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         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  67         unsigned int priority;
  68
  69         /* This context's GFP mask */
  70         gfp_t gfp_mask;
  71
  72         int may_writepage;
  73
  74         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  75          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  76          * In this context, it doesn't matter that we scan the
  77          * whole list at once. */
  78         int swap_cluster_max;
  79 };
  80
  81 /*
  82  * The list of shrinker callbacks used by to apply pressure to
  83  * ageable caches.
  84  */
  85 struct shrinker {
  86         shrinker_t              shrinker;
  87         struct list_head        list;
  88         int                     seeks;  /* seeks to recreate an obj */
  89         long                    nr;     /* objs pending delete */
  90 };
  91
  92 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  93
  94 #ifdef ARCH_HAS_PREFETCH
  95 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  96         do {                                                            \
  97                 if ((_page)->lru.prev != _base) {                       \
  98                         struct page *prev;                              \
  99                                                                         \
 100                         prev = lru_to_page(&(_page->lru));              \
 101                         prefetch(&prev->_field);                        \
 102                 }                                                       \
 103         } while (0)
 104 #else
 105 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 106 #endif
 107
 108 #ifdef ARCH_HAS_PREFETCHW
 109 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 110         do {                                                            \
 111                 if ((_page)->lru.prev != _base) {                       \
 112                         struct page *prev;                              \
 113                                                                         \
 114                         prev = lru_to_page(&(_page->lru));              \
 115                         prefetchw(&prev->_field);                       \
 116                 }                                                       \
 117         } while (0)
 118 #else
 119 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 120 #endif
 121
 122 /*
 123  * From 0 .. 100.  Higher means more swappy.
 124  */
 125 int vm_swappiness = 60;
 126 static long total_memory;
 127
 128 static LIST_HEAD(shrinker_list);
 129 static DECLARE_RWSEM(shrinker_rwsem);
 130
 131 /*
 132  * Add a shrinker callback to be called from the vm
 133  */
 134 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 135 {
 136         struct shrinker *shrinker;
 137
 138         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 139         if (shrinker) {
 140                 shrinker->shrinker = theshrinker;
 141                 shrinker->seeks = seeks;
 142                 shrinker->nr = 0;
 143                 down_write(&shrinker_rwsem);
 144                 list_add_tail(&shrinker->list, &shrinker_list);
 145                 up_write(&shrinker_rwsem);
 146         }
 147         return shrinker;
 148 }
 149 EXPORT_SYMBOL(set_shrinker);
 150
 151 /*
 152  * Remove one
 153  */
 154 void remove_shrinker(struct shrinker *shrinker)
 155 {
 156         down_write(&shrinker_rwsem);
 157         list_del(&shrinker->list);
 158         up_write(&shrinker_rwsem);
 159         kfree(shrinker);
 160 }
 161 EXPORT_SYMBOL(remove_shrinker);
 162
 163 #define SHRINK_BATCH 128
 164 /*
 165  * Call the shrink functions to age shrinkable caches
 166  *
 167  * Here we assume it costs one seek to replace a lru page and that it also
 168  * takes a seek to recreate a cache object.  With this in mind we age equal
 169  * percentages of the lru and ageable caches.  This should balance the seeks
 170  * generated by these structures.
 171  *
 172  * If the vm encounted mapped pages on the LRU it increase the pressure on
 173  * slab to avoid swapping.
 174  *
 175  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 176  *
 177  * `lru_pages' represents the number of on-LRU pages in all the zones which
 178  * are eligible for the caller's allocation attempt.  It is used for balancing
 179  * slab reclaim versus page reclaim.
 180  *
 181  * Returns the number of slab objects which we shrunk.
 182  */
 183 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
 184 {
 185         struct shrinker *shrinker;
 186         int ret = 0;
 187
 188         if (scanned == 0)
 189                 scanned = SWAP_CLUSTER_MAX;
 190
 191         if (!down_read_trylock(&shrinker_rwsem))
 192                 return 1;       /* Assume we'll be able to shrink next time */
 193
 194         list_for_each_entry(shrinker, &shrinker_list, list) {
 195                 unsigned long long delta;
 196                 unsigned long total_scan;
 197                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
 198
 199                 delta = (4 * scanned) / shrinker->seeks;
 200                 delta *= max_pass;
 201                 do_div(delta, lru_pages + 1);
 202                 shrinker->nr += delta;
 203                 if (shrinker->nr < 0) {
 204                         printk(KERN_ERR "%s: nr=%ld\n",
 205                                         __FUNCTION__, shrinker->nr);
 206                         shrinker->nr = max_pass;
 207                 }
 208
 209                 /*
 210                  * Avoid risking looping forever due to too large nr value:
 211                  * never try to free more than twice the estimate number of
 212                  * freeable entries.
