mm/vmscan.c

   1 /*
   2  *  linux/mm/vmscan.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *
   6  *  Swap reorganised 29.12.95, Stephen Tweedie.
   7  *  kswapd added: 7.1.96  sct
   8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
   9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  11  *  Multiqueue VM started 5.8.00, Rik van Riel.
  12  */
  13
  14 #include <linux/mm.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/swap.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/init.h>
  21 #include <linux/highmem.h>
  22 #include <linux/file.h>
  23 #include <linux/writeback.h>
  24 #include <linux/suspend.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
  27                                         buffer_heads_over_limit */
  28 #include <linux/mm_inline.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/rmap.h>
  32 #include <linux/topology.h>
  33 #include <linux/cpu.h>
  34 #include <linux/notifier.h>
  35 #include <linux/rwsem.h>
  36
  37 #include <asm/tlbflush.h>
  38 #include <asm/div64.h>
  39
  40 #include <linux/swapops.h>
  41 #include <linux/ckrm_mem.h>
  42 #include <linux/vs_cvirt.h>
  43
  44
  45 /* possible outcome of pageout() */
  46 typedef enum {
  47         /* failed to write page out, page is locked */
  48         PAGE_KEEP,
  49         /* move page to the active list, page is locked */
  50         PAGE_ACTIVATE,
  51         /* page has been sent to the disk successfully, page is unlocked */
  52         PAGE_SUCCESS,
  53         /* page is clean and locked */
  54         PAGE_CLEAN,
  55 } pageout_t;
  56
  57 struct scan_control {
  58         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
  59         unsigned long nr_to_scan;
  60
  61         /* Incremented by the number of inactive pages that were scanned */
  62         unsigned long nr_scanned;
  63
  64         /* Incremented by the number of pages reclaimed */
  65         unsigned long nr_reclaimed;
  66
  67         unsigned long nr_mapped;        /* From page_state */
  68
  69         /* How many pages shrink_cache() should reclaim */
  70         int nr_to_reclaim;
  71
  72         /* Ask shrink_caches, or shrink_zone to scan at this priority */
  73         unsigned int priority;
  74
  75         /* This context's GFP mask */
  76         unsigned int gfp_mask;
  77
  78         int may_writepage;
  79 };
  80
  81 /*
  82  * The list of shrinker callbacks used by to apply pressure to
  83  * ageable caches.
  84  */
  85 struct shrinker {
  86         shrinker_t              shrinker;
  87         struct list_head        list;
  88         int                     seeks;  /* seeks to recreate an obj */
  89         long                    nr;     /* objs pending delete */
  90 };
  91
  92 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  93
  94 #ifdef ARCH_HAS_PREFETCH
  95 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  96         do {                                                            \
  97                 if ((_page)->lru.prev != _base) {                       \
  98                         struct page *prev;                              \
  99                                                                         \
 100                         prev = lru_to_page(&(_page->lru));              \
 101                         prefetch(&prev->_field);                        \
 102                 }                                                       \
 103         } while (0)
 104 #else
 105 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 106 #endif
 107
 108 #ifdef ARCH_HAS_PREFETCHW
 109 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 110         do {                                                            \
 111                 if ((_page)->lru.prev != _base) {                       \
 112                         struct page *prev;                              \
 113                                                                         \
 114                         prev = lru_to_page(&(_page->lru));              \
 115                         prefetchw(&prev->_field);                       \
 116                 }                                                       \
 117         } while (0)
 118 #else
 119 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 120 #endif
 121
 122 /*
 123  * From 0 .. 100.  Higher means more swappy.
 124  */
 125 int vm_swappiness = 60;
 126 static long total_memory;
 127
 128 static LIST_HEAD(shrinker_list);
 129 static DECLARE_RWSEM(shrinker_rwsem);
 130
 131 /*
 132  * Add a shrinker callback to be called from the vm
 133  */
 134 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
 135 {
 136         struct shrinker *shrinker;
 137
 138         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
 139         if (shrinker) {
 140                 shrinker->shrinker = theshrinker;
 141                 shrinker->seeks = seeks;
 142                 shrinker->nr = 0;
 143                 down_write(&shrinker_rwsem);
 144                 list_add(&shrinker->list, &shrinker_list);
 145                 up_write(&shrinker_rwsem);
 146         }
 147         return shrinker;
 148 }
 149 EXPORT_SYMBOL(set_shrinker);
 150
 151 /*
 152  * Remove one
 153  */
 154 void remove_shrinker(struct shrinker *shrinker)
 155 {
 156         down_write(&shrinker_rwsem);
 157         list_del(&shrinker->list);
 158         up_write(&shrinker_rwsem);
 159         kfree(shrinker);
 160 }
 161 EXPORT_SYMBOL(remove_shrinker);
 162
 163 #define SHRINK_BATCH 128
 164 /*
 165  * Call the shrink functions to age shrinkable caches
 166  *
 167  * Here we assume it costs one seek to replace a lru page and that it also
 168  * takes a seek to recreate a cache object.  With this in mind we age equal
 169  * percentages of the lru and ageable caches.  This should balance the seeks
 170  * generated by these structures.
 171  *
 172  * If the vm encounted mapped pages on the LRU it increase the pressure on
 173  * slab to avoid swapping.
 174  *
 175  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 176  *
 177  * `lru_pages' represents the number of on-LRU pages in all the zones which
 178  * are eligible for the caller's allocation attempt.  It is used for balancing
 179  * slab reclaim versus page reclaim.
 180  */
 181 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
 182                         unsigned long lru_pages)
 183 {
 184         struct shrinker *shrinker;
 185
 186         if (scanned == 0)
 187                 scanned = SWAP_CLUSTER_MAX;
 188
 189         if (!down_read_trylock(&shrinker_rwsem))
 190                 return 0;
 191
 192         list_for_each_entry(shrinker, &shrinker_list, list) {
 193                 unsigned long long delta;
 194                 unsigned long total_scan;
 195
 196                 delta = (4 * scanned) / shrinker->seeks;
 197                 delta *= (*shrinker->shrinker)(0, gfp_mask);
 198                 do_div(delta, lru_pages + 1);
 199                 shrinker->nr += delta;
 200                 if (shrinker->nr < 0)
 201                         shrinker->nr = LONG_MAX;        /* It wrapped! */
 202
 203                 total_scan = shrinker->nr;
 204                 shrinker->nr = 0;
 205
 206                 while (total_scan >= SHRINK_BATCH) {
 207                         long this_scan = SHRINK_BATCH;
 208                         int shrink_ret;
 209
 210                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
 211                         if (shrink_ret == -1)
 212                                 break;
 213                         mod_page_state(slabs_scanned, this_scan);
 214                         total_scan -= this_scan;
 215
 216                         cond_resched();
 217                 }
 218
 219                 shrinker->nr += total_scan;
 220         }
 221         up_read(&shrinker_rwsem);
 222         return 0;
 223 }
 224
 225 /* Called without lock on whether page is mapped, so answer is unstable */
 226 static inline int page_mapping_inuse(struct page *page)
 227 {
 228         struct address_space *mapping;
 229
 230         /* Page is in somebody's page tables. */
 231         if (page_mapped(page))
 232                 return 1;
 233
 234         /* Be more reluctant to reclaim swapcache than pagecache */
 235         if (PageSwapCache(page))
 236                 return 1;
 237
 238         mapping = page_mapping(page);
 239         if (!mapping)
 240                 return 0;
 241
 242         /* File is mmap'd by somebody? */
 243         return mapping_mapped(mapping);
 244 }
 245
 246 static inline int is_page_cache_freeable(struct page *page)
 247 {
 248         return page_count(page) - !!PagePrivate(page) == 2;
 249 }
 250
 251 static int may_write_to_queue(struct backing_dev_info *bdi)
 252 {
 253         if (current_is_kswapd())
 254                 return 1;
 255         if (current_is_pdflush())       /* This is unlikely, but why not... */
 256                 return 1;
 257         if (!bdi_write_congested(bdi))
 258                 return 1;
 259         if (bdi == current->backing_dev_info)
 260                 return 1;
 261         return 0;
 262 }
 263
 264 /*
 265  * We detected a synchronous write error writing a page out.  Probably
 266  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 267  * fsync(), msync() or close().
