mm/vmscan.c

   1 /*
   2  *  linux/mm/vmscan.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *
   6  *  Swap reorganised 29.12.95, Stephen Tweedie.
   7  *  kswapd added: 7.1.96  sct
   8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
   9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  11  *  Multiqueue VM started 5.8.00, Rik van Riel.
  12  */
  13
  14 #include <linux/mm.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/swap.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/init.h>
  21 #include <linux/highmem.h>
  22 #include <linux/file.h>
  23 #include <linux/writeback.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  26                                         buffer_heads_over_limit */
  27 #include <linux/mm_inline.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/rmap.h>
  31 #include <linux/topology.h>
  32 #include <linux/cpu.h>
  33 #include <linux/notifier.h>
  34 #include <linux/rwsem.h>
  35
  36 #include <asm/tlbflush.h>
  37 #include <asm/div64.h>
  38
  39 #include <linux/swapops.h>
  40
  41 /* possible outcome of pageout() */
  42 typedef enum {
  43         /* failed to write page out, page is locked */
  44         PAGE_KEEP,
  45         /* move page to the active list, page is locked */
  46         PAGE_ACTIVATE,
  47         /* page has been sent to the disk successfully, page is unlocked */
  48         PAGE_SUCCESS,
  49         /* page is clean and locked */
  50         PAGE_CLEAN,
  51 } pageout_t;
  52
  53 struct scan_control {
  54         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  55         unsigned long nr_to_scan;
  56
  57         /* Incremented by the number of inactive pages that were scanned */
  58         unsigned long nr_scanned;
  59
  60         /* Incremented by the number of pages reclaimed */
  61         unsigned long nr_reclaimed;
  62
  63         unsigned long nr_mapped;        /* From page_state */
  64
  65         /* How many pages shrink_cache() should reclaim */
  66         int nr_to_reclaim;
  67
  68         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  69         unsigned int priority;
  70
  71         /* This context's GFP mask */
  72         unsigned int gfp_mask;
  73
  74         int may_writepage;
  75 };
  76
  77 /*
  78  * The list of shrinker callbacks used by to apply pressure to
  79  * ageable caches.
  80  */
  81 struct shrinker {
  82         shrinker_t              shrinker;
  83         struct list_head        list;
  84         int                     seeks;  /* seeks to recreate an obj */
  85         long                    nr;     /* objs pending delete */
  86 };
  87
  88 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  89
  90 #ifdef ARCH_HAS_PREFETCH
  91 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  92         do {                                                            \
  93                 if ((_page)->lru.prev != _base) {                       \
  94                         struct page *prev;                              \
  95                                                                         \
  96                         prev = lru_to_page(&(_page->lru));              \
  97                         prefetch(&prev->_field);                        \
  98                 }                                                       \
  99         } while (0)
 100 #else
 101 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 102 #endif
 103
 104 #ifdef ARCH_HAS_PREFETCHW
 105 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 106         do {                                                            \
 107                 if ((_page)->lru.prev != _base) {                       \
 108                         struct page *prev;                              \
 109                                                                         \
 110                         prev = lru_to_page(&(_page->lru));              \
 111                         prefetchw(&prev->_field);                       \
 112                 }                                                       \
 113         } while (0)
 114 #else
 115 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 116 #endif
 117
 118 /*
 119  * From 0 .. 100.  Higher means more swappy.
 120  */
 121 int vm_swappiness = 60;
 122 static long total_memory;
 123
 124 static LIST_HEAD(shrinker_list);
 125 static DECLARE_RWSEM(shrinker_rwsem);
 126
 127 /*
 128  * Add a shrinker callback to be called from the vm
 129  */
 130 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 131 {
 132         struct shrinker *shrinker;
 133
 134         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 135         if (shrinker) {
 136                 shrinker->shrinker = theshrinker;
 137                 shrinker->seeks = seeks;
 138                 shrinker->nr = 0;
 139                 down_write(&shrinker_rwsem);
 140                 list_add(&shrinker->list, &shrinker_list);
 141                 up_write(&shrinker_rwsem);
 142         }
 143         return shrinker;
 144 }
 145 EXPORT_SYMBOL(set_shrinker);
 146
 147 /*
 148  * Remove one
 149  */
 150 void remove_shrinker(struct shrinker *shrinker)
 151 {
 152         down_write(&shrinker_rwsem);
 153         list_del(&shrinker->list);
 154         up_write(&shrinker_rwsem);
 155         kfree(shrinker);
 156 }
 157 EXPORT_SYMBOL(remove_shrinker);
 158
 159 #define SHRINK_BATCH 128
 160 /*
 161  * Call the shrink functions to age shrinkable caches
 162  *
 163  * Here we assume it costs one seek to replace a lru page and that it also
 164  * takes a seek to recreate a cache object.  With this in mind we age equal
 165  * percentages of the lru and ageable caches.  This should balance the seeks
 166  * generated by these structures.
 167  *
 168  * If the vm encounted mapped pages on the LRU it increase the pressure on
 169  * slab to avoid swapping.
 170  *
 171  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 172  *
 173  * `lru_pages' represents the number of on-LRU pages in all the zones which
 174  * are eligible for the caller's allocation attempt.  It is used for balancing
 175  * slab reclaim versus page reclaim.
