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/vmpressure.h>
  23#include <linux/vmstat.h>
  24#include <linux/file.h>
  25#include <linux/writeback.h>
  26#include <linux/blkdev.h>
  27#include <linux/buffer_head.h>  /* for try_to_release_page(),
  28                                        buffer_heads_over_limit */
  29#include <linux/mm_inline.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#include <linux/prefetch.h>
  46
  47#include <asm/tlbflush.h>
  48#include <asm/div64.h>
  49
  50#include <linux/swapops.h>
  51#include <linux/balloon_compaction.h>
  52
  53#include "internal.h"
  54
  55#define CREATE_TRACE_POINTS
  56#include <trace/events/vmscan.h>
  57
  58struct scan_control {
  59        /* Incremented by the number of inactive pages that were scanned */
  60        unsigned long nr_scanned;
  61
  62        /* Number of pages freed so far during a call to shrink_zones() */
  63        unsigned long nr_reclaimed;
  64
  65        /* How many pages shrink_list() should reclaim */
  66        unsigned long nr_to_reclaim;
  67
  68        unsigned long hibernation_mode;
  69
  70        /* This context's GFP mask */
  71        gfp_t gfp_mask;
  72
  73        int may_writepage;
  74
  75        /* Can mapped pages be reclaimed? */
  76        int may_unmap;
  77
  78        /* Can pages be swapped as part of reclaim? */
  79        int may_swap;
  80
  81        int order;
  82
  83        /* Scan (total_size >> priority) pages at once */
  84        int priority;
  85
  86        /*
  87         * The memory cgroup that hit its limit and as a result is the
  88         * primary target of this reclaim invocation.
  89         */
  90        struct mem_cgroup *target_mem_cgroup;
  91
  92        /*
  93         * Nodemask of nodes allowed by the caller. If NULL, all nodes
  94         * are scanned.
  95         */
  96        nodemask_t      *nodemask;
  97};
  98
  99#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
 100
 101#ifdef ARCH_HAS_PREFETCH
 102#define prefetch_prev_lru_page(_page, _base, _field)                    \
 103        do {                                                            \
 104                if ((_page)->lru.prev != _base) {                       \
 105                        struct page *prev;                              \
 106                                                                        \
 107                        prev = lru_to_page(&(_page->lru));              \
 108                        prefetch(&prev->_field);                        \
 109                }                                                       \
 110        } while (0)
 111#else
 112#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 113#endif
 114
 115#ifdef ARCH_HAS_PREFETCHW
 116#define prefetchw_prev_lru_page(_page, _base, _field)                   \
 117        do {                                                            \
 118                if ((_page)->lru.prev != _base) {                       \
 119                        struct page *prev;                              \
 120                                                                        \
 121                        prev = lru_to_page(&(_page->lru));              \
 122                        prefetchw(&prev->_field);                       \
 123                }                                                       \
 124        } while (0)
 125#else
 126#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 127#endif
 128
 129/*
 130 * From 0 .. 100.  Higher means more swappy.
 131 */
 132int vm_swappiness = 60;
 133unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
 134
 135static LIST_HEAD(shrinker_list);
 136static DECLARE_RWSEM(shrinker_rwsem);
 137
 138#ifdef CONFIG_MEMCG
 139static bool global_reclaim(struct scan_control *sc)
 140{
 141        return !sc->target_mem_cgroup;
 142}
 143#else
 144static bool global_reclaim(struct scan_control *sc)
 145{
 146        return true;
 147}
 148#endif
 149
 150unsigned long zone_reclaimable_pages(struct zone *zone)
 151{
 152        int nr;
 153
 154        nr = zone_page_state(zone, NR_ACTIVE_FILE) +
 155             zone_page_state(zone, NR_INACTIVE_FILE);
 156
 157        if (get_nr_swap_pages() > 0)
 158                nr += zone_page_state(zone, NR_ACTIVE_ANON) +
 159                      zone_page_state(zone, NR_INACTIVE_ANON);
 160
 161        return nr;
 162}
 163
 164bool zone_reclaimable(struct zone *zone)
 165{
 166        return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
 167}
 168
 169static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 170{
 171        if (!mem_cgroup_disabled())
 172                return mem_cgroup_get_lru_size(lruvec, lru);
 173
 174        return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
 175}
 176
 177/*
 178 * Add a shrinker callback to be called from the vm.
 179 */
 180int register_shrinker(struct shrinker *shrinker)
 181{
 182        size_t size = sizeof(*shrinker->nr_deferred);
 183
 184        /*
 185         * If we only have one possible node in the system anyway, save
 186         * ourselves the trouble and disable NUMA aware behavior. This way we
 187         * will save memory and some small loop time later.
 188         */
 189        if (nr_node_ids == 1)
 190                shrinker->flags &= ~SHRINKER_NUMA_AWARE;
 191
 192        if (shrinker->flags & SHRINKER_NUMA_AWARE)
 193                size *= nr_node_ids;
 194
 195        shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
 196        if (!shrinker->nr_deferred)
 197                return -ENOMEM;
 198
 199        down_write(&shrinker_rwsem);
 200        list_add_tail(&shrinker->list, &shrinker_list);
 201        up_write(&shrinker_rwsem);
 202        return 0;
 203}
 204EXPORT_SYMBOL(register_shrinker);
 205
 206/*
 207 * Remove one
 208 */
 209void unregister_shrinker(struct shrinker *shrinker)
 210{
 211        down_write(&shrinker_rwsem);
 212        list_del(&shrinker->list);
 213        up_write(&shrinker_rwsem);
 214        kfree(shrinker->nr_deferred);
 215}
 216EXPORT_SYMBOL(unregister_shrinker);
 217
 218#define SHRINK_BATCH 128
 219
 220static unsigned long
 221shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
 222                 unsigned long nr_pages_scanned, unsigned long lru_pages)
 223{
 224        unsigned long freed = 0;
 225        unsigned long long delta;
 226        long total_scan;
 227        long max_pass;
 228        long nr;
 229        long new_nr;
 230        int nid = shrinkctl->nid;
 231        long batch_size = shrinker->batch ? shrinker->batch
 232                                          : SHRINK_BATCH;
 233
 234        max_pass = shrinker->count_objects(shrinker, shrinkctl);
 235        if (max_pass == 0)
 236                return 0;
 237
 238        /*
 239         * copy the current shrinker scan count into a local variable
 240         * and zero it so that other concurrent shrinker invocations
 241         * don't also do this scanning work.
 242         */
 243        nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
 244
 245        total_scan = nr;
 246        delta = (4 * nr_pages_scanned) / shrinker->seeks;
 247        delta *= max_pass;
 248        do_div(delta, lru_pages + 1);
 249        total_scan += delta;
 250        if (total_scan < 0) {
 251                printk(KERN_ERR
 252                "shrink_slab: %pF negative objects to delete nr=%ld\n",
 253                       shrinker->scan_objects, total_scan);
 254                total_scan = max_pass;
 255        }
 256
 257        /*
 258         * We need to avoid excessive windup on filesystem shrinkers
 259         * due to large numbers of GFP_NOFS allocations causing the
 260         * shrinkers to return -1 all the time. This results in a large
 261         * nr being built up so when a shrink that can do some work
 262         * comes along it empties the entire cache due to nr >>>
 263         * max_pass.  This is bad for sustaining a working set in
 264         * memory.
 265         *
 266         * Hence only allow the shrinker to scan the entire cache when
 267         * a large delta change is calculated directly.
 268         */
 269        if (delta < max_pass / 4)
 270                total_scan = min(total_scan, max_pass / 2);
 271
 272        /*
 273         * Avoid risking looping forever due to too large nr value:
 274         * never try to free more than twice the estimate number of
 275         * freeable entries.
 276         */
 277        if (total_scan > max_pass * 2)
 278                total_scan = max_pass * 2;
 279
 280        trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
 281                                nr_pages_scanned, lru_pages,
 282                                max_pass, delta, total_scan);
 283
 284        while (total_scan >= batch_size) {
 285                unsigned long ret;
 286
 287                shrinkctl->nr_to_scan = batch_size;
 288                ret = shrinker->scan_objects(shrinker, shrinkctl);
 289                if (ret == SHRINK_STOP)
 290                        break;
 291                freed += ret;
 292
 293                count_vm_events(SLABS_SCANNED, batch_size);
 294                total_scan -= batch_size;
 295
 296                cond_resched();
 297        }
 298
 299        /*
 300         * move the unused scan count back into the shrinker in a
 301         * manner that handles concurrent updates. If we exhausted the
 302         * scan, there is no need to do an update.
 303         */
 304        if (total_scan > 0)
 305                new_nr = atomic_long_add_return(total_scan,
 306                                                &shrinker->nr_deferred[nid]);
 307        else
 308                new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
 309
 310        trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
 311        return freed;
 312}
 313
 314/*
 315 * Call the shrink functions to age shrinkable caches
 316 *
 317 * Here we assume it costs one seek to replace a lru page and that it also
 318 * takes a seek to recreate a cache object.  With this in mind we age equal
 319 * percentages of the lru and ageable caches.  This should balance the seeks
 320 * generated by these structures.
 321 *
 322 * If the vm encountered mapped pages on the LRU it increase the pressure on
 323 * slab to avoid swapping.
 324 *
 325 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 326 *
 327 * `lru_pages' represents the number of on-LRU pages in all the zones which
 328 * are eligible for the caller's allocation attempt.  It is used for balancing
 329 * slab reclaim versus page reclaim.
 330 *
 331 * Returns the number of slab objects which we shrunk.
 332 */
 333unsigned long shrink_slab(struct shrink_control *shrinkctl,
 334                          unsigned long nr_pages_scanned,
 335                          unsigned long lru_pages)
 336{
 337        struct shrinker *shrinker;
 338        unsigned long freed = 0;
 339
 340        if (nr_pages_scanned == 0)
 341                nr_pages_scanned = SWAP_CLUSTER_MAX;
 342
 343        if (!down_read_trylock(&shrinker_rwsem)) {
 344                /*
 345                 * If we would return 0, our callers would understand that we
 346                 * have nothing else to shrink and give up trying. By returning
 347                 * 1 we keep it going and assume we'll be able to shrink next
 348                 * time.
 349                 */
 350                freed = 1;
 351                goto out;
 352        }
 353
 354        list_for_each_entry(shrinker, &shrinker_list, list) {
 355                for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
 356                        if (!node_online(shrinkctl->nid))
 357                                continue;
 358
 359                        if (!(shrinker->flags & SHRINKER_NUMA_AWARE) &&
 360                            (shrinkctl->nid != 0))
 361                                break;
 362
 363                        freed += shrink_slab_node(shrinkctl, shrinker,
 364                                 nr_pages_scanned, lru_pages);
 365
 366                }
 367        }
 368        up_read(&shrinker_rwsem);
 369out:
 370        cond_resched();
 371        return freed;
 372}
 373
 374static inline int is_page_cache_freeable(struct page *page)
 375{
 376        /*
 377         * A freeable page cache page is referenced only by the caller
 378         * that isolated the page, the page cache radix tree and
 379         * optional buffer heads at page->private.
 380         */
 381        return page_count(page) - page_has_private(page) == 2;
 382}
 383
 384static int may_write_to_queue(struct backing_dev_info *bdi,
 385                              struct scan_control *sc)
 386{
 387        if (current->flags & PF_SWAPWRITE)
 388                return 1;
 389        if (!bdi_write_congested(bdi))
 390                return 1;
 391        if (bdi == current->backing_dev_info)
 392                return 1;
 393        return 0;
 394}
 395
 396/*
 397 * We detected a synchronous write error writing a page out.  Probably
 398 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 399 * fsync(), msync() or close().
 400 *
 401 * The tricky part is that after writepage we cannot touch the mapping: nothing
 402 * prevents it from being freed up.  But we have a ref on the page and once
 403 * that page is locked, the mapping is pinned.
 404 *
 405 * We're allowed to run sleeping lock_page() here because we know the caller has
 406 * __GFP_FS.
 407 */
 408static void handle_write_error(struct address_space *mapping,
 409                                struct page *page, int error)
 410{
 411        lock_page(page);
 412        if (page_mapping(page) == mapping)
 413                mapping_set_error(mapping, error);
 414        unlock_page(page);
 415}
 416
 417/* possible outcome of pageout() */
 418typedef enum {
 419        /* failed to write page out, page is locked */
 420        PAGE_KEEP,
 421        /* move page to the active list, page is locked */
 422        PAGE_ACTIVATE,
 423        /* page has been sent to the disk successfully, page is unlocked */
 424        PAGE_SUCCESS,
 425        /* page is clean and locked */
 426        PAGE_CLEAN,
 427} pageout_t;
 428
 429/*
 430 * pageout is called by shrink_page_list() for each dirty page.
 431 * Calls ->writepage().
 432 */
 433static pageout_t pageout(struct page *page, struct address_space *mapping,
 434                         struct scan_control *sc)
 435{
 436        /*
 437         * If the page is dirty, only perform writeback if that write
 438         * will be non-blocking.  To prevent this allocation from being
 439         * stalled by pagecache activity.  But note that there may be
 440         * stalls if we need to run get_block().  We could test
 441         * PagePrivate for that.
 442         *
 443         * If this process is currently in __generic_file_aio_write() against
 444         * this page's queue, we can perform writeback even if that
 445         * will block.
 446         *
 447         * If the page is swapcache, write it back even if that would
 448         * block, for some throttling. This happens by accident, because
 449         * swap_backing_dev_info is bust: it doesn't reflect the
 450         * congestion state of the swapdevs.  Easy to fix, if needed.
 451         */
 452        if (!is_page_cache_freeable(page))
 453                return PAGE_KEEP;
 454        if (!mapping) {
 455                /*
 456                 * Some data journaling orphaned pages can have
 457                 * page->mapping == NULL while being dirty with clean buffers.
 458                 */
 459                if (page_has_private(page)) {
 460                        if (try_to_free_buffers(page)) {
 461                                ClearPageDirty(page);
 462                                printk("%s: orphaned page\n", __func__);
 463                                return PAGE_CLEAN;
 464                        }
 465                }
 466                return PAGE_KEEP;
 467        }
 468        if (mapping->a_ops->writepage == NULL)
 469                return PAGE_ACTIVATE;
 470        if (!may_write_to_queue(mapping->backing_dev_info, sc))
 471                return PAGE_KEEP;
 472
 473        if (clear_page_dirty_for_io(page)) {
 474                int res;
 475                struct writeback_control wbc = {
 476                        .sync_mode = WB_SYNC_NONE,
 477                        .nr_to_write = SWAP_CLUSTER_MAX,
 478                        .range_start = 0,
 479                        .range_end = LLONG_MAX,
 480                        .for_reclaim = 1,
 481                };
 482
 483                SetPageReclaim(page);
 484                res = mapping->a_ops->writepage(page, &wbc);
 485                if (res < 0)
 486                        handle_write_error(mapping, page, res);
 487                if (res == AOP_WRITEPAGE_ACTIVATE) {
 488                        ClearPageReclaim(page);
 489                        return PAGE_ACTIVATE;
 490                }
 491
 492                if (!PageWriteback(page)) {
 493                        /* synchronous write or broken a_ops? */
 494                        ClearPageReclaim(page);
 495                }
 496                trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
 497                inc_zone_page_state(page, NR_VMSCAN_WRITE);
 498                return PAGE_SUCCESS;
 499        }
 500
 501        return PAGE_CLEAN;
 502}
 503
 504/*
 505 * Same as remove_mapping, but if the page is removed from the mapping, it
 506 * gets returned with a refcount of 0.