 268  *
 269  * The tricky part is that after writepage we cannot touch the mapping: nothing
 270  * prevents it from being freed up.  But we have a ref on the page and once
 271  * that page is locked, the mapping is pinned.
 272  *
 273  * We're allowed to run sleeping lock_page() here because we know the caller has
 274  * __GFP_FS.
 275  */
 276 static void handle_write_error(struct address_space *mapping,
 277                                 struct page *page, int error)
 278 {
 279         lock_page(page);
 280         if (page_mapping(page) == mapping) {
 281                 if (error == -ENOSPC)
 282                         set_bit(AS_ENOSPC, &mapping->flags);
 283                 else
 284                         set_bit(AS_EIO, &mapping->flags);
 285         }
 286         unlock_page(page);
 287 }
 288
 289 /*
 290  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
 291  */
 292 static pageout_t pageout(struct page *page, struct address_space *mapping)
 293 {
 294         /*
 295          * If the page is dirty, only perform writeback if that write
 296          * will be non-blocking.  To prevent this allocation from being
 297          * stalled by pagecache activity.  But note that there may be
 298          * stalls if we need to run get_block().  We could test
 299          * PagePrivate for that.
 300          *
 301          * If this process is currently in generic_file_write() against
 302          * this page's queue, we can perform writeback even if that
 303          * will block.
 304          *
 305          * If the page is swapcache, write it back even if that would
 306          * block, for some throttling. This happens by accident, because
 307          * swap_backing_dev_info is bust: it doesn't reflect the
 308          * congestion state of the swapdevs.  Easy to fix, if needed.
 309          * See swapfile.c:page_queue_congested().
 310          */
 311         if (!is_page_cache_freeable(page))
 312                 return PAGE_KEEP;
 313         if (!mapping)
 314                 return PAGE_KEEP;
 315         if (mapping->a_ops->writepage == NULL)
 316                 return PAGE_ACTIVATE;
 317         if (!may_write_to_queue(mapping->backing_dev_info))
 318                 return PAGE_KEEP;
 319
 320         if (clear_page_dirty_for_io(page)) {
 321                 int res;
 322                 struct writeback_control wbc = {
 323                         .sync_mode = WB_SYNC_NONE,
 324                         .nr_to_write = SWAP_CLUSTER_MAX,
 325                         .nonblocking = 1,
 326                         .for_reclaim = 1,
 327                 };
 328
 329                 SetPageReclaim(page);
 330                 res = mapping->a_ops->writepage(page, &wbc);
 331                 if (res < 0)
 332                         handle_write_error(mapping, page, res);
 333                 if (res == WRITEPAGE_ACTIVATE) {
 334                         ClearPageReclaim(page);
 335                         return PAGE_ACTIVATE;
 336                 }
 337                 if (!PageWriteback(page)) {
 338                         /* synchronous write or broken a_ops? */
 339                         ClearPageReclaim(page);
 340                 }
 341
 342                 return PAGE_SUCCESS;
 343         }
 344
 345         return PAGE_CLEAN;
 346 }
 347
 348 /*
 349  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
 350  */
 351 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 352 {
 353         LIST_HEAD(ret_pages);
 354         struct pagevec freed_pvec;
 355         int pgactivate = 0;
 356         int reclaimed = 0;
 357
 358         cond_resched();
 359
 360         pagevec_init(&freed_pvec, 1);
 361         while (!list_empty(page_list)) {
 362                 struct address_space *mapping;
 363                 struct page *page;
 364                 int may_enter_fs;
 365                 int referenced;
 366
 367                 page = lru_to_page(page_list);
 368                 list_del(&page->lru);
 369
 370                 if (TestSetPageLocked(page))
 371                         goto keep;
 372
 373                 BUG_ON(PageActive(page));
 374
 375                 sc->nr_scanned++;
 376                 /* Double the slab pressure for mapped and swapcache pages */
 377                 if (page_mapped(page) || PageSwapCache(page))
 378                         sc->nr_scanned++;
 379
 380                 if (PageWriteback(page))
 381                         goto keep_locked;
 382
 383                 referenced = page_referenced(page, 1, sc->priority <= 0);
 384                 /* In active use or really unfreeable?  Activate it. */
 385                 if (referenced && page_mapping_inuse(page))
 386                         goto activate_locked;
 387
 388 #ifdef CONFIG_SWAP
 389                 /*
 390                  * Anonymous process memory has backing store?
 391                  * Try to allocate it some swap space here.
 392                  */
 393                 if (PageAnon(page) && !PageSwapCache(page)) {
 394                         if (!add_to_swap(page))
 395                                 goto activate_locked;
 396                 }
 397 #endif /* CONFIG_SWAP */
 398
 399                 mapping = page_mapping(page);
 400                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 401                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 402
 403                 /*
 404                  * The page is mapped into the page tables of one or more
 405                  * processes. Try to unmap it here.
 406                  */
 407                 if (page_mapped(page) && mapping) {
 408                         switch (try_to_unmap(page)) {
 409                         case SWAP_FAIL:
 410                                 goto activate_locked;
 411                         case SWAP_AGAIN:
 412                                 goto keep_locked;
 413                         case SWAP_SUCCESS:
 414                                 ; /* try to free the page below */
 415                         }
 416                 }
 417
 418                 if (PageDirty(page)) {
 419                         if (referenced)
 420                                 goto keep_locked;
 421                         if (!may_enter_fs)
 422                                 goto keep_locked;
 423                         if (laptop_mode && !sc->may_writepage)
 424                                 goto keep_locked;
 425
 426                         /* Page is dirty, try to write it out here */
 427                         switch(pageout(page, mapping)) {
 428                         case PAGE_KEEP:
 429                                 goto keep_locked;
 430                         case PAGE_ACTIVATE:
 431                                 goto activate_locked;
 432                         case PAGE_SUCCESS:
 433                                 if (PageWriteback(page) || PageDirty(page))
 434                                         goto keep;
 435                                 /*
 436                                  * A synchronous write - probably a ramdisk.  Go
 437                                  * ahead and try to reclaim the page.
