linux/mm/compaction.c
<<
>>
Prefs
   1/*
   2 * linux/mm/compaction.c
   3 *
   4 * Memory compaction for the reduction of external fragmentation. Note that
   5 * this heavily depends upon page migration to do all the real heavy
   6 * lifting
   7 *
   8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
   9 */
  10#include <linux/cpu.h>
  11#include <linux/swap.h>
  12#include <linux/migrate.h>
  13#include <linux/compaction.h>
  14#include <linux/mm_inline.h>
  15#include <linux/backing-dev.h>
  16#include <linux/sysctl.h>
  17#include <linux/sysfs.h>
  18#include <linux/page-isolation.h>
  19#include <linux/kasan.h>
  20#include <linux/kthread.h>
  21#include <linux/freezer.h>
  22#include <linux/page_owner.h>
  23#include "internal.h"
  24
  25#ifdef CONFIG_COMPACTION
  26static inline void count_compact_event(enum vm_event_item item)
  27{
  28        count_vm_event(item);
  29}
  30
  31static inline void count_compact_events(enum vm_event_item item, long delta)
  32{
  33        count_vm_events(item, delta);
  34}
  35#else
  36#define count_compact_event(item) do { } while (0)
  37#define count_compact_events(item, delta) do { } while (0)
  38#endif
  39
  40#if defined CONFIG_COMPACTION || defined CONFIG_CMA
  41
  42#define CREATE_TRACE_POINTS
  43#include <trace/events/compaction.h>
  44
  45#define block_start_pfn(pfn, order)     round_down(pfn, 1UL << (order))
  46#define block_end_pfn(pfn, order)       ALIGN((pfn) + 1, 1UL << (order))
  47#define pageblock_start_pfn(pfn)        block_start_pfn(pfn, pageblock_order)
  48#define pageblock_end_pfn(pfn)          block_end_pfn(pfn, pageblock_order)
  49
  50static unsigned long release_freepages(struct list_head *freelist)
  51{
  52        struct page *page, *next;
  53        unsigned long high_pfn = 0;
  54
  55        list_for_each_entry_safe(page, next, freelist, lru) {
  56                unsigned long pfn = page_to_pfn(page);
  57                list_del(&page->lru);
  58                __free_page(page);
  59                if (pfn > high_pfn)
  60                        high_pfn = pfn;
  61        }
  62
  63        return high_pfn;
  64}
  65
  66static void map_pages(struct list_head *list)
  67{
  68        unsigned int i, order, nr_pages;
  69        struct page *page, *next;
  70        LIST_HEAD(tmp_list);
  71
  72        list_for_each_entry_safe(page, next, list, lru) {
  73                list_del(&page->lru);
  74
  75                order = page_private(page);
  76                nr_pages = 1 << order;
  77
  78                post_alloc_hook(page, order, __GFP_MOVABLE);
  79                if (order)
  80                        split_page(page, order);
  81
  82                for (i = 0; i < nr_pages; i++) {
  83                        list_add(&page->lru, &tmp_list);
  84                        page++;
  85                }
  86        }
  87
  88        list_splice(&tmp_list, list);
  89}
  90
  91static inline bool migrate_async_suitable(int migratetype)
  92{
  93        return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
  94}
  95
  96#ifdef CONFIG_COMPACTION
  97
  98int PageMovable(struct page *page)
  99{
 100        struct address_space *mapping;
 101
 102        VM_BUG_ON_PAGE(!PageLocked(page), page);
 103        if (!__PageMovable(page))
 104                return 0;
 105
 106        mapping = page_mapping(page);
 107        if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
 108                return 1;
 109
 110        return 0;
 111}
 112EXPORT_SYMBOL(PageMovable);
 113
 114void __SetPageMovable(struct page *page, struct address_space *mapping)
 115{
 116        VM_BUG_ON_PAGE(!PageLocked(page), page);
 117        VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
 118        page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
 119}
 120EXPORT_SYMBOL(__SetPageMovable);
 121
 122void __ClearPageMovable(struct page *page)
 123{
 124        VM_BUG_ON_PAGE(!PageLocked(page), page);
 125        VM_BUG_ON_PAGE(!PageMovable(page), page);
 126        /*
 127         * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
 128         * flag so that VM can catch up released page by driver after isolation.
 129         * With it, VM migration doesn't try to put it back.
 130         */
 131        page->mapping = (void *)((unsigned long)page->mapping &
 132                                PAGE_MAPPING_MOVABLE);
 133}
 134EXPORT_SYMBOL(__ClearPageMovable);
 135
 136/* Do not skip compaction more than 64 times */
 137#define COMPACT_MAX_DEFER_SHIFT 6
 138
 139/*
 140 * Compaction is deferred when compaction fails to result in a page
 141 * allocation success. 1 << compact_defer_limit compactions are skipped up
 142 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
 143 */
 144void defer_compaction(struct zone *zone, int order)
 145{
 146        zone->compact_considered = 0;
 147        zone->compact_defer_shift++;
 148
 149        if (order < zone->compact_order_failed)
 150                zone->compact_order_failed = order;
 151
 152        if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
 153                zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
 154
 155        trace_mm_compaction_defer_compaction(zone, order);
 156}
 157
 158/* Returns true if compaction should be skipped this time */
 159bool compaction_deferred(struct zone *zone, int order)
 160{
 161        unsigned long defer_limit = 1UL << zone->compact_defer_shift;
 162
 163        if (order < zone->compact_order_failed)
 164                return false;
 165
 166        /* Avoid possible overflow */
 167        if (++zone->compact_considered > defer_limit)
 168                zone->compact_considered = defer_limit;
 169
 170        if (zone->compact_considered >= defer_limit)
 171                return false;
 172
 173        trace_mm_compaction_deferred(zone, order);
 174
 175        return true;
 176}
 177
 178/*
 179 * Update defer tracking counters after successful compaction of given order,
 180 * which means an allocation either succeeded (alloc_success == true) or is
 181 * expected to succeed.
 182 */
 183void compaction_defer_reset(struct zone *zone, int order,
 184                bool alloc_success)
 185{
 186        if (alloc_success) {
 187                zone->compact_considered = 0;
 188                zone->compact_defer_shift = 0;
 189        }
 190        if (order >= zone->compact_order_failed)
 191                zone->compact_order_failed = order + 1;
 192
 193        trace_mm_compaction_defer_reset(zone, order);
 194}
 195
 196/* Returns true if restarting compaction after many failures */
 197bool compaction_restarting(struct zone *zone, int order)
 198{
 199        if (order < zone->compact_order_failed)
 200                return false;
 201
 202        return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
 203                zone->compact_considered >= 1UL << zone->compact_defer_shift;
 204}
 205
 206/* Returns true if the pageblock should be scanned for pages to isolate. */
 207static inline bool isolation_suitable(struct compact_control *cc,
 208                                        struct page *page)
 209{
 210        if (cc->ignore_skip_hint)
 211                return true;
 212
 213        return !get_pageblock_skip(page);
 214}
 215
 216static void reset_cached_positions(struct zone *zone)
 217{
 218        zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
 219        zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
 220        zone->compact_cached_free_pfn =
 221                                pageblock_start_pfn(zone_end_pfn(zone) - 1);
 222}
 223
 224/*
 225 * This function is called to clear all cached information on pageblocks that
 226 * should be skipped for page isolation when the migrate and free page scanner
 227 * meet.
 228 */
 229static void __reset_isolation_suitable(struct zone *zone)
 230{
 231        unsigned long start_pfn = zone->zone_start_pfn;
 232        unsigned long end_pfn = zone_end_pfn(zone);
 233        unsigned long pfn;
 234
 235        zone->compact_blockskip_flush = false;
 236
 237        /* Walk the zone and mark every pageblock as suitable for isolation */
 238        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 239                struct page *page;
 240
 241                cond_resched();
 242
 243                if (!pfn_valid(pfn))
 244                        continue;
 245
 246                page = pfn_to_page(pfn);
 247                if (zone != page_zone(page))
 248                        continue;
 249
 250                clear_pageblock_skip(page);
 251        }
 252
 253        reset_cached_positions(zone);
 254}
 255
 256void reset_isolation_suitable(pg_data_t *pgdat)
 257{
 258        int zoneid;
 259
 260        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 261                struct zone *zone = &pgdat->node_zones[zoneid];
 262                if (!populated_zone(zone))
 263                        continue;
 264
 265                /* Only flush if a full compaction finished recently */
 266                if (zone->compact_blockskip_flush)
 267                        __reset_isolation_suitable(zone);
 268        }
 269}
 270
 271/*
 272 * If no pages were isolated then mark this pageblock to be skipped in the
 273 * future. The information is later cleared by __reset_isolation_suitable().
