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 #include <linux/delay.h>
  37
  38 #include <asm/tlbflush.h>
  39 #include <asm/div64.h>
  40
  41 #include <linux/swapops.h>
  42
  43 #include "internal.h"
  44
  45 struct scan_control {
  46         /* Incremented by the number of inactive pages that were scanned */
  47         unsigned long nr_scanned;
  48
  49         unsigned long nr_mapped;        /* From page_state */
  50
  51         /* This context's GFP mask */
  52         gfp_t gfp_mask;
  53
  54         int may_writepage;
  55
  56         /* Can pages be swapped as part of reclaim? */
  57         int may_swap;
  58
  59         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  60          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  61          * In this context, it doesn't matter that we scan the
  62          * whole list at once. */
  63         int swap_cluster_max;
  64
  65         int swappiness;
  66 };
  67
  68 /*
  69  * The list of shrinker callbacks used by to apply pressure to
  70  * ageable caches.
  71  */
  72 struct shrinker {
  73         shrinker_t              shrinker;
  74         struct list_head        list;
  75         int                     seeks;  /* seeks to recreate an obj */
  76         long                    nr;     /* objs pending delete */
  77 };
  78
  79 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  80
  81 #ifdef ARCH_HAS_PREFETCH
  82 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  83         do {                                                            \
  84                 if ((_page)->lru.prev != _base) {                       \
  85                         struct page *prev;                              \
  86                                                                         \
  87                         prev = lru_to_page(&(_page->lru));              \
  88                         prefetch(&prev->_field);                        \
  89                 }                                                       \
  90         } while (0)
  91 #else
  92 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
  93 #endif
  94
  95 #ifdef ARCH_HAS_PREFETCHW
  96 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
  97         do {                                                            \
  98                 if ((_page)->lru.prev != _base) {                       \
  99                         struct page *prev;                              \
 100                                                                         \
 101                         prev = lru_to_page(&(_page->lru));              \
 102                         prefetchw(&prev->_field);                       \
 103                 }                                                       \
 104         } while (0)
 105 #else
 106 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 107 #endif
 108
 109 /*
 110  * From 0 .. 100.  Higher means more swappy.
 111  */
 112 int vm_swappiness = 60;
 113 long vm_total_pages;    /* The total number of pages which the VM controls */
 114
 115 static LIST_HEAD(shrinker_list);
 116 static DECLARE_RWSEM(shrinker_rwsem);
 117
 118 /*
 119  * Add a shrinker callback to be called from the vm
 120  */
 121 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 122 {
 123         struct shrinker *shrinker;
 124
 125         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 126         if (shrinker) {
 127                 shrinker->shrinker = theshrinker;
 128                 shrinker->seeks = seeks;
 129                 shrinker->nr = 0;
 130                 down_write(&shrinker_rwsem);
 131                 list_add_tail(&shrinker->list, &shrinker_list);
 132                 up_write(&shrinker_rwsem);
 133         }
 134         return shrinker;
 135 }
 136 EXPORT_SYMBOL(set_shrinker);
 137
 138 /*
 139  * Remove one
 140  */
 141 void remove_shrinker(struct shrinker *shrinker)
 142 {
 143         down_write(&shrinker_rwsem);
 144         list_del(&shrinker->list);
 145         up_write(&shrinker_rwsem);
 146         kfree(shrinker);
 147 }
 148 EXPORT_SYMBOL(remove_shrinker);
 149
 150 #define SHRINK_BATCH 128
 151 /*
 152  * Call the shrink functions to age shrinkable caches
 153  *
 154  * Here we assume it costs one seek to replace a lru page and that it also
 155  * takes a seek to recreate a cache object.  With this in mind we age equal
 156  * percentages of the lru and ageable caches.  This should balance the seeks
 157  * generated by these structures.
 158  *
 159  * If the vm encounted mapped pages on the LRU it increase the pressure on
 160  * slab to avoid swapping.
 161  *
 162  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 163  *
 164  * `lru_pages' represents the number of on-LRU pages in all the zones which
 165  * are eligible for the caller's allocation attempt.  It is used for balancing
 166  * slab reclaim versus page reclaim.
 167  *
 168  * Returns the number of slab objects which we shrunk.
 169  */
 170 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
 171                         unsigned long lru_pages)
 172 {
 173         struct shrinker *shrinker;
 174         unsigned long ret = 0;
 175
 176         if (scanned == 0)
 177                 scanned = SWAP_CLUSTER_MAX;
 178
 179         if (!down_read_trylock(&shrinker_rwsem))
 180                 return 1;       /* Assume we'll be able to shrink next time */
 181
 182         list_for_each_entry(shrinker, &shrinker_list, list) {
 183                 unsigned long long delta;
 184                 unsigned long total_scan;
 185                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
 186
 187                 delta = (4 * scanned) / shrinker->seeks;
 188                 delta *= max_pass;
 189                 do_div(delta, lru_pages + 1);
 190                 shrinker->nr += delta;
 191                 if (shrinker->nr < 0) {
 192                         printk(KERN_ERR "%s: nr=%ld\n",
 193                                         __FUNCTION__, shrinker->nr);
 194                         shrinker->nr = max_pass;
 195                 }
 196
 197                 /*
 198                  * Avoid risking looping forever due to too large nr value:
 199                  * never try to free more than twice the estimate number of
 200                  * freeable entries.
 201                  */
 202                 if (shrinker->nr > max_pass * 2)
 203                         shrinker->nr = max_pass * 2;
 204
 205                 total_scan = shrinker->nr;
 206                 shrinker->nr = 0;
 207
 208                 while (total_scan >= SHRINK_BATCH) {
 209                         long this_scan = SHRINK_BATCH;
 210                         int shrink_ret;
 211                         int nr_before;
 212
 213                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
 214                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 215                         if (shrink_ret == -1)
 216                                 break;
 217                         if (shrink_ret < nr_before)
 218                                 ret += nr_before - shrink_ret;
 219                         mod_page_state(slabs_scanned, this_scan);
 220                         total_scan -= this_scan;
 221
 222                         cond_resched();
 223                 }
 224
 225                 shrinker->nr += total_scan;
 226         }
 227         up_read(&shrinker_rwsem);
 228         return ret;
 229 }
 230
 231 /* Called without lock on whether page is mapped, so answer is unstable */
 232 static inline int page_mapping_inuse(struct page *page)
 233 {
 234         struct address_space *mapping;
 235
 236         /* Page is in somebody's page tables. */
 237         if (page_mapped(page))
 238                 return 1;
 239
 240         /* Be more reluctant to reclaim swapcache than pagecache */
 241         if (PageSwapCache(page))
 242                 return 1;
 243
 244         mapping = page_mapping(page);
 245         if (!mapping)
 246                 return 0;
 247
 248         /* File is mmap'd by somebody? */
 249         return mapping_mapped(mapping);
 250 }
 251
 252 static inline int is_page_cache_freeable(struct page *page)
 253 {
 254         return page_count(page) - !!PagePrivate(page) == 2;
 255 }
 256
 257 static int may_write_to_queue(struct backing_dev_info *bdi)
 258 {
 259         if (current->flags & PF_SWAPWRITE)
 260                 return 1;
 261         if (!bdi_write_congested(bdi))
 262                 return 1;
 263         if (bdi == current->backing_dev_info)
 264                 return 1;
 265         return 0;
 266 }
 267
 268 /*
 269  * We detected a synchronous write error writing a page out.  Probably
 270  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 271  * fsync(), msync() or close().
