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/vmpressure.h>
  53#include <linux/mm_inline.h>
  54#include <linux/page_cgroup.h>
  55#include <linux/cpu.h>
  56#include <linux/oom.h>
  57#include "internal.h"
  58#include <net/sock.h>
  59#include <net/ip.h>
  60#include <net/tcp_memcontrol.h>
  61
  62#include <asm/uaccess.h>
  63
  64#include <trace/events/vmscan.h>
  65
  66struct cgroup_subsys mem_cgroup_subsys __read_mostly;
  67EXPORT_SYMBOL(mem_cgroup_subsys);
  68
  69#define MEM_CGROUP_RECLAIM_RETRIES      5
  70static struct mem_cgroup *root_mem_cgroup __read_mostly;
  71
  72#ifdef CONFIG_MEMCG_SWAP
  73/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
  74int do_swap_account __read_mostly;
  75
  76/* for remember boot option*/
  77#ifdef CONFIG_MEMCG_SWAP_ENABLED
  78static int really_do_swap_account __initdata = 1;
  79#else
  80static int really_do_swap_account __initdata = 0;
  81#endif
  82
  83#else
  84#define do_swap_account         0
  85#endif
  86
  87
  88/*
  89 * Statistics for memory cgroup.
  90 */
  91enum mem_cgroup_stat_index {
  92        /*
  93         * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  94         */
  95        MEM_CGROUP_STAT_CACHE,          /* # of pages charged as cache */
  96        MEM_CGROUP_STAT_RSS,            /* # of pages charged as anon rss */
  97        MEM_CGROUP_STAT_RSS_HUGE,       /* # of pages charged as anon huge */
  98        MEM_CGROUP_STAT_FILE_MAPPED,    /* # of pages charged as file rss */
  99        MEM_CGROUP_STAT_SWAP,           /* # of pages, swapped out */
 100        MEM_CGROUP_STAT_NSTATS,
 101};
 102
 103static const char * const mem_cgroup_stat_names[] = {
 104        "cache",
 105        "rss",
 106        "rss_huge",
 107        "mapped_file",
 108        "swap",
 109};
 110
 111enum mem_cgroup_events_index {
 112        MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
 113        MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
 114        MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
 115        MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
 116        MEM_CGROUP_EVENTS_NSTATS,
 117};
 118
 119static const char * const mem_cgroup_events_names[] = {
 120        "pgpgin",
 121        "pgpgout",
 122        "pgfault",
 123        "pgmajfault",
 124};
 125
 126static const char * const mem_cgroup_lru_names[] = {
 127        "inactive_anon",
 128        "active_anon",
 129        "inactive_file",
 130        "active_file",
 131        "unevictable",
 132};
 133
 134/*
 135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
 136 * it will be incremated by the number of pages. This counter is used for
 137 * for trigger some periodic events. This is straightforward and better
 138 * than using jiffies etc. to handle periodic memcg event.
 139 */
 140enum mem_cgroup_events_target {
 141        MEM_CGROUP_TARGET_THRESH,
 142        MEM_CGROUP_TARGET_SOFTLIMIT,
 143        MEM_CGROUP_TARGET_NUMAINFO,
 144        MEM_CGROUP_NTARGETS,
 145};
 146#define THRESHOLDS_EVENTS_TARGET 128
 147#define SOFTLIMIT_EVENTS_TARGET 1024
 148#define NUMAINFO_EVENTS_TARGET  1024
 149
 150struct mem_cgroup_stat_cpu {
 151        long count[MEM_CGROUP_STAT_NSTATS];
 152        unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
 153        unsigned long nr_page_events;
 154        unsigned long targets[MEM_CGROUP_NTARGETS];
 155};
 156
 157struct mem_cgroup_reclaim_iter {
 158        /*
 159         * last scanned hierarchy member. Valid only if last_dead_count
 160         * matches memcg->dead_count of the hierarchy root group.
 161         */
 162        struct mem_cgroup *last_visited;
 163        unsigned long last_dead_count;
 164
 165        /* scan generation, increased every round-trip */
 166        unsigned int generation;
 167};
 168
 169/*
 170 * per-zone information in memory controller.
 171 */
 172struct mem_cgroup_per_zone {
 173        struct lruvec           lruvec;
 174        unsigned long           lru_size[NR_LRU_LISTS];
 175
 176        struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
 177
 178        struct rb_node          tree_node;      /* RB tree node */
 179        unsigned long long      usage_in_excess;/* Set to the value by which */
 180                                                /* the soft limit is exceeded*/
 181        bool                    on_tree;
 182        struct mem_cgroup       *memcg;         /* Back pointer, we cannot */
 183                                                /* use container_of        */
 184};
 185
 186struct mem_cgroup_per_node {
 187        struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 188};
 189
 190struct mem_cgroup_lru_info {
 191        struct mem_cgroup_per_node *nodeinfo[0];
 192};
 193
 194/*
 195 * Cgroups above their limits are maintained in a RB-Tree, independent of
 196 * their hierarchy representation
 197 */
 198
 199struct mem_cgroup_tree_per_zone {
 200        struct rb_root rb_root;
 201        spinlock_t lock;
 202};
 203
 204struct mem_cgroup_tree_per_node {
 205        struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
 206};
 207
 208struct mem_cgroup_tree {
 209        struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
 210};
 211
 212static struct mem_cgroup_tree soft_limit_tree __read_mostly;
 213
 214struct mem_cgroup_threshold {
 215        struct eventfd_ctx *eventfd;
 216        u64 threshold;
 217};
 218
 219/* For threshold */
 220struct mem_cgroup_threshold_ary {
 221        /* An array index points to threshold just below or equal to usage. */
 222        int current_threshold;
 223        /* Size of entries[] */
 224        unsigned int size;
 225        /* Array of thresholds */
 226        struct mem_cgroup_threshold entries[0];
 227};
 228
 229struct mem_cgroup_thresholds {
 230        /* Primary thresholds array */
 231        struct mem_cgroup_threshold_ary *primary;
 232        /*
 233         * Spare threshold array.
 234         * This is needed to make mem_cgroup_unregister_event() "never fail".
 235         * It must be able to store at least primary->size - 1 entries.
 236         */
 237        struct mem_cgroup_threshold_ary *spare;
 238};
 239
 240/* for OOM */
 241struct mem_cgroup_eventfd_list {
 242        struct list_head list;
 243        struct eventfd_ctx *eventfd;
 244};
 245
 246static void mem_cgroup_threshold(struct mem_cgroup *memcg);
 247static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
 248
 249/*
 250 * The memory controller data structure. The memory controller controls both
 251 * page cache and RSS per cgroup. We would eventually like to provide
 252 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 253 * to help the administrator determine what knobs to tune.
 254 *
 255 * TODO: Add a water mark for the memory controller. Reclaim will begin when
 256 * we hit the water mark. May be even add a low water mark, such that
 257 * no reclaim occurs from a cgroup at it's low water mark, this is
 258 * a feature that will be implemented much later in the future.
 259 */
 260struct mem_cgroup {
 261        struct cgroup_subsys_state css;
 262        /*
 263         * the counter to account for memory usage
 264         */
 265        struct res_counter res;
 266
 267        /* vmpressure notifications */
 268        struct vmpressure vmpressure;
 269
 270        union {
 271                /*
 272                 * the counter to account for mem+swap usage.
 273                 */
 274                struct res_counter memsw;
 275
 276                /*
 277                 * rcu_freeing is used only when freeing struct mem_cgroup,
 278                 * so put it into a union to avoid wasting more memory.
 279                 * It must be disjoint from the css field.  It could be
 280                 * in a union with the res field, but res plays a much
 281                 * larger part in mem_cgroup life than memsw, and might
 282                 * be of interest, even at time of free, when debugging.
 283                 * So share rcu_head with the less interesting memsw.
 284                 */
 285                struct rcu_head rcu_freeing;
 286                /*
 287                 * We also need some space for a worker in deferred freeing.
 288                 * By the time we call it, rcu_freeing is no longer in use.
 289                 */
 290                struct work_struct work_freeing;
 291        };
 292
 293        /*
 294         * the counter to account for kernel memory usage.
 295         */
 296        struct res_counter kmem;
 297        /*
 298         * Should the accounting and control be hierarchical, per subtree?
 299         */
 300        bool use_hierarchy;
 301        unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
 302
 303        bool            oom_lock;
 304        atomic_t        under_oom;
 305
 306        atomic_t        refcnt;
 307
 308        int     swappiness;
 309        /* OOM-Killer disable */
 310        int             oom_kill_disable;
 311
 312        /* set when res.limit == memsw.limit */
 313        bool            memsw_is_minimum;
 314
 315        /* protect arrays of thresholds */
 316        struct mutex thresholds_lock;
 317
 318        /* thresholds for memory usage. RCU-protected */
 319        struct mem_cgroup_thresholds thresholds;
 320
 321        /* thresholds for mem+swap usage. RCU-protected */
 322        struct mem_cgroup_thresholds memsw_thresholds;
 323
 324        /* For oom notifier event fd */
 325        struct list_head oom_notify;
 326
 327        /*
 328         * Should we move charges of a task when a task is moved into this
 329         * mem_cgroup ? And what type of charges should we move ?
 330         */
 331        unsigned long   move_charge_at_immigrate;
 332        /*
 333         * set > 0 if pages under this cgroup are moving to other cgroup.
 334         */
 335        atomic_t        moving_account;
 336        /* taken only while moving_account > 0 */
 337        spinlock_t      move_lock;
 338        /*
 339         * percpu counter.
 340         */
 341        struct mem_cgroup_stat_cpu __percpu *stat;
 342        /*
 343         * used when a cpu is offlined or other synchronizations
 344         * See mem_cgroup_read_stat().
 345         */
 346        struct mem_cgroup_stat_cpu nocpu_base;
 347        spinlock_t pcp_counter_lock;
 348
 349        atomic_t        dead_count;
 350#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
 351        struct tcp_memcontrol tcp_mem;
 352#endif
 353#if defined(CONFIG_MEMCG_KMEM)
 354        /* analogous to slab_common's slab_caches list. per-memcg */
 355        struct list_head memcg_slab_caches;
 356        /* Not a spinlock, we can take a lot of time walking the list */
 357        struct mutex slab_caches_mutex;
 358        /* Index in the kmem_cache->memcg_params->memcg_caches array */
 359        int kmemcg_id;
 360#endif
 361
 362        int last_scanned_node;
 363#if MAX_NUMNODES > 1
 364        nodemask_t      scan_nodes;
 365        atomic_t        numainfo_events;
 366        atomic_t        numainfo_updating;
 367#endif
 368
 369        /*
 370         * Per cgroup active and inactive list, similar to the
 371         * per zone LRU lists.
 372         *
 373         * WARNING: This has to be the last element of the struct. Don't
 374         * add new fields after this point.
 375         */
 376        struct mem_cgroup_lru_info info;
 377};
 378
 379static size_t memcg_size(void)
 380{
 381        return sizeof(struct mem_cgroup) +
 382                nr_node_ids * sizeof(struct mem_cgroup_per_node);
 383}
 384
 385/* internal only representation about the status of kmem accounting. */
 386enum {
 387        KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
 388        KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
 389        KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
 390};
 391
 392/* We account when limit is on, but only after call sites are patched */
 393#define KMEM_ACCOUNTED_MASK \
 394                ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
 395
 396#ifdef CONFIG_MEMCG_KMEM
 397static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
 398{
 399        set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 400}
 401
 402static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
 403{
 404        return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
 405}
 406
 407static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
 408{
 409        set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 410}
 411
 412static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
 413{
 414        clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
 415}
 416
 417static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
 418{
 419        if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
 420                set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
 421}
 422
 423static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
 424{
 425        return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
 426                                  &memcg->kmem_account_flags);
 427}
 428#endif
 429
 430/* Stuffs for move charges at task migration. */
 431/*
 432 * Types of charges to be moved. "move_charge_at_immitgrate" and
 433 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
 434 */
 435enum move_type {
 436        MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
 437        MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
 438        NR_MOVE_TYPE,
 439};
 440
 441/* "mc" and its members are protected by cgroup_mutex */
 442static struct move_charge_struct {
 443        spinlock_t        lock; /* for from, to */
 444        struct mem_cgroup *from;
 445        struct mem_cgroup *to;
 446        unsigned long immigrate_flags;
 447        unsigned long precharge;
 448        unsigned long moved_charge;
 449        unsigned long moved_swap;
 450        struct task_struct *moving_task;        /* a task moving charges */
 451        wait_queue_head_t waitq;                /* a waitq for other context */
 452} mc = {
 453        .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
 454        .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
 455};
 456
 457static bool move_anon(void)
 458{
 459        return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
 460}
 461
 462static bool move_file(void)
 463{
 464        return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
 465}
 466
 467/*
 468 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
 469 * limit reclaim to prevent infinite loops, if they ever occur.
 470 */
 471#define MEM_CGROUP_MAX_RECLAIM_LOOPS            100
 472#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
 473
 474enum charge_type {
 475        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 476        MEM_CGROUP_CHARGE_TYPE_ANON,
 477        MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
 478        MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
 479        NR_CHARGE_TYPE,
 480};
 481
 482/* for encoding cft->private value on file */
 483enum res_type {
 484        _MEM,
 485        _MEMSWAP,
 486        _OOM_TYPE,
 487        _KMEM,
 488};
 489
 490#define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
 491#define MEMFILE_TYPE(val)       ((val) >> 16 & 0xffff)
 492#define MEMFILE_ATTR(val)       ((val) & 0xffff)
 493/* Used for OOM nofiier */
 494#define OOM_CONTROL             (0)
 495
 496/*
 497 * Reclaim flags for mem_cgroup_hierarchical_reclaim
 498 */
 499#define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
 500#define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
 501#define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
 502#define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
 503
 504/*
 505 * The memcg_create_mutex will be held whenever a new cgroup is created.
 506 * As a consequence, any change that needs to protect against new child cgroups
 507 * appearing has to hold it as well.
 508 */
 509static DEFINE_MUTEX(memcg_create_mutex);
 510
 511static void mem_cgroup_get(struct mem_cgroup *memcg);
 512static void mem_cgroup_put(struct mem_cgroup *memcg);
 513
 514static inline
 515struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
 516{
 517        return container_of(s, struct mem_cgroup, css);
 518}
 519
 520/* Some nice accessors for the vmpressure. */
 521struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
 522{
 523        if (!memcg)
 524                memcg = root_mem_cgroup;
 525        return &memcg->vmpressure;
 526}
 527
 528struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
 529{
 530        return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
 531}
 532
 533struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
 534{
 535        return &mem_cgroup_from_css(css)->vmpressure;
 536}
 537
 538static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
 539{
 540        return (memcg == root_mem_cgroup);
 541}
 542
 543/* Writing them here to avoid exposing memcg's inner layout */
 544#if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
 545
 546void sock_update_memcg(struct sock *sk)
 547{
 548        if (mem_cgroup_sockets_enabled) {
 549                struct mem_cgroup *memcg;
 550                struct cg_proto *cg_proto;
 551
 552                BUG_ON(!sk->sk_prot->proto_cgroup);
 553
 554                /* Socket cloning can throw us here with sk_cgrp already
 555                 * filled. It won't however, necessarily happen from
 556                 * process context. So the test for root memcg given
 557                 * the current task's memcg won't help us in this case.
 558                 *
 559                 * Respecting the original socket's memcg is a better
 560                 * decision in this case.
 561                 */
 562                if (sk->sk_cgrp) {
 563                        BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
 564                        mem_cgroup_get(sk->sk_cgrp->memcg);
 565                        return;
 566                }
 567
 568                rcu_read_lock();
 569                memcg = mem_cgroup_from_task(current);
 570                cg_proto = sk->sk_prot->proto_cgroup(memcg);
 571                if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
 572                        mem_cgroup_get(memcg);
 573                        sk->sk_cgrp = cg_proto;
 574                }
 575                rcu_read_unlock();
 576        }
 577}
 578EXPORT_SYMBOL(sock_update_memcg);
 579
 580void sock_release_memcg(struct sock *sk)
 581{
 582        if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
 583                struct mem_cgroup *memcg;
 584                WARN_ON(!sk->sk_cgrp->memcg);
 585                memcg = sk->sk_cgrp->memcg;
 586                mem_cgroup_put(memcg);
 587        }
 588}
 589
 590struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
 591{
 592        if (!memcg || mem_cgroup_is_root(memcg))
 593                return NULL;
 594
 595        return &memcg->tcp_mem.cg_proto;
 596}
 597EXPORT_SYMBOL(tcp_proto_cgroup);
 598
 599static void disarm_sock_keys(struct mem_cgroup *memcg)
 600{
 601        if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
 602                return;
 603        static_key_slow_dec(&memcg_socket_limit_enabled);
 604}
 605#else
 606static void disarm_sock_keys(struct mem_cgroup *memcg)
 607{
 608}
 609#endif
 610
 611#ifdef CONFIG_MEMCG_KMEM
 612/*
 613 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
 614 * There are two main reasons for not using the css_id for this:
 615 *  1) this works better in sparse environments, where we have a lot of memcgs,
 616 *     but only a few kmem-limited. Or also, if we have, for instance, 200
 617 *     memcgs, and none but the 200th is kmem-limited, we'd have to have a
 618 *     200 entry array for that.
 619 *
 620 *  2) In order not to violate the cgroup API, we would like to do all memory
 621 *     allocation in ->create(). At that point, we haven't yet allocated the
 622 *     css_id. Having a separate index prevents us from messing with the cgroup
 623 *     core for this
 624 *
 625 * The current size of the caches array is stored in
 626 * memcg_limited_groups_array_size.  It will double each time we have to
 627 * increase it.
 628 */
 629static DEFINE_IDA(kmem_limited_groups);
 630int memcg_limited_groups_array_size;
 631
 632/*
 633 * MIN_SIZE is different than 1, because we would like to avoid going through
 634 * the alloc/free process all the time. In a small machine, 4 kmem-limited
 635 * cgroups is a reasonable guess. In the future, it could be a parameter or
 636 * tunable, but that is strictly not necessary.
 637 *
 638 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
 639 * this constant directly from cgroup, but it is understandable that this is
 640 * better kept as an internal representation in cgroup.c. In any case, the
 641 * css_id space is not getting any smaller, and we don't have to necessarily
 642 * increase ours as well if it increases.
 643 */
 644#define MEMCG_CACHES_MIN_SIZE 4
 645#define MEMCG_CACHES_MAX_SIZE 65535
 646
 647/*
 648 * A lot of the calls to the cache allocation functions are expected to be
 649 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
 650 * conditional to this static branch, we'll have to allow modules that does
 651 * kmem_cache_alloc and the such to see this symbol as well
 652 */
 653struct static_key memcg_kmem_enabled_key;
 654EXPORT_SYMBOL(memcg_kmem_enabled_key);
 655
 656static void disarm_kmem_keys(struct mem_cgroup *memcg)
 657{
 658        if (memcg_kmem_is_active(memcg)) {
 659                static_key_slow_dec(&memcg_kmem_enabled_key);
 660                ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
 661        }
 662        /*
 663         * This check can't live in kmem destruction function,
 664         * since the charges will outlive the cgroup
 665         */
 666        WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
 667}
 668#else
 669static void disarm_kmem_keys(struct mem_cgroup *memcg)
 670{
 671}
 672#endif /* CONFIG_MEMCG_KMEM */
 673
 674static void disarm_static_keys(struct mem_cgroup *memcg)
 675{
 676        disarm_sock_keys(memcg);
 677        disarm_kmem_keys(memcg);
 678}
 679
 680static void drain_all_stock_async(struct mem_cgroup *memcg);
 681
 682static struct mem_cgroup_per_zone *
 683mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
 684{
 685        VM_BUG_ON((unsigned)nid >= nr_node_ids);
 686        return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
 687}
 688
 689struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
 690{
 691        return &memcg->css;
 692}
 693
 694static struct mem_cgroup_per_zone *
 695page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
 696{
 697        int nid = page_to_nid(page);
 698        int zid = page_zonenum(page);
 699
 700        return mem_cgroup_zoneinfo(memcg, nid, zid);
 701}
 702
 703static struct mem_cgroup_tree_per_zone *
 704soft_limit_tree_node_zone(int nid, int zid)
 705{
 706        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 707}
 708
 709static struct mem_cgroup_tree_per_zone *
 710soft_limit_tree_from_page(struct page *page)
 711{
 712        int nid = page_to_nid(page);
 713        int zid = page_zonenum(page);
 714
 715        return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
 716}
 717
 718static void
 719__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
 720                                struct mem_cgroup_per_zone *mz,
 721                                struct mem_cgroup_tree_per_zone *mctz,
 722                                unsigned long long new_usage_in_excess)
 723{
 724        struct rb_node **p = &mctz->rb_root.rb_node;
 725        struct rb_node *parent = NULL;
 726        struct mem_cgroup_per_zone *mz_node;
 727
 728        if (mz->on_tree)
 729                return;
 730
 731        mz->usage_in_excess = new_usage_in_excess;
 732        if (!mz->usage_in_excess)
 733                return;
 734        while (*p) {
 735                parent = *p;
 736                mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
 737                                        tree_node);
 738                if (mz->usage_in_excess < mz_node->usage_in_excess)
 739                        p = &(*p)->rb_left;
 740                /*
 741                 * We can't avoid mem cgroups that are over their soft
 742                 * limit by the same amount
 743                 */
 744                else if (mz->usage_in_excess >= mz_node->usage_in_excess)
 745                        p = &(*p)->rb_right;
 746        }
 747        rb_link_node(&mz->tree_node, parent, p);
 748        rb_insert_color(&mz->tree_node, &mctz->rb_root);
 749        mz->on_tree = true;
 750}
 751
 752static void
 753__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 754                                struct mem_cgroup_per_zone *mz,
 755                                struct mem_cgroup_tree_per_zone *mctz)
 756{
 757        if (!mz->on_tree)
 758                return;
 759        rb_erase(&mz->tree_node, &mctz->rb_root);
 760        mz->on_tree = false;
 761}
 762
 763static void
 764mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
 765                                struct mem_cgroup_per_zone *mz,
 766                                struct mem_cgroup_tree_per_zone *mctz)
 767{
 768        spin_lock(&mctz->lock);
 769        __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 770        spin_unlock(&mctz->lock);
 771}
 772
 773
 774static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
 775{
 776        unsigned long long excess;
 777        struct mem_cgroup_per_zone *mz;
 778        struct mem_cgroup_tree_per_zone *mctz;
 779        int nid = page_to_nid(page);
 780        int zid = page_zonenum(page);
 781        mctz = soft_limit_tree_from_page(page);
 782
 783        /*
 784         * Necessary to update all ancestors when hierarchy is used.