 176  */
 177 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 178                         unsigned long lru_pages)
 179 {
 180         struct shrinker *shrinker;
 181
 182         if (scanned == 0)
 183                 scanned = SWAP_CLUSTER_MAX;
 184
 185         if (!down_read_trylock(&shrinker_rwsem))
 186                 return 0;
 187
 188         list_for_each_entry(shrinker, &shrinker_list, list) {
 189                 unsigned long long delta;
 190                 unsigned long total_scan;
 191
 192                 delta = (4 * scanned) / shrinker->seeks;
 193                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 194                 do_div(delta, lru_pages + 1);
 195                 shrinker->nr += delta;
 196                 if (shrinker->nr < 0)
 197                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 198
 199                 total_scan = shrinker->nr;
 200                 shrinker->nr = 0;
 201
 202                 while (total_scan >= SHRINK_BATCH) {
 203                         long this_scan = SHRINK_BATCH;
 204                         int shrink_ret;
 205
 206                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 207                         if (shrink_ret == -1)
 208                                 break;
 209                         mod_page_state(slabs_scanned, this_scan);
 210                         total_scan -= this_scan;
 211
 212                         cond_resched();
 213                 }
 214
 215                 shrinker->nr += total_scan;
 216         }
 217         up_read(&shrinker_rwsem);
 218         return 0;
 219 }
 220
 221 /* Called without lock on whether page is mapped, so answer is unstable */
 222 static inline int page_mapping_inuse(struct page *page)
 223 {
 224         struct address_space *mapping;
 225
 226         /* Page is in somebody's page tables. */
 227         if (page_mapped(page))
 228                 return 1;
 229
 230         /* Be more reluctant to reclaim swapcache than pagecache */
 231         if (PageSwapCache(page))
 232                 return 1;
 233
 234         mapping = page_mapping(page);
 235         if (!mapping)
 236                 return 0;
 237
 238         /* File is mmap'd by somebody? */
 239         return mapping_mapped(mapping);
 240 }
 241
 242 static inline int is_page_cache_freeable(struct page *page)
 243 {
 244         return page_count(page) - !!PagePrivate(page) == 2;
 245 }
 246
 247 static int may_write_to_queue(struct backing_dev_info *bdi)
 248 {
 249         if (current_is_kswapd())
 250                 return 1;
 251         if (current_is_pdflush())       /* This is unlikely, but why not... */
 252                 return 1;
 253         if (!bdi_write_congested(bdi))
 254                 return 1;
 255         if (bdi == current->backing_dev_info)
 256                 return 1;
 257         return 0;
 258 }
 259
 260 /*
 261  * We detected a synchronous write error writing a page out.  Probably
 262  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 263  * fsync(), msync() or close().
 264  *
 265  * The tricky part is that after writepage we cannot touch the mapping: nothing
 266  * prevents it from being freed up.  But we have a ref on the page and once
 267  * that page is locked, the mapping is pinned.
 268  *
 269  * We're allowed to run sleeping lock_page() here because we know the caller has
 270  * __GFP_FS.
 271  */
 272 static void handle_write_error(struct address_space *mapping,
 273                                 struct page *page, int error)
 274 {
 275         lock_page(page);
 276         if (page_mapping(page) == mapping) {
 277                 if (error == -ENOSPC)
 278                         set_bit(AS_ENOSPC, &mapping->flags);
 279                 else
 280                         set_bit(AS_EIO, &mapping->flags);
 281         }
 282         unlock_page(page);
 283 }
 284
 285 /*
 286  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 287  */
 288 static pageout_t pageout(struct page *page, struct address_space *mapping)
 289 {
 290         /*
 291          * If the page is dirty, only perform writeback if that write
 292          * will be non-blocking.  To prevent this allocation from being
 293          * stalled by pagecache activity.  But note that there may be
 294          * stalls if we need to run get_block().  We could test
 295          * PagePrivate for that.
 296          *
 297          * If this process is currently in generic_file_write() against
 298          * this page's queue, we can perform writeback even if that
 299          * will block.
 300          *
 301          * If the page is swapcache, write it back even if that would
 302          * block, for some throttling. This happens by accident, because
 303          * swap_backing_dev_info is bust: it doesn't reflect the
 304          * congestion state of the swapdevs.  Easy to fix, if needed.
 305          * See swapfile.c:page_queue_congested().
 306          */
 307         if (!is_page_cache_freeable(page))
 308                 return PAGE_KEEP;
 309         if (!mapping)
 310                 return PAGE_KEEP;
 311         if (mapping->a_ops->writepage == NULL)
 312                 return PAGE_ACTIVATE;
 313         if (!may_write_to_queue(mapping->backing_dev_info))
 314                 return PAGE_KEEP;
 315
 316         if (clear_page_dirty_for_io(page)) {
 317                 int res;
 318                 struct writeback_control wbc = {
 319                         .sync_mode = WB_SYNC_NONE,
 320                         .nr_to_write = SWAP_CLUSTER_MAX,
 321                         .nonblocking = 1,
 322                         .for_reclaim = 1,
 323                 };
 324
 325                 SetPageReclaim(page);
 326                 res = mapping->a_ops->writepage(page, &wbc);
 327                 if (res < 0)
 328                         handle_write_error(mapping, page, res);
 329                 if (res == WRITEPAGE_ACTIVATE) {
 330                         ClearPageReclaim(page);
 331                         return PAGE_ACTIVATE;
 332                 }
 333                 if (!PageWriteback(page)) {
 334                         /* synchronous write or broken a_ops? */
 335                         ClearPageReclaim(page);
 336                 }
 337
 338                 return PAGE_SUCCESS;
 339         }
 340
 341         return PAGE_CLEAN;
 342 }
 343
 344 /*
 345  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 346  */
 347 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 348 {
 349         LIST_HEAD(ret_pages);
 350         struct pagevec freed_pvec;
 351         int pgactivate = 0;
 352         int reclaimed = 0;
 353
 354         cond_resched();
 355
 356         pagevec_init(&freed_pvec, 1);
 357         while (!list_empty(page_list)) {
 358                 struct address_space *mapping;
 359                 struct page *page;
 360                 int may_enter_fs;
 361                 int referenced;
 362
 363                 cond_resched();
 364
 365                 page = lru_to_page(page_list);
 366                 list_del(&page->lru);
 367
 368                 if (TestSetPageLocked(page))
 369                         goto keep;
 370
 371                 BUG_ON(PageActive(page));
 372
 373                 sc->nr_scanned++;
 374                 /* Double the slab pressure for mapped and swapcache pages */
 375                 if (page_mapped(page) || PageSwapCache(page))
 376                         sc->nr_scanned++;
 377
 378                 if (PageWriteback(page))
 379                         goto keep_locked;
 380
 381                 referenced = page_referenced(page, 1, sc->priority <= 0);
 382                 /* In active use or really unfreeable?  Activate it. */
 383                 if (referenced && page_mapping_inuse(page))
 384                         goto activate_locked;
 385
 386 #ifdef CONFIG_SWAP
 387                 /*
 388                  * Anonymous process memory has backing store?