 438                                  */
 439                                 if (TestSetPageLocked(page))
 440                                         goto keep;
 441                                 if (PageDirty(page) || PageWriteback(page))
 442                                         goto keep_locked;
 443                                 mapping = page_mapping(page);
 444                         case PAGE_CLEAN:
 445                                 ; /* try to free the page below */
 446                         }
 447                 }
 448
 449                 /*
 450                  * If the page has buffers, try to free the buffer mappings
 451                  * associated with this page. If we succeed we try to free
 452                  * the page as well.
 453                  *
 454                  * We do this even if the page is PageDirty().
 455                  * try_to_release_page() does not perform I/O, but it is
 456                  * possible for a page to have PageDirty set, but it is actually
 457                  * clean (all its buffers are clean).  This happens if the
 458                  * buffers were written out directly, with submit_bh(). ext3
 459                  * will do this, as well as the blockdev mapping.
 460                  * try_to_release_page() will discover that cleanness and will
 461                  * drop the buffers and mark the page clean - it can be freed.
 462                  *
 463                  * Rarely, pages can have buffers and no ->mapping.  These are
 464                  * the pages which were not successfully invalidated in
 465                  * truncate_complete_page().  We try to drop those buffers here
 466                  * and if that worked, and the page is no longer mapped into
 467                  * process address space (page_count == 1) it can be freed.
 468                  * Otherwise, leave the page on the LRU so it is swappable.
 469                  */
 470                 if (PagePrivate(page)) {
 471                         if (!try_to_release_page(page, sc->gfp_mask))
 472                                 goto activate_locked;
 473                         if (!mapping && page_count(page) == 1)
 474                                 goto free_it;
 475                 }
 476
 477                 if (!mapping)
 478                         goto keep_locked;       /* truncate got there first */
 479
 480                 spin_lock_irq(&mapping->tree_lock);
 481
 482                 /*
 483                  * The non-racy check for busy page.  It is critical to check
 484                  * PageDirty _after_ making sure that the page is freeable and
 485                  * not in use by anybody.       (pagecache + us == 2)
 486                  */
 487                 if (page_count(page) != 2 || PageDirty(page)) {
 488                         spin_unlock_irq(&mapping->tree_lock);
 489                         goto keep_locked;
 490                 }
 491
 492 #ifdef CONFIG_SWAP
 493                 if (PageSwapCache(page)) {
 494                         swp_entry_t swap = { .val = page->private };
 495                         __delete_from_swap_cache(page);
 496                         spin_unlock_irq(&mapping->tree_lock);
 497                         swap_free(swap);
 498                         __put_page(page);       /* The pagecache ref */
 499                         goto free_it;
 500                 }
 501 #endif /* CONFIG_SWAP */
 502
 503                 __remove_from_page_cache(page);
 504                 spin_unlock_irq(&mapping->tree_lock);
 505                 __put_page(page);
 506
 507 free_it:
 508                 unlock_page(page);
 509                 reclaimed++;
 510                 if (!pagevec_add(&freed_pvec, page))
 511                         __pagevec_release_nonlru(&freed_pvec);
 512                 continue;
 513
 514 activate_locked:
 515                 SetPageActive(page);
 516                 pgactivate++;
 517 keep_locked:
 518                 unlock_page(page);
 519 keep:
 520                 list_add(&page->lru, &ret_pages);
 521                 BUG_ON(PageLRU(page));
 522         }
 523         list_splice(&ret_pages, page_list);
 524         if (pagevec_count(&freed_pvec))
 525                 __pagevec_release_nonlru(&freed_pvec);
 526         mod_page_state(pgactivate, pgactivate);
 527         sc->nr_reclaimed += reclaimed;
 528         return reclaimed;
 529 }
 530
 531 /*
 532  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
 533  * a batch of pages and working on them outside the lock.  Any pages which were
 534  * not freed will be added back to the LRU.
 535  *
 536  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
 537  *
 538  * For pagecache intensive workloads, the first loop here is the hottest spot
 539  * in the kernel (apart from the copy_*_user functions).
 540  */
 541 #ifdef CONFIG_CKRM_RES_MEM
 542 static void shrink_cache(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 543 #else
 544 static void shrink_cache(struct zone *zone, struct scan_control *sc)
 545 #endif
 546 {
 547         LIST_HEAD(page_list);
 548         struct pagevec pvec;
 549         int max_scan = sc->nr_to_scan;
 550 #ifdef CONFIG_CKRM_RES_MEM
 551         struct zone *zone = ckrm_zone->zone;
 552         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 553         struct list_head *active_list = &ckrm_zone->active_list;
 554 #else
 555         struct list_head *inactive_list = &zone->inactive_list;
 556         struct list_head *active_list = &zone->active_list;
 557 #endif
 558
 559         pagevec_init(&pvec, 1);
 560
 561         lru_add_drain();
 562         spin_lock_irq(&zone->lru_lock);
 563         while (max_scan > 0) {
 564                 struct page *page;
 565                 int nr_taken = 0;
 566                 int nr_scan = 0;
 567                 int nr_freed;
 568
 569                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
 570                                 !list_empty(inactive_list)) {
 571                         page = lru_to_page(inactive_list);
 572
 573                         prefetchw_prev_lru_page(page,
 574                                                 inactive_list, flags);
 575
 576                         if (!TestClearPageLRU(page))
 577                                 BUG();
 578                         list_del(&page->lru);
 579                         if (get_page_testone(page)) {
 580                                 /*
 581                                  * It is being freed elsewhere
 582                                  */
 583                                 __put_page(page);
 584                                 SetPageLRU(page);
 585                                 list_add(&page->lru, inactive_list);
 586                                 continue;
 587                         }
 588                         list_add(&page->lru, &page_list);
 589                         nr_taken++;
 590                 }
 591                 zone->nr_inactive -= nr_taken;
 592                 ckrm_zone_dec_inactive(ckrm_zone, nr_taken);
 593                 spin_unlock_irq(&zone->lru_lock);
 594
 595                 if (nr_taken == 0)
 596                         goto done;
 597
 598                 max_scan -= nr_scan;
 599                 if (current_is_kswapd())
 600                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
 601                 else
 602                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
 603                 nr_freed = shrink_list(&page_list, sc);
 604                 if (current_is_kswapd())
 605                         mod_page_state(kswapd_steal, nr_freed);
 606                 mod_page_state_zone(zone, pgsteal, nr_freed);
 607                 sc->nr_to_reclaim -= nr_freed;
 608
 609                 spin_lock_irq(&zone->lru_lock);
 610                 /*
 611                  * Put back any unfreeable pages.
 612                  */
 613                 while (!list_empty(&page_list)) {
 614                         page = lru_to_page(&page_list);
 615                         if (TestSetPageLRU(page))
 616                                 BUG();
 617                         list_del(&page->lru);
 618                         if (PageActive(page)) {
 619                                 ckrm_zone_inc_active(ckrm_zone, 1);
 620                                 zone->nr_active++;
 621                                 list_add(&page->lru, active_list);
 622                         } else {
 623                                 ckrm_zone_inc_inactive(ckrm_zone, 1);
 624                                 zone->nr_inactive++;
 625                                 list_add(&page->lru, inactive_list);
 626                         }
 627                         if (!pagevec_add(&pvec, page)) {
 628                                 spin_unlock_irq(&zone->lru_lock);
 629                                 __pagevec_release(&pvec);
 630                                 spin_lock_irq(&zone->lru_lock);
 631                         }
 632                 }
 633         }
 634         spin_unlock_irq(&zone->lru_lock);
 635 done:
 636         pagevec_release(&pvec);
 637 }
 638
 639 /*
 640  * This moves pages from the active list to the inactive list.
