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/tlbflush.h>
  37 #include <asm/div64.h>
  38
  39 #include <linux/swapops.h>
  40
  41 /* possible outcome of pageout() */
  42 typedef enum {
  43         /* failed to write page out, page is locked */
  44         PAGE_KEEP,
  45         /* move page to the active list, page is locked */
  46         PAGE_ACTIVATE,
  47         /* page has been sent to the disk successfully, page is unlocked */
  48         PAGE_SUCCESS,
  49         /* page is clean and locked */
  50         PAGE_CLEAN,
  51 } pageout_t;
  52
  53 struct scan_control {
  54         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  55         unsigned long nr_to_scan;
  56
  57         /* Incremented by the number of inactive pages that were scanned */
  58         unsigned long nr_scanned;
  59
  60         /* Incremented by the number of pages reclaimed */
  61         unsigned long nr_reclaimed;
  62
  63         unsigned long nr_mapped;        /* From page_state */
  64
  65         /* How many pages shrink_cache() should reclaim */
  66         int nr_to_reclaim;
  67
  68         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  69         unsigned int priority;
  70
  71         /* This context's GFP mask */
  72         unsigned int gfp_mask;
  73
  74         int may_writepage;
  75 };
  76
  77 /*
  78  * The list of shrinker callbacks used by to apply pressure to
  79  * ageable caches.
  80  */
  81 struct shrinker {
  82         shrinker_t              shrinker;
  83         struct list_head        list;
  84         int                     seeks;  /* seeks to recreate an obj */
  85         long                    nr;     /* objs pending delete */
  86 };
  87
  88
  89
  90 void try_to_clip_inodes(void);
  91
  92
  93 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  94
  95 #ifdef ARCH_HAS_PREFETCH
  96 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  97         do {                                                            \
  98                 if ((_page)->lru.prev != _base) {                       \
  99                         struct page *prev;                              \
 100                                                                         \
 101                         prev = lru_to_page(&(_page->lru));              \
 102                         prefetch(&prev->_field);                        \
 103                 }                                                       \
 104         } while (0)
 105 #else
 106 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 107 #endif
 108
 109 #ifdef ARCH_HAS_PREFETCHW
 110 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 111         do {                                                            \
 112                 if ((_page)->lru.prev != _base) {                       \
 113                         struct page *prev;                              \
 114                                                                         \
 115                         prev = lru_to_page(&(_page->lru));              \
 116                         prefetchw(&prev->_field);                       \
 117                 }                                                       \
 118         } while (0)
 119 #else
 120 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 121 #endif
 122
 123 /*
 124  * From 0 .. 100.  Higher means more swappy.
 125  */
 126 int vm_swappiness = 60;
 127 static long total_memory;
 128
 129 static LIST_HEAD(shrinker_list);
 130 static DECLARE_MUTEX(shrinker_sem);
 131
 132 /*
 133  * Add a shrinker callback to be called from the vm
 134  */
 135 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 136 {
 137         struct shrinker *shrinker;
 138
 139         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 140         if (shrinker) {
 141                 shrinker->shrinker = theshrinker;
 142                 shrinker->seeks = seeks;
 143                 shrinker->nr = 0;
 144                 down(&shrinker_sem);
 145                 list_add(&shrinker->list, &shrinker_list);
 146                 up(&shrinker_sem);
 147         }
 148         return shrinker;
 149 }
 150 EXPORT_SYMBOL(set_shrinker);
 151
 152 /*
 153  * Remove one
 154  */
 155 void remove_shrinker(struct shrinker *shrinker)
 156 {
 157         down(&shrinker_sem);
 158         list_del(&shrinker->list);
 159         up(&shrinker_sem);
 160         kfree(shrinker);
 161 }
 162 EXPORT_SYMBOL(remove_shrinker);
 163
 164 #define SHRINK_BATCH 128
 165 /*
 166  * Call the shrink functions to age shrinkable caches
 167  *
 168  * Here we assume it costs one seek to replace a lru page and that it also
 169  * takes a seek to recreate a cache object.  With this in mind we age equal
 170  * percentages of the lru and ageable caches.  This should balance the seeks
 171  * generated by these structures.
 172  *
 173  * If the vm encounted mapped pages on the LRU it increase the pressure on
 174  * slab to avoid swapping.
 175  *
 176  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 177  *
 178  * `lru_pages' represents the number of on-LRU pages in all the zones which
 179  * are eligible for the caller's allocation attempt.  It is used for balancing
 180  * slab reclaim versus page reclaim.
