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/suspend.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  27                                         buffer_heads_over_limit */
  28 #include <linux/mm_inline.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/rmap.h>
  32 #include <linux/topology.h>
  33 #include <linux/cpu.h>
  34 #include <linux/notifier.h>
  35 #include <linux/rwsem.h>
  36 #include <linux/ckrm_mem.h>
  37
  38 #include <asm/tlbflush.h>
  39 #include <asm/div64.h>
  40
  41 #include <linux/swapops.h>
  42
  43 /* possible outcome of pageout() */
  44 typedef enum {
  45         /* failed to write page out, page is locked */
  46         PAGE_KEEP,
  47         /* move page to the active list, page is locked */
  48         PAGE_ACTIVATE,
  49         /* page has been sent to the disk successfully, page is unlocked */
  50         PAGE_SUCCESS,
  51         /* page is clean and locked */
  52         PAGE_CLEAN,
  53 } pageout_t;
  54
  55 struct scan_control {
  56         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  57         unsigned long nr_to_scan;
  58
  59         /* Incremented by the number of inactive pages that were scanned */
  60         unsigned long nr_scanned;
  61
  62         /* Incremented by the number of pages reclaimed */
  63         unsigned long nr_reclaimed;
  64
  65         unsigned long nr_mapped;        /* From page_state */
  66
  67         /* How many pages shrink_cache() should reclaim */
  68         int nr_to_reclaim;
  69
  70         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  71         unsigned int priority;
  72
  73         /* This context's GFP mask */
  74         unsigned int gfp_mask;
  75
  76         int may_writepage;
  77 };
  78
  79 /*
  80  * The list of shrinker callbacks used by to apply pressure to
  81  * ageable caches.
  82  */
  83 struct shrinker {
  84         shrinker_t              shrinker;
  85         struct list_head        list;
  86         int                     seeks;  /* seeks to recreate an obj */
  87         long                    nr;     /* objs pending delete */
  88 };
  89
  90 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  91
  92 #ifdef ARCH_HAS_PREFETCH
  93 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  94         do {                                                            \
  95                 if ((_page)->lru.prev != _base) {                       \
  96                         struct page *prev;                              \
  97                                                                         \
  98                         prev = lru_to_page(&(_page->lru));              \
  99                         prefetch(&prev->_field);                        \
 100                 }                                                       \
 101         } while (0)
 102 #else
 103 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 104 #endif
 105
 106 #ifdef ARCH_HAS_PREFETCHW
 107 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 108         do {                                                            \
 109                 if ((_page)->lru.prev != _base) {                       \
 110                         struct page *prev;                              \
 111                                                                         \
 112                         prev = lru_to_page(&(_page->lru));              \
 113                         prefetchw(&prev->_field);                       \
 114                 }                                                       \
 115         } while (0)
 116 #else
 117 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 118 #endif
 119
 120 /*
 121  * From 0 .. 100.  Higher means more swappy.
 122  */
 123 int vm_swappiness = 60;
 124 static long total_memory;
 125
 126 static LIST_HEAD(shrinker_list);
 127 static DECLARE_RWSEM(shrinker_rwsem);
 128
 129 /*
 130  * Add a shrinker callback to be called from the vm
 131  */
 132 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 133 {
 134         struct shrinker *shrinker;
 135
 136         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 137         if (shrinker) {
 138                 shrinker->shrinker = theshrinker;
 139                 shrinker->seeks = seeks;
 140                 shrinker->nr = 0;
 141                 down_write(&shrinker_rwsem);
 142                 list_add(&shrinker->list, &shrinker_list);
 143                 up_write(&shrinker_rwsem);
 144         }
 145         return shrinker;
 146 }
 147 EXPORT_SYMBOL(set_shrinker);
 148
 149 /*
 150  * Remove one
 151  */
 152 void remove_shrinker(struct shrinker *shrinker)
 153 {
 154         down_write(&shrinker_rwsem);
 155         list_del(&shrinker->list);
 156         up_write(&shrinker_rwsem);
 157         kfree(shrinker);
 158 }
 159 EXPORT_SYMBOL(remove_shrinker);
 160
 161 #define SHRINK_BATCH 128
 162 /*
 163  * Call the shrink functions to age shrinkable caches
 164  *
 165  * Here we assume it costs one seek to replace a lru page and that it also
 166  * takes a seek to recreate a cache object.  With this in mind we age equal
 167  * percentages of the lru and ageable caches.  This should balance the seeks
 168  * generated by these structures.
 169  *
 170  * If the vm encounted mapped pages on the LRU it increase the pressure on
 171  * slab to avoid swapping.
 172  *
 173  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 174  *
 175  * `lru_pages' represents the number of on-LRU pages in all the zones which
 176  * are eligible for the caller's allocation attempt.  It is used for balancing
 177  * slab reclaim versus page reclaim.
 178  */
 179 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 180                         unsigned long lru_pages)
 181 {
 182         struct shrinker *shrinker;
 183
 184         if (scanned == 0)
 185                 scanned = SWAP_CLUSTER_MAX;
 186
 187         if (!down_read_trylock(&shrinker_rwsem))
 188                 return 0;
 189
 190         list_for_each_entry(shrinker, &shrinker_list, list) {
 191                 unsigned long long delta;
 192                 unsigned long total_scan;
 193
 194                 delta = (4 * scanned) / shrinker->seeks;
 195                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 196                 do_div(delta, lru_pages + 1);
 197                 shrinker->nr += delta;
 198                 if (shrinker->nr < 0)
 199                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 200
 201                 total_scan = shrinker->nr;
 202                 shrinker->nr = 0;
 203
 204                 while (total_scan >= SHRINK_BATCH) {
 205                         long this_scan = SHRINK_BATCH;
 206                         int shrink_ret;
 207
 208                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 209                         if (shrink_ret == -1)
 210                                 break;
 211                         mod_page_state(slabs_scanned, this_scan);
 212                         total_scan -= this_scan;
 213
 214                         cond_resched();
 215                 }
 216
 217                 shrinker->nr += total_scan;
 218         }
 219         up_read(&shrinker_rwsem);
 220         return 0;
 221 }
 222
 223 /* Called without lock on whether page is mapped, so answer is unstable */
 224 static inline int page_mapping_inuse(struct page *page)
 225 {
 226         struct address_space *mapping;
 227
 228         /* Page is in somebody's page tables. */
 229         if (page_mapped(page))
 230                 return 1;
 231
 232         /* Be more reluctant to reclaim swapcache than pagecache */
 233         if (PageSwapCache(page))
 234                 return 1;
 235
 236         mapping = page_mapping(page);
 237         if (!mapping)
 238                 return 0;
 239
 240         /* File is mmap'd by somebody? */
 241         return mapping_mapped(mapping);
 242 }
 243
 244 static inline int is_page_cache_freeable(struct page *page)
 245 {
 246         return page_count(page) - !!PagePrivate(page) == 2;
 247 }
 248
 249 static int may_write_to_queue(struct backing_dev_info *bdi)
 250 {
 251         if (current_is_kswapd())
 252                 return 1;
 253         if (current_is_pdflush())       /* This is unlikely, but why not... */
 254                 return 1;
 255         if (!bdi_write_congested(bdi))
 256                 return 1;
 257         if (bdi == current->backing_dev_info)
 258                 return 1;
 259         return 0;
 260 }
 261
 262 /*
 263  * We detected a synchronous write error writing a page out.  Probably
 264  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 265  * fsync(), msync() or close().
