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                                /* the page is freed already. */
1118                                if (!page_count(cursor_page))
1119                                        continue;
1120                                break;
1121                        }
1122                }
1123
1124                /* If we break out of the loop above, lumpy reclaim failed */
1125                if (pfn < end_pfn)
1126                        nr_lumpy_failed++;
1127        }
1128
1129        *scanned = scan;
1130
1131        trace_mm_vmscan_lru_isolate(order,
1132                        nr_to_scan, scan,
1133                        nr_taken,
1134                        nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1135                        mode);
1136        return nr_taken;
1137}
1138
1139static unsigned long isolate_pages_global(unsigned long nr,
1140                                        struct list_head *dst,
1141                                        unsigned long *scanned, int order,
1142                                        int mode, struct zone *z,
1143                                        int active, int file)
1144{
1145        int lru = LRU_BASE;
1146        if (active)
1147                lru += LRU_ACTIVE;
1148        if (file)
1149                lru += LRU_FILE;
1150        return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1151                                                                mode, file);
1152}
1153
1154/*
1155 * clear_active_flags() is a helper for shrink_active_list(), clearing
1156 * any active bits from the pages in the list.
1157 */
1158static unsigned long clear_active_flags(struct list_head *page_list,
1159                                        unsigned int *count)
1160{
1161        int nr_active = 0;
1162        int lru;
1163        struct page *page;
1164
1165        list_for_each_entry(page, page_list, lru) {
1166                int numpages = hpage_nr_pages(page);
1167                lru = page_lru_base_type(page);
1168                if (PageActive(page)) {
1169                        lru += LRU_ACTIVE;
1170                        ClearPageActive(page);
1171                        nr_active += numpages;
1172                }
1173                if (count)
1174                        count[lru] += numpages;
1175        }
1176
1177        return nr_active;
1178}
1179
1180/**
1181 * isolate_lru_page - tries to isolate a page from its LRU list
1182 * @page: page to isolate from its LRU list
1183 *
1184 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1185 * vmstat statistic corresponding to whatever LRU list the page was on.
1186 *
1187 * Returns 0 if the page was removed from an LRU list.
1188 * Returns -EBUSY if the page was not on an LRU list.
1189 *
1190 * The returned page will have PageLRU() cleared.  If it was found on
1191 * the active list, it will have PageActive set.  If it was found on
1192 * the unevictable list, it will have the PageUnevictable bit set. That flag
1193 * may need to be cleared by the caller before letting the page go.
1194 *
1195 * The vmstat statistic corresponding to the list on which the page was
1196 * found will be decremented.
1197 *
1198 * Restrictions:
1199 * (1) Must be called with an elevated refcount on the page. This is a
1200 *     fundamentnal difference from isolate_lru_pages (which is called
1201 *     without a stable reference).
1202 * (2) the lru_lock must not be held.
1203 * (3) interrupts must be enabled.
1204 */
1205int isolate_lru_page(struct page *page)
1206{
1207        int ret = -EBUSY;
1208
1209        if (PageLRU(page)) {
1210                struct zone *zone = page_zone(page);
1211
1212                spin_lock_irq(&zone->lru_lock);
1213                if (PageLRU(page) && get_page_unless_zero(page)) {
1214                        int lru = page_lru(page);
1215                        ret = 0;
1216                        ClearPageLRU(page);
1217
1218                        del_page_from_lru_list(zone, page, lru);
1219                }
1220                spin_unlock_irq(&zone->lru_lock);
1221        }
1222        return ret;
1223}
1224
1225/*
1226 * Are there way too many processes in the direct reclaim path already?
1227 */
1228static int too_many_isolated(struct zone *zone, int file,
1229                struct scan_control *sc)
1230{
1231        unsigned long inactive, isolated;
1232
1233        if (current_is_kswapd())
1234                return 0;
1235
1236        if (!scanning_global_lru(sc))
1237                return 0;
1238
1239        if (file) {
1240                inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1241                isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1242        } else {
1243                inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1244                isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1245        }
1246
1247        return isolated > inactive;
1248}
1249
1250/*
1251 * TODO: Try merging with migrations version of putback_lru_pages
1252 */
1253static noinline_for_stack void
1254putback_lru_pages(struct zone *zone, struct scan_control *sc,
1255                                unsigned long nr_anon, unsigned long nr_file,
1256                                struct list_head *page_list)
1257{
1258        struct page *page;
1259        struct pagevec pvec;
1260        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1261
1262        pagevec_init(&pvec, 1);
1263
1264        /*
1265         * Put back any unfreeable pages.
1266         */
1267        spin_lock(&zone->lru_lock);
1268        while (!list_empty(page_list)) {
1269                int lru;
1270                page = lru_to_page(page_list);
1271                VM_BUG_ON(PageLRU(page));
1272                list_del(&page->lru);
1273                if (unlikely(!page_evictable(page, NULL))) {
1274                        spin_unlock_irq(&zone->lru_lock);
1275                        putback_lru_page(page);
1276                        spin_lock_irq(&zone->lru_lock);
1277                        continue;
1278                }
1279                SetPageLRU(page);
1280                lru = page_lru(page);
1281                add_page_to_lru_list(zone, page, lru);
1282                if (is_active_lru(lru)) {
1283                        int file = is_file_lru(lru);
1284                        int numpages = hpage_nr_pages(page);
1285                        reclaim_stat->recent_rotated[file] += numpages;
1286                }
1287                if (!pagevec_add(&pvec, page)) {
1288                        spin_unlock_irq(&zone->lru_lock);
1289                        __pagevec_release(&pvec);
1290                        spin_lock_irq(&zone->lru_lock);
1291                }
1292        }
1293        __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1294        __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1295
1296        spin_unlock_irq(&zone->lru_lock);
1297        pagevec_release(&pvec);
1298}
1299
1300static noinline_for_stack void update_isolated_counts(struct zone *zone,
1301                                        struct scan_control *sc,
1302                                        unsigned long *nr_anon,
1303                                        unsigned long *nr_file,
1304                                        struct list_head *isolated_list)
1305{
1306        unsigned long nr_active;
1307        unsigned int count[NR_LRU_LISTS] = { 0, };
1308        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1309
1310        nr_active = clear_active_flags(isolated_list, count);
1311        __count_vm_events(PGDEACTIVATE, nr_active);
1312
1313        __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1314                              -count[LRU_ACTIVE_FILE]);
1315        __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1316                              -count[LRU_INACTIVE_FILE]);
1317        __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1318                              -count[LRU_ACTIVE_ANON]);
1319        __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1320                              -count[LRU_INACTIVE_ANON]);
1321
1322        *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1323        *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1324        __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1325        __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1326
1327        reclaim_stat->recent_scanned[0] += *nr_anon;
1328        reclaim_stat->recent_scanned[1] += *nr_file;
1329}
1330
1331/*
1332 * Returns true if the caller should wait to clean dirty/writeback pages.
1333 *
1334 * If we are direct reclaiming for contiguous pages and we do not reclaim
1335 * everything in the list, try again and wait for writeback IO to complete.
1336 * This will stall high-order allocations noticeably. Only do that when really
1337 * need to free the pages under high memory pressure.
1338 */
1339static inline bool should_reclaim_stall(unsigned long nr_taken,
1340                                        unsigned long nr_freed,
1341                                        int priority,
1342                                        struct scan_control *sc)
1343{
1344        int lumpy_stall_priority;
1345
1346        /* kswapd should not stall on sync IO */
1347        if (current_is_kswapd())
1348                return false;
1349
1350        /* Only stall on lumpy reclaim */
1351        if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1352                return false;
1353
1354        /* If we have relaimed everything on the isolated list, no stall */
1355        if (nr_freed == nr_taken)
1356                return false;
1357
1358        /*
1359         * For high-order allocations, there are two stall thresholds.
1360         * High-cost allocations stall immediately where as lower
1361         * order allocations such as stacks require the scanning
1362         * priority to be much higher before stalling.
