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 /* Update the number of anon and file isolated pages in the zone */
 638 static 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 */
 654 static 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  */
 686 static unsigned long
 687 isolate_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
 870 isolate_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;
 891 isolate_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  */
 959 unsigned long
 960 isolate_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 */
1000 static bool suitable_migration_target(struct page *page)
1001 {
1002         /* If the page is a large free page, then disallow migration */
1003         if (PageBuddy(page)) {
1004                 /*
1005                  * We are checking page_order without zone->lock taken. But
1006                  * the only small danger is that we skip a potentially suitable
1007                  * pageblock, so it's not worth to check order for valid range.
1008                  */
1009                 if (page_order_unsafe(page) >= pageblock_order)
1010                         return false;
1011         }
1012
1013         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1014         if (migrate_async_suitable(get_pageblock_migratetype(page)))
1015                 return true;
1016
1017         /* Otherwise skip the block */
1018         return false;
1019 }
1020
1021 /*
1022  * Test whether the free scanner has reached the same or lower pageblock than
1023  * the migration scanner, and compaction should thus terminate.
1024  */
1025 static inline bool compact_scanners_met(struct compact_control *cc)
1026 {
1027         return (cc->free_pfn >> pageblock_order)
1028                 <= (cc->migrate_pfn >> pageblock_order);
1029 }
1030
1031 /*
1032  * Based on information in the current compact_control, find blocks
1033  * suitable for isolating free pages from and then isolate them.
1034  */
1035 static void isolate_freepages(struct compact_control *cc)
1036 {
1037         struct zone *zone = cc->zone;
1038         struct page *page;
1039         unsigned long block_start_pfn;  /* start of current pageblock */
1040         unsigned long isolate_start_pfn; /* exact pfn we start at */
1041         unsigned long block_end_pfn;    /* end of current pageblock */
1042         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1043         struct list_head *freelist = &cc->freepages;
1044
1045         /*
1046          * Initialise the free scanner. The starting point is where we last
1047          * successfully isolated from, zone-cached value, or the end of the
1048          * zone when isolating for the first time. For looping we also need
1049          * this pfn aligned down to the pageblock boundary, because we do
1050          * block_start_pfn -= pageblock_nr_pages in the for loop.
1051          * For ending point, take care when isolating in last pageblock of a
1052          * a zone which ends in the middle of a pageblock.
1053          * The low boundary is the end of the pageblock the migration scanner
1054          * is using.
1055          */
1056         isolate_start_pfn = cc->free_pfn;
1057         block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1058         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1059                                                 zone_end_pfn(zone));
1060         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1061
1062         /*
1063          * Isolate free pages until enough are available to migrate the
1064          * pages on cc->migratepages. We stop searching if the migrate
1065          * and free page scanners meet or enough free pages are isolated.
1066          */
1067         for (; block_start_pfn >= low_pfn;
1068                                 block_end_pfn = block_start_pfn,
1069                                 block_start_pfn -= pageblock_nr_pages,
1070                                 isolate_start_pfn = block_start_pfn) {
1071                 /*
1072                  * This can iterate a massively long zone without finding any
1073                  * suitable migration targets, so periodically check if we need
1074                  * to schedule, or even abort async compaction.
1075                  */
1076                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1077                                                 && compact_should_abort(cc))
1078                         break;
1079
1080                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1081                                                                         zone);
1082                 if (!page)
1083                         continue;
1084
1085                 /* Check the block is suitable for migration */
1086                 if (!suitable_migration_target(page))
1087                         continue;
1088
1089                 /* If isolation recently failed, do not retry */
1090                 if (!isolation_suitable(cc, page))
1091                         continue;
1092
1093                 /* Found a block suitable for isolating free pages from. */
1094                 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1095                                         freelist, false);
1096
1097                 /*
1098                  * If we isolated enough freepages, or aborted due to lock
1099                  * contention, terminate.
1100                  */
1101                 if ((cc->nr_freepages >= cc->nr_migratepages)
1102                                                         || cc->contended) {
1103                         if (isolate_start_pfn >= block_end_pfn) {
1104                                 /*
1105                                  * Restart at previous pageblock if more
1106                                  * freepages can be isolated next time.
1107                                  */
1108                                 isolate_start_pfn =
1109                                         block_start_pfn - pageblock_nr_pages;
1110                         }
1111                         break;
1112                 } else if (isolate_start_pfn < block_end_pfn) {
1113                         /*
1114                          * If isolation failed early, do not continue
1115                          * needlessly.