 785         * because their event counter is not touched.
 786         */
 787        for (; memcg; memcg = parent_mem_cgroup(memcg)) {
 788                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 789                excess = res_counter_soft_limit_excess(&memcg->res);
 790                /*
 791                 * We have to update the tree if mz is on RB-tree or
 792                 * mem is over its softlimit.
 793                 */
 794                if (excess || mz->on_tree) {
 795                        spin_lock(&mctz->lock);
 796                        /* if on-tree, remove it */
 797                        if (mz->on_tree)
 798                                __mem_cgroup_remove_exceeded(memcg, mz, mctz);
 799                        /*
 800                         * Insert again. mz->usage_in_excess will be updated.
 801                         * If excess is 0, no tree ops.
 802                         */
 803                        __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
 804                        spin_unlock(&mctz->lock);
 805                }
 806        }
 807}
 808
 809static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
 810{
 811        int node, zone;
 812        struct mem_cgroup_per_zone *mz;
 813        struct mem_cgroup_tree_per_zone *mctz;
 814
 815        for_each_node(node) {
 816                for (zone = 0; zone < MAX_NR_ZONES; zone++) {
 817                        mz = mem_cgroup_zoneinfo(memcg, node, zone);
 818                        mctz = soft_limit_tree_node_zone(node, zone);
 819                        mem_cgroup_remove_exceeded(memcg, mz, mctz);
 820                }
 821        }
 822}
 823
 824static struct mem_cgroup_per_zone *
 825__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 826{
 827        struct rb_node *rightmost = NULL;
 828        struct mem_cgroup_per_zone *mz;
 829
 830retry:
 831        mz = NULL;
 832        rightmost = rb_last(&mctz->rb_root);
 833        if (!rightmost)
 834                goto done;              /* Nothing to reclaim from */
 835
 836        mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
 837        /*
 838         * Remove the node now but someone else can add it back,
 839         * we will to add it back at the end of reclaim to its correct
 840         * position in the tree.
 841         */
 842        __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
 843        if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
 844                !css_tryget(&mz->memcg->css))
 845                goto retry;
 846done:
 847        return mz;
 848}
 849
 850static struct mem_cgroup_per_zone *
 851mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
 852{
 853        struct mem_cgroup_per_zone *mz;
 854
 855        spin_lock(&mctz->lock);
 856        mz = __mem_cgroup_largest_soft_limit_node(mctz);
 857        spin_unlock(&mctz->lock);
 858        return mz;
 859}
 860
 861/*
 862 * Implementation Note: reading percpu statistics for memcg.
 863 *
 864 * Both of vmstat[] and percpu_counter has threshold and do periodic
 865 * synchronization to implement "quick" read. There are trade-off between
 866 * reading cost and precision of value. Then, we may have a chance to implement
 867 * a periodic synchronizion of counter in memcg's counter.
 868 *
 869 * But this _read() function is used for user interface now. The user accounts
 870 * memory usage by memory cgroup and he _always_ requires exact value because
 871 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
 872 * have to visit all online cpus and make sum. So, for now, unnecessary
 873 * synchronization is not implemented. (just implemented for cpu hotplug)
 874 *
 875 * If there are kernel internal actions which can make use of some not-exact
 876 * value, and reading all cpu value can be performance bottleneck in some
 877 * common workload, threashold and synchonization as vmstat[] should be
 878 * implemented.
 879 */
 880static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
 881                                 enum mem_cgroup_stat_index idx)
 882{
 883        long val = 0;
 884        int cpu;
 885
 886        get_online_cpus();
 887        for_each_online_cpu(cpu)
 888                val += per_cpu(memcg->stat->count[idx], cpu);
 889#ifdef CONFIG_HOTPLUG_CPU
 890        spin_lock(&memcg->pcp_counter_lock);
 891        val += memcg->nocpu_base.count[idx];
 892        spin_unlock(&memcg->pcp_counter_lock);
 893#endif
 894        put_online_cpus();
 895        return val;
 896}
 897
 898static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
 899                                         bool charge)
 900{
 901        int val = (charge) ? 1 : -1;
 902        this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
 903}
 904
 905static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
 906                                            enum mem_cgroup_events_index idx)
 907{
 908        unsigned long val = 0;
 909        int cpu;
 910
 911        for_each_online_cpu(cpu)
 912                val += per_cpu(memcg->stat->events[idx], cpu);
 913#ifdef CONFIG_HOTPLUG_CPU
 914        spin_lock(&memcg->pcp_counter_lock);
 915        val += memcg->nocpu_base.events[idx];
 916        spin_unlock(&memcg->pcp_counter_lock);
 917#endif
 918        return val;
 919}
 920
 921static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
 922                                         struct page *page,
 923                                         bool anon, int nr_pages)
 924{
 925        preempt_disable();
 926
 927        /*
 928         * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
 929         * counted as CACHE even if it's on ANON LRU.
 930         */
 931        if (anon)
 932                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
 933                                nr_pages);
 934        else
 935                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
 936                                nr_pages);
 937
 938        if (PageTransHuge(page))
 939                __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
 940                                nr_pages);
 941
 942        /* pagein of a big page is an event. So, ignore page size */
 943        if (nr_pages > 0)
 944                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
 945        else {
 946                __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
 947                nr_pages = -nr_pages; /* for event */
 948        }
 949
 950        __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
 951
 952        preempt_enable();
 953}
 954
 955unsigned long
 956mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 957{
 958        struct mem_cgroup_per_zone *mz;
 959
 960        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
 961        return mz->lru_size[lru];
 962}
 963
 964static unsigned long
 965mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
 966                        unsigned int lru_mask)
 967{
 968        struct mem_cgroup_per_zone *mz;
 969        enum lru_list lru;
 970        unsigned long ret = 0;
 971
 972        mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 973
 974        for_each_lru(lru) {
 975                if (BIT(lru) & lru_mask)
 976                        ret += mz->lru_size[lru];
 977        }
 978        return ret;
 979}
 980
 981static unsigned long
 982mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
 983                        int nid, unsigned int lru_mask)
 984{
 985        u64 total = 0;
 986        int zid;
 987
 988        for (zid = 0; zid < MAX_NR_ZONES; zid++)
 989                total += mem_cgroup_zone_nr_lru_pages(memcg,
 990                                                nid, zid, lru_mask);
 991
 992        return total;
 993}
 994
 995static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
 996                        unsigned int lru_mask)
 997{
 998        int nid;
 999        u64 total = 0;
1000
1001        for_each_node_state(nid, N_MEMORY)
1002                total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1003        return total;
1004}
1005
1006static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1007                                       enum mem_cgroup_events_target target)
1008{
1009        unsigned long val, next;
1010
1011        val = __this_cpu_read(memcg->stat->nr_page_events);
1012        next = __this_cpu_read(memcg->stat->targets[target]);
1013        /* from time_after() in jiffies.h */
1014        if ((long)next - (long)val < 0) {
1015                switch (target) {
1016                case MEM_CGROUP_TARGET_THRESH:
1017                        next = val + THRESHOLDS_EVENTS_TARGET;
1018                        break;
1019                case MEM_CGROUP_TARGET_SOFTLIMIT:
1020                        next = val + SOFTLIMIT_EVENTS_TARGET;
1021                        break;
1022                case MEM_CGROUP_TARGET_NUMAINFO:
1023                        next = val + NUMAINFO_EVENTS_TARGET;
1024                        break;
1025                default:
1026                        break;
1027                }
1028                __this_cpu_write(memcg->stat->targets[target], next);
1029                return true;
1030        }
1031        return false;
1032}
1033
1034/*
1035 * Check events in order.
1036 *
1037 */
1038static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1039{
1040        preempt_disable();
1041        /* threshold event is triggered in finer grain than soft limit */
1042        if (unlikely(mem_cgroup_event_ratelimit(memcg,
1043                                                MEM_CGROUP_TARGET_THRESH))) {
1044                bool do_softlimit;
1045                bool do_numainfo __maybe_unused;
1046
1047                do_softlimit = mem_cgroup_event_ratelimit(memcg,
1048                                                MEM_CGROUP_TARGET_SOFTLIMIT);
1049#if MAX_NUMNODES > 1
1050                do_numainfo = mem_cgroup_event_ratelimit(memcg,
1051                                                MEM_CGROUP_TARGET_NUMAINFO);
1052#endif
1053                preempt_enable();
1054
1055                mem_cgroup_threshold(memcg);
1056                if (unlikely(do_softlimit))
1057                        mem_cgroup_update_tree(memcg, page);
1058#if MAX_NUMNODES > 1
1059                if (unlikely(do_numainfo))
1060                        atomic_inc(&memcg->numainfo_events);
1061#endif
1062        } else
1063                preempt_enable();
1064}
1065
1066struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1067{
1068        return mem_cgroup_from_css(
1069                cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1070}
1071
1072struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1073{
1074        /*
1075         * mm_update_next_owner() may clear mm->owner to NULL
1076         * if it races with swapoff, page migration, etc.
1077         * So this can be called with p == NULL.
1078         */
1079        if (unlikely(!p))
1080                return NULL;
1081
1082        return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1083}
1084
1085struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1086{
1087        struct mem_cgroup *memcg = NULL;
1088
1089        if (!mm)
1090                return NULL;
1091        /*
1092         * Because we have no locks, mm->owner's may be being moved to other
1093         * cgroup. We use css_tryget() here even if this looks
1094         * pessimistic (rather than adding locks here).
1095         */
1096        rcu_read_lock();
1097        do {
1098                memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1099                if (unlikely(!memcg))
1100                        break;
1101        } while (!css_tryget(&memcg->css));
1102        rcu_read_unlock();
1103        return memcg;
1104}
1105
1106/*
1107 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1108 * ref. count) or NULL if the whole root's subtree has been visited.
1109 *
1110 * helper function to be used by mem_cgroup_iter
1111 */
1112static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1113                struct mem_cgroup *last_visited)
1114{
1115        struct cgroup *prev_cgroup, *next_cgroup;
1116
1117        /*
1118         * Root is not visited by cgroup iterators so it needs an
1119         * explicit visit.
1120         */
1121        if (!last_visited)
1122                return root;
1123
1124        prev_cgroup = (last_visited == root) ? NULL
1125                : last_visited->css.cgroup;
1126skip_node:
1127        next_cgroup = cgroup_next_descendant_pre(
1128                        prev_cgroup, root->css.cgroup);
1129
1130        /*
1131         * Even if we found a group we have to make sure it is
1132         * alive. css && !memcg means that the groups should be
1133         * skipped and we should continue the tree walk.
1134         * last_visited css is safe to use because it is
1135         * protected by css_get and the tree walk is rcu safe.
1136         */
1137        if (next_cgroup) {
1138                struct mem_cgroup *mem = mem_cgroup_from_cont(
1139                                next_cgroup);
1140                if (css_tryget(&mem->css))
1141                        return mem;
1142                else {
1143                        prev_cgroup = next_cgroup;
1144                        goto skip_node;
1145                }
1146        }
1147
1148        return NULL;
1149}
1150
1151/**
1152 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1153 * @root: hierarchy root
1154 * @prev: previously returned memcg, NULL on first invocation
1155 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1156 *
1157 * Returns references to children of the hierarchy below @root, or
1158 * @root itself, or %NULL after a full round-trip.
1159 *
1160 * Caller must pass the return value in @prev on subsequent
1161 * invocations for reference counting, or use mem_cgroup_iter_break()
1162 * to cancel a hierarchy walk before the round-trip is complete.
1163 *
1164 * Reclaimers can specify a zone and a priority level in @reclaim to
1165 * divide up the memcgs in the hierarchy among all concurrent
1166 * reclaimers operating on the same zone and priority.
1167 */
1168struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1169                                   struct mem_cgroup *prev,
1170                                   struct mem_cgroup_reclaim_cookie *reclaim)
1171{
1172        struct mem_cgroup *memcg = NULL;
1173        struct mem_cgroup *last_visited = NULL;
1174        unsigned long uninitialized_var(dead_count);
1175
1176        if (mem_cgroup_disabled())
1177                return NULL;
1178
1179        if (!root)
1180                root = root_mem_cgroup;
1181
1182        if (prev && !reclaim)
1183                last_visited = prev;
1184
1185        if (!root->use_hierarchy && root != root_mem_cgroup) {
1186                if (prev)
1187                        goto out_css_put;
1188                return root;
1189        }
1190
1191        rcu_read_lock();
1192        while (!memcg) {
1193                struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1194
1195                if (reclaim) {
1196                        int nid = zone_to_nid(reclaim->zone);
1197                        int zid = zone_idx(reclaim->zone);
1198                        struct mem_cgroup_per_zone *mz;
1199
1200                        mz = mem_cgroup_zoneinfo(root, nid, zid);
1201                        iter = &mz->reclaim_iter[reclaim->priority];
1202                        if (prev && reclaim->generation != iter->generation) {
1203                                iter->last_visited = NULL;
1204                                goto out_unlock;
1205                        }
1206
1207                        /*
1208                         * If the dead_count mismatches, a destruction
1209                         * has happened or is happening concurrently.
1210                         * If the dead_count matches, a destruction
1211                         * might still happen concurrently, but since
1212                         * we checked under RCU, that destruction
1213                         * won't free the object until we release the
1214                         * RCU reader lock.  Thus, the dead_count
1215                         * check verifies the pointer is still valid,
1216                         * css_tryget() verifies the cgroup pointed to
1217                         * is alive.
1218                         */
1219                        dead_count = atomic_read(&root->dead_count);
1220                        if (dead_count == iter->last_dead_count) {
1221                                smp_rmb();
1222                                last_visited = iter->last_visited;
1223                                if (last_visited &&
1224                                    !css_tryget(&last_visited->css))
1225                                        last_visited = NULL;
1226                        }
1227                }
1228
1229                memcg = __mem_cgroup_iter_next(root, last_visited);
1230
1231                if (reclaim) {
1232                        if (last_visited)
1233                                css_put(&last_visited->css);
1234
1235                        iter->last_visited = memcg;
1236                        smp_wmb();
1237                        iter->last_dead_count = dead_count;
1238
1239                        if (!memcg)
1240                                iter->generation++;
1241                        else if (!prev && memcg)
1242                                reclaim->generation = iter->generation;
1243                }
1244
1245                if (prev && !memcg)
1246                        goto out_unlock;
1247        }
1248out_unlock:
1249        rcu_read_unlock();
1250out_css_put:
1251        if (prev && prev != root)
1252                css_put(&prev->css);
1253
1254        return memcg;
1255}
1256
1257/**
1258 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1259 * @root: hierarchy root
1260 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1261 */
1262void mem_cgroup_iter_break(struct mem_cgroup *root,
1263                           struct mem_cgroup *prev)
1264{
1265        if (!root)
1266                root = root_mem_cgroup;
1267        if (prev && prev != root)
1268                css_put(&prev->css);
1269}
1270
1271/*
1272 * Iteration constructs for visiting all cgroups (under a tree).  If
1273 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1274 * be used for reference counting.
1275 */
1276#define for_each_mem_cgroup_tree(iter, root)            \
1277        for (iter = mem_cgroup_iter(root, NULL, NULL);  \
1278             iter != NULL;                              \
1279             iter = mem_cgroup_iter(root, iter, NULL))
1280
1281#define for_each_mem_cgroup(iter)                       \
1282        for (iter = mem_cgroup_iter(NULL, NULL, NULL);  \
1283             iter != NULL;                              \
1284             iter = mem_cgroup_iter(NULL, iter, NULL))
1285
1286void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1287{
1288        struct mem_cgroup *memcg;
1289
1290        rcu_read_lock();
1291        memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1292        if (unlikely(!memcg))
1293                goto out;
1294
1295        switch (idx) {
1296        case PGFAULT:
1297                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1298                break;
1299        case PGMAJFAULT:
1300                this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1301                break;
1302        default:
1303                BUG();
1304        }
1305out:
1306        rcu_read_unlock();
1307}
1308EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1309
1310/**
1311 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1312 * @zone: zone of the wanted lruvec
1313 * @memcg: memcg of the wanted lruvec
1314 *
1315 * Returns the lru list vector holding pages for the given @zone and
1316 * @mem.  This can be the global zone lruvec, if the memory controller
1317 * is disabled.
1318 */
1319struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1320                                      struct mem_cgroup *memcg)
1321{
1322        struct mem_cgroup_per_zone *mz;
1323        struct lruvec *lruvec;
1324
1325        if (mem_cgroup_disabled()) {
1326                lruvec = &zone->lruvec;
1327                goto out;
1328        }
1329
1330        mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1331        lruvec = &mz->lruvec;
1332out:
1333        /*
1334         * Since a node can be onlined after the mem_cgroup was created,
1335         * we have to be prepared to initialize lruvec->zone here;
1336         * and if offlined then reonlined, we need to reinitialize it.
1337         */
1338        if (unlikely(lruvec->zone != zone))
1339                lruvec->zone = zone;
1340        return lruvec;
1341}
1342
1343/*
1344 * Following LRU functions are allowed to be used without PCG_LOCK.
1345 * Operations are called by routine of global LRU independently from memcg.
1346 * What we have to take care of here is validness of pc->mem_cgroup.
1347 *
1348 * Changes to pc->mem_cgroup happens when
1349 * 1. charge
1350 * 2. moving account
1351 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1352 * It is added to LRU before charge.
1353 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1354 * When moving account, the page is not on LRU. It's isolated.
1355 */
1356
1357/**
1358 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1359 * @page: the page
1360 * @zone: zone of the page
1361 */
1362struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1363{
1364        struct mem_cgroup_per_zone *mz;
1365        struct mem_cgroup *memcg;
1366        struct page_cgroup *pc;
1367        struct lruvec *lruvec;
1368
1369        if (mem_cgroup_disabled()) {
1370                lruvec = &zone->lruvec;
1371                goto out;
1372        }
1373
1374        pc = lookup_page_cgroup(page);
1375        memcg = pc->mem_cgroup;
1376
1377        /*
1378         * Surreptitiously switch any uncharged offlist page to root:
1379         * an uncharged page off lru does nothing to secure
1380         * its former mem_cgroup from sudden removal.
1381         *
1382         * Our caller holds lru_lock, and PageCgroupUsed is updated
1383         * under page_cgroup lock: between them, they make all uses
1384         * of pc->mem_cgroup safe.
1385         */
1386        if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1387                pc->mem_cgroup = memcg = root_mem_cgroup;
1388
1389        mz = page_cgroup_zoneinfo(memcg, page);
1390        lruvec = &mz->lruvec;
1391out:
1392        /*
1393         * Since a node can be onlined after the mem_cgroup was created,
1394         * we have to be prepared to initialize lruvec->zone here;
1395         * and if offlined then reonlined, we need to reinitialize it.
1396         */
1397        if (unlikely(lruvec->zone != zone))
1398                lruvec->zone = zone;
1399        return lruvec;
1400}
1401
1402/**
1403 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1404 * @lruvec: mem_cgroup per zone lru vector
1405 * @lru: index of lru list the page is sitting on
1406 * @nr_pages: positive when adding or negative when removing
1407 *
1408 * This function must be called when a page is added to or removed from an
1409 * lru list.
1410 */
1411void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1412                                int nr_pages)
1413{
1414        struct mem_cgroup_per_zone *mz;
1415        unsigned long *lru_size;
1416
1417        if (mem_cgroup_disabled())
1418                return;
1419
1420        mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1421        lru_size = mz->lru_size + lru;
1422        *lru_size += nr_pages;
1423        VM_BUG_ON((long)(*lru_size) < 0);
1424}
1425
1426/*
1427 * Checks whether given mem is same or in the root_mem_cgroup's
1428 * hierarchy subtree
1429 */
1430bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1431                                  struct mem_cgroup *memcg)
1432{
1433        if (root_memcg == memcg)
1434                return true;
1435        if (!root_memcg->use_hierarchy || !memcg)
1436                return false;
1437        return css_is_ancestor(&memcg->css, &root_memcg->css);
1438}
1439
1440static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1441                                       struct mem_cgroup *memcg)
1442{
1443        bool ret;
1444
1445        rcu_read_lock();
1446        ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1447        rcu_read_unlock();
1448        return ret;
1449}
1450
1451int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1452{
1453        int ret;
1454        struct mem_cgroup *curr = NULL;
1455        struct task_struct *p;
1456
1457        p = find_lock_task_mm(task);
1458        if (p) {
1459                curr = try_get_mem_cgroup_from_mm(p->mm);
1460                task_unlock(p);
1461        } else {
1462                /*
1463                 * All threads may have already detached their mm's, but the oom
1464                 * killer still needs to detect if they have already been oom
1465                 * killed to prevent needlessly killing additional tasks.
1466                 */
1467                task_lock(task);
1468                curr = mem_cgroup_from_task(task);
1469                if (curr)
1470                        css_get(&curr->css);
1471                task_unlock(task);
1472        }
1473        if (!curr)
1474                return 0;
1475        /*
1476         * We should check use_hierarchy of "memcg" not "curr". Because checking
1477         * use_hierarchy of "curr" here make this function true if hierarchy is
1478         * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1479         * hierarchy(even if use_hierarchy is disabled in "memcg").