 641  *
 642  * We move them the other way if the page is referenced by one or more
 643  * processes, from rmap.
 644  *
 645  * If the pages are mostly unmapped, the processing is fast and it is
 646  * appropriate to hold zone->lru_lock across the whole operation.  But if
 647  * the pages are mapped, the processing is slow (page_referenced()) so we
 648  * should drop zone->lru_lock around each page.  It's impossible to balance
 649  * this, so instead we remove the pages from the LRU while processing them.
 650  * It is safe to rely on PG_active against the non-LRU pages in here because
 651  * nobody will play with that bit on a non-LRU page.
 652  *
 653  * The downside is that we have to touch page->_count against each page.
 654  * But we had to alter page->flags anyway.
 655  */
 656 static void
 657 #ifdef CONFIG_CKRM_RES_MEM
 658 refill_inactive_zone(struct ckrm_zone *ckrm_zone, struct scan_control *sc)
 659 #else
 660 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 661 #endif
 662 {
 663         int pgmoved;
 664         int pgdeactivate = 0;
 665         int pgscanned = 0;
 666         int nr_pages = sc->nr_to_scan;
 667         LIST_HEAD(l_hold);      /* The pages which were snipped off */
 668         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
 669         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
 670         struct page *page;
 671         struct pagevec pvec;
 672         int reclaim_mapped = 0;
 673         long mapped_ratio;
 674         long distress;
 675         long swap_tendency;
 676 #ifdef CONFIG_CKRM_RES_MEM
 677         struct zone *zone = ckrm_zone->zone;
 678         struct list_head *active_list = &ckrm_zone->active_list;
 679         struct list_head *inactive_list = &ckrm_zone->inactive_list;
 680 #else
 681         struct list_head *active_list = &zone->active_list;
 682         struct list_head *inactive_list = &zone->inactive_list;
 683 #endif
 684
 685         lru_add_drain();
 686         pgmoved = 0;
 687         spin_lock_irq(&zone->lru_lock);
 688         while (pgscanned < nr_pages && !list_empty(active_list)) {
 689                 page = lru_to_page(active_list);
 690                 prefetchw_prev_lru_page(page, active_list, flags);
 691                 if (!TestClearPageLRU(page))
 692                         BUG();
 693                 list_del(&page->lru);
 694                 if (get_page_testone(page)) {
 695                         /*
 696                          * It was already free!  release_pages() or put_page()
 697                          * are about to remove it from the LRU and free it. So
 698                          * put the refcount back and put the page back on the
 699                          * LRU
 700                          */
 701                         __put_page(page);
 702                         SetPageLRU(page);
 703                         list_add(&page->lru, active_list);
 704                 } else {
 705                         list_add(&page->lru, &l_hold);
 706                         pgmoved++;
 707                 }
 708                 pgscanned++;
 709         }
 710         zone->pages_scanned += pgscanned;
 711         zone->nr_active -= pgmoved;
 712         ckrm_zone_dec_active(ckrm_zone, pgmoved);
 713         spin_unlock_irq(&zone->lru_lock);
 714
 715         /*
 716          * `distress' is a measure of how much trouble we're having reclaiming
 717          * pages.  0 -> no problems.  100 -> great trouble.
 718          */
 719         distress = 100 >> zone->prev_priority;
 720
 721         /*
 722          * The point of this algorithm is to decide when to start reclaiming
 723          * mapped memory instead of just pagecache.  Work out how much memory
 724          * is mapped.
 725          */
 726         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 727
 728         /*
 729          * Now decide how much we really want to unmap some pages.  The mapped
 730          * ratio is downgraded - just because there's a lot of mapped memory
 731          * doesn't necessarily mean that page reclaim isn't succeeding.
 732          *
 733          * The distress ratio is important - we don't want to start going oom.
 734          *
 735          * A 100% value of vm_swappiness overrides this algorithm altogether.
 736          */
 737         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
 738
 739         /*
 740          * Now use this metric to decide whether to start moving mapped memory
 741          * onto the inactive list.
 742          */
 743         if (swap_tendency >= 100)
 744                 reclaim_mapped = 1;
 745
 746         while (!list_empty(&l_hold)) {
 747                 page = lru_to_page(&l_hold);
 748                 list_del(&page->lru);
 749                 if (page_mapped(page)) {
 750                         if (!reclaim_mapped ||
 751                             (total_swap_pages == 0 && PageAnon(page)) ||
 752                             page_referenced(page, 0, sc->priority <= 0)) {
 753                                 list_add(&page->lru, &l_active);
 754                                 continue;
 755                         }
 756                 }
 757                 list_add(&page->lru, &l_inactive);
 758         }
 759
 760         pagevec_init(&pvec, 1);
 761         pgmoved = 0;
 762         spin_lock_irq(&zone->lru_lock);
 763         while (!list_empty(&l_inactive)) {
 764                 page = lru_to_page(&l_inactive);
 765                 prefetchw_prev_lru_page(page, &l_inactive, flags);
 766                 if (TestSetPageLRU(page))
 767                         BUG();
 768                 if (!TestClearPageActive(page))
 769                         BUG();
 770                 list_move(&page->lru, inactive_list);
 771                 pgmoved++;
 772                 if (!pagevec_add(&pvec, page)) {
 773                         ckrm_zone_inc_inactive(ckrm_zone, pgmoved);
 774                         zone->nr_inactive += pgmoved;
 775                         spin_unlock_irq(&zone->lru_lock);
 776                         pgdeactivate += pgmoved;
 777                         pgmoved = 0;
 778                         if (buffer_heads_over_limit)
 779                                 pagevec_strip(&pvec);
 780                         __pagevec_release(&pvec);
 781                         spin_lock_irq(&zone->lru_lock);
 782                 }
 783         }
 784         ckrm_zone_inc_inactive(ckrm_zone, pgmoved);
 785         zone->nr_inactive += pgmoved;
 786         pgdeactivate += pgmoved;
 787         if (buffer_heads_over_limit) {
 788                 spin_unlock_irq(&zone->lru_lock);
 789                 pagevec_strip(&pvec);
 790                 spin_lock_irq(&zone->lru_lock);
 791         }
 792
 793         pgmoved = 0;
 794         while (!list_empty(&l_active)) {
 795                 page = lru_to_page(&l_active);
 796                 prefetchw_prev_lru_page(page, &l_active, flags);
 797                 if (TestSetPageLRU(page))
 798                         BUG();
 799                 BUG_ON(!PageActive(page));
 800                 list_move(&page->lru, active_list);
 801                 pgmoved++;
 802                 if (!pagevec_add(&pvec, page)) {
 803                         ckrm_zone_inc_active(ckrm_zone, pgmoved);
 804                         zone->nr_active += pgmoved;
 805                         pgmoved = 0;
 806                         spin_unlock_irq(&zone->lru_lock);
 807                         __pagevec_release(&pvec);
 808                         spin_lock_irq(&zone->lru_lock);
 809                 }
 810         }
 811         ckrm_zone_inc_active(ckrm_zone, pgmoved);
 812         zone->nr_active += pgmoved;
 813         spin_unlock_irq(&zone->lru_lock);
 814         pagevec_release(&pvec);
 815
 816         mod_page_state_zone(zone, pgrefill, pgscanned);
 817         mod_page_state(pgdeactivate, pgdeactivate);
 818 }
 819
 820 #ifdef CONFIG_CKRM_RES_MEM
 821 static int
 822 shrink_weight(struct ckrm_zone *czone)
 823 {
 824         u64 temp;
 825         struct zone *zone = czone->zone;
 826         struct ckrm_mem_res *cls = czone->memcls;
 827         int zone_usage, zone_guar, zone_total, guar, ret, cnt;
 828
 829         zone_usage = czone->nr_active + czone->nr_inactive;
 830         czone->active_over = czone->inactive_over = 0;
 831
 832         if (zone_usage < SWAP_CLUSTER_MAX * 4)
 833                 return 0;
 834
 835         if (cls->pg_guar == CKRM_SHARE_DONTCARE) {
 836                 // no guarantee for this class. use implicit guarantee
 837                 guar = cls->impl_guar / cls->nr_dontcare;
 838         } else {
 839                 guar = cls->pg_unused / cls->nr_dontcare;
 840         }
 841         zone_total = zone->nr_active + zone->nr_inactive + zone->free_pages;
 842         temp = (u64) guar * zone_total;
 843         do_div(temp, ckrm_tot_lru_pages);
 844         zone_guar = (int) temp;
 845
 846         ret = ((zone_usage - zone_guar) > SWAP_CLUSTER_MAX) ?
