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/smp.h>
  25 #include <linux/page-flags.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/bit_spinlock.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/swap.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/fs.h>
  32 #include <linux/seq_file.h>
  33
  34 #include <asm/uaccess.h>
  35
  36 struct cgroup_subsys mem_cgroup_subsys;
  37 static const int MEM_CGROUP_RECLAIM_RETRIES = 5;
  38
  39 /*
  40  * Statistics for memory cgroup.
  41  */
  42 enum mem_cgroup_stat_index {
  43         /*
  44          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  45          */
  46         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
  47         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
  48
  49         MEM_CGROUP_STAT_NSTATS,
  50 };
  51
  52 struct mem_cgroup_stat_cpu {
  53         s64 count[MEM_CGROUP_STAT_NSTATS];
  54 } ____cacheline_aligned_in_smp;
  55
  56 struct mem_cgroup_stat {
  57         struct mem_cgroup_stat_cpu cpustat[NR_CPUS];
  58 };
  59
  60 /*
  61  * For accounting under irq disable, no need for increment preempt count.
  62  */
  63 static void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat *stat,
  64                 enum mem_cgroup_stat_index idx, int val)
  65 {
  66         int cpu = smp_processor_id();
  67         stat->cpustat[cpu].count[idx] += val;
  68 }
  69
  70 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
  71                 enum mem_cgroup_stat_index idx)
  72 {
  73         int cpu;
  74         s64 ret = 0;
  75         for_each_possible_cpu(cpu)
  76                 ret += stat->cpustat[cpu].count[idx];
  77         return ret;
  78 }
  79
  80 /*
  81  * per-zone information in memory controller.
  82  */
  83
  84 enum mem_cgroup_zstat_index {
  85         MEM_CGROUP_ZSTAT_ACTIVE,
  86         MEM_CGROUP_ZSTAT_INACTIVE,
  87
  88         NR_MEM_CGROUP_ZSTAT,
  89 };
  90
  91 struct mem_cgroup_per_zone {
  92         unsigned long count[NR_MEM_CGROUP_ZSTAT];
  93 };
  94 /* Macro for accessing counter */
  95 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
  96
  97 struct mem_cgroup_per_node {
  98         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
  99 };
 100
 101 struct mem_cgroup_lru_info {
 102         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 103 };
 104
 105 /*
 106  * The memory controller data structure. The memory controller controls both
 107  * page cache and RSS per cgroup. We would eventually like to provide
 108  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 109  * to help the administrator determine what knobs to tune.
 110  *
 111  * TODO: Add a water mark for the memory controller. Reclaim will begin when
 112  * we hit the water mark. May be even add a low water mark, such that
 113  * no reclaim occurs from a cgroup at it's low water mark, this is
 114  * a feature that will be implemented much later in the future.
 115  */
 116 struct mem_cgroup {
 117         struct cgroup_subsys_state css;
 118         /*
 119          * the counter to account for memory usage
 120          */
 121         struct res_counter res;
 122         /*
 123          * Per cgroup active and inactive list, similar to the
 124          * per zone LRU lists.
 125          * TODO: Consider making these lists per zone
 126          */
 127         struct list_head active_list;
 128         struct list_head inactive_list;
 129         struct mem_cgroup_lru_info info;
 130         /*
 131          * spin_lock to protect the per cgroup LRU
 132          */
 133         spinlock_t lru_lock;
 134         unsigned long control_type;     /* control RSS or RSS+Pagecache */
 135         /*
 136          * statistics.