 272  *
 273  * The tricky part is that after writepage we cannot touch the mapping: nothing
 274  * prevents it from being freed up.  But we have a ref on the page and once
 275  * that page is locked, the mapping is pinned.
 276  *
 277  * We're allowed to run sleeping lock_page() here because we know the caller has
 278  * __GFP_FS.
 279  */
 280 static void handle_write_error(struct address_space *mapping,
 281                                 struct page *page, int error)
 282 {
 283         lock_page(page);
 284         if (page_mapping(page) == mapping) {
 285                 if (error == -ENOSPC)
 286                         set_bit(AS_ENOSPC, &mapping->flags);
 287                 else
 288                         set_bit(AS_EIO, &mapping->flags);
 289         }
 290         unlock_page(page);
 291 }
 292
 293 /* possible outcome of pageout() */
 294 typedef enum {
 295         /* failed to write page out, page is locked */
 296         PAGE_KEEP,
 297         /* move page to the active list, page is locked */
 298         PAGE_ACTIVATE,
 299         /* page has been sent to the disk successfully, page is unlocked */
 300         PAGE_SUCCESS,
 301         /* page is clean and locked */
 302         PAGE_CLEAN,
 303 } pageout_t;
 304
 305 /*
 306  * pageout is called by shrink_page_list() for each dirty page.
 307  * Calls ->writepage().
 308  */
 309 static pageout_t pageout(struct page *page, struct address_space *mapping)
 310 {
 311         /*
 312          * If the page is dirty, only perform writeback if that write
 313          * will be non-blocking.  To prevent this allocation from being
 314          * stalled by pagecache activity.  But note that there may be
 315          * stalls if we need to run get_block().  We could test
 316          * PagePrivate for that.
 317          *
 318          * If this process is currently in generic_file_write() against
 319          * this page's queue, we can perform writeback even if that
 320          * will block.
 321          *
 322          * If the page is swapcache, write it back even if that would
 323          * block, for some throttling. This happens by accident, because
 324          * swap_backing_dev_info is bust: it doesn't reflect the
 325          * congestion state of the swapdevs.  Easy to fix, if needed.
 326          * See swapfile.c:page_queue_congested().
 327          */
 328         if (!is_page_cache_freeable(page))
 329                 return PAGE_KEEP;
 330         if (!mapping) {
 331                 /*
 332                  * Some data journaling orphaned pages can have
 333                  * page->mapping == NULL while being dirty with clean buffers.
 334                  */
 335                 if (PagePrivate(page)) {
 336                         if (try_to_free_buffers(page)) {
 337                                 ClearPageDirty(page);
 338                                 printk("%s: orphaned page\n", __FUNCTION__);
 339                                 return PAGE_CLEAN;
 340                         }
 341                 }
 342                 return PAGE_KEEP;
 343         }
 344         if (mapping->a_ops->writepage == NULL)
 345                 return PAGE_ACTIVATE;
 346         if (!may_write_to_queue(mapping->backing_dev_info))
 347                 return PAGE_KEEP;
 348
 349         if (clear_page_dirty_for_io(page)) {
 350                 int res;
 351                 struct writeback_control wbc = {
 352                         .sync_mode = WB_SYNC_NONE,
 353                         .nr_to_write = SWAP_CLUSTER_MAX,
 354                         .range_start = 0,
 355                         .range_end = LLONG_MAX,
 356                         .nonblocking = 1,
 357                         .for_reclaim = 1,
 358                 };
 359
 360                 SetPageReclaim(page);
 361                 res = mapping->a_ops->writepage(page, &wbc);
 362                 if (res < 0)
 363                         handle_write_error(mapping, page, res);
 364                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 365                         ClearPageReclaim(page);
 366                         return PAGE_ACTIVATE;
 367                 }
 368                 if (!PageWriteback(page)) {
 369                         /* synchronous write or broken a_ops? */
 370                         ClearPageReclaim(page);
 371                 }
 372
 373                 return PAGE_SUCCESS;
 374         }
 375
 376         return PAGE_CLEAN;
 377 }
 378
 379 int remove_mapping(struct address_space *mapping, struct page *page)
 380 {
 381         if (!mapping)
 382                 return 0;               /* truncate got there first */
 383
 384         write_lock_irq(&mapping->tree_lock);
 385
 386         /*
 387          * The non-racy check for busy page.  It is critical to check
 388          * PageDirty _after_ making sure that the page is freeable and
 389          * not in use by anybody.       (pagecache + us == 2)
 390          */
 391         if (unlikely(page_count(page) != 2))
 392                 goto cannot_free;
 393         smp_rmb();
 394         if (unlikely(PageDirty(page)))
 395                 goto cannot_free;
 396
 397         if (PageSwapCache(page)) {
 398                 swp_entry_t swap = { .val = page_private(page) };
 399                 __delete_from_swap_cache(page);
 400                 write_unlock_irq(&mapping->tree_lock);
 401                 swap_free(swap);
 402                 __put_page(page);       /* The pagecache ref */
 403                 return 1;
 404         }
 405
 406         __remove_from_page_cache(page);
 407         write_unlock_irq(&mapping->tree_lock);
 408         __put_page(page);
 409         return 1;
 410
 411 cannot_free:
 412         write_unlock_irq(&mapping->tree_lock);
 413         return 0;
 414 }
 415
 416 /*
 417  * shrink_page_list() returns the number of reclaimed pages
 418  */
 419 static unsigned long shrink_page_list(struct list_head *page_list,
 420                                         struct scan_control *sc)
 421 {
 422         LIST_HEAD(ret_pages);
 423         struct pagevec freed_pvec;
 424         int pgactivate = 0;
 425         unsigned long nr_reclaimed = 0;
 426
 427         cond_resched();
 428
 429         pagevec_init(&freed_pvec, 1);
 430         while (!list_empty(page_list)) {
 431                 struct address_space *mapping;
 432                 struct page *page;
 433                 int may_enter_fs;
 434                 int referenced;
 435
 436                 cond_resched();
 437
 438                 page = lru_to_page(page_list);
 439                 list_del(&page->lru);
 440
 441                 if (TestSetPageLocked(page))
 442                         goto keep;
 443
 444                 BUG_ON(PageActive(page));
 445
 446                 sc->nr_scanned++;
 447
 448                 if (!sc->may_swap && page_mapped(page))
 449                         goto keep_locked;
 450
 451                 /* Double the slab pressure for mapped and swapcache pages */
 452                 if (page_mapped(page) || PageSwapCache(page))
 453                         sc->nr_scanned++;
 454
 455                 if (PageWriteback(page))
 456                         goto keep_locked;
 457
 458                 referenced = page_referenced(page, 1);
 459                 /* In active use or really unfreeable?  Activate it. */
 460                 if (referenced && page_mapping_inuse(page))
 461                         goto activate_locked;
 462
 463 #ifdef CONFIG_SWAP
 464                 /*
 465                  * Anonymous process memory has backing store?