 507 */
 508static int __remove_mapping(struct address_space *mapping, struct page *page)
 509{
 510        BUG_ON(!PageLocked(page));
 511        BUG_ON(mapping != page_mapping(page));
 512
 513        spin_lock_irq(&mapping->tree_lock);
 514        /*
 515         * The non racy check for a busy page.
 516         *
 517         * Must be careful with the order of the tests. When someone has
 518         * a ref to the page, it may be possible that they dirty it then
 519         * drop the reference. So if PageDirty is tested before page_count
 520         * here, then the following race may occur:
 521         *
 522         * get_user_pages(&page);
 523         * [user mapping goes away]
 524         * write_to(page);
 525         *                              !PageDirty(page)    [good]
 526         * SetPageDirty(page);
 527         * put_page(page);
 528         *                              !page_count(page)   [good, discard it]
 529         *
 530         * [oops, our write_to data is lost]
 531         *
 532         * Reversing the order of the tests ensures such a situation cannot
 533         * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 534         * load is not satisfied before that of page->_count.
 535         *
 536         * Note that if SetPageDirty is always performed via set_page_dirty,
 537         * and thus under tree_lock, then this ordering is not required.
 538         */
 539        if (!page_freeze_refs(page, 2))
 540                goto cannot_free;
 541        /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 542        if (unlikely(PageDirty(page))) {
 543                page_unfreeze_refs(page, 2);
 544                goto cannot_free;
 545        }
 546
 547        if (PageSwapCache(page)) {
 548                swp_entry_t swap = { .val = page_private(page) };
 549                __delete_from_swap_cache(page);
 550                spin_unlock_irq(&mapping->tree_lock);
 551                swapcache_free(swap, page);
 552        } else {
 553                void (*freepage)(struct page *);
 554
 555                freepage = mapping->a_ops->freepage;
 556
 557                __delete_from_page_cache(page);
 558                spin_unlock_irq(&mapping->tree_lock);
 559                mem_cgroup_uncharge_cache_page(page);
 560
 561                if (freepage != NULL)
 562                        freepage(page);
 563        }
 564
 565        return 1;
 566
 567cannot_free:
 568        spin_unlock_irq(&mapping->tree_lock);
 569        return 0;
 570}
 571
 572/*
 573 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 574 * someone else has a ref on the page, abort and return 0.  If it was
 575 * successfully detached, return 1.  Assumes the caller has a single ref on
 576 * this page.
 577 */
 578int remove_mapping(struct address_space *mapping, struct page *page)
 579{
 580        if (__remove_mapping(mapping, page)) {
 581                /*
 582                 * Unfreezing the refcount with 1 rather than 2 effectively
 583                 * drops the pagecache ref for us without requiring another
 584                 * atomic operation.
 585                 */
 586                page_unfreeze_refs(page, 1);
 587                return 1;
 588        }
 589        return 0;
 590}
 591
 592/**
 593 * putback_lru_page - put previously isolated page onto appropriate LRU list
 594 * @page: page to be put back to appropriate lru list
 595 *
 596 * Add previously isolated @page to appropriate LRU list.
 597 * Page may still be unevictable for other reasons.
 598 *
 599 * lru_lock must not be held, interrupts must be enabled.
 600 */
 601void putback_lru_page(struct page *page)
 602{
 603        bool is_unevictable;
 604        int was_unevictable = PageUnevictable(page);
 605
 606        VM_BUG_ON(PageLRU(page));
 607
 608redo:
 609        ClearPageUnevictable(page);
 610
 611        if (page_evictable(page)) {
 612                /*
 613                 * For evictable pages, we can use the cache.
 614                 * In event of a race, worst case is we end up with an
 615                 * unevictable page on [in]active list.
 616                 * We know how to handle that.
 617                 */
 618                is_unevictable = false;
 619                lru_cache_add(page);
 620        } else {
 621                /*
 622                 * Put unevictable pages directly on zone's unevictable
 623                 * list.
 624                 */
 625                is_unevictable = true;
 626                add_page_to_unevictable_list(page);
 627                /*
 628                 * When racing with an mlock or AS_UNEVICTABLE clearing
 629                 * (page is unlocked) make sure that if the other thread
 630                 * does not observe our setting of PG_lru and fails
 631                 * isolation/check_move_unevictable_pages,
 632                 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
 633                 * the page back to the evictable list.
 634                 *
 635                 * The other side is TestClearPageMlocked() or shmem_lock().
 636                 */
 637                smp_mb();
 638        }
 639
 640        /*
 641         * page's status can change while we move it among lru. If an evictable
 642         * page is on unevictable list, it never be freed. To avoid that,
 643         * check after we added it to the list, again.
 644         */
 645        if (is_unevictable && page_evictable(page)) {
 646                if (!isolate_lru_page(page)) {
 647                        put_page(page);
 648                        goto redo;
 649                }
 650                /* This means someone else dropped this page from LRU
 651                 * So, it will be freed or putback to LRU again. There is
 652                 * nothing to do here.
 653                 */
 654        }
 655
 656        if (was_unevictable && !is_unevictable)
 657                count_vm_event(UNEVICTABLE_PGRESCUED);
 658        else if (!was_unevictable && is_unevictable)
 659                count_vm_event(UNEVICTABLE_PGCULLED);
 660
 661        put_page(page);         /* drop ref from isolate */
 662}
 663
 664enum page_references {
 665        PAGEREF_RECLAIM,
 666        PAGEREF_RECLAIM_CLEAN,
 667        PAGEREF_KEEP,
 668        PAGEREF_ACTIVATE,
 669};
 670
 671static enum page_references page_check_references(struct page *page,
 672                                                  struct scan_control *sc)
 673{
 674        int referenced_ptes, referenced_page;
 675        unsigned long vm_flags;
 676
 677        referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
 678                                          &vm_flags);
 679        referenced_page = TestClearPageReferenced(page);
 680
 681        /*
 682         * Mlock lost the isolation race with us.  Let try_to_unmap()
 683         * move the page to the unevictable list.
 684         */
 685        if (vm_flags & VM_LOCKED)
 686                return PAGEREF_RECLAIM;
 687
 688        if (referenced_ptes) {
 689                if (PageSwapBacked(page))
 690                        return PAGEREF_ACTIVATE;
 691                /*
 692                 * All mapped pages start out with page table
 693                 * references from the instantiating fault, so we need
 694                 * to look twice if a mapped file page is used more
 695                 * than once.
 696                 *
 697                 * Mark it and spare it for another trip around the
 698                 * inactive list.  Another page table reference will
 699                 * lead to its activation.
 700                 *
 701                 * Note: the mark is set for activated pages as well
 702                 * so that recently deactivated but used pages are
 703                 * quickly recovered.
 704                 */
 705                SetPageReferenced(page);
 706
 707                if (referenced_page || referenced_ptes > 1)
 708                        return PAGEREF_ACTIVATE;
 709
 710                /*
 711                 * Activate file-backed executable pages after first usage.
 712                 */
 713                if (vm_flags & VM_EXEC)
 714                        return PAGEREF_ACTIVATE;
 715
 716                return PAGEREF_KEEP;
 717        }
 718
 719        /* Reclaim if clean, defer dirty pages to writeback */
 720        if (referenced_page && !PageSwapBacked(page))
 721                return PAGEREF_RECLAIM_CLEAN;
 722
 723        return PAGEREF_RECLAIM;
 724}
 725
 726/* Check if a page is dirty or under writeback */
 727static void page_check_dirty_writeback(struct page *page,
 728                                       bool *dirty, bool *writeback)
 729{
 730        struct address_space *mapping;
 731
 732        /*
 733         * Anonymous pages are not handled by flushers and must be written
 734         * from reclaim context. Do not stall reclaim based on them
 735         */
 736        if (!page_is_file_cache(page)) {
 737                *dirty = false;
 738                *writeback = false;
 739                return;
 740        }
 741
 742        /* By default assume that the page flags are accurate */
 743        *dirty = PageDirty(page);
 744        *writeback = PageWriteback(page);
 745
 746        /* Verify dirty/writeback state if the filesystem supports it */
 747        if (!page_has_private(page))
 748                return;
 749
 750        mapping = page_mapping(page);
 751        if (mapping && mapping->a_ops->is_dirty_writeback)
 752                mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
 753}
 754
 755/*
 756 * shrink_page_list() returns the number of reclaimed pages
 757 */
 758static unsigned long shrink_page_list(struct list_head *page_list,
 759                                      struct zone *zone,
 760                                      struct scan_control *sc,
 761                                      enum ttu_flags ttu_flags,
 762                                      unsigned long *ret_nr_dirty,
 763                                      unsigned long *ret_nr_unqueued_dirty,
 764                                      unsigned long *ret_nr_congested,
 765                                      unsigned long *ret_nr_writeback,
 766                                      unsigned long *ret_nr_immediate,
 767                                      bool force_reclaim)
 768{
 769        LIST_HEAD(ret_pages);
 770        LIST_HEAD(free_pages);
 771        int pgactivate = 0;
 772        unsigned long nr_unqueued_dirty = 0;
 773        unsigned long nr_dirty = 0;
 774        unsigned long nr_congested = 0;
 775        unsigned long nr_reclaimed = 0;
 776        unsigned long nr_writeback = 0;
 777        unsigned long nr_immediate = 0;
 778
 779        cond_resched();
 780
 781        mem_cgroup_uncharge_start();
 782        while (!list_empty(page_list)) {
 783                struct address_space *mapping;
 784                struct page *page;
 785                int may_enter_fs;
 786                enum page_references references = PAGEREF_RECLAIM_CLEAN;
 787                bool dirty, writeback;
 788
 789                cond_resched();
 790
 791                page = lru_to_page(page_list);
 792                list_del(&page->lru);
 793
 794                if (!trylock_page(page))
 795                        goto keep;
 796
 797                VM_BUG_ON(PageActive(page));
 798                VM_BUG_ON(page_zone(page) != zone);
 799
 800                sc->nr_scanned++;
 801
 802                if (unlikely(!page_evictable(page)))
 803                        goto cull_mlocked;
 804
 805                if (!sc->may_unmap && page_mapped(page))
 806                        goto keep_locked;
 807
 808                /* Double the slab pressure for mapped and swapcache pages */
 809                if (page_mapped(page) || PageSwapCache(page))
 810                        sc->nr_scanned++;
 811
 812                may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 813                        (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 814
 815                /*
 816                 * The number of dirty pages determines if a zone is marked
 817                 * reclaim_congested which affects wait_iff_congested. kswapd
 818                 * will stall and start writing pages if the tail of the LRU
 819                 * is all dirty unqueued pages.
 820                 */
 821                page_check_dirty_writeback(page, &dirty, &writeback);
 822                if (dirty || writeback)
 823                        nr_dirty++;
 824
 825                if (dirty && !writeback)
 826                        nr_unqueued_dirty++;
 827
 828                /*
 829                 * Treat this page as congested if the underlying BDI is or if
 830                 * pages are cycling through the LRU so quickly that the
 831                 * pages marked for immediate reclaim are making it to the
 832                 * end of the LRU a second time.
 833                 */
 834                mapping = page_mapping(page);
 835                if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
 836                    (writeback && PageReclaim(page)))
 837                        nr_congested++;
 838
 839                /*
 840                 * If a page at the tail of the LRU is under writeback, there
 841                 * are three cases to consider.
 842                 *
 843                 * 1) If reclaim is encountering an excessive number of pages
 844                 *    under writeback and this page is both under writeback and
 845                 *    PageReclaim then it indicates that pages are being queued
 846                 *    for IO but are being recycled through the LRU before the
 847                 *    IO can complete. Waiting on the page itself risks an
 848                 *    indefinite stall if it is impossible to writeback the
 849                 *    page due to IO error or disconnected storage so instead
 850                 *    note that the LRU is being scanned too quickly and the
 851                 *    caller can stall after page list has been processed.
 852                 *
 853                 * 2) Global reclaim encounters a page, memcg encounters a
 854                 *    page that is not marked for immediate reclaim or
 855                 *    the caller does not have __GFP_IO. In this case mark
 856                 *    the page for immediate reclaim and continue scanning.
 857                 *
 858                 *    __GFP_IO is checked  because a loop driver thread might
 859                 *    enter reclaim, and deadlock if it waits on a page for
 860                 *    which it is needed to do the write (loop masks off
 861                 *    __GFP_IO|__GFP_FS for this reason); but more thought
 862                 *    would probably show more reasons.
 863                 *
 864                 *    Don't require __GFP_FS, since we're not going into the
 865                 *    FS, just waiting on its writeback completion. Worryingly,
 866                 *    ext4 gfs2 and xfs allocate pages with
 867                 *    grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
 868                 *    may_enter_fs here is liable to OOM on them.
 869                 *
 870                 * 3) memcg encounters a page that is not already marked
 871                 *    PageReclaim. memcg does not have any dirty pages
 872                 *    throttling so we could easily OOM just because too many
 873                 *    pages are in writeback and there is nothing else to
 874                 *    reclaim. Wait for the writeback to complete.
 875                 */
 876                if (PageWriteback(page)) {
 877                        /* Case 1 above */
 878                        if (current_is_kswapd() &&
 879                            PageReclaim(page) &&
 880                            zone_is_reclaim_writeback(zone)) {
 881                                nr_immediate++;
 882                                goto keep_locked;
 883
 884                        /* Case 2 above */
 885                        } else if (global_reclaim(sc) ||
 886                            !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
 887                                /*
 888                                 * This is slightly racy - end_page_writeback()
 889                                 * might have just cleared PageReclaim, then
 890                                 * setting PageReclaim here end up interpreted
 891                                 * as PageReadahead - but that does not matter
 892                                 * enough to care.  What we do want is for this
 893                                 * page to have PageReclaim set next time memcg
 894                                 * reclaim reaches the tests above, so it will
 895                                 * then wait_on_page_writeback() to avoid OOM;
 896                                 * and it's also appropriate in global reclaim.
 897                                 */
 898                                SetPageReclaim(page);
 899                                nr_writeback++;
 900
 901                                goto keep_locked;
 902
 903                        /* Case 3 above */
 904                        } else {
 905                                wait_on_page_writeback(page);
 906                        }
 907                }
 908
 909                if (!force_reclaim)
 910                        references = page_check_references(page, sc);
 911
 912                switch (references) {
 913                case PAGEREF_ACTIVATE:
 914                        goto activate_locked;
 915                case PAGEREF_KEEP:
 916                        goto keep_locked;
 917                case PAGEREF_RECLAIM:
 918                case PAGEREF_RECLAIM_CLEAN:
 919                        ; /* try to reclaim the page below */
 920                }
 921
 922                /*
 923                 * Anonymous process memory has backing store?
 924                 * Try to allocate it some swap space here.
 925                 */
 926                if (PageAnon(page) && !PageSwapCache(page)) {
 927                        if (!(sc->gfp_mask & __GFP_IO))
 928                                goto keep_locked;
 929                        if (!add_to_swap(page, page_list))
 930                                goto activate_locked;
 931                        may_enter_fs = 1;
 932
 933                        /* Adding to swap updated mapping */
 934                        mapping = page_mapping(page);
 935                }
 936
 937                /*
 938                 * The page is mapped into the page tables of one or more
 939                 * processes. Try to unmap it here.
