal_pending(current)) {
4483                        ret = -EINTR;
4484                        break;
4485                }
4486                /*
4487                 * Rather than hide all in some function, I do this in
4488                 * open coded manner. You see what this really does.
4489                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4490                 */
4491                mutex_lock(&set_limit_mutex);
4492                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4493                if (memlimit > val) {
4494                        ret = -EINVAL;
4495                        mutex_unlock(&set_limit_mutex);
4496                        break;
4497                }
4498                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4499                if (memswlimit < val)
4500                        enlarge = 1;
4501                ret = res_counter_set_limit(&memcg->memsw, val);
4502                if (!ret) {
4503                        if (memlimit == val)
4504                                memcg->memsw_is_minimum = true;
4505                        else
4506                                memcg->memsw_is_minimum = false;
4507                }
4508                mutex_unlock(&set_limit_mutex);
4509
4510                if (!ret)
4511                        break;
4512
4513                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4514                                   MEM_CGROUP_RECLAIM_NOSWAP |
4515                                   MEM_CGROUP_RECLAIM_SHRINK);
4516                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4517                /* Usage is reduced ? */
4518                if (curusage >= oldusage)
4519                        retry_count--;
4520                else
4521                        oldusage = curusage;
4522        }
4523        if (!ret && enlarge)
4524                memcg_oom_recover(memcg);
4525        return ret;
4526}
4527
4528unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4529                                            gfp_t gfp_mask,
4530                                            unsigned long *total_scanned)
4531{
4532        unsigned long nr_reclaimed = 0;
4533        struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4534        unsigned long reclaimed;
4535        int loop = 0;
4536        struct mem_cgroup_tree_per_zone *mctz;
4537        unsigned long long excess;
4538        unsigned long nr_scanned;
4539
4540        if (order > 0)
4541                return 0;
4542
4543        mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4544        /*
4545         * This loop can run a while, specially if mem_cgroup's continuously
4546         * keep exceeding their soft limit and putting the system under
4547         * pressure
4548         */
4549        do {
4550                if (next_mz)
4551                        mz = next_mz;
4552                else
4553                        mz = mem_cgroup_largest_soft_limit_node(mctz);
4554                if (!mz)
4555                        break;
4556
4557                nr_scanned = 0;
4558                reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4559                                                    gfp_mask, &nr_scanned);
4560                nr_reclaimed += reclaimed;
4561                *total_scanned += nr_scanned;
4562                spin_lock(&mctz->lock);
4563
4564                /*
4565                 * If we failed to reclaim anything from this memory cgroup
4566                 * it is time to move on to the next cgroup
4567                 */
4568                next_mz = NULL;
4569                if (!reclaimed) {
4570                        do {
4571                                /*
4572                                 * Loop until we find yet another one.
4573                                 *
4574                                 * By the time we get the soft_limit lock
4575                                 * again, someone might have aded the
4576                                 * group back on the RB tree. Iterate to
4577                                 * make sure we get a different mem.
4578                                 * mem_cgroup_largest_soft_limit_node returns
4579                                 * NULL if no other cgroup is present on
4580                                 * the tree
4581                                 */
4582                                next_mz =
4583                                __mem_cgroup_largest_soft_limit_node(mctz);
4584                                if (next_mz == mz)
4585                                        css_put(&next_mz->memcg->css);
4586                                else /* next_mz == NULL or other memcg */
4587                                        break;
4588                        } while (1);
4589                }
4590                __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4591                excess = res_counter_soft_limit_excess(&mz->memcg->res);
4592                /*
4593                 * One school of thought says that we should not add
4594                 * back the node to the tree if reclaim returns 0.
4595                 * But our reclaim could return 0, simply because due
4596                 * to priority we are exposing a smaller subset of
4597                 * memory to reclaim from. Consider this as a longer
4598                 * term TODO.
4599                 */
4600                /* If excess == 0, no tree ops */
4601                __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4602                spin_unlock(&mctz->lock);
4603                css_put(&mz->memcg->css);
4604                loop++;
4605                /*
4606                 * Could not reclaim anything and there are no more
4607                 * mem cgroups to try or we seem to be looping without
4608                 * reclaiming anything.
4609                 */
4610                if (!nr_reclaimed &&
4611                        (next_mz == NULL ||
4612                        loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4613                        break;
4614        } while (!nr_reclaimed);
4615        if (next_mz)
4616                css_put(&next_mz->memcg->css);
4617        return nr_reclaimed;
4618}
4619
4620/**
4621 * mem_cgroup_force_empty_list - clears LRU of a group
4622 * @memcg: group to clear
4623 * @node: NUMA node
4624 * @zid: zone id
4625 * @lru: lru to to clear
4626 *
4627 * Traverse a specified page_cgroup list and try to drop them all.  This doesn't
4628 * reclaim the pages page themselves - pages are moved to the parent (or root)
4629 * group.
