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