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 "internal.h"
  18
  19 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
  20
  21 #define CREATE_TRACE_POINTS
  22 #include <trace/events/compaction.h>
  23
  24 static unsigned long release_freepages(struct list_head *freelist)
  25 {
  26         struct page *page, *next;
  27         unsigned long count = 0;
  28
  29         list_for_each_entry_safe(page, next, freelist, lru) {
  30                 list_del(&page->lru);
  31                 __free_page(page);
  32                 count++;
  33         }
  34
  35         return count;
  36 }
  37
  38 static void map_pages(struct list_head *list)
  39 {
  40         struct page *page;
  41
  42         list_for_each_entry(page, list, lru) {
  43                 arch_alloc_page(page, 0);
  44                 kernel_map_pages(page, 1, 1);
  45         }
  46 }
  47
  48 static inline bool migrate_async_suitable(int migratetype)
  49 {
  50         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
  51 }
  52
  53 /*
  54  * Compaction requires the taking of some coarse locks that are potentially
  55  * very heavily contended. Check if the process needs to be scheduled or
  56  * if the lock is contended. For async compaction, back out in the event
  57  * if contention is severe. For sync compaction, schedule.
  58  *
  59  * Returns true if the lock is held.
  60  * Returns false if the lock is released and compaction should abort
  61  */
  62 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
  63                                       bool locked, struct compact_control *cc)
  64 {
  65         if (need_resched() || spin_is_contended(lock)) {
  66                 if (locked) {
  67                         spin_unlock_irqrestore(lock, *flags);
  68                         locked = false;
  69                 }
  70
  71                 /* async aborts if taking too long or contended */
  72                 if (!cc->sync) {
  73                         if (cc->contended)
  74                                 *cc->contended = true;
  75                         return false;
  76                 }
  77
  78                 cond_resched();
  79                 if (fatal_signal_pending(current))
  80                         return false;
  81         }
  82
  83         if (!locked)
  84                 spin_lock_irqsave(lock, *flags);
  85         return true;
  86 }
  87
  88 static inline bool compact_trylock_irqsave(spinlock_t *lock,
  89                         unsigned long *flags, struct compact_control *cc)
  90 {
  91         return compact_checklock_irqsave(lock, flags, false, cc);
  92 }
  93
  94 /*
  95  * Isolate free pages onto a private freelist. Caller must hold zone->lock.
  96  * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
  97  * pages inside of the pageblock (even though it may still end up isolating
  98  * some pages).
  99  */
 100 static unsigned long isolate_freepages_block(unsigned long blockpfn,
 101                                 unsigned long end_pfn,
 102                                 struct list_head *freelist,
 103                                 bool strict)
 104 {
 105         int nr_scanned = 0, total_isolated = 0;
 106         struct page *cursor;
 107
 108         cursor = pfn_to_page(blockpfn);
 109
 110         /* Isolate free pages. This assumes the block is valid */
 111         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
 112                 int isolated, i;
 113                 struct page *page = cursor;
 114
 115                 if (!pfn_valid_within(blockpfn)) {
 116                         if (strict)
 117                                 return 0;
 118                         continue;
 119                 }
 120                 nr_scanned++;
 121
 122                 if (!PageBuddy(page)) {
 123                         if (strict)
 124                                 return 0;
 125                         continue;
 126                 }
 127
 128                 /* Found a free page, break it into order-0 pages */
 129                 isolated = split_free_page(page);
 130                 if (!isolated && strict)
 131                         return 0;
 132                 total_isolated += isolated;
 133                 for (i = 0; i < isolated; i++) {
 134                         list_add(&page->lru, freelist);
 135                         page++;
 136                 }
 137
 138                 /* If a page was split, advance to the end of it */
 139                 if (isolated) {
 140                         blockpfn += isolated - 1;
 141                         cursor += isolated - 1;
 142                 }
 143         }
 144
 145         trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
 146         return total_isolated;
 147 }
 148
 149 /**
 150  * isolate_freepages_range() - isolate free pages.
 151  * @start_pfn: The first PFN to start isolating.
 152  * @end_pfn:   The one-past-last PFN.