 137          */
 138         struct mem_cgroup_stat stat;
 139 };
 140
 141 /*
 142  * We use the lower bit of the page->page_cgroup pointer as a bit spin
 143  * lock. We need to ensure that page->page_cgroup is atleast two
 144  * byte aligned (based on comments from Nick Piggin)
 145  */
 146 #define PAGE_CGROUP_LOCK_BIT    0x0
 147 #define PAGE_CGROUP_LOCK                (1 << PAGE_CGROUP_LOCK_BIT)
 148
 149 /*
 150  * A page_cgroup page is associated with every page descriptor. The
 151  * page_cgroup helps us identify information about the cgroup
 152  */
 153 struct page_cgroup {
 154         struct list_head lru;           /* per cgroup LRU list */
 155         struct page *page;
 156         struct mem_cgroup *mem_cgroup;
 157         atomic_t ref_cnt;               /* Helpful when pages move b/w  */
 158                                         /* mapped and cached states     */
 159         int      flags;
 160 };
 161 #define PAGE_CGROUP_FLAG_CACHE  (0x1)   /* charged as cache */
 162 #define PAGE_CGROUP_FLAG_ACTIVE (0x2)   /* page is active in this cgroup */
 163
 164 static inline int page_cgroup_nid(struct page_cgroup *pc)
 165 {
 166         return page_to_nid(pc->page);
 167 }
 168
 169 static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 170 {
 171         return page_zonenum(pc->page);
 172 }
 173
 174 enum {
 175         MEM_CGROUP_TYPE_UNSPEC = 0,
 176         MEM_CGROUP_TYPE_MAPPED,
 177         MEM_CGROUP_TYPE_CACHED,
 178         MEM_CGROUP_TYPE_ALL,
 179         MEM_CGROUP_TYPE_MAX,
 180 };
 181
 182 enum charge_type {
 183         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 184         MEM_CGROUP_CHARGE_TYPE_MAPPED,
 185 };
 186
 187
 188 /*
 189  * Always modified under lru lock. Then, not necessary to preempt_disable()
 190  */
 191 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
 192                                         bool charge)
 193 {
 194         int val = (charge)? 1 : -1;
 195         struct mem_cgroup_stat *stat = &mem->stat;
 196         VM_BUG_ON(!irqs_disabled());
 197
 198         if (flags & PAGE_CGROUP_FLAG_CACHE)
 199                 __mem_cgroup_stat_add_safe(stat,
 200                                         MEM_CGROUP_STAT_CACHE, val);
 201         else
 202                 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
 203 }
 204
 205 static inline struct mem_cgroup_per_zone *
 206 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
 207 {
 208         BUG_ON(!mem->info.nodeinfo[nid]);
 209         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
 210 }
 211
 212 static inline struct mem_cgroup_per_zone *
 213 page_cgroup_zoneinfo(struct page_cgroup *pc)
 214 {
 215         struct mem_cgroup *mem = pc->mem_cgroup;
 216         int nid = page_cgroup_nid(pc);
 217         int zid = page_cgroup_zid(pc);
 218
 219         return mem_cgroup_zoneinfo(mem, nid, zid);
 220 }
 221
 222 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
 223                                         enum mem_cgroup_zstat_index idx)
 224 {
 225         int nid, zid;
 226         struct mem_cgroup_per_zone *mz;
 227         u64 total = 0;
 228
 229         for_each_online_node(nid)
 230                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 231                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
 232                         total += MEM_CGROUP_ZSTAT(mz, idx);
 233                 }
 234         return total;
 235 }
 236
 237 static struct mem_cgroup init_mem_cgroup;
 238
 239 static inline
 240 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 241 {
 242         return container_of(cgroup_subsys_state(cont,
 243                                 mem_cgroup_subsys_id), struct mem_cgroup,
 244                                 css);
 245 }
 246
 247 static inline
 248 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 249 {
 250         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 251                                 struct mem_cgroup, css);
 252 }
 253
 254 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
 255 {
 256         struct mem_cgroup *mem;
 257
 258         mem = mem_cgroup_from_task(p);
 259         css_get(&mem->css);
 260         mm->mem_cgroup = mem;
 261 }
 262
 263 void mm_free_cgroup(struct mm_struct *mm)
 264 {
 265         css_put(&mm->mem_cgroup->css);
 266 }
 267
 268 static inline int page_cgroup_locked(struct page *page)
 269 {
 270         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
 271                                         &page->page_cgroup);
 272 }
 273
 274 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 275 {
 276         int locked;
 277
 278         /*
 279          * While resetting the page_cgroup we might not hold the
 280          * page_cgroup lock. free_hot_cold_page() is an example
 281          * of such a scenario
 282          */
 283         if (pc)
 284                 VM_BUG_ON(!page_cgroup_locked(page));
 285         locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
 286         page->page_cgroup = ((unsigned long)pc | locked);
 287 }
 288
 289 struct page_cgroup *page_get_page_cgroup(struct page *page)
 290 {
 291         return (struct page_cgroup *)
 292                 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 293 }
 294
 295 static void __always_inline lock_page_cgroup(struct page *page)
 296 {
 297         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 298         VM_BUG_ON(!page_cgroup_locked(page));
 299 }
 300
 301 static void __always_inline unlock_page_cgroup(struct page *page)
 302 {
 303         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 304 }
 305
 306 /*
 307  * Tie new page_cgroup to struct page under lock_page_cgroup()
 308  * This can fail if the page has been tied to a page_cgroup.
