mm/memcontrol.c

   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  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19
  20 #include <linux/res_counter.h>
  21 #include <linux/memcontrol.h>
  22 #include <linux/cgroup.h>
  23 #include <linux/mm.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/smp.h>
  26 #include <linux/page-flags.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/bit_spinlock.h>
  29 #include <linux/rcupdate.h>
  30 #include <linux/mutex.h>
  31 #include <linux/slab.h>
  32 #include <linux/swap.h>
  33 #include <linux/spinlock.h>
  34 #include <linux/fs.h>
  35 #include <linux/seq_file.h>
  36 #include <linux/vmalloc.h>
  37 #include <linux/mm_inline.h>
  38 #include <linux/page_cgroup.h>
  39 #include "internal.h"
  40
  41 #include <asm/uaccess.h>
  42
  43 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
  44 #define MEM_CGROUP_RECLAIM_RETRIES      5
  45
  46 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
  47 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 0 */
  48 int do_swap_account __read_mostly;
  49 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
  50 #else
  51 #define do_swap_account         (0)
  52 #endif
  53
  54 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
  55
  56 /*
  57  * Statistics for memory cgroup.
  58  */
  59 enum mem_cgroup_stat_index {
  60         /*
  61          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  62          */
  63         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
  64         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
  65         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
  66         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
  67
  68         MEM_CGROUP_STAT_NSTATS,
  69 };
  70
  71 struct mem_cgroup_stat_cpu {
  72         s64 count[MEM_CGROUP_STAT_NSTATS];
  73 } ____cacheline_aligned_in_smp;
  74
  75 struct mem_cgroup_stat {
  76         struct mem_cgroup_stat_cpu cpustat[0];
  77 };
  78
  79 /*
  80  * For accounting under irq disable, no need for increment preempt count.
  81  */
  82 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
  83                 enum mem_cgroup_stat_index idx, int val)
  84 {
  85         stat->count[idx] += val;
  86 }
  87
  88 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
  89                 enum mem_cgroup_stat_index idx)
  90 {
  91         int cpu;
  92         s64 ret = 0;
  93         for_each_possible_cpu(cpu)
  94                 ret += stat->cpustat[cpu].count[idx];
  95         return ret;
  96 }
  97
  98 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
  99 {
 100         s64 ret;
 101
 102         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
 103         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
 104         return ret;
 105 }
 106
 107 /*
 108  * per-zone information in memory controller.
 109  */
 110 struct mem_cgroup_per_zone {
 111         /*
 112          * spin_lock to protect the per cgroup LRU
 113          */
 114         struct list_head        lists[NR_LRU_LISTS];
 115         unsigned long           count[NR_LRU_LISTS];
 116
 117         struct zone_reclaim_stat reclaim_stat;
 118 };
 119 /* Macro for accessing counter */
 120 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
 121
 122 struct mem_cgroup_per_node {
 123         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 124 };
 125
 126 struct mem_cgroup_lru_info {
 127         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 128 };
 129
 130 /*
 131  * The memory controller data structure. The memory controller controls both
 132  * page cache and RSS per cgroup. We would eventually like to provide
 133  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 134  * to help the administrator determine what knobs to tune.
 135  *
 136  * TODO: Add a water mark for the memory controller. Reclaim will begin when
 137  * we hit the water mark. May be even add a low water mark, such that
 138  * no reclaim occurs from a cgroup at it's low water mark, this is
 139  * a feature that will be implemented much later in the future.
 140  */
 141 struct mem_cgroup {
 142         struct cgroup_subsys_state css;
 143         /*
 144          * the counter to account for memory usage
 145          */
 146         struct res_counter res;
 147         /*
 148          * the counter to account for mem+swap usage.
 149          */
 150         struct res_counter memsw;
 151         /*
 152          * Per cgroup active and inactive list, similar to the
 153          * per zone LRU lists.
 154          */
 155         struct mem_cgroup_lru_info info;
 156
 157         /*
 158           protect against reclaim related member.
 159         */
 160         spinlock_t reclaim_param_lock;
 161
 162         int     prev_priority;  /* for recording reclaim priority */
 163
 164         /*
 165          * While reclaiming in a hiearchy, we cache the last child we
 166          * reclaimed from.
 167          */
 168         int last_scanned_child;
 169         /*
 170          * Should the accounting and control be hierarchical, per subtree?
 171          */
 172         bool use_hierarchy;
 173         unsigned long   last_oom_jiffies;
 174         atomic_t        refcnt;
 175
 176         unsigned int    swappiness;
 177
 178         /*
 179          * statistics. This must be placed at the end of memcg.
 180          */
 181         struct mem_cgroup_stat stat;
 182 };
 183
 184 enum charge_type {
 185         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 186         MEM_CGROUP_CHARGE_TYPE_MAPPED,
 187         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
 188         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
 189         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
 190         NR_CHARGE_TYPE,
 191 };
 192
 193 /* only for here (for easy reading.) */
 194 #define PCGF_CACHE      (1UL << PCG_CACHE)
 195 #define PCGF_USED       (1UL << PCG_USED)
 196 #define PCGF_LOCK       (1UL << PCG_LOCK)
 197 static const unsigned long
 198 pcg_default_flags[NR_CHARGE_TYPE] = {
 199         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
 200         PCGF_USED | PCGF_LOCK, /* Anon */
 201         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
 202         0, /* FORCE */
 203 };
 204
 205 /* for encoding cft->private value on file */
 206 #define _MEM                    (0)
 207 #define _MEMSWAP                (1)
 208 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
 209 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
 210 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
 211
 212 static void mem_cgroup_get(struct mem_cgroup *mem);
 213 static void mem_cgroup_put(struct mem_cgroup *mem);
 214 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
 215
 216 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
 217                                          struct page_cgroup *pc,
 218                                          bool charge)
 219 {
 220         int val = (charge)? 1 : -1;
 221         struct mem_cgroup_stat *stat = &mem->stat;
 222         struct mem_cgroup_stat_cpu *cpustat;
 223         int cpu = get_cpu();
 224
 225         cpustat = &stat->cpustat[cpu];
 226         if (PageCgroupCache(pc))
 227                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
 228         else
 229                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
 230
 231         if (charge)
 232                 __mem_cgroup_stat_add_safe(cpustat,
 233                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
 234         else
 235                 __mem_cgroup_stat_add_safe(cpustat,
 236                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
 237         put_cpu();
 238 }
 239
 240 static struct mem_cgroup_per_zone *
 241 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
 242 {
 243         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
 244 }
 245
 246 static struct mem_cgroup_per_zone *
 247 page_cgroup_zoneinfo(struct page_cgroup *pc)
 248 {
 249         struct mem_cgroup *mem = pc->mem_cgroup;
 250         int nid = page_cgroup_nid(pc);
 251         int zid = page_cgroup_zid(pc);
 252
 253         if (!mem)
 254                 return NULL;
 255
 256         return mem_cgroup_zoneinfo(mem, nid, zid);
 257 }
 258
 259 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
 260                                         enum lru_list idx)
 261 {
 262         int nid, zid;
 263         struct mem_cgroup_per_zone *mz;
 264         u64 total = 0;
 265
 266         for_each_online_node(nid)
 267                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 268                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
 269                         total += MEM_CGROUP_ZSTAT(mz, idx);
 270                 }
 271         return total;
 272 }
 273
 274 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 275 {
 276         return container_of(cgroup_subsys_state(cont,
 277                                 mem_cgroup_subsys_id), struct mem_cgroup,
 278                                 css);
 279 }
 280
 281 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 282 {
 283         /*
 284          * mm_update_next_owner() may clear mm->owner to NULL
 285          * if it races with swapoff, page migration, etc.
 286          * So this can be called with p == NULL.
 287          */
 288         if (unlikely(!p))
 289                 return NULL;
 290
 291         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 292                                 struct mem_cgroup, css);
 293 }
 294
 295 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
 296 {
 297         struct mem_cgroup *mem = NULL;
 298         /*
 299          * Because we have no locks, mm->owner's may be being moved to other
 300          * cgroup. We use css_tryget() here even if this looks
 301          * pessimistic (rather than adding locks here).
 302          */
 303         rcu_read_lock();
 304         do {
 305                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
 306                 if (unlikely(!mem))
 307                         break;
 308         } while (!css_tryget(&mem->css));
 309         rcu_read_unlock();
 310         return mem;
 311 }
 312
 313 static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem)
 314 {
 315         if (!mem)
 316                 return true;
 317         return css_is_removed(&mem->css);
 318 }
 319
 320 /*
 321  * Following LRU functions are allowed to be used without PCG_LOCK.
