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/slab.h>
  30 #include <linux/swap.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/fs.h>
  33 #include <linux/seq_file.h>
  34 #include <linux/vmalloc.h>
  35
  36 #include <asm/uaccess.h>
  37
  38 struct cgroup_subsys mem_cgroup_subsys;
  39 static const int MEM_CGROUP_RECLAIM_RETRIES = 5;
  40 static struct kmem_cache *page_cgroup_cache;
  41
  42 /*
  43  * Statistics for memory cgroup.
  44  */
  45 enum mem_cgroup_stat_index {
  46         /*
  47          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  48          */
  49         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
  50         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
  51
  52         MEM_CGROUP_STAT_NSTATS,
  53 };
  54
  55 struct mem_cgroup_stat_cpu {
  56         s64 count[MEM_CGROUP_STAT_NSTATS];
  57 } ____cacheline_aligned_in_smp;
  58
  59 struct mem_cgroup_stat {
  60         struct mem_cgroup_stat_cpu cpustat[NR_CPUS];
  61 };
  62
  63 /*
  64  * For accounting under irq disable, no need for increment preempt count.
  65  */
  66 static void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat *stat,
  67                 enum mem_cgroup_stat_index idx, int val)
  68 {
  69         int cpu = smp_processor_id();
  70         stat->cpustat[cpu].count[idx] += val;
  71 }
  72
  73 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
  74                 enum mem_cgroup_stat_index idx)
  75 {
  76         int cpu;
  77         s64 ret = 0;
  78         for_each_possible_cpu(cpu)
  79                 ret += stat->cpustat[cpu].count[idx];
  80         return ret;
  81 }
  82
  83 /*
  84  * per-zone information in memory controller.
  85  */
  86
  87 enum mem_cgroup_zstat_index {
  88         MEM_CGROUP_ZSTAT_ACTIVE,
  89         MEM_CGROUP_ZSTAT_INACTIVE,
  90
  91         NR_MEM_CGROUP_ZSTAT,
  92 };
  93
  94 struct mem_cgroup_per_zone {
  95         /*
  96          * spin_lock to protect the per cgroup LRU
  97          */
  98         spinlock_t              lru_lock;
  99         struct list_head        active_list;
 100         struct list_head        inactive_list;
 101         unsigned long count[NR_MEM_CGROUP_ZSTAT];
 102 };
 103 /* Macro for accessing counter */
 104 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
 105
 106 struct mem_cgroup_per_node {
 107         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
 108 };
 109
 110 struct mem_cgroup_lru_info {
 111         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 112 };
 113
 114 /*
 115  * The memory controller data structure. The memory controller controls both
 116  * page cache and RSS per cgroup. We would eventually like to provide
 117  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 118  * to help the administrator determine what knobs to tune.
 119  *
 120  * TODO: Add a water mark for the memory controller. Reclaim will begin when
 121  * we hit the water mark. May be even add a low water mark, such that
 122  * no reclaim occurs from a cgroup at it's low water mark, this is
 123  * a feature that will be implemented much later in the future.
 124  */
 125 struct mem_cgroup {
 126         struct cgroup_subsys_state css;
 127         /*
 128          * the counter to account for memory usage
 129          */
 130         struct res_counter res;
 131         /*
 132          * Per cgroup active and inactive list, similar to the
 133          * per zone LRU lists.
 134          */
 135         struct mem_cgroup_lru_info info;
 136
 137         int     prev_priority;  /* for recording reclaim priority */
 138         /*
 139          * statistics.
 140          */
 141         struct mem_cgroup_stat stat;
 142 };
 143 static struct mem_cgroup init_mem_cgroup;
 144
 145 /*
 146  * We use the lower bit of the page->page_cgroup pointer as a bit spin
 147  * lock.  We need to ensure that page->page_cgroup is at least two
 148  * byte aligned (based on comments from Nick Piggin).  But since
 149  * bit_spin_lock doesn't actually set that lock bit in a non-debug
 150  * uniprocessor kernel, we should avoid setting it here too.
