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 #include <linux/ckrm_mem.h>
  43 #include <linux/vs_cvirt.h>
  44
  45
  46 /* possible outcome of pageout() */
  47 typedef enum {
  48         /* failed to write page out, page is locked */
  49         PAGE_KEEP,
  50         /* move page to the active list, page is locked */
  51         PAGE_ACTIVATE,
  52         /* page has been sent to the disk successfully, page is unlocked */
  53         PAGE_SUCCESS,
  54         /* page is clean and locked */
  55         PAGE_CLEAN,
  56 } pageout_t;
  57
  58 struct scan_control {
  59         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  60         unsigned long nr_to_scan;
  61
  62         /* Incremented by the number of inactive pages that were scanned */
  63         unsigned long nr_scanned;
  64
  65         /* Incremented by the number of pages reclaimed */
  66         unsigned long nr_reclaimed;
  67
  68         unsigned long nr_mapped;        /* From page_state */
  69
  70         /* How many pages shrink_cache() should reclaim */
  71         int nr_to_reclaim;
  72
  73         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  74         unsigned int priority;
  75
  76         /* This context's GFP mask */
  77         unsigned int gfp_mask;
  78
  79         int may_writepage;
  80 };
  81
  82 /*
  83  * The list of shrinker callbacks used by to apply pressure to
  84  * ageable caches.
  85  */
  86 struct shrinker {
  87         shrinker_t              shrinker;
  88         struct list_head        list;
  89         int                     seeks;  /* seeks to recreate an obj */
  90         long                    nr;     /* objs pending delete */
  91 };
  92
  93 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  94
  95 #ifdef ARCH_HAS_PREFETCH
  96 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  97         do {                                                            \
  98                 if ((_page)->lru.prev != _base) {                       \
  99                         struct page *prev;                              \
 100                                                                         \
 101                         prev = lru_to_page(&(_page->lru));              \
 102                         prefetch(&prev->_field);                        \
 103                 }                                                       \
 104         } while (0)
 105 #else
 106 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 107 #endif
 108
 109 #ifdef ARCH_HAS_PREFETCHW
 110 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 111         do {                                                            \
 112                 if ((_page)->lru.prev != _base) {                       \
 113                         struct page *prev;                              \
 114                                                                         \
 115                         prev = lru_to_page(&(_page->lru));              \
 116                         prefetchw(&prev->_field);                       \
 117                 }                                                       \
 118         } while (0)
 119 #else
 120 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 121 #endif
 122
 123 /*
 124  * From 0 .. 100.  Higher means more swappy.
 125  */
 126 int vm_swappiness = 60;
 127 static long total_memory;
 128
 129 static LIST_HEAD(shrinker_list);
 130 static DECLARE_RWSEM(shrinker_rwsem);
 131
 132 /*
 133  * Add a shrinker callback to be called from the vm
 134  */
 135 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 136 {
 137         struct shrinker *shrinker;
 138
 139         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 140         if (shrinker) {
 141                 shrinker->shrinker = theshrinker;
 142                 shrinker->seeks = seeks;
 143                 shrinker->nr = 0;
 144                 down_write(&shrinker_rwsem);
 145                 list_add(&shrinker->list, &shrinker_list);
 146                 up_write(&shrinker_rwsem);
 147         }
 148         return shrinker;
 149 }
 150 EXPORT_SYMBOL(set_shrinker);
 151
 152 /*
 153  * Remove one
 154  */
 155 void remove_shrinker(struct shrinker *shrinker)
 156 {
 157         down_write(&shrinker_rwsem);
 158         list_del(&shrinker->list);
 159         up_write(&shrinker_rwsem);
 160         kfree(shrinker);
 161 }
 162 EXPORT_SYMBOL(remove_shrinker);
 163
 164 #define SHRINK_BATCH 128
 165 /*
 166  * Call the shrink functions to age shrinkable caches
 167  *
 168  * Here we assume it costs one seek to replace a lru page and that it also
 169  * takes a seek to recreate a cache object.  With this in mind we age equal
 170  * percentages of the lru and ageable caches.  This should balance the seeks
 171  * generated by these structures.
 172  *
 173  * If the vm encounted mapped pages on the LRU it increase the pressure on
 174  * slab to avoid swapping.
 175  *
 176  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 177  *
 178  * `lru_pages' represents the number of on-LRU pages in all the zones which
 179  * are eligible for the caller's allocation attempt.  It is used for balancing
 180  * slab reclaim versus page reclaim.
 181  */
 182 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 183                         unsigned long lru_pages)
 184 {
 185         struct shrinker *shrinker;
 186
 187         if (scanned == 0)
 188                 scanned = SWAP_CLUSTER_MAX;
 189
 190         if (!down_read_trylock(&shrinker_rwsem))
 191                 return 0;
 192
 193         list_for_each_entry(shrinker, &shrinker_list, list) {
 194                 unsigned long long delta;
 195                 unsigned long total_scan;
 196
 197                 delta = (4 * scanned) / shrinker->seeks;
 198                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 199                 do_div(delta, lru_pages + 1);
 200                 shrinker->nr += delta;
 201                 if (shrinker->nr < 0)
 202                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 203
 204                 total_scan = shrinker->nr;
 205                 shrinker->nr = 0;
 206
 207                 while (total_scan >= SHRINK_BATCH) {
 208                         long this_scan = SHRINK_BATCH;
 209                         int shrink_ret;
 210
 211                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 212                         if (shrink_ret == -1)
 213                                 break;
 214                         mod_page_state(slabs_scanned, this_scan);
 215                         total_scan -= this_scan;
 216
 217                         cond_resched();
 218                 }
 219
 220                 shrinker->nr += total_scan;
 221         }
 222         up_read(&shrinker_rwsem);
 223         return 0;
 224 }
 225
 226 /* Called without lock on whether page is mapped, so answer is unstable */
 227 static inline int page_mapping_inuse(struct page *page)
 228 {
 229         struct address_space *mapping;
 230
 231         /* Page is in somebody's page tables. */
 232         if (page_mapped(page))
 233                 return 1;
 234
 235         /* Be more reluctant to reclaim swapcache than pagecache */
 236         if (PageSwapCache(page))
 237                 return 1;
 238
 239         mapping = page_mapping(page);
 240         if (!mapping)
 241                 return 0;
 242
 243         /* File is mmap'd by somebody? */
 244         return mapping_mapped(mapping);
 245 }
 246
 247 static inline int is_page_cache_freeable(struct page *page)
 248 {
 249         return page_count(page) - !!PagePrivate(page) == 2;
 250 }
 251
 252 static int may_write_to_queue(struct backing_dev_info *bdi)
 253 {
 254         if (current_is_kswapd())
 255                 return 1;
 256         if (current_is_pdflush())       /* This is unlikely, but why not... */
 257                 return 1;
 258         if (!bdi_write_congested(bdi))
 259                 return 1;
 260         if (bdi == current->backing_dev_info)
 261                 return 1;
 262         return 0;
 263 }
 264
 265 /*
 266  * We detected a synchronous write error writing a page out.  Probably
 267  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 268  * fsync(), msync() or close().
