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