 151  */
 152 #define PAGE_CGROUP_LOCK_BIT    0x0
 153 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
 154 #define PAGE_CGROUP_LOCK        (1 << PAGE_CGROUP_LOCK_BIT)
 155 #else
 156 #define PAGE_CGROUP_LOCK        0x0
 157 #endif
 158
 159 /*
 160  * A page_cgroup page is associated with every page descriptor. The
 161  * page_cgroup helps us identify information about the cgroup
 162  */
 163 struct page_cgroup {
 164         struct list_head lru;           /* per cgroup LRU list */
 165         struct page *page;
 166         struct mem_cgroup *mem_cgroup;
 167         int ref_cnt;                    /* cached, mapped, migrating */
 168         int flags;
 169 };
 170 #define PAGE_CGROUP_FLAG_CACHE  (0x1)   /* charged as cache */
 171 #define PAGE_CGROUP_FLAG_ACTIVE (0x2)   /* page is active in this cgroup */
 172
 173 static int page_cgroup_nid(struct page_cgroup *pc)
 174 {
 175         return page_to_nid(pc->page);
 176 }
 177
 178 static enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 179 {
 180         return page_zonenum(pc->page);
 181 }
 182
 183 enum charge_type {
 184         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 185         MEM_CGROUP_CHARGE_TYPE_MAPPED,
 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
 197         VM_BUG_ON(!irqs_disabled());
 198         if (flags & PAGE_CGROUP_FLAG_CACHE)
 199                 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_CACHE, val);
 200         else
 201                 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
 202 }
 203
 204 static struct mem_cgroup_per_zone *
 205 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
 206 {
 207         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
 208 }
 209
 210 static struct mem_cgroup_per_zone *
 211 page_cgroup_zoneinfo(struct page_cgroup *pc)
 212 {
 213         struct mem_cgroup *mem = pc->mem_cgroup;
 214         int nid = page_cgroup_nid(pc);
 215         int zid = page_cgroup_zid(pc);
 216
 217         return mem_cgroup_zoneinfo(mem, nid, zid);
 218 }
 219
 220 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
 221                                         enum mem_cgroup_zstat_index idx)
 222 {
 223         int nid, zid;
 224         struct mem_cgroup_per_zone *mz;
 225         u64 total = 0;
 226
 227         for_each_online_node(nid)
 228                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 229                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
 230                         total += MEM_CGROUP_ZSTAT(mz, idx);
 231                 }
 232         return total;
 233 }
 234
 235 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 236 {
 237         return container_of(cgroup_subsys_state(cont,
 238                                 mem_cgroup_subsys_id), struct mem_cgroup,
 239                                 css);
 240 }
 241
 242 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 243 {
 244         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 245                                 struct mem_cgroup, css);
 246 }
 247
 248 static inline int page_cgroup_locked(struct page *page)
 249 {
 250         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 251 }
 252
 253 static void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 254 {
 255         VM_BUG_ON(!page_cgroup_locked(page));
 256         page->page_cgroup = ((unsigned long)pc | PAGE_CGROUP_LOCK);
 257 }
 258
 259 struct page_cgroup *page_get_page_cgroup(struct page *page)
 260 {
 261         return (struct page_cgroup *) (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 262 }
 263
 264 static void lock_page_cgroup(struct page *page)
 265 {
 266         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 267 }
 268
 269 static int try_lock_page_cgroup(struct page *page)
 270 {
 271         return bit_spin_trylock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 272 }
 273
 274 static void unlock_page_cgroup(struct page *page)
 275 {
 276         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 277 }
 278
 279 static void __mem_cgroup_remove_list(struct mem_cgroup_per_zone *mz,
 280                         struct page_cgroup *pc)
 281 {
 282         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 283
 284         if (from)
 285                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 286         else
 287                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 288
 289         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, false);
 290         list_del_init(&pc->lru);
 291 }
 292
 293 static void __mem_cgroup_add_list(struct mem_cgroup_per_zone *mz,
 294                                 struct page_cgroup *pc)
 295 {
 296         int to = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 297
 298         if (!to) {
 299                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 300                 list_add(&pc->lru, &mz->inactive_list);
 301         } else {
 302                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 303                 list_add(&pc->lru, &mz->active_list);
 304         }
 305         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, true);
 306 }
 307
 308 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 309 {
 310         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 311         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 312
 313         if (from)
 314                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 315         else
 316                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 317
 318         if (active) {
 319                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 320                 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
 321                 list_move(&pc->lru, &mz->active_list);
 322         } else {
 323                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 324                 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
 325                 list_move(&pc->lru, &mz->inactive_list);
 326         }
 327 }
 328
 329 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 330 {
 331         int ret;
 332
 333         task_lock(task);
 334         ret = task->mm && mm_match_cgroup(task->mm, mem);
 335         task_unlock(task);
 336         return ret;
 337 }
 338
 339 /*
 340  * This routine assumes that the appropriate zone's lru lock is already held
 341  */
 342 void mem_cgroup_move_lists(struct page *page, bool active)
 343 {
 344         struct page_cgroup *pc;
 345         struct mem_cgroup_per_zone *mz;
 346         unsigned long flags;
 347
 348         /*
 349          * We cannot lock_page_cgroup while holding zone's lru_lock,
 350          * because other holders of lock_page_cgroup can be interrupted
 351          * with an attempt to rotate_reclaimable_page.  But we cannot
 352          * safely get to page_cgroup without it, so just try_lock it:
 353          * mem_cgroup_isolate_pages allows for page left on wrong list.
