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 "internal.h"
  19
  20 #ifdef CONFIG_COMPACTION
  21 static inline void count_compact_event(enum vm_event_item item)
  22 {
  23         count_vm_event(item);
  24 }
  25
  26 static inline void count_compact_events(enum vm_event_item item, long delta)
  27 {
  28         count_vm_events(item, delta);
  29 }
  30 #else
  31 #define count_compact_event(item) do { } while (0)
  32 #define count_compact_events(item, delta) do { } while (0)
  33 #endif
  34
  35 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
  36
  37 #define CREATE_TRACE_POINTS
  38 #include <trace/events/compaction.h>
  39
  40 static unsigned long release_freepages(struct list_head *freelist)
  41 {
  42         struct page *page, *next;
  43         unsigned long count = 0;
  44
  45         list_for_each_entry_safe(page, next, freelist, lru) {
  46                 list_del(&page->lru);
  47                 __free_page(page);
  48                 count++;
  49         }
  50
  51         return count;
  52 }
  53
  54 static void map_pages(struct list_head *list)
  55 {
  56         struct page *page;
  57
  58         list_for_each_entry(page, list, lru) {
  59                 arch_alloc_page(page, 0);
  60                 kernel_map_pages(page, 1, 1);
  61         }
  62 }
  63
  64 static inline bool migrate_async_suitable(int migratetype)
  65 {
  66         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
  67 }
  68
  69 #ifdef CONFIG_COMPACTION
  70 /* Returns true if the pageblock should be scanned for pages to isolate. */
  71 static inline bool isolation_suitable(struct compact_control *cc,
  72                                         struct page *page)
  73 {
  74         if (cc->ignore_skip_hint)
  75                 return true;
  76
  77         return !get_pageblock_skip(page);
  78 }
  79
  80 /*
  81  * This function is called to clear all cached information on pageblocks that
  82  * should be skipped for page isolation when the migrate and free page scanner
  83  * meet.
  84  */
  85 static void __reset_isolation_suitable(struct zone *zone)
  86 {
  87         unsigned long start_pfn = zone->zone_start_pfn;
  88         unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
  89         unsigned long pfn;
  90
  91         zone->compact_cached_migrate_pfn = start_pfn;
  92         zone->compact_cached_free_pfn = end_pfn;
  93         zone->compact_blockskip_flush = false;
  94
  95         /* Walk the zone and mark every pageblock as suitable for isolation */
  96         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
  97                 struct page *page;
  98
  99                 cond_resched();
 100
 101                 if (!pfn_valid(pfn))
 102                         continue;
 103
 104                 page = pfn_to_page(pfn);
 105                 if (zone != page_zone(page))
 106                         continue;
 107
 108                 clear_pageblock_skip(page);
 109         }
 110 }
 111
 112 void reset_isolation_suitable(pg_data_t *pgdat)
 113 {
 114         int zoneid;
 115
 116         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 117                 struct zone *zone = &pgdat->node_zones[zoneid];
 118                 if (!populated_zone(zone))
 119                         continue;
 120
 121                 /* Only flush if a full compaction finished recently */
 122                 if (zone->compact_blockskip_flush)
 123                         __reset_isolation_suitable(zone);
 124         }
 125 }
 126
 127 /*
 128  * If no pages were isolated then mark this pageblock to be skipped in the
 129  * future. The information is later cleared by __reset_isolation_suitable().
 130  */
 131 static void update_pageblock_skip(struct compact_control *cc,
 132                         struct page *page, unsigned long nr_isolated,
 133                         bool migrate_scanner)
 134 {
 135         struct zone *zone = cc->zone;
 136         if (!page)
 137                 return;
 138
 139         if (!nr_isolated) {
 140                 unsigned long pfn = page_to_pfn(page);
 141                 set_pageblock_skip(page);
 142
 143                 /* Update where compaction should restart */
 144                 if (migrate_scanner) {
 145                         if (!cc->finished_update_migrate &&
 146                             pfn > zone->compact_cached_migrate_pfn)
 147                                 zone->compact_cached_migrate_pfn = pfn;
 148                 } else {
 149                         if (!cc->finished_update_free &&
 150                             pfn < zone->compact_cached_free_pfn)
 151                                 zone->compact_cached_free_pfn = pfn;
 152                 }
 153         }
 154 }
 155 #else
 156 static inline bool isolation_suitable(struct compact_control *cc,
 157                                         struct page *page)
 158 {
 159         return true;
 160 }
 161
 162 static void update_pageblock_skip(struct compact_control *cc,
 163                         struct page *page, unsigned long nr_isolated,
 164                         bool migrate_scanner)
 165 {
 166 }
 167 #endif /* CONFIG_COMPACTION */
 168
 169 static inline bool should_release_lock(spinlock_t *lock)
 170 {
 171         return need_resched() || spin_is_contended(lock);
 172 }
 173
 174 /*
 175  * Compaction requires the taking of some coarse locks that are potentially
 176  * very heavily contended. Check if the process needs to be scheduled or
 177  * if the lock is contended. For async compaction, back out in the event
 178  * if contention is severe. For sync compaction, schedule.
 179  *
 180  * Returns true if the lock is held.