1363         */
1364        if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1365                lumpy_stall_priority = DEF_PRIORITY;
1366        else
1367                lumpy_stall_priority = DEF_PRIORITY / 3;
1368
1369        return priority <= lumpy_stall_priority;
1370}
1371
1372/*
1373 * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1374 * of reclaimed pages
1375 */
1376static noinline_for_stack unsigned long
1377shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1378                        struct scan_control *sc, int priority, int file)
1379{
1380        LIST_HEAD(page_list);
1381        unsigned long nr_scanned;
1382        unsigned long nr_reclaimed = 0;
1383        unsigned long nr_taken;
1384        unsigned long nr_anon;
1385        unsigned long nr_file;
1386
1387        while (unlikely(too_many_isolated(zone, file, sc))) {
1388                congestion_wait(BLK_RW_ASYNC, HZ/10);
1389
1390                /* We are about to die and free our memory. Return now. */
1391                if (fatal_signal_pending(current))
1392                        return SWAP_CLUSTER_MAX;
1393        }
1394
1395        set_reclaim_mode(priority, sc, false);
1396        lru_add_drain();
1397        spin_lock_irq(&zone->lru_lock);
1398
1399        if (scanning_global_lru(sc)) {
1400                nr_taken = isolate_pages_global(nr_to_scan,
1401                        &page_list, &nr_scanned, sc->order,
1402                        sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1403                                        ISOLATE_BOTH : ISOLATE_INACTIVE,
1404                        zone, 0, file);
1405                zone->pages_scanned += nr_scanned;
1406                if (current_is_kswapd())
1407                        __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1408                                               nr_scanned);
1409                else
1410                        __count_zone_vm_events(PGSCAN_DIRECT, zone,
1411                                               nr_scanned);
1412        } else {
1413                nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1414                        &page_list, &nr_scanned, sc->order,
1415                        sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1416                                        ISOLATE_BOTH : ISOLATE_INACTIVE,
1417                        zone, sc->mem_cgroup,
1418                        0, file);
1419                /*
1420                 * mem_cgroup_isolate_pages() keeps track of
1421                 * scanned pages on its own.
1422                 */
1423        }
1424
1425        if (nr_taken == 0) {
1426                spin_unlock_irq(&zone->lru_lock);
1427                return 0;
1428        }
1429
1430        update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1431
1432        spin_unlock_irq(&zone->lru_lock);
1433
1434        nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1435
1436        /* Check if we should syncronously wait for writeback */
1437        if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1438                set_reclaim_mode(priority, sc, true);
1439                nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1440        }
1441
1442        local_irq_disable();
1443        if (current_is_kswapd())
1444                __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1445        __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1446
1447        putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1448
1449        trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1450                zone_idx(zone),
1451                nr_scanned, nr_reclaimed,
1452                priority,
1453                trace_shrink_flags(file, sc->reclaim_mode));
1454        return nr_reclaimed;
1455}
1456
1457/*
1458 * This moves pages from the active list to the inactive list.
1459 *
1460 * We move them the other way if the page is referenced by one or more
1461 * processes, from rmap.
1462 *
1463 * If the pages are mostly unmapped, the processing is fast and it is
1464 * appropriate to hold zone->lru_lock across the whole operation.  But if
1465 * the pages are mapped, the processing is slow (page_referenced()) so we
1466 * should drop zone->lru_lock around each page.  It's impossible to balance
1467 * this, so instead we remove the pages from the LRU while processing them.
1468 * It is safe to rely on PG_active against the non-LRU pages in here because
1469 * nobody will play with that bit on a non-LRU page.
1470 *
1471 * The downside is that we have to touch page->_count against each page.
1472 * But we had to alter page->flags anyway.
1473 */
1474
1475static void move_active_pages_to_lru(struct zone *zone,
1476                                     struct list_head *list,
1477                                     enum lru_list lru)
1478{
1479        unsigned long pgmoved = 0;
1480        struct pagevec pvec;
1481        struct page *page;
1482
1483        pagevec_init(&pvec, 1);
1484
1485        while (!list_empty(list)) {
1486                page = lru_to_page(list);
1487
1488                VM_BUG_ON(PageLRU(page));
1489                SetPageLRU(page);
1490
1491                list_move(&page->lru, &zone->lru[lru].list);
1492                mem_cgroup_add_lru_list(page, lru);
1493                pgmoved += hpage_nr_pages(page);
1494
1495                if (!pagevec_add(&pvec, page) || list_empty(list)) {
1496                        spin_unlock_irq(&zone->lru_lock);
1497                        if (buffer_heads_over_limit)
1498                                pagevec_strip(&pvec);
1499                        __pagevec_release(&pvec);
1500                        spin_lock_irq(&zone->lru_lock);
1501                }
1502        }
1503        __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1504        if (!is_active_lru(lru))
1505                __count_vm_events(PGDEACTIVATE, pgmoved);
1506}
1507
1508static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1509                        struct scan_control *sc, int priority, int file)
1510{
1511        unsigned long nr_taken;
1512        unsigned long pgscanned;
1513        unsigned long vm_flags;
1514        LIST_HEAD(l_hold);      /* The pages which were snipped off */
1515        LIST_HEAD(l_active);
1516        LIST_HEAD(l_inactive);
1517        struct page *page;
1518        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1519        unsigned long nr_rotated = 0;
1520
1521        lru_add_drain();
1522        spin_lock_irq(&zone->lru_lock);
1523        if (scanning_global_lru(sc)) {
1524                nr_taken = isolate_pages_global(nr_pages, &l_hold,
1525                                                &pgscanned, sc->order,
1526                                                ISOLATE_ACTIVE, zone,
1527                                                1, file);
1528                zone->pages_scanned += pgscanned;
1529        } else {
1530                nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1531                                                &pgscanned, sc->order,
1532                                                ISOLATE_ACTIVE, zone,
1533                                                sc->mem_cgroup, 1, file);
1534                /*
1535                 * mem_cgroup_isolate_pages() keeps track of
1536                 * scanned pages on its own.
1537                 */
1538        }
1539
1540        reclaim_stat->recent_scanned[file] += nr_taken;
1541
1542        __count_zone_vm_events(PGREFILL, zone, pgscanned);
1543        if (file)
1544                __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1545        else
1546                __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1547        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1548        spin_unlock_irq(&zone->lru_lock);
1549
1550        while (!list_empty(&l_hold)) {
1551                cond_resched();
1552                page = lru_to_page(&l_hold);
1553                list_del(&page->lru);
1554
1555                if (unlikely(!page_evictable(page, NULL))) {
1556                        putback_lru_page(page);
1557                        continue;
1558                }
1559
1560                if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1561                        nr_rotated += hpage_nr_pages(page);
1562                        /*
1563                         * Identify referenced, file-backed active pages and
1564                         * give them one more trip around the active list. So
1565                         * that executable code get better chances to stay in
1566                         * memory under moderate memory pressure.  Anon pages
1567                         * are not likely to be evicted by use-once streaming
1568                         * IO, plus JVM can create lots of anon VM_EXEC pages,
1569                         * so we ignore them here.
1570                         */
1571                        if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1572                                list_add(&page->lru, &l_active);
1573                                continue;
1574                        }
1575                }
1576
1577                ClearPageActive(page);  /* we are de-activating */
1578                list_add(&page->lru, &l_inactive);
1579        }
1580
1581        /*
1582         * Move pages back to the lru list.
1583         */
1584        spin_lock_irq(&zone->lru_lock);
1585        /*
1586         * Count referenced pages from currently used mappings as rotated,
1587         * even though only some of them are actually re-activated.  This
1588         * helps balance scan pressure between file and anonymous pages in
1589         * get_scan_ratio.
1590         */
1591        reclaim_stat->recent_rotated[file] += nr_rotated;
1592
1593        move_active_pages_to_lru(zone, &l_active,
1594                                                LRU_ACTIVE + file * LRU_FILE);
1595        move_active_pages_to_lru(zone, &l_inactive,
1596                                                LRU_BASE   + file * LRU_FILE);
1597        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1598        spin_unlock_irq(&zone->lru_lock);
1599}
1600
1601#ifdef CONFIG_SWAP
1602static int inactive_anon_is_low_global(struct zone *zone)
1603{
1604        unsigned long active, inactive;
1605
1606        active = zone_page_state(zone, NR_ACTIVE_ANON);
1607        inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1608
1609        if (inactive * zone->inactive_ratio < active)
1610                return 1;
1611
1612        return 0;
1613}
1614
1615/**
1616 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1617 * @zone: zone to check
1618 * @sc:   scan control of this context
1619 *
1620 * Returns true if the zone does not have enough inactive anon pages,
1621 * meaning some active anon pages need to be deactivated.
1622 */
1623static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1624{
1625        int low;
1626
1627        /*
1628         * If we don't have swap space, anonymous page deactivation
1629         * is pointless.
1630         */
1631        if (!total_swap_pages)
1632                return 0;
1633
1634        if (scanning_global_lru(sc))
1635                low = inactive_anon_is_low_global(zone);
1636        else
1637                low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1638        return low;
1639}
1640#else
1641static inline int inactive_anon_is_low(struct zone *zone,
1642                                        struct scan_control *sc)
1643{
1644        return 0;
1645}
1646#endif
1647
1648static int inactive_file_is_low_global(struct zone *zone)
1649{
1650        unsigned long active, inactive;
1651
1652        active = zone_page_state(zone, NR_ACTIVE_FILE);
1653        inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1654
1655        return (active > inactive);
1656}
1657
1658/**
1659 * inactive_file_is_low - check if file pages need to be deactivated
1660 * @zone: zone to check
1661 * @sc:   scan control of this context
1662 *
1663 * When the system is doing streaming IO, memory pressure here
1664 * ensures that active file pages get deactivated, until more
1665 * than half of the file pages are on the inactive list.
