mm/compaction.c

   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
  26 static inline void count_compact_event(enum vm_event_item item)
  27 {
  28         count_vm_event(item);
  29 }
  30
  31 static 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
  50 static 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
  66 static 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
  91 static inline bool migrate_async_suitable(int migratetype)
  92 {
  93         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
  94 }
  95
  96 #ifdef CONFIG_COMPACTION
  97
  98 int 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 }
 112 EXPORT_SYMBOL(PageMovable);
 113
 114 void __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 }
 120 EXPORT_SYMBOL(__SetPageMovable);
 121
 122 void __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 }
 134 EXPORT_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  */
 144 void 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 */
 159 bool 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  */
 183 void 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 */
 197 bool 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. */
 207 static 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
 216 static 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  */
 229 static 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
 256 void 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  */
 275 static 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
 308 static inline bool isolation_suitable(struct compact_control *cc,
 309                                         struct page *page)
 310 {
 311         return true;
 312 }
 313
 314 static 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  */
 329 static 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  */
 359 static 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  */
 392 static 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  */
 412 static 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
 515 isolate_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  */
 570 unsigned long
 571 isolate_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 /* Similar to reclaim, but different enough that they don't share logic */
 638 static bool too_many_isolated(struct zone *zone)
 639 {
 640         unsigned long active, inactive, isolated;
 641
 642         inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
 643                         node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
 644         active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
 645                         node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
 646         isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
 647                         node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
 648
 649         return isolated > (inactive + active) / 2;
 650 }
 651
 652 /**
 653  * isolate_migratepages_block() - isolate all migrate-able pages within
 654  *                                a single pageblock
 655  * @cc:         Compaction control structure.
 656  * @low_pfn:    The first PFN to isolate
 657  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
 658  * @isolate_mode: Isolation mode to be used.
 659  *
 660  * Isolate all pages that can be migrated from the range specified by
 661  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 662  * Returns zero if there is a fatal signal pending, otherwise PFN of the
 663  * first page that was not scanned (which may be both less, equal to or more
 664  * than end_pfn).
 665  *
 666  * The pages are isolated on cc->migratepages list (not required to be empty),
 667  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
 668  * is neither read nor updated.
 669  */
 670 static unsigned long
 671 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 672                         unsigned long end_pfn, isolate_mode_t isolate_mode)
 673 {
 674         struct zone *zone = cc->zone;
 675         unsigned long nr_scanned = 0, nr_isolated = 0;
 676         struct lruvec *lruvec;
 677         unsigned long flags = 0;
 678         bool locked = false;
 679         struct page *page = NULL, *valid_page = NULL;
 680         unsigned long start_pfn = low_pfn;
 681         bool skip_on_failure = false;
 682         unsigned long next_skip_pfn = 0;
 683
 684         /*
 685          * Ensure that there are not too many pages isolated from the LRU
 686          * list by either parallel reclaimers or compaction. If there are,
 687          * delay for some time until fewer pages are isolated
 688          */
 689         while (unlikely(too_many_isolated(zone))) {
 690                 /* async migration should just abort */
 691                 if (cc->mode == MIGRATE_ASYNC)
 692                         return 0;
 693
 694                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 695
 696                 if (fatal_signal_pending(current))
 697                         return 0;
 698         }
 699
 700         if (compact_should_abort(cc))
 701                 return 0;
 702
 703         if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
 704                 skip_on_failure = true;
 705                 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 706         }
 707
 708         /* Time to isolate some pages for migration */
 709         for (; low_pfn < end_pfn; low_pfn++) {
 710
 711                 if (skip_on_failure && low_pfn >= next_skip_pfn) {
 712                         /*
 713                          * We have isolated all migration candidates in the
 714                          * previous order-aligned block, and did not skip it due
 715                          * to failure. We should migrate the pages now and
 716                          * hopefully succeed compaction.
 717                          */
 718                         if (nr_isolated)
 719                                 break;
 720
 721                         /*
 722                          * We failed to isolate in the previous order-aligned
 723                          * block. Set the new boundary to the end of the
 724                          * current block. Note we can't simply increase
 725                          * next_skip_pfn by 1 << order, as low_pfn might have
 726                          * been incremented by a higher number due to skipping
 727                          * a compound or a high-order buddy page in the
 728                          * previous loop iteration.
 729                          */
 730                         next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 731                 }
 732
 733                 /*
 734                  * Periodically drop the lock (if held) regardless of its
 735                  * contention, to give chance to IRQs. Abort async compaction
 736                  * if contended.
 737                  */
 738                 if (!(low_pfn % SWAP_CLUSTER_MAX)
 739                     && compact_unlock_should_abort(zone_lru_lock(zone), flags,
 740                                                                 &locked, cc))
 741                         break;
 742
 743                 if (!pfn_valid_within(low_pfn))
 744                         goto isolate_fail;
 745                 nr_scanned++;
 746
 747                 page = pfn_to_page(low_pfn);
 748
 749                 if (!valid_page)
 750                         valid_page = page;
 751
 752                 /*
 753                  * Skip if free. We read page order here without zone lock
 754                  * which is generally unsafe, but the race window is small and
 755                  * the worst thing that can happen is that we skip some
 756                  * potential isolation targets.
 757                  */
 758                 if (PageBuddy(page)) {
 759                         unsigned long freepage_order = page_order_unsafe(page);
 760
 761                         /*
 762                          * Without lock, we cannot be sure that what we got is
 763                          * a valid page order. Consider only values in the
 764                          * valid order range to prevent low_pfn overflow.
 765                          */
 766                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
 767                                 low_pfn += (1UL << freepage_order) - 1;
 768                         continue;
 769                 }
 770
 771                 /*
 772                  * Regardless of being on LRU, compound pages such as THP and
 773                  * hugetlbfs are not to be compacted. We can potentially save
 774                  * a lot of iterations if we skip them at once. The check is
 775                  * racy, but we can consider only valid values and the only
 776                  * danger is skipping too much.
