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                         if (!mapping && page_count(page) == 1)
 474                                 goto free_it;
 475                 }
 476
 477                 if (!mapping)
 478                         goto keep_locked;       /* truncate got there first */
 479
 480                 spin_lock_irq(&mapping->tree_lock);
 481
 482                 /*
 483                  * The non-racy check for busy page.  It is critical to check
 484                  * PageDirty _after_ making sure that the page is freeable and
 485                  * not in use by anybody.       (pagecache + us == 2)
 486                  */
 487                 if (page_count(page) != 2 || PageDirty(page)) {
 488                         spin_unlock_irq(&mapping->tree_lock);
 489                         goto keep_locked;
 490                 }
 491
 492 #ifdef CONFIG_SWAP
 493                 if (PageSwapCache(page)) {
 494                         swp_entry_t swap = { .val = page->private };
 495                         __delete_from_swap_cache(page);
 496                         spin_unlock_irq(&mapping->tree_lock);
 497                         swap_free(swap);
 498                         __put_page(page);       /* The pagecache ref */
 499                         goto free_it;
 500                 }
 501 #endif /* CONFIG_SWAP */
 502
 503                 __remove_from_page_cache(page);
 504                 spin_unlock_irq(&mapping->tree_lock);
 505                 __put_page(page);
 506
 507 free_it:
 508                 unlock_page(page);
 509                 reclaimed++;
 510                 if (!pagevec_add(&freed_pvec, page))
 511                         __pagevec_release_nonlru(&freed_pvec);
 512                 continue;
 513
 514 activate_locked:
 515                 SetPageActive(page);
 516                 pgactivate++;
 517 keep_locked:
 518                 unlock_page(page);
 519 keep:
 520                 list_add(&page->lru, &ret_pages);
 521                 BUG_ON(PageLRU(page));
 522         }
 523         list_splice(&ret_pages, page_list);
 524         if (pagevec_count(&freed_pvec))
 525                 __pagevec_release_nonlru(&freed_pvec);
 526         mod_page_state(pgactivate, pgactivate);
 527         sc->nr_reclaimed += reclaimed;
 528         return reclaimed;
 529 }
 530
 531 /*
 532  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 533  * a batch of pages and working on them outside the lock.  Any pages which were
 534  * not freed will be added back to the LRU.
 535  *
 536  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 537  *
 538  * For pagecache intensive workloads, the first loop here is the hottest spot
 539  * in the kernel (apart from the copy_*_user functions).
 540  */
 541 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 542 {
 543         LIST_HEAD(page_list);
 544         struct pagevec pvec;
 545         int max_scan = sc->nr_to_scan;
 546
 547         pagevec_init(&pvec, 1);
 548
 549         lru_add_drain();
 550         spin_lock_irq(&zone->lru_lock);
 551         while (max_scan > 0) {
 552                 struct page *page;
 553                 int nr_taken = 0;
 554                 int nr_scan = 0;
 555                 int nr_freed;
 556
 557                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 558                                 !list_empty(&zone->inactive_list)) {
 559                         page = lru_to_page(&zone->inactive_list);
 560
 561                         prefetchw_prev_lru_page(page,
 562                                                 &zone->inactive_list, flags);
 563
 564                         if (!TestClearPageLRU(page))
 565                                 BUG();
 566                         list_del(&page->lru);
 567                         if (get_page_testone(page)) {
 568                                 /*
 569                                  * It is being freed elsewhere
 570                                  */
 571                                 __put_page(page);
 572                                 SetPageLRU(page);
 573                                 list_add(&page->lru, &zone->inactive_list);
 574                                 continue;
 575                         }
 576                         list_add(&page->lru, &page_list);
 577                         nr_taken++;
 578                 }
 579                 zone->nr_inactive -= nr_taken;
 580                 zone->pages_scanned += nr_taken;
 581                 spin_unlock_irq(&zone->lru_lock);
 582
 583                 if (nr_taken == 0)
 584                         goto done;
 585
 586                 max_scan -= nr_scan;
 587                 if (current_is_kswapd())
 588                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 589                 else
 590                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 591                 nr_freed = shrink_list(&page_list, sc);
 592                 if (current_is_kswapd())
 593                         mod_page_state(kswapd_steal, nr_freed);
 594                 mod_page_state_zone(zone, pgsteal, nr_freed);
 595                 sc->nr_to_reclaim -= nr_freed;
 596
 597                 spin_lock_irq(&zone->lru_lock);
 598                 /*
 599                  * Put back any unfreeable pages.