 266  *
 267  * The tricky part is that after writepage we cannot touch the mapping: nothing
 268  * prevents it from being freed up.  But we have a ref on the page and once
 269  * that page is locked, the mapping is pinned.
 270  *
 271  * We're allowed to run sleeping lock_page() here because we know the caller has
 272  * __GFP_FS.
 273  */
 274 static void handle_write_error(struct address_space *mapping,
 275                                 struct page *page, int error)
 276 {
 277         lock_page(page);
 278         if (page_mapping(page) == mapping) {
 279                 if (error == -ENOSPC)
 280                         set_bit(AS_ENOSPC, &mapping->flags);
 281                 else
 282                         set_bit(AS_EIO, &mapping->flags);
 283         }
 284         unlock_page(page);
 285 }
 286
 287 /*
 288  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 289  */
 290 static pageout_t pageout(struct page *page, struct address_space *mapping)
 291 {
 292         /*
 293          * If the page is dirty, only perform writeback if that write
 294          * will be non-blocking.  To prevent this allocation from being
 295          * stalled by pagecache activity.  But note that there may be
 296          * stalls if we need to run get_block().  We could test
 297          * PagePrivate for that.
 298          *
 299          * If this process is currently in generic_file_write() against
 300          * this page's queue, we can perform writeback even if that
 301          * will block.
 302          *
 303          * If the page is swapcache, write it back even if that would
 304          * block, for some throttling. This happens by accident, because
 305          * swap_backing_dev_info is bust: it doesn't reflect the
 306          * congestion state of the swapdevs.  Easy to fix, if needed.
 307          * See swapfile.c:page_queue_congested().
 308          */
 309         if (!is_page_cache_freeable(page))
 310                 return PAGE_KEEP;
 311         if (!mapping)
 312                 return PAGE_KEEP;
 313         if (mapping->a_ops->writepage == NULL)
 314                 return PAGE_ACTIVATE;
 315         if (!may_write_to_queue(mapping->backing_dev_info))
 316                 return PAGE_KEEP;
 317
 318         if (clear_page_dirty_for_io(page)) {
 319                 int res;
 320                 struct writeback_control wbc = {
 321                         .sync_mode = WB_SYNC_NONE,
 322                         .nr_to_write = SWAP_CLUSTER_MAX,
 323                         .nonblocking = 1,
 324                         .for_reclaim = 1,
 325                 };
 326
 327                 SetPageReclaim(page);
 328                 res = mapping->a_ops->writepage(page, &wbc);
 329                 if (res < 0)
 330                         handle_write_error(mapping, page, res);
 331                 if (res == WRITEPAGE_ACTIVATE) {
 332                         ClearPageReclaim(page);
 333                         return PAGE_ACTIVATE;
 334                 }
 335                 if (!PageWriteback(page)) {
 336                         /* synchronous write or broken a_ops? */
 337                         ClearPageReclaim(page);
 338                 }
 339
 340                 return PAGE_SUCCESS;
 341         }
 342
 343         return PAGE_CLEAN;
 344 }
 345
 346 /*
 347  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 348  */
 349 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 350 {
 351         LIST_HEAD(ret_pages);
 352         struct pagevec freed_pvec;
 353         int pgactivate = 0;
 354         int reclaimed = 0;
 355
 356         cond_resched();
 357
 358         pagevec_init(&freed_pvec, 1);
 359         while (!list_empty(page_list)) {
 360                 struct address_space *mapping;
 361                 struct page *page;
 362                 int may_enter_fs;
 363                 int referenced;
 364
 365                 page = lru_to_page(page_list);
 366                 list_del(&page->lru);
 367
 368                 if (TestSetPageLocked(page))
 369                         goto keep;
 370
 371                 BUG_ON(PageActive(page));
 372
 373                 if (PageWriteback(page))
 374                         goto keep_locked;
 375
 376                 sc->nr_scanned++;
 377                 /* Double the slab pressure for mapped and swapcache pages */
 378                 if (page_mapped(page) || PageSwapCache(page))
 379                         sc->nr_scanned++;
 380
 381                 referenced = page_referenced(page, 1, sc->priority <= 0);
 382                 /* In active use or really unfreeable?  Activate it. */
 383                 if (referenced && page_mapping_inuse(page))
 384                         goto activate_locked;
 385
 386 #ifdef CONFIG_SWAP
 387                 /*
 388                  * Anonymous process memory has backing store?
 389                  * Try to allocate it some swap space here.
 390                  */
 391                 if (PageAnon(page) && !PageSwapCache(page)) {
 392                         if (!add_to_swap(page))
 393                                 goto activate_locked;
 394                 }
 395 #endif /* CONFIG_SWAP */
 396
 397                 mapping = page_mapping(page);
 398                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 399                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 400
 401                 /*
 402                  * The page is mapped into the page tables of one or more
 403                  * processes. Try to unmap it here.