 153  *
 154  * Non-free pages, invalid PFNs, or zone boundaries within the
 155  * [start_pfn, end_pfn) range are considered errors, cause function to
 156  * undo its actions and return zero.
 157  *
 158  * Otherwise, function returns one-past-the-last PFN of isolated page
 159  * (which may be greater then end_pfn if end fell in a middle of
 160  * a free page).
 161  */
 162 unsigned long
 163 isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
 164 {
 165         unsigned long isolated, pfn, block_end_pfn, flags;
 166         struct zone *zone = NULL;
 167         LIST_HEAD(freelist);
 168
 169         if (pfn_valid(start_pfn))
 170                 zone = page_zone(pfn_to_page(start_pfn));
 171
 172         for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
 173                 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
 174                         break;
 175
 176                 /*
 177                  * On subsequent iterations ALIGN() is actually not needed,
 178                  * but we keep it that we not to complicate the code.
 179                  */
 180                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 181                 block_end_pfn = min(block_end_pfn, end_pfn);
 182
 183                 spin_lock_irqsave(&zone->lock, flags);
 184                 isolated = isolate_freepages_block(pfn, block_end_pfn,
 185                                                    &freelist, true);
 186                 spin_unlock_irqrestore(&zone->lock, flags);
 187
 188                 /*
 189                  * In strict mode, isolate_freepages_block() returns 0 if
 190                  * there are any holes in the block (ie. invalid PFNs or
 191                  * non-free pages).
 192                  */
 193                 if (!isolated)
 194                         break;
 195
 196                 /*
 197                  * If we managed to isolate pages, it is always (1 << n) *
 198                  * pageblock_nr_pages for some non-negative n.  (Max order
 199                  * page may span two pageblocks).
 200                  */
 201         }
 202
 203         /* split_free_page does not map the pages */
 204         map_pages(&freelist);
 205
 206         if (pfn < end_pfn) {
 207                 /* Loop terminated early, cleanup. */
 208                 release_freepages(&freelist);
 209                 return 0;
 210         }
 211
 212         /* We don't use freelists for anything. */
 213         return pfn;
 214 }
 215
 216 /* Update the number of anon and file isolated pages in the zone */
 217 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
 218 {
 219         struct page *page;
 220         unsigned int count[2] = { 0, };
 221
 222         list_for_each_entry(page, &cc->migratepages, lru)
 223                 count[!!page_is_file_cache(page)]++;
 224
 225         /* If locked we can use the interrupt unsafe versions */
 226         if (locked) {
 227                 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 228                 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 229         } else {
 230                 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 231                 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 232         }
 233 }
 234
 235 /* Similar to reclaim, but different enough that they don't share logic */
 236 static bool too_many_isolated(struct zone *zone)
 237 {
 238         unsigned long active, inactive, isolated;
 239
 240         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 241                                         zone_page_state(zone, NR_INACTIVE_ANON);
 242         active = zone_page_state(zone, NR_ACTIVE_FILE) +
 243                                         zone_page_state(zone, NR_ACTIVE_ANON);
 244         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 245                                         zone_page_state(zone, NR_ISOLATED_ANON);
 246
 247         return isolated > (inactive + active) / 2;
 248 }
 249
 250 /**
 251  * isolate_migratepages_range() - isolate all migrate-able pages in range.
 252  * @zone:       Zone pages are in.
 253  * @cc:         Compaction control structure.
 254  * @low_pfn:    The first PFN of the range.
 255  * @end_pfn:    The one-past-the-last PFN of the range.
 256  *
 257  * Isolate all pages that can be migrated from the range specified by
 258  * [low_pfn, end_pfn).  Returns zero if there is a fatal signal
 259  * pending), otherwise PFN of the first page that was not scanned
 260  * (which may be both less, equal to or more then end_pfn).
 261  *
 262  * Assumes that cc->migratepages is empty and cc->nr_migratepages is
 263  * zero.
 264  *
 265  * Apart from cc->migratepages and cc->nr_migratetypes this function
 266  * does not modify any cc's fields, in particular it does not modify
 267  * (or read for that matter) cc->migrate_pfn.
