mm/memcontrol.c

   1 /* memcontrol.c - Memory Controller
   2  *
   3  * Copyright IBM Corporation, 2007
   4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
   5  *
   6  * Copyright 2007 OpenVZ SWsoft Inc
   7  * Author: Pavel Emelianov <xemul@openvz.org>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19
  20 #include <linux/res_counter.h>
  21 #include <linux/memcontrol.h>
  22 #include <linux/cgroup.h>
  23 #include <linux/mm.h>
  24 #include <linux/smp.h>
  25 #include <linux/page-flags.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/bit_spinlock.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/swap.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/fs.h>
  32 #include <linux/seq_file.h>
  33
  34 #include <asm/uaccess.h>
  35
  36 struct cgroup_subsys mem_cgroup_subsys;
  37 static const int MEM_CGROUP_RECLAIM_RETRIES = 5;
  38
  39 /*
  40  * Statistics for memory cgroup.
  41  */
  42 enum mem_cgroup_stat_index {
  43         /*
  44          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
  45          */
  46         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
  47         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
  48
  49         MEM_CGROUP_STAT_NSTATS,
  50 };
  51
  52 struct mem_cgroup_stat_cpu {
  53         s64 count[MEM_CGROUP_STAT_NSTATS];
  54 } ____cacheline_aligned_in_smp;
  55
  56 struct mem_cgroup_stat {
  57         struct mem_cgroup_stat_cpu cpustat[NR_CPUS];
  58 };
  59
  60 /*
  61  * For accounting under irq disable, no need for increment preempt count.
  62  */
  63 static void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat *stat,
  64                 enum mem_cgroup_stat_index idx, int val)
  65 {
  66         int cpu = smp_processor_id();
  67         stat->cpustat[cpu].count[idx] += val;
  68 }
  69
  70 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
  71                 enum mem_cgroup_stat_index idx)
  72 {
  73         int cpu;
  74         s64 ret = 0;
  75         for_each_possible_cpu(cpu)
  76                 ret += stat->cpustat[cpu].count[idx];
  77         return ret;
  78 }
  79
  80 /*
  81  * per-zone information in memory controller.
  82  */
  83
  84 enum mem_cgroup_zstat_index {
  85         MEM_CGROUP_ZSTAT_ACTIVE,
  86         MEM_CGROUP_ZSTAT_INACTIVE,
  87
  88         NR_MEM_CGROUP_ZSTAT,
  89 };
  90
  91 struct mem_cgroup_per_zone {
  92         unsigned long count[NR_MEM_CGROUP_ZSTAT];
  93 };
  94 /* Macro for accessing counter */
  95 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
  96
  97 struct mem_cgroup_per_node {
  98         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
  99 };
 100
 101 struct mem_cgroup_lru_info {
 102         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
 103 };
 104
 105 /*
 106  * The memory controller data structure. The memory controller controls both
 107  * page cache and RSS per cgroup. We would eventually like to provide
 108  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
 109  * to help the administrator determine what knobs to tune.
 110  *
 111  * TODO: Add a water mark for the memory controller. Reclaim will begin when
 112  * we hit the water mark. May be even add a low water mark, such that
 113  * no reclaim occurs from a cgroup at it's low water mark, this is
 114  * a feature that will be implemented much later in the future.
 115  */
 116 struct mem_cgroup {
 117         struct cgroup_subsys_state css;
 118         /*
 119          * the counter to account for memory usage
 120          */
 121         struct res_counter res;
 122         /*
 123          * Per cgroup active and inactive list, similar to the
 124          * per zone LRU lists.
 125          * TODO: Consider making these lists per zone
 126          */
 127         struct list_head active_list;
 128         struct list_head inactive_list;
 129         struct mem_cgroup_lru_info info;
 130         /*
 131          * spin_lock to protect the per cgroup LRU
 132          */
 133         spinlock_t lru_lock;
 134         unsigned long control_type;     /* control RSS or RSS+Pagecache */
 135         /*
 136          * statistics.