 777                  */
 778                 if (PageCompound(page)) {
 779                         unsigned int comp_order = compound_order(page);
 780
 781                         if (likely(comp_order < MAX_ORDER))
 782                                 low_pfn += (1UL << comp_order) - 1;
 783
 784                         goto isolate_fail;
 785                 }
 786
 787                 /*
 788                  * Check may be lockless but that's ok as we recheck later.
 789                  * It's possible to migrate LRU and non-lru movable pages.
 790                  * Skip any other type of page
 791                  */
 792                 if (!PageLRU(page)) {
 793                         /*
 794                          * __PageMovable can return false positive so we need
 795                          * to verify it under page_lock.
 796                          */
 797                         if (unlikely(__PageMovable(page)) &&
 798                                         !PageIsolated(page)) {
 799                                 if (locked) {
 800                                         spin_unlock_irqrestore(zone_lru_lock(zone),
 801                                                                         flags);
 802                                         locked = false;
 803                                 }
 804
 805                                 if (isolate_movable_page(page, isolate_mode))
 806                                         goto isolate_success;
 807                         }
 808
 809                         goto isolate_fail;
 810                 }
 811
 812                 /*
 813                  * Migration will fail if an anonymous page is pinned in memory,
 814                  * so avoid taking lru_lock and isolating it unnecessarily in an
 815                  * admittedly racy check.
 816                  */
 817                 if (!page_mapping(page) &&
 818                     page_count(page) > page_mapcount(page))
 819                         goto isolate_fail;
 820
 821                 /* If we already hold the lock, we can skip some rechecking */
 822                 if (!locked) {
 823                         locked = compact_trylock_irqsave(zone_lru_lock(zone),
 824                                                                 &flags, cc);
 825                         if (!locked)
 826                                 break;
 827
 828                         /* Recheck PageLRU and PageCompound under lock */
 829                         if (!PageLRU(page))
 830                                 goto isolate_fail;
 831
 832                         /*
 833                          * Page become compound since the non-locked check,
 834                          * and it's on LRU. It can only be a THP so the order
 835                          * is safe to read and it's 0 for tail pages.
 836                          */
 837                         if (unlikely(PageCompound(page))) {
 838                                 low_pfn += (1UL << compound_order(page)) - 1;
 839                                 goto isolate_fail;
 840                         }
 841                 }
 842
 843                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
 844
 845                 /* Try isolate the page */
 846                 if (__isolate_lru_page(page, isolate_mode) != 0)
 847                         goto isolate_fail;
 848
 849                 VM_BUG_ON_PAGE(PageCompound(page), page);
 850
 851                 /* Successfully isolated */
 852                 del_page_from_lru_list(page, lruvec, page_lru(page));
 853                 inc_node_page_state(page,
 854                                 NR_ISOLATED_ANON + page_is_file_cache(page));
 855
 856 isolate_success:
 857                 list_add(&page->lru, &cc->migratepages);
 858                 cc->nr_migratepages++;
 859                 nr_isolated++;
 860
 861                 /*
 862                  * Record where we could have freed pages by migration and not
 863                  * yet flushed them to buddy allocator.
 864                  * - this is the lowest page that was isolated and likely be
 865                  * then freed by migration.
 866                  */
 867                 if (!cc->last_migrated_pfn)
 868                         cc->last_migrated_pfn = low_pfn;
 869
 870                 /* Avoid isolating too much */
 871                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
 872                         ++low_pfn;
 873                         break;
 874                 }
 875
 876                 continue;
 877 isolate_fail:
 878                 if (!skip_on_failure)
 879                         continue;
 880
 881                 /*
 882                  * We have isolated some pages, but then failed. Release them
 883                  * instead of migrating, as we cannot form the cc->order buddy
 884                  * page anyway.
 885                  */
 886                 if (nr_isolated) {
 887                         if (locked) {
 888                                 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
 889                                 locked = false;
 890                         }
 891                         putback_movable_pages(&cc->migratepages);
 892                         cc->nr_migratepages = 0;
 893                         cc->last_migrated_pfn = 0;
 894                         nr_isolated = 0;
 895                 }
 896
 897                 if (low_pfn < next_skip_pfn) {
 898                         low_pfn = next_skip_pfn - 1;
 899                         /*
 900                          * The check near the loop beginning would have updated
 901                          * next_skip_pfn too, but this is a bit simpler.
 902                          */
 903                         next_skip_pfn += 1UL << cc->order;
 904                 }
 905         }
 906
 907         /*
 908          * The PageBuddy() check could have potentially brought us outside
 909          * the range to be scanned.
 910          */
 911         if (unlikely(low_pfn > end_pfn))
 912                 low_pfn = end_pfn;
 913
 914         if (locked)
 915                 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
 916
 917         /*
 918          * Update the pageblock-skip information and cached scanner pfn,
 919          * if the whole pageblock was scanned without isolating any page.
 920          */
 921         if (low_pfn == end_pfn)
 922                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
 923
 924         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
 925                                                 nr_scanned, nr_isolated);
 926
 927         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
 928         if (nr_isolated)
 929                 count_compact_events(COMPACTISOLATED, nr_isolated);
 930
 931         return low_pfn;
 932 }
 933
 934 /**
 935  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
 936  * @cc:        Compaction control structure.
 937  * @start_pfn: The first PFN to start isolating.
 938  * @end_pfn:   The one-past-last PFN.
 939  *
 940  * Returns zero if isolation fails fatally due to e.g. pending signal.
 941  * Otherwise, function returns one-past-the-last PFN of isolated page
 942  * (which may be greater than end_pfn if end fell in a middle of a THP page).
