linux/mm/vmscan.c
<<
>>
Prefs
   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/gfp.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/compaction.h>
  36#include <linux/notifier.h>
  37#include <linux/rwsem.h>
  38#include <linux/delay.h>
  39#include <linux/kthread.h>
  40#include <linux/freezer.h>
  41#include <linux/memcontrol.h>
  42#include <linux/delayacct.h>
  43#include <linux/sysctl.h>
  44#include <linux/oom.h>
  45
  46#include <asm/tlbflush.h>
  47#include <asm/div64.h>
  48
  49#include <linux/swapops.h>
  50
  51#include "internal.h"
  52
  53#define CREATE_TRACE_POINTS
  54#include <trace/events/vmscan.h>
  55
  56/*
  57 * reclaim_mode determines how the inactive list is shrunk
  58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
  59 * RECLAIM_MODE_ASYNC:  Do not block
  60 * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
  61 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
  62 *                      page from the LRU and reclaim all pages within a
  63 *                      naturally aligned range
  64 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
  65 *                      order-0 pages and then compact the zone
  66 */
  67typedef unsigned __bitwise__ reclaim_mode_t;
  68#define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
  69#define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
  70#define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
  71#define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
  72#define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
  73
  74struct scan_control {
  75        /* Incremented by the number of inactive pages that were scanned */
  76        unsigned long nr_scanned;
  77
  78        /* Number of pages freed so far during a call to shrink_zones() */
  79        unsigned long nr_reclaimed;
  80
  81        /* How many pages shrink_list() should reclaim */
  82        unsigned long nr_to_reclaim;
  83
  84        unsigned long hibernation_mode;
  85
  86        /* This context's GFP mask */
  87        gfp_t gfp_mask;
  88
  89        int may_writepage;
  90
  91        /* Can mapped pages be reclaimed? */
  92        int may_unmap;
  93
  94        /* Can pages be swapped as part of reclaim? */
  95        int may_swap;
  96
  97        int swappiness;
  98
  99        int order;
 100
 101        /*
 102         * Intend to reclaim enough continuous memory rather than reclaim
 103         * enough amount of memory. i.e, mode for high order allocation.
 104         */
 105        reclaim_mode_t reclaim_mode;
 106
 107        /* Which cgroup do we reclaim from */
 108        struct mem_cgroup *mem_cgroup;
 109
 110        /*
 111         * Nodemask of nodes allowed by the caller. If NULL, all nodes
 112         * are scanned.
 113         */
 114        nodemask_t      *nodemask;
 115};
 116
 117#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
 118
 119#ifdef ARCH_HAS_PREFETCH
 120#define prefetch_prev_lru_page(_page, _base, _field)                    \
 121        do {                                                            \
 122                if ((_page)->lru.prev != _base) {                       \
 123                        struct page *prev;                              \
 124                                                                        \
 125                        prev = lru_to_page(&(_page->lru));              \
 126                        prefetch(&prev->_field);                        \
 127                }                                                       \
 128        } while (0)
 129#else
 130#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 131#endif
 132
 133#ifdef ARCH_HAS_PREFETCHW
 134#define prefetchw_prev_lru_page(_page, _base, _field)                   \
 135        do {                                                            \
 136                if ((_page)->lru.prev != _base) {                       \
 137                        struct page *prev;                              \
 138                                                                        \
 139                        prev = lru_to_page(&(_page->lru));              \
 140                        prefetchw(&prev->_field);                       \
 141                }                                                       \
 142        } while (0)
 143#else
 144#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 145#endif
 146
 147/*
 148 * From 0 .. 100.  Higher means more swappy.
 149 */
 150int vm_swappiness = 60;
 151long vm_total_pages;    /* The total number of pages which the VM controls */
 152
 153static LIST_HEAD(shrinker_list);
 154static DECLARE_RWSEM(shrinker_rwsem);
 155
 156#ifdef CONFIG_CGROUP_MEM_RES_CTLR
 157#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
 158#else
 159#define scanning_global_lru(sc) (1)
 160#endif
 161
 162static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
 163                                                  struct scan_control *sc)
 164{
 165        if (!scanning_global_lru(sc))
 166                return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
 167
 168        return &zone->reclaim_stat;
 169}
 170
 171static unsigned long zone_nr_lru_pages(struct zone *zone,
 172                                struct scan_control *sc, enum lru_list lru)
 173{
 174        if (!scanning_global_lru(sc))
 175                return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
 176
 177        return zone_page_state(zone, NR_LRU_BASE + lru);
 178}
 179
 180
 181/*
 182 * Add a shrinker callback to be called from the vm
 183 */
 184void register_shrinker(struct shrinker *shrinker)
 185{
 186        shrinker->nr = 0;
 187        down_write(&shrinker_rwsem);
 188        list_add_tail(&shrinker->list, &shrinker_list);
 189        up_write(&shrinker_rwsem);
 190}
 191EXPORT_SYMBOL(register_shrinker);
 192
 193/*
 194 * Remove one
 195 */
 196void unregister_shrinker(struct shrinker *shrinker)
 197{
 198        down_write(&shrinker_rwsem);
 199        list_del(&shrinker->list);
 200        up_write(&shrinker_rwsem);
 201}
 202EXPORT_SYMBOL(unregister_shrinker);
 203
 204#define SHRINK_BATCH 128
 205/*
 206 * Call the shrink functions to age shrinkable caches
 207 *
 208 * Here we assume it costs one seek to replace a lru page and that it also
 209 * takes a seek to recreate a cache object.  With this in mind we age equal
 210 * percentages of the lru and ageable caches.  This should balance the seeks
 211 * generated by these structures.
 212 *
 213 * If the vm encountered mapped pages on the LRU it increase the pressure on
 214 * slab to avoid swapping.
 215 *
 216 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 217 *
 218 * `lru_pages' represents the number of on-LRU pages in all the zones which
 219 * are eligible for the caller's allocation attempt.  It is used for balancing
 220 * slab reclaim versus page reclaim.
 221 *
 222 * Returns the number of slab objects which we shrunk.
 223 */
 224unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
 225                        unsigned long lru_pages)
 226{
 227        struct shrinker *shrinker;
 228        unsigned long ret = 0;
 229
 230        if (scanned == 0)
 231                scanned = SWAP_CLUSTER_MAX;
 232
 233        if (!down_read_trylock(&shrinker_rwsem)) {
 234                /* Assume we'll be able to shrink next time */
 235                ret = 1;
 236                goto out;
 237        }
 238
 239        list_for_each_entry(shrinker, &shrinker_list, list) {
 240                unsigned long long delta;
 241                unsigned long total_scan;
 242                unsigned long max_pass;
 243
 244                max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
 245                delta = (4 * scanned) / shrinker->seeks;
 246                delta *= max_pass;
 247                do_div(delta, lru_pages + 1);
 248                shrinker->nr += delta;
 249                if (shrinker->nr < 0) {
 250                        printk(KERN_ERR "shrink_slab: %pF negative objects to "
 251                               "delete nr=%ld\n",
 252                               shrinker->shrink, shrinker->nr);
 253                        shrinker->nr = max_pass;
 254                }
 255
 256                /*
 257                 * Avoid risking looping forever due to too large nr value:
 258                 * never try to free more than twice the estimate number of
 259                 * freeable entries.
 260                 */
 261                if (shrinker->nr > max_pass * 2)
 262                        shrinker->nr = max_pass * 2;
 263
 264                total_scan = shrinker->nr;
 265                shrinker->nr = 0;
 266
 267                while (total_scan >= SHRINK_BATCH) {
 268                        long this_scan = SHRINK_BATCH;
 269                        int shrink_ret;
 270                        int nr_before;
 271
 272                        nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
 273                        shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
 274                                                                gfp_mask);
 275                        if (shrink_ret == -1)
 276                                break;
 277                        if (shrink_ret < nr_before)
 278                                ret += nr_before - shrink_ret;
 279                        count_vm_events(SLABS_SCANNED, this_scan);
 280                        total_scan -= this_scan;
 281
 282                        cond_resched();
 283                }
 284
 285                shrinker->nr += total_scan;
 286        }
 287        up_read(&shrinker_rwsem);
 288out:
 289        cond_resched();
 290        return ret;
 291}
 292
 293static void set_reclaim_mode(int priority, struct scan_control *sc,
 294                                   bool sync)
 295{
 296        reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
 297
 298        /*
 299         * Initially assume we are entering either lumpy reclaim or
 300         * reclaim/compaction.Depending on the order, we will either set the
 301         * sync mode or just reclaim order-0 pages later.
 302         */
 303        if (COMPACTION_BUILD)
 304                sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
 305        else
 306                sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
 307
 308        /*
 309         * Avoid using lumpy reclaim or reclaim/compaction if possible by
 310         * restricting when its set to either costly allocations or when
 311         * under memory pressure
 312         */
 313        if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
 314                sc->reclaim_mode |= syncmode;
 315        else if (sc->order && priority < DEF_PRIORITY - 2)
 316                sc->reclaim_mode |= syncmode;
 317        else
 318                sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
 319}
 320
 321static void reset_reclaim_mode(struct scan_control *sc)
 322{
 323        sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
 324}
 325
 326static inline int is_page_cache_freeable(struct page *page)
 327{
 328        /*
 329         * A freeable page cache page is referenced only by the caller
 330         * that isolated the page, the page cache radix tree and
 331         * optional buffer heads at page->private.
 332         */
 333        return page_count(page) - page_has_private(page) == 2;
 334}
 335
 336static int may_write_to_queue(struct backing_dev_info *bdi,
 337                              struct scan_control *sc)
 338{
 339        if (current->flags & PF_SWAPWRITE)
 340                return 1;
 341        if (!bdi_write_congested(bdi))
 342                return 1;
 343        if (bdi == current->backing_dev_info)
 344                return 1;
 345
 346        /* lumpy reclaim for hugepage often need a lot of write */
 347        if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
 348                return 1;
 349        return 0;
 350}
 351
 352/*
 353 * We detected a synchronous write error writing a page out.  Probably
 354 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 355 * fsync(), msync() or close().
 356 *
 357 * The tricky part is that after writepage we cannot touch the mapping: nothing
 358 * prevents it from being freed up.  But we have a ref on the page and once
 359 * that page is locked, the mapping is pinned.
 360 *
 361 * We're allowed to run sleeping lock_page() here because we know the caller has
 362 * __GFP_FS.
 363 */
 364static void handle_write_error(struct address_space *mapping,
 365                                struct page *page, int error)
 366{
 367        lock_page(page);
 368        if (page_mapping(page) == mapping)
 369                mapping_set_error(mapping, error);
 370        unlock_page(page);
 371}
 372
 373/* possible outcome of pageout() */
 374typedef enum {
 375        /* failed to write page out, page is locked */
 376        PAGE_KEEP,
 377        /* move page to the active list, page is locked */
 378        PAGE_ACTIVATE,
 379        /* page has been sent to the disk successfully, page is unlocked */
 380        PAGE_SUCCESS,
 381        /* page is clean and locked */
 382        PAGE_CLEAN,
 383} pageout_t;
 384
 385/*
 386 * pageout is called by shrink_page_list() for each dirty page.
 387 * Calls ->writepage().
 388 */
 389static pageout_t pageout(struct page *page, struct address_space *mapping,
 390                         struct scan_control *sc)
 391{
 392        /*
 393         * If the page is dirty, only perform writeback if that write
 394         * will be non-blocking.  To prevent this allocation from being
 395         * stalled by pagecache activity.  But note that there may be
 396         * stalls if we need to run get_block().  We could test
 397         * PagePrivate for that.
 398         *
 399         * If this process is currently in __generic_file_aio_write() against
 400         * this page's queue, we can perform writeback even if that
 401         * will block.
 402         *
 403         * If the page is swapcache, write it back even if that would
 404         * block, for some throttling. This happens by accident, because
 405         * swap_backing_dev_info is bust: it doesn't reflect the
 406         * congestion state of the swapdevs.  Easy to fix, if needed.
 407         */
 408        if (!is_page_cache_freeable(page))
 409                return PAGE_KEEP;
 410        if (!mapping) {
 411                /*
 412                 * Some data journaling orphaned pages can have
 413                 * page->mapping == NULL while being dirty with clean buffers.
 414                 */
 415                if (page_has_private(page)) {
 416                        if (try_to_free_buffers(page)) {
 417                                ClearPageDirty(page);
 418                                printk("%s: orphaned page\n", __func__);
 419                                return PAGE_CLEAN;
 420                        }
 421                }
 422                return PAGE_KEEP;
 423        }
 424        if (mapping->a_ops->writepage == NULL)
 425                return PAGE_ACTIVATE;
 426        if (!may_write_to_queue(mapping->backing_dev_info, sc))
 427                return PAGE_KEEP;
 428
 429        if (clear_page_dirty_for_io(page)) {
 430                int res;
 431                struct writeback_control wbc = {
 432                        .sync_mode = WB_SYNC_NONE,
 433                        .nr_to_write = SWAP_CLUSTER_MAX,
 434                        .range_start = 0,
 435                        .range_end = LLONG_MAX,
 436                        .for_reclaim = 1,
 437                };
 438
 439                SetPageReclaim(page);
 440                res = mapping->a_ops->writepage(page, &wbc);
 441                if (res < 0)
 442                        handle_write_error(mapping, page, res);
 443                if (res == AOP_WRITEPAGE_ACTIVATE) {
 444                        ClearPageReclaim(page);
 445                        return PAGE_ACTIVATE;
 446                }
 447
 448                /*
 449                 * Wait on writeback if requested to. This happens when
 450                 * direct reclaiming a large contiguous area and the
 451                 * first attempt to free a range of pages fails.