1116                          */
1117                         break;
1118                 }
1119         }
1120
1121         /* __isolate_free_page() does not map the pages */
1122         map_pages(freelist);
1123
1124         /*
1125          * Record where the free scanner will restart next time. Either we
1126          * broke from the loop and set isolate_start_pfn based on the last
1127          * call to isolate_freepages_block(), or we met the migration scanner
1128          * and the loop terminated due to isolate_start_pfn < low_pfn
1129          */
1130         cc->free_pfn = isolate_start_pfn;
1131 }
1132
1133 /*
1134  * This is a migrate-callback that "allocates" freepages by taking pages
1135  * from the isolated freelists in the block we are migrating to.
1136  */
1137 static struct page *compaction_alloc(struct page *migratepage,
1138                                         unsigned long data,
1139                                         int **result)
1140 {
1141         struct compact_control *cc = (struct compact_control *)data;
1142         struct page *freepage;
1143
1144         /*
1145          * Isolate free pages if necessary, and if we are not aborting due to
1146          * contention.
1147          */
1148         if (list_empty(&cc->freepages)) {
1149                 if (!cc->contended)
1150                         isolate_freepages(cc);
1151
1152                 if (list_empty(&cc->freepages))
1153                         return NULL;
1154         }
1155
1156         freepage = list_entry(cc->freepages.next, struct page, lru);
1157         list_del(&freepage->lru);
1158         cc->nr_freepages--;
1159
1160         return freepage;
1161 }
1162
1163 /*
1164  * This is a migrate-callback that "frees" freepages back to the isolated
1165  * freelist.  All pages on the freelist are from the same zone, so there is no
1166  * special handling needed for NUMA.
1167  */
1168 static void compaction_free(struct page *page, unsigned long data)
1169 {
1170         struct compact_control *cc = (struct compact_control *)data;
1171
1172         list_add(&page->lru, &cc->freepages);
1173         cc->nr_freepages++;
1174 }
1175
1176 /* possible outcome of isolate_migratepages */
1177 typedef enum {
1178         ISOLATE_ABORT,          /* Abort compaction now */
1179         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1180         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1181 } isolate_migrate_t;
1182
1183 /*
1184  * Allow userspace to control policy on scanning the unevictable LRU for
1185  * compactable pages.
1186  */
1187 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1188
1189 /*
1190  * Isolate all pages that can be migrated from the first suitable block,
1191  * starting at the block pointed to by the migrate scanner pfn within
1192  * compact_control.
1193  */
1194 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1195                                         struct compact_control *cc)
1196 {
1197         unsigned long block_start_pfn;
1198         unsigned long block_end_pfn;
1199         unsigned long low_pfn;
1200         struct page *page;
1201         const isolate_mode_t isolate_mode =
1202                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1203                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1204
1205         /*
1206          * Start at where we last stopped, or beginning of the zone as
1207          * initialized by compact_zone()
1208          */
1209         low_pfn = cc->migrate_pfn;
1210         block_start_pfn = pageblock_start_pfn(low_pfn);
1211         if (block_start_pfn < zone->zone_start_pfn)
1212                 block_start_pfn = zone->zone_start_pfn;
1213
1214         /* Only scan within a pageblock boundary */
1215         block_end_pfn = pageblock_end_pfn(low_pfn);
1216
1217         /*
1218          * Iterate over whole pageblocks until we find the first suitable.
1219          * Do not cross the free scanner.
1220          */
1221         for (; block_end_pfn <= cc->free_pfn;
1222                         low_pfn = block_end_pfn,
1223                         block_start_pfn = block_end_pfn,
1224                         block_end_pfn += pageblock_nr_pages) {
1225
1226                 /*
1227                  * This can potentially iterate a massively long zone with
1228                  * many pageblocks unsuitable, so periodically check if we
1229                  * need to schedule, or even abort async compaction.
1230                  */
1231                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1232                                                 && compact_should_abort(cc))
1233                         break;
1234
1235                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1236                                                                         zone);
1237                 if (!page)
1238                         continue;
1239
1240                 /* If isolation recently failed, do not retry */
1241                 if (!isolation_suitable(cc, page))
1242                         continue;
1243
1244                 /*
1245                  * For async compaction, also only scan in MOVABLE blocks.
1246                  * Async compaction is optimistic to see if the minimum amount
1247                  * of work satisfies the allocation.