 309  * If success, returns 0.
 310  */
 311 static int page_cgroup_assign_new_page_cgroup(struct page *page,
 312                                                 struct page_cgroup *pc)
 313 {
 314         int ret = 0;
 315
 316         lock_page_cgroup(page);
 317         if (!page_get_page_cgroup(page))
 318                 page_assign_page_cgroup(page, pc);
 319         else /* A page is tied to other pc. */
 320                 ret = 1;
 321         unlock_page_cgroup(page);
 322         return ret;
 323 }
 324
 325 /*
 326  * Clear page->page_cgroup member under lock_page_cgroup().
 327  * If given "pc" value is different from one page->page_cgroup,
 328  * page->cgroup is not cleared.
 329  * Returns a value of page->page_cgroup at lock taken.
 330  * A can can detect failure of clearing by following
 331  *  clear_page_cgroup(page, pc) == pc
 332  */
 333
 334 static struct page_cgroup *clear_page_cgroup(struct page *page,
 335                                                 struct page_cgroup *pc)
 336 {
 337         struct page_cgroup *ret;
 338         /* lock and clear */
 339         lock_page_cgroup(page);
 340         ret = page_get_page_cgroup(page);
 341         if (likely(ret == pc))
 342                 page_assign_page_cgroup(page, NULL);
 343         unlock_page_cgroup(page);
 344         return ret;
 345 }
 346
 347 static void __mem_cgroup_remove_list(struct page_cgroup *pc)
 348 {
 349         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 350         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 351
 352         if (from)
 353                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 354         else
 355                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 356
 357         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, false);
 358         list_del_init(&pc->lru);
 359 }
 360
 361 static void __mem_cgroup_add_list(struct page_cgroup *pc)
 362 {
 363         int to = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 364         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 365
 366         if (!to) {
 367                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 368                 list_add(&pc->lru, &pc->mem_cgroup->inactive_list);
 369         } else {
 370                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 371                 list_add(&pc->lru, &pc->mem_cgroup->active_list);
 372         }
 373         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, true);
 374 }
 375
 376 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 377 {
 378         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 379         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 380
 381         if (from)
 382                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 383         else
 384                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 385
 386         if (active) {
 387                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 388                 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
 389                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 390         } else {
 391                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 392                 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
 393                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 394         }
 395 }
 396
 397 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 398 {
 399         int ret;
 400
 401         task_lock(task);
 402         ret = task->mm && mm_cgroup(task->mm) == mem;
 403         task_unlock(task);
 404         return ret;
 405 }
 406
 407 /*
 408  * This routine assumes that the appropriate zone's lru lock is already held
 409  */
 410 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 411 {
 412         struct mem_cgroup *mem;
 413         if (!pc)
 414                 return;
 415
 416         mem = pc->mem_cgroup;
 417
 418         spin_lock(&mem->lru_lock);
 419         __mem_cgroup_move_lists(pc, active);
 420         spin_unlock(&mem->lru_lock);
 421 }
 422
 423 /*
 424  * Calculate mapped_ratio under memory controller. This will be used in
 425  * vmscan.c for deteremining we have to reclaim mapped pages.
 426  */
 427 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
 428 {
 429         long total, rss;
 430
 431         /*
 432          * usage is recorded in bytes. But, here, we assume the number of
 433          * physical pages can be represented by "long" on any arch.