 137          */
 138         struct mem_cgroup_stat stat;
 139 };
 140
 141 /*
 142  * We use the lower bit of the page->page_cgroup pointer as a bit spin
 143  * lock. We need to ensure that page->page_cgroup is atleast two
 144  * byte aligned (based on comments from Nick Piggin)
 145  */
 146 #define PAGE_CGROUP_LOCK_BIT    0x0
 147 #define PAGE_CGROUP_LOCK                (1 << PAGE_CGROUP_LOCK_BIT)
 148
 149 /*
 150  * A page_cgroup page is associated with every page descriptor. The
 151  * page_cgroup helps us identify information about the cgroup
 152  */
 153 struct page_cgroup {
 154         struct list_head lru;           /* per cgroup LRU list */
 155         struct page *page;
 156         struct mem_cgroup *mem_cgroup;
 157         atomic_t ref_cnt;               /* Helpful when pages move b/w  */
 158                                         /* mapped and cached states     */
 159         int      flags;
 160 };
 161 #define PAGE_CGROUP_FLAG_CACHE  (0x1)   /* charged as cache */
 162 #define PAGE_CGROUP_FLAG_ACTIVE (0x2)   /* page is active in this cgroup */
 163
 164 static inline int page_cgroup_nid(struct page_cgroup *pc)
 165 {
 166         return page_to_nid(pc->page);
 167 }
 168
 169 static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 170 {
 171         return page_zonenum(pc->page);
 172 }
 173
 174 enum {
 175         MEM_CGROUP_TYPE_UNSPEC = 0,
 176         MEM_CGROUP_TYPE_MAPPED,
 177         MEM_CGROUP_TYPE_CACHED,
 178         MEM_CGROUP_TYPE_ALL,
 179         MEM_CGROUP_TYPE_MAX,
 180 };
 181
 182 enum charge_type {
 183         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 184         MEM_CGROUP_CHARGE_TYPE_MAPPED,
 185 };
 186
 187
 188 /*
 189  * Always modified under lru lock. Then, not necessary to preempt_disable()
 190  */
 191 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
 192                                         bool charge)
 193 {
 194         int val = (charge)? 1 : -1;
 195         struct mem_cgroup_stat *stat = &mem->stat;
 196         VM_BUG_ON(!irqs_disabled());
 197
 198         if (flags & PAGE_CGROUP_FLAG_CACHE)
 199                 __mem_cgroup_stat_add_safe(stat,
 200                                         MEM_CGROUP_STAT_CACHE, val);
 201         else
 202                 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
 203 }
 204
 205 static inline struct mem_cgroup_per_zone *
 206 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
 207 {
 208         BUG_ON(!mem->info.nodeinfo[nid]);
 209         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
 210 }
 211
 212 static inline struct mem_cgroup_per_zone *
 213 page_cgroup_zoneinfo(struct page_cgroup *pc)
 214 {
 215         struct mem_cgroup *mem = pc->mem_cgroup;
 216         int nid = page_cgroup_nid(pc);
 217         int zid = page_cgroup_zid(pc);
 218
 219         return mem_cgroup_zoneinfo(mem, nid, zid);
 220 }
 221
 222 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
 223                                         enum mem_cgroup_zstat_index idx)
 224 {
 225         int nid, zid;
 226         struct mem_cgroup_per_zone *mz;
 227         u64 total = 0;
 228
 229         for_each_online_node(nid)
 230                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
 231                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
 232                         total += MEM_CGROUP_ZSTAT(mz, idx);
 233                 }
 234         return total;
 235 }
 236
 237 static struct mem_cgroup init_mem_cgroup;
 238
 239 static inline
 240 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 241 {
 242         return container_of(cgroup_subsys_state(cont,
 243                                 mem_cgroup_subsys_id), struct mem_cgroup,
 244                                 css);
 245 }
 246
 247 static inline
 248 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 249 {
 250         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 251                                 struct mem_cgroup, css);
 252 }
 253
 254 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
 255 {
 256         struct mem_cgroup *mem;
 257
 258         mem = mem_cgroup_from_task(p);
 259         css_get(&mem->css);
 260         mm->mem_cgroup = mem;
 261 }
 262
 263 void mm_free_cgroup(struct mm_struct *mm)
 264 {
 265         css_put(&mm->mem_cgroup->css);
 266 }
 267
 268 static inline int page_cgroup_locked(struct page *page)
 269 {
 270         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
 271                                         &page->page_cgroup);
 272 }
 273
 274 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 275 {
 276         int locked;
 277
 278         /*
 279          * While resetting the page_cgroup we might not hold the
 280          * page_cgroup lock. free_hot_cold_page() is an example
 281          * of such a scenario
 282          */
 283         if (pc)
 284                 VM_BUG_ON(!page_cgroup_locked(page));
 285         locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
 286         page->page_cgroup = ((unsigned long)pc | locked);
 287 }
 288
 289 struct page_cgroup *page_get_page_cgroup(struct page *page)
 290 {
 291         return (struct page_cgroup *)
 292                 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 293 }
 294
 295 static void __always_inline lock_page_cgroup(struct page *page)
 296 {
 297         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 298         VM_BUG_ON(!page_cgroup_locked(page));
 299 }
 300
 301 static void __always_inline unlock_page_cgroup(struct page *page)
 302 {
 303         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 304 }
 305
 306 /*
 307  * Tie new page_cgroup to struct page under lock_page_cgroup()
 308  * This can fail if the page has been tied to a page_cgroup.
