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 #define CREATE_TRACE_POINTS
  20 #include <trace/events/compaction.h>
  21
  22 /*
  23  * compact_control is used to track pages being migrated and the free pages
  24  * they are being migrated to during memory compaction. The free_pfn starts
  25  * at the end of a zone and migrate_pfn begins at the start. Movable pages
  26  * are moved to the end of a zone during a compaction run and the run
  27  * completes when free_pfn <= migrate_pfn
  28  */
  29 struct compact_control {
  30         struct list_head freepages;     /* List of free pages to migrate to */
  31         struct list_head migratepages;  /* List of pages being migrated */
  32         unsigned long nr_freepages;     /* Number of isolated free pages */
  33         unsigned long nr_migratepages;  /* Number of pages to migrate */
  34         unsigned long free_pfn;         /* isolate_freepages search base */
  35         unsigned long migrate_pfn;      /* isolate_migratepages search base */
  36         bool sync;                      /* Synchronous migration */
  37
  38         unsigned int order;             /* order a direct compactor needs */
  39         int migratetype;                /* MOVABLE, RECLAIMABLE etc */
  40         struct zone *zone;
  41 };
  42
  43 static unsigned long release_freepages(struct list_head *freelist)
  44 {
  45         struct page *page, *next;
  46         unsigned long count = 0;
  47
  48         list_for_each_entry_safe(page, next, freelist, lru) {
  49                 list_del(&page->lru);
  50                 __free_page(page);
  51                 count++;
  52         }
  53
  54         return count;
  55 }
  56
  57 /* Isolate free pages onto a private freelist. Must hold zone->lock */
  58 static unsigned long isolate_freepages_block(struct zone *zone,
  59                                 unsigned long blockpfn,
  60                                 struct list_head *freelist)
  61 {
  62         unsigned long zone_end_pfn, end_pfn;
  63         int nr_scanned = 0, total_isolated = 0;
  64         struct page *cursor;
  65
  66         /* Get the last PFN we should scan for free pages at */
  67         zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
  68         end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
  69
  70         /* Find the first usable PFN in the block to initialse page cursor */
  71         for (; blockpfn < end_pfn; blockpfn++) {
  72                 if (pfn_valid_within(blockpfn))
  73                         break;
  74         }
  75         cursor = pfn_to_page(blockpfn);
  76
  77         /* Isolate free pages. This assumes the block is valid */
  78         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
  79                 int isolated, i;
  80                 struct page *page = cursor;
  81
  82                 if (!pfn_valid_within(blockpfn))
  83                         continue;
  84                 nr_scanned++;
  85
  86                 if (!PageBuddy(page))
  87                         continue;
  88
  89                 /* Found a free page, break it into order-0 pages */
  90                 isolated = split_free_page(page);
  91                 total_isolated += isolated;
  92                 for (i = 0; i < isolated; i++) {
  93                         list_add(&page->lru, freelist);
  94                         page++;
  95                 }
  96
  97                 /* If a page was split, advance to the end of it */
  98                 if (isolated) {
  99                         blockpfn += isolated - 1;
 100                         cursor += isolated - 1;
 101                 }
 102         }
 103
 104         trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
 105         return total_isolated;
 106 }
 107
 108 /* Returns true if the page is within a block suitable for migration to */
 109 static bool suitable_migration_target(struct page *page)
 110 {
 111
 112         int migratetype = get_pageblock_migratetype(page);
 113
 114         /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
 115         if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
 116                 return false;
 117
 118         /* If the page is a large free page, then allow migration */
 119         if (PageBuddy(page) && page_order(page) >= pageblock_order)
 120                 return true;
 121
 122         /* If the block is MIGRATE_MOVABLE, allow migration */
 123         if (migratetype == MIGRATE_MOVABLE)
 124                 return true;
 125
 126         /* Otherwise skip the block */
 127         return false;
 128 }
 129
 130 /*
 131  * Based on information in the current compact_control, find blocks
 132  * suitable for isolating free pages from and then isolate them.
