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/blkdev.h>
  25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  26                                         buffer_heads_over_limit */
  27 #include <linux/mm_inline.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/rmap.h>
  31 #include <linux/topology.h>
  32 #include <linux/cpu.h>
  33 #include <linux/cpuset.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
  42 /* possible outcome of pageout() */
  43 typedef enum {
  44         /* failed to write page out, page is locked */
  45         PAGE_KEEP,
  46         /* move page to the active list, page is locked */
  47         PAGE_ACTIVATE,
  48         /* page has been sent to the disk successfully, page is unlocked */
  49         PAGE_SUCCESS,
  50         /* page is clean and locked */
  51         PAGE_CLEAN,
  52 } pageout_t;
  53
  54 struct scan_control {
  55         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  56         unsigned long nr_to_scan;
  57
  58         /* Incremented by the number of inactive pages that were scanned */
  59         unsigned long nr_scanned;
  60
  61         /* Incremented by the number of pages reclaimed */
  62         unsigned long nr_reclaimed;
  63
  64         unsigned long nr_mapped;        /* From page_state */
  65
  66         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  67         unsigned int priority;
  68
  69         /* This context's GFP mask */
  70         gfp_t gfp_mask;
  71
  72         int may_writepage;
  73
  74         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  75          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  76          * In this context, it doesn't matter that we scan the
  77          * whole list at once. */
  78         int swap_cluster_max;
  79 };
  80
  81 /*
  82  * The list of shrinker callbacks used by to apply pressure to
  83  * ageable caches.
  84  */
  85 struct shrinker {
  86         shrinker_t              shrinker;
  87         struct list_head        list;
  88         int                     seeks;  /* seeks to recreate an obj */
  89         long                    nr;     /* objs pending delete */
  90 };
  91
  92 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  93
  94 #ifdef ARCH_HAS_PREFETCH
  95 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  96         do {                                                            \
  97                 if ((_page)->lru.prev != _base) {                       \
  98                         struct page *prev;                              \
  99                                                                         \
 100                         prev = lru_to_page(&(_page->lru));              \
 101                         prefetch(&prev->_field);                        \
 102                 }                                                       \
 103         } while (0)
 104 #else
 105 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 106 #endif
 107
 108 #ifdef ARCH_HAS_PREFETCHW
 109 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 110         do {                                                            \
 111                 if ((_page)->lru.prev != _base) {                       \
 112                         struct page *prev;                              \
 113                                                                         \
 114                         prev = lru_to_page(&(_page->lru));              \
 115                         prefetchw(&prev->_field);                       \
 116                 }                                                       \
 117         } while (0)
 118 #else
 119 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 120 #endif
 121
 122 /*
 123  * From 0 .. 100.  Higher means more swappy.
 124  */
 125 int vm_swappiness = 60;
 126 static long total_memory;
 127
 128 static LIST_HEAD(shrinker_list);
 129 static DECLARE_RWSEM(shrinker_rwsem);
 130
 131 /*
 132  * Add a shrinker callback to be called from the vm
 133  */
 134 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 135 {
 136         struct shrinker *shrinker;
 137
 138         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 139         if (shrinker) {
 140                 shrinker->shrinker = theshrinker;
 141                 shrinker->seeks = seeks;
 142                 shrinker->nr = 0;
 143                 down_write(&shrinker_rwsem);
 144                 list_add_tail(&shrinker->list, &shrinker_list);
 145                 up_write(&shrinker_rwsem);
 146         }
 147         return shrinker;
 148 }
 149 EXPORT_SYMBOL(set_shrinker);
 150
 151 /*
 152  * Remove one
 153  */
 154 void remove_shrinker(struct shrinker *shrinker)
 155 {
 156         down_write(&shrinker_rwsem);
 157         list_del(&shrinker->list);
 158         up_write(&shrinker_rwsem);
 159         kfree(shrinker);
 160 }
 161 EXPORT_SYMBOL(remove_shrinker);
 162
 163 #define SHRINK_BATCH 128
 164 /*
 165  * Call the shrink functions to age shrinkable caches
 166  *
 167  * Here we assume it costs one seek to replace a lru page and that it also
 168  * takes a seek to recreate a cache object.  With this in mind we age equal
 169  * percentages of the lru and ageable caches.  This should balance the seeks
 170  * generated by these structures.
 171  *
 172  * If the vm encounted mapped pages on the LRU it increase the pressure on
 173  * slab to avoid swapping.
 174  *
 175  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 176  *
 177  * `lru_pages' represents the number of on-LRU pages in all the zones which
 178  * are eligible for the caller's allocation attempt.  It is used for balancing
 179  * slab reclaim versus page reclaim.
 180  *
 181  * Returns the number of slab objects which we shrunk.
 182  */
 183 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
 184 {
 185         struct shrinker *shrinker;
 186         int ret = 0;
 187
 188         if (scanned == 0)
 189                 scanned = SWAP_CLUSTER_MAX;
 190
 191         if (!down_read_trylock(&shrinker_rwsem))
 192                 return 1;       /* Assume we'll be able to shrink next time */
 193
 194         list_for_each_entry(shrinker, &shrinker_list, list) {
 195                 unsigned long long delta;
 196                 unsigned long total_scan;
 197                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
 198
 199                 delta = (4 * scanned) / shrinker->seeks;
 200                 delta *= max_pass;
 201                 do_div(delta, lru_pages + 1);
 202                 shrinker->nr += delta;
 203                 if (shrinker->nr < 0) {
 204                         printk(KERN_ERR "%s: nr=%ld\n",
 205                                         __FUNCTION__, shrinker->nr);
 206                         shrinker->nr = max_pass;
 207                 }
 208
 209                 /*
 210                  * Avoid risking looping forever due to too large nr value:
 211                  * never try to free more than twice the estimate number of
 212                  * freeable entries.