 309  * If success, returns 0.
 310  */
 311 static int page_cgroup_assign_new_page_cgroup(struct page *page,
 312                                                 struct page_cgroup *pc)
 313 {
 314         int ret = 0;
 315
 316         lock_page_cgroup(page);
 317         if (!page_get_page_cgroup(page))
 318                 page_assign_page_cgroup(page, pc);
 319         else /* A page is tied to other pc. */
 320                 ret = 1;
 321         unlock_page_cgroup(page);
 322         return ret;
 323 }
 324
 325 /*
 326  * Clear page->page_cgroup member under lock_page_cgroup().
 327  * If given "pc" value is different from one page->page_cgroup,
 328  * page->cgroup is not cleared.
 329  * Returns a value of page->page_cgroup at lock taken.
 330  * A can can detect failure of clearing by following
 331  *  clear_page_cgroup(page, pc) == pc
 332  */
 333
 334 static struct page_cgroup *clear_page_cgroup(struct page *page,
 335                                                 struct page_cgroup *pc)
 336 {
 337         struct page_cgroup *ret;
 338         /* lock and clear */
 339         lock_page_cgroup(page);
 340         ret = page_get_page_cgroup(page);
 341         if (likely(ret == pc))
 342                 page_assign_page_cgroup(page, NULL);
 343         unlock_page_cgroup(page);
 344         return ret;
 345 }
 346
 347 static void __mem_cgroup_remove_list(struct page_cgroup *pc)
 348 {
 349         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 350         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 351
 352         if (from)
 353                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 354         else
 355                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 356
 357         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, false);
 358         list_del_init(&pc->lru);
 359 }
 360
 361 static void __mem_cgroup_add_list(struct page_cgroup *pc)
 362 {
 363         int to = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 364         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 365
 366         if (!to) {
 367                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 368                 list_add(&pc->lru, &pc->mem_cgroup->inactive_list);
 369         } else {
 370                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 371                 list_add(&pc->lru, &pc->mem_cgroup->active_list);
 372         }
 373         mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, true);
 374 }
 375
 376 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 377 {
 378         int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
 379         struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
 380
 381         if (from)
 382                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
 383         else
 384                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
 385
 386         if (active) {
 387                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
 388                 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
 389                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 390         } else {
 391                 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
 392                 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
 393                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 394         }
 395 }
 396
 397 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 398 {
 399         int ret;
 400
 401         task_lock(task);
 402         ret = task->mm && mm_cgroup(task->mm) == mem;
 403         task_unlock(task);
 404         return ret;
 405 }
 406
 407 /*
 408  * This routine assumes that the appropriate zone's lru lock is already held
 409  */
 410 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 411 {
 412         struct mem_cgroup *mem;
 413         if (!pc)
 414                 return;
 415
 416         mem = pc->mem_cgroup;
 417
 418         spin_lock(&mem->lru_lock);
 419         __mem_cgroup_move_lists(pc, active);
 420         spin_unlock(&mem->lru_lock);
 421 }
 422
 423 /*
 424  * Calculate mapped_ratio under memory controller. This will be used in
 425  * vmscan.c for deteremining we have to reclaim mapped pages.
 426  */
 427 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
 428 {
 429         long total, rss;
 430
 431         /*
 432          * usage is recorded in bytes. But, here, we assume the number of
 433          * physical pages can be represented by "long" on any arch.
 434          */
 435         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
 436         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
 437         return (int)((rss * 100L) / total);
 438 }
 439 /*
 440  * This function is called from vmscan.c. In page reclaiming loop. balance
 441  * between active and inactive list is calculated. For memory controller
 442  * page reclaiming, we should use using mem_cgroup's imbalance rather than
 443  * zone's global lru imbalance.