 181  */
 182 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 183                         unsigned long lru_pages)
 184 {
 185         struct shrinker *shrinker;
 186
 187         if (down_trylock(&shrinker_sem))
 188                 return 0;
 189
 190         list_for_each_entry(shrinker, &shrinker_list, list) {
 191                 unsigned long long delta;
 192
 193                 delta = (4 * scanned) / shrinker->seeks;
 194                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 195                 do_div(delta, lru_pages + 1);
 196                 shrinker->nr += delta;
 197                 if (shrinker->nr < 0)
 198                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 199
 200                 if (shrinker->nr <= SHRINK_BATCH)
 201                         continue;
 202                 while (shrinker->nr) {
 203                         long this_scan = shrinker->nr;
 204                         int shrink_ret;
 205
 206                         if (this_scan > 128)
 207                                 this_scan = 128;
 208                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 209                         mod_page_state(slabs_scanned, this_scan);
 210                         shrinker->nr -= this_scan;
 211                         if (shrink_ret == -1)
 212                                 break;
 213                         cond_resched();
 214                 }
 215         }
 216         up(&shrinker_sem);
 217         return 0;
 218 }
 219
 220 /* Must be called with page's rmap lock held. */
 221 static inline int page_mapping_inuse(struct page *page)
 222 {
 223         struct address_space *mapping;
 224
 225         /* Page is in somebody's page tables. */
 226         if (page_mapped(page))
 227                 return 1;
 228
 229         /* Be more reluctant to reclaim swapcache than pagecache */
 230         if (PageSwapCache(page))
 231                 return 1;
 232
 233         mapping = page_mapping(page);
 234         if (!mapping)
 235                 return 0;
 236
 237         /* File is mmap'd by somebody? */
 238         return mapping_mapped(mapping);
 239 }
 240
 241 static inline int is_page_cache_freeable(struct page *page)
 242 {
 243         return page_count(page) - !!PagePrivate(page) == 2;
 244 }
 245
 246 static int may_write_to_queue(struct backing_dev_info *bdi)
 247 {
 248         if (current_is_kswapd())
 249                 return 1;
 250         if (current_is_pdflush())       /* This is unlikely, but why not... */
 251                 return 1;
 252         if (!bdi_write_congested(bdi))
 253                 return 1;
 254         if (bdi == current->backing_dev_info)
 255                 return 1;
 256         return 0;
 257 }
 258
 259 /*
 260  * We detected a synchronous write error writing a page out.  Probably
 261  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 262  * fsync(), msync() or close().
 263  *
 264  * The tricky part is that after writepage we cannot touch the mapping: nothing
 265  * prevents it from being freed up.  But we have a ref on the page and once
 266  * that page is locked, the mapping is pinned.
 267  *
 268  * We're allowed to run sleeping lock_page() here because we know the caller has
 269  * __GFP_FS.
 270  */
 271 static void handle_write_error(struct address_space *mapping,
 272                                 struct page *page, int error)
 273 {
 274         lock_page(page);
 275         if (page_mapping(page) == mapping) {
 276                 if (error == -ENOSPC)
 277                         set_bit(AS_ENOSPC, &mapping->flags);
 278                 else
 279                         set_bit(AS_EIO, &mapping->flags);
 280         }
 281         unlock_page(page);
 282 }
 283
 284 /*
 285  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 286  */
 287 static pageout_t pageout(struct page *page, struct address_space *mapping)
 288 {
 289         /*
 290          * If the page is dirty, only perform writeback if that write
 291          * will be non-blocking.  To prevent this allocation from being
 292          * stalled by pagecache activity.  But note that there may be
 293          * stalls if we need to run get_block().  We could test
 294          * PagePrivate for that.
 295          *
 296          * If this process is currently in generic_file_write() against
 297          * this page's queue, we can perform writeback even if that
 298          * will block.
 299          *
 300          * If the page is swapcache, write it back even if that would
 301          * block, for some throttling. This happens by accident, because
 302          * swap_backing_dev_info is bust: it doesn't reflect the
 303          * congestion state of the swapdevs.  Easy to fix, if needed.
 304          * See swapfile.c:page_queue_congested().
 305          */
 306         if (!is_page_cache_freeable(page))
 307                 return PAGE_KEEP;
 308         if (!mapping)
 309                 return PAGE_KEEP;
 310         if (mapping->a_ops->writepage == NULL)
 311                 return PAGE_ACTIVATE;
 312         if (!may_write_to_queue(mapping->backing_dev_info))
 313                 return PAGE_KEEP;
 314
 315         if (clear_page_dirty_for_io(page)) {
 316                 int res;
 317                 struct writeback_control wbc = {
 318                         .sync_mode = WB_SYNC_NONE,
 319                         .nr_to_write = SWAP_CLUSTER_MAX,
 320                         .nonblocking = 1,
 321                         .for_reclaim = 1,
 322                 };
 323
 324                 SetPageReclaim(page);
 325                 res = mapping->a_ops->writepage(page, &wbc);
 326                 if (res < 0)
 327                         handle_write_error(mapping, page, res);
 328                 if (res == WRITEPAGE_ACTIVATE) {
 329                         ClearPageReclaim(page);
 330                         return PAGE_ACTIVATE;
 331                 }
 332                 if (!PageWriteback(page)) {
 333                         /* synchronous write or broken a_ops? */
 334                         ClearPageReclaim(page);
 335                 }
 336
 337                 return PAGE_SUCCESS;
 338         }
 339
 340         return PAGE_CLEAN;
 341 }
 342
 343 /*
 344  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 345  */
 346 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 347 {
 348         LIST_HEAD(ret_pages);
 349         struct pagevec freed_pvec;
 350         int pgactivate = 0;
 351         int reclaimed = 0;
 352
 353         cond_resched();
 354
 355         pagevec_init(&freed_pvec, 1);
 356         while (!list_empty(page_list)) {
 357                 struct address_space *mapping;
 358                 struct page *page;
 359                 int may_enter_fs;
 360                 int referenced;
 361
 362                 cond_resched();
 363
 364                 page = lru_to_page(page_list);
 365                 list_del(&page->lru);
 366
 367                 if (TestSetPageLocked(page))
 368                         goto keep;
 369
 370                 BUG_ON(PageActive(page));
 371
 372                 if (PageWriteback(page))
 373                         goto keep_locked;
 374
 375                 sc->nr_scanned++;
 376                 /* Double the slab pressure for mapped and swapcache pages */
 377                 if (page_mapped(page) || PageSwapCache(page))
 378                         sc->nr_scanned++;
 379
 380                 page_map_lock(page);
 381                 referenced = page_referenced(page);
 382                 if (referenced && page_mapping_inuse(page)) {
 383                         /* In active use or really unfreeable.  Activate it. */
 384                         page_map_unlock(page);
 385                         goto activate_locked;
 386                 }
 387
 388 #ifdef CONFIG_SWAP
 389                 /*
 390                  * Anonymous process memory has backing store?