 940                 */
 941                if (page_mapped(page) && mapping) {
 942                        switch (try_to_unmap(page, ttu_flags)) {
 943                        case SWAP_FAIL:
 944                                goto activate_locked;
 945                        case SWAP_AGAIN:
 946                                goto keep_locked;
 947                        case SWAP_MLOCK:
 948                                goto cull_mlocked;
 949                        case SWAP_SUCCESS:
 950                                ; /* try to free the page below */
 951                        }
 952                }
 953
 954                if (PageDirty(page)) {
 955                        /*
 956                         * Only kswapd can writeback filesystem pages to
 957                         * avoid risk of stack overflow but only writeback
 958                         * if many dirty pages have been encountered.
 959                         */
 960                        if (page_is_file_cache(page) &&
 961                                        (!current_is_kswapd() ||
 962                                         !zone_is_reclaim_dirty(zone))) {
 963                                /*
 964                                 * Immediately reclaim when written back.
 965                                 * Similar in principal to deactivate_page()
 966                                 * except we already have the page isolated
 967                                 * and know it's dirty
 968                                 */
 969                                inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
 970                                SetPageReclaim(page);
 971
 972                                goto keep_locked;
 973                        }
 974
 975                        if (references == PAGEREF_RECLAIM_CLEAN)
 976                                goto keep_locked;
 977                        if (!may_enter_fs)
 978                                goto keep_locked;
 979                        if (!sc->may_writepage)
 980                                goto keep_locked;
 981
 982                        /* Page is dirty, try to write it out here */
 983                        switch (pageout(page, mapping, sc)) {
 984                        case PAGE_KEEP:
 985                                goto keep_locked;
 986                        case PAGE_ACTIVATE:
 987                                goto activate_locked;
 988                        case PAGE_SUCCESS:
 989                                if (PageWriteback(page))
 990                                        goto keep;
 991                                if (PageDirty(page))
 992                                        goto keep;
 993
 994                                /*
 995                                 * A synchronous write - probably a ramdisk.  Go
 996                                 * ahead and try to reclaim the page.
 997                                 */
 998                                if (!trylock_page(page))
 999                                        goto keep;
1000                                if (PageDirty(page) || PageWriteback(page))
1001                                        goto keep_locked;
1002                                mapping = page_mapping(page);
1003                        case PAGE_CLEAN:
1004                                ; /* try to free the page below */
1005                        }
1006                }
1007
1008                /*
1009                 * If the page has buffers, try to free the buffer mappings
1010                 * associated with this page. If we succeed we try to free
1011                 * the page as well.
1012                 *
1013                 * We do this even if the page is PageDirty().
1014                 * try_to_release_page() does not perform I/O, but it is
1015                 * possible for a page to have PageDirty set, but it is actually
1016                 * clean (all its buffers are clean).  This happens if the
1017                 * buffers were written out directly, with submit_bh(). ext3
1018                 * will do this, as well as the blockdev mapping.
1019                 * try_to_release_page() will discover that cleanness and will
1020                 * drop the buffers and mark the page clean - it can be freed.
1021                 *
1022                 * Rarely, pages can have buffers and no ->mapping.  These are
1023                 * the pages which were not successfully invalidated in
1024                 * truncate_complete_page().  We try to drop those buffers here
1025                 * and if that worked, and the page is no longer mapped into
1026                 * process address space (page_count == 1) it can be freed.
1027                 * Otherwise, leave the page on the LRU so it is swappable.
1028                 */
1029                if (page_has_private(page)) {
1030                        if (!try_to_release_page(page, sc->gfp_mask))
1031                                goto activate_locked;
1032                        if (!mapping && page_count(page) == 1) {
1033                                unlock_page(page);
1034                                if (put_page_testzero(page))
1035                                        goto free_it;
1036                                else {
1037                                        /*
1038                                         * rare race with speculative reference.
1039                                         * the speculative reference will free
1040                                         * this page shortly, so we may
1041                                         * increment nr_reclaimed here (and
1042                                         * leave it off the LRU).
1043                                         */
1044                                        nr_reclaimed++;
1045                                        continue;
1046                                }
1047                        }
1048                }
1049
1050                if (!mapping || !__remove_mapping(mapping, page))
1051                        goto keep_locked;
1052
1053                /*
1054                 * At this point, we have no other references and there is
1055                 * no way to pick any more up (removed from LRU, removed
1056                 * from pagecache). Can use non-atomic bitops now (and
1057                 * we obviously don't have to worry about waking up a process
1058                 * waiting on the page lock, because there are no references.
1059                 */
1060                __clear_page_locked(page);
1061free_it:
1062                nr_reclaimed++;
1063
1064                /*
1065                 * Is there need to periodically free_page_list? It would
1066                 * appear not as the counts should be low
1067                 */
1068                list_add(&page->lru, &free_pages);
1069                continue;
1070
1071cull_mlocked:
1072                if (PageSwapCache(page))
1073                        try_to_free_swap(page);
1074                unlock_page(page);
1075                putback_lru_page(page);
1076                continue;
1077
1078activate_locked:
1079                /* Not a candidate for swapping, so reclaim swap space. */
1080                if (PageSwapCache(page) && vm_swap_full())
1081                        try_to_free_swap(page);
1082                VM_BUG_ON(PageActive(page));
1083                SetPageActive(page);
1084                pgactivate++;
1085keep_locked:
1086                unlock_page(page);
1087keep:
1088                list_add(&page->lru, &ret_pages);
1089                VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1090        }
1091
1092        free_hot_cold_page_list(&free_pages, 1);
1093
1094        list_splice(&ret_pages, page_list);
1095        count_vm_events(PGACTIVATE, pgactivate);
1096        mem_cgroup_uncharge_end();
1097        *ret_nr_dirty += nr_dirty;
1098        *ret_nr_congested += nr_congested;
1099        *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1100        *ret_nr_writeback += nr_writeback;
1101        *ret_nr_immediate += nr_immediate;
1102        return nr_reclaimed;
1103}
1104
1105unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1106                                            struct list_head *page_list)
1107{
1108        struct scan_control sc = {
1109                .gfp_mask = GFP_KERNEL,
1110                .priority = DEF_PRIORITY,
1111                .may_unmap = 1,
1112        };
1113        unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1114        struct page *page, *next;
1115        LIST_HEAD(clean_pages);
1116
1117        list_for_each_entry_safe(page, next, page_list, lru) {
1118                if (page_is_file_cache(page) && !PageDirty(page) &&
1119                    !isolated_balloon_page(page)) {
1120                        ClearPageActive(page);
1121                        list_move(&page->lru, &clean_pages);
1122                }
1123        }
1124
1125        ret = shrink_page_list(&clean_pages, zone, &sc,
1126                        TTU_UNMAP|TTU_IGNORE_ACCESS,
1127                        &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1128        list_splice(&clean_pages, page_list);
1129        __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1130        return ret;
1131}
1132
1133/*
1134 * Attempt to remove the specified page from its LRU.  Only take this page
1135 * if it is of the appropriate PageActive status.  Pages which are being
1136 * freed elsewhere are also ignored.
1137 *
1138 * page:        page to consider
1139 * mode:        one of the LRU isolation modes defined above
1140 *
1141 * returns 0 on success, -ve errno on failure.
1142 */
1143int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1144{
1145        int ret = -EINVAL;
1146
1147        /* Only take pages on the LRU. */
1148        if (!PageLRU(page))
1149                return ret;
1150
1151        /* Compaction should not handle unevictable pages but CMA can do so */
1152        if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1153                return ret;
1154
1155        ret = -EBUSY;
1156
1157        /*
1158         * To minimise LRU disruption, the caller can indicate that it only
1159         * wants to isolate pages it will be able to operate on without
1160         * blocking - clean pages for the most part.
1161         *
1162         * ISOLATE_CLEAN means that only clean pages should be isolated. This
1163         * is used by reclaim when it is cannot write to backing storage
1164         *
1165         * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1166         * that it is possible to migrate without blocking
1167         */
1168        if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1169                /* All the caller can do on PageWriteback is block */
1170                if (PageWriteback(page))
1171                        return ret;
1172
1173                if (PageDirty(page)) {
1174                        struct address_space *mapping;
1175
1176                        /* ISOLATE_CLEAN means only clean pages */
1177                        if (mode & ISOLATE_CLEAN)
1178                                return ret;
1179
1180                        /*
1181                         * Only pages without mappings or that have a
1182                         * ->migratepage callback are possible to migrate
1183                         * without blocking
1184                         */
1185                        mapping = page_mapping(page);
1186                        if (mapping && !mapping->a_ops->migratepage)
1187                                return ret;
1188                }
1189        }
1190
1191        if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1192                return ret;
1193
1194        if (likely(get_page_unless_zero(page))) {
1195                /*
1196                 * Be careful not to clear PageLRU until after we're
1197                 * sure the page is not being freed elsewhere -- the
1198                 * page release code relies on it.
1199                 */
1200                ClearPageLRU(page);
1201                ret = 0;
1202        }
1203
1204        return ret;
1205}
1206
1207/*
1208 * zone->lru_lock is heavily contended.  Some of the functions that
1209 * shrink the lists perform better by taking out a batch of pages
1210 * and working on them outside the LRU lock.
1211 *
1212 * For pagecache intensive workloads, this function is the hottest
1213 * spot in the kernel (apart from copy_*_user functions).
1214 *
1215 * Appropriate locks must be held before calling this function.
1216 *
1217 * @nr_to_scan: The number of pages to look through on the list.
1218 * @lruvec:     The LRU vector to pull pages from.
1219 * @dst:        The temp list to put pages on to.
1220 * @nr_scanned: The number of pages that were scanned.
1221 * @sc:         The scan_control struct for this reclaim session
1222 * @mode:       One of the LRU isolation modes
1223 * @lru:        LRU list id for isolating
1224 *
1225 * returns how many pages were moved onto *@dst.
1226 */
1227static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1228                struct lruvec *lruvec, struct list_head *dst,
1229                unsigned long *nr_scanned, struct scan_control *sc,
1230                isolate_mode_t mode, enum lru_list lru)
1231{
1232        struct list_head *src = &lruvec->lists[lru];
1233        unsigned long nr_taken = 0;
1234        unsigned long scan;
1235
1236        for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1237                struct page *page;
1238                int nr_pages;
1239
1240                page = lru_to_page(src);
1241                prefetchw_prev_lru_page(page, src, flags);
1242
1243                VM_BUG_ON(!PageLRU(page));
1244
1245                switch (__isolate_lru_page(page, mode)) {
1246                case 0:
1247                        nr_pages = hpage_nr_pages(page);
1248                        mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1249                        list_move(&page->lru, dst);
1250                        nr_taken += nr_pages;
1251                        break;
1252
1253                case -EBUSY:
1254                        /* else it is being freed elsewhere */
1255                        list_move(&page->lru, src);
1256                        continue;
1257
1258                default:
1259                        BUG();
1260                }
1261        }
1262
1263        *nr_scanned = scan;
1264        trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1265                                    nr_taken, mode, is_file_lru(lru));
1266        return nr_taken;
1267}
1268
1269/**
1270 * isolate_lru_page - tries to isolate a page from its LRU list
1271 * @page: page to isolate from its LRU list
1272 *
1273 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1274 * vmstat statistic corresponding to whatever LRU list the page was on.
1275 *
1276 * Returns 0 if the page was removed from an LRU list.
1277 * Returns -EBUSY if the page was not on an LRU list.
1278 *
1279 * The returned page will have PageLRU() cleared.  If it was found on
1280 * the active list, it will have PageActive set.  If it was found on
1281 * the unevictable list, it will have the PageUnevictable bit set. That flag
1282 * may need to be cleared by the caller before letting the page go.
1283 *
1284 * The vmstat statistic corresponding to the list on which the page was
1285 * found will be decremented.
1286 *
1287 * Restrictions:
1288 * (1) Must be called with an elevated refcount on the page. This is a
1289 *     fundamentnal difference from isolate_lru_pages (which is called
1290 *     without a stable reference).
1291 * (2) the lru_lock must not be held.
1292 * (3) interrupts must be enabled.
1293 */
1294int isolate_lru_page(struct page *page)
1295{
1296        int ret = -EBUSY;
1297
1298        VM_BUG_ON(!page_count(page));
1299
1300        if (PageLRU(page)) {
1301                struct zone *zone = page_zone(page);
1302                struct lruvec *lruvec;
1303
1304                spin_lock_irq(&zone->lru_lock);
1305                lruvec = mem_cgroup_page_lruvec(page, zone);
1306                if (PageLRU(page)) {
1307                        int lru = page_lru(page);
1308                        get_page(page);
1309                        ClearPageLRU(page);
1310                        del_page_from_lru_list(page, lruvec, lru);
1311                        ret = 0;
1312                }
1313                spin_unlock_irq(&zone->lru_lock);
1314        }
1315        return ret;
1316}
1317
1318/*
1319 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1320 * then get resheduled. When there are massive number of tasks doing page
1321 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1322 * the LRU list will go small and be scanned faster than necessary, leading to
1323 * unnecessary swapping, thrashing and OOM.
1324 */
1325static int too_many_isolated(struct zone *zone, int file,
1326                struct scan_control *sc)
1327{
1328        unsigned long inactive, isolated;
1329
1330        if (current_is_kswapd())
1331                return 0;
1332
1333        if (!global_reclaim(sc))
1334                return 0;
1335
1336        if (file) {
1337                inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1338                isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1339        } else {
1340                inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1341                isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1342        }
1343
1344        /*
1345         * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1346         * won't get blocked by normal direct-reclaimers, forming a circular
1347         * deadlock.
1348         */
1349        if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1350                inactive >>= 3;
1351
1352        return isolated > inactive;
1353}
1354
1355static noinline_for_stack void
1356putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1357{
1358        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1359        struct zone *zone = lruvec_zone(lruvec);
1360        LIST_HEAD(pages_to_free);
1361
1362        /*
1363         * Put back any unfreeable pages.
1364         */
1365        while (!list_empty(page_list)) {
1366                struct page *page = lru_to_page(page_list);
1367                int lru;
1368
1369                VM_BUG_ON(PageLRU(page));
1370                list_del(&page->lru);
1371                if (unlikely(!page_evictable(page))) {
1372                        spin_unlock_irq(&zone->lru_lock);
1373                        putback_lru_page(page);
1374                        spin_lock_irq(&zone->lru_lock);
1375                        continue;
1376                }
1377
1378                lruvec = mem_cgroup_page_lruvec(page, zone);
1379
1380                SetPageLRU(page);
1381                lru = page_lru(page);
1382                add_page_to_lru_list(page, lruvec, lru);
1383
1384                if (is_active_lru(lru)) {
1385                        int file = is_file_lru(lru);
1386                        int numpages = hpage_nr_pages(page);
1387                        reclaim_stat->recent_rotated[file] += numpages;
1388                }
1389                if (put_page_testzero(page)) {
1390                        __ClearPageLRU(page);
1391                        __ClearPageActive(page);
1392                        del_page_from_lru_list(page, lruvec, lru);
1393
1394                        if (unlikely(PageCompound(page))) {
1395                                spin_unlock_irq(&zone->lru_lock);
1396                                (*get_compound_page_dtor(page))(page);
1397                                spin_lock_irq(&zone->lru_lock);
1398                        } else
1399                                list_add(&page->lru, &pages_to_free);
1400                }
1401        }
1402
1403        /*
1404         * To save our caller's stack, now use input list for pages to free.
