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 #ifndef AT_LIMIT_SUPPORT
  43 #warning "ckrm_at_limit disabled due to problems with memory hog tests -- seting ckrm_shrink_list_empty to true"
  44 #undef ckrm_shrink_list_empty
  45 #define ckrm_shrink_list_empty()                (1)
  46 #endif
  47
  48 /* possible outcome of pageout() */
  49 typedef enum {
  50         /* failed to write page out, page is locked */
  51         PAGE_KEEP,
  52         /* move page to the active list, page is locked */
  53         PAGE_ACTIVATE,
  54         /* page has been sent to the disk successfully, page is unlocked */
  55         PAGE_SUCCESS,
  56         /* page is clean and locked */
  57         PAGE_CLEAN,
  58 } pageout_t;
  59
  60 struct scan_control {
  61         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  62         unsigned long nr_to_scan;
  63
  64         /* Incremented by the number of inactive pages that were scanned */
  65         unsigned long nr_scanned;
  66
  67         /* Incremented by the number of pages reclaimed */
  68         unsigned long nr_reclaimed;
  69
  70         unsigned long nr_mapped;        /* From page_state */
  71
  72         /* How many pages shrink_cache() should reclaim */
  73         int nr_to_reclaim;
  74
  75         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  76         unsigned int priority;
  77
  78         /* This context's GFP mask */
  79         unsigned int gfp_mask;
  80
  81         /* Flag used by CKRM */
  82         unsigned int ckrm_flags;
  83
  84         int may_writepage;
  85 };
  86
  87 /*
  88  * The list of shrinker callbacks used by to apply pressure to
  89  * ageable caches.
  90  */
  91 struct shrinker {
  92         shrinker_t              shrinker;
  93         struct list_head        list;
  94         int                     seeks;  /* seeks to recreate an obj */
  95         long                    nr;     /* objs pending delete */
  96 };
  97
  98
  99
 100 void try_to_clip_inodes(void);
 101
 102
 103 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
 104
 105 #ifdef ARCH_HAS_PREFETCH
 106 #define prefetch_prev_lru_page(_page, _base, _field)                    \
 107         do {                                                            \
 108                 if ((_page)->lru.prev != _base) {                       \
 109                         struct page *prev;                              \
 110                                                                         \
 111                         prev = lru_to_page(&(_page->lru));              \
 112                         prefetch(&prev->_field);                        \
 113                 }                                                       \
 114         } while (0)
 115 #else
 116 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 117 #endif
 118
 119 #ifdef ARCH_HAS_PREFETCHW
 120 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 121         do {                                                            \
 122                 if ((_page)->lru.prev != _base) {                       \
 123                         struct page *prev;                              \
 124                                                                         \
 125                         prev = lru_to_page(&(_page->lru));              \
 126                         prefetchw(&prev->_field);                       \
 127                 }                                                       \
 128         } while (0)
 129 #else
 130 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 131 #endif
 132
 133 /*
 134  * From 0 .. 100.  Higher means more swappy.
 135  */
 136 int vm_swappiness = 60;
 137 static long total_memory;
 138
 139 static LIST_HEAD(shrinker_list);
 140 static DECLARE_MUTEX(shrinker_sem);
 141
 142 /*
 143  * Add a shrinker callback to be called from the vm
 144  */
 145 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 146 {
 147         struct shrinker *shrinker;
 148
 149         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 150         if (shrinker) {
 151                 shrinker->shrinker = theshrinker;
 152                 shrinker->seeks = seeks;
 153                 shrinker->nr = 0;
 154                 down(&shrinker_sem);
 155                 list_add(&shrinker->list, &shrinker_list);
 156                 up(&shrinker_sem);
 157         }
 158         return shrinker;
 159 }
 160 EXPORT_SYMBOL(set_shrinker);
 161
 162 /*
 163  * Remove one
 164  */
 165 void remove_shrinker(struct shrinker *shrinker)
 166 {
 167         down(&shrinker_sem);
 168         list_del(&shrinker->list);
 169         up(&shrinker_sem);
 170         kfree(shrinker);
 171 }
 172 EXPORT_SYMBOL(remove_shrinker);
 173
 174 #define SHRINK_BATCH 128
 175 /*
 176  * Call the shrink functions to age shrinkable caches
 177  *
 178  * Here we assume it costs one seek to replace a lru page and that it also
 179  * takes a seek to recreate a cache object.  With this in mind we age equal
 180  * percentages of the lru and ageable caches.  This should balance the seeks
 181  * generated by these structures.
 182  *
 183  * If the vm encounted mapped pages on the LRU it increase the pressure on
 184  * slab to avoid swapping.
 185  *
 186  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 187  *
 188  * `lru_pages' represents the number of on-LRU pages in all the zones which
 189  * are eligible for the caller's allocation attempt.  It is used for balancing
 190  * slab reclaim versus page reclaim.