 274 */
 275static void update_pageblock_skip(struct compact_control *cc,
 276                        struct page *page, unsigned long nr_isolated,
 277                        bool migrate_scanner)
 278{
 279        struct zone *zone = cc->zone;
 280        unsigned long pfn;
 281
 282        if (cc->ignore_skip_hint)
 283                return;
 284
 285        if (!page)
 286                return;
 287
 288        if (nr_isolated)
 289                return;
 290
 291        set_pageblock_skip(page);
 292
 293        pfn = page_to_pfn(page);
 294
 295        /* Update where async and sync compaction should restart */
 296        if (migrate_scanner) {
 297                if (pfn > zone->compact_cached_migrate_pfn[0])
 298                        zone->compact_cached_migrate_pfn[0] = pfn;
 299                if (cc->mode != MIGRATE_ASYNC &&
 300                    pfn > zone->compact_cached_migrate_pfn[1])
 301                        zone->compact_cached_migrate_pfn[1] = pfn;
 302        } else {
 303                if (pfn < zone->compact_cached_free_pfn)
 304                        zone->compact_cached_free_pfn = pfn;
 305        }
 306}
 307#else
 308static inline bool isolation_suitable(struct compact_control *cc,
 309                                        struct page *page)
 310{
 311        return true;
 312}
 313
 314static void update_pageblock_skip(struct compact_control *cc,
 315                        struct page *page, unsigned long nr_isolated,
 316                        bool migrate_scanner)
 317{
 318}
 319#endif /* CONFIG_COMPACTION */
 320
 321/*
 322 * Compaction requires the taking of some coarse locks that are potentially
 323 * very heavily contended. For async compaction, back out if the lock cannot
 324 * be taken immediately. For sync compaction, spin on the lock if needed.
 325 *
 326 * Returns true if the lock is held
 327 * Returns false if the lock is not held and compaction should abort
 328 */
 329static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
 330                                                struct compact_control *cc)
 331{
 332        if (cc->mode == MIGRATE_ASYNC) {
 333                if (!spin_trylock_irqsave(lock, *flags)) {
 334                        cc->contended = true;
 335                        return false;
 336                }
 337        } else {
 338                spin_lock_irqsave(lock, *flags);
 339        }
 340
 341        return true;
 342}
 343
 344/*
 345 * Compaction requires the taking of some coarse locks that are potentially
 346 * very heavily contended. The lock should be periodically unlocked to avoid
 347 * having disabled IRQs for a long time, even when there is nobody waiting on
 348 * the lock. It might also be that allowing the IRQs will result in
 349 * need_resched() becoming true. If scheduling is needed, async compaction
 350 * aborts. Sync compaction schedules.
 351 * Either compaction type will also abort if a fatal signal is pending.
 352 * In either case if the lock was locked, it is dropped and not regained.
 353 *
 354 * Returns true if compaction should abort due to fatal signal pending, or
 355 *              async compaction due to need_resched()
 356 * Returns false when compaction can continue (sync compaction might have
 357 *              scheduled)
 358 */
 359static bool compact_unlock_should_abort(spinlock_t *lock,
 360                unsigned long flags, bool *locked, struct compact_control *cc)
 361{
 362        if (*locked) {
 363                spin_unlock_irqrestore(lock, flags);
 364                *locked = false;
 365        }
 366
 367        if (fatal_signal_pending(current)) {
 368                cc->contended = true;
 369                return true;
 370        }
 371
 372        if (need_resched()) {
 373                if (cc->mode == MIGRATE_ASYNC) {
 374                        cc->contended = true;
 375                        return true;
 376                }
 377                cond_resched();
 378        }
 379
 380        return false;
 381}
 382
 383/*
 384 * Aside from avoiding lock contention, compaction also periodically checks
 385 * need_resched() and either schedules in sync compaction or aborts async
 386 * compaction. This is similar to what compact_unlock_should_abort() does, but
 387 * is used where no lock is concerned.
 388 *
 389 * Returns false when no scheduling was needed, or sync compaction scheduled.
 390 * Returns true when async compaction should abort.
 391 */
 392static inline bool compact_should_abort(struct compact_control *cc)
 393{
 394        /* async compaction aborts if contended */
 395        if (need_resched()) {
 396                if (cc->mode == MIGRATE_ASYNC) {
 397                        cc->contended = true;
 398                        return true;
 399                }
 400
 401                cond_resched();
 402        }
 403
 404        return false;
 405}
 406
 407/*
 408 * Isolate free pages onto a private freelist. If @strict is true, will abort
 409 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
 410 * (even though it may still end up isolating some pages).
 411 */
 412static unsigned long isolate_freepages_block(struct compact_control *cc,
 413                                unsigned long *start_pfn,
 414                                unsigned long end_pfn,
 415                                struct list_head *freelist,
 416                                bool strict)
 417{
 418        int nr_scanned = 0, total_isolated = 0;
 419        struct page *cursor, *valid_page = NULL;
 420        unsigned long flags = 0;
 421        bool locked = false;
 422        unsigned long blockpfn = *start_pfn;
 423        unsigned int order;
 424
 425        cursor = pfn_to_page(blockpfn);
 426
 427        /* Isolate free pages. */
 428        for (; blockpfn < end_pfn; blockpfn++, cursor++) {
 429                int isolated;
 430                struct page *page = cursor;
 431
 432                /*
 433                 * Periodically drop the lock (if held) regardless of its
 434                 * contention, to give chance to IRQs. Abort if fatal signal
 435                 * pending or async compaction detects need_resched()
 436                 */
 437                if (!(blockpfn % SWAP_CLUSTER_MAX)
 438                    && compact_unlock_should_abort(&cc->zone->lock, flags,
 439                                                                &locked, cc))
 440                        break;
 441
 442                nr_scanned++;
 443                if (!pfn_valid_within(blockpfn))
 444                        goto isolate_fail;
 445
 446                if (!valid_page)
 447                        valid_page = page;
 448
 449                /*
 450                 * For compound pages such as THP and hugetlbfs, we can save
 451                 * potentially a lot of iterations if we skip them at once.
 452                 * The check is racy, but we can consider only valid values
 453                 * and the only danger is skipping too much.
 454                 */
 455                if (PageCompound(page)) {
 456                        unsigned int comp_order = compound_order(page);
 457
 458                        if (likely(comp_order < MAX_ORDER)) {
 459                                blockpfn += (1UL << comp_order) - 1;
 460                                cursor += (1UL << comp_order) - 1;
 461                        }
 462
 463                        goto isolate_fail;
 464                }
 465
 466                if (!PageBuddy(page))
 467                        goto isolate_fail;
 468
 469                /*
 470                 * If we already hold the lock, we can skip some rechecking.
 471                 * Note that if we hold the lock now, checked_pageblock was
 472                 * already set in some previous iteration (or strict is true),
 473                 * so it is correct to skip the suitable migration target
 474                 * recheck as well.
 475                 */
 476                if (!locked) {
 477                        /*
 478                         * The zone lock must be held to isolate freepages.
 479                         * Unfortunately this is a very coarse lock and can be
 480                         * heavily contended if there are parallel allocations
 481                         * or parallel compactions. For async compaction do not
 482                         * spin on the lock and we acquire the lock as late as
 483                         * possible.
 484                         */
 485                        locked = compact_trylock_irqsave(&cc->zone->lock,
 486                                                                &flags, cc);
 487                        if (!locked)
 488                                break;
 489
 490                        /* Recheck this is a buddy page under lock */
 491                        if (!PageBuddy(page))
 492                                goto isolate_fail;
 493                }
 494
 495                /* Found a free page, will break it into order-0 pages */
 496                order = page_order(page);
 497                isolated = __isolate_free_page(page, order);
 498                if (!isolated)
 499                        break;
 500                set_page_private(page, order);
 501
 502                total_isolated += isolated;
 503                cc->nr_freepages += isolated;
 504                list_add_tail(&page->lru, freelist);
 505
 506                if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
 507                        blockpfn += isolated;
 508                        break;
 509                }
 510                /* Advance to the end of split page */
 511                blockpfn += isolated - 1;
 512                cursor += isolated - 1;
 513                continue;
 514
 515isolate_fail:
 516                if (strict)
 517                        break;
 518                else
 519                        continue;
 520
 521        }
 522
 523        if (locked)
 524                spin_unlock_irqrestore(&cc->zone->lock, flags);
 525
 526        /*
 527         * There is a tiny chance that we have read bogus compound_order(),
 528         * so be careful to not go outside of the pageblock.
