mm/vmscan.c

   1 /*
   2  *  linux/mm/vmscan.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *
   6  *  Swap reorganised 29.12.95, Stephen Tweedie.
   7  *  kswapd added: 7.1.96  sct
   8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
   9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  11  *  Multiqueue VM started 5.8.00, Rik van Riel.
  12  */
  13
  14 #include <linux/mm.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/swap.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/init.h>
  21 #include <linux/highmem.h>
  22 #include <linux/file.h>
  23 #include <linux/writeback.h>
  24 #include <linux/suspend.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  27                                         buffer_heads_over_limit */
  28 #include <linux/mm_inline.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/rmap.h>
  32 #include <linux/topology.h>
  33 #include <linux/cpu.h>
  34 #include <linux/notifier.h>
  35 #include <linux/rwsem.h>
  36
  37 #include <asm/tlbflush.h>
  38 #include <asm/div64.h>
  39
  40 #include <linux/swapops.h>
  41 #include <linux/ckrm_mem.h>
  42
  43 #ifndef AT_LIMIT_SUPPORT
  44 #warning "ckrm_at_limit disabled due to problems with memory hog tests -- seting ckrm_shrink_list_empty to true"
  45 #undef ckrm_shrink_list_empty
  46 #define ckrm_shrink_list_empty()                (1)
  47 #endif
  48
  49 /* possible outcome of pageout() */
  50 typedef enum {
  51         /* failed to write page out, page is locked */
  52         PAGE_KEEP,
  53         /* move page to the active list, page is locked */
  54         PAGE_ACTIVATE,
  55         /* page has been sent to the disk successfully, page is unlocked */
  56         PAGE_SUCCESS,
  57         /* page is clean and locked */
  58         PAGE_CLEAN,
  59 } pageout_t;
  60
  61 struct scan_control {
  62         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  63         unsigned long nr_to_scan;
  64
  65         /* Incremented by the number of inactive pages that were scanned */
  66         unsigned long nr_scanned;
  67
  68         /* Incremented by the number of pages reclaimed */
  69         unsigned long nr_reclaimed;
  70
  71         unsigned long nr_mapped;        /* From page_state */
  72
  73         /* How many pages shrink_cache() should reclaim */
  74         int nr_to_reclaim;
  75
  76         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  77         unsigned int priority;
  78
  79         /* This context's GFP mask */
  80         unsigned int gfp_mask;
  81
  82         /* Flag used by CKRM */
  83         unsigned int ckrm_flags;
  84
  85         int may_writepage;
  86 };
  87
  88 /*
  89  * The list of shrinker callbacks used by to apply pressure to
  90  * ageable caches.
  91  */
  92 struct shrinker {
  93         shrinker_t              shrinker;
  94         struct list_head        list;
  95         int                     seeks;  /* seeks to recreate an obj */
  96         long                    nr;     /* objs pending delete */
  97 };
  98
  99 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
 100
 101 #ifdef ARCH_HAS_PREFETCH
 102 #define prefetch_prev_lru_page(_page, _base, _field)                    \
 103         do {                                                            \
 104                 if ((_page)->lru.prev != _base) {                       \
 105                         struct page *prev;                              \
 106                                                                         \
 107                         prev = lru_to_page(&(_page->lru));              \
 108                         prefetch(&prev->_field);                        \
 109                 }                                                       \
 110         } while (0)
 111 #else
 112 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 113 #endif
 114
 115 #ifdef ARCH_HAS_PREFETCHW
 116 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 117         do {                                                            \
 118                 if ((_page)->lru.prev != _base) {                       \
 119                         struct page *prev;                              \
 120                                                                         \
 121                         prev = lru_to_page(&(_page->lru));              \
 122                         prefetchw(&prev->_field);                       \
 123                 }                                                       \
 124         } while (0)
 125 #else
 126 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 127 #endif
 128
 129 /*
 130  * From 0 .. 100.  Higher means more swappy.
 131  */
 132 int vm_swappiness = 60;
 133 static long total_memory;
 134
 135 static LIST_HEAD(shrinker_list);
 136 static DECLARE_RWSEM(shrinker_rwsem);
 137
 138 /*
 139  * Add a shrinker callback to be called from the vm
 140  */
 141 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 142 {
 143         struct shrinker *shrinker;
 144
 145         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 146         if (shrinker) {
 147                 shrinker->shrinker = theshrinker;
 148                 shrinker->seeks = seeks;
 149                 shrinker->nr = 0;
 150                 down_write(&shrinker_rwsem);
 151                 list_add(&shrinker->list, &shrinker_list);
 152                 up_write(&shrinker_rwsem);
 153         }
 154         return shrinker;
 155 }
 156 EXPORT_SYMBOL(set_shrinker);
 157
 158 /*
 159  * Remove one
 160  */
 161 void remove_shrinker(struct shrinker *shrinker)
 162 {
 163         down_write(&shrinker_rwsem);
 164         list_del(&shrinker->list);
 165         up_write(&shrinker_rwsem);
 166         kfree(shrinker);
 167 }
 168 EXPORT_SYMBOL(remove_shrinker);
 169
 170 #define SHRINK_BATCH 128
 171 /*
 172  * Call the shrink functions to age shrinkable caches
 173  *
 174  * Here we assume it costs one seek to replace a lru page and that it also
 175  * takes a seek to recreate a cache object.  With this in mind we age equal
 176  * percentages of the lru and ageable caches.  This should balance the seeks
 177  * generated by these structures.
 178  *
 179  * If the vm encounted mapped pages on the LRU it increase the pressure on
 180  * slab to avoid swapping.
 181  *
 182  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 183  *
 184  * `lru_pages' represents the number of on-LRU pages in all the zones which
 185  * are eligible for the caller's allocation attempt.  It is used for balancing
 186  * slab reclaim versus page reclaim.
