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/slab.h>
  16 #include <linux/kernel_stat.h>
  17 #include <linux/swap.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/init.h>
  20 #include <linux/highmem.h>
  21 #include <linux/file.h>
  22 #include <linux/writeback.h>
  23 #include <linux/suspend.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  26                                         buffer_heads_over_limit */
  27 #include <linux/mm_inline.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/rmap-locking.h>
  31 #include <linux/topology.h>
  32
  33 #include <asm/pgalloc.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/div64.h>
  36
  37 #include <linux/swapops.h>
  38
  39 /*
  40  * The "priority" of VM scanning is how much of the queues we will scan in one
  41  * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the
  42  * queues ("queue_length >> 12") during an aging round.
  43  */
  44 #define DEF_PRIORITY 12
  45
  46 /*
  47  * From 0 .. 100.  Higher means more swappy.
  48  */
  49 int vm_swappiness = 60;
  50 static long total_memory;
  51
  52 #ifdef ARCH_HAS_PREFETCH
  53 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  54         do {                                                            \
  55                 if ((_page)->lru.prev != _base) {                       \
  56                         struct page *prev;                              \
  57                                                                         \
  58                         prev = list_entry(_page->lru.prev,              \
  59                                         struct page, lru);              \
  60                         prefetch(&prev->_field);                        \
  61                 }                                                       \
  62         } while (0)
  63 #else
  64 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
  65 #endif
  66
  67 #ifdef ARCH_HAS_PREFETCHW
  68 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
  69         do {                                                            \
  70                 if ((_page)->lru.prev != _base) {                       \
  71                         struct page *prev;                              \
  72                                                                         \
  73                         prev = list_entry(_page->lru.prev,              \
  74                                         struct page, lru);              \
  75                         prefetchw(&prev->_field);                       \
  76                 }                                                       \
  77         } while (0)
  78 #else
  79 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
  80 #endif
  81
  82 /*
  83  * The list of shrinker callbacks used by to apply pressure to
  84  * ageable caches.
  85  */
  86 struct shrinker {
  87         shrinker_t              shrinker;
  88         struct list_head        list;
  89         int                     seeks;  /* seeks to recreate an obj */
  90         long                    nr;     /* objs pending delete */
  91 };
  92
  93 static LIST_HEAD(shrinker_list);
  94 static DECLARE_MUTEX(shrinker_sem);
  95
  96 /*
  97  * Add a shrinker callback to be called from the vm
  98  */
  99 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 100 {
 101         struct shrinker *shrinker;
 102
 103         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 104         if (shrinker) {
 105                 shrinker->shrinker = theshrinker;
 106                 shrinker->seeks = seeks;
 107                 shrinker->nr = 0;
 108                 down(&shrinker_sem);
 109                 list_add(&shrinker->list, &shrinker_list);
 110                 up(&shrinker_sem);
 111         }
 112         return shrinker;
 113 }
 114
 115 /*
 116  * Remove one
 117  */
 118 void remove_shrinker(struct shrinker *shrinker)
 119 {
 120         down(&shrinker_sem);
 121         list_del(&shrinker->list);
 122         up(&shrinker_sem);
 123         kfree(shrinker);
 124 }
 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(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 = 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                         while (nr_to_scan) {
 161                                 long this_scan = nr_to_scan;
 162
 163                                 if (this_scan > 128)
 164                                         this_scan = 128;
 165                                 (*shrinker->shrinker)(this_scan, gfp_mask);
 166                                 nr_to_scan -= this_scan;
 167                                 cond_resched();
 168                         }
 169                 }
 170         }
 171         up(&shrinker_sem);
 172         return 0;
 173 }
 174
 175 /* Must be called with page's pte_chain_lock held. */
 176 static inline int page_mapping_inuse(struct page *page)
 177 {
 178         struct address_space *mapping = page->mapping;
 179
 180         /* Page is in somebody's page tables. */
 181         if (page_mapped(page))
 182                 return 1;
 183
 184         /* XXX: does this happen ? */
 185         if (!mapping)
 186                 return 0;
 187
