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         /* Can pages be swapped as part of reclaim? */
  75         int may_swap;
  76
  77         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  78          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  79          * In this context, it doesn't matter that we scan the
  80          * whole list at once. */
  81         int swap_cluster_max;
  82 };
  83
  84 /*
  85  * The list of shrinker callbacks used by to apply pressure to
  86  * ageable caches.
  87  */
  88 struct shrinker {
  89         shrinker_t              shrinker;
  90         struct list_head        list;
  91         int                     seeks;  /* seeks to recreate an obj */
  92         long                    nr;     /* objs pending delete */
  93 };
  94
  95 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  96
  97 #ifdef ARCH_HAS_PREFETCH
  98 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  99         do {                                                            \
 100                 if ((_page)->lru.prev != _base) {                       \
 101                         struct page *prev;                              \
 102                                                                         \
 103                         prev = lru_to_page(&(_page->lru));              \
 104                         prefetch(&prev->_field);                        \
 105                 }                                                       \
 106         } while (0)
 107 #else
 108 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 109 #endif
 110
 111 #ifdef ARCH_HAS_PREFETCHW
 112 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 113         do {                                                            \
 114                 if ((_page)->lru.prev != _base) {                       \
 115                         struct page *prev;                              \
 116                                                                         \
 117                         prev = lru_to_page(&(_page->lru));              \
 118                         prefetchw(&prev->_field);                       \
 119                 }                                                       \
 120         } while (0)
 121 #else
 122 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 123 #endif
 124
 125 /*
 126  * From 0 .. 100.  Higher means more swappy.
 127  */
 128 int vm_swappiness = 60;
 129 static long total_memory;
 130
 131 static LIST_HEAD(shrinker_list);
 132 static DECLARE_RWSEM(shrinker_rwsem);
 133
 134 /*
 135  * Add a shrinker callback to be called from the vm
 136  */
 137 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 138 {
 139         struct shrinker *shrinker;
 140
 141         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 142         if (shrinker) {
 143                 shrinker->shrinker = theshrinker;
 144                 shrinker->seeks = seeks;
 145                 shrinker->nr = 0;
 146                 down_write(&shrinker_rwsem);
 147                 list_add_tail(&shrinker->list, &shrinker_list);
 148                 up_write(&shrinker_rwsem);
 149         }
 150         return shrinker;
 151 }
 152 EXPORT_SYMBOL(set_shrinker);
 153
 154 /*
 155  * Remove one
 156  */
 157 void remove_shrinker(struct shrinker *shrinker)
 158 {
 159         down_write(&shrinker_rwsem);
 160         list_del(&shrinker->list);
 161         up_write(&shrinker_rwsem);
 162         kfree(shrinker);
 163 }
 164 EXPORT_SYMBOL(remove_shrinker);
 165
 166 #define SHRINK_BATCH 128
 167 /*
 168  * Call the shrink functions to age shrinkable caches
 169  *
 170  * Here we assume it costs one seek to replace a lru page and that it also
 171  * takes a seek to recreate a cache object.  With this in mind we age equal
 172  * percentages of the lru and ageable caches.  This should balance the seeks
 173  * generated by these structures.
 174  *
 175  * If the vm encounted mapped pages on the LRU it increase the pressure on
 176  * slab to avoid swapping.
 177  *
 178  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 179  *
 180  * `lru_pages' represents the number of on-LRU pages in all the zones which
 181  * are eligible for the caller's allocation attempt.  It is used for balancing
 182  * slab reclaim versus page reclaim.
 183  *
 184  * Returns the number of slab objects which we shrunk.
 185  */
 186 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
 187 {
 188         struct shrinker *shrinker;
 189         int ret = 0;
 190
 191         if (scanned == 0)
 192                 scanned = SWAP_CLUSTER_MAX;
 193
 194         if (!down_read_trylock(&shrinker_rwsem))
 195                 return 1;       /* Assume we'll be able to shrink next time */
 196
 197         list_for_each_entry(shrinker, &shrinker_list, list) {
 198                 unsigned long long delta;
 199                 unsigned long total_scan;
 200                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
 201
 202                 delta = (4 * scanned) / shrinker->seeks;
 203                 delta *= max_pass;
 204                 do_div(delta, lru_pages + 1);
 205                 shrinker->nr += delta;
 206                 if (shrinker->nr < 0) {
 207                         printk(KERN_ERR "%s: nr=%ld\n",
 208                                         __FUNCTION__, shrinker->nr);
 209                         shrinker->nr = max_pass;
 210                 }
 211
 212                 /*
 213                  * Avoid risking looping forever due to too large nr value:
 214                  * never try to free more than twice the estimate number of
 215                  * freeable entries.
 216                  */
 217                 if (shrinker->nr > max_pass * 2)
 218                         shrinker->nr = max_pass * 2;
 219
 220                 total_scan = shrinker->nr;
 221                 shrinker->nr = 0;
 222
 223                 while (total_scan >= SHRINK_BATCH) {
 224                         long this_scan = SHRINK_BATCH;
 225                         int shrink_ret;
 226                         int nr_before;
 227
 228                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
 229                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 230                         if (shrink_ret == -1)
 231                                 break;
 232                         if (shrink_ret < nr_before)
 233                                 ret += nr_before - shrink_ret;
 234                         mod_page_state(slabs_scanned, this_scan);
 235                         total_scan -= this_scan;
 236
 237                         cond_resched();
 238                 }
 239
 240                 shrinker->nr += total_scan;
 241         }
 242         up_read(&shrinker_rwsem);
 243         return ret;
 244 }
 245
 246 /* Called without lock on whether page is mapped, so answer is unstable */
 247 static inline int page_mapping_inuse(struct page *page)
 248 {
 249         struct address_space *mapping;
 250
 251         /* Page is in somebody's page tables. */
 252         if (page_mapped(page))
 253                 return 1;
 254
 255         /* Be more reluctant to reclaim swapcache than pagecache */
 256         if (PageSwapCache(page))
 257                 return 1;
 258
 259         mapping = page_mapping(page);
 260         if (!mapping)
 261                 return 0;
 262
 263         /* File is mmap'd by somebody? */
 264         return mapping_mapped(mapping);
 265 }
 266
 267 static inline int is_page_cache_freeable(struct page *page)
 268 {
 269         return page_count(page) - !!PagePrivate(page) == 2;
 270 }
 271
 272 static int may_write_to_queue(struct backing_dev_info *bdi)
 273 {
 274         if (current->flags & PF_SWAPWRITE)
 275                 return 1;
 276         if (!bdi_write_congested(bdi))
 277                 return 1;
 278         if (bdi == current->backing_dev_info)
 279                 return 1;
 280         return 0;
 281 }
 282
 283 /*
 284  * We detected a synchronous write error writing a page out.  Probably
 285  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 286  * fsync(), msync() or close().