 213                  */
 214                 if (shrinker->nr > max_pass * 2)
 215                         shrinker->nr = max_pass * 2;
 216
 217                 total_scan = shrinker->nr;
 218                 shrinker->nr = 0;
 219
 220                 while (total_scan >= SHRINK_BATCH) {
 221                         long this_scan = SHRINK_BATCH;
 222                         int shrink_ret;
 223                         int nr_before;
 224
 225                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
 226                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 227                         if (shrink_ret == -1)
 228                                 break;
 229                         if (shrink_ret < nr_before)
 230                                 ret += nr_before - shrink_ret;
 231                         mod_page_state(slabs_scanned, this_scan);
 232                         total_scan -= this_scan;
 233
 234                         cond_resched();
 235                 }
 236
 237                 shrinker->nr += total_scan;
 238         }
 239         up_read(&shrinker_rwsem);
 240         return ret;
 241 }
 242
 243 /* Called without lock on whether page is mapped, so answer is unstable */
 244 static inline int page_mapping_inuse(struct page *page)
 245 {
 246         struct address_space *mapping;
 247
 248         /* Page is in somebody's page tables. */
 249         if (page_mapped(page))
 250                 return 1;
 251
 252         /* Be more reluctant to reclaim swapcache than pagecache */
 253         if (PageSwapCache(page))
 254                 return 1;
 255
 256         mapping = page_mapping(page);
 257         if (!mapping)
 258                 return 0;
 259
 260         /* File is mmap'd by somebody? */
 261         return mapping_mapped(mapping);
 262 }
 263
 264 static inline int is_page_cache_freeable(struct page *page)
 265 {
 266         return page_count(page) - !!PagePrivate(page) == 2;
 267 }
 268
 269 static int may_write_to_queue(struct backing_dev_info *bdi)
 270 {
 271         if (current_is_kswapd())
 272                 return 1;
 273         if (current_is_pdflush())       /* This is unlikely, but why not... */
 274                 return 1;
 275         if (!bdi_write_congested(bdi))
 276                 return 1;
 277         if (bdi == current->backing_dev_info)
 278                 return 1;
 279         return 0;
 280 }
 281
 282 /*
 283  * We detected a synchronous write error writing a page out.  Probably
 284  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 285  * fsync(), msync() or close().
 286  *
 287  * The tricky part is that after writepage we cannot touch the mapping: nothing
 288  * prevents it from being freed up.  But we have a ref on the page and once
 289  * that page is locked, the mapping is pinned.
 290  *
 291  * We're allowed to run sleeping lock_page() here because we know the caller has
 292  * __GFP_FS.
 293  */
 294 static void handle_write_error(struct address_space *mapping,
 295                                 struct page *page, int error)
 296 {
 297         lock_page(page);
 298         if (page_mapping(page) == mapping) {
 299                 if (error == -ENOSPC)
 300                         set_bit(AS_ENOSPC, &mapping->flags);
 301                 else
 302                         set_bit(AS_EIO, &mapping->flags);
 303         }
 304         unlock_page(page);
 305 }
 306
 307 /*
 308  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 309  */
 310 static pageout_t pageout(struct page *page, struct address_space *mapping)
 311 {
 312         /*
 313          * If the page is dirty, only perform writeback if that write
 314          * will be non-blocking.  To prevent this allocation from being
 315          * stalled by pagecache activity.  But note that there may be
 316          * stalls if we need to run get_block().  We could test
 317          * PagePrivate for that.
 318          *
 319          * If this process is currently in generic_file_write() against
 320          * this page's queue, we can perform writeback even if that
 321          * will block.
 322          *
 323          * If the page is swapcache, write it back even if that would
 324          * block, for some throttling. This happens by accident, because
 325          * swap_backing_dev_info is bust: it doesn't reflect the
 326          * congestion state of the swapdevs.  Easy to fix, if needed.
 327          * See swapfile.c:page_queue_congested().
 328          */
 329         if (!is_page_cache_freeable(page))
 330                 return PAGE_KEEP;
 331         if (!mapping) {
 332                 /*
 333                  * Some data journaling orphaned pages can have
 334                  * page->mapping == NULL while being dirty with clean buffers.
 335                  */
 336                 if (PagePrivate(page)) {
 337                         if (try_to_free_buffers(page)) {
 338                                 ClearPageDirty(page);
 339                                 printk("%s: orphaned page\n", __FUNCTION__);
 340                                 return PAGE_CLEAN;
 341                         }
 342                 }
 343                 return PAGE_KEEP;
 344         }
 345         if (mapping->a_ops->writepage == NULL)
 346                 return PAGE_ACTIVATE;
 347         if (!may_write_to_queue(mapping->backing_dev_info))
 348                 return PAGE_KEEP;
 349
 350         if (clear_page_dirty_for_io(page)) {
 351                 int res;
 352                 struct writeback_control wbc = {
 353                         .sync_mode = WB_SYNC_NONE,
 354                         .nr_to_write = SWAP_CLUSTER_MAX,
 355                         .nonblocking = 1,
 356                         .for_reclaim = 1,
 357                 };
 358
 359                 SetPageReclaim(page);
 360                 res = mapping->a_ops->writepage(page, &wbc);
 361                 if (res < 0)
 362                         handle_write_error(mapping, page, res);
 363                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 364                         ClearPageReclaim(page);
 365                         return PAGE_ACTIVATE;
 366                 }
 367                 if (!PageWriteback(page)) {
 368                         /* synchronous write or broken a_ops? */
 369                         ClearPageReclaim(page);
 370                 }
 371
 372                 return PAGE_SUCCESS;
 373         }
 374
 375         return PAGE_CLEAN;
 376 }
 377
 378 /*
 379  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 380  */
 381 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 382 {
 383         LIST_HEAD(ret_pages);
 384         struct pagevec freed_pvec;
 385         int pgactivate = 0;
 386         int reclaimed = 0;
 387
 388         cond_resched();
 389
 390         pagevec_init(&freed_pvec, 1);
 391         while (!list_empty(page_list)) {
 392                 struct address_space *mapping;
 393                 struct page *page;
 394                 int may_enter_fs;
 395                 int referenced;
 396
 397                 cond_resched();
 398
 399                 page = lru_to_page(page_list);
 400                 list_del(&page->lru);
 401
 402                 if (TestSetPageLocked(page))
 403                         goto keep;
 404
 405                 BUG_ON(PageActive(page));
 406
 407                 sc->nr_scanned++;
 408                 /* Double the slab pressure for mapped and swapcache pages */
 409                 if (page_mapped(page) || PageSwapCache(page))
 410                         sc->nr_scanned++;
 411
 412                 if (PageWriteback(page))
 413                         goto keep_locked;
 414
 415                 referenced = page_referenced(page, 1);
 416                 /* In active use or really unfreeable?  Activate it. */
 417                 if (referenced && page_mapping_inuse(page))
 418                         goto activate_locked;
 419
 420 #ifdef CONFIG_SWAP
 421                 /*
 422                  * Anonymous process memory has backing store?