 188         /* Be more reluctant to reclaim swapcache than pagecache */
 189         if (PageSwapCache(page))
 190                 return 1;
 191
 192         /* File is mmap'd by somebody. */
 193         if (!list_empty(&mapping->i_mmap))
 194                 return 1;
 195         if (!list_empty(&mapping->i_mmap_shared))
 196                 return 1;
 197
 198         return 0;
 199 }
 200
 201 static inline int is_page_cache_freeable(struct page *page)
 202 {
 203         return page_count(page) - !!PagePrivate(page) == 2;
 204 }
 205
 206 static int may_write_to_queue(struct backing_dev_info *bdi)
 207 {
 208         if (current_is_kswapd())
 209                 return 1;
 210         if (current_is_pdflush())       /* This is unlikely, but why not... */
 211                 return 1;
 212         if (!bdi_write_congested(bdi))
 213                 return 1;
 214         if (bdi == current->backing_dev_info)
 215                 return 1;
 216         return 0;
 217 }
 218
 219 /*
 220  * shrink_list returns the number of reclaimed pages
 221  */
 222 static int
 223 shrink_list(struct list_head *page_list, unsigned int gfp_mask,
 224                 int *max_scan, int *nr_mapped)
 225 {
 226         struct address_space *mapping;
 227         LIST_HEAD(ret_pages);
 228         struct pagevec freed_pvec;
 229         int pgactivate = 0;
 230         int ret = 0;
 231
 232         cond_resched();
 233
 234         pagevec_init(&freed_pvec, 1);
 235         while (!list_empty(page_list)) {
 236                 struct page *page;
 237                 int may_enter_fs;
 238
 239                 page = list_entry(page_list->prev, struct page, lru);
 240                 list_del(&page->lru);
 241
 242                 if (TestSetPageLocked(page))
 243                         goto keep;
 244
 245                 /* Double the slab pressure for mapped and swapcache pages */
 246                 if (page_mapped(page) || PageSwapCache(page))
 247                         (*nr_mapped)++;
 248
 249                 BUG_ON(PageActive(page));
 250                 may_enter_fs = (gfp_mask & __GFP_FS) ||
 251                                 (PageSwapCache(page) && (gfp_mask & __GFP_IO));
 252
 253                 if (PageWriteback(page))
 254                         goto keep_locked;
 255
 256                 pte_chain_lock(page);
 257                 if (page_referenced(page) && page_mapping_inuse(page)) {
 258                         /* In active use or really unfreeable.  Activate it. */
 259                         pte_chain_unlock(page);
 260                         goto activate_locked;
 261                 }
 262
 263                 mapping = page->mapping;
 264
 265 #ifdef CONFIG_SWAP
 266                 /*
 267                  * Anonymous process memory without backing store. Try to
 268                  * allocate it some swap space here.
 269                  *
 270                  * XXX: implement swap clustering ?
 271                  */
 272                 if (page_mapped(page) && !mapping && !PagePrivate(page)) {
 273                         pte_chain_unlock(page);
 274                         if (!add_to_swap(page))
 275                                 goto activate_locked;
 276                         pte_chain_lock(page);
 277                         mapping = page->mapping;
 278                 }
 279 #endif /* CONFIG_SWAP */
 280
 281                 /*
 282                  * The page is mapped into the page tables of one or more
 283                  * processes. Try to unmap it here.
 284                  */
 285                 if (page_mapped(page) && mapping) {
 286                         switch (try_to_unmap(page)) {
 287                         case SWAP_FAIL:
 288                                 pte_chain_unlock(page);
 289                                 goto activate_locked;
 290                         case SWAP_AGAIN:
 291                                 pte_chain_unlock(page);
 292                                 goto keep_locked;
 293                         case SWAP_SUCCESS:
 294                                 ; /* try to free the page below */
 295                         }
 296                 }
 297                 pte_chain_unlock(page);
 298
 299                 /*
 300                  * If the page is dirty, only perform writeback if that write
 301                  * will be non-blocking.  To prevent this allocation from being
 302                  * stalled by pagecache activity.  But note that there may be
 303                  * stalls if we need to run get_block().  We could test
 304                  * PagePrivate for that.
 305                  *
 306                  * If this process is currently in generic_file_write() against
 307                  * this page's queue, we can perform writeback even if that
 308                  * will block.
 309                  *
 310                  * If the page is swapcache, write it back even if that would
 311                  * block, for some throttling. This happens by accident, because
 312                  * swap_backing_dev_info is bust: it doesn't reflect the
 313                  * congestion state of the swapdevs.  Easy to fix, if needed.
 314                  * See swapfile.c:page_queue_congested().