 354          */
 355         if (!try_lock_page_cgroup(page))
 356                 return;
 357
 358         pc = page_get_page_cgroup(page);
 359         if (pc) {
 360                 mz = page_cgroup_zoneinfo(pc);
 361                 spin_lock_irqsave(&mz->lru_lock, flags);
 362                 __mem_cgroup_move_lists(pc, active);
 363                 spin_unlock_irqrestore(&mz->lru_lock, flags);
 364         }
 365         unlock_page_cgroup(page);
 366 }
 367
 368 /*
 369  * Calculate mapped_ratio under memory controller. This will be used in
 370  * vmscan.c for deteremining we have to reclaim mapped pages.
 371  */
 372 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
 373 {
 374         long total, rss;
 375
 376         /*
 377          * usage is recorded in bytes. But, here, we assume the number of
 378          * physical pages can be represented by "long" on any arch.
 379          */
 380         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
 381         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
 382         return (int)((rss * 100L) / total);
 383 }
 384
 385 /*
 386  * This function is called from vmscan.c. In page reclaiming loop. balance
 387  * between active and inactive list is calculated. For memory controller
 388  * page reclaiming, we should use using mem_cgroup's imbalance rather than
 389  * zone's global lru imbalance.
 390  */
 391 long mem_cgroup_reclaim_imbalance(struct mem_cgroup *mem)
 392 {
 393         unsigned long active, inactive;
 394         /* active and inactive are the number of pages. 'long' is ok.*/
 395         active = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_ACTIVE);
 396         inactive = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_INACTIVE);
 397         return (long) (active / (inactive + 1));
 398 }
 399
 400 /*
 401  * prev_priority control...this will be used in memory reclaim path.
 402  */
 403 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
 404 {
 405         return mem->prev_priority;
 406 }
 407
 408 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
 409 {
 410         if (priority < mem->prev_priority)
 411                 mem->prev_priority = priority;
 412 }
 413
 414 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
 415 {
 416         mem->prev_priority = priority;
 417 }
 418
 419 /*
 420  * Calculate # of pages to be scanned in this priority/zone.
 421  * See also vmscan.c
 422  *
 423  * priority starts from "DEF_PRIORITY" and decremented in each loop.
