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 /*
  42  * From 0 .. 100.  Higher means more swappy.
  43  */
  44 int vm_swappiness = 60;
  45 static long total_memory;
  46
  47
  48
  49 void try_to_clip_inodes(void);
  50
  51
  52 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  53
  54 #ifdef ARCH_HAS_PREFETCH
  55 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  56         do {                                                            \
  57                 if ((_page)->lru.prev != _base) {                       \
  58                         struct page *prev;                              \
  59                                                                         \
  60                         prev = lru_to_page(&(_page->lru));              \
  61                         prefetch(&prev->_field);                        \
  62                 }                                                       \
  63         } while (0)
  64 #else
  65 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
  66 #endif
  67
  68 #ifdef ARCH_HAS_PREFETCHW
  69 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
  70         do {                                                            \
  71                 if ((_page)->lru.prev != _base) {                       \
  72                         struct page *prev;                              \
  73                                                                         \
  74                         prev = lru_to_page(&(_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 EXPORT_SYMBOL(set_shrinker);
 116
 117 /*
 118  * Remove one
 119  */
 120 void remove_shrinker(struct shrinker *shrinker)
 121 {
 122         down(&shrinker_sem);
 123         list_del(&shrinker->list);
 124         up(&shrinker_sem);
 125         kfree(shrinker);
 126 }
 127
 128 EXPORT_SYMBOL(remove_shrinker);
 129
 130 #define SHRINK_BATCH 128
 131 /*
 132  * Call the shrink functions to age shrinkable caches
 133  *
 134  * Here we assume it costs one seek to replace a lru page and that it also
 135  * takes a seek to recreate a cache object.  With this in mind we age equal
 136  * percentages of the lru and ageable caches.  This should balance the seeks
 137  * generated by these structures.
 138  *
 139  * If the vm encounted mapped pages on the LRU it increase the pressure on
 140  * slab to avoid swapping.
 141  *
 142  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 143  */
 144 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask)
 145 {
 146         struct shrinker *shrinker;
 147         long pages;
 148
 149         if (down_trylock(&shrinker_sem))
 150                 return 0;
 151
 152         pages = nr_used_zone_pages();
 153         list_for_each_entry(shrinker, &shrinker_list, list) {
 154                 unsigned long long delta;
 155
 156                 delta = (4 * scanned) / shrinker->seeks;
 157                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 158                 do_div(delta, pages + 1);
 159                 shrinker->nr += delta;
 160                 if (shrinker->nr < 0)
 161                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 162
 163                 if (shrinker->nr <= SHRINK_BATCH)
 164                         continue;
 165                 while (shrinker->nr) {
 166                         long this_scan = shrinker->nr;
 167                         int shrink_ret;
 168
 169                         if (this_scan > 128)
 170                                 this_scan = 128;
 171                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 172                         mod_page_state(slabs_scanned, this_scan);
 173                         shrinker->nr -= this_scan;
 174                         if (shrink_ret == -1)
 175                                 break;
 176                         cond_resched();
 177                 }
 178         }
 179         up(&shrinker_sem);
 180         return 0;
 181 }
 182
 183 /* Must be called with page's rmap lock held. */
 184 static inline int page_mapping_inuse(struct page *page)
 185 {
 186         struct address_space *mapping;
 187
 188         /* Page is in somebody's page tables. */
 189         if (page_mapped(page))
 190                 return 1;
 191
 192         /* Be more reluctant to reclaim swapcache than pagecache */
 193         if (PageSwapCache(page))
 194                 return 1;
 195
 196         mapping = page_mapping(page);
 197         if (!mapping)
 198                 return 0;
 199
 200         /* File is mmap'd by somebody? */
 201         return mapping_mapped(mapping);
 202 }
 203
 204 static inline int is_page_cache_freeable(struct page *page)
 205 {
 206         return page_count(page) - !!PagePrivate(page) == 2;
 207 }
 208
 209 static int may_write_to_queue(struct backing_dev_info *bdi)
 210 {
 211         if (current_is_kswapd())
 212                 return 1;
 213         if (current_is_pdflush())       /* This is unlikely, but why not... */
 214                 return 1;
 215         if (!bdi_write_congested(bdi))
 216                 return 1;
 217         if (bdi == current->backing_dev_info)
 218                 return 1;
 219         return 0;
 220 }
 221
 222 /*
 223  * We detected a synchronous write error writing a page out.  Probably
 224  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 225  * fsync(), msync() or close().
 226  *
 227  * The tricky part is that after writepage we cannot touch the mapping: nothing
 228  * prevents it from being freed up.  But we have a ref on the page and once
 229  * that page is locked, the mapping is pinned.
 230  *
 231  * We're allowed to run sleeping lock_page() here because we know the caller has
 232  * __GFP_FS.