 943  */
 944 unsigned long
 945 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
 946                                                         unsigned long end_pfn)
 947 {
 948         unsigned long pfn, block_start_pfn, block_end_pfn;
 949
 950         /* Scan block by block. First and last block may be incomplete */
 951         pfn = start_pfn;
 952         block_start_pfn = pageblock_start_pfn(pfn);
 953         if (block_start_pfn < cc->zone->zone_start_pfn)
 954                 block_start_pfn = cc->zone->zone_start_pfn;
 955         block_end_pfn = pageblock_end_pfn(pfn);
 956
 957         for (; pfn < end_pfn; pfn = block_end_pfn,
 958                                 block_start_pfn = block_end_pfn,
 959                                 block_end_pfn += pageblock_nr_pages) {
 960
 961                 block_end_pfn = min(block_end_pfn, end_pfn);
 962
 963                 if (!pageblock_pfn_to_page(block_start_pfn,
 964                                         block_end_pfn, cc->zone))
 965                         continue;
 966
 967                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
 968                                                         ISOLATE_UNEVICTABLE);
 969
 970                 if (!pfn)
 971                         break;
 972
 973                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 974                         break;
 975         }
 976
 977         return pfn;
 978 }
 979
 980 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 981 #ifdef CONFIG_COMPACTION
 982
 983 /* Returns true if the page is within a block suitable for migration to */
 984 static bool suitable_migration_target(struct compact_control *cc,
 985                                                         struct page *page)
 986 {
 987         if (cc->ignore_block_suitable)
 988                 return true;
 989
 990         /* If the page is a large free page, then disallow migration */
 991         if (PageBuddy(page)) {
 992                 /*
 993                  * We are checking page_order without zone->lock taken. But
 994                  * the only small danger is that we skip a potentially suitable
 995                  * pageblock, so it's not worth to check order for valid range.
 996                  */
 997                 if (page_order_unsafe(page) >= pageblock_order)
 998                         return false;
 999         }
1000
1001         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1002         if (migrate_async_suitable(get_pageblock_migratetype(page)))
1003                 return true;
1004
1005         /* Otherwise skip the block */
1006         return false;
1007 }
1008
1009 /*
1010  * Test whether the free scanner has reached the same or lower pageblock than
1011  * the migration scanner, and compaction should thus terminate.
1012  */
1013 static inline bool compact_scanners_met(struct compact_control *cc)
1014 {
1015         return (cc->free_pfn >> pageblock_order)
1016                 <= (cc->migrate_pfn >> pageblock_order);
1017 }
1018
1019 /*
1020  * Based on information in the current compact_control, find blocks
1021  * suitable for isolating free pages from and then isolate them.
1022  */
1023 static void isolate_freepages(struct compact_control *cc)
1024 {
1025         struct zone *zone = cc->zone;
1026         struct page *page;
1027         unsigned long block_start_pfn;  /* start of current pageblock */
1028         unsigned long isolate_start_pfn; /* exact pfn we start at */
1029         unsigned long block_end_pfn;    /* end of current pageblock */
1030         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1031         struct list_head *freelist = &cc->freepages;
1032
1033         /*
1034          * Initialise the free scanner. The starting point is where we last
1035          * successfully isolated from, zone-cached value, or the end of the
1036          * zone when isolating for the first time. For looping we also need
1037          * this pfn aligned down to the pageblock boundary, because we do
1038          * block_start_pfn -= pageblock_nr_pages in the for loop.
1039          * For ending point, take care when isolating in last pageblock of a
1040          * a zone which ends in the middle of a pageblock.
1041          * The low boundary is the end of the pageblock the migration scanner
1042          * is using.
1043          */
1044         isolate_start_pfn = cc->free_pfn;
1045         block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1046         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1047                                                 zone_end_pfn(zone));
1048         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1049
1050         /*
1051          * Isolate free pages until enough are available to migrate the
1052          * pages on cc->migratepages. We stop searching if the migrate
1053          * and free page scanners meet or enough free pages are isolated.
1054          */
1055         for (; block_start_pfn >= low_pfn;
1056                                 block_end_pfn = block_start_pfn,
1057                                 block_start_pfn -= pageblock_nr_pages,
1058                                 isolate_start_pfn = block_start_pfn) {
1059                 /*
1060                  * This can iterate a massively long zone without finding any
1061                  * suitable migration targets, so periodically check if we need
1062                  * to schedule, or even abort async compaction.
1063                  */
1064                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1065                                                 && compact_should_abort(cc))
1066                         break;
1067
1068                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1069                                                                         zone);
1070                 if (!page)
1071                         continue;
1072
1073                 /* Check the block is suitable for migration */
1074                 if (!suitable_migration_target(cc, page))
1075                         continue;
1076
1077                 /* If isolation recently failed, do not retry */
1078                 if (!isolation_suitable(cc, page))
1079                         continue;
1080
1081                 /* Found a block suitable for isolating free pages from. */
1082                 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1083                                         freelist, false);
1084
1085                 /*
1086                  * If we isolated enough freepages, or aborted due to lock
1087                  * contention, terminate.
1088                  */
1089                 if ((cc->nr_freepages >= cc->nr_migratepages)
1090                                                         || cc->contended) {
1091                         if (isolate_start_pfn >= block_end_pfn) {
1092                                 /*
1093                                  * Restart at previous pageblock if more
1094                                  * freepages can be isolated next time.
1095                                  */
1096                                 isolate_start_pfn =
1097                                         block_start_pfn - pageblock_nr_pages;
1098                         }
1099                         break;
1100                 } else if (isolate_start_pfn < block_end_pfn) {
1101                         /*
1102                          * If isolation failed early, do not continue
1103                          * needlessly.
1104                          */
1105                         break;
1106                 }
1107         }
1108
1109         /* __isolate_free_page() does not map the pages */
1110         map_pages(freelist);
1111
1112         /*
1113          * Record where the free scanner will restart next time. Either we
1114          * broke from the loop and set isolate_start_pfn based on the last
1115          * call to isolate_freepages_block(), or we met the migration scanner
1116          * and the loop terminated due to isolate_start_pfn < low_pfn
1117          */
1118         cc->free_pfn = isolate_start_pfn;
1119 }
1120
1121 /*
1122  * This is a migrate-callback that "allocates" freepages by taking pages
1123  * from the isolated freelists in the block we are migrating to.
