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 static inline int page_cgroup_nid(struct page_cgroup *pc)
 139 {
 140         return page_to_nid(pc->page);
 141 }
 142
 143 static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 144 {
 145         return page_zonenum(pc->page);
 146 }
 147
 148 enum {
 149         MEM_CGROUP_TYPE_UNSPEC = 0,
 150         MEM_CGROUP_TYPE_MAPPED,
 151         MEM_CGROUP_TYPE_CACHED,
 152         MEM_CGROUP_TYPE_ALL,
 153         MEM_CGROUP_TYPE_MAX,
 154 };
 155
 156 enum charge_type {
 157         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
 158         MEM_CGROUP_CHARGE_TYPE_MAPPED,
 159 };
 160
 161 /*
 162  * Always modified under lru lock. Then, not necessary to preempt_disable()
 163  */
 164 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
 165                                         bool charge)
 166 {
 167         int val = (charge)? 1 : -1;
 168         struct mem_cgroup_stat *stat = &mem->stat;
 169         VM_BUG_ON(!irqs_disabled());
 170
 171         if (flags & PAGE_CGROUP_FLAG_CACHE)
 172                 __mem_cgroup_stat_add_safe(stat,
 173                                         MEM_CGROUP_STAT_CACHE, val);
 174         else
 175                 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
 176
 177 }
 178
 179 static struct mem_cgroup init_mem_cgroup;
 180
 181 static inline
 182 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 183 {
 184         return container_of(cgroup_subsys_state(cont,
 185                                 mem_cgroup_subsys_id), struct mem_cgroup,
 186                                 css);
 187 }
 188
 189 static inline
 190 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 191 {
 192         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 193                                 struct mem_cgroup, css);
 194 }
 195
 196 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
 197 {
 198         struct mem_cgroup *mem;
 199
 200         mem = mem_cgroup_from_task(p);
 201         css_get(&mem->css);
 202         mm->mem_cgroup = mem;
 203 }
 204
 205 void mm_free_cgroup(struct mm_struct *mm)
 206 {
 207         css_put(&mm->mem_cgroup->css);
 208 }
 209
 210 static inline int page_cgroup_locked(struct page *page)
 211 {
 212         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
 213                                         &page->page_cgroup);
 214 }
 215
 216 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 217 {
 218         int locked;
 219
 220         /*
 221          * While resetting the page_cgroup we might not hold the
 222          * page_cgroup lock. free_hot_cold_page() is an example
 223          * of such a scenario
 224          */
 225         if (pc)
 226                 VM_BUG_ON(!page_cgroup_locked(page));
 227         locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
 228         page->page_cgroup = ((unsigned long)pc | locked);
 229 }
 230
 231 struct page_cgroup *page_get_page_cgroup(struct page *page)
 232 {
 233         return (struct page_cgroup *)
 234                 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 235 }
 236
 237 static void __always_inline lock_page_cgroup(struct page *page)
 238 {
 239         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 240         VM_BUG_ON(!page_cgroup_locked(page));
 241 }
 242
 243 static void __always_inline unlock_page_cgroup(struct page *page)
 244 {
 245         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 246 }
 247
 248 /*
 249  * Tie new page_cgroup to struct page under lock_page_cgroup()
 250  * This can fail if the page has been tied to a page_cgroup.
 251  * If success, returns 0.
 252  */
 253 static int page_cgroup_assign_new_page_cgroup(struct page *page,
 254                                                 struct page_cgroup *pc)
 255 {
 256         int ret = 0;
 257
 258         lock_page_cgroup(page);
 259         if (!page_get_page_cgroup(page))
 260                 page_assign_page_cgroup(page, pc);
 261         else /* A page is tied to other pc. */
 262                 ret = 1;
 263         unlock_page_cgroup(page);
 264         return ret;
 265 }
 266
 267 /*
 268  * Clear page->page_cgroup member under lock_page_cgroup().
 269  * If given "pc" value is different from one page->page_cgroup,
 270  * page->cgroup is not cleared.
 271  * Returns a value of page->page_cgroup at lock taken.