 434          */
 435         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
 436         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
 437         return (int)((rss * 100L) / total);
 438 }
 439
 440 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 441                                         struct list_head *dst,
 442                                         unsigned long *scanned, int order,
 443                                         int mode, struct zone *z,
 444                                         struct mem_cgroup *mem_cont,
 445                                         int active)
 446 {
 447         unsigned long nr_taken = 0;
 448         struct page *page;
 449         unsigned long scan;
 450         LIST_HEAD(pc_list);
 451         struct list_head *src;
 452         struct page_cgroup *pc, *tmp;
 453
 454         if (active)
 455                 src = &mem_cont->active_list;
 456         else
 457                 src = &mem_cont->inactive_list;
 458
 459         spin_lock(&mem_cont->lru_lock);
 460         scan = 0;
 461         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 462                 if (scan >= nr_to_scan)
 463                         break;
 464                 page = pc->page;
 465                 VM_BUG_ON(!pc);
 466
 467                 if (unlikely(!PageLRU(page)))
 468                         continue;
 469
 470                 if (PageActive(page) && !active) {
 471                         __mem_cgroup_move_lists(pc, true);
 472                         continue;
 473                 }
 474                 if (!PageActive(page) && active) {
 475                         __mem_cgroup_move_lists(pc, false);
 476                         continue;
 477                 }
 478
 479                 /*
 480                  * Reclaim, per zone
 481                  * TODO: make the active/inactive lists per zone
 482                  */
 483                 if (page_zone(page) != z)
 484                         continue;
 485
 486                 scan++;
 487                 list_move(&pc->lru, &pc_list);
 488
 489                 if (__isolate_lru_page(page, mode) == 0) {
 490                         list_move(&page->lru, dst);
 491                         nr_taken++;
 492                 }
 493         }
 494
 495         list_splice(&pc_list, src);
 496         spin_unlock(&mem_cont->lru_lock);
 497
 498         *scanned = scan;
 499         return nr_taken;
 500 }
 501
 502 /*
 503  * Charge the memory controller for page usage.
 504  * Return
 505  * 0 if the charge was successful
 506  * < 0 if the cgroup is over its limit
 507  */
 508 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 509                                 gfp_t gfp_mask, enum charge_type ctype)
 510 {
 511         struct mem_cgroup *mem;
 512         struct page_cgroup *pc;
 513         unsigned long flags;
 514         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 515
 516         /*
 517          * Should page_cgroup's go to their own slab?
 518          * One could optimize the performance of the charging routine
 519          * by saving a bit in the page_flags and using it as a lock
 520          * to see if the cgroup page already has a page_cgroup associated
 521          * with it
 522          */
 523 retry:
 524         if (page) {
 525                 lock_page_cgroup(page);
 526                 pc = page_get_page_cgroup(page);
 527                 /*
 528                  * The page_cgroup exists and
 529                  * the page has already been accounted.
 530                  */
 531                 if (pc) {
 532                         if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 533                                 /* this page is under being uncharged ? */
 534                                 unlock_page_cgroup(page);
 535                                 cpu_relax();
 536                                 goto retry;
 537                         } else {
 538                                 unlock_page_cgroup(page);
 539                                 goto done;
 540                         }
 541                 }
 542                 unlock_page_cgroup(page);
 543         }
 544
 545         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 546         if (pc == NULL)
 547                 goto err;
 548
 549         /*
 550          * We always charge the cgroup the mm_struct belongs to.
 551          * The mm_struct's mem_cgroup changes on task migration if the
 552          * thread group leader migrates. It's possible that mm is not
 553          * set, if so charge the init_mm (happens for pagecache usage).
 554          */
 555         if (!mm)
 556                 mm = &init_mm;
 557
 558         rcu_read_lock();
 559         mem = rcu_dereference(mm->mem_cgroup);
 560         /*
 561          * For every charge from the cgroup, increment reference
 562          * count
 563          */
 564         css_get(&mem->css);
 565         rcu_read_unlock();
 566
 567         /*
 568          * If we created the page_cgroup, we should free it on exceeding
 569          * the cgroup limit.
 570          */
 571         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 572                 if (!(gfp_mask & __GFP_WAIT))
 573                         goto out;
 574
 575                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 576                         continue;
 577
 578                 /*
 579                  * try_to_free_mem_cgroup_pages() might not give us a full
 580                  * picture of reclaim. Some pages are reclaimed and might be
 581                  * moved to swap cache or just unmapped from the cgroup.
 582                  * Check the limit again to see if the reclaim reduced the
 583                  * current usage of the cgroup before giving up
 584                  */
 585                 if (res_counter_check_under_limit(&mem->res))
 586                         continue;
 587
 588                 if (!nr_retries--) {
 589                         mem_cgroup_out_of_memory(mem, gfp_mask);
 590                         goto out;
 591                 }
 592                 congestion_wait(WRITE, HZ/10);
 593         }
 594
 595         atomic_set(&pc->ref_cnt, 1);
 596         pc->mem_cgroup = mem;
 597         pc->page = page;
 598         pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
 599         if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
 600                 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
 601
 602         if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
 603                 /*
 604                  * Another charge has been added to this page already.