 187  */
 188 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 189                         unsigned long lru_pages)
 190 {
 191         struct shrinker *shrinker;
 192
 193         if (scanned == 0)
 194                 scanned = SWAP_CLUSTER_MAX;
 195
 196         if (!down_read_trylock(&shrinker_rwsem))
 197                 return 0;
 198
 199         list_for_each_entry(shrinker, &shrinker_list, list) {
 200                 unsigned long long delta;
 201                 unsigned long total_scan;
 202
 203                 delta = (4 * scanned) / shrinker->seeks;
 204                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 205                 do_div(delta, lru_pages + 1);
 206                 shrinker->nr += delta;
 207                 if (shrinker->nr < 0)
 208                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 209
 210                 total_scan = shrinker->nr;
 211                 shrinker->nr = 0;
 212
 213                 while (total_scan >= SHRINK_BATCH) {
 214                         long this_scan = SHRINK_BATCH;
 215                         int shrink_ret;
 216
 217                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 218                         if (shrink_ret == -1)
 219                                 break;
 220                         mod_page_state(slabs_scanned, this_scan);
 221                         total_scan -= this_scan;
 222
 223                         cond_resched();
 224                 }
 225
 226                 shrinker->nr += total_scan;
 227         }
 228         up_read(&shrinker_rwsem);
 229         return 0;
 230 }
 231
 232 /* Called without lock on whether page is mapped, so answer is unstable */
 233 static inline int page_mapping_inuse(struct page *page)
 234 {
 235         struct address_space *mapping;
 236
 237         /* Page is in somebody's page tables. */
 238         if (page_mapped(page))
 239                 return 1;
 240
 241         /* Be more reluctant to reclaim swapcache than pagecache */
 242         if (PageSwapCache(page))
 243                 return 1;
 244
 245         mapping = page_mapping(page);
 246         if (!mapping)
 247                 return 0;
 248
 249         /* File is mmap'd by somebody? */
 250         return mapping_mapped(mapping);
 251 }
 252
 253 static inline int is_page_cache_freeable(struct page *page)
 254 {
 255         return page_count(page) - !!PagePrivate(page) == 2;
 256 }
 257
 258 static int may_write_to_queue(struct backing_dev_info *bdi)
 259 {
 260         if (current_is_kswapd())
 261                 return 1;
 262         if (current_is_pdflush())       /* This is unlikely, but why not... */
 263                 return 1;
 264         if (!bdi_write_congested(bdi))
 265                 return 1;
 266         if (bdi == current->backing_dev_info)
 267                 return 1;
 268         return 0;
 269 }
 270
 271 /*
 272  * We detected a synchronous write error writing a page out.  Probably
 273  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 274  * fsync(), msync() or close().
 275  *
 276  * The tricky part is that after writepage we cannot touch the mapping: nothing
 277  * prevents it from being freed up.  But we have a ref on the page and once
 278  * that page is locked, the mapping is pinned.
 279  *
 280  * We're allowed to run sleeping lock_page() here because we know the caller has
 281  * __GFP_FS.
 282  */
 283 static void handle_write_error(struct address_space *mapping,
 284                                 struct page *page, int error)
 285 {
 286         lock_page(page);
 287         if (page_mapping(page) == mapping) {
 288                 if (error == -ENOSPC)
 289                         set_bit(AS_ENOSPC, &mapping->flags);
 290                 else
 291                         set_bit(AS_EIO, &mapping->flags);
 292         }
 293         unlock_page(page);
 294 }
 295
 296 /*
 297  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 298  */
 299 static pageout_t pageout(struct page *page, struct address_space *mapping)
 300 {
 301         /*
 302          * If the page is dirty, only perform writeback if that write
 303          * will be non-blocking.  To prevent this allocation from being
 304          * stalled by pagecache activity.  But note that there may be
 305          * stalls if we need to run get_block().  We could test
 306          * PagePrivate for that.
 307          *
 308          * If this process is currently in generic_file_write() against
 309          * this page's queue, we can perform writeback even if that
 310          * will block.
 311          *
 312          * If the page is swapcache, write it back even if that would
 313          * block, for some throttling. This happens by accident, because
 314          * swap_backing_dev_info is bust: it doesn't reflect the
 315          * congestion state of the swapdevs.  Easy to fix, if needed.
 316          * See swapfile.c:page_queue_congested().
 317          */
 318         if (!is_page_cache_freeable(page))
 319                 return PAGE_KEEP;
 320         if (!mapping)
 321                 return PAGE_KEEP;
 322         if (mapping->a_ops->writepage == NULL)
 323                 return PAGE_ACTIVATE;
 324         if (!may_write_to_queue(mapping->backing_dev_info))
 325                 return PAGE_KEEP;
 326
 327         if (clear_page_dirty_for_io(page)) {
 328                 int res;
 329                 struct writeback_control wbc = {
 330                         .sync_mode = WB_SYNC_NONE,
 331                         .nr_to_write = SWAP_CLUSTER_MAX,
 332                         .nonblocking = 1,
 333                         .for_reclaim = 1,
 334                 };
 335
 336                 SetPageReclaim(page);
 337                 res = mapping->a_ops->writepage(page, &wbc);
 338                 if (res < 0)
 339                         handle_write_error(mapping, page, res);
 340                 if (res == WRITEPAGE_ACTIVATE) {
 341                         ClearPageReclaim(page);
 342                         return PAGE_ACTIVATE;
 343                 }
 344                 if (!PageWriteback(page)) {
 345                         /* synchronous write or broken a_ops? */
 346                         ClearPageReclaim(page);
 347                 }
 348
 349                 return PAGE_SUCCESS;
 350         }
 351
 352         return PAGE_CLEAN;
 353 }
 354
 355 /*
 356  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 357  */
 358 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 359 {
 360         LIST_HEAD(ret_pages);
 361         struct pagevec freed_pvec;
 362         int pgactivate = 0;
 363         int reclaimed = 0;
 364
 365         cond_resched();
 366
 367         pagevec_init(&freed_pvec, 1);
 368         while (!list_empty(page_list)) {
 369                 struct address_space *mapping;
 370                 struct page *page;
 371                 int may_enter_fs;
 372                 int referenced;
 373
 374                 cond_resched();
 375
 376                 page = lru_to_page(page_list);
 377                 list_del(&page->lru);
 378
 379                 if (TestSetPageLocked(page))
 380                         goto keep;
 381
 382                 BUG_ON(PageActive(page));
 383
 384                 if (PageWriteback(page))
 385                         goto keep_locked;
 386
 387                 sc->nr_scanned++;
 388                 /* Double the slab pressure for mapped and swapcache pages */
 389                 if (page_mapped(page) || PageSwapCache(page))
 390                         sc->nr_scanned++;
 391
 392                 referenced = page_referenced(page, 1, sc->priority <= 0);
 393                 /* In active use or really unfreeable?  Activate it. */
 394                 if (referenced && page_mapping_inuse(page))
 395                         goto activate_locked;
 396
 397 #ifdef CONFIG_SWAP
 398                 /*
 399                  * Anonymous process memory has backing store?
