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/vmstat.h>
  23 #include <linux/file.h>
  24 #include <linux/writeback.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/cpuset.h>
  35 #include <linux/notifier.h>
  36 #include <linux/rwsem.h>
  37 #include <linux/delay.h>
  38 #include <linux/kthread.h>
  39 #include <linux/freezer.h>
  40 #include <linux/memcontrol.h>
  41
  42 #include <asm/tlbflush.h>
  43 #include <asm/div64.h>
  44
  45 #include <linux/swapops.h>
  46
  47 #include "internal.h"
  48
  49 struct scan_control {
  50         /* Incremented by the number of inactive pages that were scanned */
  51         unsigned long nr_scanned;
  52
  53         /* This context's GFP mask */
  54         gfp_t gfp_mask;
  55
  56         int may_writepage;
  57
  58         /* Can pages be swapped as part of reclaim? */
  59         int may_swap;
  60
  61         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  62          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  63          * In this context, it doesn't matter that we scan the
  64          * whole list at once. */
  65         int swap_cluster_max;
  66
  67         int swappiness;
  68
  69         int all_unreclaimable;
  70
  71         int order;
  72
  73         /*
  74          * Pages that have (or should have) IO pending.  If we run into
  75          * a lot of these, we're better off waiting a little for IO to
  76          * finish rather than scanning more pages in the VM.
  77          */
  78         int nr_io_pages;
  79
  80         /* Which cgroup do we reclaim from */
  81         struct mem_cgroup *mem_cgroup;
  82
  83         /* Pluggable isolate pages callback */
  84         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
  85                         unsigned long *scanned, int order, int mode,
  86                         struct zone *z, struct mem_cgroup *mem_cont,
  87                         int active);
  88 };
  89
  90 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  91
  92 #ifdef ARCH_HAS_PREFETCH
  93 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  94         do {                                                            \
  95                 if ((_page)->lru.prev != _base) {                       \
  96                         struct page *prev;                              \
  97                                                                         \
  98                         prev = lru_to_page(&(_page->lru));              \
  99                         prefetch(&prev->_field);                        \
 100                 }                                                       \
 101         } while (0)
 102 #else
 103 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 104 #endif
 105
 106 #ifdef ARCH_HAS_PREFETCHW
 107 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 108         do {                                                            \
 109                 if ((_page)->lru.prev != _base) {                       \
 110                         struct page *prev;                              \
 111                                                                         \
 112                         prev = lru_to_page(&(_page->lru));              \
 113                         prefetchw(&prev->_field);                       \
 114                 }                                                       \
 115         } while (0)
 116 #else
 117 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 118 #endif
 119
 120 /*
 121  * From 0 .. 100.  Higher means more swappy.
 122  */
 123 int vm_swappiness = 60;
 124 long vm_total_pages;    /* The total number of pages which the VM controls */
 125
 126 static LIST_HEAD(shrinker_list);
 127 static DECLARE_RWSEM(shrinker_rwsem);
 128
 129 #ifdef CONFIG_CGROUP_MEM_CONT
 130 #define scan_global_lru(sc)     (!(sc)->mem_cgroup)
 131 #else
 132 #define scan_global_lru(sc)     (1)
 133 #endif
 134
 135 /*
 136  * Add a shrinker callback to be called from the vm
 137  */
 138 void register_shrinker(struct shrinker *shrinker)
 139 {
 140         shrinker->nr = 0;
 141         down_write(&shrinker_rwsem);
 142         list_add_tail(&shrinker->list, &shrinker_list);
 143         up_write(&shrinker_rwsem);
 144 }
 145 EXPORT_SYMBOL(register_shrinker);
 146
 147 /*
 148  * Remove one
 149  */
 150 void unregister_shrinker(struct shrinker *shrinker)
 151 {
 152         down_write(&shrinker_rwsem);
 153         list_del(&shrinker->list);
 154         up_write(&shrinker_rwsem);
 155 }
 156 EXPORT_SYMBOL(unregister_shrinker);
 157
 158 #define SHRINK_BATCH 128
 159 /*
 160  * Call the shrink functions to age shrinkable caches
 161  *
 162  * Here we assume it costs one seek to replace a lru page and that it also
 163  * takes a seek to recreate a cache object.  With this in mind we age equal
 164  * percentages of the lru and ageable caches.  This should balance the seeks
 165  * generated by these structures.
 166  *
 167  * If the vm encountered mapped pages on the LRU it increase the pressure on
 168  * slab to avoid swapping.
 169  *
 170  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 171  *
 172  * `lru_pages' represents the number of on-LRU pages in all the zones which
 173  * are eligible for the caller's allocation attempt.  It is used for balancing
 174  * slab reclaim versus page reclaim.
 175  *
 176  * Returns the number of slab objects which we shrunk.
 177  */
 178 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
 179                         unsigned long lru_pages)
 180 {
 181         struct shrinker *shrinker;
 182         unsigned long ret = 0;
 183
 184         if (scanned == 0)
 185                 scanned = SWAP_CLUSTER_MAX;
 186
 187         if (!down_read_trylock(&shrinker_rwsem))
 188                 return 1;       /* Assume we'll be able to shrink next time */
 189
 190         list_for_each_entry(shrinker, &shrinker_list, list) {
 191                 unsigned long long delta;
 192                 unsigned long total_scan;
 193                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
 194
 195                 delta = (4 * scanned) / shrinker->seeks;
 196                 delta *= max_pass;
 197                 do_div(delta, lru_pages + 1);
 198                 shrinker->nr += delta;
 199                 if (shrinker->nr < 0) {
 200                         printk(KERN_ERR "%s: nr=%ld\n",
 201                                         __FUNCTION__, shrinker->nr);
 202                         shrinker->nr = max_pass;
 203                 }
 204
 205                 /*
 206                  * Avoid risking looping forever due to too large nr value:
 207                  * never try to free more than twice the estimate number of
 208                  * freeable entries.
 209                  */
 210                 if (shrinker->nr > max_pass * 2)
 211                         shrinker->nr = max_pass * 2;
 212
 213                 total_scan = shrinker->nr;
 214                 shrinker->nr = 0;
 215
 216                 while (total_scan >= SHRINK_BATCH) {
 217                         long this_scan = SHRINK_BATCH;
 218                         int shrink_ret;
 219                         int nr_before;
 220
 221                         nr_before = (*shrinker->shrink)(0, gfp_mask);
 222                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
 223                         if (shrink_ret == -1)
 224                                 break;
 225                         if (shrink_ret < nr_before)
 226                                 ret += nr_before - shrink_ret;
 227                         count_vm_events(SLABS_SCANNED, this_scan);
 228                         total_scan -= this_scan;
 229
 230                         cond_resched();
 231                 }
 232
 233                 shrinker->nr += total_scan;
 234         }
 235         up_read(&shrinker_rwsem);
 236         return ret;
 237 }
 238
 239 /* Called without lock on whether page is mapped, so answer is unstable */
 240 static inline int page_mapping_inuse(struct page *page)
 241 {
 242         struct address_space *mapping;
 243
 244         /* Page is in somebody's page tables. */
 245         if (page_mapped(page))
 246                 return 1;
 247
 248         /* Be more reluctant to reclaim swapcache than pagecache */
 249         if (PageSwapCache(page))
 250                 return 1;
 251
 252         mapping = page_mapping(page);
 253         if (!mapping)
 254                 return 0;
 255
 256         /* File is mmap'd by somebody? */
 257         return mapping_mapped(mapping);
 258 }
 259
 260 static inline int is_page_cache_freeable(struct page *page)
 261 {
 262         return page_count(page) - !!PagePrivate(page) == 2;
 263 }
 264
 265 static int may_write_to_queue(struct backing_dev_info *bdi)
 266 {
 267         if (current->flags & PF_SWAPWRITE)
 268                 return 1;
 269         if (!bdi_write_congested(bdi))
 270                 return 1;
 271         if (bdi == current->backing_dev_info)
 272                 return 1;
 273         return 0;
 274 }
 275
 276 /*
 277  * We detected a synchronous write error writing a page out.  Probably
 278  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 279  * fsync(), msync() or close().
 280  *
 281  * The tricky part is that after writepage we cannot touch the mapping: nothing
 282  * prevents it from being freed up.  But we have a ref on the page and once
 283  * that page is locked, the mapping is pinned.
 284  *
 285  * We're allowed to run sleeping lock_page() here because we know the caller has
 286  * __GFP_FS.