 272  * A can can detect failure of clearing by following
 273  *  clear_page_cgroup(page, pc) == pc
 274  */
 275
 276 static struct page_cgroup *clear_page_cgroup(struct page *page,
 277                                                 struct page_cgroup *pc)
 278 {
 279         struct page_cgroup *ret;
 280         /* lock and clear */
 281         lock_page_cgroup(page);
 282         ret = page_get_page_cgroup(page);
 283         if (likely(ret == pc))
 284                 page_assign_page_cgroup(page, NULL);
 285         unlock_page_cgroup(page);
 286         return ret;
 287 }
 288
 289 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 290 {
 291         if (active) {
 292                 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
 293                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 294         } else {
 295                 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
 296                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 297         }
 298 }
 299
 300 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 301 {
 302         int ret;
 303
 304         task_lock(task);
 305         ret = task->mm && mm_cgroup(task->mm) == mem;
 306         task_unlock(task);
 307         return ret;
 308 }
 309
 310 /*
 311  * This routine assumes that the appropriate zone's lru lock is already held
 312  */
 313 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 314 {
 315         struct mem_cgroup *mem;
 316         if (!pc)
 317                 return;
 318
 319         mem = pc->mem_cgroup;
 320
 321         spin_lock(&mem->lru_lock);
 322         __mem_cgroup_move_lists(pc, active);
 323         spin_unlock(&mem->lru_lock);
 324 }
 325
 326 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 327                                         struct list_head *dst,
 328                                         unsigned long *scanned, int order,
 329                                         int mode, struct zone *z,
 330                                         struct mem_cgroup *mem_cont,
 331                                         int active)
 332 {
 333         unsigned long nr_taken = 0;
 334         struct page *page;
 335         unsigned long scan;
 336         LIST_HEAD(pc_list);
 337         struct list_head *src;
 338         struct page_cgroup *pc, *tmp;
 339
 340         if (active)
 341                 src = &mem_cont->active_list;
 342         else
 343                 src = &mem_cont->inactive_list;
 344
 345         spin_lock(&mem_cont->lru_lock);
 346         scan = 0;
 347         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
 348                 if (scan >= nr_to_scan)
 349                         break;
 350                 page = pc->page;
 351                 VM_BUG_ON(!pc);
 352
 353                 if (unlikely(!PageLRU(page)))
 354                         continue;
 355
 356                 if (PageActive(page) && !active) {
 357                         __mem_cgroup_move_lists(pc, true);
 358                         continue;
 359                 }
 360                 if (!PageActive(page) && active) {
 361                         __mem_cgroup_move_lists(pc, false);
 362                         continue;
 363                 }
 364
 365                 /*
 366                  * Reclaim, per zone
 367                  * TODO: make the active/inactive lists per zone
 368                  */
 369                 if (page_zone(page) != z)
 370                         continue;
 371
 372                 scan++;
 373                 list_move(&pc->lru, &pc_list);
 374
 375                 if (__isolate_lru_page(page, mode) == 0) {
 376                         list_move(&page->lru, dst);
 377                         nr_taken++;
 378                 }
 379         }
 380
 381         list_splice(&pc_list, src);
 382         spin_unlock(&mem_cont->lru_lock);
 383
 384         *scanned = scan;
 385         return nr_taken;
 386 }
 387
 388 /*
 389  * Charge the memory controller for page usage.
 390  * Return
 391  * 0 if the charge was successful
 392  * < 0 if the cgroup is over its limit
 393  */
 394 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 395                                 gfp_t gfp_mask, enum charge_type ctype)
 396 {
 397         struct mem_cgroup *mem;
 398         struct page_cgroup *pc;
 399         unsigned long flags;
 400         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 401
 402         /*
 403          * Should page_cgroup's go to their own slab?
 404          * One could optimize the performance of the charging routine
 405          * by saving a bit in the page_flags and using it as a lock
 406          * to see if the cgroup page already has a page_cgroup associated
 407          * with it
 408          */
 409 retry:
 410         if (page) {
 411                 lock_page_cgroup(page);
 412                 pc = page_get_page_cgroup(page);
 413                 /*
 414                  * The page_cgroup exists and
 415                  * the page has already been accounted.
 416                  */
 417                 if (pc) {
 418                         if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 419                                 /* this page is under being uncharged ? */
 420                                 unlock_page_cgroup(page);
 421                                 cpu_relax();
 422                                 goto retry;
 423                         } else {
 424                                 unlock_page_cgroup(page);
 425                                 goto done;
 426                         }
 427                 }
 428                 unlock_page_cgroup(page);
 429         }
 430
 431         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 432         if (pc == NULL)
 433                 goto err;
 434
 435         /*
 436          * We always charge the cgroup the mm_struct belongs to.