 400                  * Try to allocate it some swap space here.
 401                  */
 402                 if (PageAnon(page) && !PageSwapCache(page)) {
 403                         if (!add_to_swap(page))
 404                                 goto activate_locked;
 405                 }
 406 #endif /* CONFIG_SWAP */
 407
 408                 mapping = page_mapping(page);
 409                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 410                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 411
 412                 /*
 413                  * The page is mapped into the page tables of one or more
 414                  * processes. Try to unmap it here.
 415                  */
 416                 if (page_mapped(page) && mapping) {
 417                         switch (try_to_unmap(page)) {
 418                         case SWAP_FAIL:
 419                                 goto activate_locked;
 420                         case SWAP_AGAIN:
 421                                 goto keep_locked;
 422                         case SWAP_SUCCESS:
 423                                 ; /* try to free the page below */
 424                         }
 425                 }
 426
 427                 if (PageDirty(page)) {
 428                         if (referenced)
 429                                 goto keep_locked;
 430                         if (!may_enter_fs)
 431                                 goto keep_locked;
 432                         if (laptop_mode && !sc->may_writepage)
 433                                 goto keep_locked;
 434
 435                         /* Page is dirty, try to write it out here */
 436                         switch(pageout(page, mapping)) {
 437                         case PAGE_KEEP:
 438                                 goto keep_locked;
 439                         case PAGE_ACTIVATE:
 440                                 goto activate_locked;
 441                         case PAGE_SUCCESS:
 442                                 if (PageWriteback(page) || PageDirty(page))
 443                                         goto keep;
 444                                 /*
 445                                  * A synchronous write - probably a ramdisk.  Go
 446                                  * ahead and try to reclaim the page.
 447                                  */
 448                                 if (TestSetPageLocked(page))
 449                                         goto keep;
 450                                 if (PageDirty(page) || PageWriteback(page))
 451                                         goto keep_locked;
 452                                 mapping = page_mapping(page);
 453                         case PAGE_CLEAN:
 454                                 ; /* try to free the page below */
 455                         }
 456                 }
 457
 458                 /*
 459                  * If the page has buffers, try to free the buffer mappings
 460                  * associated with this page. If we succeed we try to free
 461                  * the page as well.
 462                  *
 463                  * We do this even if the page is PageDirty().
 464                  * try_to_release_page() does not perform I/O, but it is
 465                  * possible for a page to have PageDirty set, but it is actually
 466                  * clean (all its buffers are clean).  This happens if the
 467                  * buffers were written out directly, with submit_bh(). ext3
 468                  * will do this, as well as the blockdev mapping.
 469                  * try_to_release_page() will discover that cleanness and will
 470                  * drop the buffers and mark the page clean - it can be freed.
 471                  *
 472                  * Rarely, pages can have buffers and no ->mapping.  These are
 473                  * the pages which were not successfully invalidated in
 474                  * truncate_complete_page().  We try to drop those buffers here
 475                  * and if that worked, and the page is no longer mapped into
 476                  * process address space (page_count == 1) it can be freed.
 477                  * Otherwise, leave the page on the LRU so it is swappable.
 478                  */
 479                 if (PagePrivate(page)) {
 480                         if (!try_to_release_page(page, sc->gfp_mask))
 481                                 goto activate_locked;
 482                         if (!mapping && page_count(page) == 1)
 483                                 goto free_it;
 484                 }
 485
 486                 if (!mapping)
 487                         goto keep_locked;       /* truncate got there first */
 488
 489                 spin_lock_irq(&mapping->tree_lock);
 490
 491                 /*
 492                  * The non-racy check for busy page.  It is critical to check
 493                  * PageDirty _after_ making sure that the page is freeable and
 494                  * not in use by anybody.       (pagecache + us == 2)
 495                  */
 496                 if (page_count(page) != 2 || PageDirty(page)) {
 497                         spin_unlock_irq(&mapping->tree_lock);
 498                         goto keep_locked;
 499                 }
 500
 501 #ifdef CONFIG_SWAP
 502                 if (PageSwapCache(page)) {
 503                         swp_entry_t swap = { .val = page->private };
 504                         __delete_from_swap_cache(page);
 505                         spin_unlock_irq(&mapping->tree_lock);
 506                         swap_free(swap);
 507                         __put_page(page);       /* The pagecache ref */
 508                         goto free_it;
 509                 }
 510 #endif /* CONFIG_SWAP */
 511
 512                 __remove_from_page_cache(page);
 513                 spin_unlock_irq(&mapping->tree_lock);
 514                 __put_page(page);
 515
 516 free_it:
 517                 unlock_page(page);
 518                 reclaimed++;
 519                 if (!pagevec_add(&freed_pvec, page))
 520                         __pagevec_release_nonlru(&freed_pvec);
 521                 continue;
 522
 523 activate_locked:
 524                 SetPageActive(page);
 525                 pgactivate++;
 526 keep_locked:
 527                 unlock_page(page);
 528 keep:
 529                 list_add(&page->lru, &ret_pages);
 530                 BUG_ON(PageLRU(page));
 531         }
 532         list_splice(&ret_pages, page_list);
 533         if (pagevec_count(&freed_pvec))
 534                 __pagevec_release_nonlru(&freed_pvec);
 535         mod_page_state(pgactivate, pgactivate);
 536         sc->nr_reclaimed += reclaimed;
 537         return reclaimed;
 538 }
 539
 540 /*
 541  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 542  * a batch of pages and working on them outside the lock.  Any pages which were
 543  * not freed will be added back to the LRU.
 544  *
 545  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 546  *
 547  * For pagecache intensive workloads, the first loop here is the hottest spot
 548  * in the kernel (apart from the copy_*_user functions).
