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/swap.h>
  11 #include <linux/migrate.h>
  12 #include <linux/compaction.h>
  13 #include <linux/mm_inline.h>
  14 #include <linux/backing-dev.h>
  15 #include <linux/sysctl.h>
  16 #include <linux/sysfs.h>
  17 #include <linux/balloon_compaction.h>
  18 #include <linux/page-isolation.h>
  19 #include <linux/kasan.h>
  20 #include "internal.h"
  21
  22 #ifdef CONFIG_COMPACTION
  23 static inline void count_compact_event(enum vm_event_item item)
  24 {
  25         count_vm_event(item);
  26 }
  27
  28 static inline void count_compact_events(enum vm_event_item item, long delta)
  29 {
  30         count_vm_events(item, delta);
  31 }
  32 #else
  33 #define count_compact_event(item) do { } while (0)
  34 #define count_compact_events(item, delta) do { } while (0)
  35 #endif
  36
  37 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
  38
  39 #define CREATE_TRACE_POINTS
  40 #include <trace/events/compaction.h>
  41
  42 static unsigned long release_freepages(struct list_head *freelist)
  43 {
  44         struct page *page, *next;
  45         unsigned long high_pfn = 0;
  46
  47         list_for_each_entry_safe(page, next, freelist, lru) {
  48                 unsigned long pfn = page_to_pfn(page);
  49                 list_del(&page->lru);
  50                 __free_page(page);
  51                 if (pfn > high_pfn)
  52                         high_pfn = pfn;
  53         }
  54
  55         return high_pfn;
  56 }
  57
  58 static void map_pages(struct list_head *list)
  59 {
  60         struct page *page;
  61
  62         list_for_each_entry(page, list, lru) {
  63                 arch_alloc_page(page, 0);
  64                 kernel_map_pages(page, 1, 1);
  65                 kasan_alloc_pages(page, 0);
  66         }
  67 }
  68
  69 static inline bool migrate_async_suitable(int migratetype)
  70 {
  71         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
  72 }
  73
  74 /*
  75  * Check that the whole (or subset of) a pageblock given by the interval of
  76  * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
  77  * with the migration of free compaction scanner. The scanners then need to
  78  * use only pfn_valid_within() check for arches that allow holes within
  79  * pageblocks.
  80  *
  81  * Return struct page pointer of start_pfn, or NULL if checks were not passed.
  82  *
  83  * It's possible on some configurations to have a setup like node0 node1 node0
  84  * i.e. it's possible that all pages within a zones range of pages do not
  85  * belong to a single zone. We assume that a border between node0 and node1
  86  * can occur within a single pageblock, but not a node0 node1 node0
  87  * interleaving within a single pageblock. It is therefore sufficient to check
  88  * the first and last page of a pageblock and avoid checking each individual
  89  * page in a pageblock.
  90  */
  91 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
  92                                 unsigned long end_pfn, struct zone *zone)
  93 {
  94         struct page *start_page;
  95         struct page *end_page;
  96
  97         /* end_pfn is one past the range we are checking */
  98         end_pfn--;
  99
 100         if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
 101                 return NULL;
 102
 103         start_page = pfn_to_page(start_pfn);
 104
 105         if (page_zone(start_page) != zone)
 106                 return NULL;
 107
 108         end_page = pfn_to_page(end_pfn);
 109
 110         /* This gives a shorter code than deriving page_zone(end_page) */
 111         if (page_zone_id(start_page) != page_zone_id(end_page))
 112                 return NULL;
 113
 114         return start_page;
 115 }
 116
 117 #ifdef CONFIG_COMPACTION
 118
 119 /* Do not skip compaction more than 64 times */
 120 #define COMPACT_MAX_DEFER_SHIFT 6
 121
 122 /*
 123  * Compaction is deferred when compaction fails to result in a page
 124  * allocation success. 1 << compact_defer_limit compactions are skipped up
 125  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
 126  */
 127 void defer_compaction(struct zone *zone, int order)
 128 {
 129         zone->compact_considered = 0;
 130         zone->compact_defer_shift++;
 131
 132         if (order < zone->compact_order_failed)
 133                 zone->compact_order_failed = order;
 134
 135         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
 136                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
 137
 138         trace_mm_compaction_defer_compaction(zone, order);
 139 }
 140
 141 /* Returns true if compaction should be skipped this time */
 142 bool compaction_deferred(struct zone *zone, int order)
 143 {
 144         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
 145
 146         if (order < zone->compact_order_failed)
 147                 return false;
 148
 149         /* Avoid possible overflow */
 150         if (++zone->compact_considered > defer_limit)
 151                 zone->compact_considered = defer_limit;
 152
 153         if (zone->compact_considered >= defer_limit)
 154                 return false;
 155
 156         trace_mm_compaction_deferred(zone, order);
 157
 158         return true;
 159 }
 160
 161 /*
 162  * Update defer tracking counters after successful compaction of given order,
 163  * which means an allocation either succeeded (alloc_success == true) or is
 164  * expected to succeed.
 165  */
 166 void compaction_defer_reset(struct zone *zone, int order,
 167                 bool alloc_success)
 168 {
 169         if (alloc_success) {
 170                 zone->compact_considered = 0;
 171                 zone->compact_defer_shift = 0;
 172         }
 173         if (order >= zone->compact_order_failed)
 174                 zone->compact_order_failed = order + 1;
 175
 176         trace_mm_compaction_defer_reset(zone, order);
 177 }
 178
 179 /* Returns true if restarting compaction after many failures */
 180 bool compaction_restarting(struct zone *zone, int order)
 181 {
 182         if (order < zone->compact_order_failed)
 183                 return false;
 184
 185         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
 186                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
 187 }
 188
 189 /* Returns true if the pageblock should be scanned for pages to isolate. */
 190 static inline bool isolation_suitable(struct compact_control *cc,
 191                                         struct page *page)
 192 {
 193         if (cc->ignore_skip_hint)
 194                 return true;
 195
 196         return !get_pageblock_skip(page);
 197 }
 198
 199 static void reset_cached_positions(struct zone *zone)
 200 {
 201         zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
 202         zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
 203         zone->compact_cached_free_pfn =
 204                         round_down(zone_end_pfn(zone) - 1, pageblock_nr_pages);
 205 }
 206
 207 /*
 208  * This function is called to clear all cached information on pageblocks that
 209  * should be skipped for page isolation when the migrate and free page scanner
 210  * meet.
 211  */
 212 static void __reset_isolation_suitable(struct zone *zone)
 213 {
 214         unsigned long start_pfn = zone->zone_start_pfn;
 215         unsigned long end_pfn = zone_end_pfn(zone);
 216         unsigned long pfn;
 217
 218         zone->compact_blockskip_flush = false;
 219
 220         /* Walk the zone and mark every pageblock as suitable for isolation */
 221         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 222                 struct page *page;
 223
 224                 cond_resched();
 225
 226                 if (!pfn_valid(pfn))
 227                         continue;
 228
 229                 page = pfn_to_page(pfn);
 230                 if (zone != page_zone(page))
 231                         continue;
 232
 233                 clear_pageblock_skip(page);
 234         }
 235
 236         reset_cached_positions(zone);
 237 }
 238
 239 void reset_isolation_suitable(pg_data_t *pgdat)
 240 {
 241         int zoneid;
 242
 243         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 244                 struct zone *zone = &pgdat->node_zones[zoneid];
 245                 if (!populated_zone(zone))
 246                         continue;
 247
 248                 /* Only flush if a full compaction finished recently */
 249                 if (zone->compact_blockskip_flush)
 250                         __reset_isolation_suitable(zone);
 251         }
 252 }
 253
 254 /*
 255  * If no pages were isolated then mark this pageblock to be skipped in the
 256  * future. The information is later cleared by __reset_isolation_suitable().