1480         */
1481        ret = mem_cgroup_same_or_subtree(memcg, curr);
1482        css_put(&curr->css);
1483        return ret;
1484}
1485
1486int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1487{
1488        unsigned long inactive_ratio;
1489        unsigned long inactive;
1490        unsigned long active;
1491        unsigned long gb;
1492
1493        inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1494        active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1495
1496        gb = (inactive + active) >> (30 - PAGE_SHIFT);
1497        if (gb)
1498                inactive_ratio = int_sqrt(10 * gb);
1499        else
1500                inactive_ratio = 1;
1501
1502        return inactive * inactive_ratio < active;
1503}
1504
1505#define mem_cgroup_from_res_counter(counter, member)    \
1506        container_of(counter, struct mem_cgroup, member)
1507
1508/**
1509 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1510 * @memcg: the memory cgroup
1511 *
1512 * Returns the maximum amount of memory @mem can be charged with, in
1513 * pages.
1514 */
1515static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1516{
1517        unsigned long long margin;
1518
1519        margin = res_counter_margin(&memcg->res);
1520        if (do_swap_account)
1521                margin = min(margin, res_counter_margin(&memcg->memsw));
1522        return margin >> PAGE_SHIFT;
1523}
1524
1525int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1526{
1527        struct cgroup *cgrp = memcg->css.cgroup;
1528
1529        /* root ? */
1530        if (cgrp->parent == NULL)
1531                return vm_swappiness;
1532
1533        return memcg->swappiness;
1534}
1535
1536/*
1537 * memcg->moving_account is used for checking possibility that some thread is
1538 * calling move_account(). When a thread on CPU-A starts moving pages under
1539 * a memcg, other threads should check memcg->moving_account under
1540 * rcu_read_lock(), like this:
1541 *
1542 *         CPU-A                                    CPU-B
1543 *                                              rcu_read_lock()
1544 *         memcg->moving_account+1              if (memcg->mocing_account)
1545 *                                                   take heavy locks.
1546 *         synchronize_rcu()                    update something.
1547 *                                              rcu_read_unlock()
1548 *         start move here.
1549 */
1550
1551/* for quick checking without looking up memcg */
1552atomic_t memcg_moving __read_mostly;
1553
1554static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1555{
1556        atomic_inc(&memcg_moving);
1557        atomic_inc(&memcg->moving_account);
1558        synchronize_rcu();
1559}
1560
1561static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1562{
1563        /*
1564         * Now, mem_cgroup_clear_mc() may call this function with NULL.
1565         * We check NULL in callee rather than caller.
1566         */
1567        if (memcg) {
1568                atomic_dec(&memcg_moving);
1569                atomic_dec(&memcg->moving_account);
1570        }
1571}
1572
1573/*
1574 * 2 routines for checking "mem" is under move_account() or not.
1575 *
1576 * mem_cgroup_stolen() -  checking whether a cgroup is mc.from or not. This
1577 *                        is used for avoiding races in accounting.  If true,
1578 *                        pc->mem_cgroup may be overwritten.
1579 *
1580 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1581 *                        under hierarchy of moving cgroups. This is for
1582 *                        waiting at hith-memory prressure caused by "move".
1583 */
1584
1585static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1586{
1587        VM_BUG_ON(!rcu_read_lock_held());
1588        return atomic_read(&memcg->moving_account) > 0;
1589}
1590
1591static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1592{
1593        struct mem_cgroup *from;
1594        struct mem_cgroup *to;
1595        bool ret = false;
1596        /*
1597         * Unlike task_move routines, we access mc.to, mc.from not under
1598         * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1599         */
1600        spin_lock(&mc.lock);
1601        from = mc.from;
1602        to = mc.to;
1603        if (!from)
1604                goto unlock;
1605
1606        ret = mem_cgroup_same_or_subtree(memcg, from)
1607                || mem_cgroup_same_or_subtree(memcg, to);
1608unlock:
1609        spin_unlock(&mc.lock);
1610        return ret;
1611}
1612
1613static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1614{
1615        if (mc.moving_task && current != mc.moving_task) {
1616                if (mem_cgroup_under_move(memcg)) {
1617                        DEFINE_WAIT(wait);
1618                        prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1619                        /* moving charge context might have finished. */
1620                        if (mc.moving_task)
1621                                schedule();
1622                        finish_wait(&mc.waitq, &wait);
1623                        return true;
1624                }
1625        }
1626        return false;
1627}
1628
1629/*
1630 * Take this lock when
1631 * - a code tries to modify page's memcg while it's USED.
1632 * - a code tries to modify page state accounting in a memcg.
1633 * see mem_cgroup_stolen(), too.
1634 */
1635static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1636                                  unsigned long *flags)
1637{
1638        spin_lock_irqsave(&memcg->move_lock, *flags);
1639}
1640
1641static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1642                                unsigned long *flags)
1643{
1644        spin_unlock_irqrestore(&memcg->move_lock, *flags);
1645}
1646
1647#define K(x) ((x) << (PAGE_SHIFT-10))
1648/**
1649 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1650 * @memcg: The memory cgroup that went over limit
1651 * @p: Task that is going to be killed
1652 *
1653 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1654 * enabled
1655 */
1656void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1657{
1658        struct cgroup *task_cgrp;
1659        struct cgroup *mem_cgrp;
1660        /*
1661         * Need a buffer in BSS, can't rely on allocations. The code relies
1662         * on the assumption that OOM is serialized for memory controller.
1663         * If this assumption is broken, revisit this code.
1664         */
1665        static char memcg_name[PATH_MAX];
1666        int ret;
1667        struct mem_cgroup *iter;
1668        unsigned int i;
1669
1670        if (!p)
1671                return;
1672
1673        rcu_read_lock();
1674
1675        mem_cgrp = memcg->css.cgroup;
1676        task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1677
1678        ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1679        if (ret < 0) {
1680                /*
1681                 * Unfortunately, we are unable to convert to a useful name
1682                 * But we'll still print out the usage information
1683                 */
1684                rcu_read_unlock();
1685                goto done;
1686        }
1687        rcu_read_unlock();
1688
1689        pr_info("Task in %s killed", memcg_name);
1690
1691        rcu_read_lock();
1692        ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1693        if (ret < 0) {
1694                rcu_read_unlock();
1695                goto done;
1696        }
1697        rcu_read_unlock();
1698
1699        /*
1700         * Continues from above, so we don't need an KERN_ level
1701         */
1702        pr_cont(" as a result of limit of %s\n", memcg_name);
1703done:
1704
1705        pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1706                res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1707                res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1708                res_counter_read_u64(&memcg->res, RES_FAILCNT));
1709        pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1710                res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1711                res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1712                res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1713        pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1714                res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1715                res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1716                res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1717
1718        for_each_mem_cgroup_tree(iter, memcg) {
1719                pr_info("Memory cgroup stats");
1720
1721                rcu_read_lock();
1722                ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1723                if (!ret)
1724                        pr_cont(" for %s", memcg_name);
1725                rcu_read_unlock();
1726                pr_cont(":");
1727
1728                for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1729                        if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1730                                continue;
1731                        pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1732                                K(mem_cgroup_read_stat(iter, i)));
1733                }
1734
1735                for (i = 0; i < NR_LRU_LISTS; i++)
1736                        pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1737                                K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1738
1739                pr_cont("\n");
1740        }
1741}
1742
1743/*
1744 * This function returns the number of memcg under hierarchy tree. Returns
1745 * 1(self count) if no children.
1746 */
1747static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1748{
1749        int num = 0;
1750        struct mem_cgroup *iter;
1751
1752        for_each_mem_cgroup_tree(iter, memcg)
1753                num++;
1754        return num;
1755}
1756
1757/*
1758 * Return the memory (and swap, if configured) limit for a memcg.
1759 */
1760static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1761{
1762        u64 limit;
1763
1764        limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1765
1766        /*
1767         * Do not consider swap space if we cannot swap due to swappiness
1768         */
1769        if (mem_cgroup_swappiness(memcg)) {
1770                u64 memsw;
1771
1772                limit += total_swap_pages << PAGE_SHIFT;
1773                memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1774
1775                /*
1776                 * If memsw is finite and limits the amount of swap space
1777                 * available to this memcg, return that limit.
1778                 */
1779                limit = min(limit, memsw);
1780        }
1781
1782        return limit;
1783}
1784
1785static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1786                                     int order)
1787{
1788        struct mem_cgroup *iter;
1789        unsigned long chosen_points = 0;
1790        unsigned long totalpages;
1791        unsigned int points = 0;
1792        struct task_struct *chosen = NULL;
1793
1794        /*
1795         * If current has a pending SIGKILL or is exiting, then automatically
1796         * select it.  The goal is to allow it to allocate so that it may
1797         * quickly exit and free its memory.
1798         */
1799        if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1800                set_thread_flag(TIF_MEMDIE);
1801                return;
1802        }
1803
1804        check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1805        totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1806        for_each_mem_cgroup_tree(iter, memcg) {
1807                struct cgroup *cgroup = iter->css.cgroup;
1808                struct cgroup_iter it;
1809                struct task_struct *task;
1810
1811                cgroup_iter_start(cgroup, &it);
1812                while ((task = cgroup_iter_next(cgroup, &it))) {
1813                        switch (oom_scan_process_thread(task, totalpages, NULL,
1814                                                        false)) {
1815                        case OOM_SCAN_SELECT:
1816                                if (chosen)
1817                                        put_task_struct(chosen);
1818                                chosen = task;
1819                                chosen_points = ULONG_MAX;
1820                                get_task_struct(chosen);
1821                                /* fall through */
1822                        case OOM_SCAN_CONTINUE:
1823                                continue;
1824                        case OOM_SCAN_ABORT:
1825                                cgroup_iter_end(cgroup, &it);
1826                                mem_cgroup_iter_break(memcg, iter);
1827                                if (chosen)
1828                                        put_task_struct(chosen);
1829                                return;
1830                        case OOM_SCAN_OK:
1831                                break;
1832                        };
1833                        points = oom_badness(task, memcg, NULL, totalpages);
1834                        if (points > chosen_points) {
1835                                if (chosen)
1836                                        put_task_struct(chosen);
1837                                chosen = task;
1838                                chosen_points = points;
1839                                get_task_struct(chosen);
1840                        }
1841                }
1842                cgroup_iter_end(cgroup, &it);
1843        }
1844
1845        if (!chosen)
1846                return;
1847        points = chosen_points * 1000 / totalpages;
1848        oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1849                         NULL, "Memory cgroup out of memory");
1850}
1851
1852static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1853                                        gfp_t gfp_mask,
1854                                        unsigned long flags)
1855{
1856        unsigned long total = 0;
1857        bool noswap = false;
1858        int loop;
1859
1860        if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1861                noswap = true;
1862        if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1863                noswap = true;
1864
1865        for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1866                if (loop)
1867                        drain_all_stock_async(memcg);
1868                total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1869                /*
1870                 * Allow limit shrinkers, which are triggered directly
1871                 * by userspace, to catch signals and stop reclaim
1872                 * after minimal progress, regardless of the margin.
1873                 */
1874                if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1875                        break;
1876                if (mem_cgroup_margin(memcg))
1877                        break;
1878                /*
1879                 * If nothing was reclaimed after two attempts, there
1880                 * may be no reclaimable pages in this hierarchy.
1881                 */
1882                if (loop && !total)
1883                        break;
1884        }
1885        return total;
1886}
1887
1888/**
1889 * test_mem_cgroup_node_reclaimable
1890 * @memcg: the target memcg
1891 * @nid: the node ID to be checked.
1892 * @noswap : specify true here if the user wants flle only information.
1893 *
1894 * This function returns whether the specified memcg contains any
1895 * reclaimable pages on a node. Returns true if there are any reclaimable
1896 * pages in the node.
1897 */
1898static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1899                int nid, bool noswap)
1900{
1901        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1902                return true;
1903        if (noswap || !total_swap_pages)
1904                return false;
1905        if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1906                return true;
1907        return false;
1908
1909}
1910#if MAX_NUMNODES > 1
1911
1912/*
1913 * Always updating the nodemask is not very good - even if we have an empty
1914 * list or the wrong list here, we can start from some node and traverse all
1915 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1916 *
1917 */
1918static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1919{
1920        int nid;
1921        /*
1922         * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1923         * pagein/pageout changes since the last update.
1924         */
1925        if (!atomic_read(&memcg->numainfo_events))
1926                return;
1927        if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1928                return;
1929
1930        /* make a nodemask where this memcg uses memory from */
1931        memcg->scan_nodes = node_states[N_MEMORY];
1932
1933        for_each_node_mask(nid, node_states[N_MEMORY]) {
1934
1935                if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1936                        node_clear(nid, memcg->scan_nodes);
1937        }
1938
1939        atomic_set(&memcg->numainfo_events, 0);
1940        atomic_set(&memcg->numainfo_updating, 0);
1941}
1942
1943/*
1944 * Selecting a node where we start reclaim from. Because what we need is just
1945 * reducing usage counter, start from anywhere is O,K. Considering
1946 * memory reclaim from current node, there are pros. and cons.
1947 *
1948 * Freeing memory from current node means freeing memory from a node which
1949 * we'll use or we've used. So, it may make LRU bad. And if several threads
1950 * hit limits, it will see a contention on a node. But freeing from remote
1951 * node means more costs for memory reclaim because of memory latency.
1952 *
1953 * Now, we use round-robin. Better algorithm is welcomed.
1954 */
1955int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1956{
1957        int node;
1958
1959        mem_cgroup_may_update_nodemask(memcg);
1960        node = memcg->last_scanned_node;
1961
1962        node = next_node(node, memcg->scan_nodes);
1963        if (node == MAX_NUMNODES)
1964                node = first_node(memcg->scan_nodes);
1965        /*
1966         * We call this when we hit limit, not when pages are added to LRU.
1967         * No LRU may hold pages because all pages are UNEVICTABLE or
1968         * memcg is too small and all pages are not on LRU. In that case,
1969         * we use curret node.
1970         */
1971        if (unlikely(node == MAX_NUMNODES))
1972                node = numa_node_id();
1973
1974        memcg->last_scanned_node = node;
1975        return node;
1976}
1977
1978/*
1979 * Check all nodes whether it contains reclaimable pages or not.
1980 * For quick scan, we make use of scan_nodes. This will allow us to skip
1981 * unused nodes. But scan_nodes is lazily updated and may not cotain
1982 * enough new information. We need to do double check.
1983 */
1984static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1985{
1986        int nid;
1987
1988        /*
1989         * quick check...making use of scan_node.
1990         * We can skip unused nodes.
1991         */
1992        if (!nodes_empty(memcg->scan_nodes)) {
1993                for (nid = first_node(memcg->scan_nodes);
1994                     nid < MAX_NUMNODES;
1995                     nid = next_node(nid, memcg->scan_nodes)) {
1996
1997                        if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1998                                return true;
1999                }
2000        }
2001        /*
2002         * Check rest of nodes.
2003         */
2004        for_each_node_state(nid, N_MEMORY) {
2005                if (node_isset(nid, memcg->scan_nodes))
2006                        continue;
2007                if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2008                        return true;
2009        }
2010        return false;
2011}
2012
2013#else
2014int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2015{
2016        return 0;
2017}
2018
2019static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2020{
2021        return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2022}
2023#endif
2024
2025static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2026                                   struct zone *zone,
2027                                   gfp_t gfp_mask,
2028                                   unsigned long *total_scanned)
2029{
2030        struct mem_cgroup *victim = NULL;
2031        int total = 0;
2032        int loop = 0;
2033        unsigned long excess;
2034        unsigned long nr_scanned;
2035        struct mem_cgroup_reclaim_cookie reclaim = {
2036                .zone = zone,
2037                .priority = 0,
2038        };
2039
2040        excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2041
2042        while (1) {
2043                victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2044                if (!victim) {
2045                        loop++;
2046                        if (loop >= 2) {
2047                                /*
2048                                 * If we have not been able to reclaim
2049                                 * anything, it might because there are
2050                                 * no reclaimable pages under this hierarchy
2051                                 */
2052                                if (!total)
2053                                        break;
2054                                /*
2055                                 * We want to do more targeted reclaim.
2056                                 * excess >> 2 is not to excessive so as to
2057                                 * reclaim too much, nor too less that we keep
2058                                 * coming back to reclaim from this cgroup
2059                                 */
2060                                if (total >= (excess >> 2) ||
2061                                        (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2062                                        break;
2063                        }
2064                        continue;
2065                }
2066                if (!mem_cgroup_reclaimable(victim, false))
2067                        continue;
2068                total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2069                                                     zone, &nr_scanned);
2070                *total_scanned += nr_scanned;
2071                if (!res_counter_soft_limit_excess(&root_memcg->res))
2072                        break;
2073        }
2074        mem_cgroup_iter_break(root_memcg, victim);
2075        return total;
2076}
2077
2078/*
2079 * Check OOM-Killer is already running under our hierarchy.
2080 * If someone is running, return false.
2081 * Has to be called with memcg_oom_lock
2082 */
2083static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2084{
2085        struct mem_cgroup *iter, *failed = NULL;
2086
2087        for_each_mem_cgroup_tree(iter, memcg) {
2088                if (iter->oom_lock) {
2089                        /*
2090                         * this subtree of our hierarchy is already locked
2091                         * so we cannot give a lock.
2092                         */
2093                        failed = iter;
2094                        mem_cgroup_iter_break(memcg, iter);
2095                        break;
2096                } else
2097                        iter->oom_lock = true;
2098        }
2099
2100        if (!failed)
2101                return true;
2102
2103        /*
2104         * OK, we failed to lock the whole subtree so we have to clean up
2105         * what we set up to the failing subtree
2106         */
2107        for_each_mem_cgroup_tree(iter, memcg) {
2108                if (iter == failed) {
2109                        mem_cgroup_iter_break(memcg, iter);
2110                        break;
2111                }
2112                iter->oom_lock = false;
2113        }
2114        return false;
2115}
2116
2117/*
2118 * Has to be called with memcg_oom_lock
2119 */
2120static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2121{
2122        struct mem_cgroup *iter;
2123
2124        for_each_mem_cgroup_tree(iter, memcg)
2125                iter->oom_lock = false;
2126        return 0;
2127}
2128
2129static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2130{
2131        struct mem_cgroup *iter;
2132
2133        for_each_mem_cgroup_tree(iter, memcg)
2134                atomic_inc(&iter->under_oom);
2135}
2136
2137static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2138{
2139        struct mem_cgroup *iter;
2140
2141        /*
2142         * When a new child is created while the hierarchy is under oom,
2143         * mem_cgroup_oom_lock() may not be called. We have to use
2144         * atomic_add_unless() here.
2145         */
2146        for_each_mem_cgroup_tree(iter, memcg)
2147                atomic_add_unless(&iter->under_oom, -1, 0);
2148}
2149
2150static DEFINE_SPINLOCK(memcg_oom_lock);
2151static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2152
2153struct oom_wait_info {
2154        struct mem_cgroup *memcg;
2155        wait_queue_t    wait;
2156};
2157
2158static int memcg_oom_wake_function(wait_queue_t *wait,
2159        unsigned mode, int sync, void *arg)
2160{
2161        struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2162        struct mem_cgroup *oom_wait_memcg;
2163        struct oom_wait_info *oom_wait_info;
2164
2165        oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2166        oom_wait_memcg = oom_wait_info->memcg;
2167
2168        /*
2169         * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2170         * Then we can use css_is_ancestor without taking care of RCU.
2171         */
2172        if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2173                && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2174                return 0;
2175        return autoremove_wake_function(wait, mode, sync, arg);
2176}
2177
2178static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2179{
2180        /* for filtering, pass "memcg" as argument. */
2181        __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2182}
2183
2184static void memcg_oom_recover(struct mem_cgroup *memcg)
2185{
2186        if (memcg && atomic_read(&memcg->under_oom))
2187                memcg_wakeup_oom(memcg);
2188}
2189
2190/*
2191 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2192 */
2193static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2194                                  int order)
2195{
2196        struct oom_wait_info owait;
2197        bool locked, need_to_kill;
2198
2199        owait.memcg = memcg;
2200        owait.wait.flags = 0;
2201        owait.wait.func = memcg_oom_wake_function;
2202        owait.wait.private = current;
2203        INIT_LIST_HEAD(&owait.wait.task_list);
2204        need_to_kill = true;
2205        mem_cgroup_mark_under_oom(memcg);
2206
2207        /* At first, try to OOM lock hierarchy under memcg.*/
2208        spin_lock(&memcg_oom_lock);
2209        locked = mem_cgroup_oom_lock(memcg);
2210        /*
2211         * Even if signal_pending(), we can't quit charge() loop without
2212         * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2213         * under OOM is always welcomed, use TASK_KILLABLE here.
2214         */
2215        prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2216        if (!locked || memcg->oom_kill_disable)
2217                need_to_kill = false;
2218        if (locked)
2219                mem_cgroup_oom_notify(memcg);
2220        spin_unlock(&memcg_oom_lock);
2221
2222        if (need_to_kill) {
2223                finish_wait(&memcg_oom_waitq, &owait.wait);
2224                mem_cgroup_out_of_memory(memcg, mask, order);
2225        } else {
2226                schedule();
2227                finish_wait(&memcg_oom_waitq, &owait.wait);
2228        }
2229        spin_lock(&memcg_oom_lock);
2230        if (locked)
2231                mem_cgroup_oom_unlock(memcg);
2232        memcg_wakeup_oom(memcg);
2233        spin_unlock(&memcg_oom_lock);
2234
2235        mem_cgroup_unmark_under_oom(memcg);
2236
2237        if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2238                return false;
2239        /* Give chance to dying process */
2240        schedule_timeout_uninterruptible(1);
2241        return true;
2242}
2243
2244/*
2245 * Currently used to update mapped file statistics, but the routine can be
2246 * generalized to update other statistics as well.
2247 *
2248 * Notes: Race condition
2249 *
2250 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2251 * it tends to be costly. But considering some conditions, we doesn't need
2252 * to do so _always_.
