mm/vmscan.c

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