 424  * (see include/linux/mmzone.h)
 425  */
 426
 427 long mem_cgroup_calc_reclaim_active(struct mem_cgroup *mem,
 428                                    struct zone *zone, int priority)
 429 {
 430         long nr_active;
 431         int nid = zone->zone_pgdat->node_id;
 432         int zid = zone_idx(zone);
 433         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
 434
 435         nr_active = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE);
 436         return (nr_active >> priority);
 437 }
 438
 439 long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup *mem,
 440                                         struct zone *zone, int priority)
 441 {
 442         long nr_inactive;
 443         int nid = zone->zone_pgdat->node_id;
 444         int zid = zone_idx(zone);
 445         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
 446
 447         nr_inactive = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE);
 448         return (nr_inactive >> priority);
 449 }
 450
 451 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 452                                         struct list_head *dst,
 453                                         unsigned long *scanned, int order,
 454                                         int mode, struct zone *z,
 455                                         struct mem_cgroup *mem_cont,
 456                                         int active)
 457 {
 458         unsigned long nr_taken = 0;
 459         struct page *page;
 460         unsigned long scan;
 461         LIST_HEAD(pc_list);
 462         struct list_head *src;
 463         struct page_cgroup *pc, *tmp;
 464         int nid = z->zone_pgdat->node_id;
 465         int zid = zone_idx(z);
 466         struct mem_cgroup_per_zone *mz;
 467
 468         BUG_ON(!mem_cont);
 469         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
 470         if (active)
 471                 src = &mz->active_list;
 472         else
 473                 src = &mz->inactive_list;
 474
 475
 476         spin_lock(&mz->lru_lock);
 477         scan = 0;
 478         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 479                 if (scan >= nr_to_scan)
 480                         break;
 481                 page = pc->page;
 482
 483                 if (unlikely(!PageLRU(page)))
 484                         continue;
 485
 486                 if (PageActive(page) && !active) {
 487                         __mem_cgroup_move_lists(pc, true);
 488                         continue;
 489                 }
 490                 if (!PageActive(page) && active) {
 491                         __mem_cgroup_move_lists(pc, false);
 492                         continue;
 493                 }
 494
 495                 scan++;
 496                 list_move(&pc->lru, &pc_list);
 497
 498                 if (__isolate_lru_page(page, mode) == 0) {
 499                         list_move(&page->lru, dst);
 500                         nr_taken++;
 501                 }
 502         }
 503
 504         list_splice(&pc_list, src);
 505         spin_unlock(&mz->lru_lock);
 506
 507         *scanned = scan;
 508         return nr_taken;
 509 }
 510
 511 /*
 512  * Charge the memory controller for page usage.
 513  * Return
 514  * 0 if the charge was successful
 515  * < 0 if the cgroup is over its limit
 516  */
 517 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 518                                 gfp_t gfp_mask, enum charge_type ctype)
 519 {
 520         struct mem_cgroup *mem;
 521         struct page_cgroup *pc;
 522         unsigned long flags;
 523         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 524         struct mem_cgroup_per_zone *mz;
 525
 526         if (mem_cgroup_subsys.disabled)
 527                 return 0;
 528
 529         /*
 530          * Should page_cgroup's go to their own slab?
 531          * One could optimize the performance of the charging routine
 532          * by saving a bit in the page_flags and using it as a lock
 533          * to see if the cgroup page already has a page_cgroup associated
 534          * with it
 535          */
 536 retry:
 537         lock_page_cgroup(page);
 538         pc = page_get_page_cgroup(page);
 539         /*
 540          * The page_cgroup exists and
 541          * the page has already been accounted.
 542          */
 543         if (pc) {
 544                 VM_BUG_ON(pc->page != page);
 545                 VM_BUG_ON(pc->ref_cnt <= 0);
 546
 547                 pc->ref_cnt++;
 548                 unlock_page_cgroup(page);
 549                 goto done;
 550         }
 551         unlock_page_cgroup(page);
 552
 553         pc = kmem_cache_zalloc(page_cgroup_cache, gfp_mask);
 554         if (pc == NULL)
 555                 goto err;
 556
 557         /*
 558          * We always charge the cgroup the mm_struct belongs to.
 559          * The mm_struct's mem_cgroup changes on task migration if the
 560          * thread group leader migrates. It's possible that mm is not
 561          * set, if so charge the init_mm (happens for pagecache usage).
 562          */
 563         if (!mm)
 564                 mm = &init_mm;
 565
 566         rcu_read_lock();
 567         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
 568         /*
 569          * For every charge from the cgroup, increment reference count
 570          */
 571         css_get(&mem->css);
 572         rcu_read_unlock();
 573
 574         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 575                 if (!(gfp_mask & __GFP_WAIT))
 576                         goto out;
 577
 578                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 579                         continue;
 580
 581                 /*
 582                  * try_to_free_mem_cgroup_pages() might not give us a full
 583                  * picture of reclaim. Some pages are reclaimed and might be
 584                  * moved to swap cache or just unmapped from the cgroup.