 549  */
 550 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 551 {
 552         LIST_HEAD(page_list);
 553         struct pagevec pvec;
 554         int max_scan = sc->nr_to_scan, nr_pass;
 555         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 556
 557         pagevec_init(&pvec, 1);
 558
 559         lru_add_drain();
 560         spin_lock_irq(&zone->lru_lock);
 561 redo:
 562         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 563         nr_pass = zone->nr_inactive;
 564         while (max_scan > 0) {
 565                 struct page *page;
 566                 int nr_taken = 0;
 567                 int nr_scan = 0;
 568                 int nr_freed;
 569
 570                 while (nr_pass-- && nr_scan++ < SWAP_CLUSTER_MAX &&
 571                                 !list_empty(&zone->inactive_list)) {
 572                         page = lru_to_page(&zone->inactive_list);
 573
 574                         prefetchw_prev_lru_page(page,
 575                                                 &zone->inactive_list, flags);
 576
 577                         if (!TestClearPageLRU(page))
 578                                 BUG();
 579                         list_del(&page->lru);
 580                         if (get_page_testone(page)) {
 581                                 /*
 582                                  * It is being freed elsewhere
 583                                  */
 584                                 __put_page(page);
 585                                 SetPageLRU(page);
 586                                 list_add(&page->lru, &zone->inactive_list);
 587                                 continue;
 588                         } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 589                                 __put_page(page);
 590                                 SetPageLRU(page);
 591 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 592                                 list_add_tail(&page->lru, &zone->inactive_list);
 593 #else
 594                                 list_add(&page->lru, &zone->inactive_list);
 595 #endif
 596                                 continue;
 597                         }
 598                         list_add(&page->lru, &page_list);
 599                         ckrm_mem_dec_inactive(page);
 600                         nr_taken++;
 601                 }
 602                 zone->nr_inactive -= nr_taken;
 603                 zone->pages_scanned += nr_taken;
 604                 spin_unlock_irq(&zone->lru_lock);
 605
 606                 if ((bit_flag == 0) && (nr_taken == 0))
 607                         goto done;
 608
 609                 max_scan -= nr_scan;
 610                 if (current_is_kswapd())
 611                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 612                 else
 613                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 614                 nr_freed = shrink_list(&page_list, sc);
 615                 if (current_is_kswapd())
 616                         mod_page_state(kswapd_steal, nr_freed);
 617                 mod_page_state_zone(zone, pgsteal, nr_freed);
 618                 sc->nr_to_reclaim -= nr_freed;
 619
 620                 spin_lock_irq(&zone->lru_lock);
 621                 /*
 622                  * Put back any unfreeable pages.
 623                  */
 624                 while (!list_empty(&page_list)) {
 625                         page = lru_to_page(&page_list);
 626                         if (TestSetPageLRU(page))
 627                                 BUG();
 628                         list_del(&page->lru);
 629                         if (PageActive(page))
 630                                 add_page_to_active_list(zone, page);
 631                         else
 632                                 add_page_to_inactive_list(zone, page);
 633                         if (!pagevec_add(&pvec, page)) {
 634                                 spin_unlock_irq(&zone->lru_lock);
 635                                 __pagevec_release(&pvec);
 636                                 spin_lock_irq(&zone->lru_lock);
 637                         }
 638                 }
 639                 if (ckrm_flags && (nr_pass <= 0)) {
 640                         goto redo;
 641                 }
 642         }
 643         spin_unlock_irq(&zone->lru_lock);
 644 done:
 645         pagevec_release(&pvec);
 646 }
 647
 648 /*
 649  * This moves pages from the active list to the inactive list.
 650  *
 651  * We move them the other way if the page is referenced by one or more
 652  * processes, from rmap.
 653  *
 654  * If the pages are mostly unmapped, the processing is fast and it is
 655  * appropriate to hold zone->lru_lock across the whole operation.  But if
 656  * the pages are mapped, the processing is slow (page_referenced()) so we
 657  * should drop zone->lru_lock around each page.  It's impossible to balance
 658  * this, so instead we remove the pages from the LRU while processing them.
 659  * It is safe to rely on PG_active against the non-LRU pages in here because
 660  * nobody will play with that bit on a non-LRU page.
 661  *
 662  * The downside is that we have to touch page->_count against each page.
 663  * But we had to alter page->flags anyway.
 664  */
 665 static void
 666 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 667 {
 668         int pgmoved;
 669         int pgdeactivate = 0;
 670         int pgscanned = 0;
 671         int nr_pages = sc->nr_to_scan;
 672         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 673         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 674         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 675         struct page *page;
 676         struct pagevec pvec;
 677         int reclaim_mapped = 0;
 678         long mapped_ratio;
 679         long distress;
 680         long swap_tendency;
 681         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 682         int nr_pass;
 683
 684         lru_add_drain();
 685         pgmoved = 0;
 686         spin_lock_irq(&zone->lru_lock);
 687 redo:
 688         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 689         nr_pass = zone->nr_active;
 690         while (pgscanned < nr_pages && !list_empty(&zone->active_list) &&
 691                                                 nr_pass) {
 692                 page = lru_to_page(&zone->active_list);
 693                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 694                 if (!TestClearPageLRU(page))
 695                         BUG();
 696                 list_del(&page->lru);
 697                 if (get_page_testone(page)) {
 698                         /*
 699                          * It was already free!  release_pages() or put_page()
 700                          * are about to remove it from the LRU and free it. So
 701                          * put the refcount back and put the page back on the
 702                          * LRU
 703                          */
 704                         __put_page(page);
 705                         SetPageLRU(page);
 706                         list_add(&page->lru, &zone->active_list);
 707                         pgscanned++;
 708                 } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 709                         __put_page(page);
 710                         SetPageLRU(page);
 711 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 712                         list_add_tail(&page->lru, &zone->active_list);
 713 #else
 714                         list_add(&page->lru, &zone->active_list);
 715 #endif
 716                 } else {
 717                         list_add(&page->lru, &l_hold);
 718                         ckrm_mem_dec_active(page);
 719                         pgmoved++;
 720                         pgscanned++;
 721                 }
 722                 if (!--nr_pass && ckrm_flags) {
 723                         goto redo;
 724                 }
 725         }
 726         zone->nr_active -= pgmoved;
 727         spin_unlock_irq(&zone->lru_lock);
 728
 729         /*
 730          * `distress' is a measure of how much trouble we're having reclaiming
 731          * pages.  0 -> no problems.  100 -> great trouble.
