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