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