2253 *
2254 * Considering "charge", lock_page_cgroup() is not required because all
2255 * file-stat operations happen after a page is attached to radix-tree. There
2256 * are no race with "charge".
2257 *
2258 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2259 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2260 * if there are race with "uncharge". Statistics itself is properly handled
2261 * by flags.
2262 *
2263 * Considering "move", this is an only case we see a race. To make the race
2264 * small, we check mm->moving_account and detect there are possibility of race
2265 * If there is, we take a lock.
2266 */
2267
2268void __mem_cgroup_begin_update_page_stat(struct page *page,
2269                                bool *locked, unsigned long *flags)
2270{
2271        struct mem_cgroup *memcg;
2272        struct page_cgroup *pc;
2273
2274        pc = lookup_page_cgroup(page);
2275again:
2276        memcg = pc->mem_cgroup;
2277        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2278                return;
2279        /*
2280         * If this memory cgroup is not under account moving, we don't
2281         * need to take move_lock_mem_cgroup(). Because we already hold
2282         * rcu_read_lock(), any calls to move_account will be delayed until
2283         * rcu_read_unlock() if mem_cgroup_stolen() == true.
2284         */
2285        if (!mem_cgroup_stolen(memcg))
2286                return;
2287
2288        move_lock_mem_cgroup(memcg, flags);
2289        if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2290                move_unlock_mem_cgroup(memcg, flags);
2291                goto again;
2292        }
2293        *locked = true;
2294}
2295
2296void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2297{
2298        struct page_cgroup *pc = lookup_page_cgroup(page);
2299
2300        /*
2301         * It's guaranteed that pc->mem_cgroup never changes while
2302         * lock is held because a routine modifies pc->mem_cgroup
2303         * should take move_lock_mem_cgroup().
2304         */
2305        move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2306}
2307
2308void mem_cgroup_update_page_stat(struct page *page,
2309                                 enum mem_cgroup_page_stat_item idx, int val)
2310{
2311        struct mem_cgroup *memcg;
2312        struct page_cgroup *pc = lookup_page_cgroup(page);
2313        unsigned long uninitialized_var(flags);
2314
2315        if (mem_cgroup_disabled())
2316                return;
2317
2318        memcg = pc->mem_cgroup;
2319        if (unlikely(!memcg || !PageCgroupUsed(pc)))
2320                return;
2321
2322        switch (idx) {
2323        case MEMCG_NR_FILE_MAPPED:
2324                idx = MEM_CGROUP_STAT_FILE_MAPPED;
2325                break;
2326        default:
2327                BUG();
2328        }
2329
2330        this_cpu_add(memcg->stat->count[idx], val);
2331}
2332
2333/*
2334 * size of first charge trial. "32" comes from vmscan.c's magic value.
2335 * TODO: maybe necessary to use big numbers in big irons.
2336 */
2337#define CHARGE_BATCH    32U
2338struct memcg_stock_pcp {
2339        struct mem_cgroup *cached; /* this never be root cgroup */
2340        unsigned int nr_pages;
2341        struct work_struct work;
2342        unsigned long flags;
2343#define FLUSHING_CACHED_CHARGE  0
2344};
2345static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2346static DEFINE_MUTEX(percpu_charge_mutex);
2347
2348/**
2349 * consume_stock: Try to consume stocked charge on this cpu.
2350 * @memcg: memcg to consume from.
2351 * @nr_pages: how many pages to charge.
2352 *
2353 * The charges will only happen if @memcg matches the current cpu's memcg
2354 * stock, and at least @nr_pages are available in that stock.  Failure to
2355 * service an allocation will refill the stock.
2356 *
2357 * returns true if successful, false otherwise.
2358 */
2359static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2360{
2361        struct memcg_stock_pcp *stock;
2362        bool ret = true;
2363
2364        if (nr_pages > CHARGE_BATCH)
2365                return false;
2366
2367        stock = &get_cpu_var(memcg_stock);
2368        if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2369                stock->nr_pages -= nr_pages;
2370        else /* need to call res_counter_charge */
2371                ret = false;
2372        put_cpu_var(memcg_stock);
2373        return ret;
2374}
2375
2376/*
2377 * Returns stocks cached in percpu to res_counter and reset cached information.
2378 */
2379static void drain_stock(struct memcg_stock_pcp *stock)
2380{
2381        struct mem_cgroup *old = stock->cached;
2382
2383        if (stock->nr_pages) {
2384                unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2385
2386                res_counter_uncharge(&old->res, bytes);
2387                if (do_swap_account)
2388                        res_counter_uncharge(&old->memsw, bytes);
2389                stock->nr_pages = 0;
2390        }
2391        stock->cached = NULL;
2392}
2393
2394/*
2395 * This must be called under preempt disabled or must be called by
2396 * a thread which is pinned to local cpu.
2397 */
2398static void drain_local_stock(struct work_struct *dummy)
2399{
2400        struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2401        drain_stock(stock);
2402        clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2403}
2404
2405static void __init memcg_stock_init(void)
2406{
2407        int cpu;
2408
2409        for_each_possible_cpu(cpu) {
2410                struct memcg_stock_pcp *stock =
2411                                        &per_cpu(memcg_stock, cpu);
2412                INIT_WORK(&stock->work, drain_local_stock);
2413        }
2414}
2415
2416/*
2417 * Cache charges(val) which is from res_counter, to local per_cpu area.
2418 * This will be consumed by consume_stock() function, later.
2419 */
2420static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2421{
2422        struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2423
2424        if (stock->cached != memcg) { /* reset if necessary */
2425                drain_stock(stock);
2426                stock->cached = memcg;
2427        }
2428        stock->nr_pages += nr_pages;
2429        put_cpu_var(memcg_stock);
2430}
2431
2432/*
2433 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2434 * of the hierarchy under it. sync flag says whether we should block
2435 * until the work is done.
2436 */
2437static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2438{
2439        int cpu, curcpu;
2440
2441        /* Notify other cpus that system-wide "drain" is running */
2442        get_online_cpus();
2443        curcpu = get_cpu();
2444        for_each_online_cpu(cpu) {
2445                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2446                struct mem_cgroup *memcg;
2447
2448                memcg = stock->cached;
2449                if (!memcg || !stock->nr_pages)
2450                        continue;
2451                if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2452                        continue;
2453                if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2454                        if (cpu == curcpu)
2455                                drain_local_stock(&stock->work);
2456                        else
2457                                schedule_work_on(cpu, &stock->work);
2458                }
2459        }
2460        put_cpu();
2461
2462        if (!sync)
2463                goto out;
2464
2465        for_each_online_cpu(cpu) {
2466                struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2467                if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2468                        flush_work(&stock->work);
2469        }
2470out:
2471        put_online_cpus();
2472}
2473
2474/*
2475 * Tries to drain stocked charges in other cpus. This function is asynchronous
2476 * and just put a work per cpu for draining localy on each cpu. Caller can
2477 * expects some charges will be back to res_counter later but cannot wait for
2478 * it.
2479 */
2480static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2481{
2482        /*
2483         * If someone calls draining, avoid adding more kworker runs.
2484         */
2485        if (!mutex_trylock(&percpu_charge_mutex))
2486                return;
2487        drain_all_stock(root_memcg, false);
2488        mutex_unlock(&percpu_charge_mutex);
2489}
2490
2491/* This is a synchronous drain interface. */
2492static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2493{
2494        /* called when force_empty is called */
2495        mutex_lock(&percpu_charge_mutex);
2496        drain_all_stock(root_memcg, true);
2497        mutex_unlock(&percpu_charge_mutex);
2498}
2499
2500/*
2501 * This function drains percpu counter value from DEAD cpu and
2502 * move it to local cpu. Note that this function can be preempted.
2503 */
2504static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2505{
2506        int i;
2507
2508        spin_lock(&memcg->pcp_counter_lock);
2509        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2510                long x = per_cpu(memcg->stat->count[i], cpu);
2511
2512                per_cpu(memcg->stat->count[i], cpu) = 0;
2513                memcg->nocpu_base.count[i] += x;
2514        }
2515        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2516                unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2517
2518                per_cpu(memcg->stat->events[i], cpu) = 0;
2519                memcg->nocpu_base.events[i] += x;
2520        }
2521        spin_unlock(&memcg->pcp_counter_lock);
2522}
2523
2524static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2525                                        unsigned long action,
2526                                        void *hcpu)
2527{
2528        int cpu = (unsigned long)hcpu;
2529        struct memcg_stock_pcp *stock;
2530        struct mem_cgroup *iter;
2531
2532        if (action == CPU_ONLINE)
2533                return NOTIFY_OK;
2534
2535        if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2536                return NOTIFY_OK;
2537
2538        for_each_mem_cgroup(iter)
2539                mem_cgroup_drain_pcp_counter(iter, cpu);
2540
2541        stock = &per_cpu(memcg_stock, cpu);
2542        drain_stock(stock);
2543        return NOTIFY_OK;
2544}
2545
2546
2547/* See __mem_cgroup_try_charge() for details */
2548enum {
2549        CHARGE_OK,              /* success */
2550        CHARGE_RETRY,           /* need to retry but retry is not bad */
2551        CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2552        CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2553        CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2554};
2555
2556static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2557                                unsigned int nr_pages, unsigned int min_pages,
2558                                bool oom_check)
2559{
2560        unsigned long csize = nr_pages * PAGE_SIZE;
2561        struct mem_cgroup *mem_over_limit;
2562        struct res_counter *fail_res;
2563        unsigned long flags = 0;
2564        int ret;
2565
2566        ret = res_counter_charge(&memcg->res, csize, &fail_res);
2567
2568        if (likely(!ret)) {
2569                if (!do_swap_account)
2570                        return CHARGE_OK;
2571                ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2572                if (likely(!ret))
2573                        return CHARGE_OK;
2574
2575                res_counter_uncharge(&memcg->res, csize);
2576                mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2577                flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2578        } else
2579                mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2580        /*
2581         * Never reclaim on behalf of optional batching, retry with a
2582         * single page instead.
2583         */
2584        if (nr_pages > min_pages)
2585                return CHARGE_RETRY;
2586
2587        if (!(gfp_mask & __GFP_WAIT))
2588                return CHARGE_WOULDBLOCK;
2589
2590        if (gfp_mask & __GFP_NORETRY)
2591                return CHARGE_NOMEM;
2592
2593        ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2594        if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2595                return CHARGE_RETRY;
2596        /*
2597         * Even though the limit is exceeded at this point, reclaim
2598         * may have been able to free some pages.  Retry the charge
2599         * before killing the task.
2600         *
2601         * Only for regular pages, though: huge pages are rather
2602         * unlikely to succeed so close to the limit, and we fall back
2603         * to regular pages anyway in case of failure.
2604         */
2605        if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2606                return CHARGE_RETRY;
2607
2608        /*
2609         * At task move, charge accounts can be doubly counted. So, it's
2610         * better to wait until the end of task_move if something is going on.
2611         */
2612        if (mem_cgroup_wait_acct_move(mem_over_limit))
2613                return CHARGE_RETRY;
2614
2615        /* If we don't need to call oom-killer at el, return immediately */
2616        if (!oom_check)
2617                return CHARGE_NOMEM;
2618        /* check OOM */
2619        if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2620                return CHARGE_OOM_DIE;
2621
2622        return CHARGE_RETRY;
2623}
2624
2625/*
2626 * __mem_cgroup_try_charge() does
2627 * 1. detect memcg to be charged against from passed *mm and *ptr,
2628 * 2. update res_counter
2629 * 3. call memory reclaim if necessary.
2630 *
2631 * In some special case, if the task is fatal, fatal_signal_pending() or
2632 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2633 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2634 * as possible without any hazards. 2: all pages should have a valid
2635 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2636 * pointer, that is treated as a charge to root_mem_cgroup.
2637 *
2638 * So __mem_cgroup_try_charge() will return
2639 *  0       ...  on success, filling *ptr with a valid memcg pointer.
2640 *  -ENOMEM ...  charge failure because of resource limits.
2641 *  -EINTR  ...  if thread is fatal. *ptr is filled with root_mem_cgroup.
2642 *
2643 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2644 * the oom-killer can be invoked.
2645 */
2646static int __mem_cgroup_try_charge(struct mm_struct *mm,
2647                                   gfp_t gfp_mask,
2648                                   unsigned int nr_pages,
2649                                   struct mem_cgroup **ptr,
2650                                   bool oom)
2651{
2652        unsigned int batch = max(CHARGE_BATCH, nr_pages);
2653        int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2654        struct mem_cgroup *memcg = NULL;
2655        int ret;
2656
2657        /*
2658         * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2659         * in system level. So, allow to go ahead dying process in addition to
2660         * MEMDIE process.
2661         */
2662        if (unlikely(test_thread_flag(TIF_MEMDIE)
2663                     || fatal_signal_pending(current)))
2664                goto bypass;
2665
2666        /*
2667         * We always charge the cgroup the mm_struct belongs to.
2668         * The mm_struct's mem_cgroup changes on task migration if the
2669         * thread group leader migrates. It's possible that mm is not
2670         * set, if so charge the root memcg (happens for pagecache usage).
2671         */
2672        if (!*ptr && !mm)
2673                *ptr = root_mem_cgroup;
2674again:
2675        if (*ptr) { /* css should be a valid one */
2676                memcg = *ptr;
2677                if (mem_cgroup_is_root(memcg))
2678                        goto done;
2679                if (consume_stock(memcg, nr_pages))
2680                        goto done;
2681                css_get(&memcg->css);
2682        } else {
2683                struct task_struct *p;
2684
2685                rcu_read_lock();
2686                p = rcu_dereference(mm->owner);
2687                /*
2688                 * Because we don't have task_lock(), "p" can exit.
2689                 * In that case, "memcg" can point to root or p can be NULL with
2690                 * race with swapoff. Then, we have small risk of mis-accouning.
2691                 * But such kind of mis-account by race always happens because
2692                 * we don't have cgroup_mutex(). It's overkill and we allo that
2693                 * small race, here.
2694                 * (*) swapoff at el will charge against mm-struct not against
2695                 * task-struct. So, mm->owner can be NULL.
2696                 */
2697                memcg = mem_cgroup_from_task(p);
2698                if (!memcg)
2699                        memcg = root_mem_cgroup;
2700                if (mem_cgroup_is_root(memcg)) {
2701                        rcu_read_unlock();
2702                        goto done;
2703                }
2704                if (consume_stock(memcg, nr_pages)) {
2705                        /*
2706                         * It seems dagerous to access memcg without css_get().
2707                         * But considering how consume_stok works, it's not
2708                         * necessary. If consume_stock success, some charges
2709                         * from this memcg are cached on this cpu. So, we
2710                         * don't need to call css_get()/css_tryget() before
2711                         * calling consume_stock().
2712                         */
2713                        rcu_read_unlock();
2714                        goto done;
2715                }
2716                /* after here, we may be blocked. we need to get refcnt */
2717                if (!css_tryget(&memcg->css)) {
2718                        rcu_read_unlock();
2719                        goto again;
2720                }
2721                rcu_read_unlock();
2722        }
2723
2724        do {
2725                bool oom_check;
2726
2727                /* If killed, bypass charge */
2728                if (fatal_signal_pending(current)) {
2729                        css_put(&memcg->css);
2730                        goto bypass;
2731                }
2732
2733                oom_check = false;
2734                if (oom && !nr_oom_retries) {
2735                        oom_check = true;
2736                        nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2737                }
2738
2739                ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2740                    oom_check);
2741                switch (ret) {
2742                case CHARGE_OK:
2743                        break;
2744                case CHARGE_RETRY: /* not in OOM situation but retry */
2745                        batch = nr_pages;
2746                        css_put(&memcg->css);
2747                        memcg = NULL;
2748                        goto again;
2749                case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2750                        css_put(&memcg->css);
2751                        goto nomem;
2752                case CHARGE_NOMEM: /* OOM routine works */
2753                        if (!oom) {
2754                                css_put(&memcg->css);
2755                                goto nomem;
2756                        }
2757                        /* If oom, we never return -ENOMEM */
2758                        nr_oom_retries--;
2759                        break;
2760                case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2761                        css_put(&memcg->css);
2762                        goto bypass;
2763                }
2764        } while (ret != CHARGE_OK);
2765
2766        if (batch > nr_pages)
2767                refill_stock(memcg, batch - nr_pages);
2768        css_put(&memcg->css);
2769done:
2770        *ptr = memcg;
2771        return 0;
2772nomem:
2773        *ptr = NULL;
2774        return -ENOMEM;
2775bypass:
2776        *ptr = root_mem_cgroup;
2777        return -EINTR;
2778}
2779
2780/*
2781 * Somemtimes we have to undo a charge we got by try_charge().
2782 * This function is for that and do uncharge, put css's refcnt.
2783 * gotten by try_charge().
2784 */
2785static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2786                                       unsigned int nr_pages)
2787{
2788        if (!mem_cgroup_is_root(memcg)) {
2789                unsigned long bytes = nr_pages * PAGE_SIZE;
2790
2791                res_counter_uncharge(&memcg->res, bytes);
2792                if (do_swap_account)
2793                        res_counter_uncharge(&memcg->memsw, bytes);
2794        }
2795}
2796
2797/*
2798 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2799 * This is useful when moving usage to parent cgroup.
2800 */
2801static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2802                                        unsigned int nr_pages)
2803{
2804        unsigned long bytes = nr_pages * PAGE_SIZE;
2805
2806        if (mem_cgroup_is_root(memcg))
2807                return;
2808
2809        res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2810        if (do_swap_account)
2811                res_counter_uncharge_until(&memcg->memsw,
2812                                                memcg->memsw.parent, bytes);
2813}
2814
2815/*
2816 * A helper function to get mem_cgroup from ID. must be called under
2817 * rcu_read_lock().  The caller is responsible for calling css_tryget if
2818 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2819 * called against removed memcg.)
2820 */
2821static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2822{
2823        struct cgroup_subsys_state *css;
2824
2825        /* ID 0 is unused ID */
2826        if (!id)
2827                return NULL;
2828        css = css_lookup(&mem_cgroup_subsys, id);
2829        if (!css)
2830                return NULL;
2831        return mem_cgroup_from_css(css);
2832}
2833
2834struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2835{
2836        struct mem_cgroup *memcg = NULL;
2837        struct page_cgroup *pc;
2838        unsigned short id;
2839        swp_entry_t ent;
2840
2841        VM_BUG_ON(!PageLocked(page));
2842
2843        pc = lookup_page_cgroup(page);
2844        lock_page_cgroup(pc);
2845        if (PageCgroupUsed(pc)) {
2846                memcg = pc->mem_cgroup;
2847                if (memcg && !css_tryget(&memcg->css))
2848                        memcg = NULL;
2849        } else if (PageSwapCache(page)) {
2850                ent.val = page_private(page);
2851                id = lookup_swap_cgroup_id(ent);
2852                rcu_read_lock();
2853                memcg = mem_cgroup_lookup(id);
2854                if (memcg && !css_tryget(&memcg->css))
2855                        memcg = NULL;
2856                rcu_read_unlock();
2857        }
2858        unlock_page_cgroup(pc);
2859        return memcg;
2860}
2861
2862static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2863                                       struct page *page,
2864                                       unsigned int nr_pages,
2865                                       enum charge_type ctype,
2866                                       bool lrucare)
2867{
2868        struct page_cgroup *pc = lookup_page_cgroup(page);
2869        struct zone *uninitialized_var(zone);
2870        struct lruvec *lruvec;
2871        bool was_on_lru = false;
2872        bool anon;
2873
2874        lock_page_cgroup(pc);
2875        VM_BUG_ON(PageCgroupUsed(pc));
2876        /*
2877         * we don't need page_cgroup_lock about tail pages, becase they are not
2878         * accessed by any other context at this point.
2879         */
2880
2881        /*
2882         * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2883         * may already be on some other mem_cgroup's LRU.  Take care of it.
2884         */
2885        if (lrucare) {
2886                zone = page_zone(page);
2887                spin_lock_irq(&zone->lru_lock);
2888                if (PageLRU(page)) {
2889                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2890                        ClearPageLRU(page);
2891                        del_page_from_lru_list(page, lruvec, page_lru(page));
2892                        was_on_lru = true;
2893                }
2894        }
2895
2896        pc->mem_cgroup = memcg;
2897        /*
2898         * We access a page_cgroup asynchronously without lock_page_cgroup().
2899         * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2900         * is accessed after testing USED bit. To make pc->mem_cgroup visible
2901         * before USED bit, we need memory barrier here.
2902         * See mem_cgroup_add_lru_list(), etc.
2903         */
2904        smp_wmb();
2905        SetPageCgroupUsed(pc);
2906
2907        if (lrucare) {
2908                if (was_on_lru) {
2909                        lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2910                        VM_BUG_ON(PageLRU(page));
2911                        SetPageLRU(page);
2912                        add_page_to_lru_list(page, lruvec, page_lru(page));
2913                }
2914                spin_unlock_irq(&zone->lru_lock);
2915        }
2916
2917        if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2918                anon = true;
2919        else
2920                anon = false;
2921
2922        mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2923        unlock_page_cgroup(pc);
2924
2925        /*
2926         * "charge_statistics" updated event counter. Then, check it.
2927         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2928         * if they exceeds softlimit.
2929         */
2930        memcg_check_events(memcg, page);
2931}
2932
2933static DEFINE_MUTEX(set_limit_mutex);
2934
2935#ifdef CONFIG_MEMCG_KMEM
2936static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2937{
2938        return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2939                (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2940}
2941
2942/*
2943 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2944 * in the memcg_cache_params struct.