 585                  * Check the limit again to see if the reclaim reduced the
 586                  * current usage of the cgroup before giving up
 587                  */
 588                 if (res_counter_check_under_limit(&mem->res))
 589                         continue;
 590
 591                 if (!nr_retries--) {
 592                         mem_cgroup_out_of_memory(mem, gfp_mask);
 593                         goto out;
 594                 }
 595         }
 596
 597         pc->ref_cnt = 1;
 598         pc->mem_cgroup = mem;
 599         pc->page = page;
 600         pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
 601         if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
 602                 pc->flags = PAGE_CGROUP_FLAG_CACHE;
 603
 604         lock_page_cgroup(page);
 605         if (page_get_page_cgroup(page)) {
 606                 unlock_page_cgroup(page);
 607                 /*
 608                  * Another charge has been added to this page already.
 609                  * We take lock_page_cgroup(page) again and read
 610                  * page->cgroup, increment refcnt.... just retry is OK.
 611                  */
 612                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 613                 css_put(&mem->css);
 614                 kmem_cache_free(page_cgroup_cache, pc);
 615                 goto retry;
 616         }
 617         page_assign_page_cgroup(page, pc);
 618
 619         mz = page_cgroup_zoneinfo(pc);
 620         spin_lock_irqsave(&mz->lru_lock, flags);
 621         __mem_cgroup_add_list(mz, pc);
 622         spin_unlock_irqrestore(&mz->lru_lock, flags);
 623
 624         unlock_page_cgroup(page);
 625 done:
 626         return 0;
 627 out:
 628         css_put(&mem->css);
 629         kmem_cache_free(page_cgroup_cache, pc);
 630 err:
 631         return -ENOMEM;
 632 }
 633
 634 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
 635 {
 636         return mem_cgroup_charge_common(page, mm, gfp_mask,
 637                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
 638 }
 639
 640 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 641                                 gfp_t gfp_mask)
 642 {
 643         if (!mm)
 644                 mm = &init_mm;
 645         return mem_cgroup_charge_common(page, mm, gfp_mask,
 646                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
 647 }
 648
 649 /*
 650  * Uncharging is always a welcome operation, we never complain, simply
 651  * uncharge.
 652  */
 653 void mem_cgroup_uncharge_page(struct page *page)
 654 {
 655         struct page_cgroup *pc;
 656         struct mem_cgroup *mem;
 657         struct mem_cgroup_per_zone *mz;
 658         unsigned long flags;
 659
 660         if (mem_cgroup_subsys.disabled)
 661                 return;
 662
 663         /*
 664          * Check if our page_cgroup is valid
 665          */
 666         lock_page_cgroup(page);
 667         pc = page_get_page_cgroup(page);
 668         if (!pc)
 669                 goto unlock;
 670
 671         VM_BUG_ON(pc->page != page);
 672         VM_BUG_ON(pc->ref_cnt <= 0);
 673
 674         if (--(pc->ref_cnt) == 0) {
 675                 mz = page_cgroup_zoneinfo(pc);
 676                 spin_lock_irqsave(&mz->lru_lock, flags);
 677                 __mem_cgroup_remove_list(mz, pc);
 678                 spin_unlock_irqrestore(&mz->lru_lock, flags);
 679
 680                 page_assign_page_cgroup(page, NULL);
 681                 unlock_page_cgroup(page);
 682
 683                 mem = pc->mem_cgroup;
 684                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 685                 css_put(&mem->css);
 686
 687                 kmem_cache_free(page_cgroup_cache, pc);
 688                 return;
 689         }
 690
 691 unlock:
 692         unlock_page_cgroup(page);
 693 }
 694
 695 /*
 696  * Returns non-zero if a page (under migration) has valid page_cgroup member.
 697  * Refcnt of page_cgroup is incremented.
 698  */
 699 int mem_cgroup_prepare_migration(struct page *page)
 700 {
 701         struct page_cgroup *pc;
 702
 703         if (mem_cgroup_subsys.disabled)
 704                 return 0;
 705
 706         lock_page_cgroup(page);
 707         pc = page_get_page_cgroup(page);
 708         if (pc)
 709                 pc->ref_cnt++;
 710         unlock_page_cgroup(page);
 711         return pc != NULL;
 712 }
 713
 714 void mem_cgroup_end_migration(struct page *page)
 715 {
 716         mem_cgroup_uncharge_page(page);
 717 }
 718
 719 /*
 720  * We know both *page* and *newpage* are now not-on-LRU and PG_locked.
 721  * And no race with uncharge() routines because page_cgroup for *page*
 722  * has extra one reference by mem_cgroup_prepare_migration.
