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 > SHRINK_BATCH) {
 157                         long nr_to_scan = shrinker->nr;
 158
 159                         shrinker->nr = 0;
 160                         mod_page_state(slabs_scanned, nr_to_scan);
 161                         while (nr_to_scan) {
 162                                 long this_scan = nr_to_scan;
 163
 164                                 if (this_scan > 128)
 165                                         this_scan = 128;
 166                                 (*shrinker->shrinker)(this_scan, gfp_mask);
 167                                 nr_to_scan -= this_scan;
 168                                 cond_resched();
 169                         }
 170                 }
 171         }
 172         up(&shrinker_sem);
 173         return 0;
 174 }
 175
 176 /* Must be called with page's rmap lock held. */
 177 static inline int page_mapping_inuse(struct page *page)
 178 {
 179         struct address_space *mapping;
 180
 181         /* Page is in somebody's page tables. */
 182         if (page_mapped(page))
 183                 return 1;
 184
 185         /* Be more reluctant to reclaim swapcache than pagecache */
 186         if (PageSwapCache(page))
 187                 return 1;
 188
 189         mapping = page_mapping(page);
 190         if (!mapping)
 191                 return 0;
 192
 193         /* File is mmap'd by somebody? */
 194         return mapping_mapped(mapping);
 195 }
 196
 197 static inline int is_page_cache_freeable(struct page *page)
 198 {
 199         return page_count(page) - !!PagePrivate(page) == 2;
 200 }
 201
 202 static int may_write_to_queue(struct backing_dev_info *bdi)
 203 {
 204         if (current_is_kswapd())
 205                 return 1;
 206         if (current_is_pdflush())       /* This is unlikely, but why not... */
 207                 return 1;
 208         if (!bdi_write_congested(bdi))
 209                 return 1;
 210         if (bdi == current->backing_dev_info)
 211                 return 1;
 212         return 0;
 213 }
 214
 215 /*
 216  * We detected a synchronous write error writing a page out.  Probably
 217  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 218  * fsync(), msync() or close().
 219  *
 220  * The tricky part is that after writepage we cannot touch the mapping: nothing
 221  * prevents it from being freed up.  But we have a ref on the page and once
 222  * that page is locked, the mapping is pinned.
 223  *
 224  * We're allowed to run sleeping lock_page() here because we know the caller has
 225  * __GFP_FS.
 226  */
 227 static void handle_write_error(struct address_space *mapping,
 228                                 struct page *page, int error)
 229 {
 230         lock_page(page);
 231         if (page_mapping(page) == mapping) {
 232                 if (error == -ENOSPC)
 233                         set_bit(AS_ENOSPC, &mapping->flags);
 234                 else
 235                         set_bit(AS_EIO, &mapping->flags);
 236         }
 237         unlock_page(page);
 238 }
 239
 240 /*
 241  * shrink_list returns the number of reclaimed pages
 242  */
 243 static int
 244 shrink_list(struct list_head *page_list, unsigned int gfp_mask,
 245                 int *nr_scanned, int do_writepage)
 246 {
 247         struct address_space *mapping;
 248         LIST_HEAD(ret_pages);
 249         struct pagevec freed_pvec;
 250         int pgactivate = 0;
 251         int ret = 0;
 252
 253         cond_resched();
 254
 255         pagevec_init(&freed_pvec, 1);
 256         while (!list_empty(page_list)) {
 257                 struct page *page;
 258                 int may_enter_fs;
 259                 int referenced;
 260
 261                 page = lru_to_page(page_list);
 262                 list_del(&page->lru);
 263
 264                 if (TestSetPageLocked(page))
 265                         goto keep;
 266
 267                 /* Double the slab pressure for mapped and swapcache pages */
 268                 if (page_mapped(page) || PageSwapCache(page))
 269                         (*nr_scanned)++;
 270
 271                 BUG_ON(PageActive(page));
 272
 273                 if (PageWriteback(page))
 274                         goto keep_locked;
 275
 276                 rmap_lock(page);
 277                 referenced = page_referenced(page);
 278                 if (referenced && page_mapping_inuse(page)) {
 279                         /* In active use or really unfreeable.  Activate it. */
 280                         rmap_unlock(page);
 281                         goto activate_locked;
 282                 }
 283
 284                 mapping = page_mapping(page);