 315                  */
 316                 if (PageDirty(page)) {
 317                         if (!is_page_cache_freeable(page))
 318                                 goto keep_locked;
 319                         if (!mapping)
 320                                 goto keep_locked;
 321                         if (mapping->a_ops->writepage == NULL)
 322                                 goto activate_locked;
 323                         if (!may_enter_fs)
 324                                 goto keep_locked;
 325                         if (!may_write_to_queue(mapping->backing_dev_info))
 326                                 goto keep_locked;
 327                         spin_lock(&mapping->page_lock);
 328                         if (test_clear_page_dirty(page)) {
 329                                 int res;
 330                                 struct writeback_control wbc = {
 331                                         .sync_mode = WB_SYNC_NONE,
 332                                         .nr_to_write = SWAP_CLUSTER_MAX,
 333                                         .nonblocking = 1,
 334                                         .for_reclaim = 1,
 335                                 };
 336
 337                                 list_move(&page->list, &mapping->locked_pages);
 338                                 spin_unlock(&mapping->page_lock);
 339
 340                                 SetPageReclaim(page);
 341                                 res = mapping->a_ops->writepage(page, &wbc);
 342
 343                                 if (res == WRITEPAGE_ACTIVATE) {
 344                                         ClearPageReclaim(page);
 345                                         goto activate_locked;
 346                                 }
 347                                 if (!PageWriteback(page)) {
 348                                         /* synchronous write or broken a_ops? */
 349                                         ClearPageReclaim(page);
 350                                 }
 351                                 goto keep;
 352                         }
 353                         spin_unlock(&mapping->page_lock);
 354                 }
 355
 356                 /*
 357                  * If the page has buffers, try to free the buffer mappings
 358                  * associated with this page. If we succeed we try to free
 359                  * the page as well.
 360                  *
 361                  * We do this even if the page is PageDirty().
 362                  * try_to_release_page() does not perform I/O, but it is
 363                  * possible for a page to have PageDirty set, but it is actually
 364                  * clean (all its buffers are clean).  This happens if the
 365                  * buffers were written out directly, with submit_bh(). ext3
 366                  * will do this, as well as the blockdev mapping.
 367                  * try_to_release_page() will discover that cleanness and will
 368                  * drop the buffers and mark the page clean - it can be freed.
 369                  *
 370                  * Rarely, pages can have buffers and no ->mapping.  These are
 371                  * the pages which were not successfully invalidated in
 372                  * truncate_complete_page().  We try to drop those buffers here
 373                  * and if that worked, and the page is no longer mapped into
 374                  * process address space (page_count == 0) it can be freed.
 375                  * Otherwise, leave the page on the LRU so it is swappable.
 376                  */
 377                 if (PagePrivate(page)) {
 378                         if (!try_to_release_page(page, gfp_mask))
 379                                 goto activate_locked;
 380                         if (!mapping && page_count(page) == 1)
 381                                 goto free_it;
 382                 }
 383
 384                 if (!mapping)
 385                         goto keep_locked;       /* truncate got there first */
 386
 387                 spin_lock(&mapping->page_lock);
 388
 389                 /*
 390                  * The non-racy check for busy page.  It is critical to check
 391                  * PageDirty _after_ making sure that the page is freeable and
 392                  * not in use by anybody.       (pagecache + us == 2)
 393                  */
 394                 if (page_count(page) != 2 || PageDirty(page)) {
 395                         spin_unlock(&mapping->page_lock);
 396                         goto keep_locked;
 397                 }
 398
 399 #ifdef CONFIG_SWAP
 400                 if (PageSwapCache(page)) {
 401                         swp_entry_t swap = { .val = page->index };
 402                         __delete_from_swap_cache(page);
 403                         spin_unlock(&mapping->page_lock);
 404                         swap_free(swap);
 405                         __put_page(page);       /* The pagecache ref */
 406                         goto free_it;
 407                 }
 408 #endif /* CONFIG_SWAP */
 409
 410                 __remove_from_page_cache(page);
 411                 spin_unlock(&mapping->page_lock);
 412                 __put_page(page);
 413
 414 free_it:
 415                 unlock_page(page);
 416                 ret++;
 417                 if (!pagevec_add(&freed_pvec, page))
 418                         __pagevec_release_nonlru(&freed_pvec);
 419                 continue;
 420
 421 activate_locked:
 422                 SetPageActive(page);
 423                 pgactivate++;
 424 keep_locked:
 425                 unlock_page(page);
 426 keep:
 427                 list_add(&page->lru, &ret_pages);
 428                 BUG_ON(PageLRU(page));
 429         }
 430         list_splice(&ret_pages, page_list);
 431         if (pagevec_count(&freed_pvec))
 432                 __pagevec_release_nonlru(&freed_pvec);
 433         mod_page_state(pgsteal, ret);
 434         if (current_is_kswapd())
 435                 mod_page_state(kswapd_steal, ret);
 436         mod_page_state(pgactivate, pgactivate);
 437         return ret;
 438 }
 439
 440 /*
 441  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 442  * a batch of pages and working on them outside the lock.  Any pages which were
 443  * not freed will be added back to the LRU.
 444  *
 445  * shrink_cache() is passed the number of pages to try to free, and returns
 446  * the number of pages which were reclaimed.