1666 *
1667 * Once we get to that situation, protect the system's working
1668 * set from being evicted by disabling active file page aging.
1669 *
1670 * This uses a different ratio than the anonymous pages, because
1671 * the page cache uses a use-once replacement algorithm.
1672 */
1673static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1674{
1675        int low;
1676
1677        if (scanning_global_lru(sc))
1678                low = inactive_file_is_low_global(zone);
1679        else
1680                low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1681        return low;
1682}
1683
1684static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1685                                int file)
1686{
1687        if (file)
1688                return inactive_file_is_low(zone, sc);
1689        else
1690                return inactive_anon_is_low(zone, sc);
1691}
1692
1693static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1694        struct zone *zone, struct scan_control *sc, int priority)
1695{
1696        int file = is_file_lru(lru);
1697
1698        if (is_active_lru(lru)) {
1699                if (inactive_list_is_low(zone, sc, file))
1700                    shrink_active_list(nr_to_scan, zone, sc, priority, file);
1701                return 0;
1702        }
1703
1704        return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1705}
1706
1707/*
1708 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1709 * until we collected @swap_cluster_max pages to scan.
1710 */
1711static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1712                                       unsigned long *nr_saved_scan)
1713{
1714        unsigned long nr;
1715
1716        *nr_saved_scan += nr_to_scan;
1717        nr = *nr_saved_scan;
1718
1719        if (nr >= SWAP_CLUSTER_MAX)
1720                *nr_saved_scan = 0;
1721        else
1722                nr = 0;
1723
1724        return nr;
1725}
1726
1727/*
1728 * Determine how aggressively the anon and file LRU lists should be
1729 * scanned.  The relative value of each set of LRU lists is determined
1730 * by looking at the fraction of the pages scanned we did rotate back
1731 * onto the active list instead of evict.
1732 *
1733 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1734 */
1735static void get_scan_count(struct zone *zone, struct scan_control *sc,
1736                                        unsigned long *nr, int priority)
1737{
1738        unsigned long anon, file, free;
1739        unsigned long anon_prio, file_prio;
1740        unsigned long ap, fp;
1741        struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1742        u64 fraction[2], denominator;
1743        enum lru_list l;
1744        int noswap = 0;
1745
1746        /* If we have no swap space, do not bother scanning anon pages. */
1747        if (!sc->may_swap || (nr_swap_pages <= 0)) {
1748                noswap = 1;
1749                fraction[0] = 0;
1750                fraction[1] = 1;
1751                denominator = 1;
1752                goto out;
1753        }
1754
1755        anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1756                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1757        file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1758                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1759
1760        if (scanning_global_lru(sc)) {
1761                free  = zone_page_state(zone, NR_FREE_PAGES);
1762                /* If we have very few page cache pages,
1763                   force-scan anon pages. */
1764                if (unlikely(file + free <= high_wmark_pages(zone))) {
1765                        fraction[0] = 1;
1766                        fraction[1] = 0;
1767                        denominator = 1;
1768                        goto out;
1769                }
1770        }
1771
1772        /*
1773         * With swappiness at 100, anonymous and file have the same priority.
1774         * This scanning priority is essentially the inverse of IO cost.
1775         */
1776        anon_prio = sc->swappiness;
1777        file_prio = 200 - sc->swappiness;
1778
1779        /*
1780         * OK, so we have swap space and a fair amount of page cache
1781         * pages.  We use the recently rotated / recently scanned
1782         * ratios to determine how valuable each cache is.
1783         *
1784         * Because workloads change over time (and to avoid overflow)
1785         * we keep these statistics as a floating average, which ends
1786         * up weighing recent references more than old ones.
1787         *
1788         * anon in [0], file in [1]
1789         */
1790        spin_lock_irq(&zone->lru_lock);
1791        if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1792                reclaim_stat->recent_scanned[0] /= 2;
1793                reclaim_stat->recent_rotated[0] /= 2;
1794        }
1795
1796        if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1797                reclaim_stat->recent_scanned[1] /= 2;
1798                reclaim_stat->recent_rotated[1] /= 2;
1799        }
1800
1801        /*
1802         * The amount of pressure on anon vs file pages is inversely
1803         * proportional to the fraction of recently scanned pages on
1804         * each list that were recently referenced and in active use.
1805         */
1806        ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1807        ap /= reclaim_stat->recent_rotated[0] + 1;
1808
1809        fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1810        fp /= reclaim_stat->recent_rotated[1] + 1;
1811        spin_unlock_irq(&zone->lru_lock);
1812
1813        fraction[0] = ap;
1814        fraction[1] = fp;
1815        denominator = ap + fp + 1;
1816out:
1817        for_each_evictable_lru(l) {
1818                int file = is_file_lru(l);
1819                unsigned long scan;
1820
1821                scan = zone_nr_lru_pages(zone, sc, l);
1822                if (priority || noswap) {
1823                        scan >>= priority;
1824                        scan = div64_u64(scan * fraction[file], denominator);
1825                }
1826                nr[l] = nr_scan_try_batch(scan,
1827                                          &reclaim_stat->nr_saved_scan[l]);
1828        }
1829}
1830
1831/*
1832 * Reclaim/compaction depends on a number of pages being freed. To avoid
1833 * disruption to the system, a small number of order-0 pages continue to be
1834 * rotated and reclaimed in the normal fashion. However, by the time we get
1835 * back to the allocator and call try_to_compact_zone(), we ensure that
1836 * there are enough free pages for it to be likely successful
1837 */
1838static inline bool should_continue_reclaim(struct zone *zone,
1839                                        unsigned long nr_reclaimed,
1840                                        unsigned long nr_scanned,
1841                                        struct scan_control *sc)
1842{
1843        unsigned long pages_for_compaction;
1844        unsigned long inactive_lru_pages;
1845
1846        /* If not in reclaim/compaction mode, stop */
1847        if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1848                return false;
1849
1850        /* Consider stopping depending on scan and reclaim activity */
1851        if (sc->gfp_mask & __GFP_REPEAT) {
1852                /*
1853                 * For __GFP_REPEAT allocations, stop reclaiming if the
1854                 * full LRU list has been scanned and we are still failing
1855                 * to reclaim pages. This full LRU scan is potentially
1856                 * expensive but a __GFP_REPEAT caller really wants to succeed
1857                 */
1858                if (!nr_reclaimed && !nr_scanned)
1859                        return false;
1860        } else {
1861                /*
1862                 * For non-__GFP_REPEAT allocations which can presumably
1863                 * fail without consequence, stop if we failed to reclaim
1864                 * any pages from the last SWAP_CLUSTER_MAX number of
1865                 * pages that were scanned. This will return to the
1866                 * caller faster at the risk reclaim/compaction and
1867                 * the resulting allocation attempt fails
1868                 */
1869                if (!nr_reclaimed)
1870                        return false;
1871        }
1872
1873        /*
1874         * If we have not reclaimed enough pages for compaction and the
1875         * inactive lists are large enough, continue reclaiming
1876         */
1877        pages_for_compaction = (2UL << sc->order);
1878        inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1879                                zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1880        if (sc->nr_reclaimed < pages_for_compaction &&
1881                        inactive_lru_pages > pages_for_compaction)
1882                return true;
1883
1884        /* If compaction would go ahead or the allocation would succeed, stop */
1885        switch (compaction_suitable(zone, sc->order)) {
1886        case COMPACT_PARTIAL:
1887        case COMPACT_CONTINUE:
1888                return false;
1889        default:
1890                return true;
1891        }
1892}
1893
1894/*
1895 * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1896 */
1897static void shrink_zone(int priority, struct zone *zone,
1898                                struct scan_control *sc)
1899{
1900        unsigned long nr[NR_LRU_LISTS];
1901        unsigned long nr_to_scan;
1902        enum lru_list l;
1903        unsigned long nr_reclaimed, nr_scanned;
1904        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1905
1906restart:
1907        nr_reclaimed = 0;
1908        nr_scanned = sc->nr_scanned;
1909        get_scan_count(zone, sc, nr, priority);
1910
1911        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1912                                        nr[LRU_INACTIVE_FILE]) {
1913                for_each_evictable_lru(l) {
1914                        if (nr[l]) {
1915                                nr_to_scan = min_t(unsigned long,
1916                                                   nr[l], SWAP_CLUSTER_MAX);
1917                                nr[l] -= nr_to_scan;
1918
1919                                nr_reclaimed += shrink_list(l, nr_to_scan,
1920                                                            zone, sc, priority);
1921                        }
1922                }
1923                /*
1924                 * On large memory systems, scan >> priority can become
1925                 * really large. This is fine for the starting priority;
1926                 * we want to put equal scanning pressure on each zone.