 287  *
 288  * The tricky part is that after writepage we cannot touch the mapping: nothing
 289  * prevents it from being freed up.  But we have a ref on the page and once
 290  * that page is locked, the mapping is pinned.
 291  *
 292  * We're allowed to run sleeping lock_page() here because we know the caller has
 293  * __GFP_FS.
 294  */
 295 static void handle_write_error(struct address_space *mapping,
 296                                 struct page *page, int error)
 297 {
 298         lock_page(page);
 299         if (page_mapping(page) == mapping) {
 300                 if (error == -ENOSPC)
 301                         set_bit(AS_ENOSPC, &mapping->flags);
 302                 else
 303                         set_bit(AS_EIO, &mapping->flags);
 304         }
 305         unlock_page(page);
 306 }
 307
 308 /*
 309  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 310  */
 311 static pageout_t pageout(struct page *page, struct address_space *mapping)
 312 {
 313         /*
 314          * If the page is dirty, only perform writeback if that write
 315          * will be non-blocking.  To prevent this allocation from being
 316          * stalled by pagecache activity.  But note that there may be
 317          * stalls if we need to run get_block().  We could test
 318          * PagePrivate for that.
 319          *
 320          * If this process is currently in generic_file_write() against
 321          * this page's queue, we can perform writeback even if that
 322          * will block.
 323          *
 324          * If the page is swapcache, write it back even if that would
 325          * block, for some throttling. This happens by accident, because
 326          * swap_backing_dev_info is bust: it doesn't reflect the
 327          * congestion state of the swapdevs.  Easy to fix, if needed.
 328          * See swapfile.c:page_queue_congested().
 329          */
 330         if (!is_page_cache_freeable(page))
 331                 return PAGE_KEEP;
 332         if (!mapping) {
 333                 /*
 334                  * Some data journaling orphaned pages can have
 335                  * page->mapping == NULL while being dirty with clean buffers.
 336                  */
 337                 if (PagePrivate(page)) {
 338                         if (try_to_free_buffers(page)) {
 339                                 ClearPageDirty(page);
 340                                 printk("%s: orphaned page\n", __FUNCTION__);
 341                                 return PAGE_CLEAN;
 342                         }
 343                 }
 344                 return PAGE_KEEP;
 345         }
 346         if (mapping->a_ops->writepage == NULL)
 347                 return PAGE_ACTIVATE;
 348         if (!may_write_to_queue(mapping->backing_dev_info))
 349                 return PAGE_KEEP;
 350
 351         if (clear_page_dirty_for_io(page)) {
 352                 int res;
 353                 struct writeback_control wbc = {
 354                         .sync_mode = WB_SYNC_NONE,
 355                         .nr_to_write = SWAP_CLUSTER_MAX,
 356                         .nonblocking = 1,
 357                         .for_reclaim = 1,
 358                 };
 359
 360                 SetPageReclaim(page);
 361                 res = mapping->a_ops->writepage(page, &wbc);
 362                 if (res < 0)
 363                         handle_write_error(mapping, page, res);
 364                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 365                         ClearPageReclaim(page);
 366                         return PAGE_ACTIVATE;
 367                 }
 368                 if (!PageWriteback(page)) {
 369                         /* synchronous write or broken a_ops? */
 370                         ClearPageReclaim(page);
 371                 }
 372
 373                 return PAGE_SUCCESS;
 374         }
 375
 376         return PAGE_CLEAN;
 377 }
 378
 379 static int remove_mapping(struct address_space *mapping, struct page *page)
 380 {
 381         if (!mapping)
 382                 return 0;               /* truncate got there first */
 383
 384         write_lock_irq(&mapping->tree_lock);
 385
 386         /*
 387          * The non-racy check for busy page.  It is critical to check
 388          * PageDirty _after_ making sure that the page is freeable and
 389          * not in use by anybody.       (pagecache + us == 2)
 390          */
 391         if (unlikely(page_count(page) != 2))
 392                 goto cannot_free;
 393         smp_rmb();
 394         if (unlikely(PageDirty(page)))
 395                 goto cannot_free;
 396
 397         if (PageSwapCache(page)) {
 398                 swp_entry_t swap = { .val = page_private(page) };
 399                 __delete_from_swap_cache(page);
 400                 write_unlock_irq(&mapping->tree_lock);
 401                 swap_free(swap);
 402                 __put_page(page);       /* The pagecache ref */
 403                 return 1;
 404         }
 405
 406         __remove_from_page_cache(page);
 407         write_unlock_irq(&mapping->tree_lock);
 408         __put_page(page);
 409         return 1;
 410
 411 cannot_free:
 412         write_unlock_irq(&mapping->tree_lock);
 413         return 0;
 414 }
 415
 416 /*
 417  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 418  */
 419 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 420 {
 421         LIST_HEAD(ret_pages);
 422         struct pagevec freed_pvec;
 423         int pgactivate = 0;
 424         int reclaimed = 0;
 425
 426         cond_resched();
 427
 428         pagevec_init(&freed_pvec, 1);
 429         while (!list_empty(page_list)) {
 430                 struct address_space *mapping;
 431                 struct page *page;
 432                 int may_enter_fs;
 433                 int referenced;
 434
 435                 cond_resched();
 436
 437                 page = lru_to_page(page_list);
 438                 list_del(&page->lru);
 439
 440                 if (TestSetPageLocked(page))
 441                         goto keep;
 442
 443                 BUG_ON(PageActive(page));
 444
 445                 sc->nr_scanned++;
 446                 /* Double the slab pressure for mapped and swapcache pages */
 447                 if (page_mapped(page) || PageSwapCache(page))
 448                         sc->nr_scanned++;
 449
 450                 if (PageWriteback(page))
 451                         goto keep_locked;
 452
 453                 referenced = page_referenced(page, 1);
 454                 /* In active use or really unfreeable?  Activate it. */
 455                 if (referenced && page_mapping_inuse(page))
 456                         goto activate_locked;
 457
 458 #ifdef CONFIG_SWAP
 459                 /*
 460                  * Anonymous process memory has backing store?