 322  * Operations are called by routine of global LRU independently from memcg.
 323  * What we have to take care of here is validness of pc->mem_cgroup.
 324  *
 325  * Changes to pc->mem_cgroup happens when
 326  * 1. charge
 327  * 2. moving account
 328  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
 329  * It is added to LRU before charge.
 330  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
 331  * When moving account, the page is not on LRU. It's isolated.
 332  */
 333
 334 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
 335 {
 336         struct page_cgroup *pc;
 337         struct mem_cgroup *mem;
 338         struct mem_cgroup_per_zone *mz;
 339
 340         if (mem_cgroup_disabled())
 341                 return;
 342         pc = lookup_page_cgroup(page);
 343         /* can happen while we handle swapcache. */
 344         if (list_empty(&pc->lru) || !pc->mem_cgroup)
 345                 return;
 346         /*
 347          * We don't check PCG_USED bit. It's cleared when the "page" is finally
 348          * removed from global LRU.
 349          */
 350         mz = page_cgroup_zoneinfo(pc);
 351         mem = pc->mem_cgroup;
 352         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
 353         list_del_init(&pc->lru);
 354         return;
 355 }
 356
 357 void mem_cgroup_del_lru(struct page *page)
 358 {
 359         mem_cgroup_del_lru_list(page, page_lru(page));
 360 }
 361
 362 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
 363 {
 364         struct mem_cgroup_per_zone *mz;
 365         struct page_cgroup *pc;
 366
 367         if (mem_cgroup_disabled())
 368                 return;
 369
 370         pc = lookup_page_cgroup(page);
 371         /*
 372          * Used bit is set without atomic ops but after smp_wmb().
 373          * For making pc->mem_cgroup visible, insert smp_rmb() here.
 374          */
 375         smp_rmb();
 376         /* unused page is not rotated. */
 377         if (!PageCgroupUsed(pc))
 378                 return;
 379         mz = page_cgroup_zoneinfo(pc);
 380         list_move(&pc->lru, &mz->lists[lru]);
 381 }
 382
 383 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
 384 {
 385         struct page_cgroup *pc;
 386         struct mem_cgroup_per_zone *mz;
 387
 388         if (mem_cgroup_disabled())
 389                 return;
 390         pc = lookup_page_cgroup(page);
 391         /*
 392          * Used bit is set without atomic ops but after smp_wmb().
 393          * For making pc->mem_cgroup visible, insert smp_rmb() here.
 394          */
 395         smp_rmb();
 396         if (!PageCgroupUsed(pc))
 397                 return;
 398
 399         mz = page_cgroup_zoneinfo(pc);
 400         MEM_CGROUP_ZSTAT(mz, lru) += 1;
 401         list_add(&pc->lru, &mz->lists[lru]);
 402 }
 403
 404 /*
 405  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
 406  * lru because the page may.be reused after it's fully uncharged (because of
 407  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
 408  * it again. This function is only used to charge SwapCache. It's done under
 409  * lock_page and expected that zone->lru_lock is never held.
 410  */
 411 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
 412 {
 413         unsigned long flags;
 414         struct zone *zone = page_zone(page);
 415         struct page_cgroup *pc = lookup_page_cgroup(page);
 416
 417         spin_lock_irqsave(&zone->lru_lock, flags);
 418         /*
 419          * Forget old LRU when this page_cgroup is *not* used. This Used bit
 420          * is guarded by lock_page() because the page is SwapCache.
 421          */
 422         if (!PageCgroupUsed(pc))
 423                 mem_cgroup_del_lru_list(page, page_lru(page));
 424         spin_unlock_irqrestore(&zone->lru_lock, flags);
 425 }
 426
 427 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
 428 {
 429         unsigned long flags;
 430         struct zone *zone = page_zone(page);
 431         struct page_cgroup *pc = lookup_page_cgroup(page);
 432
 433         spin_lock_irqsave(&zone->lru_lock, flags);
 434         /* link when the page is linked to LRU but page_cgroup isn't */
 435         if (PageLRU(page) && list_empty(&pc->lru))
 436                 mem_cgroup_add_lru_list(page, page_lru(page));
 437         spin_unlock_irqrestore(&zone->lru_lock, flags);
 438 }
 439
 440
 441 void mem_cgroup_move_lists(struct page *page,
 442                            enum lru_list from, enum lru_list to)
 443 {
 444         if (mem_cgroup_disabled())
 445                 return;
 446         mem_cgroup_del_lru_list(page, from);
 447         mem_cgroup_add_lru_list(page, to);
 448 }
 449
 450 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 451 {
 452         int ret;
 453
 454         task_lock(task);
 455         ret = task->mm && mm_match_cgroup(task->mm, mem);
 456         task_unlock(task);
 457         return ret;
 458 }
 459
 460 /*
 461  * Calculate mapped_ratio under memory controller. This will be used in
 462  * vmscan.c for deteremining we have to reclaim mapped pages.
 463  */
 464 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
 465 {
 466         long total, rss;
 467
 468         /*
 469          * usage is recorded in bytes. But, here, we assume the number of
 470          * physical pages can be represented by "long" on any arch.
 471          */
 472         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
 473         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
 474         return (int)((rss * 100L) / total);
 475 }
 476
 477 /*
 478  * prev_priority control...this will be used in memory reclaim path.
 479  */
 480 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
 481 {
 482         int prev_priority;
 483
 484         spin_lock(&mem->reclaim_param_lock);
 485         prev_priority = mem->prev_priority;
 486         spin_unlock(&mem->reclaim_param_lock);
 487
 488         return prev_priority;
 489 }
 490
 491 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
 492 {
 493         spin_lock(&mem->reclaim_param_lock);
 494         if (priority < mem->prev_priority)
 495                 mem->prev_priority = priority;
 496         spin_unlock(&mem->reclaim_param_lock);
 497 }
 498
 499 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
 500 {
 501         spin_lock(&mem->reclaim_param_lock);
 502         mem->prev_priority = priority;
 503         spin_unlock(&mem->reclaim_param_lock);
 504 }
 505
 506 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
 507 {
 508         unsigned long active;
 509         unsigned long inactive;
 510         unsigned long gb;
 511         unsigned long inactive_ratio;
 512
 513         inactive = mem_cgroup_get_all_zonestat(memcg, LRU_INACTIVE_ANON);
 514         active = mem_cgroup_get_all_zonestat(memcg, LRU_ACTIVE_ANON);
 515
 516         gb = (inactive + active) >> (30 - PAGE_SHIFT);
 517         if (gb)
 518                 inactive_ratio = int_sqrt(10 * gb);
 519         else
 520                 inactive_ratio = 1;
 521
 522         if (present_pages) {
 523                 present_pages[0] = inactive;
 524                 present_pages[1] = active;
 525         }
 526
 527         return inactive_ratio;
 528 }
 529
 530 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
 531 {
 532         unsigned long active;
 533         unsigned long inactive;
 534         unsigned long present_pages[2];
 535         unsigned long inactive_ratio;
 536
 537         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
 538
 539         inactive = present_pages[0];
 540         active = present_pages[1];
 541
 542         if (inactive * inactive_ratio < active)
 543                 return 1;
 544
 545         return 0;
 546 }
 547
 548 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
 549                                        struct zone *zone,
 550                                        enum lru_list lru)
 551 {
 552         int nid = zone->zone_pgdat->node_id;
 553         int zid = zone_idx(zone);
 554         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 555
 556         return MEM_CGROUP_ZSTAT(mz, lru);
 557 }
 558
 559 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
 560                                                       struct zone *zone)
 561 {
 562         int nid = zone->zone_pgdat->node_id;
 563         int zid = zone_idx(zone);
 564         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
 565
 566         return &mz->reclaim_stat;
 567 }
 568
 569 struct zone_reclaim_stat *
 570 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
 571 {
 572         struct page_cgroup *pc;
 573         struct mem_cgroup_per_zone *mz;
 574
 575         if (mem_cgroup_disabled())
 576                 return NULL;
 577
 578         pc = lookup_page_cgroup(page);
 579         /*
 580          * Used bit is set without atomic ops but after smp_wmb().
 581          * For making pc->mem_cgroup visible, insert smp_rmb() here.