 847                                 (zone_usage - zone_guar) : 0;
 848         if (ret) {
 849                 cnt = czone->nr_active - (2 * zone_guar / 3);
 850                 if (cnt > 0)
 851                         czone->active_over = cnt;
 852                 cnt = czone->active_over + czone->nr_inactive
 853                                         - zone_guar / 3;
 854                 if (cnt > 0)
 855                         czone->inactive_over = cnt;
 856         }
 857         return ret;
 858 }
 859
 860 static void
 861 shrink_ckrmzone(struct ckrm_zone *czone, struct scan_control *sc)
 862 {
 863         while (czone->shrink_active || czone->shrink_inactive) {
 864                 if (czone->shrink_active) {
 865                         sc->nr_to_scan = min(czone->shrink_active,
 866                                         (unsigned long)SWAP_CLUSTER_MAX);
 867                         czone->shrink_active -= sc->nr_to_scan;
 868                         refill_inactive_zone(czone, sc);
 869                 }
 870                 if (czone->shrink_inactive) {
 871                         sc->nr_to_scan = min(czone->shrink_inactive,
 872                                         (unsigned long)SWAP_CLUSTER_MAX);
 873                         czone->shrink_inactive -= sc->nr_to_scan;
 874                         shrink_cache(czone, sc);
 875                         if (sc->nr_to_reclaim <= 0) {
 876                                 czone->shrink_active = 0;
 877                                 czone->shrink_inactive = 0;
 878                                 break;
 879                         }
 880                 }
 881
 882                 throttle_vm_writeout();
 883         }
 884 }
 885
 886 /* insert an entry to the list and sort decendently*/
 887 static void
 888 list_add_sort(struct list_head *entry, struct list_head *head)
 889 {
 890         struct ckrm_zone *czone, *new =
 891                         list_entry(entry, struct ckrm_zone, victim_list);
 892         struct list_head* pos = head->next;
 893
 894         while (pos != head) {
 895                 czone = list_entry(pos, struct ckrm_zone, victim_list);
 896                 if (new->shrink_weight > czone->shrink_weight) {
 897                         __list_add(entry, pos->prev, pos);
 898                         return;
 899                 }
 900                 pos = pos->next;
 901         }
 902         list_add_tail(entry, head);
 903         return;
 904 }
 905
 906 static void
 907 shrink_choose_victims(struct list_head *victims,
 908                 unsigned long nr_active, unsigned long nr_inactive)
 909 {
 910         unsigned long nr;
 911         struct ckrm_zone* czone;
 912         struct list_head *pos, *next;
 913
 914         pos = victims->next;
 915         while ((pos != victims) && (nr_active || nr_inactive)) {
 916                 czone = list_entry(pos, struct ckrm_zone, victim_list);
 917
 918                 if (nr_active && czone->active_over) {
 919                         nr = min(nr_active, czone->active_over);
 920                         czone->shrink_active += nr;
 921                         czone->active_over -= nr;
 922                         nr_active -= nr;
 923                 }
 924
 925                 if (nr_inactive && czone->inactive_over) {
 926                         nr = min(nr_inactive, czone->inactive_over);
 927                         czone->shrink_inactive += nr;
 928                         czone->inactive_over -= nr;
 929                         nr_inactive -= nr;
 930                 }
 931                 pos = pos->next;
 932         }
 933
 934         pos = victims->next;
 935         while (pos != victims) {
 936                 czone = list_entry(pos, struct ckrm_zone, victim_list);
 937                 next = pos->next;
 938                 if (czone->shrink_active == 0 && czone->shrink_inactive == 0) {
 939                         list_del_init(pos);
 940                         ckrm_clear_shrink(czone);
 941                 }
 942                 pos = next;
 943         }
 944         return;
 945 }
 946
 947 static void
 948 shrink_get_victims(struct zone *zone, unsigned long nr_active,
 949                 unsigned long nr_inactive, struct list_head *victims)
 950 {
 951         struct list_head *pos;
 952         struct ckrm_mem_res *cls;
 953         struct ckrm_zone *czone;
 954         int zoneindex = zone_idx(zone);
 955
 956         if (ckrm_nr_mem_classes <= 1) {
 957                 if (ckrm_mem_root_class) {
 958                         czone = ckrm_mem_root_class->ckrm_zone + zoneindex;
 959                         if (!ckrm_test_set_shrink(czone)) {
 960                                 list_add(&czone->victim_list, victims);
 961                                 czone->shrink_active = nr_active;
 962                                 czone->shrink_inactive = nr_inactive;
 963                         }
 964                 }
 965                 return;
 966         }
 967         spin_lock_irq(&ckrm_mem_lock);
 968         list_for_each_entry(cls, &ckrm_memclass_list, mcls_list) {
 969                 czone = cls->ckrm_zone + zoneindex;
 970                 if (ckrm_test_set_shrink(czone))
 971                         continue;
 972
 973                 czone->shrink_active = 0;
 974                 czone->shrink_inactive = 0;
 975                 czone->shrink_weight = shrink_weight(czone);
 976                 if (czone->shrink_weight) {
 977                         list_add_sort(&czone->victim_list, victims);
 978                 } else {
 979                         ckrm_clear_shrink(czone);
 980                 }
 981         }
 982         pos = victims->next;
 983         while (pos != victims) {
 984                 czone = list_entry(pos, struct ckrm_zone, victim_list);
 985                 pos = pos->next;
 986         }
 987         shrink_choose_victims(victims, nr_active, nr_inactive);
 988         spin_unlock_irq(&ckrm_mem_lock);
 989         pos = victims->next;
 990         while (pos != victims) {
 991                 czone = list_entry(pos, struct ckrm_zone, victim_list);
 992                 pos = pos->next;
 993         }
 994 }
 995 #endif /* CONFIG_CKRM_RES_MEM */
 996
 997 /*
 998  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
 999  */
1000 static void
1001 shrink_zone(struct zone *zone, struct scan_control *sc)
1002 {
1003         unsigned long nr_active;
1004         unsigned long nr_inactive;
1005 #ifdef CONFIG_CKRM_RES_MEM
1006         struct ckrm_zone *czone;
1007 #endif
1008
1009         /*
1010          * Add one to `nr_to_scan' just to make sure that the kernel will
1011          * slowly sift through the active list.