 529         */
 530        if (unlikely(blockpfn > end_pfn))
 531                blockpfn = end_pfn;
 532
 533        trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
 534                                        nr_scanned, total_isolated);
 535
 536        /* Record how far we have got within the block */
 537        *start_pfn = blockpfn;
 538
 539        /*
 540         * If strict isolation is requested by CMA then check that all the
 541         * pages requested were isolated. If there were any failures, 0 is
 542         * returned and CMA will fail.
 543         */
 544        if (strict && blockpfn < end_pfn)
 545                total_isolated = 0;
 546
 547        /* Update the pageblock-skip if the whole pageblock was scanned */
 548        if (blockpfn == end_pfn)
 549                update_pageblock_skip(cc, valid_page, total_isolated, false);
 550
 551        count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
 552        if (total_isolated)
 553                count_compact_events(COMPACTISOLATED, total_isolated);
 554        return total_isolated;
 555}
 556
 557/**
 558 * isolate_freepages_range() - isolate free pages.
 559 * @start_pfn: The first PFN to start isolating.
 560 * @end_pfn:   The one-past-last PFN.
 561 *
 562 * Non-free pages, invalid PFNs, or zone boundaries within the
 563 * [start_pfn, end_pfn) range are considered errors, cause function to
 564 * undo its actions and return zero.
 565 *
 566 * Otherwise, function returns one-past-the-last PFN of isolated page
 567 * (which may be greater then end_pfn if end fell in a middle of
 568 * a free page).
 569 */
 570unsigned long
 571isolate_freepages_range(struct compact_control *cc,
 572                        unsigned long start_pfn, unsigned long end_pfn)
 573{
 574        unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
 575        LIST_HEAD(freelist);
 576
 577        pfn = start_pfn;
 578        block_start_pfn = pageblock_start_pfn(pfn);
 579        if (block_start_pfn < cc->zone->zone_start_pfn)
 580                block_start_pfn = cc->zone->zone_start_pfn;
 581        block_end_pfn = pageblock_end_pfn(pfn);
 582
 583        for (; pfn < end_pfn; pfn += isolated,
 584                                block_start_pfn = block_end_pfn,
 585                                block_end_pfn += pageblock_nr_pages) {
 586                /* Protect pfn from changing by isolate_freepages_block */
 587                unsigned long isolate_start_pfn = pfn;
 588
 589                block_end_pfn = min(block_end_pfn, end_pfn);
 590
 591                /*
 592                 * pfn could pass the block_end_pfn if isolated freepage
 593                 * is more than pageblock order. In this case, we adjust
 594                 * scanning range to right one.
 595                 */
 596                if (pfn >= block_end_pfn) {
 597                        block_start_pfn = pageblock_start_pfn(pfn);
 598                        block_end_pfn = pageblock_end_pfn(pfn);
 599                        block_end_pfn = min(block_end_pfn, end_pfn);
 600                }
 601
 602                if (!pageblock_pfn_to_page(block_start_pfn,
 603                                        block_end_pfn, cc->zone))
 604                        break;
 605
 606                isolated = isolate_freepages_block(cc, &isolate_start_pfn,
 607                                                block_end_pfn, &freelist, true);
 608
 609                /*
 610                 * In strict mode, isolate_freepages_block() returns 0 if
 611                 * there are any holes in the block (ie. invalid PFNs or
 612                 * non-free pages).
 613                 */
 614                if (!isolated)
 615                        break;
 616
 617                /*
 618                 * If we managed to isolate pages, it is always (1 << n) *
 619                 * pageblock_nr_pages for some non-negative n.  (Max order
 620                 * page may span two pageblocks).
 621                 */
 622        }
 623
 624        /* __isolate_free_page() does not map the pages */
 625        map_pages(&freelist);
 626
 627        if (pfn < end_pfn) {
 628                /* Loop terminated early, cleanup. */
 629                release_freepages(&freelist);
 630                return 0;
 631        }
 632
 633        /* We don't use freelists for anything. */
 634        return pfn;
 635}
 636
 637/* Update the number of anon and file isolated pages in the zone */
 638static void acct_isolated(struct zone *zone, struct compact_control *cc)
 639{
 640        struct page *page;
 641        unsigned int count[2] = { 0, };
 642
 643        if (list_empty(&cc->migratepages))
 644                return;
 645
 646        list_for_each_entry(page, &cc->migratepages, lru)
 647                count[!!page_is_file_cache(page)]++;
 648
 649        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, count[0]);
 650        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, count[1]);
 651}
 652
 653/* Similar to reclaim, but different enough that they don't share logic */
 654static bool too_many_isolated(struct zone *zone)
 655{
 656        unsigned long active, inactive, isolated;
 657
 658        inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
 659                        node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
 660        active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
 661                        node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
 662        isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
 663                        node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
 664
 665        return isolated > (inactive + active) / 2;
 666}
 667
 668/**
 669 * isolate_migratepages_block() - isolate all migrate-able pages within
 670 *                                a single pageblock
 671 * @cc:         Compaction control structure.
 672 * @low_pfn:    The first PFN to isolate
 673 * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
 674 * @isolate_mode: Isolation mode to be used.
 675 *
 676 * Isolate all pages that can be migrated from the range specified by
 677 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 678 * Returns zero if there is a fatal signal pending, otherwise PFN of the
 679 * first page that was not scanned (which may be both less, equal to or more
 680 * than end_pfn).
 681 *
 682 * The pages are isolated on cc->migratepages list (not required to be empty),
 683 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
 684 * is neither read nor updated.
 685 */
 686static unsigned long
 687isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 688                        unsigned long end_pfn, isolate_mode_t isolate_mode)
 689{
 690        struct zone *zone = cc->zone;
 691        unsigned long nr_scanned = 0, nr_isolated = 0;
 692        struct lruvec *lruvec;
 693        unsigned long flags = 0;
 694        bool locked = false;
 695        struct page *page = NULL, *valid_page = NULL;
 696        unsigned long start_pfn = low_pfn;
 697        bool skip_on_failure = false;
 698        unsigned long next_skip_pfn = 0;
 699
 700        /*
 701         * Ensure that there are not too many pages isolated from the LRU
 702         * list by either parallel reclaimers or compaction. If there are,
 703         * delay for some time until fewer pages are isolated
 704         */
 705        while (unlikely(too_many_isolated(zone))) {
 706                /* async migration should just abort */
 707                if (cc->mode == MIGRATE_ASYNC)
 708                        return 0;
 709
 710                congestion_wait(BLK_RW_ASYNC, HZ/10);
 711
 712                if (fatal_signal_pending(current))
 713                        return 0;
 714        }
 715
 716        if (compact_should_abort(cc))
 717                return 0;
 718
 719        if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
 720                skip_on_failure = true;
 721                next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 722        }
 723
 724        /* Time to isolate some pages for migration */
 725        for (; low_pfn < end_pfn; low_pfn++) {
 726
 727                if (skip_on_failure && low_pfn >= next_skip_pfn) {
 728                        /*
 729                         * We have isolated all migration candidates in the
 730                         * previous order-aligned block, and did not skip it due
 731                         * to failure. We should migrate the pages now and
 732                         * hopefully succeed compaction.
 733                         */
 734                        if (nr_isolated)
 735                                break;
 736
 737                        /*
 738                         * We failed to isolate in the previous order-aligned
 739                         * block. Set the new boundary to the end of the
 740                         * current block. Note we can't simply increase
 741                         * next_skip_pfn by 1 << order, as low_pfn might have
 742                         * been incremented by a higher number due to skipping
 743                         * a compound or a high-order buddy page in the
 744                         * previous loop iteration.
 745                         */
 746                        next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 747                }
 748
 749                /*
 750                 * Periodically drop the lock (if held) regardless of its
 751                 * contention, to give chance to IRQs. Abort async compaction
 752                 * if contended.
 753                 */
 754                if (!(low_pfn % SWAP_CLUSTER_MAX)
 755                    && compact_unlock_should_abort(zone_lru_lock(zone), flags,
 756                                                                &locked, cc))
 757                        break;
 758
 759                if (!pfn_valid_within(low_pfn))
 760                        goto isolate_fail;
 761                nr_scanned++;
 762
 763                page = pfn_to_page(low_pfn);
 764
 765                if (!valid_page)
 766                        valid_page = page;
 767
 768                /*
 769                 * Skip if free. We read page order here without zone lock
 770                 * which is generally unsafe, but the race window is small and
 771                 * the worst thing that can happen is that we skip some
 772                 * potential isolation targets.