 285                 may_enter_fs = (gfp_mask & __GFP_FS);
 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                         rmap_unlock(page);
 296                         if (!add_to_swap(page))
 297                                 goto activate_locked;
 298                         rmap_lock(page);
 299                 }
 300                 if (PageSwapCache(page)) {
 301                         mapping = &swapper_space;
 302                         may_enter_fs = (gfp_mask & __GFP_IO);
 303                 }
 304 #endif /* CONFIG_SWAP */
 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                                 rmap_unlock(page);
 314                                 goto activate_locked;
 315                         case SWAP_AGAIN:
 316                                 rmap_unlock(page);
 317                                 goto keep_locked;
 318                         case SWAP_SUCCESS:
 319                                 ; /* try to free the page below */
 320                         }
 321                 }
 322                 rmap_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 (page_count(page) == 0) {
 502                                 /* It is currently in pagevec_release() */
 503                                 SetPageLRU(page);
 504                                 list_add(&page->lru, &zone->inactive_list);
 505                                 continue;
 506                         }
 507                         list_add(&page->lru, &page_list);
 508                         page_cache_get(page);
 509                         nr_taken++;
 510                 }
 511                 zone->nr_inactive -= nr_taken;
 512                 zone->pages_scanned += nr_taken;
 513                 spin_unlock_irq(&zone->lru_lock);
 514
 515                 if (nr_taken == 0)
 516                         goto done;
 517
 518                 max_scan -= nr_scan;
 519                 if (current_is_kswapd())
 520                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 521                 else
 522                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 523                 nr_freed = shrink_list(&page_list, gfp_mask,
 524                                         total_scanned, do_writepage);
 525                 *total_scanned += nr_taken;
 526                 if (current_is_kswapd())
 527                         mod_page_state(kswapd_steal, nr_freed);
 528                 mod_page_state_zone(zone, pgsteal, nr_freed);
 529
 530                 ret += nr_freed;
 531                 if (nr_freed <= 0 && list_empty(&page_list))
 532                         goto done;
 533
 534                 spin_lock_irq(&zone->lru_lock);
 535                 /*
 536                  * Put back any unfreeable pages.
 537                  */
 538                 while (!list_empty(&page_list)) {
 539                         page = lru_to_page(&page_list);
 540                         if (TestSetPageLRU(page))
 541                                 BUG();
 542                         list_del(&page->lru);
 543                         if (PageActive(page))
 544                                 add_page_to_active_list(zone, page);
 545                         else
 546                                 add_page_to_inactive_list(zone, page);
 547                         if (!pagevec_add(&pvec, page)) {
 548                                 spin_unlock_irq(&zone->lru_lock);
 549                                 __pagevec_release(&pvec);
 550                                 spin_lock_irq(&zone->lru_lock);
 551                         }
 552                 }
 553         }
 554         spin_unlock_irq(&zone->lru_lock);
 555 done:
 556         pagevec_release(&pvec);
 557         return ret;
 558 }
 559
 560 /*
 561  * This moves pages from the active list to the inactive list.
 562  *
 563  * We move them the other way if the page is referenced by one or more
 564  * processes, from rmap.
 565  *
 566  * If the pages are mostly unmapped, the processing is fast and it is
 567  * appropriate to hold zone->lru_lock across the whole operation.  But if
 568  * the pages are mapped, the processing is slow (page_referenced()) so we
 569  * should drop zone->lru_lock around each page.  It's impossible to balance
 570  * this, so instead we remove the pages from the LRU while processing them.
 571  * It is safe to rely on PG_active against the non-LRU pages in here because
 572  * nobody will play with that bit on a non-LRU page.
 573  *
 574  * The downside is that we have to touch page->count against each page.
 575  * But we had to alter page->flags anyway.