 233  */
 234 static void handle_write_error(struct address_space *mapping,
 235                                 struct page *page, int error)
 236 {
 237         lock_page(page);
 238         if (page_mapping(page) == mapping) {
 239                 if (error == -ENOSPC)
 240                         set_bit(AS_ENOSPC, &mapping->flags);
 241                 else
 242                         set_bit(AS_EIO, &mapping->flags);
 243         }
 244         unlock_page(page);
 245 }
 246
 247 /* possible outcome of pageout() */
 248 typedef enum {
 249         /* failed to write page out, page is locked */
 250         PAGE_KEEP,
 251         /* move page to the active list, page is locked */
 252         PAGE_ACTIVATE,
 253         /* page has been sent to the disk successfully, page is unlocked */
 254         PAGE_SUCCESS,
 255         /* page is clean and locked */
 256         PAGE_CLEAN,
 257 } pageout_t;
 258
 259 /*
 260  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 261  */
 262 static pageout_t pageout(struct page *page, struct address_space *mapping)
 263 {
 264         /*
 265          * If the page is dirty, only perform writeback if that write
 266          * will be non-blocking.  To prevent this allocation from being
 267          * stalled by pagecache activity.  But note that there may be
 268          * stalls if we need to run get_block().  We could test
 269          * PagePrivate for that.
 270          *
 271          * If this process is currently in generic_file_write() against
 272          * this page's queue, we can perform writeback even if that
 273          * will block.
 274          *
 275          * If the page is swapcache, write it back even if that would
 276          * block, for some throttling. This happens by accident, because
 277          * swap_backing_dev_info is bust: it doesn't reflect the
 278          * congestion state of the swapdevs.  Easy to fix, if needed.
 279          * See swapfile.c:page_queue_congested().
 280          */
 281         if (!is_page_cache_freeable(page))
 282                 return PAGE_KEEP;
 283         if (!mapping)
 284                 return PAGE_KEEP;
 285         if (mapping->a_ops->writepage == NULL)
 286                 return PAGE_ACTIVATE;
 287         if (!may_write_to_queue(mapping->backing_dev_info))
 288                 return PAGE_KEEP;
 289
 290         if (clear_page_dirty_for_io(page)) {
 291                 int res;
 292                 struct writeback_control wbc = {
 293                         .sync_mode = WB_SYNC_NONE,
 294                         .nr_to_write = SWAP_CLUSTER_MAX,
 295                         .nonblocking = 1,
 296                         .for_reclaim = 1,
 297                 };
 298
 299                 SetPageReclaim(page);
 300                 res = mapping->a_ops->writepage(page, &wbc);
 301                 if (res < 0)
 302                         handle_write_error(mapping, page, res);
 303                 if (res == WRITEPAGE_ACTIVATE) {
 304                         ClearPageReclaim(page);
 305                         return PAGE_ACTIVATE;
 306                 }
 307                 if (!PageWriteback(page)) {
 308                         /* synchronous write or broken a_ops? */
 309                         ClearPageReclaim(page);
 310                 }
 311
 312                 return PAGE_SUCCESS;
 313         }
 314
 315         return PAGE_CLEAN;
 316 }
 317
 318 struct scan_control {
 319         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
 320         unsigned long nr_to_scan;
 321
 322         /* Incremented by the number of inactive pages that were scanned */
 323         unsigned long nr_scanned;
 324
 325         /* Incremented by the number of pages reclaimed */
 326         unsigned long nr_reclaimed;
 327
 328         unsigned long nr_mapped;        /* From page_state */
 329
 330         /* Ask shrink_caches, or shrink_zone to scan at this priority */
 331         unsigned int priority;
 332
 333         /* This context's GFP mask */
 334         unsigned int gfp_mask;
 335
 336         int may_writepage;
 337 };
 338
 339 /*
 340  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 341  */
 342 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 343 {
 344         LIST_HEAD(ret_pages);
 345         struct pagevec freed_pvec;
 346         int pgactivate = 0;
 347         int reclaimed = 0;
 348
 349         cond_resched();
 350
 351         pagevec_init(&freed_pvec, 1);
 352         while (!list_empty(page_list)) {
 353                 struct address_space *mapping;
 354                 struct page *page;
 355                 int may_enter_fs;
 356                 int referenced;
 357
 358                 page = lru_to_page(page_list);
 359                 list_del(&page->lru);
 360
 361                 if (TestSetPageLocked(page))
 362                         goto keep;
 363
 364                 BUG_ON(PageActive(page));
 365
 366                 if (PageWriteback(page))
 367                         goto keep_locked;
 368
 369                 sc->nr_scanned++;
 370                 /* Double the slab pressure for mapped and swapcache pages */
 371                 if (page_mapped(page) || PageSwapCache(page))
 372                         sc->nr_scanned++;
 373
 374                 page_map_lock(page);
 375                 referenced = page_referenced(page);
 376                 if (referenced && page_mapping_inuse(page)) {
 377                         /* In active use or really unfreeable.  Activate it. */
 378                         page_map_unlock(page);
 379                         goto activate_locked;
 380                 }
 381
 382 #ifdef CONFIG_SWAP
 383                 /*
 384                  * Anonymous process memory has backing store?
 385                  * Try to allocate it some swap space here.
 386                  *
 387                  * XXX: implement swap clustering ?