 447  *
 448  * For pagecache intensive workloads, the first loop here is the hottest spot
 449  * in the kernel (apart from the copy_*_user functions).
 450  */
 451 static int
 452 shrink_cache(const int nr_pages, struct zone *zone,
 453                 unsigned int gfp_mask, int max_scan, int *nr_mapped)
 454 {
 455         LIST_HEAD(page_list);
 456         struct pagevec pvec;
 457         int nr_to_process;
 458         int ret = 0;
 459
 460         /*
 461          * Try to ensure that we free `nr_pages' pages in one pass of the loop.
 462          */
 463         nr_to_process = nr_pages;
 464         if (nr_to_process < SWAP_CLUSTER_MAX)
 465                 nr_to_process = SWAP_CLUSTER_MAX;
 466
 467         pagevec_init(&pvec, 1);
 468
 469         lru_add_drain();
 470         spin_lock_irq(&zone->lru_lock);
 471         while (max_scan > 0 && ret < nr_pages) {
 472                 struct page *page;
 473                 int nr_taken = 0;
 474                 int nr_scan = 0;
 475                 int nr_freed;
 476
 477                 while (nr_scan++ < nr_to_process &&
 478                                 !list_empty(&zone->inactive_list)) {
 479                         page = list_entry(zone->inactive_list.prev,
 480                                                 struct page, lru);
 481
 482                         prefetchw_prev_lru_page(page,
 483                                                 &zone->inactive_list, flags);
 484
 485                         if (!TestClearPageLRU(page))
 486                                 BUG();
 487                         list_del(&page->lru);
 488                         if (page_count(page) == 0) {
 489                                 /* It is currently in pagevec_release() */
 490                                 SetPageLRU(page);
 491                                 list_add(&page->lru, &zone->inactive_list);
 492                                 continue;
 493                         }
 494                         list_add(&page->lru, &page_list);
 495                         page_cache_get(page);
 496                         nr_taken++;
 497                 }
 498                 zone->nr_inactive -= nr_taken;
 499                 zone->pages_scanned += nr_taken;
 500                 spin_unlock_irq(&zone->lru_lock);
 501
 502                 if (nr_taken == 0)
 503                         goto done;
 504
 505                 max_scan -= nr_scan;
 506                 mod_page_state(pgscan, nr_scan);
 507                 nr_freed = shrink_list(&page_list, gfp_mask,
 508                                         &max_scan, nr_mapped);
 509                 ret += nr_freed;
 510                 if (nr_freed <= 0 && list_empty(&page_list))
 511                         goto done;
 512
 513                 spin_lock_irq(&zone->lru_lock);
 514                 /*
 515                  * Put back any unfreeable pages.
 516                  */
 517                 while (!list_empty(&page_list)) {
 518                         page = list_entry(page_list.prev, struct page, lru);
 519                         if (TestSetPageLRU(page))
 520                                 BUG();
 521                         list_del(&page->lru);
 522                         if (PageActive(page))
 523                                 add_page_to_active_list(zone, page);
 524                         else
 525                                 add_page_to_inactive_list(zone, page);
 526                         if (!pagevec_add(&pvec, page)) {
 527                                 spin_unlock_irq(&zone->lru_lock);
 528                                 __pagevec_release(&pvec);
 529                                 spin_lock_irq(&zone->lru_lock);
 530                         }
 531                 }
 532         }
 533         spin_unlock_irq(&zone->lru_lock);
 534 done:
 535         pagevec_release(&pvec);
 536         return ret;
 537 }
 538
 539 /*
 540  * This moves pages from the active list to the inactive list.
 541  *
 542  * We move them the other way if the page is referenced by one or more
 543  * processes, from rmap.
 544  *
 545  * If the pages are mostly unmapped, the processing is fast and it is
 546  * appropriate to hold zone->lru_lock across the whole operation.  But if
 547  * the pages are mapped, the processing is slow (page_referenced()) so we
 548  * should drop zone->lru_lock around each page.  It's impossible to balance
 549  * this, so instead we remove the pages from the LRU while processing them.
 550  * It is safe to rely on PG_active against the non-LRU pages in here because
 551  * nobody will play with that bit on a non-LRU page.
 552  *
 553  * The downside is that we have to touch page->count against each page.
 554  * But we had to alter page->flags anyway.