 582          */
 583         smp_rmb();
 584         if (!PageCgroupUsed(pc))
 585                 return NULL;
 586
 587         mz = page_cgroup_zoneinfo(pc);
 588         if (!mz)
 589                 return NULL;
 590
 591         return &mz->reclaim_stat;
 592 }
 593
 594 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 595                                         struct list_head *dst,
 596                                         unsigned long *scanned, int order,
 597                                         int mode, struct zone *z,
 598                                         struct mem_cgroup *mem_cont,
 599                                         int active, int file)
 600 {
 601         unsigned long nr_taken = 0;
 602         struct page *page;
 603         unsigned long scan;
 604         LIST_HEAD(pc_list);
 605         struct list_head *src;
 606         struct page_cgroup *pc, *tmp;
 607         int nid = z->zone_pgdat->node_id;
 608         int zid = zone_idx(z);
 609         struct mem_cgroup_per_zone *mz;
 610         int lru = LRU_FILE * !!file + !!active;
 611
 612         BUG_ON(!mem_cont);
 613         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
 614         src = &mz->lists[lru];
 615
 616         scan = 0;
 617         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 618                 if (scan >= nr_to_scan)
 619                         break;
 620
 621                 page = pc->page;
 622                 if (unlikely(!PageCgroupUsed(pc)))
 623                         continue;
 624                 if (unlikely(!PageLRU(page)))
 625                         continue;
 626
 627                 scan++;
 628                 if (__isolate_lru_page(page, mode, file) == 0) {
 629                         list_move(&page->lru, dst);
 630                         nr_taken++;
 631                 }
 632         }
 633
 634         *scanned = scan;
 635         return nr_taken;
 636 }
 637
 638 #define mem_cgroup_from_res_counter(counter, member)    \
 639         container_of(counter, struct mem_cgroup, member)
 640
 641 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
 642 {
 643         if (do_swap_account) {
 644                 if (res_counter_check_under_limit(&mem->res) &&
 645                         res_counter_check_under_limit(&mem->memsw))
 646                         return true;
 647         } else
 648                 if (res_counter_check_under_limit(&mem->res))
 649                         return true;
 650         return false;
 651 }
 652
 653 static unsigned int get_swappiness(struct mem_cgroup *memcg)
 654 {
 655         struct cgroup *cgrp = memcg->css.cgroup;
 656         unsigned int swappiness;
 657
 658         /* root ? */
 659         if (cgrp->parent == NULL)
 660                 return vm_swappiness;
 661
 662         spin_lock(&memcg->reclaim_param_lock);
 663         swappiness = memcg->swappiness;
 664         spin_unlock(&memcg->reclaim_param_lock);
 665
 666         return swappiness;
 667 }
 668
 669 /*
 670  * Visit the first child (need not be the first child as per the ordering
 671  * of the cgroup list, since we track last_scanned_child) of @mem and use
 672  * that to reclaim free pages from.
 673  */
 674 static struct mem_cgroup *
 675 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
 676 {
 677         struct mem_cgroup *ret = NULL;
 678         struct cgroup_subsys_state *css;
 679         int nextid, found;
 680
 681         if (!root_mem->use_hierarchy) {
 682                 css_get(&root_mem->css);
 683                 ret = root_mem;
 684         }
 685
 686         while (!ret) {
 687                 rcu_read_lock();
 688                 nextid = root_mem->last_scanned_child + 1;
 689                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
 690                                    &found);
 691                 if (css && css_tryget(css))
 692                         ret = container_of(css, struct mem_cgroup, css);
 693
 694                 rcu_read_unlock();
 695                 /* Updates scanning parameter */
 696                 spin_lock(&root_mem->reclaim_param_lock);
 697                 if (!css) {
 698                         /* this means start scan from ID:1 */
 699                         root_mem->last_scanned_child = 0;
 700                 } else
 701                         root_mem->last_scanned_child = found;
 702                 spin_unlock(&root_mem->reclaim_param_lock);
 703         }
 704
 705         return ret;
 706 }
 707
 708 /*
 709  * Scan the hierarchy if needed to reclaim memory. We remember the last child
 710  * we reclaimed from, so that we don't end up penalizing one child extensively
 711  * based on its position in the children list.
 712  *
 713  * root_mem is the original ancestor that we've been reclaim from.
 714  *
 715  * We give up and return to the caller when we visit root_mem twice.
 716  * (other groups can be removed while we're walking....)
 717  */
 718 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
 719                                                 gfp_t gfp_mask, bool noswap)
 720 {
 721         struct mem_cgroup *victim;
 722         int ret, total = 0;
 723         int loop = 0;
 724
 725         while (loop < 2) {
 726                 victim = mem_cgroup_select_victim(root_mem);
 727                 if (victim == root_mem)
 728                         loop++;
 729                 if (!mem_cgroup_local_usage(&victim->stat)) {
 730                         /* this cgroup's local usage == 0 */
 731                         css_put(&victim->css);
 732                         continue;
 733                 }
 734                 /* we use swappiness of local cgroup */
 735                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
 736                                                    get_swappiness(victim));
 737                 css_put(&victim->css);
 738                 total += ret;
 739                 if (mem_cgroup_check_under_limit(root_mem))
 740                         return 1 + total;
 741         }
 742         return total;
 743 }
 744
 745 bool mem_cgroup_oom_called(struct task_struct *task)
 746 {
 747         bool ret = false;
 748         struct mem_cgroup *mem;
 749         struct mm_struct *mm;
 750
 751         rcu_read_lock();
 752         mm = task->mm;
 753         if (!mm)
 754                 mm = &init_mm;
 755         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
 756         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
 757                 ret = true;
 758         rcu_read_unlock();
 759         return ret;
 760 }
 761 /*
 762  * Unlike exported interface, "oom" parameter is added. if oom==true,
 763  * oom-killer can be invoked.
 764  */
 765 static int __mem_cgroup_try_charge(struct mm_struct *mm,
 766                         gfp_t gfp_mask, struct mem_cgroup **memcg,
 767                         bool oom)
 768 {
 769         struct mem_cgroup *mem, *mem_over_limit;
 770         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 771         struct res_counter *fail_res;
 772
 773         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
 774                 /* Don't account this! */
 775                 *memcg = NULL;
 776                 return 0;
 777         }
 778
 779         /*
 780          * We always charge the cgroup the mm_struct belongs to.
 781          * The mm_struct's mem_cgroup changes on task migration if the
 782          * thread group leader migrates. It's possible that mm is not
 783          * set, if so charge the init_mm (happens for pagecache usage).
 784          */
 785         mem = *memcg;
 786         if (likely(!mem)) {
 787                 mem = try_get_mem_cgroup_from_mm(mm);
 788                 *memcg = mem;
 789         } else {
 790                 css_get(&mem->css);
 791         }
 792         if (unlikely(!mem))
 793                 return 0;
 794
 795         VM_BUG_ON(mem_cgroup_is_obsolete(mem));
 796
 797         while (1) {
 798                 int ret;
 799                 bool noswap = false;
 800
 801                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
 802                 if (likely(!ret)) {
 803                         if (!do_swap_account)
 804                                 break;
 805                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
 806                                                         &fail_res);
 807                         if (likely(!ret))
 808                                 break;
 809                         /* mem+swap counter fails */
 810                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 811                         noswap = true;
 812                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
 813                                                                         memsw);
 814                 } else
 815                         /* mem counter fails */
 816                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
 817                                                                         res);
 818
 819                 if (!(gfp_mask & __GFP_WAIT))
 820                         goto nomem;
 821
 822                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
 823                                                         noswap);
 824                 if (ret)
 825                         continue;
 826
 827                 /*
 828                  * try_to_free_mem_cgroup_pages() might not give us a full
 829                  * picture of reclaim. Some pages are reclaimed and might be
 830                  * moved to swap cache or just unmapped from the cgroup.
 831                  * Check the limit again to see if the reclaim reduced the
 832                  * current usage of the cgroup before giving up
 833                  *
 834                  */
 835                 if (mem_cgroup_check_under_limit(mem_over_limit))
 836                         continue;
 837
 838                 if (!nr_retries--) {
 839                         if (oom) {
 840                                 mutex_lock(&memcg_tasklist);
 841                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
 842                                 mutex_unlock(&memcg_tasklist);
 843                                 mem_over_limit->last_oom_jiffies = jiffies;
 844                         }
 845                         goto nomem;
 846                 }
 847         }
 848         return 0;
 849 nomem:
 850         css_put(&mem->css);
 851         return -ENOMEM;
 852 }
 853
 854 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
 855 {
 856         struct mem_cgroup *mem;
 857         swp_entry_t ent;
 858
 859         if (!PageSwapCache(page))
 860                 return NULL;
 861
 862         ent.val = page_private(page);
 863         mem = lookup_swap_cgroup(ent);
 864         if (!mem)
 865                 return NULL;
 866         if (!css_tryget(&mem->css))
 867                 return NULL;
 868         return mem;
 869 }
 870
 871 /*
 872  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
 873  * USED state. If already USED, uncharge and return.
