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 /*
  20  * compact_control is used to track pages being migrated and the free pages
  21  * they are being migrated to during memory compaction. The free_pfn starts
  22  * at the end of a zone and migrate_pfn begins at the start. Movable pages
  23  * are moved to the end of a zone during a compaction run and the run
  24  * completes when free_pfn <= migrate_pfn
  25  */
  26 struct compact_control {
  27         struct list_head freepages;     /* List of free pages to migrate to */
  28         struct list_head migratepages;  /* List of pages being migrated */
  29         unsigned long nr_freepages;     /* Number of isolated free pages */
  30         unsigned long nr_migratepages;  /* Number of pages to migrate */
  31         unsigned long free_pfn;         /* isolate_freepages search base */
  32         unsigned long migrate_pfn;      /* isolate_migratepages search base */
  33
  34         /* Account for isolated anon and file pages */
  35         unsigned long nr_anon;
  36         unsigned long nr_file;
  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 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
  85                 if (!PageBuddy(page))
  86                         continue;
  87
  88                 /* Found a free page, break it into order-0 pages */
  89                 isolated = split_free_page(page);
  90                 total_isolated += isolated;
  91                 for (i = 0; i < isolated; i++) {
  92                         list_add(&page->lru, freelist);
  93                         page++;
  94                 }
  95
  96                 /* If a page was split, advance to the end of it */
  97                 if (isolated) {
  98                         blockpfn += isolated - 1;
  99                         cursor += isolated - 1;
 100                 }
 101         }
 102
 103         return total_isolated;
 104 }
 105
 106 /* Returns true if the page is within a block suitable for migration to */
 107 static bool suitable_migration_target(struct page *page)
 108 {
 109
 110         int migratetype = get_pageblock_migratetype(page);
 111
 112         /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
 113         if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
 114                 return false;
 115
 116         /* If the page is a large free page, then allow migration */
 117         if (PageBuddy(page) && page_order(page) >= pageblock_order)
 118                 return true;
 119
 120         /* If the block is MIGRATE_MOVABLE, allow migration */
 121         if (migratetype == MIGRATE_MOVABLE)
 122                 return true;
 123
 124         /* Otherwise skip the block */
 125         return false;
 126 }
 127
 128 /*
 129  * Based on information in the current compact_control, find blocks
 130  * suitable for isolating free pages from and then isolate them.
 131  */
 132 static void isolate_freepages(struct zone *zone,
 133                                 struct compact_control *cc)
 134 {
 135         struct page *page;
 136         unsigned long high_pfn, low_pfn, pfn;
 137         unsigned long flags;
 138         int nr_freepages = cc->nr_freepages;
 139         struct list_head *freelist = &cc->freepages;
 140
 141         pfn = cc->free_pfn;
 142         low_pfn = cc->migrate_pfn + pageblock_nr_pages;
 143         high_pfn = low_pfn;
 144
 145         /*
 146          * Isolate free pages until enough are available to migrate the
 147          * pages on cc->migratepages. We stop searching if the migrate
 148          * and free page scanners meet or enough free pages are isolated.
 149          */
 150         spin_lock_irqsave(&zone->lock, flags);
 151         for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
 152                                         pfn -= pageblock_nr_pages) {
 153                 unsigned long isolated;
 154
 155                 if (!pfn_valid(pfn))
 156                         continue;
 157
 158                 /*
 159                  * Check for overlapping nodes/zones. It's possible on some
 160                  * configurations to have a setup like
 161                  * node0 node1 node0
 162                  * i.e. it's possible that all pages within a zones range of
 163                  * pages do not belong to a single zone.