 555  */
 556 static void
 557 refill_inactive_zone(struct zone *zone, const int nr_pages_in,
 558                         struct page_state *ps, int priority)
 559 {
 560         int pgmoved;
 561         int pgdeactivate = 0;
 562         int nr_pages = nr_pages_in;
 563         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 564         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 565         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 566         struct page *page;
 567         struct pagevec pvec;
 568         int reclaim_mapped = 0;
 569         long mapped_ratio;
 570         long distress;
 571         long swap_tendency;
 572
 573         lru_add_drain();
 574         pgmoved = 0;
 575         spin_lock_irq(&zone->lru_lock);
 576         while (nr_pages && !list_empty(&zone->active_list)) {
 577                 page = list_entry(zone->active_list.prev, struct page, lru);
 578                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 579                 if (!TestClearPageLRU(page))
 580                         BUG();
 581                 list_del(&page->lru);
 582                 if (page_count(page) == 0) {
 583                         /* It is currently in pagevec_release() */
 584                         SetPageLRU(page);
 585                         list_add(&page->lru, &zone->active_list);
 586                 } else {
 587                         page_cache_get(page);
 588                         list_add(&page->lru, &l_hold);
 589                         pgmoved++;
 590                 }
 591                 nr_pages--;
 592         }
 593         zone->nr_active -= pgmoved;
 594         spin_unlock_irq(&zone->lru_lock);
 595
 596         /*
 597          * `distress' is a measure of how much trouble we're having reclaiming
 598          * pages.  0 -> no problems.  100 -> great trouble.
 599          */
 600         distress = 100 >> priority;
 601
 602         /*
 603          * The point of this algorithm is to decide when to start reclaiming
 604          * mapped memory instead of just pagecache.  Work out how much memory
 605          * is mapped.
 606          */
 607         mapped_ratio = (ps->nr_mapped * 100) / total_memory;
 608
 609         /*
 610          * Now decide how much we really want to unmap some pages.  The mapped
 611          * ratio is downgraded - just because there's a lot of mapped memory
 612          * doesn't necessarily mean that page reclaim isn't succeeding.
 613          *
 614          * The distress ratio is important - we don't want to start going oom.
 615          *
 616          * A 100% value of vm_swappiness overrides this algorithm altogether.
 617          */
 618         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 619
 620         /*
 621          * Now use this metric to decide whether to start moving mapped memory
 622          * onto the inactive list.
 623          */
 624         if (swap_tendency >= 100)
 625                 reclaim_mapped = 1;
 626
 627         while (!list_empty(&l_hold)) {
 628                 page = list_entry(l_hold.prev, struct page, lru);
 629                 list_del(&page->lru);
 630                 if (page_mapped(page)) {
 631                         pte_chain_lock(page);
 632                         if (page_mapped(page) && page_referenced(page)) {
 633                                 pte_chain_unlock(page);
 634                                 list_add(&page->lru, &l_active);
 635                                 continue;
 636                         }
 637                         pte_chain_unlock(page);
 638                         if (!reclaim_mapped) {
 639                                 list_add(&page->lru, &l_active);
 640                                 continue;
 641                         }
 642                 }
 643                 /*
 644                  * FIXME: need to consider page_count(page) here if/when we
 645                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 646                  */
 647                 if (total_swap_pages == 0 && !page->mapping &&
 648                                                 !PagePrivate(page)) {
 649                         list_add(&page->lru, &l_active);
 650                         continue;
 651                 }
 652                 list_add(&page->lru, &l_inactive);
 653         }
 654
 655         pagevec_init(&pvec, 1);
 656         pgmoved = 0;
 657         spin_lock_irq(&zone->lru_lock);
 658         while (!list_empty(&l_inactive)) {
 659                 page = list_entry(l_inactive.prev, struct page, lru);
 660                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 661                 if (TestSetPageLRU(page))
 662                         BUG();
 663                 if (!TestClearPageActive(page))
 664                         BUG();
 665                 list_move(&page->lru, &zone->inactive_list);
 666                 pgmoved++;
 667                 if (!pagevec_add(&pvec, page)) {
 668                         zone->nr_inactive += pgmoved;
 669                         spin_unlock_irq(&zone->lru_lock);
 670                         pgdeactivate += pgmoved;
 671                         pgmoved = 0;
 672                         if (buffer_heads_over_limit)
 673                                 pagevec_strip(&pvec);
 674                         __pagevec_release(&pvec);
 675                         spin_lock_irq(&zone->lru_lock);
 676                 }
 677         }
 678         zone->nr_inactive += pgmoved;
 679         pgdeactivate += pgmoved;
 680         if (buffer_heads_over_limit) {
 681                 spin_unlock_irq(&zone->lru_lock);
 682                 pagevec_strip(&pvec);
 683                 spin_lock_irq(&zone->lru_lock);
 684         }
 685
 686         pgmoved = 0;
 687         while (!list_empty(&l_active)) {
 688                 page = list_entry(l_active.prev, struct page, lru);
 689                 prefetchw_prev_lru_page(page, &l_active, flags);
 690                 if (TestSetPageLRU(page))
 691                         BUG();
 692                 BUG_ON(!PageActive(page));
 693                 list_move(&page->lru, &zone->active_list);
 694                 pgmoved++;
 695                 if (!pagevec_add(&pvec, page)) {
 696                         zone->nr_active += pgmoved;
 697                         pgmoved = 0;
 698                         spin_unlock_irq(&zone->lru_lock);
 699                         __pagevec_release(&pvec);
 700                         spin_lock_irq(&zone->lru_lock);
 701                 }
 702         }
 703         zone->nr_active += pgmoved;
 704         spin_unlock_irq(&zone->lru_lock);
 705         pagevec_release(&pvec);
 706
 707         mod_page_state(pgrefill, nr_pages_in - nr_pages);
 708         mod_page_state(pgdeactivate, pgdeactivate);
 709 }
 710
 711 /*
 712  * Try to reclaim `nr_pages' from this zone.  Returns the number of reclaimed
 713  * pages.  This is a basic per-zone page freer.  Used by both kswapd and
 714  * direct reclaim.