 874  */
 875
 876 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
 877                                      struct page_cgroup *pc,
 878                                      enum charge_type ctype)
 879 {
 880         /* try_charge() can return NULL to *memcg, taking care of it. */
 881         if (!mem)
 882                 return;
 883
 884         lock_page_cgroup(pc);
 885         if (unlikely(PageCgroupUsed(pc))) {
 886                 unlock_page_cgroup(pc);
 887                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 888                 if (do_swap_account)
 889                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
 890                 css_put(&mem->css);
 891                 return;
 892         }
 893         pc->mem_cgroup = mem;
 894         smp_wmb();
 895         pc->flags = pcg_default_flags[ctype];
 896
 897         mem_cgroup_charge_statistics(mem, pc, true);
 898
 899         unlock_page_cgroup(pc);
 900 }
 901
 902 /**
 903  * mem_cgroup_move_account - move account of the page
 904  * @pc: page_cgroup of the page.
 905  * @from: mem_cgroup which the page is moved from.
 906  * @to: mem_cgroup which the page is moved to. @from != @to.
 907  *
 908  * The caller must confirm following.
 909  * - page is not on LRU (isolate_page() is useful.)
 910  *
 911  * returns 0 at success,
 912  * returns -EBUSY when lock is busy or "pc" is unstable.
 913  *
 914  * This function does "uncharge" from old cgroup but doesn't do "charge" to
 915  * new cgroup. It should be done by a caller.
 916  */
 917
 918 static int mem_cgroup_move_account(struct page_cgroup *pc,
 919         struct mem_cgroup *from, struct mem_cgroup *to)
 920 {
 921         struct mem_cgroup_per_zone *from_mz, *to_mz;
 922         int nid, zid;
 923         int ret = -EBUSY;
 924
 925         VM_BUG_ON(from == to);
 926         VM_BUG_ON(PageLRU(pc->page));
 927
 928         nid = page_cgroup_nid(pc);
 929         zid = page_cgroup_zid(pc);
 930         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
 931         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
 932
 933         if (!trylock_page_cgroup(pc))
 934                 return ret;
 935
 936         if (!PageCgroupUsed(pc))
 937                 goto out;
 938
 939         if (pc->mem_cgroup != from)
 940                 goto out;
 941
 942         res_counter_uncharge(&from->res, PAGE_SIZE);
 943         mem_cgroup_charge_statistics(from, pc, false);
 944         if (do_swap_account)
 945                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
 946         css_put(&from->css);
 947
 948         css_get(&to->css);
 949         pc->mem_cgroup = to;
 950         mem_cgroup_charge_statistics(to, pc, true);
 951         ret = 0;
 952 out:
 953         unlock_page_cgroup(pc);
 954         return ret;
 955 }
 956
 957 /*
 958  * move charges to its parent.
 959  */
 960
 961 static int mem_cgroup_move_parent(struct page_cgroup *pc,
 962                                   struct mem_cgroup *child,
 963                                   gfp_t gfp_mask)
 964 {
 965         struct page *page = pc->page;
 966         struct cgroup *cg = child->css.cgroup;
 967         struct cgroup *pcg = cg->parent;
 968         struct mem_cgroup *parent;
 969         int ret;
 970
 971         /* Is ROOT ? */
 972         if (!pcg)
 973                 return -EINVAL;
 974
 975
 976         parent = mem_cgroup_from_cont(pcg);
 977
 978
 979         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
 980         if (ret || !parent)
 981                 return ret;
 982
 983         if (!get_page_unless_zero(page)) {
 984                 ret = -EBUSY;
 985                 goto uncharge;
 986         }
 987
 988         ret = isolate_lru_page(page);
 989
 990         if (ret)
 991                 goto cancel;
 992
 993         ret = mem_cgroup_move_account(pc, child, parent);
 994
 995         putback_lru_page(page);
 996         if (!ret) {
 997                 put_page(page);
 998                 /* drop extra refcnt by try_charge() */
 999                 css_put(&parent->css);
1000                 return 0;
1001         }
1002
1003 cancel:
1004         put_page(page);
1005 uncharge:
1006         /* drop extra refcnt by try_charge() */
1007         css_put(&parent->css);
1008         /* uncharge if move fails */
1009         res_counter_uncharge(&parent->res, PAGE_SIZE);
1010         if (do_swap_account)
1011                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1012         return ret;
1013 }
1014
1015 /*
1016  * Charge the memory controller for page usage.
1017  * Return
1018  * 0 if the charge was successful
1019  * < 0 if the cgroup is over its limit
1020  */
1021 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1022                                 gfp_t gfp_mask, enum charge_type ctype,
1023                                 struct mem_cgroup *memcg)
1024 {
1025         struct mem_cgroup *mem;
1026         struct page_cgroup *pc;
1027         int ret;
1028
1029         pc = lookup_page_cgroup(page);
1030         /* can happen at boot */
1031         if (unlikely(!pc))
1032                 return 0;
1033         prefetchw(pc);
1034
1035         mem = memcg;
1036         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1037         if (ret || !mem)
1038                 return ret;
1039
1040         __mem_cgroup_commit_charge(mem, pc, ctype);
1041         return 0;
1042 }
1043
1044 int mem_cgroup_newpage_charge(struct page *page,
1045                               struct mm_struct *mm, gfp_t gfp_mask)
1046 {
1047         if (mem_cgroup_disabled())
1048                 return 0;
1049         if (PageCompound(page))
1050                 return 0;
1051         /*
1052          * If already mapped, we don't have to account.
1053          * If page cache, page->mapping has address_space.
1054          * But page->mapping may have out-of-use anon_vma pointer,
1055          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1056          * is NULL.
1057          */
1058         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1059                 return 0;
1060         if (unlikely(!mm))
1061                 mm = &init_mm;
1062         return mem_cgroup_charge_common(page, mm, gfp_mask,
1063                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1064 }
1065
1066 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1067                                 gfp_t gfp_mask)
1068 {
1069         struct mem_cgroup *mem = NULL;
1070         int ret;
1071
1072         if (mem_cgroup_disabled())
1073                 return 0;
1074         if (PageCompound(page))
1075                 return 0;
1076         /*
1077          * Corner case handling. This is called from add_to_page_cache()
1078          * in usual. But some FS (shmem) precharges this page before calling it
1079          * and call add_to_page_cache() with GFP_NOWAIT.
1080          *
1081          * For GFP_NOWAIT case, the page may be pre-charged before calling
1082          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1083          * charge twice. (It works but has to pay a bit larger cost.)
1084          * And when the page is SwapCache, it should take swap information
1085          * into account. This is under lock_page() now.
1086          */
1087         if (!(gfp_mask & __GFP_WAIT)) {
1088                 struct page_cgroup *pc;
1089
1090
1091                 pc = lookup_page_cgroup(page);
1092                 if (!pc)
1093                         return 0;
1094                 lock_page_cgroup(pc);
1095                 if (PageCgroupUsed(pc)) {
1096                         unlock_page_cgroup(pc);
1097                         return 0;
1098                 }
1099                 unlock_page_cgroup(pc);
1100         }
1101
1102         if (do_swap_account && PageSwapCache(page)) {
1103                 mem = try_get_mem_cgroup_from_swapcache(page);
1104                 if (mem)
1105                         mm = NULL;
1106                   else
1107                         mem = NULL;
1108                 /* SwapCache may be still linked to LRU now. */
1109                 mem_cgroup_lru_del_before_commit_swapcache(page);
1110         }
1111
1112         if (unlikely(!mm && !mem))
1113                 mm = &init_mm;
1114
1115         if (page_is_file_cache(page))
1116                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1117                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1118
1119         ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1120                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1121         if (mem)
1122                 css_put(&mem->css);
1123         if (PageSwapCache(page))
1124                 mem_cgroup_lru_add_after_commit_swapcache(page);
1125
1126         if (do_swap_account && !ret && PageSwapCache(page)) {
1127                 swp_entry_t ent = {.val = page_private(page)};
1128                 /* avoid double counting */
1129                 mem = swap_cgroup_record(ent, NULL);
1130                 if (mem) {
1131                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1132                         mem_cgroup_put(mem);
1133                 }
1134         }
1135         return ret;
1136 }
1137
1138 /*
1139  * While swap-in, try_charge -> commit or cancel, the page is locked.