 191  */
 192 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 193                         unsigned long lru_pages)
 194 {
 195         struct shrinker *shrinker;
 196
 197         if (down_trylock(&shrinker_sem))
 198                 return 0;
 199
 200         list_for_each_entry(shrinker, &shrinker_list, list) {
 201                 unsigned long long delta;
 202
 203                 delta = (4 * scanned) / shrinker->seeks;
 204                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 205                 do_div(delta, lru_pages + 1);
 206                 shrinker->nr += delta;
 207                 if (shrinker->nr < 0)
 208                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 209
 210                 if (shrinker->nr <= SHRINK_BATCH)
 211                         continue;
 212                 while (shrinker->nr) {
 213                         long this_scan = shrinker->nr;
 214                         int shrink_ret;
 215
 216                         if (this_scan > 128)
 217                                 this_scan = 128;
 218                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 219                         mod_page_state(slabs_scanned, this_scan);
 220                         shrinker->nr -= this_scan;
 221                         if (shrink_ret == -1)
 222                                 break;
 223                         cond_resched();
 224                 }
 225         }
 226         up(&shrinker_sem);
 227         return 0;
 228 }
 229
 230 /* Must be called with page's rmap lock held. */
 231 static inline int page_mapping_inuse(struct page *page)
 232 {
 233         struct address_space *mapping;
 234
 235         /* Page is in somebody's page tables. */
 236         if (page_mapped(page))
 237                 return 1;
 238
 239         /* Be more reluctant to reclaim swapcache than pagecache */
 240         if (PageSwapCache(page))
 241                 return 1;
 242
 243         mapping = page_mapping(page);
 244         if (!mapping)
 245                 return 0;
 246
 247         /* File is mmap'd by somebody? */
 248         return mapping_mapped(mapping);
 249 }
 250
 251 static inline int is_page_cache_freeable(struct page *page)
 252 {
 253         return page_count(page) - !!PagePrivate(page) == 2;
 254 }
 255
 256 static int may_write_to_queue(struct backing_dev_info *bdi)
 257 {
 258         if (current_is_kswapd())
 259                 return 1;
 260         if (current_is_pdflush())       /* This is unlikely, but why not... */
 261                 return 1;
 262         if (!bdi_write_congested(bdi))
 263                 return 1;
 264         if (bdi == current->backing_dev_info)
 265                 return 1;
 266         return 0;
 267 }
 268
 269 /*
 270  * We detected a synchronous write error writing a page out.  Probably
 271  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 272  * fsync(), msync() or close().
 273  *
 274  * The tricky part is that after writepage we cannot touch the mapping: nothing
 275  * prevents it from being freed up.  But we have a ref on the page and once
 276  * that page is locked, the mapping is pinned.
 277  *
 278  * We're allowed to run sleeping lock_page() here because we know the caller has
 279  * __GFP_FS.
 280  */
 281 static void handle_write_error(struct address_space *mapping,
 282                                 struct page *page, int error)
 283 {
 284         lock_page(page);
 285         if (page_mapping(page) == mapping) {
 286                 if (error == -ENOSPC)
 287                         set_bit(AS_ENOSPC, &mapping->flags);
 288                 else
 289                         set_bit(AS_EIO, &mapping->flags);
 290         }
 291         unlock_page(page);
 292 }
 293
 294 /*
 295  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 296  */
 297 static pageout_t pageout(struct page *page, struct address_space *mapping)
 298 {
 299         /*
 300          * If the page is dirty, only perform writeback if that write
 301          * will be non-blocking.  To prevent this allocation from being
 302          * stalled by pagecache activity.  But note that there may be
 303          * stalls if we need to run get_block().  We could test
 304          * PagePrivate for that.
 305          *
 306          * If this process is currently in generic_file_write() against
 307          * this page's queue, we can perform writeback even if that
 308          * will block.
 309          *
 310          * If the page is swapcache, write it back even if that would
 311          * block, for some throttling. This happens by accident, because
 312          * swap_backing_dev_info is bust: it doesn't reflect the
 313          * congestion state of the swapdevs.  Easy to fix, if needed.
 314          * See swapfile.c:page_queue_congested().
 315          */
 316         if (!is_page_cache_freeable(page))
 317                 return PAGE_KEEP;
 318         if (!mapping)
 319                 return PAGE_KEEP;
 320         if (mapping->a_ops->writepage == NULL)
 321                 return PAGE_ACTIVATE;
 322         if (!may_write_to_queue(mapping->backing_dev_info))
 323                 return PAGE_KEEP;
 324
 325         if (clear_page_dirty_for_io(page)) {
 326                 int res;
 327                 struct writeback_control wbc = {
 328                         .sync_mode = WB_SYNC_NONE,
 329                         .nr_to_write = SWAP_CLUSTER_MAX,
 330                         .nonblocking = 1,
 331                         .for_reclaim = 1,
 332                 };
 333
 334                 SetPageReclaim(page);
 335                 res = mapping->a_ops->writepage(page, &wbc);
 336                 if (res < 0)
 337                         handle_write_error(mapping, page, res);
 338                 if (res == WRITEPAGE_ACTIVATE) {
 339                         ClearPageReclaim(page);
 340                         return PAGE_ACTIVATE;
 341                 }
 342                 if (!PageWriteback(page)) {
 343                         /* synchronous write or broken a_ops? */
 344                         ClearPageReclaim(page);
 345                 }
 346
 347                 return PAGE_SUCCESS;
 348         }
 349
 350         return PAGE_CLEAN;
 351 }
 352
 353 /*
 354  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 355  */
 356 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 357 {
 358         LIST_HEAD(ret_pages);
 359         struct pagevec freed_pvec;
 360         int pgactivate = 0;
 361         int reclaimed = 0;
 362
 363         cond_resched();
 364
 365         pagevec_init(&freed_pvec, 1);
 366         while (!list_empty(page_list)) {
 367                 struct address_space *mapping;
 368                 struct page *page;
 369                 int may_enter_fs;
 370                 int referenced;
 371
 372                 cond_resched();
 373
 374                 page = lru_to_page(page_list);
 375                 list_del(&page->lru);
 376
 377                 if (TestSetPageLocked(page))
 378                         goto keep;
 379
 380                 BUG_ON(PageActive(page));
 381
 382                 if (PageWriteback(page))
 383                         goto keep_locked;
 384
 385                 sc->nr_scanned++;
 386                 /* Double the slab pressure for mapped and swapcache pages */
 387                 if (page_mapped(page) || PageSwapCache(page))
 388                         sc->nr_scanned++;
 389
 390                 page_map_lock(page);
 391                 referenced = page_referenced(page);
 392                 if (referenced && page_mapping_inuse(page)) {
 393                         /* In active use or really unfreeable.  Activate it. */
 394                         page_map_unlock(page);
 395                         goto activate_locked;
 396                 }
 397
 398 #ifdef CONFIG_SWAP
 399                 /*
 400                  * Anonymous process memory has backing store?
 401                  * Try to allocate it some swap space here.
 402                  *
 403                  * XXX: implement swap clustering ?
 404                  */
 405                 if (PageAnon(page) && !PageSwapCache(page)) {
 406                         page_map_unlock(page);
 407                         if (!add_to_swap(page))
 408                                 goto activate_locked;
 409                         page_map_lock(page);
 410                 }
 411 #endif /* CONFIG_SWAP */
 412
 413                 mapping = page_mapping(page);
 414                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 415                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 416
 417                 /*
 418                  * The page is mapped into the page tables of one or more
 419                  * processes. Try to unmap it here.