 389                  * Try to allocate it some swap space here.
 390                  */
 391                 if (PageAnon(page) && !PageSwapCache(page)) {
 392                         if (!add_to_swap(page))
 393                                 goto activate_locked;
 394                 }
 395 #endif /* CONFIG_SWAP */
 396
 397                 mapping = page_mapping(page);
 398                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 399                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 400
 401                 /*
 402                  * The page is mapped into the page tables of one or more
 403                  * processes. Try to unmap it here.
 404                  */
 405                 if (page_mapped(page) && mapping) {
 406                         switch (try_to_unmap(page)) {
 407                         case SWAP_FAIL:
 408                                 goto activate_locked;
 409                         case SWAP_AGAIN:
 410                                 goto keep_locked;
 411                         case SWAP_SUCCESS:
 412                                 ; /* try to free the page below */
 413                         }
 414                 }
 415
 416                 if (PageDirty(page)) {
 417                         if (referenced)
 418                                 goto keep_locked;
 419                         if (!may_enter_fs)
 420                                 goto keep_locked;
 421                         if (laptop_mode && !sc->may_writepage)
 422                                 goto keep_locked;
 423
 424                         /* Page is dirty, try to write it out here */
 425                         switch(pageout(page, mapping)) {
 426                         case PAGE_KEEP:
 427                                 goto keep_locked;
 428                         case PAGE_ACTIVATE:
 429                                 goto activate_locked;
 430                         case PAGE_SUCCESS:
 431                                 if (PageWriteback(page) || PageDirty(page))
 432                                         goto keep;
 433                                 /*
 434                                  * A synchronous write - probably a ramdisk.  Go
 435                                  * ahead and try to reclaim the page.
 436                                  */
 437                                 if (TestSetPageLocked(page))
 438                                         goto keep;
 439                                 if (PageDirty(page) || PageWriteback(page))
 440                                         goto keep_locked;
 441                                 mapping = page_mapping(page);
 442                         case PAGE_CLEAN:
 443                                 ; /* try to free the page below */
 444                         }
 445                 }
 446
 447                 /*
 448                  * If the page has buffers, try to free the buffer mappings
 449                  * associated with this page. If we succeed we try to free
 450                  * the page as well.
 451                  *
 452                  * We do this even if the page is PageDirty().
 453                  * try_to_release_page() does not perform I/O, but it is
 454                  * possible for a page to have PageDirty set, but it is actually
 455                  * clean (all its buffers are clean).  This happens if the
 456                  * buffers were written out directly, with submit_bh(). ext3
 457                  * will do this, as well as the blockdev mapping.
 458                  * try_to_release_page() will discover that cleanness and will
 459                  * drop the buffers and mark the page clean - it can be freed.
 460                  *
 461                  * Rarely, pages can have buffers and no ->mapping.  These are
 462                  * the pages which were not successfully invalidated in
 463                  * truncate_complete_page().  We try to drop those buffers here
 464                  * and if that worked, and the page is no longer mapped into
 465                  * process address space (page_count == 1) it can be freed.
 466                  * Otherwise, leave the page on the LRU so it is swappable.