4630 */
4631static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4632                                int node, int zid, enum lru_list lru)
4633{
4634        struct lruvec *lruvec;
4635        unsigned long flags;
4636        struct list_head *list;
4637        struct page *busy;
4638        struct zone *zone;
4639
4640        zone = &NODE_DATA(node)->node_zones[zid];
4641        lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4642        list = &lruvec->lists[lru];
4643
4644        busy = NULL;
4645        do {
4646                struct page_cgroup *pc;
4647                struct page *page;
4648
4649                spin_lock_irqsave(&zone->lru_lock, flags);
4650                if (list_empty(list)) {
4651                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4652                        break;
4653                }
4654                page = list_entry(list->prev, struct page, lru);
4655                if (busy == page) {
4656                        list_move(&page->lru, list);
4657                        busy = NULL;
4658                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4659                        continue;
4660                }
4661                spin_unlock_irqrestore(&zone->lru_lock, flags);
4662
4663                pc = lookup_page_cgroup(page);
4664
4665                if (mem_cgroup_move_parent(page, pc, memcg)) {
4666                        /* found lock contention or "pc" is obsolete. */
4667                        busy = page;
4668                        cond_resched();
4669                } else
4670                        busy = NULL;
4671        } while (!list_empty(list));
4672}
4673
4674/*
4675 * make mem_cgroup's charge to be 0 if there is no task by moving
4676 * all the charges and pages to the parent.
4677 * This enables deleting this mem_cgroup.
4678 *
4679 * Caller is responsible for holding css reference on the memcg.
4680 */
4681static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4682{
4683        int node, zid;
4684        u64 usage;
4685
4686        do {
4687                /* This is for making all *used* pages to be on LRU. */
4688                lru_add_drain_all();
4689                drain_all_stock_sync(memcg);
4690                mem_cgroup_start_move(memcg);
4691                for_each_node_state(node, N_MEMORY) {
4692                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4693                                enum lru_list lru;
4694                                for_each_lru(lru) {
4695                                        mem_cgroup_force_empty_list(memcg,
4696                                                        node, zid, lru);
4697                                }
4698                        }
4699                }
4700                mem_cgroup_end_move(memcg);
4701                memcg_oom_recover(memcg);
4702                cond_resched();
4703
4704                /*
4705                 * Kernel memory may not necessarily be trackable to a specific
4706                 * process. So they are not migrated, and therefore we can't
4707                 * expect their value to drop to 0 here.
4708                 * Having res filled up with kmem only is enough.
4709                 *
4710                 * This is a safety check because mem_cgroup_force_empty_list
4711                 * could have raced with mem_cgroup_replace_page_cache callers
4712                 * so the lru seemed empty but the page could have been added
4713                 * right after the check. RES_USAGE should be safe as we always
4714                 * charge before adding to the LRU.
4715                 */
4716                usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4717                        res_counter_read_u64(&memcg->kmem, RES_USAGE);
4718        } while (usage > 0);
4719}
4720
4721/*
4722 * Reclaims as many pages from the given memcg as possible and moves
4723 * the rest to the parent.
4724 *
4725 * Caller is responsible for holding css reference for memcg.
4726 */
4727static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4728{
4729        int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4730        struct cgroup *cgrp = memcg->css.cgroup;
4731
4732        /* returns EBUSY if there is a task or if we come here twice. */
4733        if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4734                return -EBUSY;
4735
4736        /* we call try-to-free pages for make this cgroup empty */
4737        lru_add_drain_all();
4738        /* try to free all pages in this cgroup */
4739        while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4740                int progress;
4741
4742                if (signal_pending(current))
4743                        return -EINTR;
4744
4745                progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4746                                                false);
4747                if (!progress) {
4748                        nr_retries--;
4749                        /* maybe some writeback is necessary */
4750                        congestion_wait(BLK_RW_ASYNC, HZ/10);
4751                }
4752
4753        }
4754        lru_add_drain();
4755        mem_cgroup_reparent_charges(memcg);
4756
4757        return 0;
4758}
4759
4760static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4761{
4762        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4763        int ret;
4764
4765        if (mem_cgroup_is_root(memcg))
4766                return -EINVAL;
4767        css_get(&memcg->css);
4768        ret = mem_cgroup_force_empty(memcg);
4769        css_put(&memcg->css);
4770
4771        return ret;
4772}
4773
4774
4775static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4776{
4777        return mem_cgroup_from_cont(cont)->use_hierarchy;
4778}
4779
4780static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4781                                        u64 val)
4782{
4783        int retval = 0;
4784        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4785        struct cgroup *parent = cont->parent;
4786        struct mem_cgroup *parent_memcg = NULL;
4787
4788        if (parent)
4789                parent_memcg = mem_cgroup_from_cont(parent);
4790
4791        cgroup_lock();
4792
4793        if (memcg->use_hierarchy == val)
4794                goto out;
4795
4796        /*
4797         * If parent's use_hierarchy is set, we can't make any modifications
4798         * in the child subtrees. If it is unset, then the change can
4799         * occur, provided the current cgroup has no children.