 181  * Returns false if the lock is released and compaction should abort
 182  */
 183 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
 184                                       bool locked, struct compact_control *cc)
 185 {
 186         if (should_release_lock(lock)) {
 187                 if (locked) {
 188                         spin_unlock_irqrestore(lock, *flags);
 189                         locked = false;
 190                 }
 191
 192                 /* async aborts if taking too long or contended */
 193                 if (!cc->sync) {
 194                         cc->contended = true;
 195                         return false;
 196                 }
 197
 198                 cond_resched();
 199         }
 200
 201         if (!locked)
 202                 spin_lock_irqsave(lock, *flags);
 203         return true;
 204 }
 205
 206 static inline bool compact_trylock_irqsave(spinlock_t *lock,
 207                         unsigned long *flags, struct compact_control *cc)
 208 {
 209         return compact_checklock_irqsave(lock, flags, false, cc);
 210 }
 211
 212 /* Returns true if the page is within a block suitable for migration to */
 213 static bool suitable_migration_target(struct page *page)
 214 {
 215         int migratetype = get_pageblock_migratetype(page);
 216
 217         /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
 218         if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
 219                 return false;
 220
 221         /* If the page is a large free page, then allow migration */
 222         if (PageBuddy(page) && page_order(page) >= pageblock_order)
 223                 return true;
 224
 225         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
 226         if (migrate_async_suitable(migratetype))
 227                 return true;
 228
 229         /* Otherwise skip the block */
 230         return false;
 231 }
 232
 233 /*
 234  * Isolate free pages onto a private freelist. Caller must hold zone->lock.
 235  * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
 236  * pages inside of the pageblock (even though it may still end up isolating
 237  * some pages).
 238  */
 239 static unsigned long isolate_freepages_block(struct compact_control *cc,
 240                                 unsigned long blockpfn,
 241                                 unsigned long end_pfn,
 242                                 struct list_head *freelist,
 243                                 bool strict)
 244 {
 245         int nr_scanned = 0, total_isolated = 0;
 246         struct page *cursor, *valid_page = NULL;
 247         unsigned long nr_strict_required = end_pfn - blockpfn;
 248         unsigned long flags;
 249         bool locked = false;
 250
 251         cursor = pfn_to_page(blockpfn);
 252
 253         /* Isolate free pages. */
 254         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
 255                 int isolated, i;
 256                 struct page *page = cursor;
 257
 258                 nr_scanned++;
 259                 if (!pfn_valid_within(blockpfn))
 260                         continue;
 261                 if (!valid_page)
 262                         valid_page = page;
 263                 if (!PageBuddy(page))
 264                         continue;
 265
 266                 /*
 267                  * The zone lock must be held to isolate freepages.
 268                  * Unfortunately this is a very coarse lock and can be
 269                  * heavily contended if there are parallel allocations
 270                  * or parallel compactions. For async compaction do not
 271                  * spin on the lock and we acquire the lock as late as
 272                  * possible.
 273                  */
 274                 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
 275                                                                 locked, cc);
 276                 if (!locked)
 277                         break;
 278
 279                 /* Recheck this is a suitable migration target under lock */
 280                 if (!strict && !suitable_migration_target(page))
 281                         break;
 282
 283                 /* Recheck this is a buddy page under lock */
 284                 if (!PageBuddy(page))
 285                         continue;
 286
 287                 /* Found a free page, break it into order-0 pages */
 288                 isolated = split_free_page(page);
 289                 if (!isolated && strict)
 290                         break;
 291                 total_isolated += isolated;
 292                 for (i = 0; i < isolated; i++) {
 293                         list_add(&page->lru, freelist);
 294                         page++;
 295                 }
 296
 297                 /* If a page was split, advance to the end of it */
 298                 if (isolated) {
 299                         blockpfn += isolated - 1;
 300                         cursor += isolated - 1;
 301                 }
 302         }
 303
 304         trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
 305
 306         /*
 307          * If strict isolation is requested by CMA then check that all the
 308          * pages requested were isolated. If there were any failures, 0 is
 309          * returned and CMA will fail.
 310          */
 311         if (strict && nr_strict_required > total_isolated)
 312                 total_isolated = 0;
 313
 314         if (locked)
 315                 spin_unlock_irqrestore(&cc->zone->lock, flags);
 316
 317         /* Update the pageblock-skip if the whole pageblock was scanned */
 318         if (blockpfn == end_pfn)
 319                 update_pageblock_skip(cc, valid_page, total_isolated, false);
 320
 321         count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
 322         if (total_isolated)
 323                 count_compact_events(COMPACTISOLATED, total_isolated);
 324         return total_isolated;
 325 }
 326
 327 /**
 328  * isolate_freepages_range() - isolate free pages.
 329  * @start_pfn: The first PFN to start isolating.
 330  * @end_pfn:   The one-past-last PFN.
 331  *
 332  * Non-free pages, invalid PFNs, or zone boundaries within the
 333  * [start_pfn, end_pfn) range are considered errors, cause function to
 334  * undo its actions and return zero.
 335  *
 336  * Otherwise, function returns one-past-the-last PFN of isolated page
 337  * (which may be greater then end_pfn if end fell in a middle of
 338  * a free page).
