linux/mm/compaction.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * linux/mm/compaction.c
   4 *
   5 * Memory compaction for the reduction of external fragmentation. Note that
   6 * this heavily depends upon page migration to do all the real heavy
   7 * lifting
   8 *
   9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
  10 */
  11#include <linux/cpu.h>
  12#include <linux/swap.h>
  13#include <linux/migrate.h>
  14#include <linux/compaction.h>
  15#include <linux/mm_inline.h>
  16#include <linux/sched/signal.h>
  17#include <linux/backing-dev.h>
  18#include <linux/sysctl.h>
  19#include <linux/sysfs.h>
  20#include <linux/page-isolation.h>
  21#include <linux/kasan.h>
  22#include <linux/kthread.h>
  23#include <linux/freezer.h>
  24#include <linux/page_owner.h>
  25#include <linux/psi.h>
  26#include "internal.h"
  27
  28#ifdef CONFIG_COMPACTION
  29static inline void count_compact_event(enum vm_event_item item)
  30{
  31        count_vm_event(item);
  32}
  33
  34static inline void count_compact_events(enum vm_event_item item, long delta)
  35{
  36        count_vm_events(item, delta);
  37}
  38#else
  39#define count_compact_event(item) do { } while (0)
  40#define count_compact_events(item, delta) do { } while (0)
  41#endif
  42
  43#if defined CONFIG_COMPACTION || defined CONFIG_CMA
  44
  45#define CREATE_TRACE_POINTS
  46#include <trace/events/compaction.h>
  47
  48#define block_start_pfn(pfn, order)     round_down(pfn, 1UL << (order))
  49#define block_end_pfn(pfn, order)       ALIGN((pfn) + 1, 1UL << (order))
  50#define pageblock_start_pfn(pfn)        block_start_pfn(pfn, pageblock_order)
  51#define pageblock_end_pfn(pfn)          block_end_pfn(pfn, pageblock_order)
  52
  53/*
  54 * Fragmentation score check interval for proactive compaction purposes.
  55 */
  56static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
  57
  58/*
  59 * Page order with-respect-to which proactive compaction
  60 * calculates external fragmentation, which is used as
  61 * the "fragmentation score" of a node/zone.
  62 */
  63#if defined CONFIG_TRANSPARENT_HUGEPAGE
  64#define COMPACTION_HPAGE_ORDER  HPAGE_PMD_ORDER
  65#elif defined CONFIG_HUGETLBFS
  66#define COMPACTION_HPAGE_ORDER  HUGETLB_PAGE_ORDER
  67#else
  68#define COMPACTION_HPAGE_ORDER  (PMD_SHIFT - PAGE_SHIFT)
  69#endif
  70
  71static unsigned long release_freepages(struct list_head *freelist)
  72{
  73        struct page *page, *next;
  74        unsigned long high_pfn = 0;
  75
  76        list_for_each_entry_safe(page, next, freelist, lru) {
  77                unsigned long pfn = page_to_pfn(page);
  78                list_del(&page->lru);
  79                __free_page(page);
  80                if (pfn > high_pfn)
  81                        high_pfn = pfn;
  82        }
  83
  84        return high_pfn;
  85}
  86
  87static void split_map_pages(struct list_head *list)
  88{
  89        unsigned int i, order, nr_pages;
  90        struct page *page, *next;
  91        LIST_HEAD(tmp_list);
  92
  93        list_for_each_entry_safe(page, next, list, lru) {
  94                list_del(&page->lru);
  95
  96                order = page_private(page);
  97                nr_pages = 1 << order;
  98
  99                post_alloc_hook(page, order, __GFP_MOVABLE);
 100                if (order)
 101                        split_page(page, order);
 102
 103                for (i = 0; i < nr_pages; i++) {
 104                        list_add(&page->lru, &tmp_list);
 105                        page++;
 106                }
 107        }
 108
 109        list_splice(&tmp_list, list);
 110}
 111
 112#ifdef CONFIG_COMPACTION
 113
 114int PageMovable(struct page *page)
 115{
 116        struct address_space *mapping;
 117
 118        VM_BUG_ON_PAGE(!PageLocked(page), page);
 119        if (!__PageMovable(page))
 120                return 0;
 121
 122        mapping = page_mapping(page);
 123        if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
 124                return 1;
 125
 126        return 0;
 127}
 128EXPORT_SYMBOL(PageMovable);
 129
 130void __SetPageMovable(struct page *page, struct address_space *mapping)
 131{
 132        VM_BUG_ON_PAGE(!PageLocked(page), page);
 133        VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
 134        page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
 135}
 136EXPORT_SYMBOL(__SetPageMovable);
 137
 138void __ClearPageMovable(struct page *page)
 139{
 140        VM_BUG_ON_PAGE(!PageMovable(page), page);
 141        /*
 142         * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
 143         * flag so that VM can catch up released page by driver after isolation.
 144         * With it, VM migration doesn't try to put it back.
 145         */
 146        page->mapping = (void *)((unsigned long)page->mapping &
 147                                PAGE_MAPPING_MOVABLE);
 148}
 149EXPORT_SYMBOL(__ClearPageMovable);
 150
 151/* Do not skip compaction more than 64 times */
 152#define COMPACT_MAX_DEFER_SHIFT 6
 153
 154/*
 155 * Compaction is deferred when compaction fails to result in a page
 156 * allocation success. 1 << compact_defer_shift, compactions are skipped up
 157 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
 158 */
 159static void defer_compaction(struct zone *zone, int order)
 160{
 161        zone->compact_considered = 0;
 162        zone->compact_defer_shift++;
 163
 164        if (order < zone->compact_order_failed)
 165                zone->compact_order_failed = order;
 166
 167        if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
 168                zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
 169
 170        trace_mm_compaction_defer_compaction(zone, order);
 171}
 172
 173/* Returns true if compaction should be skipped this time */
 174static bool compaction_deferred(struct zone *zone, int order)
 175{
 176        unsigned long defer_limit = 1UL << zone->compact_defer_shift;
 177
 178        if (order < zone->compact_order_failed)
 179                return false;
 180
 181        /* Avoid possible overflow */
 182        if (++zone->compact_considered >= defer_limit) {
 183                zone->compact_considered = defer_limit;
 184                return false;
 185        }
 186
 187        trace_mm_compaction_deferred(zone, order);
 188
 189        return true;
 190}
 191
 192/*
 193 * Update defer tracking counters after successful compaction of given order,
 194 * which means an allocation either succeeded (alloc_success == true) or is
 195 * expected to succeed.
 196 */
 197void compaction_defer_reset(struct zone *zone, int order,
 198                bool alloc_success)
 199{
 200        if (alloc_success) {
 201                zone->compact_considered = 0;
 202                zone->compact_defer_shift = 0;
 203        }
 204        if (order >= zone->compact_order_failed)
 205                zone->compact_order_failed = order + 1;
 206
 207        trace_mm_compaction_defer_reset(zone, order);
 208}
 209
 210/* Returns true if restarting compaction after many failures */
 211static bool compaction_restarting(struct zone *zone, int order)
 212{
 213        if (order < zone->compact_order_failed)
 214                return false;
 215
 216        return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
 217                zone->compact_considered >= 1UL << zone->compact_defer_shift;
 218}
 219
 220/* Returns true if the pageblock should be scanned for pages to isolate. */
 221static inline bool isolation_suitable(struct compact_control *cc,
 222                                        struct page *page)
 223{
 224        if (cc->ignore_skip_hint)
 225                return true;
 226
 227        return !get_pageblock_skip(page);
 228}
 229
 230static void reset_cached_positions(struct zone *zone)
 231{
 232        zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
 233        zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
 234        zone->compact_cached_free_pfn =
 235                                pageblock_start_pfn(zone_end_pfn(zone) - 1);
 236}
 237
 238/*
 239 * Compound pages of >= pageblock_order should consistently be skipped until
 240 * released. It is always pointless to compact pages of such order (if they are
 241 * migratable), and the pageblocks they occupy cannot contain any free pages.
 242 */
 243static bool pageblock_skip_persistent(struct page *page)
 244{
 245        if (!PageCompound(page))
 246                return false;
 247
 248        page = compound_head(page);
 249
 250        if (compound_order(page) >= pageblock_order)
 251                return true;
 252
 253        return false;
 254}
 255
 256static bool
 257__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
 258                                                        bool check_target)
 259{
 260        struct page *page = pfn_to_online_page(pfn);
 261        struct page *block_page;
 262        struct page *end_page;
 263        unsigned long block_pfn;
 264
 265        if (!page)
 266                return false;
 267        if (zone != page_zone(page))
 268                return false;
 269        if (pageblock_skip_persistent(page))
 270                return false;
 271
 272        /*
 273         * If skip is already cleared do no further checking once the
 274         * restart points have been set.
 275         */
 276        if (check_source && check_target && !get_pageblock_skip(page))
 277                return true;
 278
 279        /*
 280         * If clearing skip for the target scanner, do not select a
 281         * non-movable pageblock as the starting point.
 282         */
 283        if (!check_source && check_target &&
 284            get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
 285                return false;
 286
 287        /* Ensure the start of the pageblock or zone is online and valid */
 288        block_pfn = pageblock_start_pfn(pfn);
 289        block_pfn = max(block_pfn, zone->zone_start_pfn);
 290        block_page = pfn_to_online_page(block_pfn);
 291        if (block_page) {
 292                page = block_page;
 293                pfn = block_pfn;
 294        }
 295
 296        /* Ensure the end of the pageblock or zone is online and valid */
 297        block_pfn = pageblock_end_pfn(pfn) - 1;
 298        block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
 299        end_page = pfn_to_online_page(block_pfn);
 300        if (!end_page)
 301                return false;
 302
 303        /*
 304         * Only clear the hint if a sample indicates there is either a
 305         * free page or an LRU page in the block. One or other condition
 306         * is necessary for the block to be a migration source/target.
 307         */
 308        do {
 309                if (pfn_valid_within(pfn)) {
 310                        if (check_source && PageLRU(page)) {
 311                                clear_pageblock_skip(page);
 312                                return true;
 313                        }
 314
 315                        if (check_target && PageBuddy(page)) {
 316                                clear_pageblock_skip(page);
 317                                return true;
 318                        }
 319                }
 320
 321                page += (1 << PAGE_ALLOC_COSTLY_ORDER);
 322                pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
 323        } while (page <= end_page);
 324
 325        return false;
 326}
 327
 328/*
 329 * This function is called to clear all cached information on pageblocks that
 330 * should be skipped for page isolation when the migrate and free page scanner
 331 * meet.
 332 */
 333static void __reset_isolation_suitable(struct zone *zone)
 334{
 335        unsigned long migrate_pfn = zone->zone_start_pfn;
 336        unsigned long free_pfn = zone_end_pfn(zone) - 1;
 337        unsigned long reset_migrate = free_pfn;
 338        unsigned long reset_free = migrate_pfn;
 339        bool source_set = false;
 340        bool free_set = false;
 341
 342        if (!zone->compact_blockskip_flush)
 343                return;
 344
 345        zone->compact_blockskip_flush = false;
 346
 347        /*
 348         * Walk the zone and update pageblock skip information. Source looks
 349         * for PageLRU while target looks for PageBuddy. When the scanner
 350         * is found, both PageBuddy and PageLRU are checked as the pageblock
 351         * is suitable as both source and target.
 352         */
 353        for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
 354                                        free_pfn -= pageblock_nr_pages) {
 355                cond_resched();
 356
 357                /* Update the migrate PFN */
 358                if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
 359                    migrate_pfn < reset_migrate) {
 360                        source_set = true;
 361                        reset_migrate = migrate_pfn;
 362                        zone->compact_init_migrate_pfn = reset_migrate;
 363                        zone->compact_cached_migrate_pfn[0] = reset_migrate;
 364                        zone->compact_cached_migrate_pfn[1] = reset_migrate;
 365                }
 366
 367                /* Update the free PFN */
 368                if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
 369                    free_pfn > reset_free) {
 370                        free_set = true;
 371                        reset_free = free_pfn;
 372                        zone->compact_init_free_pfn = reset_free;
 373                        zone->compact_cached_free_pfn = reset_free;
 374                }
 375        }
 376
 377        /* Leave no distance if no suitable block was reset */
 378        if (reset_migrate >= reset_free) {
 379                zone->compact_cached_migrate_pfn[0] = migrate_pfn;
 380                zone->compact_cached_migrate_pfn[1] = migrate_pfn;
 381                zone->compact_cached_free_pfn = free_pfn;
 382        }
 383}
 384
 385void reset_isolation_suitable(pg_data_t *pgdat)
 386{
 387        int zoneid;
 388
 389        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 390                struct zone *zone = &pgdat->node_zones[zoneid];
 391                if (!populated_zone(zone))
 392                        continue;
 393
 394                /* Only flush if a full compaction finished recently */
 395                if (zone->compact_blockskip_flush)
 396                        __reset_isolation_suitable(zone);
 397        }
 398}
 399
 400/*
 401 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
 402 * locks are not required for read/writers. Returns true if it was already set.