 269  *
 270  * The tricky part is that after writepage we cannot touch the mapping: nothing
 271  * prevents it from being freed up.  But we have a ref on the page and once
 272  * that page is locked, the mapping is pinned.
 273  *
 274  * We're allowed to run sleeping lock_page() here because we know the caller has
 275  * __GFP_FS.
 276  */
 277 static void handle_write_error(struct address_space *mapping,
 278                                 struct page *page, int error)
 279 {
 280         lock_page(page);
 281         if (page_mapping(page) == mapping) {
 282                 if (error == -ENOSPC)
 283                         set_bit(AS_ENOSPC, &mapping->flags);
 284                 else
 285                         set_bit(AS_EIO, &mapping->flags);
 286         }
 287         unlock_page(page);
 288 }
 289
 290 /*
 291  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 292  */
 293 static pageout_t pageout(struct page *page, struct address_space *mapping)
 294 {
 295         /*
 296          * If the page is dirty, only perform writeback if that write
 297          * will be non-blocking.  To prevent this allocation from being
 298          * stalled by pagecache activity.  But note that there may be
 299          * stalls if we need to run get_block().  We could test
 300          * PagePrivate for that.
 301          *
 302          * If this process is currently in generic_file_write() against
 303          * this page's queue, we can perform writeback even if that
 304          * will block.
 305          *
 306          * If the page is swapcache, write it back even if that would
 307          * block, for some throttling. This happens by accident, because
 308          * swap_backing_dev_info is bust: it doesn't reflect the
 309          * congestion state of the swapdevs.  Easy to fix, if needed.
 310          * See swapfile.c:page_queue_congested().
 311          */
 312         if (!is_page_cache_freeable(page))
 313                 return PAGE_KEEP;
 314         if (!mapping)
 315                 return PAGE_KEEP;
 316         if (mapping->a_ops->writepage == NULL)
 317                 return PAGE_ACTIVATE;
 318         if (!may_write_to_queue(mapping->backing_dev_info))
 319                 return PAGE_KEEP;
 320
 321         if (clear_page_dirty_for_io(page)) {
 322                 int res;
 323                 struct writeback_control wbc = {
 324                         .sync_mode = WB_SYNC_NONE,
 325                         .nr_to_write = SWAP_CLUSTER_MAX,
 326                         .nonblocking = 1,
 327                         .for_reclaim = 1,
 328                 };
 329
 330                 SetPageReclaim(page);
 331                 res = mapping->a_ops->writepage(page, &wbc);
 332                 if (res < 0)
 333                         handle_write_error(mapping, page, res);
 334                 if (res == WRITEPAGE_ACTIVATE) {
 335                         ClearPageReclaim(page);
 336                         return PAGE_ACTIVATE;
 337                 }
 338                 if (!PageWriteback(page)) {
 339                         /* synchronous write or broken a_ops? */
 340                         ClearPageReclaim(page);
 341                 }
 342
 343                 return PAGE_SUCCESS;
 344         }
 345
 346         return PAGE_CLEAN;
 347 }
 348
 349 /*
 350  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 351  */
 352 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 353 {
 354         LIST_HEAD(ret_pages);
 355         struct pagevec freed_pvec;
 356         int pgactivate = 0;
 357         int reclaimed = 0;
 358
 359         cond_resched();
 360
 361         pagevec_init(&freed_pvec, 1);
 362         while (!list_empty(page_list)) {
 363                 struct address_space *mapping;
 364                 struct page *page;
 365                 int may_enter_fs;
 366                 int referenced;
 367
 368                 page = lru_to_page(page_list);
 369                 list_del(&page->lru);
 370
 371                 if (TestSetPageLocked(page))
 372                         goto keep;
 373
 374                 BUG_ON(PageActive(page));
 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                 if (PageWriteback(page))
 382                         goto keep_locked;
 383
 384                 referenced = page_referenced(page, 1, sc->priority <= 0);
 385                 /* In active use or really unfreeable?  Activate it. */
 386                 if (referenced && page_mapping_inuse(page))
 387                         goto activate_locked;
 388
 389 #ifdef CONFIG_SWAP
 390                 /*
 391                  * Anonymous process memory has backing store?
 392                  * Try to allocate it some swap space here.
 393                  */
 394                 if (PageAnon(page) && !PageSwapCache(page)) {
 395                         if (!add_to_swap(page))
 396                                 goto activate_locked;
 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                                 goto activate_locked;
 412                         case SWAP_AGAIN:
 413                                 goto keep_locked;
 414                         case SWAP_SUCCESS:
 415                                 ; /* try to free the page below */
 416                         }
 417                 }
 418
 419                 if (PageDirty(page)) {
 420                         if (referenced)
 421                                 goto keep_locked;
 422                         if (!may_enter_fs)
 423                                 goto keep_locked;
 424                         if (laptop_mode && !sc->may_writepage)
 425                                 goto keep_locked;
 426
 427                         /* Page is dirty, try to write it out here */
 428                         switch(pageout(page, mapping)) {
 429                         case PAGE_KEEP:
 430                                 goto keep_locked;
 431                         case PAGE_ACTIVATE:
 432                                 goto activate_locked;
 433                         case PAGE_SUCCESS:
 434                                 if (PageWriteback(page) || PageDirty(page))
 435                                         goto keep;
 436                                 /*
 437                                  * A synchronous write - probably a ramdisk.  Go
 438                                  * ahead and try to reclaim the page.
