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         /*
  93          * spin_lock to protect the per cgroup LRU
  94          */
  95         spinlock_t              lru_lock;
  96         struct list_head        active_list;
  97         struct list_head        inactive_list;
  98         unsigned long count[NR_MEM_CGROUP_ZSTAT];
  99 };
 100 /* Macro for accessing counter */
 101 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
 102
 103 struct mem_cgroup_per_node {
 104         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 105 };
 106
 107 struct mem_cgroup_lru_info {
 108         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 109 };
 110
 111 /*
 112  * The memory controller data structure. The memory controller controls both
 113  * page cache and RSS per cgroup. We would eventually like to provide
 114  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 115  * to help the administrator determine what knobs to tune.
 116  *
 117  * TODO: Add a water mark for the memory controller. Reclaim will begin when
 118  * we hit the water mark. May be even add a low water mark, such that
 119  * no reclaim occurs from a cgroup at it's low water mark, this is
 120  * a feature that will be implemented much later in the future.
 121  */
 122 struct mem_cgroup {
 123         struct cgroup_subsys_state css;
 124         /*
 125          * the counter to account for memory usage
 126          */
 127         struct res_counter res;
 128         /*
 129          * Per cgroup active and inactive list, similar to the
 130          * per zone LRU lists.
 131          */
 132         struct mem_cgroup_lru_info info;
 133
 134         int     prev_priority;  /* for recording reclaim priority */
 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, &mz->inactive_list);
 369         } else {
 370                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 371                 list_add(&pc->lru, &mz->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, &mz->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, &mz->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 && vm_match_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_per_zone *mz;
 413         unsigned long flags;
 414
 415         if (!pc)
 416                 return;
 417
 418         mz = page_cgroup_zoneinfo(pc);
 419         spin_lock_irqsave(&mz->lru_lock, flags);
 420         __mem_cgroup_move_lists(pc, active);
 421         spin_unlock_irqrestore(&mz->lru_lock, flags);
 422 }
 423
 424 /*
 425  * Calculate mapped_ratio under memory controller. This will be used in
 426  * vmscan.c for deteremining we have to reclaim mapped pages.
 427  */
 428 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
 429 {
 430         long total, rss;
 431
 432         /*
 433          * usage is recorded in bytes. But, here, we assume the number of
 434          * physical pages can be represented by "long" on any arch.
 435          */
 436         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
 437         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
 438         return (int)((rss * 100L) / total);
 439 }
 440 /*
 441  * This function is called from vmscan.c. In page reclaiming loop. balance
 442  * between active and inactive list is calculated. For memory controller
 443  * page reclaiming, we should use using mem_cgroup's imbalance rather than
 444  * zone's global lru imbalance.
 445  */
 446 long mem_cgroup_reclaim_imbalance(struct mem_cgroup *mem)
 447 {
 448         unsigned long active, inactive;
 449         /* active and inactive are the number of pages. 'long' is ok.*/
 450         active = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_ACTIVE);
 451         inactive = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_INACTIVE);
 452         return (long) (active / (inactive + 1));
 453 }
 454
 455 /*
 456  * prev_priority control...this will be used in memory reclaim path.
 457  */
 458 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
 459 {
 460         return mem->prev_priority;
 461 }
 462
 463 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
 464 {
 465         if (priority < mem->prev_priority)
 466                 mem->prev_priority = priority;
 467 }
 468
 469 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
 470 {
 471         mem->prev_priority = priority;
 472 }
 473
 474 /*
 475  * Calculate # of pages to be scanned in this priority/zone.
 476  * See also vmscan.c
 477  *
 478  * priority starts from "DEF_PRIORITY" and decremented in each loop.