 403 */
 404static bool test_and_set_skip(struct compact_control *cc, struct page *page,
 405                                                        unsigned long pfn)
 406{
 407        bool skip;
 408
 409        /* Do no update if skip hint is being ignored */
 410        if (cc->ignore_skip_hint)
 411                return false;
 412
 413        if (!IS_ALIGNED(pfn, pageblock_nr_pages))
 414                return false;
 415
 416        skip = get_pageblock_skip(page);
 417        if (!skip && !cc->no_set_skip_hint)
 418                set_pageblock_skip(page);
 419
 420        return skip;
 421}
 422
 423static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 424{
 425        struct zone *zone = cc->zone;
 426
 427        pfn = pageblock_end_pfn(pfn);
 428
 429        /* Set for isolation rather than compaction */
 430        if (cc->no_set_skip_hint)
 431                return;
 432
 433        if (pfn > zone->compact_cached_migrate_pfn[0])
 434                zone->compact_cached_migrate_pfn[0] = pfn;
 435        if (cc->mode != MIGRATE_ASYNC &&
 436            pfn > zone->compact_cached_migrate_pfn[1])
 437                zone->compact_cached_migrate_pfn[1] = pfn;
 438}
 439
 440/*
 441 * If no pages were isolated then mark this pageblock to be skipped in the
 442 * future. The information is later cleared by __reset_isolation_suitable().
 443 */
 444static void update_pageblock_skip(struct compact_control *cc,
 445                        struct page *page, unsigned long pfn)
 446{
 447        struct zone *zone = cc->zone;
 448
 449        if (cc->no_set_skip_hint)
 450                return;
 451
 452        if (!page)
 453                return;
 454
 455        set_pageblock_skip(page);
 456
 457        /* Update where async and sync compaction should restart */
 458        if (pfn < zone->compact_cached_free_pfn)
 459                zone->compact_cached_free_pfn = pfn;
 460}
 461#else
 462static inline bool isolation_suitable(struct compact_control *cc,
 463                                        struct page *page)
 464{
 465        return true;
 466}
 467
 468static inline bool pageblock_skip_persistent(struct page *page)
 469{
 470        return false;
 471}
 472
 473static inline void update_pageblock_skip(struct compact_control *cc,
 474                        struct page *page, unsigned long pfn)
 475{
 476}
 477
 478static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 479{
 480}
 481
 482static bool test_and_set_skip(struct compact_control *cc, struct page *page,
 483                                                        unsigned long pfn)
 484{
 485        return false;
 486}
 487#endif /* CONFIG_COMPACTION */
 488
 489/*
 490 * Compaction requires the taking of some coarse locks that are potentially
 491 * very heavily contended. For async compaction, trylock and record if the
 492 * lock is contended. The lock will still be acquired but compaction will
 493 * abort when the current block is finished regardless of success rate.
 494 * Sync compaction acquires the lock.
 495 *
 496 * Always returns true which makes it easier to track lock state in callers.
 497 */
 498static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
 499                                                struct compact_control *cc)
 500        __acquires(lock)
 501{
 502        /* Track if the lock is contended in async mode */
 503        if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
 504                if (spin_trylock_irqsave(lock, *flags))
 505                        return true;
 506
 507                cc->contended = true;
 508        }
 509
 510        spin_lock_irqsave(lock, *flags);
 511        return true;
 512}
 513
 514/*
 515 * Compaction requires the taking of some coarse locks that are potentially
 516 * very heavily contended. The lock should be periodically unlocked to avoid
 517 * having disabled IRQs for a long time, even when there is nobody waiting on
 518 * the lock. It might also be that allowing the IRQs will result in
 519 * need_resched() becoming true. If scheduling is needed, async compaction
 520 * aborts. Sync compaction schedules.
 521 * Either compaction type will also abort if a fatal signal is pending.
 522 * In either case if the lock was locked, it is dropped and not regained.
 523 *
 524 * Returns true if compaction should abort due to fatal signal pending, or
 525 *              async compaction due to need_resched()
 526 * Returns false when compaction can continue (sync compaction might have
 527 *              scheduled)
 528 */
 529static bool compact_unlock_should_abort(spinlock_t *lock,
 530                unsigned long flags, bool *locked, struct compact_control *cc)
 531{
 532        if (*locked) {
 533                spin_unlock_irqrestore(lock, flags);
 534                *locked = false;
 535        }
 536
 537        if (fatal_signal_pending(current)) {
 538                cc->contended = true;
 539                return true;
 540        }
 541
 542        cond_resched();
 543
 544        return false;
 545}
 546
 547/*
 548 * Isolate free pages onto a private freelist. If @strict is true, will abort
 549 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
 550 * (even though it may still end up isolating some pages).
 551 */
 552static unsigned long isolate_freepages_block(struct compact_control *cc,
 553                                unsigned long *start_pfn,
 554                                unsigned long end_pfn,
 555                                struct list_head *freelist,
 556                                unsigned int stride,
 557                                bool strict)
 558{
 559        int nr_scanned = 0, total_isolated = 0;
 560        struct page *cursor;
 561        unsigned long flags = 0;
 562        bool locked = false;
 563        unsigned long blockpfn = *start_pfn;
 564        unsigned int order;
 565
 566        /* Strict mode is for isolation, speed is secondary */
 567        if (strict)
 568                stride = 1;
 569
 570        cursor = pfn_to_page(blockpfn);
 571
 572        /* Isolate free pages. */
 573        for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
 574                int isolated;
 575                struct page *page = cursor;
 576
 577                /*
 578                 * Periodically drop the lock (if held) regardless of its
 579                 * contention, to give chance to IRQs. Abort if fatal signal
 580                 * pending or async compaction detects need_resched()
 581                 */
 582                if (!(blockpfn % SWAP_CLUSTER_MAX)
 583                    && compact_unlock_should_abort(&cc->zone->lock, flags,
 584                                                                &locked, cc))
 585                        break;
 586
 587                nr_scanned++;
 588                if (!pfn_valid_within(blockpfn))
 589                        goto isolate_fail;
 590
 591                /*
 592                 * For compound pages such as THP and hugetlbfs, we can save
 593                 * potentially a lot of iterations if we skip them at once.
 594                 * The check is racy, but we can consider only valid values
 595                 * and the only danger is skipping too much.
 596                 */
 597                if (PageCompound(page)) {
 598                        const unsigned int order = compound_order(page);
 599
 600                        if (likely(order < MAX_ORDER)) {
 601                                blockpfn += (1UL << order) - 1;
 602                                cursor += (1UL << order) - 1;
 603                        }
 604                        goto isolate_fail;
 605                }
 606
 607                if (!PageBuddy(page))
 608                        goto isolate_fail;
 609
 610                /*
 611                 * If we already hold the lock, we can skip some rechecking.
 612                 * Note that if we hold the lock now, checked_pageblock was
 613                 * already set in some previous iteration (or strict is true),
 614                 * so it is correct to skip the suitable migration target
 615                 * recheck as well.
 616                 */
 617                if (!locked) {
 618                        locked = compact_lock_irqsave(&cc->zone->lock,
 619                                                                &flags, cc);
 620
 621                        /* Recheck this is a buddy page under lock */
 622                        if (!PageBuddy(page))
 623                                goto isolate_fail;
 624                }
 625
 626                /* Found a free page, will break it into order-0 pages */
 627                order = buddy_order(page);
 628                isolated = __isolate_free_page(page, order);
 629                if (!isolated)
 630                        break;
 631                set_page_private(page, order);
 632
 633                total_isolated += isolated;
 634                cc->nr_freepages += isolated;
 635                list_add_tail(&page->lru, freelist);
 636
 637                if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
 638                        blockpfn += isolated;
 639                        break;
 640                }
 641                /* Advance to the end of split page */
 642                blockpfn += isolated - 1;
 643                cursor += isolated - 1;
 644                continue;
 645
 646isolate_fail:
 647                if (strict)
 648                        break;
 649                else
 650                        continue;
 651
 652        }
 653
 654        if (locked)
 655                spin_unlock_irqrestore(&cc->zone->lock, flags);
 656
 657        /*
 658         * There is a tiny chance that we have read bogus compound_order(),
 659         * so be careful to not go outside of the pageblock.
 660         */
 661        if (unlikely(blockpfn > end_pfn))
 662                blockpfn = end_pfn;
 663
 664        trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
 665                                        nr_scanned, total_isolated);
 666
 667        /* Record how far we have got within the block */
 668        *start_pfn = blockpfn;
 669
 670        /*
 671         * If strict isolation is requested by CMA then check that all the
 672         * pages requested were isolated. If there were any failures, 0 is
 673         * returned and CMA will fail.
 674         */
 675        if (strict && blockpfn < end_pfn)
 676                total_isolated = 0;
 677
 678        cc->total_free_scanned += nr_scanned;
 679        if (total_isolated)
 680                count_compact_events(COMPACTISOLATED, total_isolated);
 681        return total_isolated;
 682}
 683
 684/**
 685 * isolate_freepages_range() - isolate free pages.
 686 * @cc:        Compaction control structure.
 687 * @start_pfn: The first PFN to start isolating.
 688 * @end_pfn:   The one-past-last PFN.
 689 *
 690 * Non-free pages, invalid PFNs, or zone boundaries within the
 691 * [start_pfn, end_pfn) range are considered errors, cause function to
 692 * undo its actions and return zero.
 693 *
 694 * Otherwise, function returns one-past-the-last PFN of isolated page
 695 * (which may be greater then end_pfn if end fell in a middle of
 696 * a free page).
 697 */
 698unsigned long
 699isolate_freepages_range(struct compact_control *cc,
 700                        unsigned long start_pfn, unsigned long end_pfn)
 701{
 702        unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
 703        LIST_HEAD(freelist);
 704
 705        pfn = start_pfn;
 706        block_start_pfn = pageblock_start_pfn(pfn);
 707        if (block_start_pfn < cc->zone->zone_start_pfn)
 708                block_start_pfn = cc->zone->zone_start_pfn;
 709        block_end_pfn = pageblock_end_pfn(pfn);
 710
 711        for (; pfn < end_pfn; pfn += isolated,
 712                                block_start_pfn = block_end_pfn,
 713                                block_end_pfn += pageblock_nr_pages) {
 714                /* Protect pfn from changing by isolate_freepages_block */
 715                unsigned long isolate_start_pfn = pfn;
 716
 717                block_end_pfn = min(block_end_pfn, end_pfn);
 718
 719                /*
 720                 * pfn could pass the block_end_pfn if isolated freepage
 721                 * is more than pageblock order. In this case, we adjust
 722                 * scanning range to right one.
 723                 */
 724                if (pfn >= block_end_pfn) {
 725                        block_start_pfn = pageblock_start_pfn(pfn);
 726                        block_end_pfn = pageblock_end_pfn(pfn);
 727                        block_end_pfn = min(block_end_pfn, end_pfn);
 728                }
 729
 730                if (!pageblock_pfn_to_page(block_start_pfn,
 731                                        block_end_pfn, cc->zone))
 732                        break;
 733
 734                isolated = isolate_freepages_block(cc, &isolate_start_pfn,
 735                                        block_end_pfn, &freelist, 0, true);
 736
 737                /*
 738                 * In strict mode, isolate_freepages_block() returns 0 if
 739                 * there are any holes in the block (ie. invalid PFNs or
 740                 * non-free pages).