 404                  */
 405                 if (page_mapped(page) && mapping) {
 406                         switch (try_to_unmap(page)) {
 407                         case SWAP_FAIL:
 408                                 goto activate_locked;
 409                         case SWAP_AGAIN:
 410                                 goto keep_locked;
 411                         case SWAP_SUCCESS:
 412                                 ; /* try to free the page below */
 413                         }
 414                 }
 415
 416                 if (PageDirty(page)) {
 417                         if (referenced)
 418                                 goto keep_locked;
 419                         if (!may_enter_fs)
 420                                 goto keep_locked;
 421                         if (laptop_mode && !sc->may_writepage)
 422                                 goto keep_locked;
 423
 424                         /* Page is dirty, try to write it out here */
 425                         switch(pageout(page, mapping)) {
 426                         case PAGE_KEEP:
 427                                 goto keep_locked;
 428                         case PAGE_ACTIVATE:
 429                                 goto activate_locked;
 430                         case PAGE_SUCCESS:
 431                                 if (PageWriteback(page) || PageDirty(page))
 432                                         goto keep;
 433                                 /*
 434                                  * A synchronous write - probably a ramdisk.  Go
 435                                  * ahead and try to reclaim the page.
 436                                  */
 437                                 if (TestSetPageLocked(page))
 438                                         goto keep;
 439                                 if (PageDirty(page) || PageWriteback(page))
 440                                         goto keep_locked;
 441                                 mapping = page_mapping(page);
 442                         case PAGE_CLEAN:
 443                                 ; /* try to free the page below */
 444                         }
 445                 }
 446
 447                 /*
 448                  * If the page has buffers, try to free the buffer mappings
 449                  * associated with this page. If we succeed we try to free
 450                  * the page as well.
 451                  *
 452                  * We do this even if the page is PageDirty().
 453                  * try_to_release_page() does not perform I/O, but it is
 454                  * possible for a page to have PageDirty set, but it is actually
 455                  * clean (all its buffers are clean).  This happens if the
 456                  * buffers were written out directly, with submit_bh(). ext3
 457                  * will do this, as well as the blockdev mapping.
 458                  * try_to_release_page() will discover that cleanness and will
 459                  * drop the buffers and mark the page clean - it can be freed.
 460                  *
 461                  * Rarely, pages can have buffers and no ->mapping.  These are
 462                  * the pages which were not successfully invalidated in
 463                  * truncate_complete_page().  We try to drop those buffers here
 464                  * and if that worked, and the page is no longer mapped into
 465                  * process address space (page_count == 1) it can be freed.
 466                  * Otherwise, leave the page on the LRU so it is swappable.
 467                  */
 468                 if (PagePrivate(page)) {
 469                         if (!try_to_release_page(page, sc->gfp_mask))
 470                                 goto activate_locked;
 471                         if (!mapping && page_count(page) == 1)
 472                                 goto free_it;
 473                 }
 474
 475                 if (!mapping)
 476                         goto keep_locked;       /* truncate got there first */
 477
 478                 spin_lock_irq(&mapping->tree_lock);
 479
 480                 /*
 481                  * The non-racy check for busy page.  It is critical to check
 482                  * PageDirty _after_ making sure that the page is freeable and
 483                  * not in use by anybody.       (pagecache + us == 2)
 484                  */
 485                 if (page_count(page) != 2 || PageDirty(page)) {
 486                         spin_unlock_irq(&mapping->tree_lock);
 487                         goto keep_locked;
 488                 }
 489
 490 #ifdef CONFIG_SWAP
 491                 if (PageSwapCache(page)) {
 492                         swp_entry_t swap = { .val = page->private };
 493                         __delete_from_swap_cache(page);
 494                         spin_unlock_irq(&mapping->tree_lock);
 495                         swap_free(swap);
 496                         __put_page(page);       /* The pagecache ref */
 497                         goto free_it;
 498                 }
 499 #endif /* CONFIG_SWAP */
 500
 501                 __remove_from_page_cache(page);
 502                 spin_unlock_irq(&mapping->tree_lock);
 503                 __put_page(page);
 504
 505 free_it:
 506                 unlock_page(page);
 507                 reclaimed++;
 508                 if (!pagevec_add(&freed_pvec, page))
 509                         __pagevec_release_nonlru(&freed_pvec);
 510                 continue;
 511
 512 activate_locked:
 513                 SetPageActive(page);
 514                 pgactivate++;
 515 keep_locked:
 516                 unlock_page(page);
 517 keep:
 518                 list_add(&page->lru, &ret_pages);
 519                 BUG_ON(PageLRU(page));
 520         }
 521         list_splice(&ret_pages, page_list);
 522         if (pagevec_count(&freed_pvec))
 523                 __pagevec_release_nonlru(&freed_pvec);
 524         mod_page_state(pgactivate, pgactivate);
 525         sc->nr_reclaimed += reclaimed;
 526         return reclaimed;
 527 }
 528
 529 /*
 530  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 531  * a batch of pages and working on them outside the lock.  Any pages which were
 532  * not freed will be added back to the LRU.
 533  *
 534  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 535  *
 536  * For pagecache intensive workloads, the first loop here is the hottest spot
 537  * in the kernel (apart from the copy_*_user functions).