 439                                  */
 440                                 if (TestSetPageLocked(page))
 441                                         goto keep;
 442                                 if (PageDirty(page) || PageWriteback(page))
 443                                         goto keep_locked;
 444                                 mapping = page_mapping(page);
 445                         case PAGE_CLEAN:
 446                                 ; /* try to free the page below */
 447                         }
 448                 }
 449
 450                 /*
 451                  * If the page has buffers, try to free the buffer mappings
 452                  * associated with this page. If we succeed we try to free
 453                  * the page as well.
 454                  *
 455                  * We do this even if the page is PageDirty().
 456                  * try_to_release_page() does not perform I/O, but it is
 457                  * possible for a page to have PageDirty set, but it is actually
 458                  * clean (all its buffers are clean).  This happens if the
 459                  * buffers were written out directly, with submit_bh(). ext3
 460                  * will do this, as well as the blockdev mapping.
 461                  * try_to_release_page() will discover that cleanness and will
 462                  * drop the buffers and mark the page clean - it can be freed.
 463                  *
 464                  * Rarely, pages can have buffers and no ->mapping.  These are
 465                  * the pages which were not successfully invalidated in
 466                  * truncate_complete_page().  We try to drop those buffers here
 467                  * and if that worked, and the page is no longer mapped into
 468                  * process address space (page_count == 1) it can be freed.
 469                  * Otherwise, leave the page on the LRU so it is swappable.
 470                  */
 471                 if (PagePrivate(page)) {
 472                         if (!try_to_release_page(page, sc->gfp_mask))
 473                                 goto activate_locked;
 474                         if (!mapping && page_count(page) == 1)
 475                                 goto free_it;
 476                 }
 477
 478                 if (!mapping)
 479                         goto keep_locked;       /* truncate got there first */
 480
 481                 spin_lock_irq(&mapping->tree_lock);
 482
 483                 /*
 484                  * The non-racy check for busy page.  It is critical to check
 485                  * PageDirty _after_ making sure that the page is freeable and
 486                  * not in use by anybody.       (pagecache + us == 2)
 487                  */
 488                 if (page_count(page) != 2 || PageDirty(page)) {
 489                         spin_unlock_irq(&mapping->tree_lock);
 490                         goto keep_locked;
 491                 }
 492
 493 #ifdef CONFIG_SWAP
 494                 if (PageSwapCache(page)) {
 495                         swp_entry_t swap = { .val = page->private };
 496                         __delete_from_swap_cache(page);
 497                         spin_unlock_irq(&mapping->tree_lock);
 498                         swap_free(swap);
 499                         __put_page(page);       /* The pagecache ref */
 500                         goto free_it;
 501                 }
 502 #endif /* CONFIG_SWAP */
 503
 504                 __remove_from_page_cache(page);
 505                 spin_unlock_irq(&mapping->tree_lock);
 506                 __put_page(page);
 507
 508 free_it:
 509                 unlock_page(page);
 510                 reclaimed++;
 511                 if (!pagevec_add(&freed_pvec, page))
 512                         __pagevec_release_nonlru(&freed_pvec);
 513                 continue;
 514
 515 activate_locked:
 516                 SetPageActive(page);
 517                 pgactivate++;
 518 keep_locked:
 519                 unlock_page(page);
 520 keep:
 521                 list_add(&page->lru, &ret_pages);
 522                 BUG_ON(PageLRU(page));
 523         }
 524         list_splice(&ret_pages, page_list);
 525         if (pagevec_count(&freed_pvec))
 526                 __pagevec_release_nonlru(&freed_pvec);
 527         mod_page_state(pgactivate, pgactivate);
 528         sc->nr_reclaimed += reclaimed;
 529         return reclaimed;
 530 }
 531
 532 /*
 533  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 534  * a batch of pages and working on them outside the lock.  Any pages which were
 535  * not freed will be added back to the LRU.
 536  *
 537  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 538  *
 539  * For pagecache intensive workloads, the first loop here is the hottest spot
 540  * in the kernel (apart from the copy_*_user functions).
 541  */
 542 #ifdef CONFIG_CKRM_RES_MEM
 543 static void shrink_cache(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 544 #else
 545 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 546 #endif
 547 {
 548         LIST_HEAD(page_list);
 549         struct pagevec pvec;
 550         int max_scan = sc->nr_to_scan;
 551 #ifdef CONFIG_CKRM_RES_MEM
 552         struct zone *zone = ckrm_zone->zone;
 553         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 554         struct list_head *active_list = &ckrm_zone->active_list;
 555 #else
 556         struct list_head *inactive_list = &zone->inactive_list;
 557         struct list_head *active_list = &zone->active_list;
 558 #endif
 559
 560         pagevec_init(&pvec, 1);
 561
 562         lru_add_drain();
 563         spin_lock_irq(&zone->lru_lock);
 564         while (max_scan > 0) {
 565                 struct page *page;
 566                 int nr_taken = 0;
 567                 int nr_scan = 0;
 568                 int nr_freed;
 569
 570                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 571                                 !list_empty(inactive_list)) {
 572                         page = lru_to_page(inactive_list);
 573
 574                         prefetchw_prev_lru_page(page,
 575                                                 inactive_list, flags);
 576
 577                         if (!TestClearPageLRU(page))
 578                                 BUG();
 579                         list_del(&page->lru);
 580                         if (get_page_testone(page)) {
 581                                 /*
 582                                  * It is being freed elsewhere
 583                                  */
 584                                 __put_page(page);
 585                                 SetPageLRU(page);
 586                                 list_add(&page->lru, inactive_list);
 587                                 continue;
 588                         }
 589                         list_add(&page->lru, &page_list);
 590                         nr_taken++;
 591                 }
 592                 zone->nr_inactive -= nr_taken;
 593                 ckrm_zone_sub_inactive(ckrm_zone, nr_taken);
 594                 spin_unlock_irq(&zone->lru_lock);
 595
 596                 if (nr_taken == 0)
 597                         goto done;
 598
 599                 max_scan -= nr_scan;
 600                 if (current_is_kswapd())
 601                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 602                 else
 603                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 604                 nr_freed = shrink_list(&page_list, sc);
 605                 if (current_is_kswapd())
 606                         mod_page_state(kswapd_steal, nr_freed);
 607                 mod_page_state_zone(zone, pgsteal, nr_freed);
 608                 sc->nr_to_reclaim -= nr_freed;
 609
 610                 spin_lock_irq(&zone->lru_lock);
 611                 /*
 612                  * Put back any unfreeable pages.