 423                  * Try to allocate it some swap space here.
 424                  */
 425                 if (PageAnon(page) && !PageSwapCache(page)) {
 426                         if (!add_to_swap(page))
 427                                 goto activate_locked;
 428                 }
 429 #endif /* CONFIG_SWAP */
 430
 431                 mapping = page_mapping(page);
 432                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 433                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 434
 435                 /*
 436                  * The page is mapped into the page tables of one or more
 437                  * processes. Try to unmap it here.
 438                  */
 439                 if (page_mapped(page) && mapping) {
 440                         switch (try_to_unmap(page)) {
 441                         case SWAP_FAIL:
 442                                 goto activate_locked;
 443                         case SWAP_AGAIN:
 444                                 goto keep_locked;
 445                         case SWAP_SUCCESS:
 446                                 ; /* try to free the page below */
 447                         }
 448                 }
 449
 450                 if (PageDirty(page)) {
 451                         if (referenced)
 452                                 goto keep_locked;
 453                         if (!may_enter_fs)
 454                                 goto keep_locked;
 455                         if (laptop_mode && !sc->may_writepage)
 456                                 goto keep_locked;
 457
 458                         /* Page is dirty, try to write it out here */
 459                         switch(pageout(page, mapping)) {
 460                         case PAGE_KEEP:
 461                                 goto keep_locked;
 462                         case PAGE_ACTIVATE:
 463                                 goto activate_locked;
 464                         case PAGE_SUCCESS:
 465                                 if (PageWriteback(page) || PageDirty(page))
 466                                         goto keep;
 467                                 /*
 468                                  * A synchronous write - probably a ramdisk.  Go
 469                                  * ahead and try to reclaim the page.
 470                                  */
 471                                 if (TestSetPageLocked(page))
 472                                         goto keep;
 473                                 if (PageDirty(page) || PageWriteback(page))
 474                                         goto keep_locked;
 475                                 mapping = page_mapping(page);
 476                         case PAGE_CLEAN:
 477                                 ; /* try to free the page below */
 478                         }
 479                 }
 480
 481                 /*
 482                  * If the page has buffers, try to free the buffer mappings
 483                  * associated with this page. If we succeed we try to free
 484                  * the page as well.
 485                  *
 486                  * We do this even if the page is PageDirty().
 487                  * try_to_release_page() does not perform I/O, but it is
 488                  * possible for a page to have PageDirty set, but it is actually
 489                  * clean (all its buffers are clean).  This happens if the
 490                  * buffers were written out directly, with submit_bh(). ext3
 491                  * will do this, as well as the blockdev mapping.
 492                  * try_to_release_page() will discover that cleanness and will
 493                  * drop the buffers and mark the page clean - it can be freed.
 494                  *
 495                  * Rarely, pages can have buffers and no ->mapping.  These are
 496                  * the pages which were not successfully invalidated in
 497                  * truncate_complete_page().  We try to drop those buffers here
 498                  * and if that worked, and the page is no longer mapped into
 499                  * process address space (page_count == 1) it can be freed.
 500                  * Otherwise, leave the page on the LRU so it is swappable.
 501                  */
 502                 if (PagePrivate(page)) {
 503                         if (!try_to_release_page(page, sc->gfp_mask))
 504                                 goto activate_locked;
 505                         if (!mapping && page_count(page) == 1)
 506                                 goto free_it;
 507                 }
 508
 509                 if (!mapping)
 510                         goto keep_locked;       /* truncate got there first */
 511
 512                 write_lock_irq(&mapping->tree_lock);
 513
 514                 /*
 515                  * The non-racy check for busy page.  It is critical to check
 516                  * PageDirty _after_ making sure that the page is freeable and
 517                  * not in use by anybody.       (pagecache + us == 2)
 518                  */
 519                 if (unlikely(page_count(page) != 2))
 520                         goto cannot_free;
 521                 smp_rmb();
 522                 if (unlikely(PageDirty(page)))
 523                         goto cannot_free;
 524
 525 #ifdef CONFIG_SWAP
 526                 if (PageSwapCache(page)) {
 527                         swp_entry_t swap = { .val = page_private(page) };
 528                         __delete_from_swap_cache(page);
 529                         write_unlock_irq(&mapping->tree_lock);
 530                         swap_free(swap);
 531                         __put_page(page);       /* The pagecache ref */
 532                         goto free_it;
 533                 }
 534 #endif /* CONFIG_SWAP */
 535
 536                 __remove_from_page_cache(page);
 537                 write_unlock_irq(&mapping->tree_lock);
 538                 __put_page(page);
 539
 540 free_it:
 541                 unlock_page(page);
 542                 reclaimed++;
 543                 if (!pagevec_add(&freed_pvec, page))
 544                         __pagevec_release_nonlru(&freed_pvec);
 545                 continue;
 546
 547 cannot_free:
 548                 write_unlock_irq(&mapping->tree_lock);
 549                 goto keep_locked;
 550
 551 activate_locked:
 552                 SetPageActive(page);
 553                 pgactivate++;
 554 keep_locked:
 555                 unlock_page(page);
 556 keep:
 557                 list_add(&page->lru, &ret_pages);
 558                 BUG_ON(PageLRU(page));
 559         }
 560         list_splice(&ret_pages, page_list);
 561         if (pagevec_count(&freed_pvec))
 562                 __pagevec_release_nonlru(&freed_pvec);
 563         mod_page_state(pgactivate, pgactivate);
 564         sc->nr_reclaimed += reclaimed;
 565         return reclaimed;
 566 }
 567
 568 /*
 569  * zone->lru_lock is heavily contended.  Some of the functions that
 570  * shrink the lists perform better by taking out a batch of pages
 571  * and working on them outside the LRU lock.