 287  */
 288 static void handle_write_error(struct address_space *mapping,
 289                                 struct page *page, int error)
 290 {
 291         lock_page(page);
 292         if (page_mapping(page) == mapping)
 293                 mapping_set_error(mapping, error);
 294         unlock_page(page);
 295 }
 296
 297 /* Request for sync pageout. */
 298 enum pageout_io {
 299         PAGEOUT_IO_ASYNC,
 300         PAGEOUT_IO_SYNC,
 301 };
 302
 303 /* possible outcome of pageout() */
 304 typedef enum {
 305         /* failed to write page out, page is locked */
 306         PAGE_KEEP,
 307         /* move page to the active list, page is locked */
 308         PAGE_ACTIVATE,
 309         /* page has been sent to the disk successfully, page is unlocked */
 310         PAGE_SUCCESS,
 311         /* page is clean and locked */
 312         PAGE_CLEAN,
 313 } pageout_t;
 314
 315 /*
 316  * pageout is called by shrink_page_list() for each dirty page.
 317  * Calls ->writepage().
 318  */
 319 static pageout_t pageout(struct page *page, struct address_space *mapping,
 320                                                 enum pageout_io sync_writeback)
 321 {
 322         /*
 323          * If the page is dirty, only perform writeback if that write
 324          * will be non-blocking.  To prevent this allocation from being
 325          * stalled by pagecache activity.  But note that there may be
 326          * stalls if we need to run get_block().  We could test
 327          * PagePrivate for that.
 328          *
 329          * If this process is currently in generic_file_write() against
 330          * this page's queue, we can perform writeback even if that
 331          * will block.
 332          *
 333          * If the page is swapcache, write it back even if that would
 334          * block, for some throttling. This happens by accident, because
 335          * swap_backing_dev_info is bust: it doesn't reflect the
 336          * congestion state of the swapdevs.  Easy to fix, if needed.
 337          * See swapfile.c:page_queue_congested().
 338          */
 339         if (!is_page_cache_freeable(page))
 340                 return PAGE_KEEP;
 341         if (!mapping) {
 342                 /*
 343                  * Some data journaling orphaned pages can have
 344                  * page->mapping == NULL while being dirty with clean buffers.
 345                  */
 346                 if (PagePrivate(page)) {
 347                         if (try_to_free_buffers(page)) {
 348                                 ClearPageDirty(page);
 349                                 printk("%s: orphaned page\n", __FUNCTION__);
 350                                 return PAGE_CLEAN;
 351                         }
 352                 }
 353                 return PAGE_KEEP;
 354         }
 355         if (mapping->a_ops->writepage == NULL)
 356                 return PAGE_ACTIVATE;
 357         if (!may_write_to_queue(mapping->backing_dev_info))
 358                 return PAGE_KEEP;
 359
 360         if (clear_page_dirty_for_io(page)) {
 361                 int res;
 362                 struct writeback_control wbc = {
 363                         .sync_mode = WB_SYNC_NONE,
 364                         .nr_to_write = SWAP_CLUSTER_MAX,
 365                         .range_start = 0,
 366                         .range_end = LLONG_MAX,
 367                         .nonblocking = 1,
 368                         .for_reclaim = 1,
 369                 };
 370
 371                 SetPageReclaim(page);
 372                 res = mapping->a_ops->writepage(page, &wbc);
 373                 if (res < 0)
 374                         handle_write_error(mapping, page, res);
 375                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 376                         ClearPageReclaim(page);
 377                         return PAGE_ACTIVATE;
 378                 }
 379
 380                 /*
 381                  * Wait on writeback if requested to. This happens when
 382                  * direct reclaiming a large contiguous area and the
 383                  * first attempt to free a range of pages fails.
 384                  */
 385                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
 386                         wait_on_page_writeback(page);
 387
 388                 if (!PageWriteback(page)) {
 389                         /* synchronous write or broken a_ops? */
 390                         ClearPageReclaim(page);
 391                 }
 392                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
 393                 return PAGE_SUCCESS;
 394         }
 395
 396         return PAGE_CLEAN;
 397 }
 398
 399 /*
 400  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 401  * someone else has a ref on the page, abort and return 0.  If it was
 402  * successfully detached, return 1.  Assumes the caller has a single ref on
 403  * this page.
 404  */
 405 int remove_mapping(struct address_space *mapping, struct page *page)
 406 {
 407         BUG_ON(!PageLocked(page));
 408         BUG_ON(mapping != page_mapping(page));
 409
 410         write_lock_irq(&mapping->tree_lock);
 411         /*
 412          * The non racy check for a busy page.
 413          *
 414          * Must be careful with the order of the tests. When someone has
 415          * a ref to the page, it may be possible that they dirty it then
 416          * drop the reference. So if PageDirty is tested before page_count
 417          * here, then the following race may occur:
 418          *
 419          * get_user_pages(&page);
 420          * [user mapping goes away]
 421          * write_to(page);
 422          *                              !PageDirty(page)    [good]
 423          * SetPageDirty(page);
 424          * put_page(page);
 425          *                              !page_count(page)   [good, discard it]
 426          *
 427          * [oops, our write_to data is lost]
 428          *
 429          * Reversing the order of the tests ensures such a situation cannot
 430          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 431          * load is not satisfied before that of page->_count.
 432          *
 433          * Note that if SetPageDirty is always performed via set_page_dirty,
 434          * and thus under tree_lock, then this ordering is not required.
 435          */
 436         if (unlikely(page_count(page) != 2))
 437                 goto cannot_free;
 438         smp_rmb();
 439         if (unlikely(PageDirty(page)))
 440                 goto cannot_free;
 441
 442         if (PageSwapCache(page)) {
 443                 swp_entry_t swap = { .val = page_private(page) };
 444                 __delete_from_swap_cache(page);
 445                 write_unlock_irq(&mapping->tree_lock);
 446                 swap_free(swap);
 447                 __put_page(page);       /* The pagecache ref */
 448                 return 1;
 449         }
 450
 451         __remove_from_page_cache(page);
 452         write_unlock_irq(&mapping->tree_lock);
 453         __put_page(page);
 454         return 1;
 455
 456 cannot_free:
 457         write_unlock_irq(&mapping->tree_lock);
 458         return 0;
 459 }
 460
 461 /*
 462  * shrink_page_list() returns the number of reclaimed pages
 463  */
 464 static unsigned long shrink_page_list(struct list_head *page_list,
 465                                         struct scan_control *sc,
 466                                         enum pageout_io sync_writeback)
 467 {
 468         LIST_HEAD(ret_pages);
 469         struct pagevec freed_pvec;
 470         int pgactivate = 0;
 471         unsigned long nr_reclaimed = 0;
 472
 473         cond_resched();
 474
 475         pagevec_init(&freed_pvec, 1);
 476         while (!list_empty(page_list)) {
 477                 struct address_space *mapping;
 478                 struct page *page;
 479                 int may_enter_fs;
 480                 int referenced;
 481
 482                 cond_resched();
 483
 484                 page = lru_to_page(page_list);
 485                 list_del(&page->lru);
 486
 487                 if (TestSetPageLocked(page))
 488                         goto keep;
 489
 490                 VM_BUG_ON(PageActive(page));
 491
 492                 sc->nr_scanned++;
 493
 494                 if (!sc->may_swap && page_mapped(page))
 495                         goto keep_locked;
 496
 497                 /* Double the slab pressure for mapped and swapcache pages */
 498                 if (page_mapped(page) || PageSwapCache(page))
 499                         sc->nr_scanned++;
 500
 501                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 502                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 503
 504                 if (PageWriteback(page)) {
 505                         /*
 506                          * Synchronous reclaim is performed in two passes,
 507                          * first an asynchronous pass over the list to
 508                          * start parallel writeback, and a second synchronous
 509                          * pass to wait for the IO to complete.  Wait here
 510                          * for any page for which writeback has already
 511                          * started.
 512                          */
 513                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
 514                                 wait_on_page_writeback(page);
 515                         else {
 516                                 sc->nr_io_pages++;
 517                                 goto keep_locked;
 518                         }
 519                 }
 520
 521                 referenced = page_referenced(page, 1, sc->mem_cgroup);
 522                 /* In active use or really unfreeable?  Activate it. */
 523                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
 524                                         referenced && page_mapping_inuse(page))
 525                         goto activate_locked;
 526
 527 #ifdef CONFIG_SWAP
 528                 /*
 529                  * Anonymous process memory has backing store?
 530                  * Try to allocate it some swap space here.
 531                  */
 532                 if (PageAnon(page) && !PageSwapCache(page))
 533                         if (!add_to_swap(page, GFP_ATOMIC))
 534                                 goto activate_locked;
 535 #endif /* CONFIG_SWAP */
 536
 537                 mapping = page_mapping(page);
 538
 539                 /*
 540                  * The page is mapped into the page tables of one or more
 541                  * processes. Try to unmap it here.