 732          */
 733         distress = 100 >> zone->prev_priority;
 734
 735         /*
 736          * The point of this algorithm is to decide when to start reclaiming
 737          * mapped memory instead of just pagecache.  Work out how much memory
 738          * is mapped.
 739          */
 740         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 741
 742         /*
 743          * Now decide how much we really want to unmap some pages.  The mapped
 744          * ratio is downgraded - just because there's a lot of mapped memory
 745          * doesn't necessarily mean that page reclaim isn't succeeding.
 746          *
 747          * The distress ratio is important - we don't want to start going oom.
 748          *
 749          * A 100% value of vm_swappiness overrides this algorithm altogether.
 750          */
 751         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 752
 753         /*
 754          * Now use this metric to decide whether to start moving mapped memory
 755          * onto the inactive list.
 756          */
 757         if (swap_tendency >= 100)
 758                 reclaim_mapped = 1;
 759
 760         while (!list_empty(&l_hold)) {
 761                 cond_resched();
 762                 page = lru_to_page(&l_hold);
 763                 list_del(&page->lru);
 764                 if (page_mapped(page)) {
 765                         if (!reclaim_mapped ||
 766                             (total_swap_pages == 0 && PageAnon(page)) ||
 767                             page_referenced(page, 0, sc->priority <= 0)) {
 768                                 list_add(&page->lru, &l_active);
 769                                 continue;
 770                         }
 771                 }
 772                 list_add(&page->lru, &l_inactive);
 773         }
 774
 775         pagevec_init(&pvec, 1);
 776         pgmoved = 0;
 777         spin_lock_irq(&zone->lru_lock);
 778         while (!list_empty(&l_inactive)) {
 779                 page = lru_to_page(&l_inactive);
 780                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 781                 if (TestSetPageLRU(page))
 782                         BUG();
 783                 if (!TestClearPageActive(page))
 784                         BUG();
 785                 list_move(&page->lru, &zone->inactive_list);
 786                 ckrm_mem_inc_inactive(page);
 787                 pgmoved++;
 788                 if (!pagevec_add(&pvec, page)) {
 789                         zone->nr_inactive += pgmoved;
 790                         spin_unlock_irq(&zone->lru_lock);
 791                         pgdeactivate += pgmoved;
 792                         pgmoved = 0;
 793                         if (buffer_heads_over_limit)
 794                                 pagevec_strip(&pvec);
 795                         __pagevec_release(&pvec);
 796                         spin_lock_irq(&zone->lru_lock);
 797                 }
 798         }
 799         zone->nr_inactive += pgmoved;
 800         pgdeactivate += pgmoved;
 801         if (buffer_heads_over_limit) {
 802                 spin_unlock_irq(&zone->lru_lock);
 803                 pagevec_strip(&pvec);
 804                 spin_lock_irq(&zone->lru_lock);
 805         }
 806
 807         pgmoved = 0;
 808         while (!list_empty(&l_active)) {
 809                 page = lru_to_page(&l_active);
 810                 prefetchw_prev_lru_page(page, &l_active, flags);
 811                 if (TestSetPageLRU(page))
 812                         BUG();
 813                 BUG_ON(!PageActive(page));
 814                 list_move(&page->lru, &zone->active_list);
 815                 ckrm_mem_inc_active(page);
 816                 pgmoved++;
 817                 if (!pagevec_add(&pvec, page)) {
 818                         zone->nr_active += pgmoved;
 819                         pgmoved = 0;
 820                         spin_unlock_irq(&zone->lru_lock);
 821                         __pagevec_release(&pvec);
 822                         spin_lock_irq(&zone->lru_lock);
 823                 }
 824         }
 825         zone->nr_active += pgmoved;
 826         spin_unlock_irq(&zone->lru_lock);
 827         pagevec_release(&pvec);
 828
 829         mod_page_state_zone(zone, pgrefill, pgscanned);
 830         mod_page_state(pgdeactivate, pgdeactivate);
 831 }
 832
 833 /*
 834  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 835  */
 836 static void
 837 shrink_zone(struct zone *zone, struct scan_control *sc)
 838 {
 839         unsigned long nr_active;
 840         unsigned long nr_inactive;
 841
 842         /*
 843          * Add one to `nr_to_scan' just to make sure that the kernel will
 844          * slowly sift through the active list.
 845          */
 846         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 847         nr_active = zone->nr_scan_active;
 848         if (nr_active >= SWAP_CLUSTER_MAX)
 849                 zone->nr_scan_active = 0;
 850         else
 851                 nr_active = 0;
 852
 853         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 854         nr_inactive = zone->nr_scan_inactive;
 855         if (nr_inactive >= SWAP_CLUSTER_MAX)
 856                 zone->nr_scan_inactive = 0;
 857         else
 858                 nr_inactive = 0;
 859
 860         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 861
 862         while (nr_active || nr_inactive) {
 863                 sc->ckrm_flags = ckrm_setup_reclamation();
 864                 if (nr_active) {
 865                         sc->nr_to_scan = min(nr_active,
 866                                         (unsigned long)SWAP_CLUSTER_MAX);
 867                         nr_active -= sc->nr_to_scan;
 868                         refill_inactive_zone(zone, sc);
 869                 }
 870
 871                 if (nr_inactive) {
 872                         sc->nr_to_scan = min(nr_inactive,
 873                                         (unsigned long)SWAP_CLUSTER_MAX);
 874                         nr_inactive -= sc->nr_to_scan;
 875                         shrink_cache(zone, sc);
 876                         if (sc->nr_to_reclaim <= 0)
 877                                 break;
 878                 }
 879                 ckrm_teardown_reclamation();
 880         }
 881 }
 882
 883 #if defined(CONFIG_CKRM_RES_MEM) && defined(AT_LIMIT_SUPPORT)
 884 // This function needs to be given more thought.