 437          * The mm_struct's mem_cgroup changes on task migration if the
 438          * thread group leader migrates. It's possible that mm is not
 439          * set, if so charge the init_mm (happens for pagecache usage).
 440          */
 441         if (!mm)
 442                 mm = &init_mm;
 443
 444         rcu_read_lock();
 445         mem = rcu_dereference(mm->mem_cgroup);
 446         /*
 447          * For every charge from the cgroup, increment reference
 448          * count
 449          */
 450         css_get(&mem->css);
 451         rcu_read_unlock();
 452
 453         /*
 454          * If we created the page_cgroup, we should free it on exceeding
 455          * the cgroup limit.
 456          */
 457         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 458                 if (!(gfp_mask & __GFP_WAIT))
 459                         goto out;
 460
 461                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 462                         continue;
 463
 464                 /*
 465                  * try_to_free_mem_cgroup_pages() might not give us a full
 466                  * picture of reclaim. Some pages are reclaimed and might be
 467                  * moved to swap cache or just unmapped from the cgroup.
 468                  * Check the limit again to see if the reclaim reduced the
 469                  * current usage of the cgroup before giving up
 470                  */
 471                 if (res_counter_check_under_limit(&mem->res))
 472                         continue;
 473
 474                 if (!nr_retries--) {
 475                         mem_cgroup_out_of_memory(mem, gfp_mask);
 476                         goto out;
 477                 }
 478                 congestion_wait(WRITE, HZ/10);
 479         }
 480
 481         atomic_set(&pc->ref_cnt, 1);
 482         pc->mem_cgroup = mem;
 483         pc->page = page;
 484         pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
 485         if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
 486                 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
 487
 488         if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
 489                 /*
 490                  * Another charge has been added to this page already.
 491                  * We take lock_page_cgroup(page) again and read
 492                  * page->cgroup, increment refcnt.... just retry is OK.
 493                  */
 494                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 495                 css_put(&mem->css);
 496                 kfree(pc);
 497                 if (!page)
 498                         goto done;
 499                 goto retry;
 500         }
 501
 502         spin_lock_irqsave(&mem->lru_lock, flags);
 503         /* Update statistics vector */
 504         mem_cgroup_charge_statistics(mem, pc->flags, true);
 505         list_add(&pc->lru, &mem->active_list);
 506         spin_unlock_irqrestore(&mem->lru_lock, flags);
 507
 508 done:
 509         return 0;
 510 out:
 511         css_put(&mem->css);
 512         kfree(pc);
 513 err:
 514         return -ENOMEM;
 515 }
 516
 517 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 518                         gfp_t gfp_mask)
 519 {
 520         return mem_cgroup_charge_common(page, mm, gfp_mask,
 521                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
 522 }
 523
 524 /*
 525  * See if the cached pages should be charged at all?
 526  */
 527 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 528                                 gfp_t gfp_mask)
 529 {
 530         int ret = 0;
 531         struct mem_cgroup *mem;
 532         if (!mm)
 533                 mm = &init_mm;
 534
 535         rcu_read_lock();
 536         mem = rcu_dereference(mm->mem_cgroup);
 537         css_get(&mem->css);
 538         rcu_read_unlock();
 539         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 540                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
 541                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
 542         css_put(&mem->css);
 543         return ret;
 544 }
 545
 546 /*
 547  * Uncharging is always a welcome operation, we never complain, simply
 548  * uncharge.
 549  */
 550 void mem_cgroup_uncharge(struct page_cgroup *pc)
 551 {
 552         struct mem_cgroup *mem;
 553         struct page *page;
 554         unsigned long flags;
 555
 556         /*
 557          * This can handle cases when a page is not charged at all and we
 558          * are switching between handling the control_type.
 559          */
 560         if (!pc)
 561                 return;
 562
 563         if (atomic_dec_and_test(&pc->ref_cnt)) {
 564                 page = pc->page;
 565                 /*
 566                  * get page->cgroup and clear it under lock.
 567                  * force_empty can drop page->cgroup without checking refcnt.