 773                 */
 774                if (PageBuddy(page)) {
 775                        unsigned long freepage_order = page_order_unsafe(page);
 776
 777                        /*
 778                         * Without lock, we cannot be sure that what we got is
 779                         * a valid page order. Consider only values in the
 780                         * valid order range to prevent low_pfn overflow.
 781                         */
 782                        if (freepage_order > 0 && freepage_order < MAX_ORDER)
 783                                low_pfn += (1UL << freepage_order) - 1;
 784                        continue;
 785                }
 786
 787                /*
 788                 * Regardless of being on LRU, compound pages such as THP and
 789                 * hugetlbfs are not to be compacted. We can potentially save
 790                 * a lot of iterations if we skip them at once. The check is
 791                 * racy, but we can consider only valid values and the only
 792                 * danger is skipping too much.
 793                 */
 794                if (PageCompound(page)) {
 795                        unsigned int comp_order = compound_order(page);
 796
 797                        if (likely(comp_order < MAX_ORDER))
 798                                low_pfn += (1UL << comp_order) - 1;
 799
 800                        goto isolate_fail;
 801                }
 802
 803                /*
 804                 * Check may be lockless but that's ok as we recheck later.
 805                 * It's possible to migrate LRU and non-lru movable pages.
 806                 * Skip any other type of page
 807                 */
 808                if (!PageLRU(page)) {
 809                        /*
 810                         * __PageMovable can return false positive so we need
 811                         * to verify it under page_lock.
 812                         */
 813                        if (unlikely(__PageMovable(page)) &&
 814                                        !PageIsolated(page)) {
 815                                if (locked) {
 816                                        spin_unlock_irqrestore(zone_lru_lock(zone),
 817                                                                        flags);
 818                                        locked = false;
 819                                }
 820
 821                                if (isolate_movable_page(page, isolate_mode))
 822                                        goto isolate_success;
 823                        }
 824
 825                        goto isolate_fail;
 826                }
 827
 828                /*
 829                 * Migration will fail if an anonymous page is pinned in memory,
 830                 * so avoid taking lru_lock and isolating it unnecessarily in an
 831                 * admittedly racy check.
 832                 */
 833                if (!page_mapping(page) &&
 834                    page_count(page) > page_mapcount(page))
 835                        goto isolate_fail;
 836
 837                /* If we already hold the lock, we can skip some rechecking */
 838                if (!locked) {
 839                        locked = compact_trylock_irqsave(zone_lru_lock(zone),
 840                                                                &flags, cc);
 841                        if (!locked)
 842                                break;
 843
 844                        /* Recheck PageLRU and PageCompound under lock */
 845                        if (!PageLRU(page))
 846                                goto isolate_fail;
 847
 848                        /*
 849                         * Page become compound since the non-locked check,
 850                         * and it's on LRU. It can only be a THP so the order
 851                         * is safe to read and it's 0 for tail pages.
 852                         */
 853                        if (unlikely(PageCompound(page))) {
 854                                low_pfn += (1UL << compound_order(page)) - 1;
 855                                goto isolate_fail;
 856                        }
 857                }
 858
 859                lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
 860
 861                /* Try isolate the page */
 862                if (__isolate_lru_page(page, isolate_mode) != 0)
 863                        goto isolate_fail;
 864
 865                VM_BUG_ON_PAGE(PageCompound(page), page);
 866
 867                /* Successfully isolated */
 868                del_page_from_lru_list(page, lruvec, page_lru(page));
 869
 870isolate_success:
 871                list_add(&page->lru, &cc->migratepages);
 872                cc->nr_migratepages++;
 873                nr_isolated++;
 874
 875                /*
 876                 * Record where we could have freed pages by migration and not
 877                 * yet flushed them to buddy allocator.
 878                 * - this is the lowest page that was isolated and likely be
 879                 * then freed by migration.
 880                 */
 881                if (!cc->last_migrated_pfn)
 882                        cc->last_migrated_pfn = low_pfn;
 883
 884                /* Avoid isolating too much */
 885                if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
 886                        ++low_pfn;
 887                        break;
 888                }
 889
 890                continue;
 891isolate_fail:
 892                if (!skip_on_failure)
 893                        continue;
 894
 895                /*
 896                 * We have isolated some pages, but then failed. Release them
 897                 * instead of migrating, as we cannot form the cc->order buddy
 898                 * page anyway.
 899                 */
 900                if (nr_isolated) {
 901                        if (locked) {
 902                                spin_unlock_irqrestore(zone_lru_lock(zone), flags);
 903                                locked = false;
 904                        }
 905                        acct_isolated(zone, cc);
 906                        putback_movable_pages(&cc->migratepages);
 907                        cc->nr_migratepages = 0;
 908                        cc->last_migrated_pfn = 0;
 909                        nr_isolated = 0;
 910                }
 911
 912                if (low_pfn < next_skip_pfn) {
 913                        low_pfn = next_skip_pfn - 1;
 914                        /*
 915                         * The check near the loop beginning would have updated
 916                         * next_skip_pfn too, but this is a bit simpler.
 917                         */
 918                        next_skip_pfn += 1UL << cc->order;
 919                }
 920        }
 921
 922        /*
 923         * The PageBuddy() check could have potentially brought us outside
 924         * the range to be scanned.
 925         */
 926        if (unlikely(low_pfn > end_pfn))
 927                low_pfn = end_pfn;
 928
 929        if (locked)
 930                spin_unlock_irqrestore(zone_lru_lock(zone), flags);
 931
 932        /*
 933         * Update the pageblock-skip information and cached scanner pfn,
 934         * if the whole pageblock was scanned without isolating any page.
 935         */
 936        if (low_pfn == end_pfn)
 937                update_pageblock_skip(cc, valid_page, nr_isolated, true);
 938
 939        trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
 940                                                nr_scanned, nr_isolated);
 941
 942        count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
 943        if (nr_isolated)
 944                count_compact_events(COMPACTISOLATED, nr_isolated);
 945
 946        return low_pfn;
 947}
 948
 949/**
 950 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
 951 * @cc:        Compaction control structure.
 952 * @start_pfn: The first PFN to start isolating.
 953 * @end_pfn:   The one-past-last PFN.
 954 *
 955 * Returns zero if isolation fails fatally due to e.g. pending signal.
 956 * Otherwise, function returns one-past-the-last PFN of isolated page
 957 * (which may be greater than end_pfn if end fell in a middle of a THP page).
 958 */
 959unsigned long
 960isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
 961                                                        unsigned long end_pfn)
 962{
 963        unsigned long pfn, block_start_pfn, block_end_pfn;
 964
 965        /* Scan block by block. First and last block may be incomplete */
 966        pfn = start_pfn;
 967        block_start_pfn = pageblock_start_pfn(pfn);
 968        if (block_start_pfn < cc->zone->zone_start_pfn)
 969                block_start_pfn = cc->zone->zone_start_pfn;
 970        block_end_pfn = pageblock_end_pfn(pfn);
 971
 972        for (; pfn < end_pfn; pfn = block_end_pfn,
 973                                block_start_pfn = block_end_pfn,
 974                                block_end_pfn += pageblock_nr_pages) {
 975
 976                block_end_pfn = min(block_end_pfn, end_pfn);
 977
 978                if (!pageblock_pfn_to_page(block_start_pfn,
 979                                        block_end_pfn, cc->zone))
 980                        continue;
 981
 982                pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
 983                                                        ISOLATE_UNEVICTABLE);
 984
 985                if (!pfn)
 986                        break;
 987
 988                if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 989                        break;
 990        }
 991        acct_isolated(cc->zone, cc);
 992
 993        return pfn;
 994}
 995
 996#endif /* CONFIG_COMPACTION || CONFIG_CMA */
 997#ifdef CONFIG_COMPACTION
 998
 999/* Returns true if the page is within a block suitable for migration to */
1000static bool suitable_migration_target(struct compact_control *cc,
1001                                                        struct page *page)
1002{
1003        if (cc->ignore_block_suitable)
1004                return true;
1005
1006        /* If the page is a large free page, then disallow migration */
1007        if (PageBuddy(page)) {
1008                /*
1009                 * We are checking page_order without zone->lock taken. But
1010                 * the only small danger is that we skip a potentially suitable
1011                 * pageblock, so it's not worth to check order for valid range.
1012                 */
1013                if (page_order_unsafe(page) >= pageblock_order)
1014                        return false;
1015        }
1016
1017        /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1018        if (migrate_async_suitable(get_pageblock_migratetype(page)))
1019                return true;
1020
1021        /* Otherwise skip the block */
1022        return false;
1023}
1024
1025/*
1026 * Test whether the free scanner has reached the same or lower pageblock than
1027 * the migration scanner, and compaction should thus terminate.