 268  */
 269 unsigned long
 270 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
 271                            unsigned long low_pfn, unsigned long end_pfn)
 272 {
 273         unsigned long last_pageblock_nr = 0, pageblock_nr;
 274         unsigned long nr_scanned = 0, nr_isolated = 0;
 275         struct list_head *migratelist = &cc->migratepages;
 276         isolate_mode_t mode = 0;
 277         struct lruvec *lruvec;
 278         unsigned long flags;
 279         bool locked;
 280
 281         /*
 282          * Ensure that there are not too many pages isolated from the LRU
 283          * list by either parallel reclaimers or compaction. If there are,
 284          * delay for some time until fewer pages are isolated
 285          */
 286         while (unlikely(too_many_isolated(zone))) {
 287                 /* async migration should just abort */
 288                 if (!cc->sync)
 289                         return 0;
 290
 291                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 292
 293                 if (fatal_signal_pending(current))
 294                         return 0;
 295         }
 296
 297         /* Time to isolate some pages for migration */
 298         cond_resched();
 299         spin_lock_irqsave(&zone->lru_lock, flags);
 300         locked = true;
 301         for (; low_pfn < end_pfn; low_pfn++) {
 302                 struct page *page;
 303
 304                 /* give a chance to irqs before checking need_resched() */
 305                 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
 306                         spin_unlock_irqrestore(&zone->lru_lock, flags);
 307                         locked = false;
 308                 }
 309
 310                 /* Check if it is ok to still hold the lock */
 311                 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
 312                                                                 locked, cc);
 313                 if (!locked)
 314                         break;
 315
 316                 /*
 317                  * migrate_pfn does not necessarily start aligned to a
 318                  * pageblock. Ensure that pfn_valid is called when moving
 319                  * into a new MAX_ORDER_NR_PAGES range in case of large
 320                  * memory holes within the zone
 321                  */
 322                 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
 323                         if (!pfn_valid(low_pfn)) {
 324                                 low_pfn += MAX_ORDER_NR_PAGES - 1;
 325                                 continue;
 326                         }
 327                 }
 328
 329                 if (!pfn_valid_within(low_pfn))
 330                         continue;
 331                 nr_scanned++;
 332
 333                 /*
 334                  * Get the page and ensure the page is within the same zone.
 335                  * See the comment in isolate_freepages about overlapping
 336                  * nodes. It is deliberate that the new zone lock is not taken
 337                  * as memory compaction should not move pages between nodes.
 338                  */
 339                 page = pfn_to_page(low_pfn);
 340                 if (page_zone(page) != zone)
 341                         continue;
 342
 343                 /* Skip if free */
 344                 if (PageBuddy(page))
 345                         continue;
 346
 347                 /*
 348                  * For async migration, also only scan in MOVABLE blocks. Async
 349                  * migration is optimistic to see if the minimum amount of work
 350                  * satisfies the allocation
 351                  */
 352                 pageblock_nr = low_pfn >> pageblock_order;
 353                 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
 354                     !migrate_async_suitable(get_pageblock_migratetype(page))) {
 355                         low_pfn += pageblock_nr_pages;
 356                         low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
 357                         last_pageblock_nr = pageblock_nr;
 358                         continue;
 359                 }
 360
 361                 if (!PageLRU(page))
 362                         continue;
 363
 364                 /*
 365                  * PageLRU is set, and lru_lock excludes isolation,
 366                  * splitting and collapsing (collapsing has already
 367                  * happened if PageLRU is set).