4800         *
4801         * For the root cgroup, parent_mem is NULL, we allow value to be
4802         * set if there are no children.
4803         */
4804        if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4805                                (val == 1 || val == 0)) {
4806                if (list_empty(&cont->children))
4807                        memcg->use_hierarchy = val;
4808                else
4809                        retval = -EBUSY;
4810        } else
4811                retval = -EINVAL;
4812
4813out:
4814        cgroup_unlock();
4815
4816        return retval;
4817}
4818
4819
4820static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4821                                               enum mem_cgroup_stat_index idx)
4822{
4823        struct mem_cgroup *iter;
4824        long val = 0;
4825
4826        /* Per-cpu values can be negative, use a signed accumulator */
4827        for_each_mem_cgroup_tree(iter, memcg)
4828                val += mem_cgroup_read_stat(iter, idx);
4829
4830        if (val < 0) /* race ? */
4831                val = 0;
4832        return val;
4833}
4834
4835static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4836{
4837        u64 val;
4838
4839        if (!mem_cgroup_is_root(memcg)) {
4840                if (!swap)
4841                        return res_counter_read_u64(&memcg->res, RES_USAGE);
4842                else
4843                        return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4844        }
4845
4846        val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4847        val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4848
4849        if (swap)
4850                val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4851
4852        return val << PAGE_SHIFT;
4853}
4854
4855static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4856                               struct file *file, char __user *buf,
4857                               size_t nbytes, loff_t *ppos)
4858{
4859        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4860        char str[64];
4861        u64 val;
4862        int name, len;
4863        enum res_type type;
4864
4865        type = MEMFILE_TYPE(cft->private);
4866        name = MEMFILE_ATTR(cft->private);
4867
4868        if (!do_swap_account && type == _MEMSWAP)
4869                return -EOPNOTSUPP;
4870
4871        switch (type) {
4872        case _MEM:
4873                if (name == RES_USAGE)
4874                        val = mem_cgroup_usage(memcg, false);
4875                else
4876                        val = res_counter_read_u64(&memcg->res, name);
4877                break;
4878        case _MEMSWAP:
4879                if (name == RES_USAGE)
4880                        val = mem_cgroup_usage(memcg, true);
4881                else
4882                        val = res_counter_read_u64(&memcg->memsw, name);
4883                break;
4884        case _KMEM:
4885                val = res_counter_read_u64(&memcg->kmem, name);
4886                break;
4887        default:
4888                BUG();
4889        }
4890
4891        len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4892        return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4893}
4894
4895static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4896{
4897        int ret = -EINVAL;
4898#ifdef CONFIG_MEMCG_KMEM
4899        bool must_inc_static_branch = false;
4900
4901        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4902        /*
4903         * For simplicity, we won't allow this to be disabled.  It also can't
4904         * be changed if the cgroup has children already, or if tasks had
4905         * already joined.
4906         *
4907         * If tasks join before we set the limit, a person looking at
4908         * kmem.usage_in_bytes will have no way to determine when it took
4909         * place, which makes the value quite meaningless.
4910         *
4911         * After it first became limited, changes in the value of the limit are
4912         * of course permitted.
4913         *
4914         * Taking the cgroup_lock is really offensive, but it is so far the only
4915         * way to guarantee that no children will appear. There are plenty of
4916         * other offenders, and they should all go away. Fine grained locking
4917         * is probably the way to go here. When we are fully hierarchical, we
4918         * can also get rid of the use_hierarchy check.
4919         */
4920        cgroup_lock();
4921        mutex_lock(&set_limit_mutex);
4922        if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4923                if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4924                                                !list_empty(&cont->children))) {
4925                        ret = -EBUSY;
4926                        goto out;
4927                }
4928                ret = res_counter_set_limit(&memcg->kmem, val);
4929                VM_BUG_ON(ret);
4930
4931                ret = memcg_update_cache_sizes(memcg);
4932                if (ret) {
4933                        res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4934                        goto out;
4935                }
4936                must_inc_static_branch = true;
4937                /*
4938                 * kmem charges can outlive the cgroup. In the case of slab
4939                 * pages, for instance, a page contain objects from various
4940                 * processes, so it is unfeasible to migrate them away. We
4941                 * need to reference count the memcg because of that.
4942                 */
4943                mem_cgroup_get(memcg);
4944        } else
4945                ret = res_counter_set_limit(&memcg->kmem, val);
4946out:
4947        mutex_unlock(&set_limit_mutex);
4948        cgroup_unlock();
4949
4950        /*
4951         * We are by now familiar with the fact that we can't inc the static
4952         * branch inside cgroup_lock. See disarm functions for details. A
4953         * worker here is overkill, but also wrong: After the limit is set, we
4954         * must start accounting right away. Since this operation can't fail,
4955         * we can safely defer it to here - no rollback will be needed.
