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 #include <linux/delayacct.h>
  42
  43 #include <asm/tlbflush.h>
  44 #include <asm/div64.h>
  45
  46 #include <linux/swapops.h>
  47
  48 #include "internal.h"
  49
  50 struct scan_control {
  51         /* Incremented by the number of inactive pages that were scanned */
  52         unsigned long nr_scanned;
  53
  54         /* This context's GFP mask */
  55         gfp_t gfp_mask;
  56
  57         int may_writepage;
  58
  59         /* Can pages be swapped as part of reclaim? */
  60         int may_swap;
  61
  62         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
  63          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
  64          * In this context, it doesn't matter that we scan the
  65          * whole list at once. */
  66         int swap_cluster_max;
  67
  68         int swappiness;
  69
  70         int all_unreclaimable;
  71
  72         int order;
  73
  74         /* Which cgroup do we reclaim from */
  75         struct mem_cgroup *mem_cgroup;
  76
  77         /* Pluggable isolate pages callback */
  78         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
  79                         unsigned long *scanned, int order, int mode,
  80                         struct zone *z, struct mem_cgroup *mem_cont,
  81                         int active);
  82 };
  83
  84 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  85
  86 #ifdef ARCH_HAS_PREFETCH
  87 #define prefetch_prev_lru_page(_page, _base, _field)                    \
  88         do {                                                            \
  89                 if ((_page)->lru.prev != _base) {                       \
  90                         struct page *prev;                              \
  91                                                                         \
  92                         prev = lru_to_page(&(_page->lru));              \
  93                         prefetch(&prev->_field);                        \
  94                 }                                                       \
  95         } while (0)
  96 #else
  97 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
  98 #endif
  99
 100 #ifdef ARCH_HAS_PREFETCHW
 101 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
 102         do {                                                            \
 103                 if ((_page)->lru.prev != _base) {                       \
 104                         struct page *prev;                              \
 105                                                                         \
 106                         prev = lru_to_page(&(_page->lru));              \
 107                         prefetchw(&prev->_field);                       \
 108                 }                                                       \
 109         } while (0)
 110 #else
 111 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 112 #endif
 113
 114 /*
 115  * From 0 .. 100.  Higher means more swappy.
 116  */
 117 int vm_swappiness = 60;
 118 long vm_total_pages;    /* The total number of pages which the VM controls */
 119
 120 static LIST_HEAD(shrinker_list);
 121 static DECLARE_RWSEM(shrinker_rwsem);
 122
 123 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 124 #define scan_global_lru(sc)     (!(sc)->mem_cgroup)
 125 #else
 126 #define scan_global_lru(sc)     (1)
 127 #endif
 128
 129 /*
 130  * Add a shrinker callback to be called from the vm
 131  */
 132 void register_shrinker(struct shrinker *shrinker)
 133 {
 134         shrinker->nr = 0;
 135         down_write(&shrinker_rwsem);
 136         list_add_tail(&shrinker->list, &shrinker_list);
 137         up_write(&shrinker_rwsem);
 138 }
 139 EXPORT_SYMBOL(register_shrinker);
 140
 141 /*
 142  * Remove one
 143  */
 144 void unregister_shrinker(struct shrinker *shrinker)
 145 {
 146         down_write(&shrinker_rwsem);
 147         list_del(&shrinker->list);
 148         up_write(&shrinker_rwsem);
 149 }
 150 EXPORT_SYMBOL(unregister_shrinker);
 151
 152 #define SHRINK_BATCH 128
 153 /*
 154  * Call the shrink functions to age shrinkable caches
 155  *
 156  * Here we assume it costs one seek to replace a lru page and that it also
 157  * takes a seek to recreate a cache object.  With this in mind we age equal
 158  * percentages of the lru and ageable caches.  This should balance the seeks
 159  * generated by these structures.
 160  *
 161  * If the vm encountered mapped pages on the LRU it increase the pressure on
 162  * slab to avoid swapping.
 163  *
 164  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 165  *
 166  * `lru_pages' represents the number of on-LRU pages in all the zones which
 167  * are eligible for the caller's allocation attempt.  It is used for balancing
 168  * slab reclaim versus page reclaim.
 169  *
 170  * Returns the number of slab objects which we shrunk.
 171  */
 172 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
 173                         unsigned long lru_pages)
 174 {
 175         struct shrinker *shrinker;
 176         unsigned long ret = 0;
 177
 178         if (scanned == 0)
 179                 scanned = SWAP_CLUSTER_MAX;
 180
 181         if (!down_read_trylock(&shrinker_rwsem))
 182                 return 1;       /* Assume we'll be able to shrink next time */
 183
 184         list_for_each_entry(shrinker, &shrinker_list, list) {
 185                 unsigned long long delta;
 186                 unsigned long total_scan;
 187                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
 188
 189                 delta = (4 * scanned) / shrinker->seeks;
 190                 delta *= max_pass;
 191                 do_div(delta, lru_pages + 1);
 192                 shrinker->nr += delta;
 193                 if (shrinker->nr < 0) {
 194                         printk(KERN_ERR "%s: nr=%ld\n",
 195                                         __func__, shrinker->nr);
 196                         shrinker->nr = max_pass;
 197                 }
 198
 199                 /*
 200                  * Avoid risking looping forever due to too large nr value:
 201                  * never try to free more than twice the estimate number of
 202                  * freeable entries.
 203                  */
 204                 if (shrinker->nr > max_pass * 2)
 205                         shrinker->nr = max_pass * 2;
 206
 207                 total_scan = shrinker->nr;
 208                 shrinker->nr = 0;
 209
 210                 while (total_scan >= SHRINK_BATCH) {
 211                         long this_scan = SHRINK_BATCH;
 212                         int shrink_ret;
 213                         int nr_before;
 214
 215                         nr_before = (*shrinker->shrink)(0, gfp_mask);
 216                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
 217                         if (shrink_ret == -1)
 218                                 break;
 219                         if (shrink_ret < nr_before)
 220                                 ret += nr_before - shrink_ret;
 221                         count_vm_events(SLABS_SCANNED, this_scan);
 222                         total_scan -= this_scan;
 223
 224                         cond_resched();
 225                 }
 226
 227                 shrinker->nr += total_scan;
 228         }
 229         up_read(&shrinker_rwsem);
 230         return ret;
 231 }
 232
 233 /* Called without lock on whether page is mapped, so answer is unstable */
 234 static inline int page_mapping_inuse(struct page *page)
 235 {
 236         struct address_space *mapping;
 237
 238         /* Page is in somebody's page tables. */
 239         if (page_mapped(page))
 240                 return 1;
 241
 242         /* Be more reluctant to reclaim swapcache than pagecache */
 243         if (PageSwapCache(page))
 244                 return 1;
 245
 246         mapping = page_mapping(page);
 247         if (!mapping)
 248                 return 0;
 249
 250         /* File is mmap'd by somebody? */
 251         return mapping_mapped(mapping);
 252 }
 253
 254 static inline int is_page_cache_freeable(struct page *page)
 255 {
 256         return page_count(page) - !!PagePrivate(page) == 2;
 257 }
 258
 259 static int may_write_to_queue(struct backing_dev_info *bdi)
 260 {
 261         if (current->flags & PF_SWAPWRITE)
 262                 return 1;
 263         if (!bdi_write_congested(bdi))
 264                 return 1;
 265         if (bdi == current->backing_dev_info)
 266                 return 1;
 267         return 0;
 268 }
 269
 270 /*
 271  * We detected a synchronous write error writing a page out.  Probably
 272  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 273  * fsync(), msync() or close().
 274  *
 275  * The tricky part is that after writepage we cannot touch the mapping: nothing
 276  * prevents it from being freed up.  But we have a ref on the page and once
 277  * that page is locked, the mapping is pinned.
 278  *
 279  * We're allowed to run sleeping lock_page() here because we know the caller has
 280  * __GFP_FS.
 281  */
 282 static void handle_write_error(struct address_space *mapping,
 283                                 struct page *page, int error)
 284 {
 285         lock_page(page);
 286         if (page_mapping(page) == mapping)
 287                 mapping_set_error(mapping, error);
 288         unlock_page(page);
 289 }
 290
 291 /* Request for sync pageout. */
 292 enum pageout_io {
 293         PAGEOUT_IO_ASYNC,
 294         PAGEOUT_IO_SYNC,
 295 };
 296
 297 /* possible outcome of pageout() */
 298 typedef enum {
 299         /* failed to write page out, page is locked */
 300         PAGE_KEEP,
 301         /* move page to the active list, page is locked */
 302         PAGE_ACTIVATE,
 303         /* page has been sent to the disk successfully, page is unlocked */
 304         PAGE_SUCCESS,
 305         /* page is clean and locked */
 306         PAGE_CLEAN,
 307 } pageout_t;
 308
 309 /*
 310  * pageout is called by shrink_page_list() for each dirty page.