 452                 */
 453                if (PageWriteback(page) &&
 454                    (sc->reclaim_mode & RECLAIM_MODE_SYNC))
 455                        wait_on_page_writeback(page);
 456
 457                if (!PageWriteback(page)) {
 458                        /* synchronous write or broken a_ops? */
 459                        ClearPageReclaim(page);
 460                }
 461                trace_mm_vmscan_writepage(page,
 462                        trace_reclaim_flags(page, sc->reclaim_mode));
 463                inc_zone_page_state(page, NR_VMSCAN_WRITE);
 464                return PAGE_SUCCESS;
 465        }
 466
 467        return PAGE_CLEAN;
 468}
 469
 470/*
 471 * Same as remove_mapping, but if the page is removed from the mapping, it
 472 * gets returned with a refcount of 0.
 473 */
 474static int __remove_mapping(struct address_space *mapping, struct page *page)
 475{
 476        BUG_ON(!PageLocked(page));
 477        BUG_ON(mapping != page_mapping(page));
 478
 479        spin_lock_irq(&mapping->tree_lock);
 480        /*
 481         * The non racy check for a busy page.
 482         *
 483         * Must be careful with the order of the tests. When someone has
 484         * a ref to the page, it may be possible that they dirty it then
 485         * drop the reference. So if PageDirty is tested before page_count
 486         * here, then the following race may occur:
 487         *
 488         * get_user_pages(&page);
 489         * [user mapping goes away]
 490         * write_to(page);
 491         *                              !PageDirty(page)    [good]
 492         * SetPageDirty(page);
 493         * put_page(page);
 494         *                              !page_count(page)   [good, discard it]
 495         *
 496         * [oops, our write_to data is lost]
 497         *
 498         * Reversing the order of the tests ensures such a situation cannot
 499         * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 500         * load is not satisfied before that of page->_count.
 501         *
 502         * Note that if SetPageDirty is always performed via set_page_dirty,
 503         * and thus under tree_lock, then this ordering is not required.
 504         */
 505        if (!page_freeze_refs(page, 2))
 506                goto cannot_free;
 507        /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 508        if (unlikely(PageDirty(page))) {
 509                page_unfreeze_refs(page, 2);
 510                goto cannot_free;
 511        }
 512
 513        if (PageSwapCache(page)) {
 514                swp_entry_t swap = { .val = page_private(page) };
 515                __delete_from_swap_cache(page);
 516                spin_unlock_irq(&mapping->tree_lock);
 517                swapcache_free(swap, page);
 518        } else {
 519                void (*freepage)(struct page *);
 520
 521                freepage = mapping->a_ops->freepage;
 522
 523                __delete_from_page_cache(page);
 524                spin_unlock_irq(&mapping->tree_lock);
 525                mem_cgroup_uncharge_cache_page(page);
 526
 527                if (freepage != NULL)
 528                        freepage(page);
 529        }
 530
 531        return 1;
 532
 533cannot_free:
 534        spin_unlock_irq(&mapping->tree_lock);
 535        return 0;
 536}
 537
 538/*
 539 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 540 * someone else has a ref on the page, abort and return 0.  If it was
 541 * successfully detached, return 1.  Assumes the caller has a single ref on
 542 * this page.
 543 */
 544int remove_mapping(struct address_space *mapping, struct page *page)
 545{
 546        if (__remove_mapping(mapping, page)) {
 547                /*
 548                 * Unfreezing the refcount with 1 rather than 2 effectively
 549                 * drops the pagecache ref for us without requiring another
 550                 * atomic operation.
 551                 */
 552                page_unfreeze_refs(page, 1);
 553                return 1;
 554        }
 555        return 0;
 556}
 557
 558/**
 559 * putback_lru_page - put previously isolated page onto appropriate LRU list
 560 * @page: page to be put back to appropriate lru list
 561 *
 562 * Add previously isolated @page to appropriate LRU list.
 563 * Page may still be unevictable for other reasons.
 564 *
 565 * lru_lock must not be held, interrupts must be enabled.
 566 */
 567void putback_lru_page(struct page *page)
 568{
 569        int lru;
 570        int active = !!TestClearPageActive(page);
 571        int was_unevictable = PageUnevictable(page);
 572
 573        VM_BUG_ON(PageLRU(page));
 574
 575redo:
 576        ClearPageUnevictable(page);
 577
 578        if (page_evictable(page, NULL)) {
 579                /*
 580                 * For evictable pages, we can use the cache.
 581                 * In event of a race, worst case is we end up with an
 582                 * unevictable page on [in]active list.
 583                 * We know how to handle that.
 584                 */
 585                lru = active + page_lru_base_type(page);
 586                lru_cache_add_lru(page, lru);
 587        } else {
 588                /*
 589                 * Put unevictable pages directly on zone's unevictable
 590                 * list.
 591                 */
 592                lru = LRU_UNEVICTABLE;
 593                add_page_to_unevictable_list(page);
 594                /*
 595                 * When racing with an mlock clearing (page is
 596                 * unlocked), make sure that if the other thread does
 597                 * not observe our setting of PG_lru and fails
 598                 * isolation, we see PG_mlocked cleared below and move
 599                 * the page back to the evictable list.
 600                 *
 601                 * The other side is TestClearPageMlocked().
 602                 */
 603                smp_mb();
 604        }
 605
 606        /*
 607         * page's status can change while we move it among lru. If an evictable
 608         * page is on unevictable list, it never be freed. To avoid that,
 609         * check after we added it to the list, again.
 610         */
 611        if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
 612                if (!isolate_lru_page(page)) {
 613                        put_page(page);
 614                        goto redo;
 615                }
 616                /* This means someone else dropped this page from LRU
 617                 * So, it will be freed or putback to LRU again. There is
 618                 * nothing to do here.
 619                 */
 620        }
 621
 622        if (was_unevictable && lru != LRU_UNEVICTABLE)
 623                count_vm_event(UNEVICTABLE_PGRESCUED);
 624        else if (!was_unevictable && lru == LRU_UNEVICTABLE)
 625                count_vm_event(UNEVICTABLE_PGCULLED);
 626
 627        put_page(page);         /* drop ref from isolate */
 628}
 629
 630enum page_references {
 631        PAGEREF_RECLAIM,
 632        PAGEREF_RECLAIM_CLEAN,
 633        PAGEREF_KEEP,
 634        PAGEREF_ACTIVATE,
 635};
 636
 637static enum page_references page_check_references(struct page *page,
 638                                                  struct scan_control *sc)
 639{
 640        int referenced_ptes, referenced_page;
 641        unsigned long vm_flags;
 642
 643        referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
 644        referenced_page = TestClearPageReferenced(page);
 645
 646        /* Lumpy reclaim - ignore references */
 647        if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
 648                return PAGEREF_RECLAIM;
 649
 650        /*
 651         * Mlock lost the isolation race with us.  Let try_to_unmap()
 652         * move the page to the unevictable list.
 653         */
 654        if (vm_flags & VM_LOCKED)
 655                return PAGEREF_RECLAIM;
 656
 657        if (referenced_ptes) {
 658                if (PageAnon(page))
 659                        return PAGEREF_ACTIVATE;
 660                /*
 661                 * All mapped pages start out with page table
 662                 * references from the instantiating fault, so we need
 663                 * to look twice if a mapped file page is used more
 664                 * than once.
 665                 *
 666                 * Mark it and spare it for another trip around the
 667                 * inactive list.  Another page table reference will
 668                 * lead to its activation.
 669                 *
 670                 * Note: the mark is set for activated pages as well
 671                 * so that recently deactivated but used pages are
 672                 * quickly recovered.
 673                 */
 674                SetPageReferenced(page);
 675
 676                if (referenced_page)
 677                        return PAGEREF_ACTIVATE;
 678
 679                return PAGEREF_KEEP;
 680        }
 681
 682        /* Reclaim if clean, defer dirty pages to writeback */
 683        if (referenced_page && !PageSwapBacked(page))
 684                return PAGEREF_RECLAIM_CLEAN;
 685
 686        return PAGEREF_RECLAIM;
 687}
 688
 689static noinline_for_stack void free_page_list(struct list_head *free_pages)
 690{
 691        struct pagevec freed_pvec;
 692        struct page *page, *tmp;
 693
 694        pagevec_init(&freed_pvec, 1);
 695
 696        list_for_each_entry_safe(page, tmp, free_pages, lru) {
 697                list_del(&page->lru);
 698                if (!pagevec_add(&freed_pvec, page)) {
 699                        __pagevec_free(&freed_pvec);
 700                        pagevec_reinit(&freed_pvec);
 701                }
 702        }
 703
 704        pagevec_free(&freed_pvec);
 705}
 706
 707/*
 708 * shrink_page_list() returns the number of reclaimed pages
 709 */
 710static unsigned long shrink_page_list(struct list_head *page_list,
 711                                      struct zone *zone,
 712                                      struct scan_control *sc)
 713{
 714        LIST_HEAD(ret_pages);
 715        LIST_HEAD(free_pages);
 716        int pgactivate = 0;
 717        unsigned long nr_dirty = 0;
 718        unsigned long nr_congested = 0;
 719        unsigned long nr_reclaimed = 0;
 720
 721        cond_resched();
 722
 723        while (!list_empty(page_list)) {
 724                enum page_references references;
 725                struct address_space *mapping;
 726                struct page *page;
 727                int may_enter_fs;
 728
 729                cond_resched();
 730
 731                page = lru_to_page(page_list);
 732                list_del(&page->lru);
 733
 734                if (!trylock_page(page))
 735                        goto keep;
 736
 737                VM_BUG_ON(PageActive(page));
 738                VM_BUG_ON(page_zone(page) != zone);
 739
 740                sc->nr_scanned++;
 741
 742                if (unlikely(!page_evictable(page, NULL)))
 743                        goto cull_mlocked;
 744
 745                if (!sc->may_unmap && page_mapped(page))
 746                        goto keep_locked;
 747
 748                /* Double the slab pressure for mapped and swapcache pages */
 749                if (page_mapped(page) || PageSwapCache(page))
 750                        sc->nr_scanned++;
 751
 752                may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 753                        (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 754
 755                if (PageWriteback(page)) {
 756                        /*
 757                         * Synchronous reclaim is performed in two passes,
 758                         * first an asynchronous pass over the list to
 759                         * start parallel writeback, and a second synchronous
 760                         * pass to wait for the IO to complete.  Wait here
 761                         * for any page for which writeback has already
 762                         * started.
 763                         */
 764                        if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
 765                            may_enter_fs)
 766                                wait_on_page_writeback(page);
 767                        else {
 768                                unlock_page(page);
 769                                goto keep_lumpy;
 770                        }
 771                }
 772
 773                references = page_check_references(page, sc);
 774                switch (references) {
 775                case PAGEREF_ACTIVATE:
 776                        goto activate_locked;
 777                case PAGEREF_KEEP:
 778                        goto keep_locked;
 779                case PAGEREF_RECLAIM:
 780                case PAGEREF_RECLAIM_CLEAN:
 781                        ; /* try to reclaim the page below */
 782                }
 783
 784                /*
 785                 * Anonymous process memory has backing store?
 786                 * Try to allocate it some swap space here.
 787                 */
 788                if (PageAnon(page) && !PageSwapCache(page)) {
 789                        if (!(sc->gfp_mask & __GFP_IO))
 790                                goto keep_locked;
 791                        if (!add_to_swap(page))
 792                                goto activate_locked;
 793                        may_enter_fs = 1;
 794                }
 795
 796                mapping = page_mapping(page);
 797
 798                /*
 799                 * The page is mapped into the page tables of one or more
 800                 * processes. Try to unmap it here.
 801                 */
 802                if (page_mapped(page) && mapping) {
 803                        switch (try_to_unmap(page, TTU_UNMAP)) {
 804                        case SWAP_FAIL:
 805                                goto activate_locked;
 806                        case SWAP_AGAIN:
 807                                goto keep_locked;
 808                        case SWAP_MLOCK:
 809                                goto cull_mlocked;
 810                        case SWAP_SUCCESS:
 811                                ; /* try to free the page below */
 812                        }
 813                }
 814
 815                if (PageDirty(page)) {
 816                        nr_dirty++;
 817
 818                        if (references == PAGEREF_RECLAIM_CLEAN)
 819                                goto keep_locked;
 820                        if (!may_enter_fs)
 821                                goto keep_locked;
 822                        if (!sc->may_writepage)
 823                                goto keep_locked;
 824
 825                        /* Page is dirty, try to write it out here */
 826                        switch (pageout(page, mapping, sc)) {
 827                        case PAGE_KEEP:
 828                                nr_congested++;
 829                                goto keep_locked;
 830                        case PAGE_ACTIVATE:
 831                                goto activate_locked;
 832                        case PAGE_SUCCESS:
 833                                if (PageWriteback(page))
 834                                        goto keep_lumpy;
 835                                if (PageDirty(page))
 836                                        goto keep;
 837
 838                                /*
 839                                 * A synchronous write - probably a ramdisk.  Go
 840                                 * ahead and try to reclaim the page.