 542                  */
 543                 if (page_mapped(page) && mapping) {
 544                         switch (try_to_unmap(page, 0)) {
 545                         case SWAP_FAIL:
 546                                 goto activate_locked;
 547                         case SWAP_AGAIN:
 548                                 goto keep_locked;
 549                         case SWAP_SUCCESS:
 550                                 ; /* try to free the page below */
 551                         }
 552                 }
 553
 554                 if (PageDirty(page)) {
 555                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
 556                                 goto keep_locked;
 557                         if (!may_enter_fs) {
 558                                 sc->nr_io_pages++;
 559                                 goto keep_locked;
 560                         }
 561                         if (!sc->may_writepage)
 562                                 goto keep_locked;
 563
 564                         /* Page is dirty, try to write it out here */
 565                         switch (pageout(page, mapping, sync_writeback)) {
 566                         case PAGE_KEEP:
 567                                 goto keep_locked;
 568                         case PAGE_ACTIVATE:
 569                                 goto activate_locked;
 570                         case PAGE_SUCCESS:
 571                                 if (PageWriteback(page) || PageDirty(page)) {
 572                                         sc->nr_io_pages++;
 573                                         goto keep;
 574                                 }
 575                                 /*
 576                                  * A synchronous write - probably a ramdisk.  Go
 577                                  * ahead and try to reclaim the page.
 578                                  */
 579                                 if (TestSetPageLocked(page))
 580                                         goto keep;
 581                                 if (PageDirty(page) || PageWriteback(page))
 582                                         goto keep_locked;
 583                                 mapping = page_mapping(page);
 584                         case PAGE_CLEAN:
 585                                 ; /* try to free the page below */
 586                         }
 587                 }
 588
 589                 /*
 590                  * If the page has buffers, try to free the buffer mappings
 591                  * associated with this page. If we succeed we try to free
 592                  * the page as well.
 593                  *
 594                  * We do this even if the page is PageDirty().
 595                  * try_to_release_page() does not perform I/O, but it is
 596                  * possible for a page to have PageDirty set, but it is actually
 597                  * clean (all its buffers are clean).  This happens if the
 598                  * buffers were written out directly, with submit_bh(). ext3
 599                  * will do this, as well as the blockdev mapping.
 600                  * try_to_release_page() will discover that cleanness and will
 601                  * drop the buffers and mark the page clean - it can be freed.
 602                  *
 603                  * Rarely, pages can have buffers and no ->mapping.  These are
 604                  * the pages which were not successfully invalidated in
 605                  * truncate_complete_page().  We try to drop those buffers here
 606                  * and if that worked, and the page is no longer mapped into
 607                  * process address space (page_count == 1) it can be freed.
 608                  * Otherwise, leave the page on the LRU so it is swappable.
 609                  */
 610                 if (PagePrivate(page)) {
 611                         if (!try_to_release_page(page, sc->gfp_mask))
 612                                 goto activate_locked;
 613                         if (!mapping && page_count(page) == 1)
 614                                 goto free_it;
 615                 }
 616
 617                 if (!mapping || !remove_mapping(mapping, page))
 618                         goto keep_locked;
 619
 620 free_it:
 621                 unlock_page(page);
 622                 nr_reclaimed++;
 623                 if (!pagevec_add(&freed_pvec, page))
 624                         __pagevec_release_nonlru(&freed_pvec);
 625                 continue;
 626
 627 activate_locked:
 628                 SetPageActive(page);
 629                 pgactivate++;
 630 keep_locked:
 631                 unlock_page(page);
 632 keep:
 633                 list_add(&page->lru, &ret_pages);
 634                 VM_BUG_ON(PageLRU(page));
 635         }
 636         list_splice(&ret_pages, page_list);
 637         if (pagevec_count(&freed_pvec))
 638                 __pagevec_release_nonlru(&freed_pvec);
 639         count_vm_events(PGACTIVATE, pgactivate);
 640         return nr_reclaimed;
 641 }
 642
 643 /* LRU Isolation modes. */
 644 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
 645 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
 646 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
 647
 648 /*
 649  * Attempt to remove the specified page from its LRU.  Only take this page
 650  * if it is of the appropriate PageActive status.  Pages which are being
 651  * freed elsewhere are also ignored.
 652  *
 653  * page:        page to consider
 654  * mode:        one of the LRU isolation modes defined above
 655  *
 656  * returns 0 on success, -ve errno on failure.
 657  */
 658 int __isolate_lru_page(struct page *page, int mode)
 659 {
 660         int ret = -EINVAL;
 661
 662         /* Only take pages on the LRU. */
 663         if (!PageLRU(page))
 664                 return ret;
 665
 666         /*
 667          * When checking the active state, we need to be sure we are
 668          * dealing with comparible boolean values.  Take the logical not
 669          * of each.
 670          */
 671         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
 672                 return ret;
 673
 674         ret = -EBUSY;
 675         if (likely(get_page_unless_zero(page))) {
 676                 /*
 677                  * Be careful not to clear PageLRU until after we're
 678                  * sure the page is not being freed elsewhere -- the
 679                  * page release code relies on it.
 680                  */
 681                 ClearPageLRU(page);
 682                 ret = 0;
 683         }
 684
 685         return ret;
 686 }
 687
 688 /*
 689  * zone->lru_lock is heavily contended.  Some of the functions that
 690  * shrink the lists perform better by taking out a batch of pages
 691  * and working on them outside the LRU lock.
 692  *
 693  * For pagecache intensive workloads, this function is the hottest
 694  * spot in the kernel (apart from copy_*_user functions).
 695  *
 696  * Appropriate locks must be held before calling this function.
 697  *
 698  * @nr_to_scan: The number of pages to look through on the list.
 699  * @src:        The LRU list to pull pages off.
 700  * @dst:        The temp list to put pages on to.
 701  * @scanned:    The number of pages that were scanned.
 702  * @order:      The caller's attempted allocation order
 703  * @mode:       One of the LRU isolation modes
 704  *
 705  * returns how many pages were moved onto *@dst.
 706  */
 707 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
 708                 struct list_head *src, struct list_head *dst,
 709                 unsigned long *scanned, int order, int mode)
 710 {
 711         unsigned long nr_taken = 0;
 712         unsigned long scan;
 713
 714         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 715                 struct page *page;
 716                 unsigned long pfn;
 717                 unsigned long end_pfn;
 718                 unsigned long page_pfn;
 719                 int zone_id;
 720
 721                 page = lru_to_page(src);
 722                 prefetchw_prev_lru_page(page, src, flags);
 723
 724                 VM_BUG_ON(!PageLRU(page));
 725
 726                 switch (__isolate_lru_page(page, mode)) {
 727                 case 0:
 728                         list_move(&page->lru, dst);
 729                         nr_taken++;
 730                         break;
 731
 732                 case -EBUSY:
 733                         /* else it is being freed elsewhere */
 734                         list_move(&page->lru, src);
 735                         continue;
 736
 737                 default:
 738                         BUG();
 739                 }
 740
 741                 if (!order)
 742                         continue;
 743
 744                 /*
 745                  * Attempt to take all pages in the order aligned region
 746                  * surrounding the tag page.  Only take those pages of
 747                  * the same active state as that tag page.  We may safely
 748                  * round the target page pfn down to the requested order
 749                  * as the mem_map is guarenteed valid out to MAX_ORDER,
 750                  * where that page is in a different zone we will detect
 751                  * it from its zone id and abort this block scan.
 752                  */
 753                 zone_id = page_zone_id(page);
 754                 page_pfn = page_to_pfn(page);
 755                 pfn = page_pfn & ~((1 << order) - 1);
 756                 end_pfn = pfn + (1 << order);
 757                 for (; pfn < end_pfn; pfn++) {
 758                         struct page *cursor_page;
 759
 760                         /* The target page is in the block, ignore it. */
 761                         if (unlikely(pfn == page_pfn))
 762                                 continue;
 763
 764                         /* Avoid holes within the zone. */
 765                         if (unlikely(!pfn_valid_within(pfn)))
 766                                 break;
 767
 768                         cursor_page = pfn_to_page(pfn);
 769                         /* Check that we have not crossed a zone boundary. */
 770                         if (unlikely(page_zone_id(cursor_page) != zone_id))
 771                                 continue;
 772                         switch (__isolate_lru_page(cursor_page, mode)) {
 773                         case 0:
 774                                 list_move(&cursor_page->lru, dst);
 775                                 nr_taken++;
 776                                 scan++;
 777                                 break;
 778
 779                         case -EBUSY:
 780                                 /* else it is being freed elsewhere */
 781                                 list_move(&cursor_page->lru, src);
 782                         default:
 783                                 break;
 784                         }
 785                 }
 786         }
 787
 788         *scanned = scan;
 789         return nr_taken;
 790 }
 791
 792 static unsigned long isolate_pages_global(unsigned long nr,
 793                                         struct list_head *dst,
 794                                         unsigned long *scanned, int order,
 795                                         int mode, struct zone *z,
 796                                         struct mem_cgroup *mem_cont,
 797                                         int active)
 798 {
 799         if (active)
 800                 return isolate_lru_pages(nr, &z->active_list, dst,
 801                                                 scanned, order, mode);
 802         else
 803                 return isolate_lru_pages(nr, &z->inactive_list, dst,
 804                                                 scanned, order, mode);
 805 }
 806
 807 /*
 808  * clear_active_flags() is a helper for shrink_active_list(), clearing
 809  * any active bits from the pages in the list.
