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/blkdev.h>
  25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  26                                         buffer_heads_over_limit */
  27 #include <linux/mm_inline.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/rmap.h>
  31 #include <linux/topology.h>
  32 #include <linux/cpu.h>
  33 #include <linux/cpuset.h>
  34 #include <linux/notifier.h>
  35 #include <linux/rwsem.h>
  36
  37 #include <asm/tlbflush.h>
  38 #include <asm/div64.h>
  39
  40 #include <linux/swapops.h>
  41
  42 /* possible outcome of pageout() */
  43 typedef enum {
  44         /* failed to write page out, page is locked */
  45         PAGE_KEEP,
  46         /* move page to the active list, page is locked */
  47         PAGE_ACTIVATE,
  48         /* page has been sent to the disk successfully, page is unlocked */
  49         PAGE_SUCCESS,
  50         /* page is clean and locked */
  51         PAGE_CLEAN,
  52 } pageout_t;
  53
  54 struct scan_control {
  55         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  56         unsigned long nr_to_scan;
  57
  58         /* Incremented by the number of inactive pages that were scanned */
  59         unsigned long nr_scanned;
  60
  61         /* Incremented by the number of pages reclaimed */
  62         unsigned long nr_reclaimed;
  63
  64         unsigned long nr_mapped;        /* From page_state */
  65
  66         /* How many pages shrink_cache() should reclaim */
  67         int nr_to_reclaim;
  68
  69         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  70         unsigned int priority;
  71
  72         /* This context's GFP mask */
  73         gfp_t gfp_mask;
  74
  75         int may_writepage;
  76
  77         /* Can pages be swapped as part of reclaim? */
  78         int may_swap;
  79
  80         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  81          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  82          * In this context, it doesn't matter that we scan the
  83          * whole list at once. */
  84         int swap_cluster_max;
  85 };
  86
  87 /*
  88  * The list of shrinker callbacks used by to apply pressure to
  89  * ageable caches.
  90  */
  91 struct shrinker {
  92         shrinker_t              shrinker;
  93         struct list_head        list;
  94         int                     seeks;  /* seeks to recreate an obj */
  95         long                    nr;     /* objs pending delete */
  96 };
  97
  98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  99
 100 #ifdef ARCH_HAS_PREFETCH
 101 #define prefetch_prev_lru_page(_page, _base, _field)                    \
 102         do {                                                            \
 103                 if ((_page)->lru.prev != _base) {                       \
 104                         struct page *prev;                              \
 105                                                                         \
 106                         prev = lru_to_page(&(_page->lru));              \
 107                         prefetch(&prev->_field);                        \
 108                 }                                                       \
 109         } while (0)
 110 #else
 111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 112 #endif
 113
 114 #ifdef ARCH_HAS_PREFETCHW
 115 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 116         do {                                                            \
 117                 if ((_page)->lru.prev != _base) {                       \
 118                         struct page *prev;                              \
 119                                                                         \
 120                         prev = lru_to_page(&(_page->lru));              \
 121                         prefetchw(&prev->_field);                       \
 122                 }                                                       \
 123         } while (0)
 124 #else
 125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 126 #endif
 127
 128 /*
 129  * From 0 .. 100.  Higher means more swappy.
 130  */
 131 int vm_swappiness = 60;
 132 static long total_memory;
 133
 134 static LIST_HEAD(shrinker_list);
 135 static DECLARE_RWSEM(shrinker_rwsem);
 136
 137 /*
 138  * Add a shrinker callback to be called from the vm
 139  */
 140 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 141 {
 142         struct shrinker *shrinker;
 143
 144         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 145         if (shrinker) {
 146                 shrinker->shrinker = theshrinker;
 147                 shrinker->seeks = seeks;
 148                 shrinker->nr = 0;
 149                 down_write(&shrinker_rwsem);
 150                 list_add_tail(&shrinker->list, &shrinker_list);
 151                 up_write(&shrinker_rwsem);
 152         }
 153         return shrinker;
 154 }
 155 EXPORT_SYMBOL(set_shrinker);
 156
 157 /*
 158  * Remove one
 159  */
 160 void remove_shrinker(struct shrinker *shrinker)
 161 {
 162         down_write(&shrinker_rwsem);
 163         list_del(&shrinker->list);
 164         up_write(&shrinker_rwsem);
 165         kfree(shrinker);
 166 }
 167 EXPORT_SYMBOL(remove_shrinker);
 168
 169 #define SHRINK_BATCH 128
 170 /*
 171  * Call the shrink functions to age shrinkable caches
 172  *
 173  * Here we assume it costs one seek to replace a lru page and that it also
 174  * takes a seek to recreate a cache object.  With this in mind we age equal
 175  * percentages of the lru and ageable caches.  This should balance the seeks
 176  * generated by these structures.
 177  *
 178  * If the vm encounted mapped pages on the LRU it increase the pressure on
 179  * slab to avoid swapping.
 180  *
 181  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 182  *
 183  * `lru_pages' represents the number of on-LRU pages in all the zones which
 184  * are eligible for the caller's allocation attempt.  It is used for balancing
 185  * slab reclaim versus page reclaim.
 186  *
 187  * Returns the number of slab objects which we shrunk.