4956         *
4957         * The boolean used to control this is also safe, because
4958         * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4959         * able to set it to true;
4960         */
4961        if (must_inc_static_branch) {
4962                static_key_slow_inc(&memcg_kmem_enabled_key);
4963                /*
4964                 * setting the active bit after the inc will guarantee no one
4965                 * starts accounting before all call sites are patched
4966                 */
4967                memcg_kmem_set_active(memcg);
4968        }
4969
4970#endif
4971        return ret;
4972}
4973
4974static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4975{
4976        int ret = 0;
4977        struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4978        if (!parent)
4979                goto out;
4980
4981        memcg->kmem_account_flags = parent->kmem_account_flags;
4982#ifdef CONFIG_MEMCG_KMEM
4983        /*
4984         * When that happen, we need to disable the static branch only on those
4985         * memcgs that enabled it. To achieve this, we would be forced to
4986         * complicate the code by keeping track of which memcgs were the ones
4987         * that actually enabled limits, and which ones got it from its
4988         * parents.
4989         *
4990         * It is a lot simpler just to do static_key_slow_inc() on every child
4991         * that is accounted.
4992         */
4993        if (!memcg_kmem_is_active(memcg))
4994                goto out;
4995
4996        /*
4997         * destroy(), called if we fail, will issue static_key_slow_inc() and
4998         * mem_cgroup_put() if kmem is enabled. We have to either call them
4999         * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5000         * this more consistent, since it always leads to the same destroy path
5001         */
5002        mem_cgroup_get(memcg);
5003        static_key_slow_inc(&memcg_kmem_enabled_key);
5004
5005        mutex_lock(&set_limit_mutex);
5006        ret = memcg_update_cache_sizes(memcg);
5007        mutex_unlock(&set_limit_mutex);
5008#endif
5009out:
5010        return ret;
5011}
5012
5013/*
5014 * The user of this function is...
5015 * RES_LIMIT.
5016 */
5017static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5018                            const char *buffer)
5019{
5020        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5021        enum res_type type;
5022        int name;
5023        unsigned long long val;
5024        int ret;
5025
5026        type = MEMFILE_TYPE(cft->private);
5027        name = MEMFILE_ATTR(cft->private);
5028
5029        if (!do_swap_account && type == _MEMSWAP)
5030                return -EOPNOTSUPP;
5031
5032        switch (name) {
5033        case RES_LIMIT:
5034                if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5035                        ret = -EINVAL;
5036                        break;
5037                }
5038                /* This function does all necessary parse...reuse it */
5039                ret = res_counter_memparse_write_strategy(buffer, &val);
5040                if (ret)
5041                        break;
5042                if (type == _MEM)
5043                        ret = mem_cgroup_resize_limit(memcg, val);
5044                else if (type == _MEMSWAP)
5045                        ret = mem_cgroup_resize_memsw_limit(memcg, val);
5046                else if (type == _KMEM)
5047                        ret = memcg_update_kmem_limit(cont, val);
5048                else
5049                        return -EINVAL;
5050                break;
5051        case RES_SOFT_LIMIT:
5052                ret = res_counter_memparse_write_strategy(buffer, &val);
5053                if (ret)
5054                        break;
5055                /*
5056                 * For memsw, soft limits are hard to implement in terms
5057                 * of semantics, for now, we support soft limits for
5058                 * control without swap
5059                 */
5060                if (type == _MEM)
5061                        ret = res_counter_set_soft_limit(&memcg->res, val);
5062                else
5063                        ret = -EINVAL;
5064                break;
5065        default:
5066                ret = -EINVAL; /* should be BUG() ? */
5067                break;
5068        }
5069        return ret;
5070}
5071
5072static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5073                unsigned long long *mem_limit, unsigned long long *memsw_limit)
5074{
5075        struct cgroup *cgroup;
5076        unsigned long long min_limit, min_memsw_limit, tmp;
5077
5078        min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5079        min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5080        cgroup = memcg->css.cgroup;
5081        if (!memcg->use_hierarchy)
5082                goto out;
5083
5084        while (cgroup->parent) {
5085                cgroup = cgroup->parent;
5086                memcg = mem_cgroup_from_cont(cgroup);
5087                if (!memcg->use_hierarchy)
5088                        break;
5089                tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5090                min_limit = min(min_limit, tmp);
5091                tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5092                min_memsw_limit = min(min_memsw_limit, tmp);
5093        }
5094out:
5095        *mem_limit = min_limit;
5096        *memsw_limit = min_memsw_limit;
5097}
5098
5099static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5100{
5101        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5102        int name;
5103        enum res_type type;
5104
5105        type = MEMFILE_TYPE(event);
5106        name = MEMFILE_ATTR(event);
5107
5108        if (!do_swap_account && type == _MEMSWAP)
5109                return -EOPNOTSUPP;
5110
5111        switch (name) {
5112        case RES_MAX_USAGE:
5113                if (type == _MEM)
5114                        res_counter_reset_max(&memcg->res);
5115                else if (type == _MEMSWAP)
5116                        res_counter_reset_max(&memcg->memsw);
5117                else if (type == _KMEM)
5118                        res_counter_reset_max(&memcg->kmem);
5119                else
5120                        return -EINVAL;
5121                break;
5122        case RES_FAILCNT:
5123                if (type == _MEM)
5124                        res_counter_reset_failcnt(&memcg->res);
5125                else if (type == _MEMSWAP)
5126                        res_counter_reset_failcnt(&memcg->memsw);
5127                else if (type == _KMEM)
5128                        res_counter_reset_failcnt(&memcg->kmem);
5129                else
5130                        return -EINVAL;
5131                break;
5132        }
5133
5134        return 0;
5135}
5136
5137static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5138                                        struct cftype *cft)
5139{
5140        return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5141}
5142
5143#ifdef CONFIG_MMU
5144static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5145                                        struct cftype *cft, u64 val)
5146{
5147        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5148
5149        if (val >= (1 << NR_MOVE_TYPE))
5150                return -EINVAL;
5151        /*
5152         * We check this value several times in both in can_attach() and
5153         * attach(), so we need cgroup lock to prevent this value from being
5154         * inconsistent.