 723  */
 724 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
 725 {
 726         struct page_cgroup *pc;
 727         struct mem_cgroup_per_zone *mz;
 728         unsigned long flags;
 729
 730         lock_page_cgroup(page);
 731         pc = page_get_page_cgroup(page);
 732         if (!pc) {
 733                 unlock_page_cgroup(page);
 734                 return;
 735         }
 736
 737         mz = page_cgroup_zoneinfo(pc);
 738         spin_lock_irqsave(&mz->lru_lock, flags);
 739         __mem_cgroup_remove_list(mz, pc);
 740         spin_unlock_irqrestore(&mz->lru_lock, flags);
 741
 742         page_assign_page_cgroup(page, NULL);
 743         unlock_page_cgroup(page);
 744
 745         pc->page = newpage;
 746         lock_page_cgroup(newpage);
 747         page_assign_page_cgroup(newpage, pc);
 748
 749         mz = page_cgroup_zoneinfo(pc);
 750         spin_lock_irqsave(&mz->lru_lock, flags);
 751         __mem_cgroup_add_list(mz, pc);
 752         spin_unlock_irqrestore(&mz->lru_lock, flags);
 753
 754         unlock_page_cgroup(newpage);
 755 }
 756
 757 /*
 758  * This routine traverse page_cgroup in given list and drop them all.
 759  * This routine ignores page_cgroup->ref_cnt.
 760  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
 761  */
 762 #define FORCE_UNCHARGE_BATCH    (128)
 763 static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
 764                             struct mem_cgroup_per_zone *mz,
 765                             int active)
 766 {
 767         struct page_cgroup *pc;
 768         struct page *page;
 769         int count = FORCE_UNCHARGE_BATCH;
 770         unsigned long flags;
 771         struct list_head *list;
 772
 773         if (active)
 774                 list = &mz->active_list;
 775         else
 776                 list = &mz->inactive_list;
 777
 778         spin_lock_irqsave(&mz->lru_lock, flags);
 779         while (!list_empty(list)) {
 780                 pc = list_entry(list->prev, struct page_cgroup, lru);
 781                 page = pc->page;
 782                 get_page(page);
 783                 spin_unlock_irqrestore(&mz->lru_lock, flags);
 784                 mem_cgroup_uncharge_page(page);
 785                 put_page(page);
 786                 if (--count <= 0) {
 787                         count = FORCE_UNCHARGE_BATCH;
 788                         cond_resched();
 789                 }
 790                 spin_lock_irqsave(&mz->lru_lock, flags);
 791         }
 792         spin_unlock_irqrestore(&mz->lru_lock, flags);
 793 }
 794
 795 /*
 796  * make mem_cgroup's charge to be 0 if there is no task.
 797  * This enables deleting this mem_cgroup.
 798  */
 799 static int mem_cgroup_force_empty(struct mem_cgroup *mem)
 800 {
 801         int ret = -EBUSY;
 802         int node, zid;
 803
 804         if (mem_cgroup_subsys.disabled)
 805                 return 0;
 806
 807         css_get(&mem->css);
 808         /*
 809          * page reclaim code (kswapd etc..) will move pages between
 810          * active_list <-> inactive_list while we don't take a lock.
 811          * So, we have to do loop here until all lists are empty.