1927                 * However, if the VM has a harder time of freeing pages,
1928                 * with multiple processes reclaiming pages, the total
1929                 * freeing target can get unreasonably large.
1930                 */
1931                if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1932                        break;
1933        }
1934        sc->nr_reclaimed += nr_reclaimed;
1935
1936        /*
1937         * Even if we did not try to evict anon pages at all, we want to
1938         * rebalance the anon lru active/inactive ratio.
1939         */
1940        if (inactive_anon_is_low(zone, sc))
1941                shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1942
1943        /* reclaim/compaction might need reclaim to continue */
1944        if (should_continue_reclaim(zone, nr_reclaimed,
1945                                        sc->nr_scanned - nr_scanned, sc))
1946                goto restart;
1947
1948        throttle_vm_writeout(sc->gfp_mask);
1949}
1950
1951/*
1952 * This is the direct reclaim path, for page-allocating processes.  We only
1953 * try to reclaim pages from zones which will satisfy the caller's allocation
1954 * request.
1955 *
1956 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1957 * Because:
1958 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1959 *    allocation or
1960 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1961 *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1962 *    zone defense algorithm.
1963 *
1964 * If a zone is deemed to be full of pinned pages then just give it a light
1965 * scan then give up on it.
1966 */
1967static void shrink_zones(int priority, struct zonelist *zonelist,
1968                                        struct scan_control *sc)
1969{
1970        struct zoneref *z;
1971        struct zone *zone;
1972
1973        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1974                                        gfp_zone(sc->gfp_mask), sc->nodemask) {
1975                if (!populated_zone(zone))
1976                        continue;
1977                /*
1978                 * Take care memory controller reclaiming has small influence
1979                 * to global LRU.
1980                 */
1981                if (scanning_global_lru(sc)) {
1982                        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1983                                continue;
1984                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1985                                continue;       /* Let kswapd poll it */
1986                }
1987
1988                shrink_zone(priority, zone, sc);
1989        }
1990}
1991
1992static bool zone_reclaimable(struct zone *zone)
1993{
1994        return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1995}
1996
1997/* All zones in zonelist are unreclaimable? */
1998static bool all_unreclaimable(struct zonelist *zonelist,
1999                struct scan_control *sc)
2000{
2001        struct zoneref *z;
2002        struct zone *zone;
2003
2004        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2005                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2006                if (!populated_zone(zone))
2007                        continue;
2008                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2009                        continue;
2010                if (!zone->all_unreclaimable)
2011                        return false;
2012        }
2013
2014        return true;
2015}
2016
2017/*
2018 * This is the main entry point to direct page reclaim.
2019 *
2020 * If a full scan of the inactive list fails to free enough memory then we
2021 * are "out of memory" and something needs to be killed.
2022 *
2023 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2024 * high - the zone may be full of dirty or under-writeback pages, which this
2025 * caller can't do much about.  We kick the writeback threads and take explicit
2026 * naps in the hope that some of these pages can be written.  But if the
2027 * allocating task holds filesystem locks which prevent writeout this might not
2028 * work, and the allocation attempt will fail.
2029 *
2030 * returns:     0, if no pages reclaimed
2031 *              else, the number of pages reclaimed
2032 */
2033static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2034                                        struct scan_control *sc)
2035{
2036        int priority;
2037        unsigned long total_scanned = 0;
2038        struct reclaim_state *reclaim_state = current->reclaim_state;
2039        struct zoneref *z;
2040        struct zone *zone;
2041        unsigned long writeback_threshold;
2042
2043        get_mems_allowed();
2044        delayacct_freepages_start();
2045
2046        if (scanning_global_lru(sc))
2047                count_vm_event(ALLOCSTALL);
2048
2049        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2050                sc->nr_scanned = 0;
2051                if (!priority)
2052                        disable_swap_token();
2053                shrink_zones(priority, zonelist, sc);
2054                /*
2055                 * Don't shrink slabs when reclaiming memory from
2056                 * over limit cgroups
2057                 */
2058                if (scanning_global_lru(sc)) {
2059                        unsigned long lru_pages = 0;
2060                        for_each_zone_zonelist(zone, z, zonelist,
2061                                        gfp_zone(sc->gfp_mask)) {
2062                                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2063                                        continue;
2064
2065                                lru_pages += zone_reclaimable_pages(zone);
2066                        }
2067
2068                        shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2069                        if (reclaim_state) {
2070                                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2071                                reclaim_state->reclaimed_slab = 0;
2072                        }
2073                }
2074                total_scanned += sc->nr_scanned;
2075                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2076                        goto out;
2077
2078                /*
2079                 * Try to write back as many pages as we just scanned.  This
2080                 * tends to cause slow streaming writers to write data to the
2081                 * disk smoothly, at the dirtying rate, which is nice.   But
2082                 * that's undesirable in laptop mode, where we *want* lumpy
2083                 * writeout.  So in laptop mode, write out the whole world.
2084                 */
2085                writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2086                if (total_scanned > writeback_threshold) {
2087                        wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2088                        sc->may_writepage = 1;
2089                }
2090
2091                /* Take a nap, wait for some writeback to complete */
2092                if (!sc->hibernation_mode && sc->nr_scanned &&
2093                    priority < DEF_PRIORITY - 2) {
2094                        struct zone *preferred_zone;
2095
2096                        first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2097                                                &cpuset_current_mems_allowed,
2098                                                &preferred_zone);
2099                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2100                }
2101        }
2102
2103out:
2104        delayacct_freepages_end();
2105        put_mems_allowed();
2106
2107        if (sc->nr_reclaimed)
2108                return sc->nr_reclaimed;
2109
2110        /*
2111         * As hibernation is going on, kswapd is freezed so that it can't mark
2112         * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2113         * check.
2114         */
2115        if (oom_killer_disabled)
2116                return 0;
2117
2118        /* top priority shrink_zones still had more to do? don't OOM, then */
2119        if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2120                return 1;
2121
2122        return 0;
2123}
2124
2125unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2126                                gfp_t gfp_mask, nodemask_t *nodemask)
2127{
2128        unsigned long nr_reclaimed;
2129        struct scan_control sc = {
2130                .gfp_mask = gfp_mask,
2131                .may_writepage = !laptop_mode,
2132                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2133                .may_unmap = 1,
2134                .may_swap = 1,
2135                .swappiness = vm_swappiness,
2136                .order = order,
2137                .mem_cgroup = NULL,
2138                .nodemask = nodemask,
2139        };
2140
2141        trace_mm_vmscan_direct_reclaim_begin(order,
2142                                sc.may_writepage,
2143                                gfp_mask);
2144
2145        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2146
2147        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2148
2149        return nr_reclaimed;
2150}
2151
2152#ifdef CONFIG_CGROUP_MEM_RES_CTLR
2153
2154unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2155                                                gfp_t gfp_mask, bool noswap,
2156                                                unsigned int swappiness,
2157                                                struct zone *zone)
2158{
2159        struct scan_control sc = {
2160                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2161                .may_writepage = !laptop_mode,
2162                .may_unmap = 1,
2163                .may_swap = !noswap,
2164                .swappiness = swappiness,
2165                .order = 0,
2166                .mem_cgroup = mem,
2167        };
2168        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2169                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2170
2171        trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2172                                                      sc.may_writepage,
2173                                                      sc.gfp_mask);
2174
2175        /*
2176         * NOTE: Although we can get the priority field, using it
2177         * here is not a good idea, since it limits the pages we can scan.
2178         * if we don't reclaim here, the shrink_zone from balance_pgdat
2179         * will pick up pages from other mem cgroup's as well. We hack
2180         * the priority and make it zero.
2181         */
2182        shrink_zone(0, zone, &sc);
2183
2184        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2185
2186        return sc.nr_reclaimed;
2187}
2188
2189unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2190                                           gfp_t gfp_mask,
2191                                           bool noswap,
2192                                           unsigned int swappiness)
2193{
2194        struct zonelist *zonelist;
2195        unsigned long nr_reclaimed;
2196        struct scan_control sc = {
2197                .may_writepage = !laptop_mode,
2198                .may_unmap = 1,
2199                .may_swap = !noswap,
2200                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2201                .swappiness = swappiness,
2202                .order = 0,
2203                .mem_cgroup = mem_cont,
2204                .nodemask = NULL, /* we don't care the placement */
2205        };
2206
2207        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2208                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2209        zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2210
2211        trace_mm_vmscan_memcg_reclaim_begin(0,
2212                                            sc.may_writepage,
2213                                            sc.gfp_mask);
2214
2215        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2216
2217        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2218
2219        return nr_reclaimed;
2220}
2221#endif
2222
2223/*
2224 * pgdat_balanced is used when checking if a node is balanced for high-order
2225 * allocations. Only zones that meet watermarks and are in a zone allowed
2226 * by the callers classzone_idx are added to balanced_pages. The total of
2227 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2228 * for the node to be considered balanced. Forcing all zones to be balanced
2229 * for high orders can cause excessive reclaim when there are imbalanced zones.