 368                  */
 369                 if (PageTransHuge(page)) {
 370                         low_pfn += (1 << compound_order(page)) - 1;
 371                         continue;
 372                 }
 373
 374                 if (!cc->sync)
 375                         mode |= ISOLATE_ASYNC_MIGRATE;
 376
 377                 lruvec = mem_cgroup_page_lruvec(page, zone);
 378
 379                 /* Try isolate the page */
 380                 if (__isolate_lru_page(page, mode) != 0)
 381                         continue;
 382
 383                 VM_BUG_ON(PageTransCompound(page));
 384
 385                 /* Successfully isolated */
 386                 del_page_from_lru_list(page, lruvec, page_lru(page));
 387                 list_add(&page->lru, migratelist);
 388                 cc->nr_migratepages++;
 389                 nr_isolated++;
 390
 391                 /* Avoid isolating too much */
 392                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
 393                         ++low_pfn;
 394                         break;
 395                 }
 396         }
 397
 398         acct_isolated(zone, locked, cc);
 399
 400         if (locked)
 401                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 402
 403         trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 404
 405         return low_pfn;
 406 }
 407
 408 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 409 #ifdef CONFIG_COMPACTION
 410
 411 /* Returns true if the page is within a block suitable for migration to */
 412 static bool suitable_migration_target(struct page *page)
 413 {
 414
 415         int migratetype = get_pageblock_migratetype(page);
 416
 417         /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
 418         if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
 419                 return false;
 420
 421         /* If the page is a large free page, then allow migration */
 422         if (PageBuddy(page) && page_order(page) >= pageblock_order)
 423                 return true;
 424
 425         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
 426         if (migrate_async_suitable(migratetype))
 427                 return true;
 428
 429         /* Otherwise skip the block */
 430         return false;
 431 }
 432
 433 /*
 434  * Returns the start pfn of the last page block in a zone.  This is the starting
 435  * point for full compaction of a zone.  Compaction searches for free pages from
 436  * the end of each zone, while isolate_freepages_block scans forward inside each
 437  * page block.
 438  */
 439 static unsigned long start_free_pfn(struct zone *zone)
 440 {
 441         unsigned long free_pfn;
 442         free_pfn = zone->zone_start_pfn + zone->spanned_pages;
 443         free_pfn &= ~(pageblock_nr_pages-1);
 444         return free_pfn;
 445 }
 446
 447 /*
 448  * Based on information in the current compact_control, find blocks
 449  * suitable for isolating free pages from and then isolate them.
 450  */
 451 static void isolate_freepages(struct zone *zone,
 452                                 struct compact_control *cc)
 453 {
 454         struct page *page;
 455         unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
 456         unsigned long flags;
 457         int nr_freepages = cc->nr_freepages;
 458         struct list_head *freelist = &cc->freepages;
 459
 460         /*
 461          * Initialise the free scanner. The starting point is where we last
 462          * scanned from (or the end of the zone if starting). The low point
 463          * is the end of the pageblock the migration scanner is using.
 464          */
 465         pfn = cc->free_pfn;
 466         low_pfn = cc->migrate_pfn + pageblock_nr_pages;
 467
 468         /*
 469          * Take care that if the migration scanner is at the end of the zone
 470          * that the free scanner does not accidentally move to the next zone
 471          * in the next isolation cycle.
 472          */
 473         high_pfn = min(low_pfn, pfn);
 474
 475         zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
 476
 477         /*
 478          * Isolate free pages until enough are available to migrate the
 479          * pages on cc->migratepages. We stop searching if the migrate
 480          * and free page scanners meet or enough free pages are isolated.
 481          */
 482         for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
 483                                         pfn -= pageblock_nr_pages) {
 484                 unsigned long isolated;
 485
 486                 if (!pfn_valid(pfn))
 487                         continue;
 488
 489                 /*
 490                  * Check for overlapping nodes/zones. It's possible on some
 491                  * configurations to have a setup like
 492                  * node0 node1 node0
 493                  * i.e. it's possible that all pages within a zones range of
 494                  * pages do not belong to a single zone.
