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