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