 885 // Shrink the class to be at 90% of its limit
 886 static void
 887 ckrm_shrink_class(ckrm_mem_res_t *cls)
 888 {
 889         struct scan_control sc;
 890         struct zone *zone;
 891         int zindex = 0, active_credit = 0, inactive_credit = 0;
 892
 893         if (ckrm_test_set_shrink(cls)) { // set the SHRINK bit atomically
 894                 // if it is already set somebody is working on it. so... leave
 895                 return;
 896         }
 897         sc.nr_mapped = read_page_state(nr_mapped);
 898         sc.nr_scanned = 0;
 899         sc.ckrm_flags = ckrm_get_reclaim_flags(cls);
 900         sc.nr_reclaimed = 0;
 901         sc.priority = 0; // always very high priority
 902
 903         for_each_zone(zone) {
 904                 int zone_total, zone_limit, active_limit, inactive_limit;
 905                 int active_over, inactive_over;
 906                 unsigned long nr_active, nr_inactive;
 907                 u64 temp;
 908
 909                 zone->temp_priority = zone->prev_priority;
 910                 zone->prev_priority = sc.priority;
 911
 912                 zone_total = zone->nr_active + zone->nr_inactive + zone->free_pages;
 913
 914                 temp = (u64) cls->pg_limit * zone_total;
 915                 do_div(temp, ckrm_tot_lru_pages);
 916                 zone_limit = (int) temp;
 917                 active_limit = (6 * zone_limit) / 10; // 2/3rd in active list
 918                 inactive_limit = (3 * zone_limit) / 10; // 1/3rd in inactive list
 919
 920                 active_over = cls->nr_active[zindex] - active_limit + active_credit;
 921                 inactive_over = active_over +
 922                                 (cls->nr_inactive[zindex] - inactive_limit) + inactive_credit;
 923
 924                 if (active_over > 0) {
 925                         zone->nr_scan_active += active_over + 1;
 926                         nr_active = zone->nr_scan_active;
 927                         active_credit = 0;
 928                 } else {
 929                         active_credit += active_over;
 930                         nr_active = 0;
 931                 }
 932
 933                 if (inactive_over > 0) {
 934                         zone->nr_scan_inactive += inactive_over;
 935                         nr_inactive = zone->nr_scan_inactive;
 936                         inactive_credit = 0;
 937                 } else {
 938                         inactive_credit += inactive_over;
 939                         nr_inactive = 0;
 940                 }
 941                 while (nr_active || nr_inactive) {
 942                         if (nr_active) {
 943                                 sc.nr_to_scan = min(nr_active,
 944                                                 (unsigned long)SWAP_CLUSTER_MAX);
 945                                 nr_active -= sc.nr_to_scan;
 946                                 refill_inactive_zone(zone, &sc);
 947                         }
 948
 949                         if (nr_inactive) {
 950                                 sc.nr_to_scan = min(nr_inactive,
 951                                                 (unsigned long)SWAP_CLUSTER_MAX);
 952                                 nr_inactive -= sc.nr_to_scan;
 953                                 shrink_cache(zone, &sc);
 954                                 if (sc.nr_to_reclaim <= 0)
 955                                         break;
 956                         }
 957                 }
 958                 zone->prev_priority = zone->temp_priority;
 959                 zindex++;
 960         }
 961         ckrm_clear_shrink(cls);
 962 }
 963
 964 static void
 965 ckrm_shrink_classes(void)
 966 {
 967         ckrm_mem_res_t *cls;
 968
 969         spin_lock(&ckrm_mem_lock);
 970         while (!ckrm_shrink_list_empty()) {
 971                 cls =  list_entry(ckrm_shrink_list.next, ckrm_mem_res_t,
 972                                 shrink_list);
 973                 spin_unlock(&ckrm_mem_lock);
 974                 ckrm_shrink_class(cls);
 975                 spin_lock(&ckrm_mem_lock);
 976                 list_del(&cls->shrink_list);
 977                 cls->flags &= ~MEM_AT_LIMIT;
 978         }
 979         spin_unlock(&ckrm_mem_lock);
 980         throttle_vm_writeout();
 981 }
 982
 983 #else
 984
 985 #if defined(CONFIG_CKRM_RES_MEM) && !defined(AT_LIMIT_SUPPORT)
 986 #warning "disabling ckrm_at_limit -- setting ckrm_shrink_classes to noop "
 987 #endif
 988
 989 #define ckrm_shrink_classes()   do { } while(0)
 990 #endif
 991
 992 /*
 993  * This is the direct reclaim path, for page-allocating processes.  We only
 994  * try to reclaim pages from zones which will satisfy the caller's allocation
 995  * request.
 996  *
 997  * We reclaim from a zone even if that zone is over pages_high.  Because:
 998  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 999  *    allocation or
1000  * b) The zones may be over pages_high but they must go *over* pages_high to
1001  *    satisfy the `incremental min' zone defense algorithm.
1002  *
1003  * Returns the number of reclaimed pages.
1004  *
1005  * If a zone is deemed to be full of pinned pages then just give it a light
1006  * scan then give up on it.
1007  */
1008 static void
1009 shrink_caches(struct zone **zones, struct scan_control *sc)
1010 {
1011         int i;
1012
1013         for (i = 0; zones[i] != NULL; i++) {
1014                 struct zone *zone = zones[i];
1015
1016                 if (zone->present_pages == 0)
1017                         continue;
1018
1019                 zone->temp_priority = sc->priority;
1020                 if (zone->prev_priority > sc->priority)
1021                         zone->prev_priority = sc->priority;
1022
1023                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1024                         continue;       /* Let kswapd poll it */
1025
1026                 shrink_zone(zone, sc);
1027         }
1028 }
1029
1030 /*
1031  * This is the main entry point to direct page reclaim.
1032  *
1033  * If a full scan of the inactive list fails to free enough memory then we
1034  * are "out of memory" and something needs to be killed.
1035  *
1036  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1037  * high - the zone may be full of dirty or under-writeback pages, which this
1038  * caller can't do much about.  We kick pdflush and take explicit naps in the
1039  * hope that some of these pages can be written.  But if the allocating task
1040  * holds filesystem locks which prevent writeout this might not work, and the
1041  * allocation attempt will fail.