 613                  */
 614                 while (!list_empty(&page_list)) {
 615                         page = lru_to_page(&page_list);
 616                         if (TestSetPageLRU(page))
 617                                 BUG();
 618                         list_del(&page->lru);
 619                         if (PageActive(page)) {
 620                                 ckrm_zone_add_active(ckrm_zone, 1);
 621                                 zone->nr_active++;
 622                                 list_add(&page->lru, active_list);
 623                         } else {
 624                                 ckrm_zone_add_inactive(ckrm_zone, 1);
 625                                 zone->nr_inactive++;
 626                                 list_add(&page->lru, inactive_list);
 627                         }
 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         }
 635         spin_unlock_irq(&zone->lru_lock);
 636 done:
 637         pagevec_release(&pvec);
 638 }
 639
 640 /*
 641  * This moves pages from the active list to the inactive list.
 642  *
 643  * We move them the other way if the page is referenced by one or more
 644  * processes, from rmap.
 645  *
 646  * If the pages are mostly unmapped, the processing is fast and it is
 647  * appropriate to hold zone->lru_lock across the whole operation.  But if
 648  * the pages are mapped, the processing is slow (page_referenced()) so we
 649  * should drop zone->lru_lock around each page.  It's impossible to balance
 650  * this, so instead we remove the pages from the LRU while processing them.
 651  * It is safe to rely on PG_active against the non-LRU pages in here because
 652  * nobody will play with that bit on a non-LRU page.
 653  *
 654  * The downside is that we have to touch page->_count against each page.
 655  * But we had to alter page->flags anyway.
 656  */
 657 static void
 658 #ifdef CONFIG_CKRM_RES_MEM
 659 refill_inactive_zone(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 660 #else
 661 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 662 #endif
 663 {
 664         int pgmoved;
 665         int pgdeactivate = 0;
 666         int pgscanned = 0;
 667         int nr_pages = sc->nr_to_scan;
 668         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 669         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 670         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 671         struct page *page;
 672         struct pagevec pvec;
 673         int reclaim_mapped = 0;
 674         long mapped_ratio;
 675         long distress;
 676         long swap_tendency;
 677 #ifdef CONFIG_CKRM_RES_MEM
 678         struct zone *zone = ckrm_zone->zone;
 679         struct list_head *active_list = &ckrm_zone->active_list;
 680         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 681 #else
 682         struct list_head *active_list = &zone->active_list;
 683         struct list_head *inactive_list = &zone->inactive_list;
 684 #endif
 685
 686         lru_add_drain();
 687         pgmoved = 0;
 688         spin_lock_irq(&zone->lru_lock);
 689         while (pgscanned < nr_pages && !list_empty(active_list)) {
 690                 page = lru_to_page(active_list);
 691                 prefetchw_prev_lru_page(page, active_list, flags);
 692                 if (!TestClearPageLRU(page))
 693                         BUG();
 694                 list_del(&page->lru);
 695                 if (get_page_testone(page)) {
 696                         /*
 697                          * It was already free!  release_pages() or put_page()
 698                          * are about to remove it from the LRU and free it. So
 699                          * put the refcount back and put the page back on the
 700                          * LRU
 701                          */
 702                         __put_page(page);
 703                         SetPageLRU(page);
 704                         list_add(&page->lru, active_list);
 705                 } else {
 706                         list_add(&page->lru, &l_hold);
 707                         pgmoved++;
 708                 }
 709                 pgscanned++;
 710         }
 711         zone->pages_scanned += pgscanned;
 712         zone->nr_active -= pgmoved;
 713         ckrm_zone_sub_active(ckrm_zone, pgmoved);
 714         spin_unlock_irq(&zone->lru_lock);
 715
 716         /*
 717          * `distress' is a measure of how much trouble we're having reclaiming
 718          * pages.  0 -> no problems.  100 -> great trouble.
 719          */
 720         distress = 100 >> zone->prev_priority;
 721
 722         /*
 723          * The point of this algorithm is to decide when to start reclaiming
 724          * mapped memory instead of just pagecache.  Work out how much memory
 725          * is mapped.
 726          */
 727         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 728
 729         /*
 730          * Now decide how much we really want to unmap some pages.  The mapped
 731          * ratio is downgraded - just because there's a lot of mapped memory
 732          * doesn't necessarily mean that page reclaim isn't succeeding.
 733          *
 734          * The distress ratio is important - we don't want to start going oom.
 735          *
 736          * A 100% value of vm_swappiness overrides this algorithm altogether.
 737          */
 738         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 739
 740         /*
 741          * Now use this metric to decide whether to start moving mapped memory
 742          * onto the inactive list.