 388                  */
 389                 if (PageAnon(page) && !PageSwapCache(page)) {
 390                         page_map_unlock(page);
 391                         if (!add_to_swap(page))
 392                                 goto activate_locked;
 393                         page_map_lock(page);
 394                 }
 395 #endif /* CONFIG_SWAP */
 396
 397                 mapping = page_mapping(page);
 398                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 399                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 400
 401                 /*
 402                  * The page is mapped into the page tables of one or more
 403                  * processes. Try to unmap it here.
 404                  */
 405                 if (page_mapped(page) && mapping) {
 406                         switch (try_to_unmap(page)) {
 407                         case SWAP_FAIL:
 408                                 page_map_unlock(page);
 409                                 goto activate_locked;
 410                         case SWAP_AGAIN:
 411                                 page_map_unlock(page);
 412                                 goto keep_locked;
 413                         case SWAP_SUCCESS:
 414                                 ; /* try to free the page below */
 415                         }
 416                 }
 417                 page_map_unlock(page);
 418
 419                 if (PageDirty(page)) {
 420                         if (referenced)
 421                                 goto keep_locked;
 422                         if (!may_enter_fs)
 423                                 goto keep_locked;
 424                         if (laptop_mode && !sc->may_writepage)
 425                                 goto keep_locked;
 426
 427                         /* Page is dirty, try to write it out here */
 428                         switch(pageout(page, mapping)) {
 429                         case PAGE_KEEP:
 430                                 goto keep_locked;
 431                         case PAGE_ACTIVATE:
 432                                 goto activate_locked;
 433                         case PAGE_SUCCESS:
 434                                 if (PageWriteback(page) || PageDirty(page))
 435                                         goto keep;
 436                                 /*
 437                                  * A synchronous write - probably a ramdisk.  Go
 438                                  * ahead and try to reclaim the page.
 439                                  */
 440                                 if (TestSetPageLocked(page))
 441                                         goto keep;
 442                                 if (PageDirty(page) || PageWriteback(page))
 443                                         goto keep_locked;
 444                                 mapping = page_mapping(page);
 445                         case PAGE_CLEAN:
 446                                 ; /* try to free the page below */
 447                         }
 448                 }
 449
 450                 /*
 451                  * If the page has buffers, try to free the buffer mappings
 452                  * associated with this page. If we succeed we try to free
 453                  * the page as well.
 454                  *
 455                  * We do this even if the page is PageDirty().
 456                  * try_to_release_page() does not perform I/O, but it is
 457                  * possible for a page to have PageDirty set, but it is actually
 458                  * clean (all its buffers are clean).  This happens if the
 459                  * buffers were written out directly, with submit_bh(). ext3
 460                  * will do this, as well as the blockdev mapping.
 461                  * try_to_release_page() will discover that cleanness and will
 462                  * drop the buffers and mark the page clean - it can be freed.
 463                  *
 464                  * Rarely, pages can have buffers and no ->mapping.  These are
 465                  * the pages which were not successfully invalidated in
 466                  * truncate_complete_page().  We try to drop those buffers here
 467                  * and if that worked, and the page is no longer mapped into
 468                  * process address space (page_count == 1) it can be freed.
 469                  * Otherwise, leave the page on the LRU so it is swappable.
 470                  */
 471                 if (PagePrivate(page)) {
 472                         if (!try_to_release_page(page, sc->gfp_mask))
 473                                 goto activate_locked;
 474                         if (!mapping && page_count(page) == 1)
 475                                 goto free_it;
 476                 }
 477
 478                 if (!mapping)
 479                         goto keep_locked;       /* truncate got there first */
 480
 481                 spin_lock_irq(&mapping->tree_lock);
 482
 483                 /*
 484                  * The non-racy check for busy page.  It is critical to check
 485                  * PageDirty _after_ making sure that the page is freeable and
 486                  * not in use by anybody.       (pagecache + us == 2)
 487                  */
 488                 if (page_count(page) != 2 || PageDirty(page)) {
 489                         spin_unlock_irq(&mapping->tree_lock);
 490                         goto keep_locked;
 491                 }
 492
 493 #ifdef CONFIG_SWAP
 494                 if (PageSwapCache(page)) {
 495                         swp_entry_t swap = { .val = page->private };
 496                         __delete_from_swap_cache(page);
 497                         spin_unlock_irq(&mapping->tree_lock);
 498                         swap_free(swap);
 499                         __put_page(page);       /* The pagecache ref */
 500                         goto free_it;
 501                 }
 502 #endif /* CONFIG_SWAP */
 503
 504                 __remove_from_page_cache(page);
 505                 spin_unlock_irq(&mapping->tree_lock);
 506                 __put_page(page);
 507
 508 free_it:
 509                 unlock_page(page);
 510                 reclaimed++;
 511                 if (!pagevec_add(&freed_pvec, page))
 512                         __pagevec_release_nonlru(&freed_pvec);
 513                 continue;
 514
 515 activate_locked:
 516                 SetPageActive(page);
 517                 pgactivate++;
 518 keep_locked:
 519                 unlock_page(page);
 520 keep:
 521                 list_add(&page->lru, &ret_pages);
 522                 BUG_ON(PageLRU(page));
 523         }
 524         list_splice(&ret_pages, page_list);
 525         if (pagevec_count(&freed_pvec))
 526                 __pagevec_release_nonlru(&freed_pvec);
 527         mod_page_state(pgactivate, pgactivate);
 528         sc->nr_reclaimed += reclaimed;
 529         return reclaimed;
 530 }
 531
 532 /*
 533  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 534  * a batch of pages and working on them outside the lock.  Any pages which were
 535  * not freed will be added back to the LRU.