5155         */
5156        cgroup_lock();
5157        memcg->move_charge_at_immigrate = val;
5158        cgroup_unlock();
5159
5160        return 0;
5161}
5162#else
5163static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5164                                        struct cftype *cft, u64 val)
5165{
5166        return -ENOSYS;
5167}
5168#endif
5169
5170#ifdef CONFIG_NUMA
5171static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5172                                      struct seq_file *m)
5173{
5174        int nid;
5175        unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5176        unsigned long node_nr;
5177        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5178
5179        total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5180        seq_printf(m, "total=%lu", total_nr);
5181        for_each_node_state(nid, N_MEMORY) {
5182                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5183                seq_printf(m, " N%d=%lu", nid, node_nr);
5184        }
5185        seq_putc(m, '\n');
5186
5187        file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5188        seq_printf(m, "file=%lu", file_nr);
5189        for_each_node_state(nid, N_MEMORY) {
5190                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5191                                LRU_ALL_FILE);
5192                seq_printf(m, " N%d=%lu", nid, node_nr);
5193        }
5194        seq_putc(m, '\n');
5195
5196        anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5197        seq_printf(m, "anon=%lu", anon_nr);
5198        for_each_node_state(nid, N_MEMORY) {
5199                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5200                                LRU_ALL_ANON);
5201                seq_printf(m, " N%d=%lu", nid, node_nr);
5202        }
5203        seq_putc(m, '\n');
5204
5205        unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5206        seq_printf(m, "unevictable=%lu", unevictable_nr);
5207        for_each_node_state(nid, N_MEMORY) {
5208                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5209                                BIT(LRU_UNEVICTABLE));
5210                seq_printf(m, " N%d=%lu", nid, node_nr);
5211        }
5212        seq_putc(m, '\n');
5213        return 0;
5214}
5215#endif /* CONFIG_NUMA */
5216
5217static const char * const mem_cgroup_lru_names[] = {
5218        "inactive_anon",
5219        "active_anon",
5220        "inactive_file",
5221        "active_file",
5222        "unevictable",
5223};
5224
5225static inline void mem_cgroup_lru_names_not_uptodate(void)
5226{
5227        BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5228}
5229
5230static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5231                                 struct seq_file *m)
5232{
5233        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5234        struct mem_cgroup *mi;
5235        unsigned int i;
5236
5237        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5238                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5239                        continue;
5240                seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5241                           mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5242        }
5243
5244        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5245                seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5246                           mem_cgroup_read_events(memcg, i));
5247
5248        for (i = 0; i < NR_LRU_LISTS; i++)
5249                seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5250                           mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5251
5252        /* Hierarchical information */
5253        {
5254                unsigned long long limit, memsw_limit;
5255                memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5256                seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5257                if (do_swap_account)
5258                        seq_printf(m, "hierarchical_memsw_limit %llu\n",
5259                                   memsw_limit);
5260        }
5261
5262        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5263                long long val = 0;
5264
5265                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5266                        continue;
5267                for_each_mem_cgroup_tree(mi, memcg)
5268                        val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5269                seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5270        }
5271
5272        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5273                unsigned long long val = 0;
5274
5275                for_each_mem_cgroup_tree(mi, memcg)
5276                        val += mem_cgroup_read_events(mi, i);
5277                seq_printf(m, "total_%s %llu\n",
5278                           mem_cgroup_events_names[i], val);
5279        }
5280
5281        for (i = 0; i < NR_LRU_LISTS; i++) {
5282                unsigned long long val = 0;
5283
5284                for_each_mem_cgroup_tree(mi, memcg)
5285                        val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5286                seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5287        }
5288
5289#ifdef CONFIG_DEBUG_VM
5290        {
5291                int nid, zid;
5292                struct mem_cgroup_per_zone *mz;
5293                struct zone_reclaim_stat *rstat;
5294                unsigned long recent_rotated[2] = {0, 0};
5295                unsigned long recent_scanned[2] = {0, 0};
5296
5297                for_each_online_node(nid)
5298                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5299                                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5300                                rstat = &mz->lruvec.reclaim_stat;
5301
5302                                recent_rotated[0] += rstat->recent_rotated[0];
5303                                recent_rotated[1] += rstat->recent_rotated[1];
5304                                recent_scanned[0] += rstat->recent_scanned[0];
5305                                recent_scanned[1] += rstat->recent_scanned[1];
5306                        }
5307                seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5308                seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5309                seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5310                seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5311        }