 467                  */
 468                 if (PagePrivate(page)) {
 469                         if (!try_to_release_page(page, sc->gfp_mask))
 470                                 goto activate_locked;
 471                         if (!mapping && page_count(page) == 1)
 472                                 goto free_it;
 473                 }
 474
 475                 if (!mapping)
 476                         goto keep_locked;       /* truncate got there first */
 477
 478                 spin_lock_irq(&mapping->tree_lock);
 479
 480                 /*
 481                  * The non-racy check for busy page.  It is critical to check
 482                  * PageDirty _after_ making sure that the page is freeable and
 483                  * not in use by anybody.       (pagecache + us == 2)
 484                  */
 485                 if (page_count(page) != 2 || PageDirty(page)) {
 486                         spin_unlock_irq(&mapping->tree_lock);
 487                         goto keep_locked;
 488                 }
 489
 490 #ifdef CONFIG_SWAP
 491                 if (PageSwapCache(page)) {
 492                         swp_entry_t swap = { .val = page->private };
 493                         __delete_from_swap_cache(page);
 494                         spin_unlock_irq(&mapping->tree_lock);
 495                         swap_free(swap);
 496                         __put_page(page);       /* The pagecache ref */
 497                         goto free_it;
 498                 }
 499 #endif /* CONFIG_SWAP */
 500
 501                 __remove_from_page_cache(page);
 502                 spin_unlock_irq(&mapping->tree_lock);
 503                 __put_page(page);
 504
 505 free_it:
 506                 unlock_page(page);
 507                 reclaimed++;
 508                 if (!pagevec_add(&freed_pvec, page))
 509                         __pagevec_release_nonlru(&freed_pvec);
 510                 continue;
 511
 512 activate_locked:
 513                 SetPageActive(page);
 514                 pgactivate++;
 515 keep_locked:
 516                 unlock_page(page);
 517 keep:
 518                 list_add(&page->lru, &ret_pages);
 519                 BUG_ON(PageLRU(page));
 520         }
 521         list_splice(&ret_pages, page_list);
 522         if (pagevec_count(&freed_pvec))
 523                 __pagevec_release_nonlru(&freed_pvec);
 524         mod_page_state(pgactivate, pgactivate);
 525         sc->nr_reclaimed += reclaimed;
 526         return reclaimed;
 527 }
 528
 529 /*
 530  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 531  * a batch of pages and working on them outside the lock.  Any pages which were
 532  * not freed will be added back to the LRU.
 533  *
 534  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 535  *
 536  * For pagecache intensive workloads, the first loop here is the hottest spot
 537  * in the kernel (apart from the copy_*_user functions).
 538  */
 539 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 540 {
 541         LIST_HEAD(page_list);
 542         struct pagevec pvec;
 543         int max_scan = sc->nr_to_scan;
 544
 545         pagevec_init(&pvec, 1);
 546
 547         lru_add_drain();
 548         spin_lock_irq(&zone->lru_lock);
 549         while (max_scan > 0) {
 550                 struct page *page;
 551                 int nr_taken = 0;
 552                 int nr_scan = 0;
 553                 int nr_freed;
 554
 555                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 556                                 !list_empty(&zone->inactive_list)) {
 557                         page = lru_to_page(&zone->inactive_list);
 558
 559                         prefetchw_prev_lru_page(page,
 560                                                 &zone->inactive_list, flags);
 561
 562                         if (!TestClearPageLRU(page))
 563                                 BUG();
 564                         list_del(&page->lru);
 565                         if (get_page_testone(page)) {
 566                                 /*
 567                                  * It is being freed elsewhere
 568                                  */
 569                                 __put_page(page);
 570                                 SetPageLRU(page);
 571                                 list_add(&page->lru, &zone->inactive_list);
 572                                 continue;
 573                         }
 574                         list_add(&page->lru, &page_list);
 575                         nr_taken++;
 576                 }
 577                 zone->nr_inactive -= nr_taken;
 578                 zone->pages_scanned += nr_scan;
 579                 spin_unlock_irq(&zone->lru_lock);
 580
 581                 if (nr_taken == 0)
 582                         goto done;
 583
 584                 max_scan -= nr_scan;
 585                 if (current_is_kswapd())
 586                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 587                 else
 588                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 589                 nr_freed = shrink_list(&page_list, sc);
 590                 if (current_is_kswapd())
 591                         mod_page_state(kswapd_steal, nr_freed);
 592                 mod_page_state_zone(zone, pgsteal, nr_freed);
 593                 sc->nr_to_reclaim -= nr_freed;
 594
 595                 spin_lock_irq(&zone->lru_lock);
 596                 /*
 597                  * Put back any unfreeable pages.
 598                  */
 599                 while (!list_empty(&page_list)) {
 600                         page = lru_to_page(&page_list);
 601                         if (TestSetPageLRU(page))
 602                                 BUG();
 603                         list_del(&page->lru);
 604                         if (PageActive(page))
 605                                 add_page_to_active_list(zone, page);
 606                         else
 607                                 add_page_to_inactive_list(zone, page);
 608                         if (!pagevec_add(&pvec, page)) {
 609                                 spin_unlock_irq(&zone->lru_lock);
 610                                 __pagevec_release(&pvec);
 611                                 spin_lock_irq(&zone->lru_lock);
 612                         }
 613                 }
 614         }
 615         spin_unlock_irq(&zone->lru_lock);
 616 done:
 617         pagevec_release(&pvec);
 618 }
 619
 620 /*
 621  * This moves pages from the active list to the inactive list.
 622  *
 623  * We move them the other way if the page is referenced by one or more
 624  * processes, from rmap.
 625  *
 626  * If the pages are mostly unmapped, the processing is fast and it is
 627  * appropriate to hold zone->lru_lock across the whole operation.  But if
 628  * the pages are mapped, the processing is slow (page_referenced()) so we
 629  * should drop zone->lru_lock around each page.  It's impossible to balance
 630  * this, so instead we remove the pages from the LRU while processing them.
 631  * It is safe to rely on PG_active against the non-LRU pages in here because
 632  * nobody will play with that bit on a non-LRU page.
 633  *
 634  * The downside is that we have to touch page->_count against each page.
 635  * But we had to alter page->flags anyway.
