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
  93
  94 void try_to_clip_inodes(void);
  95
  96
  97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  98
  99 #ifdef ARCH_HAS_PREFETCH
 100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
 101         do {                                                            \
 102                 if ((_page)->lru.prev != _base) {                       \
 103                         struct page *prev;                              \
 104                                                                         \
 105                         prev = lru_to_page(&(_page->lru));              \
 106                         prefetch(&prev->_field);                        \
 107                 }                                                       \
 108         } while (0)
 109 #else
 110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 111 #endif
 112
 113 #ifdef ARCH_HAS_PREFETCHW
 114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 115         do {                                                            \
 116                 if ((_page)->lru.prev != _base) {                       \
 117                         struct page *prev;                              \
 118                                                                         \
 119                         prev = lru_to_page(&(_page->lru));              \
 120                         prefetchw(&prev->_field);                       \
 121                 }                                                       \
 122         } while (0)
 123 #else
 124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 125 #endif
 126
 127 /*
 128  * From 0 .. 100.  Higher means more swappy.
 129  */
 130 int vm_swappiness = 60;
 131 static long total_memory;
 132
 133 static LIST_HEAD(shrinker_list);
 134 static DECLARE_MUTEX(shrinker_sem);
 135
 136 /*
 137  * Add a shrinker callback to be called from the vm
 138  */
 139 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 140 {
 141         struct shrinker *shrinker;
 142
 143         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 144         if (shrinker) {
 145                 shrinker->shrinker = theshrinker;
 146                 shrinker->seeks = seeks;
 147                 shrinker->nr = 0;
 148                 down(&shrinker_sem);
 149                 list_add(&shrinker->list, &shrinker_list);
 150                 up(&shrinker_sem);
 151         }
 152         return shrinker;
 153 }
 154 EXPORT_SYMBOL(set_shrinker);
 155
 156 /*
 157  * Remove one
 158  */
 159 void remove_shrinker(struct shrinker *shrinker)
 160 {
 161         down(&shrinker_sem);
 162         list_del(&shrinker->list);
 163         up(&shrinker_sem);
 164         kfree(shrinker);
 165 }
 166 EXPORT_SYMBOL(remove_shrinker);
 167
 168 #define SHRINK_BATCH 128
 169 /*
 170  * Call the shrink functions to age shrinkable caches
 171  *
 172  * Here we assume it costs one seek to replace a lru page and that it also
 173  * takes a seek to recreate a cache object.  With this in mind we age equal
 174  * percentages of the lru and ageable caches.  This should balance the seeks
 175  * generated by these structures.
 176  *
 177  * If the vm encounted mapped pages on the LRU it increase the pressure on
 178  * slab to avoid swapping.
 179  *
 180  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 181  *
 182  * `lru_pages' represents the number of on-LRU pages in all the zones which
 183  * are eligible for the caller's allocation attempt.  It is used for balancing
 184  * slab reclaim versus page reclaim.
 185  */
 186 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 187                         unsigned long lru_pages)
 188 {
 189         struct shrinker *shrinker;
 190
 191         if (down_trylock(&shrinker_sem))
 192                 return 0;
 193
 194         list_for_each_entry(shrinker, &shrinker_list, list) {
 195                 unsigned long long delta;
 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                 if (shrinker->nr <= SHRINK_BATCH)
 205                         continue;
 206                 while (shrinker->nr) {
 207                         long this_scan = shrinker->nr;
 208                         int shrink_ret;
 209
 210                         if (this_scan > 128)
 211                                 this_scan = 128;
 212                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 213                         mod_page_state(slabs_scanned, this_scan);
 214                         shrinker->nr -= this_scan;
 215                         if (shrink_ret == -1)
 216                                 break;
 217                         cond_resched();
 218                 }
 219         }
 220         up(&shrinker_sem);
 221         return 0;
 222 }
 223
 224 /* Must be called with page's rmap lock held. */
 225 static inline int page_mapping_inuse(struct page *page)
 226 {
 227         struct address_space *mapping;
 228
 229         /* Page is in somebody's page tables. */
 230         if (page_mapped(page))
 231                 return 1;
 232
 233         /* Be more reluctant to reclaim swapcache than pagecache */
 234         if (PageSwapCache(page))
 235                 return 1;
 236
 237         mapping = page_mapping(page);
 238         if (!mapping)
 239                 return 0;
 240
 241         /* File is mmap'd by somebody? */
 242         return mapping_mapped(mapping);
 243 }
 244
 245 static inline int is_page_cache_freeable(struct page *page)
 246 {
 247         return page_count(page) - !!PagePrivate(page) == 2;
 248 }
 249
 250 static int may_write_to_queue(struct backing_dev_info *bdi)
 251 {
 252         if (current_is_kswapd())
 253                 return 1;
 254         if (current_is_pdflush())       /* This is unlikely, but why not... */
 255                 return 1;
 256         if (!bdi_write_congested(bdi))
 257                 return 1;
 258         if (bdi == current->backing_dev_info)
 259                 return 1;
 260         return 0;
 261 }
 262
 263 /*
 264  * We detected a synchronous write error writing a page out.  Probably
 265  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 266  * fsync(), msync() or close().
 267  *
 268  * The tricky part is that after writepage we cannot touch the mapping: nothing
 269  * prevents it from being freed up.  But we have a ref on the page and once
 270  * that page is locked, the mapping is pinned.
 271  *
 272  * We're allowed to run sleeping lock_page() here because we know the caller has
 273  * __GFP_FS.