1028 */
1029static inline bool compact_scanners_met(struct compact_control *cc)
1030{
1031        return (cc->free_pfn >> pageblock_order)
1032                <= (cc->migrate_pfn >> pageblock_order);
1033}
1034
1035/*
1036 * Based on information in the current compact_control, find blocks
1037 * suitable for isolating free pages from and then isolate them.
1038 */
1039static void isolate_freepages(struct compact_control *cc)
1040{
1041        struct zone *zone = cc->zone;
1042        struct page *page;
1043        unsigned long block_start_pfn;  /* start of current pageblock */
1044        unsigned long isolate_start_pfn; /* exact pfn we start at */
1045        unsigned long block_end_pfn;    /* end of current pageblock */
1046        unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1047        struct list_head *freelist = &cc->freepages;
1048
1049        /*
1050         * Initialise the free scanner. The starting point is where we last
1051         * successfully isolated from, zone-cached value, or the end of the
1052         * zone when isolating for the first time. For looping we also need
1053         * this pfn aligned down to the pageblock boundary, because we do
1054         * block_start_pfn -= pageblock_nr_pages in the for loop.
1055         * For ending point, take care when isolating in last pageblock of a
1056         * a zone which ends in the middle of a pageblock.
1057         * The low boundary is the end of the pageblock the migration scanner
1058         * is using.
1059         */
1060        isolate_start_pfn = cc->free_pfn;
1061        block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1062        block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1063                                                zone_end_pfn(zone));
1064        low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1065
1066        /*
1067         * Isolate free pages until enough are available to migrate the
1068         * pages on cc->migratepages. We stop searching if the migrate
1069         * and free page scanners meet or enough free pages are isolated.
1070         */
1071        for (; block_start_pfn >= low_pfn;
1072                                block_end_pfn = block_start_pfn,
1073                                block_start_pfn -= pageblock_nr_pages,
1074                                isolate_start_pfn = block_start_pfn) {
1075                /*
1076                 * This can iterate a massively long zone without finding any
1077                 * suitable migration targets, so periodically check if we need
1078                 * to schedule, or even abort async compaction.
1079                 */
1080                if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1081                                                && compact_should_abort(cc))
1082                        break;
1083
1084                page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1085                                                                        zone);
1086                if (!page)
1087                        continue;
1088
1089                /* Check the block is suitable for migration */
1090                if (!suitable_migration_target(cc, page))
1091                        continue;
1092
1093                /* If isolation recently failed, do not retry */
1094                if (!isolation_suitable(cc, page))
1095                        continue;
1096
1097                /* Found a block suitable for isolating free pages from. */
1098                isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1099                                        freelist, false);
1100
1101                /*
1102                 * If we isolated enough freepages, or aborted due to lock
1103                 * contention, terminate.
1104                 */
1105                if ((cc->nr_freepages >= cc->nr_migratepages)
1106                                                        || cc->contended) {
1107                        if (isolate_start_pfn >= block_end_pfn) {
1108                                /*
1109                                 * Restart at previous pageblock if more
1110                                 * freepages can be isolated next time.
1111                                 */
1112                                isolate_start_pfn =
1113                                        block_start_pfn - pageblock_nr_pages;
1114                        }
1115                        break;
1116                } else if (isolate_start_pfn < block_end_pfn) {
1117                        /*
1118                         * If isolation failed early, do not continue
1119                         * needlessly.
1120                         */
1121                        break;
1122                }
1123        }
1124
1125        /* __isolate_free_page() does not map the pages */
1126        map_pages(freelist);
1127
1128        /*
1129         * Record where the free scanner will restart next time. Either we
1130         * broke from the loop and set isolate_start_pfn based on the last
1131         * call to isolate_freepages_block(), or we met the migration scanner
1132         * and the loop terminated due to isolate_start_pfn < low_pfn
1133         */
1134        cc->free_pfn = isolate_start_pfn;
1135}
1136
1137/*
1138 * This is a migrate-callback that "allocates" freepages by taking pages
1139 * from the isolated freelists in the block we are migrating to.
1140 */
1141static struct page *compaction_alloc(struct page *migratepage,
1142                                        unsigned long data,
1143                                        int **result)
1144{
1145        struct compact_control *cc = (struct compact_control *)data;
1146        struct page *freepage;
1147
1148        /*
1149         * Isolate free pages if necessary, and if we are not aborting due to
1150         * contention.
1151         */
1152        if (list_empty(&cc->freepages)) {
1153                if (!cc->contended)
1154                        isolate_freepages(cc);
1155
1156                if (list_empty(&cc->freepages))
1157                        return NULL;
1158        }
1159
1160        freepage = list_entry(cc->freepages.next, struct page, lru);
1161        list_del(&freepage->lru);
1162        cc->nr_freepages--;
1163
1164        return freepage;
1165}
1166
1167/*
1168 * This is a migrate-callback that "frees" freepages back to the isolated
1169 * freelist.  All pages on the freelist are from the same zone, so there is no
1170 * special handling needed for NUMA.
1171 */
1172static void compaction_free(struct page *page, unsigned long data)
1173{
1174        struct compact_control *cc = (struct compact_control *)data;
1175
1176        list_add(&page->lru, &cc->freepages);
1177        cc->nr_freepages++;
1178}
1179
1180/* possible outcome of isolate_migratepages */
1181typedef enum {
1182        ISOLATE_ABORT,          /* Abort compaction now */
1183        ISOLATE_NONE,           /* No pages isolated, continue scanning */
1184        ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1185} isolate_migrate_t;
1186
1187/*
1188 * Allow userspace to control policy on scanning the unevictable LRU for
1189 * compactable pages.
1190 */
1191int sysctl_compact_unevictable_allowed __read_mostly = 1;
1192
1193/*
1194 * Isolate all pages that can be migrated from the first suitable block,
1195 * starting at the block pointed to by the migrate scanner pfn within
1196 * compact_control.
1197 */
1198static isolate_migrate_t isolate_migratepages(struct zone *zone,
1199                                        struct compact_control *cc)
1200{
1201        unsigned long block_start_pfn;
1202        unsigned long block_end_pfn;
1203        unsigned long low_pfn;
1204        struct page *page;
1205        const isolate_mode_t isolate_mode =
1206                (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1207                (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1208
1209        /*
1210         * Start at where we last stopped, or beginning of the zone as
1211         * initialized by compact_zone()
1212         */
1213        low_pfn = cc->migrate_pfn;
1214        block_start_pfn = pageblock_start_pfn(low_pfn);
1215        if (block_start_pfn < zone->zone_start_pfn)
1216                block_start_pfn = zone->zone_start_pfn;
1217
1218        /* Only scan within a pageblock boundary */
1219        block_end_pfn = pageblock_end_pfn(low_pfn);
1220
1221        /*
1222         * Iterate over whole pageblocks until we find the first suitable.
1223         * Do not cross the free scanner.
1224         */
1225        for (; block_end_pfn <= cc->free_pfn;
1226                        low_pfn = block_end_pfn,
1227                        block_start_pfn = block_end_pfn,
1228                        block_end_pfn += pageblock_nr_pages) {
1229
1230                /*
1231                 * This can potentially iterate a massively long zone with
1232                 * many pageblocks unsuitable, so periodically check if we
1233                 * need to schedule, or even abort async compaction.
1234                 */
1235                if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1236                                                && compact_should_abort(cc))
1237                        break;
1238
1239                page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1240                                                                        zone);
1241                if (!page)
1242                        continue;
1243
1244                /* If isolation recently failed, do not retry */
1245                if (!isolation_suitable(cc, page))
1246                        continue;
1247
1248                /*
1249                 * For async compaction, also only scan in MOVABLE blocks.
1250                 * Async compaction is optimistic to see if the minimum amount
1251                 * of work satisfies the allocation.
1252                 */
1253                if (cc->mode == MIGRATE_ASYNC &&
1254                    !migrate_async_suitable(get_pageblock_migratetype(page)))
1255                        continue;
1256
1257                /* Perform the isolation */
1258                low_pfn = isolate_migratepages_block(cc, low_pfn,
1259                                                block_end_pfn, isolate_mode);
1260
1261                if (!low_pfn || cc->contended) {
1262                        acct_isolated(zone, cc);
1263                        return ISOLATE_ABORT;
1264                }
1265
1266                /*
1267                 * Either we isolated something and proceed with migration. Or
1268                 * we failed and compact_zone should decide if we should
1269                 * continue or not.