1248                  */
1249                 if (cc->mode == MIGRATE_ASYNC &&
1250                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1251                         continue;
1252
1253                 /* Perform the isolation */
1254                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1255                                                 block_end_pfn, isolate_mode);
1256
1257                 if (!low_pfn || cc->contended) {
1258                         acct_isolated(zone, cc);
1259                         return ISOLATE_ABORT;
1260                 }
1261
1262                 /*
1263                  * Either we isolated something and proceed with migration. Or
1264                  * we failed and compact_zone should decide if we should
1265                  * continue or not.
1266                  */
1267                 break;
1268         }
1269
1270         acct_isolated(zone, cc);
1271         /* Record where migration scanner will be restarted. */
1272         cc->migrate_pfn = low_pfn;
1273
1274         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1275 }
1276
1277 /*
1278  * order == -1 is expected when compacting via
1279  * /proc/sys/vm/compact_memory
1280  */
1281 static inline bool is_via_compact_memory(int order)
1282 {
1283         return order == -1;
1284 }
1285
1286 static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1287                             const int migratetype)
1288 {
1289         unsigned int order;
1290         unsigned long watermark;
1291
1292         if (cc->contended || fatal_signal_pending(current))
1293                 return COMPACT_CONTENDED;
1294
1295         /* Compaction run completes if the migrate and free scanner meet */
1296         if (compact_scanners_met(cc)) {
1297                 /* Let the next compaction start anew. */
1298                 reset_cached_positions(zone);
1299
1300                 /*
1301                  * Mark that the PG_migrate_skip information should be cleared
1302                  * by kswapd when it goes to sleep. kcompactd does not set the
1303                  * flag itself as the decision to be clear should be directly
1304                  * based on an allocation request.
1305                  */
1306                 if (cc->direct_compaction)
1307                         zone->compact_blockskip_flush = true;
1308
1309                 if (cc->whole_zone)
1310                         return COMPACT_COMPLETE;
1311                 else
1312                         return COMPACT_PARTIAL_SKIPPED;
1313         }
1314
1315         if (is_via_compact_memory(cc->order))
1316                 return COMPACT_CONTINUE;
1317
1318         /* Compaction run is not finished if the watermark is not met */
1319         watermark = zone->watermark[cc->alloc_flags & ALLOC_WMARK_MASK];
1320
1321         if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1322                                                         cc->alloc_flags))
1323                 return COMPACT_CONTINUE;
1324
1325         /* Direct compactor: Is a suitable page free? */
1326         for (order = cc->order; order < MAX_ORDER; order++) {
1327                 struct free_area *area = &zone->free_area[order];
1328                 bool can_steal;
1329
1330                 /* Job done if page is free of the right migratetype */
1331                 if (!list_empty(&area->free_list[migratetype]))
1332                         return COMPACT_SUCCESS;
1333
1334 #ifdef CONFIG_CMA
1335                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1336                 if (migratetype == MIGRATE_MOVABLE &&
1337                         !list_empty(&area->free_list[MIGRATE_CMA]))
1338                         return COMPACT_SUCCESS;
1339 #endif
1340                 /*
1341                  * Job done if allocation would steal freepages from
1342                  * other migratetype buddy lists.
1343                  */
1344                 if (find_suitable_fallback(area, order, migratetype,
1345                                                 true, &can_steal) != -1)
1346                         return COMPACT_SUCCESS;
1347         }
1348
1349         return COMPACT_NO_SUITABLE_PAGE;
1350 }
1351
1352 static enum compact_result compact_finished(struct zone *zone,
1353                         struct compact_control *cc,
1354                         const int migratetype)
1355 {
1356         int ret;
1357
1358         ret = __compact_finished(zone, cc, migratetype);
1359         trace_mm_compaction_finished(zone, cc->order, ret);
1360         if (ret == COMPACT_NO_SUITABLE_PAGE)
1361                 ret = COMPACT_CONTINUE;
1362
1363         return ret;
1364 }
1365
1366 /*
1367  * compaction_suitable: Is this suitable to run compaction on this zone now?
1368  * Returns
1369  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1370  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1371  *   COMPACT_CONTINUE - If compaction should run now
1372  */
1373 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1374                                         unsigned int alloc_flags,
1375                                         int classzone_idx,
1376                                         unsigned long wmark_target)
1377 {
1378         int fragindex;
1379         unsigned long watermark;
1380
1381         if (is_via_compact_memory(order))
1382                 return COMPACT_CONTINUE;
1383
1384         watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1385         /*
1386          * If watermarks for high-order allocation are already met, there
1387          * should be no need for compaction at all.