 810  */
 811 static unsigned long clear_active_flags(struct list_head *page_list)
 812 {
 813         int nr_active = 0;
 814         struct page *page;
 815
 816         list_for_each_entry(page, page_list, lru)
 817                 if (PageActive(page)) {
 818                         ClearPageActive(page);
 819                         nr_active++;
 820                 }
 821
 822         return nr_active;
 823 }
 824
 825 /*
 826  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
 827  * of reclaimed pages
 828  */
 829 static unsigned long shrink_inactive_list(unsigned long max_scan,
 830                                 struct zone *zone, struct scan_control *sc)
 831 {
 832         LIST_HEAD(page_list);
 833         struct pagevec pvec;
 834         unsigned long nr_scanned = 0;
 835         unsigned long nr_reclaimed = 0;
 836
 837         pagevec_init(&pvec, 1);
 838
 839         lru_add_drain();
 840         spin_lock_irq(&zone->lru_lock);
 841         do {
 842                 struct page *page;
 843                 unsigned long nr_taken;
 844                 unsigned long nr_scan;
 845                 unsigned long nr_freed;
 846                 unsigned long nr_active;
 847
 848                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
 849                              &page_list, &nr_scan, sc->order,
 850                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
 851                                              ISOLATE_BOTH : ISOLATE_INACTIVE,
 852                                 zone, sc->mem_cgroup, 0);
 853                 nr_active = clear_active_flags(&page_list);
 854                 __count_vm_events(PGDEACTIVATE, nr_active);
 855
 856                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
 857                 __mod_zone_page_state(zone, NR_INACTIVE,
 858                                                 -(nr_taken - nr_active));
 859                 if (scan_global_lru(sc))
 860                         zone->pages_scanned += nr_scan;
 861                 spin_unlock_irq(&zone->lru_lock);
 862
 863                 nr_scanned += nr_scan;
 864                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
 865
 866                 /*
 867                  * If we are direct reclaiming for contiguous pages and we do
 868                  * not reclaim everything in the list, try again and wait
 869                  * for IO to complete. This will stall high-order allocations
 870                  * but that should be acceptable to the caller
 871                  */
 872                 if (nr_freed < nr_taken && !current_is_kswapd() &&
 873                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
 874                         congestion_wait(WRITE, HZ/10);
 875
 876                         /*
 877                          * The attempt at page out may have made some
 878                          * of the pages active, mark them inactive again.
 879                          */
 880                         nr_active = clear_active_flags(&page_list);
 881                         count_vm_events(PGDEACTIVATE, nr_active);
 882
 883                         nr_freed += shrink_page_list(&page_list, sc,
 884                                                         PAGEOUT_IO_SYNC);
 885                 }
 886
 887                 nr_reclaimed += nr_freed;
 888                 local_irq_disable();
 889                 if (current_is_kswapd()) {
 890                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
 891                         __count_vm_events(KSWAPD_STEAL, nr_freed);
 892                 } else if (scan_global_lru(sc))
 893                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
 894
 895                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
 896
 897                 if (nr_taken == 0)
 898                         goto done;
 899
 900                 spin_lock(&zone->lru_lock);
 901                 /*
 902                  * Put back any unfreeable pages.
 903                  */
 904                 while (!list_empty(&page_list)) {
 905                         page = lru_to_page(&page_list);
 906                         VM_BUG_ON(PageLRU(page));
 907                         SetPageLRU(page);
 908                         list_del(&page->lru);
 909                         if (PageActive(page))
 910                                 add_page_to_active_list(zone, page);
 911                         else
 912                                 add_page_to_inactive_list(zone, page);
 913                         if (!pagevec_add(&pvec, page)) {
 914                                 spin_unlock_irq(&zone->lru_lock);
 915                                 __pagevec_release(&pvec);
 916                                 spin_lock_irq(&zone->lru_lock);
 917                         }
 918                 }
 919         } while (nr_scanned < max_scan);
 920         spin_unlock(&zone->lru_lock);
 921 done:
 922         local_irq_enable();
 923         pagevec_release(&pvec);
 924         return nr_reclaimed;
 925 }
 926
 927 /*
 928  * We are about to scan this zone at a certain priority level.  If that priority
 929  * level is smaller (ie: more urgent) than the previous priority, then note
 930  * that priority level within the zone.  This is done so that when the next
 931  * process comes in to scan this zone, it will immediately start out at this
 932  * priority level rather than having to build up its own scanning priority.
 933  * Here, this priority affects only the reclaim-mapped threshold.
 934  */
 935 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
 936 {
 937         if (priority < zone->prev_priority)
 938                 zone->prev_priority = priority;
 939 }
 940
 941 static inline int zone_is_near_oom(struct zone *zone)
 942 {
 943         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
 944                                 + zone_page_state(zone, NR_INACTIVE))*3;
 945 }
 946
 947 /*
 948  * Determine we should try to reclaim mapped pages.
 949  * This is called only when sc->mem_cgroup is NULL.
 950  */
 951 static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
 952                                 int priority)
 953 {
 954         long mapped_ratio;
 955         long distress;
 956         long swap_tendency;
 957         long imbalance;
 958         int reclaim_mapped = 0;
 959         int prev_priority;
 960
 961         if (scan_global_lru(sc) && zone_is_near_oom(zone))
 962                 return 1;
 963         /*
 964          * `distress' is a measure of how much trouble we're having
 965          * reclaiming pages.  0 -> no problems.  100 -> great trouble.
 966          */
 967         if (scan_global_lru(sc))
 968                 prev_priority = zone->prev_priority;
 969         else
 970                 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
 971
 972         distress = 100 >> min(prev_priority, priority);
 973
 974         /*
 975          * The point of this algorithm is to decide when to start
 976          * reclaiming mapped memory instead of just pagecache.  Work out
 977          * how much memory
 978          * is mapped.
 979          */
 980         if (scan_global_lru(sc))
 981                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
 982                                 global_page_state(NR_ANON_PAGES)) * 100) /
 983                                         vm_total_pages;
 984         else
 985                 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
 986
 987         /*
 988          * Now decide how much we really want to unmap some pages.  The
 989          * mapped ratio is downgraded - just because there's a lot of
 990          * mapped memory doesn't necessarily mean that page reclaim
 991          * isn't succeeding.
 992          *
 993          * The distress ratio is important - we don't want to start
 994          * going oom.
 995          *
 996          * A 100% value of vm_swappiness overrides this algorithm
 997          * altogether.
 998          */
 999         swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1000
1001         /*
1002          * If there's huge imbalance between active and inactive
1003          * (think active 100 times larger than inactive) we should
1004          * become more permissive, or the system will take too much
1005          * cpu before it start swapping during memory pressure.
1006          * Distress is about avoiding early-oom, this is about
1007          * making swappiness graceful despite setting it to low
1008          * values.
1009          *
1010          * Avoid div by zero with nr_inactive+1, and max resulting
1011          * value is vm_total_pages.
1012          */
1013         if (scan_global_lru(sc)) {
1014                 imbalance  = zone_page_state(zone, NR_ACTIVE);
1015                 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1016         } else
1017                 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1018
1019         /*
1020          * Reduce the effect of imbalance if swappiness is low,
1021          * this means for a swappiness very low, the imbalance
1022          * must be much higher than 100 for this logic to make
1023          * the difference.
1024          *
1025          * Max temporary value is vm_total_pages*100.
1026          */
1027         imbalance *= (vm_swappiness + 1);
1028         imbalance /= 100;
1029
1030         /*
1031          * If not much of the ram is mapped, makes the imbalance
1032          * less relevant, it's high priority we refill the inactive
1033          * list with mapped pages only in presence of high ratio of
1034          * mapped pages.
1035          *
1036          * Max temporary value is vm_total_pages*100.
1037          */
1038         imbalance *= mapped_ratio;
1039         imbalance /= 100;
1040
1041         /* apply imbalance feedback to swap_tendency */
1042         swap_tendency += imbalance;
1043
1044         /*
1045          * Now use this metric to decide whether to start moving mapped
1046          * memory onto the inactive list.
1047          */
1048         if (swap_tendency >= 100)
1049                 reclaim_mapped = 1;
1050
1051         return reclaim_mapped;
1052 }
1053
1054 /*
1055  * This moves pages from the active list to the inactive list.