 812          */
 813         while (mem->res.usage > 0) {
 814                 if (atomic_read(&mem->css.cgroup->count) > 0)
 815                         goto out;
 816                 for_each_node_state(node, N_POSSIBLE)
 817                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 818                                 struct mem_cgroup_per_zone *mz;
 819                                 mz = mem_cgroup_zoneinfo(mem, node, zid);
 820                                 /* drop all page_cgroup in active_list */
 821                                 mem_cgroup_force_empty_list(mem, mz, 1);
 822                                 /* drop all page_cgroup in inactive_list */
 823                                 mem_cgroup_force_empty_list(mem, mz, 0);
 824                         }
 825         }
 826         ret = 0;
 827 out:
 828         css_put(&mem->css);
 829         return ret;
 830 }
 831
 832 static int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 833 {
 834         *tmp = memparse(buf, &buf);
 835         if (*buf != '\0')
 836                 return -EINVAL;
 837
 838         /*
 839          * Round up the value to the closest page size
 840          */
 841         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 842         return 0;
 843 }
 844
 845 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
 846 {
 847         return res_counter_read_u64(&mem_cgroup_from_cont(cont)->res,
 848                                     cft->private);
 849 }
 850
 851 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 852                                 struct file *file, const char __user *userbuf,
 853                                 size_t nbytes, loff_t *ppos)
 854 {
 855         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 856                                 cft->private, userbuf, nbytes, ppos,
 857                                 mem_cgroup_write_strategy);
 858 }
 859
 860 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
 861 {
 862         struct mem_cgroup *mem;
 863
 864         mem = mem_cgroup_from_cont(cont);
 865         switch (event) {
 866         case RES_MAX_USAGE:
 867                 res_counter_reset_max(&mem->res);
 868                 break;
 869         case RES_FAILCNT:
 870                 res_counter_reset_failcnt(&mem->res);
 871                 break;
 872         }
 873         return 0;
 874 }
 875
 876 static int mem_force_empty_write(struct cgroup *cont, unsigned int event)
 877 {
 878         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont));
 879 }
 880
 881 static const struct mem_cgroup_stat_desc {
 882         const char *msg;
 883         u64 unit;
 884 } mem_cgroup_stat_desc[] = {
 885         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
 886         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
 887 };
 888
 889 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
 890                                  struct cgroup_map_cb *cb)
 891 {
 892         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
 893         struct mem_cgroup_stat *stat = &mem_cont->stat;
 894         int i;
 895
 896         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
 897                 s64 val;
 898
 899                 val = mem_cgroup_read_stat(stat, i);
 900                 val *= mem_cgroup_stat_desc[i].unit;
 901                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
 902         }
 903         /* showing # of active pages */
 904         {
 905                 unsigned long active, inactive;
 906
 907                 inactive = mem_cgroup_get_all_zonestat(mem_cont,
 908                                                 MEM_CGROUP_ZSTAT_INACTIVE);
 909                 active = mem_cgroup_get_all_zonestat(mem_cont,
 910                                                 MEM_CGROUP_ZSTAT_ACTIVE);
 911                 cb->fill(cb, "active", (active) * PAGE_SIZE);
 912                 cb->fill(cb, "inactive", (inactive) * PAGE_SIZE);
 913         }
 914         return 0;
 915 }
 916
 917 static struct cftype mem_cgroup_files[] = {
 918         {
 919                 .name = "usage_in_bytes",
 920                 .private = RES_USAGE,
 921                 .read_u64 = mem_cgroup_read,
 922         },
 923         {
 924                 .name = "max_usage_in_bytes",
 925                 .private = RES_MAX_USAGE,
 926                 .trigger = mem_cgroup_reset,
 927                 .read_u64 = mem_cgroup_read,
 928         },
 929         {
 930                 .name = "limit_in_bytes",
 931                 .private = RES_LIMIT,
 932                 .write = mem_cgroup_write,
 933                 .read_u64 = mem_cgroup_read,
 934         },
 935         {
 936                 .name = "failcnt",
 937                 .private = RES_FAILCNT,
 938                 .trigger = mem_cgroup_reset,
 939                 .read_u64 = mem_cgroup_read,
 940         },
 941         {
 942                 .name = "force_empty",
 943                 .trigger = mem_force_empty_write,
 944         },
 945         {
 946                 .name = "stat",
 947                 .read_map = mem_control_stat_show,
 948         },
 949 };
 950
 951 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
 952 {
 953         struct mem_cgroup_per_node *pn;
 954         struct mem_cgroup_per_zone *mz;
 955         int zone, tmp = node;
 956         /*
 957          * This routine is called against possible nodes.
 958          * But it's BUG to call kmalloc() against offline node.
 959          *
 960          * TODO: this routine can waste much memory for nodes which will
 961          *       never be onlined. It's better to use memory hotplug callback
 962          *       function.