 466                  * Try to allocate it some swap space here.
 467                  */
 468                 if (PageAnon(page) && !PageSwapCache(page))
 469                         if (!add_to_swap(page, GFP_ATOMIC))
 470                                 goto activate_locked;
 471 #endif /* CONFIG_SWAP */
 472
 473                 mapping = page_mapping(page);
 474                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 475                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 476
 477                 /*
 478                  * The page is mapped into the page tables of one or more
 479                  * processes. Try to unmap it here.
 480                  */
 481                 if (page_mapped(page) && mapping) {
 482                         switch (try_to_unmap(page, 0)) {
 483                         case SWAP_FAIL:
 484                                 goto activate_locked;
 485                         case SWAP_AGAIN:
 486                                 goto keep_locked;
 487                         case SWAP_SUCCESS:
 488                                 ; /* try to free the page below */
 489                         }
 490                 }
 491
 492                 if (PageDirty(page)) {
 493                         if (referenced)
 494                                 goto keep_locked;
 495                         if (!may_enter_fs)
 496                                 goto keep_locked;
 497                         if (!sc->may_writepage)
 498                                 goto keep_locked;
 499
 500                         /* Page is dirty, try to write it out here */
 501                         switch(pageout(page, mapping)) {
 502                         case PAGE_KEEP:
 503                                 goto keep_locked;
 504                         case PAGE_ACTIVATE:
 505                                 goto activate_locked;
 506                         case PAGE_SUCCESS:
 507                                 if (PageWriteback(page) || PageDirty(page))
 508                                         goto keep;
 509                                 /*
 510                                  * A synchronous write - probably a ramdisk.  Go
 511                                  * ahead and try to reclaim the page.
 512                                  */
 513                                 if (TestSetPageLocked(page))
 514                                         goto keep;
 515                                 if (PageDirty(page) || PageWriteback(page))
 516                                         goto keep_locked;
 517                                 mapping = page_mapping(page);
 518                         case PAGE_CLEAN:
 519                                 ; /* try to free the page below */
 520                         }
 521                 }
 522
 523                 /*
 524                  * If the page has buffers, try to free the buffer mappings
 525                  * associated with this page. If we succeed we try to free
 526                  * the page as well.
 527                  *
 528                  * We do this even if the page is PageDirty().
 529                  * try_to_release_page() does not perform I/O, but it is
 530                  * possible for a page to have PageDirty set, but it is actually
 531                  * clean (all its buffers are clean).  This happens if the
 532                  * buffers were written out directly, with submit_bh(). ext3
 533                  * will do this, as well as the blockdev mapping.
 534                  * try_to_release_page() will discover that cleanness and will
 535                  * drop the buffers and mark the page clean - it can be freed.
 536                  *
 537                  * Rarely, pages can have buffers and no ->mapping.  These are
 538                  * the pages which were not successfully invalidated in
 539                  * truncate_complete_page().  We try to drop those buffers here
 540                  * and if that worked, and the page is no longer mapped into
 541                  * process address space (page_count == 1) it can be freed.
 542                  * Otherwise, leave the page on the LRU so it is swappable.
 543                  */
 544                 if (PagePrivate(page)) {
 545                         if (!try_to_release_page(page, sc->gfp_mask))
 546                                 goto activate_locked;
 547                         if (!mapping && page_count(page) == 1)
 548                                 goto free_it;
 549                 }
 550
 551                 if (!remove_mapping(mapping, page))
 552                         goto keep_locked;
 553
 554 free_it:
 555                 unlock_page(page);
 556                 nr_reclaimed++;
 557                 if (!pagevec_add(&freed_pvec, page))
 558                         __pagevec_release_nonlru(&freed_pvec);
 559                 continue;
 560
 561 activate_locked:
 562                 SetPageActive(page);
 563                 pgactivate++;
 564 keep_locked:
 565                 unlock_page(page);
 566 keep:
 567                 list_add(&page->lru, &ret_pages);
 568                 BUG_ON(PageLRU(page));
 569         }
 570         list_splice(&ret_pages, page_list);
 571         if (pagevec_count(&freed_pvec))
 572                 __pagevec_release_nonlru(&freed_pvec);
 573         mod_page_state(pgactivate, pgactivate);
 574         return nr_reclaimed;
 575 }
 576
 577 /*
 578  * zone->lru_lock is heavily contended.  Some of the functions that
 579  * shrink the lists perform better by taking out a batch of pages
 580  * and working on them outside the LRU lock.
 581  *
 582  * For pagecache intensive workloads, this function is the hottest
 583  * spot in the kernel (apart from copy_*_user functions).
 584  *
 585  * Appropriate locks must be held before calling this function.
 586  *
 587  * @nr_to_scan: The number of pages to look through on the list.
 588  * @src:        The LRU list to pull pages off.
 589  * @dst:        The temp list to put pages on to.
 590  * @scanned:    The number of pages that were scanned.
 591  *
 592  * returns how many pages were moved onto *@dst.
 593  */
 594 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
 595                 struct list_head *src, struct list_head *dst,
 596                 unsigned long *scanned)
 597 {
 598         unsigned long nr_taken = 0;
 599         struct page *page;
 600         unsigned long scan;
 601
 602         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 603                 struct list_head *target;
 604                 page = lru_to_page(src);
 605                 prefetchw_prev_lru_page(page, src, flags);
 606
 607                 BUG_ON(!PageLRU(page));
 608
 609                 list_del(&page->lru);
 610                 target = src;
 611                 if (likely(get_page_unless_zero(page))) {
 612                         /*
 613                          * Be careful not to clear PageLRU until after we're
 614                          * sure the page is not being freed elsewhere -- the
 615                          * page release code relies on it.