 536  *
 537  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 538  *
 539  * For pagecache intensive workloads, the first loop here is the hottest spot
 540  * in the kernel (apart from the copy_*_user functions).
 541  */
 542 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 543 {
 544         LIST_HEAD(page_list);
 545         struct pagevec pvec;
 546         int max_scan = sc->nr_to_scan;
 547
 548         pagevec_init(&pvec, 1);
 549
 550         lru_add_drain();
 551         spin_lock_irq(&zone->lru_lock);
 552         while (max_scan > 0) {
 553                 struct page *page;
 554                 int nr_taken = 0;
 555                 int nr_scan = 0;
 556                 int nr_freed;
 557
 558                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 559                                 !list_empty(&zone->inactive_list)) {
 560                         page = lru_to_page(&zone->inactive_list);
 561
 562                         prefetchw_prev_lru_page(page,
 563                                                 &zone->inactive_list, flags);
 564
 565                         if (!TestClearPageLRU(page))
 566                                 BUG();
 567                         list_del(&page->lru);
 568                         if (get_page_testone(page)) {
 569                                 /*
 570                                  * It is being freed elsewhere
 571                                  */
 572                                 __put_page(page);
 573                                 SetPageLRU(page);
 574                                 list_add(&page->lru, &zone->inactive_list);
 575                                 continue;
 576                         }
 577                         list_add(&page->lru, &page_list);
 578                         nr_taken++;
 579                 }
 580                 zone->nr_inactive -= nr_taken;
 581                 zone->pages_scanned += nr_taken;
 582                 spin_unlock_irq(&zone->lru_lock);
 583
 584                 if (nr_taken == 0)
 585                         goto done;
 586
 587                 max_scan -= nr_scan;
 588                 if (current_is_kswapd())
 589                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 590                 else
 591                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 592                 nr_freed = shrink_list(&page_list, sc);
 593                 if (current_is_kswapd())
 594                         mod_page_state(kswapd_steal, nr_freed);
 595                 mod_page_state_zone(zone, pgsteal, nr_freed);
 596
 597                 spin_lock_irq(&zone->lru_lock);
 598                 /*
 599                  * Put back any unfreeable pages.
 600                  */
 601                 while (!list_empty(&page_list)) {
 602                         page = lru_to_page(&page_list);
 603                         if (TestSetPageLRU(page))
 604                                 BUG();
 605                         list_del(&page->lru);
 606                         if (PageActive(page))
 607                                 add_page_to_active_list(zone, page);
 608                         else
 609                                 add_page_to_inactive_list(zone, page);
 610                         if (!pagevec_add(&pvec, page)) {
 611                                 spin_unlock_irq(&zone->lru_lock);
 612                                 __pagevec_release(&pvec);
 613                                 spin_lock_irq(&zone->lru_lock);
 614                         }
 615                 }
 616         }
 617         spin_unlock_irq(&zone->lru_lock);
 618 done:
 619         pagevec_release(&pvec);
 620 }
 621
 622 /*
 623  * This moves pages from the active list to the inactive list.
 624  *
 625  * We move them the other way if the page is referenced by one or more
 626  * processes, from rmap.
 627  *
 628  * If the pages are mostly unmapped, the processing is fast and it is
 629  * appropriate to hold zone->lru_lock across the whole operation.  But if
 630  * the pages are mapped, the processing is slow (page_referenced()) so we
 631  * should drop zone->lru_lock around each page.  It's impossible to balance
 632  * this, so instead we remove the pages from the LRU while processing them.
 633  * It is safe to rely on PG_active against the non-LRU pages in here because
 634  * nobody will play with that bit on a non-LRU page.
 635  *
 636  * The downside is that we have to touch page->_count against each page.
 637  * But we had to alter page->flags anyway.
 638  */
 639 static void
 640 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 641 {
 642         int pgmoved;
 643         int pgdeactivate = 0;
 644         int pgscanned = 0;
 645         int nr_pages = sc->nr_to_scan;
 646         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 647         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 648         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 649         struct page *page;
 650         struct pagevec pvec;
 651         int reclaim_mapped = 0;
 652         long mapped_ratio;
 653         long distress;
 654         long swap_tendency;
 655
 656         lru_add_drain();
 657         pgmoved = 0;
 658         spin_lock_irq(&zone->lru_lock);
 659         while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
 660                 page = lru_to_page(&zone->active_list);
 661                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 662                 if (!TestClearPageLRU(page))
 663                         BUG();
 664                 list_del(&page->lru);
 665                 if (get_page_testone(page)) {
 666                         /*
 667                          * It was already free!  release_pages() or put_page()
 668                          * are about to remove it from the LRU and free it. So
 669                          * put the refcount back and put the page back on the
 670                          * LRU
 671                          */
 672                         __put_page(page);
 673                         SetPageLRU(page);
 674                         list_add(&page->lru, &zone->active_list);
 675                 } else {
 676                         list_add(&page->lru, &l_hold);
 677                         pgmoved++;
 678                 }
 679                 pgscanned++;
 680         }
 681         zone->nr_active -= pgmoved;
 682         spin_unlock_irq(&zone->lru_lock);
 683
 684         /*
 685          * `distress' is a measure of how much trouble we're having reclaiming
 686          * pages.  0 -> no problems.  100 -> great trouble.