2945 */
2946static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2947{
2948        struct kmem_cache *cachep;
2949
2950        VM_BUG_ON(p->is_root_cache);
2951        cachep = p->root_cache;
2952        return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2953}
2954
2955#ifdef CONFIG_SLABINFO
2956static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2957                                        struct seq_file *m)
2958{
2959        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2960        struct memcg_cache_params *params;
2961
2962        if (!memcg_can_account_kmem(memcg))
2963                return -EIO;
2964
2965        print_slabinfo_header(m);
2966
2967        mutex_lock(&memcg->slab_caches_mutex);
2968        list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2969                cache_show(memcg_params_to_cache(params), m);
2970        mutex_unlock(&memcg->slab_caches_mutex);
2971
2972        return 0;
2973}
2974#endif
2975
2976static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2977{
2978        struct res_counter *fail_res;
2979        struct mem_cgroup *_memcg;
2980        int ret = 0;
2981        bool may_oom;
2982
2983        ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2984        if (ret)
2985                return ret;
2986
2987        /*
2988         * Conditions under which we can wait for the oom_killer. Those are
2989         * the same conditions tested by the core page allocator
2990         */
2991        may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2992
2993        _memcg = memcg;
2994        ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2995                                      &_memcg, may_oom);
2996
2997        if (ret == -EINTR)  {
2998                /*
2999                 * __mem_cgroup_try_charge() chosed to bypass to root due to
3000                 * OOM kill or fatal signal.  Since our only options are to
3001                 * either fail the allocation or charge it to this cgroup, do
3002                 * it as a temporary condition. But we can't fail. From a
3003                 * kmem/slab perspective, the cache has already been selected,
3004                 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3005                 * our minds.
3006                 *
3007                 * This condition will only trigger if the task entered
3008                 * memcg_charge_kmem in a sane state, but was OOM-killed during
3009                 * __mem_cgroup_try_charge() above. Tasks that were already
3010                 * dying when the allocation triggers should have been already
3011                 * directed to the root cgroup in memcontrol.h
3012                 */
3013                res_counter_charge_nofail(&memcg->res, size, &fail_res);
3014                if (do_swap_account)
3015                        res_counter_charge_nofail(&memcg->memsw, size,
3016                                                  &fail_res);
3017                ret = 0;
3018        } else if (ret)
3019                res_counter_uncharge(&memcg->kmem, size);
3020
3021        return ret;
3022}
3023
3024static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3025{
3026        res_counter_uncharge(&memcg->res, size);
3027        if (do_swap_account)
3028                res_counter_uncharge(&memcg->memsw, size);
3029
3030        /* Not down to 0 */
3031        if (res_counter_uncharge(&memcg->kmem, size))
3032                return;
3033
3034        if (memcg_kmem_test_and_clear_dead(memcg))
3035                mem_cgroup_put(memcg);
3036}
3037
3038void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3039{
3040        if (!memcg)
3041                return;
3042
3043        mutex_lock(&memcg->slab_caches_mutex);
3044        list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3045        mutex_unlock(&memcg->slab_caches_mutex);
3046}
3047
3048/*
3049 * helper for acessing a memcg's index. It will be used as an index in the
3050 * child cache array in kmem_cache, and also to derive its name. This function
3051 * will return -1 when this is not a kmem-limited memcg.
3052 */
3053int memcg_cache_id(struct mem_cgroup *memcg)
3054{
3055        return memcg ? memcg->kmemcg_id : -1;
3056}
3057
3058/*
3059 * This ends up being protected by the set_limit mutex, during normal
3060 * operation, because that is its main call site.
3061 *
3062 * But when we create a new cache, we can call this as well if its parent
3063 * is kmem-limited. That will have to hold set_limit_mutex as well.
3064 */
3065int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3066{
3067        int num, ret;
3068
3069        num = ida_simple_get(&kmem_limited_groups,
3070                                0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3071        if (num < 0)
3072                return num;
3073        /*
3074         * After this point, kmem_accounted (that we test atomically in
3075         * the beginning of this conditional), is no longer 0. This
3076         * guarantees only one process will set the following boolean
3077         * to true. We don't need test_and_set because we're protected
3078         * by the set_limit_mutex anyway.
3079         */
3080        memcg_kmem_set_activated(memcg);
3081
3082        ret = memcg_update_all_caches(num+1);
3083        if (ret) {
3084                ida_simple_remove(&kmem_limited_groups, num);
3085                memcg_kmem_clear_activated(memcg);
3086                return ret;
3087        }
3088
3089        memcg->kmemcg_id = num;
3090        INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3091        mutex_init(&memcg->slab_caches_mutex);
3092        return 0;
3093}
3094
3095static size_t memcg_caches_array_size(int num_groups)
3096{
3097        ssize_t size;
3098        if (num_groups <= 0)
3099                return 0;
3100
3101        size = 2 * num_groups;
3102        if (size < MEMCG_CACHES_MIN_SIZE)
3103                size = MEMCG_CACHES_MIN_SIZE;
3104        else if (size > MEMCG_CACHES_MAX_SIZE)
3105                size = MEMCG_CACHES_MAX_SIZE;
3106
3107        return size;
3108}
3109
3110/*
3111 * We should update the current array size iff all caches updates succeed. This
3112 * can only be done from the slab side. The slab mutex needs to be held when
3113 * calling this.
3114 */
3115void memcg_update_array_size(int num)
3116{
3117        if (num > memcg_limited_groups_array_size)
3118                memcg_limited_groups_array_size = memcg_caches_array_size(num);
3119}
3120
3121static void kmem_cache_destroy_work_func(struct work_struct *w);
3122
3123int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3124{
3125        struct memcg_cache_params *cur_params = s->memcg_params;
3126
3127        VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3128
3129        if (num_groups > memcg_limited_groups_array_size) {
3130                int i;
3131                ssize_t size = memcg_caches_array_size(num_groups);
3132
3133                size *= sizeof(void *);
3134                size += sizeof(struct memcg_cache_params);
3135
3136                s->memcg_params = kzalloc(size, GFP_KERNEL);
3137                if (!s->memcg_params) {
3138                        s->memcg_params = cur_params;
3139                        return -ENOMEM;
3140                }
3141
3142                s->memcg_params->is_root_cache = true;
3143
3144                /*
3145                 * There is the chance it will be bigger than
3146                 * memcg_limited_groups_array_size, if we failed an allocation
3147                 * in a cache, in which case all caches updated before it, will
3148                 * have a bigger array.
3149                 *
3150                 * But if that is the case, the data after
3151                 * memcg_limited_groups_array_size is certainly unused
3152                 */
3153                for (i = 0; i < memcg_limited_groups_array_size; i++) {
3154                        if (!cur_params->memcg_caches[i])
3155                                continue;
3156                        s->memcg_params->memcg_caches[i] =
3157                                                cur_params->memcg_caches[i];
3158                }
3159
3160                /*
3161                 * Ideally, we would wait until all caches succeed, and only
3162                 * then free the old one. But this is not worth the extra
3163                 * pointer per-cache we'd have to have for this.
3164                 *
3165                 * It is not a big deal if some caches are left with a size
3166                 * bigger than the others. And all updates will reset this
3167                 * anyway.
3168                 */
3169                kfree(cur_params);
3170        }
3171        return 0;
3172}
3173
3174int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3175                         struct kmem_cache *root_cache)
3176{
3177        size_t size = sizeof(struct memcg_cache_params);
3178
3179        if (!memcg_kmem_enabled())
3180                return 0;
3181
3182        if (!memcg)
3183                size += memcg_limited_groups_array_size * sizeof(void *);
3184
3185        s->memcg_params = kzalloc(size, GFP_KERNEL);
3186        if (!s->memcg_params)
3187                return -ENOMEM;
3188
3189        INIT_WORK(&s->memcg_params->destroy,
3190                        kmem_cache_destroy_work_func);
3191        if (memcg) {
3192                s->memcg_params->memcg = memcg;
3193                s->memcg_params->root_cache = root_cache;
3194        } else
3195                s->memcg_params->is_root_cache = true;
3196
3197        return 0;
3198}
3199
3200void memcg_release_cache(struct kmem_cache *s)
3201{
3202        struct kmem_cache *root;
3203        struct mem_cgroup *memcg;
3204        int id;
3205
3206        /*
3207         * This happens, for instance, when a root cache goes away before we
3208         * add any memcg.
3209         */
3210        if (!s->memcg_params)
3211                return;
3212
3213        if (s->memcg_params->is_root_cache)
3214                goto out;
3215
3216        memcg = s->memcg_params->memcg;
3217        id  = memcg_cache_id(memcg);
3218
3219        root = s->memcg_params->root_cache;
3220        root->memcg_params->memcg_caches[id] = NULL;
3221
3222        mutex_lock(&memcg->slab_caches_mutex);
3223        list_del(&s->memcg_params->list);
3224        mutex_unlock(&memcg->slab_caches_mutex);
3225
3226        mem_cgroup_put(memcg);
3227out:
3228        kfree(s->memcg_params);
3229}
3230
3231/*
3232 * During the creation a new cache, we need to disable our accounting mechanism
3233 * altogether. This is true even if we are not creating, but rather just
3234 * enqueing new caches to be created.
3235 *
3236 * This is because that process will trigger allocations; some visible, like
3237 * explicit kmallocs to auxiliary data structures, name strings and internal
3238 * cache structures; some well concealed, like INIT_WORK() that can allocate
3239 * objects during debug.
3240 *
3241 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3242 * to it. This may not be a bounded recursion: since the first cache creation
3243 * failed to complete (waiting on the allocation), we'll just try to create the
3244 * cache again, failing at the same point.
3245 *
3246 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3247 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3248 * inside the following two functions.
3249 */
3250static inline void memcg_stop_kmem_account(void)
3251{
3252        VM_BUG_ON(!current->mm);
3253        current->memcg_kmem_skip_account++;
3254}
3255
3256static inline void memcg_resume_kmem_account(void)
3257{
3258        VM_BUG_ON(!current->mm);
3259        current->memcg_kmem_skip_account--;
3260}
3261
3262static void kmem_cache_destroy_work_func(struct work_struct *w)
3263{
3264        struct kmem_cache *cachep;
3265        struct memcg_cache_params *p;
3266
3267        p = container_of(w, struct memcg_cache_params, destroy);
3268
3269        cachep = memcg_params_to_cache(p);
3270
3271        /*
3272         * If we get down to 0 after shrink, we could delete right away.
3273         * However, memcg_release_pages() already puts us back in the workqueue
3274         * in that case. If we proceed deleting, we'll get a dangling
3275         * reference, and removing the object from the workqueue in that case
3276         * is unnecessary complication. We are not a fast path.
3277         *
3278         * Note that this case is fundamentally different from racing with
3279         * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3280         * kmem_cache_shrink, not only we would be reinserting a dead cache
3281         * into the queue, but doing so from inside the worker racing to
3282         * destroy it.
3283         *
3284         * So if we aren't down to zero, we'll just schedule a worker and try
3285         * again
3286         */
3287        if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3288                kmem_cache_shrink(cachep);
3289                if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3290                        return;
3291        } else
3292                kmem_cache_destroy(cachep);
3293}
3294
3295void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3296{
3297        if (!cachep->memcg_params->dead)
3298                return;
3299
3300        /*
3301         * There are many ways in which we can get here.
3302         *
3303         * We can get to a memory-pressure situation while the delayed work is
3304         * still pending to run. The vmscan shrinkers can then release all
3305         * cache memory and get us to destruction. If this is the case, we'll
3306         * be executed twice, which is a bug (the second time will execute over
3307         * bogus data). In this case, cancelling the work should be fine.
3308         *
3309         * But we can also get here from the worker itself, if
3310         * kmem_cache_shrink is enough to shake all the remaining objects and
3311         * get the page count to 0. In this case, we'll deadlock if we try to
3312         * cancel the work (the worker runs with an internal lock held, which
3313         * is the same lock we would hold for cancel_work_sync().)
3314         *
3315         * Since we can't possibly know who got us here, just refrain from
3316         * running if there is already work pending
3317         */
3318        if (work_pending(&cachep->memcg_params->destroy))
3319                return;
3320        /*
3321         * We have to defer the actual destroying to a workqueue, because
3322         * we might currently be in a context that cannot sleep.
3323         */
3324        schedule_work(&cachep->memcg_params->destroy);
3325}
3326
3327/*
3328 * This lock protects updaters, not readers. We want readers to be as fast as
3329 * they can, and they will either see NULL or a valid cache value. Our model
3330 * allow them to see NULL, in which case the root memcg will be selected.
3331 *
3332 * We need this lock because multiple allocations to the same cache from a non
3333 * will span more than one worker. Only one of them can create the cache.
3334 */
3335static DEFINE_MUTEX(memcg_cache_mutex);
3336
3337/*
3338 * Called with memcg_cache_mutex held
3339 */
3340static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3341                                         struct kmem_cache *s)
3342{
3343        struct kmem_cache *new;
3344        static char *tmp_name = NULL;
3345
3346        lockdep_assert_held(&memcg_cache_mutex);
3347
3348        /*
3349         * kmem_cache_create_memcg duplicates the given name and
3350         * cgroup_name for this name requires RCU context.
3351         * This static temporary buffer is used to prevent from
3352         * pointless shortliving allocation.
3353         */
3354        if (!tmp_name) {
3355                tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3356                if (!tmp_name)
3357                        return NULL;
3358        }
3359
3360        rcu_read_lock();
3361        snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3362                         memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3363        rcu_read_unlock();
3364
3365        new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3366                                      (s->flags & ~SLAB_PANIC), s->ctor, s);
3367
3368        if (new)
3369                new->allocflags |= __GFP_KMEMCG;
3370
3371        return new;
3372}
3373
3374static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3375                                                  struct kmem_cache *cachep)
3376{
3377        struct kmem_cache *new_cachep;
3378        int idx;
3379
3380        BUG_ON(!memcg_can_account_kmem(memcg));
3381
3382        idx = memcg_cache_id(memcg);
3383
3384        mutex_lock(&memcg_cache_mutex);
3385        new_cachep = cachep->memcg_params->memcg_caches[idx];
3386        if (new_cachep)
3387                goto out;
3388
3389        new_cachep = kmem_cache_dup(memcg, cachep);
3390        if (new_cachep == NULL) {
3391                new_cachep = cachep;
3392                goto out;
3393        }
3394
3395        mem_cgroup_get(memcg);
3396        atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3397
3398        cachep->memcg_params->memcg_caches[idx] = new_cachep;
3399        /*
3400         * the readers won't lock, make sure everybody sees the updated value,
3401         * so they won't put stuff in the queue again for no reason
3402         */
3403        wmb();
3404out:
3405        mutex_unlock(&memcg_cache_mutex);
3406        return new_cachep;
3407}
3408
3409void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3410{
3411        struct kmem_cache *c;
3412        int i;
3413
3414        if (!s->memcg_params)
3415                return;
3416        if (!s->memcg_params->is_root_cache)
3417                return;
3418
3419        /*
3420         * If the cache is being destroyed, we trust that there is no one else
3421         * requesting objects from it. Even if there are, the sanity checks in
3422         * kmem_cache_destroy should caught this ill-case.
3423         *
3424         * Still, we don't want anyone else freeing memcg_caches under our
3425         * noses, which can happen if a new memcg comes to life. As usual,
3426         * we'll take the set_limit_mutex to protect ourselves against this.
3427         */
3428        mutex_lock(&set_limit_mutex);
3429        for (i = 0; i < memcg_limited_groups_array_size; i++) {
3430                c = s->memcg_params->memcg_caches[i];
3431                if (!c)
3432                        continue;
3433
3434                /*
3435                 * We will now manually delete the caches, so to avoid races
3436                 * we need to cancel all pending destruction workers and
3437                 * proceed with destruction ourselves.
3438                 *
3439                 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3440                 * and that could spawn the workers again: it is likely that
3441                 * the cache still have active pages until this very moment.
3442                 * This would lead us back to mem_cgroup_destroy_cache.
3443                 *
3444                 * But that will not execute at all if the "dead" flag is not
3445                 * set, so flip it down to guarantee we are in control.
3446                 */
3447                c->memcg_params->dead = false;
3448                cancel_work_sync(&c->memcg_params->destroy);
3449                kmem_cache_destroy(c);
3450        }
3451        mutex_unlock(&set_limit_mutex);
3452}
3453
3454struct create_work {
3455        struct mem_cgroup *memcg;
3456        struct kmem_cache *cachep;
3457        struct work_struct work;
3458};
3459
3460static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3461{
3462        struct kmem_cache *cachep;
3463        struct memcg_cache_params *params;
3464
3465        if (!memcg_kmem_is_active(memcg))
3466                return;
3467
3468        mutex_lock(&memcg->slab_caches_mutex);
3469        list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3470                cachep = memcg_params_to_cache(params);
3471                cachep->memcg_params->dead = true;
3472                schedule_work(&cachep->memcg_params->destroy);
3473        }
3474        mutex_unlock(&memcg->slab_caches_mutex);
3475}
3476
3477static void memcg_create_cache_work_func(struct work_struct *w)
3478{
3479        struct create_work *cw;
3480
3481        cw = container_of(w, struct create_work, work);
3482        memcg_create_kmem_cache(cw->memcg, cw->cachep);
3483        /* Drop the reference gotten when we enqueued. */
3484        css_put(&cw->memcg->css);
3485        kfree(cw);
3486}
3487
3488/*
3489 * Enqueue the creation of a per-memcg kmem_cache.
3490 */
3491static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3492                                         struct kmem_cache *cachep)
3493{
3494        struct create_work *cw;
3495
3496        cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3497        if (cw == NULL) {
3498                css_put(&memcg->css);
3499                return;
3500        }
3501
3502        cw->memcg = memcg;
3503        cw->cachep = cachep;
3504
3505        INIT_WORK(&cw->work, memcg_create_cache_work_func);
3506        schedule_work(&cw->work);
3507}
3508
3509static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3510                                       struct kmem_cache *cachep)
3511{
3512        /*
3513         * We need to stop accounting when we kmalloc, because if the
3514         * corresponding kmalloc cache is not yet created, the first allocation
3515         * in __memcg_create_cache_enqueue will recurse.
3516         *
3517         * However, it is better to enclose the whole function. Depending on
3518         * the debugging options enabled, INIT_WORK(), for instance, can
3519         * trigger an allocation. This too, will make us recurse. Because at
3520         * this point we can't allow ourselves back into memcg_kmem_get_cache,
3521         * the safest choice is to do it like this, wrapping the whole function.
3522         */
3523        memcg_stop_kmem_account();
3524        __memcg_create_cache_enqueue(memcg, cachep);
3525        memcg_resume_kmem_account();
3526}
3527/*
3528 * Return the kmem_cache we're supposed to use for a slab allocation.
3529 * We try to use the current memcg's version of the cache.
3530 *
3531 * If the cache does not exist yet, if we are the first user of it,
3532 * we either create it immediately, if possible, or create it asynchronously
3533 * in a workqueue.
3534 * In the latter case, we will let the current allocation go through with
3535 * the original cache.
3536 *
3537 * Can't be called in interrupt context or from kernel threads.
3538 * This function needs to be called with rcu_read_lock() held.
3539 */
3540struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3541                                          gfp_t gfp)
3542{
3543        struct mem_cgroup *memcg;
3544        int idx;
3545
3546        VM_BUG_ON(!cachep->memcg_params);
3547        VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3548
3549        if (!current->mm || current->memcg_kmem_skip_account)
3550                return cachep;
3551
3552        rcu_read_lock();
3553        memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3554
3555        if (!memcg_can_account_kmem(memcg))
3556                goto out;
3557
3558        idx = memcg_cache_id(memcg);
3559
3560        /*
3561         * barrier to mare sure we're always seeing the up to date value.  The
3562         * code updating memcg_caches will issue a write barrier to match this.
3563         */
3564        read_barrier_depends();
3565        if (likely(cachep->memcg_params->memcg_caches[idx])) {
3566                cachep = cachep->memcg_params->memcg_caches[idx];
3567                goto out;
3568        }
3569
3570        /* The corresponding put will be done in the workqueue. */
3571        if (!css_tryget(&memcg->css))
3572                goto out;
3573        rcu_read_unlock();
3574
3575        /*
3576         * If we are in a safe context (can wait, and not in interrupt
3577         * context), we could be be predictable and return right away.
3578         * This would guarantee that the allocation being performed
3579         * already belongs in the new cache.
3580         *
3581         * However, there are some clashes that can arrive from locking.
3582         * For instance, because we acquire the slab_mutex while doing
3583         * kmem_cache_dup, this means no further allocation could happen
3584         * with the slab_mutex held.
3585         *
3586         * Also, because cache creation issue get_online_cpus(), this
3587         * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3588         * that ends up reversed during cpu hotplug. (cpuset allocates
3589         * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3590         * better to defer everything.
3591         */
3592        memcg_create_cache_enqueue(memcg, cachep);
3593        return cachep;
3594out:
3595        rcu_read_unlock();
3596        return cachep;
3597}
3598EXPORT_SYMBOL(__memcg_kmem_get_cache);
3599
3600/*
3601 * We need to verify if the allocation against current->mm->owner's memcg is
3602 * possible for the given order. But the page is not allocated yet, so we'll
3603 * need a further commit step to do the final arrangements.
3604 *
3605 * It is possible for the task to switch cgroups in this mean time, so at
3606 * commit time, we can't rely on task conversion any longer.  We'll then use
3607 * the handle argument to return to the caller which cgroup we should commit
3608 * against. We could also return the memcg directly and avoid the pointer
3609 * passing, but a boolean return value gives better semantics considering
3610 * the compiled-out case as well.
3611 *
3612 * Returning true means the allocation is possible.