 600                  */
 601                 while (!list_empty(&page_list)) {
 602                         page = lru_to_page(&page_list);
 603                         if (TestSetPageLRU(page))
 604                                 BUG();
 605                         list_del(&page->lru);
 606                         if (PageActive(page))
 607                                 add_page_to_active_list(zone, page);
 608                         else
 609                                 add_page_to_inactive_list(zone, page);
 610                         if (!pagevec_add(&pvec, page)) {
 611                                 spin_unlock_irq(&zone->lru_lock);
 612                                 __pagevec_release(&pvec);
 613                                 spin_lock_irq(&zone->lru_lock);
 614                         }
 615                 }
 616         }
 617         spin_unlock_irq(&zone->lru_lock);
 618 done:
 619         pagevec_release(&pvec);
 620 }
 621
 622 /*
 623  * This moves pages from the active list to the inactive list.
 624  *
 625  * We move them the other way if the page is referenced by one or more
 626  * processes, from rmap.
 627  *
 628  * If the pages are mostly unmapped, the processing is fast and it is
 629  * appropriate to hold zone->lru_lock across the whole operation.  But if
 630  * the pages are mapped, the processing is slow (page_referenced()) so we
 631  * should drop zone->lru_lock around each page.  It's impossible to balance
 632  * this, so instead we remove the pages from the LRU while processing them.
 633  * It is safe to rely on PG_active against the non-LRU pages in here because
 634  * nobody will play with that bit on a non-LRU page.
 635  *
 636  * The downside is that we have to touch page->_count against each page.
 637  * But we had to alter page->flags anyway.
 638  */
 639 static void
 640 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 641 {
 642         int pgmoved;
 643         int pgdeactivate = 0;
 644         int pgscanned = 0;
 645         int nr_pages = sc->nr_to_scan;
 646         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 647         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 648         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 649         struct page *page;
 650         struct pagevec pvec;
 651         int reclaim_mapped = 0;
 652         long mapped_ratio;
 653         long distress;
 654         long swap_tendency;
 655
 656         lru_add_drain();
 657         pgmoved = 0;
 658         spin_lock_irq(&zone->lru_lock);
 659         while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
 660                 page = lru_to_page(&zone->active_list);
 661                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 662                 if (!TestClearPageLRU(page))
 663                         BUG();
 664                 list_del(&page->lru);
 665                 if (get_page_testone(page)) {
 666                         /*
 667                          * It was already free!  release_pages() or put_page()
 668                          * are about to remove it from the LRU and free it. So
 669                          * put the refcount back and put the page back on the
 670                          * LRU
 671                          */
 672                         __put_page(page);
 673                         SetPageLRU(page);
 674                         list_add(&page->lru, &zone->active_list);
 675                 } else {
 676                         list_add(&page->lru, &l_hold);
 677                         pgmoved++;
 678                 }
 679                 pgscanned++;
 680         }
 681         zone->nr_active -= pgmoved;
 682         spin_unlock_irq(&zone->lru_lock);
 683
 684         /*
 685          * `distress' is a measure of how much trouble we're having reclaiming
 686          * pages.  0 -> no problems.  100 -> great trouble.
 687          */
 688         distress = 100 >> zone->prev_priority;
 689
 690         /*
 691          * The point of this algorithm is to decide when to start reclaiming
 692          * mapped memory instead of just pagecache.  Work out how much memory
 693          * is mapped.