 164                  */
 165                 page = pfn_to_page(pfn);
 166                 if (page_zone(page) != zone)
 167                         continue;
 168
 169                 /* Check the block is suitable for migration */
 170                 if (!suitable_migration_target(page))
 171                         continue;
 172
 173                 /* Found a block suitable for isolating free pages from */
 174                 isolated = isolate_freepages_block(zone, pfn, freelist);
 175                 nr_freepages += isolated;
 176
 177                 /*
 178                  * Record the highest PFN we isolated pages from. When next
 179                  * looking for free pages, the search will restart here as
 180                  * page migration may have returned some pages to the allocator
 181                  */
 182                 if (isolated)
 183                         high_pfn = max(high_pfn, pfn);
 184         }
 185         spin_unlock_irqrestore(&zone->lock, flags);
 186
 187         /* split_free_page does not map the pages */
 188         list_for_each_entry(page, freelist, lru) {
 189                 arch_alloc_page(page, 0);
 190                 kernel_map_pages(page, 1, 1);
 191         }
 192
 193         cc->free_pfn = high_pfn;
 194         cc->nr_freepages = nr_freepages;
 195 }
 196
 197 /* Update the number of anon and file isolated pages in the zone */
 198 static void acct_isolated(struct zone *zone, struct compact_control *cc)
 199 {
 200         struct page *page;
 201         unsigned int count[NR_LRU_LISTS] = { 0, };
 202
 203         list_for_each_entry(page, &cc->migratepages, lru) {
 204                 int lru = page_lru_base_type(page);
 205                 count[lru]++;
 206         }
 207
 208         cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
 209         cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
 210         __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
 211         __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
 212 }
 213
 214 /* Similar to reclaim, but different enough that they don't share logic */
 215 static bool too_many_isolated(struct zone *zone)
 216 {
 217
 218         unsigned long inactive, isolated;
 219
 220         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 221                                         zone_page_state(zone, NR_INACTIVE_ANON);
 222         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 223                                         zone_page_state(zone, NR_ISOLATED_ANON);
 224
 225         return isolated > inactive;
 226 }
 227
 228 /*
 229  * Isolate all pages that can be migrated from the block pointed to by
 230  * the migrate scanner within compact_control.
 231  */
 232 static unsigned long isolate_migratepages(struct zone *zone,
 233                                         struct compact_control *cc)
 234 {
 235         unsigned long low_pfn, end_pfn;
 236         struct list_head *migratelist = &cc->migratepages;
 237
 238         /* Do not scan outside zone boundaries */
 239         low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
 240
 241         /* Only scan within a pageblock boundary */
 242         end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
 243
 244         /* Do not cross the free scanner or scan within a memory hole */
 245         if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
 246                 cc->migrate_pfn = end_pfn;
 247                 return 0;
 248         }
 249
 250         /*
 251          * Ensure that there are not too many pages isolated from the LRU
 252          * list by either parallel reclaimers or compaction. If there are,
 253          * delay for some time until fewer pages are isolated
 254          */
 255         while (unlikely(too_many_isolated(zone))) {
 256                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 257
 258                 if (fatal_signal_pending(current))
 259                         return 0;
 260         }
 261
 262         /* Time to isolate some pages for migration */
 263         spin_lock_irq(&zone->lru_lock);
 264         for (; low_pfn < end_pfn; low_pfn++) {
 265                 struct page *page;
 266                 if (!pfn_valid_within(low_pfn))
 267                         continue;
 268
 269                 /* Get the page and skip if free */
 270                 page = pfn_to_page(low_pfn);
 271                 if (PageBuddy(page))
 272                         continue;
 273
 274                 /* Try isolate the page */
 275                 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
 276                         continue;
 277
 278                 /* Successfully isolated */
 279                 del_page_from_lru_list(zone, page, page_lru(page));
 280                 list_add(&page->lru, migratelist);
 281                 mem_cgroup_del_lru(page);
 282                 cc->nr_migratepages++;
 283
 284                 /* Avoid isolating too much */
 285                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 286                         break;
 287         }
 288
 289         acct_isolated(zone, cc);
 290
 291         spin_unlock_irq(&zone->lru_lock);
 292         cc->migrate_pfn = low_pfn;
 293
 294         return cc->nr_migratepages;
 295 }
 296
 297 /*
 298  * This is a migrate-callback that "allocates" freepages by taking pages
 299  * from the isolated freelists in the block we are migrating to.