 636  */
 637 static void
 638 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 639 {
 640         int pgmoved;
 641         int pgdeactivate = 0;
 642         int pgscanned = 0;
 643         int nr_pages = sc->nr_to_scan;
 644         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 645         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 646         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 647         struct page *page;
 648         struct pagevec pvec;
 649         int reclaim_mapped = 0;
 650         long mapped_ratio;
 651         long distress;
 652         long swap_tendency;
 653
 654         lru_add_drain();
 655         pgmoved = 0;
 656         spin_lock_irq(&zone->lru_lock);
 657         while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
 658                 page = lru_to_page(&zone->active_list);
 659                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 660                 if (!TestClearPageLRU(page))
 661                         BUG();
 662                 list_del(&page->lru);
 663                 if (get_page_testone(page)) {
 664                         /*
 665                          * It was already free!  release_pages() or put_page()
 666                          * are about to remove it from the LRU and free it. So
 667                          * put the refcount back and put the page back on the
 668                          * LRU
 669                          */
 670                         __put_page(page);
 671                         SetPageLRU(page);
 672                         list_add(&page->lru, &zone->active_list);
 673                 } else {
 674                         list_add(&page->lru, &l_hold);
 675                         pgmoved++;
 676                 }
 677                 pgscanned++;
 678         }
 679         zone->pages_scanned += pgscanned;
 680         zone->nr_active -= pgmoved;
 681         spin_unlock_irq(&zone->lru_lock);
 682
 683         /*
 684          * `distress' is a measure of how much trouble we're having reclaiming
 685          * pages.  0 -> no problems.  100 -> great trouble.
 686          */
 687         distress = 100 >> zone->prev_priority;
 688
 689         /*
 690          * The point of this algorithm is to decide when to start reclaiming
 691          * mapped memory instead of just pagecache.  Work out how much memory
 692          * is mapped.
 693          */
 694         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 695
 696         /*
 697          * Now decide how much we really want to unmap some pages.  The mapped
 698          * ratio is downgraded - just because there's a lot of mapped memory
 699          * doesn't necessarily mean that page reclaim isn't succeeding.
 700          *
 701          * The distress ratio is important - we don't want to start going oom.
 702          *
 703          * A 100% value of vm_swappiness overrides this algorithm altogether.
 704          */
 705         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 706
 707         /*
 708          * Now use this metric to decide whether to start moving mapped memory
 709          * onto the inactive list.
 710          */
 711         if (swap_tendency >= 100)
 712                 reclaim_mapped = 1;
 713
 714         while (!list_empty(&l_hold)) {
 715                 cond_resched();
 716                 page = lru_to_page(&l_hold);
 717                 list_del(&page->lru);
 718                 if (page_mapped(page)) {
 719                         if (!reclaim_mapped ||
 720                             (total_swap_pages == 0 && PageAnon(page)) ||
 721                             page_referenced(page, 0, sc->priority <= 0)) {
 722                                 list_add(&page->lru, &l_active);
 723                                 continue;
 724                         }
 725                 }
 726                 list_add(&page->lru, &l_inactive);
 727         }
 728
 729         pagevec_init(&pvec, 1);
 730         pgmoved = 0;
 731         spin_lock_irq(&zone->lru_lock);
 732         while (!list_empty(&l_inactive)) {
 733                 page = lru_to_page(&l_inactive);
 734                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 735                 if (TestSetPageLRU(page))
 736                         BUG();
 737                 if (!TestClearPageActive(page))
 738                         BUG();
 739                 list_move(&page->lru, &zone->inactive_list);
 740                 pgmoved++;
 741                 if (!pagevec_add(&pvec, page)) {
 742                         zone->nr_inactive += pgmoved;
 743                         spin_unlock_irq(&zone->lru_lock);
 744                         pgdeactivate += pgmoved;
 745                         pgmoved = 0;
 746                         if (buffer_heads_over_limit)
 747                                 pagevec_strip(&pvec);
 748                         __pagevec_release(&pvec);
 749                         spin_lock_irq(&zone->lru_lock);
 750                 }
 751         }
 752         zone->nr_inactive += pgmoved;
 753         pgdeactivate += pgmoved;
 754         if (buffer_heads_over_limit) {
 755                 spin_unlock_irq(&zone->lru_lock);
 756                 pagevec_strip(&pvec);
 757                 spin_lock_irq(&zone->lru_lock);
 758         }
 759
 760         pgmoved = 0;
 761         while (!list_empty(&l_active)) {
 762                 page = lru_to_page(&l_active);
 763                 prefetchw_prev_lru_page(page, &l_active, flags);
 764                 if (TestSetPageLRU(page))
 765                         BUG();
 766                 BUG_ON(!PageActive(page));
 767                 list_move(&page->lru, &zone->active_list);
 768                 pgmoved++;
 769                 if (!pagevec_add(&pvec, page)) {
 770                         zone->nr_active += pgmoved;
 771                         pgmoved = 0;
 772                         spin_unlock_irq(&zone->lru_lock);
 773                         __pagevec_release(&pvec);
 774                         spin_lock_irq(&zone->lru_lock);
 775                 }
 776         }
 777         zone->nr_active += pgmoved;
 778         spin_unlock_irq(&zone->lru_lock);
 779         pagevec_release(&pvec);
 780
 781         mod_page_state_zone(zone, pgrefill, pgscanned);
 782         mod_page_state(pgdeactivate, pgdeactivate);
 783 }
 784
 785 /*
 786  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 787  */
 788 static void
 789 shrink_zone(struct zone *zone, struct scan_control *sc)
 790 {
 791         unsigned long nr_active;
 792         unsigned long nr_inactive;
 793
 794         /*
 795          * Add one to `nr_to_scan' just to make sure that the kernel will
 796          * slowly sift through the active list.