 694          */
 695         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 696
 697         /*
 698          * Now decide how much we really want to unmap some pages.  The mapped
 699          * ratio is downgraded - just because there's a lot of mapped memory
 700          * doesn't necessarily mean that page reclaim isn't succeeding.
 701          *
 702          * The distress ratio is important - we don't want to start going oom.
 703          *
 704          * A 100% value of vm_swappiness overrides this algorithm altogether.
 705          */
 706         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 707
 708         /*
 709          * Now use this metric to decide whether to start moving mapped memory
 710          * onto the inactive list.
 711          */
 712         if (swap_tendency >= 100)
 713                 reclaim_mapped = 1;
 714
 715         while (!list_empty(&l_hold)) {
 716                 page = lru_to_page(&l_hold);
 717                 list_del(&page->lru);
 718                 if (page_mapped(page)) {
 719                         if (!reclaim_mapped) {
 720                                 list_add(&page->lru, &l_active);
 721                                 continue;
 722                         }
 723                         page_map_lock(page);
 724                         if (page_referenced(page)) {
 725                                 page_map_unlock(page);
 726                                 list_add(&page->lru, &l_active);
 727                                 continue;
 728                         }
 729                         page_map_unlock(page);
 730                 }
 731                 /*
 732                  * FIXME: need to consider page_count(page) here if/when we
 733                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 734                  */
 735                 if (total_swap_pages == 0 && PageAnon(page)) {
 736                         list_add(&page->lru, &l_active);
 737                         continue;
 738                 }
 739                 list_add(&page->lru, &l_inactive);
 740         }
 741
 742         pagevec_init(&pvec, 1);
 743         pgmoved = 0;
 744         spin_lock_irq(&zone->lru_lock);
 745         while (!list_empty(&l_inactive)) {
 746                 page = lru_to_page(&l_inactive);
 747                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 748                 if (TestSetPageLRU(page))
 749                         BUG();
 750                 if (!TestClearPageActive(page))
 751                         BUG();
 752                 list_move(&page->lru, &zone->inactive_list);
 753                 pgmoved++;
 754                 if (!pagevec_add(&pvec, page)) {
 755                         zone->nr_inactive += pgmoved;
 756                         spin_unlock_irq(&zone->lru_lock);
 757                         pgdeactivate += pgmoved;
 758                         pgmoved = 0;
 759                         if (buffer_heads_over_limit)
 760                                 pagevec_strip(&pvec);
 761                         __pagevec_release(&pvec);
 762                         spin_lock_irq(&zone->lru_lock);
 763                 }
 764         }
 765         zone->nr_inactive += pgmoved;
 766         pgdeactivate += pgmoved;
 767         if (buffer_heads_over_limit) {
 768                 spin_unlock_irq(&zone->lru_lock);
 769                 pagevec_strip(&pvec);
 770                 spin_lock_irq(&zone->lru_lock);
 771         }
 772
 773         pgmoved = 0;
 774         while (!list_empty(&l_active)) {
 775                 page = lru_to_page(&l_active);
 776                 prefetchw_prev_lru_page(page, &l_active, flags);
 777                 if (TestSetPageLRU(page))
 778                         BUG();
 779                 BUG_ON(!PageActive(page));
 780                 list_move(&page->lru, &zone->active_list);
 781                 pgmoved++;
 782                 if (!pagevec_add(&pvec, page)) {
 783                         zone->nr_active += pgmoved;
 784                         pgmoved = 0;
 785                         spin_unlock_irq(&zone->lru_lock);
 786                         __pagevec_release(&pvec);
 787                         spin_lock_irq(&zone->lru_lock);
 788                 }
 789         }
 790         zone->nr_active += pgmoved;
 791         spin_unlock_irq(&zone->lru_lock);
 792         pagevec_release(&pvec);
 793
 794         mod_page_state_zone(zone, pgrefill, pgscanned);
 795         mod_page_state(pgdeactivate, pgdeactivate);
 796 }
 797
 798 /*
 799  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 800  */
 801 static void
 802 shrink_zone(struct zone *zone, struct scan_control *sc)
 803 {
 804         unsigned long nr_active;
 805         unsigned long nr_inactive;
 806
 807         /*
 808          * Add one to `nr_to_scan' just to make sure that the kernel will
 809          * slowly sift through the active list.