 257  */
 258 static void update_pageblock_skip(struct compact_control *cc,
 259                         struct page *page, unsigned long nr_isolated,
 260                         bool migrate_scanner)
 261 {
 262         struct zone *zone = cc->zone;
 263         unsigned long pfn;
 264
 265         if (cc->ignore_skip_hint)
 266                 return;
 267
 268         if (!page)
 269                 return;
 270
 271         if (nr_isolated)
 272                 return;
 273
 274         set_pageblock_skip(page);
 275
 276         pfn = page_to_pfn(page);
 277
 278         /* Update where async and sync compaction should restart */
 279         if (migrate_scanner) {
 280                 if (pfn > zone->compact_cached_migrate_pfn[0])
 281                         zone->compact_cached_migrate_pfn[0] = pfn;
 282                 if (cc->mode != MIGRATE_ASYNC &&
 283                     pfn > zone->compact_cached_migrate_pfn[1])
 284                         zone->compact_cached_migrate_pfn[1] = pfn;
 285         } else {
 286                 if (pfn < zone->compact_cached_free_pfn)
 287                         zone->compact_cached_free_pfn = pfn;
 288         }
 289 }
 290 #else
 291 static inline bool isolation_suitable(struct compact_control *cc,
 292                                         struct page *page)
 293 {
 294         return true;
 295 }
 296
 297 static void update_pageblock_skip(struct compact_control *cc,
 298                         struct page *page, unsigned long nr_isolated,
 299                         bool migrate_scanner)
 300 {
 301 }
 302 #endif /* CONFIG_COMPACTION */
 303
 304 /*
 305  * Compaction requires the taking of some coarse locks that are potentially
 306  * very heavily contended. For async compaction, back out if the lock cannot
 307  * be taken immediately. For sync compaction, spin on the lock if needed.
 308  *
 309  * Returns true if the lock is held
 310  * Returns false if the lock is not held and compaction should abort
 311  */
 312 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
 313                                                 struct compact_control *cc)
 314 {
 315         if (cc->mode == MIGRATE_ASYNC) {
 316                 if (!spin_trylock_irqsave(lock, *flags)) {
 317                         cc->contended = COMPACT_CONTENDED_LOCK;
 318                         return false;
 319                 }
 320         } else {
 321                 spin_lock_irqsave(lock, *flags);
 322         }
 323
 324         return true;
 325 }
 326
 327 /*
 328  * Compaction requires the taking of some coarse locks that are potentially
 329  * very heavily contended. The lock should be periodically unlocked to avoid
 330  * having disabled IRQs for a long time, even when there is nobody waiting on
 331  * the lock. It might also be that allowing the IRQs will result in
 332  * need_resched() becoming true. If scheduling is needed, async compaction
 333  * aborts. Sync compaction schedules.
 334  * Either compaction type will also abort if a fatal signal is pending.
 335  * In either case if the lock was locked, it is dropped and not regained.
 336  *
 337  * Returns true if compaction should abort due to fatal signal pending, or
 338  *              async compaction due to need_resched()
 339  * Returns false when compaction can continue (sync compaction might have
 340  *              scheduled)
 341  */
 342 static bool compact_unlock_should_abort(spinlock_t *lock,
 343                 unsigned long flags, bool *locked, struct compact_control *cc)
 344 {
 345         if (*locked) {
 346                 spin_unlock_irqrestore(lock, flags);
 347                 *locked = false;
 348         }
 349
 350         if (fatal_signal_pending(current)) {
 351                 cc->contended = COMPACT_CONTENDED_SCHED;
 352                 return true;
 353         }
 354
 355         if (need_resched()) {
 356                 if (cc->mode == MIGRATE_ASYNC) {
 357                         cc->contended = COMPACT_CONTENDED_SCHED;
 358                         return true;
 359                 }
 360                 cond_resched();
 361         }
 362
 363         return false;
 364 }
 365
 366 /*
 367  * Aside from avoiding lock contention, compaction also periodically checks
 368  * need_resched() and either schedules in sync compaction or aborts async
 369  * compaction. This is similar to what compact_unlock_should_abort() does, but
 370  * is used where no lock is concerned.
 371  *
 372  * Returns false when no scheduling was needed, or sync compaction scheduled.
 373  * Returns true when async compaction should abort.
 374  */
 375 static inline bool compact_should_abort(struct compact_control *cc)
 376 {
 377         /* async compaction aborts if contended */
 378         if (need_resched()) {
 379                 if (cc->mode == MIGRATE_ASYNC) {
 380                         cc->contended = COMPACT_CONTENDED_SCHED;
 381                         return true;
 382                 }
 383
 384                 cond_resched();
 385         }
 386
 387         return false;
 388 }
 389
 390 /*
 391  * Isolate free pages onto a private freelist. If @strict is true, will abort
 392  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
 393  * (even though it may still end up isolating some pages).
 394  */
 395 static unsigned long isolate_freepages_block(struct compact_control *cc,
 396                                 unsigned long *start_pfn,
 397                                 unsigned long end_pfn,
 398                                 struct list_head *freelist,
 399                                 bool strict)
 400 {
 401         int nr_scanned = 0, total_isolated = 0;
 402         struct page *cursor, *valid_page = NULL;
 403         unsigned long flags = 0;
 404         bool locked = false;
 405         unsigned long blockpfn = *start_pfn;
 406
 407         cursor = pfn_to_page(blockpfn);
 408
 409         /* Isolate free pages. */
 410         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
 411                 int isolated, i;
 412                 struct page *page = cursor;
 413
 414                 /*
 415                  * Periodically drop the lock (if held) regardless of its
 416                  * contention, to give chance to IRQs. Abort if fatal signal
 417                  * pending or async compaction detects need_resched()
 418                  */
 419                 if (!(blockpfn % SWAP_CLUSTER_MAX)
 420                     && compact_unlock_should_abort(&cc->zone->lock, flags,
 421                                                                 &locked, cc))
 422                         break;
 423
 424                 nr_scanned++;
 425                 if (!pfn_valid_within(blockpfn))
 426                         goto isolate_fail;
 427
 428                 if (!valid_page)
 429                         valid_page = page;
 430
 431                 /*
 432                  * For compound pages such as THP and hugetlbfs, we can save
 433                  * potentially a lot of iterations if we skip them at once.
 434                  * The check is racy, but we can consider only valid values
 435                  * and the only danger is skipping too much.
 436                  */
 437                 if (PageCompound(page)) {
 438                         unsigned int comp_order = compound_order(page);
 439
 440                         if (likely(comp_order < MAX_ORDER)) {
 441                                 blockpfn += (1UL << comp_order) - 1;
 442                                 cursor += (1UL << comp_order) - 1;
 443                         }
 444
 445                         goto isolate_fail;
 446                 }
 447
 448                 if (!PageBuddy(page))
 449                         goto isolate_fail;
 450
 451                 /*
 452                  * If we already hold the lock, we can skip some rechecking.