1270                 */
1271                break;
1272        }
1273
1274        acct_isolated(zone, cc);
1275        /* Record where migration scanner will be restarted. */
1276        cc->migrate_pfn = low_pfn;
1277
1278        return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1279}
1280
1281/*
1282 * order == -1 is expected when compacting via
1283 * /proc/sys/vm/compact_memory
1284 */
1285static inline bool is_via_compact_memory(int order)
1286{
1287        return order == -1;
1288}
1289
1290static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1291                            const int migratetype)
1292{
1293        unsigned int order;
1294        unsigned long watermark;
1295
1296        if (cc->contended || fatal_signal_pending(current))
1297                return COMPACT_CONTENDED;
1298
1299        /* Compaction run completes if the migrate and free scanner meet */
1300        if (compact_scanners_met(cc)) {
1301                /* Let the next compaction start anew. */
1302                reset_cached_positions(zone);
1303
1304                /*
1305                 * Mark that the PG_migrate_skip information should be cleared
1306                 * by kswapd when it goes to sleep. kcompactd does not set the
1307                 * flag itself as the decision to be clear should be directly
1308                 * based on an allocation request.
1309                 */
1310                if (cc->direct_compaction)
1311                        zone->compact_blockskip_flush = true;
1312
1313                if (cc->whole_zone)
1314                        return COMPACT_COMPLETE;
1315                else
1316                        return COMPACT_PARTIAL_SKIPPED;
1317        }
1318
1319        if (is_via_compact_memory(cc->order))
1320                return COMPACT_CONTINUE;
1321
1322        /* Compaction run is not finished if the watermark is not met */
1323        watermark = zone->watermark[cc->alloc_flags & ALLOC_WMARK_MASK];
1324
1325        if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1326                                                        cc->alloc_flags))
1327                return COMPACT_CONTINUE;
1328
1329        /* Direct compactor: Is a suitable page free? */
1330        for (order = cc->order; order < MAX_ORDER; order++) {
1331                struct free_area *area = &zone->free_area[order];
1332                bool can_steal;
1333
1334                /* Job done if page is free of the right migratetype */
1335                if (!list_empty(&area->free_list[migratetype]))
1336                        return COMPACT_SUCCESS;
1337
1338#ifdef CONFIG_CMA
1339                /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1340                if (migratetype == MIGRATE_MOVABLE &&
1341                        !list_empty(&area->free_list[MIGRATE_CMA]))
1342                        return COMPACT_SUCCESS;
1343#endif
1344                /*
1345                 * Job done if allocation would steal freepages from
1346                 * other migratetype buddy lists.
1347                 */
1348                if (find_suitable_fallback(area, order, migratetype,
1349                                                true, &can_steal) != -1)
1350                        return COMPACT_SUCCESS;
1351        }
1352
1353        return COMPACT_NO_SUITABLE_PAGE;
1354}
1355
1356static enum compact_result compact_finished(struct zone *zone,
1357                        struct compact_control *cc,
1358                        const int migratetype)
1359{
1360        int ret;
1361
1362        ret = __compact_finished(zone, cc, migratetype);
1363        trace_mm_compaction_finished(zone, cc->order, ret);
1364        if (ret == COMPACT_NO_SUITABLE_PAGE)
1365                ret = COMPACT_CONTINUE;
1366
1367        return ret;
1368}
1369
1370/*
1371 * compaction_suitable: Is this suitable to run compaction on this zone now?
1372 * Returns
1373 *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1374 *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1375 *   COMPACT_CONTINUE - If compaction should run now
1376 */
1377static enum compact_result __compaction_suitable(struct zone *zone, int order,
1378                                        unsigned int alloc_flags,
1379                                        int classzone_idx,
1380                                        unsigned long wmark_target)
1381{
1382        unsigned long watermark;
1383
1384        if (is_via_compact_memory(order))
1385                return COMPACT_CONTINUE;
1386
1387        watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1388        /*
1389         * If watermarks for high-order allocation are already met, there
1390         * should be no need for compaction at all.
1391         */
1392        if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1393                                                                alloc_flags))
1394                return COMPACT_SUCCESS;
1395
1396        /*
1397         * Watermarks for order-0 must be met for compaction to be able to
1398         * isolate free pages for migration targets. This means that the
1399         * watermark and alloc_flags have to match, or be more pessimistic than
1400         * the check in __isolate_free_page(). We don't use the direct
1401         * compactor's alloc_flags, as they are not relevant for freepage
1402         * isolation. We however do use the direct compactor's classzone_idx to
1403         * skip over zones where lowmem reserves would prevent allocation even
1404         * if compaction succeeds.
1405         * For costly orders, we require low watermark instead of min for
1406         * compaction to proceed to increase its chances.
1407         * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1408         * suitable migration targets
1409         */
1410        watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1411                                low_wmark_pages(zone) : min_wmark_pages(zone);
1412        watermark += compact_gap(order);
1413        if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1414                                                ALLOC_CMA, wmark_target))
1415                return COMPACT_SKIPPED;
1416
1417        return COMPACT_CONTINUE;
1418}
1419
1420enum compact_result compaction_suitable(struct zone *zone, int order,
1421                                        unsigned int alloc_flags,
1422                                        int classzone_idx)
1423{
1424        enum compact_result ret;
1425        int fragindex;
1426
1427        ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1428                                    zone_page_state(zone, NR_FREE_PAGES));
1429        /*
1430         * fragmentation index determines if allocation failures are due to
1431         * low memory or external fragmentation
1432         *
1433         * index of -1000 would imply allocations might succeed depending on
1434         * watermarks, but we already failed the high-order watermark check
1435         * index towards 0 implies failure is due to lack of memory
1436         * index towards 1000 implies failure is due to fragmentation
1437         *
1438         * Only compact if a failure would be due to fragmentation. Also
1439         * ignore fragindex for non-costly orders where the alternative to
1440         * a successful reclaim/compaction is OOM. Fragindex and the
1441         * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1442         * excessive compaction for costly orders, but it should not be at the
1443         * expense of system stability.
1444         */
1445        if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
1446                fragindex = fragmentation_index(zone, order);
1447                if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1448                        ret = COMPACT_NOT_SUITABLE_ZONE;
1449        }
1450
1451        trace_mm_compaction_suitable(zone, order, ret);
1452        if (ret == COMPACT_NOT_SUITABLE_ZONE)
1453                ret = COMPACT_SKIPPED;
1454
1455        return ret;
1456}
1457
1458bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1459                int alloc_flags)
1460{
1461        struct zone *zone;
1462        struct zoneref *z;
1463
1464        /*
1465         * Make sure at least one zone would pass __compaction_suitable if we continue
1466         * retrying the reclaim.
1467         */
1468        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1469                                        ac->nodemask) {
1470                unsigned long available;
1471                enum compact_result compact_result;
1472
1473                /*
1474                 * Do not consider all the reclaimable memory because we do not
1475                 * want to trash just for a single high order allocation which
1476                 * is even not guaranteed to appear even if __compaction_suitable
1477                 * is happy about the watermark check.
1478                 */
1479                available = zone_reclaimable_pages(zone) / order;
1480                available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1481                compact_result = __compaction_suitable(zone, order, alloc_flags,
1482                                ac_classzone_idx(ac), available);
1483                if (compact_result != COMPACT_SKIPPED)
1484                        return true;
1485        }
1486
1487        return false;
1488}
1489
1490static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1491{
1492        enum compact_result ret;
1493        unsigned long start_pfn = zone->zone_start_pfn;
1494        unsigned long end_pfn = zone_end_pfn(zone);
1495        const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1496        const bool sync = cc->mode != MIGRATE_ASYNC;
1497
1498        ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1499                                                        cc->classzone_idx);
1500        /* Compaction is likely to fail */
1501        if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1502                return ret;
1503
1504        /* huh, compaction_suitable is returning something unexpected */
1505        VM_BUG_ON(ret != COMPACT_CONTINUE);
1506
1507        /*
1508         * Clear pageblock skip if there were failures recently and compaction
1509         * is about to be retried after being deferred.
1510         */
1511        if (compaction_restarting(zone, cc->order))
1512                __reset_isolation_suitable(zone);
1513
1514        /*
1515         * Setup to move all movable pages to the end of the zone. Used cached
1516         * information on where the scanners should start (unless we explicitly
1517         * want to compact the whole zone), but check that it is initialised
1518         * by ensuring the values are within zone boundaries.