 461                  * Try to allocate it some swap space here.
 462                  */
 463                 if (PageAnon(page) && !PageSwapCache(page)) {
 464                         if (!sc->may_swap)
 465                                 goto keep_locked;
 466                         if (!add_to_swap(page, GFP_ATOMIC))
 467                                 goto activate_locked;
 468                 }
 469 #endif /* CONFIG_SWAP */
 470
 471                 mapping = page_mapping(page);
 472                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 473                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 474
 475                 /*
 476                  * The page is mapped into the page tables of one or more
 477                  * processes. Try to unmap it here.
 478                  */
 479                 if (page_mapped(page) && mapping) {
 480                         switch (try_to_unmap(page)) {
 481                         case SWAP_FAIL:
 482                                 goto activate_locked;
 483                         case SWAP_AGAIN:
 484                                 goto keep_locked;
 485                         case SWAP_SUCCESS:
 486                                 ; /* try to free the page below */
 487                         }
 488                 }
 489
 490                 if (PageDirty(page)) {
 491                         if (referenced)
 492                                 goto keep_locked;
 493                         if (!may_enter_fs)
 494                                 goto keep_locked;
 495                         if (laptop_mode && !sc->may_writepage)
 496                                 goto keep_locked;
 497
 498                         /* Page is dirty, try to write it out here */
 499                         switch(pageout(page, mapping)) {
 500                         case PAGE_KEEP:
 501                                 goto keep_locked;
 502                         case PAGE_ACTIVATE:
 503                                 goto activate_locked;
 504                         case PAGE_SUCCESS:
 505                                 if (PageWriteback(page) || PageDirty(page))
 506                                         goto keep;
 507                                 /*
 508                                  * A synchronous write - probably a ramdisk.  Go
 509                                  * ahead and try to reclaim the page.
 510                                  */
 511                                 if (TestSetPageLocked(page))
 512                                         goto keep;
 513                                 if (PageDirty(page) || PageWriteback(page))
 514                                         goto keep_locked;
 515                                 mapping = page_mapping(page);
 516                         case PAGE_CLEAN:
 517                                 ; /* try to free the page below */
 518                         }
 519                 }
 520
 521                 /*
 522                  * If the page has buffers, try to free the buffer mappings
 523                  * associated with this page. If we succeed we try to free
 524                  * the page as well.
 525                  *
 526                  * We do this even if the page is PageDirty().
 527                  * try_to_release_page() does not perform I/O, but it is
 528                  * possible for a page to have PageDirty set, but it is actually
 529                  * clean (all its buffers are clean).  This happens if the
 530                  * buffers were written out directly, with submit_bh(). ext3
 531                  * will do this, as well as the blockdev mapping.
 532                  * try_to_release_page() will discover that cleanness and will
 533                  * drop the buffers and mark the page clean - it can be freed.
 534                  *
 535                  * Rarely, pages can have buffers and no ->mapping.  These are
 536                  * the pages which were not successfully invalidated in
 537                  * truncate_complete_page().  We try to drop those buffers here
 538                  * and if that worked, and the page is no longer mapped into
 539                  * process address space (page_count == 1) it can be freed.
 540                  * Otherwise, leave the page on the LRU so it is swappable.
 541                  */
 542                 if (PagePrivate(page)) {
 543                         if (!try_to_release_page(page, sc->gfp_mask))
 544                                 goto activate_locked;
 545                         if (!mapping && page_count(page) == 1)
 546                                 goto free_it;
 547                 }
 548
 549                 if (!remove_mapping(mapping, page))
 550                         goto keep_locked;
 551
 552 free_it:
 553                 unlock_page(page);
 554                 reclaimed++;
 555                 if (!pagevec_add(&freed_pvec, page))
 556                         __pagevec_release_nonlru(&freed_pvec);
 557                 continue;
 558
 559 activate_locked:
 560                 SetPageActive(page);
 561                 pgactivate++;
 562 keep_locked:
 563                 unlock_page(page);
 564 keep:
 565                 list_add(&page->lru, &ret_pages);
 566                 BUG_ON(PageLRU(page));
 567         }
 568         list_splice(&ret_pages, page_list);
 569         if (pagevec_count(&freed_pvec))
 570                 __pagevec_release_nonlru(&freed_pvec);
 571         mod_page_state(pgactivate, pgactivate);
 572         sc->nr_reclaimed += reclaimed;
 573         return reclaimed;
 574 }
 575
 576 #ifdef CONFIG_MIGRATION
 577 static inline void move_to_lru(struct page *page)
 578 {
 579         list_del(&page->lru);
 580         if (PageActive(page)) {
 581                 /*
 582                  * lru_cache_add_active checks that
 583                  * the PG_active bit is off.
 584                  */
 585                 ClearPageActive(page);
 586                 lru_cache_add_active(page);
 587         } else {
 588                 lru_cache_add(page);
 589         }
 590         put_page(page);
 591 }
 592
 593 /*
 594  * Add isolated pages on the list back to the LRU.
 595  *
 596  * returns the number of pages put back.
 597  */
 598 int putback_lru_pages(struct list_head *l)
 599 {
 600         struct page *page;
 601         struct page *page2;
 602         int count = 0;
 603
 604         list_for_each_entry_safe(page, page2, l, lru) {
 605                 move_to_lru(page);
 606                 count++;
 607         }
 608         return count;
 609 }
 610
 611 /*
 612  * swapout a single page
 613  * page is locked upon entry, unlocked on exit
 614  */
 615 static int swap_page(struct page *page)
 616 {
 617         struct address_space *mapping = page_mapping(page);
 618
 619         if (page_mapped(page) && mapping)
 620                 if (try_to_unmap(page) != SWAP_SUCCESS)
 621                         goto unlock_retry;
 622
 623         if (PageDirty(page)) {
 624                 /* Page is dirty, try to write it out here */
 625                 switch(pageout(page, mapping)) {
 626                 case PAGE_KEEP:
 627                 case PAGE_ACTIVATE:
 628                         goto unlock_retry;
 629
 630                 case PAGE_SUCCESS:
 631                         goto retry;
 632
 633                 case PAGE_CLEAN:
 634                         ; /* try to free the page below */
 635                 }
 636         }
 637
 638         if (PagePrivate(page)) {
 639                 if (!try_to_release_page(page, GFP_KERNEL) ||
 640                     (!mapping && page_count(page) == 1))
 641                         goto unlock_retry;
 642         }
 643
 644         if (remove_mapping(mapping, page)) {
 645                 /* Success */
 646                 unlock_page(page);
 647                 return 0;
 648         }
 649
 650 unlock_retry:
 651         unlock_page(page);
 652
 653 retry:
 654         return -EAGAIN;
 655 }
 656 /*
 657  * migrate_pages
 658  *
 659  * Two lists are passed to this function. The first list
 660  * contains the pages isolated from the LRU to be migrated.