 479  * (see include/linux/mmzone.h)
 480  */
 481
 482 long mem_cgroup_calc_reclaim_active(struct mem_cgroup *mem,
 483                                    struct zone *zone, int priority)
 484 {
 485         long nr_active;
 486         int nid = zone->zone_pgdat->node_id;
 487         int zid = zone_idx(zone);
 488         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
 489
 490         nr_active = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE);
 491         return (nr_active >> priority);
 492 }
 493
 494 long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup *mem,
 495                                         struct zone *zone, int priority)
 496 {
 497         long nr_inactive;
 498         int nid = zone->zone_pgdat->node_id;
 499         int zid = zone_idx(zone);
 500         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
 501
 502         nr_inactive = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE);
 503
 504         return (nr_inactive >> priority);
 505 }
 506
 507 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 508                                         struct list_head *dst,
 509                                         unsigned long *scanned, int order,
 510                                         int mode, struct zone *z,
 511                                         struct mem_cgroup *mem_cont,
 512                                         int active)
 513 {
 514         unsigned long nr_taken = 0;
 515         struct page *page;
 516         unsigned long scan;
 517         LIST_HEAD(pc_list);
 518         struct list_head *src;
 519         struct page_cgroup *pc, *tmp;
 520         int nid = z->zone_pgdat->node_id;
 521         int zid = zone_idx(z);
 522         struct mem_cgroup_per_zone *mz;
 523
 524         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
 525         if (active)
 526                 src = &mz->active_list;
 527         else
 528                 src = &mz->inactive_list;
 529
 530
 531         spin_lock(&mz->lru_lock);
 532         scan = 0;
 533         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 534                 if (scan >= nr_to_scan)
 535                         break;
 536                 page = pc->page;
 537                 VM_BUG_ON(!pc);
 538
 539                 if (unlikely(!PageLRU(page)))
 540                         continue;
 541
 542                 if (PageActive(page) && !active) {
 543                         __mem_cgroup_move_lists(pc, true);
 544                         continue;
 545                 }
 546                 if (!PageActive(page) && active) {
 547                         __mem_cgroup_move_lists(pc, false);
 548                         continue;
 549                 }
 550
 551                 scan++;
 552                 list_move(&pc->lru, &pc_list);
 553
 554                 if (__isolate_lru_page(page, mode) == 0) {
 555                         list_move(&page->lru, dst);
 556                         nr_taken++;
 557                 }
 558         }
 559
 560         list_splice(&pc_list, src);
 561         spin_unlock(&mz->lru_lock);
 562
 563         *scanned = scan;
 564         return nr_taken;
 565 }
 566
 567 /*
 568  * Charge the memory controller for page usage.
 569  * Return
 570  * 0 if the charge was successful
 571  * < 0 if the cgroup is over its limit
 572  */
 573 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 574                                 gfp_t gfp_mask, enum charge_type ctype)
 575 {
 576         struct mem_cgroup *mem;
 577         struct page_cgroup *pc;
 578         unsigned long flags;
 579         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 580         struct mem_cgroup_per_zone *mz;
 581
 582         /*
 583          * Should page_cgroup's go to their own slab?
 584          * One could optimize the performance of the charging routine
 585          * by saving a bit in the page_flags and using it as a lock
 586          * to see if the cgroup page already has a page_cgroup associated
 587          * with it
 588          */
 589 retry:
 590         if (page) {
 591                 lock_page_cgroup(page);
 592                 pc = page_get_page_cgroup(page);
 593                 /*
 594                  * The page_cgroup exists and
 595                  * the page has already been accounted.
 596                  */
 597                 if (pc) {
 598                         if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 599                                 /* this page is under being uncharged ? */
 600                                 unlock_page_cgroup(page);
 601                                 cpu_relax();
 602                                 goto retry;
 603                         } else {
 604                                 unlock_page_cgroup(page);
 605                                 goto done;
 606                         }
 607                 }
 608                 unlock_page_cgroup(page);
 609         }
 610
 611         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 612         if (pc == NULL)
 613                 goto err;
 614
 615         /*
 616          * We always charge the cgroup the mm_struct belongs to.
 617          * The mm_struct's mem_cgroup changes on task migration if the
 618          * thread group leader migrates. It's possible that mm is not
 619          * set, if so charge the init_mm (happens for pagecache usage).
 620          */
 621         if (!mm)
 622                 mm = &init_mm;
 623
 624         rcu_read_lock();
 625         mem = rcu_dereference(mm->mem_cgroup);
 626         /*
 627          * For every charge from the cgroup, increment reference
 628          * count
 629          */
 630         css_get(&mem->css);
 631         rcu_read_unlock();
 632
 633         /*
 634          * If we created the page_cgroup, we should free it on exceeding
 635          * the cgroup limit.
 636          */
 637         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 638                 if (!(gfp_mask & __GFP_WAIT))
 639                         goto out;
 640
 641                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 642                         continue;
 643
 644                 /*
 645                  * try_to_free_mem_cgroup_pages() might not give us a full
 646                  * picture of reclaim. Some pages are reclaimed and might be
 647                  * moved to swap cache or just unmapped from the cgroup.