1124  */
1125 static struct page *compaction_alloc(struct page *migratepage,
1126                                         unsigned long data,
1127                                         int **result)
1128 {
1129         struct compact_control *cc = (struct compact_control *)data;
1130         struct page *freepage;
1131
1132         /*
1133          * Isolate free pages if necessary, and if we are not aborting due to
1134          * contention.
1135          */
1136         if (list_empty(&cc->freepages)) {
1137                 if (!cc->contended)
1138                         isolate_freepages(cc);
1139
1140                 if (list_empty(&cc->freepages))
1141                         return NULL;
1142         }
1143
1144         freepage = list_entry(cc->freepages.next, struct page, lru);
1145         list_del(&freepage->lru);
1146         cc->nr_freepages--;
1147
1148         return freepage;
1149 }
1150
1151 /*
1152  * This is a migrate-callback that "frees" freepages back to the isolated
1153  * freelist.  All pages on the freelist are from the same zone, so there is no
1154  * special handling needed for NUMA.
1155  */
1156 static void compaction_free(struct page *page, unsigned long data)
1157 {
1158         struct compact_control *cc = (struct compact_control *)data;
1159
1160         list_add(&page->lru, &cc->freepages);
1161         cc->nr_freepages++;
1162 }
1163
1164 /* possible outcome of isolate_migratepages */
1165 typedef enum {
1166         ISOLATE_ABORT,          /* Abort compaction now */
1167         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1168         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1169 } isolate_migrate_t;
1170
1171 /*
1172  * Allow userspace to control policy on scanning the unevictable LRU for
1173  * compactable pages.
1174  */
1175 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1176
1177 /*
1178  * Isolate all pages that can be migrated from the first suitable block,
1179  * starting at the block pointed to by the migrate scanner pfn within
1180  * compact_control.
1181  */
1182 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1183                                         struct compact_control *cc)
1184 {
1185         unsigned long block_start_pfn;
1186         unsigned long block_end_pfn;
1187         unsigned long low_pfn;
1188         struct page *page;
1189         const isolate_mode_t isolate_mode =
1190                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1191                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1192
1193         /*
1194          * Start at where we last stopped, or beginning of the zone as
1195          * initialized by compact_zone()
1196          */
1197         low_pfn = cc->migrate_pfn;
1198         block_start_pfn = pageblock_start_pfn(low_pfn);
1199         if (block_start_pfn < zone->zone_start_pfn)
1200                 block_start_pfn = zone->zone_start_pfn;
1201
1202         /* Only scan within a pageblock boundary */
1203         block_end_pfn = pageblock_end_pfn(low_pfn);
1204
1205         /*
1206          * Iterate over whole pageblocks until we find the first suitable.
1207          * Do not cross the free scanner.
1208          */
1209         for (; block_end_pfn <= cc->free_pfn;
1210                         low_pfn = block_end_pfn,
1211                         block_start_pfn = block_end_pfn,
1212                         block_end_pfn += pageblock_nr_pages) {
1213
1214                 /*
1215                  * This can potentially iterate a massively long zone with
1216                  * many pageblocks unsuitable, so periodically check if we
1217                  * need to schedule, or even abort async compaction.
1218                  */
1219                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1220                                                 && compact_should_abort(cc))
1221                         break;
1222
1223                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1224                                                                         zone);
1225                 if (!page)
1226                         continue;
1227
1228                 /* If isolation recently failed, do not retry */
1229                 if (!isolation_suitable(cc, page))
1230                         continue;
1231
1232                 /*
1233                  * For async compaction, also only scan in MOVABLE blocks.
1234                  * Async compaction is optimistic to see if the minimum amount
1235                  * of work satisfies the allocation.
1236                  */
1237                 if (cc->mode == MIGRATE_ASYNC &&
1238                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1239                         continue;
1240
1241                 /* Perform the isolation */
1242                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1243                                                 block_end_pfn, isolate_mode);
1244
1245                 if (!low_pfn || cc->contended)
1246                         return ISOLATE_ABORT;
1247
1248                 /*
1249                  * Either we isolated something and proceed with migration. Or
1250                  * we failed and compact_zone should decide if we should
1251                  * continue or not.
1252                  */
1253                 break;
1254         }
1255
1256         /* Record where migration scanner will be restarted. */
1257         cc->migrate_pfn = low_pfn;
1258
1259         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1260 }
1261
1262 /*
1263  * order == -1 is expected when compacting via
1264  * /proc/sys/vm/compact_memory
1265  */
1266 static inline bool is_via_compact_memory(int order)
1267 {
1268         return order == -1;
1269 }
1270
1271 static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1272                             const int migratetype)
1273 {
1274         unsigned int order;
1275         unsigned long watermark;
1276
1277         if (cc->contended || fatal_signal_pending(current))
1278                 return COMPACT_CONTENDED;
1279
1280         /* Compaction run completes if the migrate and free scanner meet */
1281         if (compact_scanners_met(cc)) {
1282                 /* Let the next compaction start anew. */
1283                 reset_cached_positions(zone);
1284
1285                 /*
1286                  * Mark that the PG_migrate_skip information should be cleared
1287                  * by kswapd when it goes to sleep. kcompactd does not set the
1288                  * flag itself as the decision to be clear should be directly
1289                  * based on an allocation request.
1290                  */
1291                 if (cc->direct_compaction)
1292                         zone->compact_blockskip_flush = true;
1293
1294                 if (cc->whole_zone)
1295                         return COMPACT_COMPLETE;
1296                 else
1297                         return COMPACT_PARTIAL_SKIPPED;
1298         }
1299
1300         if (is_via_compact_memory(cc->order))
1301                 return COMPACT_CONTINUE;
1302
1303         /* Compaction run is not finished if the watermark is not met */
1304         watermark = zone->watermark[cc->alloc_flags & ALLOC_WMARK_MASK];
1305
1306         if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1307                                                         cc->alloc_flags))
1308                 return COMPACT_CONTINUE;
1309
1310         /* Direct compactor: Is a suitable page free? */
1311         for (order = cc->order; order < MAX_ORDER; order++) {
1312                 struct free_area *area = &zone->free_area[order];
1313                 bool can_steal;
1314
1315                 /* Job done if page is free of the right migratetype */
1316                 if (!list_empty(&area->free_list[migratetype]))
1317                         return COMPACT_SUCCESS;
1318
1319 #ifdef CONFIG_CMA
1320                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1321                 if (migratetype == MIGRATE_MOVABLE &&
1322                         !list_empty(&area->free_list[MIGRATE_CMA]))
1323                         return COMPACT_SUCCESS;
1324 #endif
1325                 /*
1326                  * Job done if allocation would steal freepages from
1327                  * other migratetype buddy lists.