1388          */
1389         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1390                                                                 alloc_flags))
1391                 return COMPACT_SUCCESS;
1392
1393         /*
1394          * Watermarks for order-0 must be met for compaction to be able to
1395          * isolate free pages for migration targets. This means that the
1396          * watermark and alloc_flags have to match, or be more pessimistic than
1397          * the check in __isolate_free_page(). We don't use the direct
1398          * compactor's alloc_flags, as they are not relevant for freepage
1399          * isolation. We however do use the direct compactor's classzone_idx to
1400          * skip over zones where lowmem reserves would prevent allocation even
1401          * if compaction succeeds.
1402          * For costly orders, we require low watermark instead of min for
1403          * compaction to proceed to increase its chances.
1404          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1405          * suitable migration targets
1406          */
1407         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1408                                 low_wmark_pages(zone) : min_wmark_pages(zone);
1409         watermark += compact_gap(order);
1410         if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1411                                                 ALLOC_CMA, wmark_target))
1412                 return COMPACT_SKIPPED;
1413
1414         /*
1415          * fragmentation index determines if allocation failures are due to
1416          * low memory or external fragmentation
1417          *
1418          * index of -1000 would imply allocations might succeed depending on
1419          * watermarks, but we already failed the high-order watermark check
1420          * index towards 0 implies failure is due to lack of memory
1421          * index towards 1000 implies failure is due to fragmentation
1422          *
1423          * Only compact if a failure would be due to fragmentation.
1424          */
1425         fragindex = fragmentation_index(zone, order);
1426         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1427                 return COMPACT_NOT_SUITABLE_ZONE;
1428
1429         return COMPACT_CONTINUE;
1430 }
1431
1432 enum compact_result compaction_suitable(struct zone *zone, int order,
1433                                         unsigned int alloc_flags,
1434                                         int classzone_idx)
1435 {
1436         enum compact_result ret;
1437
1438         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1439                                     zone_page_state(zone, NR_FREE_PAGES));
1440         trace_mm_compaction_suitable(zone, order, ret);
1441         if (ret == COMPACT_NOT_SUITABLE_ZONE)
1442                 ret = COMPACT_SKIPPED;
1443
1444         return ret;
1445 }
1446
1447 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1448                 int alloc_flags)
1449 {
1450         struct zone *zone;
1451         struct zoneref *z;
1452
1453         /*
1454          * Make sure at least one zone would pass __compaction_suitable if we continue
1455          * retrying the reclaim.
1456          */
1457         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1458                                         ac->nodemask) {
1459                 unsigned long available;
1460                 enum compact_result compact_result;
1461
1462                 /*
1463                  * Do not consider all the reclaimable memory because we do not
1464                  * want to trash just for a single high order allocation which
1465                  * is even not guaranteed to appear even if __compaction_suitable
1466                  * is happy about the watermark check.
1467                  */
1468                 available = zone_reclaimable_pages(zone) / order;
1469                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1470                 compact_result = __compaction_suitable(zone, order, alloc_flags,
1471                                 ac_classzone_idx(ac), available);
1472                 if (compact_result != COMPACT_SKIPPED &&
1473                                 compact_result != COMPACT_NOT_SUITABLE_ZONE)
1474                         return true;
1475         }
1476
1477         return false;
1478 }
1479
1480 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1481 {
1482         enum compact_result ret;
1483         unsigned long start_pfn = zone->zone_start_pfn;
1484         unsigned long end_pfn = zone_end_pfn(zone);
1485         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1486         const bool sync = cc->mode != MIGRATE_ASYNC;
1487
1488         ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1489                                                         cc->classzone_idx);
1490         /* Compaction is likely to fail */
1491         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1492                 return ret;
1493
1494         /* huh, compaction_suitable is returning something unexpected */
1495         VM_BUG_ON(ret != COMPACT_CONTINUE);
1496
1497         /*
1498          * Clear pageblock skip if there were failures recently and compaction
1499          * is about to be retried after being deferred.
1500          */
1501         if (compaction_restarting(zone, cc->order))
1502                 __reset_isolation_suitable(zone);
1503
1504         /*
1505          * Setup to move all movable pages to the end of the zone. Used cached
1506          * information on where the scanners should start (unless we explicitly
1507          * want to compact the whole zone), but check that it is initialised
1508          * by ensuring the values are within zone boundaries.