 741                 */
 742                if (!isolated)
 743                        break;
 744
 745                /*
 746                 * If we managed to isolate pages, it is always (1 << n) *
 747                 * pageblock_nr_pages for some non-negative n.  (Max order
 748                 * page may span two pageblocks).
 749                 */
 750        }
 751
 752        /* __isolate_free_page() does not map the pages */
 753        split_map_pages(&freelist);
 754
 755        if (pfn < end_pfn) {
 756                /* Loop terminated early, cleanup. */
 757                release_freepages(&freelist);
 758                return 0;
 759        }
 760
 761        /* We don't use freelists for anything. */
 762        return pfn;
 763}
 764
 765/* Similar to reclaim, but different enough that they don't share logic */
 766static bool too_many_isolated(pg_data_t *pgdat)
 767{
 768        unsigned long active, inactive, isolated;
 769
 770        inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
 771                        node_page_state(pgdat, NR_INACTIVE_ANON);
 772        active = node_page_state(pgdat, NR_ACTIVE_FILE) +
 773                        node_page_state(pgdat, NR_ACTIVE_ANON);
 774        isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
 775                        node_page_state(pgdat, NR_ISOLATED_ANON);
 776
 777        return isolated > (inactive + active) / 2;
 778}
 779
 780/**
 781 * isolate_migratepages_block() - isolate all migrate-able pages within
 782 *                                a single pageblock
 783 * @cc:         Compaction control structure.
 784 * @low_pfn:    The first PFN to isolate
 785 * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
 786 * @isolate_mode: Isolation mode to be used.
 787 *
 788 * Isolate all pages that can be migrated from the range specified by
 789 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 790 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
 791 * -ENOMEM in case we could not allocate a page, or 0.
 792 * cc->migrate_pfn will contain the next pfn to scan.
 793 *
 794 * The pages are isolated on cc->migratepages list (not required to be empty),
 795 * and cc->nr_migratepages is updated accordingly.
 796 */
 797static int
 798isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 799                        unsigned long end_pfn, isolate_mode_t isolate_mode)
 800{
 801        pg_data_t *pgdat = cc->zone->zone_pgdat;
 802        unsigned long nr_scanned = 0, nr_isolated = 0;
 803        struct lruvec *lruvec;
 804        unsigned long flags = 0;
 805        struct lruvec *locked = NULL;
 806        struct page *page = NULL, *valid_page = NULL;
 807        unsigned long start_pfn = low_pfn;
 808        bool skip_on_failure = false;
 809        unsigned long next_skip_pfn = 0;
 810        bool skip_updated = false;
 811        int ret = 0;
 812
 813        cc->migrate_pfn = low_pfn;
 814
 815        /*
 816         * Ensure that there are not too many pages isolated from the LRU
 817         * list by either parallel reclaimers or compaction. If there are,
 818         * delay for some time until fewer pages are isolated
 819         */
 820        while (unlikely(too_many_isolated(pgdat))) {
 821                /* stop isolation if there are still pages not migrated */
 822                if (cc->nr_migratepages)
 823                        return -EAGAIN;
 824
 825                /* async migration should just abort */
 826                if (cc->mode == MIGRATE_ASYNC)
 827                        return -EAGAIN;
 828
 829                congestion_wait(BLK_RW_ASYNC, HZ/10);
 830
 831                if (fatal_signal_pending(current))
 832                        return -EINTR;
 833        }
 834
 835        cond_resched();
 836
 837        if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
 838                skip_on_failure = true;
 839                next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 840        }
 841
 842        /* Time to isolate some pages for migration */
 843        for (; low_pfn < end_pfn; low_pfn++) {
 844
 845                if (skip_on_failure && low_pfn >= next_skip_pfn) {
 846                        /*
 847                         * We have isolated all migration candidates in the
 848                         * previous order-aligned block, and did not skip it due
 849                         * to failure. We should migrate the pages now and
 850                         * hopefully succeed compaction.
 851                         */
 852                        if (nr_isolated)
 853                                break;
 854
 855                        /*
 856                         * We failed to isolate in the previous order-aligned
 857                         * block. Set the new boundary to the end of the
 858                         * current block. Note we can't simply increase
 859                         * next_skip_pfn by 1 << order, as low_pfn might have
 860                         * been incremented by a higher number due to skipping
 861                         * a compound or a high-order buddy page in the
 862                         * previous loop iteration.
 863                         */
 864                        next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 865                }
 866
 867                /*
 868                 * Periodically drop the lock (if held) regardless of its
 869                 * contention, to give chance to IRQs. Abort completely if
 870                 * a fatal signal is pending.
 871                 */
 872                if (!(low_pfn % SWAP_CLUSTER_MAX)) {
 873                        if (locked) {
 874                                unlock_page_lruvec_irqrestore(locked, flags);
 875                                locked = NULL;
 876                        }
 877
 878                        if (fatal_signal_pending(current)) {
 879                                cc->contended = true;
 880                                ret = -EINTR;
 881
 882                                goto fatal_pending;
 883                        }
 884
 885                        cond_resched();
 886                }
 887
 888                if (!pfn_valid_within(low_pfn))
 889                        goto isolate_fail;
 890                nr_scanned++;
 891
 892                page = pfn_to_page(low_pfn);
 893
 894                /*
 895                 * Check if the pageblock has already been marked skipped.
 896                 * Only the aligned PFN is checked as the caller isolates
 897                 * COMPACT_CLUSTER_MAX at a time so the second call must
 898                 * not falsely conclude that the block should be skipped.
 899                 */
 900                if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
 901                        if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
 902                                low_pfn = end_pfn;
 903                                page = NULL;
 904                                goto isolate_abort;
 905                        }
 906                        valid_page = page;
 907                }
 908
 909                if (PageHuge(page) && cc->alloc_contig) {
 910                        ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
 911
 912                        /*
 913                         * Fail isolation in case isolate_or_dissolve_huge_page()
 914                         * reports an error. In case of -ENOMEM, abort right away.
 915                         */
 916                        if (ret < 0) {
 917                                 /* Do not report -EBUSY down the chain */
 918                                if (ret == -EBUSY)
 919                                        ret = 0;
 920                                low_pfn += (1UL << compound_order(page)) - 1;
 921                                goto isolate_fail;
 922                        }
 923
 924                        if (PageHuge(page)) {
 925                                /*
 926                                 * Hugepage was successfully isolated and placed
 927                                 * on the cc->migratepages list.
 928                                 */
 929                                low_pfn += compound_nr(page) - 1;
 930                                goto isolate_success_no_list;
 931                        }
 932
 933                        /*
 934                         * Ok, the hugepage was dissolved. Now these pages are
 935                         * Buddy and cannot be re-allocated because they are
 936                         * isolated. Fall-through as the check below handles
 937                         * Buddy pages.
 938                         */
 939                }
 940
 941                /*
 942                 * Skip if free. We read page order here without zone lock
 943                 * which is generally unsafe, but the race window is small and
 944                 * the worst thing that can happen is that we skip some
 945                 * potential isolation targets.
 946                 */
 947                if (PageBuddy(page)) {
 948                        unsigned long freepage_order = buddy_order_unsafe(page);
 949
 950                        /*
 951                         * Without lock, we cannot be sure that what we got is
 952                         * a valid page order. Consider only values in the
 953                         * valid order range to prevent low_pfn overflow.
 954                         */
 955                        if (freepage_order > 0 && freepage_order < MAX_ORDER)
 956                                low_pfn += (1UL << freepage_order) - 1;
 957                        continue;
 958                }
 959
 960                /*
 961                 * Regardless of being on LRU, compound pages such as THP and
 962                 * hugetlbfs are not to be compacted unless we are attempting
 963                 * an allocation much larger than the huge page size (eg CMA).
 964                 * We can potentially save a lot of iterations if we skip them
 965                 * at once. The check is racy, but we can consider only valid
 966                 * values and the only danger is skipping too much.
 967                 */
 968                if (PageCompound(page) && !cc->alloc_contig) {
 969                        const unsigned int order = compound_order(page);
 970
 971                        if (likely(order < MAX_ORDER))
 972                                low_pfn += (1UL << order) - 1;
 973                        goto isolate_fail;
 974                }
 975
 976                /*
 977                 * Check may be lockless but that's ok as we recheck later.
 978                 * It's possible to migrate LRU and non-lru movable pages.
 979                 * Skip any other type of page
 980                 */
 981                if (!PageLRU(page)) {
 982                        /*
 983                         * __PageMovable can return false positive so we need
 984                         * to verify it under page_lock.
 985                         */
 986                        if (unlikely(__PageMovable(page)) &&
 987                                        !PageIsolated(page)) {
 988                                if (locked) {
 989                                        unlock_page_lruvec_irqrestore(locked, flags);
 990                                        locked = NULL;
 991                                }
 992
 993                                if (!isolate_movable_page(page, isolate_mode))
 994                                        goto isolate_success;
 995                        }
 996
 997                        goto isolate_fail;
 998                }
 999
1000                /*
1001                 * Migration will fail if an anonymous page is pinned in memory,
1002                 * so avoid taking lru_lock and isolating it unnecessarily in an
1003                 * admittedly racy check.
1004                 */
1005                if (!page_mapping(page) &&
1006                    page_count(page) > page_mapcount(page))
1007                        goto isolate_fail;
1008
1009                /*
1010                 * Only allow to migrate anonymous pages in GFP_NOFS context
1011                 * because those do not depend on fs locks.
1012                 */
1013                if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
1014                        goto isolate_fail;
1015
1016                /*
1017                 * Be careful not to clear PageLRU until after we're
1018                 * sure the page is not being freed elsewhere -- the
1019                 * page release code relies on it.
1020                 */
1021                if (unlikely(!get_page_unless_zero(page)))
1022                        goto isolate_fail;
1023
1024                if (!__isolate_lru_page_prepare(page, isolate_mode))
1025                        goto isolate_fail_put;
1026
1027                /* Try isolate the page */
1028                if (!TestClearPageLRU(page))
1029                        goto isolate_fail_put;
1030
1031                lruvec = mem_cgroup_page_lruvec(page, pgdat);
1032
1033                /* If we already hold the lock, we can skip some rechecking */
1034                if (lruvec != locked) {
1035                        if (locked)
1036                                unlock_page_lruvec_irqrestore(locked, flags);
1037
1038                        compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1039                        locked = lruvec;
1040
1041                        lruvec_memcg_debug(lruvec, page);
1042
1043                        /* Try get exclusive access under lock */
1044                        if (!skip_updated) {
1045                                skip_updated = true;
1046                                if (test_and_set_skip(cc, page, low_pfn))
1047                                        goto isolate_abort;
1048                        }
1049
1050                        /*
1051                         * Page become compound since the non-locked check,
1052                         * and it's on LRU. It can only be a THP so the order
1053                         * is safe to read and it's 0 for tail pages.
1054                         */
1055                        if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1056                                low_pfn += compound_nr(page) - 1;
1057                                SetPageLRU(page);
1058                                goto isolate_fail_put;
1059                        }
1060                }
1061
1062                /* The whole page is taken off the LRU; skip the tail pages. */
1063                if (PageCompound(page))
1064                        low_pfn += compound_nr(page) - 1;
1065
1066                /* Successfully isolated */
1067                del_page_from_lru_list(page, lruvec);
1068                mod_node_page_state(page_pgdat(page),
1069                                NR_ISOLATED_ANON + page_is_file_lru(page),
1070                                thp_nr_pages(page));
1071
1072isolate_success:
1073                list_add(&page->lru, &cc->migratepages);
1074isolate_success_no_list:
1075                cc->nr_migratepages += compound_nr(page);
1076                nr_isolated += compound_nr(page);
1077
1078                /*
1079                 * Avoid isolating too much unless this block is being
1080                 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1081                 * or a lock is contended. For contention, isolate quickly to
1082                 * potentially remove one source of contention.
1083                 */
1084                if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1085                    !cc->rescan && !cc->contended) {
1086                        ++low_pfn;
1087                        break;
1088                }
1089
1090                continue;
1091
1092isolate_fail_put:
1093                /* Avoid potential deadlock in freeing page under lru_lock */
1094                if (locked) {
1095                        unlock_page_lruvec_irqrestore(locked, flags);
1096                        locked = NULL;
1097                }
1098                put_page(page);
1099
1100isolate_fail:
1101                if (!skip_on_failure && ret != -ENOMEM)
1102                        continue;
1103
1104                /*
1105                 * We have isolated some pages, but then failed. Release them
1106                 * instead of migrating, as we cannot form the cc->order buddy
1107                 * page anyway.