 420                  */
 421                 if (page_mapped(page) && mapping) {
 422                         switch (try_to_unmap(page)) {
 423                         case SWAP_FAIL:
 424                                 page_map_unlock(page);
 425                                 goto activate_locked;
 426                         case SWAP_AGAIN:
 427                                 page_map_unlock(page);
 428                                 goto keep_locked;
 429                         case SWAP_SUCCESS:
 430                                 ; /* try to free the page below */
 431                         }
 432                 }
 433                 page_map_unlock(page);
 434
 435                 if (PageDirty(page)) {
 436                         if (referenced)
 437                                 goto keep_locked;
 438                         if (!may_enter_fs)
 439                                 goto keep_locked;
 440                         if (laptop_mode && !sc->may_writepage)
 441                                 goto keep_locked;
 442
 443                         /* Page is dirty, try to write it out here */
 444                         switch(pageout(page, mapping)) {
 445                         case PAGE_KEEP:
 446                                 goto keep_locked;
 447                         case PAGE_ACTIVATE:
 448                                 goto activate_locked;
 449                         case PAGE_SUCCESS:
 450                                 if (PageWriteback(page) || PageDirty(page))
 451                                         goto keep;
 452                                 /*
 453                                  * A synchronous write - probably a ramdisk.  Go
 454                                  * ahead and try to reclaim the page.
 455                                  */
 456                                 if (TestSetPageLocked(page))
 457                                         goto keep;
 458                                 if (PageDirty(page) || PageWriteback(page))
 459                                         goto keep_locked;
 460                                 mapping = page_mapping(page);
 461                         case PAGE_CLEAN:
 462                                 ; /* try to free the page below */
 463                         }
 464                 }
 465
 466                 /*
 467                  * If the page has buffers, try to free the buffer mappings
 468                  * associated with this page. If we succeed we try to free
 469                  * the page as well.
 470                  *
 471                  * We do this even if the page is PageDirty().
 472                  * try_to_release_page() does not perform I/O, but it is
 473                  * possible for a page to have PageDirty set, but it is actually
 474                  * clean (all its buffers are clean).  This happens if the
 475                  * buffers were written out directly, with submit_bh(). ext3
 476                  * will do this, as well as the blockdev mapping.
 477                  * try_to_release_page() will discover that cleanness and will
 478                  * drop the buffers and mark the page clean - it can be freed.
 479                  *
 480                  * Rarely, pages can have buffers and no ->mapping.  These are
 481                  * the pages which were not successfully invalidated in
 482                  * truncate_complete_page().  We try to drop those buffers here
 483                  * and if that worked, and the page is no longer mapped into
 484                  * process address space (page_count == 1) it can be freed.
 485                  * Otherwise, leave the page on the LRU so it is swappable.
 486                  */
 487                 if (PagePrivate(page)) {
 488                         if (!try_to_release_page(page, sc->gfp_mask))
 489                                 goto activate_locked;
 490                         if (!mapping && page_count(page) == 1)
 491                                 goto free_it;
 492                 }
 493
 494                 if (!mapping)
 495                         goto keep_locked;       /* truncate got there first */
 496
 497                 spin_lock_irq(&mapping->tree_lock);
 498
 499                 /*
 500                  * The non-racy check for busy page.  It is critical to check
 501                  * PageDirty _after_ making sure that the page is freeable and
 502                  * not in use by anybody.       (pagecache + us == 2)
 503                  */
 504                 if (page_count(page) != 2 || PageDirty(page)) {
 505                         spin_unlock_irq(&mapping->tree_lock);
 506                         goto keep_locked;
 507                 }
 508
 509 #ifdef CONFIG_SWAP
 510                 if (PageSwapCache(page)) {
 511                         swp_entry_t swap = { .val = page->private };
 512                         __delete_from_swap_cache(page);
 513                         spin_unlock_irq(&mapping->tree_lock);
 514                         swap_free(swap);
 515                         __put_page(page);       /* The pagecache ref */
 516                         goto free_it;
 517                 }
 518 #endif /* CONFIG_SWAP */
 519
 520                 __remove_from_page_cache(page);
 521                 spin_unlock_irq(&mapping->tree_lock);
 522                 __put_page(page);
 523
 524 free_it:
 525                 unlock_page(page);
 526                 reclaimed++;
 527                 if (!pagevec_add(&freed_pvec, page))
 528                         __pagevec_release_nonlru(&freed_pvec);
 529                 continue;
 530
 531 activate_locked:
 532                 SetPageActive(page);
 533                 pgactivate++;
 534 keep_locked:
 535                 unlock_page(page);
 536 keep:
 537                 list_add(&page->lru, &ret_pages);
 538                 BUG_ON(PageLRU(page));
 539         }
 540         list_splice(&ret_pages, page_list);
 541         if (pagevec_count(&freed_pvec))
 542                 __pagevec_release_nonlru(&freed_pvec);
 543         mod_page_state(pgactivate, pgactivate);
 544         sc->nr_reclaimed += reclaimed;
 545         return reclaimed;
 546 }
 547
 548 /*
 549  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 550  * a batch of pages and working on them outside the lock.  Any pages which were
 551  * not freed will be added back to the LRU.
 552  *
 553  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 554  *
 555  * For pagecache intensive workloads, the first loop here is the hottest spot
 556  * in the kernel (apart from the copy_*_user functions).