 616                          */
 617                         ClearPageLRU(page);
 618                         target = dst;
 619                         nr_taken++;
 620                 } /* else it is being freed elsewhere */
 621
 622                 list_add(&page->lru, target);
 623         }
 624
 625         *scanned = scan;
 626         return nr_taken;
 627 }
 628
 629 /*
 630  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
 631  * of reclaimed pages
 632  */
 633 static unsigned long shrink_inactive_list(unsigned long max_scan,
 634                                 struct zone *zone, struct scan_control *sc)
 635 {
 636         LIST_HEAD(page_list);
 637         struct pagevec pvec;
 638         unsigned long nr_scanned = 0;
 639         unsigned long nr_reclaimed = 0;
 640
 641         pagevec_init(&pvec, 1);
 642
 643         lru_add_drain();
 644         spin_lock_irq(&zone->lru_lock);
 645         do {
 646                 struct page *page;
 647                 unsigned long nr_taken;
 648                 unsigned long nr_scan;
 649                 unsigned long nr_freed;
 650
 651                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
 652                                              &zone->inactive_list,
 653                                              &page_list, &nr_scan);
 654                 zone->nr_inactive -= nr_taken;
 655                 zone->pages_scanned += nr_scan;
 656                 spin_unlock_irq(&zone->lru_lock);
 657
 658                 nr_scanned += nr_scan;
 659                 nr_freed = shrink_page_list(&page_list, sc);
 660                 nr_reclaimed += nr_freed;
 661                 local_irq_disable();
 662                 if (current_is_kswapd()) {
 663                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 664                         __mod_page_state(kswapd_steal, nr_freed);
 665                 } else
 666                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
 667                 __mod_page_state_zone(zone, pgsteal, nr_freed);
 668
 669                 if (nr_taken == 0)
 670                         goto done;
 671
 672                 spin_lock(&zone->lru_lock);
 673                 /*
 674                  * Put back any unfreeable pages.
 675                  */
 676                 while (!list_empty(&page_list)) {
 677                         page = lru_to_page(&page_list);
 678                         BUG_ON(PageLRU(page));
 679                         SetPageLRU(page);
 680                         list_del(&page->lru);
 681                         if (PageActive(page))
 682                                 add_page_to_active_list(zone, page);
 683                         else
 684                                 add_page_to_inactive_list(zone, page);
 685                         if (!pagevec_add(&pvec, page)) {
 686                                 spin_unlock_irq(&zone->lru_lock);
 687                                 __pagevec_release(&pvec);
 688                                 spin_lock_irq(&zone->lru_lock);
 689                         }
 690                 }
 691         } while (nr_scanned < max_scan);
 692         spin_unlock(&zone->lru_lock);
 693 done:
 694         local_irq_enable();
 695         pagevec_release(&pvec);
 696         return nr_reclaimed;
 697 }
 698
 699 /*
 700  * This moves pages from the active list to the inactive list.
 701  *
 702  * We move them the other way if the page is referenced by one or more
 703  * processes, from rmap.
 704  *
 705  * If the pages are mostly unmapped, the processing is fast and it is
 706  * appropriate to hold zone->lru_lock across the whole operation.  But if
 707  * the pages are mapped, the processing is slow (page_referenced()) so we
 708  * should drop zone->lru_lock around each page.  It's impossible to balance
 709  * this, so instead we remove the pages from the LRU while processing them.
 710  * It is safe to rely on PG_active against the non-LRU pages in here because
 711  * nobody will play with that bit on a non-LRU page.
 712  *
 713  * The downside is that we have to touch page->_count against each page.
 714  * But we had to alter page->flags anyway.
 715  */
 716 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
 717                                 struct scan_control *sc)
 718 {
 719         unsigned long pgmoved;
 720         int pgdeactivate = 0;
 721         unsigned long pgscanned;
 722         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 723         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 724         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 725         struct page *page;
 726         struct pagevec pvec;
 727         int reclaim_mapped = 0;
 728
 729         if (sc->may_swap) {
 730                 long mapped_ratio;
 731                 long distress;
 732                 long swap_tendency;
 733
 734                 /*
 735                  * `distress' is a measure of how much trouble we're having
 736                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
 737                  */
 738                 distress = 100 >> zone->prev_priority;
 739
 740                 /*
 741                  * The point of this algorithm is to decide when to start
 742                  * reclaiming mapped memory instead of just pagecache.  Work out
 743                  * how much memory
 744                  * is mapped.
 745                  */
 746                 mapped_ratio = (sc->nr_mapped * 100) / vm_total_pages;
 747
 748                 /*
 749                  * Now decide how much we really want to unmap some pages.  The
 750                  * mapped ratio is downgraded - just because there's a lot of
 751                  * mapped memory doesn't necessarily mean that page reclaim
 752                  * isn't succeeding.
 753                  *
 754                  * The distress ratio is important - we don't want to start
 755                  * going oom.
 756                  *
 757                  * A 100% value of vm_swappiness overrides this algorithm
 758                  * altogether.
 759                  */
 760                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
 761
 762                 /*
 763                  * Now use this metric to decide whether to start moving mapped
 764                  * memory onto the inactive list.
 765                  */
 766                 if (swap_tendency >= 100)
 767                         reclaim_mapped = 1;
 768         }
 769
 770         lru_add_drain();
 771         spin_lock_irq(&zone->lru_lock);
 772         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
 773                                     &l_hold, &pgscanned);
 774         zone->pages_scanned += pgscanned;
 775         zone->nr_active -= pgmoved;
 776         spin_unlock_irq(&zone->lru_lock);
 777
 778         while (!list_empty(&l_hold)) {
 779                 cond_resched();
 780                 page = lru_to_page(&l_hold);
 781                 list_del(&page->lru);
 782                 if (page_mapped(page)) {
 783                         if (!reclaim_mapped ||
 784                             (total_swap_pages == 0 && PageAnon(page)) ||
 785                             page_referenced(page, 0)) {
 786                                 list_add(&page->lru, &l_active);
 787                                 continue;
 788                         }
 789                 }
 790                 list_add(&page->lru, &l_inactive);
 791         }
 792
 793         pagevec_init(&pvec, 1);
 794         pgmoved = 0;
 795         spin_lock_irq(&zone->lru_lock);
 796         while (!list_empty(&l_inactive)) {
 797                 page = lru_to_page(&l_inactive);
 798                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 799                 BUG_ON(PageLRU(page));
 800                 SetPageLRU(page);
 801                 BUG_ON(!PageActive(page));
 802                 ClearPageActive(page);
 803
 804                 list_move(&page->lru, &zone->inactive_list);
 805                 pgmoved++;
 806                 if (!pagevec_add(&pvec, page)) {
 807                         zone->nr_inactive += pgmoved;
 808                         spin_unlock_irq(&zone->lru_lock);
 809                         pgdeactivate += pgmoved;
 810                         pgmoved = 0;
 811                         if (buffer_heads_over_limit)
 812                                 pagevec_strip(&pvec);
 813                         __pagevec_release(&pvec);
 814                         spin_lock_irq(&zone->lru_lock);
 815                 }
 816         }
 817         zone->nr_inactive += pgmoved;
 818         pgdeactivate += pgmoved;
 819         if (buffer_heads_over_limit) {
 820                 spin_unlock_irq(&zone->lru_lock);
 821                 pagevec_strip(&pvec);
 822                 spin_lock_irq(&zone->lru_lock);
 823         }
 824
 825         pgmoved = 0;
 826         while (!list_empty(&l_active)) {
 827                 page = lru_to_page(&l_active);
 828                 prefetchw_prev_lru_page(page, &l_active, flags);
 829                 BUG_ON(PageLRU(page));
 830                 SetPageLRU(page);
 831                 BUG_ON(!PageActive(page));
 832                 list_move(&page->lru, &zone->active_list);
 833                 pgmoved++;
 834                 if (!pagevec_add(&pvec, page)) {
 835                         zone->nr_active += pgmoved;
 836                         pgmoved = 0;
 837                         spin_unlock_irq(&zone->lru_lock);
 838                         __pagevec_release(&pvec);
 839                         spin_lock_irq(&zone->lru_lock);
 840                 }
 841         }
 842         zone->nr_active += pgmoved;
 843         spin_unlock(&zone->lru_lock);
 844
 845         __mod_page_state_zone(zone, pgrefill, pgscanned);
 846         __mod_page_state(pgdeactivate, pgdeactivate);
 847         local_irq_enable();
 848
 849         pagevec_release(&pvec);
 850 }
 851
 852 /*
 853  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 854  */
 855 static unsigned long shrink_zone(int priority, struct zone *zone,
 856                                 struct scan_control *sc)
 857 {
 858         unsigned long nr_active;
 859         unsigned long nr_inactive;
 860         unsigned long nr_to_scan;
 861         unsigned long nr_reclaimed = 0;
 862
 863         atomic_inc(&zone->reclaim_in_progress);
 864
 865         /*
 866          * Add one to `nr_to_scan' just to make sure that the kernel will
 867          * slowly sift through the active list.