1056  *
1057  * We move them the other way if the page is referenced by one or more
1058  * processes, from rmap.
1059  *
1060  * If the pages are mostly unmapped, the processing is fast and it is
1061  * appropriate to hold zone->lru_lock across the whole operation.  But if
1062  * the pages are mapped, the processing is slow (page_referenced()) so we
1063  * should drop zone->lru_lock around each page.  It's impossible to balance
1064  * this, so instead we remove the pages from the LRU while processing them.
1065  * It is safe to rely on PG_active against the non-LRU pages in here because
1066  * nobody will play with that bit on a non-LRU page.
1067  *
1068  * The downside is that we have to touch page->_count against each page.
1069  * But we had to alter page->flags anyway.
1070  */
1071
1072
1073 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1074                                 struct scan_control *sc, int priority)
1075 {
1076         unsigned long pgmoved;
1077         int pgdeactivate = 0;
1078         unsigned long pgscanned;
1079         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1080         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1081         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1082         struct page *page;
1083         struct pagevec pvec;
1084         int reclaim_mapped = 0;
1085
1086         if (sc->may_swap)
1087                 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
1088
1089         lru_add_drain();
1090         spin_lock_irq(&zone->lru_lock);
1091         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1092                                         ISOLATE_ACTIVE, zone,
1093                                         sc->mem_cgroup, 1);
1094         /*
1095          * zone->pages_scanned is used for detect zone's oom
1096          * mem_cgroup remembers nr_scan by itself.
1097          */
1098         if (scan_global_lru(sc))
1099                 zone->pages_scanned += pgscanned;
1100
1101         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1102         spin_unlock_irq(&zone->lru_lock);
1103
1104         while (!list_empty(&l_hold)) {
1105                 cond_resched();
1106                 page = lru_to_page(&l_hold);
1107                 list_del(&page->lru);
1108                 if (page_mapped(page)) {
1109                         if (!reclaim_mapped ||
1110                             (total_swap_pages == 0 && PageAnon(page)) ||
1111                             page_referenced(page, 0, sc->mem_cgroup)) {
1112                                 list_add(&page->lru, &l_active);
1113                                 continue;
1114                         }
1115                 }
1116                 list_add(&page->lru, &l_inactive);
1117         }
1118
1119         pagevec_init(&pvec, 1);
1120         pgmoved = 0;
1121         spin_lock_irq(&zone->lru_lock);
1122         while (!list_empty(&l_inactive)) {
1123                 page = lru_to_page(&l_inactive);
1124                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1125                 VM_BUG_ON(PageLRU(page));
1126                 SetPageLRU(page);
1127                 VM_BUG_ON(!PageActive(page));
1128                 ClearPageActive(page);
1129
1130                 list_move(&page->lru, &zone->inactive_list);
1131                 mem_cgroup_move_lists(page_get_page_cgroup(page), false);
1132                 pgmoved++;
1133                 if (!pagevec_add(&pvec, page)) {
1134                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1135                         spin_unlock_irq(&zone->lru_lock);
1136                         pgdeactivate += pgmoved;
1137                         pgmoved = 0;
1138                         if (buffer_heads_over_limit)
1139                                 pagevec_strip(&pvec);
1140                         __pagevec_release(&pvec);
1141                         spin_lock_irq(&zone->lru_lock);
1142                 }
1143         }
1144         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1145         pgdeactivate += pgmoved;
1146         if (buffer_heads_over_limit) {
1147                 spin_unlock_irq(&zone->lru_lock);
1148                 pagevec_strip(&pvec);
1149                 spin_lock_irq(&zone->lru_lock);
1150         }
1151
1152         pgmoved = 0;
1153         while (!list_empty(&l_active)) {
1154                 page = lru_to_page(&l_active);
1155                 prefetchw_prev_lru_page(page, &l_active, flags);
1156                 VM_BUG_ON(PageLRU(page));
1157                 SetPageLRU(page);
1158                 VM_BUG_ON(!PageActive(page));
1159                 list_move(&page->lru, &zone->active_list);
1160                 mem_cgroup_move_lists(page_get_page_cgroup(page), true);
1161                 pgmoved++;
1162                 if (!pagevec_add(&pvec, page)) {
1163                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1164                         pgmoved = 0;
1165                         spin_unlock_irq(&zone->lru_lock);
1166                         __pagevec_release(&pvec);
1167                         spin_lock_irq(&zone->lru_lock);
1168                 }
1169         }
1170         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1171
1172         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1173         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1174         spin_unlock_irq(&zone->lru_lock);
1175
1176         pagevec_release(&pvec);
1177 }
1178
1179 /*
1180  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1181  */
1182 static unsigned long shrink_zone(int priority, struct zone *zone,
1183                                 struct scan_control *sc)
1184 {
1185         unsigned long nr_active;
1186         unsigned long nr_inactive;
1187         unsigned long nr_to_scan;
1188         unsigned long nr_reclaimed = 0;
1189
1190         if (scan_global_lru(sc)) {
1191                 /*
1192                  * Add one to nr_to_scan just to make sure that the kernel
1193                  * will slowly sift through the active list.
1194                  */
1195                 zone->nr_scan_active +=
1196                         (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1197                 nr_active = zone->nr_scan_active;
1198                 zone->nr_scan_inactive +=
1199                         (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1200                 nr_inactive = zone->nr_scan_inactive;
1201                 if (nr_inactive >= sc->swap_cluster_max)
1202                         zone->nr_scan_inactive = 0;
1203                 else
1204                         nr_inactive = 0;
1205
1206                 if (nr_active >= sc->swap_cluster_max)
1207                         zone->nr_scan_active = 0;
1208                 else
1209                         nr_active = 0;
1210         } else {
1211                 /*
1212                  * This reclaim occurs not because zone memory shortage but
1213                  * because memory controller hits its limit.
1214                  * Then, don't modify zone reclaim related data.
1215                  */
1216                 nr_active = mem_cgroup_calc_reclaim_active(sc->mem_cgroup,
1217                                         zone, priority);
1218
1219                 nr_inactive = mem_cgroup_calc_reclaim_inactive(sc->mem_cgroup,
1220                                         zone, priority);
1221         }
1222
1223
1224         while (nr_active || nr_inactive) {
1225                 if (nr_active) {
1226                         nr_to_scan = min(nr_active,
1227                                         (unsigned long)sc->swap_cluster_max);
1228                         nr_active -= nr_to_scan;
1229                         shrink_active_list(nr_to_scan, zone, sc, priority);
1230                 }
1231
1232                 if (nr_inactive) {
1233                         nr_to_scan = min(nr_inactive,
1234                                         (unsigned long)sc->swap_cluster_max);
1235                         nr_inactive -= nr_to_scan;
1236                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1237                                                                 sc);
1238                 }
1239         }
1240
1241         throttle_vm_writeout(sc->gfp_mask);
1242         return nr_reclaimed;
1243 }
1244
1245 /*
1246  * This is the direct reclaim path, for page-allocating processes.  We only
1247  * try to reclaim pages from zones which will satisfy the caller's allocation
1248  * request.
1249  *
1250  * We reclaim from a zone even if that zone is over pages_high.  Because:
1251  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1252  *    allocation or
1253  * b) The zones may be over pages_high but they must go *over* pages_high to
1254  *    satisfy the `incremental min' zone defense algorithm.
1255  *
1256  * Returns the number of reclaimed pages.
1257  *
1258  * If a zone is deemed to be full of pinned pages then just give it a light
1259  * scan then give up on it.
1260  */
1261 static unsigned long shrink_zones(int priority, struct zone **zones,
1262                                         struct scan_control *sc)
1263 {
1264         unsigned long nr_reclaimed = 0;
1265         int i;
1266
1267
1268         sc->all_unreclaimable = 1;
1269         for (i = 0; zones[i] != NULL; i++) {
1270                 struct zone *zone = zones[i];
1271
1272                 if (!populated_zone(zone))
1273                         continue;
1274                 /*
1275                  * Take care memory controller reclaiming has small influence
1276                  * to global LRU.
1277                  */
1278                 if (scan_global_lru(sc)) {
1279                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1280                                 continue;
1281                         note_zone_scanning_priority(zone, priority);
1282
1283                         if (zone_is_all_unreclaimable(zone) &&
1284                                                 priority != DEF_PRIORITY)
1285                                 continue;       /* Let kswapd poll it */
1286                         sc->all_unreclaimable = 0;
1287                 } else {
1288                         /*
1289                          * Ignore cpuset limitation here. We just want to reduce
1290                          * # of used pages by us regardless of memory shortage.