3613 */
3614bool
3615__memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3616{
3617        struct mem_cgroup *memcg;
3618        int ret;
3619
3620        *_memcg = NULL;
3621        memcg = try_get_mem_cgroup_from_mm(current->mm);
3622
3623        /*
3624         * very rare case described in mem_cgroup_from_task. Unfortunately there
3625         * isn't much we can do without complicating this too much, and it would
3626         * be gfp-dependent anyway. Just let it go
3627         */
3628        if (unlikely(!memcg))
3629                return true;
3630
3631        if (!memcg_can_account_kmem(memcg)) {
3632                css_put(&memcg->css);
3633                return true;
3634        }
3635
3636        ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3637        if (!ret)
3638                *_memcg = memcg;
3639
3640        css_put(&memcg->css);
3641        return (ret == 0);
3642}
3643
3644void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3645                              int order)
3646{
3647        struct page_cgroup *pc;
3648
3649        VM_BUG_ON(mem_cgroup_is_root(memcg));
3650
3651        /* The page allocation failed. Revert */
3652        if (!page) {
3653                memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3654                return;
3655        }
3656
3657        pc = lookup_page_cgroup(page);
3658        lock_page_cgroup(pc);
3659        pc->mem_cgroup = memcg;
3660        SetPageCgroupUsed(pc);
3661        unlock_page_cgroup(pc);
3662}
3663
3664void __memcg_kmem_uncharge_pages(struct page *page, int order)
3665{
3666        struct mem_cgroup *memcg = NULL;
3667        struct page_cgroup *pc;
3668
3669
3670        pc = lookup_page_cgroup(page);
3671        /*
3672         * Fast unlocked return. Theoretically might have changed, have to
3673         * check again after locking.
3674         */
3675        if (!PageCgroupUsed(pc))
3676                return;
3677
3678        lock_page_cgroup(pc);
3679        if (PageCgroupUsed(pc)) {
3680                memcg = pc->mem_cgroup;
3681                ClearPageCgroupUsed(pc);
3682        }
3683        unlock_page_cgroup(pc);
3684
3685        /*
3686         * We trust that only if there is a memcg associated with the page, it
3687         * is a valid allocation
3688         */
3689        if (!memcg)
3690                return;
3691
3692        VM_BUG_ON(mem_cgroup_is_root(memcg));
3693        memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3694}
3695#else
3696static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3697{
3698}
3699#endif /* CONFIG_MEMCG_KMEM */
3700
3701#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3702
3703#define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3704/*
3705 * Because tail pages are not marked as "used", set it. We're under
3706 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3707 * charge/uncharge will be never happen and move_account() is done under
3708 * compound_lock(), so we don't have to take care of races.
3709 */
3710void mem_cgroup_split_huge_fixup(struct page *head)
3711{
3712        struct page_cgroup *head_pc = lookup_page_cgroup(head);
3713        struct page_cgroup *pc;
3714        struct mem_cgroup *memcg;
3715        int i;
3716
3717        if (mem_cgroup_disabled())
3718                return;
3719
3720        memcg = head_pc->mem_cgroup;
3721        for (i = 1; i < HPAGE_PMD_NR; i++) {
3722                pc = head_pc + i;
3723                pc->mem_cgroup = memcg;
3724                smp_wmb();/* see __commit_charge() */
3725                pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3726        }
3727        __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3728                       HPAGE_PMD_NR);
3729}
3730#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3731
3732/**
3733 * mem_cgroup_move_account - move account of the page
3734 * @page: the page
3735 * @nr_pages: number of regular pages (>1 for huge pages)
3736 * @pc: page_cgroup of the page.
3737 * @from: mem_cgroup which the page is moved from.
3738 * @to: mem_cgroup which the page is moved to. @from != @to.
3739 *
3740 * The caller must confirm following.
3741 * - page is not on LRU (isolate_page() is useful.)
3742 * - compound_lock is held when nr_pages > 1
3743 *
3744 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3745 * from old cgroup.
3746 */
3747static int mem_cgroup_move_account(struct page *page,
3748                                   unsigned int nr_pages,
3749                                   struct page_cgroup *pc,
3750                                   struct mem_cgroup *from,
3751                                   struct mem_cgroup *to)
3752{
3753        unsigned long flags;
3754        int ret;
3755        bool anon = PageAnon(page);
3756
3757        VM_BUG_ON(from == to);
3758        VM_BUG_ON(PageLRU(page));
3759        /*
3760         * The page is isolated from LRU. So, collapse function
3761         * will not handle this page. But page splitting can happen.
3762         * Do this check under compound_page_lock(). The caller should
3763         * hold it.
3764         */
3765        ret = -EBUSY;
3766        if (nr_pages > 1 && !PageTransHuge(page))
3767                goto out;
3768
3769        lock_page_cgroup(pc);
3770
3771        ret = -EINVAL;
3772        if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3773                goto unlock;
3774
3775        move_lock_mem_cgroup(from, &flags);
3776
3777        if (!anon && page_mapped(page)) {
3778                /* Update mapped_file data for mem_cgroup */
3779                preempt_disable();
3780                __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3781                __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3782                preempt_enable();
3783        }
3784        mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3785
3786        /* caller should have done css_get */
3787        pc->mem_cgroup = to;
3788        mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3789        move_unlock_mem_cgroup(from, &flags);
3790        ret = 0;
3791unlock:
3792        unlock_page_cgroup(pc);
3793        /*
3794         * check events
3795         */
3796        memcg_check_events(to, page);
3797        memcg_check_events(from, page);
3798out:
3799        return ret;
3800}
3801
3802/**
3803 * mem_cgroup_move_parent - moves page to the parent group
3804 * @page: the page to move
3805 * @pc: page_cgroup of the page
3806 * @child: page's cgroup
3807 *
3808 * move charges to its parent or the root cgroup if the group has no
3809 * parent (aka use_hierarchy==0).
3810 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3811 * mem_cgroup_move_account fails) the failure is always temporary and
3812 * it signals a race with a page removal/uncharge or migration. In the
3813 * first case the page is on the way out and it will vanish from the LRU
3814 * on the next attempt and the call should be retried later.
3815 * Isolation from the LRU fails only if page has been isolated from
3816 * the LRU since we looked at it and that usually means either global
3817 * reclaim or migration going on. The page will either get back to the
3818 * LRU or vanish.
3819 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3820 * (!PageCgroupUsed) or moved to a different group. The page will
3821 * disappear in the next attempt.
3822 */
3823static int mem_cgroup_move_parent(struct page *page,
3824                                  struct page_cgroup *pc,
3825                                  struct mem_cgroup *child)
3826{
3827        struct mem_cgroup *parent;
3828        unsigned int nr_pages;
3829        unsigned long uninitialized_var(flags);
3830        int ret;
3831
3832        VM_BUG_ON(mem_cgroup_is_root(child));
3833
3834        ret = -EBUSY;
3835        if (!get_page_unless_zero(page))
3836                goto out;
3837        if (isolate_lru_page(page))
3838                goto put;
3839
3840        nr_pages = hpage_nr_pages(page);
3841
3842        parent = parent_mem_cgroup(child);
3843        /*
3844         * If no parent, move charges to root cgroup.
3845         */
3846        if (!parent)
3847                parent = root_mem_cgroup;
3848
3849        if (nr_pages > 1) {
3850                VM_BUG_ON(!PageTransHuge(page));
3851                flags = compound_lock_irqsave(page);
3852        }
3853
3854        ret = mem_cgroup_move_account(page, nr_pages,
3855                                pc, child, parent);
3856        if (!ret)
3857                __mem_cgroup_cancel_local_charge(child, nr_pages);
3858
3859        if (nr_pages > 1)
3860                compound_unlock_irqrestore(page, flags);
3861        putback_lru_page(page);
3862put:
3863        put_page(page);
3864out:
3865        return ret;
3866}
3867
3868/*
3869 * Charge the memory controller for page usage.
3870 * Return
3871 * 0 if the charge was successful
3872 * < 0 if the cgroup is over its limit
3873 */
3874static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3875                                gfp_t gfp_mask, enum charge_type ctype)
3876{
3877        struct mem_cgroup *memcg = NULL;
3878        unsigned int nr_pages = 1;
3879        bool oom = true;
3880        int ret;
3881
3882        if (PageTransHuge(page)) {
3883                nr_pages <<= compound_order(page);
3884                VM_BUG_ON(!PageTransHuge(page));
3885                /*
3886                 * Never OOM-kill a process for a huge page.  The
3887                 * fault handler will fall back to regular pages.
3888                 */
3889                oom = false;
3890        }
3891
3892        ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3893        if (ret == -ENOMEM)
3894                return ret;
3895        __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3896        return 0;
3897}
3898
3899int mem_cgroup_newpage_charge(struct page *page,
3900                              struct mm_struct *mm, gfp_t gfp_mask)
3901{
3902        if (mem_cgroup_disabled())
3903                return 0;
3904        VM_BUG_ON(page_mapped(page));
3905        VM_BUG_ON(page->mapping && !PageAnon(page));
3906        VM_BUG_ON(!mm);
3907        return mem_cgroup_charge_common(page, mm, gfp_mask,
3908                                        MEM_CGROUP_CHARGE_TYPE_ANON);
3909}
3910
3911/*
3912 * While swap-in, try_charge -> commit or cancel, the page is locked.
3913 * And when try_charge() successfully returns, one refcnt to memcg without
3914 * struct page_cgroup is acquired. This refcnt will be consumed by
3915 * "commit()" or removed by "cancel()"
3916 */
3917static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3918                                          struct page *page,
3919                                          gfp_t mask,
3920                                          struct mem_cgroup **memcgp)
3921{
3922        struct mem_cgroup *memcg;
3923        struct page_cgroup *pc;
3924        int ret;
3925
3926        pc = lookup_page_cgroup(page);
3927        /*
3928         * Every swap fault against a single page tries to charge the
3929         * page, bail as early as possible.  shmem_unuse() encounters
3930         * already charged pages, too.  The USED bit is protected by
3931         * the page lock, which serializes swap cache removal, which
3932         * in turn serializes uncharging.
3933         */
3934        if (PageCgroupUsed(pc))
3935                return 0;
3936        if (!do_swap_account)
3937                goto charge_cur_mm;
3938        memcg = try_get_mem_cgroup_from_page(page);
3939        if (!memcg)
3940                goto charge_cur_mm;
3941        *memcgp = memcg;
3942        ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3943        css_put(&memcg->css);
3944        if (ret == -EINTR)
3945                ret = 0;
3946        return ret;
3947charge_cur_mm:
3948        ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3949        if (ret == -EINTR)
3950                ret = 0;
3951        return ret;
3952}
3953
3954int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3955                                 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3956{
3957        *memcgp = NULL;
3958        if (mem_cgroup_disabled())
3959                return 0;
3960        /*
3961         * A racing thread's fault, or swapoff, may have already
3962         * updated the pte, and even removed page from swap cache: in
3963         * those cases unuse_pte()'s pte_same() test will fail; but
3964         * there's also a KSM case which does need to charge the page.
3965         */
3966        if (!PageSwapCache(page)) {
3967                int ret;
3968
3969                ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3970                if (ret == -EINTR)
3971                        ret = 0;
3972                return ret;
3973        }
3974        return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3975}
3976
3977void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3978{
3979        if (mem_cgroup_disabled())
3980                return;
3981        if (!memcg)
3982                return;
3983        __mem_cgroup_cancel_charge(memcg, 1);
3984}
3985
3986static void
3987__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3988                                        enum charge_type ctype)
3989{
3990        if (mem_cgroup_disabled())
3991                return;
3992        if (!memcg)
3993                return;
3994
3995        __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3996        /*
3997         * Now swap is on-memory. This means this page may be
3998         * counted both as mem and swap....double count.
3999         * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4000         * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4001         * may call delete_from_swap_cache() before reach here.
4002         */
4003        if (do_swap_account && PageSwapCache(page)) {
4004                swp_entry_t ent = {.val = page_private(page)};
4005                mem_cgroup_uncharge_swap(ent);
4006        }
4007}
4008
4009void mem_cgroup_commit_charge_swapin(struct page *page,
4010                                     struct mem_cgroup *memcg)
4011{
4012        __mem_cgroup_commit_charge_swapin(page, memcg,
4013                                          MEM_CGROUP_CHARGE_TYPE_ANON);
4014}
4015
4016int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4017                                gfp_t gfp_mask)
4018{
4019        struct mem_cgroup *memcg = NULL;
4020        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4021        int ret;
4022
4023        if (mem_cgroup_disabled())
4024                return 0;
4025        if (PageCompound(page))
4026                return 0;
4027
4028        if (!PageSwapCache(page))
4029                ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4030        else { /* page is swapcache/shmem */
4031                ret = __mem_cgroup_try_charge_swapin(mm, page,
4032                                                     gfp_mask, &memcg);
4033                if (!ret)
4034                        __mem_cgroup_commit_charge_swapin(page, memcg, type);
4035        }
4036        return ret;
4037}
4038
4039static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4040                                   unsigned int nr_pages,
4041                                   const enum charge_type ctype)
4042{
4043        struct memcg_batch_info *batch = NULL;
4044        bool uncharge_memsw = true;
4045
4046        /* If swapout, usage of swap doesn't decrease */
4047        if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4048                uncharge_memsw = false;
4049
4050        batch = &current->memcg_batch;
4051        /*
4052         * In usual, we do css_get() when we remember memcg pointer.
4053         * But in this case, we keep res->usage until end of a series of
4054         * uncharges. Then, it's ok to ignore memcg's refcnt.
4055         */
4056        if (!batch->memcg)
4057                batch->memcg = memcg;
4058        /*
4059         * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4060         * In those cases, all pages freed continuously can be expected to be in
4061         * the same cgroup and we have chance to coalesce uncharges.
4062         * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4063         * because we want to do uncharge as soon as possible.
4064         */
4065
4066        if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4067                goto direct_uncharge;
4068
4069        if (nr_pages > 1)
4070                goto direct_uncharge;
4071
4072        /*
4073         * In typical case, batch->memcg == mem. This means we can
4074         * merge a series of uncharges to an uncharge of res_counter.
4075         * If not, we uncharge res_counter ony by one.
4076         */
4077        if (batch->memcg != memcg)
4078                goto direct_uncharge;
4079        /* remember freed charge and uncharge it later */
4080        batch->nr_pages++;
4081        if (uncharge_memsw)
4082                batch->memsw_nr_pages++;
4083        return;
4084direct_uncharge:
4085        res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4086        if (uncharge_memsw)
4087                res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4088        if (unlikely(batch->memcg != memcg))
4089                memcg_oom_recover(memcg);
4090}
4091
4092/*
4093 * uncharge if !page_mapped(page)
4094 */
4095static struct mem_cgroup *
4096__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4097                             bool end_migration)
4098{
4099        struct mem_cgroup *memcg = NULL;
4100        unsigned int nr_pages = 1;
4101        struct page_cgroup *pc;
4102        bool anon;
4103
4104        if (mem_cgroup_disabled())
4105                return NULL;
4106
4107        if (PageTransHuge(page)) {
4108                nr_pages <<= compound_order(page);
4109                VM_BUG_ON(!PageTransHuge(page));
4110        }
4111        /*
4112         * Check if our page_cgroup is valid
4113         */
4114        pc = lookup_page_cgroup(page);
4115        if (unlikely(!PageCgroupUsed(pc)))
4116                return NULL;
4117
4118        lock_page_cgroup(pc);
4119
4120        memcg = pc->mem_cgroup;
4121
4122        if (!PageCgroupUsed(pc))
4123                goto unlock_out;
4124
4125        anon = PageAnon(page);
4126
4127        switch (ctype) {
4128        case MEM_CGROUP_CHARGE_TYPE_ANON:
4129                /*
4130                 * Generally PageAnon tells if it's the anon statistics to be
4131                 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4132                 * used before page reached the stage of being marked PageAnon.
4133                 */
4134                anon = true;
4135                /* fallthrough */
4136        case MEM_CGROUP_CHARGE_TYPE_DROP:
4137                /* See mem_cgroup_prepare_migration() */
4138                if (page_mapped(page))
4139                        goto unlock_out;
4140                /*
4141                 * Pages under migration may not be uncharged.  But
4142                 * end_migration() /must/ be the one uncharging the
4143                 * unused post-migration page and so it has to call
4144                 * here with the migration bit still set.  See the
4145                 * res_counter handling below.
4146                 */
4147                if (!end_migration && PageCgroupMigration(pc))
4148                        goto unlock_out;
4149                break;
4150        case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4151                if (!PageAnon(page)) {  /* Shared memory */
4152                        if (page->mapping && !page_is_file_cache(page))
4153                                goto unlock_out;
4154                } else if (page_mapped(page)) /* Anon */
4155                                goto unlock_out;
4156                break;
4157        default:
4158                break;
4159        }
4160
4161        mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4162
4163        ClearPageCgroupUsed(pc);
4164        /*
4165         * pc->mem_cgroup is not cleared here. It will be accessed when it's
4166         * freed from LRU. This is safe because uncharged page is expected not
4167         * to be reused (freed soon). Exception is SwapCache, it's handled by
4168         * special functions.
4169         */
4170
4171        unlock_page_cgroup(pc);
4172        /*
4173         * even after unlock, we have memcg->res.usage here and this memcg
4174         * will never be freed.
4175         */
4176        memcg_check_events(memcg, page);
4177        if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4178                mem_cgroup_swap_statistics(memcg, true);
4179                mem_cgroup_get(memcg);
4180        }
4181        /*
4182         * Migration does not charge the res_counter for the
4183         * replacement page, so leave it alone when phasing out the
4184         * page that is unused after the migration.
4185         */
4186        if (!end_migration && !mem_cgroup_is_root(memcg))
4187                mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4188
4189        return memcg;
4190
4191unlock_out:
4192        unlock_page_cgroup(pc);
4193        return NULL;
4194}
4195
4196void mem_cgroup_uncharge_page(struct page *page)
4197{
4198        /* early check. */
4199        if (page_mapped(page))
4200                return;
4201        VM_BUG_ON(page->mapping && !PageAnon(page));
4202        /*
4203         * If the page is in swap cache, uncharge should be deferred
4204         * to the swap path, which also properly accounts swap usage
4205         * and handles memcg lifetime.
4206         *
4207         * Note that this check is not stable and reclaim may add the
4208         * page to swap cache at any time after this.  However, if the
4209         * page is not in swap cache by the time page->mapcount hits
4210         * 0, there won't be any page table references to the swap
4211         * slot, and reclaim will free it and not actually write the
4212         * page to disk.
4213         */
4214        if (PageSwapCache(page))
4215                return;
4216        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4217}
4218
4219void mem_cgroup_uncharge_cache_page(struct page *page)
4220{
4221        VM_BUG_ON(page_mapped(page));
4222        VM_BUG_ON(page->mapping);
4223        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4224}
4225
4226/*
4227 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4228 * In that cases, pages are freed continuously and we can expect pages
4229 * are in the same memcg. All these calls itself limits the number of
4230 * pages freed at once, then uncharge_start/end() is called properly.
4231 * This may be called prural(2) times in a context,
4232 */
4233
4234void mem_cgroup_uncharge_start(void)
4235{
4236        current->memcg_batch.do_batch++;
4237        /* We can do nest. */
4238        if (current->memcg_batch.do_batch == 1) {
4239                current->memcg_batch.memcg = NULL;
4240                current->memcg_batch.nr_pages = 0;
4241                current->memcg_batch.memsw_nr_pages = 0;
4242        }
4243}
4244
4245void mem_cgroup_uncharge_end(void)
4246{
4247        struct memcg_batch_info *batch = &current->memcg_batch;
4248
4249        if (!batch->do_batch)
4250                return;
4251
4252        batch->do_batch--;
4253        if (batch->do_batch) /* If stacked, do nothing. */
4254                return;
4255
4256        if (!batch->memcg)
4257                return;
4258        /*
4259         * This "batch->memcg" is valid without any css_get/put etc...
4260         * bacause we hide charges behind us.
4261         */
4262        if (batch->nr_pages)
4263                res_counter_uncharge(&batch->memcg->res,
4264                                     batch->nr_pages * PAGE_SIZE);
4265        if (batch->memsw_nr_pages)
4266                res_counter_uncharge(&batch->memcg->memsw,
4267                                     batch->memsw_nr_pages * PAGE_SIZE);
4268        memcg_oom_recover(batch->memcg);
4269        /* forget this pointer (for sanity check) */
4270        batch->memcg = NULL;
4271}
4272
4273#ifdef CONFIG_SWAP
4274/*
4275 * called after __delete_from_swap_cache() and drop "page" account.
4276 * memcg information is recorded to swap_cgroup of "ent"
4277 */
4278void
4279mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4280{
4281        struct mem_cgroup *memcg;
4282        int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4283
4284        if (!swapout) /* this was a swap cache but the swap is unused ! */
4285                ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4286
4287        memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4288
4289        /*
4290         * record memcg information,  if swapout && memcg != NULL,
4291         * mem_cgroup_get() was called in uncharge().
4292         */
4293        if (do_swap_account && swapout && memcg)
4294                swap_cgroup_record(ent, css_id(&memcg->css));
4295}
4296#endif
4297
4298#ifdef CONFIG_MEMCG_SWAP
4299/*
4300 * called from swap_entry_free(). remove record in swap_cgroup and
4301 * uncharge "memsw" account.
4302 */
4303void mem_cgroup_uncharge_swap(swp_entry_t ent)
4304{
4305        struct mem_cgroup *memcg;
4306        unsigned short id;
4307
4308        if (!do_swap_account)
4309                return;
4310
4311        id = swap_cgroup_record(ent, 0);
4312        rcu_read_lock();
4313        memcg = mem_cgroup_lookup(id);
4314        if (memcg) {
4315                /*
4316                 * We uncharge this because swap is freed.
4317                 * This memcg can be obsolete one. We avoid calling css_tryget
4318                 */
4319                if (!mem_cgroup_is_root(memcg))
4320                        res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4321                mem_cgroup_swap_statistics(memcg, false);
4322                mem_cgroup_put(memcg);
4323        }
4324        rcu_read_unlock();
4325}
4326
4327/**
4328 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4329 * @entry: swap entry to be moved
4330 * @from:  mem_cgroup which the entry is moved from
4331 * @to:  mem_cgroup which the entry is moved to
4332 *
4333 * It succeeds only when the swap_cgroup's record for this entry is the same
4334 * as the mem_cgroup's id of @from.
4335 *
4336 * Returns 0 on success, -EINVAL on failure.
4337 *
4338 * The caller must have charged to @to, IOW, called res_counter_charge() about
4339 * both res and memsw, and called css_get().