 572  *
 573  * For pagecache intensive workloads, this function is the hottest
 574  * spot in the kernel (apart from copy_*_user functions).
 575  *
 576  * Appropriate locks must be held before calling this function.
 577  *
 578  * @nr_to_scan: The number of pages to look through on the list.
 579  * @src:        The LRU list to pull pages off.
 580  * @dst:        The temp list to put pages on to.
 581  * @scanned:    The number of pages that were scanned.
 582  *
 583  * returns how many pages were moved onto *@dst.
 584  */
 585 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
 586                              struct list_head *dst, int *scanned)
 587 {
 588         int nr_taken = 0;
 589         struct page *page;
 590         int scan = 0;
 591
 592         while (scan++ < nr_to_scan && !list_empty(src)) {
 593                 page = lru_to_page(src);
 594                 prefetchw_prev_lru_page(page, src, flags);
 595
 596                 switch (__isolate_lru_page(page)) {
 597                 case 1:
 598                         /* Succeeded to isolate page */
 599                         list_move(&page->lru, dst);
 600                         nr_taken++;
 601                         break;
 602                 case -ENOENT:
 603                         /* Not possible to isolate */
 604                         list_move(&page->lru, src);
 605                         break;
 606                 default:
 607                         BUG();
 608                 }
 609         }
 610
 611         *scanned = scan;
 612         return nr_taken;
 613 }
 614
 615 static void lru_add_drain_per_cpu(void *dummy)
 616 {
 617         lru_add_drain();
 618 }
 619
 620 /*
 621  * Isolate one page from the LRU lists and put it on the
 622  * indicated list. Do necessary cache draining if the
 623  * page is not on the LRU lists yet.
 624  *
 625  * Result:
 626  *  0 = page not on LRU list
 627  *  1 = page removed from LRU list and added to the specified list.
 628  * -ENOENT = page is being freed elsewhere.
 629  */
 630 int isolate_lru_page(struct page *page)
 631 {
 632         int rc = 0;
 633         struct zone *zone = page_zone(page);
 634
 635 redo:
 636         spin_lock_irq(&zone->lru_lock);
 637         rc = __isolate_lru_page(page);
 638         if (rc == 1) {
 639                 if (PageActive(page))
 640                         del_page_from_active_list(zone, page);
 641                 else
 642                         del_page_from_inactive_list(zone, page);
 643         }
 644         spin_unlock_irq(&zone->lru_lock);
 645         if (rc == 0) {
 646                 /*
 647                  * Maybe this page is still waiting for a cpu to drain it
 648                  * from one of the lru lists?
 649                  */
 650                 rc = schedule_on_each_cpu(lru_add_drain_per_cpu, NULL);
 651                 if (rc == 0 && PageLRU(page))
 652                         goto redo;
 653         }
 654         return rc;
 655 }
 656
 657 /*
 658  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 659  */
 660 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 661 {
 662         LIST_HEAD(page_list);
 663         struct pagevec pvec;
 664         int max_scan = sc->nr_to_scan;
 665
 666         pagevec_init(&pvec, 1);
 667
 668         lru_add_drain();
 669         spin_lock_irq(&zone->lru_lock);
 670         while (max_scan > 0) {
 671                 struct page *page;
 672                 int nr_taken;
 673                 int nr_scan;
 674                 int nr_freed;
 675
 676                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
 677                                              &zone->inactive_list,
 678                                              &page_list, &nr_scan);
 679                 zone->nr_inactive -= nr_taken;
 680                 zone->pages_scanned += nr_scan;
 681                 spin_unlock_irq(&zone->lru_lock);
 682
 683                 if (nr_taken == 0)
 684                         goto done;
 685
 686                 max_scan -= nr_scan;
 687                 nr_freed = shrink_list(&page_list, sc);
 688
 689                 local_irq_disable();
 690                 if (current_is_kswapd()) {
 691                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 692                         __mod_page_state(kswapd_steal, nr_freed);
 693                 } else
 694                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
 695                 __mod_page_state_zone(zone, pgsteal, nr_freed);
 696
 697                 spin_lock(&zone->lru_lock);
 698                 /*
 699                  * Put back any unfreeable pages.
 700                  */
 701                 while (!list_empty(&page_list)) {
 702                         page = lru_to_page(&page_list);
 703                         if (TestSetPageLRU(page))
 704                                 BUG();
 705                         list_del(&page->lru);
 706                         if (PageActive(page))
 707                                 add_page_to_active_list(zone, page);
 708                         else
 709                                 add_page_to_inactive_list(zone, page);
 710                         if (!pagevec_add(&pvec, page)) {
 711                                 spin_unlock_irq(&zone->lru_lock);
 712                                 __pagevec_release(&pvec);
 713                                 spin_lock_irq(&zone->lru_lock);
 714                         }
 715                 }
 716         }
 717         spin_unlock_irq(&zone->lru_lock);
 718 done:
 719         pagevec_release(&pvec);
 720 }
 721
 722 static inline void move_to_lru(struct page *page)
 723 {
 724         list_del(&page->lru);
 725         if (PageActive(page)) {
 726                 /*
 727                  * lru_cache_add_active checks that
 728                  * the PG_active bit is off.
 729                  */
 730                 ClearPageActive(page);
 731                 lru_cache_add_active(page);
 732         } else {
 733                 lru_cache_add(page);
 734         }
 735         put_page(page);
 736 }
 737
 738 /*
 739  * Add isolated pages on the list back to the LRU
 740  *
 741  * returns the number of pages put back.