1519         */
1520        if (cc->whole_zone) {
1521                cc->migrate_pfn = start_pfn;
1522                cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1523        } else {
1524                cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1525                cc->free_pfn = zone->compact_cached_free_pfn;
1526                if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1527                        cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1528                        zone->compact_cached_free_pfn = cc->free_pfn;
1529                }
1530                if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1531                        cc->migrate_pfn = start_pfn;
1532                        zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1533                        zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1534                }
1535
1536                if (cc->migrate_pfn == start_pfn)
1537                        cc->whole_zone = true;
1538        }
1539
1540        cc->last_migrated_pfn = 0;
1541
1542        trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1543                                cc->free_pfn, end_pfn, sync);
1544
1545        migrate_prep_local();
1546
1547        while ((ret = compact_finished(zone, cc, migratetype)) ==
1548                                                COMPACT_CONTINUE) {
1549                int err;
1550
1551                switch (isolate_migratepages(zone, cc)) {
1552                case ISOLATE_ABORT:
1553                        ret = COMPACT_CONTENDED;
1554                        putback_movable_pages(&cc->migratepages);
1555                        cc->nr_migratepages = 0;
1556                        goto out;
1557                case ISOLATE_NONE:
1558                        /*
1559                         * We haven't isolated and migrated anything, but
1560                         * there might still be unflushed migrations from
1561                         * previous cc->order aligned block.
1562                         */
1563                        goto check_drain;
1564                case ISOLATE_SUCCESS:
1565                        ;
1566                }
1567
1568                err = migrate_pages(&cc->migratepages, compaction_alloc,
1569                                compaction_free, (unsigned long)cc, cc->mode,
1570                                MR_COMPACTION);
1571
1572                trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1573                                                        &cc->migratepages);
1574
1575                /* All pages were either migrated or will be released */
1576                cc->nr_migratepages = 0;
1577                if (err) {
1578                        putback_movable_pages(&cc->migratepages);
1579                        /*
1580                         * migrate_pages() may return -ENOMEM when scanners meet
1581                         * and we want compact_finished() to detect it
1582                         */
1583                        if (err == -ENOMEM && !compact_scanners_met(cc)) {
1584                                ret = COMPACT_CONTENDED;
1585                                goto out;
1586                        }
1587                        /*
1588                         * We failed to migrate at least one page in the current
1589                         * order-aligned block, so skip the rest of it.
1590                         */
1591                        if (cc->direct_compaction &&
1592                                                (cc->mode == MIGRATE_ASYNC)) {
1593                                cc->migrate_pfn = block_end_pfn(
1594                                                cc->migrate_pfn - 1, cc->order);
1595                                /* Draining pcplists is useless in this case */
1596                                cc->last_migrated_pfn = 0;
1597
1598                        }
1599                }
1600
1601check_drain:
1602                /*
1603                 * Has the migration scanner moved away from the previous
1604                 * cc->order aligned block where we migrated from? If yes,
1605                 * flush the pages that were freed, so that they can merge and
1606                 * compact_finished() can detect immediately if allocation
1607                 * would succeed.
1608                 */
1609                if (cc->order > 0 && cc->last_migrated_pfn) {
1610                        int cpu;
1611                        unsigned long current_block_start =
1612                                block_start_pfn(cc->migrate_pfn, cc->order);
1613
1614                        if (cc->last_migrated_pfn < current_block_start) {
1615                                cpu = get_cpu();
1616                                lru_add_drain_cpu(cpu);
1617                                drain_local_pages(zone);
1618                                put_cpu();
1619                                /* No more flushing until we migrate again */
1620                                cc->last_migrated_pfn = 0;
1621                        }
1622                }
1623
1624        }
1625
1626out:
1627        /*
1628         * Release free pages and update where the free scanner should restart,
1629         * so we don't leave any returned pages behind in the next attempt.
1630         */
1631        if (cc->nr_freepages > 0) {
1632                unsigned long free_pfn = release_freepages(&cc->freepages);
1633
1634                cc->nr_freepages = 0;
1635                VM_BUG_ON(free_pfn == 0);
1636                /* The cached pfn is always the first in a pageblock */
1637                free_pfn = pageblock_start_pfn(free_pfn);
1638                /*
1639                 * Only go back, not forward. The cached pfn might have been
1640                 * already reset to zone end in compact_finished()
1641                 */
1642                if (free_pfn > zone->compact_cached_free_pfn)
1643                        zone->compact_cached_free_pfn = free_pfn;
1644        }
1645
1646        trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1647                                cc->free_pfn, end_pfn, sync, ret);
1648
1649        return ret;
1650}
1651
1652static enum compact_result compact_zone_order(struct zone *zone, int order,
1653                gfp_t gfp_mask, enum compact_priority prio,
1654                unsigned int alloc_flags, int classzone_idx)
1655{
1656        enum compact_result ret;
1657        struct compact_control cc = {
1658                .nr_freepages = 0,
1659                .nr_migratepages = 0,
1660                .order = order,
1661                .gfp_mask = gfp_mask,
1662                .zone = zone,
1663                .mode = (prio == COMPACT_PRIO_ASYNC) ?
1664                                        MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1665                .alloc_flags = alloc_flags,
1666                .classzone_idx = classzone_idx,
1667                .direct_compaction = true,
1668                .whole_zone = (prio == MIN_COMPACT_PRIORITY),
1669                .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1670                .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1671        };
1672        INIT_LIST_HEAD(&cc.freepages);
1673        INIT_LIST_HEAD(&cc.migratepages);
1674
1675        ret = compact_zone(zone, &cc);
1676
1677        VM_BUG_ON(!list_empty(&cc.freepages));
1678        VM_BUG_ON(!list_empty(&cc.migratepages));
1679
1680        return ret;
1681}
1682
1683int sysctl_extfrag_threshold = 500;
1684
1685/**
1686 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1687 * @gfp_mask: The GFP mask of the current allocation
1688 * @order: The order of the current allocation
1689 * @alloc_flags: The allocation flags of the current allocation
1690 * @ac: The context of current allocation
1691 * @mode: The migration mode for async, sync light, or sync migration
1692 *
1693 * This is the main entry point for direct page compaction.
1694 */
1695enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1696                unsigned int alloc_flags, const struct alloc_context *ac,
1697                enum compact_priority prio)
1698{
1699        int may_enter_fs = gfp_mask & __GFP_FS;
1700        int may_perform_io = gfp_mask & __GFP_IO;
1701        struct zoneref *z;
1702        struct zone *zone;
1703        enum compact_result rc = COMPACT_SKIPPED;
1704
1705        /* Check if the GFP flags allow compaction */
1706        if (!may_enter_fs || !may_perform_io)
1707                return COMPACT_SKIPPED;
1708
1709        trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1710
1711        /* Compact each zone in the list */
1712        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1713                                                                ac->nodemask) {
1714                enum compact_result status;
1715
1716                if (prio > MIN_COMPACT_PRIORITY
1717                                        && compaction_deferred(zone, order)) {
1718                        rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1719                        continue;
1720                }
1721
1722                status = compact_zone_order(zone, order, gfp_mask, prio,
1723                                        alloc_flags, ac_classzone_idx(ac));
1724                rc = max(status, rc);
1725
1726                /* The allocation should succeed, stop compacting */
1727                if (status == COMPACT_SUCCESS) {
1728                        /*
1729                         * We think the allocation will succeed in this zone,
1730                         * but it is not certain, hence the false. The caller
1731                         * will repeat this with true if allocation indeed
1732                         * succeeds in this zone.
1733                         */
1734                        compaction_defer_reset(zone, order, false);
1735
1736                        break;
1737                }
1738
1739                if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1740                                        status == COMPACT_PARTIAL_SKIPPED))
1741                        /*
1742                         * We think that allocation won't succeed in this zone
1743                         * so we defer compaction there. If it ends up
1744                         * succeeding after all, it will be reset.