 274  */
 275 static void handle_write_error(struct address_space *mapping,
 276                                 struct page *page, int error)
 277 {
 278         lock_page(page);
 279         if (page_mapping(page) == mapping) {
 280                 if (error == -ENOSPC)
 281                         set_bit(AS_ENOSPC, &mapping->flags);
 282                 else
 283                         set_bit(AS_EIO, &mapping->flags);
 284         }
 285         unlock_page(page);
 286 }
 287
 288 /*
 289  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 290  */
 291 static pageout_t pageout(struct page *page, struct address_space *mapping)
 292 {
 293         /*
 294          * If the page is dirty, only perform writeback if that write
 295          * will be non-blocking.  To prevent this allocation from being
 296          * stalled by pagecache activity.  But note that there may be
 297          * stalls if we need to run get_block().  We could test
 298          * PagePrivate for that.
 299          *
 300          * If this process is currently in generic_file_write() against
 301          * this page's queue, we can perform writeback even if that
 302          * will block.
 303          *
 304          * If the page is swapcache, write it back even if that would
 305          * block, for some throttling. This happens by accident, because
 306          * swap_backing_dev_info is bust: it doesn't reflect the
 307          * congestion state of the swapdevs.  Easy to fix, if needed.
 308          * See swapfile.c:page_queue_congested().
 309          */
 310         if (!is_page_cache_freeable(page))
 311                 return PAGE_KEEP;
 312         if (!mapping)
 313                 return PAGE_KEEP;
 314         if (mapping->a_ops->writepage == NULL)
 315                 return PAGE_ACTIVATE;
 316         if (!may_write_to_queue(mapping->backing_dev_info))
 317                 return PAGE_KEEP;
 318
 319         if (clear_page_dirty_for_io(page)) {
 320                 int res;
 321                 struct writeback_control wbc = {
 322                         .sync_mode = WB_SYNC_NONE,
 323                         .nr_to_write = SWAP_CLUSTER_MAX,
 324                         .nonblocking = 1,
 325                         .for_reclaim = 1,
 326                 };
 327
 328                 SetPageReclaim(page);
 329                 res = mapping->a_ops->writepage(page, &wbc);
 330                 if (res < 0)
 331                         handle_write_error(mapping, page, res);
 332                 if (res == WRITEPAGE_ACTIVATE) {
 333                         ClearPageReclaim(page);
 334                         return PAGE_ACTIVATE;
 335                 }
 336                 if (!PageWriteback(page)) {
 337                         /* synchronous write or broken a_ops? */
 338                         ClearPageReclaim(page);
 339                 }
 340
 341                 return PAGE_SUCCESS;
 342         }
 343
 344         return PAGE_CLEAN;
 345 }
 346
 347 /*
 348  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 349  */
 350 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 351 {
 352         LIST_HEAD(ret_pages);
 353         struct pagevec freed_pvec;
 354         int pgactivate = 0;
 355         int reclaimed = 0;
 356
 357         cond_resched();
 358
 359         pagevec_init(&freed_pvec, 1);
 360         while (!list_empty(page_list)) {
 361                 struct address_space *mapping;
 362                 struct page *page;
 363                 int may_enter_fs;
 364                 int referenced;
 365
 366                 cond_resched();
 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                 if (PageWriteback(page))
 377                         goto keep_locked;
 378
 379                 sc->nr_scanned++;
 380                 /* Double the slab pressure for mapped and swapcache pages */
 381                 if (page_mapped(page) || PageSwapCache(page))
 382                         sc->nr_scanned++;
 383
 384                 page_map_lock(page);
 385                 referenced = page_referenced(page);
 386                 if (referenced && page_mapping_inuse(page)) {
 387                         /* In active use or really unfreeable.  Activate it. */
 388                         page_map_unlock(page);
 389                         goto activate_locked;
 390                 }
 391
 392 #ifdef CONFIG_SWAP
 393                 /*
 394                  * Anonymous process memory has backing store?
 395                  * Try to allocate it some swap space here.
 396                  *
 397                  * XXX: implement swap clustering ?
 398                  */
 399                 if (PageAnon(page) && !PageSwapCache(page)) {
 400                         page_map_unlock(page);
 401                         if (!add_to_swap(page))
 402                                 goto activate_locked;
 403                         page_map_lock(page);
 404                 }
 405 #endif /* CONFIG_SWAP */
 406
 407                 mapping = page_mapping(page);
 408                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 409                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 410
 411                 /*
 412                  * The page is mapped into the page tables of one or more
 413                  * processes. Try to unmap it here.
 414                  */
 415                 if (page_mapped(page) && mapping) {
 416                         switch (try_to_unmap(page)) {
 417                         case SWAP_FAIL:
 418                                 page_map_unlock(page);
 419                                 goto activate_locked;
 420                         case SWAP_AGAIN:
 421                                 page_map_unlock(page);
 422                                 goto keep_locked;
 423                         case SWAP_SUCCESS:
 424                                 ; /* try to free the page below */
 425                         }
 426                 }
 427                 page_map_unlock(page);
 428
 429                 if (PageDirty(page)) {
 430                         if (referenced)
 431                                 goto keep_locked;
 432                         if (!may_enter_fs)
 433                                 goto keep_locked;
 434                         if (laptop_mode && !sc->may_writepage)
 435                                 goto keep_locked;
 436
 437                         /* Page is dirty, try to write it out here */
 438                         switch(pageout(page, mapping)) {
 439                         case PAGE_KEEP:
 440                                 goto keep_locked;
 441                         case PAGE_ACTIVATE:
 442                                 goto activate_locked;
 443                         case PAGE_SUCCESS:
 444                                 if (PageWriteback(page) || PageDirty(page))
 445                                         goto keep;
 446                                 /*
 447                                  * A synchronous write - probably a ramdisk.  Go
 448                                  * ahead and try to reclaim the page.