4340 */
4341static int mem_cgroup_move_swap_account(swp_entry_t entry,
4342                                struct mem_cgroup *from, struct mem_cgroup *to)
4343{
4344        unsigned short old_id, new_id;
4345
4346        old_id = css_id(&from->css);
4347        new_id = css_id(&to->css);
4348
4349        if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4350                mem_cgroup_swap_statistics(from, false);
4351                mem_cgroup_swap_statistics(to, true);
4352                /*
4353                 * This function is only called from task migration context now.
4354                 * It postpones res_counter and refcount handling till the end
4355                 * of task migration(mem_cgroup_clear_mc()) for performance
4356                 * improvement. But we cannot postpone mem_cgroup_get(to)
4357                 * because if the process that has been moved to @to does
4358                 * swap-in, the refcount of @to might be decreased to 0.
4359                 */
4360                mem_cgroup_get(to);
4361                return 0;
4362        }
4363        return -EINVAL;
4364}
4365#else
4366static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4367                                struct mem_cgroup *from, struct mem_cgroup *to)
4368{
4369        return -EINVAL;
4370}
4371#endif
4372
4373/*
4374 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4375 * page belongs to.
4376 */
4377void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4378                                  struct mem_cgroup **memcgp)
4379{
4380        struct mem_cgroup *memcg = NULL;
4381        unsigned int nr_pages = 1;
4382        struct page_cgroup *pc;
4383        enum charge_type ctype;
4384
4385        *memcgp = NULL;
4386
4387        if (mem_cgroup_disabled())
4388                return;
4389
4390        if (PageTransHuge(page))
4391                nr_pages <<= compound_order(page);
4392
4393        pc = lookup_page_cgroup(page);
4394        lock_page_cgroup(pc);
4395        if (PageCgroupUsed(pc)) {
4396                memcg = pc->mem_cgroup;
4397                css_get(&memcg->css);
4398                /*
4399                 * At migrating an anonymous page, its mapcount goes down
4400                 * to 0 and uncharge() will be called. But, even if it's fully
4401                 * unmapped, migration may fail and this page has to be
4402                 * charged again. We set MIGRATION flag here and delay uncharge
4403                 * until end_migration() is called
4404                 *
4405                 * Corner Case Thinking
4406                 * A)
4407                 * When the old page was mapped as Anon and it's unmap-and-freed
4408                 * while migration was ongoing.
4409                 * If unmap finds the old page, uncharge() of it will be delayed
4410                 * until end_migration(). If unmap finds a new page, it's
4411                 * uncharged when it make mapcount to be 1->0. If unmap code
4412                 * finds swap_migration_entry, the new page will not be mapped
4413                 * and end_migration() will find it(mapcount==0).
4414                 *
4415                 * B)
4416                 * When the old page was mapped but migraion fails, the kernel
4417                 * remaps it. A charge for it is kept by MIGRATION flag even
4418                 * if mapcount goes down to 0. We can do remap successfully
4419                 * without charging it again.
4420                 *
4421                 * C)
4422                 * The "old" page is under lock_page() until the end of
4423                 * migration, so, the old page itself will not be swapped-out.
4424                 * If the new page is swapped out before end_migraton, our
4425                 * hook to usual swap-out path will catch the event.
4426                 */
4427                if (PageAnon(page))
4428                        SetPageCgroupMigration(pc);
4429        }
4430        unlock_page_cgroup(pc);
4431        /*
4432         * If the page is not charged at this point,
4433         * we return here.
4434         */
4435        if (!memcg)
4436                return;
4437
4438        *memcgp = memcg;
4439        /*
4440         * We charge new page before it's used/mapped. So, even if unlock_page()
4441         * is called before end_migration, we can catch all events on this new
4442         * page. In the case new page is migrated but not remapped, new page's
4443         * mapcount will be finally 0 and we call uncharge in end_migration().
4444         */
4445        if (PageAnon(page))
4446                ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4447        else
4448                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4449        /*
4450         * The page is committed to the memcg, but it's not actually
4451         * charged to the res_counter since we plan on replacing the
4452         * old one and only one page is going to be left afterwards.
4453         */
4454        __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4455}
4456
4457/* remove redundant charge if migration failed*/
4458void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4459        struct page *oldpage, struct page *newpage, bool migration_ok)
4460{
4461        struct page *used, *unused;
4462        struct page_cgroup *pc;
4463        bool anon;
4464
4465        if (!memcg)
4466                return;
4467
4468        if (!migration_ok) {
4469                used = oldpage;
4470                unused = newpage;
4471        } else {
4472                used = newpage;
4473                unused = oldpage;
4474        }
4475        anon = PageAnon(used);
4476        __mem_cgroup_uncharge_common(unused,
4477                                     anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4478                                     : MEM_CGROUP_CHARGE_TYPE_CACHE,
4479                                     true);
4480        css_put(&memcg->css);
4481        /*
4482         * We disallowed uncharge of pages under migration because mapcount
4483         * of the page goes down to zero, temporarly.
4484         * Clear the flag and check the page should be charged.
4485         */
4486        pc = lookup_page_cgroup(oldpage);
4487        lock_page_cgroup(pc);
4488        ClearPageCgroupMigration(pc);
4489        unlock_page_cgroup(pc);
4490
4491        /*
4492         * If a page is a file cache, radix-tree replacement is very atomic
4493         * and we can skip this check. When it was an Anon page, its mapcount
4494         * goes down to 0. But because we added MIGRATION flage, it's not
4495         * uncharged yet. There are several case but page->mapcount check
4496         * and USED bit check in mem_cgroup_uncharge_page() will do enough
4497         * check. (see prepare_charge() also)
4498         */
4499        if (anon)
4500                mem_cgroup_uncharge_page(used);
4501}
4502
4503/*
4504 * At replace page cache, newpage is not under any memcg but it's on
4505 * LRU. So, this function doesn't touch res_counter but handles LRU
4506 * in correct way. Both pages are locked so we cannot race with uncharge.
4507 */
4508void mem_cgroup_replace_page_cache(struct page *oldpage,
4509                                  struct page *newpage)
4510{
4511        struct mem_cgroup *memcg = NULL;
4512        struct page_cgroup *pc;
4513        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4514
4515        if (mem_cgroup_disabled())
4516                return;
4517
4518        pc = lookup_page_cgroup(oldpage);
4519        /* fix accounting on old pages */
4520        lock_page_cgroup(pc);
4521        if (PageCgroupUsed(pc)) {
4522                memcg = pc->mem_cgroup;
4523                mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4524                ClearPageCgroupUsed(pc);
4525        }
4526        unlock_page_cgroup(pc);
4527
4528        /*
4529         * When called from shmem_replace_page(), in some cases the
4530         * oldpage has already been charged, and in some cases not.
4531         */
4532        if (!memcg)
4533                return;
4534        /*
4535         * Even if newpage->mapping was NULL before starting replacement,
4536         * the newpage may be on LRU(or pagevec for LRU) already. We lock
4537         * LRU while we overwrite pc->mem_cgroup.
4538         */
4539        __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4540}
4541
4542#ifdef CONFIG_DEBUG_VM
4543static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4544{
4545        struct page_cgroup *pc;
4546
4547        pc = lookup_page_cgroup(page);
4548        /*
4549         * Can be NULL while feeding pages into the page allocator for
4550         * the first time, i.e. during boot or memory hotplug;
4551         * or when mem_cgroup_disabled().
4552         */
4553        if (likely(pc) && PageCgroupUsed(pc))
4554                return pc;
4555        return NULL;
4556}
4557
4558bool mem_cgroup_bad_page_check(struct page *page)
4559{
4560        if (mem_cgroup_disabled())
4561                return false;
4562
4563        return lookup_page_cgroup_used(page) != NULL;
4564}
4565
4566void mem_cgroup_print_bad_page(struct page *page)
4567{
4568        struct page_cgroup *pc;
4569
4570        pc = lookup_page_cgroup_used(page);
4571        if (pc) {
4572                pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4573                         pc, pc->flags, pc->mem_cgroup);
4574        }
4575}
4576#endif
4577
4578static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4579                                unsigned long long val)
4580{
4581        int retry_count;
4582        u64 memswlimit, memlimit;
4583        int ret = 0;
4584        int children = mem_cgroup_count_children(memcg);
4585        u64 curusage, oldusage;
4586        int enlarge;
4587
4588        /*
4589         * For keeping hierarchical_reclaim simple, how long we should retry
4590         * is depends on callers. We set our retry-count to be function
4591         * of # of children which we should visit in this loop.
4592         */
4593        retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4594
4595        oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4596
4597        enlarge = 0;
4598        while (retry_count) {
4599                if (signal_pending(current)) {
4600                        ret = -EINTR;
4601                        break;
4602                }
4603                /*
4604                 * Rather than hide all in some function, I do this in
4605                 * open coded manner. You see what this really does.
4606                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4607                 */
4608                mutex_lock(&set_limit_mutex);
4609                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4610                if (memswlimit < val) {
4611                        ret = -EINVAL;
4612                        mutex_unlock(&set_limit_mutex);
4613                        break;
4614                }
4615
4616                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4617                if (memlimit < val)
4618                        enlarge = 1;
4619
4620                ret = res_counter_set_limit(&memcg->res, val);
4621                if (!ret) {
4622                        if (memswlimit == val)
4623                                memcg->memsw_is_minimum = true;
4624                        else
4625                                memcg->memsw_is_minimum = false;
4626                }
4627                mutex_unlock(&set_limit_mutex);
4628
4629                if (!ret)
4630                        break;
4631
4632                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4633                                   MEM_CGROUP_RECLAIM_SHRINK);
4634                curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4635                /* Usage is reduced ? */
4636                if (curusage >= oldusage)
4637                        retry_count--;
4638                else
4639                        oldusage = curusage;
4640        }
4641        if (!ret && enlarge)
4642                memcg_oom_recover(memcg);
4643
4644        return ret;
4645}
4646
4647static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4648                                        unsigned long long val)
4649{
4650        int retry_count;
4651        u64 memlimit, memswlimit, oldusage, curusage;
4652        int children = mem_cgroup_count_children(memcg);
4653        int ret = -EBUSY;
4654        int enlarge = 0;
4655
4656        /* see mem_cgroup_resize_res_limit */
4657        retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4658        oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4659        while (retry_count) {
4660                if (signal_pending(current)) {
4661                        ret = -EINTR;
4662                        break;
4663                }
4664                /*
4665                 * Rather than hide all in some function, I do this in
4666                 * open coded manner. You see what this really does.
4667                 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4668                 */
4669                mutex_lock(&set_limit_mutex);
4670                memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4671                if (memlimit > val) {
4672                        ret = -EINVAL;
4673                        mutex_unlock(&set_limit_mutex);
4674                        break;
4675                }
4676                memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4677                if (memswlimit < val)
4678                        enlarge = 1;
4679                ret = res_counter_set_limit(&memcg->memsw, val);
4680                if (!ret) {
4681                        if (memlimit == val)
4682                                memcg->memsw_is_minimum = true;
4683                        else
4684                                memcg->memsw_is_minimum = false;
4685                }
4686                mutex_unlock(&set_limit_mutex);
4687
4688                if (!ret)
4689                        break;
4690
4691                mem_cgroup_reclaim(memcg, GFP_KERNEL,
4692                                   MEM_CGROUP_RECLAIM_NOSWAP |
4693                                   MEM_CGROUP_RECLAIM_SHRINK);
4694                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4695                /* Usage is reduced ? */
4696                if (curusage >= oldusage)
4697                        retry_count--;
4698                else
4699                        oldusage = curusage;
4700        }
4701        if (!ret && enlarge)
4702                memcg_oom_recover(memcg);
4703        return ret;
4704}
4705
4706unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4707                                            gfp_t gfp_mask,
4708                                            unsigned long *total_scanned)
4709{
4710        unsigned long nr_reclaimed = 0;
4711        struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4712        unsigned long reclaimed;
4713        int loop = 0;
4714        struct mem_cgroup_tree_per_zone *mctz;
4715        unsigned long long excess;
4716        unsigned long nr_scanned;
4717
4718        if (order > 0)
4719                return 0;
4720
4721        mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4722        /*
4723         * This loop can run a while, specially if mem_cgroup's continuously
4724         * keep exceeding their soft limit and putting the system under
4725         * pressure
4726         */
4727        do {
4728                if (next_mz)
4729                        mz = next_mz;
4730                else
4731                        mz = mem_cgroup_largest_soft_limit_node(mctz);
4732                if (!mz)
4733                        break;
4734
4735                nr_scanned = 0;
4736                reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4737                                                    gfp_mask, &nr_scanned);
4738                nr_reclaimed += reclaimed;
4739                *total_scanned += nr_scanned;
4740                spin_lock(&mctz->lock);
4741
4742                /*
4743                 * If we failed to reclaim anything from this memory cgroup
4744                 * it is time to move on to the next cgroup
4745                 */
4746                next_mz = NULL;
4747                if (!reclaimed) {
4748                        do {
4749                                /*
4750                                 * Loop until we find yet another one.
4751                                 *
4752                                 * By the time we get the soft_limit lock
4753                                 * again, someone might have aded the
4754                                 * group back on the RB tree. Iterate to
4755                                 * make sure we get a different mem.
4756                                 * mem_cgroup_largest_soft_limit_node returns
4757                                 * NULL if no other cgroup is present on
4758                                 * the tree
4759                                 */
4760                                next_mz =
4761                                __mem_cgroup_largest_soft_limit_node(mctz);
4762                                if (next_mz == mz)
4763                                        css_put(&next_mz->memcg->css);
4764                                else /* next_mz == NULL or other memcg */
4765                                        break;
4766                        } while (1);
4767                }
4768                __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4769                excess = res_counter_soft_limit_excess(&mz->memcg->res);
4770                /*
4771                 * One school of thought says that we should not add
4772                 * back the node to the tree if reclaim returns 0.
4773                 * But our reclaim could return 0, simply because due
4774                 * to priority we are exposing a smaller subset of
4775                 * memory to reclaim from. Consider this as a longer
4776                 * term TODO.
4777                 */
4778                /* If excess == 0, no tree ops */
4779                __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4780                spin_unlock(&mctz->lock);
4781                css_put(&mz->memcg->css);
4782                loop++;
4783                /*
4784                 * Could not reclaim anything and there are no more
4785                 * mem cgroups to try or we seem to be looping without
4786                 * reclaiming anything.
4787                 */
4788                if (!nr_reclaimed &&
4789                        (next_mz == NULL ||
4790                        loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4791                        break;
4792        } while (!nr_reclaimed);
4793        if (next_mz)
4794                css_put(&next_mz->memcg->css);
4795        return nr_reclaimed;
4796}
4797
4798/**
4799 * mem_cgroup_force_empty_list - clears LRU of a group
4800 * @memcg: group to clear
4801 * @node: NUMA node
4802 * @zid: zone id
4803 * @lru: lru to to clear
4804 *
4805 * Traverse a specified page_cgroup list and try to drop them all.  This doesn't
4806 * reclaim the pages page themselves - pages are moved to the parent (or root)
4807 * group.
4808 */
4809static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4810                                int node, int zid, enum lru_list lru)
4811{
4812        struct lruvec *lruvec;
4813        unsigned long flags;
4814        struct list_head *list;
4815        struct page *busy;
4816        struct zone *zone;
4817
4818        zone = &NODE_DATA(node)->node_zones[zid];
4819        lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4820        list = &lruvec->lists[lru];
4821
4822        busy = NULL;
4823        do {
4824                struct page_cgroup *pc;
4825                struct page *page;
4826
4827                spin_lock_irqsave(&zone->lru_lock, flags);
4828                if (list_empty(list)) {
4829                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4830                        break;
4831                }
4832                page = list_entry(list->prev, struct page, lru);
4833                if (busy == page) {
4834                        list_move(&page->lru, list);
4835                        busy = NULL;
4836                        spin_unlock_irqrestore(&zone->lru_lock, flags);
4837                        continue;
4838                }
4839                spin_unlock_irqrestore(&zone->lru_lock, flags);
4840
4841                pc = lookup_page_cgroup(page);
4842
4843                if (mem_cgroup_move_parent(page, pc, memcg)) {
4844                        /* found lock contention or "pc" is obsolete. */
4845                        busy = page;
4846                        cond_resched();
4847                } else
4848                        busy = NULL;
4849        } while (!list_empty(list));
4850}
4851
4852/*
4853 * make mem_cgroup's charge to be 0 if there is no task by moving
4854 * all the charges and pages to the parent.
4855 * This enables deleting this mem_cgroup.
4856 *
4857 * Caller is responsible for holding css reference on the memcg.
4858 */
4859static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4860{
4861        int node, zid;
4862        u64 usage;
4863
4864        do {
4865                /* This is for making all *used* pages to be on LRU. */
4866                lru_add_drain_all();
4867                drain_all_stock_sync(memcg);
4868                mem_cgroup_start_move(memcg);
4869                for_each_node_state(node, N_MEMORY) {
4870                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4871                                enum lru_list lru;
4872                                for_each_lru(lru) {
4873                                        mem_cgroup_force_empty_list(memcg,
4874                                                        node, zid, lru);
4875                                }
4876                        }
4877                }
4878                mem_cgroup_end_move(memcg);
4879                memcg_oom_recover(memcg);
4880                cond_resched();
4881
4882                /*
4883                 * Kernel memory may not necessarily be trackable to a specific
4884                 * process. So they are not migrated, and therefore we can't
4885                 * expect their value to drop to 0 here.
4886                 * Having res filled up with kmem only is enough.
4887                 *
4888                 * This is a safety check because mem_cgroup_force_empty_list
4889                 * could have raced with mem_cgroup_replace_page_cache callers
4890                 * so the lru seemed empty but the page could have been added
4891                 * right after the check. RES_USAGE should be safe as we always
4892                 * charge before adding to the LRU.
4893                 */
4894                usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4895                        res_counter_read_u64(&memcg->kmem, RES_USAGE);
4896        } while (usage > 0);
4897}
4898
4899/*
4900 * This mainly exists for tests during the setting of set of use_hierarchy.
4901 * Since this is the very setting we are changing, the current hierarchy value
4902 * is meaningless
4903 */
4904static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4905{
4906        struct cgroup *pos;
4907
4908        /* bounce at first found */
4909        cgroup_for_each_child(pos, memcg->css.cgroup)
4910                return true;
4911        return false;
4912}
4913
4914/*
4915 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4916 * to be already dead (as in mem_cgroup_force_empty, for instance).  This is
4917 * from mem_cgroup_count_children(), in the sense that we don't really care how
4918 * many children we have; we only need to know if we have any.  It also counts
4919 * any memcg without hierarchy as infertile.
4920 */
4921static inline bool memcg_has_children(struct mem_cgroup *memcg)
4922{
4923        return memcg->use_hierarchy && __memcg_has_children(memcg);
4924}
4925
4926/*
4927 * Reclaims as many pages from the given memcg as possible and moves
4928 * the rest to the parent.
4929 *
4930 * Caller is responsible for holding css reference for memcg.
4931 */
4932static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4933{
4934        int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4935        struct cgroup *cgrp = memcg->css.cgroup;
4936
4937        /* returns EBUSY if there is a task or if we come here twice. */
4938        if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4939                return -EBUSY;
4940
4941        /* we call try-to-free pages for make this cgroup empty */
4942        lru_add_drain_all();
4943        /* try to free all pages in this cgroup */
4944        while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4945                int progress;
4946
4947                if (signal_pending(current))
4948                        return -EINTR;
4949
4950                progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4951                                                false);
4952                if (!progress) {
4953                        nr_retries--;
4954                        /* maybe some writeback is necessary */
4955                        congestion_wait(BLK_RW_ASYNC, HZ/10);
4956                }
4957
4958        }
4959        lru_add_drain();
4960        mem_cgroup_reparent_charges(memcg);
4961
4962        return 0;
4963}
4964
4965static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4966{
4967        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4968        int ret;
4969
4970        if (mem_cgroup_is_root(memcg))
4971                return -EINVAL;
4972        css_get(&memcg->css);
4973        ret = mem_cgroup_force_empty(memcg);
4974        css_put(&memcg->css);
4975
4976        return ret;
4977}
4978
4979
4980static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4981{
4982        return mem_cgroup_from_cont(cont)->use_hierarchy;
4983}
4984
4985static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4986                                        u64 val)
4987{
4988        int retval = 0;
4989        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4990        struct cgroup *parent = cont->parent;
4991        struct mem_cgroup *parent_memcg = NULL;
4992
4993        if (parent)
4994                parent_memcg = mem_cgroup_from_cont(parent);
4995
4996        mutex_lock(&memcg_create_mutex);
4997
4998        if (memcg->use_hierarchy == val)
4999                goto out;
5000
5001        /*
5002         * If parent's use_hierarchy is set, we can't make any modifications
5003         * in the child subtrees. If it is unset, then the change can
5004         * occur, provided the current cgroup has no children.
5005         *
5006         * For the root cgroup, parent_mem is NULL, we allow value to be
5007         * set if there are no children.