 743          */
 744         if (swap_tendency >= 100)
 745                 reclaim_mapped = 1;
 746
 747         while (!list_empty(&l_hold)) {
 748                 page = lru_to_page(&l_hold);
 749                 list_del(&page->lru);
 750                 if (page_mapped(page)) {
 751                         if (!reclaim_mapped ||
 752                             (total_swap_pages == 0 && PageAnon(page)) ||
 753                             page_referenced(page, 0, sc->priority <= 0)) {
 754                                 list_add(&page->lru, &l_active);
 755                                 continue;
 756                         }
 757                 }
 758                 list_add(&page->lru, &l_inactive);
 759         }
 760
 761         pagevec_init(&pvec, 1);
 762         pgmoved = 0;
 763         spin_lock_irq(&zone->lru_lock);
 764         while (!list_empty(&l_inactive)) {
 765                 page = lru_to_page(&l_inactive);
 766                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 767                 if (TestSetPageLRU(page))
 768                         BUG();
 769                 if (!TestClearPageActive(page))
 770                         BUG();
 771                 list_move(&page->lru, inactive_list);
 772                 pgmoved++;
 773                 if (!pagevec_add(&pvec, page)) {
 774                         zone->nr_inactive += pgmoved;
 775                         ckrm_zone_add_inactive(ckrm_zone, pgmoved);
 776                         spin_unlock_irq(&zone->lru_lock);
 777                         pgdeactivate += pgmoved;
 778                         pgmoved = 0;
 779                         if (buffer_heads_over_limit)
 780                                 pagevec_strip(&pvec);
 781                         __pagevec_release(&pvec);
 782                         spin_lock_irq(&zone->lru_lock);
 783                 }
 784         }
 785         zone->nr_inactive += pgmoved;
 786         ckrm_zone_add_inactive(ckrm_zone, pgmoved);
 787         pgdeactivate += pgmoved;
 788         if (buffer_heads_over_limit) {
 789                 spin_unlock_irq(&zone->lru_lock);
 790                 pagevec_strip(&pvec);
 791                 spin_lock_irq(&zone->lru_lock);
 792         }
 793
 794         pgmoved = 0;
 795         while (!list_empty(&l_active)) {
 796                 page = lru_to_page(&l_active);
 797                 prefetchw_prev_lru_page(page, &l_active, flags);
 798                 if (TestSetPageLRU(page))
 799                         BUG();
 800                 BUG_ON(!PageActive(page));
 801                 list_move(&page->lru, active_list);
 802                 pgmoved++;
 803                 if (!pagevec_add(&pvec, page)) {
 804                         zone->nr_active += pgmoved;
 805                         ckrm_zone_add_active(ckrm_zone, pgmoved);
 806                         pgmoved = 0;
 807                         spin_unlock_irq(&zone->lru_lock);
 808                         __pagevec_release(&pvec);
 809                         spin_lock_irq(&zone->lru_lock);
 810                 }
 811         }
 812         zone->nr_active += pgmoved;
 813         ckrm_zone_add_active(ckrm_zone, pgmoved);
 814         spin_unlock_irq(&zone->lru_lock);
 815         pagevec_release(&pvec);
 816
 817         mod_page_state_zone(zone, pgrefill, pgscanned);
 818         mod_page_state(pgdeactivate, pgdeactivate);
 819 }
 820
 821 #ifdef CONFIG_CKRM_RES_MEM
 822 static void
 823 shrink_ckrmzone(struct ckrm_zone *czone, struct scan_control *sc)
 824 {
 825         while (czone->shrink_active || czone->shrink_inactive) {
 826                 if (czone->shrink_active) {
 827                         sc->nr_to_scan = min(czone->shrink_active,
 828                                         (unsigned long)SWAP_CLUSTER_MAX);
 829                         czone->shrink_active -= sc->nr_to_scan;
 830                         refill_inactive_zone(czone, sc);
 831                 }
 832                 if (czone->shrink_inactive) {
 833                         sc->nr_to_scan = min(czone->shrink_inactive,
 834                                         (unsigned long)SWAP_CLUSTER_MAX);
 835                         czone->shrink_inactive -= sc->nr_to_scan;
 836                         shrink_cache(czone, sc);
 837                         if (sc->nr_to_reclaim <= 0) {
 838                                 czone->shrink_active = 0;
 839                                 czone->shrink_inactive = 0;
 840                                 break;
 841                         }
 842                 }
 843         }
 844 }
 845
 846 /* FIXME: This function needs to be given more thought. */
 847 static void
 848 ckrm_shrink_class(struct ckrm_mem_res *cls)
 849 {
 850         struct scan_control sc;
 851         struct zone *zone;
 852         int zindex = 0, cnt, act_credit = 0, inact_credit = 0;
 853
 854         sc.nr_mapped = read_page_state(nr_mapped);
 855         sc.nr_scanned = 0;
 856         sc.nr_reclaimed = 0;
 857         sc.priority = 0; // always very high priority
 858
 859         for_each_zone(zone) {
 860                 int zone_total, zone_limit, active_limit,
 861                                         inactive_limit, clszone_limit;
 862                 struct ckrm_zone *czone;
 863                 u64 temp;
 864
 865                 czone = &cls->ckrm_zone[zindex];
 866                 if (ckrm_test_set_shrink(czone))
 867                         continue;
 868
 869                 zone->temp_priority = zone->prev_priority;
 870                 zone->prev_priority = sc.priority;
 871
 872                 zone_total = zone->nr_active + zone->nr_inactive
 873                                                 + zone->free_pages;
 874
 875                 temp = (u64) cls->pg_limit * zone_total;
 876                 do_div(temp, ckrm_tot_lru_pages);
 877                 zone_limit = (int) temp;
 878                 clszone_limit = (ckrm_mem_shrink_to * zone_limit) / 100;
 879                 active_limit = (2 * clszone_limit) / 3; // 2/3rd in active list
 880                 inactive_limit = clszone_limit / 3; // 1/3rd in inactive list
 881
 882                 czone->shrink_active = 0;
 883                 cnt = czone->nr_active + act_credit - active_limit;
 884                 if (cnt > 0) {
 885                         czone->shrink_active = (unsigned long) cnt;
 886                         act_credit = 0;
 887                 } else {
 888                         act_credit += cnt;
 889                 }
 890
 891                 czone->shrink_inactive = 0;
 892                 cnt = czone->shrink_active + inact_credit +
 893                                         (czone->nr_inactive - inactive_limit);
 894                 if (cnt > 0) {
 895                         czone->shrink_inactive = (unsigned long) cnt;
 896                         inact_credit = 0;
 897                 } else {
 898                         inact_credit += cnt;
 899                 }
 900
 901
 902                 if (czone->shrink_active || czone->shrink_inactive) {
 903                         sc.nr_to_reclaim = czone->shrink_inactive;
 904                         shrink_ckrmzone(czone, &sc);
 905                 }
 906                 zone->prev_priority = zone->temp_priority;
 907                 zindex++;
 908                 ckrm_clear_shrink(czone);
 909         }
 910 }
 911
 912 static void
 913 ckrm_shrink_classes(void)
 914 {
 915         struct ckrm_mem_res *cls;
 916
 917         spin_lock(&ckrm_mem_lock);
 918         while (!ckrm_shrink_list_empty()) {
 919                 cls =  list_entry(ckrm_shrink_list.next, struct ckrm_mem_res,
 920                                 shrink_list);
 921                 list_del(&cls->shrink_list);
 922                 cls->flags &= ~CLS_AT_LIMIT;
 923                 spin_unlock(&ckrm_mem_lock);
 924                 ckrm_shrink_class(cls);
 925                 spin_lock(&ckrm_mem_lock);
 926         }
 927         spin_unlock(&ckrm_mem_lock);
 928 }
 929
 930 #else
 931 #define ckrm_shrink_classes()   do { } while(0)
 932 #endif
 933
 934 /*
 935  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 936  */
 937 static void
 938 shrink_zone(struct zone *zone, struct scan_control *sc)
 939 {
 940         unsigned long nr_active;
 941         unsigned long nr_inactive;
 942 #ifdef CONFIG_CKRM_RES_MEM
 943         struct ckrm_zone *czone;
 944 #endif
 945
 946         /*
 947          * Add one to `nr_to_scan' just to make sure that the kernel will
 948          * slowly sift through the active list.
