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
  36 #include <asm/tlbflush.h>
  37 #include <asm/div64.h>
  38
  39 #include <linux/swapops.h>
  40 #include <linux/ckrm_mem.h>
  41
  42 /* possible outcome of pageout() */
  43 typedef enum {
  44         /* failed to write page out, page is locked */
  45         PAGE_KEEP,
  46         /* move page to the active list, page is locked */
  47         PAGE_ACTIVATE,
  48         /* page has been sent to the disk successfully, page is unlocked */
  49         PAGE_SUCCESS,
  50         /* page is clean and locked */
  51         PAGE_CLEAN,
  52 } pageout_t;
  53
  54 struct scan_control {
  55         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  56         unsigned long nr_to_scan;
  57
  58         /* Incremented by the number of inactive pages that were scanned */
  59         unsigned long nr_scanned;
  60
  61         /* Incremented by the number of pages reclaimed */
  62         unsigned long nr_reclaimed;
  63
  64         unsigned long nr_mapped;        /* From page_state */
  65
  66         /* How many pages shrink_cache() should reclaim */
  67         int nr_to_reclaim;
  68
  69         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  70         unsigned int priority;
  71
  72         /* This context's GFP mask */
  73         unsigned int gfp_mask;
  74
  75         /* Flag used by CKRM */
  76         unsigned int ckrm_flags;
  77
  78         int may_writepage;
  79 };
  80
  81 /*
  82  * The list of shrinker callbacks used by to apply pressure to
  83  * ageable caches.
  84  */
  85 struct shrinker {
  86         shrinker_t              shrinker;
  87         struct list_head        list;
  88         int                     seeks;  /* seeks to recreate an obj */
  89         long                    nr;     /* objs pending delete */
  90 };
  91
  92 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  93
  94 #ifdef ARCH_HAS_PREFETCH
  95 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  96         do {                                                            \
  97                 if ((_page)->lru.prev != _base) {                       \
  98                         struct page *prev;                              \
  99                                                                         \
 100                         prev = lru_to_page(&(_page->lru));              \
 101                         prefetch(&prev->_field);                        \
 102                 }                                                       \
 103         } while (0)
 104 #else
 105 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 106 #endif
 107
 108 #ifdef ARCH_HAS_PREFETCHW
 109 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 110         do {                                                            \
 111                 if ((_page)->lru.prev != _base) {                       \
 112                         struct page *prev;                              \
 113                                                                         \
 114                         prev = lru_to_page(&(_page->lru));              \
 115                         prefetchw(&prev->_field);                       \
 116                 }                                                       \
 117         } while (0)
 118 #else
 119 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 120 #endif
 121
 122 /*
 123  * From 0 .. 100.  Higher means more swappy.
 124  */
 125 int vm_swappiness = 60;
 126 static long total_memory;
 127
 128 static LIST_HEAD(shrinker_list);
 129 static DECLARE_MUTEX(shrinker_sem);
 130
 131 /*
 132  * Add a shrinker callback to be called from the vm
 133  */
 134 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 135 {
 136         struct shrinker *shrinker;
 137
 138         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 139         if (shrinker) {
 140                 shrinker->shrinker = theshrinker;
 141                 shrinker->seeks = seeks;
 142                 shrinker->nr = 0;
 143                 down(&shrinker_sem);
 144                 list_add(&shrinker->list, &shrinker_list);
 145                 up(&shrinker_sem);
 146         }
 147         return shrinker;
 148 }
 149 EXPORT_SYMBOL(set_shrinker);
 150
 151 /*
 152  * Remove one
 153  */
 154 void remove_shrinker(struct shrinker *shrinker)
 155 {
 156         down(&shrinker_sem);
 157         list_del(&shrinker->list);
 158         up(&shrinker_sem);
 159         kfree(shrinker);
 160 }
 161 EXPORT_SYMBOL(remove_shrinker);
 162
 163 #define SHRINK_BATCH 128
 164 /*
 165  * Call the shrink functions to age shrinkable caches
 166  *
 167  * Here we assume it costs one seek to replace a lru page and that it also
 168  * takes a seek to recreate a cache object.  With this in mind we age equal
 169  * percentages of the lru and ageable caches.  This should balance the seeks
 170  * generated by these structures.
 171  *
 172  * If the vm encounted mapped pages on the LRU it increase the pressure on
 173  * slab to avoid swapping.
 174  *
 175  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 176  *
 177  * `lru_pages' represents the number of on-LRU pages in all the zones which
 178  * are eligible for the caller's allocation attempt.  It is used for balancing
 179  * slab reclaim versus page reclaim.
 180  */
 181 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 182                         unsigned long lru_pages)
 183 {
 184         struct shrinker *shrinker;
 185
 186         if (down_trylock(&shrinker_sem))
 187                 return 0;
 188
 189         list_for_each_entry(shrinker, &shrinker_list, list) {
 190                 unsigned long long delta;
 191
 192                 delta = (4 * scanned) / shrinker->seeks;
 193                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 194                 do_div(delta, lru_pages + 1);
 195                 shrinker->nr += delta;
 196                 if (shrinker->nr < 0)
 197                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 198
 199                 if (shrinker->nr <= SHRINK_BATCH)
 200                         continue;
 201                 while (shrinker->nr) {
 202                         long this_scan = shrinker->nr;
 203                         int shrink_ret;
 204
 205                         if (this_scan > 128)
 206                                 this_scan = 128;
 207                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 208                         mod_page_state(slabs_scanned, this_scan);
 209                         shrinker->nr -= this_scan;
 210                         if (shrink_ret == -1)
 211                                 break;
 212                         cond_resched();
 213                 }
 214         }
 215         up(&shrinker_sem);
 216         return 0;
 217 }
 218
 219 /* Must be called with page's rmap lock held. */
 220 static inline int page_mapping_inuse(struct page *page)
 221 {
 222         struct address_space *mapping;
 223
 224         /* Page is in somebody's page tables. */
 225         if (page_mapped(page))
 226                 return 1;
 227
 228         /* Be more reluctant to reclaim swapcache than pagecache */
 229         if (PageSwapCache(page))
 230                 return 1;
 231
 232         mapping = page_mapping(page);
 233         if (!mapping)
 234                 return 0;
 235
 236         /* File is mmap'd by somebody? */
 237         return mapping_mapped(mapping);
 238 }
 239
 240 static inline int is_page_cache_freeable(struct page *page)
 241 {
 242         return page_count(page) - !!PagePrivate(page) == 2;
 243 }
 244
 245 static int may_write_to_queue(struct backing_dev_info *bdi)
 246 {
 247         if (current_is_kswapd())
 248                 return 1;
 249         if (current_is_pdflush())       /* This is unlikely, but why not... */
 250                 return 1;
 251         if (!bdi_write_congested(bdi))
 252                 return 1;
 253         if (bdi == current->backing_dev_info)
 254                 return 1;
 255         return 0;
 256 }
 257
 258 /*
 259  * We detected a synchronous write error writing a page out.  Probably
 260  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 261  * fsync(), msync() or close().