 391                  * Try to allocate it some swap space here.
 392                  *
 393                  * XXX: implement swap clustering ?
 394                  */
 395                 if (PageAnon(page) && !PageSwapCache(page)) {
 396                         page_map_unlock(page);
 397                         if (!add_to_swap(page))
 398                                 goto activate_locked;
 399                         page_map_lock(page);
 400                 }
 401 #endif /* CONFIG_SWAP */
 402
 403                 mapping = page_mapping(page);
 404                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 405                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 406
 407                 /*
 408                  * The page is mapped into the page tables of one or more
 409                  * processes. Try to unmap it here.
 410                  */
 411                 if (page_mapped(page) && mapping) {
 412                         switch (try_to_unmap(page)) {
 413                         case SWAP_FAIL:
 414                                 page_map_unlock(page);
 415                                 goto activate_locked;
 416                         case SWAP_AGAIN:
 417                                 page_map_unlock(page);
 418                                 goto keep_locked;
 419                         case SWAP_SUCCESS:
 420                                 ; /* try to free the page below */
 421                         }
 422                 }
 423                 page_map_unlock(page);
 424
 425                 if (PageDirty(page)) {
 426                         if (referenced)
 427                                 goto keep_locked;
 428                         if (!may_enter_fs)
 429                                 goto keep_locked;
 430                         if (laptop_mode && !sc->may_writepage)
 431                                 goto keep_locked;
 432
 433                         /* Page is dirty, try to write it out here */
 434                         switch(pageout(page, mapping)) {
 435                         case PAGE_KEEP:
 436                                 goto keep_locked;
 437                         case PAGE_ACTIVATE:
 438                                 goto activate_locked;
 439                         case PAGE_SUCCESS:
 440                                 if (PageWriteback(page) || PageDirty(page))
 441                                         goto keep;
 442                                 /*
 443                                  * A synchronous write - probably a ramdisk.  Go
 444                                  * ahead and try to reclaim the page.
 445                                  */
 446                                 if (TestSetPageLocked(page))
 447                                         goto keep;
 448                                 if (PageDirty(page) || PageWriteback(page))
 449                                         goto keep_locked;
 450                                 mapping = page_mapping(page);
 451                         case PAGE_CLEAN:
 452                                 ; /* try to free the page below */
 453                         }
 454                 }
 455
 456                 /*
 457                  * If the page has buffers, try to free the buffer mappings
 458                  * associated with this page. If we succeed we try to free
 459                  * the page as well.
 460                  *
 461                  * We do this even if the page is PageDirty().
 462                  * try_to_release_page() does not perform I/O, but it is
 463                  * possible for a page to have PageDirty set, but it is actually
 464                  * clean (all its buffers are clean).  This happens if the
 465                  * buffers were written out directly, with submit_bh(). ext3
 466                  * will do this, as well as the blockdev mapping.
 467                  * try_to_release_page() will discover that cleanness and will
 468                  * drop the buffers and mark the page clean - it can be freed.
 469                  *
 470                  * Rarely, pages can have buffers and no ->mapping.  These are
 471                  * the pages which were not successfully invalidated in
 472                  * truncate_complete_page().  We try to drop those buffers here
 473                  * and if that worked, and the page is no longer mapped into
 474                  * process address space (page_count == 1) it can be freed.
 475                  * Otherwise, leave the page on the LRU so it is swappable.