1291                          */
1292                         sc->all_unreclaimable = 0;
1293                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1294                                                         priority);
1295                 }
1296
1297                 nr_reclaimed += shrink_zone(priority, zone, sc);
1298         }
1299
1300         return nr_reclaimed;
1301 }
1302
1303 /*
1304  * This is the main entry point to direct page reclaim.
1305  *
1306  * If a full scan of the inactive list fails to free enough memory then we
1307  * are "out of memory" and something needs to be killed.
1308  *
1309  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1310  * high - the zone may be full of dirty or under-writeback pages, which this
1311  * caller can't do much about.  We kick pdflush and take explicit naps in the
1312  * hope that some of these pages can be written.  But if the allocating task
1313  * holds filesystem locks which prevent writeout this might not work, and the
1314  * allocation attempt will fail.
1315  */
1316 static unsigned long do_try_to_free_pages(struct zone **zones, gfp_t gfp_mask,
1317                                           struct scan_control *sc)
1318 {
1319         int priority;
1320         int ret = 0;
1321         unsigned long total_scanned = 0;
1322         unsigned long nr_reclaimed = 0;
1323         struct reclaim_state *reclaim_state = current->reclaim_state;
1324         unsigned long lru_pages = 0;
1325         int i;
1326
1327         if (scan_global_lru(sc))
1328                 count_vm_event(ALLOCSTALL);
1329         /*
1330          * mem_cgroup will not do shrink_slab.
1331          */
1332         if (scan_global_lru(sc)) {
1333                 for (i = 0; zones[i] != NULL; i++) {
1334                         struct zone *zone = zones[i];
1335
1336                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1337                                 continue;
1338
1339                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1340                                         + zone_page_state(zone, NR_INACTIVE);
1341                 }
1342         }
1343
1344         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1345                 sc->nr_scanned = 0;
1346                 sc->nr_io_pages = 0;
1347                 if (!priority)
1348                         disable_swap_token();
1349                 nr_reclaimed += shrink_zones(priority, zones, sc);
1350                 /*
1351                  * Don't shrink slabs when reclaiming memory from
1352                  * over limit cgroups
1353                  */
1354                 if (scan_global_lru(sc)) {
1355                         shrink_slab(sc->nr_scanned, gfp_mask, lru_pages);
1356                         if (reclaim_state) {
1357                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1358                                 reclaim_state->reclaimed_slab = 0;
1359                         }
1360                 }
1361                 total_scanned += sc->nr_scanned;
1362                 if (nr_reclaimed >= sc->swap_cluster_max) {
1363                         ret = 1;
1364                         goto out;
1365                 }
1366
1367                 /*
1368                  * Try to write back as many pages as we just scanned.  This
1369                  * tends to cause slow streaming writers to write data to the
1370                  * disk smoothly, at the dirtying rate, which is nice.   But
1371                  * that's undesirable in laptop mode, where we *want* lumpy
1372                  * writeout.  So in laptop mode, write out the whole world.
1373                  */
1374                 if (total_scanned > sc->swap_cluster_max +
1375                                         sc->swap_cluster_max / 2) {
1376                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1377                         sc->may_writepage = 1;
1378                 }
1379
1380                 /* Take a nap, wait for some writeback to complete */
1381                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2 &&
1382                                 sc->nr_io_pages > sc->swap_cluster_max)
1383                         congestion_wait(WRITE, HZ/10);
1384         }
1385         /* top priority shrink_caches still had more to do? don't OOM, then */
1386         if (!sc->all_unreclaimable && scan_global_lru(sc))
1387                 ret = 1;
1388 out:
1389         /*
1390          * Now that we've scanned all the zones at this priority level, note
1391          * that level within the zone so that the next thread which performs
1392          * scanning of this zone will immediately start out at this priority
1393          * level.  This affects only the decision whether or not to bring
1394          * mapped pages onto the inactive list.
1395          */
1396         if (priority < 0)
1397                 priority = 0;
1398
1399         if (scan_global_lru(sc)) {
1400                 for (i = 0; zones[i] != NULL; i++) {
1401                         struct zone *zone = zones[i];
1402
1403                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1404                                 continue;
1405
1406                         zone->prev_priority = priority;
1407                 }
1408         } else
1409                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1410
1411         return ret;
1412 }
1413
1414 unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1415 {
1416         struct scan_control sc = {
1417                 .gfp_mask = gfp_mask,
1418                 .may_writepage = !laptop_mode,
1419                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1420                 .may_swap = 1,
1421                 .swappiness = vm_swappiness,
1422                 .order = order,
1423                 .mem_cgroup = NULL,
1424                 .isolate_pages = isolate_pages_global,
1425         };
1426
1427         return do_try_to_free_pages(zones, gfp_mask, &sc);
1428 }
1429
1430 #ifdef CONFIG_CGROUP_MEM_CONT
1431
1432 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1433                                                 gfp_t gfp_mask)
1434 {
1435         struct scan_control sc = {
1436                 .gfp_mask = gfp_mask,
1437                 .may_writepage = !laptop_mode,
1438                 .may_swap = 1,
1439                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1440                 .swappiness = vm_swappiness,
1441                 .order = 0,
1442                 .mem_cgroup = mem_cont,
1443                 .isolate_pages = mem_cgroup_isolate_pages,
1444         };
1445         struct zone **zones;
1446         int target_zone = gfp_zone(GFP_HIGHUSER_MOVABLE);
1447
1448         zones = NODE_DATA(numa_node_id())->node_zonelists[target_zone].zones;
1449         if (do_try_to_free_pages(zones, sc.gfp_mask, &sc))
1450                 return 1;
1451         return 0;
1452 }
1453 #endif
1454
1455 /*
1456  * For kswapd, balance_pgdat() will work across all this node's zones until
1457  * they are all at pages_high.
1458  *
1459  * Returns the number of pages which were actually freed.
1460  *
1461  * There is special handling here for zones which are full of pinned pages.
1462  * This can happen if the pages are all mlocked, or if they are all used by
1463  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1464  * What we do is to detect the case where all pages in the zone have been
1465  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1466  * dead and from now on, only perform a short scan.  Basically we're polling
1467  * the zone for when the problem goes away.
1468  *
1469  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1470  * zones which have free_pages > pages_high, but once a zone is found to have
1471  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1472  * of the number of free pages in the lower zones.  This interoperates with
1473  * the page allocator fallback scheme to ensure that aging of pages is balanced
1474  * across the zones.
1475  */
1476 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1477 {
1478         int all_zones_ok;
1479         int priority;
1480         int i;
1481         unsigned long total_scanned;
1482         unsigned long nr_reclaimed;
1483         struct reclaim_state *reclaim_state = current->reclaim_state;
1484         struct scan_control sc = {
1485                 .gfp_mask = GFP_KERNEL,
1486                 .may_swap = 1,
1487                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1488                 .swappiness = vm_swappiness,
1489                 .order = order,
1490                 .mem_cgroup = NULL,
1491                 .isolate_pages = isolate_pages_global,
1492         };
1493         /*
1494          * temp_priority is used to remember the scanning priority at which
1495          * this zone was successfully refilled to free_pages == pages_high.
1496          */
1497         int temp_priority[MAX_NR_ZONES];
1498
1499 loop_again:
1500         total_scanned = 0;
1501         nr_reclaimed = 0;
1502         sc.may_writepage = !laptop_mode;
1503         count_vm_event(PAGEOUTRUN);
1504
1505         for (i = 0; i < pgdat->nr_zones; i++)
1506                 temp_priority[i] = DEF_PRIORITY;
1507
1508         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1509                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1510                 unsigned long lru_pages = 0;
1511
1512                 /* The swap token gets in the way of swapout... */
1513                 if (!priority)
1514                         disable_swap_token();
1515
1516                 sc.nr_io_pages = 0;
1517                 all_zones_ok = 1;
1518
1519                 /*
1520                  * Scan in the highmem->dma direction for the highest
1521                  * zone which needs scanning
1522                  */
1523                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1524                         struct zone *zone = pgdat->node_zones + i;
1525
1526                         if (!populated_zone(zone))
1527                                 continue;
1528
1529                         if (zone_is_all_unreclaimable(zone) &&
1530                             priority != DEF_PRIORITY)
1531                                 continue;
1532
1533                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1534                                                0, 0)) {
1535                                 end_zone = i;
1536                                 break;
1537                         }
1538                 }
1539                 if (i < 0)
1540                         goto out;
1541
1542                 for (i = 0; i <= end_zone; i++) {
1543                         struct zone *zone = pgdat->node_zones + i;
1544
1545                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1546                                         + zone_page_state(zone, NR_INACTIVE);
1547                 }
1548
1549                 /*
1550                  * Now scan the zone in the dma->highmem direction, stopping
1551                  * at the last zone which needs scanning.