 605                  * We take lock_page_cgroup(page) again and read
 606                  * page->cgroup, increment refcnt.... just retry is OK.
 607                  */
 608                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 609                 css_put(&mem->css);
 610                 kfree(pc);
 611                 if (!page)
 612                         goto done;
 613                 goto retry;
 614         }
 615
 616         spin_lock_irqsave(&mem->lru_lock, flags);
 617         /* Update statistics vector */
 618         __mem_cgroup_add_list(pc);
 619         spin_unlock_irqrestore(&mem->lru_lock, flags);
 620
 621 done:
 622         return 0;
 623 out:
 624         css_put(&mem->css);
 625         kfree(pc);
 626 err:
 627         return -ENOMEM;
 628 }
 629
 630 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 631                         gfp_t gfp_mask)
 632 {
 633         return mem_cgroup_charge_common(page, mm, gfp_mask,
 634                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
 635 }
 636
 637 /*
 638  * See if the cached pages should be charged at all?
 639  */
 640 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 641                                 gfp_t gfp_mask)
 642 {
 643         int ret = 0;
 644         struct mem_cgroup *mem;
 645         if (!mm)
 646                 mm = &init_mm;
 647
 648         rcu_read_lock();
 649         mem = rcu_dereference(mm->mem_cgroup);
 650         css_get(&mem->css);
 651         rcu_read_unlock();
 652         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 653                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
 654                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
 655         css_put(&mem->css);
 656         return ret;
 657 }
 658
 659 /*
 660  * Uncharging is always a welcome operation, we never complain, simply
 661  * uncharge.
 662  */
 663 void mem_cgroup_uncharge(struct page_cgroup *pc)
 664 {
 665         struct mem_cgroup *mem;
 666         struct page *page;
 667         unsigned long flags;
 668
 669         /*
 670          * This can handle cases when a page is not charged at all and we
 671          * are switching between handling the control_type.
 672          */
 673         if (!pc)
 674                 return;
 675
 676         if (atomic_dec_and_test(&pc->ref_cnt)) {
 677                 page = pc->page;
 678                 /*
 679                  * get page->cgroup and clear it under lock.
 680                  * force_empty can drop page->cgroup without checking refcnt.
 681                  */
 682                 if (clear_page_cgroup(page, pc) == pc) {
 683                         mem = pc->mem_cgroup;
 684                         css_put(&mem->css);
 685                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 686                         spin_lock_irqsave(&mem->lru_lock, flags);
 687                         __mem_cgroup_remove_list(pc);
 688                         spin_unlock_irqrestore(&mem->lru_lock, flags);
 689                         kfree(pc);
 690                 }
 691         }
 692 }
 693
 694 /*
 695  * Returns non-zero if a page (under migration) has valid page_cgroup member.
 696  * Refcnt of page_cgroup is incremented.
 697  */
 698
 699 int mem_cgroup_prepare_migration(struct page *page)
 700 {
 701         struct page_cgroup *pc;
 702         int ret = 0;
 703         lock_page_cgroup(page);
 704         pc = page_get_page_cgroup(page);
 705         if (pc && atomic_inc_not_zero(&pc->ref_cnt))
 706                 ret = 1;
 707         unlock_page_cgroup(page);
 708         return ret;
 709 }
 710
 711 void mem_cgroup_end_migration(struct page *page)
 712 {
 713         struct page_cgroup *pc = page_get_page_cgroup(page);
 714         mem_cgroup_uncharge(pc);
 715 }
 716 /*
 717  * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
 718  * And no race with uncharge() routines because page_cgroup for *page*
 719  * has extra one reference by mem_cgroup_prepare_migration.