 311  * Calls ->writepage().
 312  */
 313 static pageout_t pageout(struct page *page, struct address_space *mapping,
 314                                                 enum pageout_io sync_writeback)
 315 {
 316         /*
 317          * If the page is dirty, only perform writeback if that write
 318          * will be non-blocking.  To prevent this allocation from being
 319          * stalled by pagecache activity.  But note that there may be
 320          * stalls if we need to run get_block().  We could test
 321          * PagePrivate for that.
 322          *
 323          * If this process is currently in generic_file_write() against
 324          * this page's queue, we can perform writeback even if that
 325          * will block.
 326          *
 327          * If the page is swapcache, write it back even if that would
 328          * block, for some throttling. This happens by accident, because
 329          * swap_backing_dev_info is bust: it doesn't reflect the
 330          * congestion state of the swapdevs.  Easy to fix, if needed.
 331          * See swapfile.c:page_queue_congested().
 332          */
 333         if (!is_page_cache_freeable(page))
 334                 return PAGE_KEEP;
 335         if (!mapping) {
 336                 /*
 337                  * Some data journaling orphaned pages can have
 338                  * page->mapping == NULL while being dirty with clean buffers.
 339                  */
 340                 if (PagePrivate(page)) {
 341                         if (try_to_free_buffers(page)) {
 342                                 ClearPageDirty(page);
 343                                 printk("%s: orphaned page\n", __func__);
 344                                 return PAGE_CLEAN;
 345                         }
 346                 }
 347                 return PAGE_KEEP;
 348         }
 349         if (mapping->a_ops->writepage == NULL)
 350                 return PAGE_ACTIVATE;
 351         if (!may_write_to_queue(mapping->backing_dev_info))
 352                 return PAGE_KEEP;
 353
 354         if (clear_page_dirty_for_io(page)) {
 355                 int res;
 356                 struct writeback_control wbc = {
 357                         .sync_mode = WB_SYNC_NONE,
 358                         .nr_to_write = SWAP_CLUSTER_MAX,
 359                         .range_start = 0,
 360                         .range_end = LLONG_MAX,
 361                         .nonblocking = 1,
 362                         .for_reclaim = 1,
 363                 };
 364
 365                 SetPageReclaim(page);
 366                 res = mapping->a_ops->writepage(page, &wbc);
 367                 if (res < 0)
 368                         handle_write_error(mapping, page, res);
 369                 if (res == AOP_WRITEPAGE_ACTIVATE) {
 370                         ClearPageReclaim(page);
 371                         return PAGE_ACTIVATE;
 372                 }
 373
 374                 /*
 375                  * Wait on writeback if requested to. This happens when
 376                  * direct reclaiming a large contiguous area and the
 377                  * first attempt to free a range of pages fails.
 378                  */
 379                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
 380                         wait_on_page_writeback(page);
 381
 382                 if (!PageWriteback(page)) {
 383                         /* synchronous write or broken a_ops? */
 384                         ClearPageReclaim(page);
 385                 }
 386                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
 387                 return PAGE_SUCCESS;
 388         }
 389
 390         return PAGE_CLEAN;
 391 }
 392
 393 /*
 394  * Same as remove_mapping, but if the page is removed from the mapping, it
 395  * gets returned with a refcount of 0.
 396  */
 397 static int __remove_mapping(struct address_space *mapping, struct page *page)
 398 {
 399         BUG_ON(!PageLocked(page));
 400         BUG_ON(mapping != page_mapping(page));
 401
 402         spin_lock_irq(&mapping->tree_lock);
 403         /*
 404          * The non racy check for a busy page.
 405          *
 406          * Must be careful with the order of the tests. When someone has
 407          * a ref to the page, it may be possible that they dirty it then
 408          * drop the reference. So if PageDirty is tested before page_count
 409          * here, then the following race may occur:
 410          *
 411          * get_user_pages(&page);
 412          * [user mapping goes away]
 413          * write_to(page);
 414          *                              !PageDirty(page)    [good]
 415          * SetPageDirty(page);
 416          * put_page(page);
 417          *                              !page_count(page)   [good, discard it]
 418          *
 419          * [oops, our write_to data is lost]
 420          *
 421          * Reversing the order of the tests ensures such a situation cannot
 422          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 423          * load is not satisfied before that of page->_count.
 424          *
 425          * Note that if SetPageDirty is always performed via set_page_dirty,
 426          * and thus under tree_lock, then this ordering is not required.
 427          */
 428         if (!page_freeze_refs(page, 2))
 429                 goto cannot_free;
 430         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 431         if (unlikely(PageDirty(page))) {
 432                 page_unfreeze_refs(page, 2);
 433                 goto cannot_free;
 434         }
 435
 436         if (PageSwapCache(page)) {
 437                 swp_entry_t swap = { .val = page_private(page) };
 438                 __delete_from_swap_cache(page);
 439                 spin_unlock_irq(&mapping->tree_lock);
 440                 swap_free(swap);
 441         } else {
 442                 __remove_from_page_cache(page);
 443                 spin_unlock_irq(&mapping->tree_lock);
 444         }
 445
 446         return 1;
 447
 448 cannot_free:
 449         spin_unlock_irq(&mapping->tree_lock);
 450         return 0;
 451 }
 452
 453 /*
 454  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 455  * someone else has a ref on the page, abort and return 0.  If it was
 456  * successfully detached, return 1.  Assumes the caller has a single ref on
 457  * this page.
 458  */
 459 int remove_mapping(struct address_space *mapping, struct page *page)
 460 {
 461         if (__remove_mapping(mapping, page)) {
 462                 /*
 463                  * Unfreezing the refcount with 1 rather than 2 effectively
 464                  * drops the pagecache ref for us without requiring another
 465                  * atomic operation.
 466                  */
 467                 page_unfreeze_refs(page, 1);
 468                 return 1;
 469         }
 470         return 0;
 471 }
 472
 473 /*
 474  * shrink_page_list() returns the number of reclaimed pages
 475  */
 476 static unsigned long shrink_page_list(struct list_head *page_list,
 477                                         struct scan_control *sc,
 478                                         enum pageout_io sync_writeback)
 479 {
 480         LIST_HEAD(ret_pages);
 481         struct pagevec freed_pvec;
 482         int pgactivate = 0;
 483         unsigned long nr_reclaimed = 0;
 484
 485         cond_resched();
 486
 487         pagevec_init(&freed_pvec, 1);
 488         while (!list_empty(page_list)) {
 489                 struct address_space *mapping;
 490                 struct page *page;
 491                 int may_enter_fs;
 492                 int referenced;
 493
 494                 cond_resched();
 495
 496                 page = lru_to_page(page_list);
 497                 list_del(&page->lru);
 498
 499                 if (!trylock_page(page))
 500                         goto keep;
 501
 502                 VM_BUG_ON(PageActive(page));
 503
 504                 sc->nr_scanned++;
 505
 506                 if (!sc->may_swap && page_mapped(page))
 507                         goto keep_locked;
 508
 509                 /* Double the slab pressure for mapped and swapcache pages */
 510                 if (page_mapped(page) || PageSwapCache(page))
 511                         sc->nr_scanned++;
 512
 513                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 514                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 515
 516                 if (PageWriteback(page)) {
 517                         /*
 518                          * Synchronous reclaim is performed in two passes,
 519                          * first an asynchronous pass over the list to
 520                          * start parallel writeback, and a second synchronous
 521                          * pass to wait for the IO to complete.  Wait here
 522                          * for any page for which writeback has already
 523                          * started.
 524                          */
 525                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
 526                                 wait_on_page_writeback(page);
 527                         else
 528                                 goto keep_locked;
 529                 }
 530
 531                 referenced = page_referenced(page, 1, sc->mem_cgroup);
 532                 /* In active use or really unfreeable?  Activate it. */
 533                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
 534                                         referenced && page_mapping_inuse(page))
 535                         goto activate_locked;
 536
 537 #ifdef CONFIG_SWAP
 538                 /*
 539                  * Anonymous process memory has backing store?
 540                  * Try to allocate it some swap space here.
 541                  */
 542                 if (PageAnon(page) && !PageSwapCache(page))
 543                         if (!add_to_swap(page, GFP_ATOMIC))
 544                                 goto activate_locked;
 545 #endif /* CONFIG_SWAP */
 546
 547                 mapping = page_mapping(page);
 548
 549                 /*
 550                  * The page is mapped into the page tables of one or more
 551                  * processes. Try to unmap it here.