 687          */
 688         distress = 100 >> zone->prev_priority;
 689
 690         /*
 691          * The point of this algorithm is to decide when to start reclaiming
 692          * mapped memory instead of just pagecache.  Work out how much memory
 693          * is mapped.
 694          */
 695         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 696
 697         /*
 698          * Now decide how much we really want to unmap some pages.  The mapped
 699          * ratio is downgraded - just because there's a lot of mapped memory
 700          * doesn't necessarily mean that page reclaim isn't succeeding.
 701          *
 702          * The distress ratio is important - we don't want to start going oom.
 703          *
 704          * A 100% value of vm_swappiness overrides this algorithm altogether.
 705          */
 706         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 707
 708         /*
 709          * Now use this metric to decide whether to start moving mapped memory
 710          * onto the inactive list.
 711          */
 712         if (swap_tendency >= 100)
 713                 reclaim_mapped = 1;
 714
 715         while (!list_empty(&l_hold)) {
 716                 page = lru_to_page(&l_hold);
 717                 list_del(&page->lru);
 718                 if (page_mapped(page)) {
 719                         if (!reclaim_mapped) {
 720                                 list_add(&page->lru, &l_active);
 721                                 continue;
 722                         }
 723                         page_map_lock(page);
 724                         if (page_referenced(page)) {
 725                                 page_map_unlock(page);
 726                                 list_add(&page->lru, &l_active);
 727                                 continue;
 728                         }
 729                         page_map_unlock(page);
 730                 }
 731                 /*
 732                  * FIXME: need to consider page_count(page) here if/when we
 733                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 734                  */
 735                 if (total_swap_pages == 0 && PageAnon(page)) {
 736                         list_add(&page->lru, &l_active);
 737                         continue;
 738                 }
 739                 list_add(&page->lru, &l_inactive);
 740         }
 741
 742         pagevec_init(&pvec, 1);
 743         pgmoved = 0;
 744         spin_lock_irq(&zone->lru_lock);
 745         while (!list_empty(&l_inactive)) {
 746                 page = lru_to_page(&l_inactive);
 747                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 748                 if (TestSetPageLRU(page))
 749                         BUG();
 750                 if (!TestClearPageActive(page))
 751                         BUG();
 752                 list_move(&page->lru, &zone->inactive_list);
 753                 pgmoved++;
 754                 if (!pagevec_add(&pvec, page)) {
 755                         zone->nr_inactive += pgmoved;
 756                         spin_unlock_irq(&zone->lru_lock);
 757                         pgdeactivate += pgmoved;
 758                         pgmoved = 0;
 759                         if (buffer_heads_over_limit)
 760                                 pagevec_strip(&pvec);
 761                         __pagevec_release(&pvec);
 762                         spin_lock_irq(&zone->lru_lock);
 763                 }
 764         }
 765         zone->nr_inactive += pgmoved;
 766         pgdeactivate += pgmoved;
 767         if (buffer_heads_over_limit) {
 768                 spin_unlock_irq(&zone->lru_lock);
 769                 pagevec_strip(&pvec);
 770                 spin_lock_irq(&zone->lru_lock);
 771         }
 772
 773         pgmoved = 0;
 774         while (!list_empty(&l_active)) {
 775                 page = lru_to_page(&l_active);
 776                 prefetchw_prev_lru_page(page, &l_active, flags);
 777                 if (TestSetPageLRU(page))
 778                         BUG();
 779                 BUG_ON(!PageActive(page));
 780                 list_move(&page->lru, &zone->active_list);
 781                 pgmoved++;
 782                 if (!pagevec_add(&pvec, page)) {
 783                         zone->nr_active += pgmoved;
 784                         pgmoved = 0;
 785                         spin_unlock_irq(&zone->lru_lock);
 786                         __pagevec_release(&pvec);
 787                         spin_lock_irq(&zone->lru_lock);
 788                 }
 789         }
 790         zone->nr_active += pgmoved;
 791         spin_unlock_irq(&zone->lru_lock);
 792         pagevec_release(&pvec);
 793
 794         mod_page_state_zone(zone, pgrefill, pgscanned);
 795         mod_page_state(pgdeactivate, pgdeactivate);
 796 }
 797
 798 /*
 799  * Scan `nr_pages' from this zone.  Returns the number of reclaimed pages.
 800  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 801  */
 802 static void
 803 shrink_zone(struct zone *zone, struct scan_control *sc)
 804 {
 805         unsigned long scan_active, scan_inactive;
 806         int count;
 807
 808         scan_inactive = (zone->nr_active + zone->nr_inactive) >> sc->priority;
 809
 810         /*
 811          * Try to keep the active list 2/3 of the size of the cache.  And
 812          * make sure that refill_inactive is given a decent number of pages.