 453                  * Note that if we hold the lock now, checked_pageblock was
 454                  * already set in some previous iteration (or strict is true),
 455                  * so it is correct to skip the suitable migration target
 456                  * recheck as well.
 457                  */
 458                 if (!locked) {
 459                         /*
 460                          * The zone lock must be held to isolate freepages.
 461                          * Unfortunately this is a very coarse lock and can be
 462                          * heavily contended if there are parallel allocations
 463                          * or parallel compactions. For async compaction do not
 464                          * spin on the lock and we acquire the lock as late as
 465                          * possible.
 466                          */
 467                         locked = compact_trylock_irqsave(&cc->zone->lock,
 468                                                                 &flags, cc);
 469                         if (!locked)
 470                                 break;
 471
 472                         /* Recheck this is a buddy page under lock */
 473                         if (!PageBuddy(page))
 474                                 goto isolate_fail;
 475                 }
 476
 477                 /* Found a free page, break it into order-0 pages */
 478                 isolated = split_free_page(page);
 479                 if (!isolated)
 480                         break;
 481
 482                 total_isolated += isolated;
 483                 cc->nr_freepages += isolated;
 484                 for (i = 0; i < isolated; i++) {
 485                         list_add(&page->lru, freelist);
 486                         page++;
 487                 }
 488                 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
 489                         blockpfn += isolated;
 490                         break;
 491                 }
 492                 /* Advance to the end of split page */
 493                 blockpfn += isolated - 1;
 494                 cursor += isolated - 1;
 495                 continue;
 496
 497 isolate_fail:
 498                 if (strict)
 499                         break;
 500                 else
 501                         continue;
 502
 503         }
 504
 505         if (locked)
 506                 spin_unlock_irqrestore(&cc->zone->lock, flags);
 507
 508         /*
 509          * There is a tiny chance that we have read bogus compound_order(),
 510          * so be careful to not go outside of the pageblock.
 511          */
 512         if (unlikely(blockpfn > end_pfn))
 513                 blockpfn = end_pfn;
 514
 515         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
 516                                         nr_scanned, total_isolated);
 517
 518         /* Record how far we have got within the block */
 519         *start_pfn = blockpfn;
 520
 521         /*
 522          * If strict isolation is requested by CMA then check that all the
 523          * pages requested were isolated. If there were any failures, 0 is
 524          * returned and CMA will fail.
 525          */
 526         if (strict && blockpfn < end_pfn)
 527                 total_isolated = 0;
 528
 529         /* Update the pageblock-skip if the whole pageblock was scanned */
 530         if (blockpfn == end_pfn)
 531                 update_pageblock_skip(cc, valid_page, total_isolated, false);
 532
 533         count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
 534         if (total_isolated)
 535                 count_compact_events(COMPACTISOLATED, total_isolated);
 536         return total_isolated;
 537 }
 538
 539 /**
 540  * isolate_freepages_range() - isolate free pages.
 541  * @start_pfn: The first PFN to start isolating.
 542  * @end_pfn:   The one-past-last PFN.
 543  *
 544  * Non-free pages, invalid PFNs, or zone boundaries within the
 545  * [start_pfn, end_pfn) range are considered errors, cause function to
 546  * undo its actions and return zero.
 547  *
 548  * Otherwise, function returns one-past-the-last PFN of isolated page
 549  * (which may be greater then end_pfn if end fell in a middle of
 550  * a free page).
 551  */
 552 unsigned long
 553 isolate_freepages_range(struct compact_control *cc,
 554                         unsigned long start_pfn, unsigned long end_pfn)
 555 {
 556         unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
 557         LIST_HEAD(freelist);
 558
 559         pfn = start_pfn;
 560         block_start_pfn = pfn & ~(pageblock_nr_pages - 1);
 561         if (block_start_pfn < cc->zone->zone_start_pfn)
 562                 block_start_pfn = cc->zone->zone_start_pfn;
 563         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 564
 565         for (; pfn < end_pfn; pfn += isolated,
 566                                 block_start_pfn = block_end_pfn,
 567                                 block_end_pfn += pageblock_nr_pages) {
 568                 /* Protect pfn from changing by isolate_freepages_block */
 569                 unsigned long isolate_start_pfn = pfn;
 570
 571                 block_end_pfn = min(block_end_pfn, end_pfn);
 572
 573                 /*
 574                  * pfn could pass the block_end_pfn if isolated freepage
 575                  * is more than pageblock order. In this case, we adjust
 576                  * scanning range to right one.
 577                  */
 578                 if (pfn >= block_end_pfn) {
 579                         block_start_pfn = pfn & ~(pageblock_nr_pages - 1);
 580                         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 581                         block_end_pfn = min(block_end_pfn, end_pfn);
 582                 }
 583
 584                 if (!pageblock_pfn_to_page(block_start_pfn,
 585                                         block_end_pfn, cc->zone))
 586                         break;
 587
 588                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
 589                                                 block_end_pfn, &freelist, true);
 590
 591                 /*
 592                  * In strict mode, isolate_freepages_block() returns 0 if
 593                  * there are any holes in the block (ie. invalid PFNs or
 594                  * non-free pages).
 595                  */
 596                 if (!isolated)
 597                         break;
 598
 599                 /*
 600                  * If we managed to isolate pages, it is always (1 << n) *
 601                  * pageblock_nr_pages for some non-negative n.  (Max order
 602                  * page may span two pageblocks).
 603                  */
 604         }
 605
 606         /* split_free_page does not map the pages */
 607         map_pages(&freelist);
 608
 609         if (pfn < end_pfn) {
 610                 /* Loop terminated early, cleanup. */
 611                 release_freepages(&freelist);
 612                 return 0;
 613         }
 614
 615         /* We don't use freelists for anything. */
 616         return pfn;
 617 }
 618
 619 /* Update the number of anon and file isolated pages in the zone */
 620 static void acct_isolated(struct zone *zone, struct compact_control *cc)
 621 {
 622         struct page *page;
 623         unsigned int count[2] = { 0, };
 624
 625         if (list_empty(&cc->migratepages))
 626                 return;
 627
 628         list_for_each_entry(page, &cc->migratepages, lru)
 629                 count[!!page_is_file_cache(page)]++;
 630
 631         mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 632         mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 633 }
 634
 635 /* Similar to reclaim, but different enough that they don't share logic */
 636 static bool too_many_isolated(struct zone *zone)
 637 {
 638         unsigned long active, inactive, isolated;
 639
 640         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 641                                         zone_page_state(zone, NR_INACTIVE_ANON);
 642         active = zone_page_state(zone, NR_ACTIVE_FILE) +
 643                                         zone_page_state(zone, NR_ACTIVE_ANON);
 644         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 645                                         zone_page_state(zone, NR_ISOLATED_ANON);
 646
 647         return isolated > (inactive + active) / 2;
 648 }
 649
 650 /**
 651  * isolate_migratepages_block() - isolate all migrate-able pages within
 652  *                                a single pageblock
 653  * @cc:         Compaction control structure.
 654  * @low_pfn:    The first PFN to isolate
 655  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
 656  * @isolate_mode: Isolation mode to be used.
 657  *
 658  * Isolate all pages that can be migrated from the range specified by
 659  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 660  * Returns zero if there is a fatal signal pending, otherwise PFN of the
 661  * first page that was not scanned (which may be both less, equal to or more
 662  * than end_pfn).