1108                 */
1109                if (nr_isolated) {
1110                        if (locked) {
1111                                unlock_page_lruvec_irqrestore(locked, flags);
1112                                locked = NULL;
1113                        }
1114                        putback_movable_pages(&cc->migratepages);
1115                        cc->nr_migratepages = 0;
1116                        nr_isolated = 0;
1117                }
1118
1119                if (low_pfn < next_skip_pfn) {
1120                        low_pfn = next_skip_pfn - 1;
1121                        /*
1122                         * The check near the loop beginning would have updated
1123                         * next_skip_pfn too, but this is a bit simpler.
1124                         */
1125                        next_skip_pfn += 1UL << cc->order;
1126                }
1127
1128                if (ret == -ENOMEM)
1129                        break;
1130        }
1131
1132        /*
1133         * The PageBuddy() check could have potentially brought us outside
1134         * the range to be scanned.
1135         */
1136        if (unlikely(low_pfn > end_pfn))
1137                low_pfn = end_pfn;
1138
1139        page = NULL;
1140
1141isolate_abort:
1142        if (locked)
1143                unlock_page_lruvec_irqrestore(locked, flags);
1144        if (page) {
1145                SetPageLRU(page);
1146                put_page(page);
1147        }
1148
1149        /*
1150         * Updated the cached scanner pfn once the pageblock has been scanned
1151         * Pages will either be migrated in which case there is no point
1152         * scanning in the near future or migration failed in which case the
1153         * failure reason may persist. The block is marked for skipping if
1154         * there were no pages isolated in the block or if the block is
1155         * rescanned twice in a row.
1156         */
1157        if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1158                if (valid_page && !skip_updated)
1159                        set_pageblock_skip(valid_page);
1160                update_cached_migrate(cc, low_pfn);
1161        }
1162
1163        trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1164                                                nr_scanned, nr_isolated);
1165
1166fatal_pending:
1167        cc->total_migrate_scanned += nr_scanned;
1168        if (nr_isolated)
1169                count_compact_events(COMPACTISOLATED, nr_isolated);
1170
1171        cc->migrate_pfn = low_pfn;
1172
1173        return ret;
1174}
1175
1176/**
1177 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1178 * @cc:        Compaction control structure.
1179 * @start_pfn: The first PFN to start isolating.
1180 * @end_pfn:   The one-past-last PFN.
1181 *
1182 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1183 * in case we could not allocate a page, or 0.
1184 */
1185int
1186isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1187                                                        unsigned long end_pfn)
1188{
1189        unsigned long pfn, block_start_pfn, block_end_pfn;
1190        int ret = 0;
1191
1192        /* Scan block by block. First and last block may be incomplete */
1193        pfn = start_pfn;
1194        block_start_pfn = pageblock_start_pfn(pfn);
1195        if (block_start_pfn < cc->zone->zone_start_pfn)
1196                block_start_pfn = cc->zone->zone_start_pfn;
1197        block_end_pfn = pageblock_end_pfn(pfn);
1198
1199        for (; pfn < end_pfn; pfn = block_end_pfn,
1200                                block_start_pfn = block_end_pfn,
1201                                block_end_pfn += pageblock_nr_pages) {
1202
1203                block_end_pfn = min(block_end_pfn, end_pfn);
1204
1205                if (!pageblock_pfn_to_page(block_start_pfn,
1206                                        block_end_pfn, cc->zone))
1207                        continue;
1208
1209                ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1210                                                 ISOLATE_UNEVICTABLE);
1211
1212                if (ret)
1213                        break;
1214
1215                if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1216                        break;
1217        }
1218
1219        return ret;
1220}
1221
1222#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1223#ifdef CONFIG_COMPACTION
1224
1225static bool suitable_migration_source(struct compact_control *cc,
1226                                                        struct page *page)
1227{
1228        int block_mt;
1229
1230        if (pageblock_skip_persistent(page))
1231                return false;
1232
1233        if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1234                return true;
1235
1236        block_mt = get_pageblock_migratetype(page);
1237
1238        if (cc->migratetype == MIGRATE_MOVABLE)
1239                return is_migrate_movable(block_mt);
1240        else
1241                return block_mt == cc->migratetype;
1242}
1243
1244/* Returns true if the page is within a block suitable for migration to */
1245static bool suitable_migration_target(struct compact_control *cc,
1246                                                        struct page *page)
1247{
1248        /* If the page is a large free page, then disallow migration */
1249        if (PageBuddy(page)) {
1250                /*
1251                 * We are checking page_order without zone->lock taken. But
1252                 * the only small danger is that we skip a potentially suitable
1253                 * pageblock, so it's not worth to check order for valid range.
1254                 */
1255                if (buddy_order_unsafe(page) >= pageblock_order)
1256                        return false;
1257        }
1258
1259        if (cc->ignore_block_suitable)
1260                return true;
1261
1262        /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1263        if (is_migrate_movable(get_pageblock_migratetype(page)))
1264                return true;
1265
1266        /* Otherwise skip the block */
1267        return false;
1268}
1269
1270static inline unsigned int
1271freelist_scan_limit(struct compact_control *cc)
1272{
1273        unsigned short shift = BITS_PER_LONG - 1;
1274
1275        return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1276}
1277
1278/*
1279 * Test whether the free scanner has reached the same or lower pageblock than
1280 * the migration scanner, and compaction should thus terminate.
1281 */
1282static inline bool compact_scanners_met(struct compact_control *cc)
1283{
1284        return (cc->free_pfn >> pageblock_order)
1285                <= (cc->migrate_pfn >> pageblock_order);
1286}
1287
1288/*
1289 * Used when scanning for a suitable migration target which scans freelists
1290 * in reverse. Reorders the list such as the unscanned pages are scanned
1291 * first on the next iteration of the free scanner
1292 */
1293static void
1294move_freelist_head(struct list_head *freelist, struct page *freepage)
1295{
1296        LIST_HEAD(sublist);
1297
1298        if (!list_is_last(freelist, &freepage->lru)) {
1299                list_cut_before(&sublist, freelist, &freepage->lru);
1300                if (!list_empty(&sublist))
1301                        list_splice_tail(&sublist, freelist);
1302        }
1303}
1304
1305/*
1306 * Similar to move_freelist_head except used by the migration scanner
1307 * when scanning forward. It's possible for these list operations to
1308 * move against each other if they search the free list exactly in
1309 * lockstep.
1310 */
1311static void
1312move_freelist_tail(struct list_head *freelist, struct page *freepage)
1313{
1314        LIST_HEAD(sublist);
1315
1316        if (!list_is_first(freelist, &freepage->lru)) {
1317                list_cut_position(&sublist, freelist, &freepage->lru);
1318                if (!list_empty(&sublist))
1319                        list_splice_tail(&sublist, freelist);
1320        }
1321}
1322
1323static void
1324fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1325{
1326        unsigned long start_pfn, end_pfn;
1327        struct page *page;
1328
1329        /* Do not search around if there are enough pages already */
1330        if (cc->nr_freepages >= cc->nr_migratepages)
1331                return;
1332
1333        /* Minimise scanning during async compaction */
1334        if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1335                return;
1336
1337        /* Pageblock boundaries */
1338        start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1339        end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1340
1341        page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1342        if (!page)
1343                return;
1344
1345        /* Scan before */
1346        if (start_pfn != pfn) {
1347                isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1348                if (cc->nr_freepages >= cc->nr_migratepages)
1349                        return;
1350        }
1351
1352        /* Scan after */
1353        start_pfn = pfn + nr_isolated;
1354        if (start_pfn < end_pfn)
1355                isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1356
1357        /* Skip this pageblock in the future as it's full or nearly full */
1358        if (cc->nr_freepages < cc->nr_migratepages)
1359                set_pageblock_skip(page);
1360}
1361
1362/* Search orders in round-robin fashion */
1363static int next_search_order(struct compact_control *cc, int order)
1364{
1365        order--;
1366        if (order < 0)
1367                order = cc->order - 1;
1368
1369        /* Search wrapped around? */
1370        if (order == cc->search_order) {
1371                cc->search_order--;
1372                if (cc->search_order < 0)
1373                        cc->search_order = cc->order - 1;
1374                return -1;
1375        }
1376
1377        return order;
1378}
1379
1380static unsigned long
1381fast_isolate_freepages(struct compact_control *cc)
1382{
1383        unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1384        unsigned int nr_scanned = 0;
1385        unsigned long low_pfn, min_pfn, highest = 0;
1386        unsigned long nr_isolated = 0;
1387        unsigned long distance;
1388        struct page *page = NULL;
1389        bool scan_start = false;
1390        int order;
1391
1392        /* Full compaction passes in a negative order */
1393        if (cc->order <= 0)
1394                return cc->free_pfn;
1395
1396        /*
1397         * If starting the scan, use a deeper search and use the highest
1398         * PFN found if a suitable one is not found.
1399         */
1400        if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1401                limit = pageblock_nr_pages >> 1;
1402                scan_start = true;
1403        }
1404
1405        /*
1406         * Preferred point is in the top quarter of the scan space but take
1407         * a pfn from the top half if the search is problematic.
1408         */
1409        distance = (cc->free_pfn - cc->migrate_pfn);
1410        low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1411        min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1412
1413        if (WARN_ON_ONCE(min_pfn > low_pfn))
1414                low_pfn = min_pfn;
1415
1416        /*
1417         * Search starts from the last successful isolation order or the next
1418         * order to search after a previous failure
1419         */
1420        cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1421
1422        for (order = cc->search_order;
1423             !page && order >= 0;
1424             order = next_search_order(cc, order)) {
1425                struct free_area *area = &cc->zone->free_area[order];
1426                struct list_head *freelist;
1427                struct page *freepage;
1428                unsigned long flags;
1429                unsigned int order_scanned = 0;
1430                unsigned long high_pfn = 0;
1431
1432                if (!area->nr_free)
1433                        continue;
1434
1435                spin_lock_irqsave(&cc->zone->lock, flags);
1436                freelist = &area->free_list[MIGRATE_MOVABLE];
1437                list_for_each_entry_reverse(freepage, freelist, lru) {
1438                        unsigned long pfn;
1439
1440                        order_scanned++;
1441                        nr_scanned++;
1442                        pfn = page_to_pfn(freepage);
1443
1444                        if (pfn >= highest)
1445                                highest = max(pageblock_start_pfn(pfn),
1446                                              cc->zone->zone_start_pfn);
1447
1448                        if (pfn >= low_pfn) {
1449                                cc->fast_search_fail = 0;
1450                                cc->search_order = order;
1451                                page = freepage;
1452                                break;
1453                        }
1454
1455                        if (pfn >= min_pfn && pfn > high_pfn) {
1456                                high_pfn = pfn;
1457
1458                                /* Shorten the scan if a candidate is found */
1459                                limit >>= 1;
1460                        }
1461
1462                        if (order_scanned >= limit)
1463                                break;
1464                }
1465
1466                /* Use a minimum pfn if a preferred one was not found */
1467                if (!page && high_pfn) {
1468                        page = pfn_to_page(high_pfn);
1469
1470                        /* Update freepage for the list reorder below */
1471                        freepage = page;
1472                }
1473
1474                /* Reorder to so a future search skips recent pages */
1475                move_freelist_head(freelist, freepage);
1476
1477                /* Isolate the page if available */
1478                if (page) {
1479                        if (__isolate_free_page(page, order)) {
1480                                set_page_private(page, order);
1481                                nr_isolated = 1 << order;
1482                                cc->nr_freepages += nr_isolated;
1483                                list_add_tail(&page->lru, &cc->freepages);
1484                                count_compact_events(COMPACTISOLATED, nr_isolated);
1485                        } else {
1486                                /* If isolation fails, abort the search */
1487                                order = cc->search_order + 1;
1488                                page = NULL;
1489                        }
1490                }
1491
1492                spin_unlock_irqrestore(&cc->zone->lock, flags);
1493
1494                /*
1495                 * Smaller scan on next order so the total scan ig related
1496                 * to freelist_scan_limit.