 538  */
 539 #ifdef CONFIG_CKRM_RES_MEM
 540 static void shrink_cache(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 541 #else
 542 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 543 #endif
 544 {
 545         LIST_HEAD(page_list);
 546         struct pagevec pvec;
 547         int max_scan = sc->nr_to_scan;
 548 #ifdef CONFIG_CKRM_RES_MEM
 549         struct zone *zone = ckrm_zone->zone;
 550         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 551         struct list_head *active_list = &ckrm_zone->active_list;
 552 #else
 553         struct list_head *inactive_list = &zone->inactive_list;
 554         struct list_head *active_list = &zone->active_list;
 555 #endif
 556
 557         pagevec_init(&pvec, 1);
 558
 559         lru_add_drain();
 560         spin_lock_irq(&zone->lru_lock);
 561         while (max_scan > 0) {
 562                 struct page *page;
 563                 int nr_taken = 0;
 564                 int nr_scan = 0;
 565                 int nr_freed;
 566
 567                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 568                                 !list_empty(inactive_list)) {
 569                         page = lru_to_page(inactive_list);
 570
 571                         prefetchw_prev_lru_page(page,
 572                                                 inactive_list, flags);
 573
 574                         if (!TestClearPageLRU(page))
 575                                 BUG();
 576                         list_del(&page->lru);
 577                         if (get_page_testone(page)) {
 578                                 /*
 579                                  * It is being freed elsewhere
 580                                  */
 581                                 __put_page(page);
 582                                 SetPageLRU(page);
 583                                 list_add(&page->lru, inactive_list);
 584                                 continue;
 585                         }
 586                         list_add(&page->lru, &page_list);
 587                         nr_taken++;
 588                 }
 589                 zone->nr_inactive -= nr_taken;
 590                 ckrm_zone_sub_inactive(ckrm_zone, nr_taken);
 591                 spin_unlock_irq(&zone->lru_lock);
 592
 593                 if (nr_taken == 0)
 594                         goto done;
 595
 596                 max_scan -= nr_scan;
 597                 if (current_is_kswapd())
 598                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 599                 else
 600                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 601                 nr_freed = shrink_list(&page_list, sc);
 602                 if (current_is_kswapd())
 603                         mod_page_state(kswapd_steal, nr_freed);
 604                 mod_page_state_zone(zone, pgsteal, nr_freed);
 605                 sc->nr_to_reclaim -= nr_freed;
 606
 607                 spin_lock_irq(&zone->lru_lock);
 608                 /*
 609                  * Put back any unfreeable pages.
 610                  */
 611                 while (!list_empty(&page_list)) {
 612                         page = lru_to_page(&page_list);
 613                         if (TestSetPageLRU(page))
 614                                 BUG();
 615                         list_del(&page->lru);
 616                         if (PageActive(page)) {
 617                                 ckrm_zone_add_active(ckrm_zone, 1);
 618                                 zone->nr_active++;
 619                                 list_add(&page->lru, active_list);
 620                         } else {
 621                                 ckrm_zone_add_inactive(ckrm_zone, 1);
 622                                 zone->nr_inactive++;
 623                                 list_add(&page->lru, inactive_list);
 624                         }
 625                         if (!pagevec_add(&pvec, page)) {
 626                                 spin_unlock_irq(&zone->lru_lock);
 627                                 __pagevec_release(&pvec);
 628                                 spin_lock_irq(&zone->lru_lock);
 629                         }
 630                 }
 631         }
 632         spin_unlock_irq(&zone->lru_lock);
 633 done:
 634         pagevec_release(&pvec);
 635 }
 636
 637 /*
 638  * This moves pages from the active list to the inactive list.
 639  *
 640  * We move them the other way if the page is referenced by one or more
 641  * processes, from rmap.
 642  *
 643  * If the pages are mostly unmapped, the processing is fast and it is
 644  * appropriate to hold zone->lru_lock across the whole operation.  But if
 645  * the pages are mapped, the processing is slow (page_referenced()) so we
 646  * should drop zone->lru_lock around each page.  It's impossible to balance
 647  * this, so instead we remove the pages from the LRU while processing them.
 648  * It is safe to rely on PG_active against the non-LRU pages in here because
 649  * nobody will play with that bit on a non-LRU page.
 650  *
 651  * The downside is that we have to touch page->_count against each page.
 652  * But we had to alter page->flags anyway.
 653  */
 654 static void
 655 #ifdef CONFIG_CKRM_RES_MEM
 656 refill_inactive_zone(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 657 #else
 658 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 659 #endif
 660 {
 661         int pgmoved;
 662         int pgdeactivate = 0;
 663         int pgscanned = 0;
 664         int nr_pages = sc->nr_to_scan;
 665         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 666         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 667         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 668         struct page *page;
 669         struct pagevec pvec;
 670         int reclaim_mapped = 0;
 671         long mapped_ratio;
 672         long distress;
 673         long swap_tendency;
 674 #ifdef CONFIG_CKRM_RES_MEM
 675         struct zone *zone = ckrm_zone->zone;
 676         struct list_head *active_list = &ckrm_zone->active_list;
 677         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 678 #else
 679         struct list_head *active_list = &zone->active_list;
 680         struct list_head *inactive_list = &zone->inactive_list;
 681 #endif
 682
 683         lru_add_drain();
 684         pgmoved = 0;
 685         spin_lock_irq(&zone->lru_lock);
 686         while (pgscanned < nr_pages && !list_empty(active_list)) {
 687                 page = lru_to_page(active_list);
 688                 prefetchw_prev_lru_page(page, active_list, flags);
 689                 if (!TestClearPageLRU(page))
 690                         BUG();
 691                 list_del(&page->lru);
 692                 if (get_page_testone(page)) {
 693                         /*
 694                          * It was already free!  release_pages() or put_page()
 695                          * are about to remove it from the LRU and free it. So
 696                          * put the refcount back and put the page back on the
 697                          * LRU
 698                          */
 699                         __put_page(page);
 700                         SetPageLRU(page);
 701                         list_add(&page->lru, active_list);
 702                 } else {
 703                         list_add(&page->lru, &l_hold);
 704                         pgmoved++;
 705                 }
 706                 pgscanned++;
 707         }
 708         zone->pages_scanned += pgscanned;
 709         zone->nr_active -= pgmoved;
 710         ckrm_zone_sub_active(ckrm_zone, pgmoved);
 711         spin_unlock_irq(&zone->lru_lock);
 712
 713         /*
 714          * `distress' is a measure of how much trouble we're having reclaiming
 715          * pages.  0 -> no problems.  100 -> great trouble.
 716          */
 717         distress = 100 >> zone->prev_priority;
 718
 719         /*
 720          * The point of this algorithm is to decide when to start reclaiming
 721          * mapped memory instead of just pagecache.  Work out how much memory
 722          * is mapped.
 723          */
 724         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 725
 726         /*
 727          * Now decide how much we really want to unmap some pages.  The mapped
 728          * ratio is downgraded - just because there's a lot of mapped memory
 729          * doesn't necessarily mean that page reclaim isn't succeeding.
 730          *
 731          * The distress ratio is important - we don't want to start going oom.
 732          *
 733          * A 100% value of vm_swappiness overrides this algorithm altogether.
 734          */
 735         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 736
 737         /*
 738          * Now use this metric to decide whether to start moving mapped memory
 739          * onto the inactive list.