2230 * The choice of 25% is due to
2231 *   o a 16M DMA zone that is balanced will not balance a zone on any
2232 *     reasonable sized machine
2233 *   o On all other machines, the top zone must be at least a reasonable
2234 *     percentage of the middle zones. For example, on 32-bit x86, highmem
2235 *     would need to be at least 256M for it to be balance a whole node.
2236 *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2237 *     to balance a node on its own. These seemed like reasonable ratios.
2238 */
2239static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2240                                                int classzone_idx)
2241{
2242        unsigned long present_pages = 0;
2243        int i;
2244
2245        for (i = 0; i <= classzone_idx; i++)
2246                present_pages += pgdat->node_zones[i].present_pages;
2247
2248        return balanced_pages > (present_pages >> 2);
2249}
2250
2251/* is kswapd sleeping prematurely? */
2252static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2253                                        int classzone_idx)
2254{
2255        int i;
2256        unsigned long balanced = 0;
2257        bool all_zones_ok = true;
2258
2259        /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2260        if (remaining)
2261                return true;
2262
2263        /* Check the watermark levels */
2264        for (i = 0; i < pgdat->nr_zones; i++) {
2265                struct zone *zone = pgdat->node_zones + i;
2266
2267                if (!populated_zone(zone))
2268                        continue;
2269
2270                /*
2271                 * balance_pgdat() skips over all_unreclaimable after
2272                 * DEF_PRIORITY. Effectively, it considers them balanced so
2273                 * they must be considered balanced here as well if kswapd
2274                 * is to sleep
2275                 */
2276                if (zone->all_unreclaimable) {
2277                        balanced += zone->present_pages;
2278                        continue;
2279                }
2280
2281                if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2282                                                        classzone_idx, 0))
2283                        all_zones_ok = false;
2284                else
2285                        balanced += zone->present_pages;
2286        }
2287
2288        /*
2289         * For high-order requests, the balanced zones must contain at least
2290         * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2291         * must be balanced
2292         */
2293        if (order)
2294                return !pgdat_balanced(pgdat, balanced, classzone_idx);
2295        else
2296                return !all_zones_ok;
2297}
2298
2299/*
2300 * For kswapd, balance_pgdat() will work across all this node's zones until
2301 * they are all at high_wmark_pages(zone).
2302 *
2303 * Returns the final order kswapd was reclaiming at
2304 *
2305 * There is special handling here for zones which are full of pinned pages.
2306 * This can happen if the pages are all mlocked, or if they are all used by
2307 * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2308 * What we do is to detect the case where all pages in the zone have been
2309 * scanned twice and there has been zero successful reclaim.  Mark the zone as
2310 * dead and from now on, only perform a short scan.  Basically we're polling
2311 * the zone for when the problem goes away.
2312 *
2313 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2314 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2315 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2316 * lower zones regardless of the number of free pages in the lower zones. This
2317 * interoperates with the page allocator fallback scheme to ensure that aging
2318 * of pages is balanced across the zones.
2319 */
2320static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2321                                                        int *classzone_idx)
2322{
2323        int all_zones_ok;
2324        unsigned long balanced;
2325        int priority;
2326        int i;
2327        int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2328        unsigned long total_scanned;
2329        struct reclaim_state *reclaim_state = current->reclaim_state;
2330        struct scan_control sc = {
2331                .gfp_mask = GFP_KERNEL,
2332                .may_unmap = 1,
2333                .may_swap = 1,
2334                /*
2335                 * kswapd doesn't want to be bailed out while reclaim. because
2336                 * we want to put equal scanning pressure on each zone.
2337                 */
2338                .nr_to_reclaim = ULONG_MAX,
2339                .swappiness = vm_swappiness,
2340                .order = order,
2341                .mem_cgroup = NULL,
2342        };
2343loop_again:
2344        total_scanned = 0;
2345        sc.nr_reclaimed = 0;
2346        sc.may_writepage = !laptop_mode;
2347        count_vm_event(PAGEOUTRUN);
2348
2349        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2350                unsigned long lru_pages = 0;
2351                int has_under_min_watermark_zone = 0;
2352
2353                /* The swap token gets in the way of swapout... */
2354                if (!priority)
2355                        disable_swap_token();
2356
2357                all_zones_ok = 1;
2358                balanced = 0;
2359
2360                /*
2361                 * Scan in the highmem->dma direction for the highest
2362                 * zone which needs scanning
2363                 */
2364                for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2365                        struct zone *zone = pgdat->node_zones + i;
2366
2367                        if (!populated_zone(zone))
2368                                continue;
2369
2370                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2371                                continue;
2372
2373                        /*
2374                         * Do some background aging of the anon list, to give
2375                         * pages a chance to be referenced before reclaiming.
2376                         */
2377                        if (inactive_anon_is_low(zone, &sc))
2378                                shrink_active_list(SWAP_CLUSTER_MAX, zone,
2379                                                        &sc, priority, 0);
2380
2381                        if (!zone_watermark_ok_safe(zone, order,
2382                                        high_wmark_pages(zone), 0, 0)) {
2383                                end_zone = i;
2384                                *classzone_idx = i;
2385                                break;
2386                        }
2387                }
2388                if (i < 0)
2389                        goto out;
2390
2391                for (i = 0; i <= end_zone; i++) {
2392                        struct zone *zone = pgdat->node_zones + i;
2393
2394                        lru_pages += zone_reclaimable_pages(zone);
2395                }
2396
2397                /*
2398                 * Now scan the zone in the dma->highmem direction, stopping
2399                 * at the last zone which needs scanning.
2400                 *
2401                 * We do this because the page allocator works in the opposite
2402                 * direction.  This prevents the page allocator from allocating
2403                 * pages behind kswapd's direction of progress, which would
2404                 * cause too much scanning of the lower zones.
2405                 */
2406                for (i = 0; i <= end_zone; i++) {
2407                        struct zone *zone = pgdat->node_zones + i;
2408                        int nr_slab;
2409                        unsigned long balance_gap;
2410
2411                        if (!populated_zone(zone))
2412                                continue;
2413
2414                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2415                                continue;
2416
2417                        sc.nr_scanned = 0;
2418
2419                        /*
2420                         * Call soft limit reclaim before calling shrink_zone.
2421                         * For now we ignore the return value
2422                         */
2423                        mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2424
2425                        /*
2426                         * We put equal pressure on every zone, unless
2427                         * one zone has way too many pages free
2428                         * already. The "too many pages" is defined
2429                         * as the high wmark plus a "gap" where the
2430                         * gap is either the low watermark or 1%
2431                         * of the zone, whichever is smaller.
2432                         */
2433                        balance_gap = min(low_wmark_pages(zone),
2434                                (zone->present_pages +
2435                                        KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2436                                KSWAPD_ZONE_BALANCE_GAP_RATIO);
2437                        if (!zone_watermark_ok_safe(zone, order,
2438                                        high_wmark_pages(zone) + balance_gap,
2439                                        end_zone, 0))
2440                                shrink_zone(priority, zone, &sc);
2441                        reclaim_state->reclaimed_slab = 0;
2442                        nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2443                                                lru_pages);
2444                        sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2445                        total_scanned += sc.nr_scanned;
2446
2447                        if (zone->all_unreclaimable)
2448                                continue;
2449                        if (nr_slab == 0 &&
2450                            !zone_reclaimable(zone))
2451                                zone->all_unreclaimable = 1;
2452                        /*
2453                         * If we've done a decent amount of scanning and
2454                         * the reclaim ratio is low, start doing writepage
2455                         * even in laptop mode
2456                         */
2457                        if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2458                            total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2459                                sc.may_writepage = 1;
2460
2461                        if (!zone_watermark_ok_safe(zone, order,
2462                                        high_wmark_pages(zone), end_zone, 0)) {
2463                                all_zones_ok = 0;
2464                                /*
2465                                 * We are still under min water mark.  This
2466                                 * means that we have a GFP_ATOMIC allocation
2467                                 * failure risk. Hurry up!
2468                                 */
2469                                if (!zone_watermark_ok_safe(zone, order,
2470                                            min_wmark_pages(zone), end_zone, 0))
2471                                        has_under_min_watermark_zone = 1;
2472                        } else {
2473                                /*
2474                                 * If a zone reaches its high watermark,
2475                                 * consider it to be no longer congested. It's
2476                                 * possible there are dirty pages backed by
2477                                 * congested BDIs but as pressure is relieved,
2478                                 * spectulatively avoid congestion waits
2479                                 */
2480                                zone_clear_flag(zone, ZONE_CONGESTED);
2481                                if (i <= *classzone_idx)
2482                                        balanced += zone->present_pages;
2483                        }
2484
2485                }
2486                if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2487                        break;          /* kswapd: all done */
2488                /*
2489                 * OK, kswapd is getting into trouble.  Take a nap, then take
2490                 * another pass across the zones.