 568                  */
 569                 if (clear_page_cgroup(page, pc) == pc) {
 570                         mem = pc->mem_cgroup;
 571                         css_put(&mem->css);
 572                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 573                         spin_lock_irqsave(&mem->lru_lock, flags);
 574                         list_del_init(&pc->lru);
 575                         mem_cgroup_charge_statistics(mem, pc->flags, false);
 576                         spin_unlock_irqrestore(&mem->lru_lock, flags);
 577                         kfree(pc);
 578                 }
 579         }
 580 }
 581 /*
 582  * Returns non-zero if a page (under migration) has valid page_cgroup member.
 583  * Refcnt of page_cgroup is incremented.
 584  */
 585
 586 int mem_cgroup_prepare_migration(struct page *page)
 587 {
 588         struct page_cgroup *pc;
 589         int ret = 0;
 590         lock_page_cgroup(page);
 591         pc = page_get_page_cgroup(page);
 592         if (pc && atomic_inc_not_zero(&pc->ref_cnt))
 593                 ret = 1;
 594         unlock_page_cgroup(page);
 595         return ret;
 596 }
 597
 598 void mem_cgroup_end_migration(struct page *page)
 599 {
 600         struct page_cgroup *pc = page_get_page_cgroup(page);
 601         mem_cgroup_uncharge(pc);
 602 }
 603 /*
 604  * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
 605  * And no race with uncharge() routines because page_cgroup for *page*
 606  * has extra one reference by mem_cgroup_prepare_migration.
 607  */
 608
 609 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
 610 {
 611         struct page_cgroup *pc;
 612 retry:
 613         pc = page_get_page_cgroup(page);
 614         if (!pc)
 615                 return;
 616         if (clear_page_cgroup(page, pc) != pc)
 617                 goto retry;
 618         pc->page = newpage;
 619         lock_page_cgroup(newpage);
 620         page_assign_page_cgroup(newpage, pc);
 621         unlock_page_cgroup(newpage);
 622         return;
 623 }
 624
 625 /*
 626  * This routine traverse page_cgroup in given list and drop them all.
 627  * This routine ignores page_cgroup->ref_cnt.
 628  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
 629  */
 630 #define FORCE_UNCHARGE_BATCH    (128)
 631 static void
 632 mem_cgroup_force_empty_list(struct mem_cgroup *mem, struct list_head *list)
 633 {
 634         struct page_cgroup *pc;
 635         struct page *page;
 636         int count;
 637         unsigned long flags;
 638
 639 retry:
 640         count = FORCE_UNCHARGE_BATCH;
 641         spin_lock_irqsave(&mem->lru_lock, flags);
 642
 643         while (--count && !list_empty(list)) {
 644                 pc = list_entry(list->prev, struct page_cgroup, lru);
 645                 page = pc->page;
 646                 /* Avoid race with charge */
 647                 atomic_set(&pc->ref_cnt, 0);
 648                 if (clear_page_cgroup(page, pc) == pc) {
 649                         css_put(&mem->css);
 650                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 651                         list_del_init(&pc->lru);
 652                         mem_cgroup_charge_statistics(mem, pc->flags, false);
 653                         kfree(pc);
 654                 } else  /* being uncharged ? ...do relax */
 655                         break;
 656         }
 657         spin_unlock_irqrestore(&mem->lru_lock, flags);
 658         if (!list_empty(list)) {
 659                 cond_resched();
 660                 goto retry;
 661         }
 662         return;
 663 }
 664
 665 /*
 666  * make mem_cgroup's charge to be 0 if there is no task.
 667  * This enables deleting this mem_cgroup.
 668  */
 669
 670 int mem_cgroup_force_empty(struct mem_cgroup *mem)
 671 {
 672         int ret = -EBUSY;
 673         css_get(&mem->css);
 674         /*
 675          * page reclaim code (kswapd etc..) will move pages between
 676 `        * active_list <-> inactive_list while we don't take a lock.
 677          * So, we have to do loop here until all lists are empty.