 813          *
 814          * The "scan_active + 1" here is important.  With pagecache-intensive
 815          * workloads the inactive list is huge, and `ratio' evaluates to zero
 816          * all the time.  Which pins the active list memory.  So we add one to
 817          * `scan_active' just to make sure that the kernel will slowly sift
 818          * through the active list.
 819          */
 820         if (zone->nr_active >= 4*(zone->nr_inactive*2 + 1)) {
 821                 /* Don't scan more than 4 times the inactive list scan size */
 822                 scan_active = 4*scan_inactive;
 823         } else {
 824                 unsigned long long tmp;
 825
 826                 /* Cast to long long so the multiply doesn't overflow */
 827
 828                 tmp = (unsigned long long)scan_inactive * zone->nr_active;
 829                 do_div(tmp, zone->nr_inactive*2 + 1);
 830                 scan_active = (unsigned long)tmp;
 831         }
 832
 833         atomic_add(scan_active + 1, &zone->nr_scan_active);
 834         count = atomic_read(&zone->nr_scan_active);
 835         if (count >= SWAP_CLUSTER_MAX) {
 836                 atomic_set(&zone->nr_scan_active, 0);
 837                 sc->nr_to_scan = count;
 838                 refill_inactive_zone(zone, sc);
 839         }
 840
 841         atomic_add(scan_inactive, &zone->nr_scan_inactive);
 842         count = atomic_read(&zone->nr_scan_inactive);
 843         if (count >= SWAP_CLUSTER_MAX) {
 844                 atomic_set(&zone->nr_scan_inactive, 0);
 845                 sc->nr_to_scan = count;
 846                 shrink_cache(zone, sc);
 847         }
 848 }
 849
 850 /*
 851  * This is the direct reclaim path, for page-allocating processes.  We only
 852  * try to reclaim pages from zones which will satisfy the caller's allocation
 853  * request.
 854  *
 855  * We reclaim from a zone even if that zone is over pages_high.  Because:
 856  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 857  *    allocation or
 858  * b) The zones may be over pages_high but they must go *over* pages_high to
 859  *    satisfy the `incremental min' zone defense algorithm.
 860  *
 861  * Returns the number of reclaimed pages.
 862  *
 863  * If a zone is deemed to be full of pinned pages then just give it a light
 864  * scan then give up on it.
 865  */
 866 static void
 867 shrink_caches(struct zone **zones, struct scan_control *sc)
 868 {
 869         int i;
 870
 871         for (i = 0; zones[i] != NULL; i++) {
 872                 struct zone *zone = zones[i];
 873
 874                 zone->temp_priority = sc->priority;
 875                 if (zone->prev_priority > sc->priority)
 876                         zone->prev_priority = sc->priority;
 877
 878                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 879                         continue;       /* Let kswapd poll it */
 880
 881                 shrink_zone(zone, sc);
 882         }
 883 }
 884
 885 /*
 886  * This is the main entry point to direct page reclaim.
 887  *
 888  * If a full scan of the inactive list fails to free enough memory then we
 889  * are "out of memory" and something needs to be killed.
 890  *
 891  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 892  * high - the zone may be full of dirty or under-writeback pages, which this
 893  * caller can't do much about.  We kick pdflush and take explicit naps in the
 894  * hope that some of these pages can be written.  But if the allocating task
 895  * holds filesystem locks which prevent writeout this might not work, and the
 896  * allocation attempt will fail.
 897  */
 898 int try_to_free_pages(struct zone **zones,
 899                 unsigned int gfp_mask, unsigned int order)
 900 {
 901         int priority;
 902         int ret = 0;
 903         int total_scanned = 0, total_reclaimed = 0;
 904         struct reclaim_state *reclaim_state = current->reclaim_state;
 905         struct scan_control sc;
 906         int i;
 907
 908         sc.gfp_mask = gfp_mask;
 909         sc.may_writepage = 0;
 910
 911         inc_page_state(allocstall);
 912
 913         for (i = 0; zones[i] != 0; i++)
 914                 zones[i]->temp_priority = DEF_PRIORITY;
 915
 916         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 917                 sc.nr_mapped = read_page_state(nr_mapped);
 918                 sc.nr_scanned = 0;
 919                 sc.nr_reclaimed = 0;
 920                 sc.priority = priority;
 921                 shrink_caches(zones, &sc);
 922                 shrink_slab(sc.nr_scanned, gfp_mask);
 923                 if (reclaim_state) {
 924                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
 925                         reclaim_state->reclaimed_slab = 0;
 926                 }
 927                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
 928                         ret = 1;
 929                         goto out;
 930                 }
 931                 total_scanned += sc.nr_scanned;
 932                 total_reclaimed += sc.nr_reclaimed;
 933
 934                 /*
 935                  * Try to write back as many pages as we just scanned.  This
 936                  * tends to cause slow streaming writers to write data to the
 937                  * disk smoothly, at the dirtying rate, which is nice.   But
 938                  * that's undesirable in laptop mode, where we *want* lumpy
 939                  * writeout.  So in laptop mode, write out the whole world.