 797          */
 798         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 799         nr_active = zone->nr_scan_active;
 800         if (nr_active >= SWAP_CLUSTER_MAX)
 801                 zone->nr_scan_active = 0;
 802         else
 803                 nr_active = 0;
 804
 805         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 806         nr_inactive = zone->nr_scan_inactive;
 807         if (nr_inactive >= SWAP_CLUSTER_MAX)
 808                 zone->nr_scan_inactive = 0;
 809         else
 810                 nr_inactive = 0;
 811
 812         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 813
 814         while (nr_active || nr_inactive) {
 815                 if (nr_active) {
 816                         sc->nr_to_scan = min(nr_active,
 817                                         (unsigned long)SWAP_CLUSTER_MAX);
 818                         nr_active -= sc->nr_to_scan;
 819                         refill_inactive_zone(zone, sc);
 820                 }
 821
 822                 if (nr_inactive) {
 823                         sc->nr_to_scan = min(nr_inactive,
 824                                         (unsigned long)SWAP_CLUSTER_MAX);
 825                         nr_inactive -= sc->nr_to_scan;
 826                         shrink_cache(zone, sc);
 827                         if (sc->nr_to_reclaim <= 0)
 828                                 break;
 829                 }
 830         }
 831 }
 832
 833 /*
 834  * This is the direct reclaim path, for page-allocating processes.  We only
 835  * try to reclaim pages from zones which will satisfy the caller's allocation
 836  * request.
 837  *
 838  * We reclaim from a zone even if that zone is over pages_high.  Because:
 839  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 840  *    allocation or
 841  * b) The zones may be over pages_high but they must go *over* pages_high to
 842  *    satisfy the `incremental min' zone defense algorithm.
 843  *
 844  * Returns the number of reclaimed pages.
 845  *
 846  * If a zone is deemed to be full of pinned pages then just give it a light
 847  * scan then give up on it.
 848  */
 849 static void
 850 shrink_caches(struct zone **zones, struct scan_control *sc)
 851 {
 852         int i;
 853
 854         for (i = 0; zones[i] != NULL; i++) {
 855                 struct zone *zone = zones[i];
 856
 857                 if (zone->present_pages == 0)
 858                         continue;
 859
 860                 zone->temp_priority = sc->priority;
 861                 if (zone->prev_priority > sc->priority)
 862                         zone->prev_priority = sc->priority;
 863
 864                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 865                         continue;       /* Let kswapd poll it */
 866
 867                 shrink_zone(zone, sc);
 868         }
 869 }
 870
 871 /*
 872  * This is the main entry point to direct page reclaim.
 873  *
 874  * If a full scan of the inactive list fails to free enough memory then we
 875  * are "out of memory" and something needs to be killed.
 876  *
 877  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 878  * high - the zone may be full of dirty or under-writeback pages, which this
 879  * caller can't do much about.  We kick pdflush and take explicit naps in the
 880  * hope that some of these pages can be written.  But if the allocating task
 881  * holds filesystem locks which prevent writeout this might not work, and the
 882  * allocation attempt will fail.
 883  */
 884 int try_to_free_pages(struct zone **zones,
 885                 unsigned int gfp_mask, unsigned int order)
 886 {
 887         int priority;
 888         int ret = 0;
 889         int total_scanned = 0, total_reclaimed = 0;
 890         struct reclaim_state *reclaim_state = current->reclaim_state;
 891         struct scan_control sc;
 892         unsigned long lru_pages = 0;
 893         int i;
 894
 895         sc.gfp_mask = gfp_mask;
 896         sc.may_writepage = 0;
 897
 898         inc_page_state(allocstall);
 899
 900         for (i = 0; zones[i] != NULL; i++) {
 901                 struct zone *zone = zones[i];
 902
 903                 zone->temp_priority = DEF_PRIORITY;
 904                 lru_pages += zone->nr_active + zone->nr_inactive;
 905         }
 906
 907         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 908                 sc.nr_mapped = read_page_state(nr_mapped);
 909                 sc.nr_scanned = 0;
 910                 sc.nr_reclaimed = 0;
 911                 sc.priority = priority;
 912                 shrink_caches(zones, &sc);
 913                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
 914                 if (reclaim_state) {
 915                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
 916                         reclaim_state->reclaimed_slab = 0;
 917                 }
 918                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
 919                         ret = 1;
 920                         goto out;
 921                 }
 922                 total_scanned += sc.nr_scanned;
 923                 total_reclaimed += sc.nr_reclaimed;
 924
 925                 /*
 926                  * Try to write back as many pages as we just scanned.  This
 927                  * tends to cause slow streaming writers to write data to the
 928                  * disk smoothly, at the dirtying rate, which is nice.   But
 929                  * that's undesirable in laptop mode, where we *want* lumpy
 930                  * writeout.  So in laptop mode, write out the whole world.
 931                  */
 932                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
 933                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
 934                         sc.may_writepage = 1;
 935                 }
 936
 937                 /* Take a nap, wait for some writeback to complete */
 938                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
 939                         blk_congestion_wait(WRITE, HZ/10);
 940         }
 941 out:
 942         for (i = 0; zones[i] != 0; i++)
 943                 zones[i]->prev_priority = zones[i]->temp_priority;
 944         return ret;
 945 }
 946
 947 /*
 948  * For kswapd, balance_pgdat() will work across all this node's zones until
 949  * they are all at pages_high.
 950  *
 951  * If `nr_pages' is non-zero then it is the number of pages which are to be
 952  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 953  * special.
 954  *
 955  * Returns the number of pages which were actually freed.