 661  * The second list contains new pages that the pages isolated
 662  * can be moved to. If the second list is NULL then all
 663  * pages are swapped out.
 664  *
 665  * The function returns after 10 attempts or if no pages
 666  * are movable anymore because t has become empty
 667  * or no retryable pages exist anymore.
 668  *
 669  * SIMPLIFIED VERSION: This implementation of migrate_pages
 670  * is only swapping out pages and never touches the second
 671  * list. The direct migration patchset
 672  * extends this function to avoid the use of swap.
 673  *
 674  * Return: Number of pages not migrated when "to" ran empty.
 675  */
 676 int migrate_pages(struct list_head *from, struct list_head *to,
 677                   struct list_head *moved, struct list_head *failed)
 678 {
 679         int retry;
 680         int nr_failed = 0;
 681         int pass = 0;
 682         struct page *page;
 683         struct page *page2;
 684         int swapwrite = current->flags & PF_SWAPWRITE;
 685         int rc;
 686
 687         if (!swapwrite)
 688                 current->flags |= PF_SWAPWRITE;
 689
 690 redo:
 691         retry = 0;
 692
 693         list_for_each_entry_safe(page, page2, from, lru) {
 694                 cond_resched();
 695
 696                 rc = 0;
 697                 if (page_count(page) == 1)
 698                         /* page was freed from under us. So we are done. */
 699                         goto next;
 700
 701                 /*
 702                  * Skip locked pages during the first two passes to give the
 703                  * functions holding the lock time to release the page. Later we
 704                  * use lock_page() to have a higher chance of acquiring the
 705                  * lock.
 706                  */
 707                 rc = -EAGAIN;
 708                 if (pass > 2)
 709                         lock_page(page);
 710                 else
 711                         if (TestSetPageLocked(page))
 712                                 goto next;
 713
 714                 /*
 715                  * Only wait on writeback if we have already done a pass where
 716                  * we we may have triggered writeouts for lots of pages.
 717                  */
 718                 if (pass > 0) {
 719                         wait_on_page_writeback(page);
 720                 } else {
 721                         if (PageWriteback(page))
 722                                 goto unlock_page;
 723                 }
 724
 725                 /*
 726                  * Anonymous pages must have swap cache references otherwise
 727                  * the information contained in the page maps cannot be
 728                  * preserved.
 729                  */
 730                 if (PageAnon(page) && !PageSwapCache(page)) {
 731                         if (!add_to_swap(page, GFP_KERNEL)) {
 732                                 rc = -ENOMEM;
 733                                 goto unlock_page;
 734                         }
 735                 }
 736
 737                 /*
 738                  * Page is properly locked and writeback is complete.
 739                  * Try to migrate the page.
 740                  */
 741                 rc = swap_page(page);
 742                 goto next;
 743
 744 unlock_page:
 745                 unlock_page(page);
 746
 747 next:
 748                 if (rc == -EAGAIN) {
 749                         retry++;
 750                 } else if (rc) {
 751                         /* Permanent failure */
 752                         list_move(&page->lru, failed);
 753                         nr_failed++;
 754                 } else {
 755                         /* Success */
 756                         list_move(&page->lru, moved);
 757                 }
 758         }
 759         if (retry && pass++ < 10)
 760                 goto redo;
 761
 762         if (!swapwrite)
 763                 current->flags &= ~PF_SWAPWRITE;
 764
 765         return nr_failed + retry;
 766 }
 767
 768 /*
 769  * Isolate one page from the LRU lists and put it on the
 770  * indicated list with elevated refcount.
 771  *
 772  * Result:
 773  *  0 = page not on LRU list
 774  *  1 = page removed from LRU list and added to the specified list.
 775  */
 776 int isolate_lru_page(struct page *page)
 777 {
 778         int ret = 0;
 779
 780         if (PageLRU(page)) {
 781                 struct zone *zone = page_zone(page);
 782                 spin_lock_irq(&zone->lru_lock);
 783                 if (TestClearPageLRU(page)) {
 784                         ret = 1;
 785                         get_page(page);
 786                         if (PageActive(page))
 787                                 del_page_from_active_list(zone, page);
 788                         else
 789                                 del_page_from_inactive_list(zone, page);
 790                 }
 791                 spin_unlock_irq(&zone->lru_lock);
 792         }
 793
 794         return ret;
 795 }
 796 #endif
 797
 798 /*
 799  * zone->lru_lock is heavily contended.  Some of the functions that
 800  * shrink the lists perform better by taking out a batch of pages
 801  * and working on them outside the LRU lock.
 802  *
 803  * For pagecache intensive workloads, this function is the hottest
 804  * spot in the kernel (apart from copy_*_user functions).
 805  *
 806  * Appropriate locks must be held before calling this function.
 807  *
 808  * @nr_to_scan: The number of pages to look through on the list.
 809  * @src:        The LRU list to pull pages off.
 810  * @dst:        The temp list to put pages on to.
 811  * @scanned:    The number of pages that were scanned.
 812  *
 813  * returns how many pages were moved onto *@dst.