1405         */
1406        list_splice(&pages_to_free, page_list);
1407}
1408
1409/*
1410 * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1411 * of reclaimed pages
1412 */
1413static noinline_for_stack unsigned long
1414shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1415                     struct scan_control *sc, enum lru_list lru)
1416{
1417        LIST_HEAD(page_list);
1418        unsigned long nr_scanned;
1419        unsigned long nr_reclaimed = 0;
1420        unsigned long nr_taken;
1421        unsigned long nr_dirty = 0;
1422        unsigned long nr_congested = 0;
1423        unsigned long nr_unqueued_dirty = 0;
1424        unsigned long nr_writeback = 0;
1425        unsigned long nr_immediate = 0;
1426        isolate_mode_t isolate_mode = 0;
1427        int file = is_file_lru(lru);
1428        struct zone *zone = lruvec_zone(lruvec);
1429        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1430
1431        while (unlikely(too_many_isolated(zone, file, sc))) {
1432                congestion_wait(BLK_RW_ASYNC, HZ/10);
1433
1434                /* We are about to die and free our memory. Return now. */
1435                if (fatal_signal_pending(current))
1436                        return SWAP_CLUSTER_MAX;
1437        }
1438
1439        lru_add_drain();
1440
1441        if (!sc->may_unmap)
1442                isolate_mode |= ISOLATE_UNMAPPED;
1443        if (!sc->may_writepage)
1444                isolate_mode |= ISOLATE_CLEAN;
1445
1446        spin_lock_irq(&zone->lru_lock);
1447
1448        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1449                                     &nr_scanned, sc, isolate_mode, lru);
1450
1451        __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1452        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1453
1454        if (global_reclaim(sc)) {
1455                zone->pages_scanned += nr_scanned;
1456                if (current_is_kswapd())
1457                        __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1458                else
1459                        __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1460        }
1461        spin_unlock_irq(&zone->lru_lock);
1462
1463        if (nr_taken == 0)
1464                return 0;
1465
1466        nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1467                                &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1468                                &nr_writeback, &nr_immediate,
1469                                false);
1470
1471        spin_lock_irq(&zone->lru_lock);
1472
1473        reclaim_stat->recent_scanned[file] += nr_taken;
1474
1475        if (global_reclaim(sc)) {
1476                if (current_is_kswapd())
1477                        __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1478                                               nr_reclaimed);
1479                else
1480                        __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1481                                               nr_reclaimed);
1482        }
1483
1484        putback_inactive_pages(lruvec, &page_list);
1485
1486        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1487
1488        spin_unlock_irq(&zone->lru_lock);
1489
1490        free_hot_cold_page_list(&page_list, 1);
1491
1492        /*
1493         * If reclaim is isolating dirty pages under writeback, it implies
1494         * that the long-lived page allocation rate is exceeding the page
1495         * laundering rate. Either the global limits are not being effective
1496         * at throttling processes due to the page distribution throughout
1497         * zones or there is heavy usage of a slow backing device. The
1498         * only option is to throttle from reclaim context which is not ideal
1499         * as there is no guarantee the dirtying process is throttled in the
1500         * same way balance_dirty_pages() manages.
1501         *
1502         * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1503         * of pages under pages flagged for immediate reclaim and stall if any
1504         * are encountered in the nr_immediate check below.
1505         */
1506        if (nr_writeback && nr_writeback == nr_taken)
1507                zone_set_flag(zone, ZONE_WRITEBACK);
1508
1509        /*
1510         * memcg will stall in page writeback so only consider forcibly
1511         * stalling for global reclaim
1512         */
1513        if (global_reclaim(sc)) {
1514                /*
1515                 * Tag a zone as congested if all the dirty pages scanned were
1516                 * backed by a congested BDI and wait_iff_congested will stall.
1517                 */
1518                if (nr_dirty && nr_dirty == nr_congested)
1519                        zone_set_flag(zone, ZONE_CONGESTED);
1520
1521                /*
1522                 * If dirty pages are scanned that are not queued for IO, it
1523                 * implies that flushers are not keeping up. In this case, flag
1524                 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1525                 * pages from reclaim context. It will forcibly stall in the
1526                 * next check.
1527                 */
1528                if (nr_unqueued_dirty == nr_taken)
1529                        zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1530
1531                /*
1532                 * In addition, if kswapd scans pages marked marked for
1533                 * immediate reclaim and under writeback (nr_immediate), it
1534                 * implies that pages are cycling through the LRU faster than
1535                 * they are written so also forcibly stall.
1536                 */
1537                if (nr_unqueued_dirty == nr_taken || nr_immediate)
1538                        congestion_wait(BLK_RW_ASYNC, HZ/10);
1539        }
1540
1541        /*
1542         * Stall direct reclaim for IO completions if underlying BDIs or zone
1543         * is congested. Allow kswapd to continue until it starts encountering
1544         * unqueued dirty pages or cycling through the LRU too quickly.
1545         */
1546        if (!sc->hibernation_mode && !current_is_kswapd())
1547                wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1548
1549        trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1550                zone_idx(zone),
1551                nr_scanned, nr_reclaimed,
1552                sc->priority,
1553                trace_shrink_flags(file));
1554        return nr_reclaimed;
1555}
1556
1557/*
1558 * This moves pages from the active list to the inactive list.
1559 *
1560 * We move them the other way if the page is referenced by one or more
1561 * processes, from rmap.
1562 *
1563 * If the pages are mostly unmapped, the processing is fast and it is
1564 * appropriate to hold zone->lru_lock across the whole operation.  But if
1565 * the pages are mapped, the processing is slow (page_referenced()) so we
1566 * should drop zone->lru_lock around each page.  It's impossible to balance
1567 * this, so instead we remove the pages from the LRU while processing them.
1568 * It is safe to rely on PG_active against the non-LRU pages in here because
1569 * nobody will play with that bit on a non-LRU page.
1570 *
1571 * The downside is that we have to touch page->_count against each page.
1572 * But we had to alter page->flags anyway.
1573 */
1574
1575static void move_active_pages_to_lru(struct lruvec *lruvec,
1576                                     struct list_head *list,
1577                                     struct list_head *pages_to_free,
1578                                     enum lru_list lru)
1579{
1580        struct zone *zone = lruvec_zone(lruvec);
1581        unsigned long pgmoved = 0;
1582        struct page *page;
1583        int nr_pages;
1584
1585        while (!list_empty(list)) {
1586                page = lru_to_page(list);
1587                lruvec = mem_cgroup_page_lruvec(page, zone);
1588
1589                VM_BUG_ON(PageLRU(page));
1590                SetPageLRU(page);
1591
1592                nr_pages = hpage_nr_pages(page);
1593                mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1594                list_move(&page->lru, &lruvec->lists[lru]);
1595                pgmoved += nr_pages;
1596
1597                if (put_page_testzero(page)) {
1598                        __ClearPageLRU(page);
1599                        __ClearPageActive(page);
1600                        del_page_from_lru_list(page, lruvec, lru);
1601
1602                        if (unlikely(PageCompound(page))) {
1603                                spin_unlock_irq(&zone->lru_lock);
1604                                (*get_compound_page_dtor(page))(page);
1605                                spin_lock_irq(&zone->lru_lock);
1606                        } else
1607                                list_add(&page->lru, pages_to_free);
1608                }
1609        }
1610        __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1611        if (!is_active_lru(lru))
1612                __count_vm_events(PGDEACTIVATE, pgmoved);
1613}
1614
1615static void shrink_active_list(unsigned long nr_to_scan,
1616                               struct lruvec *lruvec,
1617                               struct scan_control *sc,
1618                               enum lru_list lru)
1619{
1620        unsigned long nr_taken;
1621        unsigned long nr_scanned;
1622        unsigned long vm_flags;
1623        LIST_HEAD(l_hold);      /* The pages which were snipped off */
1624        LIST_HEAD(l_active);
1625        LIST_HEAD(l_inactive);
1626        struct page *page;
1627        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1628        unsigned long nr_rotated = 0;
1629        isolate_mode_t isolate_mode = 0;
1630        int file = is_file_lru(lru);
1631        struct zone *zone = lruvec_zone(lruvec);
1632
1633        lru_add_drain();
1634
1635        if (!sc->may_unmap)
1636                isolate_mode |= ISOLATE_UNMAPPED;
1637        if (!sc->may_writepage)
1638                isolate_mode |= ISOLATE_CLEAN;
1639
1640        spin_lock_irq(&zone->lru_lock);
1641
1642        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1643                                     &nr_scanned, sc, isolate_mode, lru);
1644        if (global_reclaim(sc))
1645                zone->pages_scanned += nr_scanned;
1646
1647        reclaim_stat->recent_scanned[file] += nr_taken;
1648
1649        __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1650        __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1651        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1652        spin_unlock_irq(&zone->lru_lock);
1653
1654        while (!list_empty(&l_hold)) {
1655                cond_resched();
1656                page = lru_to_page(&l_hold);
1657                list_del(&page->lru);
1658
1659                if (unlikely(!page_evictable(page))) {
1660                        putback_lru_page(page);
1661                        continue;
1662                }
1663
1664                if (unlikely(buffer_heads_over_limit)) {
1665                        if (page_has_private(page) && trylock_page(page)) {
1666                                if (page_has_private(page))
1667                                        try_to_release_page(page, 0);
1668                                unlock_page(page);
1669                        }
1670                }
1671
1672                if (page_referenced(page, 0, sc->target_mem_cgroup,
1673                                    &vm_flags)) {
1674                        nr_rotated += hpage_nr_pages(page);
1675                        /*
1676                         * Identify referenced, file-backed active pages and
1677                         * give them one more trip around the active list. So
1678                         * that executable code get better chances to stay in
1679                         * memory under moderate memory pressure.  Anon pages
1680                         * are not likely to be evicted by use-once streaming
1681                         * IO, plus JVM can create lots of anon VM_EXEC pages,
1682                         * so we ignore them here.
1683                         */
1684                        if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1685                                list_add(&page->lru, &l_active);
1686                                continue;
1687                        }
1688                }
1689
1690                ClearPageActive(page);  /* we are de-activating */
1691                list_add(&page->lru, &l_inactive);
1692        }
1693
1694        /*
1695         * Move pages back to the lru list.
1696         */
1697        spin_lock_irq(&zone->lru_lock);
1698        /*
1699         * Count referenced pages from currently used mappings as rotated,
1700         * even though only some of them are actually re-activated.  This
1701         * helps balance scan pressure between file and anonymous pages in
1702         * get_scan_ratio.
1703         */
1704        reclaim_stat->recent_rotated[file] += nr_rotated;
1705
1706        move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1707        move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1708        __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1709        spin_unlock_irq(&zone->lru_lock);
1710
1711        free_hot_cold_page_list(&l_hold, 1);
1712}
1713
1714#ifdef CONFIG_SWAP
1715static int inactive_anon_is_low_global(struct zone *zone)
1716{
1717        unsigned long active, inactive;
1718
1719        active = zone_page_state(zone, NR_ACTIVE_ANON);
1720        inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1721
1722        if (inactive * zone->inactive_ratio < active)
1723                return 1;
1724
1725        return 0;
1726}
1727
1728/**
1729 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1730 * @lruvec: LRU vector to check
1731 *
1732 * Returns true if the zone does not have enough inactive anon pages,
1733 * meaning some active anon pages need to be deactivated.
1734 */
1735static int inactive_anon_is_low(struct lruvec *lruvec)
1736{
1737        /*
1738         * If we don't have swap space, anonymous page deactivation
1739         * is pointless.
1740         */
1741        if (!total_swap_pages)
1742                return 0;
1743
1744        if (!mem_cgroup_disabled())
1745                return mem_cgroup_inactive_anon_is_low(lruvec);
1746
1747        return inactive_anon_is_low_global(lruvec_zone(lruvec));
1748}
1749#else
1750static inline int inactive_anon_is_low(struct lruvec *lruvec)
1751{
1752        return 0;
1753}
1754#endif
1755
1756/**
1757 * inactive_file_is_low - check if file pages need to be deactivated
1758 * @lruvec: LRU vector to check
1759 *
1760 * When the system is doing streaming IO, memory pressure here
1761 * ensures that active file pages get deactivated, until more
1762 * than half of the file pages are on the inactive list.
1763 *
1764 * Once we get to that situation, protect the system's working
1765 * set from being evicted by disabling active file page aging.
1766 *
1767 * This uses a different ratio than the anonymous pages, because
1768 * the page cache uses a use-once replacement algorithm.
1769 */
1770static int inactive_file_is_low(struct lruvec *lruvec)
1771{
1772        unsigned long inactive;
1773        unsigned long active;
1774
1775        inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1776        active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1777
1778        return active > inactive;
1779}
1780
1781static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1782{
1783        if (is_file_lru(lru))
1784                return inactive_file_is_low(lruvec);
1785        else
1786                return inactive_anon_is_low(lruvec);
1787}
1788
1789static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1790                                 struct lruvec *lruvec, struct scan_control *sc)
1791{
1792        if (is_active_lru(lru)) {
1793                if (inactive_list_is_low(lruvec, lru))
1794                        shrink_active_list(nr_to_scan, lruvec, sc, lru);
1795                return 0;
1796        }
1797
1798        return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1799}
1800
1801static int vmscan_swappiness(struct scan_control *sc)
1802{
1803        if (global_reclaim(sc))
1804                return vm_swappiness;
1805        return mem_cgroup_swappiness(sc->target_mem_cgroup);
1806}
1807
1808enum scan_balance {
1809        SCAN_EQUAL,
1810        SCAN_FRACT,
1811        SCAN_ANON,
1812        SCAN_FILE,
1813};
1814
1815/*
1816 * Determine how aggressively the anon and file LRU lists should be
1817 * scanned.  The relative value of each set of LRU lists is determined
1818 * by looking at the fraction of the pages scanned we did rotate back
1819 * onto the active list instead of evict.
1820 *
1821 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1822 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1823 */
1824static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1825                           unsigned long *nr)
1826{
1827        struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1828        u64 fraction[2];
1829        u64 denominator = 0;    /* gcc */
1830        struct zone *zone = lruvec_zone(lruvec);
1831        unsigned long anon_prio, file_prio;
1832        enum scan_balance scan_balance;
1833        unsigned long anon, file, free;
1834        bool force_scan = false;
1835        unsigned long ap, fp;
1836        enum lru_list lru;
1837
1838        /*
1839         * If the zone or memcg is small, nr[l] can be 0.  This
1840         * results in no scanning on this priority and a potential
1841         * priority drop.  Global direct reclaim can go to the next
1842         * zone and tends to have no problems. Global kswapd is for
1843         * zone balancing and it needs to scan a minimum amount. When
1844         * reclaiming for a memcg, a priority drop can cause high
1845         * latencies, so it's better to scan a minimum amount there as
1846         * well.
1847         */
1848        if (current_is_kswapd() && !zone_reclaimable(zone))
1849                force_scan = true;
1850        if (!global_reclaim(sc))
1851                force_scan = true;
1852
1853        /* If we have no swap space, do not bother scanning anon pages. */
1854        if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1855                scan_balance = SCAN_FILE;
1856                goto out;
1857        }
1858
1859        /*
1860         * Global reclaim will swap to prevent OOM even with no
1861         * swappiness, but memcg users want to use this knob to
1862         * disable swapping for individual groups completely when
1863         * using the memory controller's swap limit feature would be
1864         * too expensive.
1865         */
1866        if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1867                scan_balance = SCAN_FILE;
1868                goto out;
1869        }
1870
1871        /*
1872         * Do not apply any pressure balancing cleverness when the
1873         * system is close to OOM, scan both anon and file equally
1874         * (unless the swappiness setting disagrees with swapping).