 841                                 */
 842                                if (!trylock_page(page))
 843                                        goto keep;
 844                                if (PageDirty(page) || PageWriteback(page))
 845                                        goto keep_locked;
 846                                mapping = page_mapping(page);
 847                        case PAGE_CLEAN:
 848                                ; /* try to free the page below */
 849                        }
 850                }
 851
 852                /*
 853                 * If the page has buffers, try to free the buffer mappings
 854                 * associated with this page. If we succeed we try to free
 855                 * the page as well.
 856                 *
 857                 * We do this even if the page is PageDirty().
 858                 * try_to_release_page() does not perform I/O, but it is
 859                 * possible for a page to have PageDirty set, but it is actually
 860                 * clean (all its buffers are clean).  This happens if the
 861                 * buffers were written out directly, with submit_bh(). ext3
 862                 * will do this, as well as the blockdev mapping.
 863                 * try_to_release_page() will discover that cleanness and will
 864                 * drop the buffers and mark the page clean - it can be freed.
 865                 *
 866                 * Rarely, pages can have buffers and no ->mapping.  These are
 867                 * the pages which were not successfully invalidated in
 868                 * truncate_complete_page().  We try to drop those buffers here
 869                 * and if that worked, and the page is no longer mapped into
 870                 * process address space (page_count == 1) it can be freed.
 871                 * Otherwise, leave the page on the LRU so it is swappable.
 872                 */
 873                if (page_has_private(page)) {
 874                        if (!try_to_release_page(page, sc->gfp_mask))
 875                                goto activate_locked;
 876                        if (!mapping && page_count(page) == 1) {
 877                                unlock_page(page);
 878                                if (put_page_testzero(page))
 879                                        goto free_it;
 880                                else {
 881                                        /*
 882                                         * rare race with speculative reference.
 883                                         * the speculative reference will free
 884                                         * this page shortly, so we may
 885                                         * increment nr_reclaimed here (and
 886                                         * leave it off the LRU).
 887                                         */
 888                                        nr_reclaimed++;
 889                                        continue;
 890                                }
 891                        }
 892                }
 893
 894                if (!mapping || !__remove_mapping(mapping, page))
 895                        goto keep_locked;
 896
 897                /*
 898                 * At this point, we have no other references and there is
 899                 * no way to pick any more up (removed from LRU, removed
 900                 * from pagecache). Can use non-atomic bitops now (and
 901                 * we obviously don't have to worry about waking up a process
 902                 * waiting on the page lock, because there are no references.
 903                 */
 904                __clear_page_locked(page);
 905free_it:
 906                nr_reclaimed++;
 907
 908                /*
 909                 * Is there need to periodically free_page_list? It would
 910                 * appear not as the counts should be low
 911                 */
 912                list_add(&page->lru, &free_pages);
 913                continue;
 914
 915cull_mlocked:
 916                if (PageSwapCache(page))
 917                        try_to_free_swap(page);
 918                unlock_page(page);
 919                putback_lru_page(page);
 920                reset_reclaim_mode(sc);
 921                continue;
 922
 923activate_locked:
 924                /* Not a candidate for swapping, so reclaim swap space. */
 925                if (PageSwapCache(page) && vm_swap_full())
 926                        try_to_free_swap(page);
 927                VM_BUG_ON(PageActive(page));
 928                SetPageActive(page);
 929                pgactivate++;
 930keep_locked:
 931                unlock_page(page);
 932keep:
 933                reset_reclaim_mode(sc);
 934keep_lumpy:
 935                list_add(&page->lru, &ret_pages);
 936                VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
 937        }
 938
 939        /*
 940         * Tag a zone as congested if all the dirty pages encountered were
 941         * backed by a congested BDI. In this case, reclaimers should just
 942         * back off and wait for congestion to clear because further reclaim
 943         * will encounter the same problem
 944         */
 945        if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
 946                zone_set_flag(zone, ZONE_CONGESTED);
 947
 948        free_page_list(&free_pages);
 949
 950        list_splice(&ret_pages, page_list);
 951        count_vm_events(PGACTIVATE, pgactivate);
 952        return nr_reclaimed;
 953}
 954
 955/*
 956 * Attempt to remove the specified page from its LRU.  Only take this page
 957 * if it is of the appropriate PageActive status.  Pages which are being
 958 * freed elsewhere are also ignored.
 959 *
 960 * page:        page to consider
 961 * mode:        one of the LRU isolation modes defined above
 962 *
 963 * returns 0 on success, -ve errno on failure.
 964 */
 965int __isolate_lru_page(struct page *page, int mode, int file)
 966{
 967        int ret = -EINVAL;
 968
 969        /* Only take pages on the LRU. */
 970        if (!PageLRU(page))
 971                return ret;
 972
 973        /*
 974         * When checking the active state, we need to be sure we are
 975         * dealing with comparible boolean values.  Take the logical not
 976         * of each.
 977         */
 978        if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
 979                return ret;
 980
 981        if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
 982                return ret;
 983
 984        /*
 985         * When this function is being called for lumpy reclaim, we
 986         * initially look into all LRU pages, active, inactive and
 987         * unevictable; only give shrink_page_list evictable pages.
 988         */
 989        if (PageUnevictable(page))
 990                return ret;
 991
 992        ret = -EBUSY;
 993
 994        if (likely(get_page_unless_zero(page))) {
 995                /*
 996                 * Be careful not to clear PageLRU until after we're
 997                 * sure the page is not being freed elsewhere -- the
 998                 * page release code relies on it.
 999                 */
1000                ClearPageLRU(page);
1001                ret = 0;
1002        }
1003
1004        return ret;
1005}
1006
1007/*
1008 * zone->lru_lock is heavily contended.  Some of the functions that
1009 * shrink the lists perform better by taking out a batch of pages
1010 * and working on them outside the LRU lock.
1011 *
1012 * For pagecache intensive workloads, this function is the hottest
1013 * spot in the kernel (apart from copy_*_user functions).
1014 *
1015 * Appropriate locks must be held before calling this function.
1016 *
1017 * @nr_to_scan: The number of pages to look through on the list.
1018 * @src:        The LRU list to pull pages off.
1019 * @dst:        The temp list to put pages on to.
1020 * @scanned:    The number of pages that were scanned.
1021 * @order:      The caller's attempted allocation order
1022 * @mode:       One of the LRU isolation modes
1023 * @file:       True [1] if isolating file [!anon] pages
1024 *
1025 * returns how many pages were moved onto *@dst.
1026 */
1027static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1028                struct list_head *src, struct list_head *dst,
1029                unsigned long *scanned, int order, int mode, int file)
1030{
1031        unsigned long nr_taken = 0;
1032        unsigned long nr_lumpy_taken = 0;
1033        unsigned long nr_lumpy_dirty = 0;
1034        unsigned long nr_lumpy_failed = 0;
1035        unsigned long scan;
1036
1037        for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1038                struct page *page;
1039                unsigned long pfn;
1040                unsigned long end_pfn;
1041                unsigned long page_pfn;
1042                int zone_id;
1043
1044                page = lru_to_page(src);
1045                prefetchw_prev_lru_page(page, src, flags);
1046
1047                VM_BUG_ON(!PageLRU(page));
1048
1049                switch (__isolate_lru_page(page, mode, file)) {
1050                case 0:
1051                        list_move(&page->lru, dst);
1052                        mem_cgroup_del_lru(page);
1053                        nr_taken += hpage_nr_pages(page);
1054                        break;
1055
1056                case -EBUSY:
1057                        /* else it is being freed elsewhere */
1058                        list_move(&page->lru, src);
1059                        mem_cgroup_rotate_lru_list(page, page_lru(page));
1060                        continue;
1061
1062                default:
1063                        BUG();
1064                }
1065
1066                if (!order)
1067                        continue;
1068
1069                /*
1070                 * Attempt to take all pages in the order aligned region
1071                 * surrounding the tag page.  Only take those pages of
1072                 * the same active state as that tag page.  We may safely
1073                 * round the target page pfn down to the requested order
1074                 * as the mem_map is guaranteed valid out to MAX_ORDER,
1075                 * where that page is in a different zone we will detect
1076                 * it from its zone id and abort this block scan.
1077                 */
1078                zone_id = page_zone_id(page);
1079                page_pfn = page_to_pfn(page);
1080                pfn = page_pfn & ~((1 << order) - 1);
1081                end_pfn = pfn + (1 << order);
1082                for (; pfn < end_pfn; pfn++) {
1083                        struct page *cursor_page;
1084
1085                        /* The target page is in the block, ignore it. */
1086                        if (unlikely(pfn == page_pfn))
1087                                continue;
1088
1089                        /* Avoid holes within the zone. */
1090                        if (unlikely(!pfn_valid_within(pfn)))
1091                                break;
1092
1093                        cursor_page = pfn_to_page(pfn);
1094
1095                        /* Check that we have not crossed a zone boundary. */
1096                        if (unlikely(page_zone_id(cursor_page) != zone_id))
1097                                break;
1098
1099                        /*
1100                         * If we don't have enough swap space, reclaiming of
1101                         * anon page which don't already have a swap slot is
1102                         * pointless.
1103                         */
1104                        if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1105                            !PageSwapCache(cursor_page))
1106                                break;
1107
1108                        if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1109                                list_move(&cursor_page->lru, dst);
1110                                mem_cgroup_del_lru(cursor_page);
1111                                nr_taken += hpage_nr_pages(page);
1112                                nr_lumpy_taken++;
1113                                if (PageDirty(cursor_page))
1114                                        nr_lumpy_dirty++;
1115                                scan++;
1116                        } else {
1117                                /*
1118                                 * Check if the page is freed already.
1119                                 *
1120                                 * We can't use page_count() as that
1121                                 * requires compound_head and we don't
1122                                 * have a pin on the page here. If a
1123                                 * page is tail, we may or may not
1124                                 * have isolated the head, so assume
1125                                 * it's not free, it'd be tricky to
1126                                 * track the head status without a
1127                                 * page pin.
1128                                 */
1129                                if (!PageTail(cursor_page) &&
1130                                    !atomic_read(&cursor_page->_count))
1131                                        continue;
1132                                break;
1133                        }
1134                }
1135
1136                /* If we break out of the loop above, lumpy reclaim failed */
1137                if (pfn < end_pfn)
1138                        nr_lumpy_failed++;
1139        }
1140
1141        *scanned = scan;
1142
1143        trace_mm_vmscan_lru_isolate(order,
1144                        nr_to_scan, scan,
1145                        nr_taken,
1146                        nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1147                        mode);
1148        return nr_taken;
1149}
1150
1151static unsigned long isolate_pages_global(unsigned long nr,
1152                                        struct list_head *dst,
1153                                        unsigned long *scanned, int order,
1154                                        int mode, struct zone *z,
1155                                        int active, int file)
1156{
1157        int lru = LRU_BASE;
1158        if (active)
1159                lru += LRU_ACTIVE;
1160        if (file)
1161                lru += LRU_FILE;
1162        return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1163                                                                mode, file);
1164}
1165
1166/*
1167 * clear_active_flags() is a helper for shrink_active_list(), clearing
1168 * any active bits from the pages in the list.
1169 */
1170static unsigned long clear_active_flags(struct list_head *page_list,
1171                                        unsigned int *count)
1172{
1173        int nr_active = 0;
1174        int lru;
1175        struct page *page;
1176
1177        list_for_each_entry(page, page_list, lru) {
1178                int numpages = hpage_nr_pages(page);
1179                lru = page_lru_base_type(page);
1180                if (PageActive(page)) {
1181                        lru += LRU_ACTIVE;
1182                        ClearPageActive(page);
1183                        nr_active += numpages;
1184                }
1185                if (count)
1186                        count[lru] += numpages;
1187        }
1188
1189        return nr_active;
1190}
1191
1192/**
1193 * isolate_lru_page - tries to isolate a page from its LRU list
1194 * @page: page to isolate from its LRU list
1195 *
1196 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1197 * vmstat statistic corresponding to whatever LRU list the page was on.
1198 *
1199 * Returns 0 if the page was removed from an LRU list.
1200 * Returns -EBUSY if the page was not on an LRU list.
1201 *
1202 * The returned page will have PageLRU() cleared.  If it was found on
1203 * the active list, it will have PageActive set.  If it was found on
1204 * the unevictable list, it will have the PageUnevictable bit set. That flag
1205 * may need to be cleared by the caller before letting the page go.
1206 *
1207 * The vmstat statistic corresponding to the list on which the page was
1208 * found will be decremented.
1209 *
1210 * Restrictions:
1211 * (1) Must be called with an elevated refcount on the page. This is a
1212 *     fundamentnal difference from isolate_lru_pages (which is called
1213 *     without a stable reference).
1214 * (2) the lru_lock must not be held.
1215 * (3) interrupts must be enabled.
1216 */
1217int isolate_lru_page(struct page *page)
1218{
1219        int ret = -EBUSY;
1220
1221        if (PageLRU(page)) {
1222                struct zone *zone = page_zone(page);
1223
1224                spin_lock_irq(&zone->lru_lock);
1225                if (PageLRU(page) && get_page_unless_zero(page)) {
1226                        int lru = page_lru(page);
1227                        ret = 0;
1228                        ClearPageLRU(page);
1229
1230                        del_page_from_lru_list(zone, page, lru);
1231                }
1232                spin_unlock_irq(&zone->lru_lock);
1233        }
1234        return ret;
1235}
1236
1237/*
1238 * Are there way too many processes in the direct reclaim path already?