 557  */
 558 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 559 {
 560         LIST_HEAD(page_list);
 561         struct pagevec pvec;
 562         int max_scan = sc->nr_to_scan, nr_pass;
 563         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 564
 565         pagevec_init(&pvec, 1);
 566
 567         lru_add_drain();
 568         spin_lock_irq(&zone->lru_lock);
 569 redo:
 570         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 571         nr_pass = zone->nr_inactive;
 572         while (max_scan > 0) {
 573                 struct page *page;
 574                 int nr_taken = 0;
 575                 int nr_scan = 0;
 576                 int nr_freed;
 577
 578                 while (nr_pass-- && nr_scan++ < SWAP_CLUSTER_MAX &&
 579                                 !list_empty(&zone->inactive_list)) {
 580                         page = lru_to_page(&zone->inactive_list);
 581
 582                         prefetchw_prev_lru_page(page,
 583                                                 &zone->inactive_list, flags);
 584
 585                         if (!TestClearPageLRU(page))
 586                                 BUG();
 587                         list_del(&page->lru);
 588                         if (get_page_testone(page)) {
 589                                 /*
 590                                  * It is being freed elsewhere
 591                                  */
 592                                 __put_page(page);
 593                                 SetPageLRU(page);
 594                                 list_add(&page->lru, &zone->inactive_list);
 595                                 continue;
 596                         } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 597                                 __put_page(page);
 598                                 SetPageLRU(page);
 599 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 600                                 list_add_tail(&page->lru, &zone->inactive_list);
 601 #else
 602                                 list_add(&page->lru, &zone->inactive_list);
 603 #endif
 604                                 continue;
 605                         }
 606                         list_add(&page->lru, &page_list);
 607                         ckrm_mem_dec_inactive(page);
 608                         nr_taken++;
 609                 }
 610                 zone->nr_inactive -= nr_taken;
 611                 zone->pages_scanned += nr_taken;
 612                 spin_unlock_irq(&zone->lru_lock);
 613
 614                 if ((bit_flag == 0) && (nr_taken == 0))
 615                         goto done;
 616
 617                 max_scan -= nr_scan;
 618                 if (current_is_kswapd())
 619                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 620                 else
 621                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 622                 nr_freed = shrink_list(&page_list, sc);
 623                 if (current_is_kswapd())
 624                         mod_page_state(kswapd_steal, nr_freed);
 625                 mod_page_state_zone(zone, pgsteal, nr_freed);
 626                 sc->nr_to_reclaim -= nr_freed;
 627
 628                 spin_lock_irq(&zone->lru_lock);
 629                 /*
 630                  * Put back any unfreeable pages.
 631                  */
 632                 while (!list_empty(&page_list)) {
 633                         page = lru_to_page(&page_list);
 634                         if (TestSetPageLRU(page))
 635                                 BUG();
 636                         list_del(&page->lru);
 637                         if (PageActive(page))
 638                                 add_page_to_active_list(zone, page);
 639                         else
 640                                 add_page_to_inactive_list(zone, page);
 641                         if (!pagevec_add(&pvec, page)) {
 642                                 spin_unlock_irq(&zone->lru_lock);
 643                                 __pagevec_release(&pvec);
 644                                 spin_lock_irq(&zone->lru_lock);
 645                         }
 646                 }
 647                 if (ckrm_flags && (nr_pass <= 0)) {
 648                         goto redo;
 649                 }
 650         }
 651         spin_unlock_irq(&zone->lru_lock);
 652 done:
 653         pagevec_release(&pvec);
 654 }
 655
 656 /*
 657  * This moves pages from the active list to the inactive list.
 658  *
 659  * We move them the other way if the page is referenced by one or more
 660  * processes, from rmap.
 661  *
 662  * If the pages are mostly unmapped, the processing is fast and it is
 663  * appropriate to hold zone->lru_lock across the whole operation.  But if
 664  * the pages are mapped, the processing is slow (page_referenced()) so we
 665  * should drop zone->lru_lock around each page.  It's impossible to balance
 666  * this, so instead we remove the pages from the LRU while processing them.
 667  * It is safe to rely on PG_active against the non-LRU pages in here because
 668  * nobody will play with that bit on a non-LRU page.
 669  *
 670  * The downside is that we have to touch page->_count against each page.
 671  * But we had to alter page->flags anyway.
 672  */
 673 static void
 674 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 675 {
 676         int pgmoved;
 677         int pgdeactivate = 0;
 678         int pgscanned = 0;
 679         int nr_pages = sc->nr_to_scan;
 680         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 681         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 682         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 683         struct page *page;
 684         struct pagevec pvec;
 685         int reclaim_mapped = 0;
 686         long mapped_ratio;
 687         long distress;
 688         long swap_tendency;
 689         unsigned int ckrm_flags = sc->ckrm_flags, bit_flag;
 690         int nr_pass;
 691
 692         lru_add_drain();
 693         pgmoved = 0;
 694         spin_lock_irq(&zone->lru_lock);
 695 redo:
 696         ckrm_get_reclaim_bits(&ckrm_flags, &bit_flag);
 697         nr_pass = zone->nr_active;
 698         while (pgscanned < nr_pages && !list_empty(&zone->active_list) &&
 699                                                 nr_pass) {
 700                 page = lru_to_page(&zone->active_list);
 701                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
 702                 if (!TestClearPageLRU(page))
 703                         BUG();
 704                 list_del(&page->lru);
 705                 if (get_page_testone(page)) {
 706                         /*
 707                          * It was already free!  release_pages() or put_page()
 708                          * are about to remove it from the LRU and free it. So
 709                          * put the refcount back and put the page back on the
 710                          * LRU
 711                          */
 712                         __put_page(page);
 713                         SetPageLRU(page);
 714                         list_add(&page->lru, &zone->active_list);
 715                         pgscanned++;
 716                 } else if (bit_flag && !ckrm_kick_page(page, bit_flag)) {
 717                         __put_page(page);
 718                         SetPageLRU(page);
 719 #ifdef CONFIG_CKRM_MEM_LRUORDER_CHANGE
 720                         list_add_tail(&page->lru, &zone->active_list);
 721 #else
 722                         list_add(&page->lru, &zone->active_list);
 723 #endif
 724                 } else {
 725                         list_add(&page->lru, &l_hold);
 726                         ckrm_mem_dec_active(page);
 727                         pgmoved++;
 728                         pgscanned++;
 729                 }
 730                 if (!--nr_pass && ckrm_flags) {
 731                         goto redo;
 732                 }
 733         }
 734         zone->nr_active -= pgmoved;
 735         spin_unlock_irq(&zone->lru_lock);
 736
 737         /*
 738          * `distress' is a measure of how much trouble we're having reclaiming
 739          * pages.  0 -> no problems.  100 -> great trouble.