 449                                  */
 450                                 if (TestSetPageLocked(page))
 451                                         goto keep;
 452                                 if (PageDirty(page) || PageWriteback(page))
 453                                         goto keep_locked;
 454                                 mapping = page_mapping(page);
 455                         case PAGE_CLEAN:
 456                                 ; /* try to free the page below */
 457                         }
 458                 }
 459
 460                 /*
 461                  * If the page has buffers, try to free the buffer mappings
 462                  * associated with this page. If we succeed we try to free
 463                  * the page as well.
 464                  *
 465                  * We do this even if the page is PageDirty().
 466                  * try_to_release_page() does not perform I/O, but it is
 467                  * possible for a page to have PageDirty set, but it is actually
 468                  * clean (all its buffers are clean).  This happens if the
 469                  * buffers were written out directly, with submit_bh(). ext3
 470                  * will do this, as well as the blockdev mapping.
 471                  * try_to_release_page() will discover that cleanness and will
 472                  * drop the buffers and mark the page clean - it can be freed.
 473                  *
 474                  * Rarely, pages can have buffers and no ->mapping.  These are
 475                  * the pages which were not successfully invalidated in
 476                  * truncate_complete_page().  We try to drop those buffers here
 477                  * and if that worked, and the page is no longer mapped into
 478                  * process address space (page_count == 1) it can be freed.
 479                  * Otherwise, leave the page on the LRU so it is swappable.
 480                  */
 481                 if (PagePrivate(page)) {
 482                         if (!try_to_release_page(page, sc->gfp_mask))
 483                                 goto activate_locked;
 484                         if (!mapping && page_count(page) == 1)
 485                                 goto free_it;
 486                 }
 487
 488                 if (!mapping)
 489                         goto keep_locked;       /* truncate got there first */
 490
 491                 spin_lock_irq(&mapping->tree_lock);
 492
 493                 /*
 494                  * The non-racy check for busy page.  It is critical to check
 495                  * PageDirty _after_ making sure that the page is freeable and
 496                  * not in use by anybody.       (pagecache + us == 2)
 497                  */
 498                 if (page_count(page) != 2 || PageDirty(page)) {
 499                         spin_unlock_irq(&mapping->tree_lock);
 500                         goto keep_locked;
 501                 }
 502
 503 #ifdef CONFIG_SWAP
 504                 if (PageSwapCache(page)) {
 505                         swp_entry_t swap = { .val = page->private };
 506                         __delete_from_swap_cache(page);
 507                         spin_unlock_irq(&mapping->tree_lock);
 508                         swap_free(swap);
 509                         __put_page(page);       /* The pagecache ref */
 510                         goto free_it;
 511                 }
 512 #endif /* CONFIG_SWAP */
 513
 514                 __remove_from_page_cache(page);
 515                 spin_unlock_irq(&mapping->tree_lock);
 516                 __put_page(page);
 517
 518 free_it:
 519                 unlock_page(page);
 520                 reclaimed++;
 521                 if (!pagevec_add(&freed_pvec, page))
 522                         __pagevec_release_nonlru(&freed_pvec);
 523                 continue;
 524
 525 activate_locked:
 526                 SetPageActive(page);
 527                 pgactivate++;
 528 keep_locked:
 529                 unlock_page(page);
 530 keep:
 531                 list_add(&page->lru, &ret_pages);
 532                 BUG_ON(PageLRU(page));
 533         }
 534         list_splice(&ret_pages, page_list);
 535         if (pagevec_count(&freed_pvec))
 536                 __pagevec_release_nonlru(&freed_pvec);
 537         mod_page_state(pgactivate, pgactivate);
 538         sc->nr_reclaimed += reclaimed;
 539         return reclaimed;
 540 }
 541
 542 /*
 543  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 544  * a batch of pages and working on them outside the lock.  Any pages which were
 545  * not freed will be added back to the LRU.
 546  *
 547  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 548  *
 549  * For pagecache intensive workloads, the first loop here is the hottest spot
 550  * in the kernel (apart from the copy_*_user functions).
 551  */
 552 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 553 {
 554         LIST_HEAD(page_list);
 555         struct pagevec pvec;
 556         int max_scan = sc->nr_to_scan, nr_pass;
 557         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 558
 559         pagevec_init(&pvec, 1);
 560
 561         lru_add_drain();
 562         spin_lock_irq(&zone->lru_lock);
 563 redo:
 564         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 565         nr_pass = zone->nr_inactive;
 566         while (max_scan > 0) {
 567                 struct page *page;
 568                 int nr_taken = 0;
 569                 int nr_scan = 0;
 570                 int nr_freed;
 571
 572                 while (nr_pass-- && nr_scan++ < SWAP_CLUSTER_MAX &&
 573                                 !list_empty(&zone->inactive_list)) {
 574                         page = lru_to_page(&zone->inactive_list);
 575
 576                         prefetchw_prev_lru_page(page,
 577                                                 &zone->inactive_list, flags);
 578
 579                         if (!TestClearPageLRU(page))
 580                                 BUG();
 581                         list_del(&page->lru);
 582                         if (get_page_testone(page)) {
 583                                 /*
 584                                  * It is being freed elsewhere
 585                                  */
 586                                 __put_page(page);
 587                                 SetPageLRU(page);
 588                                 list_add(&page->lru, &zone->inactive_list);
 589                                 continue;
 590                         } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 591                                 __put_page(page);
 592                                 SetPageLRU(page);
 593 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 594                                 list_add_tail(&page->lru, &zone->inactive_list);
 595 #else
 596                                 list_add(&page->lru, &zone->inactive_list);
 597 #endif
 598                                 continue;
 599                         }
 600                         list_add(&page->lru, &page_list);
 601                         ckrm_mem_dec_inactive(page);
 602                         nr_taken++;
 603                 }
 604                 zone->nr_inactive -= nr_taken;
 605                 zone->pages_scanned += nr_taken;
 606                 spin_unlock_irq(&zone->lru_lock);
 607
 608                 if ((bit_flag == 0) && (nr_taken == 0))
 609                         goto done;
 610
 611                 max_scan -= nr_scan;
 612                 if (current_is_kswapd())
 613                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 614                 else
 615                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 616                 nr_freed = shrink_list(&page_list, sc);
 617                 if (current_is_kswapd())
 618                         mod_page_state(kswapd_steal, nr_freed);
 619                 mod_page_state_zone(zone, pgsteal, nr_freed);
 620                 sc->nr_to_reclaim -= nr_freed;
 621
 622                 spin_lock_irq(&zone->lru_lock);
 623                 /*
 624                  * Put back any unfreeable pages.