 956  *
 957  * There is special handling here for zones which are full of pinned pages.
 958  * This can happen if the pages are all mlocked, or if they are all used by
 959  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 960  * What we do is to detect the case where all pages in the zone have been
 961  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 962  * dead and from now on, only perform a short scan.  Basically we're polling
 963  * the zone for when the problem goes away.
 964  *
 965  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
 966  * zones which have free_pages > pages_high, but once a zone is found to have
 967  * free_pages <= pages_high, we scan that zone and the lower zones regardless
 968  * of the number of free pages in the lower zones.  This interoperates with
 969  * the page allocator fallback scheme to ensure that aging of pages is balanced
 970  * across the zones.
 971  */
 972 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
 973 {
 974         int to_free = nr_pages;
 975         int all_zones_ok;
 976         int priority;
 977         int i;
 978         int total_scanned, total_reclaimed;
 979         struct reclaim_state *reclaim_state = current->reclaim_state;
 980         struct scan_control sc;
 981
 982 loop_again:
 983         total_scanned = 0;
 984         total_reclaimed = 0;
 985         sc.gfp_mask = GFP_KERNEL;
 986         sc.may_writepage = 0;
 987         sc.nr_mapped = read_page_state(nr_mapped);
 988
 989         inc_page_state(pageoutrun);
 990
 991         for (i = 0; i < pgdat->nr_zones; i++) {
 992                 struct zone *zone = pgdat->node_zones + i;
 993
 994                 zone->temp_priority = DEF_PRIORITY;
 995         }
 996
 997         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 998                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
 999                 unsigned long lru_pages = 0;
1000
1001                 all_zones_ok = 1;
1002
1003                 if (nr_pages == 0) {
1004                         /*
1005                          * Scan in the highmem->dma direction for the highest
1006                          * zone which needs scanning
1007                          */
1008                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1009                                 struct zone *zone = pgdat->node_zones + i;
1010
1011                                 if (zone->present_pages == 0)
1012                                         continue;
1013
1014                                 if (zone->all_unreclaimable &&
1015                                                 priority != DEF_PRIORITY)
1016                                         continue;
1017
1018                                 if (!zone_watermark_ok(zone, order,
1019                                                 zone->pages_high, 0, 0, 0)) {
1020                                         end_zone = i;
1021                                         goto scan;
1022                                 }
1023                         }
1024                         goto out;
1025                 } else {
1026                         end_zone = pgdat->nr_zones - 1;
1027                 }
1028 scan:
1029                 for (i = 0; i <= end_zone; i++) {
1030                         struct zone *zone = pgdat->node_zones + i;
1031
1032                         lru_pages += zone->nr_active + zone->nr_inactive;
1033                 }
1034
1035                 /*
1036                  * Now scan the zone in the dma->highmem direction, stopping
1037                  * at the last zone which needs scanning.
1038                  *
1039                  * We do this because the page allocator works in the opposite
1040                  * direction.  This prevents the page allocator from allocating
1041                  * pages behind kswapd's direction of progress, which would
1042                  * cause too much scanning of the lower zones.
1043                  */
1044                 for (i = 0; i <= end_zone; i++) {
1045                         struct zone *zone = pgdat->node_zones + i;
1046
1047                         if (zone->present_pages == 0)
1048                                 continue;
1049
1050                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1051                                 continue;
1052
1053                         if (nr_pages == 0) {    /* Not software suspend */
1054                                 if (!zone_watermark_ok(zone, order,
1055                                                 zone->pages_high, end_zone, 0, 0))
1056                                         all_zones_ok = 0;
1057                         }
1058                         zone->temp_priority = priority;
1059                         if (zone->prev_priority > priority)
1060                                 zone->prev_priority = priority;
1061                         sc.nr_scanned = 0;
1062                         sc.nr_reclaimed = 0;
1063                         sc.priority = priority;
1064                         shrink_zone(zone, &sc);
1065                         reclaim_state->reclaimed_slab = 0;
1066                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1067                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1068                         total_reclaimed += sc.nr_reclaimed;
1069                         total_scanned += sc.nr_scanned;
1070                         if (zone->all_unreclaimable)
1071                                 continue;
1072                         if (zone->pages_scanned >= (zone->nr_active +
1073                                                         zone->nr_inactive) * 4)
1074                                 zone->all_unreclaimable = 1;
1075                         /*
1076                          * If we've done a decent amount of scanning and
1077                          * the reclaim ratio is low, start doing writepage
1078                          * even in laptop mode
1079                          */
1080                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1081                             total_scanned > total_reclaimed+total_reclaimed/2)
1082                                 sc.may_writepage = 1;
1083                 }
1084                 if (nr_pages && to_free > total_reclaimed)
1085                         continue;       /* swsusp: need to do more work */
1086                 if (all_zones_ok)
1087                         break;          /* kswapd: all done */
1088                 /*
1089                  * OK, kswapd is getting into trouble.  Take a nap, then take
1090                  * another pass across the zones.
1091                  */
1092                 if (total_scanned && priority < DEF_PRIORITY - 2)
1093                         blk_congestion_wait(WRITE, HZ/10);
1094
1095                 /*
1096                  * We do this so kswapd doesn't build up large priorities for
1097                  * example when it is freeing in parallel with allocators. It
1098                  * matches the direct reclaim path behaviour in terms of impact
1099                  * on zone->*_priority.