 810          */
 811         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 812         nr_active = zone->nr_scan_active;
 813         if (nr_active >= SWAP_CLUSTER_MAX)
 814                 zone->nr_scan_active = 0;
 815         else
 816                 nr_active = 0;
 817
 818         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 819         nr_inactive = zone->nr_scan_inactive;
 820         if (nr_inactive >= SWAP_CLUSTER_MAX)
 821                 zone->nr_scan_inactive = 0;
 822         else
 823                 nr_inactive = 0;
 824
 825         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 826
 827         while (nr_active || nr_inactive) {
 828                 if (nr_active) {
 829                         sc->nr_to_scan = min(nr_active,
 830                                         (unsigned long)SWAP_CLUSTER_MAX);
 831                         nr_active -= sc->nr_to_scan;
 832                         refill_inactive_zone(zone, sc);
 833                 }
 834
 835                 if (nr_inactive) {
 836                         sc->nr_to_scan = min(nr_inactive,
 837                                         (unsigned long)SWAP_CLUSTER_MAX);
 838                         nr_inactive -= sc->nr_to_scan;
 839                         shrink_cache(zone, sc);
 840                         if (sc->nr_to_reclaim <= 0)
 841                                 break;
 842                 }
 843         }
 844 }
 845
 846 /*
 847  * This is the direct reclaim path, for page-allocating processes.  We only
 848  * try to reclaim pages from zones which will satisfy the caller's allocation
 849  * request.
 850  *
 851  * We reclaim from a zone even if that zone is over pages_high.  Because:
 852  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 853  *    allocation or
 854  * b) The zones may be over pages_high but they must go *over* pages_high to
 855  *    satisfy the `incremental min' zone defense algorithm.
 856  *
 857  * Returns the number of reclaimed pages.
 858  *
 859  * If a zone is deemed to be full of pinned pages then just give it a light
 860  * scan then give up on it.
 861  */
 862 static void
 863 shrink_caches(struct zone **zones, struct scan_control *sc)
 864 {
 865         int i;
 866
 867         for (i = 0; zones[i] != NULL; i++) {
 868                 struct zone *zone = zones[i];
 869
 870                 zone->temp_priority = sc->priority;
 871                 if (zone->prev_priority > sc->priority)
 872                         zone->prev_priority = sc->priority;
 873
 874                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 875                         continue;       /* Let kswapd poll it */
 876
 877                 shrink_zone(zone, sc);
 878         }
 879 }
 880
 881 /*
 882  * This is the main entry point to direct page reclaim.
 883  *
 884  * If a full scan of the inactive list fails to free enough memory then we
 885  * are "out of memory" and something needs to be killed.
 886  *
 887  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 888  * high - the zone may be full of dirty or under-writeback pages, which this
 889  * caller can't do much about.  We kick pdflush and take explicit naps in the
 890  * hope that some of these pages can be written.  But if the allocating task
 891  * holds filesystem locks which prevent writeout this might not work, and the
 892  * allocation attempt will fail.
