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