1012          */
1013         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1014         nr_active = zone->nr_scan_active;
1015         if (nr_active >= SWAP_CLUSTER_MAX)
1016                 zone->nr_scan_active = 0;
1017         else
1018                 nr_active = 0;
1019
1020         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1021         nr_inactive = zone->nr_scan_inactive;
1022         if (nr_inactive >= SWAP_CLUSTER_MAX)
1023                 zone->nr_scan_inactive = 0;
1024         else
1025                 nr_inactive = 0;
1026
1027         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
1028
1029 #ifdef CONFIG_CKRM_RES_MEM
1030         if (nr_active || nr_inactive) {
1031                 struct list_head *pos, *next;
1032                 LIST_HEAD(victims);
1033
1034                 shrink_get_victims(zone, nr_active, nr_inactive, &victims);
1035                 pos = victims.next;
1036                 while (pos != &victims) {
1037                         czone = list_entry(pos, struct ckrm_zone, victim_list);
1038                         next = pos->next;
1039                         list_del_init(pos);
1040                         ckrm_clear_shrink(czone);
1041                         sc->nr_to_reclaim = czone->shrink_inactive;
1042                         shrink_ckrmzone(czone, sc);
1043                         pos = next;
1044                 }
1045         }
1046 #else
1047         while (nr_active || nr_inactive) {
1048                 if (nr_active) {
1049                         sc->nr_to_scan = min(nr_active,
1050                                         (unsigned long)SWAP_CLUSTER_MAX);
1051                         nr_active -= sc->nr_to_scan;
1052                         refill_inactive_zone(zone, sc);
1053                 }
1054
1055                 if (nr_inactive) {
1056                         sc->nr_to_scan = min(nr_inactive,
1057                                         (unsigned long)SWAP_CLUSTER_MAX);
1058                         nr_inactive -= sc->nr_to_scan;
1059                         shrink_cache(zone, sc);
1060                         if (sc->nr_to_reclaim <= 0)
1061                                 break;
1062                 }
1063         }
1064 #endif
1065 }
1066
1067 #ifdef CONFIG_CKRM_RES_MEM
1068 // This function needs to be given more thought.
1069 // Shrink the class to be at shrink_to%" of its limit
1070 static void
1071 ckrm_shrink_class(struct ckrm_mem_res *cls)
1072 {
1073         struct scan_control sc;
1074         struct zone *zone;
1075         int zindex = 0, cnt, act_credit = 0, inact_credit = 0;
1076         int shrink_to = ckrm_mem_get_shrink_to();
1077
1078         sc.nr_mapped = read_page_state(nr_mapped);
1079         sc.nr_scanned = 0;
1080         sc.nr_reclaimed = 0;
1081         sc.priority = 0; // always very high priority
1082
1083         check_memclass(cls, "bef_shnk_cls");
1084         for_each_zone(zone) {
1085                 int zone_total, zone_limit, active_limit,
1086                                         inactive_limit, clszone_limit;
1087                 struct ckrm_zone *czone;
1088                 u64 temp;
1089
1090                 czone = &cls->ckrm_zone[zindex];
1091                 if (ckrm_test_set_shrink(czone))
1092                         continue;
1093
1094                 zone->temp_priority = zone->prev_priority;
1095                 zone->prev_priority = sc.priority;
1096
1097                 zone_total = zone->nr_active + zone->nr_inactive
1098                                                 + zone->free_pages;
1099
1100                 temp = (u64) cls->pg_limit * zone_total;
1101                 do_div(temp, ckrm_tot_lru_pages);
1102                 zone_limit = (int) temp;
1103                 clszone_limit = (shrink_to * zone_limit) / 100;
1104                 active_limit = (2 * clszone_limit) / 3; // 2/3rd in active list
1105                 inactive_limit = clszone_limit / 3; // 1/3rd in inactive list
1106
1107                 czone->shrink_active = 0;
1108                 cnt = czone->nr_active + act_credit - active_limit;
1109                 if (cnt > 0) {
1110                         czone->shrink_active = (unsigned long) cnt;
1111                 } else {
1112                         act_credit += cnt;
1113                 }
1114
1115                 czone->shrink_inactive = 0;
1116                 cnt = czone->shrink_active + inact_credit +
1117                                         (czone->nr_inactive - inactive_limit);
1118                 if (cnt > 0) {
1119                         czone->shrink_inactive = (unsigned long) cnt;
1120                 } else {
1121                         inact_credit += cnt;
1122                 }
1123
1124
1125                 if (czone->shrink_active || czone->shrink_inactive) {
1126                         sc.nr_to_reclaim = czone->shrink_inactive;
1127                         shrink_ckrmzone(czone, &sc);
1128                 }
1129                 zone->prev_priority = zone->temp_priority;
1130                 zindex++;
1131                 ckrm_clear_shrink(czone);
1132         }
1133         check_memclass(cls, "aft_shnk_cls");
1134 }
1135
1136 static void
1137 ckrm_shrink_classes(void)
1138 {
1139         struct ckrm_mem_res *cls;
1140
1141         spin_lock_irq(&ckrm_mem_lock);
1142         while (!ckrm_shrink_list_empty()) {
1143                 cls =  list_entry(ckrm_shrink_list.next, struct ckrm_mem_res,
1144                                 shrink_list);
1145                 list_del(&cls->shrink_list);
1146                 cls->flags &= ~MEM_AT_LIMIT;
1147                 spin_unlock_irq(&ckrm_mem_lock);
1148                 ckrm_shrink_class(cls);
1149                 spin_lock_irq(&ckrm_mem_lock);
1150         }
1151         spin_unlock_irq(&ckrm_mem_lock);
1152 }
1153
1154 #else
1155 #define ckrm_shrink_classes()   do { } while(0)
1156 #endif
1157
1158 /*
1159  * This is the direct reclaim path, for page-allocating processes.  We only
1160  * try to reclaim pages from zones which will satisfy the caller's allocation
1161  * request.
1162  *
1163  * We reclaim from a zone even if that zone is over pages_high.  Because:
1164  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1165  *    allocation or
1166  * b) The zones may be over pages_high but they must go *over* pages_high to
1167  *    satisfy the `incremental min' zone defense algorithm.