1875         */
1876        if (!sc->priority && vmscan_swappiness(sc)) {
1877                scan_balance = SCAN_EQUAL;
1878                goto out;
1879        }
1880
1881        anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1882                get_lru_size(lruvec, LRU_INACTIVE_ANON);
1883        file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1884                get_lru_size(lruvec, LRU_INACTIVE_FILE);
1885
1886        /*
1887         * If it's foreseeable that reclaiming the file cache won't be
1888         * enough to get the zone back into a desirable shape, we have
1889         * to swap.  Better start now and leave the - probably heavily
1890         * thrashing - remaining file pages alone.
1891         */
1892        if (global_reclaim(sc)) {
1893                free = zone_page_state(zone, NR_FREE_PAGES);
1894                if (unlikely(file + free <= high_wmark_pages(zone))) {
1895                        scan_balance = SCAN_ANON;
1896                        goto out;
1897                }
1898        }
1899
1900        /*
1901         * There is enough inactive page cache, do not reclaim
1902         * anything from the anonymous working set right now.
1903         */
1904        if (!inactive_file_is_low(lruvec)) {
1905                scan_balance = SCAN_FILE;
1906                goto out;
1907        }
1908
1909        scan_balance = SCAN_FRACT;
1910
1911        /*
1912         * With swappiness at 100, anonymous and file have the same priority.
1913         * This scanning priority is essentially the inverse of IO cost.
1914         */
1915        anon_prio = vmscan_swappiness(sc);
1916        file_prio = 200 - anon_prio;
1917
1918        /*
1919         * OK, so we have swap space and a fair amount of page cache
1920         * pages.  We use the recently rotated / recently scanned
1921         * ratios to determine how valuable each cache is.
1922         *
1923         * Because workloads change over time (and to avoid overflow)
1924         * we keep these statistics as a floating average, which ends
1925         * up weighing recent references more than old ones.
1926         *
1927         * anon in [0], file in [1]
1928         */
1929        spin_lock_irq(&zone->lru_lock);
1930        if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1931                reclaim_stat->recent_scanned[0] /= 2;
1932                reclaim_stat->recent_rotated[0] /= 2;
1933        }
1934
1935        if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1936                reclaim_stat->recent_scanned[1] /= 2;
1937                reclaim_stat->recent_rotated[1] /= 2;
1938        }
1939
1940        /*
1941         * The amount of pressure on anon vs file pages is inversely
1942         * proportional to the fraction of recently scanned pages on
1943         * each list that were recently referenced and in active use.
1944         */
1945        ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1946        ap /= reclaim_stat->recent_rotated[0] + 1;
1947
1948        fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1949        fp /= reclaim_stat->recent_rotated[1] + 1;
1950        spin_unlock_irq(&zone->lru_lock);
1951
1952        fraction[0] = ap;
1953        fraction[1] = fp;
1954        denominator = ap + fp + 1;
1955out:
1956        for_each_evictable_lru(lru) {
1957                int file = is_file_lru(lru);
1958                unsigned long size;
1959                unsigned long scan;
1960
1961                size = get_lru_size(lruvec, lru);
1962                scan = size >> sc->priority;
1963
1964                if (!scan && force_scan)
1965                        scan = min(size, SWAP_CLUSTER_MAX);
1966
1967                switch (scan_balance) {
1968                case SCAN_EQUAL:
1969                        /* Scan lists relative to size */
1970                        break;
1971                case SCAN_FRACT:
1972                        /*
1973                         * Scan types proportional to swappiness and
1974                         * their relative recent reclaim efficiency.
1975                         */
1976                        scan = div64_u64(scan * fraction[file], denominator);
1977                        break;
1978                case SCAN_FILE:
1979                case SCAN_ANON:
1980                        /* Scan one type exclusively */
1981                        if ((scan_balance == SCAN_FILE) != file)
1982                                scan = 0;
1983                        break;
1984                default:
1985                        /* Look ma, no brain */
1986                        BUG();
1987                }
1988                nr[lru] = scan;
1989        }
1990}
1991
1992/*
1993 * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1994 */
1995static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1996{
1997        unsigned long nr[NR_LRU_LISTS];
1998        unsigned long targets[NR_LRU_LISTS];
1999        unsigned long nr_to_scan;
2000        enum lru_list lru;
2001        unsigned long nr_reclaimed = 0;
2002        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2003        struct blk_plug plug;
2004        bool scan_adjusted = false;
2005
2006        get_scan_count(lruvec, sc, nr);
2007
2008        /* Record the original scan target for proportional adjustments later */
2009        memcpy(targets, nr, sizeof(nr));
2010
2011        blk_start_plug(&plug);
2012        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2013                                        nr[LRU_INACTIVE_FILE]) {
2014                unsigned long nr_anon, nr_file, percentage;
2015                unsigned long nr_scanned;
2016
2017                for_each_evictable_lru(lru) {
2018                        if (nr[lru]) {
2019                                nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2020                                nr[lru] -= nr_to_scan;
2021
2022                                nr_reclaimed += shrink_list(lru, nr_to_scan,
2023                                                            lruvec, sc);
2024                        }
2025                }
2026
2027                if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2028                        continue;
2029
2030                /*
2031                 * For global direct reclaim, reclaim only the number of pages
2032                 * requested. Less care is taken to scan proportionally as it
2033                 * is more important to minimise direct reclaim stall latency
2034                 * than it is to properly age the LRU lists.
2035                 */
2036                if (global_reclaim(sc) && !current_is_kswapd())
2037                        break;
2038
2039                /*
2040                 * For kswapd and memcg, reclaim at least the number of pages
2041                 * requested. Ensure that the anon and file LRUs shrink
2042                 * proportionally what was requested by get_scan_count(). We
2043                 * stop reclaiming one LRU and reduce the amount scanning
2044                 * proportional to the original scan target.
2045                 */
2046                nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2047                nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2048
2049                if (nr_file > nr_anon) {
2050                        unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2051                                                targets[LRU_ACTIVE_ANON] + 1;
2052                        lru = LRU_BASE;
2053                        percentage = nr_anon * 100 / scan_target;
2054                } else {
2055                        unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2056                                                targets[LRU_ACTIVE_FILE] + 1;
2057                        lru = LRU_FILE;
2058                        percentage = nr_file * 100 / scan_target;
2059                }
2060
2061                /* Stop scanning the smaller of the LRU */
2062                nr[lru] = 0;
2063                nr[lru + LRU_ACTIVE] = 0;
2064
2065                /*
2066                 * Recalculate the other LRU scan count based on its original
2067                 * scan target and the percentage scanning already complete
2068                 */
2069                lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2070                nr_scanned = targets[lru] - nr[lru];
2071                nr[lru] = targets[lru] * (100 - percentage) / 100;
2072                nr[lru] -= min(nr[lru], nr_scanned);
2073
2074                lru += LRU_ACTIVE;
2075                nr_scanned = targets[lru] - nr[lru];
2076                nr[lru] = targets[lru] * (100 - percentage) / 100;
2077                nr[lru] -= min(nr[lru], nr_scanned);
2078
2079                scan_adjusted = true;
2080        }
2081        blk_finish_plug(&plug);
2082        sc->nr_reclaimed += nr_reclaimed;
2083
2084        /*
2085         * Even if we did not try to evict anon pages at all, we want to
2086         * rebalance the anon lru active/inactive ratio.
2087         */
2088        if (inactive_anon_is_low(lruvec))
2089                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2090                                   sc, LRU_ACTIVE_ANON);
2091
2092        throttle_vm_writeout(sc->gfp_mask);
2093}
2094
2095/* Use reclaim/compaction for costly allocs or under memory pressure */
2096static bool in_reclaim_compaction(struct scan_control *sc)
2097{
2098        if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2099                        (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2100                         sc->priority < DEF_PRIORITY - 2))
2101                return true;
2102
2103        return false;
2104}
2105
2106/*
2107 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2108 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2109 * true if more pages should be reclaimed such that when the page allocator
2110 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2111 * It will give up earlier than that if there is difficulty reclaiming pages.
2112 */
2113static inline bool should_continue_reclaim(struct zone *zone,
2114                                        unsigned long nr_reclaimed,
2115                                        unsigned long nr_scanned,
2116                                        struct scan_control *sc)
2117{
2118        unsigned long pages_for_compaction;
2119        unsigned long inactive_lru_pages;
2120
2121        /* If not in reclaim/compaction mode, stop */
2122        if (!in_reclaim_compaction(sc))
2123                return false;
2124
2125        /* Consider stopping depending on scan and reclaim activity */
2126        if (sc->gfp_mask & __GFP_REPEAT) {
2127                /*
2128                 * For __GFP_REPEAT allocations, stop reclaiming if the
2129                 * full LRU list has been scanned and we are still failing
2130                 * to reclaim pages. This full LRU scan is potentially
2131                 * expensive but a __GFP_REPEAT caller really wants to succeed
2132                 */
2133                if (!nr_reclaimed && !nr_scanned)
2134                        return false;
2135        } else {
2136                /*
2137                 * For non-__GFP_REPEAT allocations which can presumably
2138                 * fail without consequence, stop if we failed to reclaim
2139                 * any pages from the last SWAP_CLUSTER_MAX number of
2140                 * pages that were scanned. This will return to the
2141                 * caller faster at the risk reclaim/compaction and
2142                 * the resulting allocation attempt fails
2143                 */
2144                if (!nr_reclaimed)
2145                        return false;
2146        }
2147
2148        /*
2149         * If we have not reclaimed enough pages for compaction and the
2150         * inactive lists are large enough, continue reclaiming
2151         */
2152        pages_for_compaction = (2UL << sc->order);
2153        inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2154        if (get_nr_swap_pages() > 0)
2155                inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2156        if (sc->nr_reclaimed < pages_for_compaction &&
2157                        inactive_lru_pages > pages_for_compaction)
2158                return true;
2159
2160        /* If compaction would go ahead or the allocation would succeed, stop */
2161        switch (compaction_suitable(zone, sc->order)) {
2162        case COMPACT_PARTIAL:
2163        case COMPACT_CONTINUE:
2164                return false;
2165        default:
2166                return true;
2167        }
2168}
2169
2170static void shrink_zone(struct zone *zone, struct scan_control *sc)
2171{
2172        unsigned long nr_reclaimed, nr_scanned;
2173
2174        do {
2175                struct mem_cgroup *root = sc->target_mem_cgroup;
2176                struct mem_cgroup_reclaim_cookie reclaim = {
2177                        .zone = zone,
2178                        .priority = sc->priority,
2179                };
2180                struct mem_cgroup *memcg;
2181
2182                nr_reclaimed = sc->nr_reclaimed;
2183                nr_scanned = sc->nr_scanned;
2184
2185                memcg = mem_cgroup_iter(root, NULL, &reclaim);
2186                do {
2187                        struct lruvec *lruvec;
2188
2189                        lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2190
2191                        shrink_lruvec(lruvec, sc);
2192
2193                        /*
2194                         * Direct reclaim and kswapd have to scan all memory
2195                         * cgroups to fulfill the overall scan target for the
2196                         * zone.
2197                         *
2198                         * Limit reclaim, on the other hand, only cares about
2199                         * nr_to_reclaim pages to be reclaimed and it will
2200                         * retry with decreasing priority if one round over the
2201                         * whole hierarchy is not sufficient.
2202                         */
2203                        if (!global_reclaim(sc) &&
2204                                        sc->nr_reclaimed >= sc->nr_to_reclaim) {
2205                                mem_cgroup_iter_break(root, memcg);
2206                                break;
2207                        }
2208                        memcg = mem_cgroup_iter(root, memcg, &reclaim);
2209                } while (memcg);
2210
2211                vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2212                           sc->nr_scanned - nr_scanned,
2213                           sc->nr_reclaimed - nr_reclaimed);
2214
2215        } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2216                                         sc->nr_scanned - nr_scanned, sc));
2217}
2218
2219/* Returns true if compaction should go ahead for a high-order request */
2220static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2221{
2222        unsigned long balance_gap, watermark;
2223        bool watermark_ok;
2224
2225        /* Do not consider compaction for orders reclaim is meant to satisfy */
2226        if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2227                return false;
2228
2229        /*
2230         * Compaction takes time to run and there are potentially other
2231         * callers using the pages just freed. Continue reclaiming until
2232         * there is a buffer of free pages available to give compaction
2233         * a reasonable chance of completing and allocating the page
2234         */
2235        balance_gap = min(low_wmark_pages(zone),
2236                (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2237                        KSWAPD_ZONE_BALANCE_GAP_RATIO);
2238        watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2239        watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2240
2241        /*
2242         * If compaction is deferred, reclaim up to a point where
2243         * compaction will have a chance of success when re-enabled
2244         */
2245        if (compaction_deferred(zone, sc->order))
2246                return watermark_ok;
2247
2248        /* If compaction is not ready to start, keep reclaiming */
2249        if (!compaction_suitable(zone, sc->order))
2250                return false;
2251
2252        return watermark_ok;
2253}
2254
2255/*
2256 * This is the direct reclaim path, for page-allocating processes.  We only
2257 * try to reclaim pages from zones which will satisfy the caller's allocation
2258 * request.
2259 *
2260 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2261 * Because:
2262 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2263 *    allocation or
2264 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2265 *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2266 *    zone defense algorithm.
2267 *
2268 * If a zone is deemed to be full of pinned pages then just give it a light
2269 * scan then give up on it.
2270 *
2271 * This function returns true if a zone is being reclaimed for a costly
2272 * high-order allocation and compaction is ready to begin. This indicates to
2273 * the caller that it should consider retrying the allocation instead of
2274 * further reclaim.
2275 */
2276static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2277{
2278        struct zoneref *z;
2279        struct zone *zone;
2280        unsigned long nr_soft_reclaimed;
2281        unsigned long nr_soft_scanned;
2282        bool aborted_reclaim = false;
2283
2284        /*
2285         * If the number of buffer_heads in the machine exceeds the maximum
2286         * allowed level, force direct reclaim to scan the highmem zone as
2287         * highmem pages could be pinning lowmem pages storing buffer_heads
2288         */
2289        if (buffer_heads_over_limit)
2290                sc->gfp_mask |= __GFP_HIGHMEM;
2291
2292        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2293                                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2294                if (!populated_zone(zone))
2295                        continue;
2296                /*
2297                 * Take care memory controller reclaiming has small influence
2298                 * to global LRU.
2299                 */
2300                if (global_reclaim(sc)) {
2301                        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2302                                continue;
2303                        if (sc->priority != DEF_PRIORITY &&
2304                            !zone_reclaimable(zone))
2305                                continue;       /* Let kswapd poll it */
2306                        if (IS_ENABLED(CONFIG_COMPACTION)) {
2307                                /*
2308                                 * If we already have plenty of memory free for
2309                                 * compaction in this zone, don't free any more.
2310                                 * Even though compaction is invoked for any
2311                                 * non-zero order, only frequent costly order
2312                                 * reclamation is disruptive enough to become a
2313                                 * noticeable problem, like transparent huge
2314                                 * page allocations.
2315                                 */
2316                                if (compaction_ready(zone, sc)) {
2317                                        aborted_reclaim = true;
2318                                        continue;
2319                                }
2320                        }
2321                        /*
2322                         * This steals pages from memory cgroups over softlimit
2323                         * and returns the number of reclaimed pages and
2324                         * scanned pages. This works for global memory pressure
2325                         * and balancing, not for a memcg's limit.