 648                  * Check the limit again to see if the reclaim reduced the
 649                  * current usage of the cgroup before giving up
 650                  */
 651                 if (res_counter_check_under_limit(&mem->res))
 652                         continue;
 653
 654                 if (!nr_retries--) {
 655                         mem_cgroup_out_of_memory(mem, gfp_mask);
 656                         goto out;
 657                 }
 658                 congestion_wait(WRITE, HZ/10);
 659         }
 660
 661         atomic_set(&pc->ref_cnt, 1);
 662         pc->mem_cgroup = mem;
 663         pc->page = page;
 664         pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
 665         if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
 666                 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
 667
 668         if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
 669                 /*
 670                  * Another charge has been added to this page already.
 671                  * We take lock_page_cgroup(page) again and read
 672                  * page->cgroup, increment refcnt.... just retry is OK.
 673                  */
 674                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 675                 css_put(&mem->css);
 676                 kfree(pc);
 677                 if (!page)
 678                         goto done;
 679                 goto retry;
 680         }
 681
 682         mz = page_cgroup_zoneinfo(pc);
 683         spin_lock_irqsave(&mz->lru_lock, flags);
 684         /* Update statistics vector */
 685         __mem_cgroup_add_list(pc);
 686         spin_unlock_irqrestore(&mz->lru_lock, flags);
 687
 688 done:
 689         return 0;
 690 out:
 691         css_put(&mem->css);
 692         kfree(pc);
 693 err:
 694         return -ENOMEM;
 695 }
 696
 697 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 698                         gfp_t gfp_mask)
 699 {
 700         return mem_cgroup_charge_common(page, mm, gfp_mask,
 701                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
 702 }
 703
 704 /*
 705  * See if the cached pages should be charged at all?
 706  */
 707 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 708                                 gfp_t gfp_mask)
 709 {
 710         int ret = 0;
 711         if (!mm)
 712                 mm = &init_mm;
 713
 714         ret = mem_cgroup_charge_common(page, mm, gfp_mask,
 715                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
 716         return ret;
 717 }
 718
 719 /*
 720  * Uncharging is always a welcome operation, we never complain, simply
 721  * uncharge. This routine should be called with lock_page_cgroup held
 722  */
 723 void mem_cgroup_uncharge(struct page_cgroup *pc)
 724 {
 725         struct mem_cgroup *mem;
 726         struct mem_cgroup_per_zone *mz;
 727         struct page *page;
 728         unsigned long flags;
 729
 730         /*
 731          * Check if our page_cgroup is valid
 732          */
 733         if (!pc)
 734                 return;
 735
 736         if (atomic_dec_and_test(&pc->ref_cnt)) {
 737                 page = pc->page;
 738                 mz = page_cgroup_zoneinfo(pc);
 739                 /*
 740                  * get page->cgroup and clear it under lock.
 741                  * force_empty can drop page->cgroup without checking refcnt.
 742                  */
 743                 unlock_page_cgroup(page);
 744                 if (clear_page_cgroup(page, pc) == pc) {
 745                         mem = pc->mem_cgroup;
 746                         css_put(&mem->css);
 747                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 748                         spin_lock_irqsave(&mz->lru_lock, flags);
 749                         __mem_cgroup_remove_list(pc);
 750                         spin_unlock_irqrestore(&mz->lru_lock, flags);
 751                         kfree(pc);
 752                 }
 753                 lock_page_cgroup(page);
 754         }
 755 }
 756
 757 void mem_cgroup_uncharge_page(struct page *page)
 758 {
 759         lock_page_cgroup(page);
 760         mem_cgroup_uncharge(page_get_page_cgroup(page));
 761         unlock_page_cgroup(page);
 762 }
 763
 764 /*
 765  * Returns non-zero if a page (under migration) has valid page_cgroup member.
 766  * Refcnt of page_cgroup is incremented.
 767  */
 768
 769 int mem_cgroup_prepare_migration(struct page *page)
 770 {
 771         struct page_cgroup *pc;
 772         int ret = 0;
 773         lock_page_cgroup(page);
 774         pc = page_get_page_cgroup(page);
 775         if (pc && atomic_inc_not_zero(&pc->ref_cnt))
 776                 ret = 1;
 777         unlock_page_cgroup(page);
 778         return ret;
 779 }
 780
 781 void mem_cgroup_end_migration(struct page *page)
 782 {
 783         struct page_cgroup *pc;
 784
 785         lock_page_cgroup(page);
 786         pc = page_get_page_cgroup(page);
 787         mem_cgroup_uncharge(pc);
 788         unlock_page_cgroup(page);
 789 }
 790 /*
 791  * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
 792  * And no race with uncharge() routines because page_cgroup for *page*
 793  * has extra one reference by mem_cgroup_prepare_migration.