1140  * And when try_charge() successfully returns, one refcnt to memcg without
1141  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1142  * "commit()" or removed by "cancel()"
1143  */
1144 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1145                                  struct page *page,
1146                                  gfp_t mask, struct mem_cgroup **ptr)
1147 {
1148         struct mem_cgroup *mem;
1149         int ret;
1150
1151         if (mem_cgroup_disabled())
1152                 return 0;
1153
1154         if (!do_swap_account)
1155                 goto charge_cur_mm;
1156         /*
1157          * A racing thread's fault, or swapoff, may have already updated
1158          * the pte, and even removed page from swap cache: return success
1159          * to go on to do_swap_page()'s pte_same() test, which should fail.
1160          */
1161         if (!PageSwapCache(page))
1162                 return 0;
1163         mem = try_get_mem_cgroup_from_swapcache(page);
1164         if (!mem)
1165                 goto charge_cur_mm;
1166         *ptr = mem;
1167         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1168         /* drop extra refcnt from tryget */
1169         css_put(&mem->css);
1170         return ret;
1171 charge_cur_mm:
1172         if (unlikely(!mm))
1173                 mm = &init_mm;
1174         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1175 }
1176
1177 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1178 {
1179         struct page_cgroup *pc;
1180
1181         if (mem_cgroup_disabled())
1182                 return;
1183         if (!ptr)
1184                 return;
1185         pc = lookup_page_cgroup(page);
1186         mem_cgroup_lru_del_before_commit_swapcache(page);
1187         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1188         mem_cgroup_lru_add_after_commit_swapcache(page);
1189         /*
1190          * Now swap is on-memory. This means this page may be
1191          * counted both as mem and swap....double count.
1192          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1193          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1194          * may call delete_from_swap_cache() before reach here.
1195          */
1196         if (do_swap_account && PageSwapCache(page)) {
1197                 swp_entry_t ent = {.val = page_private(page)};
1198                 struct mem_cgroup *memcg;
1199                 memcg = swap_cgroup_record(ent, NULL);
1200                 if (memcg) {
1201                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1202                         mem_cgroup_put(memcg);
1203                 }
1204
1205         }
1206         /* add this page(page_cgroup) to the LRU we want. */
1207
1208 }
1209
1210 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1211 {
1212         if (mem_cgroup_disabled())
1213                 return;
1214         if (!mem)
1215                 return;
1216         res_counter_uncharge(&mem->res, PAGE_SIZE);
1217         if (do_swap_account)
1218                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1219         css_put(&mem->css);
1220 }
1221
1222
1223 /*
1224  * uncharge if !page_mapped(page)
1225  */
1226 static struct mem_cgroup *
1227 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1228 {
1229         struct page_cgroup *pc;
1230         struct mem_cgroup *mem = NULL;
1231         struct mem_cgroup_per_zone *mz;
1232
1233         if (mem_cgroup_disabled())
1234                 return NULL;
1235
1236         if (PageSwapCache(page))
1237                 return NULL;
1238
1239         /*
1240          * Check if our page_cgroup is valid
1241          */
1242         pc = lookup_page_cgroup(page);
1243         if (unlikely(!pc || !PageCgroupUsed(pc)))
1244                 return NULL;
1245
1246         lock_page_cgroup(pc);
1247
1248         mem = pc->mem_cgroup;
1249
1250         if (!PageCgroupUsed(pc))
1251                 goto unlock_out;
1252
1253         switch (ctype) {
1254         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1255                 if (page_mapped(page))
1256                         goto unlock_out;
1257                 break;
1258         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1259                 if (!PageAnon(page)) {  /* Shared memory */
1260                         if (page->mapping && !page_is_file_cache(page))
1261                                 goto unlock_out;
1262                 } else if (page_mapped(page)) /* Anon */
1263                                 goto unlock_out;
1264                 break;
1265         default:
1266                 break;
1267         }
1268
1269         res_counter_uncharge(&mem->res, PAGE_SIZE);
1270         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1271                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1272         mem_cgroup_charge_statistics(mem, pc, false);
1273
1274         ClearPageCgroupUsed(pc);
1275         /*
1276          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1277          * freed from LRU. This is safe because uncharged page is expected not
1278          * to be reused (freed soon). Exception is SwapCache, it's handled by
1279          * special functions.
1280          */
1281
1282         mz = page_cgroup_zoneinfo(pc);
1283         unlock_page_cgroup(pc);
1284
1285         /* at swapout, this memcg will be accessed to record to swap */
1286         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1287                 css_put(&mem->css);
1288
1289         return mem;
1290
1291 unlock_out:
1292         unlock_page_cgroup(pc);
1293         return NULL;
1294 }
1295
1296 void mem_cgroup_uncharge_page(struct page *page)
1297 {
1298         /* early check. */
1299         if (page_mapped(page))
1300                 return;
1301         if (page->mapping && !PageAnon(page))
1302                 return;
1303         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1304 }
1305
1306 void mem_cgroup_uncharge_cache_page(struct page *page)
1307 {
1308         VM_BUG_ON(page_mapped(page));
1309         VM_BUG_ON(page->mapping);
1310         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1311 }
1312
1313 /*
1314  * called from __delete_from_swap_cache() and drop "page" account.
1315  * memcg information is recorded to swap_cgroup of "ent"
1316  */
1317 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1318 {
1319         struct mem_cgroup *memcg;
1320
1321         memcg = __mem_cgroup_uncharge_common(page,
1322                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1323         /* record memcg information */
1324         if (do_swap_account && memcg) {
1325                 swap_cgroup_record(ent, memcg);
1326                 mem_cgroup_get(memcg);
1327         }
1328         if (memcg)
1329                 css_put(&memcg->css);
1330 }
1331
1332 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1333 /*
1334  * called from swap_entry_free(). remove record in swap_cgroup and
1335  * uncharge "memsw" account.
1336  */
1337 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1338 {
1339         struct mem_cgroup *memcg;
1340
1341         if (!do_swap_account)
1342                 return;
1343
1344         memcg = swap_cgroup_record(ent, NULL);
1345         if (memcg) {
1346                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1347                 mem_cgroup_put(memcg);
1348         }
1349 }
1350 #endif
1351
1352 /*
1353  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1354  * page belongs to.
1355  */
1356 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1357 {
1358         struct page_cgroup *pc;
1359         struct mem_cgroup *mem = NULL;
1360         int ret = 0;
1361
1362         if (mem_cgroup_disabled())
1363                 return 0;
1364
1365         pc = lookup_page_cgroup(page);
1366         lock_page_cgroup(pc);
1367         if (PageCgroupUsed(pc)) {
1368                 mem = pc->mem_cgroup;
1369                 css_get(&mem->css);
1370         }
1371         unlock_page_cgroup(pc);
1372
1373         if (mem) {
1374                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1375                 css_put(&mem->css);
1376         }
1377         *ptr = mem;
1378         return ret;
1379 }
1380
1381 /* remove redundant charge if migration failed*/
1382 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1383                 struct page *oldpage, struct page *newpage)
1384 {
1385         struct page *target, *unused;
1386         struct page_cgroup *pc;
1387         enum charge_type ctype;
1388
1389         if (!mem)
1390                 return;
1391
1392         /* at migration success, oldpage->mapping is NULL. */
1393         if (oldpage->mapping) {
1394                 target = oldpage;
1395                 unused = NULL;
1396         } else {
1397                 target = newpage;
1398                 unused = oldpage;
1399         }
1400
1401         if (PageAnon(target))
1402                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1403         else if (page_is_file_cache(target))
1404                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1405         else
1406                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1407
1408         /* unused page is not on radix-tree now. */
1409         if (unused)
1410                 __mem_cgroup_uncharge_common(unused, ctype);
1411
1412         pc = lookup_page_cgroup(target);
1413         /*
1414          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1415          * So, double-counting is effectively avoided.
1416          */
1417         __mem_cgroup_commit_charge(mem, pc, ctype);
1418
1419         /*
1420          * Both of oldpage and newpage are still under lock_page().