 720  */
 721
 722 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
 723 {
 724         struct page_cgroup *pc;
 725         struct mem_cgroup *mem;
 726         unsigned long flags;
 727 retry:
 728         pc = page_get_page_cgroup(page);
 729         if (!pc)
 730                 return;
 731         mem = pc->mem_cgroup;
 732         if (clear_page_cgroup(page, pc) != pc)
 733                 goto retry;
 734
 735         spin_lock_irqsave(&mem->lru_lock, flags);
 736
 737         __mem_cgroup_remove_list(pc);
 738         pc->page = newpage;
 739         lock_page_cgroup(newpage);
 740         page_assign_page_cgroup(newpage, pc);
 741         unlock_page_cgroup(newpage);
 742         __mem_cgroup_add_list(pc);
 743
 744         spin_unlock_irqrestore(&mem->lru_lock, flags);
 745         return;
 746 }
 747
 748 /*
 749  * This routine traverse page_cgroup in given list and drop them all.
 750  * This routine ignores page_cgroup->ref_cnt.
 751  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
 752  */
 753 #define FORCE_UNCHARGE_BATCH    (128)
 754 static void
 755 mem_cgroup_force_empty_list(struct mem_cgroup *mem, struct list_head *list)
 756 {
 757         struct page_cgroup *pc;
 758         struct page *page;
 759         int count;
 760         unsigned long flags;
 761
 762 retry:
 763         count = FORCE_UNCHARGE_BATCH;
 764         spin_lock_irqsave(&mem->lru_lock, flags);
 765
 766         while (--count && !list_empty(list)) {
 767                 pc = list_entry(list->prev, struct page_cgroup, lru);
 768                 page = pc->page;
 769                 /* Avoid race with charge */
 770                 atomic_set(&pc->ref_cnt, 0);
 771                 if (clear_page_cgroup(page, pc) == pc) {
 772                         css_put(&mem->css);
 773                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 774                         __mem_cgroup_remove_list(pc);
 775                         kfree(pc);
 776                 } else  /* being uncharged ? ...do relax */
 777                         break;
 778         }
 779         spin_unlock_irqrestore(&mem->lru_lock, flags);
 780         if (!list_empty(list)) {
 781                 cond_resched();
 782                 goto retry;
 783         }
 784         return;
 785 }
 786
 787 /*
 788  * make mem_cgroup's charge to be 0 if there is no task.
 789  * This enables deleting this mem_cgroup.
 790  */
 791
 792 int mem_cgroup_force_empty(struct mem_cgroup *mem)
 793 {
 794         int ret = -EBUSY;
 795         css_get(&mem->css);
 796         /*
 797          * page reclaim code (kswapd etc..) will move pages between
 798 `        * active_list <-> inactive_list while we don't take a lock.
 799          * So, we have to do loop here until all lists are empty.
 800          */
 801         while (!(list_empty(&mem->active_list) &&
 802                  list_empty(&mem->inactive_list))) {
 803                 if (atomic_read(&mem->css.cgroup->count) > 0)
 804                         goto out;
 805                 /* drop all page_cgroup in active_list */
 806                 mem_cgroup_force_empty_list(mem, &mem->active_list);
 807                 /* drop all page_cgroup in inactive_list */
 808                 mem_cgroup_force_empty_list(mem, &mem->inactive_list);
 809         }
 810         ret = 0;
 811 out:
 812         css_put(&mem->css);
 813         return ret;
 814 }
 815
 816
 817
 818 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 819 {
 820         *tmp = memparse(buf, &buf);
 821         if (*buf != '\0')
 822                 return -EINVAL;
 823
 824         /*
 825          * Round up the value to the closest page size
 826          */
 827         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 828         return 0;
 829 }
 830
 831 static ssize_t mem_cgroup_read(struct cgroup *cont,
 832                         struct cftype *cft, struct file *file,
 833                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 834 {
 835         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 836                                 cft->private, userbuf, nbytes, ppos,
 837                                 NULL);
 838 }
 839
 840 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 841                                 struct file *file, const char __user *userbuf,
 842                                 size_t nbytes, loff_t *ppos)
 843 {
 844         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 845                                 cft->private, userbuf, nbytes, ppos,
 846                                 mem_cgroup_write_strategy);
 847 }
 848
 849 static ssize_t mem_control_type_write(struct cgroup *cont,
 850                         struct cftype *cft, struct file *file,
 851                         const char __user *userbuf,