 742  */
 743 int putback_lru_pages(struct list_head *l)
 744 {
 745         struct page *page;
 746         struct page *page2;
 747         int count = 0;
 748
 749         list_for_each_entry_safe(page, page2, l, lru) {
 750                 move_to_lru(page);
 751                 count++;
 752         }
 753         return count;
 754 }
 755
 756 /*
 757  * This moves pages from the active list to the inactive list.
 758  *
 759  * We move them the other way if the page is referenced by one or more
 760  * processes, from rmap.
 761  *
 762  * If the pages are mostly unmapped, the processing is fast and it is
 763  * appropriate to hold zone->lru_lock across the whole operation.  But if
 764  * the pages are mapped, the processing is slow (page_referenced()) so we
 765  * should drop zone->lru_lock around each page.  It's impossible to balance
 766  * this, so instead we remove the pages from the LRU while processing them.
 767  * It is safe to rely on PG_active against the non-LRU pages in here because
 768  * nobody will play with that bit on a non-LRU page.
 769  *
 770  * The downside is that we have to touch page->_count against each page.
 771  * But we had to alter page->flags anyway.
 772  */
 773 static void
 774 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 775 {
 776         int pgmoved;
 777         int pgdeactivate = 0;
 778         int pgscanned;
 779         int nr_pages = sc->nr_to_scan;
 780         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 781         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 782         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 783         struct page *page;
 784         struct pagevec pvec;
 785         int reclaim_mapped = 0;
 786         long mapped_ratio;
 787         long distress;
 788         long swap_tendency;
 789
 790         lru_add_drain();
 791         spin_lock_irq(&zone->lru_lock);
 792         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
 793                                     &l_hold, &pgscanned);
 794         zone->pages_scanned += pgscanned;
 795         zone->nr_active -= pgmoved;
 796         spin_unlock_irq(&zone->lru_lock);
 797
 798         /*
 799          * `distress' is a measure of how much trouble we're having reclaiming
 800          * pages.  0 -> no problems.  100 -> great trouble.
 801          */
 802         distress = 100 >> zone->prev_priority;
 803
 804         /*
 805          * The point of this algorithm is to decide when to start reclaiming
 806          * mapped memory instead of just pagecache.  Work out how much memory
 807          * is mapped.
 808          */
 809         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 810
 811         /*
 812          * Now decide how much we really want to unmap some pages.  The mapped
 813          * ratio is downgraded - just because there's a lot of mapped memory
 814          * doesn't necessarily mean that page reclaim isn't succeeding.
 815          *
 816          * The distress ratio is important - we don't want to start going oom.
 817          *
 818          * A 100% value of vm_swappiness overrides this algorithm altogether.
 819          */
 820         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 821
 822         /*
 823          * Now use this metric to decide whether to start moving mapped memory
 824          * onto the inactive list.
 825          */
 826         if (swap_tendency >= 100)
 827                 reclaim_mapped = 1;
 828
 829         while (!list_empty(&l_hold)) {
 830                 cond_resched();
 831                 page = lru_to_page(&l_hold);
 832                 list_del(&page->lru);
 833                 if (page_mapped(page)) {
 834                         if (!reclaim_mapped ||
 835                             (total_swap_pages == 0 && PageAnon(page)) ||
 836                             page_referenced(page, 0)) {
 837                                 list_add(&page->lru, &l_active);
 838                                 continue;
 839                         }
 840                 }
 841                 list_add(&page->lru, &l_inactive);
 842         }
 843
 844         pagevec_init(&pvec, 1);
 845         pgmoved = 0;
 846         spin_lock_irq(&zone->lru_lock);
 847         while (!list_empty(&l_inactive)) {
 848                 page = lru_to_page(&l_inactive);
 849                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 850                 if (TestSetPageLRU(page))
 851                         BUG();
 852                 if (!TestClearPageActive(page))
 853                         BUG();
 854                 list_move(&page->lru, &zone->inactive_list);
 855                 pgmoved++;
 856                 if (!pagevec_add(&pvec, page)) {
 857                         zone->nr_inactive += pgmoved;
 858                         spin_unlock_irq(&zone->lru_lock);
 859                         pgdeactivate += pgmoved;
 860                         pgmoved = 0;
 861                         if (buffer_heads_over_limit)
 862                                 pagevec_strip(&pvec);
 863                         __pagevec_release(&pvec);
 864                         spin_lock_irq(&zone->lru_lock);
 865                 }
 866         }
 867         zone->nr_inactive += pgmoved;
 868         pgdeactivate += pgmoved;
 869         if (buffer_heads_over_limit) {
 870                 spin_unlock_irq(&zone->lru_lock);
 871                 pagevec_strip(&pvec);
 872                 spin_lock_irq(&zone->lru_lock);
 873         }
 874
 875         pgmoved = 0;
 876         while (!list_empty(&l_active)) {
 877                 page = lru_to_page(&l_active);
 878                 prefetchw_prev_lru_page(page, &l_active, flags);
 879                 if (TestSetPageLRU(page))
 880                         BUG();
 881                 BUG_ON(!PageActive(page));
 882                 list_move(&page->lru, &zone->active_list);
 883                 pgmoved++;
 884                 if (!pagevec_add(&pvec, page)) {
 885                         zone->nr_active += pgmoved;
 886                         pgmoved = 0;
 887                         spin_unlock_irq(&zone->lru_lock);
 888                         __pagevec_release(&pvec);
 889                         spin_lock_irq(&zone->lru_lock);
 890                 }
 891         }
 892         zone->nr_active += pgmoved;
 893         spin_unlock(&zone->lru_lock);
 894
 895         __mod_page_state_zone(zone, pgrefill, pgscanned);
 896         __mod_page_state(pgdeactivate, pgdeactivate);
 897         local_irq_enable();
 898
 899         pagevec_release(&pvec);
 900 }
 901
 902 /*
 903  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 904  */
 905 static void
 906 shrink_zone(struct zone *zone, struct scan_control *sc)
 907 {
 908         unsigned long nr_active;
 909         unsigned long nr_inactive;
 910
 911         atomic_inc(&zone->reclaim_in_progress);
 912
 913         /*
 914          * Add one to `nr_to_scan' just to make sure that the kernel will
 915          * slowly sift through the active list.