 552                  */
 553                 if (page_mapped(page) && mapping) {
 554                         switch (try_to_unmap(page, 0)) {
 555                         case SWAP_FAIL:
 556                                 goto activate_locked;
 557                         case SWAP_AGAIN:
 558                                 goto keep_locked;
 559                         case SWAP_SUCCESS:
 560                                 ; /* try to free the page below */
 561                         }
 562                 }
 563
 564                 if (PageDirty(page)) {
 565                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
 566                                 goto keep_locked;
 567                         if (!may_enter_fs)
 568                                 goto keep_locked;
 569                         if (!sc->may_writepage)
 570                                 goto keep_locked;
 571
 572                         /* Page is dirty, try to write it out here */
 573                         switch (pageout(page, mapping, sync_writeback)) {
 574                         case PAGE_KEEP:
 575                                 goto keep_locked;
 576                         case PAGE_ACTIVATE:
 577                                 goto activate_locked;
 578                         case PAGE_SUCCESS:
 579                                 if (PageWriteback(page) || PageDirty(page))
 580                                         goto keep;
 581                                 /*
 582                                  * A synchronous write - probably a ramdisk.  Go
 583                                  * ahead and try to reclaim the page.
 584                                  */
 585                                 if (!trylock_page(page))
 586                                         goto keep;
 587                                 if (PageDirty(page) || PageWriteback(page))
 588                                         goto keep_locked;
 589                                 mapping = page_mapping(page);
 590                         case PAGE_CLEAN:
 591                                 ; /* try to free the page below */
 592                         }
 593                 }
 594
 595                 /*
 596                  * If the page has buffers, try to free the buffer mappings
 597                  * associated with this page. If we succeed we try to free
 598                  * the page as well.
 599                  *
 600                  * We do this even if the page is PageDirty().
 601                  * try_to_release_page() does not perform I/O, but it is
 602                  * possible for a page to have PageDirty set, but it is actually
 603                  * clean (all its buffers are clean).  This happens if the
 604                  * buffers were written out directly, with submit_bh(). ext3
 605                  * will do this, as well as the blockdev mapping.
 606                  * try_to_release_page() will discover that cleanness and will
 607                  * drop the buffers and mark the page clean - it can be freed.
 608                  *
 609                  * Rarely, pages can have buffers and no ->mapping.  These are
 610                  * the pages which were not successfully invalidated in
 611                  * truncate_complete_page().  We try to drop those buffers here
 612                  * and if that worked, and the page is no longer mapped into
 613                  * process address space (page_count == 1) it can be freed.
 614                  * Otherwise, leave the page on the LRU so it is swappable.
 615                  */
 616                 if (PagePrivate(page)) {
 617                         if (!try_to_release_page(page, sc->gfp_mask))
 618                                 goto activate_locked;
 619                         if (!mapping && page_count(page) == 1) {
 620                                 unlock_page(page);
 621                                 if (put_page_testzero(page))
 622                                         goto free_it;
 623                                 else {
 624                                         /*
 625                                          * rare race with speculative reference.
 626                                          * the speculative reference will free
 627                                          * this page shortly, so we may
 628                                          * increment nr_reclaimed here (and
 629                                          * leave it off the LRU).
 630                                          */
 631                                         nr_reclaimed++;
 632                                         continue;
 633                                 }
 634                         }
 635                 }
 636
 637                 if (!mapping || !__remove_mapping(mapping, page))
 638                         goto keep_locked;
 639
 640                 unlock_page(page);
 641 free_it:
 642                 nr_reclaimed++;
 643                 if (!pagevec_add(&freed_pvec, page)) {
 644                         __pagevec_free(&freed_pvec);
 645                         pagevec_reinit(&freed_pvec);
 646                 }
 647                 continue;
 648
 649 activate_locked:
 650                 SetPageActive(page);
 651                 pgactivate++;
 652 keep_locked:
 653                 unlock_page(page);
 654 keep:
 655                 list_add(&page->lru, &ret_pages);
 656                 VM_BUG_ON(PageLRU(page));
 657         }
 658         list_splice(&ret_pages, page_list);
 659         if (pagevec_count(&freed_pvec))
 660                 __pagevec_free(&freed_pvec);
 661         count_vm_events(PGACTIVATE, pgactivate);
 662         return nr_reclaimed;
 663 }
 664
 665 /* LRU Isolation modes. */
 666 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
 667 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
 668 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
 669
 670 /*
 671  * Attempt to remove the specified page from its LRU.  Only take this page
 672  * if it is of the appropriate PageActive status.  Pages which are being
 673  * freed elsewhere are also ignored.
 674  *
 675  * page:        page to consider
 676  * mode:        one of the LRU isolation modes defined above
 677  *
 678  * returns 0 on success, -ve errno on failure.
 679  */
 680 int __isolate_lru_page(struct page *page, int mode)
 681 {
 682         int ret = -EINVAL;
 683
 684         /* Only take pages on the LRU. */
 685         if (!PageLRU(page))
 686                 return ret;
 687
 688         /*
 689          * When checking the active state, we need to be sure we are
 690          * dealing with comparible boolean values.  Take the logical not
 691          * of each.
 692          */
 693         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
 694                 return ret;
 695
 696         ret = -EBUSY;
 697         if (likely(get_page_unless_zero(page))) {
 698                 /*
 699                  * Be careful not to clear PageLRU until after we're
 700                  * sure the page is not being freed elsewhere -- the
 701                  * page release code relies on it.
 702                  */
 703                 ClearPageLRU(page);
 704                 ret = 0;
 705         }
 706
 707         return ret;
 708 }
 709
 710 /*
 711  * zone->lru_lock is heavily contended.  Some of the functions that
 712  * shrink the lists perform better by taking out a batch of pages
 713  * and working on them outside the LRU lock.
 714  *
 715  * For pagecache intensive workloads, this function is the hottest
 716  * spot in the kernel (apart from copy_*_user functions).
 717  *
 718  * Appropriate locks must be held before calling this function.
 719  *
 720  * @nr_to_scan: The number of pages to look through on the list.
 721  * @src:        The LRU list to pull pages off.
 722  * @dst:        The temp list to put pages on to.
 723  * @scanned:    The number of pages that were scanned.
 724  * @order:      The caller's attempted allocation order
 725  * @mode:       One of the LRU isolation modes
 726  *
 727  * returns how many pages were moved onto *@dst.
 728  */
 729 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
 730                 struct list_head *src, struct list_head *dst,
 731                 unsigned long *scanned, int order, int mode)
 732 {
 733         unsigned long nr_taken = 0;
 734         unsigned long scan;
 735
 736         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 737                 struct page *page;
 738                 unsigned long pfn;
 739                 unsigned long end_pfn;
 740                 unsigned long page_pfn;
 741                 int zone_id;
 742
 743                 page = lru_to_page(src);
 744                 prefetchw_prev_lru_page(page, src, flags);
 745
 746                 VM_BUG_ON(!PageLRU(page));
 747
 748                 switch (__isolate_lru_page(page, mode)) {
 749                 case 0:
 750                         list_move(&page->lru, dst);
 751                         nr_taken++;
 752                         break;
 753
 754                 case -EBUSY:
 755                         /* else it is being freed elsewhere */
 756                         list_move(&page->lru, src);
 757                         continue;
 758
 759                 default:
 760                         BUG();
 761                 }
 762
 763                 if (!order)
 764                         continue;
 765
 766                 /*
 767                  * Attempt to take all pages in the order aligned region
 768                  * surrounding the tag page.  Only take those pages of
 769                  * the same active state as that tag page.  We may safely
 770                  * round the target page pfn down to the requested order
 771                  * as the mem_map is guarenteed valid out to MAX_ORDER,
 772                  * where that page is in a different zone we will detect
 773                  * it from its zone id and abort this block scan.