 262  *
 263  * The tricky part is that after writepage we cannot touch the mapping: nothing
 264  * prevents it from being freed up.  But we have a ref on the page and once
 265  * that page is locked, the mapping is pinned.
 266  *
 267  * We're allowed to run sleeping lock_page() here because we know the caller has
 268  * __GFP_FS.
 269  */
 270 static void handle_write_error(struct address_space *mapping,
 271                                 struct page *page, int error)
 272 {
 273         lock_page(page);
 274         if (page_mapping(page) == mapping) {
 275                 if (error == -ENOSPC)
 276                         set_bit(AS_ENOSPC, &mapping->flags);
 277                 else
 278                         set_bit(AS_EIO, &mapping->flags);
 279         }
 280         unlock_page(page);
 281 }
 282
 283 /*
 284  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 285  */
 286 static pageout_t pageout(struct page *page, struct address_space *mapping)
 287 {
 288         /*
 289          * If the page is dirty, only perform writeback if that write
 290          * will be non-blocking.  To prevent this allocation from being
 291          * stalled by pagecache activity.  But note that there may be
 292          * stalls if we need to run get_block().  We could test
 293          * PagePrivate for that.
 294          *
 295          * If this process is currently in generic_file_write() against
 296          * this page's queue, we can perform writeback even if that
 297          * will block.
 298          *
 299          * If the page is swapcache, write it back even if that would
 300          * block, for some throttling. This happens by accident, because
 301          * swap_backing_dev_info is bust: it doesn't reflect the
 302          * congestion state of the swapdevs.  Easy to fix, if needed.
 303          * See swapfile.c:page_queue_congested().
 304          */
 305         if (!is_page_cache_freeable(page))
 306                 return PAGE_KEEP;
 307         if (!mapping)
 308                 return PAGE_KEEP;
 309         if (mapping->a_ops->writepage == NULL)
 310                 return PAGE_ACTIVATE;
 311         if (!may_write_to_queue(mapping->backing_dev_info))
 312                 return PAGE_KEEP;
 313
 314         if (clear_page_dirty_for_io(page)) {
 315                 int res;
 316                 struct writeback_control wbc = {
 317                         .sync_mode = WB_SYNC_NONE,
 318                         .nr_to_write = SWAP_CLUSTER_MAX,
 319                         .nonblocking = 1,
 320                         .for_reclaim = 1,
 321                 };
 322
 323                 SetPageReclaim(page);
 324                 res = mapping->a_ops->writepage(page, &wbc);
 325                 if (res < 0)
 326                         handle_write_error(mapping, page, res);
 327                 if (res == WRITEPAGE_ACTIVATE) {
 328                         ClearPageReclaim(page);
 329                         return PAGE_ACTIVATE;
 330                 }
 331                 if (!PageWriteback(page)) {
 332                         /* synchronous write or broken a_ops? */
 333                         ClearPageReclaim(page);
 334                 }
 335
 336                 return PAGE_SUCCESS;
 337         }
 338
 339         return PAGE_CLEAN;
 340 }
 341
 342 /*
 343  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 344  */
 345 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 346 {
 347         LIST_HEAD(ret_pages);
 348         struct pagevec freed_pvec;
 349         int pgactivate = 0;
 350         int reclaimed = 0;
 351
 352         cond_resched();
 353
 354         pagevec_init(&freed_pvec, 1);
 355         while (!list_empty(page_list)) {
 356                 struct address_space *mapping;
 357                 struct page *page;
 358                 int may_enter_fs;
 359                 int referenced;
 360
 361                 page = lru_to_page(page_list);
 362                 list_del(&page->lru);
 363
 364                 if (TestSetPageLocked(page))
 365                         goto keep;
 366
 367                 BUG_ON(PageActive(page));
 368
 369                 if (PageWriteback(page))
 370                         goto keep_locked;
 371
 372                 sc->nr_scanned++;
 373                 /* Double the slab pressure for mapped and swapcache pages */
 374                 if (page_mapped(page) || PageSwapCache(page))
 375                         sc->nr_scanned++;
 376
 377                 page_map_lock(page);
 378                 referenced = page_referenced(page);
 379                 if (referenced && page_mapping_inuse(page)) {
 380                         /* In active use or really unfreeable.  Activate it. */
 381                         page_map_unlock(page);
 382                         goto activate_locked;
 383                 }
 384
 385 #ifdef CONFIG_SWAP
 386                 /*
 387                  * Anonymous process memory has backing store?
 388                  * Try to allocate it some swap space here.
 389                  *
 390                  * XXX: implement swap clustering ?
