mm/vmscan.c

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