 715  */
 716 static int
 717 shrink_zone(struct zone *zone, int max_scan, unsigned int gfp_mask,
 718         const int nr_pages, int *nr_mapped, struct page_state *ps, int priority)
 719 {
 720         unsigned long ratio;
 721
 722         /*
 723          * Try to keep the active list 2/3 of the size of the cache.  And
 724          * make sure that refill_inactive is given a decent number of pages.
 725          *
 726          * The "ratio+1" here is important.  With pagecache-intensive workloads
 727          * the inactive list is huge, and `ratio' evaluates to zero all the
 728          * time.  Which pins the active list memory.  So we add one to `ratio'
 729          * just to make sure that the kernel will slowly sift through the
 730          * active list.
 731          */
 732         ratio = (unsigned long)nr_pages * zone->nr_active /
 733                                 ((zone->nr_inactive | 1) * 2);
 734         atomic_add(ratio+1, &zone->refill_counter);
 735         if (atomic_read(&zone->refill_counter) > SWAP_CLUSTER_MAX) {
 736                 int count;
 737
 738                 /*
 739                  * Don't try to bring down too many pages in one attempt.
 740                  * If this fails, the caller will increase `priority' and
 741                  * we'll try again, with an increased chance of reclaiming
 742                  * mapped memory.
 743                  */
 744                 count = atomic_read(&zone->refill_counter);
 745                 if (count > SWAP_CLUSTER_MAX * 4)
 746                         count = SWAP_CLUSTER_MAX * 4;
 747                 atomic_sub(count, &zone->refill_counter);
 748                 refill_inactive_zone(zone, count, ps, priority);
 749         }
 750         return shrink_cache(nr_pages, zone, gfp_mask,
 751                                 max_scan, nr_mapped);
 752 }
 753
 754 /*
 755  * This is the direct reclaim path, for page-allocating processes.  We only
 756  * try to reclaim pages from zones which will satisfy the caller's allocation
 757  * request.
 758  *
 759  * We reclaim from a zone even if that zone is over pages_high.  Because:
 760  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 761  *    allocation or
 762  * b) The zones may be over pages_high but they must go *over* pages_high to
 763  *    satisfy the `incremental min' zone defense algorithm.
 764  *
 765  * Returns the number of reclaimed pages.
 766  *
 767  * If a zone is deemed to be full of pinned pages then just give it a light
 768  * scan then give up on it.
 769  */
 770 static int
 771 shrink_caches(struct zone *classzone, int priority, int *total_scanned,
 772                 int gfp_mask, int nr_pages, struct page_state *ps)
 773 {
 774         struct zone *first_classzone;
 775         struct zone *zone;
 776         int ret = 0;
 777
 778         first_classzone = classzone->zone_pgdat->node_zones;
 779         for (zone = classzone; zone >= first_classzone; zone--) {
 780                 int to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX);
 781                 int nr_mapped = 0;
 782                 int max_scan;
 783
 784                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
 785                         continue;       /* Let kswapd poll it */
 786
 787                 /*
 788                  * If we cannot reclaim `nr_pages' pages by scanning twice
 789                  * that many pages then fall back to the next zone.
 790                  */
 791                 max_scan = zone->nr_inactive >> priority;
 792                 if (max_scan < to_reclaim * 2)
 793                         max_scan = to_reclaim * 2;
 794                 ret += shrink_zone(zone, max_scan, gfp_mask,
 795                                 to_reclaim, &nr_mapped, ps, priority);
 796                 *total_scanned += max_scan + nr_mapped;
 797                 if (ret >= nr_pages)
 798                         break;
 799         }
 800         return ret;
 801 }
 802
 803 /*
 804  * This is the main entry point to direct page reclaim.