2326                         */
2327                        nr_soft_scanned = 0;
2328                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2329                                                sc->order, sc->gfp_mask,
2330                                                &nr_soft_scanned);
2331                        sc->nr_reclaimed += nr_soft_reclaimed;
2332                        sc->nr_scanned += nr_soft_scanned;
2333                        /* need some check for avoid more shrink_zone() */
2334                }
2335
2336                shrink_zone(zone, sc);
2337        }
2338
2339        return aborted_reclaim;
2340}
2341
2342/* All zones in zonelist are unreclaimable? */
2343static bool all_unreclaimable(struct zonelist *zonelist,
2344                struct scan_control *sc)
2345{
2346        struct zoneref *z;
2347        struct zone *zone;
2348
2349        for_each_zone_zonelist_nodemask(zone, z, zonelist,
2350                        gfp_zone(sc->gfp_mask), sc->nodemask) {
2351                if (!populated_zone(zone))
2352                        continue;
2353                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2354                        continue;
2355                if (zone_reclaimable(zone))
2356                        return false;
2357        }
2358
2359        return true;
2360}
2361
2362/*
2363 * This is the main entry point to direct page reclaim.
2364 *
2365 * If a full scan of the inactive list fails to free enough memory then we
2366 * are "out of memory" and something needs to be killed.
2367 *
2368 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2369 * high - the zone may be full of dirty or under-writeback pages, which this
2370 * caller can't do much about.  We kick the writeback threads and take explicit
2371 * naps in the hope that some of these pages can be written.  But if the
2372 * allocating task holds filesystem locks which prevent writeout this might not
2373 * work, and the allocation attempt will fail.
2374 *
2375 * returns:     0, if no pages reclaimed
2376 *              else, the number of pages reclaimed
2377 */
2378static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2379                                        struct scan_control *sc,
2380                                        struct shrink_control *shrink)
2381{
2382        unsigned long total_scanned = 0;
2383        struct reclaim_state *reclaim_state = current->reclaim_state;
2384        struct zoneref *z;
2385        struct zone *zone;
2386        unsigned long writeback_threshold;
2387        bool aborted_reclaim;
2388
2389        delayacct_freepages_start();
2390
2391        if (global_reclaim(sc))
2392                count_vm_event(ALLOCSTALL);
2393
2394        do {
2395                vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2396                                sc->priority);
2397                sc->nr_scanned = 0;
2398                aborted_reclaim = shrink_zones(zonelist, sc);
2399
2400                /*
2401                 * Don't shrink slabs when reclaiming memory from over limit
2402                 * cgroups but do shrink slab at least once when aborting
2403                 * reclaim for compaction to avoid unevenly scanning file/anon
2404                 * LRU pages over slab pages.
2405                 */
2406                if (global_reclaim(sc)) {
2407                        unsigned long lru_pages = 0;
2408
2409                        nodes_clear(shrink->nodes_to_scan);
2410                        for_each_zone_zonelist(zone, z, zonelist,
2411                                        gfp_zone(sc->gfp_mask)) {
2412                                if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2413                                        continue;
2414
2415                                lru_pages += zone_reclaimable_pages(zone);
2416                                node_set(zone_to_nid(zone),
2417                                         shrink->nodes_to_scan);
2418                        }
2419
2420                        shrink_slab(shrink, sc->nr_scanned, lru_pages);
2421                        if (reclaim_state) {
2422                                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2423                                reclaim_state->reclaimed_slab = 0;
2424                        }
2425                }
2426                total_scanned += sc->nr_scanned;
2427                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2428                        goto out;
2429
2430                /*
2431                 * If we're getting trouble reclaiming, start doing
2432                 * writepage even in laptop mode.
2433                 */
2434                if (sc->priority < DEF_PRIORITY - 2)
2435                        sc->may_writepage = 1;
2436
2437                /*
2438                 * Try to write back as many pages as we just scanned.  This
2439                 * tends to cause slow streaming writers to write data to the
2440                 * disk smoothly, at the dirtying rate, which is nice.   But
2441                 * that's undesirable in laptop mode, where we *want* lumpy
2442                 * writeout.  So in laptop mode, write out the whole world.
2443                 */
2444                writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2445                if (total_scanned > writeback_threshold) {
2446                        wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2447                                                WB_REASON_TRY_TO_FREE_PAGES);
2448                        sc->may_writepage = 1;
2449                }
2450        } while (--sc->priority >= 0 && !aborted_reclaim);
2451
2452out:
2453        delayacct_freepages_end();
2454
2455        if (sc->nr_reclaimed)
2456                return sc->nr_reclaimed;
2457
2458        /*
2459         * As hibernation is going on, kswapd is freezed so that it can't mark
2460         * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2461         * check.
2462         */
2463        if (oom_killer_disabled)
2464                return 0;
2465
2466        /* Aborted reclaim to try compaction? don't OOM, then */
2467        if (aborted_reclaim)
2468                return 1;
2469
2470        /* top priority shrink_zones still had more to do? don't OOM, then */
2471        if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2472                return 1;
2473
2474        return 0;
2475}
2476
2477static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2478{
2479        struct zone *zone;
2480        unsigned long pfmemalloc_reserve = 0;
2481        unsigned long free_pages = 0;
2482        int i;
2483        bool wmark_ok;
2484
2485        for (i = 0; i <= ZONE_NORMAL; i++) {
2486                zone = &pgdat->node_zones[i];
2487                pfmemalloc_reserve += min_wmark_pages(zone);
2488                free_pages += zone_page_state(zone, NR_FREE_PAGES);
2489        }
2490
2491        wmark_ok = free_pages > pfmemalloc_reserve / 2;
2492
2493        /* kswapd must be awake if processes are being throttled */
2494        if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2495                pgdat->classzone_idx = min(pgdat->classzone_idx,
2496                                                (enum zone_type)ZONE_NORMAL);
2497                wake_up_interruptible(&pgdat->kswapd_wait);
2498        }
2499
2500        return wmark_ok;
2501}
2502
2503/*
2504 * Throttle direct reclaimers if backing storage is backed by the network
2505 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2506 * depleted. kswapd will continue to make progress and wake the processes
2507 * when the low watermark is reached.
2508 *
2509 * Returns true if a fatal signal was delivered during throttling. If this
2510 * happens, the page allocator should not consider triggering the OOM killer.
2511 */
2512static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2513                                        nodemask_t *nodemask)
2514{
2515        struct zone *zone;
2516        int high_zoneidx = gfp_zone(gfp_mask);
2517        pg_data_t *pgdat;
2518
2519        /*
2520         * Kernel threads should not be throttled as they may be indirectly
2521         * responsible for cleaning pages necessary for reclaim to make forward
2522         * progress. kjournald for example may enter direct reclaim while
2523         * committing a transaction where throttling it could forcing other
2524         * processes to block on log_wait_commit().
2525         */
2526        if (current->flags & PF_KTHREAD)
2527                goto out;
2528
2529        /*
2530         * If a fatal signal is pending, this process should not throttle.
2531         * It should return quickly so it can exit and free its memory
2532         */
2533        if (fatal_signal_pending(current))
2534                goto out;
2535
2536        /* Check if the pfmemalloc reserves are ok */
2537        first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2538        pgdat = zone->zone_pgdat;
2539        if (pfmemalloc_watermark_ok(pgdat))
2540                goto out;
2541
2542        /* Account for the throttling */
2543        count_vm_event(PGSCAN_DIRECT_THROTTLE);
2544
2545        /*
2546         * If the caller cannot enter the filesystem, it's possible that it
2547         * is due to the caller holding an FS lock or performing a journal
2548         * transaction in the case of a filesystem like ext[3|4]. In this case,
2549         * it is not safe to block on pfmemalloc_wait as kswapd could be
2550         * blocked waiting on the same lock. Instead, throttle for up to a
2551         * second before continuing.
2552         */
2553        if (!(gfp_mask & __GFP_FS)) {
2554                wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2555                        pfmemalloc_watermark_ok(pgdat), HZ);
2556
2557                goto check_pending;
2558        }
2559
2560        /* Throttle until kswapd wakes the process */
2561        wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2562                pfmemalloc_watermark_ok(pgdat));
2563
2564check_pending:
2565        if (fatal_signal_pending(current))
2566                return true;
2567
2568out:
2569        return false;
2570}
2571
2572unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2573                                gfp_t gfp_mask, nodemask_t *nodemask)
2574{
2575        unsigned long nr_reclaimed;
2576        struct scan_control sc = {
2577                .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2578                .may_writepage = !laptop_mode,
2579                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2580                .may_unmap = 1,
2581                .may_swap = 1,
2582                .order = order,
2583                .priority = DEF_PRIORITY,
2584                .target_mem_cgroup = NULL,
2585                .nodemask = nodemask,
2586        };
2587        struct shrink_control shrink = {
2588                .gfp_mask = sc.gfp_mask,
2589        };
2590
2591        /*
2592         * Do not enter reclaim if fatal signal was delivered while throttled.
2593         * 1 is returned so that the page allocator does not OOM kill at this
2594         * point.
2595         */
2596        if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2597                return 1;
2598
2599        trace_mm_vmscan_direct_reclaim_begin(order,
2600                                sc.may_writepage,
2601                                gfp_mask);
2602
2603        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2604
2605        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2606
2607        return nr_reclaimed;
2608}
2609
2610#ifdef CONFIG_MEMCG
2611
2612unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2613                                                gfp_t gfp_mask, bool noswap,
2614                                                struct zone *zone,
2615                                                unsigned long *nr_scanned)
2616{
2617        struct scan_control sc = {
2618                .nr_scanned = 0,
2619                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2620                .may_writepage = !laptop_mode,
2621                .may_unmap = 1,
2622                .may_swap = !noswap,
2623                .order = 0,
2624                .priority = 0,
2625                .target_mem_cgroup = memcg,
2626        };
2627        struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2628
2629        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2630                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2631
2632        trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2633                                                      sc.may_writepage,
2634                                                      sc.gfp_mask);
2635
2636        /*
2637         * NOTE: Although we can get the priority field, using it
2638         * here is not a good idea, since it limits the pages we can scan.
2639         * if we don't reclaim here, the shrink_zone from balance_pgdat
2640         * will pick up pages from other mem cgroup's as well. We hack
2641         * the priority and make it zero.
2642         */
2643        shrink_lruvec(lruvec, &sc);
2644
2645        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2646
2647        *nr_scanned = sc.nr_scanned;
2648        return sc.nr_reclaimed;
2649}
2650
2651unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2652                                           gfp_t gfp_mask,
2653                                           bool noswap)
2654{
2655        struct zonelist *zonelist;
2656        unsigned long nr_reclaimed;
2657        int nid;
2658        struct scan_control sc = {
2659                .may_writepage = !laptop_mode,
2660                .may_unmap = 1,
2661                .may_swap = !noswap,
2662                .nr_to_reclaim = SWAP_CLUSTER_MAX,
2663                .order = 0,
2664                .priority = DEF_PRIORITY,
2665                .target_mem_cgroup = memcg,
2666                .nodemask = NULL, /* we don't care the placement */
2667                .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2668                                (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2669        };
2670        struct shrink_control shrink = {
2671                .gfp_mask = sc.gfp_mask,
2672        };
2673
2674        /*
2675         * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2676         * take care of from where we get pages. So the node where we start the
2677         * scan does not need to be the current node.
2678         */
2679        nid = mem_cgroup_select_victim_node(memcg);
2680
2681        zonelist = NODE_DATA(nid)->node_zonelists;
2682
2683        trace_mm_vmscan_memcg_reclaim_begin(0,
2684                                            sc.may_writepage,
2685                                            sc.gfp_mask);
2686
2687        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2688
2689        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2690
2691        return nr_reclaimed;
2692}
2693#endif
2694
2695static void age_active_anon(struct zone *zone, struct scan_control *sc)
2696{
2697        struct mem_cgroup *memcg;
2698
2699        if (!total_swap_pages)
2700                return;
2701
2702        memcg = mem_cgroup_iter(NULL, NULL, NULL);
2703        do {
2704                struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2705
2706                if (inactive_anon_is_low(lruvec))
2707                        shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2708                                           sc, LRU_ACTIVE_ANON);
2709
2710                memcg = mem_cgroup_iter(NULL, memcg, NULL);
2711        } while (memcg);
2712}
2713
2714static bool zone_balanced(struct zone *zone, int order,
2715                          unsigned long balance_gap, int classzone_idx)
2716{
2717        if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2718                                    balance_gap, classzone_idx, 0))
2719                return false;
2720
2721        if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2722            !compaction_suitable(zone, order))
2723                return false;
2724
2725        return true;
2726}
2727
2728/*
2729 * pgdat_balanced() is used when checking if a node is balanced.
2730 *
2731 * For order-0, all zones must be balanced!
2732 *
2733 * For high-order allocations only zones that meet watermarks and are in a
2734 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2735 * total of balanced pages must be at least 25% of the zones allowed by
2736 * classzone_idx for the node to be considered balanced. Forcing all zones to
2737 * be balanced for high orders can cause excessive reclaim when there are
2738 * imbalanced zones.
2739 * The choice of 25% is due to
2740 *   o a 16M DMA zone that is balanced will not balance a zone on any
2741 *     reasonable sized machine
2742 *   o On all other machines, the top zone must be at least a reasonable
2743 *     percentage of the middle zones. For example, on 32-bit x86, highmem
2744 *     would need to be at least 256M for it to be balance a whole node.
2745 *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2746 *     to balance a node on its own. These seemed like reasonable ratios.
2747 */
2748static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2749{
2750        unsigned long managed_pages = 0;
2751        unsigned long balanced_pages = 0;
2752        int i;
2753
2754        /* Check the watermark levels */
2755        for (i = 0; i <= classzone_idx; i++) {
2756                struct zone *zone = pgdat->node_zones + i;
2757
2758                if (!populated_zone(zone))
2759                        continue;
2760
2761                managed_pages += zone->managed_pages;
2762
2763                /*
2764                 * A special case here:
2765                 *
2766                 * balance_pgdat() skips over all_unreclaimable after
2767                 * DEF_PRIORITY. Effectively, it considers them balanced so
2768                 * they must be considered balanced here as well!
2769                 */
2770                if (!zone_reclaimable(zone)) {
2771                        balanced_pages += zone->managed_pages;
2772                        continue;
2773                }
2774
2775                if (zone_balanced(zone, order, 0, i))
2776                        balanced_pages += zone->managed_pages;
2777                else if (!order)
2778                        return false;
2779        }
2780
2781        if (order)
2782                return balanced_pages >= (managed_pages >> 2);
2783        else
2784                return true;
2785}
2786
2787/*
2788 * Prepare kswapd for sleeping. This verifies that there are no processes
2789 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2790 *
2791 * Returns true if kswapd is ready to sleep
2792 */
2793static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2794                                        int classzone_idx)
2795{
2796        /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2797        if (remaining)
2798                return false;
2799
2800        /*
2801         * There is a potential race between when kswapd checks its watermarks
2802         * and a process gets throttled. There is also a potential race if
2803         * processes get throttled, kswapd wakes, a large process exits therby
2804         * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2805         * is going to sleep, no process should be sleeping on pfmemalloc_wait
2806         * so wake them now if necessary. If necessary, processes will wake
2807         * kswapd and get throttled again
2808         */
2809        if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2810                wake_up(&pgdat->pfmemalloc_wait);
2811                return false;
2812        }
2813
2814        return pgdat_balanced(pgdat, order, classzone_idx);
2815}
2816
2817/*
2818 * kswapd shrinks the zone by the number of pages required to reach
2819 * the high watermark.