1328                  */
1329                 if (find_suitable_fallback(area, order, migratetype,
1330                                                 true, &can_steal) != -1)
1331                         return COMPACT_SUCCESS;
1332         }
1333
1334         return COMPACT_NO_SUITABLE_PAGE;
1335 }
1336
1337 static enum compact_result compact_finished(struct zone *zone,
1338                         struct compact_control *cc,
1339                         const int migratetype)
1340 {
1341         int ret;
1342
1343         ret = __compact_finished(zone, cc, migratetype);
1344         trace_mm_compaction_finished(zone, cc->order, ret);
1345         if (ret == COMPACT_NO_SUITABLE_PAGE)
1346                 ret = COMPACT_CONTINUE;
1347
1348         return ret;
1349 }
1350
1351 /*
1352  * compaction_suitable: Is this suitable to run compaction on this zone now?
1353  * Returns
1354  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1355  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1356  *   COMPACT_CONTINUE - If compaction should run now
1357  */
1358 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1359                                         unsigned int alloc_flags,
1360                                         int classzone_idx,
1361                                         unsigned long wmark_target)
1362 {
1363         unsigned long watermark;
1364
1365         if (is_via_compact_memory(order))
1366                 return COMPACT_CONTINUE;
1367
1368         watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1369         /*
1370          * If watermarks for high-order allocation are already met, there
1371          * should be no need for compaction at all.
1372          */
1373         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1374                                                                 alloc_flags))
1375                 return COMPACT_SUCCESS;
1376
1377         /*
1378          * Watermarks for order-0 must be met for compaction to be able to
1379          * isolate free pages for migration targets. This means that the
1380          * watermark and alloc_flags have to match, or be more pessimistic than
1381          * the check in __isolate_free_page(). We don't use the direct
1382          * compactor's alloc_flags, as they are not relevant for freepage
1383          * isolation. We however do use the direct compactor's classzone_idx to
1384          * skip over zones where lowmem reserves would prevent allocation even
1385          * if compaction succeeds.
1386          * For costly orders, we require low watermark instead of min for
1387          * compaction to proceed to increase its chances.
1388          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1389          * suitable migration targets
1390          */
1391         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1392                                 low_wmark_pages(zone) : min_wmark_pages(zone);
1393         watermark += compact_gap(order);
1394         if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1395                                                 ALLOC_CMA, wmark_target))
1396                 return COMPACT_SKIPPED;
1397
1398         return COMPACT_CONTINUE;
1399 }
1400
1401 enum compact_result compaction_suitable(struct zone *zone, int order,
1402                                         unsigned int alloc_flags,
1403                                         int classzone_idx)
1404 {
1405         enum compact_result ret;
1406         int fragindex;
1407
1408         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1409                                     zone_page_state(zone, NR_FREE_PAGES));
1410         /*
1411          * fragmentation index determines if allocation failures are due to
1412          * low memory or external fragmentation
1413          *
1414          * index of -1000 would imply allocations might succeed depending on
1415          * watermarks, but we already failed the high-order watermark check
1416          * index towards 0 implies failure is due to lack of memory
1417          * index towards 1000 implies failure is due to fragmentation
1418          *
1419          * Only compact if a failure would be due to fragmentation. Also
1420          * ignore fragindex for non-costly orders where the alternative to
1421          * a successful reclaim/compaction is OOM. Fragindex and the
1422          * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1423          * excessive compaction for costly orders, but it should not be at the
1424          * expense of system stability.
1425          */
1426         if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
1427                 fragindex = fragmentation_index(zone, order);
1428                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1429                         ret = COMPACT_NOT_SUITABLE_ZONE;
1430         }
1431
1432         trace_mm_compaction_suitable(zone, order, ret);
1433         if (ret == COMPACT_NOT_SUITABLE_ZONE)
1434                 ret = COMPACT_SKIPPED;
1435
1436         return ret;
1437 }
1438
1439 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1440                 int alloc_flags)
1441 {
1442         struct zone *zone;
1443         struct zoneref *z;
1444
1445         /*
1446          * Make sure at least one zone would pass __compaction_suitable if we continue
1447          * retrying the reclaim.
1448          */
1449         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1450                                         ac->nodemask) {
1451                 unsigned long available;
1452                 enum compact_result compact_result;
1453
1454                 /*
1455                  * Do not consider all the reclaimable memory because we do not
1456                  * want to trash just for a single high order allocation which
1457                  * is even not guaranteed to appear even if __compaction_suitable
1458                  * is happy about the watermark check.
1459                  */
1460                 available = zone_reclaimable_pages(zone) / order;
1461                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1462                 compact_result = __compaction_suitable(zone, order, alloc_flags,
1463                                 ac_classzone_idx(ac), available);
1464                 if (compact_result != COMPACT_SKIPPED)
1465                         return true;
1466         }
1467
1468         return false;
1469 }
1470
1471 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1472 {
1473         enum compact_result ret;
1474         unsigned long start_pfn = zone->zone_start_pfn;
1475         unsigned long end_pfn = zone_end_pfn(zone);
1476         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1477         const bool sync = cc->mode != MIGRATE_ASYNC;
1478
1479         ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1480                                                         cc->classzone_idx);
1481         /* Compaction is likely to fail */
1482         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1483                 return ret;
1484
1485         /* huh, compaction_suitable is returning something unexpected */
1486         VM_BUG_ON(ret != COMPACT_CONTINUE);
1487
1488         /*
1489          * Clear pageblock skip if there were failures recently and compaction
1490          * is about to be retried after being deferred.