 916          */
 917         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 918         nr_active = zone->nr_scan_active;
 919         if (nr_active >= sc->swap_cluster_max)
 920                 zone->nr_scan_active = 0;
 921         else
 922                 nr_active = 0;
 923
 924         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 925         nr_inactive = zone->nr_scan_inactive;
 926         if (nr_inactive >= sc->swap_cluster_max)
 927                 zone->nr_scan_inactive = 0;
 928         else
 929                 nr_inactive = 0;
 930
 931         while (nr_active || nr_inactive) {
 932                 if (nr_active) {
 933                         sc->nr_to_scan = min(nr_active,
 934                                         (unsigned long)sc->swap_cluster_max);
 935                         nr_active -= sc->nr_to_scan;
 936                         refill_inactive_zone(zone, sc);
 937                 }
 938
 939                 if (nr_inactive) {
 940                         sc->nr_to_scan = min(nr_inactive,
 941                                         (unsigned long)sc->swap_cluster_max);
 942                         nr_inactive -= sc->nr_to_scan;
 943                         shrink_cache(zone, sc);
 944                 }
 945         }
 946
 947         throttle_vm_writeout();
 948
 949         atomic_dec(&zone->reclaim_in_progress);
 950 }
 951
 952 /*
 953  * This is the direct reclaim path, for page-allocating processes.  We only
 954  * try to reclaim pages from zones which will satisfy the caller's allocation
 955  * request.
 956  *
 957  * We reclaim from a zone even if that zone is over pages_high.  Because:
 958  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 959  *    allocation or
 960  * b) The zones may be over pages_high but they must go *over* pages_high to
 961  *    satisfy the `incremental min' zone defense algorithm.
 962  *
 963  * Returns the number of reclaimed pages.
 964  *
 965  * If a zone is deemed to be full of pinned pages then just give it a light
 966  * scan then give up on it.
 967  */
 968 static void
 969 shrink_caches(struct zone **zones, struct scan_control *sc)
 970 {
 971         int i;
 972
 973         for (i = 0; zones[i] != NULL; i++) {
 974                 struct zone *zone = zones[i];
 975
 976                 if (!populated_zone(zone))
 977                         continue;
 978
 979                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 980                         continue;
 981
 982                 zone->temp_priority = sc->priority;
 983                 if (zone->prev_priority > sc->priority)
 984                         zone->prev_priority = sc->priority;
 985
 986                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
 987                         continue;       /* Let kswapd poll it */
 988
 989                 shrink_zone(zone, sc);
 990         }
 991 }
 992
 993 /*
 994  * This is the main entry point to direct page reclaim.
 995  *
 996  * If a full scan of the inactive list fails to free enough memory then we
 997  * are "out of memory" and something needs to be killed.
 998  *
 999  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1000  * high - the zone may be full of dirty or under-writeback pages, which this
1001  * caller can't do much about.  We kick pdflush and take explicit naps in the
1002  * hope that some of these pages can be written.  But if the allocating task
1003  * holds filesystem locks which prevent writeout this might not work, and the
1004  * allocation attempt will fail.
1005  */
1006 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1007 {
1008         int priority;
1009         int ret = 0;
1010         int total_scanned = 0, total_reclaimed = 0;
1011         struct reclaim_state *reclaim_state = current->reclaim_state;
1012         struct scan_control sc;
1013         unsigned long lru_pages = 0;
1014         int i;
1015
1016         sc.gfp_mask = gfp_mask;
1017         sc.may_writepage = 0;
1018
1019         inc_page_state(allocstall);
1020
1021         for (i = 0; zones[i] != NULL; i++) {
1022                 struct zone *zone = zones[i];
1023
1024                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1025                         continue;
1026
1027                 zone->temp_priority = DEF_PRIORITY;
1028                 lru_pages += zone->nr_active + zone->nr_inactive;
1029         }
1030
1031         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1032                 sc.nr_mapped = read_page_state(nr_mapped);
1033                 sc.nr_scanned = 0;
1034                 sc.nr_reclaimed = 0;
1035                 sc.priority = priority;
1036                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1037                 if (!priority)
1038                         disable_swap_token();
1039                 shrink_caches(zones, &sc);
1040                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1041                 if (reclaim_state) {
1042                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1043                         reclaim_state->reclaimed_slab = 0;
1044                 }
1045                 total_scanned += sc.nr_scanned;
1046                 total_reclaimed += sc.nr_reclaimed;
1047                 if (total_reclaimed >= sc.swap_cluster_max) {
1048                         ret = 1;
1049                         goto out;
1050                 }
1051
1052                 /*
1053                  * Try to write back as many pages as we just scanned.  This
1054                  * tends to cause slow streaming writers to write data to the
1055                  * disk smoothly, at the dirtying rate, which is nice.   But
1056                  * that's undesirable in laptop mode, where we *want* lumpy
1057                  * writeout.  So in laptop mode, write out the whole world.
1058                  */
1059                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1060                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1061                         sc.may_writepage = 1;
1062                 }
1063
1064                 /* Take a nap, wait for some writeback to complete */
1065                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1066                         blk_congestion_wait(WRITE, HZ/10);
1067         }
1068 out:
1069         for (i = 0; zones[i] != 0; i++) {
1070                 struct zone *zone = zones[i];
1071
1072                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1073                         continue;
1074
1075                 zone->prev_priority = zone->temp_priority;
1076         }
1077         return ret;
1078 }
1079
1080 /*
1081  * For kswapd, balance_pgdat() will work across all this node's zones until
1082  * they are all at pages_high.