 940                  */
 941                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
 942                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
 943                         sc.may_writepage = 1;
 944                 }
 945
 946                 /* Take a nap, wait for some writeback to complete */
 947                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
 948                         blk_congestion_wait(WRITE, HZ/10);
 949         }
 950         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 951                 out_of_memory();
 952 out:
 953         for (i = 0; zones[i] != 0; i++)
 954                 zones[i]->prev_priority = zones[i]->temp_priority;
 955         return ret;
 956 }
 957
 958 /*
 959  * For kswapd, balance_pgdat() will work across all this node's zones until
 960  * they are all at pages_high.
 961  *
 962  * If `nr_pages' is non-zero then it is the number of pages which are to be
 963  * reclaimed, regardless of the zone occupancies.  This is a software suspend
 964  * special.
 965  *
 966  * Returns the number of pages which were actually freed.
 967  *
 968  * There is special handling here for zones which are full of pinned pages.
 969  * This can happen if the pages are all mlocked, or if they are all used by
 970  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
 971  * What we do is to detect the case where all pages in the zone have been
 972  * scanned twice and there has been zero successful reclaim.  Mark the zone as
 973  * dead and from now on, only perform a short scan.  Basically we're polling
 974  * the zone for when the problem goes away.
 975  *
 976  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
 977  * zones which have free_pages > pages_high, but once a zone is found to have
 978  * free_pages <= pages_high, we scan that zone and the lower zones regardless
 979  * of the number of free pages in the lower zones.  This interoperates with
 980  * the page allocator fallback scheme to ensure that aging of pages is balanced
 981  * across the zones.
 982  */
 983 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
 984 {
 985         int to_free = nr_pages;
 986         int priority;
 987         int i;
 988         int total_scanned = 0, total_reclaimed = 0;
 989         struct reclaim_state *reclaim_state = current->reclaim_state;
 990         struct scan_control sc;
 991
 992         sc.gfp_mask = GFP_KERNEL;
 993         sc.may_writepage = 0;
 994         sc.nr_mapped = read_page_state(nr_mapped);
 995
 996         inc_page_state(pageoutrun);
 997
 998         for (i = 0; i < pgdat->nr_zones; i++) {
 999                 struct zone *zone = pgdat->node_zones + i;
1000
1001                 zone->temp_priority = DEF_PRIORITY;
1002         }
1003
1004         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1005                 int all_zones_ok = 1;
1006                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1007
1008
1009                 if (nr_pages == 0) {
1010                         /*
1011                          * Scan in the highmem->dma direction for the highest
1012                          * zone which needs scanning
1013                          */
1014                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1015                                 struct zone *zone = pgdat->node_zones + i;
1016
1017                                 if (zone->all_unreclaimable &&
1018                                                 priority != DEF_PRIORITY)
1019                                         continue;
1020
1021                                 if (zone->free_pages <= zone->pages_high) {
1022                                         end_zone = i;
1023                                         goto scan;
1024                                 }
1025                         }
1026                         goto out;
1027                 } else {
1028                         end_zone = pgdat->nr_zones - 1;
1029                 }
1030 scan:
1031                 /*
1032                  * Now scan the zone in the dma->highmem direction, stopping
1033                  * at the last zone which needs scanning.
1034                  *
1035                  * We do this because the page allocator works in the opposite
1036                  * direction.  This prevents the page allocator from allocating
1037                  * pages behind kswapd's direction of progress, which would
1038                  * cause too much scanning of the lower zones.
1039                  */
1040                 for (i = 0; i <= end_zone; i++) {
1041                         struct zone *zone = pgdat->node_zones + i;
1042
1043                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1044                                 continue;
1045
1046                         if (nr_pages == 0) {    /* Not software suspend */
1047                                 if (zone->free_pages <= zone->pages_high)
1048                                         all_zones_ok = 0;
1049                         }
1050                         zone->temp_priority = priority;
1051                         if (zone->prev_priority > priority)
1052                                 zone->prev_priority = priority;
1053                         sc.nr_scanned = 0;
1054                         sc.nr_reclaimed = 0;
1055                         sc.priority = priority;
1056                         shrink_zone(zone, &sc);
1057                         reclaim_state->reclaimed_slab = 0;
1058                         shrink_slab(sc.nr_scanned, GFP_KERNEL);
1059                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1060                         total_reclaimed += sc.nr_reclaimed;
1061                         if (zone->all_unreclaimable)
1062                                 continue;
1063                         if (zone->pages_scanned > zone->present_pages * 2)
1064                                 zone->all_unreclaimable = 1;
1065                         /*
1066                          * If we've done a decent amount of scanning and
1067                          * the reclaim ratio is low, start doing writepage
1068                          * even in laptop mode
1069                          */
1070                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1071                             total_scanned > total_reclaimed+total_reclaimed/2)
1072                                 sc.may_writepage = 1;
1073                 }
1074                 if (nr_pages && to_free > total_reclaimed)
1075                         continue;       /* swsusp: need to do more work */
1076                 if (all_zones_ok)
1077                         break;          /* kswapd: all done */
1078                 /*
1079                  * OK, kswapd is getting into trouble.  Take a nap, then take
1080                  * another pass across the zones.