 625                  */
 626                 while (!list_empty(&page_list)) {
 627                         page = lru_to_page(&page_list);
 628                         if (TestSetPageLRU(page))
 629                                 BUG();
 630                         list_del(&page->lru);
 631                         if (PageActive(page))
 632                                 add_page_to_active_list(zone, page);
 633                         else
 634                                 add_page_to_inactive_list(zone, page);
 635                         if (!pagevec_add(&pvec, page)) {
 636                                 spin_unlock_irq(&zone->lru_lock);
 637                                 __pagevec_release(&pvec);
 638                                 spin_lock_irq(&zone->lru_lock);
 639                         }
 640                 }
 641                 if (ckrm_flags && (nr_pass <= 0)) {
 642                         goto redo;
 643                 }
 644         }
 645         spin_unlock_irq(&zone->lru_lock);
 646 done:
 647         pagevec_release(&pvec);
 648 }
 649
 650 /*
 651  * This moves pages from the active list to the inactive list.
 652  *
 653  * We move them the other way if the page is referenced by one or more
 654  * processes, from rmap.
 655  *
 656  * If the pages are mostly unmapped, the processing is fast and it is
 657  * appropriate to hold zone->lru_lock across the whole operation.  But if
 658  * the pages are mapped, the processing is slow (page_referenced()) so we
 659  * should drop zone->lru_lock around each page.  It's impossible to balance
 660  * this, so instead we remove the pages from the LRU while processing them.
 661  * It is safe to rely on PG_active against the non-LRU pages in here because
 662  * nobody will play with that bit on a non-LRU page.
 663  *
 664  * The downside is that we have to touch page->_count against each page.
 665  * But we had to alter page->flags anyway.
 666  */
 667 static void
 668 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 669 {
 670         int pgmoved;
 671         int pgdeactivate = 0;
 672         int pgscanned = 0;
 673         int nr_pages = sc->nr_to_scan;
 674         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 675         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 676         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 677         struct page *page;
 678         struct pagevec pvec;
 679         int reclaim_mapped = 0;
 680         long mapped_ratio;
 681         long distress;
 682         long swap_tendency;
 683         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 684         int nr_pass;
 685
 686         lru_add_drain();
 687         pgmoved = 0;
 688         spin_lock_irq(&zone->lru_lock);
 689 redo:
 690         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 691         nr_pass = zone->nr_active;
 692         while (pgscanned < nr_pages && !list_empty(&zone->active_list) &&
 693                                                 nr_pass) {
 694                 page = lru_to_page(&zone->active_list);
 695                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 696                 if (!TestClearPageLRU(page))
 697                         BUG();
 698                 list_del(&page->lru);
 699                 if (get_page_testone(page)) {
 700                         /*
 701                          * It was already free!  release_pages() or put_page()
 702                          * are about to remove it from the LRU and free it. So
 703                          * put the refcount back and put the page back on the
 704                          * LRU
 705                          */
 706                         __put_page(page);
 707                         SetPageLRU(page);
 708                         list_add(&page->lru, &zone->active_list);
 709                         pgscanned++;
 710                 } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 711                         __put_page(page);
 712                         SetPageLRU(page);
 713 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 714                         list_add_tail(&page->lru, &zone->active_list);
 715 #else
 716                         list_add(&page->lru, &zone->active_list);
 717 #endif
 718                 } else {
 719                         list_add(&page->lru, &l_hold);
 720                         ckrm_mem_dec_active(page);
 721                         pgmoved++;
 722                 pgscanned++;
 723         }
 724                 if (!--nr_pass && ckrm_flags) {
 725                         goto redo;
 726                 }
 727         }
 728         zone->nr_active -= pgmoved;
 729         spin_unlock_irq(&zone->lru_lock);
 730
 731         /*
 732          * `distress' is a measure of how much trouble we're having reclaiming
 733          * pages.  0 -> no problems.  100 -> great trouble.
 734          */
 735         distress = 100 >> zone->prev_priority;
 736
 737         /*
 738          * The point of this algorithm is to decide when to start reclaiming
 739          * mapped memory instead of just pagecache.  Work out how much memory
 740          * is mapped.
 741          */
 742         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 743
 744         /*
 745          * Now decide how much we really want to unmap some pages.  The mapped
 746          * ratio is downgraded - just because there's a lot of mapped memory
 747          * doesn't necessarily mean that page reclaim isn't succeeding.
 748          *
 749          * The distress ratio is important - we don't want to start going oom.
 750          *
 751          * A 100% value of vm_swappiness overrides this algorithm altogether.
 752          */
 753         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 754
 755         /*
 756          * Now use this metric to decide whether to start moving mapped memory
 757          * onto the inactive list.