 213                  */
 214                 if (shrinker->nr > max_pass * 2)
 215                         shrinker->nr = max_pass * 2;
 216
 217                 total_scan = shrinker->nr;
 218                 shrinker->nr = 0;
 219
 220                 while (total_scan >= SHRINK_BATCH) {
 221                         long this_scan = SHRINK_BATCH;
 222                         int shrink_ret;
 223                         int nr_before;
 224
 225                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
 226                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 227                         if (shrink_ret == -1)
 228                                 break;
 229                         if (shrink_ret < nr_before)
 230                                 ret += nr_before - shrink_ret;
 231                         mod_page_state(slabs_scanned, this_scan);
 232                         total_scan -= this_scan;
 233
 234                         cond_resched();
 235                 }
 236
 237                 shrinker->nr += total_scan;
 238         }
 239         up_read(&shrinker_rwsem);
 240         return ret;
 241 }
 242
 243 /* Called without lock on whether page is mapped, so answer is unstable */
 244 static inline int page_mapping_inuse(struct page *page)
 245 {
 246         struct address_space *mapping;
 247
 248         /* Page is in somebody's page tables. */
 249         if (page_mapped(page))
 250                 return 1;
 251
 252         /* Be more reluctant to reclaim swapcache than pagecache */
 253         if (PageSwapCache(page))
 254                 return 1;
 255
 256         mapping = page_mapping(page);
 257         if (!mapping)
 258                 return 0;
 259
 260         /* File is mmap'd by somebody? */
 261         return mapping_mapped(mapping);
 262 }
 263
 264 static inline int is_page_cache_freeable(struct page *page)
 265 {
 266         return page_count(page) - !!PagePrivate(page) == 2;
 267 }
 268
 269 static int may_write_to_queue(struct backing_dev_info *bdi)
 270 {
 271         if (current->flags & PF_SWAPWRITE)
 272                 return 1;
 273         if (!bdi_write_congested(bdi))
 274                 return 1;
 275         if (bdi == current->backing_dev_info)
 276                 return 1;
 277         return 0;
 278 }
 279
 280 /*
 281  * We detected a synchronous write error writing a page out.  Probably
 282  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 283  * fsync(), msync() or close().
 284  *
 285  * The tricky part is that after writepage we cannot touch the mapping: nothing
 286  * prevents it from being freed up.  But we have a ref on the page and once
 287  * that page is locked, the mapping is pinned.
 288  *
 289  * We're allowed to run sleeping lock_page() here because we know the caller has
 290  * __GFP_FS.
 291  */
 292 static void handle_write_error(struct address_space *mapping,
 293                                 struct page *page, int error)
 294 {
 295         lock_page(page);
 296         if (page_mapping(page) == mapping) {
 297                 if (error == -ENOSPC)
 298                         set_bit(AS_ENOSPC, &mapping->flags);
 299                 else
 300                         set_bit(AS_EIO, &mapping->flags);
 301         }
 302         unlock_page(page);
 303 }
 304
 305 /*
 306  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 307  */
 308 static pageout_t pageout(struct page *page, struct address_space *mapping)
 309 {
 310         /*
 311          * If the page is dirty, only perform writeback if that write
 312          * will be non-blocking.  To prevent this allocation from being
 313          * stalled by pagecache activity.  But note that there may be
 314          * stalls if we need to run get_block().  We could test
 315          * PagePrivate for that.
 316          *
 317          * If this process is currently in generic_file_write() against
 318          * this page's queue, we can perform writeback even if that
 319          * will block.
 320          *
 321          * If the page is swapcache, write it back even if that would
 322          * block, for some throttling. This happens by accident, because
 323          * swap_backing_dev_info is bust: it doesn't reflect the
 324          * congestion state of the swapdevs.  Easy to fix, if needed.
 325          * See swapfile.c:page_queue_congested().
 326          */
 327         if (!is_page_cache_freeable(page))
 328                 return PAGE_KEEP;
 329         if (!mapping) {
 330                 /*
 331                  * Some data journaling orphaned pages can have
 332                  * page->mapping == NULL while being dirty with clean buffers.
 333                  */
 334                 if (PagePrivate(page)) {
 335                         if (try_to_free_buffers(page)) {
 336                                 ClearPageDirty(page);
 337                                 printk("%s: orphaned page\n", __FUNCTION__);
 338                                 return PAGE_CLEAN;
 339                         }
 340                 }
 341                 return PAGE_KEEP;
 342         }
 343         if (mapping->a_ops->writepage == NULL)
 344                 return PAGE_ACTIVATE;
 345         if (!may_write_to_queue(mapping->backing_dev_info))
 346                 return PAGE_KEEP;
 347
 348         if (clear_page_dirty_for_io(page)) {
 349                 int res;
 350                 struct writeback_control wbc = {
 351                         .sync_mode = WB_SYNC_NONE,
 352                         .nr_to_write = SWAP_CLUSTER_MAX,
 353                         .nonblocking = 1,
 354                         .for_reclaim = 1,
 355                 };
 356
 357                 SetPageReclaim(page);
 358                 res = mapping->a_ops->writepage(page, &wbc);
 359                 if (res < 0)
 360                         handle_write_error(mapping, page, res);
 361                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 362                         ClearPageReclaim(page);
 363                         return PAGE_ACTIVATE;
 364                 }
 365                 if (!PageWriteback(page)) {
 366                         /* synchronous write or broken a_ops? */
 367                         ClearPageReclaim(page);
 368                 }
 369
 370                 return PAGE_SUCCESS;
 371         }
 372
 373         return PAGE_CLEAN;
 374 }
 375
 376 static int remove_mapping(struct address_space *mapping, struct page *page)
 377 {
 378         if (!mapping)
 379                 return 0;               /* truncate got there first */
 380
 381         write_lock_irq(&mapping->tree_lock);
 382
 383         /*
 384          * The non-racy check for busy page.  It is critical to check
 385          * PageDirty _after_ making sure that the page is freeable and
 386          * not in use by anybody.       (pagecache + us == 2)
 387          */
 388         if (unlikely(page_count(page) != 2))
 389                 goto cannot_free;
 390         smp_rmb();
 391         if (unlikely(PageDirty(page)))
 392                 goto cannot_free;
 393
 394         if (PageSwapCache(page)) {
 395                 swp_entry_t swap = { .val = page_private(page) };
 396                 __delete_from_swap_cache(page);
 397                 write_unlock_irq(&mapping->tree_lock);
 398                 swap_free(swap);
 399                 __put_page(page);       /* The pagecache ref */
 400                 return 1;
 401         }
 402
 403         __remove_from_page_cache(page);
 404         write_unlock_irq(&mapping->tree_lock);
 405         __put_page(page);
 406         return 1;
 407
 408 cannot_free:
 409         write_unlock_irq(&mapping->tree_lock);
 410         return 0;
 411 }
 412
 413 /*
 414  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 415  */
 416 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 417 {
 418         LIST_HEAD(ret_pages);
 419         struct pagevec freed_pvec;
 420         int pgactivate = 0;
 421         int reclaimed = 0;
 422
 423         cond_resched();
 424
 425         pagevec_init(&freed_pvec, 1);
 426         while (!list_empty(page_list)) {
 427                 struct address_space *mapping;
 428                 struct page *page;
 429                 int may_enter_fs;
 430                 int referenced;
 431
 432                 cond_resched();
 433
 434                 page = lru_to_page(page_list);
 435                 list_del(&page->lru);
 436
 437                 if (TestSetPageLocked(page))
 438                         goto keep;
 439
 440                 BUG_ON(PageActive(page));
 441
 442                 sc->nr_scanned++;
 443                 /* Double the slab pressure for mapped and swapcache pages */
 444                 if (page_mapped(page) || PageSwapCache(page))
 445                         sc->nr_scanned++;
 446
 447                 if (PageWriteback(page))
 448                         goto keep_locked;
 449
 450                 referenced = page_referenced(page, 1);
 451                 /* In active use or really unfreeable?  Activate it. */
 452                 if (referenced && page_mapping_inuse(page))
 453                         goto activate_locked;
 454
 455 #ifdef CONFIG_SWAP
 456                 /*
 457                  * Anonymous process memory has backing store?