1497                 */
1498                if (order_scanned >= limit)
1499                        limit = min(1U, limit >> 1);
1500        }
1501
1502        if (!page) {
1503                cc->fast_search_fail++;
1504                if (scan_start) {
1505                        /*
1506                         * Use the highest PFN found above min. If one was
1507                         * not found, be pessimistic for direct compaction
1508                         * and use the min mark.
1509                         */
1510                        if (highest) {
1511                                page = pfn_to_page(highest);
1512                                cc->free_pfn = highest;
1513                        } else {
1514                                if (cc->direct_compaction && pfn_valid(min_pfn)) {
1515                                        page = pageblock_pfn_to_page(min_pfn,
1516                                                min(pageblock_end_pfn(min_pfn),
1517                                                    zone_end_pfn(cc->zone)),
1518                                                cc->zone);
1519                                        cc->free_pfn = min_pfn;
1520                                }
1521                        }
1522                }
1523        }
1524
1525        if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1526                highest -= pageblock_nr_pages;
1527                cc->zone->compact_cached_free_pfn = highest;
1528        }
1529
1530        cc->total_free_scanned += nr_scanned;
1531        if (!page)
1532                return cc->free_pfn;
1533
1534        low_pfn = page_to_pfn(page);
1535        fast_isolate_around(cc, low_pfn, nr_isolated);
1536        return low_pfn;
1537}
1538
1539/*
1540 * Based on information in the current compact_control, find blocks
1541 * suitable for isolating free pages from and then isolate them.
1542 */
1543static void isolate_freepages(struct compact_control *cc)
1544{
1545        struct zone *zone = cc->zone;
1546        struct page *page;
1547        unsigned long block_start_pfn;  /* start of current pageblock */
1548        unsigned long isolate_start_pfn; /* exact pfn we start at */
1549        unsigned long block_end_pfn;    /* end of current pageblock */
1550        unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1551        struct list_head *freelist = &cc->freepages;
1552        unsigned int stride;
1553
1554        /* Try a small search of the free lists for a candidate */
1555        isolate_start_pfn = fast_isolate_freepages(cc);
1556        if (cc->nr_freepages)
1557                goto splitmap;
1558
1559        /*
1560         * Initialise the free scanner. The starting point is where we last
1561         * successfully isolated from, zone-cached value, or the end of the
1562         * zone when isolating for the first time. For looping we also need
1563         * this pfn aligned down to the pageblock boundary, because we do
1564         * block_start_pfn -= pageblock_nr_pages in the for loop.
1565         * For ending point, take care when isolating in last pageblock of a
1566         * zone which ends in the middle of a pageblock.
1567         * The low boundary is the end of the pageblock the migration scanner
1568         * is using.
1569         */
1570        isolate_start_pfn = cc->free_pfn;
1571        block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1572        block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1573                                                zone_end_pfn(zone));
1574        low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1575        stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1576
1577        /*
1578         * Isolate free pages until enough are available to migrate the
1579         * pages on cc->migratepages. We stop searching if the migrate
1580         * and free page scanners meet or enough free pages are isolated.
1581         */
1582        for (; block_start_pfn >= low_pfn;
1583                                block_end_pfn = block_start_pfn,
1584                                block_start_pfn -= pageblock_nr_pages,
1585                                isolate_start_pfn = block_start_pfn) {
1586                unsigned long nr_isolated;
1587
1588                /*
1589                 * This can iterate a massively long zone without finding any
1590                 * suitable migration targets, so periodically check resched.
1591                 */
1592                if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1593                        cond_resched();
1594
1595                page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1596                                                                        zone);
1597                if (!page)
1598                        continue;
1599
1600                /* Check the block is suitable for migration */
1601                if (!suitable_migration_target(cc, page))
1602                        continue;
1603
1604                /* If isolation recently failed, do not retry */
1605                if (!isolation_suitable(cc, page))
1606                        continue;
1607
1608                /* Found a block suitable for isolating free pages from. */
1609                nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1610                                        block_end_pfn, freelist, stride, false);
1611
1612                /* Update the skip hint if the full pageblock was scanned */
1613                if (isolate_start_pfn == block_end_pfn)
1614                        update_pageblock_skip(cc, page, block_start_pfn);
1615
1616                /* Are enough freepages isolated? */
1617                if (cc->nr_freepages >= cc->nr_migratepages) {
1618                        if (isolate_start_pfn >= block_end_pfn) {
1619                                /*
1620                                 * Restart at previous pageblock if more
1621                                 * freepages can be isolated next time.
1622                                 */
1623                                isolate_start_pfn =
1624                                        block_start_pfn - pageblock_nr_pages;
1625                        }
1626                        break;
1627                } else if (isolate_start_pfn < block_end_pfn) {
1628                        /*
1629                         * If isolation failed early, do not continue
1630                         * needlessly.
1631                         */
1632                        break;
1633                }
1634
1635                /* Adjust stride depending on isolation */
1636                if (nr_isolated) {
1637                        stride = 1;
1638                        continue;
1639                }
1640                stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1641        }
1642
1643        /*
1644         * Record where the free scanner will restart next time. Either we
1645         * broke from the loop and set isolate_start_pfn based on the last
1646         * call to isolate_freepages_block(), or we met the migration scanner
1647         * and the loop terminated due to isolate_start_pfn < low_pfn
1648         */
1649        cc->free_pfn = isolate_start_pfn;
1650
1651splitmap:
1652        /* __isolate_free_page() does not map the pages */
1653        split_map_pages(freelist);
1654}
1655
1656/*
1657 * This is a migrate-callback that "allocates" freepages by taking pages
1658 * from the isolated freelists in the block we are migrating to.
1659 */
1660static struct page *compaction_alloc(struct page *migratepage,
1661                                        unsigned long data)
1662{
1663        struct compact_control *cc = (struct compact_control *)data;
1664        struct page *freepage;
1665
1666        if (list_empty(&cc->freepages)) {
1667                isolate_freepages(cc);
1668
1669                if (list_empty(&cc->freepages))
1670                        return NULL;
1671        }
1672
1673        freepage = list_entry(cc->freepages.next, struct page, lru);
1674        list_del(&freepage->lru);
1675        cc->nr_freepages--;
1676
1677        return freepage;
1678}
1679
1680/*
1681 * This is a migrate-callback that "frees" freepages back to the isolated
1682 * freelist.  All pages on the freelist are from the same zone, so there is no
1683 * special handling needed for NUMA.
1684 */
1685static void compaction_free(struct page *page, unsigned long data)
1686{
1687        struct compact_control *cc = (struct compact_control *)data;
1688
1689        list_add(&page->lru, &cc->freepages);
1690        cc->nr_freepages++;
1691}
1692
1693/* possible outcome of isolate_migratepages */
1694typedef enum {
1695        ISOLATE_ABORT,          /* Abort compaction now */
1696        ISOLATE_NONE,           /* No pages isolated, continue scanning */
1697        ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1698} isolate_migrate_t;
1699
1700/*
1701 * Allow userspace to control policy on scanning the unevictable LRU for
1702 * compactable pages.
1703 */
1704#ifdef CONFIG_PREEMPT_RT
1705int sysctl_compact_unevictable_allowed __read_mostly = 0;
1706#else
1707int sysctl_compact_unevictable_allowed __read_mostly = 1;
1708#endif
1709
1710static inline void
1711update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1712{
1713        if (cc->fast_start_pfn == ULONG_MAX)
1714                return;
1715
1716        if (!cc->fast_start_pfn)
1717                cc->fast_start_pfn = pfn;
1718
1719        cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1720}
1721
1722static inline unsigned long
1723reinit_migrate_pfn(struct compact_control *cc)
1724{
1725        if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1726                return cc->migrate_pfn;
1727
1728        cc->migrate_pfn = cc->fast_start_pfn;
1729        cc->fast_start_pfn = ULONG_MAX;
1730
1731        return cc->migrate_pfn;
1732}
1733
1734/*
1735 * Briefly search the free lists for a migration source that already has
1736 * some free pages to reduce the number of pages that need migration
1737 * before a pageblock is free.
1738 */
1739static unsigned long fast_find_migrateblock(struct compact_control *cc)
1740{
1741        unsigned int limit = freelist_scan_limit(cc);
1742        unsigned int nr_scanned = 0;
1743        unsigned long distance;
1744        unsigned long pfn = cc->migrate_pfn;
1745        unsigned long high_pfn;
1746        int order;
1747        bool found_block = false;
1748
1749        /* Skip hints are relied on to avoid repeats on the fast search */
1750        if (cc->ignore_skip_hint)
1751                return pfn;
1752
1753        /*
1754         * If the migrate_pfn is not at the start of a zone or the start
1755         * of a pageblock then assume this is a continuation of a previous
1756         * scan restarted due to COMPACT_CLUSTER_MAX.
1757         */
1758        if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1759                return pfn;
1760
1761        /*
1762         * For smaller orders, just linearly scan as the number of pages
1763         * to migrate should be relatively small and does not necessarily
1764         * justify freeing up a large block for a small allocation.
1765         */
1766        if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1767                return pfn;
1768
1769        /*
1770         * Only allow kcompactd and direct requests for movable pages to
1771         * quickly clear out a MOVABLE pageblock for allocation. This
1772         * reduces the risk that a large movable pageblock is freed for
1773         * an unmovable/reclaimable small allocation.
1774         */
1775        if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1776                return pfn;
1777
1778        /*
1779         * When starting the migration scanner, pick any pageblock within the
1780         * first half of the search space. Otherwise try and pick a pageblock
1781         * within the first eighth to reduce the chances that a migration
1782         * target later becomes a source.
1783         */
1784        distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1785        if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1786                distance >>= 2;
1787        high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1788
1789        for (order = cc->order - 1;
1790             order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1791             order--) {
1792                struct free_area *area = &cc->zone->free_area[order];
1793                struct list_head *freelist;
1794                unsigned long flags;
1795                struct page *freepage;
1796
1797                if (!area->nr_free)
1798                        continue;
1799
1800                spin_lock_irqsave(&cc->zone->lock, flags);
1801                freelist = &area->free_list[MIGRATE_MOVABLE];
1802                list_for_each_entry(freepage, freelist, lru) {
1803                        unsigned long free_pfn;
1804
1805                        if (nr_scanned++ >= limit) {
1806                                move_freelist_tail(freelist, freepage);
1807                                break;
1808                        }
1809
1810                        free_pfn = page_to_pfn(freepage);
1811                        if (free_pfn < high_pfn) {
1812                                /*
1813                                 * Avoid if skipped recently. Ideally it would
1814                                 * move to the tail but even safe iteration of
1815                                 * the list assumes an entry is deleted, not
1816                                 * reordered.
1817                                 */
1818                                if (get_pageblock_skip(freepage))
1819                                        continue;
1820
1821                                /* Reorder to so a future search skips recent pages */
1822                                move_freelist_tail(freelist, freepage);
1823
1824                                update_fast_start_pfn(cc, free_pfn);
1825                                pfn = pageblock_start_pfn(free_pfn);
1826                                cc->fast_search_fail = 0;
1827                                found_block = true;
1828                                set_pageblock_skip(freepage);
1829                                break;
1830                        }
1831                }
1832                spin_unlock_irqrestore(&cc->zone->lock, flags);
1833        }
1834
1835        cc->total_migrate_scanned += nr_scanned;
1836
1837        /*
1838         * If fast scanning failed then use a cached entry for a page block
1839         * that had free pages as the basis for starting a linear scan.
1840         */
1841        if (!found_block) {
1842                cc->fast_search_fail++;
1843                pfn = reinit_migrate_pfn(cc);
1844        }
1845        return pfn;
1846}
1847
1848/*
1849 * Isolate all pages that can be migrated from the first suitable block,
1850 * starting at the block pointed to by the migrate scanner pfn within
1851 * compact_control.
1852 */
1853static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1854{
1855        unsigned long block_start_pfn;
1856        unsigned long block_end_pfn;
1857        unsigned long low_pfn;
1858        struct page *page;
1859        const isolate_mode_t isolate_mode =
1860                (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1861                (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1862        bool fast_find_block;
1863
1864        /*
1865         * Start at where we last stopped, or beginning of the zone as
1866         * initialized by compact_zone(). The first failure will use
1867         * the lowest PFN as the starting point for linear scanning.