 740          */
 741         distress = 100 >> zone->prev_priority;
 742
 743         /*
 744          * The point of this algorithm is to decide when to start reclaiming
 745          * mapped memory instead of just pagecache.  Work out how much memory
 746          * is mapped.
 747          */
 748         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 749
 750         /*
 751          * Now decide how much we really want to unmap some pages.  The mapped
 752          * ratio is downgraded - just because there's a lot of mapped memory
 753          * doesn't necessarily mean that page reclaim isn't succeeding.
 754          *
 755          * The distress ratio is important - we don't want to start going oom.
 756          *
 757          * A 100% value of vm_swappiness overrides this algorithm altogether.
 758          */
 759         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 760
 761         /*
 762          * Now use this metric to decide whether to start moving mapped memory
 763          * onto the inactive list.
 764          */
 765         if (swap_tendency >= 100)
 766                 reclaim_mapped = 1;
 767
 768         while (!list_empty(&l_hold)) {
 769                 cond_resched();
 770                 page = lru_to_page(&l_hold);
 771                 list_del(&page->lru);
 772                 if (page_mapped(page)) {
 773                         if (!reclaim_mapped) {
 774                                 list_add(&page->lru, &l_active);
 775                                 continue;
 776                         }
 777                         page_map_lock(page);
 778                         if (page_referenced(page)) {
 779                                 page_map_unlock(page);
 780                                 list_add(&page->lru, &l_active);
 781                                 continue;
 782                         }
 783                         page_map_unlock(page);
 784                 }
 785                 /*
 786                  * FIXME: need to consider page_count(page) here if/when we
 787                  * reap orphaned pages via the LRU (Daniel's locking stuff)
 788                  */
 789                 if (total_swap_pages == 0 && PageAnon(page)) {
 790                         list_add(&page->lru, &l_active);
 791                         continue;
 792                 }
 793                 list_add(&page->lru, &l_inactive);
 794         }
 795
 796         pagevec_init(&pvec, 1);
 797         pgmoved = 0;
 798         spin_lock_irq(&zone->lru_lock);
 799         while (!list_empty(&l_inactive)) {
 800                 page = lru_to_page(&l_inactive);
 801                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 802                 if (TestSetPageLRU(page))
 803                         BUG();
 804                 if (!TestClearPageActive(page))
 805                         BUG();
 806                 list_move(&page->lru, &zone->inactive_list);
 807                 ckrm_mem_inc_inactive(page);
 808                 pgmoved++;
 809                 if (!pagevec_add(&pvec, page)) {
 810                         zone->nr_inactive += pgmoved;
 811                         spin_unlock_irq(&zone->lru_lock);
 812                         pgdeactivate += pgmoved;
 813                         pgmoved = 0;
 814                         if (buffer_heads_over_limit)
 815                                 pagevec_strip(&pvec);
 816                         __pagevec_release(&pvec);
 817                         spin_lock_irq(&zone->lru_lock);
 818                 }
 819         }
 820         zone->nr_inactive += pgmoved;
 821         pgdeactivate += pgmoved;
 822         if (buffer_heads_over_limit) {
 823                 spin_unlock_irq(&zone->lru_lock);
 824                 pagevec_strip(&pvec);
 825                 spin_lock_irq(&zone->lru_lock);
 826         }
 827
 828         pgmoved = 0;
 829         while (!list_empty(&l_active)) {
 830                 page = lru_to_page(&l_active);
 831                 prefetchw_prev_lru_page(page, &l_active, flags);
 832                 if (TestSetPageLRU(page))
 833                         BUG();
 834                 BUG_ON(!PageActive(page));
 835                 list_move(&page->lru, &zone->active_list);
 836                 ckrm_mem_inc_active(page);
 837                 pgmoved++;
 838                 if (!pagevec_add(&pvec, page)) {
 839                         zone->nr_active += pgmoved;
 840                         pgmoved = 0;
 841                         spin_unlock_irq(&zone->lru_lock);
 842                         __pagevec_release(&pvec);
 843                         spin_lock_irq(&zone->lru_lock);
 844                 }
 845         }
 846         zone->nr_active += pgmoved;
 847         spin_unlock_irq(&zone->lru_lock);
 848         pagevec_release(&pvec);
 849
 850         mod_page_state_zone(zone, pgrefill, pgscanned);
 851         mod_page_state(pgdeactivate, pgdeactivate);
 852 }
 853
 854 /*
 855  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 856  */
 857 static void
 858 shrink_zone(struct zone *zone, struct scan_control *sc)
 859 {
 860         unsigned long nr_active;
 861         unsigned long nr_inactive;
 862
 863         /*
 864          * Add one to `nr_to_scan' just to make sure that the kernel will
 865          * slowly sift through the active list.
 866          */
 867         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
 868         nr_active = zone->nr_scan_active;
 869         if (nr_active >= SWAP_CLUSTER_MAX)
 870                 zone->nr_scan_active = 0;
 871         else
 872                 nr_active = 0;
 873
 874         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
 875         nr_inactive = zone->nr_scan_inactive;
 876         if (nr_inactive >= SWAP_CLUSTER_MAX)
 877                 zone->nr_scan_inactive = 0;
 878         else
 879                 nr_inactive = 0;
 880
 881         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
 882
 883         while (nr_active || nr_inactive) {
 884                 sc->ckrm_flags = ckrm_setup_reclamation();
 885                 if (nr_active) {
 886                         sc->nr_to_scan = min(nr_active,
 887                                         (unsigned long)SWAP_CLUSTER_MAX);
 888                         nr_active -= sc->nr_to_scan;
 889                         refill_inactive_zone(zone, sc);
 890                 }
 891
 892                 if (nr_inactive) {
 893                         sc->nr_to_scan = min(nr_inactive,
 894                                         (unsigned long)SWAP_CLUSTER_MAX);
 895                         nr_inactive -= sc->nr_to_scan;
 896                         shrink_cache(zone, sc);
 897                         if (sc->nr_to_reclaim <= 0)
 898                                 break;
 899                 }
 900                 ckrm_teardown_reclamation();
 901         }
 902 }
 903
 904 #if defined(CONFIG_CKRM_RES_MEM) && defined(AT_LIMIT_SUPPORT)
 905 // This function needs to be given more thought.