1100                  */
1101                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1102                         break;
1103         }
1104 out:
1105         for (i = 0; i < pgdat->nr_zones; i++) {
1106                 struct zone *zone = pgdat->node_zones + i;
1107
1108                 zone->prev_priority = zone->temp_priority;
1109         }
1110         if (!all_zones_ok) {
1111                 cond_resched();
1112                 goto loop_again;
1113         }
1114
1115         return total_reclaimed;
1116 }
1117
1118 /*
1119  * The background pageout daemon, started as a kernel thread
1120  * from the init process.
1121  *
1122  * This basically trickles out pages so that we have _some_
1123  * free memory available even if there is no other activity
1124  * that frees anything up. This is needed for things like routing
1125  * etc, where we otherwise might have all activity going on in
1126  * asynchronous contexts that cannot page things out.
1127  *
1128  * If there are applications that are active memory-allocators
1129  * (most normal use), this basically shouldn't matter.
1130  */
1131 static int kswapd(void *p)
1132 {
1133         unsigned long order;
1134         pg_data_t *pgdat = (pg_data_t*)p;
1135         struct task_struct *tsk = current;
1136         DEFINE_WAIT(wait);
1137         struct reclaim_state reclaim_state = {
1138                 .reclaimed_slab = 0,
1139         };
1140         cpumask_t cpumask;
1141
1142         daemonize("kswapd%d", pgdat->node_id);
1143         cpumask = node_to_cpumask(pgdat->node_id);
1144         if (!cpus_empty(cpumask))
1145                 set_cpus_allowed(tsk, cpumask);
1146         current->reclaim_state = &reclaim_state;
1147
1148         /*
1149          * Tell the memory management that we're a "memory allocator",
1150          * and that if we need more memory we should get access to it
1151          * regardless (see "__alloc_pages()"). "kswapd" should
1152          * never get caught in the normal page freeing logic.
1153          *
1154          * (Kswapd normally doesn't need memory anyway, but sometimes
1155          * you need a small amount of memory in order to be able to
1156          * page out something else, and this flag essentially protects
1157          * us from recursively trying to free more memory as we're
1158          * trying to free the first piece of memory in the first place).
1159          */
1160         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1161
1162         order = 0;
1163         for ( ; ; ) {
1164                 unsigned long new_order;
1165                 if (current->flags & PF_FREEZE)
1166                         refrigerator(PF_FREEZE);
1167
1168                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1169                 new_order = pgdat->kswapd_max_order;
1170                 pgdat->kswapd_max_order = 0;
1171                 if (order < new_order) {
1172                         /*
1173                          * Don't sleep if someone wants a larger 'order'
1174                          * allocation
1175                          */
1176                         order = new_order;
1177                 } else {
1178                         schedule();
1179                         order = pgdat->kswapd_max_order;
1180                 }
1181                 finish_wait(&pgdat->kswapd_wait, &wait);
1182
1183                 balance_pgdat(pgdat, 0, order);
1184         }
1185         return 0;
1186 }
1187
1188 /*
1189  * A zone is low on free memory, so wake its kswapd task to service it.
1190  */
1191 void wakeup_kswapd(struct zone *zone, int order)
1192 {
1193         pg_data_t *pgdat;
1194
1195         if (zone->present_pages == 0)
1196                 return;
1197
1198         pgdat = zone->zone_pgdat;
1199         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0, 0))
1200                 return;
1201         if (pgdat->kswapd_max_order < order)
1202                 pgdat->kswapd_max_order = order;
1203         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1204                 return;
1205         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1206 }
1207
1208 #ifdef CONFIG_PM
1209 /*
1210  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1211  * pages.
1212  */
1213 int shrink_all_memory(int nr_pages)
1214 {
1215         pg_data_t *pgdat;
1216         int nr_to_free = nr_pages;
1217         int ret = 0;
1218         struct reclaim_state reclaim_state = {
1219                 .reclaimed_slab = 0,
1220         };
1221
1222         current->reclaim_state = &reclaim_state;
1223         for_each_pgdat(pgdat) {
1224                 int freed;
1225                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1226                 ret += freed;
1227                 nr_to_free -= freed;
1228                 if (nr_to_free <= 0)
1229                         break;
1230         }
1231         current->reclaim_state = NULL;
1232         return ret;
1233 }
1234 #endif
1235
1236 #ifdef CONFIG_HOTPLUG_CPU
1237 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1238    not required for correctness.  So if the last cpu in a node goes
1239    away, we get changed to run anywhere: as the first one comes back,
1240    restore their cpu bindings. */
1241 static int __devinit cpu_callback(struct notifier_block *nfb,
1242                                   unsigned long action,
1243                                   void *hcpu)
1244 {
1245         pg_data_t *pgdat;
1246         cpumask_t mask;
1247
1248         if (action == CPU_ONLINE) {
1249                 for_each_pgdat(pgdat) {
1250                         mask = node_to_cpumask(pgdat->node_id);
1251                         if (any_online_cpu(mask) != NR_CPUS)
1252                                 /* One of our CPUs online: restore mask */
1253                                 set_cpus_allowed(pgdat->kswapd, mask);
1254                 }
1255         }
1256         return NOTIFY_OK;
1257 }
1258 #endif /* CONFIG_HOTPLUG_CPU */
1259
1260 static int __init kswapd_init(void)
1261 {
1262         pg_data_t *pgdat;
1263         swap_setup();
1264         for_each_pgdat(pgdat)
1265                 pgdat->kswapd
1266                 = find_task_by_real_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1267         total_memory = nr_free_pagecache_pages();
1268         hotcpu_notifier(cpu_callback, 0);
1269         return 0;
1270 }
1271
1272 module_init(kswapd_init)