 740          */
 741         if (swap_tendency >= 100)
 742                 reclaim_mapped = 1;
 743
 744         while (!list_empty(&l_hold)) {
 745                 page = lru_to_page(&l_hold);
 746                 list_del(&page->lru);
 747                 if (page_mapped(page)) {
 748                         if (!reclaim_mapped ||
 749                             (total_swap_pages == 0 && PageAnon(page)) ||
 750                             page_referenced(page, 0, sc->priority <= 0)) {
 751                                 list_add(&page->lru, &l_active);
 752                                 continue;
 753                         }
 754                 }
 755                 list_add(&page->lru, &l_inactive);
 756         }
 757
 758         pagevec_init(&pvec, 1);
 759         pgmoved = 0;
 760         spin_lock_irq(&zone->lru_lock);
 761         while (!list_empty(&l_inactive)) {
 762                 page = lru_to_page(&l_inactive);
 763                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 764                 if (TestSetPageLRU(page))
 765                         BUG();
 766                 if (!TestClearPageActive(page))
 767                         BUG();
 768                 list_move(&page->lru, inactive_list);
 769                 pgmoved++;
 770                 if (!pagevec_add(&pvec, page)) {
 771                         zone->nr_inactive += pgmoved;
 772                         ckrm_zone_add_inactive(ckrm_zone, pgmoved);
 773                         spin_unlock_irq(&zone->lru_lock);
 774                         pgdeactivate += pgmoved;
 775                         pgmoved = 0;
 776                         if (buffer_heads_over_limit)
 777                                 pagevec_strip(&pvec);
 778                         __pagevec_release(&pvec);
 779                         spin_lock_irq(&zone->lru_lock);
 780                 }
 781         }
 782         zone->nr_inactive += pgmoved;
 783         ckrm_zone_add_inactive(ckrm_zone, pgmoved);
 784         pgdeactivate += pgmoved;
 785         if (buffer_heads_over_limit) {
 786                 spin_unlock_irq(&zone->lru_lock);
 787                 pagevec_strip(&pvec);
 788                 spin_lock_irq(&zone->lru_lock);
 789         }
 790
 791         pgmoved = 0;
 792         while (!list_empty(&l_active)) {
 793                 page = lru_to_page(&l_active);
 794                 prefetchw_prev_lru_page(page, &l_active, flags);
 795                 if (TestSetPageLRU(page))
 796                         BUG();
 797                 BUG_ON(!PageActive(page));
 798                 list_move(&page->lru, active_list);
 799                 pgmoved++;
 800                 if (!pagevec_add(&pvec, page)) {
 801                         zone->nr_active += pgmoved;
 802                         ckrm_zone_add_active(ckrm_zone, pgmoved);
 803                         pgmoved = 0;
 804                         spin_unlock_irq(&zone->lru_lock);
 805                         __pagevec_release(&pvec);
 806                         spin_lock_irq(&zone->lru_lock);
 807                 }
 808         }
 809         zone->nr_active += pgmoved;
 810         ckrm_zone_add_active(ckrm_zone, pgmoved);
 811         spin_unlock_irq(&zone->lru_lock);
 812         pagevec_release(&pvec);
 813
 814         mod_page_state_zone(zone, pgrefill, pgscanned);
 815         mod_page_state(pgdeactivate, pgdeactivate);
 816 }
 817
 818 #ifdef CONFIG_CKRM_RES_MEM
 819 static void
 820 shrink_ckrmzone(struct ckrm_zone *czone, struct scan_control *sc)
 821 {
 822         while (czone->shrink_active || czone->shrink_inactive) {
 823                 if (czone->shrink_active) {
 824                         sc->nr_to_scan = min(czone->shrink_active,
 825                                         (unsigned long)SWAP_CLUSTER_MAX);
 826                         czone->shrink_active -= sc->nr_to_scan;
 827                         refill_inactive_zone(czone, sc);
 828                 }
 829                 if (czone->shrink_inactive) {
 830                         sc->nr_to_scan = min(czone->shrink_inactive,
 831                                         (unsigned long)SWAP_CLUSTER_MAX);
 832                         czone->shrink_inactive -= sc->nr_to_scan;
 833                         shrink_cache(czone, sc);
 834                         if (sc->nr_to_reclaim <= 0) {
 835                                 czone->shrink_active = 0;
 836                                 czone->shrink_inactive = 0;
 837                                 break;
 838                         }
 839                 }
 840         }
 841 }
 842
 843 /* FIXME: This function needs to be given more thought. */
 844 static void
 845 ckrm_shrink_class(struct ckrm_mem_res *cls)
 846 {
 847         struct scan_control sc;
 848         struct zone *zone;
 849         int zindex = 0, cnt, act_credit = 0, inact_credit = 0;
 850
 851         sc.nr_mapped = read_page_state(nr_mapped);
 852         sc.nr_scanned = 0;
 853         sc.nr_reclaimed = 0;
 854         sc.priority = 0; // always very high priority
 855
 856         for_each_zone(zone) {
 857                 int zone_total, zone_limit, active_limit,
 858                                         inactive_limit, clszone_limit;
 859                 struct ckrm_zone *czone;
 860                 u64 temp;
 861
 862                 czone = &cls->ckrm_zone[zindex];
 863                 if (ckrm_test_set_shrink(czone))
 864                         continue;
 865
 866                 zone->temp_priority = zone->prev_priority;
 867                 zone->prev_priority = sc.priority;
 868
 869                 zone_total = zone->nr_active + zone->nr_inactive
 870                                                 + zone->free_pages;
 871
 872                 temp = (u64) cls->pg_limit * zone_total;
 873                 do_div(temp, ckrm_tot_lru_pages);
 874                 zone_limit = (int) temp;
 875                 clszone_limit = (ckrm_mem_shrink_to * zone_limit) / 100;
 876                 active_limit = (2 * clszone_limit) / 3; // 2/3rd in active list
 877                 inactive_limit = clszone_limit / 3; // 1/3rd in inactive list
 878
 879                 czone->shrink_active = 0;
 880                 cnt = czone->nr_active + act_credit - active_limit;
 881                 if (cnt > 0) {
 882                         czone->shrink_active = (unsigned long) cnt;
 883                         act_credit = 0;
 884                 } else {