 963          */
 964         if (!node_state(node, N_NORMAL_MEMORY))
 965                 tmp = -1;
 966         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
 967         if (!pn)
 968                 return 1;
 969
 970         mem->info.nodeinfo[node] = pn;
 971         memset(pn, 0, sizeof(*pn));
 972
 973         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
 974                 mz = &pn->zoneinfo[zone];
 975                 INIT_LIST_HEAD(&mz->active_list);
 976                 INIT_LIST_HEAD(&mz->inactive_list);
 977                 spin_lock_init(&mz->lru_lock);
 978         }
 979         return 0;
 980 }
 981
 982 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
 983 {
 984         kfree(mem->info.nodeinfo[node]);
 985 }
 986
 987 static struct mem_cgroup *mem_cgroup_alloc(void)
 988 {
 989         struct mem_cgroup *mem;
 990
 991         if (sizeof(*mem) < PAGE_SIZE)
 992                 mem = kmalloc(sizeof(*mem), GFP_KERNEL);
 993         else
 994                 mem = vmalloc(sizeof(*mem));
 995
 996         if (mem)
 997                 memset(mem, 0, sizeof(*mem));
 998         return mem;
 999 }
1000
1001 static void mem_cgroup_free(struct mem_cgroup *mem)
1002 {
1003         if (sizeof(*mem) < PAGE_SIZE)
1004                 kfree(mem);
1005         else
1006                 vfree(mem);
1007 }
1008
1009
1010 static struct cgroup_subsys_state *
1011 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1012 {
1013         struct mem_cgroup *mem;
1014         int node;
1015
1016         if (unlikely((cont->parent) == NULL)) {
1017                 mem = &init_mem_cgroup;
1018                 page_cgroup_cache = KMEM_CACHE(page_cgroup, SLAB_PANIC);
1019         } else {
1020                 mem = mem_cgroup_alloc();
1021                 if (!mem)
1022                         return ERR_PTR(-ENOMEM);
1023         }
1024
1025         res_counter_init(&mem->res);
1026
1027         for_each_node_state(node, N_POSSIBLE)
1028                 if (alloc_mem_cgroup_per_zone_info(mem, node))
1029                         goto free_out;
1030
1031         return &mem->css;
1032 free_out:
1033         for_each_node_state(node, N_POSSIBLE)
1034                 free_mem_cgroup_per_zone_info(mem, node);
1035         if (cont->parent != NULL)
1036                 mem_cgroup_free(mem);
1037         return ERR_PTR(-ENOMEM);
1038 }
1039
1040 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1041                                         struct cgroup *cont)
1042 {
1043         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1044         mem_cgroup_force_empty(mem);
1045 }
1046
1047 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
1048                                 struct cgroup *cont)
1049 {
1050         int node;
1051         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1052
1053         for_each_node_state(node, N_POSSIBLE)
1054                 free_mem_cgroup_per_zone_info(mem, node);
1055
1056         mem_cgroup_free(mem_cgroup_from_cont(cont));
1057 }
1058
1059 static int mem_cgroup_populate(struct cgroup_subsys *ss,
1060                                 struct cgroup *cont)
1061 {
1062         if (mem_cgroup_subsys.disabled)
1063                 return 0;
1064         return cgroup_add_files(cont, ss, mem_cgroup_files,
1065                                         ARRAY_SIZE(mem_cgroup_files));
1066 }
1067
1068 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
1069                                 struct cgroup *cont,
1070                                 struct cgroup *old_cont,
1071                                 struct task_struct *p)
1072 {
1073         struct mm_struct *mm;
1074         struct mem_cgroup *mem, *old_mem;
1075
1076         if (mem_cgroup_subsys.disabled)
1077                 return;
1078
1079         mm = get_task_mm(p);
1080         if (mm == NULL)
1081                 return;
1082
1083         mem = mem_cgroup_from_cont(cont);
1084         old_mem = mem_cgroup_from_cont(old_cont);
1085
1086         if (mem == old_mem)
1087                 goto out;
1088
1089         /*
1090          * Only thread group leaders are allowed to migrate, the mm_struct is
1091          * in effect owned by the leader
1092          */
1093         if (!thread_group_leader(p))
1094                 goto out;
1095
1096 out:
1097         mmput(mm);
1098 }
1099
1100 struct cgroup_subsys mem_cgroup_subsys = {
1101         .name = "memory",
1102         .subsys_id = mem_cgroup_subsys_id,
1103         .create = mem_cgroup_create,
1104         .pre_destroy = mem_cgroup_pre_destroy,
1105         .destroy = mem_cgroup_destroy,
1106         .populate = mem_cgroup_populate,
1107         .attach = mem_cgroup_move_task,
1108         .early_init = 0,
1109 };