 893  */
 894 int try_to_free_pages(struct zone **zones,
 895                 unsigned int gfp_mask, unsigned int order)
 896 {
 897         int priority;
 898         int ret = 0;
 899         int total_scanned = 0, total_reclaimed = 0;
 900         struct reclaim_state *reclaim_state = current->reclaim_state;
 901         struct scan_control sc;
 902         unsigned long lru_pages = 0;
 903         int i;
 904
 905         sc.gfp_mask = gfp_mask;
 906         sc.may_writepage = 0;
 907
 908         inc_page_state(allocstall);
 909
 910         for (i = 0; zones[i] != NULL; i++) {
 911                 struct zone *zone = zones[i];
 912
 913                 zone->temp_priority = DEF_PRIORITY;
 914                 lru_pages += zone->nr_active + zone->nr_inactive;
 915         }
 916
 917         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 918                 sc.nr_mapped = read_page_state(nr_mapped);
 919                 sc.nr_scanned = 0;
 920                 sc.nr_reclaimed = 0;
 921                 sc.priority = priority;
 922                 shrink_caches(zones, &sc);
 923                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
 924                 if (reclaim_state) {
 925                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
 926                         reclaim_state->reclaimed_slab = 0;
 927                 }
 928                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
 929                         ret = 1;
 930                         goto out;
 931                 }
 932                 total_scanned += sc.nr_scanned;
 933                 total_reclaimed += sc.nr_reclaimed;
 934
 935                 /*
 936                  * Try to write back as many pages as we just scanned.  This
 937                  * tends to cause slow streaming writers to write data to the
 938                  * disk smoothly, at the dirtying rate, which is nice.   But
 939                  * that's undesirable in laptop mode, where we *want* lumpy
 940                  * writeout.  So in laptop mode, write out the whole world.
 941                  */
 942                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
 943                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
 944                         sc.may_writepage = 1;
 945                 }
 946
 947                 /* Take a nap, wait for some writeback to complete */
 948                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
 949                         blk_congestion_wait(WRITE, HZ/10);
 950         }
 951         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 952                 out_of_memory(gfp_mask);
 953 out:
 954         for (i = 0; zones[i] != 0; i++)
 955                 zones[i]->prev_priority = zones[i]->temp_priority;
 956         return ret;
 957 }
 958
 959 /*
 960  * For kswapd, balance_pgdat() will work across all this node's zones until
 961  * they are all at pages_high.
 962  *
 963  * If `nr_pages' is non-zero then it is the number of pages which are to be
 964  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 965  * special.
 966  *
 967  * Returns the number of pages which were actually freed.
 968  *
 969  * There is special handling here for zones which are full of pinned pages.
 970  * This can happen if the pages are all mlocked, or if they are all used by
 971  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 972  * What we do is to detect the case where all pages in the zone have been
 973  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 974  * dead and from now on, only perform a short scan.  Basically we're polling
 975  * the zone for when the problem goes away.
 976  *
 977  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
 978  * zones which have free_pages > pages_high, but once a zone is found to have
 979  * free_pages <= pages_high, we scan that zone and the lower zones regardless
 980  * of the number of free pages in the lower zones.  This interoperates with
 981  * the page allocator fallback scheme to ensure that aging of pages is balanced
 982  * across the zones.
 983  */
 984 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
 985 {
 986         int to_free = nr_pages;
 987         int priority;
 988         int i;
 989         int total_scanned = 0, total_reclaimed = 0;
 990         struct reclaim_state *reclaim_state = current->reclaim_state;
 991         struct scan_control sc;
 992
 993         sc.gfp_mask = GFP_KERNEL;
 994         sc.may_writepage = 0;
 995         sc.nr_mapped = read_page_state(nr_mapped);
 996
 997         inc_page_state(pageoutrun);
 998
 999         for (i = 0; i < pgdat->nr_zones; i++) {
1000                 struct zone *zone = pgdat->node_zones + i;
1001
1002                 zone->temp_priority = DEF_PRIORITY;
1003         }
1004
1005         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1006                 int all_zones_ok = 1;
1007                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1008                 unsigned long lru_pages = 0;
1009
1010                 if (nr_pages == 0) {
1011                         /*
1012                          * Scan in the highmem->dma direction for the highest
1013                          * zone which needs scanning
1014                          */
1015                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1016                                 struct zone *zone = pgdat->node_zones + i;
1017
1018                                 if (zone->all_unreclaimable &&
1019                                                 priority != DEF_PRIORITY)
1020                                         continue;
1021
1022                                 if (zone->free_pages <= zone->pages_high) {
1023                                         end_zone = i;
1024                                         goto scan;
1025                                 }
1026                         }
1027                         goto out;
1028                 } else {
1029                         end_zone = pgdat->nr_zones - 1;
1030                 }
1031 scan:
1032                 for (i = 0; i <= end_zone; i++) {
1033                         struct zone *zone = pgdat->node_zones + i;
1034
1035                         lru_pages += zone->nr_active + zone->nr_inactive;
1036                 }
1037
1038                 /*
1039                  * Now scan the zone in the dma->highmem direction, stopping
1040                  * at the last zone which needs scanning.