 906 // Shrink the class to be at 90% of its limit
 907 static void
 908 ckrm_shrink_class(ckrm_mem_res_t *cls)
 909 {
 910         struct scan_control sc;
 911         struct zone *zone;
 912         int zindex = 0, active_credit = 0, inactive_credit = 0;
 913
 914         if (ckrm_test_set_shrink(cls)) { // set the SHRINK bit atomically
 915                 // if it is already set somebody is working on it. so... leave
 916                 return;
 917         }
 918         sc.nr_mapped = read_page_state(nr_mapped);
 919         sc.nr_scanned = 0;
 920         sc.ckrm_flags = ckrm_get_reclaim_flags(cls);
 921         sc.nr_reclaimed = 0;
 922         sc.priority = 0; // always very high priority
 923
 924         for_each_zone(zone) {
 925                 int zone_total, zone_limit, active_limit, inactive_limit;
 926                 int active_over, inactive_over;
 927                 unsigned long nr_active, nr_inactive;
 928                 u64 temp;
 929
 930                 zone->temp_priority = zone->prev_priority;
 931                 zone->prev_priority = sc.priority;
 932
 933                 zone_total = zone->nr_active + zone->nr_inactive + zone->free_pages;
 934
 935                 temp = (u64) cls->pg_limit * zone_total;
 936                 do_div(temp, ckrm_tot_lru_pages);
 937                 zone_limit = (int) temp;
 938                 active_limit = (6 * zone_limit) / 10; // 2/3rd in active list
 939                 inactive_limit = (3 * zone_limit) / 10; // 1/3rd in inactive list
 940
 941                 active_over = cls->nr_active[zindex] - active_limit + active_credit;
 942                 inactive_over = active_over +
 943                                 (cls->nr_inactive[zindex] - inactive_limit) + inactive_credit;
 944
 945                 if (active_over > 0) {
 946                         zone->nr_scan_active += active_over + 1;
 947                         nr_active = zone->nr_scan_active;
 948                         active_credit = 0;
 949                 } else {
 950                         active_credit += active_over;
 951                         nr_active = 0;
 952                 }
 953
 954                 if (inactive_over > 0) {
 955                         zone->nr_scan_inactive += inactive_over;
 956                         nr_inactive = zone->nr_scan_inactive;
 957                         inactive_credit = 0;
 958                 } else {
 959                         inactive_credit += inactive_over;
 960                         nr_inactive = 0;
 961                 }
 962                 while (nr_active || nr_inactive) {
 963                         if (nr_active) {
 964                                 sc.nr_to_scan = min(nr_active,
 965                                                 (unsigned long)SWAP_CLUSTER_MAX);
 966                                 nr_active -= sc.nr_to_scan;
 967                                 refill_inactive_zone(zone, &sc);
 968                         }
 969
 970                         if (nr_inactive) {
 971                                 sc.nr_to_scan = min(nr_inactive,
 972                                                 (unsigned long)SWAP_CLUSTER_MAX);
 973                                 nr_inactive -= sc.nr_to_scan;
 974                                 shrink_cache(zone, &sc);
 975                                 if (sc.nr_to_reclaim <= 0)
 976                                         break;
 977                         }
 978                 }
 979                 zone->prev_priority = zone->temp_priority;
 980                 zindex++;
 981         }
 982         ckrm_clear_shrink(cls);
 983 }
 984
 985 static void
 986 ckrm_shrink_classes(void)
 987 {
 988         ckrm_mem_res_t *cls;
 989
 990         spin_lock(&ckrm_mem_lock);
 991         while (!ckrm_shrink_list_empty()) {
 992                 cls =  list_entry(ckrm_shrink_list.next, ckrm_mem_res_t,
 993                                 shrink_list);
 994                 spin_unlock(&ckrm_mem_lock);
 995                 ckrm_shrink_class(cls);
 996                 spin_lock(&ckrm_mem_lock);
 997                 list_del(&cls->shrink_list);
 998                 cls->flags &= ~MEM_AT_LIMIT;
 999         }
1000         spin_unlock(&ckrm_mem_lock);
1001 }
1002
1003 #else
1004
1005 #if defined(CONFIG_CKRM_RES_MEM) && !defined(AT_LIMIT_SUPPORT)
1006 #warning "disabling ckrm_at_limit -- setting ckrm_shrink_classes to noop "
1007 #endif
1008
1009 #define ckrm_shrink_classes()   do { } while(0)
1010 #endif
1011
1012 /*
1013  * This is the direct reclaim path, for page-allocating processes.  We only
1014  * try to reclaim pages from zones which will satisfy the caller's allocation
1015  * request.
1016  *
1017  * We reclaim from a zone even if that zone is over pages_high.  Because:
1018  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1019  *    allocation or
1020  * b) The zones may be over pages_high but they must go *over* pages_high to
1021  *    satisfy the `incremental min' zone defense algorithm.
1022  *
1023  * Returns the number of reclaimed pages.
1024  *
1025  * If a zone is deemed to be full of pinned pages then just give it a light
1026  * scan then give up on it.
1027  */
1028 static void
1029 shrink_caches(struct zone **zones, struct scan_control *sc)
1030 {
1031         int i;
1032
1033         for (i = 0; zones[i] != NULL; i++) {
1034                 struct zone *zone = zones[i];
1035
1036                 zone->temp_priority = sc->priority;
1037                 if (zone->prev_priority > sc->priority)
1038                         zone->prev_priority = sc->priority;
1039
1040                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1041                         continue;       /* Let kswapd poll it */
1042
1043                 shrink_zone(zone, sc);
1044         }
1045 }
1046
1047 /*
1048  * This is the main entry point to direct page reclaim.