 476                  */
 477                 if (PagePrivate(page)) {
 478                         if (!try_to_release_page(page, sc->gfp_mask))
 479                                 goto activate_locked;
 480                         if (!mapping && page_count(page) == 1)
 481                                 goto free_it;
 482                 }
 483
 484                 if (!mapping)
 485                         goto keep_locked;       /* truncate got there first */
 486
 487                 spin_lock_irq(&mapping->tree_lock);
 488
 489                 /*
 490                  * The non-racy check for busy page.  It is critical to check
 491                  * PageDirty _after_ making sure that the page is freeable and
 492                  * not in use by anybody.       (pagecache + us == 2)
 493                  */
 494                 if (page_count(page) != 2 || PageDirty(page)) {
 495                         spin_unlock_irq(&mapping->tree_lock);
 496                         goto keep_locked;
 497                 }
 498
 499 #ifdef CONFIG_SWAP
 500                 if (PageSwapCache(page)) {
 501                         swp_entry_t swap = { .val = page->private };
 502                         __delete_from_swap_cache(page);
 503                         spin_unlock_irq(&mapping->tree_lock);
 504                         swap_free(swap);
 505                         __put_page(page);       /* The pagecache ref */
 506                         goto free_it;
 507                 }
 508 #endif /* CONFIG_SWAP */
 509
 510                 __remove_from_page_cache(page);
 511                 spin_unlock_irq(&mapping->tree_lock);
 512                 __put_page(page);
 513
 514 free_it:
 515                 unlock_page(page);
 516                 reclaimed++;
 517                 if (!pagevec_add(&freed_pvec, page))
 518                         __pagevec_release_nonlru(&freed_pvec);
 519                 continue;
 520
 521 activate_locked:
 522                 SetPageActive(page);
 523                 pgactivate++;
 524 keep_locked:
 525                 unlock_page(page);
 526 keep:
 527                 list_add(&page->lru, &ret_pages);
 528                 BUG_ON(PageLRU(page));
 529         }
 530         list_splice(&ret_pages, page_list);
 531         if (pagevec_count(&freed_pvec))
 532                 __pagevec_release_nonlru(&freed_pvec);
 533         mod_page_state(pgactivate, pgactivate);
 534         sc->nr_reclaimed += reclaimed;
 535         return reclaimed;
 536 }
 537
 538 /*
 539  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 540  * a batch of pages and working on them outside the lock.  Any pages which were
 541  * not freed will be added back to the LRU.
 542  *
 543  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 544  *
 545  * For pagecache intensive workloads, the first loop here is the hottest spot
 546  * in the kernel (apart from the copy_*_user functions).
 547  */
 548 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 549 {
 550         LIST_HEAD(page_list);
 551         struct pagevec pvec;
 552         int max_scan = sc->nr_to_scan;
 553
 554         pagevec_init(&pvec, 1);
 555
 556         lru_add_drain();
 557         spin_lock_irq(&zone->lru_lock);
 558         while (max_scan > 0) {
 559                 struct page *page;
 560                 int nr_taken = 0;
 561                 int nr_scan = 0;
 562                 int nr_freed;
 563
 564                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 565                                 !list_empty(&zone->inactive_list)) {
 566                         page = lru_to_page(&zone->inactive_list);
 567
 568                         prefetchw_prev_lru_page(page,
 569                                                 &zone->inactive_list, flags);
 570
 571                         if (!TestClearPageLRU(page))
 572                                 BUG();
 573                         list_del(&page->lru);
 574                         if (get_page_testone(page)) {
 575                                 /*
 576                                  * It is being freed elsewhere
 577                                  */
 578                                 __put_page(page);
 579                                 SetPageLRU(page);
 580                                 list_add(&page->lru, &zone->inactive_list);
 581                                 continue;
 582                         }
 583                         list_add(&page->lru, &page_list);
 584                         nr_taken++;
 585                 }
 586                 zone->nr_inactive -= nr_taken;
 587                 zone->pages_scanned += nr_taken;
 588                 spin_unlock_irq(&zone->lru_lock);
 589
 590                 if (nr_taken == 0)
 591                         goto done;
 592
 593                 max_scan -= nr_scan;
 594                 if (current_is_kswapd())
 595                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 596                 else
 597                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 598                 nr_freed = shrink_list(&page_list, sc);
 599                 if (current_is_kswapd())
 600                         mod_page_state(kswapd_steal, nr_freed);
 601                 mod_page_state_zone(zone, pgsteal, nr_freed);
 602                 sc->nr_to_reclaim -= nr_freed;
 603
 604                 spin_lock_irq(&zone->lru_lock);
 605                 /*
 606                  * Put back any unfreeable pages.
 607                  */
 608                 while (!list_empty(&page_list)) {
 609                         page = lru_to_page(&page_list);
 610                         if (TestSetPageLRU(page))
 611                                 BUG();
 612                         list_del(&page->lru);
 613                         if (PageActive(page))
 614                                 add_page_to_active_list(zone, page);
 615                         else
 616                                 add_page_to_inactive_list(zone, page);
 617                         if (!pagevec_add(&pvec, page)) {
 618                                 spin_unlock_irq(&zone->lru_lock);
 619                                 __pagevec_release(&pvec);
 620                                 spin_lock_irq(&zone->lru_lock);
 621                         }
 622                 }
 623         }
 624         spin_unlock_irq(&zone->lru_lock);
 625 done:
 626         pagevec_release(&pvec);
 627 }
 628
 629 /*
 630  * This moves pages from the active list to the inactive list.
 631  *
 632  * We move them the other way if the page is referenced by one or more
 633  * processes, from rmap.
 634  *
 635  * If the pages are mostly unmapped, the processing is fast and it is
 636  * appropriate to hold zone->lru_lock across the whole operation.  But if
 637  * the pages are mapped, the processing is slow (page_referenced()) so we
 638  * should drop zone->lru_lock around each page.  It's impossible to balance
 639  * this, so instead we remove the pages from the LRU while processing them.
 640  * It is safe to rely on PG_active against the non-LRU pages in here because
 641  * nobody will play with that bit on a non-LRU page.
 642  *
 643  * The downside is that we have to touch page->_count against each page.
 644  * But we had to alter page->flags anyway.