5312#endif
5313
5314        return 0;
5315}
5316
5317static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5318{
5319        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5320
5321        return mem_cgroup_swappiness(memcg);
5322}
5323
5324static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5325                                       u64 val)
5326{
5327        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5328        struct mem_cgroup *parent;
5329
5330        if (val > 100)
5331                return -EINVAL;
5332
5333        if (cgrp->parent == NULL)
5334                return -EINVAL;
5335
5336        parent = mem_cgroup_from_cont(cgrp->parent);
5337
5338        cgroup_lock();
5339
5340        /* If under hierarchy, only empty-root can set this value */
5341        if ((parent->use_hierarchy) ||
5342            (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5343                cgroup_unlock();
5344                return -EINVAL;
5345        }
5346
5347        memcg->swappiness = val;
5348
5349        cgroup_unlock();
5350
5351        return 0;
5352}
5353
5354static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5355{
5356        struct mem_cgroup_threshold_ary *t;
5357        u64 usage;
5358        int i;
5359
5360        rcu_read_lock();
5361        if (!swap)
5362                t = rcu_dereference(memcg->thresholds.primary);
5363        else
5364                t = rcu_dereference(memcg->memsw_thresholds.primary);
5365
5366        if (!t)
5367                goto unlock;
5368
5369        usage = mem_cgroup_usage(memcg, swap);
5370
5371        /*
5372         * current_threshold points to threshold just below or equal to usage.
5373         * If it's not true, a threshold was crossed after last
5374         * call of __mem_cgroup_threshold().
5375         */
5376        i = t->current_threshold;
5377
5378        /*
5379         * Iterate backward over array of thresholds starting from
5380         * current_threshold and check if a threshold is crossed.
5381         * If none of thresholds below usage is crossed, we read
5382         * only one element of the array here.
5383         */
5384        for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5385                eventfd_signal(t->entries[i].eventfd, 1);
5386
5387        /* i = current_threshold + 1 */
5388        i++;
5389
5390        /*
5391         * Iterate forward over array of thresholds starting from
5392         * current_threshold+1 and check if a threshold is crossed.
5393         * If none of thresholds above usage is crossed, we read
5394         * only one element of the array here.
5395         */
5396        for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5397                eventfd_signal(t->entries[i].eventfd, 1);
5398
5399        /* Update current_threshold */
5400        t->current_threshold = i - 1;
5401unlock:
5402        rcu_read_unlock();
5403}
5404
5405static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5406{
5407        while (memcg) {
5408                __mem_cgroup_threshold(memcg, false);
5409                if (do_swap_account)
5410                        __mem_cgroup_threshold(memcg, true);
5411
5412                memcg = parent_mem_cgroup(memcg);
5413        }
5414}
5415
5416static int compare_thresholds(const void *a, const void *b)
5417{
5418        const struct mem_cgroup_threshold *_a = a;
5419        const struct mem_cgroup_threshold *_b = b;
5420
5421        return _a->threshold - _b->threshold;
5422}
5423
5424static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5425{
5426        struct mem_cgroup_eventfd_list *ev;
5427
5428        list_for_each_entry(ev, &memcg->oom_notify, list)
5429                eventfd_signal(ev->eventfd, 1);
5430        return 0;
5431}
5432
5433static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5434{
5435        struct mem_cgroup *iter;
5436
5437        for_each_mem_cgroup_tree(iter, memcg)
5438                mem_cgroup_oom_notify_cb(iter);
5439}
5440
5441static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5442        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5443{
5444        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5445        struct mem_cgroup_thresholds *thresholds;
5446        struct mem_cgroup_threshold_ary *new;
5447        enum res_type type = MEMFILE_TYPE(cft->private);
5448        u64 threshold, usage;
5449        int i, size, ret;
5450
5451        ret = res_counter_memparse_write_strategy(args, &threshold);
5452        if (ret)
5453                return ret;
5454
5455        mutex_lock(&memcg->thresholds_lock);
5456
5457        if (type == _MEM)
5458                thresholds = &memcg->thresholds;
5459        else if (type == _MEMSWAP)
5460                thresholds = &memcg->memsw_thresholds;
5461        else
5462                BUG();
5463
5464        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5465
5466        /* Check if a threshold crossed before adding a new one */
5467        if (thresholds->primary)
5468                __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5469
5470        size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5471
5472        /* Allocate memory for new array of thresholds */
5473        new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5474                        GFP_KERNEL);
5475        if (!new) {
5476                ret = -ENOMEM;
5477                goto unlock;
5478        }
5479        new->size = size;
5480
5481        /* Copy thresholds (if any) to new array */
5482        if (thresholds->primary) {
5483                memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5484                                sizeof(struct mem_cgroup_threshold));
5485        }
5486
5487        /* Add new threshold */
5488        new->entries[size - 1].eventfd = eventfd;
5489        new->entries[size - 1].threshold = threshold;
5490
5491        /* Sort thresholds. Registering of new threshold isn't time-critical */
5492        sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5493                        compare_thresholds, NULL);
5494
5495        /* Find current threshold */
5496        new->current_threshold = -1;
5497        for (i = 0; i < size; i++) {
5498                if (new->entries[i].threshold <= usage) {
5499                        /*
5500                         * new->current_threshold will not be used until
5501                         * rcu_assign_pointer(), so it's safe to increment
5502                         * it here.