 852                         size_t nbytes, loff_t *pos)
 853 {
 854         int ret;
 855         char *buf, *end;
 856         unsigned long tmp;
 857         struct mem_cgroup *mem;
 858
 859         mem = mem_cgroup_from_cont(cont);
 860         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 861         ret = -ENOMEM;
 862         if (buf == NULL)
 863                 goto out;
 864
 865         buf[nbytes] = 0;
 866         ret = -EFAULT;
 867         if (copy_from_user(buf, userbuf, nbytes))
 868                 goto out_free;
 869
 870         ret = -EINVAL;
 871         tmp = simple_strtoul(buf, &end, 10);
 872         if (*end != '\0')
 873                 goto out_free;
 874
 875         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 876                 goto out_free;
 877
 878         mem->control_type = tmp;
 879         ret = nbytes;
 880 out_free:
 881         kfree(buf);
 882 out:
 883         return ret;
 884 }
 885
 886 static ssize_t mem_control_type_read(struct cgroup *cont,
 887                                 struct cftype *cft,
 888                                 struct file *file, char __user *userbuf,
 889                                 size_t nbytes, loff_t *ppos)
 890 {
 891         unsigned long val;
 892         char buf[64], *s;
 893         struct mem_cgroup *mem;
 894
 895         mem = mem_cgroup_from_cont(cont);
 896         s = buf;
 897         val = mem->control_type;
 898         s += sprintf(s, "%lu\n", val);
 899         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 900                         ppos, buf, s - buf);
 901 }
 902
 903
 904 static ssize_t mem_force_empty_write(struct cgroup *cont,
 905                                 struct cftype *cft, struct file *file,
 906                                 const char __user *userbuf,
 907                                 size_t nbytes, loff_t *ppos)
 908 {
 909         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
 910         int ret;
 911         ret = mem_cgroup_force_empty(mem);
 912         if (!ret)
 913                 ret = nbytes;
 914         return ret;
 915 }
 916
 917 /*
 918  * Note: This should be removed if cgroup supports write-only file.
 919  */
 920
 921 static ssize_t mem_force_empty_read(struct cgroup *cont,
 922                                 struct cftype *cft,
 923                                 struct file *file, char __user *userbuf,
 924                                 size_t nbytes, loff_t *ppos)
 925 {
 926         return -EINVAL;
 927 }
 928
 929
 930 static const struct mem_cgroup_stat_desc {
 931         const char *msg;
 932         u64 unit;
 933 } mem_cgroup_stat_desc[] = {
 934         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
 935         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
 936 };
 937
 938 static int mem_control_stat_show(struct seq_file *m, void *arg)
 939 {
 940         struct cgroup *cont = m->private;
 941         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
 942         struct mem_cgroup_stat *stat = &mem_cont->stat;
 943         int i;
 944
 945         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
 946                 s64 val;
 947
 948                 val = mem_cgroup_read_stat(stat, i);
 949                 val *= mem_cgroup_stat_desc[i].unit;
 950                 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
 951                                 (long long)val);
 952         }
 953         /* showing # of active pages */
 954         {
 955                 unsigned long active, inactive;
 956
 957                 inactive = mem_cgroup_get_all_zonestat(mem_cont,
 958                                                 MEM_CGROUP_ZSTAT_INACTIVE);
 959                 active = mem_cgroup_get_all_zonestat(mem_cont,
 960                                                 MEM_CGROUP_ZSTAT_ACTIVE);
 961                 seq_printf(m, "active %ld\n", (active) * PAGE_SIZE);
 962                 seq_printf(m, "inactive %ld\n", (inactive) * PAGE_SIZE);
 963         }
 964         return 0;
 965 }
 966
 967 static const struct file_operations mem_control_stat_file_operations = {
 968         .read = seq_read,
 969         .llseek = seq_lseek,
 970         .release = single_release,
 971 };
 972
 973 static int mem_control_stat_open(struct inode *unused, struct file *file)
 974 {
 975         /* XXX __d_cont */
 976         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
 977
 978         file->f_op = &mem_control_stat_file_operations;
 979         return single_open(file, mem_control_stat_show, cont);
 980 }
 981
 982
 983
 984 static struct cftype mem_cgroup_files[] = {
 985         {
 986                 .name = "usage_in_bytes",
 987                 .private = RES_USAGE,
 988                 .read = mem_cgroup_read,
 989         },
 990         {