1509          */
1510         if (cc->whole_zone) {
1511                 cc->migrate_pfn = start_pfn;
1512                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1513         } else {
1514                 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1515                 cc->free_pfn = zone->compact_cached_free_pfn;
1516                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1517                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1518                         zone->compact_cached_free_pfn = cc->free_pfn;
1519                 }
1520                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1521                         cc->migrate_pfn = start_pfn;
1522                         zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1523                         zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1524                 }
1525
1526                 if (cc->migrate_pfn == start_pfn)
1527                         cc->whole_zone = true;
1528         }
1529
1530         cc->last_migrated_pfn = 0;
1531
1532         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1533                                 cc->free_pfn, end_pfn, sync);
1534
1535         migrate_prep_local();
1536
1537         while ((ret = compact_finished(zone, cc, migratetype)) ==
1538                                                 COMPACT_CONTINUE) {
1539                 int err;
1540
1541                 switch (isolate_migratepages(zone, cc)) {
1542                 case ISOLATE_ABORT:
1543                         ret = COMPACT_CONTENDED;
1544                         putback_movable_pages(&cc->migratepages);
1545                         cc->nr_migratepages = 0;
1546                         goto out;
1547                 case ISOLATE_NONE:
1548                         /*
1549                          * We haven't isolated and migrated anything, but
1550                          * there might still be unflushed migrations from
1551                          * previous cc->order aligned block.
1552                          */
1553                         goto check_drain;
1554                 case ISOLATE_SUCCESS:
1555                         ;
1556                 }
1557
1558                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1559                                 compaction_free, (unsigned long)cc, cc->mode,
1560                                 MR_COMPACTION);
1561
1562                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1563                                                         &cc->migratepages);
1564
1565                 /* All pages were either migrated or will be released */
1566                 cc->nr_migratepages = 0;
1567                 if (err) {
1568                         putback_movable_pages(&cc->migratepages);
1569                         /*
1570                          * migrate_pages() may return -ENOMEM when scanners meet
1571                          * and we want compact_finished() to detect it
1572                          */
1573                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
1574                                 ret = COMPACT_CONTENDED;
1575                                 goto out;
1576                         }
1577                         /*
1578                          * We failed to migrate at least one page in the current
1579                          * order-aligned block, so skip the rest of it.
1580                          */
1581                         if (cc->direct_compaction &&
1582                                                 (cc->mode == MIGRATE_ASYNC)) {
1583                                 cc->migrate_pfn = block_end_pfn(
1584                                                 cc->migrate_pfn - 1, cc->order);
1585                                 /* Draining pcplists is useless in this case */
1586                                 cc->last_migrated_pfn = 0;
1587
1588                         }
1589                 }
1590
1591 check_drain:
1592                 /*
1593                  * Has the migration scanner moved away from the previous
1594                  * cc->order aligned block where we migrated from? If yes,
1595                  * flush the pages that were freed, so that they can merge and
1596                  * compact_finished() can detect immediately if allocation
1597                  * would succeed.
1598                  */
1599                 if (cc->order > 0 && cc->last_migrated_pfn) {
1600                         int cpu;
1601                         unsigned long current_block_start =
1602                                 block_start_pfn(cc->migrate_pfn, cc->order);
1603
1604                         if (cc->last_migrated_pfn < current_block_start) {
1605                                 cpu = get_cpu();
1606                                 lru_add_drain_cpu(cpu);
1607                                 drain_local_pages(zone);
1608                                 put_cpu();
1609                                 /* No more flushing until we migrate again */
1610                                 cc->last_migrated_pfn = 0;
1611                         }
1612                 }
1613
1614         }
1615
1616 out:
1617         /*
1618          * Release free pages and update where the free scanner should restart,
1619          * so we don't leave any returned pages behind in the next attempt.