 576  */
 577 static void
 578 refill_inactive_zone(struct zone *zone, const int nr_pages_in,
 579                         struct page_state *ps)
 580 {
 581         int pgmoved;
 582         int pgdeactivate = 0;
 583         int nr_pages = nr_pages_in;
 584         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 585         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 586         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 587         struct page *page;
 588         struct pagevec pvec;
 589         int reclaim_mapped = 0;
 590         long mapped_ratio;
 591         long distress;
 592         long swap_tendency;
 593
 594         lru_add_drain();
 595         pgmoved = 0;
 596         spin_lock_irq(&zone->lru_lock);
 597         while (nr_pages && !list_empty(&zone->active_list)) {
 598                 page = lru_to_page(&zone->active_list);
 599                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 600                 if (!TestClearPageLRU(page))
 601                         BUG();
 602                 list_del(&page->lru);
 603                 if (page_count(page) == 0) {
 604                         /* It is currently in pagevec_release() */
 605                         SetPageLRU(page);
 606                         list_add(&page->lru, &zone->active_list);
 607                 } else {
 608                         page_cache_get(page);
 609                         list_add(&page->lru, &l_hold);
 610                         pgmoved++;
 611                 }
 612                 nr_pages--;
 613         }
 614         zone->nr_active -= pgmoved;
 615         spin_unlock_irq(&zone->lru_lock);
 616
 617         /*
 618          * `distress' is a measure of how much trouble we're having reclaiming
 619          * pages.  0 -> no problems.  100 -> great trouble.
 620          */
 621         distress = 100 >> zone->prev_priority;
 622
 623         /*
 624          * The point of this algorithm is to decide when to start reclaiming
 625          * mapped memory instead of just pagecache.  Work out how much memory
 626          * is mapped.
 627          */
 628         mapped_ratio = (ps->nr_mapped * 100) / total_memory;
 629
 630         /*
 631          * Now decide how much we really want to unmap some pages.  The mapped
 632          * ratio is downgraded - just because there's a lot of mapped memory
 633          * doesn't necessarily mean that page reclaim isn't succeeding.
 634          *
 635          * The distress ratio is important - we don't want to start going oom.
 636          *
 637          * A 100% value of vm_swappiness overrides this algorithm altogether.
 638          */
 639         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 640
 641         /*
 642          * Now use this metric to decide whether to start moving mapped memory
 643          * onto the inactive list.
 644          */
 645         if (swap_tendency >= 100)
 646                 reclaim_mapped = 1;
 647
 648         while (!list_empty(&l_hold)) {
 649                 page = lru_to_page(&l_hold);
 650                 list_del(&page->lru);
 651                 if (page_mapped(page)) {
 652                         if (!reclaim_mapped) {
 653                                 list_add(&page->lru, &l_active);
 654                                 continue;
 655                         }
 656                         rmap_lock(page);
 657                         if (page_referenced(page)) {
 658                                 rmap_unlock(page);
 659                                 list_add(&page->lru, &l_active);
 660                                 continue;
 661                         }
 662                         rmap_unlock(page);
 663                 }
 664                 /*
 665                  * FIXME: need to consider page_count(page) here if/when we
 666                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 667                  */
 668                 if (total_swap_pages == 0 && PageAnon(page)) {
 669                         list_add(&page->lru, &l_active);
 670                         continue;
 671                 }
 672                 list_add(&page->lru, &l_inactive);
 673         }
 674
 675         pagevec_init(&pvec, 1);
 676         pgmoved = 0;
 677         spin_lock_irq(&zone->lru_lock);
 678         while (!list_empty(&l_inactive)) {
 679                 page = lru_to_page(&l_inactive);
 680                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 681                 if (TestSetPageLRU(page))
 682                         BUG();
 683                 if (!TestClearPageActive(page))
 684                         BUG();
 685                 list_move(&page->lru, &zone->inactive_list);
 686                 pgmoved++;
 687                 if (!pagevec_add(&pvec, page)) {
 688                         zone->nr_inactive += pgmoved;
 689                         spin_unlock_irq(&zone->lru_lock);
 690                         pgdeactivate += pgmoved;
 691                         pgmoved = 0;
 692                         if (buffer_heads_over_limit)
 693                                 pagevec_strip(&pvec);
 694                         __pagevec_release(&pvec);
 695                         spin_lock_irq(&zone->lru_lock);
 696                 }
 697         }
 698         zone->nr_inactive += pgmoved;
 699         pgdeactivate += pgmoved;
 700         if (buffer_heads_over_limit) {
 701                 spin_unlock_irq(&zone->lru_lock);
 702                 pagevec_strip(&pvec);
 703                 spin_lock_irq(&zone->lru_lock);
 704         }
 705
 706         pgmoved = 0;
 707         while (!list_empty(&l_active)) {
 708                 page = lru_to_page(&l_active);
 709                 prefetchw_prev_lru_page(page, &l_active, flags);
 710                 if (TestSetPageLRU(page))
 711                         BUG();
 712                 BUG_ON(!PageActive(page));
 713                 list_move(&page->lru, &zone->active_list);
 714                 pgmoved++;
 715                 if (!pagevec_add(&pvec, page)) {
 716                         zone->nr_active += pgmoved;
 717                         pgmoved = 0;
 718                         spin_unlock_irq(&zone->lru_lock);
 719                         __pagevec_release(&pvec);
 720                         spin_lock_irq(&zone->lru_lock);
 721                 }
 722         }
 723         zone->nr_active += pgmoved;
 724         spin_unlock_irq(&zone->lru_lock);
 725         pagevec_release(&pvec);
 726
 727         mod_page_state_zone(zone, pgrefill, nr_pages_in - nr_pages);
 728         mod_page_state(pgdeactivate, pgdeactivate);
 729 }
 730
 731 /*
 732  * Scan `nr_pages' from this zone.  Returns the number of reclaimed pages.