1239 */
1240static int too_many_isolated(struct zone *zone, int file,
1241                struct scan_control *sc)
1242{
1243        unsigned long inactive, isolated;
1244
1245        if (current_is_kswapd())
1246                return 0;
1247
1248        if (!scanning_global_lru(sc))
1249                return 0;
1250
1251        if (file) {
1252                inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1253                isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1254        } else {
1255                inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1256                isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1257        }
1258
1259        return isolated > inactive;
1260}
1261
1262/*
1263 * TODO: Try merging with migrations version of putback_lru_pages
1264 */
1265static noinline_for_stack void
1266putback_lru_pages(struct zone *zone, struct scan_control *sc,
1267                                unsigned long nr_anon, unsigned long nr_file,
1268                                struct list_head *page_list)
1269{
1270        struct page *page;
1271        struct pagevec pvec;
1272        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1273
1274        pagevec_init(&pvec, 1);
1275
1276        /*
1277         * Put back any unfreeable pages.
1278         */
1279        spin_lock(&zone->lru_lock);
1280        while (!list_empty(page_list)) {
1281                int lru;
1282                page = lru_to_page(page_list);
1283                VM_BUG_ON(PageLRU(page));
1284                list_del(&page->lru);
1285                if (unlikely(!page_evictable(page, NULL))) {
1286                        spin_unlock_irq(&zone->lru_lock);
1287                        putback_lru_page(page);
1288                        spin_lock_irq(&zone->lru_lock);
1289                        continue;
1290                }
1291                SetPageLRU(page);
1292                lru = page_lru(page);
1293                add_page_to_lru_list(zone, page, lru);
1294                if (is_active_lru(lru)) {
1295                        int file = is_file_lru(lru);
1296                        int numpages = hpage_nr_pages(page);
1297                        reclaim_stat->recent_rotated[file] += numpages;
1298                }
1299                if (!pagevec_add(&pvec, page)) {
1300                        spin_unlock_irq(&zone->lru_lock);
1301                        __pagevec_release(&pvec);
1302                        spin_lock_irq(&zone->lru_lock);
1303                }
1304        }
1305        __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1306        __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1307
1308        spin_unlock_irq(&zone->lru_lock);
1309        pagevec_release(&pvec);
1310}
1311
1312static noinline_for_stack void update_isolated_counts(struct zone *zone,
1313                                        struct scan_control *sc,
1314                                        unsigned long *nr_anon,
1315                                        unsigned long *nr_file,
1316                                        struct list_head *isolated_list)
1317{
1318        unsigned long nr_active;
1319        unsigned int count[NR_LRU_LISTS] = { 0, };
1320        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1321
1322        nr_active = clear_active_flags(isolated_list, count);
1323        __count_vm_events(PGDEACTIVATE, nr_active);
1324
1325        __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1326                              -count[LRU_ACTIVE_FILE]);
1327        __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1328                              -count[LRU_INACTIVE_FILE]);
1329        __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1330                              -count[LRU_ACTIVE_ANON]);
1331        __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1332                              -count[LRU_INACTIVE_ANON]);
1333
1334        *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1335        *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1336        __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1337        __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1338
1339        reclaim_stat->recent_scanned[0] += *nr_anon;
1340        reclaim_stat->recent_scanned[1] += *nr_file;
1341}
1342
1343/*
1344 * Returns true if the caller should wait to clean dirty/writeback pages.
1345 *
1346 * If we are direct reclaiming for contiguous pages and we do not reclaim
1347 * everything in the list, try again and wait for writeback IO to complete.
1348 * This will stall high-order allocations noticeably. Only do that when really
1349 * need to free the pages under high memory pressure.
1350 */
1351static inline bool should_reclaim_stall(unsigned long nr_taken,
1352                                        unsigned long nr_freed,
1353                                        int priority,
1354                                        struct scan_control *sc)
1355{
1356        int lumpy_stall_priority;
1357
1358        /* kswapd should not stall on sync IO */
1359        if (current_is_kswapd())
1360                return false;
1361
1362        /* Only stall on lumpy reclaim */
1363        if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1364                return false;
1365
1366        /* If we have relaimed everything on the isolated list, no stall */
1367        if (nr_freed == nr_taken)
1368                return false;
1369
1370        /*
1371         * For high-order allocations, there are two stall thresholds.
1372         * High-cost allocations stall immediately where as lower
1373         * order allocations such as stacks require the scanning
1374         * priority to be much higher before stalling.
1375         */
1376        if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1377                lumpy_stall_priority = DEF_PRIORITY;
1378        else
1379                lumpy_stall_priority = DEF_PRIORITY / 3;
1380
1381        return priority <= lumpy_stall_priority;
1382}
1383
1384/*
1385 * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1386 * of reclaimed pages
1387 */
1388static noinline_for_stack unsigned long
1389shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1390                        struct scan_control *sc, int priority, int file)
1391{
1392        LIST_HEAD(page_list);
1393        unsigned long nr_scanned;
1394        unsigned long nr_reclaimed = 0;
1395        unsigned long nr_taken;
1396        unsigned long nr_anon;
1397        unsigned long nr_file;
1398
1399        while (unlikely(too_many_isolated(zone, file, sc))) {
1400                congestion_wait(BLK_RW_ASYNC, HZ/10);
1401
1402                /* We are about to die and free our memory. Return now. */
1403                if (fatal_signal_pending(current))
1404                        return SWAP_CLUSTER_MAX;
1405        }
1406
1407        set_reclaim_mode(priority, sc, false);
1408        lru_add_drain();
1409        spin_lock_irq(&zone->lru_lock);
1410
1411        if (scanning_global_lru(sc)) {
1412                nr_taken = isolate_pages_global(nr_to_scan,
1413                        &page_list, &nr_scanned, sc->order,
1414                        sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1415                                        ISOLATE_BOTH : ISOLATE_INACTIVE,
1416                        zone, 0, file);
1417                zone->pages_scanned += nr_scanned;
1418                if (current_is_kswapd())
1419                        __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1420                                               nr_scanned);
1421                else
1422                        __count_zone_vm_events(PGSCAN_DIRECT, zone,
1423                                               nr_scanned);
1424        } else {
1425                nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1426                        &page_list, &nr_scanned, sc->order,
1427                        sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1428                                        ISOLATE_BOTH : ISOLATE_INACTIVE,
1429                        zone, sc->mem_cgroup,
1430                        0, file);
1431                /*
1432                 * mem_cgroup_isolate_pages() keeps track of
1433                 * scanned pages on its own.
1434                 */
1435        }
1436
1437        if (nr_taken == 0) {
1438                spin_unlock_irq(&zone->lru_lock);
1439                return 0;
1440        }
1441
1442        update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1443
1444        spin_unlock_irq(&zone->lru_lock);
1445
1446        nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1447
1448        /* Check if we should syncronously wait for writeback */
1449        if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1450                set_reclaim_mode(priority, sc, true);
1451                nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1452        }
1453
1454        local_irq_disable();
1455        if (current_is_kswapd())
1456                __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1457        __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1458
1459        putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1460
1461        trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1462                zone_idx(zone),
1463                nr_scanned, nr_reclaimed,
1464                priority,
1465                trace_shrink_flags(file, sc->reclaim_mode));
1466        return nr_reclaimed;
1467}
1468
1469/*
1470 * This moves pages from the active list to the inactive list.
1471 *
1472 * We move them the other way if the page is referenced by one or more
1473 * processes, from rmap.
1474 *
1475 * If the pages are mostly unmapped, the processing is fast and it is
1476 * appropriate to hold zone->lru_lock across the whole operation.  But if
1477 * the pages are mapped, the processing is slow (page_referenced()) so we
1478 * should drop zone->lru_lock around each page.  It's impossible to balance
1479 * this, so instead we remove the pages from the LRU while processing them.
1480 * It is safe to rely on PG_active against the non-LRU pages in here because
1481 * nobody will play with that bit on a non-LRU page.
1482 *
1483 * The downside is that we have to touch page->_count against each page.
1484 * But we had to alter page->flags anyway.
1485 */
1486
1487static void move_active_pages_to_lru(struct zone *zone,
1488                                     struct list_head *list,
1489                                     enum lru_list lru)
1490{
1491        unsigned long pgmoved = 0;
1492        struct pagevec pvec;
1493        struct page *page;
1494
1495        pagevec_init(&pvec, 1);
1496
1497        while (!list_empty(list)) {
1498                page = lru_to_page(list);
1499
1500                VM_BUG_ON(PageLRU(page));
1501                SetPageLRU(page);
1502
1503                list_move(&page->lru, &zone->lru[lru].list);
1504                mem_cgroup_add_lru_list(page, lru);
1505                pgmoved += hpage_nr_pages(page);
1506
1507                if (!pagevec_add(&pvec, page) || list_empty(list)) {
1508                        spin_unlock_irq(&zone->lru_lock);
1509                        if (buffer_heads_over_limit)
1510                                pagevec_strip(&pvec);
1511                        __pagevec_release(&pvec);
1512                        spin_lock_irq(&zone->lru_lock);
1513                }
1514        }
1515        __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1516        if (!is_active_lru(lru))
1517                __count_vm_events(PGDEACTIVATE, pgmoved);
1518}
1519
1520static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1521                        struct scan_control *sc, int priority, int file)
1522{
1523        unsigned long nr_taken;
1524        unsigned long pgscanned;
1525        unsigned long vm_flags;
1526        LIST_HEAD(l_hold);      /* The pages which were snipped off */
1527        LIST_HEAD(l_active);
1528        LIST_HEAD(l_inactive);
1529        struct page *page;
1530        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1531        unsigned long nr_rotated = 0;
1532
1533        lru_add_drain();
1534        spin_lock_irq(&zone->lru_lock);
1535        if (scanning_global_lru(sc)) {
1536                nr_taken = isolate_pages_global(nr_pages, &l_hold,
1537                                                &pgscanned, sc->order,
1538                                                ISOLATE_ACTIVE, zone,
1539                                                1, file);
1540                zone->pages_scanned += pgscanned;
1541        } else {
1542                nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1543                                                &pgscanned, sc->order,
1544                                                ISOLATE_ACTIVE, zone,
1545                                                sc->mem_cgroup, 1, file);
1546                /*
1547                 * mem_cgroup_isolate_pages() keeps track of
1548                 * scanned pages on its own.
1549                 */
1550        }
1551
1552        reclaim_stat->recent_scanned[file] += nr_taken;
1553
1554        __count_zone_vm_events(PGREFILL, zone, pgscanned);
1555        if (file)
1556                __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1557        else
1558                __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1559        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1560        spin_unlock_irq(&zone->lru_lock);
1561
1562        while (!list_empty(&l_hold)) {
1563                cond_resched();
1564                page = lru_to_page(&l_hold);
1565                list_del(&page->lru);
1566
1567                if (unlikely(!page_evictable(page, NULL))) {
1568                        putback_lru_page(page);
1569                        continue;
1570                }
1571
1572                if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1573                        nr_rotated += hpage_nr_pages(page);
1574                        /*
1575                         * Identify referenced, file-backed active pages and
1576                         * give them one more trip around the active list. So
1577                         * that executable code get better chances to stay in
1578                         * memory under moderate memory pressure.  Anon pages
1579                         * are not likely to be evicted by use-once streaming
1580                         * IO, plus JVM can create lots of anon VM_EXEC pages,
1581                         * so we ignore them here.
1582                         */
1583                        if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1584                                list_add(&page->lru, &l_active);
1585                                continue;
1586                        }
1587                }
1588
1589                ClearPageActive(page);  /* we are de-activating */
1590                list_add(&page->lru, &l_inactive);
1591        }
1592
1593        /*
1594         * Move pages back to the lru list.
1595         */
1596        spin_lock_irq(&zone->lru_lock);
1597        /*
1598         * Count referenced pages from currently used mappings as rotated,
1599         * even though only some of them are actually re-activated.  This
1600         * helps balance scan pressure between file and anonymous pages in
1601         * get_scan_ratio.
1602         */
1603        reclaim_stat->recent_rotated[file] += nr_rotated;
1604
1605        move_active_pages_to_lru(zone, &l_active,
1606                                                LRU_ACTIVE + file * LRU_FILE);
1607        move_active_pages_to_lru(zone, &l_inactive,
1608                                                LRU_BASE   + file * LRU_FILE);
1609        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1610        spin_unlock_irq(&zone->lru_lock);
1611}
1612
1613#ifdef CONFIG_SWAP
1614static int inactive_anon_is_low_global(struct zone *zone)
1615{
1616        unsigned long active, inactive;
1617
1618        active = zone_page_state(zone, NR_ACTIVE_ANON);
1619        inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1620
1621        if (inactive * zone->inactive_ratio < active)
1622                return 1;
1623
1624        return 0;
1625}
1626
1627/**
1628 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1629 * @zone: zone to check
1630 * @sc:   scan control of this context
1631 *
1632 * Returns true if the zone does not have enough inactive anon pages,
1633 * meaning some active anon pages need to be deactivated.
1634 */
1635static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1636{
1637        int low;
1638
1639        /*
1640         * If we don't have swap space, anonymous page deactivation
1641         * is pointless.