2491                 */
2492                if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2493                        if (has_under_min_watermark_zone)
2494                                count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2495                        else
2496                                congestion_wait(BLK_RW_ASYNC, HZ/10);
2497                }
2498
2499                /*
2500                 * We do this so kswapd doesn't build up large priorities for
2501                 * example when it is freeing in parallel with allocators. It
2502                 * matches the direct reclaim path behaviour in terms of impact
2503                 * on zone->*_priority.
2504                 */
2505                if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2506                        break;
2507        }
2508out:
2509
2510        /*
2511         * order-0: All zones must meet high watermark for a balanced node
2512         * high-order: Balanced zones must make up at least 25% of the node
2513         *             for the node to be balanced
2514         */
2515        if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2516                cond_resched();
2517
2518                try_to_freeze();
2519
2520                /*
2521                 * Fragmentation may mean that the system cannot be
2522                 * rebalanced for high-order allocations in all zones.
2523                 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2524                 * it means the zones have been fully scanned and are still
2525                 * not balanced. For high-order allocations, there is
2526                 * little point trying all over again as kswapd may
2527                 * infinite loop.
2528                 *
2529                 * Instead, recheck all watermarks at order-0 as they
2530                 * are the most important. If watermarks are ok, kswapd will go
2531                 * back to sleep. High-order users can still perform direct
2532                 * reclaim if they wish.
2533                 */
2534                if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2535                        order = sc.order = 0;
2536
2537                goto loop_again;
2538        }
2539
2540        /*
2541         * If kswapd was reclaiming at a higher order, it has the option of
2542         * sleeping without all zones being balanced. Before it does, it must
2543         * ensure that the watermarks for order-0 on *all* zones are met and
2544         * that the congestion flags are cleared. The congestion flag must
2545         * be cleared as kswapd is the only mechanism that clears the flag
2546         * and it is potentially going to sleep here.
2547         */
2548        if (order) {
2549                for (i = 0; i <= end_zone; i++) {
2550                        struct zone *zone = pgdat->node_zones + i;
2551
2552                        if (!populated_zone(zone))
2553                                continue;
2554
2555                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2556                                continue;
2557
2558                        /* Confirm the zone is balanced for order-0 */
2559                        if (!zone_watermark_ok(zone, 0,
2560                                        high_wmark_pages(zone), 0, 0)) {
2561                                order = sc.order = 0;
2562                                goto loop_again;
2563                        }
2564
2565                        /* If balanced, clear the congested flag */
2566                        zone_clear_flag(zone, ZONE_CONGESTED);
2567                }
2568        }
2569
2570        /*
2571         * Return the order we were reclaiming at so sleeping_prematurely()
2572         * makes a decision on the order we were last reclaiming at. However,
2573         * if another caller entered the allocator slow path while kswapd
2574         * was awake, order will remain at the higher level
2575         */
2576        *classzone_idx = end_zone;
2577        return order;
2578}
2579
2580static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2581{
2582        long remaining = 0;
2583        DEFINE_WAIT(wait);
2584
2585        if (freezing(current) || kthread_should_stop())
2586                return;
2587
2588        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2589
2590        /* Try to sleep for a short interval */
2591        if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2592                remaining = schedule_timeout(HZ/10);
2593                finish_wait(&pgdat->kswapd_wait, &wait);
2594                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2595        }
2596
2597        /*
2598         * After a short sleep, check if it was a premature sleep. If not, then
2599         * go fully to sleep until explicitly woken up.
2600         */
2601        if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2602                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2603
2604                /*
2605                 * vmstat counters are not perfectly accurate and the estimated
2606                 * value for counters such as NR_FREE_PAGES can deviate from the
2607                 * true value by nr_online_cpus * threshold. To avoid the zone
2608                 * watermarks being breached while under pressure, we reduce the
2609                 * per-cpu vmstat threshold while kswapd is awake and restore
2610                 * them before going back to sleep.
2611                 */
2612                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2613                schedule();
2614                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2615        } else {
2616                if (remaining)
2617                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2618                else
2619                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2620        }
2621        finish_wait(&pgdat->kswapd_wait, &wait);
2622}
2623
2624/*
2625 * The background pageout daemon, started as a kernel thread
2626 * from the init process.
2627 *
2628 * This basically trickles out pages so that we have _some_
2629 * free memory available even if there is no other activity
2630 * that frees anything up. This is needed for things like routing
2631 * etc, where we otherwise might have all activity going on in
2632 * asynchronous contexts that cannot page things out.
2633 *
2634 * If there are applications that are active memory-allocators
2635 * (most normal use), this basically shouldn't matter.
2636 */
2637static int kswapd(void *p)
2638{
2639        unsigned long order;
2640        int classzone_idx;
2641        pg_data_t *pgdat = (pg_data_t*)p;
2642        struct task_struct *tsk = current;
2643
2644        struct reclaim_state reclaim_state = {
2645                .reclaimed_slab = 0,
2646        };
2647        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2648
2649        lockdep_set_current_reclaim_state(GFP_KERNEL);
2650
2651        if (!cpumask_empty(cpumask))
2652                set_cpus_allowed_ptr(tsk, cpumask);
2653        current->reclaim_state = &reclaim_state;
2654
2655        /*
2656         * Tell the memory management that we're a "memory allocator",
2657         * and that if we need more memory we should get access to it
2658         * regardless (see "__alloc_pages()"). "kswapd" should
2659         * never get caught in the normal page freeing logic.
2660         *
2661         * (Kswapd normally doesn't need memory anyway, but sometimes
2662         * you need a small amount of memory in order to be able to
2663         * page out something else, and this flag essentially protects
2664         * us from recursively trying to free more memory as we're
2665         * trying to free the first piece of memory in the first place).
2666         */
2667        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2668        set_freezable();
2669
2670        order = 0;
2671        classzone_idx = MAX_NR_ZONES - 1;
2672        for ( ; ; ) {
2673                unsigned long new_order;
2674                int new_classzone_idx;
2675                int ret;
2676
2677                new_order = pgdat->kswapd_max_order;
2678                new_classzone_idx = pgdat->classzone_idx;
2679                pgdat->kswapd_max_order = 0;
2680                pgdat->classzone_idx = MAX_NR_ZONES - 1;
2681                if (order < new_order || classzone_idx > new_classzone_idx) {
2682                        /*
2683                         * Don't sleep if someone wants a larger 'order'
2684                         * allocation or has tigher zone constraints
2685                         */
2686                        order = new_order;
2687                        classzone_idx = new_classzone_idx;
2688                } else {
2689                        kswapd_try_to_sleep(pgdat, order, classzone_idx);
2690                        order = pgdat->kswapd_max_order;
2691                        classzone_idx = pgdat->classzone_idx;
2692                        pgdat->kswapd_max_order = 0;
2693                        pgdat->classzone_idx = MAX_NR_ZONES - 1;
2694                }
2695
2696                ret = try_to_freeze();
2697                if (kthread_should_stop())
2698                        break;
2699
2700                /*
2701                 * We can speed up thawing tasks if we don't call balance_pgdat
2702                 * after returning from the refrigerator
2703                 */
2704                if (!ret) {
2705                        trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2706                        order = balance_pgdat(pgdat, order, &classzone_idx);
2707                }
2708        }
2709        return 0;
2710}
2711
2712/*
2713 * A zone is low on free memory, so wake its kswapd task to service it.
2714 */
2715void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2716{
2717        pg_data_t *pgdat;
2718
2719        if (!populated_zone(zone))
2720                return;
2721
2722        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2723                return;
2724        pgdat = zone->zone_pgdat;
2725        if (pgdat->kswapd_max_order < order) {
2726                pgdat->kswapd_max_order = order;
2727                pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2728        }
2729        if (!waitqueue_active(&pgdat->kswapd_wait))
2730                return;
2731        if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2732                return;
2733
2734        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2735        wake_up_interruptible(&pgdat->kswapd_wait);
2736}
2737
2738/*
2739 * The reclaimable count would be mostly accurate.
2740 * The less reclaimable pages may be
2741 * - mlocked pages, which will be moved to unevictable list when encountered
2742 * - mapped pages, which may require several travels to be reclaimed
2743 * - dirty pages, which is not "instantly" reclaimable
2744 */
2745unsigned long global_reclaimable_pages(void)
2746{
2747        int nr;
2748
2749        nr = global_page_state(NR_ACTIVE_FILE) +
2750             global_page_state(NR_INACTIVE_FILE);
2751
2752        if (nr_swap_pages > 0)
2753                nr += global_page_state(NR_ACTIVE_ANON) +
2754                      global_page_state(NR_INACTIVE_ANON);
2755
2756        return nr;
2757}
2758
2759unsigned long zone_reclaimable_pages(struct zone *zone)
2760{
2761        int nr;
2762
2763        nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2764             zone_page_state(zone, NR_INACTIVE_FILE);
2765
2766        if (nr_swap_pages > 0)
2767                nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2768                      zone_page_state(zone, NR_INACTIVE_ANON);
2769
2770        return nr;
2771}
2772
2773#ifdef CONFIG_HIBERNATION
2774/*
2775 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2776 * freed pages.