 733  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 734  */
 735 static int
 736 shrink_zone(struct zone *zone, int max_scan, unsigned int gfp_mask,
 737                 int *total_scanned, struct page_state *ps, int do_writepage)
 738 {
 739         unsigned long ratio;
 740         int count;
 741
 742         /*
 743          * Try to keep the active list 2/3 of the size of the cache.  And
 744          * make sure that refill_inactive is given a decent number of pages.
 745          *
 746          * The "ratio+1" here is important.  With pagecache-intensive workloads
 747          * the inactive list is huge, and `ratio' evaluates to zero all the
 748          * time.  Which pins the active list memory.  So we add one to `ratio'
 749          * just to make sure that the kernel will slowly sift through the
 750          * active list.
 751          */
 752         ratio = (unsigned long)SWAP_CLUSTER_MAX * zone->nr_active /
 753                                 ((zone->nr_inactive | 1) * 2);
 754
 755         atomic_add(ratio+1, &zone->nr_scan_active);
 756         count = atomic_read(&zone->nr_scan_active);
 757         if (count >= SWAP_CLUSTER_MAX) {
 758                 atomic_set(&zone->nr_scan_active, 0);
 759                 refill_inactive_zone(zone, count, ps);
 760         }
 761
 762         atomic_add(max_scan, &zone->nr_scan_inactive);
 763         count = atomic_read(&zone->nr_scan_inactive);
 764         if (count >= SWAP_CLUSTER_MAX) {
 765                 atomic_set(&zone->nr_scan_inactive, 0);
 766                 return shrink_cache(zone, gfp_mask, count,
 767                                         total_scanned, do_writepage);
 768         }
 769         return 0;
 770 }
 771
 772 /*
 773  * This is the direct reclaim path, for page-allocating processes.  We only
 774  * try to reclaim pages from zones which will satisfy the caller's allocation
 775  * request.
 776  *
 777  * We reclaim from a zone even if that zone is over pages_high.  Because:
 778  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 779  *    allocation or
 780  * b) The zones may be over pages_high but they must go *over* pages_high to
 781  *    satisfy the `incremental min' zone defense algorithm.
 782  *
 783  * Returns the number of reclaimed pages.
 784  *
 785  * If a zone is deemed to be full of pinned pages then just give it a light
 786  * scan then give up on it.
 787  */
 788 static int
 789 shrink_caches(struct zone **zones, int priority, int *total_scanned,
 790                 int gfp_mask, struct page_state *ps, int do_writepage)
 791 {
 792         int ret = 0;
 793         int i;
 794
 795         for (i = 0; zones[i] != NULL; i++) {
 796                 struct zone *zone = zones[i];
 797                 int max_scan;
 798
 799                 if (zone->free_pages < zone->pages_high)
 800                         zone->temp_priority = priority;
 801
 802                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
 803                         continue;       /* Let kswapd poll it */
 804
 805                 max_scan = zone->nr_inactive >> priority;
 806                 ret += shrink_zone(zone, max_scan, gfp_mask,
 807                                         total_scanned, ps, do_writepage);
 808         }
 809         return ret;
 810 }
 811
 812 /*
 813  * This is the main entry point to direct page reclaim.
