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