1620          */
1621         if (cc->nr_freepages > 0) {
1622                 unsigned long free_pfn = release_freepages(&cc->freepages);
1623
1624                 cc->nr_freepages = 0;
1625                 VM_BUG_ON(free_pfn == 0);
1626                 /* The cached pfn is always the first in a pageblock */
1627                 free_pfn = pageblock_start_pfn(free_pfn);
1628                 /*
1629                  * Only go back, not forward. The cached pfn might have been
1630                  * already reset to zone end in compact_finished()
1631                  */
1632                 if (free_pfn > zone->compact_cached_free_pfn)
1633                         zone->compact_cached_free_pfn = free_pfn;
1634         }
1635
1636         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1637                                 cc->free_pfn, end_pfn, sync, ret);
1638
1639         return ret;
1640 }
1641
1642 static enum compact_result compact_zone_order(struct zone *zone, int order,
1643                 gfp_t gfp_mask, enum compact_priority prio,
1644                 unsigned int alloc_flags, int classzone_idx)
1645 {
1646         enum compact_result ret;
1647         struct compact_control cc = {
1648                 .nr_freepages = 0,
1649                 .nr_migratepages = 0,
1650                 .order = order,
1651                 .gfp_mask = gfp_mask,
1652                 .zone = zone,
1653                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1654                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1655                 .alloc_flags = alloc_flags,
1656                 .classzone_idx = classzone_idx,
1657                 .direct_compaction = true,
1658                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
1659                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY)
1660         };
1661         INIT_LIST_HEAD(&cc.freepages);
1662         INIT_LIST_HEAD(&cc.migratepages);
1663
1664         ret = compact_zone(zone, &cc);
1665
1666         VM_BUG_ON(!list_empty(&cc.freepages));
1667         VM_BUG_ON(!list_empty(&cc.migratepages));
1668
1669         return ret;
1670 }
1671
1672 int sysctl_extfrag_threshold = 500;
1673
1674 /**
1675  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1676  * @gfp_mask: The GFP mask of the current allocation
1677  * @order: The order of the current allocation
1678  * @alloc_flags: The allocation flags of the current allocation
1679  * @ac: The context of current allocation
1680  * @mode: The migration mode for async, sync light, or sync migration
1681  *
1682  * This is the main entry point for direct page compaction.
1683  */
1684 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1685                 unsigned int alloc_flags, const struct alloc_context *ac,
1686                 enum compact_priority prio)
1687 {
1688         int may_enter_fs = gfp_mask & __GFP_FS;
1689         int may_perform_io = gfp_mask & __GFP_IO;
1690         struct zoneref *z;
1691         struct zone *zone;
1692         enum compact_result rc = COMPACT_SKIPPED;
1693
1694         /* Check if the GFP flags allow compaction */
1695         if (!may_enter_fs || !may_perform_io)
1696                 return COMPACT_SKIPPED;
1697
1698         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1699
1700         /* Compact each zone in the list */
1701         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1702                                                                 ac->nodemask) {
1703                 enum compact_result status;
1704
1705                 if (prio > MIN_COMPACT_PRIORITY
1706                                         && compaction_deferred(zone, order)) {
1707                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1708                         continue;
1709                 }
1710
1711                 status = compact_zone_order(zone, order, gfp_mask, prio,
1712                                         alloc_flags, ac_classzone_idx(ac));
1713                 rc = max(status, rc);
1714
1715                 /* The allocation should succeed, stop compacting */
1716                 if (status == COMPACT_SUCCESS) {
1717                         /*
1718                          * We think the allocation will succeed in this zone,
1719                          * but it is not certain, hence the false. The caller
1720                          * will repeat this with true if allocation indeed
1721                          * succeeds in this zone.
1722                          */
1723                         compaction_defer_reset(zone, order, false);
1724
1725                         break;
1726                 }
1727
1728                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1729                                         status == COMPACT_PARTIAL_SKIPPED))
1730                         /*
1731                          * We think that allocation won't succeed in this zone
1732                          * so we defer compaction there. If it ends up
1733                          * succeeding after all, it will be reset.