 885                         act_credit += cnt;
 886                 }
 887
 888                 czone->shrink_inactive = 0;
 889                 cnt = czone->shrink_active + inact_credit +
 890                                         (czone->nr_inactive - inactive_limit);
 891                 if (cnt > 0) {
 892                         czone->shrink_inactive = (unsigned long) cnt;
 893                         inact_credit = 0;
 894                 } else {
 895                         inact_credit += cnt;
 896                 }
 897
 898
 899                 if (czone->shrink_active || czone->shrink_inactive) {
 900                         sc.nr_to_reclaim = czone->shrink_inactive;
 901                         shrink_ckrmzone(czone, &sc);
 902                 }
 903                 zone->prev_priority = zone->temp_priority;
 904                 zindex++;
 905                 ckrm_clear_shrink(czone);
 906         }
 907 }
 908
 909 static void
 910 ckrm_shrink_classes(void)
 911 {
 912         struct ckrm_mem_res *cls;
 913
 914         spin_lock(&ckrm_mem_lock);
 915         while (!ckrm_shrink_list_empty()) {
 916                 cls =  list_entry(ckrm_shrink_list.next, struct ckrm_mem_res,
 917                                 shrink_list);
 918                 list_del(&cls->shrink_list);
 919                 cls->flags &= ~CLS_AT_LIMIT;
 920                 spin_unlock(&ckrm_mem_lock);
 921                 ckrm_shrink_class(cls);
 922                 spin_lock(&ckrm_mem_lock);
 923         }
 924         spin_unlock(&ckrm_mem_lock);
 925 }
 926
 927 #else
 928 #define ckrm_shrink_classes()   do { } while(0)
 929 #endif
 930
 931 /*
 932  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 933  */
 934 static void
 935 shrink_zone(struct zone *zone, struct scan_control *sc)
 936 {
 937         unsigned long nr_active;
 938         unsigned long nr_inactive;
 939 #ifdef CONFIG_CKRM_RES_MEM
 940         struct ckrm_zone *czone;
 941 #endif
 942
 943
 944         /*
 945          * Add one to `nr_to_scan' just to make sure that the kernel will
 946          * slowly sift through the active list.
 947          */
 948         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 949         nr_active = zone->nr_scan_active;
 950         if (nr_active >= SWAP_CLUSTER_MAX)
 951                 zone->nr_scan_active = 0;
 952         else
 953                 nr_active = 0;
 954
 955         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 956         nr_inactive = zone->nr_scan_inactive;
 957         if (nr_inactive >= SWAP_CLUSTER_MAX)
 958                 zone->nr_scan_inactive = 0;
 959         else
 960                 nr_inactive = 0;
 961
 962         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 963
 964 #ifdef CONFIG_CKRM_RES_MEM
 965         if (nr_active || nr_inactive) {
 966                 struct list_head *pos, *next;
 967                 LIST_HEAD(victims);
 968
 969                 shrink_get_victims(zone, nr_active, nr_inactive, &victims);
 970                 pos = victims.next;
 971                 while (pos != &victims) {
 972                         czone = list_entry(pos, struct ckrm_zone, victim_list);
 973                         next = pos->next;
 974                         list_del_init(pos);
 975                         sc->nr_to_reclaim = czone->shrink_inactive;
 976                         shrink_ckrmzone(czone, sc);
 977                         ckrm_clear_shrink(czone);
 978                         pos = next;
 979                 }
 980         }
 981 #else
 982         while (nr_active || nr_inactive) {
 983                 if (nr_active) {
 984                         sc->nr_to_scan = min(nr_active,
 985                                         (unsigned long)SWAP_CLUSTER_MAX);
 986                         nr_active -= sc->nr_to_scan;
 987                         refill_inactive_zone(zone, sc);
 988                 }
 989
 990                 if (nr_inactive) {
 991                         sc->nr_to_scan = min(nr_inactive,
 992                                         (unsigned long)SWAP_CLUSTER_MAX);
 993                         nr_inactive -= sc->nr_to_scan;
 994                         shrink_cache(zone, sc);
 995                         if (sc->nr_to_reclaim <= 0)
 996                                 break;
 997                 }
 998         }
 999 #endif
1000 }
1001
1002 /*
1003  * This is the direct reclaim path, for page-allocating processes.  We only
1004  * try to reclaim pages from zones which will satisfy the caller's allocation
1005  * request.
1006  *
1007  * We reclaim from a zone even if that zone is over pages_high.  Because:
1008  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1009  *    allocation or
1010  * b) The zones may be over pages_high but they must go *over* pages_high to
1011  *    satisfy the `incremental min' zone defense algorithm.
1012  *
1013  * Returns the number of reclaimed pages.
1014  *
1015  * If a zone is deemed to be full of pinned pages then just give it a light
1016  * scan then give up on it.
1017  */
1018 static void
1019 shrink_caches(struct zone **zones, struct scan_control *sc)
1020 {
1021         int i;
1022
1023         for (i = 0; zones[i] != NULL; i++) {
1024                 struct zone *zone = zones[i];
1025
1026                 if (zone->present_pages == 0)
1027                         continue;
1028
1029                 zone->temp_priority = sc->priority;
1030                 if (zone->prev_priority > sc->priority)
1031                         zone->prev_priority = sc->priority;
1032
1033                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1034                         continue;       /* Let kswapd poll it */
1035
1036                 shrink_zone(zone, sc);
1037         }
1038 }
1039
1040 /*
1041  * This is the main entry point to direct page reclaim.
1042  *
1043  * If a full scan of the inactive list fails to free enough memory then we
1044  * are "out of memory" and something needs to be killed.
1045  *
1046  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1047  * high - the zone may be full of dirty or under-writeback pages, which this
1048  * caller can't do much about.  We kick pdflush and take explicit naps in the
1049  * hope that some of these pages can be written.  But if the allocating task
1050  * holds filesystem locks which prevent writeout this might not work, and the
1051  * allocation attempt will fail.