1081                  */
1082                 if (total_scanned && priority < DEF_PRIORITY - 2)
1083                         blk_congestion_wait(WRITE, HZ/10);
1084         }
1085 out:
1086         for (i = 0; i < pgdat->nr_zones; i++) {
1087                 struct zone *zone = pgdat->node_zones + i;
1088
1089                 zone->prev_priority = zone->temp_priority;
1090         }
1091         return total_reclaimed;
1092 }
1093
1094 /*
1095  * The background pageout daemon, started as a kernel thread
1096  * from the init process.
1097  *
1098  * This basically trickles out pages so that we have _some_
1099  * free memory available even if there is no other activity
1100  * that frees anything up. This is needed for things like routing
1101  * etc, where we otherwise might have all activity going on in
1102  * asynchronous contexts that cannot page things out.
1103  *
1104  * If there are applications that are active memory-allocators
1105  * (most normal use), this basically shouldn't matter.
1106  */
1107 int kswapd(void *p)
1108 {
1109         pg_data_t *pgdat = (pg_data_t*)p;
1110         struct task_struct *tsk = current;
1111         DEFINE_WAIT(wait);
1112         struct reclaim_state reclaim_state = {
1113                 .reclaimed_slab = 0,
1114         };
1115         cpumask_t cpumask;
1116
1117         daemonize("kswapd%d", pgdat->node_id);
1118         cpumask = node_to_cpumask(pgdat->node_id);
1119         if (!cpus_empty(cpumask))
1120                 set_cpus_allowed(tsk, cpumask);
1121         current->reclaim_state = &reclaim_state;
1122
1123         /*
1124          * Tell the memory management that we're a "memory allocator",
1125          * and that if we need more memory we should get access to it
1126          * regardless (see "__alloc_pages()"). "kswapd" should
1127          * never get caught in the normal page freeing logic.
1128          *
1129          * (Kswapd normally doesn't need memory anyway, but sometimes
1130          * you need a small amount of memory in order to be able to
1131          * page out something else, and this flag essentially protects
1132          * us from recursively trying to free more memory as we're
1133          * trying to free the first piece of memory in the first place).
1134          */
1135         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1136
1137         for ( ; ; ) {
1138                 if (current->flags & PF_FREEZE)
1139                         refrigerator(PF_FREEZE);
1140                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1141                 schedule();
1142                 finish_wait(&pgdat->kswapd_wait, &wait);
1143                 try_to_clip_inodes();
1144
1145                 balance_pgdat(pgdat, 0);
1146         }
1147 }
1148
1149 /*
1150  * A zone is low on free memory, so wake its kswapd task to service it.
1151  */
1152 void wakeup_kswapd(struct zone *zone)
1153 {
1154         if (zone->free_pages > zone->pages_low)
1155                 return;
1156         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1157                 return;
1158         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1159 }
1160
1161 #ifdef CONFIG_PM
1162 /*
1163  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1164  * pages.
1165  */
1166 int shrink_all_memory(int nr_pages)
1167 {
1168         pg_data_t *pgdat;
1169         int nr_to_free = nr_pages;
1170         int ret = 0;
1171         struct reclaim_state reclaim_state = {
1172                 .reclaimed_slab = 0,
1173         };
1174
1175         current->reclaim_state = &reclaim_state;
1176         for_each_pgdat(pgdat) {
1177                 int freed;
1178                 freed = balance_pgdat(pgdat, nr_to_free);
1179                 ret += freed;
1180                 nr_to_free -= freed;
1181                 if (nr_to_free <= 0)
1182                         break;
1183         }
1184         current->reclaim_state = NULL;
1185         return ret;
1186 }
1187 #endif
1188
1189 #ifdef CONFIG_HOTPLUG_CPU
1190 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1191    not required for correctness.  So if the last cpu in a node goes
1192    away, we get changed to run anywhere: as the first one comes back,
1193    restore their cpu bindings. */
1194 static int __devinit cpu_callback(struct notifier_block *nfb,
1195                                   unsigned long action,
1196                                   void *hcpu)
1197 {
1198         pg_data_t *pgdat;
1199         cpumask_t mask;
1200
1201         if (action == CPU_ONLINE) {
1202                 for_each_pgdat(pgdat) {
1203                         mask = node_to_cpumask(pgdat->node_id);
1204                         if (any_online_cpu(mask) != NR_CPUS)
1205                                 /* One of our CPUs online: restore mask */
1206                                 set_cpus_allowed(pgdat->kswapd, mask);
1207                 }
1208         }
1209         return NOTIFY_OK;
1210 }
1211 #endif /* CONFIG_HOTPLUG_CPU */
1212
1213 static int __init kswapd_init(void)
1214 {
1215         pg_data_t *pgdat;
1216         swap_setup();
1217         for_each_pgdat(pgdat)
1218                 pgdat->kswapd
1219                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1220         total_memory = nr_free_pagecache_pages();
1221         hotcpu_notifier(cpu_callback, 0);
1222         return 0;
1223 }
1224
1225 module_init(kswapd_init)