mm/vmscan.c

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