 805  *
 806  * If a full scan of the inactive list fails to free enough memory then we
 807  * are "out of memory" and something needs to be killed.
 808  *
 809  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 810  * high - the zone may be full of dirty or under-writeback pages, which this
 811  * caller can't do much about.  So for !__GFP_FS callers, we just perform a
 812  * small LRU walk and if that didn't work out, fail the allocation back to the
 813  * caller.  GFP_NOFS allocators need to know how to deal with it.  Kicking
 814  * bdflush, waiting and retrying will work.
 815  *
 816  * This is a fairly lame algorithm - it can result in excessive CPU burning and
 817  * excessive rotation of the inactive list, which is _supposed_ to be an LRU,
 818  * yes?
 819  */
 820 int try_to_free_pages(struct zone *cz,
 821                 unsigned int gfp_mask, unsigned int order)
 822 {
 823         int priority;
 824         int ret = 0;
 825         const int nr_pages = SWAP_CLUSTER_MAX;
 826         int nr_reclaimed = 0;
 827         struct reclaim_state *reclaim_state = current->reclaim_state;
 828
 829         inc_page_state(allocstall);
 830
 831         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 832                 int total_scanned = 0;
 833                 struct page_state ps;
 834
 835                 get_page_state(&ps);
 836                 nr_reclaimed += shrink_caches(cz, priority, &total_scanned,
 837                                                 gfp_mask, nr_pages, &ps);
 838                 if (nr_reclaimed >= nr_pages) {
 839                         ret = 1;
 840                         goto out;
 841                 }
 842                 if (!(gfp_mask & __GFP_FS))
 843                         break;          /* Let the caller handle it */
 844                 /*
 845                  * Try to write back as many pages as we just scanned.  Not
 846                  * sure if that makes sense, but it's an attempt to avoid
 847                  * creating IO storms unnecessarily
 848                  */
 849                 wakeup_bdflush(total_scanned);
 850
 851                 /* Take a nap, wait for some writeback to complete */
 852                 blk_congestion_wait(WRITE, HZ/10);
 853                 if (cz - cz->zone_pgdat->node_zones < ZONE_HIGHMEM) {
 854                         shrink_slab(total_scanned, gfp_mask);
 855                         if (reclaim_state) {
 856                                 nr_reclaimed += reclaim_state->reclaimed_slab;
 857                                 reclaim_state->reclaimed_slab = 0;
 858                         }
 859                 }
 860         }
 861         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 862                 out_of_memory();
 863 out:
 864         return ret;
 865 }
 866
 867 /*
 868  * For kswapd, balance_pgdat() will work across all this node's zones until
 869  * they are all at pages_high.
 870  *
 871  * If `nr_pages' is non-zero then it is the number of pages which are to be
 872  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 873  * special.
 874  *
 875  * Returns the number of pages which were actually freed.
 876  *
 877  * There is special handling here for zones which are full of pinned pages.
 878  * This can happen if the pages are all mlocked, or if they are all used by
 879  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 880  * What we do is to detect the case where all pages in the zone have been
 881  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 882  * dead and from now on, only perform a short scan.  Basically we're polling
 883  * the zone for when the problem goes away.
 884  */
 885 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, struct page_state *ps)
 886 {
 887         int to_free = nr_pages;
 888         int priority;
 889         int i;
 890         struct reclaim_state *reclaim_state = current->reclaim_state;
 891
 892         inc_page_state(pageoutrun);
 893
 894         for (priority = DEF_PRIORITY; priority; priority--) {
 895                 int all_zones_ok = 1;
 896
 897                 for (i = 0; i < pgdat->nr_zones; i++) {
 898                         struct zone *zone = pgdat->node_zones + i;
 899                         int nr_mapped = 0;
 900                         int max_scan;
 901                         int to_reclaim;
 902
 903                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
 904                                 continue;
 905
 906                         if (nr_pages && to_free > 0) {  /* Software suspend */
 907                                 to_reclaim = min(to_free, SWAP_CLUSTER_MAX*8);
 908                         } else {                        /* Zone balancing */
 909                                 to_reclaim = zone->pages_high-zone->free_pages;
 910                                 if (to_reclaim <= 0)
 911                                         continue;
 912                         }
 913                         all_zones_ok = 0;
 914                         max_scan = zone->nr_inactive >> priority;
 915                         if (max_scan < to_reclaim * 2)
 916                                 max_scan = to_reclaim * 2;
 917                         if (max_scan < SWAP_CLUSTER_MAX)
 918                                 max_scan = SWAP_CLUSTER_MAX;
 919                         to_free -= shrink_zone(zone, max_scan, GFP_KERNEL,
 920                                         to_reclaim, &nr_mapped, ps, priority);
 921                         if (i < ZONE_HIGHMEM) {
 922                                 reclaim_state->reclaimed_slab = 0;
 923                                 shrink_slab(max_scan + nr_mapped, GFP_KERNEL);
 924                                 to_free += reclaim_state->reclaimed_slab;
 925                         }
 926                         if (zone->all_unreclaimable)
 927                                 continue;
 928                         if (zone->pages_scanned > zone->present_pages * 2)
 929                                 zone->all_unreclaimable = 1;
 930                 }
 931                 if (all_zones_ok)
 932                         break;
 933                 blk_congestion_wait(WRITE, HZ/10);
 934         }
 935         return nr_pages - to_free;
 936 }
 937
 938 /*
 939  * The background pageout daemon, started as a kernel thread
 940  * from the init process.