 758          */
 759         if (swap_tendency >= 100)
 760                 reclaim_mapped = 1;
 761
 762         while (!list_empty(&l_hold)) {
 763                 cond_resched();
 764                 page = lru_to_page(&l_hold);
 765                 list_del(&page->lru);
 766                 if (page_mapped(page)) {
 767                         if (!reclaim_mapped) {
 768                                 list_add(&page->lru, &l_active);
 769                                 continue;
 770                         }
 771                         page_map_lock(page);
 772                         if (page_referenced(page)) {
 773                                 page_map_unlock(page);
 774                                 list_add(&page->lru, &l_active);
 775                                 continue;
 776                         }
 777                         page_map_unlock(page);
 778                 }
 779                 /*
 780                  * FIXME: need to consider page_count(page) here if/when we
 781                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 782                  */
 783                 if (total_swap_pages == 0 && PageAnon(page)) {
 784                         list_add(&page->lru, &l_active);
 785                         continue;
 786                 }
 787                 list_add(&page->lru, &l_inactive);
 788         }
 789
 790         pagevec_init(&pvec, 1);
 791         pgmoved = 0;
 792         spin_lock_irq(&zone->lru_lock);
 793         while (!list_empty(&l_inactive)) {
 794                 page = lru_to_page(&l_inactive);
 795                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 796                 if (TestSetPageLRU(page))
 797                         BUG();
 798                 if (!TestClearPageActive(page))
 799                         BUG();
 800                 list_move(&page->lru, &zone->inactive_list);
 801                 ckrm_mem_inc_inactive(page);
 802                 pgmoved++;
 803                 if (!pagevec_add(&pvec, page)) {
 804                         zone->nr_inactive += pgmoved;
 805                         spin_unlock_irq(&zone->lru_lock);
 806                         pgdeactivate += pgmoved;
 807                         pgmoved = 0;
 808                         if (buffer_heads_over_limit)
 809                                 pagevec_strip(&pvec);
 810                         __pagevec_release(&pvec);
 811                         spin_lock_irq(&zone->lru_lock);
 812                 }
 813         }
 814         zone->nr_inactive += pgmoved;
 815         pgdeactivate += pgmoved;
 816         if (buffer_heads_over_limit) {
 817                 spin_unlock_irq(&zone->lru_lock);
 818                 pagevec_strip(&pvec);
 819                 spin_lock_irq(&zone->lru_lock);
 820         }
 821
 822         pgmoved = 0;
 823         while (!list_empty(&l_active)) {
 824                 page = lru_to_page(&l_active);
 825                 prefetchw_prev_lru_page(page, &l_active, flags);
 826                 if (TestSetPageLRU(page))
 827                         BUG();
 828                 BUG_ON(!PageActive(page));
 829                 list_move(&page->lru, &zone->active_list);
 830                 ckrm_mem_inc_active(page);
 831                 pgmoved++;
 832                 if (!pagevec_add(&pvec, page)) {
 833                         zone->nr_active += pgmoved;
 834                         pgmoved = 0;
 835                         spin_unlock_irq(&zone->lru_lock);
 836                         __pagevec_release(&pvec);
 837                         spin_lock_irq(&zone->lru_lock);
 838                 }
 839         }
 840         zone->nr_active += pgmoved;
 841         spin_unlock_irq(&zone->lru_lock);
 842         pagevec_release(&pvec);
 843
 844         mod_page_state_zone(zone, pgrefill, pgscanned);
 845         mod_page_state(pgdeactivate, pgdeactivate);
 846 }
 847
 848 /*
 849  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 850  */
 851 static void
 852 shrink_zone(struct zone *zone, struct scan_control *sc)
 853 {
 854         unsigned long nr_active;
 855         unsigned long nr_inactive;
 856
 857         /*
 858          * Add one to `nr_to_scan' just to make sure that the kernel will
 859          * slowly sift through the active list.
 860          */
 861         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 862         nr_active = zone->nr_scan_active;
 863         if (nr_active >= SWAP_CLUSTER_MAX)
 864                 zone->nr_scan_active = 0;
 865         else
 866                 nr_active = 0;
 867
 868         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 869         nr_inactive = zone->nr_scan_inactive;
 870         if (nr_inactive >= SWAP_CLUSTER_MAX)
 871                 zone->nr_scan_inactive = 0;
 872         else
 873                 nr_inactive = 0;
 874
 875         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 876
 877         while (nr_active || nr_inactive) {
 878                 sc->ckrm_flags = ckrm_setup_reclamation();
 879                 if (nr_active) {
 880                         sc->nr_to_scan = min(nr_active,
 881                                         (unsigned long)SWAP_CLUSTER_MAX);
 882                         nr_active -= sc->nr_to_scan;
 883                         refill_inactive_zone(zone, sc);
 884                 }
 885
 886                 if (nr_inactive) {
 887                         sc->nr_to_scan = min(nr_inactive,
 888                                         (unsigned long)SWAP_CLUSTER_MAX);
 889                         nr_inactive -= sc->nr_to_scan;
 890                         shrink_cache(zone, sc);
 891                         if (sc->nr_to_reclaim <= 0)
 892                                 break;
 893                 }
 894                 ckrm_teardown_reclamation();
 895         }
 896 }
 897
 898 #ifdef CONFIG_CKRM_RES_MEM
 899 // This function needs to be given more thought.
