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