1049  *
1050  * If a full scan of the inactive list fails to free enough memory then we
1051  * are "out of memory" and something needs to be killed.
1052  *
1053  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1054  * high - the zone may be full of dirty or under-writeback pages, which this
1055  * caller can't do much about.  We kick pdflush and take explicit naps in the
1056  * hope that some of these pages can be written.  But if the allocating task
1057  * holds filesystem locks which prevent writeout this might not work, and the
1058  * allocation attempt will fail.
1059  */
1060 int try_to_free_pages(struct zone **zones,
1061                 unsigned int gfp_mask, unsigned int order)
1062 {
1063         int priority;
1064         int ret = 0;
1065         int total_scanned = 0, total_reclaimed = 0;
1066         struct reclaim_state *reclaim_state = current->reclaim_state;
1067         struct scan_control sc;
1068         unsigned long lru_pages = 0;
1069         int i;
1070
1071         sc.gfp_mask = gfp_mask;
1072         sc.may_writepage = 0;
1073
1074         inc_page_state(allocstall);
1075
1076         for (i = 0; zones[i] != NULL; i++) {
1077                 struct zone *zone = zones[i];
1078
1079                 zone->temp_priority = DEF_PRIORITY;
1080                 lru_pages += zone->nr_active + zone->nr_inactive;
1081         }
1082
1083         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1084                 sc.nr_mapped = read_page_state(nr_mapped);
1085                 sc.nr_scanned = 0;
1086                 sc.nr_reclaimed = 0;
1087                 sc.priority = priority;
1088                 shrink_caches(zones, &sc);
1089                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1090                 if (reclaim_state) {
1091                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1092                         reclaim_state->reclaimed_slab = 0;
1093                 }
1094                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1095                         ret = 1;
1096                         goto out;
1097                 }
1098                 total_scanned += sc.nr_scanned;
1099                 total_reclaimed += sc.nr_reclaimed;
1100
1101                 /*
1102                  * Try to write back as many pages as we just scanned.  This
1103                  * tends to cause slow streaming writers to write data to the
1104                  * disk smoothly, at the dirtying rate, which is nice.   But
1105                  * that's undesirable in laptop mode, where we *want* lumpy
1106                  * writeout.  So in laptop mode, write out the whole world.
1107                  */
1108                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1109                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1110                         sc.may_writepage = 1;
1111                 }
1112
1113                 /* Take a nap, wait for some writeback to complete */
1114                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1115                         blk_congestion_wait(WRITE, HZ/10);
1116         }
1117         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1118                 out_of_memory(gfp_mask);
1119 out:
1120         for (i = 0; zones[i] != 0; i++)
1121                 zones[i]->prev_priority = zones[i]->temp_priority;
1122         return ret;
1123 }
1124
1125 /*
1126  * For kswapd, balance_pgdat() will work across all this node's zones until
1127  * they are all at pages_high.
1128  *
1129  * If `nr_pages' is non-zero then it is the number of pages which are to be
1130  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1131  * special.
1132  *
1133  * Returns the number of pages which were actually freed.
1134  *
1135  * There is special handling here for zones which are full of pinned pages.
1136  * This can happen if the pages are all mlocked, or if they are all used by
1137  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1138  * What we do is to detect the case where all pages in the zone have been
1139  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1140  * dead and from now on, only perform a short scan.  Basically we're polling
1141  * the zone for when the problem goes away.
1142  *
1143  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1144  * zones which have free_pages > pages_high, but once a zone is found to have
1145  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1146  * of the number of free pages in the lower zones.  This interoperates with
1147  * the page allocator fallback scheme to ensure that aging of pages is balanced
1148  * across the zones.
1149  */
1150 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1151 {
1152         int to_free = nr_pages;
1153         int priority;
1154         int i;
1155         int total_scanned = 0, total_reclaimed = 0;
1156         struct reclaim_state *reclaim_state = current->reclaim_state;
1157         struct scan_control sc;
1158
1159         sc.gfp_mask = GFP_KERNEL;
1160         sc.may_writepage = 0;
1161         sc.nr_mapped = read_page_state(nr_mapped);
1162
1163         inc_page_state(pageoutrun);
1164
1165         for (i = 0; i < pgdat->nr_zones; i++) {
1166                 struct zone *zone = pgdat->node_zones + i;
1167
1168                 zone->temp_priority = DEF_PRIORITY;
1169         }
1170
1171         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1172                 int all_zones_ok = 1;
1173                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1174                 unsigned long lru_pages = 0;
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->all_unreclaimable &&
1185                                                 priority != DEF_PRIORITY)
1186                                         continue;
1187
1188                                 if (zone->free_pages <= zone->pages_high) {
1189                                         end_zone = i;
1190                                         goto scan;
1191                                 }
1192                         }
1193                         goto out;
1194                 } else {
1195                         end_zone = pgdat->nr_zones - 1;
1196                 }
1197 scan:
1198                 for (i = 0; i <= end_zone; i++) {
1199                         struct zone *zone = pgdat->node_zones + i;
1200
1201                         lru_pages += zone->nr_active + zone->nr_inactive;
1202                 }
1203
1204                 /*
1205                  * Now scan the zone in the dma->highmem direction, stopping
1206                  * at the last zone which needs scanning.
1207                  *
1208                  * We do this because the page allocator works in the opposite
1209                  * direction.  This prevents the page allocator from allocating
1210                  * pages behind kswapd's direction of progress, which would
1211                  * cause too much scanning of the lower zones.