 794  */
 795
 796 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
 797 {
 798         struct page_cgroup *pc;
 799         struct mem_cgroup *mem;
 800         unsigned long flags;
 801         struct mem_cgroup_per_zone *mz;
 802 retry:
 803         pc = page_get_page_cgroup(page);
 804         if (!pc)
 805                 return;
 806         mem = pc->mem_cgroup;
 807         mz = page_cgroup_zoneinfo(pc);
 808         if (clear_page_cgroup(page, pc) != pc)
 809                 goto retry;
 810         spin_lock_irqsave(&mz->lru_lock, flags);
 811
 812         __mem_cgroup_remove_list(pc);
 813         spin_unlock_irqrestore(&mz->lru_lock, flags);
 814
 815         pc->page = newpage;
 816         lock_page_cgroup(newpage);
 817         page_assign_page_cgroup(newpage, pc);
 818         unlock_page_cgroup(newpage);
 819
 820         mz = page_cgroup_zoneinfo(pc);
 821         spin_lock_irqsave(&mz->lru_lock, flags);
 822         __mem_cgroup_add_list(pc);
 823         spin_unlock_irqrestore(&mz->lru_lock, flags);
 824         return;
 825 }
 826
 827 /*
 828  * This routine traverse page_cgroup in given list and drop them all.
 829  * This routine ignores page_cgroup->ref_cnt.
 830  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
 831  */
 832 #define FORCE_UNCHARGE_BATCH    (128)
 833 static void
 834 mem_cgroup_force_empty_list(struct mem_cgroup *mem,
 835                             struct mem_cgroup_per_zone *mz,
 836                             int active)
 837 {
 838         struct page_cgroup *pc;
 839         struct page *page;
 840         int count;
 841         unsigned long flags;
 842         struct list_head *list;
 843
 844         if (active)
 845                 list = &mz->active_list;
 846         else
 847                 list = &mz->inactive_list;
 848
 849         if (list_empty(list))
 850                 return;
 851 retry:
 852         count = FORCE_UNCHARGE_BATCH;
 853         spin_lock_irqsave(&mz->lru_lock, flags);
 854
 855         while (--count && !list_empty(list)) {
 856                 pc = list_entry(list->prev, struct page_cgroup, lru);
 857                 page = pc->page;
 858                 /* Avoid race with charge */
 859                 atomic_set(&pc->ref_cnt, 0);
 860                 if (clear_page_cgroup(page, pc) == pc) {
 861                         css_put(&mem->css);
 862                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 863                         __mem_cgroup_remove_list(pc);
 864                         kfree(pc);
 865                 } else  /* being uncharged ? ...do relax */
 866                         break;
 867         }
 868         spin_unlock_irqrestore(&mz->lru_lock, flags);
 869         if (!list_empty(list)) {
 870                 cond_resched();
 871                 goto retry;
 872         }
 873         return;
 874 }
 875
 876 /*
 877  * make mem_cgroup's charge to be 0 if there is no task.
 878  * This enables deleting this mem_cgroup.
 879  */
 880
 881 int mem_cgroup_force_empty(struct mem_cgroup *mem)
 882 {
 883         int ret = -EBUSY;
 884         int node, zid;
 885         css_get(&mem->css);
 886         /*
 887          * page reclaim code (kswapd etc..) will move pages between
 888 `        * active_list <-> inactive_list while we don't take a lock.
 889          * So, we have to do loop here until all lists are empty.