 868          */
 869         zone->nr_scan_active += (zone->nr_active >> priority) + 1;
 870         nr_active = zone->nr_scan_active;
 871         if (nr_active >= sc->swap_cluster_max)
 872                 zone->nr_scan_active = 0;
 873         else
 874                 nr_active = 0;
 875
 876         zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
 877         nr_inactive = zone->nr_scan_inactive;
 878         if (nr_inactive >= sc->swap_cluster_max)
 879                 zone->nr_scan_inactive = 0;
 880         else
 881                 nr_inactive = 0;
 882
 883         while (nr_active || nr_inactive) {
 884                 if (nr_active) {
 885                         nr_to_scan = min(nr_active,
 886                                         (unsigned long)sc->swap_cluster_max);
 887                         nr_active -= nr_to_scan;
 888                         shrink_active_list(nr_to_scan, zone, sc);
 889                 }
 890
 891                 if (nr_inactive) {
 892                         nr_to_scan = min(nr_inactive,
 893                                         (unsigned long)sc->swap_cluster_max);
 894                         nr_inactive -= nr_to_scan;
 895                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
 896                                                                 sc);
 897                 }
 898         }
 899
 900         throttle_vm_writeout();
 901
 902         atomic_dec(&zone->reclaim_in_progress);
 903         return nr_reclaimed;
 904 }
 905
 906 /*
 907  * This is the direct reclaim path, for page-allocating processes.  We only
 908  * try to reclaim pages from zones which will satisfy the caller's allocation
 909  * request.
 910  *
 911  * We reclaim from a zone even if that zone is over pages_high.  Because:
 912  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
 913  *    allocation or
 914  * b) The zones may be over pages_high but they must go *over* pages_high to
 915  *    satisfy the `incremental min' zone defense algorithm.
 916  *
 917  * Returns the number of reclaimed pages.
 918  *
 919  * If a zone is deemed to be full of pinned pages then just give it a light
 920  * scan then give up on it.
 921  */
 922 static unsigned long shrink_zones(int priority, struct zone **zones,
 923                                         struct scan_control *sc)
 924 {
 925         unsigned long nr_reclaimed = 0;
 926         int i;
 927
 928         for (i = 0; zones[i] != NULL; i++) {
 929                 struct zone *zone = zones[i];
 930
 931                 if (!populated_zone(zone))
 932                         continue;
 933
 934                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 935                         continue;
 936
 937                 zone->temp_priority = priority;
 938                 if (zone->prev_priority > priority)
 939                         zone->prev_priority = priority;
 940
 941                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
 942                         continue;       /* Let kswapd poll it */
 943
 944                 nr_reclaimed += shrink_zone(priority, zone, sc);
 945         }
 946         return nr_reclaimed;
 947 }
 948
 949 /*
 950  * This is the main entry point to direct page reclaim.
 951  *
 952  * If a full scan of the inactive list fails to free enough memory then we
 953  * are "out of memory" and something needs to be killed.
 954  *
 955  * If the caller is !__GFP_FS then the probability of a failure is reasonably
 956  * high - the zone may be full of dirty or under-writeback pages, which this
 957  * caller can't do much about.  We kick pdflush and take explicit naps in the
 958  * hope that some of these pages can be written.  But if the allocating task
 959  * holds filesystem locks which prevent writeout this might not work, and the
 960  * allocation attempt will fail.
 961  */
 962 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
 963 {
 964         int priority;
 965         int ret = 0;
 966         unsigned long total_scanned = 0;
 967         unsigned long nr_reclaimed = 0;
 968         struct reclaim_state *reclaim_state = current->reclaim_state;
 969         unsigned long lru_pages = 0;
 970         int i;
 971         struct scan_control sc = {
 972                 .gfp_mask = gfp_mask,
 973                 .may_writepage = !laptop_mode,
 974                 .swap_cluster_max = SWAP_CLUSTER_MAX,
 975                 .may_swap = 1,
 976                 .swappiness = vm_swappiness,
 977         };
 978
 979         inc_page_state(allocstall);
 980
 981         for (i = 0; zones[i] != NULL; i++) {
 982                 struct zone *zone = zones[i];
 983
 984                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
 985                         continue;
 986
 987                 zone->temp_priority = DEF_PRIORITY;
 988                 lru_pages += zone->nr_active + zone->nr_inactive;
 989         }
 990
 991         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
 992                 sc.nr_mapped = read_page_state(nr_mapped);
 993                 sc.nr_scanned = 0;
 994                 if (!priority)
 995                         disable_swap_token();
 996                 nr_reclaimed += shrink_zones(priority, zones, &sc);
 997                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
 998                 if (reclaim_state) {
 999                         nr_reclaimed += reclaim_state->reclaimed_slab;
1000                         reclaim_state->reclaimed_slab = 0;
1001                 }
1002                 total_scanned += sc.nr_scanned;
1003                 if (nr_reclaimed >= sc.swap_cluster_max) {
1004                         ret = 1;
1005                         goto out;
1006                 }
1007
1008                 /*
1009                  * Try to write back as many pages as we just scanned.  This
1010                  * tends to cause slow streaming writers to write data to the
1011                  * disk smoothly, at the dirtying rate, which is nice.   But
1012                  * that's undesirable in laptop mode, where we *want* lumpy
1013                  * writeout.  So in laptop mode, write out the whole world.