 949          */
 950         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 951         nr_active = zone->nr_scan_active;
 952         if (nr_active >= SWAP_CLUSTER_MAX)
 953                 zone->nr_scan_active = 0;
 954         else
 955                 nr_active = 0;
 956
 957         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 958         nr_inactive = zone->nr_scan_inactive;
 959         if (nr_inactive >= SWAP_CLUSTER_MAX)
 960                 zone->nr_scan_inactive = 0;
 961         else
 962                 nr_inactive = 0;
 963
 964         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 965
 966 #ifdef CONFIG_CKRM_RES_MEM
 967         if (nr_active || nr_inactive) {
 968                 struct list_head *pos, *next;
 969                 LIST_HEAD(victims);
 970
 971                 shrink_get_victims(zone, nr_active, nr_inactive, &victims);
 972                 pos = victims.next;
 973                 while (pos != &victims) {
 974                         czone = list_entry(pos, struct ckrm_zone, victim_list);
 975                         next = pos->next;
 976                         list_del_init(pos);
 977                         sc->nr_to_reclaim = czone->shrink_inactive;
 978                         shrink_ckrmzone(czone, sc);
 979                         ckrm_clear_shrink(czone);
 980                         pos = next;
 981                 }
 982         }
 983 #else
 984         while (nr_active || nr_inactive) {
 985                 if (nr_active) {
 986                         sc->nr_to_scan = min(nr_active,
 987                                         (unsigned long)SWAP_CLUSTER_MAX);
 988                         nr_active -= sc->nr_to_scan;
 989                         refill_inactive_zone(zone, sc);
 990                 }
 991
 992                 if (nr_inactive) {
 993                         sc->nr_to_scan = min(nr_inactive,
 994                                         (unsigned long)SWAP_CLUSTER_MAX);
 995                         nr_inactive -= sc->nr_to_scan;
 996                         shrink_cache(zone, sc);
 997                         if (sc->nr_to_reclaim <= 0)
 998                                 break;
 999                 }
1000         }
1001 #endif
1002 }
1003
1004 /*
1005  * This is the direct reclaim path, for page-allocating processes.  We only
1006  * try to reclaim pages from zones which will satisfy the caller's allocation
1007  * request.
1008  *
1009  * We reclaim from a zone even if that zone is over pages_high.  Because:
1010  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1011  *    allocation or
1012  * b) The zones may be over pages_high but they must go *over* pages_high to
1013  *    satisfy the `incremental min' zone defense algorithm.
1014  *
1015  * Returns the number of reclaimed pages.
1016  *
1017  * If a zone is deemed to be full of pinned pages then just give it a light
1018  * scan then give up on it.
1019  */
1020 static void
1021 shrink_caches(struct zone **zones, struct scan_control *sc)
1022 {
1023         int i;
1024
1025         for (i = 0; zones[i] != NULL; i++) {
1026                 struct zone *zone = zones[i];
1027
1028                 if (zone->present_pages == 0)
1029                         continue;
1030
1031                 zone->temp_priority = sc->priority;
1032                 if (zone->prev_priority > sc->priority)
1033                         zone->prev_priority = sc->priority;
1034
1035                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1036                         continue;       /* Let kswapd poll it */
1037
1038                 shrink_zone(zone, sc);
1039         }
1040 }
1041
1042 /*
1043  * This is the main entry point to direct page reclaim.
1044  *
1045  * If a full scan of the inactive list fails to free enough memory then we
1046  * are "out of memory" and something needs to be killed.
1047  *
1048  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1049  * high - the zone may be full of dirty or under-writeback pages, which this
1050  * caller can't do much about.  We kick pdflush and take explicit naps in the
1051  * hope that some of these pages can be written.  But if the allocating task
1052  * holds filesystem locks which prevent writeout this might not work, and the
1053  * allocation attempt will fail.