1642         */
1643        if (!total_swap_pages)
1644                return 0;
1645
1646        if (scanning_global_lru(sc))
1647                low = inactive_anon_is_low_global(zone);
1648        else
1649                low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1650        return low;
1651}
1652#else
1653static inline int inactive_anon_is_low(struct zone *zone,
1654                                        struct scan_control *sc)
1655{
1656        return 0;
1657}
1658#endif
1659
1660static int inactive_file_is_low_global(struct zone *zone)
1661{
1662        unsigned long active, inactive;
1663
1664        active = zone_page_state(zone, NR_ACTIVE_FILE);
1665        inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1666
1667        return (active > inactive);
1668}
1669
1670/**
1671 * inactive_file_is_low - check if file pages need to be deactivated
1672 * @zone: zone to check
1673 * @sc:   scan control of this context
1674 *
1675 * When the system is doing streaming IO, memory pressure here
1676 * ensures that active file pages get deactivated, until more
1677 * than half of the file pages are on the inactive list.
1678 *
1679 * Once we get to that situation, protect the system's working
1680 * set from being evicted by disabling active file page aging.
1681 *
1682 * This uses a different ratio than the anonymous pages, because
1683 * the page cache uses a use-once replacement algorithm.
1684 */
1685static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1686{
1687        int low;
1688
1689        if (scanning_global_lru(sc))
1690                low = inactive_file_is_low_global(zone);
1691        else
1692                low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1693        return low;
1694}
1695
1696static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1697                                int file)
1698{
1699        if (file)
1700                return inactive_file_is_low(zone, sc);
1701        else
1702                return inactive_anon_is_low(zone, sc);
1703}
1704
1705static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1706        struct zone *zone, struct scan_control *sc, int priority)
1707{
1708        int file = is_file_lru(lru);
1709
1710        if (is_active_lru(lru)) {
1711                if (inactive_list_is_low(zone, sc, file))
1712                    shrink_active_list(nr_to_scan, zone, sc, priority, file);
1713                return 0;
1714        }
1715
1716        return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1717}
1718
1719/*
1720 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1721 * until we collected @swap_cluster_max pages to scan.
1722 */
1723static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1724                                       unsigned long *nr_saved_scan)
1725{
1726        unsigned long nr;
1727
1728        *nr_saved_scan += nr_to_scan;
1729        nr = *nr_saved_scan;
1730
1731        if (nr >= SWAP_CLUSTER_MAX)
1732                *nr_saved_scan = 0;
1733        else
1734                nr = 0;
1735
1736        return nr;
1737}
1738
1739/*
1740 * Determine how aggressively the anon and file LRU lists should be
1741 * scanned.  The relative value of each set of LRU lists is determined
1742 * by looking at the fraction of the pages scanned we did rotate back
1743 * onto the active list instead of evict.
1744 *
1745 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1746 */
1747static void get_scan_count(struct zone *zone, struct scan_control *sc,
1748                                        unsigned long *nr, int priority)
1749{
1750        unsigned long anon, file, free;
1751        unsigned long anon_prio, file_prio;
1752        unsigned long ap, fp;
1753        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1754        u64 fraction[2], denominator;
1755        enum lru_list l;
1756        int noswap = 0;
1757
1758        /* If we have no swap space, do not bother scanning anon pages. */
1759        if (!sc->may_swap || (nr_swap_pages <= 0)) {
1760                noswap = 1;
1761                fraction[0] = 0;
1762                fraction[1] = 1;
1763                denominator = 1;
1764                goto out;
1765        }
1766
1767        anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1768                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1769        file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1770                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1771
1772        if (scanning_global_lru(sc)) {
1773                free  = zone_page_state(zone, NR_FREE_PAGES);
1774                /* If we have very few page cache pages,
1775                   force-scan anon pages. */
1776                if (unlikely(file + free <= high_wmark_pages(zone))) {
1777                        fraction[0] = 1;
1778                        fraction[1] = 0;
1779                        denominator = 1;
1780                        goto out;
1781                }
1782        }
1783
1784        /*
1785         * With swappiness at 100, anonymous and file have the same priority.
1786         * This scanning priority is essentially the inverse of IO cost.
1787         */
1788        anon_prio = sc->swappiness;
1789        file_prio = 200 - sc->swappiness;
1790
1791        /*
1792         * OK, so we have swap space and a fair amount of page cache
1793         * pages.  We use the recently rotated / recently scanned
1794         * ratios to determine how valuable each cache is.
1795         *
1796         * Because workloads change over time (and to avoid overflow)
1797         * we keep these statistics as a floating average, which ends
1798         * up weighing recent references more than old ones.
1799         *
1800         * anon in [0], file in [1]
1801         */
1802        spin_lock_irq(&zone->lru_lock);
1803        if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1804                reclaim_stat->recent_scanned[0] /= 2;
1805                reclaim_stat->recent_rotated[0] /= 2;
1806        }
1807
1808        if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1809                reclaim_stat->recent_scanned[1] /= 2;
1810                reclaim_stat->recent_rotated[1] /= 2;
1811        }
1812
1813        /*
1814         * The amount of pressure on anon vs file pages is inversely
1815         * proportional to the fraction of recently scanned pages on
1816         * each list that were recently referenced and in active use.
1817         */
1818        ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1819        ap /= reclaim_stat->recent_rotated[0] + 1;
1820
1821        fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1822        fp /= reclaim_stat->recent_rotated[1] + 1;
1823        spin_unlock_irq(&zone->lru_lock);
1824
1825        fraction[0] = ap;
1826        fraction[1] = fp;
1827        denominator = ap + fp + 1;
1828out:
1829        for_each_evictable_lru(l) {
1830                int file = is_file_lru(l);
1831                unsigned long scan;
1832
1833                scan = zone_nr_lru_pages(zone, sc, l);
1834                if (priority || noswap) {
1835                        scan >>= priority;
1836                        scan = div64_u64(scan * fraction[file], denominator);
1837                }
1838                nr[l] = nr_scan_try_batch(scan,
1839                                          &reclaim_stat->nr_saved_scan[l]);
1840        }
1841}
1842
1843/*
1844 * Reclaim/compaction depends on a number of pages being freed. To avoid
1845 * disruption to the system, a small number of order-0 pages continue to be
1846 * rotated and reclaimed in the normal fashion. However, by the time we get
1847 * back to the allocator and call try_to_compact_zone(), we ensure that
1848 * there are enough free pages for it to be likely successful
1849 */
1850static inline bool should_continue_reclaim(struct zone *zone,
1851                                        unsigned long nr_reclaimed,
1852                                        unsigned long nr_scanned,
1853                                        struct scan_control *sc)
1854{
1855        unsigned long pages_for_compaction;
1856        unsigned long inactive_lru_pages;
1857
1858        /* If not in reclaim/compaction mode, stop */
1859        if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1860                return false;
1861
1862        /* Consider stopping depending on scan and reclaim activity */
1863        if (sc->gfp_mask & __GFP_REPEAT) {
1864                /*
1865                 * For __GFP_REPEAT allocations, stop reclaiming if the
1866                 * full LRU list has been scanned and we are still failing
1867                 * to reclaim pages. This full LRU scan is potentially
1868                 * expensive but a __GFP_REPEAT caller really wants to succeed
1869                 */
1870                if (!nr_reclaimed && !nr_scanned)
1871                        return false;
1872        } else {
1873                /*
1874                 * For non-__GFP_REPEAT allocations which can presumably
1875                 * fail without consequence, stop if we failed to reclaim
1876                 * any pages from the last SWAP_CLUSTER_MAX number of
1877                 * pages that were scanned. This will return to the
1878                 * caller faster at the risk reclaim/compaction and
1879                 * the resulting allocation attempt fails
1880                 */
1881                if (!nr_reclaimed)
1882                        return false;
1883        }
1884
1885        /*
1886         * If we have not reclaimed enough pages for compaction and the
1887         * inactive lists are large enough, continue reclaiming
1888         */
1889        pages_for_compaction = (2UL << sc->order);
1890        inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1891                                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1892        if (sc->nr_reclaimed < pages_for_compaction &&
1893                        inactive_lru_pages > pages_for_compaction)
1894                return true;
1895
1896        /* If compaction would go ahead or the allocation would succeed, stop */
1897        switch (compaction_suitable(zone, sc->order)) {
1898        case COMPACT_PARTIAL:
1899        case COMPACT_CONTINUE:
1900                return false;
1901        default:
1902                return true;
1903        }
1904}
1905
1906/*
1907 * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1908 */
1909static void shrink_zone(int priority, struct zone *zone,
1910                                struct scan_control *sc)
1911{
1912        unsigned long nr[NR_LRU_LISTS];
1913        unsigned long nr_to_scan;
1914        enum lru_list l;
1915        unsigned long nr_reclaimed, nr_scanned;
1916        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1917
1918restart:
1919        nr_reclaimed = 0;
1920        nr_scanned = sc->nr_scanned;
1921        get_scan_count(zone, sc, nr, priority);
1922
1923        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1924                                        nr[LRU_INACTIVE_FILE]) {
1925                for_each_evictable_lru(l) {
1926                        if (nr[l]) {
1927                                nr_to_scan = min_t(unsigned long,
1928                                                   nr[l], SWAP_CLUSTER_MAX);
1929                                nr[l] -= nr_to_scan;
1930
1931                                nr_reclaimed += shrink_list(l, nr_to_scan,
1932                                                            zone, sc, priority);
1933                        }
1934                }
1935                /*
1936                 * On large memory systems, scan >> priority can become
1937                 * really large. This is fine for the starting priority;
1938                 * we want to put equal scanning pressure on each zone.
1939                 * However, if the VM has a harder time of freeing pages,
1940                 * with multiple processes reclaiming pages, the total
1941                 * freeing target can get unreasonably large.
1942                 */
1943                if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1944                        break;
1945        }
1946        sc->nr_reclaimed += nr_reclaimed;
1947
1948        /*
1949         * Even if we did not try to evict anon pages at all, we want to
1950         * rebalance the anon lru active/inactive ratio.
1951         */
1952        if (inactive_anon_is_low(zone, sc))
1953                shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1954
1955        /* reclaim/compaction might need reclaim to continue */
1956        if (should_continue_reclaim(zone, nr_reclaimed,
1957                                        sc->nr_scanned - nr_scanned, sc))
1958                goto restart;
1959
1960        throttle_vm_writeout(sc->gfp_mask);
1961}
1962
1963/*
1964 * This is the direct reclaim path, for page-allocating processes.  We only
1965 * try to reclaim pages from zones which will satisfy the caller's allocation
1966 * request.
1967 *
1968 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1969 * Because:
1970 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1971 *    allocation or
1972 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1973 *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1974 *    zone defense algorithm.
1975 *
1976 * If a zone is deemed to be full of pinned pages then just give it a light
1977 * scan then give up on it.
1978 */
1979static void shrink_zones(int priority, struct zonelist *zonelist,
1980                                        struct scan_control *sc)
1981{
1982        struct zoneref *z;
1983        struct zone *zone;
1984
1985        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1986                                        gfp_zone(sc->gfp_mask), sc->nodemask) {
1987                if (!populated_zone(zone))
1988                        continue;
1989                /*
1990                 * Take care memory controller reclaiming has small influence
1991                 * to global LRU.
1992                 */
1993                if (scanning_global_lru(sc)) {
1994                        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1995                                continue;
1996                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1997                                continue;       /* Let kswapd poll it */
1998                }
1999
2000                shrink_zone(priority, zone, sc);
2001        }
2002}
2003
2004static bool zone_reclaimable(struct zone *zone)
2005{
2006        return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2007}
2008
2009/* All zones in zonelist are unreclaimable? */
2010static bool all_unreclaimable(struct zonelist *zonelist,
2011                struct scan_control *sc)
2012{
2013        struct zoneref *z;
2014        struct zone *zone;
2015
2016        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2017                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2018                if (!populated_zone(zone))
2019                        continue;
2020                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2021                        continue;
2022                if (!zone->all_unreclaimable)
2023                        return false;
2024        }
2025
2026        return true;
2027}
2028
2029/*
2030 * This is the main entry point to direct page reclaim.
2031 *
2032 * If a full scan of the inactive list fails to free enough memory then we
2033 * are "out of memory" and something needs to be killed.
2034 *
2035 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2036 * high - the zone may be full of dirty or under-writeback pages, which this
2037 * caller can't do much about.  We kick the writeback threads and take explicit
2038 * naps in the hope that some of these pages can be written.  But if the
2039 * allocating task holds filesystem locks which prevent writeout this might not
2040 * work, and the allocation attempt will fail.
2041 *
2042 * returns:     0, if no pages reclaimed
2043 *              else, the number of pages reclaimed
2044 */
2045static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2046                                        struct scan_control *sc)
2047{
2048        int priority;
2049        unsigned long total_scanned = 0;
2050        struct reclaim_state *reclaim_state = current->reclaim_state;
2051        struct zoneref *z;
2052        struct zone *zone;
2053        unsigned long writeback_threshold;
2054
2055        get_mems_allowed();
2056        delayacct_freepages_start();
2057
2058        if (scanning_global_lru(sc))
2059                count_vm_event(ALLOCSTALL);
2060
2061        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2062                sc->nr_scanned = 0;
2063                if (!priority)
2064                        disable_swap_token();
2065                shrink_zones(priority, zonelist, sc);
2066                /*
2067                 * Don't shrink slabs when reclaiming memory from
2068                 * over limit cgroups
2069                 */
2070                if (scanning_global_lru(sc)) {
2071                        unsigned long lru_pages = 0;
2072                        for_each_zone_zonelist(zone, z, zonelist,
2073                                        gfp_zone(sc->gfp_mask)) {
2074                                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2075                                        continue;
2076
2077                                lru_pages += zone_reclaimable_pages(zone);
2078                        }
2079
2080                        shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2081                        if (reclaim_state) {
2082                                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2083                                reclaim_state->reclaimed_slab = 0;
2084                        }
2085                }
2086                total_scanned += sc->nr_scanned;
2087                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2088                        goto out;
2089
2090                /*
2091                 * Try to write back as many pages as we just scanned.  This
2092                 * tends to cause slow streaming writers to write data to the
2093                 * disk smoothly, at the dirtying rate, which is nice.   But
2094                 * that's undesirable in laptop mode, where we *want* lumpy
2095                 * writeout.  So in laptop mode, write out the whole world.