1552                  *
1553                  * We do this because the page allocator works in the opposite
1554                  * direction.  This prevents the page allocator from allocating
1555                  * pages behind kswapd's direction of progress, which would
1556                  * cause too much scanning of the lower zones.
1557                  */
1558                 for (i = 0; i <= end_zone; i++) {
1559                         struct zone *zone = pgdat->node_zones + i;
1560                         int nr_slab;
1561
1562                         if (!populated_zone(zone))
1563                                 continue;
1564
1565                         if (zone_is_all_unreclaimable(zone) &&
1566                                         priority != DEF_PRIORITY)
1567                                 continue;
1568
1569                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1570                                                end_zone, 0))
1571                                 all_zones_ok = 0;
1572                         temp_priority[i] = priority;
1573                         sc.nr_scanned = 0;
1574                         note_zone_scanning_priority(zone, priority);
1575                         /*
1576                          * We put equal pressure on every zone, unless one
1577                          * zone has way too many pages free already.
1578                          */
1579                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1580                                                 end_zone, 0))
1581                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1582                         reclaim_state->reclaimed_slab = 0;
1583                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1584                                                 lru_pages);
1585                         nr_reclaimed += reclaim_state->reclaimed_slab;
1586                         total_scanned += sc.nr_scanned;
1587                         if (zone_is_all_unreclaimable(zone))
1588                                 continue;
1589                         if (nr_slab == 0 && zone->pages_scanned >=
1590                                 (zone_page_state(zone, NR_ACTIVE)
1591                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1592                                         zone_set_flag(zone,
1593                                                       ZONE_ALL_UNRECLAIMABLE);
1594                         /*
1595                          * If we've done a decent amount of scanning and
1596                          * the reclaim ratio is low, start doing writepage
1597                          * even in laptop mode
1598                          */
1599                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1600                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1601                                 sc.may_writepage = 1;
1602                 }
1603                 if (all_zones_ok)
1604                         break;          /* kswapd: all done */
1605                 /*
1606                  * OK, kswapd is getting into trouble.  Take a nap, then take
1607                  * another pass across the zones.
1608                  */
1609                 if (total_scanned && priority < DEF_PRIORITY - 2 &&
1610                                         sc.nr_io_pages > sc.swap_cluster_max)
1611                         congestion_wait(WRITE, HZ/10);
1612
1613                 /*
1614                  * We do this so kswapd doesn't build up large priorities for
1615                  * example when it is freeing in parallel with allocators. It
1616                  * matches the direct reclaim path behaviour in terms of impact
1617                  * on zone->*_priority.
1618                  */
1619                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1620                         break;
1621         }
1622 out:
1623         /*
1624          * Note within each zone the priority level at which this zone was
1625          * brought into a happy state.  So that the next thread which scans this
1626          * zone will start out at that priority level.
1627          */
1628         for (i = 0; i < pgdat->nr_zones; i++) {
1629                 struct zone *zone = pgdat->node_zones + i;
1630
1631                 zone->prev_priority = temp_priority[i];
1632         }
1633         if (!all_zones_ok) {
1634                 cond_resched();
1635
1636                 try_to_freeze();
1637
1638                 goto loop_again;
1639         }
1640
1641         return nr_reclaimed;
1642 }
1643
1644 /*
1645  * The background pageout daemon, started as a kernel thread
1646  * from the init process.
1647  *
1648  * This basically trickles out pages so that we have _some_
1649  * free memory available even if there is no other activity
1650  * that frees anything up. This is needed for things like routing
1651  * etc, where we otherwise might have all activity going on in
1652  * asynchronous contexts that cannot page things out.
1653  *
1654  * If there are applications that are active memory-allocators
1655  * (most normal use), this basically shouldn't matter.
1656  */
1657 static int kswapd(void *p)
1658 {
1659         unsigned long order;
1660         pg_data_t *pgdat = (pg_data_t*)p;
1661         struct task_struct *tsk = current;
1662         DEFINE_WAIT(wait);
1663         struct reclaim_state reclaim_state = {
1664                 .reclaimed_slab = 0,
1665         };
1666         cpumask_t cpumask;
1667
1668         cpumask = node_to_cpumask(pgdat->node_id);
1669         if (!cpus_empty(cpumask))
1670                 set_cpus_allowed(tsk, cpumask);
1671         current->reclaim_state = &reclaim_state;
1672
1673         /*
1674          * Tell the memory management that we're a "memory allocator",
1675          * and that if we need more memory we should get access to it
1676          * regardless (see "__alloc_pages()"). "kswapd" should
1677          * never get caught in the normal page freeing logic.
1678          *
1679          * (Kswapd normally doesn't need memory anyway, but sometimes
1680          * you need a small amount of memory in order to be able to
1681          * page out something else, and this flag essentially protects
1682          * us from recursively trying to free more memory as we're
1683          * trying to free the first piece of memory in the first place).
1684          */
1685         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1686         set_freezable();
1687
1688         order = 0;
1689         for ( ; ; ) {
1690                 unsigned long new_order;
1691
1692                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1693                 new_order = pgdat->kswapd_max_order;
1694                 pgdat->kswapd_max_order = 0;
1695                 if (order < new_order) {
1696                         /*
1697                          * Don't sleep if someone wants a larger 'order'
1698                          * allocation
1699                          */
1700                         order = new_order;
1701                 } else {
1702                         if (!freezing(current))
1703                                 schedule();
1704
1705                         order = pgdat->kswapd_max_order;
1706                 }
1707                 finish_wait(&pgdat->kswapd_wait, &wait);
1708
1709                 if (!try_to_freeze()) {
1710                         /* We can speed up thawing tasks if we don't call
1711                          * balance_pgdat after returning from the refrigerator
1712                          */
1713                         balance_pgdat(pgdat, order);
1714                 }
1715         }
1716         return 0;
1717 }
1718
1719 /*
1720  * A zone is low on free memory, so wake its kswapd task to service it.
1721  */
1722 void wakeup_kswapd(struct zone *zone, int order)
1723 {
1724         pg_data_t *pgdat;
1725
1726         if (!populated_zone(zone))
1727                 return;
1728
1729         pgdat = zone->zone_pgdat;
1730         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1731                 return;
1732         if (pgdat->kswapd_max_order < order)
1733                 pgdat->kswapd_max_order = order;
1734         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1735                 return;
1736         if (!waitqueue_active(&pgdat->kswapd_wait))
1737                 return;
1738         wake_up_interruptible(&pgdat->kswapd_wait);
1739 }
1740
1741 #ifdef CONFIG_PM
1742 /*
1743  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1744  * from LRU lists system-wide, for given pass and priority, and returns the
1745  * number of reclaimed pages
1746  *
1747  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1748  */
1749 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1750                                       int pass, struct scan_control *sc)
1751 {
1752         struct zone *zone;
1753         unsigned long nr_to_scan, ret = 0;
1754
1755         for_each_zone(zone) {
1756
1757                 if (!populated_zone(zone))
1758                         continue;
1759
1760                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1761                         continue;
1762
1763                 /* For pass = 0 we don't shrink the active list */
1764                 if (pass > 0) {
1765                         zone->nr_scan_active +=
1766                                 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1767                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1768                                 zone->nr_scan_active = 0;
1769                                 nr_to_scan = min(nr_pages,
1770                                         zone_page_state(zone, NR_ACTIVE));
1771                                 shrink_active_list(nr_to_scan, zone, sc, prio);
1772                         }
1773                 }
1774
1775                 zone->nr_scan_inactive +=
1776                         (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1777                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1778                         zone->nr_scan_inactive = 0;
1779                         nr_to_scan = min(nr_pages,
1780                                 zone_page_state(zone, NR_INACTIVE));
1781                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1782                         if (ret >= nr_pages)
1783                                 return ret;
1784                 }
1785         }
1786
1787         return ret;
1788 }
1789
1790 static unsigned long count_lru_pages(void)
1791 {
1792         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1793 }
1794
1795 /*
1796  * Try to free `nr_pages' of memory, system-wide, and return the number of
1797  * freed pages.