 814  *
 815  * If a full scan of the inactive list fails to free enough memory then we
 816  * are "out of memory" and something needs to be killed.
 817  *
 818  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 819  * high - the zone may be full of dirty or under-writeback pages, which this
 820  * caller can't do much about.  So for !__GFP_FS callers, we just perform a
 821  * small LRU walk and if that didn't work out, fail the allocation back to the
 822  * caller.  GFP_NOFS allocators need to know how to deal with it.  Kicking
 823  * bdflush, waiting and retrying will work.
 824  *
 825  * This is a fairly lame algorithm - it can result in excessive CPU burning and
 826  * excessive rotation of the inactive list, which is _supposed_ to be an LRU,
 827  * yes?
 828  */
 829 int try_to_free_pages(struct zone **zones,
 830                 unsigned int gfp_mask, unsigned int order)
 831 {
 832         int priority;
 833         int ret = 0;
 834         int nr_reclaimed = 0;
 835         struct reclaim_state *reclaim_state = current->reclaim_state;
 836         int i;
 837         unsigned long total_scanned = 0;
 838         int do_writepage = 0;
 839
 840         inc_page_state(allocstall);
 841
 842         for (i = 0; zones[i] != 0; i++)
 843                 zones[i]->temp_priority = DEF_PRIORITY;
 844
 845         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 846                 int scanned = 0;
 847                 struct page_state ps;
 848
 849                 get_page_state(&ps);
 850                 nr_reclaimed += shrink_caches(zones, priority, &scanned,
 851                                                 gfp_mask, &ps, do_writepage);
 852                 shrink_slab(scanned, gfp_mask);
 853                 if (reclaim_state) {
 854                         nr_reclaimed += reclaim_state->reclaimed_slab;
 855                         reclaim_state->reclaimed_slab = 0;
 856                 }
 857                 if (nr_reclaimed >= SWAP_CLUSTER_MAX) {
 858                         ret = 1;
 859                         goto out;
 860                 }
 861                 if (!(gfp_mask & __GFP_FS))
 862                         break;          /* Let the caller handle it */
 863                 /*
 864                  * Try to write back as many pages as we just scanned.  This
 865                  * tends to cause slow streaming writers to write data to the
 866                  * disk smoothly, at the dirtying rate, which is nice.   But
 867                  * that's undesirable in laptop mode, where we *want* lumpy
 868                  * writeout.  So in laptop mode, write out the whole world.
 869                  */
 870                 total_scanned += scanned;
 871                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
 872                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
 873                         do_writepage = 1;
 874                 }
 875
 876                 /* Take a nap, wait for some writeback to complete */
 877                 if (scanned && priority < DEF_PRIORITY - 2)
 878                         blk_congestion_wait(WRITE, HZ/10);
 879         }
 880         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 881                 out_of_memory();
 882 out:
 883         for (i = 0; zones[i] != 0; i++)
 884                 zones[i]->prev_priority = zones[i]->temp_priority;
 885         return ret;
 886 }
 887
 888 /*
 889  * For kswapd, balance_pgdat() will work across all this node's zones until
 890  * they are all at pages_high.
 891  *
 892  * If `nr_pages' is non-zero then it is the number of pages which are to be
 893  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 894  * special.
 895  *
 896  * Returns the number of pages which were actually freed.
 897  *
 898  * There is special handling here for zones which are full of pinned pages.
 899  * This can happen if the pages are all mlocked, or if they are all used by
 900  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 901  * What we do is to detect the case where all pages in the zone have been
 902  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 903  * dead and from now on, only perform a short scan.  Basically we're polling
 904  * the zone for when the problem goes away.
 905  *
 906  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
 907  * zones which have free_pages > pages_high, but once a zone is found to have
 908  * free_pages <= pages_high, we scan that zone and the lower zones regardless
 909  * of the number of free pages in the lower zones.  This interoperates with
 910  * the page allocator fallback scheme to ensure that aging of pages is balanced
 911  * across the zones.