 774                  */
 775                 zone_id = page_zone_id(page);
 776                 page_pfn = page_to_pfn(page);
 777                 pfn = page_pfn & ~((1 << order) - 1);
 778                 end_pfn = pfn + (1 << order);
 779                 for (; pfn < end_pfn; pfn++) {
 780                         struct page *cursor_page;
 781
 782                         /* The target page is in the block, ignore it. */
 783                         if (unlikely(pfn == page_pfn))
 784                                 continue;
 785
 786                         /* Avoid holes within the zone. */
 787                         if (unlikely(!pfn_valid_within(pfn)))
 788                                 break;
 789
 790                         cursor_page = pfn_to_page(pfn);
 791                         /* Check that we have not crossed a zone boundary. */
 792                         if (unlikely(page_zone_id(cursor_page) != zone_id))
 793                                 continue;
 794                         switch (__isolate_lru_page(cursor_page, mode)) {
 795                         case 0:
 796                                 list_move(&cursor_page->lru, dst);
 797                                 nr_taken++;
 798                                 scan++;
 799                                 break;
 800
 801                         case -EBUSY:
 802                                 /* else it is being freed elsewhere */
 803                                 list_move(&cursor_page->lru, src);
 804                         default:
 805                                 break;
 806                         }
 807                 }
 808         }
 809
 810         *scanned = scan;
 811         return nr_taken;
 812 }
 813
 814 static unsigned long isolate_pages_global(unsigned long nr,
 815                                         struct list_head *dst,
 816                                         unsigned long *scanned, int order,
 817                                         int mode, struct zone *z,
 818                                         struct mem_cgroup *mem_cont,
 819                                         int active)
 820 {
 821         if (active)
 822                 return isolate_lru_pages(nr, &z->active_list, dst,
 823                                                 scanned, order, mode);
 824         else
 825                 return isolate_lru_pages(nr, &z->inactive_list, dst,
 826                                                 scanned, order, mode);
 827 }
 828
 829 /*
 830  * clear_active_flags() is a helper for shrink_active_list(), clearing
 831  * any active bits from the pages in the list.
 832  */
 833 static unsigned long clear_active_flags(struct list_head *page_list)
 834 {
 835         int nr_active = 0;
 836         struct page *page;
 837
 838         list_for_each_entry(page, page_list, lru)
 839                 if (PageActive(page)) {
 840                         ClearPageActive(page);
 841                         nr_active++;
 842                 }
 843
 844         return nr_active;
 845 }
 846
 847 /*
 848  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
 849  * of reclaimed pages
 850  */
 851 static unsigned long shrink_inactive_list(unsigned long max_scan,
 852                                 struct zone *zone, struct scan_control *sc)
 853 {
 854         LIST_HEAD(page_list);
 855         struct pagevec pvec;
 856         unsigned long nr_scanned = 0;
 857         unsigned long nr_reclaimed = 0;
 858
 859         pagevec_init(&pvec, 1);
 860
 861         lru_add_drain();
 862         spin_lock_irq(&zone->lru_lock);
 863         do {
 864                 struct page *page;
 865                 unsigned long nr_taken;
 866                 unsigned long nr_scan;
 867                 unsigned long nr_freed;
 868                 unsigned long nr_active;
 869
 870                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
 871                              &page_list, &nr_scan, sc->order,
 872                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
 873                                              ISOLATE_BOTH : ISOLATE_INACTIVE,
 874                                 zone, sc->mem_cgroup, 0);
 875                 nr_active = clear_active_flags(&page_list);
 876                 __count_vm_events(PGDEACTIVATE, nr_active);
 877
 878                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
 879                 __mod_zone_page_state(zone, NR_INACTIVE,
 880                                                 -(nr_taken - nr_active));
 881                 if (scan_global_lru(sc))
 882                         zone->pages_scanned += nr_scan;
 883                 spin_unlock_irq(&zone->lru_lock);
 884
 885                 nr_scanned += nr_scan;
 886                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
 887
 888                 /*
 889                  * If we are direct reclaiming for contiguous pages and we do
 890                  * not reclaim everything in the list, try again and wait
 891                  * for IO to complete. This will stall high-order allocations
 892                  * but that should be acceptable to the caller
 893                  */
 894                 if (nr_freed < nr_taken && !current_is_kswapd() &&
 895                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
 896                         congestion_wait(WRITE, HZ/10);
 897
 898                         /*
 899                          * The attempt at page out may have made some
 900                          * of the pages active, mark them inactive again.
 901                          */
 902                         nr_active = clear_active_flags(&page_list);
 903                         count_vm_events(PGDEACTIVATE, nr_active);
 904
 905                         nr_freed += shrink_page_list(&page_list, sc,
 906                                                         PAGEOUT_IO_SYNC);
 907                 }
 908
 909                 nr_reclaimed += nr_freed;
 910                 local_irq_disable();
 911                 if (current_is_kswapd()) {
 912                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
 913                         __count_vm_events(KSWAPD_STEAL, nr_freed);
 914                 } else if (scan_global_lru(sc))
 915                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
 916
 917                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
 918
 919                 if (nr_taken == 0)
 920                         goto done;
 921
 922                 spin_lock(&zone->lru_lock);
 923                 /*
 924                  * Put back any unfreeable pages.
 925                  */
 926                 while (!list_empty(&page_list)) {
 927                         page = lru_to_page(&page_list);
 928                         VM_BUG_ON(PageLRU(page));
 929                         SetPageLRU(page);
 930                         list_del(&page->lru);
 931                         if (PageActive(page))
 932                                 add_page_to_active_list(zone, page);
 933                         else
 934                                 add_page_to_inactive_list(zone, page);
 935                         if (!pagevec_add(&pvec, page)) {
 936                                 spin_unlock_irq(&zone->lru_lock);
 937                                 __pagevec_release(&pvec);
 938                                 spin_lock_irq(&zone->lru_lock);
 939                         }
 940                 }
 941         } while (nr_scanned < max_scan);
 942         spin_unlock(&zone->lru_lock);
 943 done:
 944         local_irq_enable();
 945         pagevec_release(&pvec);
 946         return nr_reclaimed;
 947 }
 948
 949 /*
 950  * We are about to scan this zone at a certain priority level.  If that priority
 951  * level is smaller (ie: more urgent) than the previous priority, then note
 952  * that priority level within the zone.  This is done so that when the next
 953  * process comes in to scan this zone, it will immediately start out at this
 954  * priority level rather than having to build up its own scanning priority.
 955  * Here, this priority affects only the reclaim-mapped threshold.
 956  */
 957 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
 958 {
 959         if (priority < zone->prev_priority)
 960                 zone->prev_priority = priority;
 961 }
 962
 963 static inline int zone_is_near_oom(struct zone *zone)
 964 {
 965         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
 966                                 + zone_page_state(zone, NR_INACTIVE))*3;
 967 }
 968
 969 /*
 970  * Determine we should try to reclaim mapped pages.
 971  * This is called only when sc->mem_cgroup is NULL.
 972  */
 973 static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
 974                                 int priority)
 975 {
 976         long mapped_ratio;
 977         long distress;
 978         long swap_tendency;
 979         long imbalance;
 980         int reclaim_mapped = 0;
 981         int prev_priority;
 982
 983         if (scan_global_lru(sc) && zone_is_near_oom(zone))
 984                 return 1;
 985         /*
 986          * `distress' is a measure of how much trouble we're having
 987          * reclaiming pages.  0 -> no problems.  100 -> great trouble.
 988          */
 989         if (scan_global_lru(sc))
 990                 prev_priority = zone->prev_priority;
 991         else
 992                 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
 993
 994         distress = 100 >> min(prev_priority, priority);
 995
 996         /*
 997          * The point of this algorithm is to decide when to start
 998          * reclaiming mapped memory instead of just pagecache.  Work out
 999          * how much memory
1000          * is mapped.
1001          */
1002         if (scan_global_lru(sc))
1003                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
1004                                 global_page_state(NR_ANON_PAGES)) * 100) /
1005                                         vm_total_pages;
1006         else
1007                 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
1008
1009         /*
1010          * Now decide how much we really want to unmap some pages.  The
1011          * mapped ratio is downgraded - just because there's a lot of
1012          * mapped memory doesn't necessarily mean that page reclaim
1013          * isn't succeeding.
1014          *
1015          * The distress ratio is important - we don't want to start
1016          * going oom.
1017          *
1018          * A 100% value of vm_swappiness overrides this algorithm
1019          * altogether.
1020          */
1021         swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1022
1023         /*
1024          * If there's huge imbalance between active and inactive
1025          * (think active 100 times larger than inactive) we should
1026          * become more permissive, or the system will take too much
1027          * cpu before it start swapping during memory pressure.
1028          * Distress is about avoiding early-oom, this is about
1029          * making swappiness graceful despite setting it to low
1030          * values.
1031          *
1032          * Avoid div by zero with nr_inactive+1, and max resulting
1033          * value is vm_total_pages.
1034          */
1035         if (scan_global_lru(sc)) {
1036                 imbalance  = zone_page_state(zone, NR_ACTIVE);
1037                 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1038         } else
1039                 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1040
1041         /*
1042          * Reduce the effect of imbalance if swappiness is low,
1043          * this means for a swappiness very low, the imbalance
1044          * must be much higher than 100 for this logic to make
1045          * the difference.
1046          *
1047          * Max temporary value is vm_total_pages*100.