2820 *
2821 * Returns true if kswapd scanned at least the requested number of pages to
2822 * reclaim or if the lack of progress was due to pages under writeback.
2823 * This is used to determine if the scanning priority needs to be raised.
2824 */
2825static bool kswapd_shrink_zone(struct zone *zone,
2826                               int classzone_idx,
2827                               struct scan_control *sc,
2828                               unsigned long lru_pages,
2829                               unsigned long *nr_attempted)
2830{
2831        int testorder = sc->order;
2832        unsigned long balance_gap;
2833        struct reclaim_state *reclaim_state = current->reclaim_state;
2834        struct shrink_control shrink = {
2835                .gfp_mask = sc->gfp_mask,
2836        };
2837        bool lowmem_pressure;
2838
2839        /* Reclaim above the high watermark. */
2840        sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2841
2842        /*
2843         * Kswapd reclaims only single pages with compaction enabled. Trying
2844         * too hard to reclaim until contiguous free pages have become
2845         * available can hurt performance by evicting too much useful data
2846         * from memory. Do not reclaim more than needed for compaction.
2847         */
2848        if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2849                        compaction_suitable(zone, sc->order) !=
2850                                COMPACT_SKIPPED)
2851                testorder = 0;
2852
2853        /*
2854         * We put equal pressure on every zone, unless one zone has way too
2855         * many pages free already. The "too many pages" is defined as the
2856         * high wmark plus a "gap" where the gap is either the low
2857         * watermark or 1% of the zone, whichever is smaller.
2858         */
2859        balance_gap = min(low_wmark_pages(zone),
2860                (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2861                KSWAPD_ZONE_BALANCE_GAP_RATIO);
2862
2863        /*
2864         * If there is no low memory pressure or the zone is balanced then no
2865         * reclaim is necessary
2866         */
2867        lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2868        if (!lowmem_pressure && zone_balanced(zone, testorder,
2869                                                balance_gap, classzone_idx))
2870                return true;
2871
2872        shrink_zone(zone, sc);
2873        nodes_clear(shrink.nodes_to_scan);
2874        node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2875
2876        reclaim_state->reclaimed_slab = 0;
2877        shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2878        sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2879
2880        /* Account for the number of pages attempted to reclaim */
2881        *nr_attempted += sc->nr_to_reclaim;
2882
2883        zone_clear_flag(zone, ZONE_WRITEBACK);
2884
2885        /*
2886         * If a zone reaches its high watermark, consider it to be no longer
2887         * congested. It's possible there are dirty pages backed by congested
2888         * BDIs but as pressure is relieved, speculatively avoid congestion
2889         * waits.
2890         */
2891        if (zone_reclaimable(zone) &&
2892            zone_balanced(zone, testorder, 0, classzone_idx)) {
2893                zone_clear_flag(zone, ZONE_CONGESTED);
2894                zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2895        }
2896
2897        return sc->nr_scanned >= sc->nr_to_reclaim;
2898}
2899
2900/*
2901 * For kswapd, balance_pgdat() will work across all this node's zones until
2902 * they are all at high_wmark_pages(zone).
2903 *
2904 * Returns the final order kswapd was reclaiming at
2905 *
2906 * There is special handling here for zones which are full of pinned pages.
2907 * This can happen if the pages are all mlocked, or if they are all used by
2908 * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2909 * What we do is to detect the case where all pages in the zone have been
2910 * scanned twice and there has been zero successful reclaim.  Mark the zone as
2911 * dead and from now on, only perform a short scan.  Basically we're polling
2912 * the zone for when the problem goes away.
2913 *
2914 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2915 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2916 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2917 * lower zones regardless of the number of free pages in the lower zones. This
2918 * interoperates with the page allocator fallback scheme to ensure that aging
2919 * of pages is balanced across the zones.
2920 */
2921static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2922                                                        int *classzone_idx)
2923{
2924        int i;
2925        int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2926        unsigned long nr_soft_reclaimed;
2927        unsigned long nr_soft_scanned;
2928        struct scan_control sc = {
2929                .gfp_mask = GFP_KERNEL,
2930                .priority = DEF_PRIORITY,
2931                .may_unmap = 1,
2932                .may_swap = 1,
2933                .may_writepage = !laptop_mode,
2934                .order = order,
2935                .target_mem_cgroup = NULL,
2936        };
2937        count_vm_event(PAGEOUTRUN);
2938
2939        do {
2940                unsigned long lru_pages = 0;
2941                unsigned long nr_attempted = 0;
2942                bool raise_priority = true;
2943                bool pgdat_needs_compaction = (order > 0);
2944
2945                sc.nr_reclaimed = 0;
2946
2947                /*
2948                 * Scan in the highmem->dma direction for the highest
2949                 * zone which needs scanning
2950                 */
2951                for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2952                        struct zone *zone = pgdat->node_zones + i;
2953
2954                        if (!populated_zone(zone))
2955                                continue;
2956
2957                        if (sc.priority != DEF_PRIORITY &&
2958                            !zone_reclaimable(zone))
2959                                continue;
2960
2961                        /*
2962                         * Do some background aging of the anon list, to give
2963                         * pages a chance to be referenced before reclaiming.
2964                         */
2965                        age_active_anon(zone, &sc);
2966
2967                        /*
2968                         * If the number of buffer_heads in the machine
2969                         * exceeds the maximum allowed level and this node
2970                         * has a highmem zone, force kswapd to reclaim from
2971                         * it to relieve lowmem pressure.
2972                         */
2973                        if (buffer_heads_over_limit && is_highmem_idx(i)) {
2974                                end_zone = i;
2975                                break;
2976                        }
2977
2978                        if (!zone_balanced(zone, order, 0, 0)) {
2979                                end_zone = i;
2980                                break;
2981                        } else {
2982                                /*
2983                                 * If balanced, clear the dirty and congested
2984                                 * flags
2985                                 */
2986                                zone_clear_flag(zone, ZONE_CONGESTED);
2987                                zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2988                        }
2989                }
2990
2991                if (i < 0)
2992                        goto out;
2993
2994                for (i = 0; i <= end_zone; i++) {
2995                        struct zone *zone = pgdat->node_zones + i;
2996
2997                        if (!populated_zone(zone))
2998                                continue;
2999
3000                        lru_pages += zone_reclaimable_pages(zone);
3001
3002                        /*
3003                         * If any zone is currently balanced then kswapd will
3004                         * not call compaction as it is expected that the
3005                         * necessary pages are already available.
3006                         */
3007                        if (pgdat_needs_compaction &&
3008                                        zone_watermark_ok(zone, order,
3009                                                low_wmark_pages(zone),
3010                                                *classzone_idx, 0))
3011                                pgdat_needs_compaction = false;
3012                }
3013
3014                /*
3015                 * If we're getting trouble reclaiming, start doing writepage
3016                 * even in laptop mode.
3017                 */
3018                if (sc.priority < DEF_PRIORITY - 2)
3019                        sc.may_writepage = 1;
3020
3021                /*
3022                 * Now scan the zone in the dma->highmem direction, stopping
3023                 * at the last zone which needs scanning.
3024                 *
3025                 * We do this because the page allocator works in the opposite
3026                 * direction.  This prevents the page allocator from allocating
3027                 * pages behind kswapd's direction of progress, which would
3028                 * cause too much scanning of the lower zones.
3029                 */
3030                for (i = 0; i <= end_zone; i++) {
3031                        struct zone *zone = pgdat->node_zones + i;
3032
3033                        if (!populated_zone(zone))
3034                                continue;
3035
3036                        if (sc.priority != DEF_PRIORITY &&
3037                            !zone_reclaimable(zone))
3038                                continue;
3039
3040                        sc.nr_scanned = 0;
3041
3042                        nr_soft_scanned = 0;
3043                        /*
3044                         * Call soft limit reclaim before calling shrink_zone.
3045                         */
3046                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3047                                                        order, sc.gfp_mask,
3048                                                        &nr_soft_scanned);
3049                        sc.nr_reclaimed += nr_soft_reclaimed;
3050
3051                        /*
3052                         * There should be no need to raise the scanning
3053                         * priority if enough pages are already being scanned
3054                         * that that high watermark would be met at 100%
3055                         * efficiency.
3056                         */
3057                        if (kswapd_shrink_zone(zone, end_zone, &sc,
3058                                        lru_pages, &nr_attempted))
3059                                raise_priority = false;
3060                }
3061
3062                /*
3063                 * If the low watermark is met there is no need for processes
3064                 * to be throttled on pfmemalloc_wait as they should not be
3065                 * able to safely make forward progress. Wake them
3066                 */
3067                if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3068                                pfmemalloc_watermark_ok(pgdat))
3069                        wake_up(&pgdat->pfmemalloc_wait);
3070
3071                /*
3072                 * Fragmentation may mean that the system cannot be rebalanced
3073                 * for high-order allocations in all zones. If twice the
3074                 * allocation size has been reclaimed and the zones are still
3075                 * not balanced then recheck the watermarks at order-0 to
3076                 * prevent kswapd reclaiming excessively. Assume that a
3077                 * process requested a high-order can direct reclaim/compact.
3078                 */
3079                if (order && sc.nr_reclaimed >= 2UL << order)
3080                        order = sc.order = 0;
3081
3082                /* Check if kswapd should be suspending */
3083                if (try_to_freeze() || kthread_should_stop())
3084                        break;
3085
3086                /*
3087                 * Compact if necessary and kswapd is reclaiming at least the
3088                 * high watermark number of pages as requsted
3089                 */
3090                if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3091                        compact_pgdat(pgdat, order);
3092
3093                /*
3094                 * Raise priority if scanning rate is too low or there was no
3095                 * progress in reclaiming pages
3096                 */
3097                if (raise_priority || !sc.nr_reclaimed)
3098                        sc.priority--;
3099        } while (sc.priority >= 1 &&
3100                 !pgdat_balanced(pgdat, order, *classzone_idx));
3101
3102out:
3103        /*
3104         * Return the order we were reclaiming at so prepare_kswapd_sleep()
3105         * makes a decision on the order we were last reclaiming at. However,
3106         * if another caller entered the allocator slow path while kswapd
3107         * was awake, order will remain at the higher level
3108         */
3109        *classzone_idx = end_zone;
3110        return order;
3111}
3112
3113static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3114{
3115        long remaining = 0;
3116        DEFINE_WAIT(wait);
3117
3118        if (freezing(current) || kthread_should_stop())
3119                return;
3120
3121        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3122
3123        /* Try to sleep for a short interval */
3124        if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3125                remaining = schedule_timeout(HZ/10);
3126                finish_wait(&pgdat->kswapd_wait, &wait);
3127                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3128        }
3129
3130        /*
3131         * After a short sleep, check if it was a premature sleep. If not, then
3132         * go fully to sleep until explicitly woken up.
3133         */
3134        if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3135                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3136
3137                /*
3138                 * vmstat counters are not perfectly accurate and the estimated
3139                 * value for counters such as NR_FREE_PAGES can deviate from the
3140                 * true value by nr_online_cpus * threshold. To avoid the zone
3141                 * watermarks being breached while under pressure, we reduce the
3142                 * per-cpu vmstat threshold while kswapd is awake and restore
3143                 * them before going back to sleep.
3144                 */
3145                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3146
3147                /*
3148                 * Compaction records what page blocks it recently failed to
3149                 * isolate pages from and skips them in the future scanning.
3150                 * When kswapd is going to sleep, it is reasonable to assume
3151                 * that pages and compaction may succeed so reset the cache.
3152                 */
3153                reset_isolation_suitable(pgdat);
3154
3155                if (!kthread_should_stop())
3156                        schedule();
3157
3158                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3159        } else {
3160                if (remaining)
3161                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3162                else
3163                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3164        }
3165        finish_wait(&pgdat->kswapd_wait, &wait);
3166}
3167
3168/*
3169 * The background pageout daemon, started as a kernel thread
3170 * from the init process.
3171 *
3172 * This basically trickles out pages so that we have _some_
3173 * free memory available even if there is no other activity
3174 * that frees anything up. This is needed for things like routing
3175 * etc, where we otherwise might have all activity going on in
3176 * asynchronous contexts that cannot page things out.
3177 *
3178 * If there are applications that are active memory-allocators
3179 * (most normal use), this basically shouldn't matter.
3180 */
3181static int kswapd(void *p)
3182{
3183        unsigned long order, new_order;
3184        unsigned balanced_order;
3185        int classzone_idx, new_classzone_idx;
3186        int balanced_classzone_idx;
3187        pg_data_t *pgdat = (pg_data_t*)p;
3188        struct task_struct *tsk = current;
3189
3190        struct reclaim_state reclaim_state = {
3191                .reclaimed_slab = 0,
3192        };
3193        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3194
3195        lockdep_set_current_reclaim_state(GFP_KERNEL);
3196
3197        if (!cpumask_empty(cpumask))
3198                set_cpus_allowed_ptr(tsk, cpumask);
3199        current->reclaim_state = &reclaim_state;
3200
3201        /*
3202         * Tell the memory management that we're a "memory allocator",
3203         * and that if we need more memory we should get access to it
3204         * regardless (see "__alloc_pages()"). "kswapd" should
3205         * never get caught in the normal page freeing logic.
3206         *
3207         * (Kswapd normally doesn't need memory anyway, but sometimes
3208         * you need a small amount of memory in order to be able to
3209         * page out something else, and this flag essentially protects
3210         * us from recursively trying to free more memory as we're
3211         * trying to free the first piece of memory in the first place).
3212         */
3213        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3214        set_freezable();
3215
3216        order = new_order = 0;
3217        balanced_order = 0;
3218        classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3219        balanced_classzone_idx = classzone_idx;
3220        for ( ; ; ) {
3221                bool ret;
3222
3223                /*
3224                 * If the last balance_pgdat was unsuccessful it's unlikely a
3225                 * new request of a similar or harder type will succeed soon
3226                 * so consider going to sleep on the basis we reclaimed at
3227                 */
3228                if (balanced_classzone_idx >= new_classzone_idx &&
3229                                        balanced_order == new_order) {
3230                        new_order = pgdat->kswapd_max_order;
3231                        new_classzone_idx = pgdat->classzone_idx;
3232                        pgdat->kswapd_max_order =  0;
3233                        pgdat->classzone_idx = pgdat->nr_zones - 1;
3234                }
3235
3236                if (order < new_order || classzone_idx > new_classzone_idx) {
3237                        /*
3238                         * Don't sleep if someone wants a larger 'order'
3239                         * allocation or has tigher zone constraints
3240                         */
3241                        order = new_order;
3242                        classzone_idx = new_classzone_idx;
3243                } else {
3244                        kswapd_try_to_sleep(pgdat, balanced_order,
3245                                                balanced_classzone_idx);
3246                        order = pgdat->kswapd_max_order;
3247                        classzone_idx = pgdat->classzone_idx;
3248                        new_order = order;
3249                        new_classzone_idx = classzone_idx;
3250                        pgdat->kswapd_max_order = 0;
3251                        pgdat->classzone_idx = pgdat->nr_zones - 1;
3252                }
3253
3254                ret = try_to_freeze();
3255                if (kthread_should_stop())
3256                        break;
3257
3258                /*
3259                 * We can speed up thawing tasks if we don't call balance_pgdat
3260                 * after returning from the refrigerator
3261                 */
3262                if (!ret) {
3263                        trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3264                        balanced_classzone_idx = classzone_idx;
3265                        balanced_order = balance_pgdat(pgdat, order,
3266                                                &balanced_classzone_idx);
3267                }
3268        }
3269
3270        current->reclaim_state = NULL;
3271        return 0;
3272}
3273
3274/*
3275 * A zone is low on free memory, so wake its kswapd task to service it.