 391                  */
 392                 if (PageAnon(page) && !PageSwapCache(page)) {
 393                         page_map_unlock(page);
 394                         if (!add_to_swap(page))
 395                                 goto activate_locked;
 396                         page_map_lock(page);
 397                 }
 398 #endif /* CONFIG_SWAP */
 399
 400                 mapping = page_mapping(page);
 401                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 402                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 403
 404                 /*
 405                  * The page is mapped into the page tables of one or more
 406                  * processes. Try to unmap it here.
 407                  */
 408                 if (page_mapped(page) && mapping) {
 409                         switch (try_to_unmap(page)) {
 410                         case SWAP_FAIL:
 411                                 page_map_unlock(page);
 412                                 goto activate_locked;
 413                         case SWAP_AGAIN:
 414                                 page_map_unlock(page);
 415                                 goto keep_locked;
 416                         case SWAP_SUCCESS:
 417                                 ; /* try to free the page below */
 418                         }
 419                 }
 420                 page_map_unlock(page);
 421
 422                 if (PageDirty(page)) {
 423                         if (referenced)
 424                                 goto keep_locked;
 425                         if (!may_enter_fs)
 426                                 goto keep_locked;
 427                         if (laptop_mode && !sc->may_writepage)
 428                                 goto keep_locked;
 429
 430                         /* Page is dirty, try to write it out here */
 431                         switch(pageout(page, mapping)) {
 432                         case PAGE_KEEP:
 433                                 goto keep_locked;
 434                         case PAGE_ACTIVATE:
 435                                 goto activate_locked;
 436                         case PAGE_SUCCESS:
 437                                 if (PageWriteback(page) || PageDirty(page))
 438                                         goto keep;
 439                                 /*
 440                                  * A synchronous write - probably a ramdisk.  Go
 441                                  * ahead and try to reclaim the page.
 442                                  */
 443                                 if (TestSetPageLocked(page))
 444                                         goto keep;
 445                                 if (PageDirty(page) || PageWriteback(page))
 446                                         goto keep_locked;
 447                                 mapping = page_mapping(page);
 448                         case PAGE_CLEAN:
 449                                 ; /* try to free the page below */
 450                         }
 451                 }
 452
 453                 /*
 454                  * If the page has buffers, try to free the buffer mappings
 455                  * associated with this page. If we succeed we try to free
 456                  * the page as well.
 457                  *
 458                  * We do this even if the page is PageDirty().
 459                  * try_to_release_page() does not perform I/O, but it is
 460                  * possible for a page to have PageDirty set, but it is actually
 461                  * clean (all its buffers are clean).  This happens if the
 462                  * buffers were written out directly, with submit_bh(). ext3
 463                  * will do this, as well as the blockdev mapping.
 464                  * try_to_release_page() will discover that cleanness and will
 465                  * drop the buffers and mark the page clean - it can be freed.
 466                  *
 467                  * Rarely, pages can have buffers and no ->mapping.  These are
 468                  * the pages which were not successfully invalidated in
 469                  * truncate_complete_page().  We try to drop those buffers here
 470                  * and if that worked, and the page is no longer mapped into
 471                  * process address space (page_count == 1) it can be freed.
 472                  * Otherwise, leave the page on the LRU so it is swappable.
 473                  */
 474                 if (PagePrivate(page)) {
 475                         if (!try_to_release_page(page, sc->gfp_mask))
 476                                 goto activate_locked;
 477                         if (!mapping && page_count(page) == 1)
 478                                 goto free_it;
 479                 }
 480
 481                 if (!mapping)
 482                         goto keep_locked;       /* truncate got there first */
 483
 484                 spin_lock_irq(&mapping->tree_lock);
 485
 486                 /*
 487                  * The non-racy check for busy page.  It is critical to check
 488                  * PageDirty _after_ making sure that the page is freeable and
 489                  * not in use by anybody.       (pagecache + us == 2)
 490                  */
 491                 if (page_count(page) != 2 || PageDirty(page)) {
 492                         spin_unlock_irq(&mapping->tree_lock);
 493                         goto keep_locked;
 494                 }
 495
 496 #ifdef CONFIG_SWAP
 497                 if (PageSwapCache(page)) {
 498                         swp_entry_t swap = { .val = page->private };
 499                         __delete_from_swap_cache(page);
 500                         spin_unlock_irq(&mapping->tree_lock);
 501                         swap_free(swap);
 502                         __put_page(page);       /* The pagecache ref */
 503                         goto free_it;
 504                 }
 505 #endif /* CONFIG_SWAP */
 506
 507                 __remove_from_page_cache(page);
 508                 spin_unlock_irq(&mapping->tree_lock);
 509                 __put_page(page);
 510
 511 free_it:
 512                 unlock_page(page);
 513                 reclaimed++;
 514                 if (!pagevec_add(&freed_pvec, page))
 515                         __pagevec_release_nonlru(&freed_pvec);
 516                 continue;
 517
 518 activate_locked:
 519                 SetPageActive(page);
 520                 pgactivate++;
 521 keep_locked:
 522                 unlock_page(page);
 523 keep:
 524                 list_add(&page->lru, &ret_pages);
 525                 BUG_ON(PageLRU(page));
 526         }
 527         list_splice(&ret_pages, page_list);
 528         if (pagevec_count(&freed_pvec))
 529                 __pagevec_release_nonlru(&freed_pvec);
 530         mod_page_state(pgactivate, pgactivate);
 531         sc->nr_reclaimed += reclaimed;
 532         return reclaimed;
 533 }
 534
 535 /*
 536  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 537  * a batch of pages and working on them outside the lock.  Any pages which were
 538  * not freed will be added back to the LRU.
 539  *
 540  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 541  *
 542  * For pagecache intensive workloads, the first loop here is the hottest spot
 543  * in the kernel (apart from the copy_*_user functions).