 941  *
 942  * This basically trickles out pages so that we have _some_
 943  * free memory available even if there is no other activity
 944  * that frees anything up. This is needed for things like routing
 945  * etc, where we otherwise might have all activity going on in
 946  * asynchronous contexts that cannot page things out.
 947  *
 948  * If there are applications that are active memory-allocators
 949  * (most normal use), this basically shouldn't matter.
 950  */
 951 int kswapd(void *p)
 952 {
 953         pg_data_t *pgdat = (pg_data_t*)p;
 954         struct task_struct *tsk = current;
 955         DEFINE_WAIT(wait);
 956         struct reclaim_state reclaim_state = {
 957                 .reclaimed_slab = 0,
 958         };
 959         unsigned long cpumask;
 960
 961         daemonize("kswapd%d", pgdat->node_id);
 962         cpumask = node_to_cpumask(pgdat->node_id);
 963         if (cpumask)
 964                 set_cpus_allowed(tsk, cpumask);
 965         current->reclaim_state = &reclaim_state;
 966
 967         /*
 968          * Tell the memory management that we're a "memory allocator",
 969          * and that if we need more memory we should get access to it
 970          * regardless (see "__alloc_pages()"). "kswapd" should
 971          * never get caught in the normal page freeing logic.
 972          *
 973          * (Kswapd normally doesn't need memory anyway, but sometimes
 974          * you need a small amount of memory in order to be able to
 975          * page out something else, and this flag essentially protects
 976          * us from recursively trying to free more memory as we're
 977          * trying to free the first piece of memory in the first place).
 978          */
 979         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
 980
 981         for ( ; ; ) {
 982                 struct page_state ps;
 983
 984                 if (current->flags & PF_FREEZE)
 985                         refrigerator(PF_IOTHREAD);
 986                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
 987                 schedule();
 988                 finish_wait(&pgdat->kswapd_wait, &wait);
 989                 get_page_state(&ps);
 990                 balance_pgdat(pgdat, 0, &ps);
 991         }
 992 }
 993
 994 /*
 995  * A zone is low on free memory, so wake its kswapd task to service it.
 996  */
 997 void wakeup_kswapd(struct zone *zone)
 998 {
 999         if (zone->free_pages > zone->pages_low)
1000                 return;
1001         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1002                 return;
1003         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1004 }
1005
1006 #ifdef CONFIG_SOFTWARE_SUSPEND
1007 /*
1008  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1009  * pages.
1010  */
1011 int shrink_all_memory(int nr_pages)
1012 {
1013         pg_data_t *pgdat;
1014         int nr_to_free = nr_pages;
1015         int ret = 0;
1016         struct reclaim_state reclaim_state = {
1017                 .reclaimed_slab = 0,
1018         };
1019
1020         current->reclaim_state = &reclaim_state;
1021         for_each_pgdat(pgdat) {
1022                 int freed;
1023                 struct page_state ps;
1024
1025                 get_page_state(&ps);
1026                 freed = balance_pgdat(pgdat, nr_to_free, &ps);
1027                 ret += freed;
1028                 nr_to_free -= freed;
1029                 if (nr_to_free <= 0)
1030                         break;
1031         }
1032         current->reclaim_state = NULL;
1033         return ret;
1034 }
1035 #endif
1036
1037 static int __init kswapd_init(void)
1038 {
1039         pg_data_t *pgdat;
1040         swap_setup();
1041         for_each_pgdat(pgdat)
1042                 kernel_thread(kswapd, pgdat, CLONE_KERNEL);
1043         total_memory = nr_free_pagecache_pages();
1044         return 0;
1045 }
1046
1047 module_init(kswapd_init)