1048          */
1049         imbalance *= (vm_swappiness + 1);
1050         imbalance /= 100;
1051
1052         /*
1053          * If not much of the ram is mapped, makes the imbalance
1054          * less relevant, it's high priority we refill the inactive
1055          * list with mapped pages only in presence of high ratio of
1056          * mapped pages.
1057          *
1058          * Max temporary value is vm_total_pages*100.
1059          */
1060         imbalance *= mapped_ratio;
1061         imbalance /= 100;
1062
1063         /* apply imbalance feedback to swap_tendency */
1064         swap_tendency += imbalance;
1065
1066         /*
1067          * Now use this metric to decide whether to start moving mapped
1068          * memory onto the inactive list.
1069          */
1070         if (swap_tendency >= 100)
1071                 reclaim_mapped = 1;
1072
1073         return reclaim_mapped;
1074 }
1075
1076 /*
1077  * This moves pages from the active list to the inactive list.
1078  *
1079  * We move them the other way if the page is referenced by one or more
1080  * processes, from rmap.
1081  *
1082  * If the pages are mostly unmapped, the processing is fast and it is
1083  * appropriate to hold zone->lru_lock across the whole operation.  But if
1084  * the pages are mapped, the processing is slow (page_referenced()) so we
1085  * should drop zone->lru_lock around each page.  It's impossible to balance
1086  * this, so instead we remove the pages from the LRU while processing them.
1087  * It is safe to rely on PG_active against the non-LRU pages in here because
1088  * nobody will play with that bit on a non-LRU page.
1089  *
1090  * The downside is that we have to touch page->_count against each page.
1091  * But we had to alter page->flags anyway.
1092  */
1093
1094
1095 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1096                                 struct scan_control *sc, int priority)
1097 {
1098         unsigned long pgmoved;
1099         int pgdeactivate = 0;
1100         unsigned long pgscanned;
1101         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1102         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1103         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1104         struct page *page;
1105         struct pagevec pvec;
1106         int reclaim_mapped = 0;
1107
1108         if (sc->may_swap)
1109                 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
1110
1111         lru_add_drain();
1112         spin_lock_irq(&zone->lru_lock);
1113         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1114                                         ISOLATE_ACTIVE, zone,
1115                                         sc->mem_cgroup, 1);
1116         /*
1117          * zone->pages_scanned is used for detect zone's oom
1118          * mem_cgroup remembers nr_scan by itself.
1119          */
1120         if (scan_global_lru(sc))
1121                 zone->pages_scanned += pgscanned;
1122
1123         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1124         spin_unlock_irq(&zone->lru_lock);
1125
1126         while (!list_empty(&l_hold)) {
1127                 cond_resched();
1128                 page = lru_to_page(&l_hold);
1129                 list_del(&page->lru);
1130                 if (page_mapped(page)) {
1131                         if (!reclaim_mapped ||
1132                             (total_swap_pages == 0 && PageAnon(page)) ||
1133                             page_referenced(page, 0, sc->mem_cgroup)) {
1134                                 list_add(&page->lru, &l_active);
1135                                 continue;
1136                         }
1137                 }
1138                 list_add(&page->lru, &l_inactive);
1139         }
1140
1141         pagevec_init(&pvec, 1);
1142         pgmoved = 0;
1143         spin_lock_irq(&zone->lru_lock);
1144         while (!list_empty(&l_inactive)) {
1145                 page = lru_to_page(&l_inactive);
1146                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1147                 VM_BUG_ON(PageLRU(page));
1148                 SetPageLRU(page);
1149                 VM_BUG_ON(!PageActive(page));
1150                 ClearPageActive(page);
1151
1152                 list_move(&page->lru, &zone->inactive_list);
1153                 mem_cgroup_move_lists(page, false);
1154                 pgmoved++;
1155                 if (!pagevec_add(&pvec, page)) {
1156                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1157                         spin_unlock_irq(&zone->lru_lock);
1158                         pgdeactivate += pgmoved;
1159                         pgmoved = 0;
1160                         if (buffer_heads_over_limit)
1161                                 pagevec_strip(&pvec);
1162                         __pagevec_release(&pvec);
1163                         spin_lock_irq(&zone->lru_lock);
1164                 }
1165         }
1166         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1167         pgdeactivate += pgmoved;
1168         if (buffer_heads_over_limit) {
1169                 spin_unlock_irq(&zone->lru_lock);
1170                 pagevec_strip(&pvec);
1171                 spin_lock_irq(&zone->lru_lock);
1172         }
1173
1174         pgmoved = 0;
1175         while (!list_empty(&l_active)) {
1176                 page = lru_to_page(&l_active);
1177                 prefetchw_prev_lru_page(page, &l_active, flags);
1178                 VM_BUG_ON(PageLRU(page));
1179                 SetPageLRU(page);
1180                 VM_BUG_ON(!PageActive(page));
1181
1182                 list_move(&page->lru, &zone->active_list);
1183                 mem_cgroup_move_lists(page, true);
1184                 pgmoved++;
1185                 if (!pagevec_add(&pvec, page)) {
1186                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1187                         pgmoved = 0;
1188                         spin_unlock_irq(&zone->lru_lock);
1189                         __pagevec_release(&pvec);
1190                         spin_lock_irq(&zone->lru_lock);
1191                 }
1192         }
1193         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1194
1195         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1196         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1197         spin_unlock_irq(&zone->lru_lock);
1198
1199         pagevec_release(&pvec);
1200 }
1201
1202 /*
1203  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1204  */
1205 static unsigned long shrink_zone(int priority, struct zone *zone,
1206                                 struct scan_control *sc)
1207 {
1208         unsigned long nr_active;
1209         unsigned long nr_inactive;
1210         unsigned long nr_to_scan;
1211         unsigned long nr_reclaimed = 0;
1212
1213         if (scan_global_lru(sc)) {
1214                 /*
1215                  * Add one to nr_to_scan just to make sure that the kernel
1216                  * will slowly sift through the active list.
1217                  */
1218                 zone->nr_scan_active +=
1219                         (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1220                 nr_active = zone->nr_scan_active;
1221                 zone->nr_scan_inactive +=
1222                         (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1223                 nr_inactive = zone->nr_scan_inactive;
1224                 if (nr_inactive >= sc->swap_cluster_max)
1225                         zone->nr_scan_inactive = 0;
1226                 else
1227                         nr_inactive = 0;
1228
1229                 if (nr_active >= sc->swap_cluster_max)
1230                         zone->nr_scan_active = 0;
1231                 else
1232                         nr_active = 0;
1233         } else {
1234                 /*
1235                  * This reclaim occurs not because zone memory shortage but
1236                  * because memory controller hits its limit.
1237                  * Then, don't modify zone reclaim related data.
1238                  */
1239                 nr_active = mem_cgroup_calc_reclaim_active(sc->mem_cgroup,
1240                                         zone, priority);
1241
1242                 nr_inactive = mem_cgroup_calc_reclaim_inactive(sc->mem_cgroup,
1243                                         zone, priority);
1244         }
1245
1246
1247         while (nr_active || nr_inactive) {
1248                 if (nr_active) {
1249                         nr_to_scan = min(nr_active,
1250                                         (unsigned long)sc->swap_cluster_max);
1251                         nr_active -= nr_to_scan;
1252                         shrink_active_list(nr_to_scan, zone, sc, priority);
1253                 }
1254
1255                 if (nr_inactive) {
1256                         nr_to_scan = min(nr_inactive,
1257                                         (unsigned long)sc->swap_cluster_max);
1258                         nr_inactive -= nr_to_scan;
1259                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1260                                                                 sc);
1261                 }
1262         }
1263
1264         throttle_vm_writeout(sc->gfp_mask);
1265         return nr_reclaimed;
1266 }
1267
1268 /*
1269  * This is the direct reclaim path, for page-allocating processes.  We only
1270  * try to reclaim pages from zones which will satisfy the caller's allocation
1271  * request.
1272  *
1273  * We reclaim from a zone even if that zone is over pages_high.  Because:
1274  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1275  *    allocation or
1276  * b) The zones may be over pages_high but they must go *over* pages_high to
1277  *    satisfy the `incremental min' zone defense algorithm.
1278  *
1279  * Returns the number of reclaimed pages.
1280  *
1281  * If a zone is deemed to be full of pinned pages then just give it a light
1282  * scan then give up on it.
1283  */
1284 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1285                                         struct scan_control *sc)
1286 {
1287         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1288         unsigned long nr_reclaimed = 0;
1289         struct zoneref *z;
1290         struct zone *zone;
1291
1292         sc->all_unreclaimable = 1;
1293         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1294                 if (!populated_zone(zone))
1295                         continue;
1296                 /*
1297                  * Take care memory controller reclaiming has small influence
1298                  * to global LRU.