2096                 */
2097                writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2098                if (total_scanned > writeback_threshold) {
2099                        wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2100                        sc->may_writepage = 1;
2101                }
2102
2103                /* Take a nap, wait for some writeback to complete */
2104                if (!sc->hibernation_mode && sc->nr_scanned &&
2105                    priority < DEF_PRIORITY - 2) {
2106                        struct zone *preferred_zone;
2107
2108                        first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2109                                                &cpuset_current_mems_allowed,
2110                                                &preferred_zone);
2111                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2112                }
2113        }
2114
2115out:
2116        delayacct_freepages_end();
2117        put_mems_allowed();
2118
2119        if (sc->nr_reclaimed)
2120                return sc->nr_reclaimed;
2121
2122        /*
2123         * As hibernation is going on, kswapd is freezed so that it can't mark
2124         * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2125         * check.
2126         */
2127        if (oom_killer_disabled)
2128                return 0;
2129
2130        /* top priority shrink_zones still had more to do? don't OOM, then */
2131        if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2132                return 1;
2133
2134        return 0;
2135}
2136
2137unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2138                                gfp_t gfp_mask, nodemask_t *nodemask)
2139{
2140        unsigned long nr_reclaimed;
2141        struct scan_control sc = {
2142                .gfp_mask = gfp_mask,
2143                .may_writepage = !laptop_mode,
2144                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2145                .may_unmap = 1,
2146                .may_swap = 1,
2147                .swappiness = vm_swappiness,
2148                .order = order,
2149                .mem_cgroup = NULL,
2150                .nodemask = nodemask,
2151        };
2152
2153        trace_mm_vmscan_direct_reclaim_begin(order,
2154                                sc.may_writepage,
2155                                gfp_mask);
2156
2157        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2158
2159        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2160
2161        return nr_reclaimed;
2162}
2163
2164#ifdef CONFIG_CGROUP_MEM_RES_CTLR
2165
2166unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2167                                                gfp_t gfp_mask, bool noswap,
2168                                                unsigned int swappiness,
2169                                                struct zone *zone)
2170{
2171        struct scan_control sc = {
2172                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2173                .may_writepage = !laptop_mode,
2174                .may_unmap = 1,
2175                .may_swap = !noswap,
2176                .swappiness = swappiness,
2177                .order = 0,
2178                .mem_cgroup = mem,
2179        };
2180        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2181                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2182
2183        trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2184                                                      sc.may_writepage,
2185                                                      sc.gfp_mask);
2186
2187        /*
2188         * NOTE: Although we can get the priority field, using it
2189         * here is not a good idea, since it limits the pages we can scan.
2190         * if we don't reclaim here, the shrink_zone from balance_pgdat
2191         * will pick up pages from other mem cgroup's as well. We hack
2192         * the priority and make it zero.
2193         */
2194        shrink_zone(0, zone, &sc);
2195
2196        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2197
2198        return sc.nr_reclaimed;
2199}
2200
2201unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2202                                           gfp_t gfp_mask,
2203                                           bool noswap,
2204                                           unsigned int swappiness)
2205{
2206        struct zonelist *zonelist;
2207        unsigned long nr_reclaimed;
2208        struct scan_control sc = {
2209                .may_writepage = !laptop_mode,
2210                .may_unmap = 1,
2211                .may_swap = !noswap,
2212                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2213                .swappiness = swappiness,
2214                .order = 0,
2215                .mem_cgroup = mem_cont,
2216                .nodemask = NULL, /* we don't care the placement */
2217        };
2218
2219        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2220                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2221        zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2222
2223        trace_mm_vmscan_memcg_reclaim_begin(0,
2224                                            sc.may_writepage,
2225                                            sc.gfp_mask);
2226
2227        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2228
2229        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2230
2231        return nr_reclaimed;
2232}
2233#endif
2234
2235/*
2236 * pgdat_balanced is used when checking if a node is balanced for high-order
2237 * allocations. Only zones that meet watermarks and are in a zone allowed
2238 * by the callers classzone_idx are added to balanced_pages. The total of
2239 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2240 * for the node to be considered balanced. Forcing all zones to be balanced
2241 * for high orders can cause excessive reclaim when there are imbalanced zones.
2242 * The choice of 25% is due to
2243 *   o a 16M DMA zone that is balanced will not balance a zone on any
2244 *     reasonable sized machine
2245 *   o On all other machines, the top zone must be at least a reasonable
2246 *     percentage of the middle zones. For example, on 32-bit x86, highmem
2247 *     would need to be at least 256M for it to be balance a whole node.
2248 *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2249 *     to balance a node on its own. These seemed like reasonable ratios.
2250 */
2251static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2252                                                int classzone_idx)
2253{
2254        unsigned long present_pages = 0;
2255        int i;
2256
2257        for (i = 0; i <= classzone_idx; i++)
2258                present_pages += pgdat->node_zones[i].present_pages;
2259
2260        /* A special case here: if zone has no page, we think it's balanced */
2261        return balanced_pages >= (present_pages >> 2);
2262}
2263
2264/* is kswapd sleeping prematurely? */
2265static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2266                                        int classzone_idx)
2267{
2268        int i;
2269        unsigned long balanced = 0;
2270        bool all_zones_ok = true;
2271
2272        /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2273        if (remaining)
2274                return true;
2275
2276        /* Check the watermark levels */
2277        for (i = 0; i <= classzone_idx; i++) {
2278                struct zone *zone = pgdat->node_zones + i;
2279
2280                if (!populated_zone(zone))
2281                        continue;
2282
2283                /*
2284                 * balance_pgdat() skips over all_unreclaimable after
2285                 * DEF_PRIORITY. Effectively, it considers them balanced so
2286                 * they must be considered balanced here as well if kswapd
2287                 * is to sleep
2288                 */
2289                if (zone->all_unreclaimable) {
2290                        balanced += zone->present_pages;
2291                        continue;
2292                }
2293
2294                if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2295                                                        i, 0))
2296                        all_zones_ok = false;
2297                else
2298                        balanced += zone->present_pages;
2299        }
2300
2301        /*
2302         * For high-order requests, the balanced zones must contain at least
2303         * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2304         * must be balanced
2305         */
2306        if (order)
2307                return !pgdat_balanced(pgdat, balanced, classzone_idx);
2308        else
2309                return !all_zones_ok;
2310}
2311
2312/*
2313 * For kswapd, balance_pgdat() will work across all this node's zones until
2314 * they are all at high_wmark_pages(zone).
2315 *
2316 * Returns the final order kswapd was reclaiming at
2317 *
2318 * There is special handling here for zones which are full of pinned pages.
2319 * This can happen if the pages are all mlocked, or if they are all used by
2320 * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2321 * What we do is to detect the case where all pages in the zone have been
2322 * scanned twice and there has been zero successful reclaim.  Mark the zone as
2323 * dead and from now on, only perform a short scan.  Basically we're polling
2324 * the zone for when the problem goes away.
2325 *
2326 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2327 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2328 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2329 * lower zones regardless of the number of free pages in the lower zones. This
2330 * interoperates with the page allocator fallback scheme to ensure that aging
2331 * of pages is balanced across the zones.
2332 */
2333static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2334                                                        int *classzone_idx)
2335{
2336        int all_zones_ok;
2337        unsigned long balanced;
2338        int priority;
2339        int i;
2340        int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2341        unsigned long total_scanned;
2342        struct reclaim_state *reclaim_state = current->reclaim_state;
2343        struct scan_control sc = {
2344                .gfp_mask = GFP_KERNEL,
2345                .may_unmap = 1,
2346                .may_swap = 1,
2347                /*
2348                 * kswapd doesn't want to be bailed out while reclaim. because
2349                 * we want to put equal scanning pressure on each zone.
2350                 */
2351                .nr_to_reclaim = ULONG_MAX,
2352                .swappiness = vm_swappiness,
2353                .order = order,
2354                .mem_cgroup = NULL,
2355        };
2356loop_again:
2357        total_scanned = 0;
2358        sc.nr_reclaimed = 0;
2359        sc.may_writepage = !laptop_mode;
2360        count_vm_event(PAGEOUTRUN);
2361
2362        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2363                unsigned long lru_pages = 0;
2364                int has_under_min_watermark_zone = 0;
2365
2366                /* The swap token gets in the way of swapout... */
2367                if (!priority)
2368                        disable_swap_token();
2369
2370                all_zones_ok = 1;
2371                balanced = 0;
2372
2373                /*
2374                 * Scan in the highmem->dma direction for the highest
2375                 * zone which needs scanning
2376                 */
2377                for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2378                        struct zone *zone = pgdat->node_zones + i;
2379
2380                        if (!populated_zone(zone))
2381                                continue;
2382
2383                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2384                                continue;
2385
2386                        /*
2387                         * Do some background aging of the anon list, to give
2388                         * pages a chance to be referenced before reclaiming.
2389                         */
2390                        if (inactive_anon_is_low(zone, &sc))
2391                                shrink_active_list(SWAP_CLUSTER_MAX, zone,
2392                                                        &sc, priority, 0);
2393
2394                        if (!zone_watermark_ok_safe(zone, order,
2395                                        high_wmark_pages(zone), 0, 0)) {
2396                                end_zone = i;
2397                                break;
2398                        }
2399                }
2400                if (i < 0)
2401                        goto out;
2402
2403                for (i = 0; i <= end_zone; i++) {
2404                        struct zone *zone = pgdat->node_zones + i;
2405
2406                        lru_pages += zone_reclaimable_pages(zone);
2407                }
2408
2409                /*
2410                 * Now scan the zone in the dma->highmem direction, stopping
2411                 * at the last zone which needs scanning.
2412                 *
2413                 * We do this because the page allocator works in the opposite
2414                 * direction.  This prevents the page allocator from allocating
2415                 * pages behind kswapd's direction of progress, which would
2416                 * cause too much scanning of the lower zones.
2417                 */
2418                for (i = 0; i <= end_zone; i++) {
2419                        struct zone *zone = pgdat->node_zones + i;
2420                        int nr_slab;
2421                        unsigned long balance_gap;
2422
2423                        if (!populated_zone(zone))
2424                                continue;
2425
2426                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2427                                continue;
2428
2429                        sc.nr_scanned = 0;
2430
2431                        /*
2432                         * Call soft limit reclaim before calling shrink_zone.
2433                         * For now we ignore the return value
2434                         */
2435                        mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2436
2437                        /*
2438                         * We put equal pressure on every zone, unless
2439                         * one zone has way too many pages free
2440                         * already. The "too many pages" is defined
2441                         * as the high wmark plus a "gap" where the
2442                         * gap is either the low watermark or 1%
2443                         * of the zone, whichever is smaller.
2444                         */
2445                        balance_gap = min(low_wmark_pages(zone),
2446                                (zone->present_pages +
2447                                        KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2448                                KSWAPD_ZONE_BALANCE_GAP_RATIO);
2449                        if (!zone_watermark_ok_safe(zone, order,
2450                                        high_wmark_pages(zone) + balance_gap,
2451                                        end_zone, 0)) {
2452                                shrink_zone(priority, zone, &sc);
2453
2454                                reclaim_state->reclaimed_slab = 0;
2455                                nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2456                                                        lru_pages);
2457                                sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2458                                total_scanned += sc.nr_scanned;
2459
2460                                if (nr_slab == 0 && !zone_reclaimable(zone))
2461                                        zone->all_unreclaimable = 1;
2462                        }
2463
2464                        /*
2465                         * If we've done a decent amount of scanning and
2466                         * the reclaim ratio is low, start doing writepage
2467                         * even in laptop mode
2468                         */
2469                        if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2470                            total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2471                                sc.may_writepage = 1;
2472
2473                        if (zone->all_unreclaimable) {
2474                                if (end_zone && end_zone == i)
2475                                        end_zone--;
2476                                continue;
2477                        }
2478
2479                        if (!zone_watermark_ok_safe(zone, order,
2480                                        high_wmark_pages(zone), end_zone, 0)) {
2481                                all_zones_ok = 0;
2482                                /*
2483                                 * We are still under min water mark.  This
2484                                 * means that we have a GFP_ATOMIC allocation
2485                                 * failure risk. Hurry up!