 663  *
 664  * The pages are isolated on cc->migratepages list (not required to be empty),
 665  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
 666  * is neither read nor updated.
 667  */
 668 static unsigned long
 669 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 670                         unsigned long end_pfn, isolate_mode_t isolate_mode)
 671 {
 672         struct zone *zone = cc->zone;
 673         unsigned long nr_scanned = 0, nr_isolated = 0;
 674         struct list_head *migratelist = &cc->migratepages;
 675         struct lruvec *lruvec;
 676         unsigned long flags = 0;
 677         bool locked = false;
 678         struct page *page = NULL, *valid_page = NULL;
 679         unsigned long start_pfn = low_pfn;
 680
 681         /*
 682          * Ensure that there are not too many pages isolated from the LRU
 683          * list by either parallel reclaimers or compaction. If there are,
 684          * delay for some time until fewer pages are isolated
 685          */
 686         while (unlikely(too_many_isolated(zone))) {
 687                 /* async migration should just abort */
 688                 if (cc->mode == MIGRATE_ASYNC)
 689                         return 0;
 690
 691                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 692
 693                 if (fatal_signal_pending(current))
 694                         return 0;
 695         }
 696
 697         if (compact_should_abort(cc))
 698                 return 0;
 699
 700         /* Time to isolate some pages for migration */
 701         for (; low_pfn < end_pfn; low_pfn++) {
 702                 bool is_lru;
 703
 704                 /*
 705                  * Periodically drop the lock (if held) regardless of its
 706                  * contention, to give chance to IRQs. Abort async compaction
 707                  * if contended.
 708                  */
 709                 if (!(low_pfn % SWAP_CLUSTER_MAX)
 710                     && compact_unlock_should_abort(&zone->lru_lock, flags,
 711                                                                 &locked, cc))
 712                         break;
 713
 714                 if (!pfn_valid_within(low_pfn))
 715                         continue;
 716                 nr_scanned++;
 717
 718                 page = pfn_to_page(low_pfn);
 719
 720                 if (!valid_page)
 721                         valid_page = page;
 722
 723                 /*
 724                  * Skip if free. We read page order here without zone lock
 725                  * which is generally unsafe, but the race window is small and
 726                  * the worst thing that can happen is that we skip some
 727                  * potential isolation targets.
 728                  */
 729                 if (PageBuddy(page)) {
 730                         unsigned long freepage_order = page_order_unsafe(page);
 731
 732                         /*
 733                          * Without lock, we cannot be sure that what we got is
 734                          * a valid page order. Consider only values in the
 735                          * valid order range to prevent low_pfn overflow.
 736                          */
 737                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
 738                                 low_pfn += (1UL << freepage_order) - 1;
 739                         continue;
 740                 }
 741
 742                 /*
 743                  * Check may be lockless but that's ok as we recheck later.
 744                  * It's possible to migrate LRU pages and balloon pages
 745                  * Skip any other type of page
 746                  */
 747                 is_lru = PageLRU(page);
 748                 if (!is_lru) {
 749                         if (unlikely(balloon_page_movable(page))) {
 750                                 if (balloon_page_isolate(page)) {
 751                                         /* Successfully isolated */
 752                                         goto isolate_success;
 753                                 }
 754                         }
 755                 }
 756
 757                 /*
 758                  * Regardless of being on LRU, compound pages such as THP and
 759                  * hugetlbfs are not to be compacted. We can potentially save
 760                  * a lot of iterations if we skip them at once. The check is
 761                  * racy, but we can consider only valid values and the only
 762                  * danger is skipping too much.
 763                  */
 764                 if (PageCompound(page)) {
 765                         unsigned int comp_order = compound_order(page);
 766
 767                         if (likely(comp_order < MAX_ORDER))
 768                                 low_pfn += (1UL << comp_order) - 1;
 769
 770                         continue;
 771                 }
 772
 773                 if (!is_lru)
 774                         continue;
 775
 776                 /*
 777                  * Migration will fail if an anonymous page is pinned in memory,
 778                  * so avoid taking lru_lock and isolating it unnecessarily in an
 779                  * admittedly racy check.
 780                  */
 781                 if (!page_mapping(page) &&
 782                     page_count(page) > page_mapcount(page))
 783                         continue;
 784
 785                 /* If we already hold the lock, we can skip some rechecking */
 786                 if (!locked) {
 787                         locked = compact_trylock_irqsave(&zone->lru_lock,
 788                                                                 &flags, cc);
 789                         if (!locked)
 790                                 break;
 791
 792                         /* Recheck PageLRU and PageCompound under lock */
 793                         if (!PageLRU(page))
 794                                 continue;
 795
 796                         /*
 797                          * Page become compound since the non-locked check,
 798                          * and it's on LRU. It can only be a THP so the order
 799                          * is safe to read and it's 0 for tail pages.
 800                          */
 801                         if (unlikely(PageCompound(page))) {
 802                                 low_pfn += (1UL << compound_order(page)) - 1;
 803                                 continue;
 804                         }
 805                 }
 806
 807                 lruvec = mem_cgroup_page_lruvec(page, zone);
 808
 809                 /* Try isolate the page */
 810                 if (__isolate_lru_page(page, isolate_mode) != 0)
 811                         continue;
 812
 813                 VM_BUG_ON_PAGE(PageCompound(page), page);
 814
 815                 /* Successfully isolated */
 816                 del_page_from_lru_list(page, lruvec, page_lru(page));
 817
 818 isolate_success:
 819                 list_add(&page->lru, migratelist);
 820                 cc->nr_migratepages++;
 821                 nr_isolated++;
 822
 823                 /* Avoid isolating too much */
 824                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
 825                         ++low_pfn;
 826                         break;
 827                 }
 828         }
 829
 830         /*
 831          * The PageBuddy() check could have potentially brought us outside
 832          * the range to be scanned.
 833          */
 834         if (unlikely(low_pfn > end_pfn))
 835                 low_pfn = end_pfn;
 836
 837         if (locked)
 838                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 839
 840         /*
 841          * Update the pageblock-skip information and cached scanner pfn,
 842          * if the whole pageblock was scanned without isolating any page.
 843          */
 844         if (low_pfn == end_pfn)
 845                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
 846
 847         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
 848                                                 nr_scanned, nr_isolated);
 849
 850         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
 851         if (nr_isolated)
 852                 count_compact_events(COMPACTISOLATED, nr_isolated);
 853
 854         return low_pfn;
 855 }
 856
 857 /**
 858  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
 859  * @cc:        Compaction control structure.
 860  * @start_pfn: The first PFN to start isolating.
 861  * @end_pfn:   The one-past-last PFN.
 862  *
 863  * Returns zero if isolation fails fatally due to e.g. pending signal.
 864  * Otherwise, function returns one-past-the-last PFN of isolated page
 865  * (which may be greater than end_pfn if end fell in a middle of a THP page).