1299                  */
1300                 if (scan_global_lru(sc)) {
1301                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1302                                 continue;
1303                         note_zone_scanning_priority(zone, priority);
1304
1305                         if (zone_is_all_unreclaimable(zone) &&
1306                                                 priority != DEF_PRIORITY)
1307                                 continue;       /* Let kswapd poll it */
1308                         sc->all_unreclaimable = 0;
1309                 } else {
1310                         /*
1311                          * Ignore cpuset limitation here. We just want to reduce
1312                          * # of used pages by us regardless of memory shortage.
1313                          */
1314                         sc->all_unreclaimable = 0;
1315                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1316                                                         priority);
1317                 }
1318
1319                 nr_reclaimed += shrink_zone(priority, zone, sc);
1320         }
1321
1322         return nr_reclaimed;
1323 }
1324
1325 /*
1326  * This is the main entry point to direct page reclaim.
1327  *
1328  * If a full scan of the inactive list fails to free enough memory then we
1329  * are "out of memory" and something needs to be killed.
1330  *
1331  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1332  * high - the zone may be full of dirty or under-writeback pages, which this
1333  * caller can't do much about.  We kick pdflush and take explicit naps in the
1334  * hope that some of these pages can be written.  But if the allocating task
1335  * holds filesystem locks which prevent writeout this might not work, and the
1336  * allocation attempt will fail.
1337  *
1338  * returns:     0, if no pages reclaimed
1339  *              else, the number of pages reclaimed
1340  */
1341 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1342                                         struct scan_control *sc)
1343 {
1344         int priority;
1345         unsigned long ret = 0;
1346         unsigned long total_scanned = 0;
1347         unsigned long nr_reclaimed = 0;
1348         struct reclaim_state *reclaim_state = current->reclaim_state;
1349         unsigned long lru_pages = 0;
1350         struct zoneref *z;
1351         struct zone *zone;
1352         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1353
1354         delayacct_freepages_start();
1355
1356         if (scan_global_lru(sc))
1357                 count_vm_event(ALLOCSTALL);
1358         /*
1359          * mem_cgroup will not do shrink_slab.
1360          */
1361         if (scan_global_lru(sc)) {
1362                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1363
1364                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1365                                 continue;
1366
1367                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1368                                         + zone_page_state(zone, NR_INACTIVE);
1369                 }
1370         }
1371
1372         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1373                 sc->nr_scanned = 0;
1374                 if (!priority)
1375                         disable_swap_token();
1376                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1377                 /*
1378                  * Don't shrink slabs when reclaiming memory from
1379                  * over limit cgroups
1380                  */
1381                 if (scan_global_lru(sc)) {
1382                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1383                         if (reclaim_state) {
1384                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1385                                 reclaim_state->reclaimed_slab = 0;
1386                         }
1387                 }
1388                 total_scanned += sc->nr_scanned;
1389                 if (nr_reclaimed >= sc->swap_cluster_max) {
1390                         ret = nr_reclaimed;
1391                         goto out;
1392                 }
1393
1394                 /*
1395                  * Try to write back as many pages as we just scanned.  This
1396                  * tends to cause slow streaming writers to write data to the
1397                  * disk smoothly, at the dirtying rate, which is nice.   But
1398                  * that's undesirable in laptop mode, where we *want* lumpy
1399                  * writeout.  So in laptop mode, write out the whole world.
1400                  */
1401                 if (total_scanned > sc->swap_cluster_max +
1402                                         sc->swap_cluster_max / 2) {
1403                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1404                         sc->may_writepage = 1;
1405                 }
1406
1407                 /* Take a nap, wait for some writeback to complete */
1408                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1409                         congestion_wait(WRITE, HZ/10);
1410         }
1411         /* top priority shrink_zones still had more to do? don't OOM, then */
1412         if (!sc->all_unreclaimable && scan_global_lru(sc))
1413                 ret = nr_reclaimed;
1414 out:
1415         /*
1416          * Now that we've scanned all the zones at this priority level, note
1417          * that level within the zone so that the next thread which performs
1418          * scanning of this zone will immediately start out at this priority
1419          * level.  This affects only the decision whether or not to bring
1420          * mapped pages onto the inactive list.
1421          */
1422         if (priority < 0)
1423                 priority = 0;
1424
1425         if (scan_global_lru(sc)) {
1426                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1427
1428                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1429                                 continue;
1430
1431                         zone->prev_priority = priority;
1432                 }
1433         } else
1434                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1435
1436         delayacct_freepages_end();
1437
1438         return ret;
1439 }
1440
1441 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1442                                                                 gfp_t gfp_mask)
1443 {
1444         struct scan_control sc = {
1445                 .gfp_mask = gfp_mask,
1446                 .may_writepage = !laptop_mode,
1447                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1448                 .may_swap = 1,
1449                 .swappiness = vm_swappiness,
1450                 .order = order,
1451                 .mem_cgroup = NULL,
1452                 .isolate_pages = isolate_pages_global,
1453         };
1454
1455         return do_try_to_free_pages(zonelist, &sc);
1456 }
1457
1458 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1459
1460 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1461                                                 gfp_t gfp_mask)
1462 {
1463         struct scan_control sc = {
1464                 .may_writepage = !laptop_mode,
1465                 .may_swap = 1,
1466                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1467                 .swappiness = vm_swappiness,
1468                 .order = 0,
1469                 .mem_cgroup = mem_cont,
1470                 .isolate_pages = mem_cgroup_isolate_pages,
1471         };
1472         struct zonelist *zonelist;
1473
1474         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1475                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1476         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1477         return do_try_to_free_pages(zonelist, &sc);
1478 }
1479 #endif
1480
1481 /*
1482  * For kswapd, balance_pgdat() will work across all this node's zones until
1483  * they are all at pages_high.
1484  *
1485  * Returns the number of pages which were actually freed.
1486  *
1487  * There is special handling here for zones which are full of pinned pages.
1488  * This can happen if the pages are all mlocked, or if they are all used by
1489  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1490  * What we do is to detect the case where all pages in the zone have been
1491  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1492  * dead and from now on, only perform a short scan.  Basically we're polling
1493  * the zone for when the problem goes away.
1494  *
1495  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1496  * zones which have free_pages > pages_high, but once a zone is found to have
1497  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1498  * of the number of free pages in the lower zones.  This interoperates with
1499  * the page allocator fallback scheme to ensure that aging of pages is balanced
1500  * across the zones.
1501  */
1502 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1503 {
1504         int all_zones_ok;
1505         int priority;
1506         int i;
1507         unsigned long total_scanned;
1508         unsigned long nr_reclaimed;
1509         struct reclaim_state *reclaim_state = current->reclaim_state;
1510         struct scan_control sc = {
1511                 .gfp_mask = GFP_KERNEL,
1512                 .may_swap = 1,
1513                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1514                 .swappiness = vm_swappiness,
1515                 .order = order,
1516                 .mem_cgroup = NULL,
1517                 .isolate_pages = isolate_pages_global,
1518         };
1519         /*
1520          * temp_priority is used to remember the scanning priority at which
1521          * this zone was successfully refilled to free_pages == pages_high.
1522          */
1523         int temp_priority[MAX_NR_ZONES];
1524
1525 loop_again:
1526         total_scanned = 0;
1527         nr_reclaimed = 0;
1528         sc.may_writepage = !laptop_mode;
1529         count_vm_event(PAGEOUTRUN);
1530
1531         for (i = 0; i < pgdat->nr_zones; i++)
1532                 temp_priority[i] = DEF_PRIORITY;
1533
1534         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1535                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1536                 unsigned long lru_pages = 0;
1537
1538                 /* The swap token gets in the way of swapout... */
1539                 if (!priority)
1540                         disable_swap_token();
1541
1542                 all_zones_ok = 1;
1543
1544                 /*
1545                  * Scan in the highmem->dma direction for the highest
1546                  * zone which needs scanning
1547                  */
1548                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1549                         struct zone *zone = pgdat->node_zones + i;
1550
1551                         if (!populated_zone(zone))
1552                                 continue;
1553
1554                         if (zone_is_all_unreclaimable(zone) &&
1555                             priority != DEF_PRIORITY)
1556                                 continue;
1557
1558                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1559                                                0, 0)) {
1560                                 end_zone = i;
1561                                 break;
1562                         }
1563                 }
1564                 if (i < 0)
1565                         goto out;
1566
1567                 for (i = 0; i <= end_zone; i++) {
1568                         struct zone *zone = pgdat->node_zones + i;
1569
1570                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1571                                         + zone_page_state(zone, NR_INACTIVE);
1572                 }
1573
1574                 /*
1575                  * Now scan the zone in the dma->highmem direction, stopping
1576                  * at the last zone which needs scanning.