 444  */
 445 long mem_cgroup_reclaim_imbalance(struct mem_cgroup *mem)
 446 {
 447         unsigned long active, inactive;
 448         /* active and inactive are the number of pages. 'long' is ok.*/
 449         active = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_ACTIVE);
 450         inactive = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_INACTIVE);
 451         return (long) (active / (inactive + 1));
 452 }
 453
 454 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 455                                         struct list_head *dst,
 456                                         unsigned long *scanned, int order,
 457                                         int mode, struct zone *z,
 458                                         struct mem_cgroup *mem_cont,
 459                                         int active)
 460 {
 461         unsigned long nr_taken = 0;
 462         struct page *page;
 463         unsigned long scan;
 464         LIST_HEAD(pc_list);
 465         struct list_head *src;
 466         struct page_cgroup *pc, *tmp;
 467
 468         if (active)
 469                 src = &mem_cont->active_list;
 470         else
 471                 src = &mem_cont->inactive_list;
 472
 473         spin_lock(&mem_cont->lru_lock);
 474         scan = 0;
 475         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 476                 if (scan >= nr_to_scan)
 477                         break;
 478                 page = pc->page;
 479                 VM_BUG_ON(!pc);
 480
 481                 if (unlikely(!PageLRU(page)))
 482                         continue;
 483
 484                 if (PageActive(page) && !active) {
 485                         __mem_cgroup_move_lists(pc, true);
 486                         continue;
 487                 }
 488                 if (!PageActive(page) && active) {
 489                         __mem_cgroup_move_lists(pc, false);
 490                         continue;
 491                 }
 492
 493                 /*
 494                  * Reclaim, per zone
 495                  * TODO: make the active/inactive lists per zone
 496                  */
 497                 if (page_zone(page) != z)
 498                         continue;
 499
 500                 scan++;
 501                 list_move(&pc->lru, &pc_list);
 502
 503                 if (__isolate_lru_page(page, mode) == 0) {
 504                         list_move(&page->lru, dst);
 505                         nr_taken++;
 506                 }
 507         }
 508
 509         list_splice(&pc_list, src);
 510         spin_unlock(&mem_cont->lru_lock);
 511
 512         *scanned = scan;
 513         return nr_taken;
 514 }
 515
 516 /*
 517  * Charge the memory controller for page usage.
 518  * Return
 519  * 0 if the charge was successful
 520  * < 0 if the cgroup is over its limit
 521  */
 522 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 523                                 gfp_t gfp_mask, enum charge_type ctype)
 524 {
 525         struct mem_cgroup *mem;
 526         struct page_cgroup *pc;
 527         unsigned long flags;
 528         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 529
 530         /*
 531          * Should page_cgroup's go to their own slab?
 532          * One could optimize the performance of the charging routine
 533          * by saving a bit in the page_flags and using it as a lock
 534          * to see if the cgroup page already has a page_cgroup associated
 535          * with it
 536          */
 537 retry:
 538         if (page) {
 539                 lock_page_cgroup(page);
 540                 pc = page_get_page_cgroup(page);
 541                 /*
 542                  * The page_cgroup exists and
 543                  * the page has already been accounted.
 544                  */
 545                 if (pc) {
 546                         if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 547                                 /* this page is under being uncharged ? */
 548                                 unlock_page_cgroup(page);
 549                                 cpu_relax();
 550                                 goto retry;
 551                         } else {
 552                                 unlock_page_cgroup(page);
 553                                 goto done;
 554                         }
 555                 }
 556                 unlock_page_cgroup(page);
 557         }
 558
 559         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 560         if (pc == NULL)
 561                 goto err;
 562
 563         /*
 564          * We always charge the cgroup the mm_struct belongs to.
 565          * The mm_struct's mem_cgroup changes on task migration if the
 566          * thread group leader migrates. It's possible that mm is not
 567          * set, if so charge the init_mm (happens for pagecache usage).
 568          */
 569         if (!mm)
 570                 mm = &init_mm;
 571
 572         rcu_read_lock();
 573         mem = rcu_dereference(mm->mem_cgroup);
 574         /*
 575          * For every charge from the cgroup, increment reference
 576          * count
 577          */
 578         css_get(&mem->css);
 579         rcu_read_unlock();
 580
 581         /*
 582          * If we created the page_cgroup, we should free it on exceeding
 583          * the cgroup limit.