 912  */
 913 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, struct page_state *ps)
 914 {
 915         int to_free = nr_pages;
 916         int priority;
 917         int i;
 918         struct reclaim_state *reclaim_state = current->reclaim_state;
 919         unsigned long total_scanned = 0;
 920         unsigned long total_reclaimed = 0;
 921         int do_writepage = 0;
 922
 923         inc_page_state(pageoutrun);
 924
 925         for (i = 0; i < pgdat->nr_zones; i++) {
 926                 struct zone *zone = pgdat->node_zones + i;
 927
 928                 zone->temp_priority = DEF_PRIORITY;
 929         }
 930
 931         for (priority = DEF_PRIORITY; priority; priority--) {
 932                 int all_zones_ok = 1;
 933                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
 934
 935
 936                 if (nr_pages == 0) {
 937                         /*
 938                          * Scan in the highmem->dma direction for the highest
 939                          * zone which needs scanning
 940                          */
 941                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
 942                                 struct zone *zone = pgdat->node_zones + i;
 943
 944                                 if (zone->all_unreclaimable &&
 945                                                 priority != DEF_PRIORITY)
 946                                         continue;
 947
 948                                 if (zone->free_pages <= zone->pages_high) {
 949                                         end_zone = i;
 950                                         goto scan;
 951                                 }
 952                         }
 953                         goto out;
 954                 } else {
 955                         end_zone = pgdat->nr_zones - 1;
 956                 }
 957 scan:
 958                 /*
 959                  * Now scan the zone in the dma->highmem direction, stopping
 960                  * at the last zone which needs scanning.
 961                  *
 962                  * We do this because the page allocator works in the opposite
 963                  * direction.  This prevents the page allocator from allocating
 964                  * pages behind kswapd's direction of progress, which would
 965                  * cause too much scanning of the lower zones.
 966                  */
 967                 for (i = 0; i <= end_zone; i++) {
 968                         struct zone *zone = pgdat->node_zones + i;
 969                         int max_scan;
 970                         int reclaimed;
 971                         int scanned = 0;
 972
 973                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
 974                                 continue;
 975
 976                         if (nr_pages == 0) {    /* Not software suspend */
 977                                 if (zone->free_pages <= zone->pages_high)
 978                                         all_zones_ok = 0;
 979                         }
 980                         zone->temp_priority = priority;
 981                         max_scan = zone->nr_inactive >> priority;
 982                         reclaimed = shrink_zone(zone, max_scan, GFP_KERNEL,
 983                                         &scanned, ps, do_writepage);
 984                         total_scanned += scanned;
 985                         reclaim_state->reclaimed_slab = 0;
 986                         shrink_slab(scanned, GFP_KERNEL);
 987                         reclaimed += reclaim_state->reclaimed_slab;
 988                         total_reclaimed += reclaimed;
 989                         to_free -= reclaimed;
 990                         if (zone->all_unreclaimable)
 991                                 continue;
 992                         if (zone->pages_scanned > zone->present_pages * 2)
 993                                 zone->all_unreclaimable = 1;
 994                         /*
 995                          * If we've done a decent amount of scanning and
 996                          * the reclaim ratio is low, start doing writepage
 997                          * even in laptop mode
 998                          */
 999                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1000                             total_scanned > total_reclaimed+total_reclaimed/2)
1001                                 do_writepage = 1;
1002                 }
1003                 if (nr_pages && to_free > 0)
1004                         continue;       /* swsusp: need to do more work */
1005                 if (all_zones_ok)
1006                         break;          /* kswapd: all done */
1007                 /*
1008                  * OK, kswapd is getting into trouble.  Take a nap, then take
1009                  * another pass across the zones.
1010                  */
1011                 if (total_scanned && priority < DEF_PRIORITY - 2)
1012                         blk_congestion_wait(WRITE, HZ/10);
1013         }
1014 out:
1015         for (i = 0; i < pgdat->nr_zones; i++) {
1016                 struct zone *zone = pgdat->node_zones + i;
1017
1018                 zone->prev_priority = zone->temp_priority;
1019         }
1020         return total_reclaimed;
1021 }
1022
1023 /*
1024  * The background pageout daemon, started as a kernel thread
1025  * from the init process.
1026  *
1027  * This basically trickles out pages so that we have _some_
1028  * free memory available even if there is no other activity
1029  * that frees anything up. This is needed for things like routing
1030  * etc, where we otherwise might have all activity going on in
1031  * asynchronous contexts that cannot page things out.
1032  *
1033  * If there are applications that are active memory-allocators
1034  * (most normal use), this basically shouldn't matter.