1491          */
1492         if (compaction_restarting(zone, cc->order))
1493                 __reset_isolation_suitable(zone);
1494
1495         /*
1496          * Setup to move all movable pages to the end of the zone. Used cached
1497          * information on where the scanners should start (unless we explicitly
1498          * want to compact the whole zone), but check that it is initialised
1499          * by ensuring the values are within zone boundaries.
1500          */
1501         if (cc->whole_zone) {
1502                 cc->migrate_pfn = start_pfn;
1503                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1504         } else {
1505                 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1506                 cc->free_pfn = zone->compact_cached_free_pfn;
1507                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1508                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1509                         zone->compact_cached_free_pfn = cc->free_pfn;
1510                 }
1511                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1512                         cc->migrate_pfn = start_pfn;
1513                         zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1514                         zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1515                 }
1516
1517                 if (cc->migrate_pfn == start_pfn)
1518                         cc->whole_zone = true;
1519         }
1520
1521         cc->last_migrated_pfn = 0;
1522
1523         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1524                                 cc->free_pfn, end_pfn, sync);
1525
1526         migrate_prep_local();
1527
1528         while ((ret = compact_finished(zone, cc, migratetype)) ==
1529                                                 COMPACT_CONTINUE) {
1530                 int err;
1531
1532                 switch (isolate_migratepages(zone, cc)) {
1533                 case ISOLATE_ABORT:
1534                         ret = COMPACT_CONTENDED;
1535                         putback_movable_pages(&cc->migratepages);
1536                         cc->nr_migratepages = 0;
1537                         goto out;
1538                 case ISOLATE_NONE:
1539                         /*
1540                          * We haven't isolated and migrated anything, but
1541                          * there might still be unflushed migrations from
1542                          * previous cc->order aligned block.
1543                          */
1544                         goto check_drain;
1545                 case ISOLATE_SUCCESS:
1546                         ;
1547                 }
1548
1549                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1550                                 compaction_free, (unsigned long)cc, cc->mode,
1551                                 MR_COMPACTION);
1552
1553                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1554                                                         &cc->migratepages);
1555
1556                 /* All pages were either migrated or will be released */
1557                 cc->nr_migratepages = 0;
1558                 if (err) {
1559                         putback_movable_pages(&cc->migratepages);
1560                         /*
1561                          * migrate_pages() may return -ENOMEM when scanners meet
1562                          * and we want compact_finished() to detect it
1563                          */
1564                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
1565                                 ret = COMPACT_CONTENDED;
1566                                 goto out;
1567                         }
1568                         /*
1569                          * We failed to migrate at least one page in the current
1570                          * order-aligned block, so skip the rest of it.
1571                          */
1572                         if (cc->direct_compaction &&
1573                                                 (cc->mode == MIGRATE_ASYNC)) {
1574                                 cc->migrate_pfn = block_end_pfn(
1575                                                 cc->migrate_pfn - 1, cc->order);
1576                                 /* Draining pcplists is useless in this case */
1577                                 cc->last_migrated_pfn = 0;
1578
1579                         }
1580                 }
1581
1582 check_drain:
1583                 /*
1584                  * Has the migration scanner moved away from the previous
1585                  * cc->order aligned block where we migrated from? If yes,
1586                  * flush the pages that were freed, so that they can merge and
1587                  * compact_finished() can detect immediately if allocation
1588                  * would succeed.
1589                  */
1590                 if (cc->order > 0 && cc->last_migrated_pfn) {
1591                         int cpu;
1592                         unsigned long current_block_start =
1593                                 block_start_pfn(cc->migrate_pfn, cc->order);
1594
1595                         if (cc->last_migrated_pfn < current_block_start) {
1596                                 cpu = get_cpu();
1597                                 lru_add_drain_cpu(cpu);
1598                                 drain_local_pages(zone);
1599                                 put_cpu();
1600                                 /* No more flushing until we migrate again */
1601                                 cc->last_migrated_pfn = 0;
1602                         }
1603                 }
1604
1605         }
1606
1607 out:
1608         /*
1609          * Release free pages and update where the free scanner should restart,
1610          * so we don't leave any returned pages behind in the next attempt.
1611          */
1612         if (cc->nr_freepages > 0) {
1613                 unsigned long free_pfn = release_freepages(&cc->freepages);
1614
1615                 cc->nr_freepages = 0;
1616                 VM_BUG_ON(free_pfn == 0);
1617                 /* The cached pfn is always the first in a pageblock */
1618                 free_pfn = pageblock_start_pfn(free_pfn);
1619                 /*
1620                  * Only go back, not forward. The cached pfn might have been
1621                  * already reset to zone end in compact_finished()
1622                  */
1623                 if (free_pfn > zone->compact_cached_free_pfn)
1624                         zone->compact_cached_free_pfn = free_pfn;
1625         }
1626
1627         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1628                                 cc->free_pfn, end_pfn, sync, ret);
1629
1630         return ret;
1631 }
1632
1633 static enum compact_result compact_zone_order(struct zone *zone, int order,
1634                 gfp_t gfp_mask, enum compact_priority prio,
1635                 unsigned int alloc_flags, int classzone_idx)
1636 {
1637         enum compact_result ret;
1638         struct compact_control cc = {
1639                 .nr_freepages = 0,
1640                 .nr_migratepages = 0,
1641                 .order = order,
1642                 .gfp_mask = gfp_mask,
1643                 .zone = zone,
1644                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1645                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1646                 .alloc_flags = alloc_flags,
1647                 .classzone_idx = classzone_idx,
1648                 .direct_compaction = true,
1649                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
1650                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1651                 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1652         };
1653         INIT_LIST_HEAD(&cc.freepages);
1654         INIT_LIST_HEAD(&cc.migratepages);
1655
1656         ret = compact_zone(zone, &cc);
1657
1658         VM_BUG_ON(!list_empty(&cc.freepages));
1659         VM_BUG_ON(!list_empty(&cc.migratepages));
1660
1661         return ret;
1662 }
1663
1664 int sysctl_extfrag_threshold = 500;
1665
1666 /**
1667  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1668  * @gfp_mask: The GFP mask of the current allocation
1669  * @order: The order of the current allocation
1670  * @alloc_flags: The allocation flags of the current allocation
1671  * @ac: The context of current allocation
1672  * @mode: The migration mode for async, sync light, or sync migration
1673  *
1674  * This is the main entry point for direct page compaction.