1083  *
1084  * If `nr_pages' is non-zero then it is the number of pages which are to be
1085  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1086  * special.
1087  *
1088  * Returns the number of pages which were actually freed.
1089  *
1090  * There is special handling here for zones which are full of pinned pages.
1091  * This can happen if the pages are all mlocked, or if they are all used by
1092  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1093  * What we do is to detect the case where all pages in the zone have been
1094  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1095  * dead and from now on, only perform a short scan.  Basically we're polling
1096  * the zone for when the problem goes away.
1097  *
1098  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1099  * zones which have free_pages > pages_high, but once a zone is found to have
1100  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1101  * of the number of free pages in the lower zones.  This interoperates with
1102  * the page allocator fallback scheme to ensure that aging of pages is balanced
1103  * across the zones.
1104  */
1105 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1106 {
1107         int to_free = nr_pages;
1108         int all_zones_ok;
1109         int priority;
1110         int i;
1111         int total_scanned, total_reclaimed;
1112         struct reclaim_state *reclaim_state = current->reclaim_state;
1113         struct scan_control sc;
1114
1115 loop_again:
1116         total_scanned = 0;
1117         total_reclaimed = 0;
1118         sc.gfp_mask = GFP_KERNEL;
1119         sc.may_writepage = 0;
1120         sc.nr_mapped = read_page_state(nr_mapped);
1121
1122         inc_page_state(pageoutrun);
1123
1124         for (i = 0; i < pgdat->nr_zones; i++) {
1125                 struct zone *zone = pgdat->node_zones + i;
1126
1127                 zone->temp_priority = DEF_PRIORITY;
1128         }
1129
1130         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1131                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1132                 unsigned long lru_pages = 0;
1133
1134                 /* The swap token gets in the way of swapout... */
1135                 if (!priority)
1136                         disable_swap_token();
1137
1138                 all_zones_ok = 1;
1139
1140                 if (nr_pages == 0) {
1141                         /*
1142                          * Scan in the highmem->dma direction for the highest
1143                          * zone which needs scanning
1144                          */
1145                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1146                                 struct zone *zone = pgdat->node_zones + i;
1147
1148                                 if (!populated_zone(zone))
1149                                         continue;
1150
1151                                 if (zone->all_unreclaimable &&
1152                                                 priority != DEF_PRIORITY)
1153                                         continue;
1154
1155                                 if (!zone_watermark_ok(zone, order,
1156                                                 zone->pages_high, 0, 0)) {
1157                                         end_zone = i;
1158                                         goto scan;
1159                                 }
1160                         }
1161                         goto out;
1162                 } else {
1163                         end_zone = pgdat->nr_zones - 1;
1164                 }
1165 scan:
1166                 for (i = 0; i <= end_zone; i++) {
1167                         struct zone *zone = pgdat->node_zones + i;
1168
1169                         lru_pages += zone->nr_active + zone->nr_inactive;
1170                 }
1171
1172                 /*
1173                  * Now scan the zone in the dma->highmem direction, stopping
1174                  * at the last zone which needs scanning.
1175                  *
1176                  * We do this because the page allocator works in the opposite
1177                  * direction.  This prevents the page allocator from allocating
1178                  * pages behind kswapd's direction of progress, which would
1179                  * cause too much scanning of the lower zones.
1180                  */
1181                 for (i = 0; i <= end_zone; i++) {
1182                         struct zone *zone = pgdat->node_zones + i;
1183                         int nr_slab;
1184
1185                         if (!populated_zone(zone))
1186                                 continue;
1187
1188                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1189                                 continue;
1190
1191                         if (nr_pages == 0) {    /* Not software suspend */
1192                                 if (!zone_watermark_ok(zone, order,
1193                                                 zone->pages_high, end_zone, 0))
1194                                         all_zones_ok = 0;
1195                         }
1196                         zone->temp_priority = priority;
1197                         if (zone->prev_priority > priority)
1198                                 zone->prev_priority = priority;
1199                         sc.nr_scanned = 0;
1200                         sc.nr_reclaimed = 0;
1201                         sc.priority = priority;
1202                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1203                         atomic_inc(&zone->reclaim_in_progress);
1204                         shrink_zone(zone, &sc);
1205                         atomic_dec(&zone->reclaim_in_progress);
1206                         reclaim_state->reclaimed_slab = 0;
1207                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1208                                                 lru_pages);
1209                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1210                         total_reclaimed += sc.nr_reclaimed;
1211                         total_scanned += sc.nr_scanned;
1212                         if (zone->all_unreclaimable)
1213                                 continue;
1214                         if (nr_slab == 0 && zone->pages_scanned >=
1215                                     (zone->nr_active + zone->nr_inactive) * 4)
1216                                 zone->all_unreclaimable = 1;
1217                         /*
1218                          * If we've done a decent amount of scanning and
1219                          * the reclaim ratio is low, start doing writepage
1220                          * even in laptop mode
1221                          */
1222                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1223                             total_scanned > total_reclaimed+total_reclaimed/2)
1224                                 sc.may_writepage = 1;
1225                 }
1226                 if (nr_pages && to_free > total_reclaimed)
1227                         continue;       /* swsusp: need to do more work */
1228                 if (all_zones_ok)
1229                         break;          /* kswapd: all done */
1230                 /*
1231                  * OK, kswapd is getting into trouble.  Take a nap, then take
1232                  * another pass across the zones.
1233                  */
1234                 if (total_scanned && priority < DEF_PRIORITY - 2)
1235                         blk_congestion_wait(WRITE, HZ/10);
1236
1237                 /*
1238                  * We do this so kswapd doesn't build up large priorities for
1239                  * example when it is freeing in parallel with allocators. It
1240                  * matches the direct reclaim path behaviour in terms of impact
1241                  * on zone->*_priority.