 133  */
 134 static void isolate_freepages(struct zone *zone,
 135                                 struct compact_control *cc)
 136 {
 137         struct page *page;
 138         unsigned long high_pfn, low_pfn, pfn;
 139         unsigned long flags;
 140         int nr_freepages = cc->nr_freepages;
 141         struct list_head *freelist = &cc->freepages;
 142
 143         /*
 144          * Initialise the free scanner. The starting point is where we last
 145          * scanned from (or the end of the zone if starting). The low point
 146          * is the end of the pageblock the migration scanner is using.
 147          */
 148         pfn = cc->free_pfn;
 149         low_pfn = cc->migrate_pfn + pageblock_nr_pages;
 150
 151         /*
 152          * Take care that if the migration scanner is at the end of the zone
 153          * that the free scanner does not accidentally move to the next zone
 154          * in the next isolation cycle.
 155          */
 156         high_pfn = min(low_pfn, pfn);
 157
 158         /*
 159          * Isolate free pages until enough are available to migrate the
 160          * pages on cc->migratepages. We stop searching if the migrate
 161          * and free page scanners meet or enough free pages are isolated.
 162          */
 163         for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
 164                                         pfn -= pageblock_nr_pages) {
 165                 unsigned long isolated;
 166
 167                 if (!pfn_valid(pfn))
 168                         continue;
 169
 170                 /*
 171                  * Check for overlapping nodes/zones. It's possible on some
 172                  * configurations to have a setup like
 173                  * node0 node1 node0
 174                  * i.e. it's possible that all pages within a zones range of
 175                  * pages do not belong to a single zone.
 176                  */
 177                 page = pfn_to_page(pfn);
 178                 if (page_zone(page) != zone)
 179                         continue;
 180
 181                 /* Check the block is suitable for migration */
 182                 if (!suitable_migration_target(page))
 183                         continue;
 184
 185                 /*
 186                  * Found a block suitable for isolating free pages from. Now
 187                  * we disabled interrupts, double check things are ok and
 188                  * isolate the pages. This is to minimise the time IRQs
 189                  * are disabled
 190                  */
 191                 isolated = 0;
 192                 spin_lock_irqsave(&zone->lock, flags);
 193                 if (suitable_migration_target(page)) {
 194                         isolated = isolate_freepages_block(zone, pfn, freelist);
 195                         nr_freepages += isolated;
 196                 }
 197                 spin_unlock_irqrestore(&zone->lock, flags);
 198
 199                 /*
 200                  * Record the highest PFN we isolated pages from. When next
 201                  * looking for free pages, the search will restart here as
 202                  * page migration may have returned some pages to the allocator
 203                  */
 204                 if (isolated)
 205                         high_pfn = max(high_pfn, pfn);
 206         }
 207
 208         /* split_free_page does not map the pages */
 209         list_for_each_entry(page, freelist, lru) {
 210                 arch_alloc_page(page, 0);
 211                 kernel_map_pages(page, 1, 1);
 212         }
 213
 214         cc->free_pfn = high_pfn;
 215         cc->nr_freepages = nr_freepages;
 216 }
 217
 218 /* Update the number of anon and file isolated pages in the zone */
 219 static void acct_isolated(struct zone *zone, struct compact_control *cc)
 220 {
 221         struct page *page;
 222         unsigned int count[2] = { 0, };
 223
 224         list_for_each_entry(page, &cc->migratepages, lru)
 225                 count[!!page_is_file_cache(page)]++;
 226
 227         __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
 228         __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 229 }
 230
 231 /* Similar to reclaim, but different enough that they don't share logic */
 232 static bool too_many_isolated(struct zone *zone)
 233 {
 234         unsigned long active, inactive, isolated;
 235
 236         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 237                                         zone_page_state(zone, NR_INACTIVE_ANON);
 238         active = zone_page_state(zone, NR_ACTIVE_FILE) +
 239                                         zone_page_state(zone, NR_ACTIVE_ANON);
 240         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 241                                         zone_page_state(zone, NR_ISOLATED_ANON);
 242
 243         return isolated > (inactive + active) / 2;
 244 }
 245
 246 /* possible outcome of isolate_migratepages */
 247 typedef enum {
 248         ISOLATE_ABORT,          /* Abort compaction now */
 249         ISOLATE_NONE,           /* No pages isolated, continue scanning */
 250         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
 251 } isolate_migrate_t;
 252
 253 /*
 254  * Isolate all pages that can be migrated from the block pointed to by
 255  * the migrate scanner within compact_control.