1868         */
1869        low_pfn = fast_find_migrateblock(cc);
1870        block_start_pfn = pageblock_start_pfn(low_pfn);
1871        if (block_start_pfn < cc->zone->zone_start_pfn)
1872                block_start_pfn = cc->zone->zone_start_pfn;
1873
1874        /*
1875         * fast_find_migrateblock marks a pageblock skipped so to avoid
1876         * the isolation_suitable check below, check whether the fast
1877         * search was successful.
1878         */
1879        fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1880
1881        /* Only scan within a pageblock boundary */
1882        block_end_pfn = pageblock_end_pfn(low_pfn);
1883
1884        /*
1885         * Iterate over whole pageblocks until we find the first suitable.
1886         * Do not cross the free scanner.
1887         */
1888        for (; block_end_pfn <= cc->free_pfn;
1889                        fast_find_block = false,
1890                        cc->migrate_pfn = low_pfn = block_end_pfn,
1891                        block_start_pfn = block_end_pfn,
1892                        block_end_pfn += pageblock_nr_pages) {
1893
1894                /*
1895                 * This can potentially iterate a massively long zone with
1896                 * many pageblocks unsuitable, so periodically check if we
1897                 * need to schedule.
1898                 */
1899                if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1900                        cond_resched();
1901
1902                page = pageblock_pfn_to_page(block_start_pfn,
1903                                                block_end_pfn, cc->zone);
1904                if (!page)
1905                        continue;
1906
1907                /*
1908                 * If isolation recently failed, do not retry. Only check the
1909                 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1910                 * to be visited multiple times. Assume skip was checked
1911                 * before making it "skip" so other compaction instances do
1912                 * not scan the same block.
1913                 */
1914                if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1915                    !fast_find_block && !isolation_suitable(cc, page))
1916                        continue;
1917
1918                /*
1919                 * For async compaction, also only scan in MOVABLE blocks
1920                 * without huge pages. Async compaction is optimistic to see
1921                 * if the minimum amount of work satisfies the allocation.
1922                 * The cached PFN is updated as it's possible that all
1923                 * remaining blocks between source and target are unsuitable
1924                 * and the compaction scanners fail to meet.
1925                 */
1926                if (!suitable_migration_source(cc, page)) {
1927                        update_cached_migrate(cc, block_end_pfn);
1928                        continue;
1929                }
1930
1931                /* Perform the isolation */
1932                if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
1933                                                isolate_mode))
1934                        return ISOLATE_ABORT;
1935
1936                /*
1937                 * Either we isolated something and proceed with migration. Or
1938                 * we failed and compact_zone should decide if we should
1939                 * continue or not.
1940                 */
1941                break;
1942        }
1943
1944        return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1945}
1946
1947/*
1948 * order == -1 is expected when compacting via
1949 * /proc/sys/vm/compact_memory
1950 */
1951static inline bool is_via_compact_memory(int order)
1952{
1953        return order == -1;
1954}
1955
1956static bool kswapd_is_running(pg_data_t *pgdat)
1957{
1958        return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1959}
1960
1961/*
1962 * A zone's fragmentation score is the external fragmentation wrt to the
1963 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
1964 */
1965static unsigned int fragmentation_score_zone(struct zone *zone)
1966{
1967        return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1968}
1969
1970/*
1971 * A weighted zone's fragmentation score is the external fragmentation
1972 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
1973 * returns a value in the range [0, 100].
1974 *
1975 * The scaling factor ensures that proactive compaction focuses on larger
1976 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1977 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1978 * and thus never exceeds the high threshold for proactive compaction.
1979 */
1980static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
1981{
1982        unsigned long score;
1983
1984        score = zone->present_pages * fragmentation_score_zone(zone);
1985        return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1986}
1987
1988/*
1989 * The per-node proactive (background) compaction process is started by its
1990 * corresponding kcompactd thread when the node's fragmentation score
1991 * exceeds the high threshold. The compaction process remains active till
1992 * the node's score falls below the low threshold, or one of the back-off
1993 * conditions is met.
1994 */
1995static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1996{
1997        unsigned int score = 0;
1998        int zoneid;
1999
2000        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2001                struct zone *zone;
2002
2003                zone = &pgdat->node_zones[zoneid];
2004                score += fragmentation_score_zone_weighted(zone);
2005        }
2006
2007        return score;
2008}
2009
2010static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
2011{
2012        unsigned int wmark_low;
2013
2014        /*
2015         * Cap the low watermark to avoid excessive compaction
2016         * activity in case a user sets the proactiveness tunable
2017         * close to 100 (maximum).
2018         */
2019        wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2020        return low ? wmark_low : min(wmark_low + 10, 100U);
2021}
2022
2023static bool should_proactive_compact_node(pg_data_t *pgdat)
2024{
2025        int wmark_high;
2026
2027        if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2028                return false;
2029
2030        wmark_high = fragmentation_score_wmark(pgdat, false);
2031        return fragmentation_score_node(pgdat) > wmark_high;
2032}
2033
2034static enum compact_result __compact_finished(struct compact_control *cc)
2035{
2036        unsigned int order;
2037        const int migratetype = cc->migratetype;
2038        int ret;
2039
2040        /* Compaction run completes if the migrate and free scanner meet */
2041        if (compact_scanners_met(cc)) {
2042                /* Let the next compaction start anew. */
2043                reset_cached_positions(cc->zone);
2044
2045                /*
2046                 * Mark that the PG_migrate_skip information should be cleared
2047                 * by kswapd when it goes to sleep. kcompactd does not set the
2048                 * flag itself as the decision to be clear should be directly
2049                 * based on an allocation request.
2050                 */
2051                if (cc->direct_compaction)
2052                        cc->zone->compact_blockskip_flush = true;
2053
2054                if (cc->whole_zone)
2055                        return COMPACT_COMPLETE;
2056                else
2057                        return COMPACT_PARTIAL_SKIPPED;
2058        }
2059
2060        if (cc->proactive_compaction) {
2061                int score, wmark_low;
2062                pg_data_t *pgdat;
2063
2064                pgdat = cc->zone->zone_pgdat;
2065                if (kswapd_is_running(pgdat))
2066                        return COMPACT_PARTIAL_SKIPPED;
2067
2068                score = fragmentation_score_zone(cc->zone);
2069                wmark_low = fragmentation_score_wmark(pgdat, true);
2070
2071                if (score > wmark_low)
2072                        ret = COMPACT_CONTINUE;
2073                else
2074                        ret = COMPACT_SUCCESS;
2075
2076                goto out;
2077        }
2078
2079        if (is_via_compact_memory(cc->order))
2080                return COMPACT_CONTINUE;
2081
2082        /*
2083         * Always finish scanning a pageblock to reduce the possibility of
2084         * fallbacks in the future. This is particularly important when
2085         * migration source is unmovable/reclaimable but it's not worth
2086         * special casing.
2087         */
2088        if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2089                return COMPACT_CONTINUE;
2090
2091        /* Direct compactor: Is a suitable page free? */
2092        ret = COMPACT_NO_SUITABLE_PAGE;
2093        for (order = cc->order; order < MAX_ORDER; order++) {
2094                struct free_area *area = &cc->zone->free_area[order];
2095                bool can_steal;
2096
2097                /* Job done if page is free of the right migratetype */
2098                if (!free_area_empty(area, migratetype))
2099                        return COMPACT_SUCCESS;
2100
2101#ifdef CONFIG_CMA
2102                /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2103                if (migratetype == MIGRATE_MOVABLE &&
2104                        !free_area_empty(area, MIGRATE_CMA))
2105                        return COMPACT_SUCCESS;
2106#endif
2107                /*
2108                 * Job done if allocation would steal freepages from
2109                 * other migratetype buddy lists.
2110                 */
2111                if (find_suitable_fallback(area, order, migratetype,
2112                                                true, &can_steal) != -1) {
2113
2114                        /* movable pages are OK in any pageblock */
2115                        if (migratetype == MIGRATE_MOVABLE)
2116                                return COMPACT_SUCCESS;
2117
2118                        /*
2119                         * We are stealing for a non-movable allocation. Make
2120                         * sure we finish compacting the current pageblock
2121                         * first so it is as free as possible and we won't
2122                         * have to steal another one soon. This only applies
2123                         * to sync compaction, as async compaction operates
2124                         * on pageblocks of the same migratetype.
2125                         */
2126                        if (cc->mode == MIGRATE_ASYNC ||
2127                                        IS_ALIGNED(cc->migrate_pfn,
2128                                                        pageblock_nr_pages)) {
2129                                return COMPACT_SUCCESS;
2130                        }
2131
2132                        ret = COMPACT_CONTINUE;
2133                        break;
2134                }
2135        }
2136
2137out:
2138        if (cc->contended || fatal_signal_pending(current))
2139                ret = COMPACT_CONTENDED;
2140
2141        return ret;
2142}
2143
2144static enum compact_result compact_finished(struct compact_control *cc)
2145{
2146        int ret;
2147
2148        ret = __compact_finished(cc);
2149        trace_mm_compaction_finished(cc->zone, cc->order, ret);
2150        if (ret == COMPACT_NO_SUITABLE_PAGE)
2151                ret = COMPACT_CONTINUE;
2152
2153        return ret;
2154}
2155
2156static enum compact_result __compaction_suitable(struct zone *zone, int order,
2157                                        unsigned int alloc_flags,
2158                                        int highest_zoneidx,
2159                                        unsigned long wmark_target)
2160{
2161        unsigned long watermark;
2162
2163        if (is_via_compact_memory(order))
2164                return COMPACT_CONTINUE;
2165
2166        watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2167        /*
2168         * If watermarks for high-order allocation are already met, there
2169         * should be no need for compaction at all.
2170         */
2171        if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2172                                                                alloc_flags))
2173                return COMPACT_SUCCESS;
2174
2175        /*
2176         * Watermarks for order-0 must be met for compaction to be able to
2177         * isolate free pages for migration targets. This means that the
2178         * watermark and alloc_flags have to match, or be more pessimistic than
2179         * the check in __isolate_free_page(). We don't use the direct
2180         * compactor's alloc_flags, as they are not relevant for freepage
2181         * isolation. We however do use the direct compactor's highest_zoneidx
2182         * to skip over zones where lowmem reserves would prevent allocation
2183         * even if compaction succeeds.
2184         * For costly orders, we require low watermark instead of min for
2185         * compaction to proceed to increase its chances.
2186         * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2187         * suitable migration targets
2188         */
2189        watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2190                                low_wmark_pages(zone) : min_wmark_pages(zone);
2191        watermark += compact_gap(order);
2192        if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2193                                                ALLOC_CMA, wmark_target))
2194                return COMPACT_SKIPPED;
2195
2196        return COMPACT_CONTINUE;
2197}
2198
2199/*
2200 * compaction_suitable: Is this suitable to run compaction on this zone now?
2201 * Returns
2202 *   COMPACT_SKIPPED  - If there are too few free pages for compaction
2203 *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
2204 *   COMPACT_CONTINUE - If compaction should run now
2205 */
2206enum compact_result compaction_suitable(struct zone *zone, int order,
2207                                        unsigned int alloc_flags,
2208                                        int highest_zoneidx)
2209{
2210        enum compact_result ret;
2211        int fragindex;
2212
2213        ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2214                                    zone_page_state(zone, NR_FREE_PAGES));
2215        /*
2216         * fragmentation index determines if allocation failures are due to
2217         * low memory or external fragmentation
2218         *
2219         * index of -1000 would imply allocations might succeed depending on
2220         * watermarks, but we already failed the high-order watermark check
2221         * index towards 0 implies failure is due to lack of memory
2222         * index towards 1000 implies failure is due to fragmentation
2223         *
2224         * Only compact if a failure would be due to fragmentation. Also
2225         * ignore fragindex for non-costly orders where the alternative to
2226         * a successful reclaim/compaction is OOM. Fragindex and the
2227         * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2228         * excessive compaction for costly orders, but it should not be at the
2229         * expense of system stability.