1168  *
1169  * Returns the number of reclaimed pages.
1170  *
1171  * If a zone is deemed to be full of pinned pages then just give it a light
1172  * scan then give up on it.
1173  */
1174 static void
1175 shrink_caches(struct zone **zones, struct scan_control *sc)
1176 {
1177         int i;
1178
1179         for (i = 0; zones[i] != NULL; i++) {
1180                 struct zone *zone = zones[i];
1181
1182                 if (zone->present_pages == 0)
1183                         continue;
1184
1185                 zone->temp_priority = sc->priority;
1186                 if (zone->prev_priority > sc->priority)
1187                         zone->prev_priority = sc->priority;
1188
1189                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1190                         continue;       /* Let kswapd poll it */
1191
1192                 shrink_zone(zone, sc);
1193         }
1194 }
1195
1196 /*
1197  * This is the main entry point to direct page reclaim.
1198  *
1199  * If a full scan of the inactive list fails to free enough memory then we
1200  * are "out of memory" and something needs to be killed.
1201  *
1202  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1203  * high - the zone may be full of dirty or under-writeback pages, which this
1204  * caller can't do much about.  We kick pdflush and take explicit naps in the
1205  * hope that some of these pages can be written.  But if the allocating task
1206  * holds filesystem locks which prevent writeout this might not work, and the
1207  * allocation attempt will fail.
1208  */
1209 int try_to_free_pages(struct zone **zones,
1210                 unsigned int gfp_mask, unsigned int order)
1211 {
1212         int priority;
1213         int ret = 0;
1214         int total_scanned = 0, total_reclaimed = 0;
1215         struct reclaim_state *reclaim_state = current->reclaim_state;
1216         struct scan_control sc;
1217         unsigned long lru_pages = 0;
1218         int i;
1219
1220         sc.gfp_mask = gfp_mask;
1221         sc.may_writepage = 0;
1222
1223         inc_page_state(allocstall);
1224
1225         for (i = 0; zones[i] != NULL; i++) {
1226                 struct zone *zone = zones[i];
1227
1228                 zone->temp_priority = DEF_PRIORITY;
1229                 lru_pages += zone->nr_active + zone->nr_inactive;
1230         }
1231
1232         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1233                 sc.nr_mapped = read_page_state(nr_mapped);
1234                 sc.nr_scanned = 0;
1235                 sc.nr_reclaimed = 0;
1236                 sc.priority = priority;
1237                 shrink_caches(zones, &sc);
1238                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1239                 if (reclaim_state) {
1240                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1241                         reclaim_state->reclaimed_slab = 0;
1242                 }
1243                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
1244                         ret = 1;
1245                         goto out;
1246                 }
1247                 total_scanned += sc.nr_scanned;
1248                 total_reclaimed += sc.nr_reclaimed;
1249
1250                 /*
1251                  * Try to write back as many pages as we just scanned.  This
1252                  * tends to cause slow streaming writers to write data to the
1253                  * disk smoothly, at the dirtying rate, which is nice.   But
1254                  * that's undesirable in laptop mode, where we *want* lumpy
1255                  * writeout.  So in laptop mode, write out the whole world.
1256                  */
1257                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
1258                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
1259                         sc.may_writepage = 1;
1260                 }
1261
1262                 /* Take a nap, wait for some writeback to complete */
1263                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1264                         blk_congestion_wait(WRITE, HZ/10);
1265         }
1266         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
1267                 out_of_memory(gfp_mask);
1268 out:
1269         for (i = 0; zones[i] != 0; i++)
1270                 zones[i]->prev_priority = zones[i]->temp_priority;
1271         return ret;
1272 }
1273
1274 /*
1275  * For kswapd, balance_pgdat() will work across all this node's zones until
1276  * they are all at pages_high.
1277  *
1278  * If `nr_pages' is non-zero then it is the number of pages which are to be
1279  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1280  * special.
1281  *
1282  * Returns the number of pages which were actually freed.
1283  *
1284  * There is special handling here for zones which are full of pinned pages.
1285  * This can happen if the pages are all mlocked, or if they are all used by
1286  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1287  * What we do is to detect the case where all pages in the zone have been
1288  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1289  * dead and from now on, only perform a short scan.  Basically we're polling
1290  * the zone for when the problem goes away.
1291  *
1292  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1293  * zones which have free_pages > pages_high, but once a zone is found to have
1294  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1295  * of the number of free pages in the lower zones.  This interoperates with
1296  * the page allocator fallback scheme to ensure that aging of pages is balanced
1297  * across the zones.
1298  */
1299 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
1300 {
1301         int to_free = nr_pages;
1302         int all_zones_ok;
1303         int priority;
1304         int i;
1305         int total_scanned, total_reclaimed;
1306         struct reclaim_state *reclaim_state = current->reclaim_state;
1307         struct scan_control sc;
1308
1309 loop_again:
1310         total_scanned = 0;
1311         total_reclaimed = 0;
1312         sc.gfp_mask = GFP_KERNEL;
1313         sc.may_writepage = 0;
1314         sc.nr_mapped = read_page_state(nr_mapped);
1315
1316         inc_page_state(pageoutrun);
1317
1318         for (i = 0; i < pgdat->nr_zones; i++) {
1319                 struct zone *zone = pgdat->node_zones + i;
1320
1321                 zone->temp_priority = DEF_PRIORITY;
1322         }
1323
1324         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1325                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1326                 unsigned long lru_pages = 0;
1327
1328                 all_zones_ok = 1;
1329
1330                 if (nr_pages == 0) {
1331                         /*
1332                          * Scan in the highmem->dma direction for the highest
1333                          * zone which needs scanning
1334                          */
1335                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1336                                 struct zone *zone = pgdat->node_zones + i;
1337
1338                                 if (zone->present_pages == 0)
1339                                         continue;
1340
1341                                 if (zone->all_unreclaimable &&
1342                                                 priority != DEF_PRIORITY)
1343                                         continue;
1344
1345                                 if (zone->free_pages <= zone->pages_high) {
1346                                         end_zone = i;
1347                                         goto scan;
1348                                 }
1349                         }
1350                         goto out;
1351                 } else {
1352                         end_zone = pgdat->nr_zones - 1;
1353                 }
1354 scan:
1355                 for (i = 0; i <= end_zone; i++) {
1356                         struct zone *zone = pgdat->node_zones + i;
1357
1358                         lru_pages += zone->nr_active + zone->nr_inactive;
1359                 }
1360
1361                 /*
1362                  * Now scan the zone in the dma->highmem direction, stopping
1363                  * at the last zone which needs scanning.
1364                  *
1365                  * We do this because the page allocator works in the opposite
1366                  * direction.  This prevents the page allocator from allocating
1367                  * pages behind kswapd's direction of progress, which would
1368                  * cause too much scanning of the lower zones.