 814  */
 815 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
 816                              struct list_head *dst, int *scanned)
 817 {
 818         int nr_taken = 0;
 819         struct page *page;
 820         int scan = 0;
 821
 822         while (scan++ < nr_to_scan && !list_empty(src)) {
 823                 page = lru_to_page(src);
 824                 prefetchw_prev_lru_page(page, src, flags);
 825
 826                 if (!TestClearPageLRU(page))
 827                         BUG();
 828                 list_del(&page->lru);
 829                 if (get_page_testone(page)) {
 830                         /*
 831                          * It is being freed elsewhere
 832                          */
 833                         __put_page(page);
 834                         SetPageLRU(page);
 835                         list_add(&page->lru, src);
 836                         continue;
 837                 } else {
 838                         list_add(&page->lru, dst);
 839                         nr_taken++;
 840                 }
 841         }
 842
 843         *scanned = scan;
 844         return nr_taken;
 845 }
 846
 847 /*
 848  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 849  */
 850 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 851 {
 852         LIST_HEAD(page_list);
 853         struct pagevec pvec;
 854         int max_scan = sc->nr_to_scan;
 855
 856         pagevec_init(&pvec, 1);
 857
 858         lru_add_drain();
 859         spin_lock_irq(&zone->lru_lock);
 860         while (max_scan > 0) {
 861                 struct page *page;
 862                 int nr_taken;
 863                 int nr_scan;
 864                 int nr_freed;
 865
 866                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
 867                                              &zone->inactive_list,
 868                                              &page_list, &nr_scan);
 869                 zone->nr_inactive -= nr_taken;
 870                 zone->pages_scanned += nr_scan;
 871                 spin_unlock_irq(&zone->lru_lock);
 872
 873                 if (nr_taken == 0)
 874                         goto done;
 875
 876                 max_scan -= nr_scan;
 877                 nr_freed = shrink_list(&page_list, sc);
 878
 879                 local_irq_disable();
 880                 if (current_is_kswapd()) {
 881                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 882                         __mod_page_state(kswapd_steal, nr_freed);
 883                 } else
 884                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
 885                 __mod_page_state_zone(zone, pgsteal, nr_freed);
 886
 887                 spin_lock(&zone->lru_lock);
 888                 /*
 889                  * Put back any unfreeable pages.
 890                  */
 891                 while (!list_empty(&page_list)) {
 892                         page = lru_to_page(&page_list);
 893                         if (TestSetPageLRU(page))
 894                                 BUG();
 895                         list_del(&page->lru);
 896                         if (PageActive(page))
 897                                 add_page_to_active_list(zone, page);
 898                         else
 899                                 add_page_to_inactive_list(zone, page);
 900                         if (!pagevec_add(&pvec, page)) {
 901                                 spin_unlock_irq(&zone->lru_lock);
 902                                 __pagevec_release(&pvec);
 903                                 spin_lock_irq(&zone->lru_lock);
 904                         }
 905                 }
 906         }
 907         spin_unlock_irq(&zone->lru_lock);
 908 done:
 909         pagevec_release(&pvec);
 910 }
 911
 912 /*
 913  * This moves pages from the active list to the inactive list.
 914  *
 915  * We move them the other way if the page is referenced by one or more
 916  * processes, from rmap.
 917  *
 918  * If the pages are mostly unmapped, the processing is fast and it is
 919  * appropriate to hold zone->lru_lock across the whole operation.  But if
 920  * the pages are mapped, the processing is slow (page_referenced()) so we
 921  * should drop zone->lru_lock around each page.  It's impossible to balance
 922  * this, so instead we remove the pages from the LRU while processing them.
 923  * It is safe to rely on PG_active against the non-LRU pages in here because
 924  * nobody will play with that bit on a non-LRU page.
 925  *
 926  * The downside is that we have to touch page->_count against each page.
 927  * But we had to alter page->flags anyway.
 928  */
 929 static void
 930 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 931 {
 932         int pgmoved;
 933         int pgdeactivate = 0;
 934         int pgscanned;
 935         int nr_pages = sc->nr_to_scan;
 936         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 937         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 938         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 939         struct page *page;
 940         struct pagevec pvec;
 941         int reclaim_mapped = 0;
 942         long mapped_ratio;
 943         long distress;
 944         long swap_tendency;
 945
 946         lru_add_drain();
 947         spin_lock_irq(&zone->lru_lock);
 948         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
 949                                     &l_hold, &pgscanned);
 950         zone->pages_scanned += pgscanned;
 951         zone->nr_active -= pgmoved;
 952         spin_unlock_irq(&zone->lru_lock);
 953
 954         /*
 955          * `distress' is a measure of how much trouble we're having reclaiming
 956          * pages.  0 -> no problems.  100 -> great trouble.
 957          */
 958         distress = 100 >> zone->prev_priority;
 959
 960         /*
 961          * The point of this algorithm is to decide when to start reclaiming
 962          * mapped memory instead of just pagecache.  Work out how much memory
 963          * is mapped.
 964          */
 965         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 966
 967         /*
 968          * Now decide how much we really want to unmap some pages.  The mapped
 969          * ratio is downgraded - just because there's a lot of mapped memory
 970          * doesn't necessarily mean that page reclaim isn't succeeding.
 971          *
 972          * The distress ratio is important - we don't want to start going oom.
 973          *
 974          * A 100% value of vm_swappiness overrides this algorithm altogether.
 975          */
 976         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 977
 978         /*
 979          * Now use this metric to decide whether to start moving mapped memory
 980          * onto the inactive list.