 458                  * Try to allocate it some swap space here.
 459                  */
 460                 if (PageAnon(page) && !PageSwapCache(page)) {
 461                         if (!add_to_swap(page))
 462                                 goto activate_locked;
 463                 }
 464 #endif /* CONFIG_SWAP */
 465
 466                 mapping = page_mapping(page);
 467                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 468                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 469
 470                 /*
 471                  * The page is mapped into the page tables of one or more
 472                  * processes. Try to unmap it here.
 473                  */
 474                 if (page_mapped(page) && mapping) {
 475                         switch (try_to_unmap(page)) {
 476                         case SWAP_FAIL:
 477                                 goto activate_locked;
 478                         case SWAP_AGAIN:
 479                                 goto keep_locked;
 480                         case SWAP_SUCCESS:
 481                                 ; /* try to free the page below */
 482                         }
 483                 }
 484
 485                 if (PageDirty(page)) {
 486                         if (referenced)
 487                                 goto keep_locked;
 488                         if (!may_enter_fs)
 489                                 goto keep_locked;
 490                         if (laptop_mode && !sc->may_writepage)
 491                                 goto keep_locked;
 492
 493                         /* Page is dirty, try to write it out here */
 494                         switch(pageout(page, mapping)) {
 495                         case PAGE_KEEP:
 496                                 goto keep_locked;
 497                         case PAGE_ACTIVATE:
 498                                 goto activate_locked;
 499                         case PAGE_SUCCESS:
 500                                 if (PageWriteback(page) || PageDirty(page))
 501                                         goto keep;
 502                                 /*
 503                                  * A synchronous write - probably a ramdisk.  Go
 504                                  * ahead and try to reclaim the page.
 505                                  */
 506                                 if (TestSetPageLocked(page))
 507                                         goto keep;
 508                                 if (PageDirty(page) || PageWriteback(page))
 509                                         goto keep_locked;
 510                                 mapping = page_mapping(page);
 511                         case PAGE_CLEAN:
 512                                 ; /* try to free the page below */
 513                         }
 514                 }
 515
 516                 /*
 517                  * If the page has buffers, try to free the buffer mappings
 518                  * associated with this page. If we succeed we try to free
 519                  * the page as well.
 520                  *
 521                  * We do this even if the page is PageDirty().
 522                  * try_to_release_page() does not perform I/O, but it is
 523                  * possible for a page to have PageDirty set, but it is actually
 524                  * clean (all its buffers are clean).  This happens if the
 525                  * buffers were written out directly, with submit_bh(). ext3
 526                  * will do this, as well as the blockdev mapping.
 527                  * try_to_release_page() will discover that cleanness and will
 528                  * drop the buffers and mark the page clean - it can be freed.
 529                  *
 530                  * Rarely, pages can have buffers and no ->mapping.  These are
 531                  * the pages which were not successfully invalidated in
 532                  * truncate_complete_page().  We try to drop those buffers here
 533                  * and if that worked, and the page is no longer mapped into
 534                  * process address space (page_count == 1) it can be freed.
 535                  * Otherwise, leave the page on the LRU so it is swappable.
 536                  */
 537                 if (PagePrivate(page)) {
 538                         if (!try_to_release_page(page, sc->gfp_mask))
 539                                 goto activate_locked;
 540                         if (!mapping && page_count(page) == 1)
 541                                 goto free_it;
 542                 }
 543
 544                 if (!remove_mapping(mapping, page))
 545                         goto keep_locked;
 546
 547 free_it:
 548                 unlock_page(page);
 549                 reclaimed++;
 550                 if (!pagevec_add(&freed_pvec, page))
 551                         __pagevec_release_nonlru(&freed_pvec);
 552                 continue;
 553
 554 activate_locked:
 555                 SetPageActive(page);
 556                 pgactivate++;
 557 keep_locked:
 558                 unlock_page(page);
 559 keep:
 560                 list_add(&page->lru, &ret_pages);
 561                 BUG_ON(PageLRU(page));
 562         }
 563         list_splice(&ret_pages, page_list);
 564         if (pagevec_count(&freed_pvec))
 565                 __pagevec_release_nonlru(&freed_pvec);
 566         mod_page_state(pgactivate, pgactivate);
 567         sc->nr_reclaimed += reclaimed;
 568         return reclaimed;
 569 }
 570
 571 /*
 572  * swapout a single page
 573  * page is locked upon entry, unlocked on exit
 574  *
 575  * return codes:
 576  *      0 = complete
 577  *      1 = retry
 578  */
 579 static int swap_page(struct page *page)
 580 {
 581         struct address_space *mapping = page_mapping(page);
 582
 583         if (page_mapped(page) && mapping)
 584                 if (try_to_unmap(page) != SWAP_SUCCESS)
 585                         goto unlock_retry;
 586
 587         if (PageDirty(page)) {
 588                 /* Page is dirty, try to write it out here */
 589                 switch(pageout(page, mapping)) {
 590                 case PAGE_KEEP:
 591                 case PAGE_ACTIVATE:
 592                         goto unlock_retry;
 593
 594                 case PAGE_SUCCESS:
 595                         goto retry;
 596
 597                 case PAGE_CLEAN:
 598                         ; /* try to free the page below */
 599                 }
 600         }
 601
 602         if (PagePrivate(page)) {
 603                 if (!try_to_release_page(page, GFP_KERNEL) ||
 604                     (!mapping && page_count(page) == 1))
 605                         goto unlock_retry;
 606         }
 607
 608         if (remove_mapping(mapping, page)) {
 609                 /* Success */
 610                 unlock_page(page);
 611                 return 0;
 612         }
 613
 614 unlock_retry:
 615         unlock_page(page);
 616
 617 retry:
 618         return 1;
 619 }
 620 /*
 621  * migrate_pages
 622  *
 623  * Two lists are passed to this function. The first list
 624  * contains the pages isolated from the LRU to be migrated.