1421          * Then, we don't have to care about race in radix-tree.
1422          * But we have to be careful that this page is unmapped or not.
1423          *
1424          * There is a case for !page_mapped(). At the start of
1425          * migration, oldpage was mapped. But now, it's zapped.
1426          * But we know *target* page is not freed/reused under us.
1427          * mem_cgroup_uncharge_page() does all necessary checks.
1428          */
1429         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1430                 mem_cgroup_uncharge_page(target);
1431 }
1432
1433 /*
1434  * A call to try to shrink memory usage under specified resource controller.
1435  * This is typically used for page reclaiming for shmem for reducing side
1436  * effect of page allocation from shmem, which is used by some mem_cgroup.
1437  */
1438 int mem_cgroup_shrink_usage(struct page *page,
1439                             struct mm_struct *mm,
1440                             gfp_t gfp_mask)
1441 {
1442         struct mem_cgroup *mem = NULL;
1443         int progress = 0;
1444         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1445
1446         if (mem_cgroup_disabled())
1447                 return 0;
1448         if (page)
1449                 mem = try_get_mem_cgroup_from_swapcache(page);
1450         if (!mem && mm)
1451                 mem = try_get_mem_cgroup_from_mm(mm);
1452         if (unlikely(!mem))
1453                 return 0;
1454
1455         do {
1456                 progress = mem_cgroup_hierarchical_reclaim(mem, gfp_mask, true);
1457                 progress += mem_cgroup_check_under_limit(mem);
1458         } while (!progress && --retry);
1459
1460         css_put(&mem->css);
1461         if (!retry)
1462                 return -ENOMEM;
1463         return 0;
1464 }
1465
1466 static DEFINE_MUTEX(set_limit_mutex);
1467
1468 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1469                                 unsigned long long val)
1470 {
1471
1472         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1473         int progress;
1474         u64 memswlimit;
1475         int ret = 0;
1476
1477         while (retry_count) {
1478                 if (signal_pending(current)) {
1479                         ret = -EINTR;
1480                         break;
1481                 }
1482                 /*
1483                  * Rather than hide all in some function, I do this in
1484                  * open coded manner. You see what this really does.
1485                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1486                  */
1487                 mutex_lock(&set_limit_mutex);
1488                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1489                 if (memswlimit < val) {
1490                         ret = -EINVAL;
1491                         mutex_unlock(&set_limit_mutex);
1492                         break;
1493                 }
1494                 ret = res_counter_set_limit(&memcg->res, val);
1495                 mutex_unlock(&set_limit_mutex);
1496
1497                 if (!ret)
1498                         break;
1499
1500                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1501                                                            false);
1502                 if (!progress)                  retry_count--;
1503         }
1504
1505         return ret;
1506 }
1507
1508 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1509                                 unsigned long long val)
1510 {
1511         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1512         u64 memlimit, oldusage, curusage;
1513         int ret;
1514
1515         if (!do_swap_account)
1516                 return -EINVAL;
1517
1518         while (retry_count) {
1519                 if (signal_pending(current)) {
1520                         ret = -EINTR;
1521                         break;
1522                 }
1523                 /*
1524                  * Rather than hide all in some function, I do this in
1525                  * open coded manner. You see what this really does.
1526                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1527                  */
1528                 mutex_lock(&set_limit_mutex);
1529                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1530                 if (memlimit > val) {
1531                         ret = -EINVAL;
1532                         mutex_unlock(&set_limit_mutex);
1533                         break;
1534                 }
1535                 ret = res_counter_set_limit(&memcg->memsw, val);
1536                 mutex_unlock(&set_limit_mutex);
1537
1538                 if (!ret)
1539                         break;
1540
1541                 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1542                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true);
1543                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1544                 if (curusage >= oldusage)
1545                         retry_count--;
1546         }
1547         return ret;
1548 }
1549
1550 /*
1551  * This routine traverse page_cgroup in given list and drop them all.
1552  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1553  */
1554 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1555                                 int node, int zid, enum lru_list lru)
1556 {
1557         struct zone *zone;
1558         struct mem_cgroup_per_zone *mz;
1559         struct page_cgroup *pc, *busy;
1560         unsigned long flags, loop;
1561         struct list_head *list;
1562         int ret = 0;
1563
1564         zone = &NODE_DATA(node)->node_zones[zid];
1565         mz = mem_cgroup_zoneinfo(mem, node, zid);
1566         list = &mz->lists[lru];
1567
1568         loop = MEM_CGROUP_ZSTAT(mz, lru);
1569         /* give some margin against EBUSY etc...*/
1570         loop += 256;
1571         busy = NULL;
1572         while (loop--) {
1573                 ret = 0;
1574                 spin_lock_irqsave(&zone->lru_lock, flags);
1575                 if (list_empty(list)) {
1576                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1577                         break;
1578                 }
1579                 pc = list_entry(list->prev, struct page_cgroup, lru);
1580                 if (busy == pc) {
1581                         list_move(&pc->lru, list);
1582                         busy = 0;
1583                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1584                         continue;
1585                 }
1586                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1587
1588                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1589                 if (ret == -ENOMEM)
1590                         break;
1591
1592                 if (ret == -EBUSY || ret == -EINVAL) {
1593                         /* found lock contention or "pc" is obsolete. */
1594                         busy = pc;
1595                         cond_resched();
1596                 } else
1597                         busy = NULL;
1598         }
1599
1600         if (!ret && !list_empty(list))
1601                 return -EBUSY;
1602         return ret;
1603 }
1604
1605 /*
1606  * make mem_cgroup's charge to be 0 if there is no task.
1607  * This enables deleting this mem_cgroup.
1608  */
1609 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1610 {
1611         int ret;
1612         int node, zid, shrink;
1613         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1614         struct cgroup *cgrp = mem->css.cgroup;
1615
1616         css_get(&mem->css);
1617
1618         shrink = 0;
1619         /* should free all ? */
1620         if (free_all)
1621                 goto try_to_free;
1622 move_account:
1623         while (mem->res.usage > 0) {
1624                 ret = -EBUSY;
1625                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1626                         goto out;
1627                 ret = -EINTR;
1628                 if (signal_pending(current))
1629                         goto out;
1630                 /* This is for making all *used* pages to be on LRU. */
1631                 lru_add_drain_all();
1632                 ret = 0;
1633                 for_each_node_state(node, N_HIGH_MEMORY) {
1634                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1635                                 enum lru_list l;
1636                                 for_each_lru(l) {
1637                                         ret = mem_cgroup_force_empty_list(mem,
1638                                                         node, zid, l);
1639                                         if (ret)
1640                                                 break;
1641                                 }
1642                         }
1643                         if (ret)
1644                                 break;
1645                 }
1646                 /* it seems parent cgroup doesn't have enough mem */
1647                 if (ret == -ENOMEM)
1648                         goto try_to_free;
1649                 cond_resched();
1650         }
1651         ret = 0;
1652 out:
1653         css_put(&mem->css);
1654         return ret;
1655
1656 try_to_free:
1657         /* returns EBUSY if there is a task or if we come here twice. */
1658         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1659                 ret = -EBUSY;
1660                 goto out;
1661         }
1662         /* we call try-to-free pages for make this cgroup empty */
1663         lru_add_drain_all();
1664         /* try to free all pages in this cgroup */
1665         shrink = 1;
1666         while (nr_retries && mem->res.usage > 0) {
1667                 int progress;
1668
1669                 if (signal_pending(current)) {
1670                         ret = -EINTR;
1671                         goto out;
1672                 }
1673                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1674                                                 false, get_swappiness(mem));
1675                 if (!progress) {
1676                         nr_retries--;
1677                         /* maybe some writeback is necessary */
1678                         congestion_wait(WRITE, HZ/10);
1679                 }
1680
1681         }
1682         lru_add_drain();
1683         /* try move_account...there may be some *locked* pages. */
1684         if (mem->res.usage)
1685                 goto move_account;
1686         ret = 0;
1687         goto out;
1688 }
1689
1690 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1691 {
1692         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1693 }
1694
1695
1696 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1697 {
1698         return mem_cgroup_from_cont(cont)->use_hierarchy;
1699 }
1700
1701 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1702                                         u64 val)
1703 {
1704         int retval = 0;
1705         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1706         struct cgroup *parent = cont->parent;
1707         struct mem_cgroup *parent_mem = NULL;
1708
1709         if (parent)
1710                 parent_mem = mem_cgroup_from_cont(parent);
1711
1712         cgroup_lock();
1713         /*
1714          * If parent's use_hiearchy is set, we can't make any modifications
1715          * in the child subtrees. If it is unset, then the change can
1716          * occur, provided the current cgroup has no children.