 981          */
 982         if (swap_tendency >= 100)
 983                 reclaim_mapped = 1;
 984
 985         while (!list_empty(&l_hold)) {
 986                 cond_resched();
 987                 page = lru_to_page(&l_hold);
 988                 list_del(&page->lru);
 989                 if (page_mapped(page)) {
 990                         if (!reclaim_mapped ||
 991                             (total_swap_pages == 0 && PageAnon(page)) ||
 992                             page_referenced(page, 0)) {
 993                                 list_add(&page->lru, &l_active);
 994                                 continue;
 995                         }
 996                 }
 997                 list_add(&page->lru, &l_inactive);
 998         }
 999
1000         pagevec_init(&pvec, 1);
1001         pgmoved = 0;
1002         spin_lock_irq(&zone->lru_lock);
1003         while (!list_empty(&l_inactive)) {
1004                 page = lru_to_page(&l_inactive);
1005                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1006                 if (TestSetPageLRU(page))
1007                         BUG();
1008                 if (!TestClearPageActive(page))
1009                         BUG();
1010                 list_move(&page->lru, &zone->inactive_list);
1011                 pgmoved++;
1012                 if (!pagevec_add(&pvec, page)) {
1013                         zone->nr_inactive += pgmoved;
1014                         spin_unlock_irq(&zone->lru_lock);
1015                         pgdeactivate += pgmoved;
1016                         pgmoved = 0;
1017                         if (buffer_heads_over_limit)
1018                                 pagevec_strip(&pvec);
1019                         __pagevec_release(&pvec);
1020                         spin_lock_irq(&zone->lru_lock);
1021                 }
1022         }
1023         zone->nr_inactive += pgmoved;
1024         pgdeactivate += pgmoved;
1025         if (buffer_heads_over_limit) {
1026                 spin_unlock_irq(&zone->lru_lock);
1027                 pagevec_strip(&pvec);
1028                 spin_lock_irq(&zone->lru_lock);
1029         }
1030
1031         pgmoved = 0;
1032         while (!list_empty(&l_active)) {
1033                 page = lru_to_page(&l_active);
1034                 prefetchw_prev_lru_page(page, &l_active, flags);
1035                 if (TestSetPageLRU(page))
1036                         BUG();
1037                 BUG_ON(!PageActive(page));
1038                 list_move(&page->lru, &zone->active_list);
1039                 pgmoved++;
1040                 if (!pagevec_add(&pvec, page)) {
1041                         zone->nr_active += pgmoved;
1042                         pgmoved = 0;
1043                         spin_unlock_irq(&zone->lru_lock);
1044                         __pagevec_release(&pvec);
1045                         spin_lock_irq(&zone->lru_lock);
1046                 }
1047         }
1048         zone->nr_active += pgmoved;
1049         spin_unlock(&zone->lru_lock);
1050
1051         __mod_page_state_zone(zone, pgrefill, pgscanned);
1052         __mod_page_state(pgdeactivate, pgdeactivate);
1053         local_irq_enable();
1054
1055         pagevec_release(&pvec);
1056 }
1057
1058 /*
1059  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1060  */
1061 static void
1062 shrink_zone(struct zone *zone, struct scan_control *sc)
1063 {
1064         unsigned long nr_active;
1065         unsigned long nr_inactive;
1066
1067         atomic_inc(&zone->reclaim_in_progress);
1068
1069         /*
1070          * Add one to `nr_to_scan' just to make sure that the kernel will
1071          * slowly sift through the active list.
1072          */
1073         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1074         nr_active = zone->nr_scan_active;
1075         if (nr_active >= sc->swap_cluster_max)
1076                 zone->nr_scan_active = 0;
1077         else
1078                 nr_active = 0;
1079
1080         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1081         nr_inactive = zone->nr_scan_inactive;
1082         if (nr_inactive >= sc->swap_cluster_max)
1083                 zone->nr_scan_inactive = 0;
1084         else
1085                 nr_inactive = 0;
1086
1087         while (nr_active || nr_inactive) {
1088                 if (nr_active) {
1089                         sc->nr_to_scan = min(nr_active,
1090                                         (unsigned long)sc->swap_cluster_max);
1091                         nr_active -= sc->nr_to_scan;
1092                         refill_inactive_zone(zone, sc);
1093                 }
1094
1095                 if (nr_inactive) {
1096                         sc->nr_to_scan = min(nr_inactive,
1097                                         (unsigned long)sc->swap_cluster_max);
1098                         nr_inactive -= sc->nr_to_scan;
1099                         shrink_cache(zone, sc);
1100                 }
1101         }
1102
1103         throttle_vm_writeout();
1104
1105         atomic_dec(&zone->reclaim_in_progress);
1106 }
1107
1108 /*
1109  * This is the direct reclaim path, for page-allocating processes.  We only
1110  * try to reclaim pages from zones which will satisfy the caller's allocation
1111  * request.
1112  *
1113  * We reclaim from a zone even if that zone is over pages_high.  Because:
1114  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1115  *    allocation or
1116  * b) The zones may be over pages_high but they must go *over* pages_high to
1117  *    satisfy the `incremental min' zone defense algorithm.
1118  *
1119  * Returns the number of reclaimed pages.
1120  *
1121  * If a zone is deemed to be full of pinned pages then just give it a light
1122  * scan then give up on it.
1123  */
1124 static void
1125 shrink_caches(struct zone **zones, struct scan_control *sc)
1126 {
1127         int i;
1128
1129         for (i = 0; zones[i] != NULL; i++) {
1130                 struct zone *zone = zones[i];
1131
1132                 if (!populated_zone(zone))
1133                         continue;
1134
1135                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1136                         continue;
1137
1138                 zone->temp_priority = sc->priority;
1139                 if (zone->prev_priority > sc->priority)
1140                         zone->prev_priority = sc->priority;
1141
1142                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1143                         continue;       /* Let kswapd poll it */
1144
1145                 shrink_zone(zone, sc);
1146         }
1147 }
1148
1149 /*
1150  * This is the main entry point to direct page reclaim.
1151  *
1152  * If a full scan of the inactive list fails to free enough memory then we
1153  * are "out of memory" and something needs to be killed.
1154  *
1155  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1156  * high - the zone may be full of dirty or under-writeback pages, which this
1157  * caller can't do much about.  We kick pdflush and take explicit naps in the
1158  * hope that some of these pages can be written.  But if the allocating task
1159  * holds filesystem locks which prevent writeout this might not work, and the
1160  * allocation attempt will fail.
1161  */
1162 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1163 {
1164         int priority;
1165         int ret = 0;
1166         int total_scanned = 0, total_reclaimed = 0;
1167         struct reclaim_state *reclaim_state = current->reclaim_state;
1168         struct scan_control sc;
1169         unsigned long lru_pages = 0;
1170         int i;
1171
1172         sc.gfp_mask = gfp_mask;
1173         sc.may_writepage = 0;
1174         sc.may_swap = 1;
1175
1176         inc_page_state(allocstall);
1177
1178         for (i = 0; zones[i] != NULL; i++) {
1179                 struct zone *zone = zones[i];
1180
1181                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1182                         continue;
1183
1184                 zone->temp_priority = DEF_PRIORITY;
1185                 lru_pages += zone->nr_active + zone->nr_inactive;
1186         }
1187
1188         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1189                 sc.nr_mapped = read_page_state(nr_mapped);
1190                 sc.nr_scanned = 0;
1191                 sc.nr_reclaimed = 0;
1192                 sc.priority = priority;
1193                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1194                 if (!priority)
1195                         disable_swap_token();
1196                 shrink_caches(zones, &sc);
1197                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1198                 if (reclaim_state) {
1199                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1200                         reclaim_state->reclaimed_slab = 0;
1201                 }
1202                 total_scanned += sc.nr_scanned;
1203                 total_reclaimed += sc.nr_reclaimed;
1204                 if (total_reclaimed >= sc.swap_cluster_max) {
1205                         ret = 1;
1206                         goto out;
1207                 }
1208
1209                 /*
1210                  * Try to write back as many pages as we just scanned.  This
1211                  * tends to cause slow streaming writers to write data to the
1212                  * disk smoothly, at the dirtying rate, which is nice.   But
1213                  * that's undesirable in laptop mode, where we *want* lumpy
1214                  * writeout.  So in laptop mode, write out the whole world.