 625  * The second list contains new pages that the pages isolated
 626  * can be moved to. If the second list is NULL then all
 627  * pages are swapped out.
 628  *
 629  * The function returns after 10 attempts or if no pages
 630  * are movable anymore because t has become empty
 631  * or no retryable pages exist anymore.
 632  *
 633  * SIMPLIFIED VERSION: This implementation of migrate_pages
 634  * is only swapping out pages and never touches the second
 635  * list. The direct migration patchset
 636  * extends this function to avoid the use of swap.
 637  */
 638 int migrate_pages(struct list_head *l, struct list_head *t)
 639 {
 640         int retry;
 641         LIST_HEAD(failed);
 642         int nr_failed = 0;
 643         int pass = 0;
 644         struct page *page;
 645         struct page *page2;
 646         int swapwrite = current->flags & PF_SWAPWRITE;
 647
 648         if (!swapwrite)
 649                 current->flags |= PF_SWAPWRITE;
 650
 651 redo:
 652         retry = 0;
 653
 654         list_for_each_entry_safe(page, page2, l, lru) {
 655                 cond_resched();
 656
 657                 /*
 658                  * Skip locked pages during the first two passes to give the
 659                  * functions holding the lock time to release the page. Later we use
 660                  * lock_page to have a higher chance of acquiring the lock.
 661                  */
 662                 if (pass > 2)
 663                         lock_page(page);
 664                 else
 665                         if (TestSetPageLocked(page))
 666                                 goto retry_later;
 667
 668                 /*
 669                  * Only wait on writeback if we have already done a pass where
 670                  * we we may have triggered writeouts for lots of pages.
 671                  */
 672                 if (pass > 0)
 673                         wait_on_page_writeback(page);
 674                 else
 675                         if (PageWriteback(page)) {
 676                                 unlock_page(page);
 677                                 goto retry_later;
 678                         }
 679
 680 #ifdef CONFIG_SWAP
 681                 if (PageAnon(page) && !PageSwapCache(page)) {
 682                         if (!add_to_swap(page)) {
 683                                 unlock_page(page);
 684                                 list_move(&page->lru, &failed);
 685                                 nr_failed++;
 686                                 continue;
 687                         }
 688                 }
 689 #endif /* CONFIG_SWAP */
 690
 691                 /*
 692                  * Page is properly locked and writeback is complete.
 693                  * Try to migrate the page.
 694                  */
 695                 if (swap_page(page)) {
 696 retry_later:
 697                         retry++;
 698                 }
 699         }
 700         if (retry && pass++ < 10)
 701                 goto redo;
 702
 703         if (!swapwrite)
 704                 current->flags &= ~PF_SWAPWRITE;
 705
 706         if (!list_empty(&failed))
 707                 list_splice(&failed, l);
 708
 709         return nr_failed + retry;
 710 }
 711
 712 /*
 713  * zone->lru_lock is heavily contended.  Some of the functions that
 714  * shrink the lists perform better by taking out a batch of pages
 715  * and working on them outside the LRU lock.
 716  *
 717  * For pagecache intensive workloads, this function is the hottest
 718  * spot in the kernel (apart from copy_*_user functions).
 719  *
 720  * Appropriate locks must be held before calling this function.
 721  *
 722  * @nr_to_scan: The number of pages to look through on the list.
 723  * @src:        The LRU list to pull pages off.
 724  * @dst:        The temp list to put pages on to.
 725  * @scanned:    The number of pages that were scanned.
 726  *
 727  * returns how many pages were moved onto *@dst.
 728  */
 729 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
 730                              struct list_head *dst, int *scanned)
 731 {
 732         int nr_taken = 0;
 733         struct page *page;
 734         int scan = 0;
 735
 736         while (scan++ < nr_to_scan && !list_empty(src)) {
 737                 page = lru_to_page(src);
 738                 prefetchw_prev_lru_page(page, src, flags);
 739
 740                 switch (__isolate_lru_page(page)) {
 741                 case 1:
 742                         /* Succeeded to isolate page */
 743                         list_move(&page->lru, dst);
 744                         nr_taken++;
 745                         break;
 746                 case -ENOENT:
 747                         /* Not possible to isolate */
 748                         list_move(&page->lru, src);
 749                         break;
 750                 default:
 751                         BUG();
 752                 }
 753         }
 754
 755         *scanned = scan;
 756         return nr_taken;
 757 }
 758
 759 static void lru_add_drain_per_cpu(void *dummy)
 760 {
 761         lru_add_drain();
 762 }
 763
 764 /*
 765  * Isolate one page from the LRU lists and put it on the
 766  * indicated list. Do necessary cache draining if the
 767  * page is not on the LRU lists yet.
 768  *
 769  * Result:
 770  *  0 = page not on LRU list
 771  *  1 = page removed from LRU list and added to the specified list.
 772  * -ENOENT = page is being freed elsewhere.
 773  */
 774 int isolate_lru_page(struct page *page)
 775 {
 776         int rc = 0;
 777         struct zone *zone = page_zone(page);
 778
 779 redo:
 780         spin_lock_irq(&zone->lru_lock);
 781         rc = __isolate_lru_page(page);
 782         if (rc == 1) {
 783                 if (PageActive(page))
 784                         del_page_from_active_list(zone, page);
 785                 else
 786                         del_page_from_inactive_list(zone, page);
 787         }
 788         spin_unlock_irq(&zone->lru_lock);
 789         if (rc == 0) {
 790                 /*
 791                  * Maybe this page is still waiting for a cpu to drain it
 792                  * from one of the lru lists?