 900 // Shrink the class to be at 90% of its limit
 901 static void
 902 ckrm_shrink_class(ckrm_mem_res_t *cls)
 903 {
 904         struct scan_control sc;
 905         struct zone *zone;
 906         int zindex = 0, active_credit = 0, inactive_credit = 0;
 907
 908         if (ckrm_test_set_shrink(cls)) { // set the SHRINK bit atomically
 909                 // if it is already set somebody is working on it. so... leave
 910                 return;
 911         }
 912         sc.nr_mapped = read_page_state(nr_mapped);
 913         sc.nr_scanned = 0;
 914         sc.ckrm_flags = ckrm_get_reclaim_flags(cls);
 915         sc.nr_reclaimed = 0;
 916         sc.priority = 0; // always very high priority
 917
 918         for_each_zone(zone) {
 919                 int zone_total, zone_limit, active_limit, inactive_limit;
 920                 int active_over, inactive_over;
 921                 unsigned long nr_active, nr_inactive;
 922                 u64 temp;
 923
 924                 zone->temp_priority = zone->prev_priority;
 925                 zone->prev_priority = sc.priority;
 926
 927                 zone_total = zone->nr_active + zone->nr_inactive + zone->free_pages;
 928
 929                 temp = (u64) cls->pg_limit * zone_total;
 930                 do_div(temp, ckrm_tot_lru_pages);
 931                 zone_limit = (int) temp;
 932                 active_limit = (6 * zone_limit) / 10; // 2/3rd in active list
 933                 inactive_limit = (3 * zone_limit) / 10; // 1/3rd in inactive list
 934
 935                 active_over = cls->nr_active[zindex] - active_limit + active_credit;
 936                 inactive_over = active_over +
 937                                 (cls->nr_inactive[zindex] - inactive_limit) + inactive_credit;
 938
 939                 if (active_over > 0) {
 940                         zone->nr_scan_active += active_over + 1;
 941                         nr_active = zone->nr_scan_active;
 942                         active_credit = 0;
 943                 } else {
 944                         active_credit += active_over;
 945                         nr_active = 0;
 946                 }
 947
 948                 if (inactive_over > 0) {
 949                         zone->nr_scan_inactive += inactive_over;
 950                         nr_inactive = zone->nr_scan_inactive;
 951                         inactive_credit = 0;
 952                 } else {
 953                         inactive_credit += inactive_over;
 954                         nr_inactive = 0;
 955                 }
 956                 while (nr_active || nr_inactive) {
 957                         if (nr_active) {
 958                                 sc.nr_to_scan = min(nr_active,
 959                                                 (unsigned long)SWAP_CLUSTER_MAX);
 960                                 nr_active -= sc.nr_to_scan;
 961                                 refill_inactive_zone(zone, &sc);
 962                         }
 963
 964                         if (nr_inactive) {
 965                                 sc.nr_to_scan = min(nr_inactive,
 966                                                 (unsigned long)SWAP_CLUSTER_MAX);
 967                                 nr_inactive -= sc.nr_to_scan;
 968                                 shrink_cache(zone, &sc);
 969                                 if (sc.nr_to_reclaim <= 0)
 970                                         break;
 971                         }
 972                 }
 973                 zone->prev_priority = zone->temp_priority;
 974                 zindex++;
 975         }
 976         ckrm_clear_shrink(cls);
 977 }
 978
 979 static void
 980 ckrm_shrink_classes(void)
 981 {
 982         ckrm_mem_res_t *cls;
 983
 984         spin_lock(&ckrm_mem_lock);
 985         while (!ckrm_shrink_list_empty()) {
 986                 cls =  list_entry(ckrm_shrink_list.next, ckrm_mem_res_t,
 987                                 shrink_list);
 988                 spin_unlock(&ckrm_mem_lock);
 989                 ckrm_shrink_class(cls);
 990                 spin_lock(&ckrm_mem_lock);
 991                 list_del(&cls->shrink_list);
 992                 cls->flags &= ~MEM_AT_LIMIT;
 993         }
 994         spin_unlock(&ckrm_mem_lock);
 995 }
 996
 997 #else
 998 #define ckrm_shrink_classes()   do { } while(0)
 999 #endif
1000
1001 /*
1002  * This is the direct reclaim path, for page-allocating processes.  We only
1003  * try to reclaim pages from zones which will satisfy the caller's allocation
1004  * request.
1005  *
1006  * We reclaim from a zone even if that zone is over pages_high.  Because:
1007  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1008  *    allocation or
1009  * b) The zones may be over pages_high but they must go *over* pages_high to
1010  *    satisfy the `incremental min' zone defense algorithm.
1011  *
1012  * Returns the number of reclaimed pages.
1013  *
1014  * If a zone is deemed to be full of pinned pages then just give it a light
1015  * scan then give up on it.
1016  */
1017 static void
1018 shrink_caches(struct zone **zones, struct scan_control *sc)
1019 {
1020         int i;
1021
1022         for (i = 0; zones[i] != NULL; i++) {
1023                 struct zone *zone = zones[i];
1024
1025                 zone->temp_priority = sc->priority;
1026                 if (zone->prev_priority > sc->priority)
1027                         zone->prev_priority = sc->priority;
1028
1029                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1030                         continue;       /* Let kswapd poll it */
1031
1032                 shrink_zone(zone, sc);
1033         }
1034 }
1035
1036 /*
1037  * This is the main entry point to direct page reclaim.
1038  *
1039  * If a full scan of the inactive list fails to free enough memory then we
1040  * are "out of memory" and something needs to be killed.
1041  *
1042  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1043  * high - the zone may be full of dirty or under-writeback pages, which this
1044  * caller can't do much about.  We kick pdflush and take explicit naps in the
1045  * hope that some of these pages can be written.  But if the allocating task
1046  * holds filesystem locks which prevent writeout this might not work, and the
1047  * allocation attempt will fail.