1041                  *
1042                  * We do this because the page allocator works in the opposite
1043                  * direction.  This prevents the page allocator from allocating
1044                  * pages behind kswapd's direction of progress, which would
1045                  * cause too much scanning of the lower zones.
1046                  */
1047                 for (i = 0; i <= end_zone; i++) {
1048                         struct zone *zone = pgdat->node_zones + i;
1049
1050                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1051                                 continue;
1052
1053                         if (nr_pages == 0) {    /* Not software suspend */
1054                                 if (zone->free_pages <= zone->pages_high)
1055                                         all_zones_ok = 0;
1056                         }
1057                         zone->temp_priority = priority;
1058                         if (zone->prev_priority > priority)
1059                                 zone->prev_priority = priority;
1060                         sc.nr_scanned = 0;
1061                         sc.nr_reclaimed = 0;
1062                         sc.priority = priority;
1063                         shrink_zone(zone, &sc);
1064                         reclaim_state->reclaimed_slab = 0;
1065                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1066                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1067                         total_reclaimed += sc.nr_reclaimed;
1068                         if (zone->all_unreclaimable)
1069                                 continue;
1070                         if (zone->pages_scanned > zone->present_pages * 2)
1071                                 zone->all_unreclaimable = 1;
1072                         /*
1073                          * If we've done a decent amount of scanning and
1074                          * the reclaim ratio is low, start doing writepage
1075                          * even in laptop mode
1076                          */
1077                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1078                             total_scanned > total_reclaimed+total_reclaimed/2)
1079                                 sc.may_writepage = 1;
1080                 }
1081                 if (nr_pages && to_free > total_reclaimed)
1082                         continue;       /* swsusp: need to do more work */
1083                 if (all_zones_ok)
1084                         break;          /* kswapd: all done */
1085                 /*
1086                  * OK, kswapd is getting into trouble.  Take a nap, then take
1087                  * another pass across the zones.
1088                  */
1089                 if (total_scanned && priority < DEF_PRIORITY - 2)
1090                         blk_congestion_wait(WRITE, HZ/10);
1091         }
1092 out:
1093         for (i = 0; i < pgdat->nr_zones; i++) {
1094                 struct zone *zone = pgdat->node_zones + i;
1095
1096                 zone->prev_priority = zone->temp_priority;
1097         }
1098         return total_reclaimed;
1099 }
1100
1101 /*
1102  * The background pageout daemon, started as a kernel thread
1103  * from the init process.
1104  *
1105  * This basically trickles out pages so that we have _some_
1106  * free memory available even if there is no other activity
1107  * that frees anything up. This is needed for things like routing
1108  * etc, where we otherwise might have all activity going on in
1109  * asynchronous contexts that cannot page things out.
1110  *
1111  * If there are applications that are active memory-allocators
1112  * (most normal use), this basically shouldn't matter.