 256  */
 257 static isolate_migrate_t isolate_migratepages(struct zone *zone,
 258                                         struct compact_control *cc)
 259 {
 260         unsigned long low_pfn, end_pfn;
 261         unsigned long last_pageblock_nr = 0, pageblock_nr;
 262         unsigned long nr_scanned = 0, nr_isolated = 0;
 263         struct list_head *migratelist = &cc->migratepages;
 264         isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE;
 265
 266         /* Do not scan outside zone boundaries */
 267         low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
 268
 269         /* Only scan within a pageblock boundary */
 270         end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
 271
 272         /* Do not cross the free scanner or scan within a memory hole */
 273         if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
 274                 cc->migrate_pfn = end_pfn;
 275                 return ISOLATE_NONE;
 276         }
 277
 278         /*
 279          * Ensure that there are not too many pages isolated from the LRU
 280          * list by either parallel reclaimers or compaction. If there are,
 281          * delay for some time until fewer pages are isolated
 282          */
 283         while (unlikely(too_many_isolated(zone))) {
 284                 /* async migration should just abort */
 285                 if (!cc->sync)
 286                         return ISOLATE_ABORT;
 287
 288                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 289
 290                 if (fatal_signal_pending(current))
 291                         return ISOLATE_ABORT;
 292         }
 293
 294         /* Time to isolate some pages for migration */
 295         cond_resched();
 296         spin_lock_irq(&zone->lru_lock);
 297         for (; low_pfn < end_pfn; low_pfn++) {
 298                 struct page *page;
 299                 bool locked = true;
 300
 301                 /* give a chance to irqs before checking need_resched() */
 302                 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
 303                         spin_unlock_irq(&zone->lru_lock);
 304                         locked = false;
 305                 }
 306                 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
 307                         if (locked)
 308                                 spin_unlock_irq(&zone->lru_lock);
 309                         cond_resched();
 310                         spin_lock_irq(&zone->lru_lock);
 311                         if (fatal_signal_pending(current))
 312                                 break;
 313                 } else if (!locked)
 314                         spin_lock_irq(&zone->lru_lock);
 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                                 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
 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_CLEAN;
 376
 377                 /* Try isolate the page */
 378                 if (__isolate_lru_page(page, mode, 0) != 0)
 379                         continue;
 380
 381                 VM_BUG_ON(PageTransCompound(page));
 382
 383                 /* Successfully isolated */
 384                 del_page_from_lru_list(zone, page, page_lru(page));
 385                 list_add(&page->lru, migratelist);
 386                 cc->nr_migratepages++;
 387                 nr_isolated++;
 388
 389                 /* Avoid isolating too much */
 390                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 391                         break;
 392         }
 393
 394         acct_isolated(zone, cc);
 395
 396         spin_unlock_irq(&zone->lru_lock);
 397         cc->migrate_pfn = low_pfn;
 398
 399         trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 400
 401         return ISOLATE_SUCCESS;
 402 }
 403
 404 /*
 405  * This is a migrate-callback that "allocates" freepages by taking pages
 406  * from the isolated freelists in the block we are migrating to.