1215                  */
1216                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1217                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1218                         sc.may_writepage = 1;
1219                 }
1220
1221                 /* Take a nap, wait for some writeback to complete */
1222                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1223                         blk_congestion_wait(WRITE, HZ/10);
1224         }
1225 out:
1226         for (i = 0; zones[i] != 0; i++) {
1227                 struct zone *zone = zones[i];
1228
1229                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1230                         continue;
1231
1232                 zone->prev_priority = zone->temp_priority;
1233         }
1234         return ret;
1235 }
1236
1237 /*
1238  * For kswapd, balance_pgdat() will work across all this node's zones until
1239  * they are all at pages_high.
1240  *
1241  * If `nr_pages' is non-zero then it is the number of pages which are to be
1242  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1243  * special.
1244  *
1245  * Returns the number of pages which were actually freed.
1246  *
1247  * There is special handling here for zones which are full of pinned pages.
1248  * This can happen if the pages are all mlocked, or if they are all used by
1249  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1250  * What we do is to detect the case where all pages in the zone have been
1251  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1252  * dead and from now on, only perform a short scan.  Basically we're polling
1253  * the zone for when the problem goes away.
1254  *
1255  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1256  * zones which have free_pages > pages_high, but once a zone is found to have
1257  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1258  * of the number of free pages in the lower zones.  This interoperates with
1259  * the page allocator fallback scheme to ensure that aging of pages is balanced
1260  * across the zones.
1261  */
1262 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1263 {
1264         int to_free = nr_pages;
1265         int all_zones_ok;
1266         int priority;
1267         int i;
1268         int total_scanned, total_reclaimed;
1269         struct reclaim_state *reclaim_state = current->reclaim_state;
1270         struct scan_control sc;
1271
1272 loop_again:
1273         total_scanned = 0;
1274         total_reclaimed = 0;
1275         sc.gfp_mask = GFP_KERNEL;
1276         sc.may_writepage = 0;
1277         sc.may_swap = 1;
1278         sc.nr_mapped = read_page_state(nr_mapped);
1279
1280         inc_page_state(pageoutrun);
1281
1282         for (i = 0; i < pgdat->nr_zones; i++) {
1283                 struct zone *zone = pgdat->node_zones + i;
1284
1285                 zone->temp_priority = DEF_PRIORITY;
1286         }
1287
1288         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1289                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1290                 unsigned long lru_pages = 0;
1291
1292                 /* The swap token gets in the way of swapout... */
1293                 if (!priority)
1294                         disable_swap_token();
1295
1296                 all_zones_ok = 1;
1297
1298                 if (nr_pages == 0) {
1299                         /*
1300                          * Scan in the highmem->dma direction for the highest
1301                          * zone which needs scanning
1302                          */
1303                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1304                                 struct zone *zone = pgdat->node_zones + i;
1305
1306                                 if (!populated_zone(zone))
1307                                         continue;
1308
1309                                 if (zone->all_unreclaimable &&
1310                                                 priority != DEF_PRIORITY)
1311                                         continue;
1312
1313                                 if (!zone_watermark_ok(zone, order,
1314                                                 zone->pages_high, 0, 0)) {
1315                                         end_zone = i;
1316                                         goto scan;
1317                                 }
1318                         }
1319                         goto out;
1320                 } else {
1321                         end_zone = pgdat->nr_zones - 1;
1322                 }
1323 scan:
1324                 for (i = 0; i <= end_zone; i++) {
1325                         struct zone *zone = pgdat->node_zones + i;
1326
1327                         lru_pages += zone->nr_active + zone->nr_inactive;
1328                 }
1329
1330                 /*
1331                  * Now scan the zone in the dma->highmem direction, stopping
1332                  * at the last zone which needs scanning.
1333                  *
1334                  * We do this because the page allocator works in the opposite
1335                  * direction.  This prevents the page allocator from allocating
1336                  * pages behind kswapd's direction of progress, which would
1337                  * cause too much scanning of the lower zones.
1338                  */
1339                 for (i = 0; i <= end_zone; i++) {
1340                         struct zone *zone = pgdat->node_zones + i;
1341                         int nr_slab;
1342
1343                         if (!populated_zone(zone))
1344                                 continue;
1345
1346                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1347                                 continue;
1348
1349                         if (nr_pages == 0) {    /* Not software suspend */
1350                                 if (!zone_watermark_ok(zone, order,
1351                                                 zone->pages_high, end_zone, 0))
1352                                         all_zones_ok = 0;
1353                         }
1354                         zone->temp_priority = priority;
1355                         if (zone->prev_priority > priority)
1356                                 zone->prev_priority = priority;
1357                         sc.nr_scanned = 0;
1358                         sc.nr_reclaimed = 0;
1359                         sc.priority = priority;
1360                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1361                         atomic_inc(&zone->reclaim_in_progress);
1362                         shrink_zone(zone, &sc);
1363                         atomic_dec(&zone->reclaim_in_progress);
1364                         reclaim_state->reclaimed_slab = 0;
1365                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1366                                                 lru_pages);
1367                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1368                         total_reclaimed += sc.nr_reclaimed;
1369                         total_scanned += sc.nr_scanned;
1370                         if (zone->all_unreclaimable)
1371                                 continue;
1372                         if (nr_slab == 0 && zone->pages_scanned >=
1373                                     (zone->nr_active + zone->nr_inactive) * 4)
1374                                 zone->all_unreclaimable = 1;
1375                         /*
1376                          * If we've done a decent amount of scanning and
1377                          * the reclaim ratio is low, start doing writepage
1378                          * even in laptop mode
1379                          */
1380                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1381                             total_scanned > total_reclaimed+total_reclaimed/2)
1382                                 sc.may_writepage = 1;
1383                 }
1384                 if (nr_pages && to_free > total_reclaimed)
1385                         continue;       /* swsusp: need to do more work */
1386                 if (all_zones_ok)
1387                         break;          /* kswapd: all done */
1388                 /*
1389                  * OK, kswapd is getting into trouble.  Take a nap, then take
1390                  * another pass across the zones.