1052  */
1053 int try_to_free_pages(struct zone **zones,
1054                 unsigned int gfp_mask, unsigned int order)
1055 {
1056         int priority;
1057         int ret = 0;
1058         int total_scanned = 0, total_reclaimed = 0;
1059         struct reclaim_state *reclaim_state = current->reclaim_state;
1060         struct scan_control sc;
1061         unsigned long lru_pages = 0;
1062         int i;
1063
1064         sc.gfp_mask = gfp_mask;
1065         sc.may_writepage = 0;
1066
1067         inc_page_state(allocstall);
1068
1069         for (i = 0; zones[i] != NULL; i++) {
1070                 struct zone *zone = zones[i];
1071
1072                 zone->temp_priority = DEF_PRIORITY;
1073                 lru_pages += zone->nr_active + zone->nr_inactive;
1074         }
1075
1076         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1077                 sc.nr_mapped = read_page_state(nr_mapped);
1078                 sc.nr_scanned = 0;
1079                 sc.nr_reclaimed = 0;
1080                 sc.priority = priority;
1081                 shrink_caches(zones, &sc);
1082                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1083                 if (reclaim_state) {
1084                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1085                         reclaim_state->reclaimed_slab = 0;
1086                 }
1087                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1088                         ret = 1;
1089                         goto out;
1090                 }
1091                 total_scanned += sc.nr_scanned;
1092                 total_reclaimed += sc.nr_reclaimed;
1093
1094                 /*
1095                  * Try to write back as many pages as we just scanned.  This
1096                  * tends to cause slow streaming writers to write data to the
1097                  * disk smoothly, at the dirtying rate, which is nice.   But
1098                  * that's undesirable in laptop mode, where we *want* lumpy
1099                  * writeout.  So in laptop mode, write out the whole world.
1100                  */
1101                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1102                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1103                         sc.may_writepage = 1;
1104                 }
1105
1106                 /* Take a nap, wait for some writeback to complete */
1107                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1108                         blk_congestion_wait(WRITE, HZ/10);
1109         }
1110         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1111                 out_of_memory(gfp_mask);
1112 out:
1113         for (i = 0; zones[i] != 0; i++)
1114                 zones[i]->prev_priority = zones[i]->temp_priority;
1115         return ret;
1116 }
1117
1118 /*
1119  * For kswapd, balance_pgdat() will work across all this node's zones until
1120  * they are all at pages_high.
1121  *
1122  * If `nr_pages' is non-zero then it is the number of pages which are to be
1123  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1124  * special.
1125  *
1126  * Returns the number of pages which were actually freed.
1127  *
1128  * There is special handling here for zones which are full of pinned pages.
1129  * This can happen if the pages are all mlocked, or if they are all used by
1130  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1131  * What we do is to detect the case where all pages in the zone have been
1132  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1133  * dead and from now on, only perform a short scan.  Basically we're polling
1134  * the zone for when the problem goes away.
1135  *
1136  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1137  * zones which have free_pages > pages_high, but once a zone is found to have
1138  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1139  * of the number of free pages in the lower zones.  This interoperates with
1140  * the page allocator fallback scheme to ensure that aging of pages is balanced
1141  * across the zones.
1142  */
1143 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1144 {
1145         int to_free = nr_pages;
1146         int all_zones_ok;
1147         int priority;
1148         int i;
1149         int total_scanned, total_reclaimed;
1150         struct reclaim_state *reclaim_state = current->reclaim_state;
1151         struct scan_control sc;
1152
1153 loop_again:
1154         total_scanned = 0;
1155         total_reclaimed = 0;
1156         sc.gfp_mask = GFP_KERNEL;
1157         sc.may_writepage = 0;
1158         sc.nr_mapped = read_page_state(nr_mapped);
1159
1160         inc_page_state(pageoutrun);
1161
1162         for (i = 0; i < pgdat->nr_zones; i++) {
1163                 struct zone *zone = pgdat->node_zones + i;
1164
1165                 zone->temp_priority = DEF_PRIORITY;
1166         }
1167
1168         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1169                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1170                 unsigned long lru_pages = 0;
1171
1172                 all_zones_ok = 1;
1173
1174                 if (nr_pages == 0) {
1175                         /*
1176                          * Scan in the highmem->dma direction for the highest
1177                          * zone which needs scanning
1178                          */
1179                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1180                                 struct zone *zone = pgdat->node_zones + i;
1181
1182                                 if (zone->present_pages == 0)
1183                                         continue;
1184
1185                                 if (zone->all_unreclaimable &&
1186                                                 priority != DEF_PRIORITY)
1187                                         continue;
1188
1189                                 if (zone->free_pages <= zone->pages_high) {
1190                                         end_zone = i;
1191                                         goto scan;
1192                                 }
1193                         }
1194                         goto out;
1195                 } else {
1196                         end_zone = pgdat->nr_zones - 1;
1197                 }
1198 scan:
1199                 for (i = 0; i <= end_zone; i++) {
1200                         struct zone *zone = pgdat->node_zones + i;
1201
1202                         lru_pages += zone->nr_active + zone->nr_inactive;
1203                 }
1204
1205                 /*
1206                  * Now scan the zone in the dma->highmem direction, stopping
1207                  * at the last zone which needs scanning.
1208                  *
1209                  * We do this because the page allocator works in the opposite
1210                  * direction.  This prevents the page allocator from allocating
1211                  * pages behind kswapd's direction of progress, which would
1212                  * cause too much scanning of the lower zones.
1213                  */
1214                 for (i = 0; i <= end_zone; i++) {
1215                         struct zone *zone = pgdat->node_zones + i;
1216
1217                         if (zone->present_pages == 0)
1218                                 continue;
1219
1220                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1221                                 continue;
1222
1223                         if (nr_pages == 0) {    /* Not software suspend */
1224                                 if (zone->free_pages <= zone->pages_high)
1225                                         all_zones_ok = 0;
1226                         }
1227                         zone->temp_priority = priority;
1228                         if (zone->prev_priority > priority)
1229                                 zone->prev_priority = priority;
1230                         sc.nr_scanned = 0;
1231                         sc.nr_reclaimed = 0;
1232                         sc.priority = priority;
1233                         shrink_zone(zone, &sc);
1234                         reclaim_state->reclaimed_slab = 0;
1235                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1236                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1237                         total_reclaimed += sc.nr_reclaimed;
1238                         if (zone->all_unreclaimable)
1239                                 continue;
1240                         if (zone->pages_scanned >= (zone->nr_active +
1241                                                         zone->nr_inactive) * 4)
1242                                 zone->all_unreclaimable = 1;
1243                         /*
1244                          * If we've done a decent amount of scanning and
1245                          * the reclaim ratio is low, start doing writepage
1246                          * even in laptop mode
1247                          */
1248                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1249                             total_scanned > total_reclaimed+total_reclaimed/2)
1250                                 sc.may_writepage = 1;
1251                 }
1252                 if (nr_pages && to_free > total_reclaimed)
1253                         continue;       /* swsusp: need to do more work */
1254                 if (all_zones_ok)
1255                         break;          /* kswapd: all done */
1256                 /*
1257                  * OK, kswapd is getting into trouble.  Take a nap, then take
1258                  * another pass across the zones.