 407  */
 408 static struct page *compaction_alloc(struct page *migratepage,
 409                                         unsigned long data,
 410                                         int **result)
 411 {
 412         struct compact_control *cc = (struct compact_control *)data;
 413         struct page *freepage;
 414
 415         /* Isolate free pages if necessary */
 416         if (list_empty(&cc->freepages)) {
 417                 isolate_freepages(cc->zone, cc);
 418
 419                 if (list_empty(&cc->freepages))
 420                         return NULL;
 421         }
 422
 423         freepage = list_entry(cc->freepages.next, struct page, lru);
 424         list_del(&freepage->lru);
 425         cc->nr_freepages--;
 426
 427         return freepage;
 428 }
 429
 430 /*
 431  * We cannot control nr_migratepages and nr_freepages fully when migration is
 432  * running as migrate_pages() has no knowledge of compact_control. When
 433  * migration is complete, we count the number of pages on the lists by hand.
 434  */
 435 static void update_nr_listpages(struct compact_control *cc)
 436 {
 437         int nr_migratepages = 0;
 438         int nr_freepages = 0;
 439         struct page *page;
 440
 441         list_for_each_entry(page, &cc->migratepages, lru)
 442                 nr_migratepages++;
 443         list_for_each_entry(page, &cc->freepages, lru)
 444                 nr_freepages++;
 445
 446         cc->nr_migratepages = nr_migratepages;
 447         cc->nr_freepages = nr_freepages;
 448 }
 449
 450 static int compact_finished(struct zone *zone,
 451                             struct compact_control *cc)
 452 {
 453         unsigned int order;
 454         unsigned long watermark;
 455
 456         if (fatal_signal_pending(current))
 457                 return COMPACT_PARTIAL;
 458
 459         /* Compaction run completes if the migrate and free scanner meet */
 460         if (cc->free_pfn <= cc->migrate_pfn)
 461                 return COMPACT_COMPLETE;
 462
 463         /*
 464          * order == -1 is expected when compacting via
 465          * /proc/sys/vm/compact_memory
 466          */
 467         if (cc->order == -1)
 468                 return COMPACT_CONTINUE;
 469
 470         /* Compaction run is not finished if the watermark is not met */
 471         watermark = low_wmark_pages(zone);
 472         watermark += (1 << cc->order);
 473
 474         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
 475                 return COMPACT_CONTINUE;
 476
 477         /* Direct compactor: Is a suitable page free? */
 478         for (order = cc->order; order < MAX_ORDER; order++) {
 479                 /* Job done if page is free of the right migratetype */
 480                 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
 481                         return COMPACT_PARTIAL;
 482
 483                 /* Job done if allocation would set block type */
 484                 if (order >= pageblock_order && zone->free_area[order].nr_free)
 485                         return COMPACT_PARTIAL;
 486         }
 487
 488         return COMPACT_CONTINUE;
 489 }
 490
 491 /*
 492  * compaction_suitable: Is this suitable to run compaction on this zone now?
 493  * Returns
 494  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
 495  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
 496  *   COMPACT_CONTINUE - If compaction should run now
 497  */
 498 unsigned long compaction_suitable(struct zone *zone, int order)
 499 {
 500         int fragindex;
 501         unsigned long watermark;
 502
 503         /*
 504          * order == -1 is expected when compacting via
 505          * /proc/sys/vm/compact_memory
 506          */
 507         if (order == -1)
 508                 return COMPACT_CONTINUE;
 509
 510         /*
 511          * Watermarks for order-0 must be met for compaction. Note the 2UL.
 512          * This is because during migration, copies of pages need to be
 513          * allocated and for a short time, the footprint is higher
 514          */
 515         watermark = low_wmark_pages(zone) + (2UL << order);
 516         if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
 517                 return COMPACT_SKIPPED;
 518
 519         /*
 520          * fragmentation index determines if allocation failures are due to
 521          * low memory or external fragmentation
 522          *
 523          * index of -1000 implies allocations might succeed depending on
 524          * watermarks
 525          * index towards 0 implies failure is due to lack of memory
 526          * index towards 1000 implies failure is due to fragmentation
 527          *
 528          * Only compact if a failure would be due to fragmentation.