 495                  */
 496                 page = pfn_to_page(pfn);
 497                 if (page_zone(page) != zone)
 498                         continue;
 499
 500                 /* Check the block is suitable for migration */
 501                 if (!suitable_migration_target(page))
 502                         continue;
 503
 504                 /*
 505                  * Found a block suitable for isolating free pages from. Now
 506                  * we disabled interrupts, double check things are ok and
 507                  * isolate the pages. This is to minimise the time IRQs
 508                  * are disabled
 509                  */
 510                 isolated = 0;
 511
 512                 /*
 513                  * The zone lock must be held to isolate freepages. This
 514                  * unfortunately this is a very coarse lock and can be
 515                  * heavily contended if there are parallel allocations
 516                  * or parallel compactions. For async compaction do not
 517                  * spin on the lock
 518                  */
 519                 if (!compact_trylock_irqsave(&zone->lock, &flags, cc))
 520                         break;
 521                 if (suitable_migration_target(page)) {
 522                         end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
 523                         isolated = isolate_freepages_block(pfn, end_pfn,
 524                                                            freelist, false);
 525                         nr_freepages += isolated;
 526                 }
 527                 spin_unlock_irqrestore(&zone->lock, flags);
 528
 529                 /*
 530                  * Record the highest PFN we isolated pages from. When next
 531                  * looking for free pages, the search will restart here as
 532                  * page migration may have returned some pages to the allocator
 533                  */
 534                 if (isolated) {
 535                         high_pfn = max(high_pfn, pfn);
 536
 537                         /*
 538                          * If the free scanner has wrapped, update
 539                          * compact_cached_free_pfn to point to the highest
 540                          * pageblock with free pages. This reduces excessive
 541                          * scanning of full pageblocks near the end of the
 542                          * zone
 543                          */
 544                         if (cc->order > 0 && cc->wrapped)
 545                                 zone->compact_cached_free_pfn = high_pfn;
 546                 }
 547         }
 548
 549         /* split_free_page does not map the pages */
 550         map_pages(freelist);
 551
 552         cc->free_pfn = high_pfn;
 553         cc->nr_freepages = nr_freepages;
 554
 555         /* If compact_cached_free_pfn is reset then set it now */
 556         if (cc->order > 0 && !cc->wrapped &&
 557                         zone->compact_cached_free_pfn == start_free_pfn(zone))
 558                 zone->compact_cached_free_pfn = high_pfn;
 559 }
 560
 561 /*
 562  * This is a migrate-callback that "allocates" freepages by taking pages
 563  * from the isolated freelists in the block we are migrating to.
 564  */
 565 static struct page *compaction_alloc(struct page *migratepage,
 566                                         unsigned long data,
 567                                         int **result)
 568 {
 569         struct compact_control *cc = (struct compact_control *)data;
 570         struct page *freepage;
 571
 572         /* Isolate free pages if necessary */
 573         if (list_empty(&cc->freepages)) {
 574                 isolate_freepages(cc->zone, cc);
 575
 576                 if (list_empty(&cc->freepages))
 577                         return NULL;
 578         }
 579
 580         freepage = list_entry(cc->freepages.next, struct page, lru);
 581         list_del(&freepage->lru);
 582         cc->nr_freepages--;
 583
 584         return freepage;
 585 }
 586
 587 /*
 588  * We cannot control nr_migratepages and nr_freepages fully when migration is
 589  * running as migrate_pages() has no knowledge of compact_control. When
 590  * migration is complete, we count the number of pages on the lists by hand.
 591  */
 592 static void update_nr_listpages(struct compact_control *cc)
 593 {
 594         int nr_migratepages = 0;
 595         int nr_freepages = 0;
 596         struct page *page;
 597
 598         list_for_each_entry(page, &cc->migratepages, lru)
 599                 nr_migratepages++;
 600         list_for_each_entry(page, &cc->freepages, lru)
 601                 nr_freepages++;
 602
 603         cc->nr_migratepages = nr_migratepages;
 604         cc->nr_freepages = nr_freepages;
 605 }
 606
 607 /* possible outcome of isolate_migratepages */
 608 typedef enum {
 609         ISOLATE_ABORT,          /* Abort compaction now */
 610         ISOLATE_NONE,           /* No pages isolated, continue scanning */
 611         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
 612 } isolate_migrate_t;
 613
 614 /*
 615  * Isolate all pages that can be migrated from the block pointed to by
 616  * the migrate scanner within compact_control.
 617  */
 618 static isolate_migrate_t isolate_migratepages(struct zone *zone,
 619                                         struct compact_control *cc)
 620 {
 621         unsigned long low_pfn, end_pfn;
 622
 623         /* Do not scan outside zone boundaries */
 624         low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
 625
 626         /* Only scan within a pageblock boundary */
 627         end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
 628
 629         /* Do not cross the free scanner or scan within a memory hole */
 630         if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
 631                 cc->migrate_pfn = end_pfn;
 632                 return ISOLATE_NONE;
 633         }
 634
 635         /* Perform the isolation */
 636         low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
 637         if (!low_pfn)
 638                 return ISOLATE_ABORT;
 639
 640         cc->migrate_pfn = low_pfn;
 641
 642         return ISOLATE_SUCCESS;
 643 }
 644
 645 static int compact_finished(struct zone *zone,
 646                             struct compact_control *cc)
 647 {
 648         unsigned int order;
 649         unsigned long watermark;
 650
 651         if (fatal_signal_pending(current))
 652                 return COMPACT_PARTIAL;
 653
 654         /*
 655          * A full (order == -1) compaction run starts at the beginning and
 656          * end of a zone; it completes when the migrate and free scanner meet.