 678          */
 679         while (!(list_empty(&mem->active_list) &&
 680                  list_empty(&mem->inactive_list))) {
 681                 if (atomic_read(&mem->css.cgroup->count) > 0)
 682                         goto out;
 683                 /* drop all page_cgroup in active_list */
 684                 mem_cgroup_force_empty_list(mem, &mem->active_list);
 685                 /* drop all page_cgroup in inactive_list */
 686                 mem_cgroup_force_empty_list(mem, &mem->inactive_list);
 687         }
 688         ret = 0;
 689 out:
 690         css_put(&mem->css);
 691         return ret;
 692 }
 693
 694
 695
 696 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 697 {
 698         *tmp = memparse(buf, &buf);
 699         if (*buf != '\0')
 700                 return -EINVAL;
 701
 702         /*
 703          * Round up the value to the closest page size
 704          */
 705         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 706         return 0;
 707 }
 708
 709 static ssize_t mem_cgroup_read(struct cgroup *cont,
 710                         struct cftype *cft, struct file *file,
 711                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 712 {
 713         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 714                                 cft->private, userbuf, nbytes, ppos,
 715                                 NULL);
 716 }
 717
 718 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 719                                 struct file *file, const char __user *userbuf,
 720                                 size_t nbytes, loff_t *ppos)
 721 {
 722         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 723                                 cft->private, userbuf, nbytes, ppos,
 724                                 mem_cgroup_write_strategy);
 725 }
 726
 727 static ssize_t mem_control_type_write(struct cgroup *cont,
 728                         struct cftype *cft, struct file *file,
 729                         const char __user *userbuf,
 730                         size_t nbytes, loff_t *pos)
 731 {
 732         int ret;
 733         char *buf, *end;
 734         unsigned long tmp;
 735         struct mem_cgroup *mem;
 736
 737         mem = mem_cgroup_from_cont(cont);
 738         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 739         ret = -ENOMEM;
 740         if (buf == NULL)
 741                 goto out;
 742
 743         buf[nbytes] = 0;
 744         ret = -EFAULT;
 745         if (copy_from_user(buf, userbuf, nbytes))
 746                 goto out_free;
 747
 748         ret = -EINVAL;
 749         tmp = simple_strtoul(buf, &end, 10);
 750         if (*end != '\0')
 751                 goto out_free;
 752
 753         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 754                 goto out_free;
 755
 756         mem->control_type = tmp;
 757         ret = nbytes;
 758 out_free:
 759         kfree(buf);
 760 out:
 761         return ret;
 762 }
 763
 764 static ssize_t mem_control_type_read(struct cgroup *cont,
 765                                 struct cftype *cft,
 766                                 struct file *file, char __user *userbuf,
 767                                 size_t nbytes, loff_t *ppos)
 768 {
 769         unsigned long val;
 770         char buf[64], *s;
 771         struct mem_cgroup *mem;
 772
 773         mem = mem_cgroup_from_cont(cont);
 774         s = buf;
 775         val = mem->control_type;
 776         s += sprintf(s, "%lu\n", val);
 777         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 778                         ppos, buf, s - buf);
 779 }
 780
 781
 782 static ssize_t mem_force_empty_write(struct cgroup *cont,
 783                                 struct cftype *cft, struct file *file,
 784                                 const char __user *userbuf,
 785                                 size_t nbytes, loff_t *ppos)
 786 {
 787         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
 788         int ret;
 789         ret = mem_cgroup_force_empty(mem);
 790         if (!ret)
 791                 ret = nbytes;
 792         return ret;
 793 }
 794
 795 /*
 796  * Note: This should be removed if cgroup supports write-only file.