 584          */
 585         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 586                 if (!(gfp_mask & __GFP_WAIT))
 587                         goto out;
 588
 589                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 590                         continue;
 591
 592                 /*
 593                  * try_to_free_mem_cgroup_pages() might not give us a full
 594                  * picture of reclaim. Some pages are reclaimed and might be
 595                  * moved to swap cache or just unmapped from the cgroup.
 596                  * Check the limit again to see if the reclaim reduced the
 597                  * current usage of the cgroup before giving up
 598                  */
 599                 if (res_counter_check_under_limit(&mem->res))
 600                         continue;
 601
 602                 if (!nr_retries--) {
 603                         mem_cgroup_out_of_memory(mem, gfp_mask);
 604                         goto out;
 605                 }
 606                 congestion_wait(WRITE, HZ/10);
 607         }
 608
 609         atomic_set(&pc->ref_cnt, 1);
 610         pc->mem_cgroup = mem;
 611         pc->page = page;
 612         pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
 613         if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
 614                 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
 615
 616         if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
 617                 /*
 618                  * Another charge has been added to this page already.
 619                  * We take lock_page_cgroup(page) again and read
 620                  * page->cgroup, increment refcnt.... just retry is OK.
 621                  */
 622                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 623                 css_put(&mem->css);
 624                 kfree(pc);
 625                 if (!page)
 626                         goto done;
 627                 goto retry;
 628         }
 629
 630         spin_lock_irqsave(&mem->lru_lock, flags);
 631         /* Update statistics vector */
 632         __mem_cgroup_add_list(pc);
 633         spin_unlock_irqrestore(&mem->lru_lock, flags);
 634
 635 done:
 636         return 0;
 637 out:
 638         css_put(&mem->css);
 639         kfree(pc);
 640 err:
 641         return -ENOMEM;
 642 }
 643
 644 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 645                         gfp_t gfp_mask)
 646 {
 647         return mem_cgroup_charge_common(page, mm, gfp_mask,
 648                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
 649 }
 650
 651 /*
 652  * See if the cached pages should be charged at all?
 653  */
 654 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 655                                 gfp_t gfp_mask)
 656 {
 657         int ret = 0;
 658         struct mem_cgroup *mem;
 659         if (!mm)
 660                 mm = &init_mm;
 661
 662         rcu_read_lock();
 663         mem = rcu_dereference(mm->mem_cgroup);
 664         css_get(&mem->css);
 665         rcu_read_unlock();
 666         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 667                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
 668                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
 669         css_put(&mem->css);
 670         return ret;
 671 }
 672
 673 /*
 674  * Uncharging is always a welcome operation, we never complain, simply
 675  * uncharge.
 676  */
 677 void mem_cgroup_uncharge(struct page_cgroup *pc)
 678 {
 679         struct mem_cgroup *mem;
 680         struct page *page;
 681         unsigned long flags;
 682
 683         /*
 684          * This can handle cases when a page is not charged at all and we
 685          * are switching between handling the control_type.
 686          */
 687         if (!pc)
 688                 return;
 689
 690         if (atomic_dec_and_test(&pc->ref_cnt)) {
 691                 page = pc->page;
 692                 /*
 693                  * get page->cgroup and clear it under lock.
 694                  * force_empty can drop page->cgroup without checking refcnt.
 695                  */
 696                 if (clear_page_cgroup(page, pc) == pc) {
 697                         mem = pc->mem_cgroup;
 698                         css_put(&mem->css);
 699                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 700                         spin_lock_irqsave(&mem->lru_lock, flags);
 701                         __mem_cgroup_remove_list(pc);
 702                         spin_unlock_irqrestore(&mem->lru_lock, flags);
 703                         kfree(pc);
 704                 }
 705         }
 706 }
 707
 708 /*
 709  * Returns non-zero if a page (under migration) has valid page_cgroup member.
 710  * Refcnt of page_cgroup is incremented.
 711  */
 712
 713 int mem_cgroup_prepare_migration(struct page *page)
 714 {
 715         struct page_cgroup *pc;
 716         int ret = 0;
 717         lock_page_cgroup(page);
 718         pc = page_get_page_cgroup(page);
 719         if (pc && atomic_inc_not_zero(&pc->ref_cnt))
 720                 ret = 1;
 721         unlock_page_cgroup(page);
 722         return ret;
 723 }
 724
 725 void mem_cgroup_end_migration(struct page *page)
 726 {
 727         struct page_cgroup *pc = page_get_page_cgroup(page);
 728         mem_cgroup_uncharge(pc);
 729 }
 730 /*
 731  * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
 732  * And no race with uncharge() routines because page_cgroup for *page*
 733  * has extra one reference by mem_cgroup_prepare_migration.