1042  */
1043 int try_to_free_pages(struct zone **zones,
1044                 unsigned int gfp_mask, unsigned int order)
1045 {
1046         int priority;
1047         int ret = 0;
1048         int total_scanned = 0, total_reclaimed = 0;
1049         struct reclaim_state *reclaim_state = current->reclaim_state;
1050         struct scan_control sc;
1051         unsigned long lru_pages = 0;
1052         int i;
1053
1054         sc.gfp_mask = gfp_mask;
1055         sc.may_writepage = 0;
1056
1057         inc_page_state(allocstall);
1058
1059         for (i = 0; zones[i] != NULL; i++) {
1060                 struct zone *zone = zones[i];
1061
1062                 zone->temp_priority = DEF_PRIORITY;
1063                 lru_pages += zone->nr_active + zone->nr_inactive;
1064         }
1065
1066         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1067                 sc.nr_mapped = read_page_state(nr_mapped);
1068                 sc.nr_scanned = 0;
1069                 sc.nr_reclaimed = 0;
1070                 sc.priority = priority;
1071                 shrink_caches(zones, &sc);
1072                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1073                 if (reclaim_state) {
1074                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1075                         reclaim_state->reclaimed_slab = 0;
1076                 }
1077                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1078                         ret = 1;
1079                         goto out;
1080                 }
1081                 total_scanned += sc.nr_scanned;
1082                 total_reclaimed += sc.nr_reclaimed;
1083
1084                 /*
1085                  * Try to write back as many pages as we just scanned.  This
1086                  * tends to cause slow streaming writers to write data to the
1087                  * disk smoothly, at the dirtying rate, which is nice.   But
1088                  * that's undesirable in laptop mode, where we *want* lumpy
1089                  * writeout.  So in laptop mode, write out the whole world.
1090                  */
1091                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1092                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1093                         sc.may_writepage = 1;
1094                 }
1095
1096                 /* Take a nap, wait for some writeback to complete */
1097                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1098                         blk_congestion_wait(WRITE, HZ/10);
1099         }
1100         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1101                 out_of_memory(gfp_mask);
1102 out:
1103         for (i = 0; zones[i] != 0; i++)
1104                 zones[i]->prev_priority = zones[i]->temp_priority;
1105         return ret;
1106 }
1107
1108 /*
1109  * For kswapd, balance_pgdat() will work across all this node's zones until
1110  * they are all at pages_high.
1111  *
1112  * If `nr_pages' is non-zero then it is the number of pages which are to be
1113  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1114  * special.
1115  *
1116  * Returns the number of pages which were actually freed.
1117  *
1118  * There is special handling here for zones which are full of pinned pages.
1119  * This can happen if the pages are all mlocked, or if they are all used by
1120  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1121  * What we do is to detect the case where all pages in the zone have been
1122  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1123  * dead and from now on, only perform a short scan.  Basically we're polling
1124  * the zone for when the problem goes away.
1125  *
1126  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1127  * zones which have free_pages > pages_high, but once a zone is found to have
1128  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1129  * of the number of free pages in the lower zones.  This interoperates with
1130  * the page allocator fallback scheme to ensure that aging of pages is balanced
1131  * across the zones.
1132  */
1133 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1134 {
1135         int to_free = nr_pages;
1136         int all_zones_ok;
1137         int priority;
1138         int i;
1139         int total_scanned, total_reclaimed;
1140         struct reclaim_state *reclaim_state = current->reclaim_state;
1141         struct scan_control sc;
1142
1143 loop_again:
1144         total_scanned = 0;
1145         total_reclaimed = 0;
1146         sc.gfp_mask = GFP_KERNEL;
1147         sc.may_writepage = 0;
1148         sc.nr_mapped = read_page_state(nr_mapped);
1149
1150         inc_page_state(pageoutrun);
1151
1152         for (i = 0; i < pgdat->nr_zones; i++) {
1153                 struct zone *zone = pgdat->node_zones + i;
1154
1155                 zone->temp_priority = DEF_PRIORITY;
1156         }
1157
1158         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1159                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1160                 unsigned long lru_pages = 0;
1161
1162                 all_zones_ok = 1;
1163
1164                 if (nr_pages == 0) {
1165                         /*
1166                          * Scan in the highmem->dma direction for the highest
1167                          * zone which needs scanning
1168                          */
1169                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1170                                 struct zone *zone = pgdat->node_zones + i;
1171
1172                                 if (zone->present_pages == 0)
1173                                         continue;
1174
1175                                 if (zone->all_unreclaimable &&
1176                                                 priority != DEF_PRIORITY)
1177                                         continue;
1178
1179                                 if (zone->free_pages <= zone->pages_high) {
1180                                         end_zone = i;
1181                                         goto scan;
1182                                 }
1183                         }
1184                         goto out;
1185                 } else {
1186                         end_zone = pgdat->nr_zones - 1;
1187                 }
1188 scan:
1189                 for (i = 0; i <= end_zone; i++) {
1190                         struct zone *zone = pgdat->node_zones + i;
1191
1192                         lru_pages += zone->nr_active + zone->nr_inactive;
1193                 }
1194
1195                 /*
1196                  * Now scan the zone in the dma->highmem direction, stopping
1197                  * at the last zone which needs scanning.
1198                  *
1199                  * We do this because the page allocator works in the opposite
1200                  * direction.  This prevents the page allocator from allocating
1201                  * pages behind kswapd's direction of progress, which would
1202                  * cause too much scanning of the lower zones.
1203                  */
1204                 for (i = 0; i <= end_zone; i++) {
1205                         struct zone *zone = pgdat->node_zones + i;
1206
1207                         if (zone->present_pages == 0)
1208                                 continue;
1209
1210                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1211                                 continue;
1212
1213                         if (nr_pages == 0) {    /* Not software suspend */
1214                                 if (zone->free_pages <= zone->pages_high)
1215                                         all_zones_ok = 0;
1216                         }
1217                         zone->temp_priority = priority;
1218                         if (zone->prev_priority > priority)
1219                                 zone->prev_priority = priority;
1220                         sc.nr_scanned = 0;
1221                         sc.nr_reclaimed = 0;
1222                         sc.priority = priority;
1223                         shrink_zone(zone, &sc);
1224                         reclaim_state->reclaimed_slab = 0;
1225                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1226                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1227                         total_reclaimed += sc.nr_reclaimed;
1228                         if (zone->all_unreclaimable)
1229                                 continue;
1230                         if (zone->pages_scanned >= (zone->nr_active +
1231                                                         zone->nr_inactive) * 4)
1232                                 zone->all_unreclaimable = 1;
1233                         /*
1234                          * If we've done a decent amount of scanning and
1235                          * the reclaim ratio is low, start doing writepage
1236                          * even in laptop mode
1237                          */
1238                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1239                             total_scanned > total_reclaimed+total_reclaimed/2)
1240                                 sc.may_writepage = 1;
1241                 }
1242                 if (nr_pages && to_free > total_reclaimed)
1243                         continue;       /* swsusp: need to do more work */
1244                 if (all_zones_ok)
1245                         break;          /* kswapd: all done */
1246                 /*
1247                  * OK, kswapd is getting into trouble.  Take a nap, then take
1248                  * another pass across the zones.