 544  */
 545 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 546 {
 547         LIST_HEAD(page_list);
 548         struct pagevec pvec;
 549         int max_scan = sc->nr_to_scan, nr_pass;
 550         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 551
 552         pagevec_init(&pvec, 1);
 553
 554         lru_add_drain();
 555         spin_lock_irq(&zone->lru_lock);
 556 redo:
 557         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 558         nr_pass = zone->nr_inactive;
 559         while (max_scan > 0) {
 560                 struct page *page;
 561                 int nr_taken = 0;
 562                 int nr_scan = 0;
 563                 int nr_freed;
 564
 565                 while (nr_pass-- && nr_scan++ < SWAP_CLUSTER_MAX &&
 566                                 !list_empty(&zone->inactive_list)) {
 567                         page = lru_to_page(&zone->inactive_list);
 568
 569                         prefetchw_prev_lru_page(page,
 570                                                 &zone->inactive_list, flags);
 571
 572                         if (!TestClearPageLRU(page))
 573                                 BUG();
 574                         list_del(&page->lru);
 575                         if (get_page_testone(page)) {
 576                                 /*
 577                                  * It is being freed elsewhere
 578                                  */
 579                                 __put_page(page);
 580                                 SetPageLRU(page);
 581                                 list_add(&page->lru, &zone->inactive_list);
 582                                 continue;
 583                         } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 584                                 __put_page(page);
 585                                 SetPageLRU(page);
 586 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 587                                 list_add_tail(&page->lru, &zone->inactive_list);
 588 #else
 589                                 list_add(&page->lru, &zone->inactive_list);
 590 #endif
 591                                 continue;
 592                         }
 593                         list_add(&page->lru, &page_list);
 594                         ckrm_mem_dec_inactive(page);
 595                         nr_taken++;
 596                 }
 597                 zone->nr_inactive -= nr_taken;
 598                 zone->pages_scanned += nr_taken;
 599                 spin_unlock_irq(&zone->lru_lock);
 600
 601                 if ((bit_flag == 0) && (nr_taken == 0))
 602                         goto done;
 603
 604                 max_scan -= nr_scan;
 605                 if (current_is_kswapd())
 606                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 607                 else
 608                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 609                 nr_freed = shrink_list(&page_list, sc);
 610                 if (current_is_kswapd())
 611                         mod_page_state(kswapd_steal, nr_freed);
 612                 mod_page_state_zone(zone, pgsteal, nr_freed);
 613                 sc->nr_to_reclaim -= nr_freed;
 614
 615                 spin_lock_irq(&zone->lru_lock);
 616                 /*
 617                  * Put back any unfreeable pages.
 618                  */
 619                 while (!list_empty(&page_list)) {
 620                         page = lru_to_page(&page_list);
 621                         if (TestSetPageLRU(page))
 622                                 BUG();
 623                         list_del(&page->lru);
 624                         if (PageActive(page))
 625                                 add_page_to_active_list(zone, page);
 626                         else
 627                                 add_page_to_inactive_list(zone, page);
 628                         if (!pagevec_add(&pvec, page)) {
 629                                 spin_unlock_irq(&zone->lru_lock);
 630                                 __pagevec_release(&pvec);
 631                                 spin_lock_irq(&zone->lru_lock);
 632                         }
 633                 }
 634                 if (ckrm_flags && (nr_pass <= 0)) {
 635                         goto redo;
 636                 }
 637         }
 638         spin_unlock_irq(&zone->lru_lock);
 639 done:
 640         pagevec_release(&pvec);
 641 }
 642
 643 /*
 644  * This moves pages from the active list to the inactive list.
 645  *
 646  * We move them the other way if the page is referenced by one or more
 647  * processes, from rmap.
 648  *
 649  * If the pages are mostly unmapped, the processing is fast and it is
 650  * appropriate to hold zone->lru_lock across the whole operation.  But if
 651  * the pages are mapped, the processing is slow (page_referenced()) so we
 652  * should drop zone->lru_lock around each page.  It's impossible to balance
 653  * this, so instead we remove the pages from the LRU while processing them.
 654  * It is safe to rely on PG_active against the non-LRU pages in here because
 655  * nobody will play with that bit on a non-LRU page.
 656  *
 657  * The downside is that we have to touch page->_count against each page.
 658  * But we had to alter page->flags anyway.
 659  */
 660 static void
 661 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 662 {
 663         int pgmoved;
 664         int pgdeactivate = 0;
 665         int pgscanned = 0;
 666         int nr_pages = sc->nr_to_scan;
 667         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 668         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 669         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 670         struct page *page;
 671         struct pagevec pvec;
 672         int reclaim_mapped = 0;
 673         long mapped_ratio;
 674         long distress;
 675         long swap_tendency;
 676         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 677         int nr_pass;
 678
 679         lru_add_drain();
 680         pgmoved = 0;
 681         spin_lock_irq(&zone->lru_lock);
 682 redo:
 683         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 684         nr_pass = zone->nr_active;
 685         while (pgscanned < nr_pages && !list_empty(&zone->active_list) &&
 686                                                 nr_pass) {
 687                 page = lru_to_page(&zone->active_list);
 688                 prefetchw_prev_lru_page(page, &zone->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, &zone->active_list);
 702                         pgscanned++;
 703                 } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 704                         __put_page(page);
 705                         SetPageLRU(page);
 706 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 707                         list_add_tail(&page->lru, &zone->active_list);
 708 #else
 709                         list_add(&page->lru, &zone->active_list);
 710 #endif
 711                 } else {
 712                         list_add(&page->lru, &l_hold);
 713                         ckrm_mem_dec_active(page);
 714                         pgmoved++;
 715                 pgscanned++;
 716         }
 717                 if (!--nr_pass && ckrm_flags) {
 718                         goto redo;
 719                 }
 720         }
 721         zone->nr_active -= pgmoved;
 722         spin_unlock_irq(&zone->lru_lock);
 723
 724         /*
 725          * `distress' is a measure of how much trouble we're having reclaiming
 726          * pages.  0 -> no problems.  100 -> great trouble.
 727          */
 728         distress = 100 >> zone->prev_priority;
 729
 730         /*
 731          * The point of this algorithm is to decide when to start reclaiming
 732          * mapped memory instead of just pagecache.  Work out how much memory
 733          * is mapped.