1391                  */
1392                 if (total_scanned && priority < DEF_PRIORITY - 2)
1393                         blk_congestion_wait(WRITE, HZ/10);
1394
1395                 /*
1396                  * We do this so kswapd doesn't build up large priorities for
1397                  * example when it is freeing in parallel with allocators. It
1398                  * matches the direct reclaim path behaviour in terms of impact
1399                  * on zone->*_priority.
1400                  */
1401                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1402                         break;
1403         }
1404 out:
1405         for (i = 0; i < pgdat->nr_zones; i++) {
1406                 struct zone *zone = pgdat->node_zones + i;
1407
1408                 zone->prev_priority = zone->temp_priority;
1409         }
1410         if (!all_zones_ok) {
1411                 cond_resched();
1412                 goto loop_again;
1413         }
1414
1415         return total_reclaimed;
1416 }
1417
1418 /*
1419  * The background pageout daemon, started as a kernel thread
1420  * from the init process.
1421  *
1422  * This basically trickles out pages so that we have _some_
1423  * free memory available even if there is no other activity
1424  * that frees anything up. This is needed for things like routing
1425  * etc, where we otherwise might have all activity going on in
1426  * asynchronous contexts that cannot page things out.
1427  *
1428  * If there are applications that are active memory-allocators
1429  * (most normal use), this basically shouldn't matter.
1430  */
1431 static int kswapd(void *p)
1432 {
1433         unsigned long order;
1434         pg_data_t *pgdat = (pg_data_t*)p;
1435         struct task_struct *tsk = current;
1436         DEFINE_WAIT(wait);
1437         struct reclaim_state reclaim_state = {
1438                 .reclaimed_slab = 0,
1439         };
1440         cpumask_t cpumask;
1441
1442         daemonize("kswapd%d", pgdat->node_id);
1443         cpumask = node_to_cpumask(pgdat->node_id);
1444         if (!cpus_empty(cpumask))
1445                 set_cpus_allowed(tsk, cpumask);
1446         current->reclaim_state = &reclaim_state;
1447
1448         /*
1449          * Tell the memory management that we're a "memory allocator",
1450          * and that if we need more memory we should get access to it
1451          * regardless (see "__alloc_pages()"). "kswapd" should
1452          * never get caught in the normal page freeing logic.
1453          *
1454          * (Kswapd normally doesn't need memory anyway, but sometimes
1455          * you need a small amount of memory in order to be able to
1456          * page out something else, and this flag essentially protects
1457          * us from recursively trying to free more memory as we're
1458          * trying to free the first piece of memory in the first place).
1459          */
1460         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1461
1462         order = 0;
1463         for ( ; ; ) {
1464                 unsigned long new_order;
1465
1466                 try_to_freeze();
1467
1468                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1469                 new_order = pgdat->kswapd_max_order;
1470                 pgdat->kswapd_max_order = 0;
1471                 if (order < new_order) {
1472                         /*
1473                          * Don't sleep if someone wants a larger 'order'
1474                          * allocation
1475                          */
1476                         order = new_order;
1477                 } else {
1478                         schedule();
1479                         order = pgdat->kswapd_max_order;
1480                 }
1481                 finish_wait(&pgdat->kswapd_wait, &wait);
1482
1483                 balance_pgdat(pgdat, 0, order);
1484         }
1485         return 0;
1486 }
1487
1488 /*
1489  * A zone is low on free memory, so wake its kswapd task to service it.
1490  */
1491 void wakeup_kswapd(struct zone *zone, int order)
1492 {
1493         pg_data_t *pgdat;
1494
1495         if (!populated_zone(zone))
1496                 return;
1497
1498         pgdat = zone->zone_pgdat;
1499         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1500                 return;
1501         if (pgdat->kswapd_max_order < order)
1502                 pgdat->kswapd_max_order = order;
1503         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1504                 return;
1505         if (!waitqueue_active(&pgdat->kswapd_wait))
1506                 return;
1507         wake_up_interruptible(&pgdat->kswapd_wait);
1508 }
1509
1510 #ifdef CONFIG_PM
1511 /*
1512  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1513  * pages.
1514  */
1515 int shrink_all_memory(int nr_pages)
1516 {
1517         pg_data_t *pgdat;
1518         int nr_to_free = nr_pages;
1519         int ret = 0;
1520         struct reclaim_state reclaim_state = {
1521                 .reclaimed_slab = 0,
1522         };
1523
1524         current->reclaim_state = &reclaim_state;
1525         for_each_pgdat(pgdat) {
1526                 int freed;
1527                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1528                 ret += freed;
1529                 nr_to_free -= freed;
1530                 if (nr_to_free <= 0)
1531                         break;
1532         }
1533         current->reclaim_state = NULL;
1534         return ret;
1535 }
1536 #endif
1537
1538 #ifdef CONFIG_HOTPLUG_CPU
1539 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1540    not required for correctness.  So if the last cpu in a node goes
1541    away, we get changed to run anywhere: as the first one comes back,
1542    restore their cpu bindings. */
1543 static int __devinit cpu_callback(struct notifier_block *nfb,
1544                                   unsigned long action,
1545                                   void *hcpu)
1546 {
1547         pg_data_t *pgdat;
1548         cpumask_t mask;
1549
1550         if (action == CPU_ONLINE) {
1551                 for_each_pgdat(pgdat) {
1552                         mask = node_to_cpumask(pgdat->node_id);
1553                         if (any_online_cpu(mask) != NR_CPUS)
1554                                 /* One of our CPUs online: restore mask */
1555                                 set_cpus_allowed(pgdat->kswapd, mask);
1556                 }
1557         }
1558         return NOTIFY_OK;
1559 }
1560 #endif /* CONFIG_HOTPLUG_CPU */
1561
1562 static int __init kswapd_init(void)
1563 {
1564         pg_data_t *pgdat;
1565         swap_setup();
1566         for_each_pgdat(pgdat)
1567                 pgdat->kswapd
1568                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1569         total_memory = nr_free_pagecache_pages();
1570         hotcpu_notifier(cpu_callback, 0);
1571         return 0;
1572 }
1573
1574 module_init(kswapd_init)