2230         */
2231        if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2232                fragindex = fragmentation_index(zone, order);
2233                if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2234                        ret = COMPACT_NOT_SUITABLE_ZONE;
2235        }
2236
2237        trace_mm_compaction_suitable(zone, order, ret);
2238        if (ret == COMPACT_NOT_SUITABLE_ZONE)
2239                ret = COMPACT_SKIPPED;
2240
2241        return ret;
2242}
2243
2244bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2245                int alloc_flags)
2246{
2247        struct zone *zone;
2248        struct zoneref *z;
2249
2250        /*
2251         * Make sure at least one zone would pass __compaction_suitable if we continue
2252         * retrying the reclaim.
2253         */
2254        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2255                                ac->highest_zoneidx, ac->nodemask) {
2256                unsigned long available;
2257                enum compact_result compact_result;
2258
2259                /*
2260                 * Do not consider all the reclaimable memory because we do not
2261                 * want to trash just for a single high order allocation which
2262                 * is even not guaranteed to appear even if __compaction_suitable
2263                 * is happy about the watermark check.
2264                 */
2265                available = zone_reclaimable_pages(zone) / order;
2266                available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2267                compact_result = __compaction_suitable(zone, order, alloc_flags,
2268                                ac->highest_zoneidx, available);
2269                if (compact_result != COMPACT_SKIPPED)
2270                        return true;
2271        }
2272
2273        return false;
2274}
2275
2276static enum compact_result
2277compact_zone(struct compact_control *cc, struct capture_control *capc)
2278{
2279        enum compact_result ret;
2280        unsigned long start_pfn = cc->zone->zone_start_pfn;
2281        unsigned long end_pfn = zone_end_pfn(cc->zone);
2282        unsigned long last_migrated_pfn;
2283        const bool sync = cc->mode != MIGRATE_ASYNC;
2284        bool update_cached;
2285
2286        /*
2287         * These counters track activities during zone compaction.  Initialize
2288         * them before compacting a new zone.
2289         */
2290        cc->total_migrate_scanned = 0;
2291        cc->total_free_scanned = 0;
2292        cc->nr_migratepages = 0;
2293        cc->nr_freepages = 0;
2294        INIT_LIST_HEAD(&cc->freepages);
2295        INIT_LIST_HEAD(&cc->migratepages);
2296
2297        cc->migratetype = gfp_migratetype(cc->gfp_mask);
2298        ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2299                                                        cc->highest_zoneidx);
2300        /* Compaction is likely to fail */
2301        if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2302                return ret;
2303
2304        /* huh, compaction_suitable is returning something unexpected */
2305        VM_BUG_ON(ret != COMPACT_CONTINUE);
2306
2307        /*
2308         * Clear pageblock skip if there were failures recently and compaction
2309         * is about to be retried after being deferred.
2310         */
2311        if (compaction_restarting(cc->zone, cc->order))
2312                __reset_isolation_suitable(cc->zone);
2313
2314        /*
2315         * Setup to move all movable pages to the end of the zone. Used cached
2316         * information on where the scanners should start (unless we explicitly
2317         * want to compact the whole zone), but check that it is initialised
2318         * by ensuring the values are within zone boundaries.
2319         */
2320        cc->fast_start_pfn = 0;
2321        if (cc->whole_zone) {
2322                cc->migrate_pfn = start_pfn;
2323                cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2324        } else {
2325                cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2326                cc->free_pfn = cc->zone->compact_cached_free_pfn;
2327                if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2328                        cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2329                        cc->zone->compact_cached_free_pfn = cc->free_pfn;
2330                }
2331                if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2332                        cc->migrate_pfn = start_pfn;
2333                        cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2334                        cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2335                }
2336
2337                if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2338                        cc->whole_zone = true;
2339        }
2340
2341        last_migrated_pfn = 0;
2342
2343        /*
2344         * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2345         * the basis that some migrations will fail in ASYNC mode. However,
2346         * if the cached PFNs match and pageblocks are skipped due to having
2347         * no isolation candidates, then the sync state does not matter.
2348         * Until a pageblock with isolation candidates is found, keep the
2349         * cached PFNs in sync to avoid revisiting the same blocks.
2350         */
2351        update_cached = !sync &&
2352                cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2353
2354        trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2355                                cc->free_pfn, end_pfn, sync);
2356
2357        /* lru_add_drain_all could be expensive with involving other CPUs */
2358        lru_add_drain();
2359
2360        while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2361                int err;
2362                unsigned long iteration_start_pfn = cc->migrate_pfn;
2363
2364                /*
2365                 * Avoid multiple rescans which can happen if a page cannot be
2366                 * isolated (dirty/writeback in async mode) or if the migrated
2367                 * pages are being allocated before the pageblock is cleared.
2368                 * The first rescan will capture the entire pageblock for
2369                 * migration. If it fails, it'll be marked skip and scanning
2370                 * will proceed as normal.
2371                 */
2372                cc->rescan = false;
2373                if (pageblock_start_pfn(last_migrated_pfn) ==
2374                    pageblock_start_pfn(iteration_start_pfn)) {
2375                        cc->rescan = true;
2376                }
2377
2378                switch (isolate_migratepages(cc)) {
2379                case ISOLATE_ABORT:
2380                        ret = COMPACT_CONTENDED;
2381                        putback_movable_pages(&cc->migratepages);
2382                        cc->nr_migratepages = 0;
2383                        goto out;
2384                case ISOLATE_NONE:
2385                        if (update_cached) {
2386                                cc->zone->compact_cached_migrate_pfn[1] =
2387                                        cc->zone->compact_cached_migrate_pfn[0];
2388                        }
2389
2390                        /*
2391                         * We haven't isolated and migrated anything, but
2392                         * there might still be unflushed migrations from
2393                         * previous cc->order aligned block.
2394                         */
2395                        goto check_drain;
2396                case ISOLATE_SUCCESS:
2397                        update_cached = false;
2398                        last_migrated_pfn = iteration_start_pfn;
2399                }
2400
2401                err = migrate_pages(&cc->migratepages, compaction_alloc,
2402                                compaction_free, (unsigned long)cc, cc->mode,
2403                                MR_COMPACTION);
2404
2405                trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2406                                                        &cc->migratepages);
2407
2408                /* All pages were either migrated or will be released */
2409                cc->nr_migratepages = 0;
2410                if (err) {
2411                        putback_movable_pages(&cc->migratepages);
2412                        /*
2413                         * migrate_pages() may return -ENOMEM when scanners meet
2414                         * and we want compact_finished() to detect it
2415                         */
2416                        if (err == -ENOMEM && !compact_scanners_met(cc)) {
2417                                ret = COMPACT_CONTENDED;
2418                                goto out;
2419                        }
2420                        /*
2421                         * We failed to migrate at least one page in the current
2422                         * order-aligned block, so skip the rest of it.
2423                         */
2424                        if (cc->direct_compaction &&
2425                                                (cc->mode == MIGRATE_ASYNC)) {
2426                                cc->migrate_pfn = block_end_pfn(
2427                                                cc->migrate_pfn - 1, cc->order);
2428                                /* Draining pcplists is useless in this case */
2429                                last_migrated_pfn = 0;
2430                        }
2431                }
2432
2433check_drain:
2434                /*
2435                 * Has the migration scanner moved away from the previous
2436                 * cc->order aligned block where we migrated from? If yes,
2437                 * flush the pages that were freed, so that they can merge and
2438                 * compact_finished() can detect immediately if allocation
2439                 * would succeed.
2440                 */
2441                if (cc->order > 0 && last_migrated_pfn) {
2442                        unsigned long current_block_start =
2443                                block_start_pfn(cc->migrate_pfn, cc->order);
2444
2445                        if (last_migrated_pfn < current_block_start) {
2446                                lru_add_drain_cpu_zone(cc->zone);
2447                                /* No more flushing until we migrate again */
2448                                last_migrated_pfn = 0;
2449                        }
2450                }
2451
2452                /* Stop if a page has been captured */
2453                if (capc && capc->page) {
2454                        ret = COMPACT_SUCCESS;
2455                        break;
2456                }
2457        }
2458
2459out:
2460        /*
2461         * Release free pages and update where the free scanner should restart,
2462         * so we don't leave any returned pages behind in the next attempt.
2463         */
2464        if (cc->nr_freepages > 0) {
2465                unsigned long free_pfn = release_freepages(&cc->freepages);
2466
2467                cc->nr_freepages = 0;
2468                VM_BUG_ON(free_pfn == 0);
2469                /* The cached pfn is always the first in a pageblock */
2470                free_pfn = pageblock_start_pfn(free_pfn);
2471                /*
2472                 * Only go back, not forward. The cached pfn might have been
2473                 * already reset to zone end in compact_finished()
2474                 */
2475                if (free_pfn > cc->zone->compact_cached_free_pfn)
2476                        cc->zone->compact_cached_free_pfn = free_pfn;
2477        }
2478
2479        count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2480        count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2481
2482        trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2483                                cc->free_pfn, end_pfn, sync, ret);
2484
2485        return ret;
2486}
2487
2488static enum compact_result compact_zone_order(struct zone *zone, int order,
2489                gfp_t gfp_mask, enum compact_priority prio,
2490                unsigned int alloc_flags, int highest_zoneidx,
2491                struct page **capture)
2492{
2493        enum compact_result ret;
2494        struct compact_control cc = {
2495                .order = order,
2496                .search_order = order,
2497                .gfp_mask = gfp_mask,
2498                .zone = zone,
2499                .mode = (prio == COMPACT_PRIO_ASYNC) ?
2500                                        MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2501                .alloc_flags = alloc_flags,
2502                .highest_zoneidx = highest_zoneidx,
2503                .direct_compaction = true,
2504                .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2505                .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2506                .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2507        };
2508        struct capture_control capc = {
2509                .cc = &cc,
2510                .page = NULL,
2511        };
2512
2513        /*
2514         * Make sure the structs are really initialized before we expose the
2515         * capture control, in case we are interrupted and the interrupt handler
2516         * frees a page.
2517         */
2518        barrier();
2519        WRITE_ONCE(current->capture_control, &capc);
2520
2521        ret = compact_zone(&cc, &capc);
2522
2523        VM_BUG_ON(!list_empty(&cc.freepages));
2524        VM_BUG_ON(!list_empty(&cc.migratepages));
2525
2526        /*
2527         * Make sure we hide capture control first before we read the captured
2528         * page pointer, otherwise an interrupt could free and capture a page
2529         * and we would leak it.
2530         */
2531        WRITE_ONCE(current->capture_control, NULL);
2532        *capture = READ_ONCE(capc.page);
2533        /*
2534         * Technically, it is also possible that compaction is skipped but
2535         * the page is still captured out of luck(IRQ came and freed the page).
2536         * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2537         * the COMPACT[STALL|FAIL] when compaction is skipped.
2538         */
2539        if (*capture)
2540                ret = COMPACT_SUCCESS;
2541
2542        return ret;
2543}
2544
2545int sysctl_extfrag_threshold = 500;
2546
2547/**
2548 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2549 * @gfp_mask: The GFP mask of the current allocation
2550 * @order: The order of the current allocation
2551 * @alloc_flags: The allocation flags of the current allocation
2552 * @ac: The context of current allocation
2553 * @prio: Determines how hard direct compaction should try to succeed
2554 * @capture: Pointer to free page created by compaction will be stored here
2555 *
2556 * This is the main entry point for direct page compaction.
2557 */
2558enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2559                unsigned int alloc_flags, const struct alloc_context *ac,
2560                enum compact_priority prio, struct page **capture)
2561{
2562        int may_perform_io = gfp_mask & __GFP_IO;
2563        struct zoneref *z;
2564        struct zone *zone;
2565        enum compact_result rc = COMPACT_SKIPPED;
2566
2567        /*
2568         * Check if the GFP flags allow compaction - GFP_NOIO is really
2569         * tricky context because the migration might require IO
2570         */
2571        if (!may_perform_io)
2572                return COMPACT_SKIPPED;
2573
2574        trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2575
2576        /* Compact each zone in the list */
2577        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2578                                        ac->highest_zoneidx, ac->nodemask) {
2579                enum compact_result status;
2580
2581                if (prio > MIN_COMPACT_PRIORITY
2582                                        && compaction_deferred(zone, order)) {
2583                        rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2584                        continue;
2585                }
2586
2587                status = compact_zone_order(zone, order, gfp_mask, prio,
2588                                alloc_flags, ac->highest_zoneidx, capture);
2589                rc = max(status, rc);
2590
2591                /* The allocation should succeed, stop compacting */
2592                if (status == COMPACT_SUCCESS) {
2593                        /*
2594                         * We think the allocation will succeed in this zone,
2595                         * but it is not certain, hence the false. The caller
2596                         * will repeat this with true if allocation indeed
2597                         * succeeds in this zone.