1212                  */
1213                 for (i = 0; i <= end_zone; i++) {
1214                         struct zone *zone = pgdat->node_zones + i;
1215
1216                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1217                                 continue;
1218
1219                         if (nr_pages == 0) {    /* Not software suspend */
1220                                 if (zone->free_pages <= zone->pages_high)
1221                                         all_zones_ok = 0;
1222                         }
1223                         zone->temp_priority = priority;
1224                         if (zone->prev_priority > priority)
1225                                 zone->prev_priority = priority;
1226                         sc.nr_scanned = 0;
1227                         sc.nr_reclaimed = 0;
1228                         sc.priority = priority;
1229                         shrink_zone(zone, &sc);
1230                         reclaim_state->reclaimed_slab = 0;
1231                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1232                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1233                         total_reclaimed += sc.nr_reclaimed;
1234                         if (zone->all_unreclaimable)
1235                                 continue;
1236                         if (zone->pages_scanned > zone->present_pages * 2)
1237                                 zone->all_unreclaimable = 1;
1238                         /*
1239                          * If we've done a decent amount of scanning and
1240                          * the reclaim ratio is low, start doing writepage
1241                          * even in laptop mode
1242                          */
1243                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1244                             total_scanned > total_reclaimed+total_reclaimed/2)
1245                                 sc.may_writepage = 1;
1246                 }
1247                 if (nr_pages && to_free > total_reclaimed)
1248                         continue;       /* swsusp: need to do more work */
1249                 if (all_zones_ok)
1250                         break;          /* kswapd: all done */
1251                 /*
1252                  * OK, kswapd is getting into trouble.  Take a nap, then take
1253                  * another pass across the zones.
1254                  */
1255                 if (total_scanned && priority < DEF_PRIORITY - 2)
1256                         blk_congestion_wait(WRITE, HZ/10);
1257         }
1258 out:
1259         for (i = 0; i < pgdat->nr_zones; i++) {
1260                 struct zone *zone = pgdat->node_zones + i;
1261
1262                 zone->prev_priority = zone->temp_priority;
1263         }
1264         return total_reclaimed;
1265 }
1266
1267 /*
1268  * The background pageout daemon, started as a kernel thread
1269  * from the init process.
1270  *
1271  * This basically trickles out pages so that we have _some_
1272  * free memory available even if there is no other activity
1273  * that frees anything up. This is needed for things like routing
1274  * etc, where we otherwise might have all activity going on in
1275  * asynchronous contexts that cannot page things out.
1276  *
1277  * If there are applications that are active memory-allocators
1278  * (most normal use), this basically shouldn't matter.
1279  */
1280 static int kswapd(void *p)
1281 {
1282         pg_data_t *pgdat = (pg_data_t*)p;
1283         struct task_struct *tsk = current;
1284         DEFINE_WAIT(wait);
1285         struct reclaim_state reclaim_state = {
1286                 .reclaimed_slab = 0,
1287         };
1288         cpumask_t cpumask;
1289
1290         daemonize("kswapd%d", pgdat->node_id);
1291         cpumask = node_to_cpumask(pgdat->node_id);
1292         if (!cpus_empty(cpumask))
1293                 set_cpus_allowed(tsk, cpumask);
1294         current->reclaim_state = &reclaim_state;
1295
1296         /*
1297          * Tell the memory management that we're a "memory allocator",
1298          * and that if we need more memory we should get access to it
1299          * regardless (see "__alloc_pages()"). "kswapd" should
1300          * never get caught in the normal page freeing logic.
1301          *
1302          * (Kswapd normally doesn't need memory anyway, but sometimes
1303          * you need a small amount of memory in order to be able to
1304          * page out something else, and this flag essentially protects
1305          * us from recursively trying to free more memory as we're
1306          * trying to free the first piece of memory in the first place).
1307          */
1308         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1309
1310         for ( ; ; ) {
1311                 if (current->flags & PF_FREEZE)
1312                         refrigerator(PF_FREEZE);
1313                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1314                 schedule();
1315                 finish_wait(&pgdat->kswapd_wait, &wait);
1316                 try_to_clip_inodes();
1317
1318                 if (!ckrm_shrink_list_empty())
1319                         ckrm_shrink_classes();
1320                 else
1321                 balance_pgdat(pgdat, 0);
1322         }
1323         return 0;
1324 }
1325
1326 /*
1327  * A zone is low on free memory, so wake its kswapd task to service it.
1328  */
1329 void wakeup_kswapd(struct zone *zone)
1330 {
1331         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1332                 return;
1333         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1334                 return;
1335         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1336 }
1337
1338 #ifdef CONFIG_PM
1339 /*
1340  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1341  * pages.
1342  */
1343 int shrink_all_memory(int nr_pages)
1344 {
1345         pg_data_t *pgdat;
1346         int nr_to_free = nr_pages;
1347         int ret = 0;
1348         struct reclaim_state reclaim_state = {
1349                 .reclaimed_slab = 0,
1350         };
1351
1352         current->reclaim_state = &reclaim_state;
1353         for_each_pgdat(pgdat) {
1354                 int freed;
1355                 freed = balance_pgdat(pgdat, nr_to_free);
1356                 ret += freed;
1357                 nr_to_free -= freed;
1358                 if (nr_to_free <= 0)
1359                         break;
1360         }
1361         current->reclaim_state = NULL;
1362         return ret;
1363 }
1364 #endif
1365
1366 #ifdef CONFIG_HOTPLUG_CPU
1367 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1368    not required for correctness.  So if the last cpu in a node goes
1369    away, we get changed to run anywhere: as the first one comes back,
1370    restore their cpu bindings. */
1371 static int __devinit cpu_callback(struct notifier_block *nfb,
1372                                   unsigned long action,
1373                                   void *hcpu)
1374 {
1375         pg_data_t *pgdat;
1376         cpumask_t mask;
1377
1378         if (action == CPU_ONLINE) {
1379                 for_each_pgdat(pgdat) {
1380                         mask = node_to_cpumask(pgdat->node_id);
1381                         if (any_online_cpu(mask) != NR_CPUS)
1382                                 /* One of our CPUs online: restore mask */
1383                                 set_cpus_allowed(pgdat->kswapd, mask);
1384                 }
1385         }
1386         return NOTIFY_OK;
1387 }
1388 #endif /* CONFIG_HOTPLUG_CPU */
1389
1390 static int __init kswapd_init(void)
1391 {
1392         pg_data_t *pgdat;
1393         swap_setup();
1394         for_each_pgdat(pgdat)
1395                 pgdat->kswapd
1396                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1397         total_memory = nr_free_pagecache_pages();
1398         hotcpu_notifier(cpu_callback, 0);
1399         return 0;
1400 }
1401
1402 module_init(kswapd_init)