 734  */
 735
 736 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
 737 {
 738         struct page_cgroup *pc;
 739         struct mem_cgroup *mem;
 740         unsigned long flags;
 741 retry:
 742         pc = page_get_page_cgroup(page);
 743         if (!pc)
 744                 return;
 745         mem = pc->mem_cgroup;
 746         if (clear_page_cgroup(page, pc) != pc)
 747                 goto retry;
 748
 749         spin_lock_irqsave(&mem->lru_lock, flags);
 750
 751         __mem_cgroup_remove_list(pc);
 752         pc->page = newpage;
 753         lock_page_cgroup(newpage);
 754         page_assign_page_cgroup(newpage, pc);
 755         unlock_page_cgroup(newpage);
 756         __mem_cgroup_add_list(pc);
 757
 758         spin_unlock_irqrestore(&mem->lru_lock, flags);
 759         return;
 760 }
 761
 762 /*
 763  * This routine traverse page_cgroup in given list and drop them all.
 764  * This routine ignores page_cgroup->ref_cnt.
 765  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
 766  */
 767 #define FORCE_UNCHARGE_BATCH    (128)
 768 static void
 769 mem_cgroup_force_empty_list(struct mem_cgroup *mem, struct list_head *list)
 770 {
 771         struct page_cgroup *pc;
 772         struct page *page;
 773         int count;
 774         unsigned long flags;
 775
 776 retry:
 777         count = FORCE_UNCHARGE_BATCH;
 778         spin_lock_irqsave(&mem->lru_lock, flags);
 779
 780         while (--count && !list_empty(list)) {
 781                 pc = list_entry(list->prev, struct page_cgroup, lru);
 782                 page = pc->page;
 783                 /* Avoid race with charge */
 784                 atomic_set(&pc->ref_cnt, 0);
 785                 if (clear_page_cgroup(page, pc) == pc) {
 786                         css_put(&mem->css);
 787                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 788                         __mem_cgroup_remove_list(pc);
 789                         kfree(pc);
 790                 } else  /* being uncharged ? ...do relax */
 791                         break;
 792         }
 793         spin_unlock_irqrestore(&mem->lru_lock, flags);
 794         if (!list_empty(list)) {
 795                 cond_resched();
 796                 goto retry;
 797         }
 798         return;
 799 }
 800
 801 /*
 802  * make mem_cgroup's charge to be 0 if there is no task.
 803  * This enables deleting this mem_cgroup.
 804  */
 805
 806 int mem_cgroup_force_empty(struct mem_cgroup *mem)
 807 {
 808         int ret = -EBUSY;
 809         css_get(&mem->css);
 810         /*
 811          * page reclaim code (kswapd etc..) will move pages between
 812 `        * active_list <-> inactive_list while we don't take a lock.
 813          * So, we have to do loop here until all lists are empty.
 814          */
 815         while (!(list_empty(&mem->active_list) &&
 816                  list_empty(&mem->inactive_list))) {
 817                 if (atomic_read(&mem->css.cgroup->count) > 0)
 818                         goto out;
 819                 /* drop all page_cgroup in active_list */
 820                 mem_cgroup_force_empty_list(mem, &mem->active_list);
 821                 /* drop all page_cgroup in inactive_list */
 822                 mem_cgroup_force_empty_list(mem, &mem->inactive_list);
 823         }
 824         ret = 0;
 825 out:
 826         css_put(&mem->css);
 827         return ret;
 828 }
 829
 830
 831
 832 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 ssize_t mem_cgroup_read(struct cgroup *cont,
 846                         struct cftype *cft, struct file *file,
 847                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 848 {
 849         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 850                                 cft->private, userbuf, nbytes, ppos,
 851                                 NULL);
 852 }
 853
 854 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 855                                 struct file *file, const char __user *userbuf,
 856                                 size_t nbytes, loff_t *ppos)
 857 {
 858         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 859                                 cft->private, userbuf, nbytes, ppos,
 860                                 mem_cgroup_write_strategy);
 861 }
 862
 863 static ssize_t mem_control_type_write(struct cgroup *cont,
 864                         struct cftype *cft, struct file *file,
 865                         const char __user *userbuf,
 866                         size_t nbytes, loff_t *pos)
 867 {
 868         int ret;
 869         char *buf, *end;
 870         unsigned long tmp;
 871         struct mem_cgroup *mem;
 872
 873         mem = mem_cgroup_from_cont(cont);