 339  */
 340 unsigned long
 341 isolate_freepages_range(struct compact_control *cc,
 342                         unsigned long start_pfn, unsigned long end_pfn)
 343 {
 344         unsigned long isolated, pfn, block_end_pfn;
 345         LIST_HEAD(freelist);
 346
 347         for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
 348                 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
 349                         break;
 350
 351                 /*
 352                  * On subsequent iterations ALIGN() is actually not needed,
 353                  * but we keep it that we not to complicate the code.
 354                  */
 355                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 356                 block_end_pfn = min(block_end_pfn, end_pfn);
 357
 358                 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
 359                                                    &freelist, true);
 360
 361                 /*
 362                  * In strict mode, isolate_freepages_block() returns 0 if
 363                  * there are any holes in the block (ie. invalid PFNs or
 364                  * non-free pages).
 365                  */
 366                 if (!isolated)
 367                         break;
 368
 369                 /*
 370                  * If we managed to isolate pages, it is always (1 << n) *
 371                  * pageblock_nr_pages for some non-negative n.  (Max order
 372                  * page may span two pageblocks).
 373                  */
 374         }
 375
 376         /* split_free_page does not map the pages */
 377         map_pages(&freelist);
 378
 379         if (pfn < end_pfn) {
 380                 /* Loop terminated early, cleanup. */
 381                 release_freepages(&freelist);
 382                 return 0;
 383         }
 384
 385         /* We don't use freelists for anything. */
 386         return pfn;
 387 }
 388
 389 /* Update the number of anon and file isolated pages in the zone */
 390 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
 391 {
 392         struct page *page;
 393         unsigned int count[2] = { 0, };
 394
 395         list_for_each_entry(page, &cc->migratepages, lru)
 396                 count[!!page_is_file_cache(page)]++;
 397
 398         /* If locked we can use the interrupt unsafe versions */
 399         if (locked) {
 400                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 401                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 402         } else {
 403                 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 404                 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 405         }
 406 }
 407
 408 /* Similar to reclaim, but different enough that they don't share logic */
 409 static bool too_many_isolated(struct zone *zone)
 410 {
 411         unsigned long active, inactive, isolated;
 412
 413         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 414                                         zone_page_state(zone, NR_INACTIVE_ANON);
 415         active = zone_page_state(zone, NR_ACTIVE_FILE) +
 416                                         zone_page_state(zone, NR_ACTIVE_ANON);
 417         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 418                                         zone_page_state(zone, NR_ISOLATED_ANON);
 419
 420         return isolated > (inactive + active) / 2;
 421 }
 422
 423 /**
 424  * isolate_migratepages_range() - isolate all migrate-able pages in range.
 425  * @zone:       Zone pages are in.
 426  * @cc:         Compaction control structure.
 427  * @low_pfn:    The first PFN of the range.
 428  * @end_pfn:    The one-past-the-last PFN of the range.
 429  * @unevictable: true if it allows to isolate unevictable pages
 430  *
 431  * Isolate all pages that can be migrated from the range specified by
 432  * [low_pfn, end_pfn).  Returns zero if there is a fatal signal
 433  * pending), otherwise PFN of the first page that was not scanned
 434  * (which may be both less, equal to or more then end_pfn).
 435  *
 436  * Assumes that cc->migratepages is empty and cc->nr_migratepages is
 437  * zero.
 438  *
 439  * Apart from cc->migratepages and cc->nr_migratetypes this function
 440  * does not modify any cc's fields, in particular it does not modify
 441  * (or read for that matter) cc->migrate_pfn.
 442  */
 443 unsigned long
 444 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
 445                 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
 446 {
 447         unsigned long last_pageblock_nr = 0, pageblock_nr;
 448         unsigned long nr_scanned = 0, nr_isolated = 0;
 449         struct list_head *migratelist = &cc->migratepages;
 450         isolate_mode_t mode = 0;
 451         struct lruvec *lruvec;
 452         unsigned long flags;
 453         bool locked = false;
 454         struct page *page = NULL, *valid_page = NULL;
 455
 456         /*
 457          * Ensure that there are not too many pages isolated from the LRU
 458          * list by either parallel reclaimers or compaction. If there are,
 459          * delay for some time until fewer pages are isolated
 460          */
 461         while (unlikely(too_many_isolated(zone))) {
 462                 /* async migration should just abort */
 463                 if (!cc->sync)
 464                         return 0;
 465
 466                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 467
 468                 if (fatal_signal_pending(current))
 469                         return 0;
 470         }
 471
 472         /* Time to isolate some pages for migration */
 473         cond_resched();
 474         for (; low_pfn < end_pfn; low_pfn++) {
 475                 /* give a chance to irqs before checking need_resched() */
 476                 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
 477                         if (should_release_lock(&zone->lru_lock)) {
 478                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 479                                 locked = false;
 480                         }
 481                 }
 482
 483                 /*
 484                  * migrate_pfn does not necessarily start aligned to a
 485                  * pageblock. Ensure that pfn_valid is called when moving
 486                  * into a new MAX_ORDER_NR_PAGES range in case of large
 487                  * memory holes within the zone
 488                  */
 489                 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
 490                         if (!pfn_valid(low_pfn)) {
 491                                 low_pfn += MAX_ORDER_NR_PAGES - 1;
 492                                 continue;
 493                         }
 494                 }
 495
 496                 if (!pfn_valid_within(low_pfn))
 497                         continue;
 498                 nr_scanned++;
 499
 500                 /*
 501                  * Get the page and ensure the page is within the same zone.