1675  */
1676 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1677                 unsigned int alloc_flags, const struct alloc_context *ac,
1678                 enum compact_priority prio)
1679 {
1680         int may_enter_fs = gfp_mask & __GFP_FS;
1681         int may_perform_io = gfp_mask & __GFP_IO;
1682         struct zoneref *z;
1683         struct zone *zone;
1684         enum compact_result rc = COMPACT_SKIPPED;
1685
1686         /* Check if the GFP flags allow compaction */
1687         if (!may_enter_fs || !may_perform_io)
1688                 return COMPACT_SKIPPED;
1689
1690         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1691
1692         /* Compact each zone in the list */
1693         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1694                                                                 ac->nodemask) {
1695                 enum compact_result status;
1696
1697                 if (prio > MIN_COMPACT_PRIORITY
1698                                         && compaction_deferred(zone, order)) {
1699                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1700                         continue;
1701                 }
1702
1703                 status = compact_zone_order(zone, order, gfp_mask, prio,
1704                                         alloc_flags, ac_classzone_idx(ac));
1705                 rc = max(status, rc);
1706
1707                 /* The allocation should succeed, stop compacting */
1708                 if (status == COMPACT_SUCCESS) {
1709                         /*
1710                          * We think the allocation will succeed in this zone,
1711                          * but it is not certain, hence the false. The caller
1712                          * will repeat this with true if allocation indeed
1713                          * succeeds in this zone.
1714                          */
1715                         compaction_defer_reset(zone, order, false);
1716
1717                         break;
1718                 }
1719
1720                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1721                                         status == COMPACT_PARTIAL_SKIPPED))
1722                         /*
1723                          * We think that allocation won't succeed in this zone
1724                          * so we defer compaction there. If it ends up
1725                          * succeeding after all, it will be reset.
1726                          */
1727                         defer_compaction(zone, order);
1728
1729                 /*
1730                  * We might have stopped compacting due to need_resched() in
1731                  * async compaction, or due to a fatal signal detected. In that
1732                  * case do not try further zones
1733                  */
1734                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1735                                         || fatal_signal_pending(current))
1736                         break;
1737         }
1738
1739         return rc;
1740 }
1741
1742
1743 /* Compact all zones within a node */
1744 static void compact_node(int nid)
1745 {
1746         pg_data_t *pgdat = NODE_DATA(nid);
1747         int zoneid;
1748         struct zone *zone;
1749         struct compact_control cc = {
1750                 .order = -1,
1751                 .mode = MIGRATE_SYNC,
1752                 .ignore_skip_hint = true,
1753                 .whole_zone = true,
1754         };
1755
1756
1757         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1758
1759                 zone = &pgdat->node_zones[zoneid];
1760                 if (!populated_zone(zone))
1761                         continue;
1762
1763                 cc.nr_freepages = 0;
1764                 cc.nr_migratepages = 0;
1765                 cc.zone = zone;
1766                 INIT_LIST_HEAD(&cc.freepages);
1767                 INIT_LIST_HEAD(&cc.migratepages);
1768
1769                 compact_zone(zone, &cc);
1770
1771                 VM_BUG_ON(!list_empty(&cc.freepages));
1772                 VM_BUG_ON(!list_empty(&cc.migratepages));
1773         }
1774 }
1775
1776 /* Compact all nodes in the system */
1777 static void compact_nodes(void)
1778 {
1779         int nid;
1780
1781         /* Flush pending updates to the LRU lists */
1782         lru_add_drain_all();
1783
1784         for_each_online_node(nid)
1785                 compact_node(nid);
1786 }
1787
1788 /* The written value is actually unused, all memory is compacted */
1789 int sysctl_compact_memory;
1790
1791 /*
1792  * This is the entry point for compacting all nodes via
1793  * /proc/sys/vm/compact_memory
1794  */
1795 int sysctl_compaction_handler(struct ctl_table *table, int write,
1796                         void __user *buffer, size_t *length, loff_t *ppos)
1797 {
1798         if (write)
1799                 compact_nodes();
1800
1801         return 0;
1802 }
1803
1804 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1805                         void __user *buffer, size_t *length, loff_t *ppos)
1806 {
1807         proc_dointvec_minmax(table, write, buffer, length, ppos);
1808
1809         return 0;
1810 }
1811
1812 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1813 static ssize_t sysfs_compact_node(struct device *dev,
1814                         struct device_attribute *attr,
1815                         const char *buf, size_t count)
1816 {
1817         int nid = dev->id;
1818
1819         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1820                 /* Flush pending updates to the LRU lists */
1821                 lru_add_drain_all();
1822
1823                 compact_node(nid);
1824         }
1825
1826         return count;
1827 }
1828 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1829
1830 int compaction_register_node(struct node *node)
1831 {
1832         return device_create_file(&node->dev, &dev_attr_compact);
1833 }
1834
1835 void compaction_unregister_node(struct node *node)
1836 {
1837         return device_remove_file(&node->dev, &dev_attr_compact);
1838 }
1839 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1840
1841 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1842 {
1843         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1844 }
1845
1846 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1847 {
1848         int zoneid;
1849         struct zone *zone;
1850         enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1851
1852         for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1853                 zone = &pgdat->node_zones[zoneid];
1854
1855                 if (!populated_zone(zone))
1856                         continue;
1857
1858                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1859                                         classzone_idx) == COMPACT_CONTINUE)
1860                         return true;
1861         }
1862
1863         return false;
1864 }
1865
1866 static void kcompactd_do_work(pg_data_t *pgdat)
1867 {
1868         /*
1869          * With no special task, compact all zones so that a page of requested
1870          * order is allocatable.