 645  */
 646 static void
 647 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 648 {
 649         int pgmoved;
 650         int pgdeactivate = 0;
 651         int pgscanned = 0;
 652         int nr_pages = sc->nr_to_scan;
 653         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 654         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 655         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 656         struct page *page;
 657         struct pagevec pvec;
 658         int reclaim_mapped = 0;
 659         long mapped_ratio;
 660         long distress;
 661         long swap_tendency;
 662
 663         lru_add_drain();
 664         pgmoved = 0;
 665         spin_lock_irq(&zone->lru_lock);
 666         while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
 667                 page = lru_to_page(&zone->active_list);
 668                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 669                 if (!TestClearPageLRU(page))
 670                         BUG();
 671                 list_del(&page->lru);
 672                 if (get_page_testone(page)) {
 673                         /*
 674                          * It was already free!  release_pages() or put_page()
 675                          * are about to remove it from the LRU and free it. So
 676                          * put the refcount back and put the page back on the
 677                          * LRU
 678                          */
 679                         __put_page(page);
 680                         SetPageLRU(page);
 681                         list_add(&page->lru, &zone->active_list);
 682                 } else {
 683                         list_add(&page->lru, &l_hold);
 684                         pgmoved++;
 685                 }
 686                 pgscanned++;
 687         }
 688         zone->nr_active -= pgmoved;
 689         spin_unlock_irq(&zone->lru_lock);
 690
 691         /*
 692          * `distress' is a measure of how much trouble we're having reclaiming
 693          * pages.  0 -> no problems.  100 -> great trouble.
 694          */
 695         distress = 100 >> zone->prev_priority;
 696
 697         /*
 698          * The point of this algorithm is to decide when to start reclaiming
 699          * mapped memory instead of just pagecache.  Work out how much memory
 700          * is mapped.
 701          */
 702         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 703
 704         /*
 705          * Now decide how much we really want to unmap some pages.  The mapped
 706          * ratio is downgraded - just because there's a lot of mapped memory
 707          * doesn't necessarily mean that page reclaim isn't succeeding.
 708          *
 709          * The distress ratio is important - we don't want to start going oom.
 710          *
 711          * A 100% value of vm_swappiness overrides this algorithm altogether.
 712          */
 713         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 714
 715         /*
 716          * Now use this metric to decide whether to start moving mapped memory
 717          * onto the inactive list.
 718          */
 719         if (swap_tendency >= 100)
 720                 reclaim_mapped = 1;
 721
 722         while (!list_empty(&l_hold)) {
 723                 cond_resched();
 724                 page = lru_to_page(&l_hold);
 725                 list_del(&page->lru);
 726                 if (page_mapped(page)) {
 727                         if (!reclaim_mapped) {
 728                                 list_add(&page->lru, &l_active);
 729                                 continue;
 730                         }
 731                         page_map_lock(page);
 732                         if (page_referenced(page)) {
 733                                 page_map_unlock(page);
 734                                 list_add(&page->lru, &l_active);
 735                                 continue;
 736                         }
 737                         page_map_unlock(page);
 738                 }
 739                 /*
 740                  * FIXME: need to consider page_count(page) here if/when we
 741                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 742                  */
 743                 if (total_swap_pages == 0 && PageAnon(page)) {
 744                         list_add(&page->lru, &l_active);
 745                         continue;
 746                 }
 747                 list_add(&page->lru, &l_inactive);
 748         }
 749
 750         pagevec_init(&pvec, 1);
 751         pgmoved = 0;
 752         spin_lock_irq(&zone->lru_lock);
 753         while (!list_empty(&l_inactive)) {
 754                 page = lru_to_page(&l_inactive);
 755                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 756                 if (TestSetPageLRU(page))
 757                         BUG();
 758                 if (!TestClearPageActive(page))
 759                         BUG();
 760                 list_move(&page->lru, &zone->inactive_list);
 761                 pgmoved++;
 762                 if (!pagevec_add(&pvec, page)) {
 763                         zone->nr_inactive += pgmoved;
 764                         spin_unlock_irq(&zone->lru_lock);
 765                         pgdeactivate += pgmoved;
 766                         pgmoved = 0;
 767                         if (buffer_heads_over_limit)
 768                                 pagevec_strip(&pvec);
 769                         __pagevec_release(&pvec);
 770                         spin_lock_irq(&zone->lru_lock);
 771                 }
 772         }
 773         zone->nr_inactive += pgmoved;
 774         pgdeactivate += pgmoved;
 775         if (buffer_heads_over_limit) {
 776                 spin_unlock_irq(&zone->lru_lock);
 777                 pagevec_strip(&pvec);
 778                 spin_lock_irq(&zone->lru_lock);
 779         }
 780
 781         pgmoved = 0;
 782         while (!list_empty(&l_active)) {
 783                 page = lru_to_page(&l_active);
 784                 prefetchw_prev_lru_page(page, &l_active, flags);
 785                 if (TestSetPageLRU(page))
 786                         BUG();
 787                 BUG_ON(!PageActive(page));
 788                 list_move(&page->lru, &zone->active_list);
 789                 pgmoved++;
 790                 if (!pagevec_add(&pvec, page)) {
 791                         zone->nr_active += pgmoved;
 792                         pgmoved = 0;
 793                         spin_unlock_irq(&zone->lru_lock);
 794                         __pagevec_release(&pvec);
 795                         spin_lock_irq(&zone->lru_lock);
 796                 }
 797         }
 798         zone->nr_active += pgmoved;
 799         spin_unlock_irq(&zone->lru_lock);
 800         pagevec_release(&pvec);
 801
 802         mod_page_state_zone(zone, pgrefill, pgscanned);
 803         mod_page_state(pgdeactivate, pgdeactivate);
 804 }
 805
 806 /*
 807  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 808  */
 809 static void
 810 shrink_zone(struct zone *zone, struct scan_control *sc)
 811 {
 812         unsigned long nr_active;
 813         unsigned long nr_inactive;
 814
 815         /*
 816          * Add one to `nr_to_scan' just to make sure that the kernel will
 817          * slowly sift through the active list.