1048  */
1049 int try_to_free_pages(struct zone **zones,
1050                 unsigned int gfp_mask, unsigned int order)
1051 {
1052         int priority;
1053         int ret = 0;
1054         int total_scanned = 0, total_reclaimed = 0;
1055         struct reclaim_state *reclaim_state = current->reclaim_state;
1056         struct scan_control sc;
1057         unsigned long lru_pages = 0;
1058         int i;
1059
1060         sc.gfp_mask = gfp_mask;
1061         sc.may_writepage = 0;
1062
1063         inc_page_state(allocstall);
1064
1065         for (i = 0; zones[i] != NULL; i++) {
1066                 struct zone *zone = zones[i];
1067
1068                 zone->temp_priority = DEF_PRIORITY;
1069                 lru_pages += zone->nr_active + zone->nr_inactive;
1070         }
1071
1072         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1073                 sc.nr_mapped = read_page_state(nr_mapped);
1074                 sc.nr_scanned = 0;
1075                 sc.nr_reclaimed = 0;
1076                 sc.priority = priority;
1077                 shrink_caches(zones, &sc);
1078                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1079                 if (reclaim_state) {
1080                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1081                         reclaim_state->reclaimed_slab = 0;
1082                 }
1083                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1084                         ret = 1;
1085                         goto out;
1086                 }
1087                 total_scanned += sc.nr_scanned;
1088                 total_reclaimed += sc.nr_reclaimed;
1089
1090                 /*
1091                  * Try to write back as many pages as we just scanned.  This
1092                  * tends to cause slow streaming writers to write data to the
1093                  * disk smoothly, at the dirtying rate, which is nice.   But
1094                  * that's undesirable in laptop mode, where we *want* lumpy
1095                  * writeout.  So in laptop mode, write out the whole world.
1096                  */
1097                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1098                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1099                         sc.may_writepage = 1;
1100                 }
1101
1102                 /* Take a nap, wait for some writeback to complete */
1103                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1104                         blk_congestion_wait(WRITE, HZ/10);
1105         }
1106         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1107                 out_of_memory(gfp_mask);
1108 out:
1109         for (i = 0; zones[i] != 0; i++)
1110                 zones[i]->prev_priority = zones[i]->temp_priority;
1111         return ret;
1112 }
1113
1114 /*
1115  * For kswapd, balance_pgdat() will work across all this node's zones until
1116  * they are all at pages_high.
1117  *
1118  * If `nr_pages' is non-zero then it is the number of pages which are to be
1119  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1120  * special.
1121  *
1122  * Returns the number of pages which were actually freed.
1123  *
1124  * There is special handling here for zones which are full of pinned pages.
1125  * This can happen if the pages are all mlocked, or if they are all used by
1126  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1127  * What we do is to detect the case where all pages in the zone have been
1128  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1129  * dead and from now on, only perform a short scan.  Basically we're polling
1130  * the zone for when the problem goes away.
1131  *
1132  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1133  * zones which have free_pages > pages_high, but once a zone is found to have
1134  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1135  * of the number of free pages in the lower zones.  This interoperates with
1136  * the page allocator fallback scheme to ensure that aging of pages is balanced
1137  * across the zones.
1138  */
1139 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1140 {
1141         int to_free = nr_pages;
1142         int priority;
1143         int i;
1144         int total_scanned = 0, total_reclaimed = 0;
1145         struct reclaim_state *reclaim_state = current->reclaim_state;
1146         struct scan_control sc;
1147
1148         sc.gfp_mask = GFP_KERNEL;
1149         sc.may_writepage = 0;
1150         sc.nr_mapped = read_page_state(nr_mapped);
1151
1152         inc_page_state(pageoutrun);
1153
1154         for (i = 0; i < pgdat->nr_zones; i++) {
1155                 struct zone *zone = pgdat->node_zones + i;
1156
1157                 zone->temp_priority = DEF_PRIORITY;
1158         }
1159
1160         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1161                 int all_zones_ok = 1;
1162                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1163                 unsigned long lru_pages = 0;
1164
1165                 if (nr_pages == 0) {
1166                         /*
1167                          * Scan in the highmem->dma direction for the highest
1168                          * zone which needs scanning
1169                          */
1170                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1171                                 struct zone *zone = pgdat->node_zones + i;
1172
1173                                 if (zone->all_unreclaimable &&
1174                                                 priority != DEF_PRIORITY)
1175                                         continue;
1176
1177                                 if (zone->free_pages <= zone->pages_high) {
1178                                         end_zone = i;
1179                                         goto scan;
1180                                 }
1181                         }
1182                         goto out;
1183                 } else {
1184                         end_zone = pgdat->nr_zones - 1;
1185                 }
1186 scan:
1187                 for (i = 0; i <= end_zone; i++) {
1188                         struct zone *zone = pgdat->node_zones + i;
1189
1190                         lru_pages += zone->nr_active + zone->nr_inactive;
1191                 }
1192
1193                 /*
1194                  * Now scan the zone in the dma->highmem direction, stopping
1195                  * at the last zone which needs scanning.
1196                  *
1197                  * We do this because the page allocator works in the opposite
1198                  * direction.  This prevents the page allocator from allocating
1199                  * pages behind kswapd's direction of progress, which would
1200                  * cause too much scanning of the lower zones.
1201                  */
1202                 for (i = 0; i <= end_zone; i++) {
1203                         struct zone *zone = pgdat->node_zones + i;
1204
1205                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1206                                 continue;
1207
1208                         if (nr_pages == 0) {    /* Not software suspend */
1209                                 if (zone->free_pages <= zone->pages_high)
1210                                         all_zones_ok = 0;
1211                         }
1212                         zone->temp_priority = priority;
1213                         if (zone->prev_priority > priority)
1214                                 zone->prev_priority = priority;
1215                         sc.nr_scanned = 0;
1216                         sc.nr_reclaimed = 0;
1217                         sc.priority = priority;
1218                         shrink_zone(zone, &sc);
1219                         reclaim_state->reclaimed_slab = 0;
1220                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1221                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1222                         total_reclaimed += sc.nr_reclaimed;
1223                         if (zone->all_unreclaimable)
1224                                 continue;
1225                         if (zone->pages_scanned > zone->present_pages * 2)
1226                                 zone->all_unreclaimable = 1;
1227                         /*
1228                          * If we've done a decent amount of scanning and
1229                          * the reclaim ratio is low, start doing writepage
1230                          * even in laptop mode
1231                          */
1232                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1233                             total_scanned > total_reclaimed+total_reclaimed/2)
1234                                 sc.may_writepage = 1;
1235                 }
1236                 if (nr_pages && to_free > total_reclaimed)
1237                         continue;       /* swsusp: need to do more work */
1238                 if (all_zones_ok)
1239                         break;          /* kswapd: all done */
1240                 /*
1241                  * OK, kswapd is getting into trouble.  Take a nap, then take
1242                  * another pass across the zones.