5503                         */
5504                        ++new->current_threshold;
5505                } else
5506                        break;
5507        }
5508
5509        /* Free old spare buffer and save old primary buffer as spare */
5510        kfree(thresholds->spare);
5511        thresholds->spare = thresholds->primary;
5512
5513        rcu_assign_pointer(thresholds->primary, new);
5514
5515        /* To be sure that nobody uses thresholds */
5516        synchronize_rcu();
5517
5518unlock:
5519        mutex_unlock(&memcg->thresholds_lock);
5520
5521        return ret;
5522}
5523
5524static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5525        struct cftype *cft, struct eventfd_ctx *eventfd)
5526{
5527        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5528        struct mem_cgroup_thresholds *thresholds;
5529        struct mem_cgroup_threshold_ary *new;
5530        enum res_type type = MEMFILE_TYPE(cft->private);
5531        u64 usage;
5532        int i, j, size;
5533
5534        mutex_lock(&memcg->thresholds_lock);
5535        if (type == _MEM)
5536                thresholds = &memcg->thresholds;
5537        else if (type == _MEMSWAP)
5538                thresholds = &memcg->memsw_thresholds;
5539        else
5540                BUG();
5541
5542        if (!thresholds->primary)
5543                goto unlock;
5544
5545        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5546
5547        /* Check if a threshold crossed before removing */
5548        __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5549
5550        /* Calculate new number of threshold */
5551        size = 0;
5552        for (i = 0; i < thresholds->primary->size; i++) {
5553                if (thresholds->primary->entries[i].eventfd != eventfd)
5554                        size++;
5555        }
5556
5557        new = thresholds->spare;
5558
5559        /* Set thresholds array to NULL if we don't have thresholds */
5560        if (!size) {
5561                kfree(new);
5562                new = NULL;
5563                goto swap_buffers;
5564        }
5565
5566        new->size = size;
5567
5568        /* Copy thresholds and find current threshold */
5569        new->current_threshold = -1;
5570        for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5571                if (thresholds->primary->entries[i].eventfd == eventfd)
5572                        continue;
5573
5574                new->entries[j] = thresholds->primary->entries[i];
5575                if (new->entries[j].threshold <= usage) {
5576                        /*
5577                         * new->current_threshold will not be used
5578                         * until rcu_assign_pointer(), so it's safe to increment
5579                         * it here.
5580                         */
5581                        ++new->current_threshold;
5582                }
5583                j++;
5584        }
5585
5586swap_buffers:
5587        /* Swap primary and spare array */
5588        thresholds->spare = thresholds->primary;
5589        /* If all events are unregistered, free the spare array */
5590        if (!new) {
5591                kfree(thresholds->spare);
5592                thresholds->spare = NULL;
5593        }
5594
5595        rcu_assign_pointer(thresholds->primary, new);
5596
5597        /* To be sure that nobody uses thresholds */
5598        synchronize_rcu();
5599unlock:
5600        mutex_unlock(&memcg->thresholds_lock);
5601}
5602
5603static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5604        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5605{
5606        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5607        struct mem_cgroup_eventfd_list *event;
5608        enum res_type type = MEMFILE_TYPE(cft->private);
5609
5610        BUG_ON(type != _OOM_TYPE);
5611        event = kmalloc(sizeof(*event), GFP_KERNEL);
5612        if (!event)
5613                return -ENOMEM;
5614
5615        spin_lock(&memcg_oom_lock);
5616
5617        event->eventfd = eventfd;
5618        list_add(&event->list, &memcg->oom_notify);
5619
5620        /* already in OOM ? */
5621        if (atomic_read(&memcg->under_oom))
5622                eventfd_signal(eventfd, 1);
5623        spin_unlock(&memcg_oom_lock);
5624
5625        return 0;
5626}
5627
5628static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5629        struct cftype *cft, struct eventfd_ctx *eventfd)
5630{
5631        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5632        struct mem_cgroup_eventfd_list *ev, *tmp;
5633        enum res_type type = MEMFILE_TYPE(cft->private);
5634
5635        BUG_ON(type != _OOM_TYPE);
5636
5637        spin_lock(&memcg_oom_lock);
5638
5639        list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5640                if (ev->eventfd == eventfd) {
5641                        list_del(&ev->list);
5642                        kfree(ev);
5643                }
5644        }
5645
5646        spin_unlock(&memcg_oom_lock);
5647}
5648
5649static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5650        struct cftype *cft,  struct cgroup_map_cb *cb)
5651{
5652        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5653
5654        cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5655
5656        if (atomic_read(&memcg->under_oom))