1113  */
1114 static int kswapd(void *p)
1115 {
1116         pg_data_t *pgdat = (pg_data_t*)p;
1117         struct task_struct *tsk = current;
1118         DEFINE_WAIT(wait);
1119         struct reclaim_state reclaim_state = {
1120                 .reclaimed_slab = 0,
1121         };
1122         cpumask_t cpumask;
1123
1124         daemonize("kswapd%d", pgdat->node_id);
1125         cpumask = node_to_cpumask(pgdat->node_id);
1126         if (!cpus_empty(cpumask))
1127                 set_cpus_allowed(tsk, cpumask);
1128         current->reclaim_state = &reclaim_state;
1129
1130         /*
1131          * Tell the memory management that we're a "memory allocator",
1132          * and that if we need more memory we should get access to it
1133          * regardless (see "__alloc_pages()"). "kswapd" should
1134          * never get caught in the normal page freeing logic.
1135          *
1136          * (Kswapd normally doesn't need memory anyway, but sometimes
1137          * you need a small amount of memory in order to be able to
1138          * page out something else, and this flag essentially protects
1139          * us from recursively trying to free more memory as we're
1140          * trying to free the first piece of memory in the first place).
1141          */
1142         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1143
1144         for ( ; ; ) {
1145                 if (current->flags & PF_FREEZE)
1146                         refrigerator(PF_FREEZE);
1147                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1148                 schedule();
1149                 finish_wait(&pgdat->kswapd_wait, &wait);
1150
1151                 balance_pgdat(pgdat, 0);
1152         }
1153         return 0;
1154 }
1155
1156 /*
1157  * A zone is low on free memory, so wake its kswapd task to service it.
1158  */
1159 void wakeup_kswapd(struct zone *zone)
1160 {
1161         if (zone->free_pages > zone->pages_low)
1162                 return;
1163         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1164                 return;
1165         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1166 }
1167
1168 #ifdef CONFIG_PM
1169 /*
1170  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1171  * pages.
1172  */
1173 int shrink_all_memory(int nr_pages)
1174 {
1175         pg_data_t *pgdat;
1176         int nr_to_free = nr_pages;
1177         int ret = 0;
1178         struct reclaim_state reclaim_state = {
1179                 .reclaimed_slab = 0,
1180         };
1181
1182         current->reclaim_state = &reclaim_state;
1183         for_each_pgdat(pgdat) {
1184                 int freed;
1185                 freed = balance_pgdat(pgdat, nr_to_free);
1186                 ret += freed;
1187                 nr_to_free -= freed;
1188                 if (nr_to_free <= 0)
1189                         break;
1190         }
1191         current->reclaim_state = NULL;
1192         return ret;
1193 }
1194 #endif
1195
1196 #ifdef CONFIG_HOTPLUG_CPU
1197 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1198    not required for correctness.  So if the last cpu in a node goes
1199    away, we get changed to run anywhere: as the first one comes back,
1200    restore their cpu bindings. */
1201 static int __devinit cpu_callback(struct notifier_block *nfb,
1202                                   unsigned long action,
1203                                   void *hcpu)
1204 {
1205         pg_data_t *pgdat;
1206         cpumask_t mask;
1207
1208         if (action == CPU_ONLINE) {
1209                 for_each_pgdat(pgdat) {
1210                         mask = node_to_cpumask(pgdat->node_id);
1211                         if (any_online_cpu(mask) != NR_CPUS)
1212                                 /* One of our CPUs online: restore mask */
1213                                 set_cpus_allowed(pgdat->kswapd, mask);
1214                 }
1215         }
1216         return NOTIFY_OK;
1217 }
1218 #endif /* CONFIG_HOTPLUG_CPU */
1219
1220 static int __init kswapd_init(void)
1221 {
1222         pg_data_t *pgdat;
1223         swap_setup();
1224         for_each_pgdat(pgdat)
1225                 pgdat->kswapd
1226                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1227         total_memory = nr_free_pagecache_pages();
1228         hotcpu_notifier(cpu_callback, 0);
1229         return 0;
1230 }
1231
1232 module_init(kswapd_init)