1717          *
1718          * For the root cgroup, parent_mem is NULL, we allow value to be
1719          * set if there are no children.
1720          */
1721         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1722                                 (val == 1 || val == 0)) {
1723                 if (list_empty(&cont->children))
1724                         mem->use_hierarchy = val;
1725                 else
1726                         retval = -EBUSY;
1727         } else
1728                 retval = -EINVAL;
1729         cgroup_unlock();
1730
1731         return retval;
1732 }
1733
1734 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1735 {
1736         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1737         u64 val = 0;
1738         int type, name;
1739
1740         type = MEMFILE_TYPE(cft->private);
1741         name = MEMFILE_ATTR(cft->private);
1742         switch (type) {
1743         case _MEM:
1744                 val = res_counter_read_u64(&mem->res, name);
1745                 break;
1746         case _MEMSWAP:
1747                 if (do_swap_account)
1748                         val = res_counter_read_u64(&mem->memsw, name);
1749                 break;
1750         default:
1751                 BUG();
1752                 break;
1753         }
1754         return val;
1755 }
1756 /*
1757  * The user of this function is...
1758  * RES_LIMIT.
1759  */
1760 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1761                             const char *buffer)
1762 {
1763         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1764         int type, name;
1765         unsigned long long val;
1766         int ret;
1767
1768         type = MEMFILE_TYPE(cft->private);
1769         name = MEMFILE_ATTR(cft->private);
1770         switch (name) {
1771         case RES_LIMIT:
1772                 /* This function does all necessary parse...reuse it */
1773                 ret = res_counter_memparse_write_strategy(buffer, &val);
1774                 if (ret)
1775                         break;
1776                 if (type == _MEM)
1777                         ret = mem_cgroup_resize_limit(memcg, val);
1778                 else
1779                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1780                 break;
1781         default:
1782                 ret = -EINVAL; /* should be BUG() ? */
1783                 break;
1784         }
1785         return ret;
1786 }
1787
1788 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1789                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1790 {
1791         struct cgroup *cgroup;
1792         unsigned long long min_limit, min_memsw_limit, tmp;
1793
1794         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1795         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1796         cgroup = memcg->css.cgroup;
1797         if (!memcg->use_hierarchy)
1798                 goto out;
1799
1800         while (cgroup->parent) {
1801                 cgroup = cgroup->parent;
1802                 memcg = mem_cgroup_from_cont(cgroup);
1803                 if (!memcg->use_hierarchy)
1804                         break;
1805                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
1806                 min_limit = min(min_limit, tmp);
1807                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1808                 min_memsw_limit = min(min_memsw_limit, tmp);
1809         }
1810 out:
1811         *mem_limit = min_limit;
1812         *memsw_limit = min_memsw_limit;
1813         return;
1814 }
1815
1816 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1817 {
1818         struct mem_cgroup *mem;
1819         int type, name;
1820
1821         mem = mem_cgroup_from_cont(cont);
1822         type = MEMFILE_TYPE(event);
1823         name = MEMFILE_ATTR(event);
1824         switch (name) {
1825         case RES_MAX_USAGE:
1826                 if (type == _MEM)
1827                         res_counter_reset_max(&mem->res);
1828                 else
1829                         res_counter_reset_max(&mem->memsw);
1830                 break;
1831         case RES_FAILCNT:
1832                 if (type == _MEM)
1833                         res_counter_reset_failcnt(&mem->res);
1834                 else
1835                         res_counter_reset_failcnt(&mem->memsw);
1836                 break;
1837         }
1838         return 0;
1839 }
1840
1841 static const struct mem_cgroup_stat_desc {
1842         const char *msg;
1843         u64 unit;
1844 } mem_cgroup_stat_desc[] = {
1845         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
1846         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
1847         [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
1848         [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
1849 };
1850
1851 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
1852                                  struct cgroup_map_cb *cb)
1853 {
1854         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
1855         struct mem_cgroup_stat *stat = &mem_cont->stat;
1856         int i;
1857
1858         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
1859                 s64 val;
1860
1861                 val = mem_cgroup_read_stat(stat, i);
1862                 val *= mem_cgroup_stat_desc[i].unit;
1863                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
1864         }
1865         /* showing # of active pages */
1866         {
1867                 unsigned long active_anon, inactive_anon;
1868                 unsigned long active_file, inactive_file;
1869                 unsigned long unevictable;
1870
1871                 inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
1872                                                 LRU_INACTIVE_ANON);
1873                 active_anon = mem_cgroup_get_all_zonestat(mem_cont,
1874                                                 LRU_ACTIVE_ANON);
1875                 inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
1876                                                 LRU_INACTIVE_FILE);
1877                 active_file = mem_cgroup_get_all_zonestat(mem_cont,
1878                                                 LRU_ACTIVE_FILE);
1879                 unevictable = mem_cgroup_get_all_zonestat(mem_cont,
1880                                                         LRU_UNEVICTABLE);
1881
1882                 cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
1883                 cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
1884                 cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
1885                 cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
1886                 cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
1887
1888         }
1889         {
1890                 unsigned long long limit, memsw_limit;
1891                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
1892                 cb->fill(cb, "hierarchical_memory_limit", limit);
1893                 if (do_swap_account)
1894                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
1895         }
1896
1897 #ifdef CONFIG_DEBUG_VM
1898         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
1899
1900         {
1901                 int nid, zid;
1902                 struct mem_cgroup_per_zone *mz;
1903                 unsigned long recent_rotated[2] = {0, 0};
1904                 unsigned long recent_scanned[2] = {0, 0};
1905
1906                 for_each_online_node(nid)
1907                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1908                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1909
1910                                 recent_rotated[0] +=
1911                                         mz->reclaim_stat.recent_rotated[0];
1912                                 recent_rotated[1] +=
1913                                         mz->reclaim_stat.recent_rotated[1];
1914                                 recent_scanned[0] +=
1915                                         mz->reclaim_stat.recent_scanned[0];
1916                                 recent_scanned[1] +=
1917                                         mz->reclaim_stat.recent_scanned[1];
1918                         }
1919                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
1920                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
1921                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
1922                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
1923         }
1924 #endif
1925
1926         return 0;
1927 }
1928
1929 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
1930 {
1931         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1932
1933         return get_swappiness(memcg);
1934 }
1935
1936 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
1937                                        u64 val)
1938 {
1939         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1940         struct mem_cgroup *parent;
1941
1942         if (val > 100)
1943                 return -EINVAL;
1944
1945         if (cgrp->parent == NULL)
1946                 return -EINVAL;
1947
1948         parent = mem_cgroup_from_cont(cgrp->parent);
1949
1950         cgroup_lock();
1951
1952         /* If under hierarchy, only empty-root can set this value */
1953         if ((parent->use_hierarchy) ||
1954             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
1955                 cgroup_unlock();
1956                 return -EINVAL;
1957         }
1958
1959         spin_lock(&memcg->reclaim_param_lock);
1960         memcg->swappiness = val;
1961         spin_unlock(&memcg->reclaim_param_lock);
1962
1963         cgroup_unlock();
1964
1965         return 0;
1966 }
1967
1968
1969 static struct cftype mem_cgroup_files[] = {
1970         {
1971                 .name = "usage_in_bytes",
1972                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
1973                 .read_u64 = mem_cgroup_read,
1974         },
1975         {
1976                 .name = "max_usage_in_bytes",
1977                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
1978                 .trigger = mem_cgroup_reset,
1979                 .read_u64 = mem_cgroup_read,
1980         },
1981         {
1982                 .name = "limit_in_bytes",
1983                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
1984                 .write_string = mem_cgroup_write,
1985                 .read_u64 = mem_cgroup_read,
1986         },
1987         {
1988                 .name = "failcnt",
1989                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
1990                 .trigger = mem_cgroup_reset,
1991                 .read_u64 = mem_cgroup_read,
1992         },
1993         {
1994                 .name = "stat",
1995                 .read_map = mem_control_stat_show,
1996         },
1997         {
1998                 .name = "force_empty",
1999                 .trigger = mem_cgroup_force_empty_write,
2000         },
2001         {
2002                 .name = "use_hierarchy",
2003                 .write_u64 = mem_cgroup_hierarchy_write,
2004                 .read_u64 = mem_cgroup_hierarchy_read,
2005         },
2006         {
2007                 .name = "swappiness",
2008                 .read_u64 = mem_cgroup_swappiness_read,
2009                 .write_u64 = mem_cgroup_swappiness_write,
2010         },
2011 };
2012
2013 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2014 static struct cftype memsw_cgroup_files[] = {
2015         {
2016                 .name = "memsw.usage_in_bytes",
2017                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2018                 .read_u64 = mem_cgroup_read,
2019         },
2020         {
2021                 .name = "memsw.max_usage_in_bytes",
2022                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2023                 .trigger = mem_cgroup_reset,
2024                 .read_u64 = mem_cgroup_read,
2025         },
2026         {
2027                 .name = "memsw.limit_in_bytes",
2028                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2029                 .write_string = mem_cgroup_write,
2030                 .read_u64 = mem_cgroup_read,
2031         },
2032         {
2033                 .name = "memsw.failcnt",
2034                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2035                 .trigger = mem_cgroup_reset,
2036                 .read_u64 = mem_cgroup_read,
2037         },
2038 };
2039
2040 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2041 {
2042         if (!do_swap_account)
2043                 return 0;
2044         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2045                                 ARRAY_SIZE(memsw_cgroup_files));
2046 };
2047 #else
2048 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2049 {
2050         return 0;
2051 }
2052 #endif
2053
2054 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2055 {
2056         struct mem_cgroup_per_node *pn;
2057         struct mem_cgroup_per_zone *mz;
2058         enum lru_list l;
2059         int zone, tmp = node;
2060         /*
2061          * This routine is called against possible nodes.