 188  */
 189 static int shrink_slab(unsigned long scanned, gfp_t gfp_mask,
 190                         unsigned long lru_pages)
 191 {
 192         struct shrinker *shrinker;
 193         int ret = 0;
 194
 195         if (scanned == 0)
 196                 scanned = SWAP_CLUSTER_MAX;
 197
 198         if (!down_read_trylock(&shrinker_rwsem))
 199                 return 1;       /* Assume we'll be able to shrink next time */
 200
 201         list_for_each_entry(shrinker, &shrinker_list, list) {
 202                 unsigned long long delta;
 203                 unsigned long total_scan;
 204
 205                 delta = (4 * scanned) / shrinker->seeks;
 206                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 207                 do_div(delta, lru_pages + 1);
 208                 shrinker->nr += delta;
 209                 if (shrinker->nr < 0)
 210                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 211
 212                 total_scan = shrinker->nr;
 213                 shrinker->nr = 0;
 214
 215                 while (total_scan >= SHRINK_BATCH) {
 216                         long this_scan = SHRINK_BATCH;
 217                         int shrink_ret;
 218                         int nr_before;
 219
 220                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
 221                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 222                         if (shrink_ret == -1)
 223                                 break;
 224                         if (shrink_ret < nr_before)
 225                                 ret += nr_before - shrink_ret;
 226                         mod_page_state(slabs_scanned, this_scan);
 227                         total_scan -= this_scan;
 228
 229                         cond_resched();
 230                 }
 231
 232                 shrinker->nr += total_scan;
 233         }
 234         up_read(&shrinker_rwsem);
 235         return ret;
 236 }
 237
 238 /* Called without lock on whether page is mapped, so answer is unstable */
 239 static inline int page_mapping_inuse(struct page *page)
 240 {
 241         struct address_space *mapping;
 242
 243         /* Page is in somebody's page tables. */
 244         if (page_mapped(page))
 245                 return 1;
 246
 247         /* Be more reluctant to reclaim swapcache than pagecache */
 248         if (PageSwapCache(page))
 249                 return 1;
 250
 251         mapping = page_mapping(page);
 252         if (!mapping)
 253                 return 0;
 254
 255         /* File is mmap'd by somebody? */
 256         return mapping_mapped(mapping);
 257 }
 258
 259 static inline int is_page_cache_freeable(struct page *page)
 260 {
 261         return page_count(page) - !!PagePrivate(page) == 2;
 262 }
 263
 264 static int may_write_to_queue(struct backing_dev_info *bdi)
 265 {
 266         if (current_is_kswapd())
 267                 return 1;
 268         if (current_is_pdflush())       /* This is unlikely, but why not... */
 269                 return 1;
 270         if (!bdi_write_congested(bdi))
 271                 return 1;
 272         if (bdi == current->backing_dev_info)
 273                 return 1;
 274         return 0;
 275 }
 276
 277 /*
 278  * We detected a synchronous write error writing a page out.  Probably
 279  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 280  * fsync(), msync() or close().
 281  *
 282  * The tricky part is that after writepage we cannot touch the mapping: nothing
 283  * prevents it from being freed up.  But we have a ref on the page and once
 284  * that page is locked, the mapping is pinned.
 285  *
 286  * We're allowed to run sleeping lock_page() here because we know the caller has
 287  * __GFP_FS.
 288  */
 289 static void handle_write_error(struct address_space *mapping,
 290                                 struct page *page, int error)
 291 {
 292         lock_page(page);
 293         if (page_mapping(page) == mapping) {
 294                 if (error == -ENOSPC)
 295                         set_bit(AS_ENOSPC, &mapping->flags);
 296                 else
 297                         set_bit(AS_EIO, &mapping->flags);
 298         }
 299         unlock_page(page);
 300 }
 301
 302 /*
 303  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 304  */
 305 static pageout_t pageout(struct page *page, struct address_space *mapping)
 306 {
 307         /*
 308          * If the page is dirty, only perform writeback if that write
 309          * will be non-blocking.  To prevent this allocation from being
 310          * stalled by pagecache activity.  But note that there may be
 311          * stalls if we need to run get_block().  We could test
 312          * PagePrivate for that.
 313          *
 314          * If this process is currently in generic_file_write() against
 315          * this page's queue, we can perform writeback even if that
 316          * will block.
 317          *
 318          * If the page is swapcache, write it back even if that would
 319          * block, for some throttling. This happens by accident, because
 320          * swap_backing_dev_info is bust: it doesn't reflect the
 321          * congestion state of the swapdevs.  Easy to fix, if needed.
 322          * See swapfile.c:page_queue_congested().
 323          */
 324         if (!is_page_cache_freeable(page))
 325                 return PAGE_KEEP;
 326         if (!mapping) {
 327                 /*
 328                  * Some data journaling orphaned pages can have
 329                  * page->mapping == NULL while being dirty with clean buffers.
 330                  */
 331                 if (PagePrivate(page)) {
 332                         if (try_to_free_buffers(page)) {
 333                                 ClearPageDirty(page);
 334                                 printk("%s: orphaned page\n", __FUNCTION__);
 335                                 return PAGE_CLEAN;
 336                         }
 337                 }
 338                 return PAGE_KEEP;
 339         }
 340         if (mapping->a_ops->writepage == NULL)
 341                 return PAGE_ACTIVATE;
 342         if (!may_write_to_queue(mapping->backing_dev_info))
 343                 return PAGE_KEEP;
 344
 345         if (clear_page_dirty_for_io(page)) {
 346                 int res;
 347                 struct writeback_control wbc = {
 348                         .sync_mode = WB_SYNC_NONE,
 349                         .nr_to_write = SWAP_CLUSTER_MAX,
 350                         .nonblocking = 1,
 351                         .for_reclaim = 1,
 352                 };
 353
 354                 SetPageReclaim(page);
 355                 res = mapping->a_ops->writepage(page, &wbc);
 356                 if (res < 0)
 357                         handle_write_error(mapping, page, res);
 358                 if (res == WRITEPAGE_ACTIVATE) {
 359                         ClearPageReclaim(page);
 360                         return PAGE_ACTIVATE;
 361                 }
 362                 if (!PageWriteback(page)) {
 363                         /* synchronous write or broken a_ops? */
 364                         ClearPageReclaim(page);
 365                 }
 366
 367                 return PAGE_SUCCESS;
 368         }
 369
 370         return PAGE_CLEAN;
 371 }
 372
 373 /*
 374  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 375  */
 376 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 377 {
 378         LIST_HEAD(ret_pages);
 379         struct pagevec freed_pvec;
 380         int pgactivate = 0;
 381         int reclaimed = 0;
 382
 383         cond_resched();
 384
 385         pagevec_init(&freed_pvec, 1);
 386         while (!list_empty(page_list)) {
 387                 struct address_space *mapping;
 388                 struct page *page;
 389                 int may_enter_fs;
 390                 int referenced;
 391
 392                 cond_resched();
 393
 394                 page = lru_to_page(page_list);
 395                 list_del(&page->lru);
 396
 397                 if (TestSetPageLocked(page))
 398                         goto keep;
 399
 400                 BUG_ON(PageActive(page));
 401
 402                 sc->nr_scanned++;
 403                 /* Double the slab pressure for mapped and swapcache pages */
 404                 if (page_mapped(page) || PageSwapCache(page))
 405                         sc->nr_scanned++;
 406
 407                 if (PageWriteback(page))
 408                         goto keep_locked;
 409
 410                 referenced = page_referenced(page, 1, sc->priority <= 0);
 411                 /* In active use or really unfreeable?  Activate it. */
 412                 if (referenced && page_mapping_inuse(page))
 413                         goto activate_locked;
 414
 415 #ifdef CONFIG_SWAP
 416                 /*
 417                  * Anonymous process memory has backing store?