 529          */
 530         fragindex = fragmentation_index(zone, order);
 531         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
 532                 return COMPACT_SKIPPED;
 533
 534         if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
 535             0, 0))
 536                 return COMPACT_PARTIAL;
 537
 538         return COMPACT_CONTINUE;
 539 }
 540
 541 static int compact_zone(struct zone *zone, struct compact_control *cc)
 542 {
 543         int ret;
 544
 545         ret = compaction_suitable(zone, cc->order);
 546         switch (ret) {
 547         case COMPACT_PARTIAL:
 548         case COMPACT_SKIPPED:
 549                 /* Compaction is likely to fail */
 550                 return ret;
 551         case COMPACT_CONTINUE:
 552                 /* Fall through to compaction */
 553                 ;
 554         }
 555
 556         /* Setup to move all movable pages to the end of the zone */
 557         cc->migrate_pfn = zone->zone_start_pfn;
 558         cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
 559         cc->free_pfn &= ~(pageblock_nr_pages-1);
 560
 561         migrate_prep_local();
 562
 563         while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
 564                 unsigned long nr_migrate, nr_remaining;
 565                 int err;
 566
 567                 switch (isolate_migratepages(zone, cc)) {
 568                 case ISOLATE_ABORT:
 569                         ret = COMPACT_PARTIAL;
 570                         goto out;
 571                 case ISOLATE_NONE:
 572                         continue;
 573                 case ISOLATE_SUCCESS:
 574                         ;
 575                 }
 576
 577                 nr_migrate = cc->nr_migratepages;
 578                 err = migrate_pages(&cc->migratepages, compaction_alloc,
 579                                 (unsigned long)cc, false,
 580                                 cc->sync);
 581                 update_nr_listpages(cc);
 582                 nr_remaining = cc->nr_migratepages;
 583
 584                 count_vm_event(COMPACTBLOCKS);
 585                 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
 586                 if (nr_remaining)
 587                         count_vm_events(COMPACTPAGEFAILED, nr_remaining);
 588                 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
 589                                                 nr_remaining);
 590
 591                 /* Release LRU pages not migrated */
 592                 if (err) {
 593                         putback_lru_pages(&cc->migratepages);
 594                         cc->nr_migratepages = 0;
 595                         if (err == -ENOMEM) {
 596                                 ret = COMPACT_PARTIAL;
 597                                 goto out;
 598                         }
 599                 }
 600         }
 601
 602 out:
 603         /* Release free pages and check accounting */
 604         cc->nr_freepages -= release_freepages(&cc->freepages);
 605         VM_BUG_ON(cc->nr_freepages != 0);
 606
 607         return ret;
 608 }
 609
 610 static unsigned long compact_zone_order(struct zone *zone,
 611                                  int order, gfp_t gfp_mask,
 612                                  bool sync)
 613 {
 614         struct compact_control cc = {
 615                 .nr_freepages = 0,
 616                 .nr_migratepages = 0,
 617                 .order = order,
 618                 .migratetype = allocflags_to_migratetype(gfp_mask),
 619                 .zone = zone,
 620                 .sync = sync,
 621         };
 622         INIT_LIST_HEAD(&cc.freepages);
 623         INIT_LIST_HEAD(&cc.migratepages);
 624
 625         return compact_zone(zone, &cc);
 626 }
 627
 628 int sysctl_extfrag_threshold = 500;
 629
 630 /**
 631  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
 632  * @zonelist: The zonelist used for the current allocation
 633  * @order: The order of the current allocation
 634  * @gfp_mask: The GFP mask of the current allocation
 635  * @nodemask: The allowed nodes to allocate from
 636  * @sync: Whether migration is synchronous or not
 637  *
 638  * This is the main entry point for direct page compaction.