 657          * A partial (order > 0) compaction can start with the free scanner
 658          * at a random point in the zone, and may have to restart.
 659          */
 660         if (cc->free_pfn <= cc->migrate_pfn) {
 661                 if (cc->order > 0 && !cc->wrapped) {
 662                         /* We started partway through; restart at the end. */
 663                         unsigned long free_pfn = start_free_pfn(zone);
 664                         zone->compact_cached_free_pfn = free_pfn;
 665                         cc->free_pfn = free_pfn;
 666                         cc->wrapped = 1;
 667                         return COMPACT_CONTINUE;
 668                 }
 669                 return COMPACT_COMPLETE;
 670         }
 671
 672         /* We wrapped around and ended up where we started. */
 673         if (cc->wrapped && cc->free_pfn <= cc->start_free_pfn)
 674                 return COMPACT_COMPLETE;
 675
 676         /*
 677          * order == -1 is expected when compacting via
 678          * /proc/sys/vm/compact_memory
 679          */
 680         if (cc->order == -1)
 681                 return COMPACT_CONTINUE;
 682
 683         /* Compaction run is not finished if the watermark is not met */
 684         watermark = low_wmark_pages(zone);
 685         watermark += (1 << cc->order);
 686
 687         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
 688                 return COMPACT_CONTINUE;
 689
 690         /* Direct compactor: Is a suitable page free? */
 691         for (order = cc->order; order < MAX_ORDER; order++) {
 692                 /* Job done if page is free of the right migratetype */
 693                 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
 694                         return COMPACT_PARTIAL;
 695
 696                 /* Job done if allocation would set block type */
 697                 if (order >= pageblock_order && zone->free_area[order].nr_free)
 698                         return COMPACT_PARTIAL;
 699         }
 700
 701         return COMPACT_CONTINUE;
 702 }
 703
 704 /*
 705  * compaction_suitable: Is this suitable to run compaction on this zone now?
 706  * Returns
 707  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
 708  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
 709  *   COMPACT_CONTINUE - If compaction should run now
 710  */
 711 unsigned long compaction_suitable(struct zone *zone, int order)
 712 {
 713         int fragindex;
 714         unsigned long watermark;
 715
 716         /*
 717          * order == -1 is expected when compacting via
 718          * /proc/sys/vm/compact_memory
 719          */
 720         if (order == -1)
 721                 return COMPACT_CONTINUE;
 722
 723         /*
 724          * Watermarks for order-0 must be met for compaction. Note the 2UL.
 725          * This is because during migration, copies of pages need to be
 726          * allocated and for a short time, the footprint is higher
 727          */
 728         watermark = low_wmark_pages(zone) + (2UL << order);
 729         if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
 730                 return COMPACT_SKIPPED;
 731
 732         /*
 733          * fragmentation index determines if allocation failures are due to
 734          * low memory or external fragmentation
 735          *
 736          * index of -1000 implies allocations might succeed depending on
 737          * watermarks
 738          * index towards 0 implies failure is due to lack of memory
 739          * index towards 1000 implies failure is due to fragmentation
 740          *
 741          * Only compact if a failure would be due to fragmentation.