 418                  * Try to allocate it some swap space here.
 419                  */
 420                 if (PageAnon(page) && !PageSwapCache(page) && sc->may_swap) {
 421                         if (!add_to_swap(page))
 422                                 goto activate_locked;
 423                 }
 424 #endif /* CONFIG_SWAP */
 425
 426                 mapping = page_mapping(page);
 427                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 428                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 429
 430                 /*
 431                  * The page is mapped into the page tables of one or more
 432                  * processes. Try to unmap it here.
 433                  */
 434                 if (page_mapped(page) && mapping) {
 435                         switch (try_to_unmap(page)) {
 436                         case SWAP_FAIL:
 437                                 goto activate_locked;
 438                         case SWAP_AGAIN:
 439                                 goto keep_locked;
 440                         case SWAP_SUCCESS:
 441                                 ; /* try to free the page below */
 442                         }
 443                 }
 444
 445                 if (PageDirty(page)) {
 446                         if (referenced)
 447                                 goto keep_locked;
 448                         if (!may_enter_fs)
 449                                 goto keep_locked;
 450                         if (laptop_mode && !sc->may_writepage)
 451                                 goto keep_locked;
 452
 453                         /* Page is dirty, try to write it out here */
 454                         switch(pageout(page, mapping)) {
 455                         case PAGE_KEEP:
 456                                 goto keep_locked;
 457                         case PAGE_ACTIVATE:
 458                                 goto activate_locked;
 459                         case PAGE_SUCCESS:
 460                                 if (PageWriteback(page) || PageDirty(page))
 461                                         goto keep;
 462                                 /*
 463                                  * A synchronous write - probably a ramdisk.  Go
 464                                  * ahead and try to reclaim the page.
 465                                  */
 466                                 if (TestSetPageLocked(page))
 467                                         goto keep;
 468                                 if (PageDirty(page) || PageWriteback(page))
 469                                         goto keep_locked;
 470                                 mapping = page_mapping(page);
 471                         case PAGE_CLEAN:
 472                                 ; /* try to free the page below */
 473                         }
 474                 }
 475
 476                 /*
 477                  * If the page has buffers, try to free the buffer mappings
 478                  * associated with this page. If we succeed we try to free
 479                  * the page as well.
 480                  *
 481                  * We do this even if the page is PageDirty().
 482                  * try_to_release_page() does not perform I/O, but it is
 483                  * possible for a page to have PageDirty set, but it is actually
 484                  * clean (all its buffers are clean).  This happens if the
 485                  * buffers were written out directly, with submit_bh(). ext3
 486                  * will do this, as well as the blockdev mapping.
 487                  * try_to_release_page() will discover that cleanness and will
 488                  * drop the buffers and mark the page clean - it can be freed.
 489                  *
 490                  * Rarely, pages can have buffers and no ->mapping.  These are
 491                  * the pages which were not successfully invalidated in
 492                  * truncate_complete_page().  We try to drop those buffers here
 493                  * and if that worked, and the page is no longer mapped into
 494                  * process address space (page_count == 1) it can be freed.
 495                  * Otherwise, leave the page on the LRU so it is swappable.
 496                  */
 497                 if (PagePrivate(page)) {
 498                         if (!try_to_release_page(page, sc->gfp_mask))
 499                                 goto activate_locked;
 500                         if (!mapping && page_count(page) == 1)
 501                                 goto free_it;
 502                 }
 503
 504                 if (!mapping)
 505                         goto keep_locked;       /* truncate got there first */
 506
 507                 write_lock_irq(&mapping->tree_lock);
 508
 509                 /*
 510                  * The non-racy check for busy page.  It is critical to check
 511                  * PageDirty _after_ making sure that the page is freeable and
 512                  * not in use by anybody.       (pagecache + us == 2)
 513                  */
 514                 if (unlikely(page_count(page) != 2))
 515                         goto cannot_free;
 516                 smp_rmb();
 517                 if (unlikely(PageDirty(page)))
 518                         goto cannot_free;
 519
 520 #ifdef CONFIG_SWAP
 521                 if (PageSwapCache(page)) {
 522                         swp_entry_t swap = { .val = page->private };
 523                         __delete_from_swap_cache(page);
 524                         write_unlock_irq(&mapping->tree_lock);
 525                         swap_free(swap);
 526                         __put_page(page);       /* The pagecache ref */
 527                         goto free_it;
 528                 }
 529 #endif /* CONFIG_SWAP */
 530
 531                 __remove_from_page_cache(page);
 532                 write_unlock_irq(&mapping->tree_lock);
 533                 __put_page(page);
 534
 535 free_it:
 536                 unlock_page(page);
 537                 reclaimed++;
 538                 if (!pagevec_add(&freed_pvec, page))
 539                         __pagevec_release_nonlru(&freed_pvec);
 540                 continue;
 541
 542 cannot_free:
 543                 write_unlock_irq(&mapping->tree_lock);
 544                 goto keep_locked;
 545
 546 activate_locked:
 547                 SetPageActive(page);
 548                 pgactivate++;
 549 keep_locked:
 550                 unlock_page(page);
 551 keep:
 552                 list_add(&page->lru, &ret_pages);
 553                 BUG_ON(PageLRU(page));
 554         }
 555         list_splice(&ret_pages, page_list);
 556         if (pagevec_count(&freed_pvec))
 557                 __pagevec_release_nonlru(&freed_pvec);
 558         mod_page_state(pgactivate, pgactivate);
 559         sc->nr_reclaimed += reclaimed;
 560         return reclaimed;
 561 }
 562
 563 /*
 564  * zone->lru_lock is heavily contended.  Some of the functions that
 565  * shrink the lists perform better by taking out a batch of pages
 566  * and working on them outside the LRU lock.
