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