1369                  */
1370                 for (i = 0; i <= end_zone; i++) {
1371                         struct zone *zone = pgdat->node_zones + i;
1372
1373                         if (zone->present_pages == 0)
1374                                 continue;
1375
1376                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1377                                 continue;
1378
1379                         if (nr_pages == 0) {    /* Not software suspend */
1380                                 if (zone->free_pages <= zone->pages_high)
1381                                         all_zones_ok = 0;
1382                         }
1383                         zone->temp_priority = priority;
1384                         if (zone->prev_priority > priority)
1385                                 zone->prev_priority = priority;
1386                         sc.nr_scanned = 0;
1387                         sc.nr_reclaimed = 0;
1388                         sc.priority = priority;
1389                         shrink_zone(zone, &sc);
1390                         reclaim_state->reclaimed_slab = 0;
1391                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1392                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1393                         total_reclaimed += sc.nr_reclaimed;
1394                         total_scanned += sc.nr_scanned;
1395                         if (zone->all_unreclaimable)
1396                                 continue;
1397                         if (zone->pages_scanned >= (zone->nr_active +
1398                                                         zone->nr_inactive) * 4)
1399                                 zone->all_unreclaimable = 1;
1400                         /*
1401                          * If we've done a decent amount of scanning and
1402                          * the reclaim ratio is low, start doing writepage
1403                          * even in laptop mode
1404                          */
1405                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1406                             total_scanned > total_reclaimed+total_reclaimed/2)
1407                                 sc.may_writepage = 1;
1408                 }
1409                 if (nr_pages && to_free > total_reclaimed)
1410                         continue;       /* swsusp: need to do more work */
1411                 if (all_zones_ok)
1412                         break;          /* kswapd: all done */
1413                 /*
1414                  * OK, kswapd is getting into trouble.  Take a nap, then take
1415                  * another pass across the zones.
1416                  */
1417                 if (total_scanned && priority < DEF_PRIORITY - 2)
1418                         blk_congestion_wait(WRITE, HZ/10);
1419
1420                 /*
1421                  * We do this so kswapd doesn't build up large priorities for
1422                  * example when it is freeing in parallel with allocators. It
1423                  * matches the direct reclaim path behaviour in terms of impact
1424                  * on zone->*_priority.
1425                  */
1426                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1427                         break;
1428         }
1429 out:
1430         for (i = 0; i < pgdat->nr_zones; i++) {
1431                 struct zone *zone = pgdat->node_zones + i;
1432
1433                 zone->prev_priority = zone->temp_priority;
1434         }
1435         if (!all_zones_ok) {
1436                 cond_resched();
1437                 goto loop_again;
1438         }
1439
1440         return total_reclaimed;
1441 }
1442
1443 /*
1444  * The background pageout daemon, started as a kernel thread
1445  * from the init process.
1446  *
1447  * This basically trickles out pages so that we have _some_
1448  * free memory available even if there is no other activity
1449  * that frees anything up. This is needed for things like routing
1450  * etc, where we otherwise might have all activity going on in
1451  * asynchronous contexts that cannot page things out.
1452  *
1453  * If there are applications that are active memory-allocators
1454  * (most normal use), this basically shouldn't matter.
1455  */
1456 static int kswapd(void *p)
1457 {
1458         pg_data_t *pgdat = (pg_data_t*)p;
1459         struct task_struct *tsk = current;
1460         DEFINE_WAIT(wait);
1461         struct reclaim_state reclaim_state = {
1462                 .reclaimed_slab = 0,
1463         };
1464         cpumask_t cpumask;
1465
1466         daemonize("kswapd%d", pgdat->node_id);
1467         cpumask = node_to_cpumask(pgdat->node_id);
1468         if (!cpus_empty(cpumask))
1469                 set_cpus_allowed(tsk, cpumask);
1470         current->reclaim_state = &reclaim_state;
1471
1472         /*
1473          * Tell the memory management that we're a "memory allocator",
1474          * and that if we need more memory we should get access to it
1475          * regardless (see "__alloc_pages()"). "kswapd" should
1476          * never get caught in the normal page freeing logic.
1477          *
1478          * (Kswapd normally doesn't need memory anyway, but sometimes
1479          * you need a small amount of memory in order to be able to
1480          * page out something else, and this flag essentially protects
1481          * us from recursively trying to free more memory as we're
1482          * trying to free the first piece of memory in the first place).
1483          */
1484         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1485
1486         for ( ; ; ) {
1487                 if (current->flags & PF_FREEZE)
1488                         refrigerator(PF_FREEZE);
1489                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1490                 schedule();
1491                 finish_wait(&pgdat->kswapd_wait, &wait);
1492
1493                 if (!ckrm_shrink_list_empty())
1494                         ckrm_shrink_classes();
1495                 else
1496                         balance_pgdat(pgdat, 0);
1497         }
1498         return 0;
1499 }
1500
1501 /*
1502  * A zone is low on free memory, so wake its kswapd task to service it.
1503  */
1504 void wakeup_kswapd(struct zone *zone)
1505 {
1506         if (zone->present_pages == 0)
1507                 return;
1508         if ((zone->free_pages > zone->pages_low) && ckrm_shrink_list_empty())
1509                 return;
1510         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1511                 return;
1512         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1513 }
1514
1515 #ifdef CONFIG_PM
1516 /*
1517  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1518  * pages.
1519  */
1520 int shrink_all_memory(int nr_pages)
1521 {
1522         pg_data_t *pgdat;
1523         int nr_to_free = nr_pages;
1524         int ret = 0;
1525         struct reclaim_state reclaim_state = {
1526                 .reclaimed_slab = 0,
1527         };
1528
1529         current->reclaim_state = &reclaim_state;
1530         for_each_pgdat(pgdat) {
1531                 int freed;
1532                 freed = balance_pgdat(pgdat, nr_to_free);
1533                 ret += freed;
1534                 nr_to_free -= freed;
1535                 if (nr_to_free <= 0)
1536                         break;
1537         }
1538         current->reclaim_state = NULL;
1539         return ret;
1540 }
1541 #endif
1542
1543 #ifdef CONFIG_HOTPLUG_CPU
1544 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1545    not required for correctness.  So if the last cpu in a node goes
1546    away, we get changed to run anywhere: as the first one comes back,
1547    restore their cpu bindings. */
1548 static int __devinit cpu_callback(struct notifier_block *nfb,
1549                                   unsigned long action,
1550                                   void *hcpu)
1551 {
1552         pg_data_t *pgdat;
1553         cpumask_t mask;
1554
1555         if (action == CPU_ONLINE) {
1556                 for_each_pgdat(pgdat) {
1557                         mask = node_to_cpumask(pgdat->node_id);
1558                         if (any_online_cpu(mask) != NR_CPUS)
1559                                 /* One of our CPUs online: restore mask */
1560                                 set_cpus_allowed(pgdat->kswapd, mask);
1561                 }
1562         }
1563         return NOTIFY_OK;
1564 }
1565 #endif /* CONFIG_HOTPLUG_CPU */
1566
1567 static int __init kswapd_init(void)
1568 {
1569         pg_data_t *pgdat;
1570         swap_setup();
1571         for_each_pgdat(pgdat)
1572                 pgdat->kswapd
1573                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1574         total_memory = nr_free_pagecache_pages();
1575         hotcpu_notifier(cpu_callback, 0);
1576         return 0;
1577 }
1578
1579 module_init(kswapd_init)