 866  */
 867 unsigned long
 868 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
 869                                                         unsigned long end_pfn)
 870 {
 871         unsigned long pfn, block_start_pfn, block_end_pfn;
 872
 873         /* Scan block by block. First and last block may be incomplete */
 874         pfn = start_pfn;
 875         block_start_pfn = pfn & ~(pageblock_nr_pages - 1);
 876         if (block_start_pfn < cc->zone->zone_start_pfn)
 877                 block_start_pfn = cc->zone->zone_start_pfn;
 878         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 879
 880         for (; pfn < end_pfn; pfn = block_end_pfn,
 881                                 block_start_pfn = block_end_pfn,
 882                                 block_end_pfn += pageblock_nr_pages) {
 883
 884                 block_end_pfn = min(block_end_pfn, end_pfn);
 885
 886                 if (!pageblock_pfn_to_page(block_start_pfn,
 887                                         block_end_pfn, cc->zone))
 888                         continue;
 889
 890                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
 891                                                         ISOLATE_UNEVICTABLE);
 892
 893                 if (!pfn)
 894                         break;
 895
 896                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 897                         break;
 898         }
 899         acct_isolated(cc->zone, cc);
 900
 901         return pfn;
 902 }
 903
 904 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 905 #ifdef CONFIG_COMPACTION
 906
 907 /* Returns true if the page is within a block suitable for migration to */
 908 static bool suitable_migration_target(struct page *page)
 909 {
 910         /* If the page is a large free page, then disallow migration */
 911         if (PageBuddy(page)) {
 912                 /*
 913                  * We are checking page_order without zone->lock taken. But
 914                  * the only small danger is that we skip a potentially suitable
 915                  * pageblock, so it's not worth to check order for valid range.
 916                  */
 917                 if (page_order_unsafe(page) >= pageblock_order)
 918                         return false;
 919         }
 920
 921         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
 922         if (migrate_async_suitable(get_pageblock_migratetype(page)))
 923                 return true;
 924
 925         /* Otherwise skip the block */
 926         return false;
 927 }
 928
 929 /*
 930  * Test whether the free scanner has reached the same or lower pageblock than
 931  * the migration scanner, and compaction should thus terminate.
 932  */
 933 static inline bool compact_scanners_met(struct compact_control *cc)
 934 {
 935         return (cc->free_pfn >> pageblock_order)
 936                 <= (cc->migrate_pfn >> pageblock_order);
 937 }
 938
 939 /*
 940  * Based on information in the current compact_control, find blocks
 941  * suitable for isolating free pages from and then isolate them.
 942  */
 943 static void isolate_freepages(struct compact_control *cc)
 944 {
 945         struct zone *zone = cc->zone;
 946         struct page *page;
 947         unsigned long block_start_pfn;  /* start of current pageblock */
 948         unsigned long isolate_start_pfn; /* exact pfn we start at */
 949         unsigned long block_end_pfn;    /* end of current pageblock */
 950         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
 951         struct list_head *freelist = &cc->freepages;
 952
 953         /*
 954          * Initialise the free scanner. The starting point is where we last
 955          * successfully isolated from, zone-cached value, or the end of the
 956          * zone when isolating for the first time. For looping we also need
 957          * this pfn aligned down to the pageblock boundary, because we do
 958          * block_start_pfn -= pageblock_nr_pages in the for loop.
 959          * For ending point, take care when isolating in last pageblock of a
 960          * a zone which ends in the middle of a pageblock.
 961          * The low boundary is the end of the pageblock the migration scanner
 962          * is using.
 963          */
 964         isolate_start_pfn = cc->free_pfn;
 965         block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
 966         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
 967                                                 zone_end_pfn(zone));
 968         low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
 969
 970         /*
 971          * Isolate free pages until enough are available to migrate the
 972          * pages on cc->migratepages. We stop searching if the migrate
 973          * and free page scanners meet or enough free pages are isolated.
 974          */
 975         for (; block_start_pfn >= low_pfn;
 976                                 block_end_pfn = block_start_pfn,
 977                                 block_start_pfn -= pageblock_nr_pages,
 978                                 isolate_start_pfn = block_start_pfn) {
 979                 /*
 980                  * This can iterate a massively long zone without finding any
 981                  * suitable migration targets, so periodically check if we need
 982                  * to schedule, or even abort async compaction.
 983                  */
 984                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
 985                                                 && compact_should_abort(cc))
 986                         break;
 987
 988                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
 989                                                                         zone);
 990                 if (!page)
 991                         continue;
 992
 993                 /* Check the block is suitable for migration */
 994                 if (!suitable_migration_target(page))
 995                         continue;
 996
 997                 /* If isolation recently failed, do not retry */
 998                 if (!isolation_suitable(cc, page))
 999                         continue;
1000
1001                 /* Found a block suitable for isolating free pages from. */
1002                 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1003                                         freelist, false);
1004
1005                 /*
1006                  * If we isolated enough freepages, or aborted due to lock
1007                  * contention, terminate.
1008                  */
1009                 if ((cc->nr_freepages >= cc->nr_migratepages)
1010                                                         || cc->contended) {
1011                         if (isolate_start_pfn >= block_end_pfn) {
1012                                 /*
1013                                  * Restart at previous pageblock if more
1014                                  * freepages can be isolated next time.
1015                                  */
1016                                 isolate_start_pfn =
1017                                         block_start_pfn - pageblock_nr_pages;
1018                         }
1019                         break;
1020                 } else if (isolate_start_pfn < block_end_pfn) {
1021                         /*
1022                          * If isolation failed early, do not continue
1023                          * needlessly.
1024                          */
1025                         break;
1026                 }
1027         }
1028
1029         /* split_free_page does not map the pages */
1030         map_pages(freelist);
1031
1032         /*
1033          * Record where the free scanner will restart next time. Either we
1034          * broke from the loop and set isolate_start_pfn based on the last
1035          * call to isolate_freepages_block(), or we met the migration scanner
1036          * and the loop terminated due to isolate_start_pfn < low_pfn
1037          */
1038         cc->free_pfn = isolate_start_pfn;
1039 }
1040
1041 /*
1042  * This is a migrate-callback that "allocates" freepages by taking pages
1043  * from the isolated freelists in the block we are migrating to.
1044  */
1045 static struct page *compaction_alloc(struct page *migratepage,
1046                                         unsigned long data,
1047                                         int **result)
1048 {
1049         struct compact_control *cc = (struct compact_control *)data;
1050         struct page *freepage;
1051
1052         /*
1053          * Isolate free pages if necessary, and if we are not aborting due to
1054          * contention.
1055          */
1056         if (list_empty(&cc->freepages)) {
1057                 if (!cc->contended)
1058                         isolate_freepages(cc);
1059
1060                 if (list_empty(&cc->freepages))
1061                         return NULL;
1062         }
1063
1064         freepage = list_entry(cc->freepages.next, struct page, lru);
1065         list_del(&freepage->lru);
1066         cc->nr_freepages--;
1067
1068         return freepage;
1069 }
1070
1071 /*
1072  * This is a migrate-callback that "frees" freepages back to the isolated
1073  * freelist.  All pages on the freelist are from the same zone, so there is no
1074  * special handling needed for NUMA.
1075  */
1076 static void compaction_free(struct page *page, unsigned long data)
1077 {
1078         struct compact_control *cc = (struct compact_control *)data;
1079
1080         list_add(&page->lru, &cc->freepages);
1081         cc->nr_freepages++;
1082 }
1083
1084 /* possible outcome of isolate_migratepages */
1085 typedef enum {
1086         ISOLATE_ABORT,          /* Abort compaction now */
1087         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1088         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1089 } isolate_migrate_t;
1090
1091 /*
1092  * Allow userspace to control policy on scanning the unevictable LRU for
1093  * compactable pages.