1035  */
1036 int kswapd(void *p)
1037 {
1038         pg_data_t *pgdat = (pg_data_t*)p;
1039         struct task_struct *tsk = current;
1040         DEFINE_WAIT(wait);
1041         struct reclaim_state reclaim_state = {
1042                 .reclaimed_slab = 0,
1043         };
1044         cpumask_t cpumask;
1045
1046         daemonize("kswapd%d", pgdat->node_id);
1047         cpumask = node_to_cpumask(pgdat->node_id);
1048         if (!cpus_empty(cpumask))
1049                 set_cpus_allowed(tsk, cpumask);
1050         current->reclaim_state = &reclaim_state;
1051
1052         /*
1053          * Tell the memory management that we're a "memory allocator",
1054          * and that if we need more memory we should get access to it
1055          * regardless (see "__alloc_pages()"). "kswapd" should
1056          * never get caught in the normal page freeing logic.
1057          *
1058          * (Kswapd normally doesn't need memory anyway, but sometimes
1059          * you need a small amount of memory in order to be able to
1060          * page out something else, and this flag essentially protects
1061          * us from recursively trying to free more memory as we're
1062          * trying to free the first piece of memory in the first place).
1063          */
1064         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1065
1066         for ( ; ; ) {
1067                 struct page_state ps;
1068
1069                 if (current->flags & PF_FREEZE)
1070                         refrigerator(PF_FREEZE);
1071                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1072                 schedule();
1073                 finish_wait(&pgdat->kswapd_wait, &wait);
1074                 get_page_state(&ps);
1075                 balance_pgdat(pgdat, 0, &ps);
1076         }
1077 }
1078
1079 /*
1080  * A zone is low on free memory, so wake its kswapd task to service it.
1081  */
1082 void wakeup_kswapd(struct zone *zone)
1083 {
1084         if (zone->free_pages > zone->pages_low)
1085                 return;
1086         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1087                 return;
1088         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1089 }
1090
1091 #ifdef CONFIG_PM
1092 /*
1093  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1094  * pages.
1095  */
1096 int shrink_all_memory(int nr_pages)
1097 {
1098         pg_data_t *pgdat;
1099         int nr_to_free = nr_pages;
1100         int ret = 0;
1101         struct reclaim_state reclaim_state = {
1102                 .reclaimed_slab = 0,
1103         };
1104
1105         current->reclaim_state = &reclaim_state;
1106         for_each_pgdat(pgdat) {
1107                 int freed;
1108                 struct page_state ps;
1109
1110                 get_page_state(&ps);
1111                 freed = balance_pgdat(pgdat, nr_to_free, &ps);
1112                 ret += freed;
1113                 nr_to_free -= freed;
1114                 if (nr_to_free <= 0)
1115                         break;
1116         }
1117         current->reclaim_state = NULL;
1118         return ret;
1119 }
1120 #endif
1121
1122 #ifdef CONFIG_HOTPLUG_CPU
1123 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1124    not required for correctness.  So if the last cpu in a node goes
1125    away, we get changed to run anywhere: as the first one comes back,
1126    restore their cpu bindings. */
1127 static int __devinit cpu_callback(struct notifier_block *nfb,
1128                                   unsigned long action,
1129                                   void *hcpu)
1130 {
1131         pg_data_t *pgdat;
1132         cpumask_t mask;
1133
1134         if (action == CPU_ONLINE) {
1135                 for_each_pgdat(pgdat) {
1136                         mask = node_to_cpumask(pgdat->node_id);
1137                         if (any_online_cpu(mask) != NR_CPUS)
1138                                 /* One of our CPUs online: restore mask */
1139                                 set_cpus_allowed(pgdat->kswapd, mask);
1140                 }
1141         }
1142         return NOTIFY_OK;
1143 }
1144 #endif /* CONFIG_HOTPLUG_CPU */
1145
1146 static int __init kswapd_init(void)
1147 {
1148         pg_data_t *pgdat;
1149         swap_setup();
1150         for_each_pgdat(pgdat)
1151                 pgdat->kswapd
1152                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1153         total_memory = nr_free_pagecache_pages();
1154         hotcpu_notifier(cpu_callback, 0);
1155         return 0;
1156 }
1157
1158 module_init(kswapd_init)