 991                 .name = "limit_in_bytes",
 992                 .private = RES_LIMIT,
 993                 .write = mem_cgroup_write,
 994                 .read = mem_cgroup_read,
 995         },
 996         {
 997                 .name = "failcnt",
 998                 .private = RES_FAILCNT,
 999                 .read = mem_cgroup_read,
1000         },
1001         {
1002                 .name = "control_type",
1003                 .write = mem_control_type_write,
1004                 .read = mem_control_type_read,
1005         },
1006         {
1007                 .name = "force_empty",
1008                 .write = mem_force_empty_write,
1009                 .read = mem_force_empty_read,
1010         },
1011         {
1012                 .name = "stat",
1013                 .open = mem_control_stat_open,
1014         },
1015 };
1016
1017 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1018 {
1019         struct mem_cgroup_per_node *pn;
1020
1021         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, node);
1022         if (!pn)
1023                 return 1;
1024         mem->info.nodeinfo[node] = pn;
1025         memset(pn, 0, sizeof(*pn));
1026         return 0;
1027 }
1028
1029 static struct mem_cgroup init_mem_cgroup;
1030
1031 static struct cgroup_subsys_state *
1032 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1033 {
1034         struct mem_cgroup *mem;
1035         int node;
1036
1037         if (unlikely((cont->parent) == NULL)) {
1038                 mem = &init_mem_cgroup;
1039                 init_mm.mem_cgroup = mem;
1040         } else
1041                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
1042
1043         if (mem == NULL)
1044                 return NULL;
1045
1046         res_counter_init(&mem->res);
1047         INIT_LIST_HEAD(&mem->active_list);
1048         INIT_LIST_HEAD(&mem->inactive_list);
1049         spin_lock_init(&mem->lru_lock);
1050         mem->control_type = MEM_CGROUP_TYPE_ALL;
1051         memset(&mem->info, 0, sizeof(mem->info));
1052
1053         for_each_node_state(node, N_POSSIBLE)
1054                 if (alloc_mem_cgroup_per_zone_info(mem, node))
1055                         goto free_out;
1056
1057         return &mem->css;
1058 free_out:
1059         for_each_node_state(node, N_POSSIBLE)
1060                 kfree(mem->info.nodeinfo[node]);
1061         if (cont->parent != NULL)
1062                 kfree(mem);
1063         return NULL;
1064 }
1065
1066 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1067                                         struct cgroup *cont)
1068 {
1069         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1070         mem_cgroup_force_empty(mem);
1071 }
1072
1073 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
1074                                 struct cgroup *cont)
1075 {
1076         int node;
1077         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1078
1079         for_each_node_state(node, N_POSSIBLE)
1080                 kfree(mem->info.nodeinfo[node]);
1081
1082         kfree(mem_cgroup_from_cont(cont));
1083 }
1084
1085 static int mem_cgroup_populate(struct cgroup_subsys *ss,
1086                                 struct cgroup *cont)
1087 {
1088         return cgroup_add_files(cont, ss, mem_cgroup_files,
1089                                         ARRAY_SIZE(mem_cgroup_files));
1090 }
1091
1092 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
1093                                 struct cgroup *cont,
1094                                 struct cgroup *old_cont,
1095                                 struct task_struct *p)
1096 {
1097         struct mm_struct *mm;
1098         struct mem_cgroup *mem, *old_mem;
1099
1100         mm = get_task_mm(p);
1101         if (mm == NULL)
1102                 return;
1103
1104         mem = mem_cgroup_from_cont(cont);
1105         old_mem = mem_cgroup_from_cont(old_cont);
1106
1107         if (mem == old_mem)
1108                 goto out;
1109
1110         /*
1111          * Only thread group leaders are allowed to migrate, the mm_struct is
1112          * in effect owned by the leader
1113          */
1114         if (p->tgid != p->pid)
1115                 goto out;
1116
1117         css_get(&mem->css);
1118         rcu_assign_pointer(mm->mem_cgroup, mem);
1119         css_put(&old_mem->css);
1120
1121 out:
1122         mmput(mm);
1123         return;
1124 }
1125
1126 struct cgroup_subsys mem_cgroup_subsys = {
1127         .name = "memory",
1128         .subsys_id = mem_cgroup_subsys_id,
1129         .create = mem_cgroup_create,
1130         .pre_destroy = mem_cgroup_pre_destroy,
1131         .destroy = mem_cgroup_destroy,
1132         .populate = mem_cgroup_populate,
1133         .attach = mem_cgroup_move_task,
1134         .early_init = 0,
1135 };