2486                                 */
2487                                if (!zone_watermark_ok_safe(zone, order,
2488                                            min_wmark_pages(zone), end_zone, 0))
2489                                        has_under_min_watermark_zone = 1;
2490                        } else {
2491                                /*
2492                                 * If a zone reaches its high watermark,
2493                                 * consider it to be no longer congested. It's
2494                                 * possible there are dirty pages backed by
2495                                 * congested BDIs but as pressure is relieved,
2496                                 * spectulatively avoid congestion waits
2497                                 */
2498                                zone_clear_flag(zone, ZONE_CONGESTED);
2499                                if (i <= *classzone_idx)
2500                                        balanced += zone->present_pages;
2501                        }
2502
2503                }
2504                if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2505                        break;          /* kswapd: all done */
2506                /*
2507                 * OK, kswapd is getting into trouble.  Take a nap, then take
2508                 * another pass across the zones.
2509                 */
2510                if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2511                        if (has_under_min_watermark_zone)
2512                                count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2513                        else
2514                                congestion_wait(BLK_RW_ASYNC, HZ/10);
2515                }
2516
2517                /*
2518                 * We do this so kswapd doesn't build up large priorities for
2519                 * example when it is freeing in parallel with allocators. It
2520                 * matches the direct reclaim path behaviour in terms of impact
2521                 * on zone->*_priority.
2522                 */
2523                if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2524                        break;
2525        }
2526out:
2527
2528        /*
2529         * order-0: All zones must meet high watermark for a balanced node
2530         * high-order: Balanced zones must make up at least 25% of the node
2531         *             for the node to be balanced
2532         */
2533        if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2534                cond_resched();
2535
2536                try_to_freeze();
2537
2538                /*
2539                 * Fragmentation may mean that the system cannot be
2540                 * rebalanced for high-order allocations in all zones.
2541                 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2542                 * it means the zones have been fully scanned and are still
2543                 * not balanced. For high-order allocations, there is
2544                 * little point trying all over again as kswapd may
2545                 * infinite loop.
2546                 *
2547                 * Instead, recheck all watermarks at order-0 as they
2548                 * are the most important. If watermarks are ok, kswapd will go
2549                 * back to sleep. High-order users can still perform direct
2550                 * reclaim if they wish.
2551                 */
2552                if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2553                        order = sc.order = 0;
2554
2555                goto loop_again;
2556        }
2557
2558        /*
2559         * If kswapd was reclaiming at a higher order, it has the option of
2560         * sleeping without all zones being balanced. Before it does, it must
2561         * ensure that the watermarks for order-0 on *all* zones are met and
2562         * that the congestion flags are cleared. The congestion flag must
2563         * be cleared as kswapd is the only mechanism that clears the flag
2564         * and it is potentially going to sleep here.
2565         */
2566        if (order) {
2567                for (i = 0; i <= end_zone; i++) {
2568                        struct zone *zone = pgdat->node_zones + i;
2569
2570                        if (!populated_zone(zone))
2571                                continue;
2572
2573                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2574                                continue;
2575
2576                        /* Confirm the zone is balanced for order-0 */
2577                        if (!zone_watermark_ok(zone, 0,
2578                                        high_wmark_pages(zone), 0, 0)) {
2579                                order = sc.order = 0;
2580                                goto loop_again;
2581                        }
2582
2583                        /* If balanced, clear the congested flag */
2584                        zone_clear_flag(zone, ZONE_CONGESTED);
2585                }
2586        }
2587
2588        /*
2589         * Return the order we were reclaiming at so sleeping_prematurely()
2590         * makes a decision on the order we were last reclaiming at. However,
2591         * if another caller entered the allocator slow path while kswapd
2592         * was awake, order will remain at the higher level
2593         */
2594        *classzone_idx = end_zone;
2595        return order;
2596}
2597
2598static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2599{
2600        long remaining = 0;
2601        DEFINE_WAIT(wait);
2602
2603        if (freezing(current) || kthread_should_stop())
2604                return;
2605
2606        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2607
2608        /* Try to sleep for a short interval */
2609        if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2610                remaining = schedule_timeout(HZ/10);
2611                finish_wait(&pgdat->kswapd_wait, &wait);
2612                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2613        }
2614
2615        /*
2616         * After a short sleep, check if it was a premature sleep. If not, then
2617         * go fully to sleep until explicitly woken up.
2618         */
2619        if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2620                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2621
2622                /*
2623                 * vmstat counters are not perfectly accurate and the estimated
2624                 * value for counters such as NR_FREE_PAGES can deviate from the
2625                 * true value by nr_online_cpus * threshold. To avoid the zone
2626                 * watermarks being breached while under pressure, we reduce the
2627                 * per-cpu vmstat threshold while kswapd is awake and restore
2628                 * them before going back to sleep.
2629                 */
2630                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2631                schedule();
2632                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2633        } else {
2634                if (remaining)
2635                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2636                else
2637                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2638        }
2639        finish_wait(&pgdat->kswapd_wait, &wait);
2640}
2641
2642/*
2643 * The background pageout daemon, started as a kernel thread
2644 * from the init process.
2645 *
2646 * This basically trickles out pages so that we have _some_
2647 * free memory available even if there is no other activity
2648 * that frees anything up. This is needed for things like routing
2649 * etc, where we otherwise might have all activity going on in
2650 * asynchronous contexts that cannot page things out.
2651 *
2652 * If there are applications that are active memory-allocators
2653 * (most normal use), this basically shouldn't matter.
2654 */
2655static int kswapd(void *p)
2656{
2657        unsigned long order, new_order;
2658        int classzone_idx, new_classzone_idx;
2659        pg_data_t *pgdat = (pg_data_t*)p;
2660        struct task_struct *tsk = current;
2661
2662        struct reclaim_state reclaim_state = {
2663                .reclaimed_slab = 0,
2664        };
2665        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2666
2667        lockdep_set_current_reclaim_state(GFP_KERNEL);
2668
2669        if (!cpumask_empty(cpumask))
2670                set_cpus_allowed_ptr(tsk, cpumask);
2671        current->reclaim_state = &reclaim_state;
2672
2673        /*
2674         * Tell the memory management that we're a "memory allocator",
2675         * and that if we need more memory we should get access to it
2676         * regardless (see "__alloc_pages()"). "kswapd" should
2677         * never get caught in the normal page freeing logic.
2678         *
2679         * (Kswapd normally doesn't need memory anyway, but sometimes
2680         * you need a small amount of memory in order to be able to
2681         * page out something else, and this flag essentially protects
2682         * us from recursively trying to free more memory as we're
2683         * trying to free the first piece of memory in the first place).
2684         */
2685        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2686        set_freezable();
2687
2688        order = new_order = 0;
2689        classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2690        for ( ; ; ) {
2691                int ret;
2692
2693                /*
2694                 * If the last balance_pgdat was unsuccessful it's unlikely a
2695                 * new request of a similar or harder type will succeed soon
2696                 * so consider going to sleep on the basis we reclaimed at
2697                 */
2698                if (classzone_idx >= new_classzone_idx && order == new_order) {
2699                        new_order = pgdat->kswapd_max_order;
2700                        new_classzone_idx = pgdat->classzone_idx;
2701                        pgdat->kswapd_max_order =  0;
2702                        pgdat->classzone_idx = pgdat->nr_zones - 1;
2703                }
2704
2705                if (order < new_order || classzone_idx > new_classzone_idx) {
2706                        /*
2707                         * Don't sleep if someone wants a larger 'order'
2708                         * allocation or has tigher zone constraints
2709                         */
2710                        order = new_order;
2711                        classzone_idx = new_classzone_idx;
2712                } else {
2713                        kswapd_try_to_sleep(pgdat, order, classzone_idx);
2714                        order = pgdat->kswapd_max_order;
2715                        classzone_idx = pgdat->classzone_idx;
2716                        pgdat->kswapd_max_order = 0;
2717                        pgdat->classzone_idx = pgdat->nr_zones - 1;
2718                }
2719
2720                ret = try_to_freeze();
2721                if (kthread_should_stop())
2722                        break;
2723
2724                /*
2725                 * We can speed up thawing tasks if we don't call balance_pgdat
2726                 * after returning from the refrigerator
2727                 */
2728                if (!ret) {
2729                        trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2730                        order = balance_pgdat(pgdat, order, &classzone_idx);
2731                }
2732        }
2733        return 0;
2734}
2735
2736/*
2737 * A zone is low on free memory, so wake its kswapd task to service it.
2738 */
2739void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2740{
2741        pg_data_t *pgdat;
2742
2743        if (!populated_zone(zone))
2744                return;
2745
2746        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2747                return;
2748        pgdat = zone->zone_pgdat;
2749        if (pgdat->kswapd_max_order < order) {
2750                pgdat->kswapd_max_order = order;
2751                pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2752        }
2753        if (!waitqueue_active(&pgdat->kswapd_wait))
2754                return;
2755        if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2756                return;
2757
2758        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2759        wake_up_interruptible(&pgdat->kswapd_wait);
2760}
2761
2762/*
2763 * The reclaimable count would be mostly accurate.
2764 * The less reclaimable pages may be
2765 * - mlocked pages, which will be moved to unevictable list when encountered
2766 * - mapped pages, which may require several travels to be reclaimed
2767 * - dirty pages, which is not "instantly" reclaimable
2768 */
2769unsigned long global_reclaimable_pages(void)
2770{
2771        int nr;
2772
2773        nr = global_page_state(NR_ACTIVE_FILE) +
2774             global_page_state(NR_INACTIVE_FILE);
2775
2776        if (nr_swap_pages > 0)
2777                nr += global_page_state(NR_ACTIVE_ANON) +
2778                      global_page_state(NR_INACTIVE_ANON);
2779
2780        return nr;
2781}
2782
2783unsigned long zone_reclaimable_pages(struct zone *zone)
2784{
2785        int nr;
2786
2787        nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2788             zone_page_state(zone, NR_INACTIVE_FILE);
2789
2790        if (nr_swap_pages > 0)
2791                nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2792                      zone_page_state(zone, NR_INACTIVE_ANON);
2793
2794        return nr;
2795}
2796
2797#ifdef CONFIG_HIBERNATION
2798/*
2799 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2800 * freed pages.
2801 *
2802 * Rather than trying to age LRUs the aim is to preserve the overall
2803 * LRU order by reclaiming preferentially
2804 * inactive > active > active referenced > active mapped
2805 */
2806unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2807{
2808        struct reclaim_state reclaim_state;
2809        struct scan_control sc = {
2810                .gfp_mask = GFP_HIGHUSER_MOVABLE,
2811                .may_swap = 1,
2812                .may_unmap = 1,
2813                .may_writepage = 1,
2814                .nr_to_reclaim = nr_to_reclaim,
2815                .hibernation_mode = 1,
2816                .swappiness = vm_swappiness,
2817                .order = 0,
2818        };
2819        struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2820        struct task_struct *p = current;
2821        unsigned long nr_reclaimed;
2822
2823        p->flags |= PF_MEMALLOC;
2824        lockdep_set_current_reclaim_state(sc.gfp_mask);
2825        reclaim_state.reclaimed_slab = 0;
2826        p->reclaim_state = &reclaim_state;
2827
2828        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2829
2830        p->reclaim_state = NULL;
2831        lockdep_clear_current_reclaim_state();
2832        p->flags &= ~PF_MEMALLOC;
2833
2834        return nr_reclaimed;
2835}
2836#endif /* CONFIG_HIBERNATION */
2837
2838/* It's optimal to keep kswapds on the same CPUs as their memory, but
2839   not required for correctness.  So if the last cpu in a node goes
2840   away, we get changed to run anywhere: as the first one comes back,
2841   restore their cpu bindings. */
2842static int __devinit cpu_callback(struct notifier_block *nfb,
2843                                  unsigned long action, void *hcpu)
2844{
2845        int nid;
2846
2847        if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2848                for_each_node_state(nid, N_HIGH_MEMORY) {
2849                        pg_data_t *pgdat = NODE_DATA(nid);
2850                        const struct cpumask *mask;
2851
2852                        mask = cpumask_of_node(pgdat->node_id);
2853
2854                        if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2855                                /* One of our CPUs online: restore mask */
2856                                set_cpus_allowed_ptr(pgdat->kswapd, mask);
2857                }
2858        }
2859        return NOTIFY_OK;
2860}
2861
2862/*
2863 * This kswapd start function will be called by init and node-hot-add.
2864 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2865 */
2866int kswapd_run(int nid)
2867{
2868        pg_data_t *pgdat = NODE_DATA(nid);
2869        int ret = 0;
2870
2871        if (pgdat->kswapd)
2872                return 0;
2873
2874        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2875        if (IS_ERR(pgdat->kswapd)) {
2876                /* failure at boot is fatal */
2877                BUG_ON(system_state == SYSTEM_BOOTING);
2878                printk("Failed to start kswapd on node %d\n",nid);
2879                ret = -1;
2880        }
2881        return ret;
2882}
2883
2884/*
2885 * Called by memory hotplug when all memory in a node is offlined.
2886 */
2887void kswapd_stop(int nid)
2888{
2889        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2890
2891        if (kswapd)
2892                kthread_stop(kswapd);
2893}
2894
2895static int __init kswapd_init(void)
2896{
2897        int nid;
2898
2899        swap_setup();
2900        for_each_node_state(nid, N_HIGH_MEMORY)
2901                kswapd_run(nid);
2902        hotcpu_notifier(cpu_callback, 0);
2903        return 0;
2904}
2905
2906module_init(kswapd_init)
2907
2908#ifdef CONFIG_NUMA
2909/*
2910 * Zone reclaim mode
2911 *
2912 * If non-zero call zone_reclaim when the number of free pages falls below
2913 * the watermarks.