 567  *
 568  * For pagecache intensive workloads, this function is the hottest
 569  * spot in the kernel (apart from copy_*_user functions).
 570  *
 571  * Appropriate locks must be held before calling this function.
 572  *
 573  * @nr_to_scan: The number of pages to look through on the list.
 574  * @src:        The LRU list to pull pages off.
 575  * @dst:        The temp list to put pages on to.
 576  * @scanned:    The number of pages that were scanned.
 577  *
 578  * returns how many pages were moved onto *@dst.
 579  */
 580 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
 581                              struct list_head *dst, int *scanned)
 582 {
 583         int nr_taken = 0;
 584         struct page *page;
 585         int scan = 0;
 586
 587         while (scan++ < nr_to_scan && !list_empty(src)) {
 588                 page = lru_to_page(src);
 589                 prefetchw_prev_lru_page(page, src, flags);
 590
 591                 if (!TestClearPageLRU(page))
 592                         BUG();
 593                 list_del(&page->lru);
 594                 if (get_page_testone(page)) {
 595                         /*
 596                          * It is being freed elsewhere
 597                          */
 598                         __put_page(page);
 599                         SetPageLRU(page);
 600                         list_add(&page->lru, src);
 601                         continue;
 602                 } else {
 603                         list_add(&page->lru, dst);
 604                         nr_taken++;
 605                 }
 606         }
 607
 608         *scanned = scan;
 609         return nr_taken;
 610 }
 611
 612 /*
 613  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 614  */
 615 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 616 {
 617         LIST_HEAD(page_list);
 618         struct pagevec pvec;
 619         int max_scan = sc->nr_to_scan;
 620
 621         pagevec_init(&pvec, 1);
 622
 623         lru_add_drain();
 624         spin_lock_irq(&zone->lru_lock);
 625         while (max_scan > 0) {
 626                 struct page *page;
 627                 int nr_taken;
 628                 int nr_scan;
 629                 int nr_freed;
 630
 631                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
 632                                              &zone->inactive_list,
 633                                              &page_list, &nr_scan);
 634                 zone->nr_inactive -= nr_taken;
 635                 zone->pages_scanned += nr_scan;
 636                 spin_unlock_irq(&zone->lru_lock);
 637
 638                 if (nr_taken == 0)
 639                         goto done;
 640
 641                 max_scan -= nr_scan;
 642                 if (current_is_kswapd())
 643                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 644                 else
 645                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 646                 nr_freed = shrink_list(&page_list, sc);
 647                 if (current_is_kswapd())
 648                         mod_page_state(kswapd_steal, nr_freed);
 649                 mod_page_state_zone(zone, pgsteal, nr_freed);
 650                 sc->nr_to_reclaim -= nr_freed;
 651
 652                 spin_lock_irq(&zone->lru_lock);
 653                 /*
 654                  * Put back any unfreeable pages.
 655                  */
 656                 while (!list_empty(&page_list)) {
 657                         page = lru_to_page(&page_list);
 658                         if (TestSetPageLRU(page))
 659                                 BUG();
 660                         list_del(&page->lru);
 661                         if (PageActive(page))
 662                                 add_page_to_active_list(zone, page);
 663                         else
 664                                 add_page_to_inactive_list(zone, page);
 665                         if (!pagevec_add(&pvec, page)) {
 666                                 spin_unlock_irq(&zone->lru_lock);
 667                                 __pagevec_release(&pvec);
 668                                 spin_lock_irq(&zone->lru_lock);
 669                         }
 670                 }
 671         }
 672         spin_unlock_irq(&zone->lru_lock);
 673 done:
 674         pagevec_release(&pvec);
 675 }
 676
 677 /*
 678  * This moves pages from the active list to the inactive list.
 679  *
 680  * We move them the other way if the page is referenced by one or more
 681  * processes, from rmap.
 682  *
 683  * If the pages are mostly unmapped, the processing is fast and it is
 684  * appropriate to hold zone->lru_lock across the whole operation.  But if
 685  * the pages are mapped, the processing is slow (page_referenced()) so we
 686  * should drop zone->lru_lock around each page.  It's impossible to balance
 687  * this, so instead we remove the pages from the LRU while processing them.
 688  * It is safe to rely on PG_active against the non-LRU pages in here because
 689  * nobody will play with that bit on a non-LRU page.
 690  *
 691  * The downside is that we have to touch page->_count against each page.
 692  * But we had to alter page->flags anyway.
 693  */
 694 static void
 695 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 696 {
 697         int pgmoved;
 698         int pgdeactivate = 0;
 699         int pgscanned;
 700         int nr_pages = sc->nr_to_scan;
 701         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 702         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 703         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 704         struct page *page;
 705         struct pagevec pvec;
 706         int reclaim_mapped = 0;
 707         long mapped_ratio;
 708         long distress;
 709         long swap_tendency;
 710
 711         lru_add_drain();
 712         spin_lock_irq(&zone->lru_lock);
 713         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
 714                                     &l_hold, &pgscanned);
 715         zone->pages_scanned += pgscanned;
 716         zone->nr_active -= pgmoved;
 717         spin_unlock_irq(&zone->lru_lock);
 718
 719         /*
 720          * `distress' is a measure of how much trouble we're having reclaiming
 721          * pages.  0 -> no problems.  100 -> great trouble.
