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