 874         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 875         ret = -ENOMEM;
 876         if (buf == NULL)
 877                 goto out;
 878
 879         buf[nbytes] = 0;
 880         ret = -EFAULT;
 881         if (copy_from_user(buf, userbuf, nbytes))
 882                 goto out_free;
 883
 884         ret = -EINVAL;
 885         tmp = simple_strtoul(buf, &end, 10);
 886         if (*end != '\0')
 887                 goto out_free;
 888
 889         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 890                 goto out_free;
 891
 892         mem->control_type = tmp;
 893         ret = nbytes;
 894 out_free:
 895         kfree(buf);
 896 out:
 897         return ret;
 898 }
 899
 900 static ssize_t mem_control_type_read(struct cgroup *cont,
 901                                 struct cftype *cft,
 902                                 struct file *file, char __user *userbuf,
 903                                 size_t nbytes, loff_t *ppos)
 904 {
 905         unsigned long val;
 906         char buf[64], *s;
 907         struct mem_cgroup *mem;
 908
 909         mem = mem_cgroup_from_cont(cont);
 910         s = buf;
 911         val = mem->control_type;
 912         s += sprintf(s, "%lu\n", val);
 913         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 914                         ppos, buf, s - buf);
 915 }
 916
 917
 918 static ssize_t mem_force_empty_write(struct cgroup *cont,
 919                                 struct cftype *cft, struct file *file,
 920                                 const char __user *userbuf,
 921                                 size_t nbytes, loff_t *ppos)
 922 {
 923         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
 924         int ret;
 925         ret = mem_cgroup_force_empty(mem);
 926         if (!ret)
 927                 ret = nbytes;
 928         return ret;
 929 }
 930
 931 /*
 932  * Note: This should be removed if cgroup supports write-only file.
 933  */
 934
 935 static ssize_t mem_force_empty_read(struct cgroup *cont,
 936                                 struct cftype *cft,
 937                                 struct file *file, char __user *userbuf,
 938                                 size_t nbytes, loff_t *ppos)
 939 {
 940         return -EINVAL;
 941 }
 942
 943
 944 static const struct mem_cgroup_stat_desc {
 945         const char *msg;
 946         u64 unit;
 947 } mem_cgroup_stat_desc[] = {
 948         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
 949         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
 950 };
 951
 952 static int mem_control_stat_show(struct seq_file *m, void *arg)
 953 {
 954         struct cgroup *cont = m->private;
 955         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
 956         struct mem_cgroup_stat *stat = &mem_cont->stat;
 957         int i;
 958
 959         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
 960                 s64 val;
 961
 962                 val = mem_cgroup_read_stat(stat, i);
 963                 val *= mem_cgroup_stat_desc[i].unit;
 964                 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
 965                                 (long long)val);
 966         }
 967         /* showing # of active pages */
 968         {
 969                 unsigned long active, inactive;
 970
 971                 inactive = mem_cgroup_get_all_zonestat(mem_cont,
 972                                                 MEM_CGROUP_ZSTAT_INACTIVE);
 973                 active = mem_cgroup_get_all_zonestat(mem_cont,
 974                                                 MEM_CGROUP_ZSTAT_ACTIVE);
 975                 seq_printf(m, "active %ld\n", (active) * PAGE_SIZE);
 976                 seq_printf(m, "inactive %ld\n", (inactive) * PAGE_SIZE);
 977         }
 978         return 0;
 979 }
 980
 981 static const struct file_operations mem_control_stat_file_operations = {
 982         .read = seq_read,
 983         .llseek = seq_lseek,
 984         .release = single_release,
 985 };
 986
 987 static int mem_control_stat_open(struct inode *unused, struct file *file)
 988 {
 989         /* XXX __d_cont */
 990         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
 991
 992         file->f_op = &mem_control_stat_file_operations;
 993         return single_open(file, mem_control_stat_show, cont);
 994 }
 995
 996
 997
 998 static struct cftype mem_cgroup_files[] = {
 999         {
1000                 .name = "usage_in_bytes",
1001                 .private = RES_USAGE,
1002                 .read = mem_cgroup_read,
1003         },
1004         {
1005                 .name = "limit_in_bytes",
1006                 .private = RES_LIMIT,