 793                  */
 794                 rc = schedule_on_each_cpu(lru_add_drain_per_cpu, NULL);
 795                 if (rc == 0 && PageLRU(page))
 796                         goto redo;
 797         }
 798         return rc;
 799 }
 800
 801 /*
 802  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 803  */
 804 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 805 {
 806         LIST_HEAD(page_list);
 807         struct pagevec pvec;
 808         int max_scan = sc->nr_to_scan;
 809
 810         pagevec_init(&pvec, 1);
 811
 812         lru_add_drain();
 813         spin_lock_irq(&zone->lru_lock);
 814         while (max_scan > 0) {
 815                 struct page *page;
 816                 int nr_taken;
 817                 int nr_scan;
 818                 int nr_freed;
 819
 820                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
 821                                              &zone->inactive_list,
 822                                              &page_list, &nr_scan);
 823                 zone->nr_inactive -= nr_taken;
 824                 zone->pages_scanned += nr_scan;
 825                 spin_unlock_irq(&zone->lru_lock);
 826
 827                 if (nr_taken == 0)
 828                         goto done;
 829
 830                 max_scan -= nr_scan;
 831                 nr_freed = shrink_list(&page_list, sc);
 832
 833                 local_irq_disable();
 834                 if (current_is_kswapd()) {
 835                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 836                         __mod_page_state(kswapd_steal, nr_freed);
 837                 } else
 838                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
 839                 __mod_page_state_zone(zone, pgsteal, nr_freed);
 840
 841                 spin_lock(&zone->lru_lock);
 842                 /*
 843                  * Put back any unfreeable pages.
 844                  */
 845                 while (!list_empty(&page_list)) {
 846                         page = lru_to_page(&page_list);
 847                         if (TestSetPageLRU(page))
 848                                 BUG();
 849                         list_del(&page->lru);
 850                         if (PageActive(page))
 851                                 add_page_to_active_list(zone, page);
 852                         else
 853                                 add_page_to_inactive_list(zone, page);
 854                         if (!pagevec_add(&pvec, page)) {
 855                                 spin_unlock_irq(&zone->lru_lock);
 856                                 __pagevec_release(&pvec);
 857                                 spin_lock_irq(&zone->lru_lock);
 858                         }
 859                 }
 860         }
 861         spin_unlock_irq(&zone->lru_lock);
 862 done:
 863         pagevec_release(&pvec);
 864 }
 865
 866 static inline void move_to_lru(struct page *page)
 867 {
 868         list_del(&page->lru);
 869         if (PageActive(page)) {
 870                 /*
 871                  * lru_cache_add_active checks that
 872                  * the PG_active bit is off.
 873                  */
 874                 ClearPageActive(page);
 875                 lru_cache_add_active(page);
 876         } else {
 877                 lru_cache_add(page);
 878         }
 879         put_page(page);
 880 }
 881
 882 /*
 883  * Add isolated pages on the list back to the LRU
 884  *
 885  * returns the number of pages put back.
 886  */
 887 int putback_lru_pages(struct list_head *l)
 888 {
 889         struct page *page;
 890         struct page *page2;
 891         int count = 0;
 892
 893         list_for_each_entry_safe(page, page2, l, lru) {
 894                 move_to_lru(page);
 895                 count++;
 896         }
 897         return count;
 898 }
 899
 900 /*
 901  * This moves pages from the active list to the inactive list.
 902  *
 903  * We move them the other way if the page is referenced by one or more
 904  * processes, from rmap.
 905  *
 906  * If the pages are mostly unmapped, the processing is fast and it is
 907  * appropriate to hold zone->lru_lock across the whole operation.  But if
 908  * the pages are mapped, the processing is slow (page_referenced()) so we
 909  * should drop zone->lru_lock around each page.  It's impossible to balance
 910  * this, so instead we remove the pages from the LRU while processing them.
 911  * It is safe to rely on PG_active against the non-LRU pages in here because
 912  * nobody will play with that bit on a non-LRU page.
 913  *
 914  * The downside is that we have to touch page->_count against each page.
 915  * But we had to alter page->flags anyway.
 916  */
 917 static void
 918 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 919 {
 920         int pgmoved;
 921         int pgdeactivate = 0;
 922         int pgscanned;
 923         int nr_pages = sc->nr_to_scan;
 924         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 925         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 926         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 927         struct page *page;
 928         struct pagevec pvec;
 929         int reclaim_mapped = 0;
 930         long mapped_ratio;
 931         long distress;
 932         long swap_tendency;
 933
 934         lru_add_drain();
 935         spin_lock_irq(&zone->lru_lock);
 936         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
 937                                     &l_hold, &pgscanned);
 938         zone->pages_scanned += pgscanned;
 939         zone->nr_active -= pgmoved;
 940         spin_unlock_irq(&zone->lru_lock);
 941
 942         /*
 943          * `distress' is a measure of how much trouble we're having reclaiming
 944          * pages.  0 -> no problems.  100 -> great trouble.
 945          */
 946         distress = 100 >> zone->prev_priority;
 947
 948         /*
 949          * The point of this algorithm is to decide when to start reclaiming
 950          * mapped memory instead of just pagecache.  Work out how much memory
 951          * is mapped.
 952          */
 953         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 954
 955         /*
 956          * Now decide how much we really want to unmap some pages.  The mapped
 957          * ratio is downgraded - just because there's a lot of mapped memory
 958          * doesn't necessarily mean that page reclaim isn't succeeding.
 959          *
 960          * The distress ratio is important - we don't want to start going oom.
 961          *
 962          * A 100% value of vm_swappiness overrides this algorithm altogether.
 963          */
 964         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 965
 966         /*
 967          * Now use this metric to decide whether to start moving mapped memory
 968          * onto the inactive list.
