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