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