1094  */
1095 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1096
1097 /*
1098  * Isolate all pages that can be migrated from the first suitable block,
1099  * starting at the block pointed to by the migrate scanner pfn within
1100  * compact_control.
1101  */
1102 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1103                                         struct compact_control *cc)
1104 {
1105         unsigned long block_start_pfn;
1106         unsigned long block_end_pfn;
1107         unsigned long low_pfn;
1108         unsigned long isolate_start_pfn;
1109         struct page *page;
1110         const isolate_mode_t isolate_mode =
1111                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1112                 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1113
1114         /*
1115          * Start at where we last stopped, or beginning of the zone as
1116          * initialized by compact_zone()
1117          */
1118         low_pfn = cc->migrate_pfn;
1119         block_start_pfn = cc->migrate_pfn & ~(pageblock_nr_pages - 1);
1120         if (block_start_pfn < zone->zone_start_pfn)
1121                 block_start_pfn = zone->zone_start_pfn;
1122
1123         /* Only scan within a pageblock boundary */
1124         block_end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1125
1126         /*
1127          * Iterate over whole pageblocks until we find the first suitable.
1128          * Do not cross the free scanner.
1129          */
1130         for (; block_end_pfn <= cc->free_pfn;
1131                         low_pfn = block_end_pfn,
1132                         block_start_pfn = block_end_pfn,
1133                         block_end_pfn += pageblock_nr_pages) {
1134
1135                 /*
1136                  * This can potentially iterate a massively long zone with
1137                  * many pageblocks unsuitable, so periodically check if we
1138                  * need to schedule, or even abort async compaction.
1139                  */
1140                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1141                                                 && compact_should_abort(cc))
1142                         break;
1143
1144                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1145                                                                         zone);
1146                 if (!page)
1147                         continue;
1148
1149                 /* If isolation recently failed, do not retry */
1150                 if (!isolation_suitable(cc, page))
1151                         continue;
1152
1153                 /*
1154                  * For async compaction, also only scan in MOVABLE blocks.
1155                  * Async compaction is optimistic to see if the minimum amount
1156                  * of work satisfies the allocation.
1157                  */
1158                 if (cc->mode == MIGRATE_ASYNC &&
1159                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1160                         continue;
1161
1162                 /* Perform the isolation */
1163                 isolate_start_pfn = low_pfn;
1164                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1165                                                 block_end_pfn, isolate_mode);
1166
1167                 if (!low_pfn || cc->contended) {
1168                         acct_isolated(zone, cc);
1169                         return ISOLATE_ABORT;
1170                 }
1171
1172                 /*
1173                  * Record where we could have freed pages by migration and not
1174                  * yet flushed them to buddy allocator.
1175                  * - this is the lowest page that could have been isolated and
1176                  * then freed by migration.
1177                  */
1178                 if (cc->nr_migratepages && !cc->last_migrated_pfn)
1179                         cc->last_migrated_pfn = isolate_start_pfn;
1180
1181                 /*
1182                  * Either we isolated something and proceed with migration. Or
1183                  * we failed and compact_zone should decide if we should
1184                  * continue or not.
1185                  */
1186                 break;
1187         }
1188
1189         acct_isolated(zone, cc);
1190         /* Record where migration scanner will be restarted. */
1191         cc->migrate_pfn = low_pfn;
1192
1193         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1194 }
1195
1196 /*
1197  * order == -1 is expected when compacting via
1198  * /proc/sys/vm/compact_memory
1199  */
1200 static inline bool is_via_compact_memory(int order)
1201 {
1202         return order == -1;
1203 }
1204
1205 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1206                             const int migratetype)
1207 {
1208         unsigned int order;
1209         unsigned long watermark;
1210
1211         if (cc->contended || fatal_signal_pending(current))
1212                 return COMPACT_CONTENDED;
1213
1214         /* Compaction run completes if the migrate and free scanner meet */
1215         if (compact_scanners_met(cc)) {
1216                 /* Let the next compaction start anew. */
1217                 reset_cached_positions(zone);
1218
1219                 /*
1220                  * Mark that the PG_migrate_skip information should be cleared
1221                  * by kswapd when it goes to sleep. kswapd does not set the
1222                  * flag itself as the decision to be clear should be directly
1223                  * based on an allocation request.
1224                  */
1225                 if (!current_is_kswapd())
1226                         zone->compact_blockskip_flush = true;
1227
1228                 return COMPACT_COMPLETE;
1229         }
1230
1231         if (is_via_compact_memory(cc->order))
1232                 return COMPACT_CONTINUE;
1233
1234         /* Compaction run is not finished if the watermark is not met */
1235         watermark = low_wmark_pages(zone);
1236
1237         if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1238                                                         cc->alloc_flags))
1239                 return COMPACT_CONTINUE;
1240
1241         /* Direct compactor: Is a suitable page free? */
1242         for (order = cc->order; order < MAX_ORDER; order++) {
1243                 struct free_area *area = &zone->free_area[order];
1244                 bool can_steal;
1245
1246                 /* Job done if page is free of the right migratetype */
1247                 if (!list_empty(&area->free_list[migratetype]))
1248                         return COMPACT_PARTIAL;
1249
1250 #ifdef CONFIG_CMA
1251                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1252                 if (migratetype == MIGRATE_MOVABLE &&
1253                         !list_empty(&area->free_list[MIGRATE_CMA]))
1254                         return COMPACT_PARTIAL;
1255 #endif
1256                 /*
1257                  * Job done if allocation would steal freepages from
1258                  * other migratetype buddy lists.
1259                  */
1260                 if (find_suitable_fallback(area, order, migratetype,
1261                                                 true, &can_steal) != -1)
1262                         return COMPACT_PARTIAL;
1263         }
1264
1265         return COMPACT_NO_SUITABLE_PAGE;
1266 }
1267
1268 static int compact_finished(struct zone *zone, struct compact_control *cc,
1269                             const int migratetype)
1270 {
1271         int ret;
1272
1273         ret = __compact_finished(zone, cc, migratetype);
1274         trace_mm_compaction_finished(zone, cc->order, ret);
1275         if (ret == COMPACT_NO_SUITABLE_PAGE)
1276                 ret = COMPACT_CONTINUE;
1277
1278         return ret;
1279 }
1280
1281 /*
1282  * compaction_suitable: Is this suitable to run compaction on this zone now?
1283  * Returns
1284  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1285  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
1286  *   COMPACT_CONTINUE - If compaction should run now
1287  */
1288 static unsigned long __compaction_suitable(struct zone *zone, int order,
1289                                         int alloc_flags, int classzone_idx)
1290 {
1291         int fragindex;
1292         unsigned long watermark;
1293
1294         if (is_via_compact_memory(order))
1295                 return COMPACT_CONTINUE;
1296
1297         watermark = low_wmark_pages(zone);
1298         /*
1299          * If watermarks for high-order allocation are already met, there
1300          * should be no need for compaction at all.
1301          */
1302         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1303                                                                 alloc_flags))
1304                 return COMPACT_PARTIAL;
1305
1306         /*
1307          * Watermarks for order-0 must be met for compaction. Note the 2UL.
1308          * This is because during migration, copies of pages need to be
1309          * allocated and for a short time, the footprint is higher
1310          */
1311         watermark += (2UL << order);
1312         if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1313                 return COMPACT_SKIPPED;
1314
1315         /*
1316          * fragmentation index determines if allocation failures are due to
1317          * low memory or external fragmentation
1318          *
1319          * index of -1000 would imply allocations might succeed depending on
1320          * watermarks, but we already failed the high-order watermark check
1321          * index towards 0 implies failure is due to lack of memory
1322          * index towards 1000 implies failure is due to fragmentation
1323          *
1324          * Only compact if a failure would be due to fragmentation.