5008         */
5009        if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5010                                (val == 1 || val == 0)) {
5011                if (!__memcg_has_children(memcg))
5012                        memcg->use_hierarchy = val;
5013                else
5014                        retval = -EBUSY;
5015        } else
5016                retval = -EINVAL;
5017
5018out:
5019        mutex_unlock(&memcg_create_mutex);
5020
5021        return retval;
5022}
5023
5024
5025static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5026                                               enum mem_cgroup_stat_index idx)
5027{
5028        struct mem_cgroup *iter;
5029        long val = 0;
5030
5031        /* Per-cpu values can be negative, use a signed accumulator */
5032        for_each_mem_cgroup_tree(iter, memcg)
5033                val += mem_cgroup_read_stat(iter, idx);
5034
5035        if (val < 0) /* race ? */
5036                val = 0;
5037        return val;
5038}
5039
5040static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5041{
5042        u64 val;
5043
5044        if (!mem_cgroup_is_root(memcg)) {
5045                if (!swap)
5046                        return res_counter_read_u64(&memcg->res, RES_USAGE);
5047                else
5048                        return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5049        }
5050
5051        /*
5052         * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5053         * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5054         */
5055        val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5056        val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5057
5058        if (swap)
5059                val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5060
5061        return val << PAGE_SHIFT;
5062}
5063
5064static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5065                               struct file *file, char __user *buf,
5066                               size_t nbytes, loff_t *ppos)
5067{
5068        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5069        char str[64];
5070        u64 val;
5071        int name, len;
5072        enum res_type type;
5073
5074        type = MEMFILE_TYPE(cft->private);
5075        name = MEMFILE_ATTR(cft->private);
5076
5077        switch (type) {
5078        case _MEM:
5079                if (name == RES_USAGE)
5080                        val = mem_cgroup_usage(memcg, false);
5081                else
5082                        val = res_counter_read_u64(&memcg->res, name);
5083                break;
5084        case _MEMSWAP:
5085                if (name == RES_USAGE)
5086                        val = mem_cgroup_usage(memcg, true);
5087                else
5088                        val = res_counter_read_u64(&memcg->memsw, name);
5089                break;
5090        case _KMEM:
5091                val = res_counter_read_u64(&memcg->kmem, name);
5092                break;
5093        default:
5094                BUG();
5095        }
5096
5097        len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5098        return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5099}
5100
5101static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5102{
5103        int ret = -EINVAL;
5104#ifdef CONFIG_MEMCG_KMEM
5105        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5106        /*
5107         * For simplicity, we won't allow this to be disabled.  It also can't
5108         * be changed if the cgroup has children already, or if tasks had
5109         * already joined.
5110         *
5111         * If tasks join before we set the limit, a person looking at
5112         * kmem.usage_in_bytes will have no way to determine when it took
5113         * place, which makes the value quite meaningless.
5114         *
5115         * After it first became limited, changes in the value of the limit are
5116         * of course permitted.
5117         */
5118        mutex_lock(&memcg_create_mutex);
5119        mutex_lock(&set_limit_mutex);
5120        if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5121                if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5122                        ret = -EBUSY;
5123                        goto out;
5124                }
5125                ret = res_counter_set_limit(&memcg->kmem, val);
5126                VM_BUG_ON(ret);
5127
5128                ret = memcg_update_cache_sizes(memcg);
5129                if (ret) {
5130                        res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5131                        goto out;
5132                }
5133                static_key_slow_inc(&memcg_kmem_enabled_key);
5134                /*
5135                 * setting the active bit after the inc will guarantee no one
5136                 * starts accounting before all call sites are patched
5137                 */
5138                memcg_kmem_set_active(memcg);
5139
5140                /*
5141                 * kmem charges can outlive the cgroup. In the case of slab
5142                 * pages, for instance, a page contain objects from various
5143                 * processes, so it is unfeasible to migrate them away. We
5144                 * need to reference count the memcg because of that.
5145                 */
5146                mem_cgroup_get(memcg);
5147        } else
5148                ret = res_counter_set_limit(&memcg->kmem, val);
5149out:
5150        mutex_unlock(&set_limit_mutex);
5151        mutex_unlock(&memcg_create_mutex);
5152#endif
5153        return ret;
5154}
5155
5156#ifdef CONFIG_MEMCG_KMEM
5157static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5158{
5159        int ret = 0;
5160        struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5161        if (!parent)
5162                goto out;
5163
5164        memcg->kmem_account_flags = parent->kmem_account_flags;
5165        /*
5166         * When that happen, we need to disable the static branch only on those
5167         * memcgs that enabled it. To achieve this, we would be forced to
5168         * complicate the code by keeping track of which memcgs were the ones
5169         * that actually enabled limits, and which ones got it from its
5170         * parents.
5171         *
5172         * It is a lot simpler just to do static_key_slow_inc() on every child
5173         * that is accounted.
5174         */
5175        if (!memcg_kmem_is_active(memcg))
5176                goto out;
5177
5178        /*
5179         * destroy(), called if we fail, will issue static_key_slow_inc() and
5180         * mem_cgroup_put() if kmem is enabled. We have to either call them
5181         * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5182         * this more consistent, since it always leads to the same destroy path
5183         */
5184        mem_cgroup_get(memcg);
5185        static_key_slow_inc(&memcg_kmem_enabled_key);
5186
5187        mutex_lock(&set_limit_mutex);
5188        ret = memcg_update_cache_sizes(memcg);
5189        mutex_unlock(&set_limit_mutex);
5190out:
5191        return ret;
5192}
5193#endif /* CONFIG_MEMCG_KMEM */
5194
5195/*
5196 * The user of this function is...
5197 * RES_LIMIT.
5198 */
5199static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5200                            const char *buffer)
5201{
5202        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5203        enum res_type type;
5204        int name;
5205        unsigned long long val;
5206        int ret;
5207
5208        type = MEMFILE_TYPE(cft->private);
5209        name = MEMFILE_ATTR(cft->private);
5210
5211        switch (name) {
5212        case RES_LIMIT:
5213                if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5214                        ret = -EINVAL;
5215                        break;
5216                }
5217                /* This function does all necessary parse...reuse it */
5218                ret = res_counter_memparse_write_strategy(buffer, &val);
5219                if (ret)
5220                        break;
5221                if (type == _MEM)
5222                        ret = mem_cgroup_resize_limit(memcg, val);
5223                else if (type == _MEMSWAP)
5224                        ret = mem_cgroup_resize_memsw_limit(memcg, val);
5225                else if (type == _KMEM)
5226                        ret = memcg_update_kmem_limit(cont, val);
5227                else
5228                        return -EINVAL;
5229                break;
5230        case RES_SOFT_LIMIT:
5231                ret = res_counter_memparse_write_strategy(buffer, &val);
5232                if (ret)
5233                        break;
5234                /*
5235                 * For memsw, soft limits are hard to implement in terms
5236                 * of semantics, for now, we support soft limits for
5237                 * control without swap
5238                 */
5239                if (type == _MEM)
5240                        ret = res_counter_set_soft_limit(&memcg->res, val);
5241                else
5242                        ret = -EINVAL;
5243                break;
5244        default:
5245                ret = -EINVAL; /* should be BUG() ? */
5246                break;
5247        }
5248        return ret;
5249}
5250
5251static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5252                unsigned long long *mem_limit, unsigned long long *memsw_limit)
5253{
5254        struct cgroup *cgroup;
5255        unsigned long long min_limit, min_memsw_limit, tmp;
5256
5257        min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5258        min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5259        cgroup = memcg->css.cgroup;
5260        if (!memcg->use_hierarchy)
5261                goto out;
5262
5263        while (cgroup->parent) {
5264                cgroup = cgroup->parent;
5265                memcg = mem_cgroup_from_cont(cgroup);
5266                if (!memcg->use_hierarchy)
5267                        break;
5268                tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5269                min_limit = min(min_limit, tmp);
5270                tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5271                min_memsw_limit = min(min_memsw_limit, tmp);
5272        }
5273out:
5274        *mem_limit = min_limit;
5275        *memsw_limit = min_memsw_limit;
5276}
5277
5278static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5279{
5280        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5281        int name;
5282        enum res_type type;
5283
5284        type = MEMFILE_TYPE(event);
5285        name = MEMFILE_ATTR(event);
5286
5287        switch (name) {
5288        case RES_MAX_USAGE:
5289                if (type == _MEM)
5290                        res_counter_reset_max(&memcg->res);
5291                else if (type == _MEMSWAP)
5292                        res_counter_reset_max(&memcg->memsw);
5293                else if (type == _KMEM)
5294                        res_counter_reset_max(&memcg->kmem);
5295                else
5296                        return -EINVAL;
5297                break;
5298        case RES_FAILCNT:
5299                if (type == _MEM)
5300                        res_counter_reset_failcnt(&memcg->res);
5301                else if (type == _MEMSWAP)
5302                        res_counter_reset_failcnt(&memcg->memsw);
5303                else if (type == _KMEM)
5304                        res_counter_reset_failcnt(&memcg->kmem);
5305                else
5306                        return -EINVAL;
5307                break;
5308        }
5309
5310        return 0;
5311}
5312
5313static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5314                                        struct cftype *cft)
5315{
5316        return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5317}
5318
5319#ifdef CONFIG_MMU
5320static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5321                                        struct cftype *cft, u64 val)
5322{
5323        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5324
5325        if (val >= (1 << NR_MOVE_TYPE))
5326                return -EINVAL;
5327
5328        /*
5329         * No kind of locking is needed in here, because ->can_attach() will
5330         * check this value once in the beginning of the process, and then carry
5331         * on with stale data. This means that changes to this value will only
5332         * affect task migrations starting after the change.
5333         */
5334        memcg->move_charge_at_immigrate = val;
5335        return 0;
5336}
5337#else
5338static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5339                                        struct cftype *cft, u64 val)
5340{
5341        return -ENOSYS;
5342}
5343#endif
5344
5345#ifdef CONFIG_NUMA
5346static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5347                                      struct seq_file *m)
5348{
5349        int nid;
5350        unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5351        unsigned long node_nr;
5352        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5353
5354        total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5355        seq_printf(m, "total=%lu", total_nr);
5356        for_each_node_state(nid, N_MEMORY) {
5357                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5358                seq_printf(m, " N%d=%lu", nid, node_nr);
5359        }
5360        seq_putc(m, '\n');
5361
5362        file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5363        seq_printf(m, "file=%lu", file_nr);
5364        for_each_node_state(nid, N_MEMORY) {
5365                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5366                                LRU_ALL_FILE);
5367                seq_printf(m, " N%d=%lu", nid, node_nr);
5368        }
5369        seq_putc(m, '\n');
5370
5371        anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5372        seq_printf(m, "anon=%lu", anon_nr);
5373        for_each_node_state(nid, N_MEMORY) {
5374                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5375                                LRU_ALL_ANON);
5376                seq_printf(m, " N%d=%lu", nid, node_nr);
5377        }
5378        seq_putc(m, '\n');
5379
5380        unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5381        seq_printf(m, "unevictable=%lu", unevictable_nr);
5382        for_each_node_state(nid, N_MEMORY) {
5383                node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5384                                BIT(LRU_UNEVICTABLE));
5385                seq_printf(m, " N%d=%lu", nid, node_nr);
5386        }
5387        seq_putc(m, '\n');
5388        return 0;
5389}
5390#endif /* CONFIG_NUMA */
5391
5392static inline void mem_cgroup_lru_names_not_uptodate(void)
5393{
5394        BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5395}
5396
5397static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5398                                 struct seq_file *m)
5399{
5400        struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5401        struct mem_cgroup *mi;
5402        unsigned int i;
5403
5404        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5405                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5406                        continue;
5407                seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5408                           mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5409        }
5410
5411        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5412                seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5413                           mem_cgroup_read_events(memcg, i));
5414
5415        for (i = 0; i < NR_LRU_LISTS; i++)
5416                seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5417                           mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5418
5419        /* Hierarchical information */
5420        {
5421                unsigned long long limit, memsw_limit;
5422                memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5423                seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5424                if (do_swap_account)
5425                        seq_printf(m, "hierarchical_memsw_limit %llu\n",
5426                                   memsw_limit);
5427        }
5428
5429        for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5430                long long val = 0;
5431
5432                if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5433                        continue;
5434                for_each_mem_cgroup_tree(mi, memcg)
5435                        val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5436                seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5437        }
5438
5439        for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5440                unsigned long long val = 0;
5441
5442                for_each_mem_cgroup_tree(mi, memcg)
5443                        val += mem_cgroup_read_events(mi, i);
5444                seq_printf(m, "total_%s %llu\n",
5445                           mem_cgroup_events_names[i], val);
5446        }
5447
5448        for (i = 0; i < NR_LRU_LISTS; i++) {
5449                unsigned long long val = 0;
5450
5451                for_each_mem_cgroup_tree(mi, memcg)
5452                        val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5453                seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5454        }
5455
5456#ifdef CONFIG_DEBUG_VM
5457        {
5458                int nid, zid;
5459                struct mem_cgroup_per_zone *mz;
5460                struct zone_reclaim_stat *rstat;
5461                unsigned long recent_rotated[2] = {0, 0};
5462                unsigned long recent_scanned[2] = {0, 0};
5463
5464                for_each_online_node(nid)
5465                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5466                                mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5467                                rstat = &mz->lruvec.reclaim_stat;
5468
5469                                recent_rotated[0] += rstat->recent_rotated[0];
5470                                recent_rotated[1] += rstat->recent_rotated[1];
5471                                recent_scanned[0] += rstat->recent_scanned[0];
5472                                recent_scanned[1] += rstat->recent_scanned[1];
5473                        }
5474                seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5475                seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5476                seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5477                seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5478        }
5479#endif
5480
5481        return 0;
5482}
5483
5484static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5485{
5486        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5487
5488        return mem_cgroup_swappiness(memcg);
5489}
5490
5491static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5492                                       u64 val)
5493{
5494        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5495        struct mem_cgroup *parent;
5496
5497        if (val > 100)
5498                return -EINVAL;
5499
5500        if (cgrp->parent == NULL)
5501                return -EINVAL;
5502
5503        parent = mem_cgroup_from_cont(cgrp->parent);
5504
5505        mutex_lock(&memcg_create_mutex);
5506
5507        /* If under hierarchy, only empty-root can set this value */
5508        if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5509                mutex_unlock(&memcg_create_mutex);
5510                return -EINVAL;
5511        }
5512
5513        memcg->swappiness = val;
5514
5515        mutex_unlock(&memcg_create_mutex);
5516
5517        return 0;
5518}
5519
5520static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5521{
5522        struct mem_cgroup_threshold_ary *t;
5523        u64 usage;
5524        int i;
5525
5526        rcu_read_lock();
5527        if (!swap)
5528                t = rcu_dereference(memcg->thresholds.primary);
5529        else
5530                t = rcu_dereference(memcg->memsw_thresholds.primary);
5531
5532        if (!t)
5533                goto unlock;
5534
5535        usage = mem_cgroup_usage(memcg, swap);
5536
5537        /*
5538         * current_threshold points to threshold just below or equal to usage.
5539         * If it's not true, a threshold was crossed after last
5540         * call of __mem_cgroup_threshold().
5541         */
5542        i = t->current_threshold;
5543
5544        /*
5545         * Iterate backward over array of thresholds starting from
5546         * current_threshold and check if a threshold is crossed.
5547         * If none of thresholds below usage is crossed, we read
5548         * only one element of the array here.
5549         */
5550        for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5551                eventfd_signal(t->entries[i].eventfd, 1);
5552
5553        /* i = current_threshold + 1 */
5554        i++;
5555
5556        /*
5557         * Iterate forward over array of thresholds starting from
5558         * current_threshold+1 and check if a threshold is crossed.
5559         * If none of thresholds above usage is crossed, we read
5560         * only one element of the array here.
5561         */
5562        for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5563                eventfd_signal(t->entries[i].eventfd, 1);
5564
5565        /* Update current_threshold */
5566        t->current_threshold = i - 1;
5567unlock:
5568        rcu_read_unlock();
5569}
5570
5571static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5572{
5573        while (memcg) {
5574                __mem_cgroup_threshold(memcg, false);
5575                if (do_swap_account)
5576                        __mem_cgroup_threshold(memcg, true);
5577
5578                memcg = parent_mem_cgroup(memcg);
5579        }
5580}
5581
5582static int compare_thresholds(const void *a, const void *b)
5583{
5584        const struct mem_cgroup_threshold *_a = a;
5585        const struct mem_cgroup_threshold *_b = b;
5586
5587        return _a->threshold - _b->threshold;
5588}
5589
5590static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5591{
5592        struct mem_cgroup_eventfd_list *ev;
5593
5594        list_for_each_entry(ev, &memcg->oom_notify, list)
5595                eventfd_signal(ev->eventfd, 1);
5596        return 0;
5597}
5598
5599static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5600{
5601        struct mem_cgroup *iter;
5602
5603        for_each_mem_cgroup_tree(iter, memcg)
5604                mem_cgroup_oom_notify_cb(iter);
5605}
5606
5607static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5608        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5609{
5610        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5611        struct mem_cgroup_thresholds *thresholds;
5612        struct mem_cgroup_threshold_ary *new;
5613        enum res_type type = MEMFILE_TYPE(cft->private);
5614        u64 threshold, usage;
5615        int i, size, ret;
5616
5617        ret = res_counter_memparse_write_strategy(args, &threshold);
5618        if (ret)
5619                return ret;
5620
5621        mutex_lock(&memcg->thresholds_lock);
5622
5623        if (type == _MEM)
5624                thresholds = &memcg->thresholds;
5625        else if (type == _MEMSWAP)
5626                thresholds = &memcg->memsw_thresholds;
5627        else
5628                BUG();
5629
5630        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5631
5632        /* Check if a threshold crossed before adding a new one */
5633        if (thresholds->primary)
5634                __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5635
5636        size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5637
5638        /* Allocate memory for new array of thresholds */
5639        new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5640                        GFP_KERNEL);
5641        if (!new) {
5642                ret = -ENOMEM;
5643                goto unlock;
5644        }
5645        new->size = size;
5646
5647        /* Copy thresholds (if any) to new array */
5648        if (thresholds->primary) {
5649                memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5650                                sizeof(struct mem_cgroup_threshold));
5651        }
5652
5653        /* Add new threshold */
5654        new->entries[size - 1].eventfd = eventfd;
5655        new->entries[size - 1].threshold = threshold;
5656
5657        /* Sort thresholds. Registering of new threshold isn't time-critical */
5658        sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5659                        compare_thresholds, NULL);
5660
5661        /* Find current threshold */
5662        new->current_threshold = -1;
5663        for (i = 0; i < size; i++) {
5664                if (new->entries[i].threshold <= usage) {
5665                        /*
5666                         * new->current_threshold will not be used until
5667                         * rcu_assign_pointer(), so it's safe to increment
5668                         * it here.
5669                         */
5670                        ++new->current_threshold;
5671                } else
5672                        break;
5673        }
5674
5675        /* Free old spare buffer and save old primary buffer as spare */
5676        kfree(thresholds->spare);
5677        thresholds->spare = thresholds->primary;
5678
5679        rcu_assign_pointer(thresholds->primary, new);
5680
5681        /* To be sure that nobody uses thresholds */
5682        synchronize_rcu();
5683
5684unlock:
5685        mutex_unlock(&memcg->thresholds_lock);
5686
5687        return ret;
5688}
5689
5690static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5691        struct cftype *cft, struct eventfd_ctx *eventfd)
5692{
5693        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5694        struct mem_cgroup_thresholds *thresholds;
5695        struct mem_cgroup_threshold_ary *new;
5696        enum res_type type = MEMFILE_TYPE(cft->private);
5697        u64 usage;
5698        int i, j, size;
5699
5700        mutex_lock(&memcg->thresholds_lock);
5701        if (type == _MEM)
5702                thresholds = &memcg->thresholds;
5703        else if (type == _MEMSWAP)
5704                thresholds = &memcg->memsw_thresholds;
5705        else
5706                BUG();
5707
5708        if (!thresholds->primary)
5709                goto unlock;
5710
5711        usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5712
5713        /* Check if a threshold crossed before removing */
5714        __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5715
5716        /* Calculate new number of threshold */
5717        size = 0;
5718        for (i = 0; i < thresholds->primary->size; i++) {
5719                if (thresholds->primary->entries[i].eventfd != eventfd)
5720                        size++;
5721        }
5722
5723        new = thresholds->spare;
5724
5725        /* Set thresholds array to NULL if we don't have thresholds */
5726        if (!size) {
5727                kfree(new);
5728                new = NULL;
5729                goto swap_buffers;
5730        }
5731
5732        new->size = size;
5733
5734        /* Copy thresholds and find current threshold */
5735        new->current_threshold = -1;
5736        for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5737                if (thresholds->primary->entries[i].eventfd == eventfd)
5738                        continue;
5739
5740                new->entries[j] = thresholds->primary->entries[i];
5741                if (new->entries[j].threshold <= usage) {
5742                        /*
5743                         * new->current_threshold will not be used
5744                         * until rcu_assign_pointer(), so it's safe to increment
5745                         * it here.
5746                         */
5747                        ++new->current_threshold;
5748                }
5749                j++;
5750        }
5751
5752swap_buffers:
5753        /* Swap primary and spare array */
5754        thresholds->spare = thresholds->primary;
5755        /* If all events are unregistered, free the spare array */
5756        if (!new) {
5757                kfree(thresholds->spare);
5758                thresholds->spare = NULL;
5759        }
5760
5761        rcu_assign_pointer(thresholds->primary, new);
5762
5763        /* To be sure that nobody uses thresholds */
5764        synchronize_rcu();
5765unlock:
5766        mutex_unlock(&memcg->thresholds_lock);
5767}
5768
5769static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5770        struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5771{
5772        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5773        struct mem_cgroup_eventfd_list *event;
5774        enum res_type type = MEMFILE_TYPE(cft->private);
5775
5776        BUG_ON(type != _OOM_TYPE);
5777        event = kmalloc(sizeof(*event), GFP_KERNEL);
5778        if (!event)
5779                return -ENOMEM;
5780
5781        spin_lock(&memcg_oom_lock);
5782
5783        event->eventfd = eventfd;
5784        list_add(&event->list, &memcg->oom_notify);
5785
5786        /* already in OOM ? */
5787        if (atomic_read(&memcg->under_oom))
5788                eventfd_signal(eventfd, 1);
5789        spin_unlock(&memcg_oom_lock);
5790
5791        return 0;
5792}
5793
5794static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5795        struct cftype *cft, struct eventfd_ctx *eventfd)
5796{
5797        struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5798        struct mem_cgroup_eventfd_list *ev, *tmp;
5799        enum res_type type = MEMFILE_TYPE(cft->private);
5800
5801        BUG_ON(type != _OOM_TYPE);
5802
5803        spin_lock(&memcg_oom_lock);
5804
5805        list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5806                if (ev->eventfd ==