 818          */
 819         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 820         nr_active = zone->nr_scan_active;
 821         if (nr_active >= SWAP_CLUSTER_MAX)
 822                 zone->nr_scan_active = 0;
 823         else
 824                 nr_active = 0;
 825
 826         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 827         nr_inactive = zone->nr_scan_inactive;
 828         if (nr_inactive >= SWAP_CLUSTER_MAX)
 829                 zone->nr_scan_inactive = 0;
 830         else
 831                 nr_inactive = 0;
 832
 833         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 834
 835         while (nr_active || nr_inactive) {
 836                 if (nr_active) {
 837                         sc->nr_to_scan = min(nr_active,
 838                                         (unsigned long)SWAP_CLUSTER_MAX);
 839                         nr_active -= sc->nr_to_scan;
 840                         refill_inactive_zone(zone, sc);
 841                 }
 842
 843                 if (nr_inactive) {
 844                         sc->nr_to_scan = min(nr_inactive,
 845                                         (unsigned long)SWAP_CLUSTER_MAX);
 846                         nr_inactive -= sc->nr_to_scan;
 847                         shrink_cache(zone, sc);
 848                         if (sc->nr_to_reclaim <= 0)
 849                                 break;
 850                 }
 851         }
 852 }
 853
 854 /*
 855  * This is the direct reclaim path, for page-allocating processes.  We only
 856  * try to reclaim pages from zones which will satisfy the caller's allocation
 857  * request.
 858  *
 859  * We reclaim from a zone even if that zone is over pages_high.  Because:
 860  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 861  *    allocation or
 862  * b) The zones may be over pages_high but they must go *over* pages_high to
 863  *    satisfy the `incremental min' zone defense algorithm.
 864  *
 865  * Returns the number of reclaimed pages.
 866  *
 867  * If a zone is deemed to be full of pinned pages then just give it a light
 868  * scan then give up on it.
 869  */
 870 static void
 871 shrink_caches(struct zone **zones, struct scan_control *sc)
 872 {
 873         int i;
 874
 875         for (i = 0; zones[i] != NULL; i++) {
 876                 struct zone *zone = zones[i];
 877
 878                 zone->temp_priority = sc->priority;
 879                 if (zone->prev_priority > sc->priority)
 880                         zone->prev_priority = sc->priority;
 881
 882                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 883                         continue;       /* Let kswapd poll it */
 884
 885                 shrink_zone(zone, sc);
 886         }
 887 }
 888
 889 /*
 890  * This is the main entry point to direct page reclaim.
 891  *
 892  * If a full scan of the inactive list fails to free enough memory then we
 893  * are "out of memory" and something needs to be killed.
 894  *
 895  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 896  * high - the zone may be full of dirty or under-writeback pages, which this
 897  * caller can't do much about.  We kick pdflush and take explicit naps in the
 898  * hope that some of these pages can be written.  But if the allocating task
 899  * holds filesystem locks which prevent writeout this might not work, and the
 900  * allocation attempt will fail.
 901  */
 902 int try_to_free_pages(struct zone **zones,
 903                 unsigned int gfp_mask, unsigned int order)
 904 {
 905         int priority;
 906         int ret = 0;
 907         int total_scanned = 0, total_reclaimed = 0;
 908         struct reclaim_state *reclaim_state = current->reclaim_state;
 909         struct scan_control sc;
 910         unsigned long lru_pages = 0;
 911         int i;
 912
 913         sc.gfp_mask = gfp_mask;
 914         sc.may_writepage = 0;
 915
 916         inc_page_state(allocstall);
 917
 918         for (i = 0; zones[i] != NULL; i++) {
 919                 struct zone *zone = zones[i];
 920
 921                 zone->temp_priority = DEF_PRIORITY;
 922                 lru_pages += zone->nr_active + zone->nr_inactive;
 923         }
 924
 925         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 926                 sc.nr_mapped = read_page_state(nr_mapped);
 927                 sc.nr_scanned = 0;
 928                 sc.nr_reclaimed = 0;
 929                 sc.priority = priority;
 930                 shrink_caches(zones, &sc);
 931                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
 932                 if (reclaim_state) {
 933                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
 934                         reclaim_state->reclaimed_slab = 0;
 935                 }
 936                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
 937                         ret = 1;
 938                         goto out;
 939                 }
 940                 total_scanned += sc.nr_scanned;
 941                 total_reclaimed += sc.nr_reclaimed;
 942
 943                 /*
 944                  * Try to write back as many pages as we just scanned.  This
 945                  * tends to cause slow streaming writers to write data to the
 946                  * disk smoothly, at the dirtying rate, which is nice.   But
 947                  * that's undesirable in laptop mode, where we *want* lumpy
 948                  * writeout.  So in laptop mode, write out the whole world.