1325          */
1326         fragindex = fragmentation_index(zone, order);
1327         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1328                 return COMPACT_NOT_SUITABLE_ZONE;
1329
1330         return COMPACT_CONTINUE;
1331 }
1332
1333 unsigned long compaction_suitable(struct zone *zone, int order,
1334                                         int alloc_flags, int classzone_idx)
1335 {
1336         unsigned long ret;
1337
1338         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1339         trace_mm_compaction_suitable(zone, order, ret);
1340         if (ret == COMPACT_NOT_SUITABLE_ZONE)
1341                 ret = COMPACT_SKIPPED;
1342
1343         return ret;
1344 }
1345
1346 static int compact_zone(struct zone *zone, struct compact_control *cc)
1347 {
1348         int ret;
1349         unsigned long start_pfn = zone->zone_start_pfn;
1350         unsigned long end_pfn = zone_end_pfn(zone);
1351         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1352         const bool sync = cc->mode != MIGRATE_ASYNC;
1353
1354         ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1355                                                         cc->classzone_idx);
1356         switch (ret) {
1357         case COMPACT_PARTIAL:
1358         case COMPACT_SKIPPED:
1359                 /* Compaction is likely to fail */
1360                 return ret;
1361         case COMPACT_CONTINUE:
1362                 /* Fall through to compaction */
1363                 ;
1364         }
1365
1366         /*
1367          * Clear pageblock skip if there were failures recently and compaction
1368          * is about to be retried after being deferred. kswapd does not do
1369          * this reset as it'll reset the cached information when going to sleep.
1370          */
1371         if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1372                 __reset_isolation_suitable(zone);
1373
1374         /*
1375          * Setup to move all movable pages to the end of the zone. Used cached
1376          * information on where the scanners should start but check that it
1377          * is initialised by ensuring the values are within zone boundaries.
1378          */
1379         cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1380         cc->free_pfn = zone->compact_cached_free_pfn;
1381         if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1382                 cc->free_pfn = round_down(end_pfn - 1, pageblock_nr_pages);
1383                 zone->compact_cached_free_pfn = cc->free_pfn;
1384         }
1385         if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1386                 cc->migrate_pfn = start_pfn;
1387                 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1388                 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1389         }
1390         cc->last_migrated_pfn = 0;
1391
1392         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1393                                 cc->free_pfn, end_pfn, sync);
1394
1395         migrate_prep_local();
1396
1397         while ((ret = compact_finished(zone, cc, migratetype)) ==
1398                                                 COMPACT_CONTINUE) {
1399                 int err;
1400
1401                 switch (isolate_migratepages(zone, cc)) {
1402                 case ISOLATE_ABORT:
1403                         ret = COMPACT_CONTENDED;
1404                         putback_movable_pages(&cc->migratepages);
1405                         cc->nr_migratepages = 0;
1406                         goto out;
1407                 case ISOLATE_NONE:
1408                         /*
1409                          * We haven't isolated and migrated anything, but
1410                          * there might still be unflushed migrations from
1411                          * previous cc->order aligned block.
1412                          */
1413                         goto check_drain;
1414                 case ISOLATE_SUCCESS:
1415                         ;
1416                 }
1417
1418                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1419                                 compaction_free, (unsigned long)cc, cc->mode,
1420                                 MR_COMPACTION);
1421
1422                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1423                                                         &cc->migratepages);
1424
1425                 /* All pages were either migrated or will be released */
1426                 cc->nr_migratepages = 0;
1427                 if (err) {
1428                         putback_movable_pages(&cc->migratepages);
1429                         /*
1430                          * migrate_pages() may return -ENOMEM when scanners meet
1431                          * and we want compact_finished() to detect it
1432                          */
1433                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
1434                                 ret = COMPACT_CONTENDED;
1435                                 goto out;
1436                         }
1437                 }
1438
1439 check_drain:
1440                 /*
1441                  * Has the migration scanner moved away from the previous
1442                  * cc->order aligned block where we migrated from? If yes,
1443                  * flush the pages that were freed, so that they can merge and
1444                  * compact_finished() can detect immediately if allocation
1445                  * would succeed.
1446                  */
1447                 if (cc->order > 0 && cc->last_migrated_pfn) {
1448                         int cpu;
1449                         unsigned long current_block_start =
1450                                 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1451
1452                         if (cc->last_migrated_pfn < current_block_start) {
1453                                 cpu = get_cpu();
1454                                 lru_add_drain_cpu(cpu);
1455                                 drain_local_pages(zone);
1456                                 put_cpu();
1457                                 /* No more flushing until we migrate again */
1458                                 cc->last_migrated_pfn = 0;
1459                         }
1460                 }
1461
1462         }
1463
1464 out:
1465         /*
1466          * Release free pages and update where the free scanner should restart,
1467          * so we don't leave any returned pages behind in the next attempt.
1468          */
1469         if (cc->nr_freepages > 0) {
1470                 unsigned long free_pfn = release_freepages(&cc->freepages);
1471
1472                 cc->nr_freepages = 0;
1473                 VM_BUG_ON(free_pfn == 0);
1474                 /* The cached pfn is always the first in a pageblock */
1475                 free_pfn &= ~(pageblock_nr_pages-1);
1476                 /*
1477                  * Only go back, not forward. The cached pfn might have been
1478                  * already reset to zone end in compact_finished()
1479                  */
1480                 if (free_pfn > zone->compact_cached_free_pfn)
1481                         zone->compact_cached_free_pfn = free_pfn;
1482         }
1483
1484         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1485                                 cc->free_pfn, end_pfn, sync, ret);
1486
1487         if (ret == COMPACT_CONTENDED)
1488                 ret = COMPACT_PARTIAL;
1489
1490         return ret;
1491 }
1492
1493 static unsigned long compact_zone_order(struct zone *zone, int order,
1494                 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1495                 int alloc_flags, int classzone_idx)
1496 {
1497         unsigned long ret;
1498         struct compact_control cc = {
1499                 .nr_freepages = 0,
1500                 .nr_migratepages = 0,
1501                 .order = order,
1502                 .gfp_mask = gfp_mask,
1503                 .zone = zone,
1504                 .mode = mode,
1505                 .alloc_flags = alloc_flags,
1506                 .classzone_idx = classzone_idx,
1507         };
1508         INIT_LIST_HEAD(&cc.freepages);
1509         INIT_LIST_HEAD(&cc.migratepages);
1510
1511         ret = compact_zone(zone, &cc);
1512
1513         VM_BUG_ON(!list_empty(&cc.freepages));
1514         VM_BUG_ON(!list_empty(&cc.migratepages));
1515
1516         *contended = cc.contended;
1517         return ret;
1518 }
1519
1520 int sysctl_extfrag_threshold = 500;
1521
1522 /**
1523  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1524  * @gfp_mask: The GFP mask of the current allocation
1525  * @order: The order of the current allocation
1526  * @alloc_flags: The allocation flags of the current allocation
1527  * @ac: The context of current allocation
1528  * @mode: The migration mode for async, sync light, or sync migration
1529  * @contended: Return value that determines if compaction was aborted due to
1530  *             need_resched() or lock contention
1531  *
1532  * This is the main entry point for direct page compaction.