 502                  * See the comment in isolate_freepages about overlapping
 503                  * nodes. It is deliberate that the new zone lock is not taken
 504                  * as memory compaction should not move pages between nodes.
 505                  */
 506                 page = pfn_to_page(low_pfn);
 507                 if (page_zone(page) != zone)
 508                         continue;
 509
 510                 if (!valid_page)
 511                         valid_page = page;
 512
 513                 /* If isolation recently failed, do not retry */
 514                 pageblock_nr = low_pfn >> pageblock_order;
 515                 if (!isolation_suitable(cc, page))
 516                         goto next_pageblock;
 517
 518                 /* Skip if free */
 519                 if (PageBuddy(page))
 520                         continue;
 521
 522                 /*
 523                  * For async migration, also only scan in MOVABLE blocks. Async
 524                  * migration is optimistic to see if the minimum amount of work
 525                  * satisfies the allocation
 526                  */
 527                 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
 528                     !migrate_async_suitable(get_pageblock_migratetype(page))) {
 529                         cc->finished_update_migrate = true;
 530                         goto next_pageblock;
 531                 }
 532
 533                 /*
 534                  * Check may be lockless but that's ok as we recheck later.
 535                  * It's possible to migrate LRU pages and balloon pages
 536                  * Skip any other type of page
 537                  */
 538                 if (!PageLRU(page)) {
 539                         if (unlikely(balloon_page_movable(page))) {
 540                                 if (locked && balloon_page_isolate(page)) {
 541                                         /* Successfully isolated */
 542                                         cc->finished_update_migrate = true;
 543                                         list_add(&page->lru, migratelist);
 544                                         cc->nr_migratepages++;
 545                                         nr_isolated++;
 546                                         goto check_compact_cluster;
 547                                 }
 548                         }
 549                         continue;
 550                 }
 551
 552                 /*
 553                  * PageLRU is set. lru_lock normally excludes isolation
 554                  * splitting and collapsing (collapsing has already happened
 555                  * if PageLRU is set) but the lock is not necessarily taken
 556                  * here and it is wasteful to take it just to check transhuge.
 557                  * Check TransHuge without lock and skip the whole pageblock if
 558                  * it's either a transhuge or hugetlbfs page, as calling
 559                  * compound_order() without preventing THP from splitting the
 560                  * page underneath us may return surprising results.
 561                  */
 562                 if (PageTransHuge(page)) {
 563                         if (!locked)
 564                                 goto next_pageblock;
 565                         low_pfn += (1 << compound_order(page)) - 1;
 566                         continue;
 567                 }
 568
 569                 /* Check if it is ok to still hold the lock */
 570                 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
 571                                                                 locked, cc);
 572                 if (!locked || fatal_signal_pending(current))
 573                         break;
 574
 575                 /* Recheck PageLRU and PageTransHuge under lock */
 576                 if (!PageLRU(page))
 577                         continue;
 578                 if (PageTransHuge(page)) {
 579                         low_pfn += (1 << compound_order(page)) - 1;
 580                         continue;
 581                 }
 582
 583                 if (!cc->sync)
 584                         mode |= ISOLATE_ASYNC_MIGRATE;
 585
 586                 if (unevictable)
 587                         mode |= ISOLATE_UNEVICTABLE;
 588
 589                 lruvec = mem_cgroup_page_lruvec(page, zone);
 590
 591                 /* Try isolate the page */
 592                 if (__isolate_lru_page(page, mode) != 0)
 593                         continue;
 594
 595                 VM_BUG_ON(PageTransCompound(page));
 596
 597                 /* Successfully isolated */
 598                 cc->finished_update_migrate = true;
 599                 del_page_from_lru_list(page, lruvec, page_lru(page));
 600                 list_add(&page->lru, migratelist);
 601                 cc->nr_migratepages++;
 602                 nr_isolated++;
 603
 604 check_compact_cluster:
 605                 /* Avoid isolating too much */
 606                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
 607                         ++low_pfn;
 608                         break;
 609                 }
 610
 611                 continue;
 612
 613 next_pageblock:
 614                 low_pfn += pageblock_nr_pages;
 615                 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
 616                 last_pageblock_nr = pageblock_nr;
 617         }
 618
 619         acct_isolated(zone, locked, cc);
 620
 621         if (locked)
 622                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 623
 624         /* Update the pageblock-skip if the whole pageblock was scanned */
 625         if (low_pfn == end_pfn)
 626                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
 627
 628         trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 629
 630         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
 631         if (nr_isolated)
 632                 count_compact_events(COMPACTISOLATED, nr_isolated);
 633
 634         return low_pfn;
 635 }
 636
 637 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 638 #ifdef CONFIG_COMPACTION
 639 /*
 640  * Based on information in the current compact_control, find blocks
 641  * suitable for isolating free pages from and then isolate them.
 642  */
 643 static void isolate_freepages(struct zone *zone,
 644                                 struct compact_control *cc)
 645 {
 646         struct page *page;
 647         unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
 648         int nr_freepages = cc->nr_freepages;
 649         struct list_head *freelist = &cc->freepages;
 650
 651         /*
 652          * Initialise the free scanner. The starting point is where we last
 653          * scanned from (or the end of the zone if starting). The low point
 654          * is the end of the pageblock the migration scanner is using.