 742          */
 743         fragindex = fragmentation_index(zone, order);
 744         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
 745                 return COMPACT_SKIPPED;
 746
 747         if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
 748             0, 0))
 749                 return COMPACT_PARTIAL;
 750
 751         return COMPACT_CONTINUE;
 752 }
 753
 754 static int compact_zone(struct zone *zone, struct compact_control *cc)
 755 {
 756         int ret;
 757
 758         ret = compaction_suitable(zone, cc->order);
 759         switch (ret) {
 760         case COMPACT_PARTIAL:
 761         case COMPACT_SKIPPED:
 762                 /* Compaction is likely to fail */
 763                 return ret;
 764         case COMPACT_CONTINUE:
 765                 /* Fall through to compaction */
 766                 ;
 767         }
 768
 769         /* Setup to move all movable pages to the end of the zone */
 770         cc->migrate_pfn = zone->zone_start_pfn;
 771
 772         if (cc->order > 0) {
 773                 /* Incremental compaction. Start where the last one stopped. */
 774                 cc->free_pfn = zone->compact_cached_free_pfn;
 775                 cc->start_free_pfn = cc->free_pfn;
 776         } else {
 777                 /* Order == -1 starts at the end of the zone. */
 778                 cc->free_pfn = start_free_pfn(zone);
 779         }
 780
 781         migrate_prep_local();
 782
 783         while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
 784                 unsigned long nr_migrate, nr_remaining;
 785                 int err;
 786
 787                 switch (isolate_migratepages(zone, cc)) {
 788                 case ISOLATE_ABORT:
 789                         ret = COMPACT_PARTIAL;
 790                         goto out;
 791                 case ISOLATE_NONE:
 792                         continue;
 793                 case ISOLATE_SUCCESS:
 794                         ;
 795                 }
 796
 797                 nr_migrate = cc->nr_migratepages;
 798                 err = migrate_pages(&cc->migratepages, compaction_alloc,
 799                                 (unsigned long)cc, false,
 800                                 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
 801                 update_nr_listpages(cc);
 802                 nr_remaining = cc->nr_migratepages;
 803
 804                 count_vm_event(COMPACTBLOCKS);
 805                 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
 806                 if (nr_remaining)
 807                         count_vm_events(COMPACTPAGEFAILED, nr_remaining);
 808                 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
 809                                                 nr_remaining);
 810
 811                 /* Release LRU pages not migrated */
 812                 if (err) {
 813                         putback_lru_pages(&cc->migratepages);
 814                         cc->nr_migratepages = 0;
 815                         if (err == -ENOMEM) {
 816                                 ret = COMPACT_PARTIAL;
 817                                 goto out;
 818                         }
 819                 }
 820         }
 821
 822 out:
 823         /* Release free pages and check accounting */
 824         cc->nr_freepages -= release_freepages(&cc->freepages);
 825         VM_BUG_ON(cc->nr_freepages != 0);
 826
 827         return ret;
 828 }
 829
 830 static unsigned long compact_zone_order(struct zone *zone,
 831                                  int order, gfp_t gfp_mask,
 832                                  bool sync, bool *contended)
 833 {
 834         struct compact_control cc = {
 835                 .nr_freepages = 0,
 836                 .nr_migratepages = 0,
 837                 .order = order,
 838                 .migratetype = allocflags_to_migratetype(gfp_mask),
 839                 .zone = zone,
 840                 .sync = sync,
 841                 .contended = contended,
 842         };
 843         INIT_LIST_HEAD(&cc.freepages);
 844         INIT_LIST_HEAD(&cc.migratepages);
 845
 846         return compact_zone(zone, &cc);
 847 }
 848
 849 int sysctl_extfrag_threshold = 500;
 850
 851 /**
 852  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
 853  * @zonelist: The zonelist used for the current allocation
 854  * @order: The order of the current allocation
 855  * @gfp_mask: The GFP mask of the current allocation
 856  * @nodemask: The allowed nodes to allocate from
 857  * @sync: Whether migration is synchronous or not
 858  *
 859  * This is the main entry point for direct page compaction.