 722          */
 723         distress = 100 >> zone->prev_priority;
 724
 725         /*
 726          * The point of this algorithm is to decide when to start reclaiming
 727          * mapped memory instead of just pagecache.  Work out how much memory
 728          * is mapped.
 729          */
 730         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 731
 732         /*
 733          * Now decide how much we really want to unmap some pages.  The mapped
 734          * ratio is downgraded - just because there's a lot of mapped memory
 735          * doesn't necessarily mean that page reclaim isn't succeeding.
 736          *
 737          * The distress ratio is important - we don't want to start going oom.
 738          *
 739          * A 100% value of vm_swappiness overrides this algorithm altogether.
 740          */
 741         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 742
 743         /*
 744          * Now use this metric to decide whether to start moving mapped memory
 745          * onto the inactive list.
 746          */
 747         if (swap_tendency >= 100)
 748                 reclaim_mapped = 1;
 749
 750         while (!list_empty(&l_hold)) {
 751                 cond_resched();
 752                 page = lru_to_page(&l_hold);
 753                 list_del(&page->lru);
 754                 if (page_mapped(page)) {
 755                         if (!reclaim_mapped ||
 756                             (total_swap_pages == 0 && PageAnon(page)) ||
 757                             page_referenced(page, 0, sc->priority <= 0)) {
 758                                 list_add(&page->lru, &l_active);
 759                                 continue;
 760                         }
 761                 }
 762                 list_add(&page->lru, &l_inactive);
 763         }
 764
 765         pagevec_init(&pvec, 1);
 766         pgmoved = 0;
 767         spin_lock_irq(&zone->lru_lock);
 768         while (!list_empty(&l_inactive)) {
 769                 page = lru_to_page(&l_inactive);
 770                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 771                 if (TestSetPageLRU(page))
 772                         BUG();
 773                 if (!TestClearPageActive(page))
 774                         BUG();
 775                 list_move(&page->lru, &zone->inactive_list);
 776                 pgmoved++;
 777                 if (!pagevec_add(&pvec, page)) {
 778                         zone->nr_inactive += pgmoved;
 779                         spin_unlock_irq(&zone->lru_lock);
 780                         pgdeactivate += pgmoved;
 781                         pgmoved = 0;
 782                         if (buffer_heads_over_limit)
 783                                 pagevec_strip(&pvec);
 784                         __pagevec_release(&pvec);
 785                         spin_lock_irq(&zone->lru_lock);
 786                 }
 787         }
 788         zone->nr_inactive += pgmoved;
 789         pgdeactivate += pgmoved;
 790         if (buffer_heads_over_limit) {
 791                 spin_unlock_irq(&zone->lru_lock);
 792                 pagevec_strip(&pvec);
 793                 spin_lock_irq(&zone->lru_lock);
 794         }
 795
 796         pgmoved = 0;
 797         while (!list_empty(&l_active)) {
 798                 page = lru_to_page(&l_active);
 799                 prefetchw_prev_lru_page(page, &l_active, flags);
 800                 if (TestSetPageLRU(page))
 801                         BUG();
 802                 BUG_ON(!PageActive(page));
 803                 list_move(&page->lru, &zone->active_list);
 804                 pgmoved++;
 805                 if (!pagevec_add(&pvec, page)) {
 806                         zone->nr_active += pgmoved;
 807                         pgmoved = 0;
 808                         spin_unlock_irq(&zone->lru_lock);
 809                         __pagevec_release(&pvec);
 810                         spin_lock_irq(&zone->lru_lock);
 811                 }
 812         }
 813         zone->nr_active += pgmoved;
 814         spin_unlock_irq(&zone->lru_lock);
 815         pagevec_release(&pvec);
 816
 817         mod_page_state_zone(zone, pgrefill, pgscanned);
 818         mod_page_state(pgdeactivate, pgdeactivate);
 819 }
 820
 821 /*
 822  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 823  */
 824 static void
 825 shrink_zone(struct zone *zone, struct scan_control *sc)
 826 {
 827         unsigned long nr_active;
 828         unsigned long nr_inactive;
 829
 830         atomic_inc(&zone->reclaim_in_progress);
 831
 832         /*
 833          * Add one to `nr_to_scan' just to make sure that the kernel will
 834          * slowly sift through the active list.
 835          */
 836         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 837         nr_active = zone->nr_scan_active;
 838         if (nr_active >= sc->swap_cluster_max)
 839                 zone->nr_scan_active = 0;
 840         else
 841                 nr_active = 0;
 842
 843         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 844         nr_inactive = zone->nr_scan_inactive;
 845         if (nr_inactive >= sc->swap_cluster_max)
 846                 zone->nr_scan_inactive = 0;
 847         else
 848                 nr_inactive = 0;
 849
 850         sc->nr_to_reclaim = sc->swap_cluster_max;
 851
 852         while (nr_active || nr_inactive) {
 853                 if (nr_active) {
 854                         sc->nr_to_scan = min(nr_active,
 855                                         (unsigned long)sc->swap_cluster_max);
 856                         nr_active -= sc->nr_to_scan;
 857                         refill_inactive_zone(zone, sc);
 858                 }
 859
 860                 if (nr_inactive) {
 861                         sc->nr_to_scan = min(nr_inactive,
 862                                         (unsigned long)sc->swap_cluster_max);
 863                         nr_inactive -= sc->nr_to_scan;
 864                         shrink_cache(zone, sc);
 865                         if (sc->nr_to_reclaim <= 0)
 866                                 break;
 867                 }
 868         }
 869
 870         throttle_vm_writeout();
 871
 872         atomic_dec(&zone->reclaim_in_progress);
 873 }
 874
 875 /*
 876  * This is the direct reclaim path, for page-allocating processes.  We only
 877  * try to reclaim pages from zones which will satisfy the caller's allocation
 878  * request.