 949                  */
 950                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
 951                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
 952                         sc.may_writepage = 1;
 953                 }
 954
 955                 /* Take a nap, wait for some writeback to complete */
 956                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
 957                         blk_congestion_wait(WRITE, HZ/10);
 958         }
 959         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 960                 out_of_memory(gfp_mask);
 961 out:
 962         for (i = 0; zones[i] != 0; i++)
 963                 zones[i]->prev_priority = zones[i]->temp_priority;
 964         return ret;
 965 }
 966
 967 /*
 968  * For kswapd, balance_pgdat() will work across all this node's zones until
 969  * they are all at pages_high.
 970  *
 971  * If `nr_pages' is non-zero then it is the number of pages which are to be
 972  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 973  * special.
 974  *
 975  * Returns the number of pages which were actually freed.
 976  *
 977  * There is special handling here for zones which are full of pinned pages.
 978  * This can happen if the pages are all mlocked, or if they are all used by
 979  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 980  * What we do is to detect the case where all pages in the zone have been
 981  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 982  * dead and from now on, only perform a short scan.  Basically we're polling
 983  * the zone for when the problem goes away.
 984  *
 985  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
 986  * zones which have free_pages > pages_high, but once a zone is found to have
 987  * free_pages <= pages_high, we scan that zone and the lower zones regardless
 988  * of the number of free pages in the lower zones.  This interoperates with
 989  * the page allocator fallback scheme to ensure that aging of pages is balanced
 990  * across the zones.
 991  */
 992 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
 993 {
 994         int to_free = nr_pages;
 995         int priority;
 996         int i;
 997         int total_scanned = 0, total_reclaimed = 0;
 998         struct reclaim_state *reclaim_state = current->reclaim_state;
 999         struct scan_control sc;
1000
1001         sc.gfp_mask = GFP_KERNEL;
1002         sc.may_writepage = 0;
1003         sc.nr_mapped = read_page_state(nr_mapped);
1004
1005         inc_page_state(pageoutrun);
1006
1007         for (i = 0; i < pgdat->nr_zones; i++) {
1008                 struct zone *zone = pgdat->node_zones + i;
1009
1010                 zone->temp_priority = DEF_PRIORITY;
1011         }
1012
1013         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1014                 int all_zones_ok = 1;
1015                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1016                 unsigned long lru_pages = 0;
1017
1018                 if (nr_pages == 0) {
1019                         /*
1020                          * Scan in the highmem->dma direction for the highest
1021                          * zone which needs scanning
1022                          */
1023                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1024                                 struct zone *zone = pgdat->node_zones + i;
1025
1026                                 if (zone->all_unreclaimable &&
1027                                                 priority != DEF_PRIORITY)
1028                                         continue;
1029
1030                                 if (zone->free_pages <= zone->pages_high) {
1031                                         end_zone = i;
1032                                         goto scan;
1033                                 }
1034                         }
1035                         goto out;
1036                 } else {
1037                         end_zone = pgdat->nr_zones - 1;
1038                 }
1039 scan:
1040                 for (i = 0; i <= end_zone; i++) {
1041                         struct zone *zone = pgdat->node_zones + i;
1042
1043                         lru_pages += zone->nr_active + zone->nr_inactive;
1044                 }
1045
1046                 /*
1047                  * Now scan the zone in the dma->highmem direction, stopping
1048                  * at the last zone which needs scanning.
1049                  *
1050                  * We do this because the page allocator works in the opposite
1051                  * direction.  This prevents the page allocator from allocating
1052                  * pages behind kswapd's direction of progress, which would
1053                  * cause too much scanning of the lower zones.
1054                  */
1055                 for (i = 0; i <= end_zone; i++) {
1056                         struct zone *zone = pgdat->node_zones + i;
1057
1058                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1059                                 continue;
1060
1061                         if (nr_pages == 0) {    /* Not software suspend */
1062                                 if (zone->free_pages <= zone->pages_high)
1063                                         all_zones_ok = 0;
1064                         }
1065                         zone->temp_priority = priority;
1066                         if (zone->prev_priority > priority)
1067                                 zone->prev_priority = priority;
1068                         sc.nr_scanned = 0;
1069                         sc.nr_reclaimed = 0;
1070                         sc.priority = priority;
1071                         shrink_zone(zone, &sc);
1072                         reclaim_state->reclaimed_slab = 0;
1073                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1074                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1075                         total_reclaimed += sc.nr_reclaimed;
1076                         if (zone->all_unreclaimable)
1077                                 continue;
1078                         if (zone->pages_scanned > zone->present_pages * 2)
1079                                 zone->all_unreclaimable = 1;
1080                         /*
1081                          * If we've done a decent amount of scanning and
1082                          * the reclaim ratio is low, start doing writepage
1083                          * even in laptop mode
1084                          */
1085                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1086                             total_scanned > total_reclaimed+total_reclaimed/2)
1087                                 sc.may_writepage = 1;
1088                 }
1089                 if (nr_pages && to_free > total_reclaimed)
1090                         continue;       /* swsusp: need to do more work */
1091                 if (all_zones_ok)
1092                         break;          /* kswapd: all done */
1093                 /*
1094                  * OK, kswapd is getting into trouble.  Take a nap, then take
1095                  * another pass across the zones.