2777 *
2778 * Rather than trying to age LRUs the aim is to preserve the overall
2779 * LRU order by reclaiming preferentially
2780 * inactive > active > active referenced > active mapped
2781 */
2782unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2783{
2784        struct reclaim_state reclaim_state;
2785        struct scan_control sc = {
2786                .gfp_mask = GFP_HIGHUSER_MOVABLE,
2787                .may_swap = 1,
2788                .may_unmap = 1,
2789                .may_writepage = 1,
2790                .nr_to_reclaim = nr_to_reclaim,
2791                .hibernation_mode = 1,
2792                .swappiness = vm_swappiness,
2793                .order = 0,
2794        };
2795        struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2796        struct task_struct *p = current;
2797        unsigned long nr_reclaimed;
2798
2799        p->flags |= PF_MEMALLOC;
2800        lockdep_set_current_reclaim_state(sc.gfp_mask);
2801        reclaim_state.reclaimed_slab = 0;
2802        p->reclaim_state = &reclaim_state;
2803
2804        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2805
2806        p->reclaim_state = NULL;
2807        lockdep_clear_current_reclaim_state();
2808        p->flags &= ~PF_MEMALLOC;
2809
2810        return nr_reclaimed;
2811}
2812#endif /* CONFIG_HIBERNATION */
2813
2814/* It's optimal to keep kswapds on the same CPUs as their memory, but
2815   not required for correctness.  So if the last cpu in a node goes
2816   away, we get changed to run anywhere: as the first one comes back,
2817   restore their cpu bindings. */
2818static int __devinit cpu_callback(struct notifier_block *nfb,
2819                                  unsigned long action, void *hcpu)
2820{
2821        int nid;
2822
2823        if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2824                for_each_node_state(nid, N_HIGH_MEMORY) {
2825                        pg_data_t *pgdat = NODE_DATA(nid);
2826                        const struct cpumask *mask;
2827
2828                        mask = cpumask_of_node(pgdat->node_id);
2829
2830                        if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2831                                /* One of our CPUs online: restore mask */
2832                                set_cpus_allowed_ptr(pgdat->kswapd, mask);
2833                }
2834        }
2835        return NOTIFY_OK;
2836}
2837
2838/*
2839 * This kswapd start function will be called by init and node-hot-add.
2840 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2841 */
2842int kswapd_run(int nid)
2843{
2844        pg_data_t *pgdat = NODE_DATA(nid);
2845        int ret = 0;
2846
2847        if (pgdat->kswapd)
2848                return 0;
2849
2850        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2851        if (IS_ERR(pgdat->kswapd)) {
2852                /* failure at boot is fatal */
2853                BUG_ON(system_state == SYSTEM_BOOTING);
2854                printk("Failed to start kswapd on node %d\n",nid);
2855                ret = -1;
2856        }
2857        return ret;
2858}
2859
2860/*
2861 * Called by memory hotplug when all memory in a node is offlined.
2862 */
2863void kswapd_stop(int nid)
2864{
2865        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2866
2867        if (kswapd)
2868                kthread_stop(kswapd);
2869}
2870
2871static int __init kswapd_init(void)
2872{
2873        int nid;
2874
2875        swap_setup();
2876        for_each_node_state(nid, N_HIGH_MEMORY)
2877                kswapd_run(nid);
2878        hotcpu_notifier(cpu_callback, 0);
2879        return 0;
2880}
2881
2882module_init(kswapd_init)
2883
2884#ifdef CONFIG_NUMA
2885/*
2886 * Zone reclaim mode
2887 *
2888 * If non-zero call zone_reclaim when the number of free pages falls below
2889 * the watermarks.
2890 */
2891int zone_reclaim_mode __read_mostly;
2892
2893#define RECLAIM_OFF 0
2894#define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2895#define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2896#define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2897
2898/*
2899 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2900 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2901 * a zone.
2902 */
2903#define ZONE_RECLAIM_PRIORITY 4
2904
2905/*
2906 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2907 * occur.
2908 */
2909int sysctl_min_unmapped_ratio = 1;
2910
2911/*
2912 * If the number of slab pages in a zone grows beyond this percentage then
2913 * slab reclaim needs to occur.
2914 */
2915int sysctl_min_slab_ratio = 5;
2916
2917static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2918{
2919        unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2920        unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2921                zone_page_state(zone, NR_ACTIVE_FILE);
2922
2923        /*
2924         * It's possible for there to be more file mapped pages than
2925         * accounted for by the pages on the file LRU lists because
2926         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2927         */
2928        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2929}
2930
2931/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2932static long zone_pagecache_reclaimable(struct zone *zone)
2933{
2934        long nr_pagecache_reclaimable;
2935        long delta = 0;
2936
2937        /*
2938         * If RECLAIM_SWAP is set, then all file pages are considered
2939         * potentially reclaimable. Otherwise, we have to worry about
2940         * pages like swapcache and zone_unmapped_file_pages() provides
2941         * a better estimate
2942         */
2943        if (zone_reclaim_mode & RECLAIM_SWAP)
2944                nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2945        else
2946                nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2947
2948        /* If we can't clean pages, remove dirty pages from consideration */
2949        if (!(zone_reclaim_mode & RECLAIM_WRITE))
2950                delta += zone_page_state(zone, NR_FILE_DIRTY);
2951
2952        /* Watch for any possible underflows due to delta */
2953        if (unlikely(delta > nr_pagecache_reclaimable))
2954                delta = nr_pagecache_reclaimable;
2955
2956        return nr_pagecache_reclaimable - delta;
2957}
2958
2959/*
2960 * Try to free up some pages from this zone through reclaim.
2961 */
2962static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2963{
2964        /* Minimum pages needed in order to stay on node */
2965        const unsigned long nr_pages = 1 << order;
2966        struct task_struct *p = current;
2967        struct reclaim_state reclaim_state;
2968        int priority;
2969        struct scan_control sc = {
2970                .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2971                .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2972                .may_swap = 1,
2973                .nr_to_reclaim = max_t(unsigned long, nr_pages,
2974                                       SWAP_CLUSTER_MAX),
2975                .gfp_mask = gfp_mask,
2976                .swappiness = vm_swappiness,
2977                .order = order,
2978        };
2979        unsigned long nr_slab_pages0, nr_slab_pages1;
2980
2981        cond_resched();
2982        /*
2983         * We need to be able to allocate from the reserves for RECLAIM_SWAP
2984         * and we also need to be able to write out pages for RECLAIM_WRITE
2985         * and RECLAIM_SWAP.
2986         */
2987        p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2988        lockdep_set_current_reclaim_state(gfp_mask);
2989        reclaim_state.reclaimed_slab = 0;
2990        p->reclaim_state = &reclaim_state;
2991
2992        if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2993                /*
2994                 * Free memory by calling shrink zone with increasing
2995                 * priorities until we have enough memory freed.
2996                 */
2997                priority = ZONE_RECLAIM_PRIORITY;
2998                do {
2999                        shrink_zone(priority, zone, &sc);
3000                        priority--;
3001                } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3002        }
3003
3004        nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3005        if (nr_slab_pages0 > zone->min_slab_pages) {
3006                /*
3007                 * shrink_slab() does not currently allow us to determine how
3008                 * many pages were freed in this zone. So we take the current
3009                 * number of slab pages and shake the slab until it is reduced
3010                 * by the same nr_pages that we used for reclaiming unmapped
3011                 * pages.
3012                 *
3013                 * Note that shrink_slab will free memory on all zones and may
3014                 * take a long time.
3015                 */
3016                for (;;) {
3017                        unsigned long lru_pages = zone_reclaimable_pages(zone);
3018
3019                        /* No reclaimable slab or very low memory pressure */
3020                        if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3021                                break;
3022
3023                        /* Freed enough memory */
3024                        nr_slab_pages1 = zone_page_state(zone,
3025                                                        NR_SLAB_RECLAIMABLE);
3026                        if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3027                                break;
3028                }
3029
3030                /*
3031                 * Update nr_reclaimed by the number of slab pages we
3032                 * reclaimed from this zone.
3033                 */
3034                nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3035                if (nr_slab_pages1 < nr_slab_pages0)
3036                        sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3037        }
3038
3039        p->reclaim_state = NULL;
3040        current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3041        lockdep_clear_current_reclaim_state();
3042        return sc.nr_reclaimed >= nr_pages;
3043}
3044
3045int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3046{
3047        int node_id;
3048        int ret;
3049
3050        /*
3051         * Zone reclaim reclaims unmapped file backed pages and
3052         * slab pages if we are over the defined limits.