 879  *
 880  * We reclaim from a zone even if that zone is over pages_high.  Because:
 881  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 882  *    allocation or
 883  * b) The zones may be over pages_high but they must go *over* pages_high to
 884  *    satisfy the `incremental min' zone defense algorithm.
 885  *
 886  * Returns the number of reclaimed pages.
 887  *
 888  * If a zone is deemed to be full of pinned pages then just give it a light
 889  * scan then give up on it.
 890  */
 891 static void
 892 shrink_caches(struct zone **zones, struct scan_control *sc)
 893 {
 894         int i;
 895
 896         for (i = 0; zones[i] != NULL; i++) {
 897                 struct zone *zone = zones[i];
 898
 899                 if (zone->present_pages == 0)
 900                         continue;
 901
 902                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 903                         continue;
 904
 905                 zone->temp_priority = sc->priority;
 906                 if (zone->prev_priority > sc->priority)
 907                         zone->prev_priority = sc->priority;
 908
 909                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 910                         continue;       /* Let kswapd poll it */
 911
 912                 shrink_zone(zone, sc);
 913         }
 914 }
 915
 916 /*
 917  * This is the main entry point to direct page reclaim.
 918  *
 919  * If a full scan of the inactive list fails to free enough memory then we
 920  * are "out of memory" and something needs to be killed.
 921  *
 922  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 923  * high - the zone may be full of dirty or under-writeback pages, which this
 924  * caller can't do much about.  We kick pdflush and take explicit naps in the
 925  * hope that some of these pages can be written.  But if the allocating task
 926  * holds filesystem locks which prevent writeout this might not work, and the
 927  * allocation attempt will fail.
 928  */
 929 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
 930 {
 931         int priority;
 932         int ret = 0;
 933         int total_scanned = 0, total_reclaimed = 0;
 934         struct reclaim_state *reclaim_state = current->reclaim_state;
 935         struct scan_control sc;
 936         unsigned long lru_pages = 0;
 937         int i;
 938
 939         sc.gfp_mask = gfp_mask;
 940         sc.may_writepage = 0;
 941         sc.may_swap = 1;
 942
 943         inc_page_state(allocstall);
 944
 945         for (i = 0; zones[i] != NULL; i++) {
 946                 struct zone *zone = zones[i];
 947
 948                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 949                         continue;
 950
 951                 zone->temp_priority = DEF_PRIORITY;
 952                 lru_pages += zone->nr_active + zone->nr_inactive;
 953         }
 954
 955         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 956                 sc.nr_mapped = read_page_state(nr_mapped);
 957                 sc.nr_scanned = 0;
 958                 sc.nr_reclaimed = 0;
 959                 sc.priority = priority;
 960                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
 961                 shrink_caches(zones, &sc);
 962                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
 963                 if (reclaim_state) {
 964                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
 965                         reclaim_state->reclaimed_slab = 0;
 966                 }
 967                 total_scanned += sc.nr_scanned;
 968                 total_reclaimed += sc.nr_reclaimed;
 969                 if (total_reclaimed >= sc.swap_cluster_max) {
 970                         ret = 1;
 971                         goto out;
 972                 }
 973
 974                 /*
 975                  * Try to write back as many pages as we just scanned.  This
 976                  * tends to cause slow streaming writers to write data to the
 977                  * disk smoothly, at the dirtying rate, which is nice.   But
 978                  * that's undesirable in laptop mode, where we *want* lumpy
 979                  * writeout.  So in laptop mode, write out the whole world.
 980                  */
 981                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
 982                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
 983                         sc.may_writepage = 1;
 984                 }
 985
 986                 /* Take a nap, wait for some writeback to complete */
 987                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
 988                         blk_congestion_wait(WRITE, HZ/10);
 989         }
 990 out:
 991         for (i = 0; zones[i] != 0; i++) {
 992                 struct zone *zone = zones[i];
 993
 994                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 995                         continue;
 996
 997                 zone->prev_priority = zone->temp_priority;
 998         }
 999         return ret;
1000 }
1001
1002 /*
1003  * For kswapd, balance_pgdat() will work across all this node's zones until
1004  * they are all at pages_high.
1005  *
1006  * If `nr_pages' is non-zero then it is the number of pages which are to be
1007  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1008  * special.
1009  *
1010  * Returns the number of pages which were actually freed.
1011  *
1012  * There is special handling here for zones which are full of pinned pages.
1013  * This can happen if the pages are all mlocked, or if they are all used by
1014  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1015  * What we do is to detect the case where all pages in the zone have been
1016  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1017  * dead and from now on, only perform a short scan.  Basically we're polling
1018  * the zone for when the problem goes away.
1019  *
1020  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1021  * zones which have free_pages > pages_high, but once a zone is found to have
1022  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1023  * of the number of free pages in the lower zones.  This interoperates with
1024  * the page allocator fallback scheme to ensure that aging of pages is balanced
1025  * across the zones.
1026  */
1027 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1028 {
1029         int to_free = nr_pages;
1030         int all_zones_ok;
1031         int priority;
1032         int i;
1033         int total_scanned, total_reclaimed;
1034         struct reclaim_state *reclaim_state = current->reclaim_state;
1035         struct scan_control sc;
1036
1037 loop_again:
1038         total_scanned = 0;
1039         total_reclaimed = 0;
1040         sc.gfp_mask = GFP_KERNEL;
1041         sc.may_writepage = 0;
1042         sc.may_swap = 1;
1043         sc.nr_mapped = read_page_state(nr_mapped);
1044
1045         inc_page_state(pageoutrun);
1046
1047         for (i = 0; i < pgdat->nr_zones; i++) {
1048                 struct zone *zone = pgdat->node_zones + i;
1049
1050                 zone->temp_priority = DEF_PRIORITY;
1051         }
1052
1053         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1054                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1055                 unsigned long lru_pages = 0;
1056
1057                 all_zones_ok = 1;
1058
1059                 if (nr_pages == 0) {
1060                         /*
1061                          * Scan in the highmem->dma direction for the highest
1062                          * zone which needs scanning
1063                          */
1064                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1065                                 struct zone *zone = pgdat->node_zones + i;
1066
1067                                 if (zone->present_pages == 0)
1068                                         continue;
1069
1070                                 if (zone->all_unreclaimable &&
1071                                                 priority != DEF_PRIORITY)
1072                                         continue;
1073
1074                                 if (!zone_watermark_ok(zone, order,
1075                                                 zone->pages_high, 0, 0, 0)) {
1076                                         end_zone = i;
1077                                         goto scan;
1078                                 }
1079                         }
1080                         goto out;
1081                 } else {
1082                         end_zone = pgdat->nr_zones - 1;
1083                 }
1084 scan:
1085                 for (i = 0; i <= end_zone; i++) {
1086                         struct zone *zone = pgdat->node_zones + i;
1087
1088                         lru_pages += zone->nr_active + zone->nr_inactive;
1089                 }
1090
1091                 /*
1092                  * Now scan the zone in the dma->highmem direction, stopping
1093                  * at the last zone which needs scanning.