1242                  */
1243                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1244                         break;
1245         }
1246 out:
1247         for (i = 0; i < pgdat->nr_zones; i++) {
1248                 struct zone *zone = pgdat->node_zones + i;
1249
1250                 zone->prev_priority = zone->temp_priority;
1251         }
1252         if (!all_zones_ok) {
1253                 cond_resched();
1254                 goto loop_again;
1255         }
1256
1257         return total_reclaimed;
1258 }
1259
1260 /*
1261  * The background pageout daemon, started as a kernel thread
1262  * from the init process.
1263  *
1264  * This basically trickles out pages so that we have _some_
1265  * free memory available even if there is no other activity
1266  * that frees anything up. This is needed for things like routing
1267  * etc, where we otherwise might have all activity going on in
1268  * asynchronous contexts that cannot page things out.
1269  *
1270  * If there are applications that are active memory-allocators
1271  * (most normal use), this basically shouldn't matter.
1272  */
1273 static int kswapd(void *p)
1274 {
1275         unsigned long order;
1276         pg_data_t *pgdat = (pg_data_t*)p;
1277         struct task_struct *tsk = current;
1278         DEFINE_WAIT(wait);
1279         struct reclaim_state reclaim_state = {
1280                 .reclaimed_slab = 0,
1281         };
1282         cpumask_t cpumask;
1283
1284         daemonize("kswapd%d", pgdat->node_id);
1285         cpumask = node_to_cpumask(pgdat->node_id);
1286         if (!cpus_empty(cpumask))
1287                 set_cpus_allowed(tsk, cpumask);
1288         current->reclaim_state = &reclaim_state;
1289
1290         /*
1291          * Tell the memory management that we're a "memory allocator",
1292          * and that if we need more memory we should get access to it
1293          * regardless (see "__alloc_pages()"). "kswapd" should
1294          * never get caught in the normal page freeing logic.
1295          *
1296          * (Kswapd normally doesn't need memory anyway, but sometimes
1297          * you need a small amount of memory in order to be able to
1298          * page out something else, and this flag essentially protects
1299          * us from recursively trying to free more memory as we're
1300          * trying to free the first piece of memory in the first place).
1301          */
1302         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1303
1304         order = 0;
1305         for ( ; ; ) {
1306                 unsigned long new_order;
1307
1308                 try_to_freeze();
1309
1310                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1311                 new_order = pgdat->kswapd_max_order;
1312                 pgdat->kswapd_max_order = 0;
1313                 if (order < new_order) {
1314                         /*
1315                          * Don't sleep if someone wants a larger 'order'
1316                          * allocation
1317                          */
1318                         order = new_order;
1319                 } else {
1320                         schedule();
1321                         order = pgdat->kswapd_max_order;
1322                 }
1323                 finish_wait(&pgdat->kswapd_wait, &wait);
1324
1325                 balance_pgdat(pgdat, 0, order);
1326         }
1327         return 0;
1328 }
1329
1330 /*
1331  * A zone is low on free memory, so wake its kswapd task to service it.
1332  */
1333 void wakeup_kswapd(struct zone *zone, int order)
1334 {
1335         pg_data_t *pgdat;
1336
1337         if (!populated_zone(zone))
1338                 return;
1339
1340         pgdat = zone->zone_pgdat;
1341         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1342                 return;
1343         if (pgdat->kswapd_max_order < order)
1344                 pgdat->kswapd_max_order = order;
1345         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1346                 return;
1347         if (!waitqueue_active(&pgdat->kswapd_wait))
1348                 return;
1349         wake_up_interruptible(&pgdat->kswapd_wait);
1350 }
1351
1352 #ifdef CONFIG_PM
1353 /*
1354  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1355  * pages.
1356  */
1357 int shrink_all_memory(int nr_pages)
1358 {
1359         pg_data_t *pgdat;
1360         int nr_to_free = nr_pages;
1361         int ret = 0;
1362         struct reclaim_state reclaim_state = {
1363                 .reclaimed_slab = 0,
1364         };
1365
1366         current->reclaim_state = &reclaim_state;
1367         for_each_pgdat(pgdat) {
1368                 int freed;
1369                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1370                 ret += freed;
1371                 nr_to_free -= freed;
1372                 if (nr_to_free <= 0)
1373                         break;
1374         }
1375         current->reclaim_state = NULL;
1376         return ret;
1377 }
1378 #endif
1379
1380 #ifdef CONFIG_HOTPLUG_CPU
1381 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1382    not required for correctness.  So if the last cpu in a node goes
1383    away, we get changed to run anywhere: as the first one comes back,
1384    restore their cpu bindings. */
1385 static int __devinit cpu_callback(struct notifier_block *nfb,
1386                                   unsigned long action,
1387                                   void *hcpu)
1388 {
1389         pg_data_t *pgdat;
1390         cpumask_t mask;
1391
1392         if (action == CPU_ONLINE) {
1393                 for_each_pgdat(pgdat) {
1394                         mask = node_to_cpumask(pgdat->node_id);
1395                         if (any_online_cpu(mask) != NR_CPUS)
1396                                 /* One of our CPUs online: restore mask */
1397                                 set_cpus_allowed(pgdat->kswapd, mask);
1398                 }
1399         }
1400         return NOTIFY_OK;
1401 }
1402 #endif /* CONFIG_HOTPLUG_CPU */
1403
1404 static int __init kswapd_init(void)
1405 {
1406         pg_data_t *pgdat;
1407         swap_setup();
1408         for_each_pgdat(pgdat)
1409                 pgdat->kswapd
1410                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1411         total_memory = nr_free_pagecache_pages();
1412         hotcpu_notifier(cpu_callback, 0);
1413         return 0;
1414 }
1415
1416 module_init(kswapd_init)