2598                         */
2599                        compaction_defer_reset(zone, order, false);
2600
2601                        break;
2602                }
2603
2604                if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2605                                        status == COMPACT_PARTIAL_SKIPPED))
2606                        /*
2607                         * We think that allocation won't succeed in this zone
2608                         * so we defer compaction there. If it ends up
2609                         * succeeding after all, it will be reset.
2610                         */
2611                        defer_compaction(zone, order);
2612
2613                /*
2614                 * We might have stopped compacting due to need_resched() in
2615                 * async compaction, or due to a fatal signal detected. In that
2616                 * case do not try further zones
2617                 */
2618                if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2619                                        || fatal_signal_pending(current))
2620                        break;
2621        }
2622
2623        return rc;
2624}
2625
2626/*
2627 * Compact all zones within a node till each zone's fragmentation score
2628 * reaches within proactive compaction thresholds (as determined by the
2629 * proactiveness tunable).
2630 *
2631 * It is possible that the function returns before reaching score targets
2632 * due to various back-off conditions, such as, contention on per-node or
2633 * per-zone locks.
2634 */
2635static void proactive_compact_node(pg_data_t *pgdat)
2636{
2637        int zoneid;
2638        struct zone *zone;
2639        struct compact_control cc = {
2640                .order = -1,
2641                .mode = MIGRATE_SYNC_LIGHT,
2642                .ignore_skip_hint = true,
2643                .whole_zone = true,
2644                .gfp_mask = GFP_KERNEL,
2645                .proactive_compaction = true,
2646        };
2647
2648        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2649                zone = &pgdat->node_zones[zoneid];
2650                if (!populated_zone(zone))
2651                        continue;
2652
2653                cc.zone = zone;
2654
2655                compact_zone(&cc, NULL);
2656
2657                VM_BUG_ON(!list_empty(&cc.freepages));
2658                VM_BUG_ON(!list_empty(&cc.migratepages));
2659        }
2660}
2661
2662/* Compact all zones within a node */
2663static void compact_node(int nid)
2664{
2665        pg_data_t *pgdat = NODE_DATA(nid);
2666        int zoneid;
2667        struct zone *zone;
2668        struct compact_control cc = {
2669                .order = -1,
2670                .mode = MIGRATE_SYNC,
2671                .ignore_skip_hint = true,
2672                .whole_zone = true,
2673                .gfp_mask = GFP_KERNEL,
2674        };
2675
2676
2677        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2678
2679                zone = &pgdat->node_zones[zoneid];
2680                if (!populated_zone(zone))
2681                        continue;
2682
2683                cc.zone = zone;
2684
2685                compact_zone(&cc, NULL);
2686
2687                VM_BUG_ON(!list_empty(&cc.freepages));
2688                VM_BUG_ON(!list_empty(&cc.migratepages));
2689        }
2690}
2691
2692/* Compact all nodes in the system */
2693static void compact_nodes(void)
2694{
2695        int nid;
2696
2697        /* Flush pending updates to the LRU lists */
2698        lru_add_drain_all();
2699
2700        for_each_online_node(nid)
2701                compact_node(nid);
2702}
2703
2704/*
2705 * Tunable for proactive compaction. It determines how
2706 * aggressively the kernel should compact memory in the
2707 * background. It takes values in the range [0, 100].
2708 */
2709unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2710
2711/*
2712 * This is the entry point for compacting all nodes via
2713 * /proc/sys/vm/compact_memory
2714 */
2715int sysctl_compaction_handler(struct ctl_table *table, int write,
2716                        void *buffer, size_t *length, loff_t *ppos)
2717{
2718        if (write)
2719                compact_nodes();
2720
2721        return 0;
2722}
2723
2724#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2725static ssize_t sysfs_compact_node(struct device *dev,
2726                        struct device_attribute *attr,
2727                        const char *buf, size_t count)
2728{
2729        int nid = dev->id;
2730
2731        if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2732                /* Flush pending updates to the LRU lists */
2733                lru_add_drain_all();
2734
2735                compact_node(nid);
2736        }
2737
2738        return count;
2739}
2740static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2741
2742int compaction_register_node(struct node *node)
2743{
2744        return device_create_file(&node->dev, &dev_attr_compact);
2745}
2746
2747void compaction_unregister_node(struct node *node)
2748{
2749        return device_remove_file(&node->dev, &dev_attr_compact);
2750}
2751#endif /* CONFIG_SYSFS && CONFIG_NUMA */
2752
2753static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2754{
2755        return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2756}
2757
2758static bool kcompactd_node_suitable(pg_data_t *pgdat)
2759{
2760        int zoneid;
2761        struct zone *zone;
2762        enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2763
2764        for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2765                zone = &pgdat->node_zones[zoneid];
2766
2767                if (!populated_zone(zone))
2768                        continue;
2769
2770                if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2771                                        highest_zoneidx) == COMPACT_CONTINUE)
2772                        return true;
2773        }
2774
2775        return false;
2776}
2777
2778static void kcompactd_do_work(pg_data_t *pgdat)
2779{
2780        /*
2781         * With no special task, compact all zones so that a page of requested
2782         * order is allocatable.
2783         */
2784        int zoneid;
2785        struct zone *zone;
2786        struct compact_control cc = {
2787                .order = pgdat->kcompactd_max_order,
2788                .search_order = pgdat->kcompactd_max_order,
2789                .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2790                .mode = MIGRATE_SYNC_LIGHT,
2791                .ignore_skip_hint = false,
2792                .gfp_mask = GFP_KERNEL,
2793        };
2794        trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2795                                                        cc.highest_zoneidx);
2796        count_compact_event(KCOMPACTD_WAKE);
2797
2798        for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2799                int status;
2800
2801                zone = &pgdat->node_zones[zoneid];
2802                if (!populated_zone(zone))
2803                        continue;
2804
2805                if (compaction_deferred(zone, cc.order))
2806                        continue;
2807
2808                if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2809                                                        COMPACT_CONTINUE)
2810                        continue;
2811
2812                if (kthread_should_stop())
2813                        return;
2814
2815                cc.zone = zone;
2816                status = compact_zone(&cc, NULL);
2817
2818                if (status == COMPACT_SUCCESS) {
2819                        compaction_defer_reset(zone, cc.order, false);
2820                } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2821                        /*
2822                         * Buddy pages may become stranded on pcps that could
2823                         * otherwise coalesce on the zone's free area for
2824                         * order >= cc.order.  This is ratelimited by the
2825                         * upcoming deferral.
2826                         */
2827                        drain_all_pages(zone);
2828
2829                        /*
2830                         * We use sync migration mode here, so we defer like
2831                         * sync direct compaction does.
2832                         */
2833                        defer_compaction(zone, cc.order);
2834                }
2835
2836                count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2837                                     cc.total_migrate_scanned);
2838                count_compact_events(KCOMPACTD_FREE_SCANNED,
2839                                     cc.total_free_scanned);
2840
2841                VM_BUG_ON(!list_empty(&cc.freepages));
2842                VM_BUG_ON(!list_empty(&cc.migratepages));
2843        }
2844
2845        /*
2846         * Regardless of success, we are done until woken up next. But remember
2847         * the requested order/highest_zoneidx in case it was higher/tighter
2848         * than our current ones
2849         */
2850        if (pgdat->kcompactd_max_order <= cc.order)
2851                pgdat->kcompactd_max_order = 0;
2852        if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2853                pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2854}
2855
2856void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2857{
2858        if (!order)
2859                return;
2860
2861        if (pgdat->kcompactd_max_order < order)
2862                pgdat->kcompactd_max_order = order;
2863
2864        if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2865                pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2866
2867        /*
2868         * Pairs with implicit barrier in wait_event_freezable()
2869         * such that wakeups are not missed.
2870         */
2871        if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2872                return;
2873
2874        if (!kcompactd_node_suitable(pgdat))
2875                return;
2876
2877        trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2878                                                        highest_zoneidx);
2879        wake_up_interruptible(&pgdat->kcompactd_wait);
2880}
2881
2882/*
2883 * The background compaction daemon, started as a kernel thread
2884 * from the init process.
2885 */
2886static int kcompactd(void *p)
2887{
2888        pg_data_t *pgdat = (pg_data_t *)p;
2889        struct task_struct *tsk = current;
2890        unsigned int proactive_defer = 0;
2891
2892        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2893
2894        if (!cpumask_empty(cpumask))
2895                set_cpus_allowed_ptr(tsk, cpumask);
2896
2897        set_freezable();
2898
2899        pgdat->kcompactd_max_order = 0;
2900        pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2901
2902        while (!kthread_should_stop()) {
2903                unsigned long pflags;
2904
2905                trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2906                if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2907                        kcompactd_work_requested(pgdat),
2908                        msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2909
2910                        psi_memstall_enter(&pflags);
2911                        kcompactd_do_work(pgdat);
2912                        psi_memstall_leave(&pflags);
2913                        continue;
2914                }
2915
2916                /* kcompactd wait timeout */
2917                if (should_proactive_compact_node(pgdat)) {
2918                        unsigned int prev_score, score;
2919
2920                        if (proactive_defer) {
2921                                proactive_defer--;
2922                                continue;
2923                        }
2924                        prev_score = fragmentation_score_node(pgdat);
2925                        proactive_compact_node(pgdat);
2926                        score = fragmentation_score_node(pgdat);
2927                        /*
2928                         * Defer proactive compaction if the fragmentation
2929                         * score did not go down i.e. no progress made.
2930                         */
2931                        proactive_defer = score < prev_score ?
2932                                        0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2933                }
2934        }
2935
2936        return 0;
2937}
2938
2939/*
2940 * This kcompactd start function will be called by init and node-hot-add.
2941 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2942 */
2943int kcompactd_run(int nid)
2944{
2945        pg_data_t *pgdat = NODE_DATA(nid);
2946        int ret = 0;
2947
2948        if (pgdat->kcompactd)
2949                return 0;
2950
2951        pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2952        if (IS_ERR(pgdat->kcompactd)) {
2953                pr_err("Failed to start kcompactd on node %d\n", nid);
2954                ret = PTR_ERR(pgdat->kcompactd);
2955                pgdat->kcompactd = NULL;
2956        }
2957        return ret;
2958}
2959
2960/*
2961 * Called by memory hotplug when all memory in a node is offlined. Caller must
2962 * hold mem_hotplug_begin/end().
2963 */
2964void kcompactd_stop(int nid)
2965{
2966        struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2967
2968        if (kcompactd) {
2969                kthread_stop(kcompactd);
2970                NODE_DATA(nid)->kcompactd = NULL;
2971        }
2972}
2973
2974/*
2975 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2976 * not required for correctness. So if the last cpu in a node goes
2977 * away, we get changed to run anywhere: as the first one comes back,
2978 * restore their cpu bindings.
2979 */
2980static int kcompactd_cpu_online(unsigned int cpu)
2981{
2982        int nid;
2983
2984        for_each_node_state(nid, N_MEMORY) {
2985                pg_data_t *pgdat = NODE_DATA(nid);
2986                const struct cpumask *mask;
2987
2988                mask = cpumask_of_node(pgdat->node_id);
2989
2990                if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2991                        /* One of our CPUs online: restore mask */
2992                        set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2993        }
2994        return 0;
2995}
2996
2997static int __init kcompactd_init(void)
2998{
2999        int nid;
3000        int ret;
3001
3002        ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3003                                        "mm/compaction:online",
3004                                        kcompactd_cpu_online, NULL);
3005        if (ret < 0) {
3006                pr_err("kcompactd: failed to register hotplug callbacks.\n");
3007                return ret;
3008        }
3009
3010        for_each_node_state(nid, N_MEMORY)
3011                kcompactd_run(nid);
3012        return 0;
3013}
3014subsys_initcall(kcompactd_init)
3015
3016#endif /* CONFIG_COMPACTION */
3017