1094                  *
1095                  * We do this because the page allocator works in the opposite
1096                  * direction.  This prevents the page allocator from allocating
1097                  * pages behind kswapd's direction of progress, which would
1098                  * cause too much scanning of the lower zones.
1099                  */
1100                 for (i = 0; i <= end_zone; i++) {
1101                         struct zone *zone = pgdat->node_zones + i;
1102                         int nr_slab;
1103
1104                         if (zone->present_pages == 0)
1105                                 continue;
1106
1107                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1108                                 continue;
1109
1110                         if (nr_pages == 0) {    /* Not software suspend */
1111                                 if (!zone_watermark_ok(zone, order,
1112                                                 zone->pages_high, end_zone, 0, 0))
1113                                         all_zones_ok = 0;
1114                         }
1115                         zone->temp_priority = priority;
1116                         if (zone->prev_priority > priority)
1117                                 zone->prev_priority = priority;
1118                         sc.nr_scanned = 0;
1119                         sc.nr_reclaimed = 0;
1120                         sc.priority = priority;
1121                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1122                         atomic_inc(&zone->reclaim_in_progress);
1123                         shrink_zone(zone, &sc);
1124                         atomic_dec(&zone->reclaim_in_progress);
1125                         reclaim_state->reclaimed_slab = 0;
1126                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1127                                                 lru_pages);
1128                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1129                         total_reclaimed += sc.nr_reclaimed;
1130                         total_scanned += sc.nr_scanned;
1131                         if (zone->all_unreclaimable)
1132                                 continue;
1133                         if (nr_slab == 0 && zone->pages_scanned >=
1134                                     (zone->nr_active + zone->nr_inactive) * 4)
1135                                 zone->all_unreclaimable = 1;
1136                         /*
1137                          * If we've done a decent amount of scanning and
1138                          * the reclaim ratio is low, start doing writepage
1139                          * even in laptop mode
1140                          */
1141                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1142                             total_scanned > total_reclaimed+total_reclaimed/2)
1143                                 sc.may_writepage = 1;
1144                 }
1145                 if (nr_pages && to_free > total_reclaimed)
1146                         continue;       /* swsusp: need to do more work */
1147                 if (all_zones_ok)
1148                         break;          /* kswapd: all done */
1149                 /*
1150                  * OK, kswapd is getting into trouble.  Take a nap, then take
1151                  * another pass across the zones.
1152                  */
1153                 if (total_scanned && priority < DEF_PRIORITY - 2)
1154                         blk_congestion_wait(WRITE, HZ/10);
1155
1156                 /*
1157                  * We do this so kswapd doesn't build up large priorities for
1158                  * example when it is freeing in parallel with allocators. It
1159                  * matches the direct reclaim path behaviour in terms of impact
1160                  * on zone->*_priority.
1161                  */
1162                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1163                         break;
1164         }
1165 out:
1166         for (i = 0; i < pgdat->nr_zones; i++) {
1167                 struct zone *zone = pgdat->node_zones + i;
1168
1169                 zone->prev_priority = zone->temp_priority;
1170         }
1171         if (!all_zones_ok) {
1172                 cond_resched();
1173                 goto loop_again;
1174         }
1175
1176         return total_reclaimed;
1177 }
1178
1179 /*
1180  * The background pageout daemon, started as a kernel thread
1181  * from the init process.
1182  *
1183  * This basically trickles out pages so that we have _some_
1184  * free memory available even if there is no other activity
1185  * that frees anything up. This is needed for things like routing
1186  * etc, where we otherwise might have all activity going on in
1187  * asynchronous contexts that cannot page things out.
1188  *
1189  * If there are applications that are active memory-allocators
1190  * (most normal use), this basically shouldn't matter.
1191  */
1192 static int kswapd(void *p)
1193 {
1194         unsigned long order;
1195         pg_data_t *pgdat = (pg_data_t*)p;
1196         struct task_struct *tsk = current;
1197         DEFINE_WAIT(wait);
1198         struct reclaim_state reclaim_state = {
1199                 .reclaimed_slab = 0,
1200         };
1201         cpumask_t cpumask;
1202
1203         daemonize("kswapd%d", pgdat->node_id);
1204         cpumask = node_to_cpumask(pgdat->node_id);
1205         if (!cpus_empty(cpumask))
1206                 set_cpus_allowed(tsk, cpumask);
1207         current->reclaim_state = &reclaim_state;
1208
1209         /*
1210          * Tell the memory management that we're a "memory allocator",
1211          * and that if we need more memory we should get access to it
1212          * regardless (see "__alloc_pages()"). "kswapd" should
1213          * never get caught in the normal page freeing logic.
1214          *
1215          * (Kswapd normally doesn't need memory anyway, but sometimes
1216          * you need a small amount of memory in order to be able to
1217          * page out something else, and this flag essentially protects
1218          * us from recursively trying to free more memory as we're
1219          * trying to free the first piece of memory in the first place).