 655          */
 656         pfn = cc->free_pfn;
 657         low_pfn = cc->migrate_pfn + pageblock_nr_pages;
 658
 659         /*
 660          * Take care that if the migration scanner is at the end of the zone
 661          * that the free scanner does not accidentally move to the next zone
 662          * in the next isolation cycle.
 663          */
 664         high_pfn = min(low_pfn, pfn);
 665
 666         zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
 667
 668         /*
 669          * Isolate free pages until enough are available to migrate the
 670          * pages on cc->migratepages. We stop searching if the migrate
 671          * and free page scanners meet or enough free pages are isolated.
 672          */
 673         for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
 674                                         pfn -= pageblock_nr_pages) {
 675                 unsigned long isolated;
 676
 677                 if (!pfn_valid(pfn))
 678                         continue;
 679
 680                 /*
 681                  * Check for overlapping nodes/zones. It's possible on some
 682                  * configurations to have a setup like
 683                  * node0 node1 node0
 684                  * i.e. it's possible that all pages within a zones range of
 685                  * pages do not belong to a single zone.
 686                  */
 687                 page = pfn_to_page(pfn);
 688                 if (page_zone(page) != zone)
 689                         continue;
 690
 691                 /* Check the block is suitable for migration */
 692                 if (!suitable_migration_target(page))
 693                         continue;
 694
 695                 /* If isolation recently failed, do not retry */
 696                 if (!isolation_suitable(cc, page))
 697                         continue;
 698
 699                 /* Found a block suitable for isolating free pages from */
 700                 isolated = 0;
 701
 702                 /*
 703                  * As pfn may not start aligned, pfn+pageblock_nr_page
 704                  * may cross a MAX_ORDER_NR_PAGES boundary and miss
 705                  * a pfn_valid check. Ensure isolate_freepages_block()
 706                  * only scans within a pageblock
 707                  */
 708                 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 709                 end_pfn = min(end_pfn, zone_end_pfn);
 710                 isolated = isolate_freepages_block(cc, pfn, end_pfn,
 711                                                    freelist, false);
 712                 nr_freepages += isolated;
 713
 714                 /*
 715                  * Record the highest PFN we isolated pages from. When next
 716                  * looking for free pages, the search will restart here as
 717                  * page migration may have returned some pages to the allocator
 718                  */
 719                 if (isolated) {
 720                         cc->finished_update_free = true;
 721                         high_pfn = max(high_pfn, pfn);
 722                 }
 723         }
 724
 725         /* split_free_page does not map the pages */
 726         map_pages(freelist);
 727
 728         cc->free_pfn = high_pfn;
 729         cc->nr_freepages = nr_freepages;
 730 }
 731
 732 /*
 733  * This is a migrate-callback that "allocates" freepages by taking pages
 734  * from the isolated freelists in the block we are migrating to.
 735  */
 736 static struct page *compaction_alloc(struct page *migratepage,
 737                                         unsigned long data,
 738                                         int **result)
 739 {
 740         struct compact_control *cc = (struct compact_control *)data;
 741         struct page *freepage;
 742
 743         /* Isolate free pages if necessary */
 744         if (list_empty(&cc->freepages)) {
 745                 isolate_freepages(cc->zone, cc);
 746
 747                 if (list_empty(&cc->freepages))
 748                         return NULL;
 749         }
 750
 751         freepage = list_entry(cc->freepages.next, struct page, lru);
 752         list_del(&freepage->lru);
 753         cc->nr_freepages--;
 754
 755         return freepage;
 756 }
 757
 758 /*
 759  * We cannot control nr_migratepages and nr_freepages fully when migration is
 760  * running as migrate_pages() has no knowledge of compact_control. When
 761  * migration is complete, we count the number of pages on the lists by hand.
 762  */
 763 static void update_nr_listpages(struct compact_control *cc)
 764 {
 765         int nr_migratepages = 0;
 766         int nr_freepages = 0;
 767         struct page *page;
 768
 769         list_for_each_entry(page, &cc->migratepages, lru)
 770                 nr_migratepages++;
 771         list_for_each_entry(page, &cc->freepages, lru)
 772                 nr_freepages++;
 773
 774         cc->nr_migratepages = nr_migratepages;
 775         cc->nr_freepages = nr_freepages;
 776 }
 777
 778 /* possible outcome of isolate_migratepages */
 779 typedef enum {
 780         ISOLATE_ABORT,          /* Abort compaction now */
 781         ISOLATE_NONE,           /* No pages isolated, continue scanning */
 782         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
 783 } isolate_migrate_t;
 784
 785 /*
 786  * Isolate all pages that can be migrated from the block pointed to by
 787  * the migrate scanner within compact_control.