1096                  */
1097                 if (total_scanned && priority < DEF_PRIORITY - 2)
1098                         blk_congestion_wait(WRITE, HZ/10);
1099         }
1100 out:
1101         for (i = 0; i < pgdat->nr_zones; i++) {
1102                 struct zone *zone = pgdat->node_zones + i;
1103
1104                 zone->prev_priority = zone->temp_priority;
1105         }
1106         return total_reclaimed;
1107 }
1108
1109 /*
1110  * The background pageout daemon, started as a kernel thread
1111  * from the init process.
1112  *
1113  * This basically trickles out pages so that we have _some_
1114  * free memory available even if there is no other activity
1115  * that frees anything up. This is needed for things like routing
1116  * etc, where we otherwise might have all activity going on in
1117  * asynchronous contexts that cannot page things out.
1118  *
1119  * If there are applications that are active memory-allocators
1120  * (most normal use), this basically shouldn't matter.
1121  */
1122 static int kswapd(void *p)
1123 {
1124         pg_data_t *pgdat = (pg_data_t*)p;
1125         struct task_struct *tsk = current;
1126         DEFINE_WAIT(wait);
1127         struct reclaim_state reclaim_state = {
1128                 .reclaimed_slab = 0,
1129         };
1130         cpumask_t cpumask;
1131
1132         daemonize("kswapd%d", pgdat->node_id);
1133         cpumask = node_to_cpumask(pgdat->node_id);
1134         if (!cpus_empty(cpumask))
1135                 set_cpus_allowed(tsk, cpumask);
1136         current->reclaim_state = &reclaim_state;
1137
1138         /*
1139          * Tell the memory management that we're a "memory allocator",
1140          * and that if we need more memory we should get access to it
1141          * regardless (see "__alloc_pages()"). "kswapd" should
1142          * never get caught in the normal page freeing logic.
1143          *
1144          * (Kswapd normally doesn't need memory anyway, but sometimes
1145          * you need a small amount of memory in order to be able to
1146          * page out something else, and this flag essentially protects
1147          * us from recursively trying to free more memory as we're
1148          * trying to free the first piece of memory in the first place).
1149          */
1150         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1151
1152         for ( ; ; ) {
1153                 if (current->flags & PF_FREEZE)
1154                         refrigerator(PF_FREEZE);
1155                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1156                 schedule();
1157                 finish_wait(&pgdat->kswapd_wait, &wait);
1158                 try_to_clip_inodes();
1159
1160                 balance_pgdat(pgdat, 0);
1161         }
1162         return 0;
1163 }
1164
1165 /*
1166  * A zone is low on free memory, so wake its kswapd task to service it.
1167  */
1168 void wakeup_kswapd(struct zone *zone)
1169 {
1170         if (zone->free_pages > zone->pages_low)
1171                 return;
1172         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1173                 return;
1174         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1175 }
1176
1177 #ifdef CONFIG_PM
1178 /*
1179  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1180  * pages.
1181  */
1182 int shrink_all_memory(int nr_pages)
1183 {
1184         pg_data_t *pgdat;
1185         int nr_to_free = nr_pages;
1186         int ret = 0;
1187         struct reclaim_state reclaim_state = {
1188                 .reclaimed_slab = 0,
1189         };
1190
1191         current->reclaim_state = &reclaim_state;
1192         for_each_pgdat(pgdat) {
1193                 int freed;
1194                 freed = balance_pgdat(pgdat, nr_to_free);
1195                 ret += freed;
1196                 nr_to_free -= freed;
1197                 if (nr_to_free <= 0)
1198                         break;
1199         }
1200         current->reclaim_state = NULL;
1201         return ret;
1202 }
1203 #endif
1204
1205 #ifdef CONFIG_HOTPLUG_CPU
1206 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1207    not required for correctness.  So if the last cpu in a node goes
1208    away, we get changed to run anywhere: as the first one comes back,
1209    restore their cpu bindings. */
1210 static int __devinit cpu_callback(struct notifier_block *nfb,
1211                                   unsigned long action,
1212                                   void *hcpu)
1213 {
1214         pg_data_t *pgdat;
1215         cpumask_t mask;
1216
1217         if (action == CPU_ONLINE) {
1218                 for_each_pgdat(pgdat) {
1219                         mask = node_to_cpumask(pgdat->node_id);
1220                         if (any_online_cpu(mask) != NR_CPUS)
1221                                 /* One of our CPUs online: restore mask */
1222                                 set_cpus_allowed(pgdat->kswapd, mask);
1223                 }
1224         }
1225         return NOTIFY_OK;
1226 }
1227 #endif /* CONFIG_HOTPLUG_CPU */
1228
1229 static int __init kswapd_init(void)
1230 {
1231         pg_data_t *pgdat;
1232         swap_setup();
1233         for_each_pgdat(pgdat)
1234                 pgdat->kswapd
1235                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1236         total_memory = nr_free_pagecache_pages();
1237         hotcpu_notifier(cpu_callback, 0);
1238         return 0;
1239 }
1240
1241 module_init(kswapd_init)