 969          */
 970         if (swap_tendency >= 100)
 971                 reclaim_mapped = 1;
 972
 973         while (!list_empty(&l_hold)) {
 974                 cond_resched();
 975                 page = lru_to_page(&l_hold);
 976                 list_del(&page->lru);
 977                 if (page_mapped(page)) {
 978                         if (!reclaim_mapped ||
 979                             (total_swap_pages == 0 && PageAnon(page)) ||
 980                             page_referenced(page, 0)) {
 981                                 list_add(&page->lru, &l_active);
 982                                 continue;
 983                         }
 984                 }
 985                 list_add(&page->lru, &l_inactive);
 986         }
 987
 988         pagevec_init(&pvec, 1);
 989         pgmoved = 0;
 990         spin_lock_irq(&zone->lru_lock);
 991         while (!list_empty(&l_inactive)) {
 992                 page = lru_to_page(&l_inactive);
 993                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 994                 if (TestSetPageLRU(page))
 995                         BUG();
 996                 if (!TestClearPageActive(page))
 997                         BUG();
 998                 list_move(&page->lru, &zone->inactive_list);
 999                 pgmoved++;
1000                 if (!pagevec_add(&pvec, page)) {
1001                         zone->nr_inactive += pgmoved;
1002                         spin_unlock_irq(&zone->lru_lock);
1003                         pgdeactivate += pgmoved;
1004                         pgmoved = 0;
1005                         if (buffer_heads_over_limit)
1006                                 pagevec_strip(&pvec);
1007                         __pagevec_release(&pvec);
1008                         spin_lock_irq(&zone->lru_lock);
1009                 }
1010         }
1011         zone->nr_inactive += pgmoved;
1012         pgdeactivate += pgmoved;
1013         if (buffer_heads_over_limit) {
1014                 spin_unlock_irq(&zone->lru_lock);
1015                 pagevec_strip(&pvec);
1016                 spin_lock_irq(&zone->lru_lock);
1017         }
1018
1019         pgmoved = 0;
1020         while (!list_empty(&l_active)) {
1021                 page = lru_to_page(&l_active);
1022                 prefetchw_prev_lru_page(page, &l_active, flags);
1023                 if (TestSetPageLRU(page))
1024                         BUG();
1025                 BUG_ON(!PageActive(page));
1026                 list_move(&page->lru, &zone->active_list);
1027                 pgmoved++;
1028                 if (!pagevec_add(&pvec, page)) {
1029                         zone->nr_active += pgmoved;
1030                         pgmoved = 0;
1031                         spin_unlock_irq(&zone->lru_lock);
1032                         __pagevec_release(&pvec);
1033                         spin_lock_irq(&zone->lru_lock);
1034                 }
1035         }
1036         zone->nr_active += pgmoved;
1037         spin_unlock(&zone->lru_lock);
1038
1039         __mod_page_state_zone(zone, pgrefill, pgscanned);
1040         __mod_page_state(pgdeactivate, pgdeactivate);
1041         local_irq_enable();
1042
1043         pagevec_release(&pvec);
1044 }
1045
1046 /*
1047  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1048  */
1049 static void
1050 shrink_zone(struct zone *zone, struct scan_control *sc)
1051 {
1052         unsigned long nr_active;
1053         unsigned long nr_inactive;
1054
1055         atomic_inc(&zone->reclaim_in_progress);
1056
1057         /*
1058          * Add one to `nr_to_scan' just to make sure that the kernel will
1059          * slowly sift through the active list.
1060          */
1061         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1062         nr_active = zone->nr_scan_active;
1063         if (nr_active >= sc->swap_cluster_max)
1064                 zone->nr_scan_active = 0;
1065         else
1066                 nr_active = 0;
1067
1068         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1069         nr_inactive = zone->nr_scan_inactive;
1070         if (nr_inactive >= sc->swap_cluster_max)
1071                 zone->nr_scan_inactive = 0;
1072         else
1073                 nr_inactive = 0;
1074
1075         while (nr_active || nr_inactive) {
1076                 if (nr_active) {
1077                         sc->nr_to_scan = min(nr_active,
1078                                         (unsigned long)sc->swap_cluster_max);
1079                         nr_active -= sc->nr_to_scan;
1080                         refill_inactive_zone(zone, sc);
1081                 }
1082
1083                 if (nr_inactive) {
1084                         sc->nr_to_scan = min(nr_inactive,
1085                                         (unsigned long)sc->swap_cluster_max);
1086                         nr_inactive -= sc->nr_to_scan;
1087                         shrink_cache(zone, sc);
1088                 }
1089         }
1090
1091         throttle_vm_writeout();
1092
1093         atomic_dec(&zone->reclaim_in_progress);
1094 }
1095
1096 /*
1097  * This is the direct reclaim path, for page-allocating processes.  We only
1098  * try to reclaim pages from zones which will satisfy the caller's allocation
1099  * request.
1100  *
1101  * We reclaim from a zone even if that zone is over pages_high.  Because:
1102  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1103  *    allocation or
1104  * b) The zones may be over pages_high but they must go *over* pages_high to
1105  *    satisfy the `incremental min' zone defense algorithm.
1106  *
1107  * Returns the number of reclaimed pages.
1108  *
1109  * If a zone is deemed to be full of pinned pages then just give it a light
1110  * scan then give up on it.
1111  */
1112 static void
1113 shrink_caches(struct zone **zones, struct scan_control *sc)
1114 {
1115         int i;
1116
1117         for (i = 0; zones[i] != NULL; i++) {
1118                 struct zone *zone = zones[i];
1119
1120                 if (!populated_zone(zone))
1121                         continue;
1122
1123                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1124                         continue;
1125
1126                 zone->temp_priority = sc->priority;
1127                 if (zone->prev_priority > sc->priority)
1128                         zone->prev_priority = sc->priority;
1129
1130                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1131                         continue;       /* Let kswapd poll it */
1132
1133                 shrink_zone(zone, sc);
1134         }
1135 }
1136
1137 /*
1138  * This is the main entry point to direct page reclaim.
1139  *
1140  * If a full scan of the inactive list fails to free enough memory then we
1141  * are "out of memory" and something needs to be killed.
1142  *
1143  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1144  * high - the zone may be full of dirty or under-writeback pages, which this
1145  * caller can't do much about.  We kick pdflush and take explicit naps in the
1146  * hope that some of these pages can be written.  But if the allocating task
1147  * holds filesystem locks which prevent writeout this might not work, and the
1148  * allocation attempt will fail.
1149  */
1150 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1151 {
1152         int priority;
1153         int ret = 0;
1154         int total_scanned = 0, total_reclaimed = 0;
1155         struct reclaim_state *reclaim_state = current->reclaim_state;
1156         struct scan_control sc;
1157         unsigned long lru_pages = 0;
1158         int i;
1159
1160         sc.gfp_mask = gfp_mask;
1161         sc.may_writepage = 0;
1162
1163         inc_page_state(allocstall);
1164
1165         for (i = 0; zones[i] != NULL; i++) {
1166                 struct zone *zone = zones[i];
1167
1168                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1169                         continue;
1170
1171                 zone->temp_priority = DEF_PRIORITY;
1172                 lru_pages += zone->nr_active + zone->nr_inactive;
1173         }
1174
1175         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1176                 sc.nr_mapped = read_page_state(nr_mapped);
1177                 sc.nr_scanned = 0;
1178                 sc.nr_reclaimed = 0;
1179                 sc.priority = priority;
1180                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1181                 if (!priority)
1182                         disable_swap_token();
1183                 shrink_caches(zones, &sc);
1184                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1185                 if (reclaim_state) {
1186                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1187                         reclaim_state->reclaimed_slab = 0;
1188                 }
1189                 total_scanned += sc.nr_scanned;
1190                 total_reclaimed += sc.nr_reclaimed;
1191                 if (total_reclaimed >= sc.swap_cluster_max) {
1192                         ret = 1;
1193                         goto out;
1194                 }
1195
1196                 /*
1197                  * Try to write back as many pages as we just scanned.  This
1198                  * tends to cause slow streaming writers to write data to the
1199                  * disk smoothly, at the dirtying rate, which is nice.   But
1200                  * that's undesirable in laptop mode, where we *want* lumpy
1201                  * writeout.  So in laptop mode, write out the whole world.