1054  */
1055 int try_to_free_pages(struct zone **zones,
1056                 unsigned int gfp_mask, unsigned int order)
1057 {
1058         int priority;
1059         int ret = 0;
1060         int total_scanned = 0, total_reclaimed = 0;
1061         struct reclaim_state *reclaim_state = current->reclaim_state;
1062         struct scan_control sc;
1063         unsigned long lru_pages = 0;
1064         int i;
1065
1066         sc.gfp_mask = gfp_mask;
1067         sc.may_writepage = 0;
1068
1069         inc_page_state(allocstall);
1070
1071         for (i = 0; zones[i] != NULL; i++) {
1072                 struct zone *zone = zones[i];
1073
1074                 zone->temp_priority = DEF_PRIORITY;
1075                 lru_pages += zone->nr_active + zone->nr_inactive;
1076         }
1077
1078         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1079                 sc.nr_mapped = read_page_state(nr_mapped);
1080                 sc.nr_scanned = 0;
1081                 sc.nr_reclaimed = 0;
1082                 sc.priority = priority;
1083                 shrink_caches(zones, &sc);
1084                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1085                 if (reclaim_state) {
1086                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1087                         reclaim_state->reclaimed_slab = 0;
1088                 }
1089                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1090                         ret = 1;
1091                         goto out;
1092                 }
1093                 total_scanned += sc.nr_scanned;
1094                 total_reclaimed += sc.nr_reclaimed;
1095
1096                 /*
1097                  * Try to write back as many pages as we just scanned.  This
1098                  * tends to cause slow streaming writers to write data to the
1099                  * disk smoothly, at the dirtying rate, which is nice.   But
1100                  * that's undesirable in laptop mode, where we *want* lumpy
1101                  * writeout.  So in laptop mode, write out the whole world.
1102                  */
1103                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1104                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1105                         sc.may_writepage = 1;
1106                 }
1107
1108                 /* Take a nap, wait for some writeback to complete */
1109                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1110                         blk_congestion_wait(WRITE, HZ/10);
1111         }
1112         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1113                 out_of_memory(gfp_mask);
1114 out:
1115         for (i = 0; zones[i] != 0; i++)
1116                 zones[i]->prev_priority = zones[i]->temp_priority;
1117         return ret;
1118 }
1119
1120 /*
1121  * For kswapd, balance_pgdat() will work across all this node's zones until
1122  * they are all at pages_high.
1123  *
1124  * If `nr_pages' is non-zero then it is the number of pages which are to be
1125  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1126  * special.
1127  *
1128  * Returns the number of pages which were actually freed.
1129  *
1130  * There is special handling here for zones which are full of pinned pages.
1131  * This can happen if the pages are all mlocked, or if they are all used by
1132  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1133  * What we do is to detect the case where all pages in the zone have been
1134  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1135  * dead and from now on, only perform a short scan.  Basically we're polling
1136  * the zone for when the problem goes away.
1137  *
1138  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1139  * zones which have free_pages > pages_high, but once a zone is found to have
1140  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1141  * of the number of free pages in the lower zones.  This interoperates with
1142  * the page allocator fallback scheme to ensure that aging of pages is balanced
1143  * across the zones.
1144  */
1145 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1146 {
1147         int to_free = nr_pages;
1148         int all_zones_ok;
1149         int priority;
1150         int i;
1151         int total_scanned, total_reclaimed;
1152         struct reclaim_state *reclaim_state = current->reclaim_state;
1153         struct scan_control sc;
1154
1155 loop_again:
1156         total_scanned = 0;
1157         total_reclaimed = 0;
1158         sc.gfp_mask = GFP_KERNEL;
1159         sc.may_writepage = 0;
1160         sc.nr_mapped = read_page_state(nr_mapped);
1161
1162         inc_page_state(pageoutrun);
1163
1164         for (i = 0; i < pgdat->nr_zones; i++) {
1165                 struct zone *zone = pgdat->node_zones + i;
1166
1167                 zone->temp_priority = DEF_PRIORITY;
1168         }
1169
1170         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1171                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1172                 unsigned long lru_pages = 0;
1173
1174                 all_zones_ok = 1;
1175
1176                 if (nr_pages == 0) {
1177                         /*
1178                          * Scan in the highmem->dma direction for the highest
1179                          * zone which needs scanning
1180                          */
1181                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1182                                 struct zone *zone = pgdat->node_zones + i;
1183
1184                                 if (zone->present_pages == 0)
1185                                         continue;
1186
1187                                 if (zone->all_unreclaimable &&
1188                                                 priority != DEF_PRIORITY)
1189                                         continue;
1190
1191                                 if (zone->free_pages <= zone->pages_high) {
1192                                         end_zone = i;
1193                                         goto scan;
1194                                 }
1195                         }
1196                         goto out;
1197                 } else {
1198                         end_zone = pgdat->nr_zones - 1;
1199                 }
1200 scan:
1201                 for (i = 0; i <= end_zone; i++) {
1202                         struct zone *zone = pgdat->node_zones + i;
1203
1204                         lru_pages += zone->nr_active + zone->nr_inactive;
1205                 }
1206
1207                 /*
1208                  * Now scan the zone in the dma->highmem direction, stopping
1209                  * at the last zone which needs scanning.
1210                  *
1211                  * We do this because the page allocator works in the opposite
1212                  * direction.  This prevents the page allocator from allocating
1213                  * pages behind kswapd's direction of progress, which would
1214                  * cause too much scanning of the lower zones.
1215                  */
1216                 for (i = 0; i <= end_zone; i++) {
1217                         struct zone *zone = pgdat->node_zones + i;
1218
1219                         if (zone->present_pages == 0)
1220                                 continue;
1221
1222                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1223                                 continue;
1224
1225                         if (nr_pages == 0) {    /* Not software suspend */
1226                                 if (zone->free_pages <= zone->pages_high)
1227                                         all_zones_ok = 0;
1228                         }
1229                         zone->temp_priority = priority;
1230                         if (zone->prev_priority > priority)
1231                                 zone->prev_priority = priority;
1232                         sc.nr_scanned = 0;
1233                         sc.nr_reclaimed = 0;
1234                         sc.priority = priority;
1235                         shrink_zone(zone, &sc);
1236                         reclaim_state->reclaimed_slab = 0;
1237                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1238                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1239                         total_reclaimed += sc.nr_reclaimed;
1240                         total_scanned += sc.nr_scanned;
1241                         if (zone->all_unreclaimable)
1242                                 continue;
1243                         if (zone->pages_scanned >= (zone->nr_active +
1244                                                         zone->nr_inactive) * 4)
1245                                 zone->all_unreclaimable = 1;
1246                         /*
1247                          * If we've done a decent amount of scanning and
1248                          * the reclaim ratio is low, start doing writepage
1249                          * even in laptop mode
1250                          */
1251                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1252                             total_scanned > total_reclaimed+total_reclaimed/2)
1253                                 sc.may_writepage = 1;
1254                 }
1255                 if (nr_pages && to_free > total_reclaimed)
1256                         continue;       /* swsusp: need to do more work */
1257                 if (all_zones_ok)
1258                         break;          /* kswapd: all done */
1259                 /*
1260                  * OK, kswapd is getting into trouble.  Take a nap, then take
1261                  * another pass across the zones.