 734          */
 735         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 736
 737         /*
 738          * Now decide how much we really want to unmap some pages.  The mapped
 739          * ratio is downgraded - just because there's a lot of mapped memory
 740          * doesn't necessarily mean that page reclaim isn't succeeding.
 741          *
 742          * The distress ratio is important - we don't want to start going oom.
 743          *
 744          * A 100% value of vm_swappiness overrides this algorithm altogether.
 745          */
 746         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 747
 748         /*
 749          * Now use this metric to decide whether to start moving mapped memory
 750          * onto the inactive list.
 751          */
 752         if (swap_tendency >= 100)
 753                 reclaim_mapped = 1;
 754
 755         while (!list_empty(&l_hold)) {
 756                 page = lru_to_page(&l_hold);
 757                 list_del(&page->lru);
 758                 if (page_mapped(page)) {
 759                         if (!reclaim_mapped) {
 760                                 list_add(&page->lru, &l_active);
 761                                 continue;
 762                         }
 763                         page_map_lock(page);
 764                         if (page_referenced(page)) {
 765                                 page_map_unlock(page);
 766                                 list_add(&page->lru, &l_active);
 767                                 continue;
 768                         }
 769                         page_map_unlock(page);
 770                 }
 771                 /*
 772                  * FIXME: need to consider page_count(page) here if/when we
 773                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 774                  */
 775                 if (total_swap_pages == 0 && PageAnon(page)) {
 776                         list_add(&page->lru, &l_active);
 777                         continue;
 778                 }
 779                 list_add(&page->lru, &l_inactive);
 780         }
 781
 782         pagevec_init(&pvec, 1);
 783         pgmoved = 0;
 784         spin_lock_irq(&zone->lru_lock);
 785         while (!list_empty(&l_inactive)) {
 786                 page = lru_to_page(&l_inactive);
 787                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 788                 if (TestSetPageLRU(page))
 789                         BUG();
 790                 if (!TestClearPageActive(page))
 791                         BUG();
 792                 list_move(&page->lru, &zone->inactive_list);
 793                 ckrm_mem_inc_inactive(page);
 794                 pgmoved++;
 795                 if (!pagevec_add(&pvec, page)) {
 796                         zone->nr_inactive += pgmoved;
 797                         spin_unlock_irq(&zone->lru_lock);
 798                         pgdeactivate += pgmoved;
 799                         pgmoved = 0;
 800                         if (buffer_heads_over_limit)
 801                                 pagevec_strip(&pvec);
 802                         __pagevec_release(&pvec);
 803                         spin_lock_irq(&zone->lru_lock);
 804                 }
 805         }
 806         zone->nr_inactive += pgmoved;
 807         pgdeactivate += pgmoved;
 808         if (buffer_heads_over_limit) {
 809                 spin_unlock_irq(&zone->lru_lock);
 810                 pagevec_strip(&pvec);
 811                 spin_lock_irq(&zone->lru_lock);
 812         }
 813
 814         pgmoved = 0;
 815         while (!list_empty(&l_active)) {
 816                 page = lru_to_page(&l_active);
 817                 prefetchw_prev_lru_page(page, &l_active, flags);
 818                 if (TestSetPageLRU(page))
 819                         BUG();
 820                 BUG_ON(!PageActive(page));
 821                 list_move(&page->lru, &zone->active_list);
 822                 ckrm_mem_inc_active(page);
 823                 pgmoved++;
 824                 if (!pagevec_add(&pvec, page)) {
 825                         zone->nr_active += pgmoved;
 826                         pgmoved = 0;
 827                         spin_unlock_irq(&zone->lru_lock);
 828                         __pagevec_release(&pvec);
 829                         spin_lock_irq(&zone->lru_lock);
 830                 }
 831         }
 832         zone->nr_active += pgmoved;
 833         spin_unlock_irq(&zone->lru_lock);
 834         pagevec_release(&pvec);
 835
 836         mod_page_state_zone(zone, pgrefill, pgscanned);
 837         mod_page_state(pgdeactivate, pgdeactivate);
 838 }
 839
 840 /*
 841  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 842  */
 843 static void
 844 shrink_zone(struct zone *zone, struct scan_control *sc)
 845 {
 846         unsigned long nr_active;
 847         unsigned long nr_inactive;
 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 >> sc->priority) + 1;
 854         nr_active = zone->nr_scan_active;
 855         if (nr_active >= SWAP_CLUSTER_MAX)
 856                 zone->nr_scan_active = 0;
 857         else
 858                 nr_active = 0;
 859
 860         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 861         nr_inactive = zone->nr_scan_inactive;
 862         if (nr_inactive >= SWAP_CLUSTER_MAX)
 863                 zone->nr_scan_inactive = 0;
 864         else
 865                 nr_inactive = 0;
 866
 867         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 868
 869         while (nr_active || nr_inactive) {
 870                 sc->ckrm_flags = ckrm_setup_reclamation();
 871                 if (nr_active) {
 872                         sc->nr_to_scan = min(nr_active,
 873                                         (unsigned long)SWAP_CLUSTER_MAX);
 874                         nr_active -= sc->nr_to_scan;
 875                         refill_inactive_zone(zone, sc);
 876                 }
 877
 878                 if (nr_inactive) {
 879                         sc->nr_to_scan = min(nr_inactive,
 880                                         (unsigned long)SWAP_CLUSTER_MAX);
 881                         nr_inactive -= sc->nr_to_scan;
 882                         shrink_cache(zone, sc);
 883                         if (sc->nr_to_reclaim <= 0)
 884                                 break;
 885                 }
 886                 ckrm_teardown_reclamation();
 887         }
 888 }
 889
 890 #ifdef CONFIG_CKRM_RES_MEM
 891 // This function needs to be given more thought.