2062          * But it's BUG to call kmalloc() against offline node.
2063          *
2064          * TODO: this routine can waste much memory for nodes which will
2065          *       never be onlined. It's better to use memory hotplug callback
2066          *       function.
2067          */
2068         if (!node_state(node, N_NORMAL_MEMORY))
2069                 tmp = -1;
2070         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2071         if (!pn)
2072                 return 1;
2073
2074         mem->info.nodeinfo[node] = pn;
2075         memset(pn, 0, sizeof(*pn));
2076
2077         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2078                 mz = &pn->zoneinfo[zone];
2079                 for_each_lru(l)
2080                         INIT_LIST_HEAD(&mz->lists[l]);
2081         }
2082         return 0;
2083 }
2084
2085 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2086 {
2087         kfree(mem->info.nodeinfo[node]);
2088 }
2089
2090 static int mem_cgroup_size(void)
2091 {
2092         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2093         return sizeof(struct mem_cgroup) + cpustat_size;
2094 }
2095
2096 static struct mem_cgroup *mem_cgroup_alloc(void)
2097 {
2098         struct mem_cgroup *mem;
2099         int size = mem_cgroup_size();
2100
2101         if (size < PAGE_SIZE)
2102                 mem = kmalloc(size, GFP_KERNEL);
2103         else
2104                 mem = vmalloc(size);
2105
2106         if (mem)
2107                 memset(mem, 0, size);
2108         return mem;
2109 }
2110
2111 /*
2112  * At destroying mem_cgroup, references from swap_cgroup can remain.
2113  * (scanning all at force_empty is too costly...)
2114  *
2115  * Instead of clearing all references at force_empty, we remember
2116  * the number of reference from swap_cgroup and free mem_cgroup when
2117  * it goes down to 0.
2118  *
2119  * Removal of cgroup itself succeeds regardless of refs from swap.
2120  */
2121
2122 static void __mem_cgroup_free(struct mem_cgroup *mem)
2123 {
2124         int node;
2125
2126         free_css_id(&mem_cgroup_subsys, &mem->css);
2127
2128         for_each_node_state(node, N_POSSIBLE)
2129                 free_mem_cgroup_per_zone_info(mem, node);
2130
2131         if (mem_cgroup_size() < PAGE_SIZE)
2132                 kfree(mem);
2133         else
2134                 vfree(mem);
2135 }
2136
2137 static void mem_cgroup_get(struct mem_cgroup *mem)
2138 {
2139         atomic_inc(&mem->refcnt);
2140 }
2141
2142 static void mem_cgroup_put(struct mem_cgroup *mem)
2143 {
2144         if (atomic_dec_and_test(&mem->refcnt)) {
2145                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2146                 __mem_cgroup_free(mem);
2147                 if (parent)
2148                         mem_cgroup_put(parent);
2149         }
2150 }
2151
2152 /*
2153  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2154  */
2155 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2156 {
2157         if (!mem->res.parent)
2158                 return NULL;
2159         return mem_cgroup_from_res_counter(mem->res.parent, res);
2160 }
2161
2162 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2163 static void __init enable_swap_cgroup(void)
2164 {
2165         if (!mem_cgroup_disabled() && really_do_swap_account)
2166                 do_swap_account = 1;
2167 }
2168 #else
2169 static void __init enable_swap_cgroup(void)
2170 {
2171 }
2172 #endif
2173
2174 static struct cgroup_subsys_state * __ref
2175 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2176 {
2177         struct mem_cgroup *mem, *parent;
2178         long error = -ENOMEM;
2179         int node;
2180
2181         mem = mem_cgroup_alloc();
2182         if (!mem)
2183                 return ERR_PTR(error);
2184
2185         for_each_node_state(node, N_POSSIBLE)
2186                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2187                         goto free_out;
2188         /* root ? */
2189         if (cont->parent == NULL) {
2190                 enable_swap_cgroup();
2191                 parent = NULL;
2192         } else {
2193                 parent = mem_cgroup_from_cont(cont->parent);
2194                 mem->use_hierarchy = parent->use_hierarchy;
2195         }
2196
2197         if (parent && parent->use_hierarchy) {
2198                 res_counter_init(&mem->res, &parent->res);
2199                 res_counter_init(&mem->memsw, &parent->memsw);
2200                 /*
2201                  * We increment refcnt of the parent to ensure that we can
2202                  * safely access it on res_counter_charge/uncharge.
2203                  * This refcnt will be decremented when freeing this
2204                  * mem_cgroup(see mem_cgroup_put).
2205                  */
2206                 mem_cgroup_get(parent);
2207         } else {
2208                 res_counter_init(&mem->res, NULL);
2209                 res_counter_init(&mem->memsw, NULL);
2210         }
2211         mem->last_scanned_child = 0;
2212         spin_lock_init(&mem->reclaim_param_lock);
2213
2214         if (parent)
2215                 mem->swappiness = get_swappiness(parent);
2216         atomic_set(&mem->refcnt, 1);
2217         return &mem->css;
2218 free_out:
2219         __mem_cgroup_free(mem);
2220         return ERR_PTR(error);
2221 }
2222
2223 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2224                                         struct cgroup *cont)
2225 {
2226         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2227
2228         return mem_cgroup_force_empty(mem, false);
2229 }
2230
2231 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2232                                 struct cgroup *cont)
2233 {
2234         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2235
2236         mem_cgroup_put(mem);
2237 }
2238
2239 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2240                                 struct cgroup *cont)
2241 {
2242         int ret;
2243
2244         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2245                                 ARRAY_SIZE(mem_cgroup_files));
2246
2247         if (!ret)
2248                 ret = register_memsw_files(cont, ss);
2249         return ret;
2250 }
2251
2252 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2253                                 struct cgroup *cont,
2254                                 struct cgroup *old_cont,
2255                                 struct task_struct *p)
2256 {
2257         mutex_lock(&memcg_tasklist);
2258         /*
2259          * FIXME: It's better to move charges of this process from old
2260          * memcg to new memcg. But it's just on TODO-List now.
2261          */
2262         mutex_unlock(&memcg_tasklist);
2263 }
2264
2265 struct cgroup_subsys mem_cgroup_subsys = {
2266         .name = "memory",
2267         .subsys_id = mem_cgroup_subsys_id,
2268         .create = mem_cgroup_create,
2269         .pre_destroy = mem_cgroup_pre_destroy,
2270         .destroy = mem_cgroup_destroy,
2271         .populate = mem_cgroup_populate,
2272         .attach = mem_cgroup_move_task,
2273         .early_init = 0,
2274         .use_id = 1,
2275 };
2276
2277 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2278
2279 static int __init disable_swap_account(char *s)
2280 {
2281         really_do_swap_account = 0;
2282         return 1;
2283 }
2284 __setup("noswapaccount", disable_swap_account);
2285 #endif