 890          */
 891         while (mem->res.usage > 0) {
 892                 if (atomic_read(&mem->css.cgroup->count) > 0)
 893                         goto out;
 894                 for_each_node_state(node, N_POSSIBLE)
 895                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 896                                 struct mem_cgroup_per_zone *mz;
 897                                 mz = mem_cgroup_zoneinfo(mem, node, zid);
 898                                 /* drop all page_cgroup in active_list */
 899                                 mem_cgroup_force_empty_list(mem, mz, 1);
 900                                 /* drop all page_cgroup in inactive_list */
 901                                 mem_cgroup_force_empty_list(mem, mz, 0);
 902                         }
 903         }
 904         ret = 0;
 905 out:
 906         css_put(&mem->css);
 907         return ret;
 908 }
 909
 910
 911
 912 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 913 {
 914         *tmp = memparse(buf, &buf);
 915         if (*buf != '\0')
 916                 return -EINVAL;
 917
 918         /*
 919          * Round up the value to the closest page size
 920          */
 921         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 922         return 0;
 923 }
 924
 925 static ssize_t mem_cgroup_read(struct cgroup *cont,
 926                         struct cftype *cft, struct file *file,
 927                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 928 {
 929         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 930                                 cft->private, userbuf, nbytes, ppos,
 931                                 NULL);
 932 }
 933
 934 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 935                                 struct file *file, const char __user *userbuf,
 936                                 size_t nbytes, loff_t *ppos)
 937 {
 938         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 939                                 cft->private, userbuf, nbytes, ppos,
 940                                 mem_cgroup_write_strategy);
 941 }
 942
 943 static ssize_t mem_force_empty_write(struct cgroup *cont,
 944                                 struct cftype *cft, struct file *file,
 945                                 const char __user *userbuf,
 946                                 size_t nbytes, loff_t *ppos)
 947 {
 948         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
 949         int ret;
 950         ret = mem_cgroup_force_empty(mem);
 951         if (!ret)
 952                 ret = nbytes;
 953         return ret;
 954 }
 955
 956 /*
 957  * Note: This should be removed if cgroup supports write-only file.
 958  */
 959
 960 static ssize_t mem_force_empty_read(struct cgroup *cont,
 961                                 struct cftype *cft,
 962                                 struct file *file, char __user *userbuf,
 963                                 size_t nbytes, loff_t *ppos)
 964 {
 965         return -EINVAL;
 966 }
 967
 968
 969 static const struct mem_cgroup_stat_desc {
 970         const char *msg;
 971         u64 unit;
 972 } mem_cgroup_stat_desc[] = {
 973         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
 974         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
 975 };
 976
 977 static int mem_control_stat_show(struct seq_file *m, void *arg)
 978 {
 979         struct cgroup *cont = m->private;
 980         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
 981         struct mem_cgroup_stat *stat = &mem_cont->stat;
 982         int i;
 983
 984         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
 985                 s64 val;
 986
 987                 val = mem_cgroup_read_stat(stat, i);
 988                 val *= mem_cgroup_stat_desc[i].unit;
 989                 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
 990                                 (long long)val);
 991         }
 992         /* showing # of active pages */
 993         {
 994                 unsigned long active, inactive;
 995
 996                 inactive = mem_cgroup_get_all_zonestat(mem_cont,
 997                                                 MEM_CGROUP_ZSTAT_INACTIVE);
 998                 active = mem_cgroup_get_all_zonestat(mem_cont,
 999                                                 MEM_CGROUP_ZSTAT_ACTIVE);
1000                 seq_printf(m, "active %ld\n", (active) * PAGE_SIZE);
1001                 seq_printf(m, "inactive %ld\n", (inactive) * PAGE_SIZE);
1002         }
1003         return 0;
1004 }
1005
1006 static const struct file_operations mem_control_stat_file_operations = {
1007         .read = seq_read,
1008         .llseek = seq_lseek,
1009         .release = single_release,
1010 };
1011
1012 static int mem_control_stat_open(struct inode *unused, struct file *file)
1013 {
1014         /* XXX __d_cont */
1015         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
1016
1017         file->f_op = &mem_control_stat_file_operations;
1018         return single_open(file, mem_control_stat_show, cont);
1019 }
1020
1021
1022
1023 static struct cftype mem_cgroup_files[] = {
1024         {
1025                 .name = "usage_in_bytes",
1026                 .private = RES_USAGE,
1027                 .read = mem_cgroup_read,
1028         },
1029         {
1030                 .name = "limit_in_bytes",
1031                 .private = RES_LIMIT,
1032                 .write = mem_cgroup_write,
1033                 .read = mem_cgroup_read,
1034         },
1035         {
1036                 .name = "failcnt",
1037                 .private = RES_FAILCNT,
1038                 .read = mem_cgroup_read,
1039         },
1040         {
1041                 .name = "force_empty",
1042                 .write = mem_force_empty_write,
1043                 .read = mem_force_empty_read,
1044         },
1045         {
1046                 .name = "stat",
1047                 .open = mem_control_stat_open,
1048         },
1049 };
1050
1051 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1052 {
1053         struct mem_cgroup_per_node *pn;
1054         struct mem_cgroup_per_zone *mz;
1055         int zone;
1056         /*
1057          * This routine is called against possible nodes.