 788  */
 789 static isolate_migrate_t isolate_migratepages(struct zone *zone,
 790                                         struct compact_control *cc)
 791 {
 792         unsigned long low_pfn, end_pfn;
 793
 794         /* Do not scan outside zone boundaries */
 795         low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
 796
 797         /* Only scan within a pageblock boundary */
 798         end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
 799
 800         /* Do not cross the free scanner or scan within a memory hole */
 801         if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
 802                 cc->migrate_pfn = end_pfn;
 803                 return ISOLATE_NONE;
 804         }
 805
 806         /* Perform the isolation */
 807         low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
 808         if (!low_pfn || cc->contended)
 809                 return ISOLATE_ABORT;
 810
 811         cc->migrate_pfn = low_pfn;
 812
 813         return ISOLATE_SUCCESS;
 814 }
 815
 816 static int compact_finished(struct zone *zone,
 817                             struct compact_control *cc)
 818 {
 819         unsigned long watermark;
 820
 821         if (fatal_signal_pending(current))
 822                 return COMPACT_PARTIAL;
 823
 824         /* Compaction run completes if the migrate and free scanner meet */
 825         if (cc->free_pfn <= cc->migrate_pfn) {
 826                 /*
 827                  * Mark that the PG_migrate_skip information should be cleared
 828                  * by kswapd when it goes to sleep. kswapd does not set the
 829                  * flag itself as the decision to be clear should be directly
 830                  * based on an allocation request.
 831                  */
 832                 if (!current_is_kswapd())
 833                         zone->compact_blockskip_flush = true;
 834
 835                 return COMPACT_COMPLETE;
 836         }
 837
 838         /*
 839          * order == -1 is expected when compacting via
 840          * /proc/sys/vm/compact_memory
 841          */
 842         if (cc->order == -1)
 843                 return COMPACT_CONTINUE;
 844
 845         /* Compaction run is not finished if the watermark is not met */
 846         watermark = low_wmark_pages(zone);
 847         watermark += (1 << cc->order);
 848
 849         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
 850                 return COMPACT_CONTINUE;
 851
 852         /* Direct compactor: Is a suitable page free? */
 853         if (cc->page) {
 854                 /* Was a suitable page captured? */
 855                 if (*cc->page)
 856                         return COMPACT_PARTIAL;
 857         } else {
 858                 unsigned int order;
 859                 for (order = cc->order; order < MAX_ORDER; order++) {
 860                         struct free_area *area = &zone->free_area[cc->order];
 861                         /* Job done if page is free of the right migratetype */
 862                         if (!list_empty(&area->free_list[cc->migratetype]))
 863                                 return COMPACT_PARTIAL;
 864
 865                         /* Job done if allocation would set block type */
 866                         if (cc->order >= pageblock_order && area->nr_free)
 867                                 return COMPACT_PARTIAL;
 868                 }
 869         }
 870
 871         return COMPACT_CONTINUE;
 872 }
 873
 874 /*
 875  * compaction_suitable: Is this suitable to run compaction on this zone now?
 876  * Returns
 877  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
 878  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
 879  *   COMPACT_CONTINUE - If compaction should run now
 880  */
 881 unsigned long compaction_suitable(struct zone *zone, int order)
 882 {
 883         int fragindex;
 884         unsigned long watermark;
 885
 886         /*
 887          * order == -1 is expected when compacting via
 888          * /proc/sys/vm/compact_memory
 889          */
 890         if (order == -1)
 891                 return COMPACT_CONTINUE;
 892
 893         /*
 894          * Watermarks for order-0 must be met for compaction. Note the 2UL.
 895          * This is because during migration, copies of pages need to be
 896          * allocated and for a short time, the footprint is higher
 897          */
 898         watermark = low_wmark_pages(zone) + (2UL << order);
 899         if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
 900                 return COMPACT_SKIPPED;
 901
 902         /*
 903          * fragmentation index determines if allocation failures are due to
 904          * low memory or external fragmentation
 905          *
 906          * index of -1000 implies allocations might succeed depending on
 907          * watermarks
 908          * index towards 0 implies failure is due to lack of memory
 909          * index towards 1000 implies failure is due to fragmentation
 910          *
 911          * Only compact if a failure would be due to fragmentation.
 912          */
 913         fragindex = fragmentation_index(zone, order);
 914         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
 915                 return COMPACT_SKIPPED;
 916
 917         if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
 918             0, 0))
 919                 return COMPACT_PARTIAL;
 920
 921         return COMPACT_CONTINUE;
 922 }
 923
 924 static void compact_capture_page(struct compact_control *cc)
 925 {
 926         unsigned long flags;
 927         int mtype, mtype_low, mtype_high;
 928
 929         if (!cc->page || *cc->page)
 930                 return;
 931
 932         /*
 933          * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
 934          * regardless of the migratetype of the freelist is is captured from.
 935          * This is fine because the order for a high-order MIGRATE_MOVABLE
 936          * allocation is typically at least a pageblock size and overall
 937          * fragmentation is not impaired. Other allocation types must
 938          * capture pages from their own migratelist because otherwise they
 939          * could pollute other pageblocks like MIGRATE_MOVABLE with
 940          * difficult to move pages and making fragmentation worse overall.