1202                  */
1203                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1204                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1205                         sc.may_writepage = 1;
1206                 }
1207
1208                 /* Take a nap, wait for some writeback to complete */
1209                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1210                         blk_congestion_wait(WRITE, HZ/10);
1211         }
1212 out:
1213         for (i = 0; zones[i] != 0; i++) {
1214                 struct zone *zone = zones[i];
1215
1216                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1217                         continue;
1218
1219                 zone->prev_priority = zone->temp_priority;
1220         }
1221         return ret;
1222 }
1223
1224 /*
1225  * For kswapd, balance_pgdat() will work across all this node's zones until
1226  * they are all at pages_high.
1227  *
1228  * If `nr_pages' is non-zero then it is the number of pages which are to be
1229  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1230  * special.
1231  *
1232  * Returns the number of pages which were actually freed.
1233  *
1234  * There is special handling here for zones which are full of pinned pages.
1235  * This can happen if the pages are all mlocked, or if they are all used by
1236  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1237  * What we do is to detect the case where all pages in the zone have been
1238  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1239  * dead and from now on, only perform a short scan.  Basically we're polling
1240  * the zone for when the problem goes away.
1241  *
1242  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1243  * zones which have free_pages > pages_high, but once a zone is found to have
1244  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1245  * of the number of free pages in the lower zones.  This interoperates with
1246  * the page allocator fallback scheme to ensure that aging of pages is balanced
1247  * across the zones.
1248  */
1249 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1250 {
1251         int to_free = nr_pages;
1252         int all_zones_ok;
1253         int priority;
1254         int i;
1255         int total_scanned, total_reclaimed;
1256         struct reclaim_state *reclaim_state = current->reclaim_state;
1257         struct scan_control sc;
1258
1259 loop_again:
1260         total_scanned = 0;
1261         total_reclaimed = 0;
1262         sc.gfp_mask = GFP_KERNEL;
1263         sc.may_writepage = 0;
1264         sc.nr_mapped = read_page_state(nr_mapped);
1265
1266         inc_page_state(pageoutrun);
1267
1268         for (i = 0; i < pgdat->nr_zones; i++) {
1269                 struct zone *zone = pgdat->node_zones + i;
1270
1271                 zone->temp_priority = DEF_PRIORITY;
1272         }
1273
1274         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1275                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1276                 unsigned long lru_pages = 0;
1277
1278                 /* The swap token gets in the way of swapout... */
1279                 if (!priority)
1280                         disable_swap_token();
1281
1282                 all_zones_ok = 1;
1283
1284                 if (nr_pages == 0) {
1285                         /*
1286                          * Scan in the highmem->dma direction for the highest
1287                          * zone which needs scanning
1288                          */
1289                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1290                                 struct zone *zone = pgdat->node_zones + i;
1291
1292                                 if (!populated_zone(zone))
1293                                         continue;
1294
1295                                 if (zone->all_unreclaimable &&
1296                                                 priority != DEF_PRIORITY)
1297                                         continue;
1298
1299                                 if (!zone_watermark_ok(zone, order,
1300                                                 zone->pages_high, 0, 0)) {
1301                                         end_zone = i;
1302                                         goto scan;
1303                                 }
1304                         }
1305                         goto out;
1306                 } else {
1307                         end_zone = pgdat->nr_zones - 1;
1308                 }
1309 scan:
1310                 for (i = 0; i <= end_zone; i++) {
1311                         struct zone *zone = pgdat->node_zones + i;
1312
1313                         lru_pages += zone->nr_active + zone->nr_inactive;
1314                 }
1315
1316                 /*
1317                  * Now scan the zone in the dma->highmem direction, stopping
1318                  * at the last zone which needs scanning.
1319                  *
1320                  * We do this because the page allocator works in the opposite
1321                  * direction.  This prevents the page allocator from allocating
1322                  * pages behind kswapd's direction of progress, which would
1323                  * cause too much scanning of the lower zones.
1324                  */
1325                 for (i = 0; i <= end_zone; i++) {
1326                         struct zone *zone = pgdat->node_zones + i;
1327                         int nr_slab;
1328
1329                         if (!populated_zone(zone))
1330                                 continue;
1331
1332                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1333                                 continue;
1334
1335                         if (nr_pages == 0) {    /* Not software suspend */
1336                                 if (!zone_watermark_ok(zone, order,
1337                                                 zone->pages_high, end_zone, 0))
1338                                         all_zones_ok = 0;
1339                         }
1340                         zone->temp_priority = priority;
1341                         if (zone->prev_priority > priority)
1342                                 zone->prev_priority = priority;
1343                         sc.nr_scanned = 0;
1344                         sc.nr_reclaimed = 0;
1345                         sc.priority = priority;
1346                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1347                         atomic_inc(&zone->reclaim_in_progress);
1348                         shrink_zone(zone, &sc);
1349                         atomic_dec(&zone->reclaim_in_progress);
1350                         reclaim_state->reclaimed_slab = 0;
1351                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1352                                                 lru_pages);
1353                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1354                         total_reclaimed += sc.nr_reclaimed;
1355                         total_scanned += sc.nr_scanned;
1356                         if (zone->all_unreclaimable)
1357                                 continue;
1358                         if (nr_slab == 0 && zone->pages_scanned >=
1359                                     (zone->nr_active + zone->nr_inactive) * 4)
1360                                 zone->all_unreclaimable = 1;
1361                         /*
1362                          * If we've done a decent amount of scanning and
1363                          * the reclaim ratio is low, start doing writepage
1364                          * even in laptop mode
1365                          */
1366                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1367                             total_scanned > total_reclaimed+total_reclaimed/2)
1368                                 sc.may_writepage = 1;
1369                 }
1370                 if (nr_pages && to_free > total_reclaimed)
1371                         continue;       /* swsusp: need to do more work */
1372                 if (all_zones_ok)
1373                         break;          /* kswapd: all done */
1374                 /*
1375                  * OK, kswapd is getting into trouble.  Take a nap, then take
1376                  * another pass across the zones.
