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