1577                  *
1578                  * We do this because the page allocator works in the opposite
1579                  * direction.  This prevents the page allocator from allocating
1580                  * pages behind kswapd's direction of progress, which would
1581                  * cause too much scanning of the lower zones.
1582                  */
1583                 for (i = 0; i <= end_zone; i++) {
1584                         struct zone *zone = pgdat->node_zones + i;
1585                         int nr_slab;
1586
1587                         if (!populated_zone(zone))
1588                                 continue;
1589
1590                         if (zone_is_all_unreclaimable(zone) &&
1591                                         priority != DEF_PRIORITY)
1592                                 continue;
1593
1594                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1595                                                end_zone, 0))
1596                                 all_zones_ok = 0;
1597                         temp_priority[i] = priority;
1598                         sc.nr_scanned = 0;
1599                         note_zone_scanning_priority(zone, priority);
1600                         /*
1601                          * We put equal pressure on every zone, unless one
1602                          * zone has way too many pages free already.
1603                          */
1604                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1605                                                 end_zone, 0))
1606                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1607                         reclaim_state->reclaimed_slab = 0;
1608                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1609                                                 lru_pages);
1610                         nr_reclaimed += reclaim_state->reclaimed_slab;
1611                         total_scanned += sc.nr_scanned;
1612                         if (zone_is_all_unreclaimable(zone))
1613                                 continue;
1614                         if (nr_slab == 0 && zone->pages_scanned >=
1615                                 (zone_page_state(zone, NR_ACTIVE)
1616                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1617                                         zone_set_flag(zone,
1618                                                       ZONE_ALL_UNRECLAIMABLE);
1619                         /*
1620                          * If we've done a decent amount of scanning and
1621                          * the reclaim ratio is low, start doing writepage
1622                          * even in laptop mode
1623                          */
1624                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1625                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1626                                 sc.may_writepage = 1;
1627                 }
1628                 if (all_zones_ok)
1629                         break;          /* kswapd: all done */
1630                 /*
1631                  * OK, kswapd is getting into trouble.  Take a nap, then take
1632                  * another pass across the zones.
1633                  */
1634                 if (total_scanned && priority < DEF_PRIORITY - 2)
1635                         congestion_wait(WRITE, HZ/10);
1636
1637                 /*
1638                  * We do this so kswapd doesn't build up large priorities for
1639                  * example when it is freeing in parallel with allocators. It
1640                  * matches the direct reclaim path behaviour in terms of impact
1641                  * on zone->*_priority.
1642                  */
1643                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1644                         break;
1645         }
1646 out:
1647         /*
1648          * Note within each zone the priority level at which this zone was
1649          * brought into a happy state.  So that the next thread which scans this
1650          * zone will start out at that priority level.
1651          */
1652         for (i = 0; i < pgdat->nr_zones; i++) {
1653                 struct zone *zone = pgdat->node_zones + i;
1654
1655                 zone->prev_priority = temp_priority[i];
1656         }
1657         if (!all_zones_ok) {
1658                 cond_resched();
1659
1660                 try_to_freeze();
1661
1662                 goto loop_again;
1663         }
1664
1665         return nr_reclaimed;
1666 }
1667
1668 /*
1669  * The background pageout daemon, started as a kernel thread
1670  * from the init process.
1671  *
1672  * This basically trickles out pages so that we have _some_
1673  * free memory available even if there is no other activity
1674  * that frees anything up. This is needed for things like routing
1675  * etc, where we otherwise might have all activity going on in
1676  * asynchronous contexts that cannot page things out.
1677  *
1678  * If there are applications that are active memory-allocators
1679  * (most normal use), this basically shouldn't matter.
1680  */
1681 static int kswapd(void *p)
1682 {
1683         unsigned long order;
1684         pg_data_t *pgdat = (pg_data_t*)p;
1685         struct task_struct *tsk = current;
1686         DEFINE_WAIT(wait);
1687         struct reclaim_state reclaim_state = {
1688                 .reclaimed_slab = 0,
1689         };
1690         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1691
1692         if (!cpus_empty(*cpumask))
1693                 set_cpus_allowed_ptr(tsk, cpumask);
1694         current->reclaim_state = &reclaim_state;
1695
1696         /*
1697          * Tell the memory management that we're a "memory allocator",
1698          * and that if we need more memory we should get access to it
1699          * regardless (see "__alloc_pages()"). "kswapd" should
1700          * never get caught in the normal page freeing logic.
1701          *
1702          * (Kswapd normally doesn't need memory anyway, but sometimes
1703          * you need a small amount of memory in order to be able to
1704          * page out something else, and this flag essentially protects
1705          * us from recursively trying to free more memory as we're
1706          * trying to free the first piece of memory in the first place).
1707          */
1708         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1709         set_freezable();
1710
1711         order = 0;
1712         for ( ; ; ) {
1713                 unsigned long new_order;
1714
1715                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1716                 new_order = pgdat->kswapd_max_order;
1717                 pgdat->kswapd_max_order = 0;
1718                 if (order < new_order) {
1719                         /*
1720                          * Don't sleep if someone wants a larger 'order'
1721                          * allocation
1722                          */
1723                         order = new_order;
1724                 } else {
1725                         if (!freezing(current))
1726                                 schedule();
1727
1728                         order = pgdat->kswapd_max_order;
1729                 }
1730                 finish_wait(&pgdat->kswapd_wait, &wait);
1731
1732                 if (!try_to_freeze()) {
1733                         /* We can speed up thawing tasks if we don't call
1734                          * balance_pgdat after returning from the refrigerator
1735                          */
1736                         balance_pgdat(pgdat, order);
1737                 }
1738         }
1739         return 0;
1740 }
1741
1742 /*
1743  * A zone is low on free memory, so wake its kswapd task to service it.
1744  */
1745 void wakeup_kswapd(struct zone *zone, int order)
1746 {
1747         pg_data_t *pgdat;
1748
1749         if (!populated_zone(zone))
1750                 return;
1751
1752         pgdat = zone->zone_pgdat;
1753         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1754                 return;
1755         if (pgdat->kswapd_max_order < order)
1756                 pgdat->kswapd_max_order = order;
1757         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1758                 return;
1759         if (!waitqueue_active(&pgdat->kswapd_wait))
1760                 return;
1761         wake_up_interruptible(&pgdat->kswapd_wait);
1762 }
1763
1764 #ifdef CONFIG_PM
1765 /*
1766  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1767  * from LRU lists system-wide, for given pass and priority, and returns the
1768  * number of reclaimed pages
1769  *
1770  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1771  */
1772 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1773                                       int pass, struct scan_control *sc)
1774 {
1775         struct zone *zone;
1776         unsigned long nr_to_scan, ret = 0;
1777
1778         for_each_zone(zone) {
1779
1780                 if (!populated_zone(zone))
1781                         continue;
1782
1783                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1784                         continue;
1785
1786                 /* For pass = 0 we don't shrink the active list */
1787                 if (pass > 0) {
1788                         zone->nr_scan_active +=
1789                                 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1790                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1791                                 zone->nr_scan_active = 0;
1792                                 nr_to_scan = min(nr_pages,
1793                                         zone_page_state(zone, NR_ACTIVE));
1794                                 shrink_active_list(nr_to_scan, zone, sc, prio);
1795                         }
1796                 }
1797
1798                 zone->nr_scan_inactive +=
1799                         (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1800                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1801                         zone->nr_scan_inactive = 0;
1802                         nr_to_scan = min(nr_pages,
1803                                 zone_page_state(zone, NR_INACTIVE));
1804                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1805                         if (ret >= nr_pages)
1806                                 return ret;
1807                 }
1808         }
1809
1810         return ret;
1811 }
1812
1813 static unsigned long count_lru_pages(void)
1814 {
1815         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1816 }
1817
1818 /*
1819  * Try to free `nr_pages' of memory, system-wide, and return the number of
1820  * freed pages.