1745                         */
1746                        defer_compaction(zone, order);
1747
1748                /*
1749                 * We might have stopped compacting due to need_resched() in
1750                 * async compaction, or due to a fatal signal detected. In that
1751                 * case do not try further zones
1752                 */
1753                if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1754                                        || fatal_signal_pending(current))
1755                        break;
1756        }
1757
1758        return rc;
1759}
1760
1761
1762/* Compact all zones within a node */
1763static void compact_node(int nid)
1764{
1765        pg_data_t *pgdat = NODE_DATA(nid);
1766        int zoneid;
1767        struct zone *zone;
1768        struct compact_control cc = {
1769                .order = -1,
1770                .mode = MIGRATE_SYNC,
1771                .ignore_skip_hint = true,
1772                .whole_zone = true,
1773        };
1774
1775
1776        for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1777
1778                zone = &pgdat->node_zones[zoneid];
1779                if (!populated_zone(zone))
1780                        continue;
1781
1782                cc.nr_freepages = 0;
1783                cc.nr_migratepages = 0;
1784                cc.zone = zone;
1785                INIT_LIST_HEAD(&cc.freepages);
1786                INIT_LIST_HEAD(&cc.migratepages);
1787
1788                compact_zone(zone, &cc);
1789
1790                VM_BUG_ON(!list_empty(&cc.freepages));
1791                VM_BUG_ON(!list_empty(&cc.migratepages));
1792        }
1793}
1794
1795/* Compact all nodes in the system */
1796static void compact_nodes(void)
1797{
1798        int nid;
1799
1800        /* Flush pending updates to the LRU lists */
1801        lru_add_drain_all();
1802
1803        for_each_online_node(nid)
1804                compact_node(nid);
1805}
1806
1807/* The written value is actually unused, all memory is compacted */
1808int sysctl_compact_memory;
1809
1810/*
1811 * This is the entry point for compacting all nodes via
1812 * /proc/sys/vm/compact_memory
1813 */
1814int sysctl_compaction_handler(struct ctl_table *table, int write,
1815                        void __user *buffer, size_t *length, loff_t *ppos)
1816{
1817        if (write)
1818                compact_nodes();
1819
1820        return 0;
1821}
1822
1823int sysctl_extfrag_handler(struct ctl_table *table, int write,
1824                        void __user *buffer, size_t *length, loff_t *ppos)
1825{
1826        proc_dointvec_minmax(table, write, buffer, length, ppos);
1827
1828        return 0;
1829}
1830
1831#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1832static ssize_t sysfs_compact_node(struct device *dev,
1833                        struct device_attribute *attr,
1834                        const char *buf, size_t count)
1835{
1836        int nid = dev->id;
1837
1838        if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1839                /* Flush pending updates to the LRU lists */
1840                lru_add_drain_all();
1841
1842                compact_node(nid);
1843        }
1844
1845        return count;
1846}
1847static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1848
1849int compaction_register_node(struct node *node)
1850{
1851        return device_create_file(&node->dev, &dev_attr_compact);
1852}
1853
1854void compaction_unregister_node(struct node *node)
1855{
1856        return device_remove_file(&node->dev, &dev_attr_compact);
1857}
1858#endif /* CONFIG_SYSFS && CONFIG_NUMA */
1859
1860static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1861{
1862        return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1863}
1864
1865static bool kcompactd_node_suitable(pg_data_t *pgdat)
1866{
1867        int zoneid;
1868        struct zone *zone;
1869        enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1870
1871        for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1872                zone = &pgdat->node_zones[zoneid];
1873
1874                if (!populated_zone(zone))
1875                        continue;
1876
1877                if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1878                                        classzone_idx) == COMPACT_CONTINUE)
1879                        return true;
1880        }
1881
1882        return false;
1883}
1884
1885static void kcompactd_do_work(pg_data_t *pgdat)
1886{
1887        /*
1888         * With no special task, compact all zones so that a page of requested
1889         * order is allocatable.
1890         */
1891        int zoneid;
1892        struct zone *zone;
1893        struct compact_control cc = {
1894                .order = pgdat->kcompactd_max_order,
1895                .classzone_idx = pgdat->kcompactd_classzone_idx,
1896                .mode = MIGRATE_SYNC_LIGHT,
1897                .ignore_skip_hint = true,
1898
1899        };
1900        trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1901                                                        cc.classzone_idx);
1902        count_vm_event(KCOMPACTD_WAKE);
1903
1904        for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1905                int status;
1906
1907                zone = &pgdat->node_zones[zoneid];
1908                if (!populated_zone(zone))
1909                        continue;
1910
1911                if (compaction_deferred(zone, cc.order))
1912                        continue;
1913
1914                if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1915                                                        COMPACT_CONTINUE)
1916                        continue;
1917
1918                cc.nr_freepages = 0;
1919                cc.nr_migratepages = 0;
1920                cc.zone = zone;
1921                INIT_LIST_HEAD(&cc.freepages);
1922                INIT_LIST_HEAD(&cc.migratepages);
1923
1924                if (kthread_should_stop())
1925                        return;
1926                status = compact_zone(zone, &cc);
1927
1928                if (status == COMPACT_SUCCESS) {
1929                        compaction_defer_reset(zone, cc.order, false);
1930                } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1931                        /*
1932                         * We use sync migration mode here, so we defer like
1933                         * sync direct compaction does.
1934                         */
1935                        defer_compaction(zone, cc.order);
1936                }
1937
1938                VM_BUG_ON(!list_empty(&cc.freepages));
1939                VM_BUG_ON(!list_empty(&cc.migratepages));
1940        }
1941
1942        /*
1943         * Regardless of success, we are done until woken up next. But remember
1944         * the requested order/classzone_idx in case it was higher/tighter than
1945         * our current ones
1946         */
1947        if (pgdat->kcompactd_max_order <= cc.order)
1948                pgdat->kcompactd_max_order = 0;
1949        if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1950                pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1951}
1952
1953void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1954{
1955        if (!order)
1956                return;
1957
1958        if (pgdat->kcompactd_max_order < order)
1959                pgdat->kcompactd_max_order = order;
1960
1961        if (pgdat->kcompactd_classzone_idx > classzone_idx)
1962                pgdat->kcompactd_classzone_idx = classzone_idx;
1963
1964        if (!waitqueue_active(&pgdat->kcompactd_wait))
1965                return;
1966
1967        if (!kcompactd_node_suitable(pgdat))
1968                return;
1969
1970        trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1971                                                        classzone_idx);
1972        wake_up_interruptible(&pgdat->kcompactd_wait);
1973}
1974
1975/*
1976 * The background compaction daemon, started as a kernel thread
1977 * from the init process.
1978 */
1979static int kcompactd(void *p)
1980{
1981        pg_data_t *pgdat = (pg_data_t*)p;
1982        struct task_struct *tsk = current;
1983
1984        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1985
1986        if (!cpumask_empty(cpumask))
1987                set_cpus_allowed_ptr(tsk, cpumask);
1988
1989        set_freezable();
1990
1991        pgdat->kcompactd_max_order = 0;
1992        pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1993
1994        while (!kthread_should_stop()) {
1995                trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1996                wait_event_freezable(pgdat->kcompactd_wait,
1997                                kcompactd_work_requested(pgdat));
1998
1999                kcompactd_do_work(pgdat);
2000        }
2001
2002        return 0;
2003}
2004
2005/*
2006 * This kcompactd start function will be called by init and node-hot-add.
2007 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2008 */
2009int kcompactd_run(int nid)
2010{
2011        pg_data_t *pgdat = NODE_DATA(nid);
2012        int ret = 0;
2013
2014        if (pgdat->kcompactd)
2015                return 0;
2016
2017        pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2018        if (IS_ERR(pgdat->kcompactd)) {
2019                pr_err("Failed to start kcompactd on node %d\n", nid);
2020                ret = PTR_ERR(pgdat->kcompactd);
2021                pgdat->kcompactd = NULL;
2022        }
2023        return ret;
2024}
2025
2026/*
2027 * Called by memory hotplug when all memory in a node is offlined. Caller must
2028 * hold mem_hotplug_begin/end().
2029 */
2030void kcompactd_stop(int nid)
2031{
2032        struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2033
2034        if (kcompactd) {
2035                kthread_stop(kcompactd);
2036                NODE_DATA(nid)->kcompactd = NULL;
2037        }
2038}
2039
2040/*
2041 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2042 * not required for correctness. So if the last cpu in a node goes
2043 * away, we get changed to run anywhere: as the first one comes back,
2044 * restore their cpu bindings.
2045 */
2046static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2047                        void *hcpu)
2048{
2049        int nid;
2050
2051        if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2052                for_each_node_state(nid, N_MEMORY) {
2053                        pg_data_t *pgdat = NODE_DATA(nid);
2054                        const struct cpumask *mask;
2055
2056                        mask = cpumask_of_node(pgdat->node_id);
2057
2058                        if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2059                                /* One of our CPUs online: restore mask */
2060                                set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2061                }
2062        }
2063        return NOTIFY_OK;
2064}
2065
2066static int __init kcompactd_init(void)
2067{
2068        int nid;
2069
2070        for_each_node_state(nid, N_MEMORY)
2071                kcompactd_run(nid);
2072        hotcpu_notifier(cpu_callback, 0);
2073        return 0;
2074}
2075subsys_initcall(kcompactd_init)
2076
2077#endif /* CONFIG_COMPACTION */
2078