1014                  */
1015                 if (total_scanned > sc.swap_cluster_max +
1016                                         sc.swap_cluster_max / 2) {
1017                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1018                         sc.may_writepage = 1;
1019                 }
1020
1021                 /* Take a nap, wait for some writeback to complete */
1022                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1023                         blk_congestion_wait(WRITE, HZ/10);
1024         }
1025 out:
1026         for (i = 0; zones[i] != 0; i++) {
1027                 struct zone *zone = zones[i];
1028
1029                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1030                         continue;
1031
1032                 zone->prev_priority = zone->temp_priority;
1033         }
1034         return ret;
1035 }
1036
1037 /*
1038  * For kswapd, balance_pgdat() will work across all this node's zones until
1039  * they are all at pages_high.
1040  *
1041  * Returns the number of pages which were actually freed.
1042  *
1043  * There is special handling here for zones which are full of pinned pages.
1044  * This can happen if the pages are all mlocked, or if they are all used by
1045  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1046  * What we do is to detect the case where all pages in the zone have been
1047  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1048  * dead and from now on, only perform a short scan.  Basically we're polling
1049  * the zone for when the problem goes away.
1050  *
1051  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1052  * zones which have free_pages > pages_high, but once a zone is found to have
1053  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1054  * of the number of free pages in the lower zones.  This interoperates with
1055  * the page allocator fallback scheme to ensure that aging of pages is balanced
1056  * across the zones.
1057  */
1058 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1059 {
1060         int all_zones_ok;
1061         int priority;
1062         int i;
1063         unsigned long total_scanned;
1064         unsigned long nr_reclaimed;
1065         struct reclaim_state *reclaim_state = current->reclaim_state;
1066         struct scan_control sc = {
1067                 .gfp_mask = GFP_KERNEL,
1068                 .may_swap = 1,
1069                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1070                 .swappiness = vm_swappiness,
1071         };
1072
1073 loop_again:
1074         total_scanned = 0;
1075         nr_reclaimed = 0;
1076         sc.may_writepage = !laptop_mode;
1077         sc.nr_mapped = read_page_state(nr_mapped);
1078
1079         inc_page_state(pageoutrun);
1080
1081         for (i = 0; i < pgdat->nr_zones; i++) {
1082                 struct zone *zone = pgdat->node_zones + i;
1083
1084                 zone->temp_priority = DEF_PRIORITY;
1085         }
1086
1087         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1088                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1089                 unsigned long lru_pages = 0;
1090
1091                 /* The swap token gets in the way of swapout... */
1092                 if (!priority)
1093                         disable_swap_token();
1094
1095                 all_zones_ok = 1;
1096
1097                 /*
1098                  * Scan in the highmem->dma direction for the highest
1099                  * zone which needs scanning
1100                  */
1101                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1102                         struct zone *zone = pgdat->node_zones + i;
1103
1104                         if (!populated_zone(zone))
1105                                 continue;
1106
1107                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1108                                 continue;
1109
1110                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1111                                                0, 0)) {
1112                                 end_zone = i;
1113                                 goto scan;
1114                         }
1115                 }
1116                 goto out;
1117 scan:
1118                 for (i = 0; i <= end_zone; i++) {
1119                         struct zone *zone = pgdat->node_zones + i;
1120
1121                         lru_pages += zone->nr_active + zone->nr_inactive;
1122                 }
1123
1124                 /*
1125                  * Now scan the zone in the dma->highmem direction, stopping
1126                  * at the last zone which needs scanning.
1127                  *
1128                  * We do this because the page allocator works in the opposite
1129                  * direction.  This prevents the page allocator from allocating
1130                  * pages behind kswapd's direction of progress, which would
1131                  * cause too much scanning of the lower zones.
1132                  */
1133                 for (i = 0; i <= end_zone; i++) {
1134                         struct zone *zone = pgdat->node_zones + i;
1135                         int nr_slab;
1136
1137                         if (!populated_zone(zone))
1138                                 continue;
1139
1140                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1141                                 continue;
1142
1143                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1144                                                end_zone, 0))
1145                                 all_zones_ok = 0;
1146                         zone->temp_priority = priority;
1147                         if (zone->prev_priority > priority)
1148                                 zone->prev_priority = priority;
1149                         sc.nr_scanned = 0;
1150                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1151                         reclaim_state->reclaimed_slab = 0;
1152                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1153                                                 lru_pages);
1154                         nr_reclaimed += reclaim_state->reclaimed_slab;
1155                         total_scanned += sc.nr_scanned;
1156                         if (zone->all_unreclaimable)
1157                                 continue;
1158                         if (nr_slab == 0 && zone->pages_scanned >=
1159                                     (zone->nr_active + zone->nr_inactive) * 4)
1160                                 zone->all_unreclaimable = 1;
1161                         /*
1162                          * If we've done a decent amount of scanning and
1163                          * the reclaim ratio is low, start doing writepage
1164                          * even in laptop mode
1165                          */
1166                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1167                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1168                                 sc.may_writepage = 1;
1169                 }
1170                 if (all_zones_ok)
1171                         break;          /* kswapd: all done */
1172                 /*
1173                  * OK, kswapd is getting into trouble.  Take a nap, then take
1174                  * another pass across the zones.
1175                  */
1176                 if (total_scanned && priority < DEF_PRIORITY - 2)
1177                         blk_congestion_wait(WRITE, HZ/10);
1178
1179                 /*
1180                  * We do this so kswapd doesn't build up large priorities for
1181                  * example when it is freeing in parallel with allocators. It
1182                  * matches the direct reclaim path behaviour in terms of impact
1183                  * on zone->*_priority.
1184                  */
1185                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1186                         break;
1187         }
1188 out:
1189         for (i = 0; i < pgdat->nr_zones; i++) {
1190                 struct zone *zone = pgdat->node_zones + i;
1191
1192                 zone->prev_priority = zone->temp_priority;
1193         }
1194         if (!all_zones_ok) {
1195                 cond_resched();
1196                 goto loop_again;
1197         }
1198
1199         return nr_reclaimed;
1200 }
1201
1202 /*
1203  * The background pageout daemon, started as a kernel thread
1204  * from the init process.
1205  *
1206  * This basically trickles out pages so that we have _some_
1207  * free memory available even if there is no other activity
1208  * that frees anything up. This is needed for things like routing
1209  * etc, where we otherwise might have all activity going on in
1210  * asynchronous contexts that cannot page things out.