1798  *
1799  * Rather than trying to age LRUs the aim is to preserve the overall
1800  * LRU order by reclaiming preferentially
1801  * inactive > active > active referenced > active mapped
1802  */
1803 unsigned long shrink_all_memory(unsigned long nr_pages)
1804 {
1805         unsigned long lru_pages, nr_slab;
1806         unsigned long ret = 0;
1807         int pass;
1808         struct reclaim_state reclaim_state;
1809         struct scan_control sc = {
1810                 .gfp_mask = GFP_KERNEL,
1811                 .may_swap = 0,
1812                 .swap_cluster_max = nr_pages,
1813                 .may_writepage = 1,
1814                 .swappiness = vm_swappiness,
1815                 .isolate_pages = isolate_pages_global,
1816         };
1817
1818         current->reclaim_state = &reclaim_state;
1819
1820         lru_pages = count_lru_pages();
1821         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1822         /* If slab caches are huge, it's better to hit them first */
1823         while (nr_slab >= lru_pages) {
1824                 reclaim_state.reclaimed_slab = 0;
1825                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1826                 if (!reclaim_state.reclaimed_slab)
1827                         break;
1828
1829                 ret += reclaim_state.reclaimed_slab;
1830                 if (ret >= nr_pages)
1831                         goto out;
1832
1833                 nr_slab -= reclaim_state.reclaimed_slab;
1834         }
1835
1836         /*
1837          * We try to shrink LRUs in 5 passes:
1838          * 0 = Reclaim from inactive_list only
1839          * 1 = Reclaim from active list but don't reclaim mapped
1840          * 2 = 2nd pass of type 1
1841          * 3 = Reclaim mapped (normal reclaim)
1842          * 4 = 2nd pass of type 3
1843          */
1844         for (pass = 0; pass < 5; pass++) {
1845                 int prio;
1846
1847                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1848                 if (pass > 2) {
1849                         sc.may_swap = 1;
1850                         sc.swappiness = 100;
1851                 }
1852
1853                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1854                         unsigned long nr_to_scan = nr_pages - ret;
1855
1856                         sc.nr_scanned = 0;
1857                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1858                         if (ret >= nr_pages)
1859                                 goto out;
1860
1861                         reclaim_state.reclaimed_slab = 0;
1862                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1863                                         count_lru_pages());
1864                         ret += reclaim_state.reclaimed_slab;
1865                         if (ret >= nr_pages)
1866                                 goto out;
1867
1868                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1869                                 congestion_wait(WRITE, HZ / 10);
1870                 }
1871         }
1872
1873         /*
1874          * If ret = 0, we could not shrink LRUs, but there may be something
1875          * in slab caches
1876          */
1877         if (!ret) {
1878                 do {
1879                         reclaim_state.reclaimed_slab = 0;
1880                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1881                         ret += reclaim_state.reclaimed_slab;
1882                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1883         }
1884
1885 out:
1886         current->reclaim_state = NULL;
1887
1888         return ret;
1889 }
1890 #endif
1891
1892 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1893    not required for correctness.  So if the last cpu in a node goes
1894    away, we get changed to run anywhere: as the first one comes back,
1895    restore their cpu bindings. */
1896 static int __devinit cpu_callback(struct notifier_block *nfb,
1897                                   unsigned long action, void *hcpu)
1898 {
1899         pg_data_t *pgdat;
1900         cpumask_t mask;
1901         int nid;
1902
1903         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1904                 for_each_node_state(nid, N_HIGH_MEMORY) {
1905                         pgdat = NODE_DATA(nid);
1906                         mask = node_to_cpumask(pgdat->node_id);
1907                         if (any_online_cpu(mask) != NR_CPUS)
1908                                 /* One of our CPUs online: restore mask */
1909                                 set_cpus_allowed(pgdat->kswapd, mask);
1910                 }
1911         }
1912         return NOTIFY_OK;
1913 }
1914
1915 /*
1916  * This kswapd start function will be called by init and node-hot-add.
1917  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1918  */
1919 int kswapd_run(int nid)
1920 {
1921         pg_data_t *pgdat = NODE_DATA(nid);
1922         int ret = 0;
1923
1924         if (pgdat->kswapd)
1925                 return 0;
1926
1927         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1928         if (IS_ERR(pgdat->kswapd)) {
1929                 /* failure at boot is fatal */
1930                 BUG_ON(system_state == SYSTEM_BOOTING);
1931                 printk("Failed to start kswapd on node %d\n",nid);
1932                 ret = -1;
1933         }
1934         return ret;
1935 }
1936
1937 static int __init kswapd_init(void)
1938 {
1939         int nid;
1940
1941         swap_setup();
1942         for_each_node_state(nid, N_HIGH_MEMORY)
1943                 kswapd_run(nid);
1944         hotcpu_notifier(cpu_callback, 0);
1945         return 0;
1946 }
1947
1948 module_init(kswapd_init)
1949
1950 #ifdef CONFIG_NUMA
1951 /*
1952  * Zone reclaim mode
1953  *
1954  * If non-zero call zone_reclaim when the number of free pages falls below
1955  * the watermarks.
1956  */
1957 int zone_reclaim_mode __read_mostly;
1958
1959 #define RECLAIM_OFF 0
1960 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1961 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1962 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1963
1964 /*
1965  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1966  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1967  * a zone.
1968  */
1969 #define ZONE_RECLAIM_PRIORITY 4
1970
1971 /*
1972  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1973  * occur.
1974  */
1975 int sysctl_min_unmapped_ratio = 1;
1976
1977 /*
1978  * If the number of slab pages in a zone grows beyond this percentage then
1979  * slab reclaim needs to occur.
1980  */
1981 int sysctl_min_slab_ratio = 5;
1982
1983 /*
1984  * Try to free up some pages from this zone through reclaim.
1985  */
1986 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1987 {
1988         /* Minimum pages needed in order to stay on node */
1989         const unsigned long nr_pages = 1 << order;
1990         struct task_struct *p = current;
1991         struct reclaim_state reclaim_state;
1992         int priority;
1993         unsigned long nr_reclaimed = 0;
1994         struct scan_control sc = {
1995                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1996                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1997                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1998                                         SWAP_CLUSTER_MAX),
1999                 .gfp_mask = gfp_mask,
2000                 .swappiness = vm_swappiness,
2001                 .isolate_pages = isolate_pages_global,
2002         };
2003         unsigned long slab_reclaimable;
2004
2005         disable_swap_token();
2006         cond_resched();
2007         /*
2008          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2009          * and we also need to be able to write out pages for RECLAIM_WRITE
2010          * and RECLAIM_SWAP.
2011          */
2012         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2013         reclaim_state.reclaimed_slab = 0;
2014         p->reclaim_state = &reclaim_state;
2015
2016         if (zone_page_state(zone, NR_FILE_PAGES) -
2017                 zone_page_state(zone, NR_FILE_MAPPED) >
2018                 zone->min_unmapped_pages) {
2019                 /*
2020                  * Free memory by calling shrink zone with increasing
2021                  * priorities until we have enough memory freed.
2022                  */
2023                 priority = ZONE_RECLAIM_PRIORITY;
2024                 do {
2025                         note_zone_scanning_priority(zone, priority);
2026                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2027                         priority--;
2028                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2029         }
2030
2031         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2032         if (slab_reclaimable > zone->min_slab_pages) {
2033                 /*
2034                  * shrink_slab() does not currently allow us to determine how
2035                  * many pages were freed in this zone. So we take the current
2036                  * number of slab pages and shake the slab until it is reduced
2037                  * by the same nr_pages that we used for reclaiming unmapped
2038                  * pages.
2039                  *
2040                  * Note that shrink_slab will free memory on all zones and may
2041                  * take a long time.
2042                  */
2043                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2044                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2045                                 slab_reclaimable - nr_pages)
2046                         ;
2047
2048                 /*
2049                  * Update nr_reclaimed by the number of slab pages we
2050                  * reclaimed from this zone.
2051                  */
2052                 nr_reclaimed += slab_reclaimable -
2053                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2054         }
2055
2056         p->reclaim_state = NULL;
2057         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2058         return nr_reclaimed >= nr_pages;
2059 }
2060
2061 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2062 {
2063         int node_id;
2064         int ret;
2065
2066         /*
2067          * Zone reclaim reclaims unmapped file backed pages and
2068          * slab pages if we are over the defined limits.
2069          *
2070          * A small portion of unmapped file backed pages is needed for
2071          * file I/O otherwise pages read by file I/O will be immediately
2072          * thrown out if the zone is overallocated. So we do not reclaim
2073          * if less than a specified percentage of the zone is used by
2074          * unmapped file backed pages.
2075          */
2076         if (zone_page_state(zone, NR_FILE_PAGES) -
2077             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2078             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2079                         <= zone->min_slab_pages)
2080                 return 0;
2081
2082         if (zone_is_all_unreclaimable(zone))
2083                 return 0;
2084
2085         /*
2086          * Do not scan if the allocation should not be delayed.
2087          */
2088         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2089                         return 0;
2090
2091         /*
2092          * Only run zone reclaim on the local zone or on zones that do not
2093          * have associated processors. This will favor the local processor
2094          * over remote processors and spread off node memory allocations
2095          * as wide as possible.
2096          */
2097         node_id = zone_to_nid(zone);
2098         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2099                 return 0;
2100
2101         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2102                 return 0;
2103         ret = __zone_reclaim(zone, gfp_mask, order);
2104         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2105
2106         return ret;
2107 }
2108 #endif