1262                  */
1263                 if (total_scanned && priority < DEF_PRIORITY - 2)
1264                         blk_congestion_wait(WRITE, HZ/10);
1265
1266                 /*
1267                  * We do this so kswapd doesn't build up large priorities for
1268                  * example when it is freeing in parallel with allocators. It
1269                  * matches the direct reclaim path behaviour in terms of impact
1270                  * on zone->*_priority.
1271                  */
1272                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1273                         break;
1274         }
1275 out:
1276         for (i = 0; i < pgdat->nr_zones; i++) {
1277                 struct zone *zone = pgdat->node_zones + i;
1278
1279                 zone->prev_priority = zone->temp_priority;
1280         }
1281         if (!all_zones_ok) {
1282                 cond_resched();
1283                 goto loop_again;
1284         }
1285
1286         return total_reclaimed;
1287 }
1288
1289 /*
1290  * The background pageout daemon, started as a kernel thread
1291  * from the init process.
1292  *
1293  * This basically trickles out pages so that we have _some_
1294  * free memory available even if there is no other activity
1295  * that frees anything up. This is needed for things like routing
1296  * etc, where we otherwise might have all activity going on in
1297  * asynchronous contexts that cannot page things out.
1298  *
1299  * If there are applications that are active memory-allocators
1300  * (most normal use), this basically shouldn't matter.
1301  */
1302 static int kswapd(void *p)
1303 {
1304         pg_data_t *pgdat = (pg_data_t*)p;
1305         struct task_struct *tsk = current;
1306         DEFINE_WAIT(wait);
1307         struct reclaim_state reclaim_state = {
1308                 .reclaimed_slab = 0,
1309         };
1310         cpumask_t cpumask;
1311
1312         daemonize("kswapd%d", pgdat->node_id);
1313         cpumask = node_to_cpumask(pgdat->node_id);
1314         if (!cpus_empty(cpumask))
1315                 set_cpus_allowed(tsk, cpumask);
1316         current->reclaim_state = &reclaim_state;
1317
1318         /*
1319          * Tell the memory management that we're a "memory allocator",
1320          * and that if we need more memory we should get access to it
1321          * regardless (see "__alloc_pages()"). "kswapd" should
1322          * never get caught in the normal page freeing logic.
1323          *
1324          * (Kswapd normally doesn't need memory anyway, but sometimes
1325          * you need a small amount of memory in order to be able to
1326          * page out something else, and this flag essentially protects
1327          * us from recursively trying to free more memory as we're
1328          * trying to free the first piece of memory in the first place).
1329          */
1330         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1331
1332         for ( ; ; ) {
1333                 if (current->flags & PF_FREEZE)
1334                         refrigerator(PF_FREEZE);
1335                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1336                 schedule();
1337                 finish_wait(&pgdat->kswapd_wait, &wait);
1338
1339                 if (!ckrm_shrink_list_empty())
1340                         ckrm_shrink_classes();
1341                 else
1342                         balance_pgdat(pgdat, 0);
1343         }
1344         return 0;
1345 }
1346
1347 /*
1348  * A zone is low on free memory, so wake its kswapd task to service it.
1349  */
1350 void wakeup_kswapd(struct zone *zone)
1351 {
1352         if (zone->present_pages == 0)
1353                 return;
1354         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1355                 return;
1356         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1357                 return;
1358         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1359 }
1360
1361 #ifdef CONFIG_PM
1362 /*
1363  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1364  * pages.
1365  */
1366 int shrink_all_memory(int nr_pages)
1367 {
1368         pg_data_t *pgdat;
1369         int nr_to_free = nr_pages;
1370         int ret = 0;
1371         struct reclaim_state reclaim_state = {
1372                 .reclaimed_slab = 0,
1373         };
1374
1375         current->reclaim_state = &reclaim_state;
1376         for_each_pgdat(pgdat) {
1377                 int freed;
1378                 freed = balance_pgdat(pgdat, nr_to_free);
1379                 ret += freed;
1380                 nr_to_free -= freed;
1381                 if (nr_to_free <= 0)
1382                         break;
1383         }
1384         current->reclaim_state = NULL;
1385         return ret;
1386 }
1387 #endif
1388
1389 #ifdef CONFIG_HOTPLUG_CPU
1390 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1391    not required for correctness.  So if the last cpu in a node goes
1392    away, we get changed to run anywhere: as the first one comes back,
1393    restore their cpu bindings. */
1394 static int __devinit cpu_callback(struct notifier_block *nfb,
1395                                   unsigned long action,
1396                                   void *hcpu)
1397 {
1398         pg_data_t *pgdat;
1399         cpumask_t mask;
1400
1401         if (action == CPU_ONLINE) {
1402                 for_each_pgdat(pgdat) {
1403                         mask = node_to_cpumask(pgdat->node_id);
1404                         if (any_online_cpu(mask) != NR_CPUS)
1405                                 /* One of our CPUs online: restore mask */
1406                                 set_cpus_allowed(pgdat->kswapd, mask);
1407                 }
1408         }
1409         return NOTIFY_OK;
1410 }
1411 #endif /* CONFIG_HOTPLUG_CPU */
1412
1413 static int __init kswapd_init(void)
1414 {
1415         pg_data_t *pgdat;
1416         swap_setup();
1417         for_each_pgdat(pgdat)
1418                 pgdat->kswapd
1419                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1420         total_memory = nr_free_pagecache_pages();
1421         hotcpu_notifier(cpu_callback, 0);
1422         return 0;
1423 }
1424
1425 module_init(kswapd_init)