1058          * But it's BUG to call kmalloc() against offline node.
1059          *
1060          * TODO: this routine can waste much memory for nodes which will
1061          *       never be onlined. It's better to use memory hotplug callback
1062          *       function.
1063          */
1064         if (node_state(node, N_HIGH_MEMORY))
1065                 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, node);
1066         else
1067                 pn = kmalloc(sizeof(*pn), GFP_KERNEL);
1068         if (!pn)
1069                 return 1;
1070
1071         mem->info.nodeinfo[node] = pn;
1072         memset(pn, 0, sizeof(*pn));
1073
1074         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
1075                 mz = &pn->zoneinfo[zone];
1076                 INIT_LIST_HEAD(&mz->active_list);
1077                 INIT_LIST_HEAD(&mz->inactive_list);
1078                 spin_lock_init(&mz->lru_lock);
1079         }
1080         return 0;
1081 }
1082
1083 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1084 {
1085         kfree(mem->info.nodeinfo[node]);
1086 }
1087
1088
1089 static struct mem_cgroup init_mem_cgroup;
1090
1091 static struct cgroup_subsys_state *
1092 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1093 {
1094         struct mem_cgroup *mem;
1095         int node;
1096
1097         if (unlikely((cont->parent) == NULL)) {
1098                 mem = &init_mem_cgroup;
1099                 init_mm.mem_cgroup = mem;
1100         } else
1101                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
1102
1103         if (mem == NULL)
1104                 return NULL;
1105
1106         res_counter_init(&mem->res);
1107
1108         memset(&mem->info, 0, sizeof(mem->info));
1109
1110         for_each_node_state(node, N_POSSIBLE)
1111                 if (alloc_mem_cgroup_per_zone_info(mem, node))
1112                         goto free_out;
1113
1114         return &mem->css;
1115 free_out:
1116         for_each_node_state(node, N_POSSIBLE)
1117                 free_mem_cgroup_per_zone_info(mem, node);
1118         if (cont->parent != NULL)
1119                 kfree(mem);
1120         return NULL;
1121 }
1122
1123 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1124                                         struct cgroup *cont)
1125 {
1126         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1127         mem_cgroup_force_empty(mem);
1128 }
1129
1130 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
1131                                 struct cgroup *cont)
1132 {
1133         int node;
1134         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1135
1136         for_each_node_state(node, N_POSSIBLE)
1137                 free_mem_cgroup_per_zone_info(mem, node);
1138
1139         kfree(mem_cgroup_from_cont(cont));
1140 }
1141
1142 static int mem_cgroup_populate(struct cgroup_subsys *ss,
1143                                 struct cgroup *cont)
1144 {
1145         return cgroup_add_files(cont, ss, mem_cgroup_files,
1146                                         ARRAY_SIZE(mem_cgroup_files));
1147 }
1148
1149 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
1150                                 struct cgroup *cont,
1151                                 struct cgroup *old_cont,
1152                                 struct task_struct *p)
1153 {
1154         struct mm_struct *mm;
1155         struct mem_cgroup *mem, *old_mem;
1156
1157         mm = get_task_mm(p);
1158         if (mm == NULL)
1159                 return;
1160
1161         mem = mem_cgroup_from_cont(cont);
1162         old_mem = mem_cgroup_from_cont(old_cont);
1163
1164         if (mem == old_mem)
1165                 goto out;
1166
1167         /*
1168          * Only thread group leaders are allowed to migrate, the mm_struct is
1169          * in effect owned by the leader
1170          */
1171         if (p->tgid != p->pid)
1172                 goto out;
1173
1174         css_get(&mem->css);
1175         rcu_assign_pointer(mm->mem_cgroup, mem);
1176         css_put(&old_mem->css);
1177
1178 out:
1179         mmput(mm);
1180         return;
1181 }
1182
1183 struct cgroup_subsys mem_cgroup_subsys = {
1184         .name = "memory",
1185         .subsys_id = mem_cgroup_subsys_id,
1186         .create = mem_cgroup_create,
1187         .pre_destroy = mem_cgroup_pre_destroy,
1188         .destroy = mem_cgroup_destroy,
1189         .populate = mem_cgroup_populate,
1190         .attach = mem_cgroup_move_task,
1191         .early_init = 0,
1192 };