 892 // Shrink the class to be at 90% of its limit
 893 static void
 894 ckrm_shrink_class(ckrm_mem_res_t *cls)
 895 {
 896         struct scan_control sc;
 897         struct zone *zone;
 898         int zindex = 0, active_credit = 0, inactive_credit = 0;
 899
 900         if (ckrm_test_set_shrink(cls)) { // set the SHRINK bit atomically
 901                 // if it is already set somebody is working on it. so... leave
 902                 return;
 903         }
 904         sc.nr_mapped = read_page_state(nr_mapped);
 905         sc.nr_scanned = 0;
 906         sc.ckrm_flags = ckrm_get_reclaim_flags(cls);
 907         sc.nr_reclaimed = 0;
 908         sc.priority = 0; // always very high priority
 909
 910         for_each_zone(zone) {
 911                 int zone_total, zone_limit, active_limit, inactive_limit;
 912                 int active_over, inactive_over;
 913                 unsigned long nr_active, nr_inactive;
 914                 u64 temp;
 915
 916                 zone->temp_priority = zone->prev_priority;
 917                 zone->prev_priority = sc.priority;
 918
 919                 zone_total = zone->nr_active + zone->nr_inactive + zone->free_pages;
 920
 921                 temp = (u64) cls->pg_limit * zone_total;
 922                 do_div(temp, ckrm_tot_lru_pages);
 923                 zone_limit = (int) temp;
 924                 active_limit = (6 * zone_limit) / 10; // 2/3rd in active list
 925                 inactive_limit = (3 * zone_limit) / 10; // 1/3rd in inactive list
 926
 927                 active_over = cls->nr_active[zindex] - active_limit + active_credit;
 928                 inactive_over = active_over +
 929                                 (cls->nr_inactive[zindex] - inactive_limit) + inactive_credit;
 930
 931                 if (active_over > 0) {
 932                         zone->nr_scan_active += active_over + 1;
 933                         nr_active = zone->nr_scan_active;
 934                         active_credit = 0;
 935                 } else {
 936                         active_credit += active_over;
 937                         nr_active = 0;
 938                 }
 939
 940                 if (inactive_over > 0) {
 941                         zone->nr_scan_inactive += inactive_over;
 942                         nr_inactive = zone->nr_scan_inactive;
 943                         inactive_credit = 0;
 944                 } else {
 945                         inactive_credit += inactive_over;
 946                         nr_inactive = 0;
 947                 }
 948                 while (nr_active || nr_inactive) {
 949                         if (nr_active) {
 950                                 sc.nr_to_scan = min(nr_active,
 951                                                 (unsigned long)SWAP_CLUSTER_MAX);
 952                                 nr_active -= sc.nr_to_scan;
 953                                 refill_inactive_zone(zone, &sc);
 954                         }
 955
 956                         if (nr_inactive) {
 957                                 sc.nr_to_scan = min(nr_inactive,
 958                                                 (unsigned long)SWAP_CLUSTER_MAX);
 959                                 nr_inactive -= sc.nr_to_scan;
 960                                 shrink_cache(zone, &sc);
 961                                 if (sc.nr_to_reclaim <= 0)
 962                                         break;
 963                         }
 964                 }
 965                 zone->prev_priority = zone->temp_priority;
 966                 zindex++;
 967         }
 968         ckrm_clear_shrink(cls);
 969 }
 970
 971 static void
 972 ckrm_shrink_classes(void)
 973 {
 974         ckrm_mem_res_t *cls;
 975
 976         spin_lock(&ckrm_mem_lock);
 977         while (!ckrm_shrink_list_empty()) {
 978                 cls =  list_entry(ckrm_shrink_list.next, ckrm_mem_res_t,
 979                                 shrink_list);
 980                 spin_unlock(&ckrm_mem_lock);
 981                 ckrm_shrink_class(cls);
 982                 spin_lock(&ckrm_mem_lock);
 983                 list_del(&cls->shrink_list);
 984                 cls->flags &= ~MEM_AT_LIMIT;
 985         }
 986         spin_unlock(&ckrm_mem_lock);
 987 }
 988
 989 #else
 990 #define ckrm_shrink_classes()   do { } while(0)
 991 #endif
 992
 993 /*
 994  * This is the direct reclaim path, for page-allocating processes.  We only
 995  * try to reclaim pages from zones which will satisfy the caller's allocation
 996  * request.
 997  *
 998  * We reclaim from a zone even if that zone is over pages_high.  Because:
 999  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1000  *    allocation or
1001  * b) The zones may be over pages_high but they must go *over* pages_high to
1002  *    satisfy the `incremental min' zone defense algorithm.
1003  *
1004  * Returns the number of reclaimed pages.
1005  *
1006  * If a zone is deemed to be full of pinned pages then just give it a light
1007  * scan then give up on it.
1008  */
1009 static void
1010 shrink_caches(struct zone **zones, struct scan_control *sc)
1011 {
1012         int i;
1013
1014         for (i = 0; zones[i] != NULL; i++) {
1015                 struct zone *zone = zones[i];
1016
1017                 zone->temp_priority = sc->priority;
1018                 if (zone->prev_priority > sc->priority)
1019                         zone->prev_priority = sc->priority;
1020
1021                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1022                         continue;       /* Let kswapd poll it */
1023
1024                 shrink_zone(zone, sc);
1025         }
1026 }
1027
1028 /*
1029  * This is the main entry point to direct page reclaim.
1030  *
1031  * If a full scan of the inactive list fails to free enough memory then we
1032  * are "out of memory" and something needs to be killed.
1033  *
1034  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1035  * high - the zone may be full of dirty or under-writeback pages, which this
1036  * caller can't do much about.  We kick pdflush and take explicit naps in the
1037  * hope that some of these pages can be written.  But if the allocating task
1038  * holds filesystem locks which prevent writeout this might not work, and the
1039  * allocation attempt will fail.