1734                          */
1735                         defer_compaction(zone, order);
1736
1737                 /*
1738                  * We might have stopped compacting due to need_resched() in
1739                  * async compaction, or due to a fatal signal detected. In that
1740                  * case do not try further zones
1741                  */
1742                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1743                                         || fatal_signal_pending(current))
1744                         break;
1745         }
1746
1747         return rc;
1748 }
1749
1750
1751 /* Compact all zones within a node */
1752 static void compact_node(int nid)
1753 {
1754         pg_data_t *pgdat = NODE_DATA(nid);
1755         int zoneid;
1756         struct zone *zone;
1757         struct compact_control cc = {
1758                 .order = -1,
1759                 .mode = MIGRATE_SYNC,
1760                 .ignore_skip_hint = true,
1761                 .whole_zone = true,
1762         };
1763
1764
1765         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1766
1767                 zone = &pgdat->node_zones[zoneid];
1768                 if (!populated_zone(zone))
1769                         continue;
1770
1771                 cc.nr_freepages = 0;
1772                 cc.nr_migratepages = 0;
1773                 cc.zone = zone;
1774                 INIT_LIST_HEAD(&cc.freepages);
1775                 INIT_LIST_HEAD(&cc.migratepages);
1776
1777                 compact_zone(zone, &cc);
1778
1779                 VM_BUG_ON(!list_empty(&cc.freepages));
1780                 VM_BUG_ON(!list_empty(&cc.migratepages));
1781         }
1782 }
1783
1784 /* Compact all nodes in the system */
1785 static void compact_nodes(void)
1786 {
1787         int nid;
1788
1789         /* Flush pending updates to the LRU lists */
1790         lru_add_drain_all();
1791
1792         for_each_online_node(nid)
1793                 compact_node(nid);
1794 }
1795
1796 /* The written value is actually unused, all memory is compacted */
1797 int sysctl_compact_memory;
1798
1799 /*
1800  * This is the entry point for compacting all nodes via
1801  * /proc/sys/vm/compact_memory
1802  */
1803 int sysctl_compaction_handler(struct ctl_table *table, int write,
1804                         void __user *buffer, size_t *length, loff_t *ppos)
1805 {
1806         if (write)
1807                 compact_nodes();
1808
1809         return 0;
1810 }
1811
1812 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1813                         void __user *buffer, size_t *length, loff_t *ppos)
1814 {
1815         proc_dointvec_minmax(table, write, buffer, length, ppos);
1816
1817         return 0;
1818 }
1819
1820 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1821 static ssize_t sysfs_compact_node(struct device *dev,
1822                         struct device_attribute *attr,
1823                         const char *buf, size_t count)
1824 {
1825         int nid = dev->id;
1826
1827         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1828                 /* Flush pending updates to the LRU lists */
1829                 lru_add_drain_all();
1830
1831                 compact_node(nid);
1832         }
1833
1834         return count;
1835 }
1836 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1837
1838 int compaction_register_node(struct node *node)
1839 {
1840         return device_create_file(&node->dev, &dev_attr_compact);
1841 }
1842
1843 void compaction_unregister_node(struct node *node)
1844 {
1845         return device_remove_file(&node->dev, &dev_attr_compact);
1846 }
1847 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1848
1849 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1850 {
1851         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1852 }
1853
1854 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1855 {
1856         int zoneid;
1857         struct zone *zone;
1858         enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1859
1860         for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1861                 zone = &pgdat->node_zones[zoneid];
1862
1863                 if (!populated_zone(zone))
1864                         continue;
1865
1866                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1867                                         classzone_idx) == COMPACT_CONTINUE)
1868                         return true;
1869         }
1870
1871         return false;
1872 }
1873
1874 static void kcompactd_do_work(pg_data_t *pgdat)
1875 {
1876         /*
1877          * With no special task, compact all zones so that a page of requested
1878          * order is allocatable.
1879          */
1880         int zoneid;
1881         struct zone *zone;
1882         struct compact_control cc = {
1883                 .order = pgdat->kcompactd_max_order,
1884                 .classzone_idx = pgdat->kcompactd_classzone_idx,
1885                 .mode = MIGRATE_SYNC_LIGHT,
1886                 .ignore_skip_hint = true,
1887
1888         };
1889         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1890                                                         cc.classzone_idx);
1891         count_vm_event(KCOMPACTD_WAKE);
1892
1893         for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1894                 int status;
1895
1896                 zone = &pgdat->node_zones[zoneid];
1897                 if (!populated_zone(zone))
1898                         continue;
1899
1900                 if (compaction_deferred(zone, cc.order))
1901                         continue;
1902
1903                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1904                                                         COMPACT_CONTINUE)
1905                         continue;
1906
1907                 cc.nr_freepages = 0;
1908                 cc.nr_migratepages = 0;
1909                 cc.zone = zone;
1910                 INIT_LIST_HEAD(&cc.freepages);
1911                 INIT_LIST_HEAD(&cc.migratepages);
1912
1913                 if (kthread_should_stop())
1914                         return;
1915                 status = compact_zone(zone, &cc);
1916
1917                 if (status == COMPACT_SUCCESS) {
1918                         compaction_defer_reset(zone, cc.order, false);
1919                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1920                         /*
1921                          * We use sync migration mode here, so we defer like
1922                          * sync direct compaction does.