1243                  */
1244                 if (total_scanned && priority < DEF_PRIORITY - 2)
1245                         blk_congestion_wait(WRITE, HZ/10);
1246         }
1247 out:
1248         for (i = 0; i < pgdat->nr_zones; i++) {
1249                 struct zone *zone = pgdat->node_zones + i;
1250
1251                 zone->prev_priority = zone->temp_priority;
1252         }
1253         return total_reclaimed;
1254 }
1255
1256 /*
1257  * The background pageout daemon, started as a kernel thread
1258  * from the init process.
1259  *
1260  * This basically trickles out pages so that we have _some_
1261  * free memory available even if there is no other activity
1262  * that frees anything up. This is needed for things like routing
1263  * etc, where we otherwise might have all activity going on in
1264  * asynchronous contexts that cannot page things out.
1265  *
1266  * If there are applications that are active memory-allocators
1267  * (most normal use), this basically shouldn't matter.
1268  */
1269 static int kswapd(void *p)
1270 {
1271         pg_data_t *pgdat = (pg_data_t*)p;
1272         struct task_struct *tsk = current;
1273         DEFINE_WAIT(wait);
1274         struct reclaim_state reclaim_state = {
1275                 .reclaimed_slab = 0,
1276         };
1277         cpumask_t cpumask;
1278
1279         daemonize("kswapd%d", pgdat->node_id);
1280         cpumask = node_to_cpumask(pgdat->node_id);
1281         if (!cpus_empty(cpumask))
1282                 set_cpus_allowed(tsk, cpumask);
1283         current->reclaim_state = &reclaim_state;
1284
1285         /*
1286          * Tell the memory management that we're a "memory allocator",
1287          * and that if we need more memory we should get access to it
1288          * regardless (see "__alloc_pages()"). "kswapd" should
1289          * never get caught in the normal page freeing logic.
1290          *
1291          * (Kswapd normally doesn't need memory anyway, but sometimes
1292          * you need a small amount of memory in order to be able to
1293          * page out something else, and this flag essentially protects
1294          * us from recursively trying to free more memory as we're
1295          * trying to free the first piece of memory in the first place).
1296          */
1297         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1298
1299         for ( ; ; ) {
1300                 if (current->flags & PF_FREEZE)
1301                         refrigerator(PF_FREEZE);
1302                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1303                 schedule();
1304                 finish_wait(&pgdat->kswapd_wait, &wait);
1305                 try_to_clip_inodes();
1306
1307                 if (!ckrm_shrink_list_empty())
1308                         ckrm_shrink_classes();
1309                 else
1310                 balance_pgdat(pgdat, 0);
1311         }
1312         return 0;
1313 }
1314
1315 /*
1316  * A zone is low on free memory, so wake its kswapd task to service it.
1317  */
1318 void wakeup_kswapd(struct zone *zone)
1319 {
1320         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1321                 return;
1322         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1323                 return;
1324         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1325 }
1326
1327 #ifdef CONFIG_PM
1328 /*
1329  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1330  * pages.
1331  */
1332 int shrink_all_memory(int nr_pages)
1333 {
1334         pg_data_t *pgdat;
1335         int nr_to_free = nr_pages;
1336         int ret = 0;
1337         struct reclaim_state reclaim_state = {
1338                 .reclaimed_slab = 0,
1339         };
1340
1341         current->reclaim_state = &reclaim_state;
1342         for_each_pgdat(pgdat) {
1343                 int freed;
1344                 freed = balance_pgdat(pgdat, nr_to_free);
1345                 ret += freed;
1346                 nr_to_free -= freed;
1347                 if (nr_to_free <= 0)
1348                         break;
1349         }
1350         current->reclaim_state = NULL;
1351         return ret;
1352 }
1353 #endif
1354
1355 #ifdef CONFIG_HOTPLUG_CPU
1356 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1357    not required for correctness.  So if the last cpu in a node goes
1358    away, we get changed to run anywhere: as the first one comes back,
1359    restore their cpu bindings. */
1360 static int __devinit cpu_callback(struct notifier_block *nfb,
1361                                   unsigned long action,
1362                                   void *hcpu)
1363 {
1364         pg_data_t *pgdat;
1365         cpumask_t mask;
1366
1367         if (action == CPU_ONLINE) {
1368                 for_each_pgdat(pgdat) {
1369                         mask = node_to_cpumask(pgdat->node_id);
1370                         if (any_online_cpu(mask) != NR_CPUS)
1371                                 /* One of our CPUs online: restore mask */
1372                                 set_cpus_allowed(pgdat->kswapd, mask);
1373                 }
1374         }
1375         return NOTIFY_OK;
1376 }
1377 #endif /* CONFIG_HOTPLUG_CPU */
1378
1379 static int __init kswapd_init(void)
1380 {
1381         pg_data_t *pgdat;
1382         swap_setup();
1383         for_each_pgdat(pgdat)
1384                 pgdat->kswapd
1385                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1386         total_memory = nr_free_pagecache_pages();
1387         hotcpu_notifier(cpu_callback, 0);
1388         return 0;
1389 }
1390
1391 module_init(kswapd_init)