1871          */
1872         int zoneid;
1873         struct zone *zone;
1874         struct compact_control cc = {
1875                 .order = pgdat->kcompactd_max_order,
1876                 .classzone_idx = pgdat->kcompactd_classzone_idx,
1877                 .mode = MIGRATE_SYNC_LIGHT,
1878                 .ignore_skip_hint = true,
1879
1880         };
1881         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1882                                                         cc.classzone_idx);
1883         count_vm_event(KCOMPACTD_WAKE);
1884
1885         for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1886                 int status;
1887
1888                 zone = &pgdat->node_zones[zoneid];
1889                 if (!populated_zone(zone))
1890                         continue;
1891
1892                 if (compaction_deferred(zone, cc.order))
1893                         continue;
1894
1895                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1896                                                         COMPACT_CONTINUE)
1897                         continue;
1898
1899                 cc.nr_freepages = 0;
1900                 cc.nr_migratepages = 0;
1901                 cc.zone = zone;
1902                 INIT_LIST_HEAD(&cc.freepages);
1903                 INIT_LIST_HEAD(&cc.migratepages);
1904
1905                 if (kthread_should_stop())
1906                         return;
1907                 status = compact_zone(zone, &cc);
1908
1909                 if (status == COMPACT_SUCCESS) {
1910                         compaction_defer_reset(zone, cc.order, false);
1911                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1912                         /*
1913                          * We use sync migration mode here, so we defer like
1914                          * sync direct compaction does.
1915                          */
1916                         defer_compaction(zone, cc.order);
1917                 }
1918
1919                 VM_BUG_ON(!list_empty(&cc.freepages));
1920                 VM_BUG_ON(!list_empty(&cc.migratepages));
1921         }
1922
1923         /*
1924          * Regardless of success, we are done until woken up next. But remember
1925          * the requested order/classzone_idx in case it was higher/tighter than
1926          * our current ones
1927          */
1928         if (pgdat->kcompactd_max_order <= cc.order)
1929                 pgdat->kcompactd_max_order = 0;
1930         if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1931                 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1932 }
1933
1934 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1935 {
1936         if (!order)
1937                 return;
1938
1939         if (pgdat->kcompactd_max_order < order)
1940                 pgdat->kcompactd_max_order = order;
1941
1942         if (pgdat->kcompactd_classzone_idx > classzone_idx)
1943                 pgdat->kcompactd_classzone_idx = classzone_idx;
1944
1945         if (!waitqueue_active(&pgdat->kcompactd_wait))
1946                 return;
1947
1948         if (!kcompactd_node_suitable(pgdat))
1949                 return;
1950
1951         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1952                                                         classzone_idx);
1953         wake_up_interruptible(&pgdat->kcompactd_wait);
1954 }
1955
1956 /*
1957  * The background compaction daemon, started as a kernel thread
1958  * from the init process.
1959  */
1960 static int kcompactd(void *p)
1961 {
1962         pg_data_t *pgdat = (pg_data_t*)p;
1963         struct task_struct *tsk = current;
1964
1965         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1966
1967         if (!cpumask_empty(cpumask))
1968                 set_cpus_allowed_ptr(tsk, cpumask);
1969
1970         set_freezable();
1971
1972         pgdat->kcompactd_max_order = 0;
1973         pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1974
1975         while (!kthread_should_stop()) {
1976                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1977                 wait_event_freezable(pgdat->kcompactd_wait,
1978                                 kcompactd_work_requested(pgdat));
1979
1980                 kcompactd_do_work(pgdat);
1981         }
1982
1983         return 0;
1984 }
1985
1986 /*
1987  * This kcompactd start function will be called by init and node-hot-add.
1988  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
1989  */
1990 int kcompactd_run(int nid)
1991 {
1992         pg_data_t *pgdat = NODE_DATA(nid);
1993         int ret = 0;
1994
1995         if (pgdat->kcompactd)
1996                 return 0;
1997
1998         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
1999         if (IS_ERR(pgdat->kcompactd)) {
2000                 pr_err("Failed to start kcompactd on node %d\n", nid);
2001                 ret = PTR_ERR(pgdat->kcompactd);
2002                 pgdat->kcompactd = NULL;
2003         }
2004         return ret;
2005 }
2006
2007 /*
2008  * Called by memory hotplug when all memory in a node is offlined. Caller must
2009  * hold mem_hotplug_begin/end().
2010  */
2011 void kcompactd_stop(int nid)
2012 {
2013         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2014
2015         if (kcompactd) {
2016                 kthread_stop(kcompactd);
2017                 NODE_DATA(nid)->kcompactd = NULL;
2018         }
2019 }
2020
2021 /*
2022  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2023  * not required for correctness. So if the last cpu in a node goes
2024  * away, we get changed to run anywhere: as the first one comes back,
2025  * restore their cpu bindings.
2026  */
2027 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2028                         void *hcpu)
2029 {
2030         int nid;
2031
2032         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2033                 for_each_node_state(nid, N_MEMORY) {
2034                         pg_data_t *pgdat = NODE_DATA(nid);
2035                         const struct cpumask *mask;
2036
2037                         mask = cpumask_of_node(pgdat->node_id);
2038
2039                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2040                                 /* One of our CPUs online: restore mask */
2041                                 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2042                 }
2043         }
2044         return NOTIFY_OK;
2045 }
2046
2047 static int __init kcompactd_init(void)
2048 {
2049         int nid;
2050
2051         for_each_node_state(nid, N_MEMORY)
2052                 kcompactd_run(nid);
2053         hotcpu_notifier(cpu_callback, 0);
2054         return 0;
2055 }
2056 subsys_initcall(kcompactd_init)
2057
2058 #endif /* CONFIG_COMPACTION */