1249                  */
1250                 if (total_scanned && priority < DEF_PRIORITY - 2)
1251                         blk_congestion_wait(WRITE, HZ/10);
1252
1253                 /*
1254                  * We do this so kswapd doesn't build up large priorities for
1255                  * example when it is freeing in parallel with allocators. It
1256                  * matches the direct reclaim path behaviour in terms of impact
1257                  * on zone->*_priority.
1258                  */
1259                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1260                         break;
1261         }
1262 out:
1263         for (i = 0; i < pgdat->nr_zones; i++) {
1264                 struct zone *zone = pgdat->node_zones + i;
1265
1266                 zone->prev_priority = zone->temp_priority;
1267         }
1268         if (!all_zones_ok) {
1269                 cond_resched();
1270                 goto loop_again;
1271         }
1272
1273         return total_reclaimed;
1274 }
1275
1276 /*
1277  * The background pageout daemon, started as a kernel thread
1278  * from the init process.
1279  *
1280  * This basically trickles out pages so that we have _some_
1281  * free memory available even if there is no other activity
1282  * that frees anything up. This is needed for things like routing
1283  * etc, where we otherwise might have all activity going on in
1284  * asynchronous contexts that cannot page things out.
1285  *
1286  * If there are applications that are active memory-allocators
1287  * (most normal use), this basically shouldn't matter.
1288  */
1289 static int kswapd(void *p)
1290 {
1291         pg_data_t *pgdat = (pg_data_t*)p;
1292         struct task_struct *tsk = current;
1293         DEFINE_WAIT(wait);
1294         struct reclaim_state reclaim_state = {
1295                 .reclaimed_slab = 0,
1296         };
1297         cpumask_t cpumask;
1298
1299         daemonize("kswapd%d", pgdat->node_id);
1300         cpumask = node_to_cpumask(pgdat->node_id);
1301         if (!cpus_empty(cpumask))
1302                 set_cpus_allowed(tsk, cpumask);
1303         current->reclaim_state = &reclaim_state;
1304
1305         /*
1306          * Tell the memory management that we're a "memory allocator",
1307          * and that if we need more memory we should get access to it
1308          * regardless (see "__alloc_pages()"). "kswapd" should
1309          * never get caught in the normal page freeing logic.
1310          *
1311          * (Kswapd normally doesn't need memory anyway, but sometimes
1312          * you need a small amount of memory in order to be able to
1313          * page out something else, and this flag essentially protects
1314          * us from recursively trying to free more memory as we're
1315          * trying to free the first piece of memory in the first place).
1316          */
1317         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1318
1319         for ( ; ; ) {
1320                 if (current->flags & PF_FREEZE)
1321                         refrigerator(PF_FREEZE);
1322                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1323                 schedule();
1324                 finish_wait(&pgdat->kswapd_wait, &wait);
1325
1326                 if (!ckrm_shrink_list_empty())
1327                         ckrm_shrink_classes();
1328                 else
1329                 balance_pgdat(pgdat, 0);
1330         }
1331         return 0;
1332 }
1333
1334 /*
1335  * A zone is low on free memory, so wake its kswapd task to service it.
1336  */
1337 void wakeup_kswapd(struct zone *zone)
1338 {
1339         if (zone->present_pages == 0)
1340                 return;
1341         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1342                 return;
1343         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1344                 return;
1345         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1346 }
1347
1348 #ifdef CONFIG_PM
1349 /*
1350  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1351  * pages.
1352  */
1353 int shrink_all_memory(int nr_pages)
1354 {
1355         pg_data_t *pgdat;
1356         int nr_to_free = nr_pages;
1357         int ret = 0;
1358         struct reclaim_state reclaim_state = {
1359                 .reclaimed_slab = 0,
1360         };
1361
1362         current->reclaim_state = &reclaim_state;
1363         for_each_pgdat(pgdat) {
1364                 int freed;
1365                 freed = balance_pgdat(pgdat, nr_to_free);
1366                 ret += freed;
1367                 nr_to_free -= freed;
1368                 if (nr_to_free <= 0)
1369                         break;
1370         }
1371         current->reclaim_state = NULL;
1372         return ret;
1373 }
1374 #endif
1375
1376 #ifdef CONFIG_HOTPLUG_CPU
1377 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1378    not required for correctness.  So if the last cpu in a node goes
1379    away, we get changed to run anywhere: as the first one comes back,
1380    restore their cpu bindings. */
1381 static int __devinit cpu_callback(struct notifier_block *nfb,
1382                                   unsigned long action,
1383                                   void *hcpu)
1384 {
1385         pg_data_t *pgdat;
1386         cpumask_t mask;
1387
1388         if (action == CPU_ONLINE) {
1389                 for_each_pgdat(pgdat) {
1390                         mask = node_to_cpumask(pgdat->node_id);
1391                         if (any_online_cpu(mask) != NR_CPUS)
1392                                 /* One of our CPUs online: restore mask */
1393                                 set_cpus_allowed(pgdat->kswapd, mask);
1394                 }
1395         }
1396         return NOTIFY_OK;
1397 }
1398 #endif /* CONFIG_HOTPLUG_CPU */
1399
1400 static int __init kswapd_init(void)
1401 {
1402         pg_data_t *pgdat;
1403         swap_setup();
1404         for_each_pgdat(pgdat)
1405                 pgdat->kswapd
1406                 = find_task_by_real_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1407         total_memory = nr_free_pagecache_pages();
1408         hotcpu_notifier(cpu_callback, 0);
1409         return 0;
1410 }
1411
1412 module_init(kswapd_init)