 797  */
 798
 799 static ssize_t mem_force_empty_read(struct cgroup *cont,
 800                                 struct cftype *cft,
 801                                 struct file *file, char __user *userbuf,
 802                                 size_t nbytes, loff_t *ppos)
 803 {
 804         return -EINVAL;
 805 }
 806
 807
 808 static const struct mem_cgroup_stat_desc {
 809         const char *msg;
 810         u64 unit;
 811 } mem_cgroup_stat_desc[] = {
 812         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
 813         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
 814 };
 815
 816 static int mem_control_stat_show(struct seq_file *m, void *arg)
 817 {
 818         struct cgroup *cont = m->private;
 819         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
 820         struct mem_cgroup_stat *stat = &mem_cont->stat;
 821         int i;
 822
 823         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
 824                 s64 val;
 825
 826                 val = mem_cgroup_read_stat(stat, i);
 827                 val *= mem_cgroup_stat_desc[i].unit;
 828                 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
 829                                 (long long)val);
 830         }
 831         return 0;
 832 }
 833
 834 static const struct file_operations mem_control_stat_file_operations = {
 835         .read = seq_read,
 836         .llseek = seq_lseek,
 837         .release = single_release,
 838 };
 839
 840 static int mem_control_stat_open(struct inode *unused, struct file *file)
 841 {
 842         /* XXX __d_cont */
 843         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
 844
 845         file->f_op = &mem_control_stat_file_operations;
 846         return single_open(file, mem_control_stat_show, cont);
 847 }
 848
 849
 850
 851 static struct cftype mem_cgroup_files[] = {
 852         {
 853                 .name = "usage_in_bytes",
 854                 .private = RES_USAGE,
 855                 .read = mem_cgroup_read,
 856         },
 857         {
 858                 .name = "limit_in_bytes",
 859                 .private = RES_LIMIT,
 860                 .write = mem_cgroup_write,
 861                 .read = mem_cgroup_read,
 862         },
 863         {
 864                 .name = "failcnt",
 865                 .private = RES_FAILCNT,
 866                 .read = mem_cgroup_read,
 867         },
 868         {
 869                 .name = "control_type",
 870                 .write = mem_control_type_write,
 871                 .read = mem_control_type_read,
 872         },
 873         {
 874                 .name = "force_empty",
 875                 .write = mem_force_empty_write,
 876                 .read = mem_force_empty_read,
 877         },
 878         {
 879                 .name = "stat",
 880                 .open = mem_control_stat_open,
 881         },
 882 };
 883
 884 static struct mem_cgroup init_mem_cgroup;
 885
 886 static struct cgroup_subsys_state *
 887 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 888 {
 889         struct mem_cgroup *mem;
 890
 891         if (unlikely((cont->parent) == NULL)) {
 892                 mem = &init_mem_cgroup;
 893                 init_mm.mem_cgroup = mem;
 894         } else
 895                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
 896
 897         if (mem == NULL)
 898                 return NULL;
 899
 900         res_counter_init(&mem->res);
 901         INIT_LIST_HEAD(&mem->active_list);
 902         INIT_LIST_HEAD(&mem->inactive_list);
 903         spin_lock_init(&mem->lru_lock);
 904         mem->control_type = MEM_CGROUP_TYPE_ALL;
 905         return &mem->css;
 906 }
 907
 908 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
 909                                         struct cgroup *cont)
 910 {
 911         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
 912         mem_cgroup_force_empty(mem);
 913 }
 914
 915 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
 916                                 struct cgroup *cont)
 917 {
 918         kfree(mem_cgroup_from_cont(cont));
 919 }
 920
 921 static int mem_cgroup_populate(struct cgroup_subsys *ss,
 922                                 struct cgroup *cont)
 923 {
 924         return cgroup_add_files(cont, ss, mem_cgroup_files,
 925                                         ARRAY_SIZE(mem_cgroup_files));
 926 }
 927
 928 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
 929                                 struct cgroup *cont,
 930                                 struct cgroup *old_cont,
 931                                 struct task_struct *p)
 932 {
 933         struct mm_struct *mm;
 934         struct mem_cgroup *mem, *old_mem;
 935
 936         mm = get_task_mm(p);
 937         if (mm == NULL)
 938                 return;
 939
 940         mem = mem_cgroup_from_cont(cont);
 941         old_mem = mem_cgroup_from_cont(old_cont);
 942
 943         if (mem == old_mem)
 944                 goto out;
 945
 946         /*
 947          * Only thread group leaders are allowed to migrate, the mm_struct is
 948          * in effect owned by the leader
 949          */
 950         if (p->tgid != p->pid)
 951                 goto out;
 952
 953         css_get(&mem->css);
 954         rcu_assign_pointer(mm->mem_cgroup, mem);
 955         css_put(&old_mem->css);
 956
 957 out:
 958         mmput(mm);
 959         return;
 960 }
 961
 962 struct cgroup_subsys mem_cgroup_subsys = {
 963         .name = "memory",
 964         .subsys_id = mem_cgroup_subsys_id,
 965         .create = mem_cgroup_create,
 966         .pre_destroy = mem_cgroup_pre_destroy,
 967         .destroy = mem_cgroup_destroy,
 968         .populate = mem_cgroup_populate,
 969         .attach = mem_cgroup_move_task,
 970         .early_init = 1,
 971 };