1259                  */
1260                 if (total_scanned && priority < DEF_PRIORITY - 2)
1261                         blk_congestion_wait(WRITE, HZ/10);
1262
1263                 /*
1264                  * We do this so kswapd doesn't build up large priorities for
1265                  * example when it is freeing in parallel with allocators. It
1266                  * matches the direct reclaim path behaviour in terms of impact
1267                  * on zone->*_priority.
1268                  */
1269                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1270                         break;
1271         }
1272 out:
1273         for (i = 0; i < pgdat->nr_zones; i++) {
1274                 struct zone *zone = pgdat->node_zones + i;
1275
1276                 zone->prev_priority = zone->temp_priority;
1277         }
1278         if (!all_zones_ok) {
1279                 cond_resched();
1280                 goto loop_again;
1281         }
1282
1283         return total_reclaimed;
1284 }
1285
1286 /*
1287  * The background pageout daemon, started as a kernel thread
1288  * from the init process.
1289  *
1290  * This basically trickles out pages so that we have _some_
1291  * free memory available even if there is no other activity
1292  * that frees anything up. This is needed for things like routing
1293  * etc, where we otherwise might have all activity going on in
1294  * asynchronous contexts that cannot page things out.
1295  *
1296  * If there are applications that are active memory-allocators
1297  * (most normal use), this basically shouldn't matter.
1298  */
1299 static int kswapd(void *p)
1300 {
1301         pg_data_t *pgdat = (pg_data_t*)p;
1302         struct task_struct *tsk = current;
1303         DEFINE_WAIT(wait);
1304         struct reclaim_state reclaim_state = {
1305                 .reclaimed_slab = 0,
1306         };
1307         cpumask_t cpumask;
1308
1309         daemonize("kswapd%d", pgdat->node_id);
1310         cpumask = node_to_cpumask(pgdat->node_id);
1311         if (!cpus_empty(cpumask))
1312                 set_cpus_allowed(tsk, cpumask);
1313         current->reclaim_state = &reclaim_state;
1314
1315         /*
1316          * Tell the memory management that we're a "memory allocator",
1317          * and that if we need more memory we should get access to it
1318          * regardless (see "__alloc_pages()"). "kswapd" should
1319          * never get caught in the normal page freeing logic.
1320          *
1321          * (Kswapd normally doesn't need memory anyway, but sometimes
1322          * you need a small amount of memory in order to be able to
1323          * page out something else, and this flag essentially protects
1324          * us from recursively trying to free more memory as we're
1325          * trying to free the first piece of memory in the first place).
1326          */
1327         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1328
1329         for ( ; ; ) {
1330                 if (current->flags & PF_FREEZE)
1331                         refrigerator(PF_FREEZE);
1332                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1333                 schedule();
1334                 finish_wait(&pgdat->kswapd_wait, &wait);
1335
1336                 if (!ckrm_shrink_list_empty())
1337                         ckrm_shrink_classes();
1338                 else
1339                         balance_pgdat(pgdat, 0);
1340         }
1341         return 0;
1342 }
1343
1344 /*
1345  * A zone is low on free memory, so wake its kswapd task to service it.
1346  */
1347 void wakeup_kswapd(struct zone *zone)
1348 {
1349         if (zone->present_pages == 0)
1350                 return;
1351         if (zone->free_pages > zone->pages_low)
1352                 return;
1353         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1354                 return;
1355         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1356 }
1357
1358 #ifdef CONFIG_PM
1359 /*
1360  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1361  * pages.
1362  */
1363 int shrink_all_memory(int nr_pages)
1364 {
1365         pg_data_t *pgdat;
1366         int nr_to_free = nr_pages;
1367         int ret = 0;
1368         struct reclaim_state reclaim_state = {
1369                 .reclaimed_slab = 0,
1370         };
1371
1372         current->reclaim_state = &reclaim_state;
1373         for_each_pgdat(pgdat) {
1374                 int freed;
1375                 freed = balance_pgdat(pgdat, nr_to_free);
1376                 ret += freed;
1377                 nr_to_free -= freed;
1378                 if (nr_to_free <= 0)
1379                         break;
1380         }
1381         current->reclaim_state = NULL;
1382         return ret;
1383 }
1384 #endif
1385
1386 #ifdef CONFIG_HOTPLUG_CPU
1387 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1388    not required for correctness.  So if the last cpu in a node goes
1389    away, we get changed to run anywhere: as the first one comes back,
1390    restore their cpu bindings. */
1391 static int __devinit cpu_callback(struct notifier_block *nfb,
1392                                   unsigned long action,
1393                                   void *hcpu)
1394 {
1395         pg_data_t *pgdat;
1396         cpumask_t mask;
1397
1398         if (action == CPU_ONLINE) {
1399                 for_each_pgdat(pgdat) {
1400                         mask = node_to_cpumask(pgdat->node_id);
1401                         if (any_online_cpu(mask) != NR_CPUS)
1402                                 /* One of our CPUs online: restore mask */
1403                                 set_cpus_allowed(pgdat->kswapd, mask);
1404                 }
1405         }
1406         return NOTIFY_OK;
1407 }
1408 #endif /* CONFIG_HOTPLUG_CPU */
1409
1410 static int __init kswapd_init(void)
1411 {
1412         pg_data_t *pgdat;
1413         swap_setup();
1414         for_each_pgdat(pgdat)
1415                 pgdat->kswapd
1416                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1417         total_memory = nr_free_pagecache_pages();
1418         hotcpu_notifier(cpu_callback, 0);
1419         return 0;
1420 }
1421
1422 module_init(kswapd_init)