1377                  */
1378                 if (total_scanned && priority < DEF_PRIORITY - 2)
1379                         blk_congestion_wait(WRITE, HZ/10);
1380
1381                 /*
1382                  * We do this so kswapd doesn't build up large priorities for
1383                  * example when it is freeing in parallel with allocators. It
1384                  * matches the direct reclaim path behaviour in terms of impact
1385                  * on zone->*_priority.
1386                  */
1387                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1388                         break;
1389         }
1390 out:
1391         for (i = 0; i < pgdat->nr_zones; i++) {
1392                 struct zone *zone = pgdat->node_zones + i;
1393
1394                 zone->prev_priority = zone->temp_priority;
1395         }
1396         if (!all_zones_ok) {
1397                 cond_resched();
1398                 goto loop_again;
1399         }
1400
1401         return total_reclaimed;
1402 }
1403
1404 /*
1405  * The background pageout daemon, started as a kernel thread
1406  * from the init process.
1407  *
1408  * This basically trickles out pages so that we have _some_
1409  * free memory available even if there is no other activity
1410  * that frees anything up. This is needed for things like routing
1411  * etc, where we otherwise might have all activity going on in
1412  * asynchronous contexts that cannot page things out.
1413  *
1414  * If there are applications that are active memory-allocators
1415  * (most normal use), this basically shouldn't matter.
1416  */
1417 static int kswapd(void *p)
1418 {
1419         unsigned long order;
1420         pg_data_t *pgdat = (pg_data_t*)p;
1421         struct task_struct *tsk = current;
1422         DEFINE_WAIT(wait);
1423         struct reclaim_state reclaim_state = {
1424                 .reclaimed_slab = 0,
1425         };
1426         cpumask_t cpumask;
1427
1428         daemonize("kswapd%d", pgdat->node_id);
1429         cpumask = node_to_cpumask(pgdat->node_id);
1430         if (!cpus_empty(cpumask))
1431                 set_cpus_allowed(tsk, cpumask);
1432         current->reclaim_state = &reclaim_state;
1433
1434         /*
1435          * Tell the memory management that we're a "memory allocator",
1436          * and that if we need more memory we should get access to it
1437          * regardless (see "__alloc_pages()"). "kswapd" should
1438          * never get caught in the normal page freeing logic.
1439          *
1440          * (Kswapd normally doesn't need memory anyway, but sometimes
1441          * you need a small amount of memory in order to be able to
1442          * page out something else, and this flag essentially protects
1443          * us from recursively trying to free more memory as we're
1444          * trying to free the first piece of memory in the first place).
1445          */
1446         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1447
1448         order = 0;
1449         for ( ; ; ) {
1450                 unsigned long new_order;
1451
1452                 try_to_freeze();
1453
1454                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1455                 new_order = pgdat->kswapd_max_order;
1456                 pgdat->kswapd_max_order = 0;
1457                 if (order < new_order) {
1458                         /*
1459                          * Don't sleep if someone wants a larger 'order'
1460                          * allocation
1461                          */
1462                         order = new_order;
1463                 } else {
1464                         schedule();
1465                         order = pgdat->kswapd_max_order;
1466                 }
1467                 finish_wait(&pgdat->kswapd_wait, &wait);
1468
1469                 balance_pgdat(pgdat, 0, order);
1470         }
1471         return 0;
1472 }
1473
1474 /*
1475  * A zone is low on free memory, so wake its kswapd task to service it.
1476  */
1477 void wakeup_kswapd(struct zone *zone, int order)
1478 {
1479         pg_data_t *pgdat;
1480
1481         if (!populated_zone(zone))
1482                 return;
1483
1484         pgdat = zone->zone_pgdat;
1485         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1486                 return;
1487         if (pgdat->kswapd_max_order < order)
1488                 pgdat->kswapd_max_order = order;
1489         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1490                 return;
1491         if (!waitqueue_active(&pgdat->kswapd_wait))
1492                 return;
1493         wake_up_interruptible(&pgdat->kswapd_wait);
1494 }
1495
1496 #ifdef CONFIG_PM
1497 /*
1498  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1499  * pages.
1500  */
1501 int shrink_all_memory(int nr_pages)
1502 {
1503         pg_data_t *pgdat;
1504         int nr_to_free = nr_pages;
1505         int ret = 0;
1506         struct reclaim_state reclaim_state = {
1507                 .reclaimed_slab = 0,
1508         };
1509
1510         current->reclaim_state = &reclaim_state;
1511         for_each_pgdat(pgdat) {
1512                 int freed;
1513                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1514                 ret += freed;
1515                 nr_to_free -= freed;
1516                 if (nr_to_free <= 0)
1517                         break;
1518         }
1519         current->reclaim_state = NULL;
1520         return ret;
1521 }
1522 #endif
1523
1524 #ifdef CONFIG_HOTPLUG_CPU
1525 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1526    not required for correctness.  So if the last cpu in a node goes
1527    away, we get changed to run anywhere: as the first one comes back,
1528    restore their cpu bindings. */
1529 static int __devinit cpu_callback(struct notifier_block *nfb,
1530                                   unsigned long action,
1531                                   void *hcpu)
1532 {
1533         pg_data_t *pgdat;
1534         cpumask_t mask;
1535
1536         if (action == CPU_ONLINE) {
1537                 for_each_pgdat(pgdat) {
1538                         mask = node_to_cpumask(pgdat->node_id);
1539                         if (any_online_cpu(mask) != NR_CPUS)
1540                                 /* One of our CPUs online: restore mask */
1541                                 set_cpus_allowed(pgdat->kswapd, mask);
1542                 }
1543         }
1544         return NOTIFY_OK;
1545 }
1546 #endif /* CONFIG_HOTPLUG_CPU */
1547
1548 static int __init kswapd_init(void)
1549 {
1550         pg_data_t *pgdat;
1551         swap_setup();
1552         for_each_pgdat(pgdat)
1553                 pgdat->kswapd
1554                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1555         total_memory = nr_free_pagecache_pages();
1556         hotcpu_notifier(cpu_callback, 0);
1557         return 0;
1558 }
1559
1560 module_init(kswapd_init)