1040  */
1041 int try_to_free_pages(struct zone **zones,
1042                 unsigned int gfp_mask, unsigned int order)
1043 {
1044         int priority;
1045         int ret = 0;
1046         int total_scanned = 0, total_reclaimed = 0;
1047         struct reclaim_state *reclaim_state = current->reclaim_state;
1048         struct scan_control sc;
1049         unsigned long lru_pages = 0;
1050         int i;
1051
1052         sc.gfp_mask = gfp_mask;
1053         sc.may_writepage = 0;
1054
1055         inc_page_state(allocstall);
1056
1057         for (i = 0; zones[i] != NULL; i++) {
1058                 struct zone *zone = zones[i];
1059
1060                 zone->temp_priority = DEF_PRIORITY;
1061                 lru_pages += zone->nr_active + zone->nr_inactive;
1062         }
1063
1064         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1065                 sc.nr_mapped = read_page_state(nr_mapped);
1066                 sc.nr_scanned = 0;
1067                 sc.nr_reclaimed = 0;
1068                 sc.priority = priority;
1069                 shrink_caches(zones, &sc);
1070                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1071                 if (reclaim_state) {
1072                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1073                         reclaim_state->reclaimed_slab = 0;
1074                 }
1075                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1076                         ret = 1;
1077                         goto out;
1078                 }
1079                 total_scanned += sc.nr_scanned;
1080                 total_reclaimed += sc.nr_reclaimed;
1081
1082                 /*
1083                  * Try to write back as many pages as we just scanned.  This
1084                  * tends to cause slow streaming writers to write data to the
1085                  * disk smoothly, at the dirtying rate, which is nice.   But
1086                  * that's undesirable in laptop mode, where we *want* lumpy
1087                  * writeout.  So in laptop mode, write out the whole world.
1088                  */
1089                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1090                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1091                         sc.may_writepage = 1;
1092                 }
1093
1094                 /* Take a nap, wait for some writeback to complete */
1095                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1096                         blk_congestion_wait(WRITE, HZ/10);
1097         }
1098         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1099                 out_of_memory(gfp_mask);
1100 out:
1101         for (i = 0; zones[i] != 0; i++)
1102                 zones[i]->prev_priority = zones[i]->temp_priority;
1103         return ret;
1104 }
1105
1106 /*
1107  * For kswapd, balance_pgdat() will work across all this node's zones until
1108  * they are all at pages_high.
1109  *
1110  * If `nr_pages' is non-zero then it is the number of pages which are to be
1111  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1112  * special.
1113  *
1114  * Returns the number of pages which were actually freed.
1115  *
1116  * There is special handling here for zones which are full of pinned pages.
1117  * This can happen if the pages are all mlocked, or if they are all used by
1118  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1119  * What we do is to detect the case where all pages in the zone have been
1120  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1121  * dead and from now on, only perform a short scan.  Basically we're polling
1122  * the zone for when the problem goes away.
1123  *
1124  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1125  * zones which have free_pages > pages_high, but once a zone is found to have
1126  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1127  * of the number of free pages in the lower zones.  This interoperates with
1128  * the page allocator fallback scheme to ensure that aging of pages is balanced
1129  * across the zones.
1130  */
1131 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1132 {
1133         int to_free = nr_pages;
1134         int priority;
1135         int i;
1136         int total_scanned = 0, total_reclaimed = 0;
1137         struct reclaim_state *reclaim_state = current->reclaim_state;
1138         struct scan_control sc;
1139
1140         sc.gfp_mask = GFP_KERNEL;
1141         sc.may_writepage = 0;
1142         sc.nr_mapped = read_page_state(nr_mapped);
1143
1144         inc_page_state(pageoutrun);
1145
1146         for (i = 0; i < pgdat->nr_zones; i++) {
1147                 struct zone *zone = pgdat->node_zones + i;
1148
1149                 zone->temp_priority = DEF_PRIORITY;
1150         }
1151
1152         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1153                 int all_zones_ok = 1;
1154                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1155                 unsigned long lru_pages = 0;
1156
1157                 if (nr_pages == 0) {
1158                         /*
1159                          * Scan in the highmem->dma direction for the highest
1160                          * zone which needs scanning
1161                          */
1162                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1163                                 struct zone *zone = pgdat->node_zones + i;
1164
1165                                 if (zone->all_unreclaimable &&
1166                                                 priority != DEF_PRIORITY)
1167                                         continue;
1168
1169                                 if (zone->free_pages <= zone->pages_high) {
1170                                         end_zone = i;
1171                                         goto scan;
1172                                 }
1173                         }
1174                         goto out;
1175                 } else {
1176                         end_zone = pgdat->nr_zones - 1;
1177                 }
1178 scan:
1179                 for (i = 0; i <= end_zone; i++) {
1180                         struct zone *zone = pgdat->node_zones + i;
1181
1182                         lru_pages += zone->nr_active + zone->nr_inactive;
1183                 }
1184
1185                 /*
1186                  * Now scan the zone in the dma->highmem direction, stopping
1187                  * at the last zone which needs scanning.
1188                  *
1189                  * We do this because the page allocator works in the opposite
1190                  * direction.  This prevents the page allocator from allocating
1191                  * pages behind kswapd's direction of progress, which would
1192                  * cause too much scanning of the lower zones.