5657                cb->fill(cb, "under_oom", 1);
5658        else
5659                cb->fill(cb, "under_oom", 0);
5660        return 0;
5661}
5662
5663static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5664        struct cftype *cft, u64 val)
5665{
5666        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5667        struct mem_cgroup *parent;
5668
5669        /* cannot set to root cgroup and only 0 and 1 are allowed */
5670        if (!cgrp->parent || !((val == 0) || (val == 1)))
5671                return -EINVAL;
5672
5673        parent = mem_cgroup_from_cont(cgrp->parent);
5674
5675        cgroup_lock();
5676        /* oom-kill-disable is a flag for subhierarchy. */
5677        if ((parent->use_hierarchy) ||
5678            (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5679                cgroup_unlock();
5680                return -EINVAL;
5681        }
5682        memcg->oom_kill_disable = val;
5683        if (!val)
5684                memcg_oom_recover(memcg);
5685        cgroup_unlock();
5686        return 0;
5687}
5688
5689#ifdef CONFIG_MEMCG_KMEM
5690static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5691{
5692        int ret;
5693
5694        memcg->kmemcg_id = -1;
5695        ret = memcg_propagate_kmem(memcg);
5696        if (ret)
5697                return ret;
5698
5699        return mem_cgroup_sockets_init(memcg, ss);
5700};
5701
5702static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5703{
5704        mem_cgroup_sockets_destroy(memcg);
5705
5706        memcg_kmem_mark_dead(memcg);
5707
5708        if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5709                return;
5710
5711        /*
5712         * Charges already down to 0, undo mem_cgroup_get() done in the charge
5713         * path here, being careful not to race with memcg_uncharge_kmem: it is
5714         * possible that the charges went down to 0 between mark_dead and the
5715         * res_counter read, so in that case, we don't need the put
5716         */
5717        if (memcg_kmem_test_and_clear_dead(memcg))
5718                mem_cgroup_put(memcg);
5719}
5720#else
5721static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5722{
5723        return 0;
5724}
5725
5726static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5727{
5728}
5729#endif
5730
5731static struct cftype mem_cgroup_files[] = {
5732        {
5733                .name = "usage_in_bytes",
5734                .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5735                .read = mem_cgroup_read,
5736                .register_event = mem_cgroup_usage_register_event,
5737                .unregister_event = mem_cgroup_usage_unregister_event,
5738        },
5739        {
5740                .name = "max_usage_in_bytes",
5741                .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5742                .trigger = mem_cgroup_reset,
5743                .read = mem_cgroup_read,
5744        },
5745        {
5746                .name = "limit_in_bytes",
5747                .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5748                .write_string = mem_cgroup_write,
5749                .read = mem_cgroup_read,
5750        },
5751        {
5752                .name = "soft_limit_in_bytes",
5753                .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5754                .write_string = mem_cgroup_write,
5755                .read = mem_cgroup_read,
5756        },
5757        {
5758                .name = "failcnt",
5759                .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5760                .trigger = mem_cgroup_reset,
5761                .read = mem_cgroup_read,
5762        },
5763        {
5764                .name = "stat",
5765                .read_seq_string = memcg_stat_show,
5766        },
5767        {
5768                .name = "force_empty",
5769                .trigger = mem_cgroup_force_empty_write,
5770        },
5771        {
5772                .name = "use_hierarchy",
5773                .write_u64 = mem_cgroup_hierarchy_write,
5774                .read_u64 = mem_cgroup_hierarchy_read,
5775        },
5776        {
5777                .name = "swappiness",
5778                .read_u64 = mem_cgroup_swappiness_read,
5779                .write_u64 = mem_cgroup_swappiness_write,
5780        },
5781        {
5782                .name = "move_charge_at_immigrate",
5783                .read_u64 = mem_cgroup_move_charge_read,
5784                .write_u64 = mem_cgroup_move_charge_write,
5785        },
5786        {
5787                .name = "oom_control",
5788                .read_map = mem_cgroup_oom_control_read,
5789                .write_u64 = mem_cgroup_oom_control_write,
5790                .register_event = mem_cgroup_oom_register_event,
5791                .unregister_event = mem_cgroup_oom_unregister_event,
5792                .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5793        },
5794#ifdef CONFIG_NUMA
5795        {
5796                .name = "numa_stat",
5797                .read_seq_string = memcg_numa_stat_show,
5798        },
5799#endif
5800#ifdef CONFIG_MEMCG_SWAP
5801        {
5802                .name = "memsw.usage_in_bytes",
5803                .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5804                .read = mem_cgroup_read,
5805