1220          */
1221         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1222
1223         order = 0;
1224         for ( ; ; ) {
1225                 unsigned long new_order;
1226
1227                 try_to_freeze();
1228
1229                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1230                 new_order = pgdat->kswapd_max_order;
1231                 pgdat->kswapd_max_order = 0;
1232                 if (order < new_order) {
1233                         /*
1234                          * Don't sleep if someone wants a larger 'order'
1235                          * allocation
1236                          */
1237                         order = new_order;
1238                 } else {
1239                         schedule();
1240                         order = pgdat->kswapd_max_order;
1241                 }
1242                 finish_wait(&pgdat->kswapd_wait, &wait);
1243
1244                 balance_pgdat(pgdat, 0, order);
1245         }
1246         return 0;
1247 }
1248
1249 /*
1250  * A zone is low on free memory, so wake its kswapd task to service it.
1251  */
1252 void wakeup_kswapd(struct zone *zone, int order)
1253 {
1254         pg_data_t *pgdat;
1255
1256         if (zone->present_pages == 0)
1257                 return;
1258
1259         pgdat = zone->zone_pgdat;
1260         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0, 0))
1261                 return;
1262         if (pgdat->kswapd_max_order < order)
1263                 pgdat->kswapd_max_order = order;
1264         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1265                 return;
1266         if (!waitqueue_active(&pgdat->kswapd_wait))
1267                 return;
1268         wake_up_interruptible(&pgdat->kswapd_wait);
1269 }
1270
1271 #ifdef CONFIG_PM
1272 /*
1273  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1274  * pages.
1275  */
1276 int shrink_all_memory(int nr_pages)
1277 {
1278         pg_data_t *pgdat;
1279         int nr_to_free = nr_pages;
1280         int ret = 0;
1281         struct reclaim_state reclaim_state = {
1282                 .reclaimed_slab = 0,
1283         };
1284
1285         current->reclaim_state = &reclaim_state;
1286         for_each_pgdat(pgdat) {
1287                 int freed;
1288                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1289                 ret += freed;
1290                 nr_to_free -= freed;
1291                 if (nr_to_free <= 0)
1292                         break;
1293         }
1294         current->reclaim_state = NULL;
1295         return ret;
1296 }
1297 #endif
1298
1299 #ifdef CONFIG_HOTPLUG_CPU
1300 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1301    not required for correctness.  So if the last cpu in a node goes
1302    away, we get changed to run anywhere: as the first one comes back,
1303    restore their cpu bindings. */
1304 static int __devinit cpu_callback(struct notifier_block *nfb,
1305                                   unsigned long action,
1306                                   void *hcpu)
1307 {
1308         pg_data_t *pgdat;
1309         cpumask_t mask;
1310
1311         if (action == CPU_ONLINE) {
1312                 for_each_pgdat(pgdat) {
1313                         mask = node_to_cpumask(pgdat->node_id);
1314                         if (any_online_cpu(mask) != NR_CPUS)
1315                                 /* One of our CPUs online: restore mask */
1316                                 set_cpus_allowed(pgdat->kswapd, mask);
1317                 }
1318         }
1319         return NOTIFY_OK;
1320 }
1321 #endif /* CONFIG_HOTPLUG_CPU */
1322
1323 static int __init kswapd_init(void)
1324 {
1325         pg_data_t *pgdat;
1326         swap_setup();
1327         for_each_pgdat(pgdat)
1328                 pgdat->kswapd
1329                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1330         total_memory = nr_free_pagecache_pages();
1331         hotcpu_notifier(cpu_callback, 0);
1332         return 0;
1333 }
1334
1335 module_init(kswapd_init)
1336
1337
1338 /*
1339  * Try to free up some pages from this zone through reclaim.
1340  */
1341 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1342 {
1343         struct scan_control sc;
1344         int nr_pages = 1 << order;
1345         int total_reclaimed = 0;
1346
1347         /* The reclaim may sleep, so don't do it if sleep isn't allowed */
1348         if (!(gfp_mask & __GFP_WAIT))
1349                 return 0;
1350         if (zone->all_unreclaimable)
1351                 return 0;
1352
1353         sc.gfp_mask = gfp_mask;
1354         sc.may_writepage = 0;
1355         sc.may_swap = 0;
1356         sc.nr_mapped = read_page_state(nr_mapped);
1357         sc.nr_scanned = 0;
1358         sc.nr_reclaimed = 0;
1359         /* scan at the highest priority */
1360         sc.priority = 0;
1361
1362         if (nr_pages > SWAP_CLUSTER_MAX)
1363                 sc.swap_cluster_max = nr_pages;
1364         else
1365                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1366
1367         /* Don't reclaim the zone if there are other reclaimers active */
1368         if (atomic_read(&zone->reclaim_in_progress) > 0)
1369                 goto out;
1370
1371         shrink_zone(zone, &sc);
1372         total_reclaimed = sc.nr_reclaimed;
1373
1374  out:
1375         return total_reclaimed;
1376 }
1377
1378 asmlinkage long sys_set_zone_reclaim(unsigned int node, unsigned int zone,
1379                                      unsigned int state)
1380 {
1381         struct zone *z;
1382         int i;
1383
1384         if (!capable(CAP_SYS_ADMIN))
1385                 return -EACCES;
1386
1387         if (node >= MAX_NUMNODES || !node_online(node))
1388                 return -EINVAL;
1389
1390         /* This will break if we ever add more zones */
1391         if (!(zone & (1<<ZONE_DMA|1<<ZONE_NORMAL|1<<ZONE_HIGHMEM)))
1392                 return -EINVAL;
1393
1394         for (i = 0; i < MAX_NR_ZONES; i++) {
1395                 if (!(zone & 1<<i))
1396                         continue;
1397
1398                 z = &NODE_DATA(node)->node_zones[i];
1399
1400                 if (state)
1401                         z->reclaim_pages = 1;
1402                 else
1403                         z->reclaim_pages = 0;
1404         }
1405
1406         return 0;
1407 }