1211  *
1212  * If there are applications that are active memory-allocators
1213  * (most normal use), this basically shouldn't matter.
1214  */
1215 static int kswapd(void *p)
1216 {
1217         unsigned long order;
1218         pg_data_t *pgdat = (pg_data_t*)p;
1219         struct task_struct *tsk = current;
1220         DEFINE_WAIT(wait);
1221         struct reclaim_state reclaim_state = {
1222                 .reclaimed_slab = 0,
1223         };
1224         cpumask_t cpumask;
1225
1226         daemonize("kswapd%d", pgdat->node_id);
1227         cpumask = node_to_cpumask(pgdat->node_id);
1228         if (!cpus_empty(cpumask))
1229                 set_cpus_allowed(tsk, cpumask);
1230         current->reclaim_state = &reclaim_state;
1231
1232         /*
1233          * Tell the memory management that we're a "memory allocator",
1234          * and that if we need more memory we should get access to it
1235          * regardless (see "__alloc_pages()"). "kswapd" should
1236          * never get caught in the normal page freeing logic.
1237          *
1238          * (Kswapd normally doesn't need memory anyway, but sometimes
1239          * you need a small amount of memory in order to be able to
1240          * page out something else, and this flag essentially protects
1241          * us from recursively trying to free more memory as we're
1242          * trying to free the first piece of memory in the first place).
1243          */
1244         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1245
1246         order = 0;
1247         for ( ; ; ) {
1248                 unsigned long new_order;
1249
1250                 try_to_freeze();
1251
1252                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1253                 new_order = pgdat->kswapd_max_order;
1254                 pgdat->kswapd_max_order = 0;
1255                 if (order < new_order) {
1256                         /*
1257                          * Don't sleep if someone wants a larger 'order'
1258                          * allocation
1259                          */
1260                         order = new_order;
1261                 } else {
1262                         schedule();
1263                         order = pgdat->kswapd_max_order;
1264                 }
1265                 finish_wait(&pgdat->kswapd_wait, &wait);
1266
1267                 balance_pgdat(pgdat, order);
1268         }
1269         return 0;
1270 }
1271
1272 /*
1273  * A zone is low on free memory, so wake its kswapd task to service it.
1274  */
1275 void wakeup_kswapd(struct zone *zone, int order)
1276 {
1277         pg_data_t *pgdat;
1278
1279         if (!populated_zone(zone))
1280                 return;
1281
1282         pgdat = zone->zone_pgdat;
1283         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1284                 return;
1285         if (pgdat->kswapd_max_order < order)
1286                 pgdat->kswapd_max_order = order;
1287         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1288                 return;
1289         if (!waitqueue_active(&pgdat->kswapd_wait))
1290                 return;
1291         wake_up_interruptible(&pgdat->kswapd_wait);
1292 }
1293
1294 #ifdef CONFIG_PM
1295 /*
1296  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1297  * from LRU lists system-wide, for given pass and priority, and returns the
1298  * number of reclaimed pages
1299  *
1300  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1301  */
1302 static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1303                                       int prio, struct scan_control *sc)
1304 {
1305         struct zone *zone;
1306         unsigned long nr_to_scan, ret = 0;
1307
1308         for_each_zone(zone) {
1309
1310                 if (!populated_zone(zone))
1311                         continue;
1312
1313                 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1314                         continue;
1315
1316                 /* For pass = 0 we don't shrink the active list */
1317                 if (pass > 0) {
1318                         zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1319                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1320                                 zone->nr_scan_active = 0;
1321                                 nr_to_scan = min(nr_pages, zone->nr_active);
1322                                 shrink_active_list(nr_to_scan, zone, sc);
1323                         }
1324                 }
1325
1326                 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1327                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1328                         zone->nr_scan_inactive = 0;
1329                         nr_to_scan = min(nr_pages, zone->nr_inactive);
1330                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1331                         if (ret >= nr_pages)
1332                                 return ret;
1333                 }
1334         }
1335
1336         return ret;
1337 }
1338
1339 /*
1340  * Try to free `nr_pages' of memory, system-wide, and return the number of
1341  * freed pages.
1342  *
1343  * Rather than trying to age LRUs the aim is to preserve the overall
1344  * LRU order by reclaiming preferentially
1345  * inactive > active > active referenced > active mapped
1346  */
1347 unsigned long shrink_all_memory(unsigned long nr_pages)
1348 {
1349         unsigned long lru_pages, nr_slab;
1350         unsigned long ret = 0;
1351         int pass;
1352         struct reclaim_state reclaim_state;
1353         struct zone *zone;
1354         struct scan_control sc = {
1355                 .gfp_mask = GFP_KERNEL,
1356                 .may_swap = 0,
1357                 .swap_cluster_max = nr_pages,
1358                 .may_writepage = 1,
1359                 .swappiness = vm_swappiness,
1360         };
1361
1362         current->reclaim_state = &reclaim_state;
1363
1364         lru_pages = 0;
1365         for_each_zone(zone)
1366                 lru_pages += zone->nr_active + zone->nr_inactive;
1367
1368         nr_slab = read_page_state(nr_slab);
1369         /* If slab caches are huge, it's better to hit them first */
1370         while (nr_slab >= lru_pages) {
1371                 reclaim_state.reclaimed_slab = 0;
1372                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1373                 if (!reclaim_state.reclaimed_slab)
1374                         break;
1375
1376                 ret += reclaim_state.reclaimed_slab;
1377                 if (ret >= nr_pages)
1378                         goto out;
1379
1380                 nr_slab -= reclaim_state.reclaimed_slab;
1381         }
1382
1383         /*
1384          * We try to shrink LRUs in 5 passes:
1385          * 0 = Reclaim from inactive_list only
1386          * 1 = Reclaim from active list but don't reclaim mapped
1387          * 2 = 2nd pass of type 1
1388          * 3 = Reclaim mapped (normal reclaim)
1389          * 4 = 2nd pass of type 3
1390          */
1391         for (pass = 0; pass < 5; pass++) {
1392                 int prio;
1393
1394                 /* Needed for shrinking slab caches later on */
1395                 if (!lru_pages)
1396                         for_each_zone(zone) {
1397                                 lru_pages += zone->nr_active;
1398                                 lru_pages += zone->nr_inactive;
1399                         }
1400
1401                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1402                 if (pass > 2) {
1403                         sc.may_swap = 1;
1404                         sc.swappiness = 100;
1405                 }
1406
1407                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1408                         unsigned long nr_to_scan = nr_pages - ret;
1409
1410                         sc.nr_mapped = read_page_state(nr_mapped);
1411                         sc.nr_scanned = 0;
1412
1413                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1414                         if (ret >= nr_pages)
1415                                 goto out;
1416
1417                         reclaim_state.reclaimed_slab = 0;
1418                         shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1419                         ret += reclaim_state.reclaimed_slab;