 941          */
 942         if (cc->migratetype == MIGRATE_MOVABLE) {
 943                 mtype_low = 0;
 944                 mtype_high = MIGRATE_PCPTYPES;
 945         } else {
 946                 mtype_low = cc->migratetype;
 947                 mtype_high = cc->migratetype + 1;
 948         }
 949
 950         /* Speculatively examine the free lists without zone lock */
 951         for (mtype = mtype_low; mtype < mtype_high; mtype++) {
 952                 int order;
 953                 for (order = cc->order; order < MAX_ORDER; order++) {
 954                         struct page *page;
 955                         struct free_area *area;
 956                         area = &(cc->zone->free_area[order]);
 957                         if (list_empty(&area->free_list[mtype]))
 958                                 continue;
 959
 960                         /* Take the lock and attempt capture of the page */
 961                         if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
 962                                 return;
 963                         if (!list_empty(&area->free_list[mtype])) {
 964                                 page = list_entry(area->free_list[mtype].next,
 965                                                         struct page, lru);
 966                                 if (capture_free_page(page, cc->order, mtype)) {
 967                                         spin_unlock_irqrestore(&cc->zone->lock,
 968                                                                         flags);
 969                                         *cc->page = page;
 970                                         return;
 971                                 }
 972                         }
 973                         spin_unlock_irqrestore(&cc->zone->lock, flags);
 974                 }
 975         }
 976 }
 977
 978 static int compact_zone(struct zone *zone, struct compact_control *cc)
 979 {
 980         int ret;
 981         unsigned long start_pfn = zone->zone_start_pfn;
 982         unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
 983
 984         ret = compaction_suitable(zone, cc->order);
 985         switch (ret) {
 986         case COMPACT_PARTIAL:
 987         case COMPACT_SKIPPED:
 988                 /* Compaction is likely to fail */
 989                 return ret;
 990         case COMPACT_CONTINUE:
 991                 /* Fall through to compaction */
 992                 ;
 993         }
 994
 995         /*
 996          * Setup to move all movable pages to the end of the zone. Used cached
 997          * information on where the scanners should start but check that it
 998          * is initialised by ensuring the values are within zone boundaries.
 999          */
1000         cc->migrate_pfn = zone->compact_cached_migrate_pfn;
1001         cc->free_pfn = zone->compact_cached_free_pfn;
1002         if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1003                 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1004                 zone->compact_cached_free_pfn = cc->free_pfn;
1005         }
1006         if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1007                 cc->migrate_pfn = start_pfn;
1008                 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
1009         }
1010
1011         /*
1012          * Clear pageblock skip if there were failures recently and compaction
1013          * is about to be retried after being deferred. kswapd does not do
1014          * this reset as it'll reset the cached information when going to sleep.
1015          */
1016         if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1017                 __reset_isolation_suitable(zone);
1018
1019         migrate_prep_local();
1020
1021         while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1022                 unsigned long nr_migrate, nr_remaining;
1023                 int err;
1024
1025                 switch (isolate_migratepages(zone, cc)) {
1026                 case ISOLATE_ABORT:
1027                         ret = COMPACT_PARTIAL;
1028                         putback_movable_pages(&cc->migratepages);
1029                         cc->nr_migratepages = 0;
1030                         goto out;
1031                 case ISOLATE_NONE:
1032                         continue;
1033                 case ISOLATE_SUCCESS:
1034                         ;
1035                 }
1036
1037                 nr_migrate = cc->nr_migratepages;
1038                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1039                                 (unsigned long)cc, false,
1040                                 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1041                                 MR_COMPACTION);
1042                 update_nr_listpages(cc);
1043                 nr_remaining = cc->nr_migratepages;
1044
1045                 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1046                                                 nr_remaining);
1047
1048                 /* Release isolated pages not migrated */
1049                 if (err) {
1050                         putback_movable_pages(&cc->migratepages);
1051                         cc->nr_migratepages = 0;
1052                         if (err == -ENOMEM) {
1053                                 ret = COMPACT_PARTIAL;
1054                                 goto out;
1055                         }
1056                 }
1057
1058                 /* Capture a page now if it is a suitable size */
1059                 compact_capture_page(cc);
1060         }
1061
1062 out:
1063         /* Release free pages and check accounting */
1064         cc->nr_freepages -= release_freepages(&cc->freepages);
1065         VM_BUG_ON(cc->nr_freepages != 0);
1066
1067         return ret;
1068 }
1069
1070 static unsigned long compact_zone_order(struct zone *zone,
1071                                  int order, gfp_t gfp_mask,
1072                                  bool sync, bool *contended,
1073                                  struct page **page)
1074 {
1075         unsigned long ret;
1076         struct compact_control cc = {
1077                 .nr_freepages = 0,
1078                 .nr_migratepages = 0,
1079                 .order = order,
1080                 .migratetype = allocflags_to_migratetype(gfp_mask),
1081                 .zone = zone,
1082                 .sync = sync,
1083                 .page = page,
1084         };
1085         INIT_LIST_HEAD(&cc.freepages);
1086         INIT_LIST_HEAD(&cc.migratepages);
1087
1088         ret = compact_zone(zone, &cc);
1089
1090         VM_BUG_ON(!list_empty(&cc.freepages));
1091         VM_BUG_ON(!list_empty(&cc.migratepages));
1092
1093         *contended = cc.contended;
1094         return ret;
1095 }
1096
1097 int sysctl_extfrag_threshold = 500;
1098
1099 /**
1100  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1101  * @zonelist: The zonelist used for the current allocation
1102  * @order: The order of the current allocation
1103  * @gfp_mask: The GFP mask of the current allocation
1104  * @nodemask: The allowed nodes to allocate from
1105  * @sync: Whether migration is synchronous or not
1106  * @contended: Return value that is true if compaction was aborted due to lock contention
1107  * @page: Optionally capture a free page of the requested order during compaction
1108  *
1109  * This is the main entry point for direct page compaction.