1193                  */
1194                 for (i = 0; i <= end_zone; i++) {
1195                         struct zone *zone = pgdat->node_zones + i;
1196
1197                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1198                                 continue;
1199
1200                         if (nr_pages == 0) {    /* Not software suspend */
1201                                 if (zone->free_pages <= zone->pages_high)
1202                                         all_zones_ok = 0;
1203                         }
1204                         zone->temp_priority = priority;
1205                         if (zone->prev_priority > priority)
1206                                 zone->prev_priority = priority;
1207                         sc.nr_scanned = 0;
1208                         sc.nr_reclaimed = 0;
1209                         sc.priority = priority;
1210                         shrink_zone(zone, &sc);
1211                         reclaim_state->reclaimed_slab = 0;
1212                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1213                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1214                         total_reclaimed += sc.nr_reclaimed;
1215                         if (zone->all_unreclaimable)
1216                                 continue;
1217                         if (zone->pages_scanned > zone->present_pages * 2)
1218                                 zone->all_unreclaimable = 1;
1219                         /*
1220                          * If we've done a decent amount of scanning and
1221                          * the reclaim ratio is low, start doing writepage
1222                          * even in laptop mode
1223                          */
1224                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1225                             total_scanned > total_reclaimed+total_reclaimed/2)
1226                                 sc.may_writepage = 1;
1227                 }
1228                 if (nr_pages && to_free > total_reclaimed)
1229                         continue;       /* swsusp: need to do more work */
1230                 if (all_zones_ok)
1231                         break;          /* kswapd: all done */
1232                 /*
1233                  * OK, kswapd is getting into trouble.  Take a nap, then take
1234                  * another pass across the zones.
1235                  */
1236                 if (total_scanned && priority < DEF_PRIORITY - 2)
1237                         blk_congestion_wait(WRITE, HZ/10);
1238         }
1239 out:
1240         for (i = 0; i < pgdat->nr_zones; i++) {
1241                 struct zone *zone = pgdat->node_zones + i;
1242
1243                 zone->prev_priority = zone->temp_priority;
1244         }
1245         return total_reclaimed;
1246 }
1247
1248 /*
1249  * The background pageout daemon, started as a kernel thread
1250  * from the init process.
1251  *
1252  * This basically trickles out pages so that we have _some_
1253  * free memory available even if there is no other activity
1254  * that frees anything up. This is needed for things like routing
1255  * etc, where we otherwise might have all activity going on in
1256  * asynchronous contexts that cannot page things out.
1257  *
1258  * If there are applications that are active memory-allocators
1259  * (most normal use), this basically shouldn't matter.
1260  */
1261 static int kswapd(void *p)
1262 {
1263         pg_data_t *pgdat = (pg_data_t*)p;
1264         struct task_struct *tsk = current;
1265         DEFINE_WAIT(wait);
1266         struct reclaim_state reclaim_state = {
1267                 .reclaimed_slab = 0,
1268         };
1269         cpumask_t cpumask;
1270
1271         daemonize("kswapd%d", pgdat->node_id);
1272         cpumask = node_to_cpumask(pgdat->node_id);
1273         if (!cpus_empty(cpumask))
1274                 set_cpus_allowed(tsk, cpumask);
1275         current->reclaim_state = &reclaim_state;
1276
1277         /*
1278          * Tell the memory management that we're a "memory allocator",
1279          * and that if we need more memory we should get access to it
1280          * regardless (see "__alloc_pages()"). "kswapd" should
1281          * never get caught in the normal page freeing logic.
1282          *
1283          * (Kswapd normally doesn't need memory anyway, but sometimes
1284          * you need a small amount of memory in order to be able to
1285          * page out something else, and this flag essentially protects
1286          * us from recursively trying to free more memory as we're
1287          * trying to free the first piece of memory in the first place).
1288          */
1289         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1290
1291         for ( ; ; ) {
1292                 if (current->flags & PF_FREEZE)
1293                         refrigerator(PF_FREEZE);
1294                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1295                 schedule();
1296                 finish_wait(&pgdat->kswapd_wait, &wait);
1297
1298                 if (!ckrm_shrink_list_empty())
1299                         ckrm_shrink_classes();
1300                 else
1301                 balance_pgdat(pgdat, 0);
1302         }
1303         return 0;
1304 }
1305
1306 /*
1307  * A zone is low on free memory, so wake its kswapd task to service it.
1308  */
1309 void wakeup_kswapd(struct zone *zone)
1310 {
1311         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1312                 return;
1313         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1314                 return;
1315         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1316 }
1317
1318 #ifdef CONFIG_PM
1319 /*
1320  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1321  * pages.
1322  */
1323 int shrink_all_memory(int nr_pages)
1324 {
1325         pg_data_t *pgdat;
1326         int nr_to_free = nr_pages;
1327         int ret = 0;
1328         struct reclaim_state reclaim_state = {
1329                 .reclaimed_slab = 0,
1330         };
1331
1332         current->reclaim_state = &reclaim_state;
1333         for_each_pgdat(pgdat) {
1334                 int freed;
1335                 freed = balance_pgdat(pgdat, nr_to_free);
1336                 ret += freed;
1337                 nr_to_free -= freed;
1338                 if (nr_to_free <= 0)
1339                         break;
1340         }
1341         current->reclaim_state = NULL;
1342         return ret;
1343 }
1344 #endif
1345
1346 #ifdef CONFIG_HOTPLUG_CPU
1347 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1348    not required for correctness.  So if the last cpu in a node goes
1349    away, we get changed to run anywhere: as the first one comes back,
1350    restore their cpu bindings. */
1351 static int __devinit cpu_callback(struct notifier_block *nfb,
1352                                   unsigned long action,
1353                                   void *hcpu)
1354 {
1355         pg_data_t *pgdat;
1356         cpumask_t mask;
1357
1358         if (action == CPU_ONLINE) {
1359                 for_each_pgdat(pgdat) {
1360                         mask = node_to_cpumask(pgdat->node_id);
1361                         if (any_online_cpu(mask) != NR_CPUS)
1362                                 /* One of our CPUs online: restore mask */
1363                                 set_cpus_allowed(pgdat->kswapd, mask);
1364                 }
1365         }
1366         return NOTIFY_OK;
1367 }
1368 #endif /* CONFIG_HOTPLUG_CPU */
1369
1370 static int __init kswapd_init(void)
1371 {
1372         pg_data_t *pgdat;
1373         swap_setup();
1374         for_each_pgdat(pgdat)
1375                 pgdat->kswapd
1376                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1377         total_memory = nr_free_pagecache_pages();
1378         hotcpu_notifier(cpu_callback, 0);
1379         return 0;
1380 }
1381
1382 module_init(kswapd_init)