1923                          */
1924                         defer_compaction(zone, cc.order);
1925                 }
1926
1927                 VM_BUG_ON(!list_empty(&cc.freepages));
1928                 VM_BUG_ON(!list_empty(&cc.migratepages));
1929         }
1930
1931         /*
1932          * Regardless of success, we are done until woken up next. But remember
1933          * the requested order/classzone_idx in case it was higher/tighter than
1934          * our current ones
1935          */
1936         if (pgdat->kcompactd_max_order <= cc.order)
1937                 pgdat->kcompactd_max_order = 0;
1938         if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1939                 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1940 }
1941
1942 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1943 {
1944         if (!order)
1945                 return;
1946
1947         if (pgdat->kcompactd_max_order < order)
1948                 pgdat->kcompactd_max_order = order;
1949
1950         if (pgdat->kcompactd_classzone_idx > classzone_idx)
1951                 pgdat->kcompactd_classzone_idx = classzone_idx;
1952
1953         if (!waitqueue_active(&pgdat->kcompactd_wait))
1954                 return;
1955
1956         if (!kcompactd_node_suitable(pgdat))
1957                 return;
1958
1959         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1960                                                         classzone_idx);
1961         wake_up_interruptible(&pgdat->kcompactd_wait);
1962 }
1963
1964 /*
1965  * The background compaction daemon, started as a kernel thread
1966  * from the init process.
1967  */
1968 static int kcompactd(void *p)
1969 {
1970         pg_data_t *pgdat = (pg_data_t*)p;
1971         struct task_struct *tsk = current;
1972
1973         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1974
1975         if (!cpumask_empty(cpumask))
1976                 set_cpus_allowed_ptr(tsk, cpumask);
1977
1978         set_freezable();
1979
1980         pgdat->kcompactd_max_order = 0;
1981         pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1982
1983         while (!kthread_should_stop()) {
1984                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1985                 wait_event_freezable(pgdat->kcompactd_wait,
1986                                 kcompactd_work_requested(pgdat));
1987
1988                 kcompactd_do_work(pgdat);
1989         }
1990
1991         return 0;
1992 }
1993
1994 /*
1995  * This kcompactd start function will be called by init and node-hot-add.
1996  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
1997  */
1998 int kcompactd_run(int nid)
1999 {
2000         pg_data_t *pgdat = NODE_DATA(nid);
2001         int ret = 0;
2002
2003         if (pgdat->kcompactd)
2004                 return 0;
2005
2006         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2007         if (IS_ERR(pgdat->kcompactd)) {
2008                 pr_err("Failed to start kcompactd on node %d\n", nid);
2009                 ret = PTR_ERR(pgdat->kcompactd);
2010                 pgdat->kcompactd = NULL;
2011         }
2012         return ret;
2013 }
2014
2015 /*
2016  * Called by memory hotplug when all memory in a node is offlined. Caller must
2017  * hold mem_hotplug_begin/end().
2018  */
2019 void kcompactd_stop(int nid)
2020 {
2021         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2022
2023         if (kcompactd) {
2024                 kthread_stop(kcompactd);
2025                 NODE_DATA(nid)->kcompactd = NULL;
2026         }
2027 }
2028
2029 /*
2030  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2031  * not required for correctness. So if the last cpu in a node goes
2032  * away, we get changed to run anywhere: as the first one comes back,
2033  * restore their cpu bindings.
2034  */
2035 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2036                         void *hcpu)
2037 {
2038         int nid;
2039
2040         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2041                 for_each_node_state(nid, N_MEMORY) {
2042                         pg_data_t *pgdat = NODE_DATA(nid);
2043                         const struct cpumask *mask;
2044
2045                         mask = cpumask_of_node(pgdat->node_id);
2046
2047                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2048                                 /* One of our CPUs online: restore mask */
2049                                 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2050                 }
2051         }
2052         return NOTIFY_OK;
2053 }
2054
2055 static int __init kcompactd_init(void)
2056 {
2057         int nid;
2058
2059         for_each_node_state(nid, N_MEMORY)
2060                 kcompactd_run(nid);
2061         hotcpu_notifier(cpu_callback, 0);
2062         return 0;
2063 }
2064 subsys_initcall(kcompactd_init)
2065
2066 #endif /* CONFIG_COMPACTION */