2914 */
2915int zone_reclaim_mode __read_mostly;
2916
2917#define RECLAIM_OFF 0
2918#define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2919#define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2920#define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2921
2922/*
2923 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2924 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2925 * a zone.
2926 */
2927#define ZONE_RECLAIM_PRIORITY 4
2928
2929/*
2930 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2931 * occur.
2932 */
2933int sysctl_min_unmapped_ratio = 1;
2934
2935/*
2936 * If the number of slab pages in a zone grows beyond this percentage then
2937 * slab reclaim needs to occur.
2938 */
2939int sysctl_min_slab_ratio = 5;
2940
2941static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2942{
2943        unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2944        unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2945                zone_page_state(zone, NR_ACTIVE_FILE);
2946
2947        /*
2948         * It's possible for there to be more file mapped pages than
2949         * accounted for by the pages on the file LRU lists because
2950         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2951         */
2952        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2953}
2954
2955/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2956static long zone_pagecache_reclaimable(struct zone *zone)
2957{
2958        long nr_pagecache_reclaimable;
2959        long delta = 0;
2960
2961        /*
2962         * If RECLAIM_SWAP is set, then all file pages are considered
2963         * potentially reclaimable. Otherwise, we have to worry about
2964         * pages like swapcache and zone_unmapped_file_pages() provides
2965         * a better estimate
2966         */
2967        if (zone_reclaim_mode & RECLAIM_SWAP)
2968                nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2969        else
2970                nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2971
2972        /* If we can't clean pages, remove dirty pages from consideration */
2973        if (!(zone_reclaim_mode & RECLAIM_WRITE))
2974                delta += zone_page_state(zone, NR_FILE_DIRTY);
2975
2976        /* Watch for any possible underflows due to delta */
2977        if (unlikely(delta > nr_pagecache_reclaimable))
2978                delta = nr_pagecache_reclaimable;
2979
2980        return nr_pagecache_reclaimable - delta;
2981}
2982
2983/*
2984 * Try to free up some pages from this zone through reclaim.
2985 */
2986static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2987{
2988        /* Minimum pages needed in order to stay on node */
2989        const unsigned long nr_pages = 1 << order;
2990        struct task_struct *p = current;
2991        struct reclaim_state reclaim_state;
2992        int priority;
2993        struct scan_control sc = {
2994                .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2995                .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2996                .may_swap = 1,
2997                .nr_to_reclaim = max_t(unsigned long, nr_pages,
2998                                       SWAP_CLUSTER_MAX),
2999                .gfp_mask = gfp_mask,
3000                .swappiness = vm_swappiness,
3001                .order = order,
3002        };
3003        unsigned long nr_slab_pages0, nr_slab_pages1;
3004
3005        cond_resched();
3006        /*
3007         * We need to be able to allocate from the reserves for RECLAIM_SWAP
3008         * and we also need to be able to write out pages for RECLAIM_WRITE
3009         * and RECLAIM_SWAP.
3010         */
3011        p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3012        lockdep_set_current_reclaim_state(gfp_mask);
3013        reclaim_state.reclaimed_slab = 0;
3014        p->reclaim_state = &reclaim_state;
3015
3016        if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3017                /*
3018                 * Free memory by calling shrink zone with increasing
3019                 * priorities until we have enough memory freed.
3020                 */
3021                priority = ZONE_RECLAIM_PRIORITY;
3022                do {
3023                        shrink_zone(priority, zone, &sc);
3024                        priority--;
3025                } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3026        }
3027
3028        nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3029        if (nr_slab_pages0 > zone->min_slab_pages) {
3030                /*
3031                 * shrink_slab() does not currently allow us to determine how
3032                 * many pages were freed in this zone. So we take the current
3033                 * number of slab pages and shake the slab until it is reduced
3034                 * by the same nr_pages that we used for reclaiming unmapped
3035                 * pages.
3036                 *
3037                 * Note that shrink_slab will free memory on all zones and may
3038                 * take a long time.
3039                 */
3040                for (;;) {
3041                        unsigned long lru_pages = zone_reclaimable_pages(zone);
3042
3043                        /* No reclaimable slab or very low memory pressure */
3044                        if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3045                                break;
3046
3047                        /* Freed enough memory */
3048                        nr_slab_pages1 = zone_page_state(zone,
3049                                                        NR_SLAB_RECLAIMABLE);
3050                        if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3051                                break;
3052                }
3053
3054                /*
3055                 * Update nr_reclaimed by the number of slab pages we
3056                 * reclaimed from this zone.
3057                 */
3058                nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3059                if (nr_slab_pages1 < nr_slab_pages0)
3060                        sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3061        }
3062
3063        p->reclaim_state = NULL;
3064        current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3065        lockdep_clear_current_reclaim_state();
3066        return sc.nr_reclaimed >= nr_pages;
3067}
3068
3069int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3070{
3071        int node_id;
3072        int ret;
3073
3074        /*
3075         * Zone reclaim reclaims unmapped file backed pages and
3076         * slab pages if we are over the defined limits.
3077         *
3078         * A small portion of unmapped file backed pages is needed for
3079         * file I/O otherwise pages read by file I/O will be immediately
3080         * thrown out if the zone is overallocated. So we do not reclaim
3081         * if less than a specified percentage of the zone is used by
3082         * unmapped file backed pages.
3083         */
3084        if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3085            zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3086                return ZONE_RECLAIM_FULL;
3087
3088        if (zone->all_unreclaimable)
3089                return ZONE_RECLAIM_FULL;
3090
3091        /*
3092         * Do not scan if the allocation should not be delayed.
3093         */
3094        if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3095                return ZONE_RECLAIM_NOSCAN;
3096
3097        /*
3098         * Only run zone reclaim on the local zone or on zones that do not
3099         * have associated processors. This will favor the local processor
3100         * over remote processors and spread off node memory allocations
3101         * as wide as possible.
3102         */
3103        node_id = zone_to_nid(zone);
3104        if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3105                return ZONE_RECLAIM_NOSCAN;
3106
3107        if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3108                return ZONE_RECLAIM_NOSCAN;
3109
3110        ret = __zone_reclaim(zone, gfp_mask, order);
3111        zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3112
3113        if (!ret)
3114                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3115
3116        return ret;
3117}
3118#endif
3119
3120/*
3121 * page_evictable - test whether a page is evictable
3122 * @page: the page to test
3123 * @vma: the VMA in which the page is or will be mapped, may be NULL
3124 *
3125 * Test whether page is evictable--i.e., should be placed on active/inactive
3126 * lists vs unevictable list.  The vma argument is !NULL when called from the
3127 * fault path to determine how to instantate a new page.
3128 *
3129 * Reasons page might not be evictable:
3130 * (1) page's mapping marked unevictable
3131 * (2) page is part of an mlocked VMA
3132 *
3133 */
3134int page_evictable(struct page *page, struct vm_area_struct *vma)
3135{
3136
3137        if (mapping_unevictable(page_mapping(page)))
3138                return 0;
3139
3140        if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3141                return 0;
3142
3143        return 1;
3144}
3145
3146/**
3147 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3148 * @page: page to check evictability and move to appropriate lru list
3149 * @zone: zone page is in
3150 *
3151 * Checks a page for evictability and moves the page to the appropriate
3152 * zone lru list.
3153 *
3154 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3155 * have PageUnevictable set.
3156 */
3157static void check_move_unevictable_page(struct page *page, struct zone *zone)
3158{
3159        VM_BUG_ON(PageActive(page));
3160
3161retry:
3162        ClearPageUnevictable(page);
3163        if (page_evictable(page, NULL)) {
3164                enum lru_list l = page_lru_base_type(page);
3165
3166                __dec_zone_state(zone, NR_UNEVICTABLE);
3167                list_move(&page->lru, &zone->lru[l].list);
3168                mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3169                __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3170                __count_vm_event(UNEVICTABLE_PGRESCUED);
3171        } else {
3172                /*
3173                 * rotate unevictable list
3174                 */
3175                SetPageUnevictable(page);
3176                list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3177                mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3178                if (page_evictable(page, NULL))
3179                        goto retry;
3180        }
3181}
3182
3183/**
3184 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3185 * @mapping: struct address_space to scan for evictable pages
3186 *
3187 * Scan all pages in mapping.  Check unevictable pages for
3188 * evictability and move them to the appropriate zone lru list.
3189 */
3190void scan_mapping_unevictable_pages(struct address_space *mapping)
3191{
3192        pgoff_t next = 0;
3193        pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3194                         PAGE_CACHE_SHIFT;
3195        struct zone *zone;
3196        struct pagevec pvec;
3197
3198        if (mapping->nrpages == 0)
3199                return;
3200
3201        pagevec_init(&pvec, 0);
3202        while (next < end &&
3203                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3204                int i;
3205                int pg_scanned = 0;
3206
3207                zone = NULL;
3208
3209                for (i = 0; i < pagevec_count(&pvec); i++) {
3210                        struct page *page = pvec.pages[i];
3211                        pgoff_t page_index = page->index;
3212                        struct zone *pagezone = page_zone(page);
3213
3214                        pg_scanned++;
3215                        if (page_index > next)
3216                                next = page_index;
3217                        next++;
3218
3219                        if (pagezone != zone) {
3220                                if (zone)
3221                                        spin_unlock_irq(&zone->lru_lock);
3222                                zone = pagezone;
3223                                spin_lock_irq(&zone->lru_lock);
3224                        }
3225
3226                        if (PageLRU(page) && PageUnevictable(page))
3227                                check_move_unevictable_page(page, zone);
3228                }
3229                if (zone)
3230                        spin_unlock_irq(&zone->lru_lock);
3231                pagevec_release(&pvec);
3232
3233                count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3234        }
3235
3236}
3237
3238/**
3239 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3240 * @zone - zone of which to scan the unevictable list
3241 *
3242 * Scan @zone's unevictable LRU lists to check for pages that have become
3243 * evictable.  Move those that have to @zone's inactive list where they
3244 * become candidates for reclaim, unless shrink_inactive_zone() decides
3245 * to reactivate them.  Pages that are still unevictable are rotated
3246 * back onto @zone's unevictable list.
3247 */
3248#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3249static void scan_zone_unevictable_pages(struct zone *zone)
3250{
3251        struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3252        unsigned long scan;
3253        unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3254
3255        while (nr_to_scan > 0) {
3256                unsigned long batch_size = min(nr_to_scan,
3257                                                SCAN_UNEVICTABLE_BATCH_SIZE);
3258
3259                spin_lock_irq(&zone->lru_lock);
3260                for (scan = 0;  scan < batch_size; scan++) {
3261                        struct page *page = lru_to_page(l_unevictable);
3262
3263                        if (!trylock_page(page))
3264                                continue;
3265
3266                        prefetchw_prev_lru_page(page, l_unevictable, flags);
3267
3268                        if (likely(PageLRU(page) && PageUnevictable(page)))
3269                                check_move_unevictable_page(page, zone);
3270
3271                        unlock_page(page);
3272                }
3273                spin_unlock_irq(&zone->lru_lock);
3274
3275                nr_to_scan -= batch_size;
3276        }
3277}
3278
3279
3280/**
3281 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3282 *
3283 * A really big hammer:  scan all zones' unevictable LRU lists to check for
3284 * pages that have become evictable.  Move those back to the zones'
3285 * inactive list where they become candidates for reclaim.
3286 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3287 * and we add swap to the system.  As such, it runs in the context of a task
3288 * that has possibly/probably made some previously unevictable pages
3289 * evictable.
3290 */
3291static void scan_all_zones_unevictable_pages(void)
3292{
3293        struct zone *zone;
3294
3295        for_each_zone(zone) {
3296                scan_zone_unevictable_pages(zone);
3297        }
3298}
3299
3300/*
3301 * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3302 * all nodes' unevictable lists for evictable pages
3303 */
3304unsigned long scan_unevictable_pages;
3305
3306int scan_unevictable_handler(struct ctl_table *table, int write,
3307                           void __user *buffer,
3308                           size_t *length, loff_t *ppos)
3309{
3310        proc_doulongvec_minmax(table, write, buffer, length, ppos);
3311
3312        if (write && *(unsigned long *)table->data)
3313                scan_all_zones_unevictable_pages();
3314
3315        scan_unevictable_pages = 0;
3316        return 0;
3317}
3318
3319#ifdef CONFIG_NUMA
3320/*
3321 * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3322 * a specified node's per zone unevictable lists for evictable pages.
3323 */
3324
3325static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3326                                          struct sysdev_attribute *attr,
3327                                          char *buf)
3328{
3329        return sprintf(buf, "0\n");     /* always zero; should fit... */
3330}
3331
3332static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3333                                           struct sysdev_attribute *attr,
3334                                        const char *buf, size_t count)
3335{
3336        struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3337        struct zone *zone;
3338        unsigned long res;
3339        unsigned long req = strict_strtoul(buf, 10, &res);
3340
3341        if (!req)
3342                return 1;       /* zero is no-op */
3343
3344        for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3345                if (!populated_zone(zone))
3346                        continue;
3347                scan_zone_unevictable_pages(zone);
3348        }
3349        return 1;
3350}
3351
3352
3353static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3354                        read_scan_unevictable_node,
3355                        write_scan_unevictable_node);
3356
3357int scan_unevictable_register_node(struct node *node)
3358{
3359        return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3360}
3361
3362void scan_unevictable_unregister_node(struct node *node)
3363{
3364        sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3365}
3366#endif
3367