 860  */
 861 unsigned long try_to_compact_pages(struct zonelist *zonelist,
 862                         int order, gfp_t gfp_mask, nodemask_t *nodemask,
 863                         bool sync, bool *contended)
 864 {
 865         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 866         int may_enter_fs = gfp_mask & __GFP_FS;
 867         int may_perform_io = gfp_mask & __GFP_IO;
 868         struct zoneref *z;
 869         struct zone *zone;
 870         int rc = COMPACT_SKIPPED;
 871
 872         /*
 873          * Check whether it is worth even starting compaction. The order check is
 874          * made because an assumption is made that the page allocator can satisfy
 875          * the "cheaper" orders without taking special steps
 876          */
 877         if (!order || !may_enter_fs || !may_perform_io)
 878                 return rc;
 879
 880         count_vm_event(COMPACTSTALL);
 881
 882         /* Compact each zone in the list */
 883         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
 884                                                                 nodemask) {
 885                 int status;
 886
 887                 status = compact_zone_order(zone, order, gfp_mask, sync,
 888                                                 contended);
 889                 rc = max(status, rc);
 890
 891                 /* If a normal allocation would succeed, stop compacting */
 892                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
 893                         break;
 894         }
 895
 896         return rc;
 897 }
 898
 899
 900 /* Compact all zones within a node */
 901 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
 902 {
 903         int zoneid;
 904         struct zone *zone;
 905
 906         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 907
 908                 zone = &pgdat->node_zones[zoneid];
 909                 if (!populated_zone(zone))
 910                         continue;
 911
 912                 cc->nr_freepages = 0;
 913                 cc->nr_migratepages = 0;
 914                 cc->zone = zone;
 915                 INIT_LIST_HEAD(&cc->freepages);
 916                 INIT_LIST_HEAD(&cc->migratepages);
 917
 918                 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
 919                         compact_zone(zone, cc);
 920
 921                 if (cc->order > 0) {
 922                         int ok = zone_watermark_ok(zone, cc->order,
 923                                                 low_wmark_pages(zone), 0, 0);
 924                         if (ok && cc->order >= zone->compact_order_failed)
 925                                 zone->compact_order_failed = cc->order + 1;
 926                         /* Currently async compaction is never deferred. */
 927                         else if (!ok && cc->sync)
 928                                 defer_compaction(zone, cc->order);
 929                 }
 930
 931                 VM_BUG_ON(!list_empty(&cc->freepages));
 932                 VM_BUG_ON(!list_empty(&cc->migratepages));
 933         }
 934
 935         return 0;
 936 }
 937
 938 int compact_pgdat(pg_data_t *pgdat, int order)
 939 {
 940         struct compact_control cc = {
 941                 .order = order,
 942                 .sync = false,
 943         };
 944
 945         return __compact_pgdat(pgdat, &cc);
 946 }
 947
 948 static int compact_node(int nid)
 949 {
 950         struct compact_control cc = {
 951                 .order = -1,
 952                 .sync = true,
 953         };
 954
 955         return __compact_pgdat(NODE_DATA(nid), &cc);
 956 }
 957
 958 /* Compact all nodes in the system */
 959 static int compact_nodes(void)
 960 {
 961         int nid;
 962
 963         /* Flush pending updates to the LRU lists */
 964         lru_add_drain_all();
 965
 966         for_each_online_node(nid)
 967                 compact_node(nid);
 968
 969         return COMPACT_COMPLETE;
 970 }
 971
 972 /* The written value is actually unused, all memory is compacted */
 973 int sysctl_compact_memory;
 974
 975 /* This is the entry point for compacting all nodes via /proc/sys/vm */
 976 int sysctl_compaction_handler(struct ctl_table *table, int write,
 977                         void __user *buffer, size_t *length, loff_t *ppos)
 978 {
 979         if (write)
 980                 return compact_nodes();
 981
 982         return 0;
 983 }
 984
 985 int sysctl_extfrag_handler(struct ctl_table *table, int write,
 986                         void __user *buffer, size_t *length, loff_t *ppos)
 987 {
 988         proc_dointvec_minmax(table, write, buffer, length, ppos);
 989
 990         return 0;
 991 }
 992
 993 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
 994 ssize_t sysfs_compact_node(struct device *dev,
 995                         struct device_attribute *attr,
 996                         const char *buf, size_t count)
 997 {
 998         int nid = dev->id;
 999
1000         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1001                 /* Flush pending updates to the LRU lists */
1002                 lru_add_drain_all();
1003
1004                 compact_node(nid);
1005         }
1006
1007         return count;
1008 }
1009 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1010
1011 int compaction_register_node(struct node *node)
1012 {
1013         return device_create_file(&node->dev, &dev_attr_compact);
1014 }
1015
1016 void compaction_unregister_node(struct node *node)
1017 {
1018         return device_remove_file(&node->dev, &dev_attr_compact);
1019 }
1020 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1021
1022 #endif /* CONFIG_COMPACTION */