1007                 .write = mem_cgroup_write,
1008                 .read = mem_cgroup_read,
1009         },
1010         {
1011                 .name = "failcnt",
1012                 .private = RES_FAILCNT,
1013                 .read = mem_cgroup_read,
1014         },
1015         {
1016                 .name = "control_type",
1017                 .write = mem_control_type_write,
1018                 .read = mem_control_type_read,
1019         },
1020         {
1021                 .name = "force_empty",
1022                 .write = mem_force_empty_write,
1023                 .read = mem_force_empty_read,
1024         },
1025         {
1026                 .name = "stat",
1027                 .open = mem_control_stat_open,
1028         },
1029 };
1030
1031 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1032 {
1033         struct mem_cgroup_per_node *pn;
1034
1035         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, node);
1036         if (!pn)
1037                 return 1;
1038         mem->info.nodeinfo[node] = pn;
1039         memset(pn, 0, sizeof(*pn));
1040         return 0;
1041 }
1042
1043 static struct mem_cgroup init_mem_cgroup;
1044
1045 static struct cgroup_subsys_state *
1046 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1047 {
1048         struct mem_cgroup *mem;
1049         int node;
1050
1051         if (unlikely((cont->parent) == NULL)) {
1052                 mem = &init_mem_cgroup;
1053                 init_mm.mem_cgroup = mem;
1054         } else
1055                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
1056
1057         if (mem == NULL)
1058                 return NULL;
1059
1060         res_counter_init(&mem->res);
1061         INIT_LIST_HEAD(&mem->active_list);
1062         INIT_LIST_HEAD(&mem->inactive_list);
1063         spin_lock_init(&mem->lru_lock);
1064         mem->control_type = MEM_CGROUP_TYPE_ALL;
1065         memset(&mem->info, 0, sizeof(mem->info));
1066
1067         for_each_node_state(node, N_POSSIBLE)
1068                 if (alloc_mem_cgroup_per_zone_info(mem, node))
1069                         goto free_out;
1070
1071         return &mem->css;
1072 free_out:
1073         for_each_node_state(node, N_POSSIBLE)
1074                 kfree(mem->info.nodeinfo[node]);
1075         if (cont->parent != NULL)
1076                 kfree(mem);
1077         return NULL;
1078 }
1079
1080 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1081                                         struct cgroup *cont)
1082 {
1083         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1084         mem_cgroup_force_empty(mem);
1085 }
1086
1087 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
1088                                 struct cgroup *cont)
1089 {
1090         int node;
1091         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1092
1093         for_each_node_state(node, N_POSSIBLE)
1094                 kfree(mem->info.nodeinfo[node]);
1095
1096         kfree(mem_cgroup_from_cont(cont));
1097 }
1098
1099 static int mem_cgroup_populate(struct cgroup_subsys *ss,
1100                                 struct cgroup *cont)
1101 {
1102         return cgroup_add_files(cont, ss, mem_cgroup_files,
1103                                         ARRAY_SIZE(mem_cgroup_files));
1104 }
1105
1106 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
1107                                 struct cgroup *cont,
1108                                 struct cgroup *old_cont,
1109                                 struct task_struct *p)
1110 {
1111         struct mm_struct *mm;
1112         struct mem_cgroup *mem, *old_mem;
1113
1114         mm = get_task_mm(p);
1115         if (mm == NULL)
1116                 return;
1117
1118         mem = mem_cgroup_from_cont(cont);
1119         old_mem = mem_cgroup_from_cont(old_cont);
1120
1121         if (mem == old_mem)
1122                 goto out;
1123
1124         /*
1125          * Only thread group leaders are allowed to migrate, the mm_struct is
1126          * in effect owned by the leader
1127          */
1128         if (p->tgid != p->pid)
1129                 goto out;
1130
1131         css_get(&mem->css);
1132         rcu_assign_pointer(mm->mem_cgroup, mem);
1133         css_put(&old_mem->css);
1134
1135 out:
1136         mmput(mm);
1137         return;
1138 }
1139
1140 struct cgroup_subsys mem_cgroup_subsys = {
1141         .name = "memory",
1142         .subsys_id = mem_cgroup_subsys_id,
1143         .create = mem_cgroup_create,
1144         .pre_destroy = mem_cgroup_pre_destroy,
1145         .destroy = mem_cgroup_destroy,
1146         .populate = mem_cgroup_populate,
1147         .attach = mem_cgroup_move_task,
1148         .early_init = 0,
1149 };