1110  */
1111 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1112                         int order, gfp_t gfp_mask, nodemask_t *nodemask,
1113                         bool sync, bool *contended, struct page **page)
1114 {
1115         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1116         int may_enter_fs = gfp_mask & __GFP_FS;
1117         int may_perform_io = gfp_mask & __GFP_IO;
1118         struct zoneref *z;
1119         struct zone *zone;
1120         int rc = COMPACT_SKIPPED;
1121         int alloc_flags = 0;
1122
1123         /* Check if the GFP flags allow compaction */
1124         if (!order || !may_enter_fs || !may_perform_io)
1125                 return rc;
1126
1127         count_compact_event(COMPACTSTALL);
1128
1129 #ifdef CONFIG_CMA
1130         if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1131                 alloc_flags |= ALLOC_CMA;
1132 #endif
1133         /* Compact each zone in the list */
1134         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1135                                                                 nodemask) {
1136                 int status;
1137
1138                 status = compact_zone_order(zone, order, gfp_mask, sync,
1139                                                 contended, page);
1140                 rc = max(status, rc);
1141
1142                 /* If a normal allocation would succeed, stop compacting */
1143                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1144                                       alloc_flags))
1145                         break;
1146         }
1147
1148         return rc;
1149 }
1150
1151
1152 /* Compact all zones within a node */
1153 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1154 {
1155         int zoneid;
1156         struct zone *zone;
1157
1158         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1159
1160                 zone = &pgdat->node_zones[zoneid];
1161                 if (!populated_zone(zone))
1162                         continue;
1163
1164                 cc->nr_freepages = 0;
1165                 cc->nr_migratepages = 0;
1166                 cc->zone = zone;
1167                 INIT_LIST_HEAD(&cc->freepages);
1168                 INIT_LIST_HEAD(&cc->migratepages);
1169
1170                 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1171                         compact_zone(zone, cc);
1172
1173                 if (cc->order > 0) {
1174                         int ok = zone_watermark_ok(zone, cc->order,
1175                                                 low_wmark_pages(zone), 0, 0);
1176                         if (ok && cc->order >= zone->compact_order_failed)
1177                                 zone->compact_order_failed = cc->order + 1;
1178                         /* Currently async compaction is never deferred. */
1179                         else if (!ok && cc->sync)
1180                                 defer_compaction(zone, cc->order);
1181                 }
1182
1183                 VM_BUG_ON(!list_empty(&cc->freepages));
1184                 VM_BUG_ON(!list_empty(&cc->migratepages));
1185         }
1186
1187         return 0;
1188 }
1189
1190 int compact_pgdat(pg_data_t *pgdat, int order)
1191 {
1192         struct compact_control cc = {
1193                 .order = order,
1194                 .sync = false,
1195                 .page = NULL,
1196         };
1197
1198         return __compact_pgdat(pgdat, &cc);
1199 }
1200
1201 static int compact_node(int nid)
1202 {
1203         struct compact_control cc = {
1204                 .order = -1,
1205                 .sync = true,
1206                 .page = NULL,
1207         };
1208
1209         return __compact_pgdat(NODE_DATA(nid), &cc);
1210 }
1211
1212 /* Compact all nodes in the system */
1213 static int compact_nodes(void)
1214 {
1215         int nid;
1216
1217         /* Flush pending updates to the LRU lists */
1218         lru_add_drain_all();
1219
1220         for_each_online_node(nid)
1221                 compact_node(nid);
1222
1223         return COMPACT_COMPLETE;
1224 }
1225
1226 /* The written value is actually unused, all memory is compacted */
1227 int sysctl_compact_memory;
1228
1229 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1230 int sysctl_compaction_handler(struct ctl_table *table, int write,
1231                         void __user *buffer, size_t *length, loff_t *ppos)
1232 {
1233         if (write)
1234                 return compact_nodes();
1235
1236         return 0;
1237 }
1238
1239 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1240                         void __user *buffer, size_t *length, loff_t *ppos)
1241 {
1242         proc_dointvec_minmax(table, write, buffer, length, ppos);
1243
1244         return 0;
1245 }
1246
1247 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1248 ssize_t sysfs_compact_node(struct device *dev,
1249                         struct device_attribute *attr,
1250                         const char *buf, size_t count)
1251 {
1252         int nid = dev->id;
1253
1254         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1255                 /* Flush pending updates to the LRU lists */
1256                 lru_add_drain_all();
1257
1258                 compact_node(nid);
1259         }
1260
1261         return count;
1262 }
1263 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1264
1265 int compaction_register_node(struct node *node)
1266 {
1267         return device_create_file(&node->dev, &dev_attr_compact);
1268 }
1269
1270 void compaction_unregister_node(struct node *node)
1271 {
1272         return device_remove_file(&node->dev, &dev_attr_compact);
1273 }
1274 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1275
1276 #endif /* CONFIG_COMPACTION */