1821  *
1822  * Rather than trying to age LRUs the aim is to preserve the overall
1823  * LRU order by reclaiming preferentially
1824  * inactive > active > active referenced > active mapped
1825  */
1826 unsigned long shrink_all_memory(unsigned long nr_pages)
1827 {
1828         unsigned long lru_pages, nr_slab;
1829         unsigned long ret = 0;
1830         int pass;
1831         struct reclaim_state reclaim_state;
1832         struct scan_control sc = {
1833                 .gfp_mask = GFP_KERNEL,
1834                 .may_swap = 0,
1835                 .swap_cluster_max = nr_pages,
1836                 .may_writepage = 1,
1837                 .swappiness = vm_swappiness,
1838                 .isolate_pages = isolate_pages_global,
1839         };
1840
1841         current->reclaim_state = &reclaim_state;
1842
1843         lru_pages = count_lru_pages();
1844         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1845         /* If slab caches are huge, it's better to hit them first */
1846         while (nr_slab >= lru_pages) {
1847                 reclaim_state.reclaimed_slab = 0;
1848                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1849                 if (!reclaim_state.reclaimed_slab)
1850                         break;
1851
1852                 ret += reclaim_state.reclaimed_slab;
1853                 if (ret >= nr_pages)
1854                         goto out;
1855
1856                 nr_slab -= reclaim_state.reclaimed_slab;
1857         }
1858
1859         /*
1860          * We try to shrink LRUs in 5 passes:
1861          * 0 = Reclaim from inactive_list only
1862          * 1 = Reclaim from active list but don't reclaim mapped
1863          * 2 = 2nd pass of type 1
1864          * 3 = Reclaim mapped (normal reclaim)
1865          * 4 = 2nd pass of type 3
1866          */
1867         for (pass = 0; pass < 5; pass++) {
1868                 int prio;
1869
1870                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1871                 if (pass > 2) {
1872                         sc.may_swap = 1;
1873                         sc.swappiness = 100;
1874                 }
1875
1876                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1877                         unsigned long nr_to_scan = nr_pages - ret;
1878
1879                         sc.nr_scanned = 0;
1880                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1881                         if (ret >= nr_pages)
1882                                 goto out;
1883
1884                         reclaim_state.reclaimed_slab = 0;
1885                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1886                                         count_lru_pages());
1887                         ret += reclaim_state.reclaimed_slab;
1888                         if (ret >= nr_pages)
1889                                 goto out;
1890
1891                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1892                                 congestion_wait(WRITE, HZ / 10);
1893                 }
1894         }
1895
1896         /*
1897          * If ret = 0, we could not shrink LRUs, but there may be something
1898          * in slab caches
1899          */
1900         if (!ret) {
1901                 do {
1902                         reclaim_state.reclaimed_slab = 0;
1903                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1904                         ret += reclaim_state.reclaimed_slab;
1905                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1906         }
1907
1908 out:
1909         current->reclaim_state = NULL;
1910
1911         return ret;
1912 }
1913 #endif
1914
1915 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1916    not required for correctness.  So if the last cpu in a node goes
1917    away, we get changed to run anywhere: as the first one comes back,
1918    restore their cpu bindings. */
1919 static int __devinit cpu_callback(struct notifier_block *nfb,
1920                                   unsigned long action, void *hcpu)
1921 {
1922         int nid;
1923
1924         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1925                 for_each_node_state(nid, N_HIGH_MEMORY) {
1926                         pg_data_t *pgdat = NODE_DATA(nid);
1927                         node_to_cpumask_ptr(mask, pgdat->node_id);
1928
1929                         if (any_online_cpu(*mask) < nr_cpu_ids)
1930                                 /* One of our CPUs online: restore mask */
1931                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1932                 }
1933         }
1934         return NOTIFY_OK;
1935 }
1936
1937 /*
1938  * This kswapd start function will be called by init and node-hot-add.
1939  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1940  */
1941 int kswapd_run(int nid)
1942 {
1943         pg_data_t *pgdat = NODE_DATA(nid);
1944         int ret = 0;
1945
1946         if (pgdat->kswapd)
1947                 return 0;
1948
1949         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1950         if (IS_ERR(pgdat->kswapd)) {
1951                 /* failure at boot is fatal */
1952                 BUG_ON(system_state == SYSTEM_BOOTING);
1953                 printk("Failed to start kswapd on node %d\n",nid);
1954                 ret = -1;
1955         }
1956         return ret;
1957 }
1958
1959 static int __init kswapd_init(void)
1960 {
1961         int nid;
1962
1963         swap_setup();
1964         for_each_node_state(nid, N_HIGH_MEMORY)
1965                 kswapd_run(nid);
1966         hotcpu_notifier(cpu_callback, 0);
1967         return 0;
1968 }
1969
1970 module_init(kswapd_init)
1971
1972 #ifdef CONFIG_NUMA
1973 /*
1974  * Zone reclaim mode
1975  *
1976  * If non-zero call zone_reclaim when the number of free pages falls below
1977  * the watermarks.
1978  */
1979 int zone_reclaim_mode __read_mostly;
1980
1981 #define RECLAIM_OFF 0
1982 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
1983 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1984 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1985
1986 /*
1987  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1988  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1989  * a zone.
1990  */
1991 #define ZONE_RECLAIM_PRIORITY 4
1992
1993 /*
1994  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1995  * occur.
1996  */
1997 int sysctl_min_unmapped_ratio = 1;
1998
1999 /*
2000  * If the number of slab pages in a zone grows beyond this percentage then
2001  * slab reclaim needs to occur.
2002  */
2003 int sysctl_min_slab_ratio = 5;
2004
2005 /*
2006  * Try to free up some pages from this zone through reclaim.
2007  */
2008 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2009 {
2010         /* Minimum pages needed in order to stay on node */
2011         const unsigned long nr_pages = 1 << order;
2012         struct task_struct *p = current;
2013         struct reclaim_state reclaim_state;
2014         int priority;
2015         unsigned long nr_reclaimed = 0;
2016         struct scan_control sc = {
2017                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2018                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2019                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2020                                         SWAP_CLUSTER_MAX),
2021                 .gfp_mask = gfp_mask,
2022                 .swappiness = vm_swappiness,
2023                 .isolate_pages = isolate_pages_global,
2024         };
2025         unsigned long slab_reclaimable;
2026
2027         disable_swap_token();
2028         cond_resched();
2029         /*
2030          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2031          * and we also need to be able to write out pages for RECLAIM_WRITE
2032          * and RECLAIM_SWAP.
2033          */
2034         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2035         reclaim_state.reclaimed_slab = 0;
2036         p->reclaim_state = &reclaim_state;
2037
2038         if (zone_page_state(zone, NR_FILE_PAGES) -
2039                 zone_page_state(zone, NR_FILE_MAPPED) >
2040                 zone->min_unmapped_pages) {
2041                 /*
2042                  * Free memory by calling shrink zone with increasing
2043                  * priorities until we have enough memory freed.
2044                  */
2045                 priority = ZONE_RECLAIM_PRIORITY;
2046                 do {
2047                         note_zone_scanning_priority(zone, priority);
2048                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2049                         priority--;
2050                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2051         }
2052
2053         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2054         if (slab_reclaimable > zone->min_slab_pages) {
2055                 /*
2056                  * shrink_slab() does not currently allow us to determine how
2057                  * many pages were freed in this zone. So we take the current
2058                  * number of slab pages and shake the slab until it is reduced
2059                  * by the same nr_pages that we used for reclaiming unmapped
2060                  * pages.
2061                  *
2062                  * Note that shrink_slab will free memory on all zones and may
2063                  * take a long time.
2064                  */
2065                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2066                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2067                                 slab_reclaimable - nr_pages)
2068                         ;
2069
2070                 /*
2071                  * Update nr_reclaimed by the number of slab pages we
2072                  * reclaimed from this zone.
2073                  */
2074                 nr_reclaimed += slab_reclaimable -
2075                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2076         }
2077
2078         p->reclaim_state = NULL;
2079         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2080         return nr_reclaimed >= nr_pages;
2081 }
2082
2083 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2084 {
2085         int node_id;
2086         int ret;
2087
2088         /*
2089          * Zone reclaim reclaims unmapped file backed pages and
2090          * slab pages if we are over the defined limits.
2091          *
2092          * A small portion of unmapped file backed pages is needed for
2093          * file I/O otherwise pages read by file I/O will be immediately
2094          * thrown out if the zone is overallocated. So we do not reclaim
2095          * if less than a specified percentage of the zone is used by
2096          * unmapped file backed pages.
2097          */
2098         if (zone_page_state(zone, NR_FILE_PAGES) -
2099             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2100             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2101                         <= zone->min_slab_pages)
2102                 return 0;
2103
2104         if (zone_is_all_unreclaimable(zone))
2105                 return 0;
2106
2107         /*
2108          * Do not scan if the allocation should not be delayed.
2109          */
2110         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2111                         return 0;
2112
2113         /*
2114          * Only run zone reclaim on the local zone or on zones that do not
2115          * have associated processors. This will favor the local processor
2116          * over remote processors and spread off node memory allocations
2117          * as wide as possible.
2118          */
2119         node_id = zone_to_nid(zone);
2120         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2121                 return 0;
2122
2123         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2124                 return 0;
2125         ret = __zone_reclaim(zone, gfp_mask, order);
2126         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2127
2128         return ret;
2129 }
2130 #endif