3053         *
3054         * A small portion of unmapped file backed pages is needed for
3055         * file I/O otherwise pages read by file I/O will be immediately
3056         * thrown out if the zone is overallocated. So we do not reclaim
3057         * if less than a specified percentage of the zone is used by
3058         * unmapped file backed pages.
3059         */
3060        if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3061            zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3062                return ZONE_RECLAIM_FULL;
3063
3064        if (zone->all_unreclaimable)
3065                return ZONE_RECLAIM_FULL;
3066
3067        /*
3068         * Do not scan if the allocation should not be delayed.
3069         */
3070        if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3071                return ZONE_RECLAIM_NOSCAN;
3072
3073        /*
3074         * Only run zone reclaim on the local zone or on zones that do not
3075         * have associated processors. This will favor the local processor
3076         * over remote processors and spread off node memory allocations
3077         * as wide as possible.
3078         */
3079        node_id = zone_to_nid(zone);
3080        if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3081                return ZONE_RECLAIM_NOSCAN;
3082
3083        if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3084                return ZONE_RECLAIM_NOSCAN;
3085
3086        ret = __zone_reclaim(zone, gfp_mask, order);
3087        zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3088
3089        if (!ret)
3090                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3091
3092        return ret;
3093}
3094#endif
3095
3096/*
3097 * page_evictable - test whether a page is evictable
3098 * @page: the page to test
3099 * @vma: the VMA in which the page is or will be mapped, may be NULL
3100 *
3101 * Test whether page is evictable--i.e., should be placed on active/inactive
3102 * lists vs unevictable list.  The vma argument is !NULL when called from the
3103 * fault path to determine how to instantate a new page.
3104 *
3105 * Reasons page might not be evictable:
3106 * (1) page's mapping marked unevictable
3107 * (2) page is part of an mlocked VMA
3108 *
3109 */
3110int page_evictable(struct page *page, struct vm_area_struct *vma)
3111{
3112
3113        if (mapping_unevictable(page_mapping(page)))
3114                return 0;
3115
3116        if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3117                return 0;
3118
3119        return 1;
3120}
3121
3122/**
3123 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3124 * @page: page to check evictability and move to appropriate lru list
3125 * @zone: zone page is in
3126 *
3127 * Checks a page for evictability and moves the page to the appropriate
3128 * zone lru list.
3129 *
3130 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3131 * have PageUnevictable set.
3132 */
3133static void check_move_unevictable_page(struct page *page, struct zone *zone)
3134{
3135        VM_BUG_ON(PageActive(page));
3136
3137retry:
3138        ClearPageUnevictable(page);
3139        if (page_evictable(page, NULL)) {
3140                enum lru_list l = page_lru_base_type(page);
3141
3142                __dec_zone_state(zone, NR_UNEVICTABLE);
3143                list_move(&page->lru, &zone->lru[l].list);
3144                mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3145                __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3146                __count_vm_event(UNEVICTABLE_PGRESCUED);
3147        } else {
3148                /*
3149                 * rotate unevictable list
3150                 */
3151                SetPageUnevictable(page);
3152                list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3153                mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3154                if (page_evictable(page, NULL))
3155                        goto retry;
3156        }
3157}
3158
3159/**
3160 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3161 * @mapping: struct address_space to scan for evictable pages
3162 *
3163 * Scan all pages in mapping.  Check unevictable pages for
3164 * evictability and move them to the appropriate zone lru list.
3165 */
3166void scan_mapping_unevictable_pages(struct address_space *mapping)
3167{
3168        pgoff_t next = 0;
3169        pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3170                         PAGE_CACHE_SHIFT;
3171        struct zone *zone;
3172        struct pagevec pvec;
3173
3174        if (mapping->nrpages == 0)
3175                return;
3176
3177        pagevec_init(&pvec, 0);
3178        while (next < end &&
3179                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3180                int i;
3181                int pg_scanned = 0;
3182
3183                zone = NULL;
3184
3185                for (i = 0; i < pagevec_count(&pvec); i++) {
3186                        struct page *page = pvec.pages[i];
3187                        pgoff_t page_index = page->index;
3188                        struct zone *pagezone = page_zone(page);
3189
3190                        pg_scanned++;
3191                        if (page_index > next)
3192                                next = page_index;
3193                        next++;
3194
3195                        if (pagezone != zone) {
3196                                if (zone)
3197                                        spin_unlock_irq(&zone->lru_lock);
3198                                zone = pagezone;
3199                                spin_lock_irq(&zone->lru_lock);
3200                        }
3201
3202                        if (PageLRU(page) && PageUnevictable(page))
3203                                check_move_unevictable_page(page, zone);
3204                }
3205                if (zone)
3206                        spin_unlock_irq(&zone->lru_lock);
3207                pagevec_release(&pvec);
3208
3209                count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3210        }
3211
3212}
3213
3214/**
3215 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3216 * @zone - zone of which to scan the unevictable list
3217 *
3218 * Scan @zone's unevictable LRU lists to check for pages that have become
3219 * evictable.  Move those that have to @zone's inactive list where they
3220 * become candidates for reclaim, unless shrink_inactive_zone() decides
3221 * to reactivate them.  Pages that are still unevictable are rotated
3222 * back onto @zone's unevictable list.
3223 */
3224#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3225static void scan_zone_unevictable_pages(struct zone *zone)
3226{
3227        struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3228        unsigned long scan;
3229        unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3230
3231        while (nr_to_scan > 0) {
3232                unsigned long batch_size = min(nr_to_scan,
3233                                                SCAN_UNEVICTABLE_BATCH_SIZE);
3234
3235                spin_lock_irq(&zone->lru_lock);
3236                for (scan = 0;  scan < batch_size; scan++) {
3237                        struct page *page = lru_to_page(l_unevictable);
3238
3239                        if (!trylock_page(page))
3240                                continue;
3241
3242                        prefetchw_prev_lru_page(page, l_unevictable, flags);
3243
3244                        if (likely(PageLRU(page) && PageUnevictable(page)))
3245                                check_move_unevictable_page(page, zone);
3246
3247                        unlock_page(page);
3248                }
3249                spin_unlock_irq(&zone->lru_lock);
3250
3251                nr_to_scan -= batch_size;
3252        }
3253}
3254
3255
3256/**
3257 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3258 *
3259 * A really big hammer:  scan all zones' unevictable LRU lists to check for
3260 * pages that have become evictable.  Move those back to the zones'
3261 * inactive list where they become candidates for reclaim.
3262 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3263 * and we add swap to the system.  As such, it runs in the context of a task
3264 * that has possibly/probably made some previously unevictable pages
3265 * evictable.
3266 */
3267static void scan_all_zones_unevictable_pages(void)
3268{
3269        struct zone *zone;
3270
3271        for_each_zone(zone) {
3272                scan_zone_unevictable_pages(zone);
3273        }
3274}
3275
3276/*
3277 * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3278 * all nodes' unevictable lists for evictable pages
3279 */
3280unsigned long scan_unevictable_pages;
3281
3282int scan_unevictable_handler(struct ctl_table *table, int write,
3283                           void __user *buffer,
3284                           size_t *length, loff_t *ppos)
3285{
3286        proc_doulongvec_minmax(table, write, buffer, length, ppos);
3287
3288        if (write && *(unsigned long *)table->data)
3289                scan_all_zones_unevictable_pages();
3290
3291        scan_unevictable_pages = 0;
3292        return 0;
3293}
3294
3295#ifdef CONFIG_NUMA
3296/*
3297 * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3298 * a specified node's per zone unevictable lists for evictable pages.
3299 */
3300
3301static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3302                                          struct sysdev_attribute *attr,
3303                                          char *buf)
3304{
3305        return sprintf(buf, "0\n");     /* always zero; should fit... */
3306}
3307
3308static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3309                                           struct sysdev_attribute *attr,
3310                                        const char *buf, size_t count)
3311{
3312        struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3313        struct zone *zone;
3314        unsigned long res;
3315        unsigned long req = strict_strtoul(buf, 10, &res);
3316
3317        if (!req)
3318                return 1;       /* zero is no-op */
3319
3320        for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3321                if (!populated_zone(zone))
3322                        continue;
3323                scan_zone_unevictable_pages(zone);
3324        }
3325        return 1;
3326}
3327
3328
3329static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3330                        read_scan_unevictable_node,
3331                        write_scan_unevictable_node);
3332
3333int scan_unevictable_register_node(struct node *node)
3334{
3335        return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3336}
3337
3338void scan_unevictable_unregister_node(struct node *node)
3339{
3340        sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3341}
3342#endif
3343