3276 */
3277void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3278{
3279        pg_data_t *pgdat;
3280
3281        if (!populated_zone(zone))
3282                return;
3283
3284        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3285                return;
3286        pgdat = zone->zone_pgdat;
3287        if (pgdat->kswapd_max_order < order) {
3288                pgdat->kswapd_max_order = order;
3289                pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3290        }
3291        if (!waitqueue_active(&pgdat->kswapd_wait))
3292                return;
3293        if (zone_balanced(zone, order, 0, 0))
3294                return;
3295
3296        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3297        wake_up_interruptible(&pgdat->kswapd_wait);
3298}
3299
3300/*
3301 * The reclaimable count would be mostly accurate.
3302 * The less reclaimable pages may be
3303 * - mlocked pages, which will be moved to unevictable list when encountered
3304 * - mapped pages, which may require several travels to be reclaimed
3305 * - dirty pages, which is not "instantly" reclaimable
3306 */
3307unsigned long global_reclaimable_pages(void)
3308{
3309        int nr;
3310
3311        nr = global_page_state(NR_ACTIVE_FILE) +
3312             global_page_state(NR_INACTIVE_FILE);
3313
3314        if (get_nr_swap_pages() > 0)
3315                nr += global_page_state(NR_ACTIVE_ANON) +
3316                      global_page_state(NR_INACTIVE_ANON);
3317
3318        return nr;
3319}
3320
3321#ifdef CONFIG_HIBERNATION
3322/*
3323 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3324 * freed pages.
3325 *
3326 * Rather than trying to age LRUs the aim is to preserve the overall
3327 * LRU order by reclaiming preferentially
3328 * inactive > active > active referenced > active mapped
3329 */
3330unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3331{
3332        struct reclaim_state reclaim_state;
3333        struct scan_control sc = {
3334                .gfp_mask = GFP_HIGHUSER_MOVABLE,
3335                .may_swap = 1,
3336                .may_unmap = 1,
3337                .may_writepage = 1,
3338                .nr_to_reclaim = nr_to_reclaim,
3339                .hibernation_mode = 1,
3340                .order = 0,
3341                .priority = DEF_PRIORITY,
3342        };
3343        struct shrink_control shrink = {
3344                .gfp_mask = sc.gfp_mask,
3345        };
3346        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3347        struct task_struct *p = current;
3348        unsigned long nr_reclaimed;
3349
3350        p->flags |= PF_MEMALLOC;
3351        lockdep_set_current_reclaim_state(sc.gfp_mask);
3352        reclaim_state.reclaimed_slab = 0;
3353        p->reclaim_state = &reclaim_state;
3354
3355        nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3356
3357        p->reclaim_state = NULL;
3358        lockdep_clear_current_reclaim_state();
3359        p->flags &= ~PF_MEMALLOC;
3360
3361        return nr_reclaimed;
3362}
3363#endif /* CONFIG_HIBERNATION */
3364
3365/* It's optimal to keep kswapds on the same CPUs as their memory, but
3366   not required for correctness.  So if the last cpu in a node goes
3367   away, we get changed to run anywhere: as the first one comes back,
3368   restore their cpu bindings. */
3369static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3370                        void *hcpu)
3371{
3372        int nid;
3373
3374        if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3375                for_each_node_state(nid, N_MEMORY) {
3376                        pg_data_t *pgdat = NODE_DATA(nid);
3377                        const struct cpumask *mask;
3378
3379                        mask = cpumask_of_node(pgdat->node_id);
3380
3381                        if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3382                                /* One of our CPUs online: restore mask */
3383                                set_cpus_allowed_ptr(pgdat->kswapd, mask);
3384                }
3385        }
3386        return NOTIFY_OK;
3387}
3388
3389/*
3390 * This kswapd start function will be called by init and node-hot-add.
3391 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3392 */
3393int kswapd_run(int nid)
3394{
3395        pg_data_t *pgdat = NODE_DATA(nid);
3396        int ret = 0;
3397
3398        if (pgdat->kswapd)
3399                return 0;
3400
3401        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3402        if (IS_ERR(pgdat->kswapd)) {
3403                /* failure at boot is fatal */
3404                BUG_ON(system_state == SYSTEM_BOOTING);
3405                pr_err("Failed to start kswapd on node %d\n", nid);
3406                ret = PTR_ERR(pgdat->kswapd);
3407                pgdat->kswapd = NULL;
3408        }
3409        return ret;
3410}
3411
3412/*
3413 * Called by memory hotplug when all memory in a node is offlined.  Caller must
3414 * hold lock_memory_hotplug().
3415 */
3416void kswapd_stop(int nid)
3417{
3418        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3419
3420        if (kswapd) {
3421                kthread_stop(kswapd);
3422                NODE_DATA(nid)->kswapd = NULL;
3423        }
3424}
3425
3426static int __init kswapd_init(void)
3427{
3428        int nid;
3429
3430        swap_setup();
3431        for_each_node_state(nid, N_MEMORY)
3432                kswapd_run(nid);
3433        hotcpu_notifier(cpu_callback, 0);
3434        return 0;
3435}
3436
3437module_init(kswapd_init)
3438
3439#ifdef CONFIG_NUMA
3440/*
3441 * Zone reclaim mode
3442 *
3443 * If non-zero call zone_reclaim when the number of free pages falls below
3444 * the watermarks.
3445 */
3446int zone_reclaim_mode __read_mostly;
3447
3448#define RECLAIM_OFF 0
3449#define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3450#define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3451#define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3452
3453/*
3454 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3455 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3456 * a zone.
3457 */
3458#define ZONE_RECLAIM_PRIORITY 4
3459
3460/*
3461 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3462 * occur.
3463 */
3464int sysctl_min_unmapped_ratio = 1;
3465
3466/*
3467 * If the number of slab pages in a zone grows beyond this percentage then
3468 * slab reclaim needs to occur.
3469 */
3470int sysctl_min_slab_ratio = 5;
3471
3472static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3473{
3474        unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3475        unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3476                zone_page_state(zone, NR_ACTIVE_FILE);
3477
3478        /*
3479         * It's possible for there to be more file mapped pages than
3480         * accounted for by the pages on the file LRU lists because
3481         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3482         */
3483        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3484}
3485
3486/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3487static long zone_pagecache_reclaimable(struct zone *zone)
3488{
3489        long nr_pagecache_reclaimable;
3490        long delta = 0;
3491
3492        /*
3493         * If RECLAIM_SWAP is set, then all file pages are considered
3494         * potentially reclaimable. Otherwise, we have to worry about
3495         * pages like swapcache and zone_unmapped_file_pages() provides
3496         * a better estimate
3497         */
3498        if (zone_reclaim_mode & RECLAIM_SWAP)
3499                nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3500        else
3501                nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3502
3503        /* If we can't clean pages, remove dirty pages from consideration */
3504        if (!(zone_reclaim_mode & RECLAIM_WRITE))
3505                delta += zone_page_state(zone, NR_FILE_DIRTY);
3506
3507        /* Watch for any possible underflows due to delta */
3508        if (unlikely(delta > nr_pagecache_reclaimable))
3509                delta = nr_pagecache_reclaimable;
3510
3511        return nr_pagecache_reclaimable - delta;
3512}
3513
3514/*
3515 * Try to free up some pages from this zone through reclaim.
3516 */
3517static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3518{
3519        /* Minimum pages needed in order to stay on node */
3520        const unsigned long nr_pages = 1 << order;
3521        struct task_struct *p = current;
3522        struct reclaim_state reclaim_state;
3523        struct scan_control sc = {
3524                .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3525                .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3526                .may_swap = 1,
3527                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3528                .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3529                .order = order,
3530                .priority = ZONE_RECLAIM_PRIORITY,
3531        };
3532        struct shrink_control shrink = {
3533                .gfp_mask = sc.gfp_mask,
3534        };
3535        unsigned long nr_slab_pages0, nr_slab_pages1;
3536
3537        cond_resched();
3538        /*
3539         * We need to be able to allocate from the reserves for RECLAIM_SWAP
3540         * and we also need to be able to write out pages for RECLAIM_WRITE
3541         * and RECLAIM_SWAP.
3542         */
3543        p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3544        lockdep_set_current_reclaim_state(gfp_mask);
3545        reclaim_state.reclaimed_slab = 0;
3546        p->reclaim_state = &reclaim_state;
3547
3548        if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3549                /*
3550                 * Free memory by calling shrink zone with increasing
3551                 * priorities until we have enough memory freed.
3552                 */
3553                do {
3554                        shrink_zone(zone, &sc);
3555                } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3556        }
3557
3558        nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3559        if (nr_slab_pages0 > zone->min_slab_pages) {
3560                /*
3561                 * shrink_slab() does not currently allow us to determine how
3562                 * many pages were freed in this zone. So we take the current
3563                 * number of slab pages and shake the slab until it is reduced
3564                 * by the same nr_pages that we used for reclaiming unmapped
3565                 * pages.
3566                 */
3567                nodes_clear(shrink.nodes_to_scan);
3568                node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3569                for (;;) {
3570                        unsigned long lru_pages = zone_reclaimable_pages(zone);
3571
3572                        /* No reclaimable slab or very low memory pressure */
3573                        if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3574                                break;
3575
3576                        /* Freed enough memory */
3577                        nr_slab_pages1 = zone_page_state(zone,
3578                                                        NR_SLAB_RECLAIMABLE);
3579                        if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3580                                break;
3581                }
3582
3583                /*
3584                 * Update nr_reclaimed by the number of slab pages we
3585                 * reclaimed from this zone.
3586                 */
3587                nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3588                if (nr_slab_pages1 < nr_slab_pages0)
3589                        sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3590        }
3591
3592        p->reclaim_state = NULL;
3593        current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3594        lockdep_clear_current_reclaim_state();
3595        return sc.nr_reclaimed >= nr_pages;
3596}
3597
3598int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3599{
3600        int node_id;
3601        int ret;
3602
3603        /*
3604         * Zone reclaim reclaims unmapped file backed pages and
3605         * slab pages if we are over the defined limits.
3606         *
3607         * A small portion of unmapped file backed pages is needed for
3608         * file I/O otherwise pages read by file I/O will be immediately
3609         * thrown out if the zone is overallocated. So we do not reclaim
3610         * if less than a specified percentage of the zone is used by
3611         * unmapped file backed pages.
3612         */
3613        if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3614            zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3615                return ZONE_RECLAIM_FULL;
3616
3617        if (!zone_reclaimable(zone))
3618                return ZONE_RECLAIM_FULL;
3619
3620        /*
3621         * Do not scan if the allocation should not be delayed.
3622         */
3623        if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3624                return ZONE_RECLAIM_NOSCAN;
3625
3626        /*
3627         * Only run zone reclaim on the local zone or on zones that do not
3628         * have associated processors. This will favor the local processor
3629         * over remote processors and spread off node memory allocations
3630         * as wide as possible.
3631         */
3632        node_id = zone_to_nid(zone);
3633        if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3634                return ZONE_RECLAIM_NOSCAN;
3635
3636        if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3637                return ZONE_RECLAIM_NOSCAN;
3638
3639        ret = __zone_reclaim(zone, gfp_mask, order);
3640        zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3641
3642        if (!ret)
3643                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3644
3645        return ret;
3646}
3647#endif
3648
3649/*
3650 * page_evictable - test whether a page is evictable
3651 * @page: the page to test
3652 *
3653 * Test whether page is evictable--i.e., should be placed on active/inactive
3654 * lists vs unevictable list.
3655 *
3656 * Reasons page might not be evictable:
3657 * (1) page's mapping marked unevictable
3658 * (2) page is part of an mlocked VMA
3659 *
3660 */
3661int page_evictable(struct page *page)
3662{
3663        return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3664}
3665
3666#ifdef CONFIG_SHMEM
3667/**
3668 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3669 * @pages:      array of pages to check
3670 * @nr_pages:   number of pages to check
3671 *
3672 * Checks pages for evictability and moves them to the appropriate lru list.
3673 *
3674 * This function is only used for SysV IPC SHM_UNLOCK.
3675 */
3676void check_move_unevictable_pages(struct page **pages, int nr_pages)
3677{
3678        struct lruvec *lruvec;
3679        struct zone *zone = NULL;
3680        int pgscanned = 0;
3681        int pgrescued = 0;
3682        int i;
3683
3684        for (i = 0; i < nr_pages; i++) {
3685                struct page *page = pages[i];
3686                struct zone *pagezone;
3687
3688                pgscanned++;
3689                pagezone = page_zone(page);
3690                if (pagezone != zone) {
3691                        if (zone)
3692                                spin_unlock_irq(&zone->lru_lock);
3693                        zone = pagezone;
3694                        spin_lock_irq(&zone->lru_lock);
3695                }
3696                lruvec = mem_cgroup_page_lruvec(page, zone);
3697
3698                if (!PageLRU(page) || !PageUnevictable(page))
3699                        continue;
3700
3701                if (page_evictable(page)) {
3702                        enum lru_list lru = page_lru_base_type(page);
3703
3704                        VM_BUG_ON(PageActive(page));
3705                        ClearPageUnevictable(page);
3706                        del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3707                        add_page_to_lru_list(page, lruvec, lru);
3708                        pgrescued++;
3709                }
3710        }
3711
3712        if (zone) {
3713                __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3714                __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3715                spin_unlock_irq(&zone->lru_lock);
3716        }
3717}
3718#endif /* CONFIG_SHMEM */
3719
3720static void warn_scan_unevictable_pages(void)
3721{
3722        printk_once(KERN_WARNING
3723                    "%s: The scan_unevictable_pages sysctl/node-interface has been "
3724                    "disabled for lack of a legitimate use case.  If you have "
3725                    "one, please send an email to linux-mm@kvack.org.\n",
3726                    current->comm);
3727}
3728
3729/*
3730 * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3731 * all nodes' unevictable lists for evictable pages
3732 */
3733unsigned long scan_unevictable_pages;
3734
3735int scan_unevictable_handler(struct ctl_table *table, int write,
3736                           void __user *buffer,
3737                           size_t *length, loff_t *ppos)
3738{
3739        warn_scan_unevictable_pages();
3740        proc_doulongvec_minmax(table, write, buffer, length, ppos);
3741        scan_unevictable_pages = 0;
3742        return 0;
3743}
3744
3745#ifdef CONFIG_NUMA
3746/*
3747 * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3748 * a specified node's per zone unevictable lists for evictable pages.
3749 */
3750
3751static ssize_t read_scan_unevictable_node(struct device *dev,
3752                                          struct device_attribute *attr,
3753                                          char *buf)
3754{
3755        warn_scan_unevictable_pages();
3756        return sprintf(buf, "0\n");     /* always zero; should fit... */
3757}
3758
3759static ssize_t write_scan_unevictable_node(struct device *dev,
3760                                           struct device_attribute *attr,
3761                                        const char *buf, size_t count)
3762{
3763        warn_scan_unevictable_pages();
3764        return 1;
3765}
3766
3767
3768static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3769                        read_scan_unevictable_node,
3770                        write_scan_unevictable_node);
3771
3772int scan_unevictable_register_node(struct node *node)
3773{
3774        return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3775}
3776
3777void scan_unevictable_unregister_node(struct node *node)
3778{
3779        device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3780}
3781#endif
3782
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.