1420                         if (ret >= nr_pages)
1421                                 goto out;
1422
1423                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1424                                 blk_congestion_wait(WRITE, HZ / 10);
1425                 }
1426
1427                 lru_pages = 0;
1428         }
1429
1430         /*
1431          * If ret = 0, we could not shrink LRUs, but there may be something
1432          * in slab caches
1433          */
1434         if (!ret)
1435                 do {
1436                         reclaim_state.reclaimed_slab = 0;
1437                         shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1438                         ret += reclaim_state.reclaimed_slab;
1439                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1440
1441 out:
1442         current->reclaim_state = NULL;
1443
1444         return ret;
1445 }
1446 #endif
1447
1448 #ifdef CONFIG_HOTPLUG_CPU
1449 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1450    not required for correctness.  So if the last cpu in a node goes
1451    away, we get changed to run anywhere: as the first one comes back,
1452    restore their cpu bindings. */
1453 static int cpu_callback(struct notifier_block *nfb,
1454                                   unsigned long action, void *hcpu)
1455 {
1456         pg_data_t *pgdat;
1457         cpumask_t mask;
1458
1459         if (action == CPU_ONLINE) {
1460                 for_each_online_pgdat(pgdat) {
1461                         mask = node_to_cpumask(pgdat->node_id);
1462                         if (any_online_cpu(mask) != NR_CPUS)
1463                                 /* One of our CPUs online: restore mask */
1464                                 set_cpus_allowed(pgdat->kswapd, mask);
1465                 }
1466         }
1467         return NOTIFY_OK;
1468 }
1469 #endif /* CONFIG_HOTPLUG_CPU */
1470
1471 static int __init kswapd_init(void)
1472 {
1473         pg_data_t *pgdat;
1474
1475         swap_setup();
1476         for_each_online_pgdat(pgdat) {
1477                 pid_t pid;
1478
1479                 pid = kernel_thread(kswapd, pgdat, CLONE_KERNEL);
1480                 BUG_ON(pid < 0);
1481                 read_lock(&tasklist_lock);
1482                 pgdat->kswapd = find_task_by_pid(pid);
1483                 read_unlock(&tasklist_lock);
1484         }
1485         hotcpu_notifier(cpu_callback, 0);
1486         return 0;
1487 }
1488
1489 module_init(kswapd_init)
1490
1491 #ifdef CONFIG_NUMA
1492 /*
1493  * Zone reclaim mode
1494  *
1495  * If non-zero call zone_reclaim when the number of free pages falls below
1496  * the watermarks.
1497  *
1498  * In the future we may add flags to the mode. However, the page allocator
1499  * should only have to check that zone_reclaim_mode != 0 before calling
1500  * zone_reclaim().
1501  */
1502 int zone_reclaim_mode __read_mostly;
1503
1504 #define RECLAIM_OFF 0
1505 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1506 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1507 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1508 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1509
1510 /*
1511  * Mininum time between zone reclaim scans
1512  */
1513 int zone_reclaim_interval __read_mostly = 30*HZ;
1514
1515 /*
1516  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1517  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1518  * a zone.
1519  */
1520 #define ZONE_RECLAIM_PRIORITY 4
1521
1522 /*
1523  * Try to free up some pages from this zone through reclaim.
1524  */
1525 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1526 {
1527         /* Minimum pages needed in order to stay on node */
1528         const unsigned long nr_pages = 1 << order;
1529         struct task_struct *p = current;
1530         struct reclaim_state reclaim_state;
1531         int priority;
1532         unsigned long nr_reclaimed = 0;
1533         struct scan_control sc = {
1534                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1535                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1536                 .nr_mapped = read_page_state(nr_mapped),
1537                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1538                                         SWAP_CLUSTER_MAX),
1539                 .gfp_mask = gfp_mask,
1540                 .swappiness = vm_swappiness,
1541         };
1542
1543         disable_swap_token();
1544         cond_resched();
1545         /*
1546          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1547          * and we also need to be able to write out pages for RECLAIM_WRITE
1548          * and RECLAIM_SWAP.
1549          */
1550         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1551         reclaim_state.reclaimed_slab = 0;
1552         p->reclaim_state = &reclaim_state;
1553
1554         /*
1555          * Free memory by calling shrink zone with increasing priorities
1556          * until we have enough memory freed.
1557          */
1558         priority = ZONE_RECLAIM_PRIORITY;
1559         do {
1560                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1561                 priority--;
1562         } while (priority >= 0 && nr_reclaimed < nr_pages);
1563
1564         if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1565                 /*
1566                  * shrink_slab() does not currently allow us to determine how
1567                  * many pages were freed in this zone. So we just shake the slab
1568                  * a bit and then go off node for this particular allocation
1569                  * despite possibly having freed enough memory to allocate in
1570                  * this zone.  If we freed local memory then the next
1571                  * allocations will be local again.
1572                  *
1573                  * shrink_slab will free memory on all zones and may take
1574                  * a long time.
1575                  */
1576                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1577         }
1578
1579         p->reclaim_state = NULL;
1580         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1581
1582         if (nr_reclaimed == 0) {
1583                 /*
1584                  * We were unable to reclaim enough pages to stay on node.  We
1585                  * now allow off node accesses for a certain time period before
1586                  * trying again to reclaim pages from the local zone.
1587                  */
1588                 zone->last_unsuccessful_zone_reclaim = jiffies;
1589         }
1590
1591         return nr_reclaimed >= nr_pages;
1592 }
1593
1594 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1595 {
1596         cpumask_t mask;
1597         int node_id;
1598
1599         /*
1600          * Do not reclaim if there was a recent unsuccessful attempt at zone
1601          * reclaim.  In that case we let allocations go off node for the
1602          * zone_reclaim_interval.  Otherwise we would scan for each off-node
1603          * page allocation.
1604          */
1605         if (time_before(jiffies,
1606                 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1607                         return 0;
1608
1609         /*
1610          * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1611          * not have reclaimable pages and if we should not delay the allocation
1612          * then do not scan.
1613          */
1614         if (!(gfp_mask & __GFP_WAIT) ||
1615                 zone->all_unreclaimable ||
1616                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1617                 (current->flags & PF_MEMALLOC))
1618                         return 0;
1619
1620         /*
1621          * Only run zone reclaim on the local zone or on zones that do not
1622          * have associated processors. This will favor the local processor
1623          * over remote processors and spread off node memory allocations
1624          * as wide as possible.
1625          */
1626         node_id = zone->zone_pgdat->node_id;
1627         mask = node_to_cpumask(node_id);
1628         if (!cpus_empty(mask) && node_id != numa_node_id())
1629                 return 0;
1630         return __zone_reclaim(zone, gfp_mask, order);
1631 }
1632 #endif