1533  */
1534 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1535                         int alloc_flags, const struct alloc_context *ac,
1536                         enum migrate_mode mode, int *contended)
1537 {
1538         int may_enter_fs = gfp_mask & __GFP_FS;
1539         int may_perform_io = gfp_mask & __GFP_IO;
1540         struct zoneref *z;
1541         struct zone *zone;
1542         int rc = COMPACT_DEFERRED;
1543         int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1544
1545         *contended = COMPACT_CONTENDED_NONE;
1546
1547         /* Check if the GFP flags allow compaction */
1548         if (!order || !may_enter_fs || !may_perform_io)
1549                 return COMPACT_SKIPPED;
1550
1551         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1552
1553         /* Compact each zone in the list */
1554         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1555                                                                 ac->nodemask) {
1556                 int status;
1557                 int zone_contended;
1558
1559                 if (compaction_deferred(zone, order))
1560                         continue;
1561
1562                 status = compact_zone_order(zone, order, gfp_mask, mode,
1563                                 &zone_contended, alloc_flags,
1564                                 ac->classzone_idx);
1565                 rc = max(status, rc);
1566                 /*
1567                  * It takes at least one zone that wasn't lock contended
1568                  * to clear all_zones_contended.
1569                  */
1570                 all_zones_contended &= zone_contended;
1571
1572                 /* If a normal allocation would succeed, stop compacting */
1573                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1574                                         ac->classzone_idx, alloc_flags)) {
1575                         /*
1576                          * We think the allocation will succeed in this zone,
1577                          * but it is not certain, hence the false. The caller
1578                          * will repeat this with true if allocation indeed
1579                          * succeeds in this zone.
1580                          */
1581                         compaction_defer_reset(zone, order, false);
1582                         /*
1583                          * It is possible that async compaction aborted due to
1584                          * need_resched() and the watermarks were ok thanks to
1585                          * somebody else freeing memory. The allocation can
1586                          * however still fail so we better signal the
1587                          * need_resched() contention anyway (this will not
1588                          * prevent the allocation attempt).
1589                          */
1590                         if (zone_contended == COMPACT_CONTENDED_SCHED)
1591                                 *contended = COMPACT_CONTENDED_SCHED;
1592
1593                         goto break_loop;
1594                 }
1595
1596                 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1597                         /*
1598                          * We think that allocation won't succeed in this zone
1599                          * so we defer compaction there. If it ends up
1600                          * succeeding after all, it will be reset.
1601                          */
1602                         defer_compaction(zone, order);
1603                 }
1604
1605                 /*
1606                  * We might have stopped compacting due to need_resched() in
1607                  * async compaction, or due to a fatal signal detected. In that
1608                  * case do not try further zones and signal need_resched()
1609                  * contention.
1610                  */
1611                 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1612                                         || fatal_signal_pending(current)) {
1613                         *contended = COMPACT_CONTENDED_SCHED;
1614                         goto break_loop;
1615                 }
1616
1617                 continue;
1618 break_loop:
1619                 /*
1620                  * We might not have tried all the zones, so  be conservative
1621                  * and assume they are not all lock contended.
1622                  */
1623                 all_zones_contended = 0;
1624                 break;
1625         }
1626
1627         /*
1628          * If at least one zone wasn't deferred or skipped, we report if all
1629          * zones that were tried were lock contended.
1630          */
1631         if (rc > COMPACT_SKIPPED && all_zones_contended)
1632                 *contended = COMPACT_CONTENDED_LOCK;
1633
1634         return rc;
1635 }
1636
1637
1638 /* Compact all zones within a node */
1639 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1640 {
1641         int zoneid;
1642         struct zone *zone;
1643
1644         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1645
1646                 zone = &pgdat->node_zones[zoneid];
1647                 if (!populated_zone(zone))
1648                         continue;
1649
1650                 cc->nr_freepages = 0;
1651                 cc->nr_migratepages = 0;
1652                 cc->zone = zone;
1653                 INIT_LIST_HEAD(&cc->freepages);
1654                 INIT_LIST_HEAD(&cc->migratepages);
1655
1656                 /*
1657                  * When called via /proc/sys/vm/compact_memory
1658                  * this makes sure we compact the whole zone regardless of
1659                  * cached scanner positions.
1660                  */
1661                 if (is_via_compact_memory(cc->order))
1662                         __reset_isolation_suitable(zone);
1663
1664                 if (is_via_compact_memory(cc->order) ||
1665                                 !compaction_deferred(zone, cc->order))
1666                         compact_zone(zone, cc);
1667
1668                 if (cc->order > 0) {
1669                         if (zone_watermark_ok(zone, cc->order,
1670                                                 low_wmark_pages(zone), 0, 0))
1671                                 compaction_defer_reset(zone, cc->order, false);
1672                 }
1673
1674                 VM_BUG_ON(!list_empty(&cc->freepages));
1675                 VM_BUG_ON(!list_empty(&cc->migratepages));
1676         }
1677 }
1678
1679 void compact_pgdat(pg_data_t *pgdat, int order)
1680 {
1681         struct compact_control cc = {
1682                 .order = order,
1683                 .mode = MIGRATE_ASYNC,
1684         };
1685
1686         if (!order)
1687                 return;
1688
1689         __compact_pgdat(pgdat, &cc);
1690 }
1691
1692 static void compact_node(int nid)
1693 {
1694         struct compact_control cc = {
1695                 .order = -1,
1696                 .mode = MIGRATE_SYNC,
1697                 .ignore_skip_hint = true,
1698         };
1699
1700         __compact_pgdat(NODE_DATA(nid), &cc);
1701 }
1702
1703 /* Compact all nodes in the system */
1704 static void compact_nodes(void)
1705 {
1706         int nid;
1707
1708         /* Flush pending updates to the LRU lists */
1709         lru_add_drain_all();
1710
1711         for_each_online_node(nid)
1712                 compact_node(nid);
1713 }
1714
1715 /* The written value is actually unused, all memory is compacted */
1716 int sysctl_compact_memory;
1717
1718 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1719 int sysctl_compaction_handler(struct ctl_table *table, int write,
1720                         void __user *buffer, size_t *length, loff_t *ppos)
1721 {
1722         if (write)
1723                 compact_nodes();
1724
1725         return 0;
1726 }
1727
1728 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1729                         void __user *buffer, size_t *length, loff_t *ppos)
1730 {
1731         proc_dointvec_minmax(table, write, buffer, length, ppos);
1732
1733         return 0;
1734 }
1735
1736 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1737 static ssize_t sysfs_compact_node(struct device *dev,
1738                         struct device_attribute *attr,
1739                         const char *buf, size_t count)
1740 {
1741         int nid = dev->id;
1742
1743         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1744                 /* Flush pending updates to the LRU lists */
1745                 lru_add_drain_all();
1746
1747                 compact_node(nid);
1748         }
1749
1750         return count;
1751 }
1752 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1753
1754 int compaction_register_node(struct node *node)
1755 {
1756         return device_create_file(&node->dev, &dev_attr_compact);
1757 }
1758
1759 void compaction_unregister_node(struct node *node)
1760 {
1761         return device_remove_file(&node->dev, &dev_attr_compact);
1762 }
1763 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1764
1765 #endif /* CONFIG_COMPACTION */