 639  */
 640 unsigned long try_to_compact_pages(struct zonelist *zonelist,
 641                         int order, gfp_t gfp_mask, nodemask_t *nodemask,
 642                         bool sync)
 643 {
 644         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 645         int may_enter_fs = gfp_mask & __GFP_FS;
 646         int may_perform_io = gfp_mask & __GFP_IO;
 647         struct zoneref *z;
 648         struct zone *zone;
 649         int rc = COMPACT_SKIPPED;
 650
 651         /*
 652          * Check whether it is worth even starting compaction. The order check is
 653          * made because an assumption is made that the page allocator can satisfy
 654          * the "cheaper" orders without taking special steps
 655          */
 656         if (!order || !may_enter_fs || !may_perform_io)
 657                 return rc;
 658
 659         count_vm_event(COMPACTSTALL);
 660
 661         /* Compact each zone in the list */
 662         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
 663                                                                 nodemask) {
 664                 int status;
 665
 666                 status = compact_zone_order(zone, order, gfp_mask, sync);
 667                 rc = max(status, rc);
 668
 669                 /* If a normal allocation would succeed, stop compacting */
 670                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
 671                         break;
 672         }
 673
 674         return rc;
 675 }
 676
 677
 678 /* Compact all zones within a node */
 679 static int compact_node(int nid)
 680 {
 681         int zoneid;
 682         pg_data_t *pgdat;
 683         struct zone *zone;
 684
 685         if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
 686                 return -EINVAL;
 687         pgdat = NODE_DATA(nid);
 688
 689         /* Flush pending updates to the LRU lists */
 690         lru_add_drain_all();
 691
 692         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 693                 struct compact_control cc = {
 694                         .nr_freepages = 0,
 695                         .nr_migratepages = 0,
 696                         .order = -1,
 697                 };
 698
 699                 zone = &pgdat->node_zones[zoneid];
 700                 if (!populated_zone(zone))
 701                         continue;
 702
 703                 cc.zone = zone;
 704                 INIT_LIST_HEAD(&cc.freepages);
 705                 INIT_LIST_HEAD(&cc.migratepages);
 706
 707                 compact_zone(zone, &cc);
 708
 709                 VM_BUG_ON(!list_empty(&cc.freepages));
 710                 VM_BUG_ON(!list_empty(&cc.migratepages));
 711         }
 712
 713         return 0;
 714 }
 715
 716 /* Compact all nodes in the system */
 717 static int compact_nodes(void)
 718 {
 719         int nid;
 720
 721         for_each_online_node(nid)
 722                 compact_node(nid);
 723
 724         return COMPACT_COMPLETE;
 725 }
 726
 727 /* The written value is actually unused, all memory is compacted */
 728 int sysctl_compact_memory;
 729
 730 /* This is the entry point for compacting all nodes via /proc/sys/vm */
 731 int sysctl_compaction_handler(struct ctl_table *table, int write,
 732                         void __user *buffer, size_t *length, loff_t *ppos)
 733 {
 734         if (write)
 735                 return compact_nodes();
 736
 737         return 0;
 738 }
 739
 740 int sysctl_extfrag_handler(struct ctl_table *table, int write,
 741                         void __user *buffer, size_t *length, loff_t *ppos)
 742 {
 743         proc_dointvec_minmax(table, write, buffer, length, ppos);
 744
 745         return 0;
 746 }
 747
 748 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
 749 ssize_t sysfs_compact_node(struct sys_device *dev,
 750                         struct sysdev_attribute *attr,
 751                         const char *buf, size_t count)
 752 {
 753         compact_node(dev->id);
 754
 755         return count;
 756 }
 757 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
 758
 759 int compaction_register_node(struct node *node)
 760 {
 761         return sysdev_create_file(&node->sysdev, &attr_compact);
 762 }
 763
 764 void compaction_unregister_node(struct node *node)
 765 {
 766         return sysdev_remove_file(&node->sysdev, &attr_compact);
 767 }
 768 #endif /* CONFIG_SYSFS && CONFIG_NUMA */