 300  */
 301 static struct page *compaction_alloc(struct page *migratepage,
 302                                         unsigned long data,
 303                                         int **result)
 304 {
 305         struct compact_control *cc = (struct compact_control *)data;
 306         struct page *freepage;
 307
 308         /* Isolate free pages if necessary */
 309         if (list_empty(&cc->freepages)) {
 310                 isolate_freepages(cc->zone, cc);
 311
 312                 if (list_empty(&cc->freepages))
 313                         return NULL;
 314         }
 315
 316         freepage = list_entry(cc->freepages.next, struct page, lru);
 317         list_del(&freepage->lru);
 318         cc->nr_freepages--;
 319
 320         return freepage;
 321 }
 322
 323 /*
 324  * We cannot control nr_migratepages and nr_freepages fully when migration is
 325  * running as migrate_pages() has no knowledge of compact_control. When
 326  * migration is complete, we count the number of pages on the lists by hand.
 327  */
 328 static void update_nr_listpages(struct compact_control *cc)
 329 {
 330         int nr_migratepages = 0;
 331         int nr_freepages = 0;
 332         struct page *page;
 333
 334         list_for_each_entry(page, &cc->migratepages, lru)
 335                 nr_migratepages++;
 336         list_for_each_entry(page, &cc->freepages, lru)
 337                 nr_freepages++;
 338
 339         cc->nr_migratepages = nr_migratepages;
 340         cc->nr_freepages = nr_freepages;
 341 }
 342
 343 static int compact_finished(struct zone *zone,
 344                                                 struct compact_control *cc)
 345 {
 346         unsigned int order;
 347         unsigned long watermark = low_wmark_pages(zone) + (1 << cc->order);
 348
 349         if (fatal_signal_pending(current))
 350                 return COMPACT_PARTIAL;
 351
 352         /* Compaction run completes if the migrate and free scanner meet */
 353         if (cc->free_pfn <= cc->migrate_pfn)
 354                 return COMPACT_COMPLETE;
 355
 356         /* Compaction run is not finished if the watermark is not met */
 357         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
 358                 return COMPACT_CONTINUE;
 359
 360         if (cc->order == -1)
 361                 return COMPACT_CONTINUE;
 362
 363         /* Direct compactor: Is a suitable page free? */
 364         for (order = cc->order; order < MAX_ORDER; order++) {
 365                 /* Job done if page is free of the right migratetype */
 366                 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
 367                         return COMPACT_PARTIAL;
 368
 369                 /* Job done if allocation would set block type */
 370                 if (order >= pageblock_order && zone->free_area[order].nr_free)
 371                         return COMPACT_PARTIAL;
 372         }
 373
 374         return COMPACT_CONTINUE;
 375 }
 376
 377 static int compact_zone(struct zone *zone, struct compact_control *cc)
 378 {
 379         int ret;
 380
 381         /* Setup to move all movable pages to the end of the zone */
 382         cc->migrate_pfn = zone->zone_start_pfn;
 383         cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
 384         cc->free_pfn &= ~(pageblock_nr_pages-1);
 385
 386         migrate_prep_local();
 387
 388         while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
 389                 unsigned long nr_migrate, nr_remaining;
 390
 391                 if (!isolate_migratepages(zone, cc))
 392                         continue;
 393
 394                 nr_migrate = cc->nr_migratepages;
 395                 migrate_pages(&cc->migratepages, compaction_alloc,
 396                                                 (unsigned long)cc, 0);
 397                 update_nr_listpages(cc);
 398                 nr_remaining = cc->nr_migratepages;
 399
 400                 count_vm_event(COMPACTBLOCKS);
 401                 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
 402                 if (nr_remaining)
 403                         count_vm_events(COMPACTPAGEFAILED, nr_remaining);
 404
 405                 /* Release LRU pages not migrated */
 406                 if (!list_empty(&cc->migratepages)) {
 407                         putback_lru_pages(&cc->migratepages);
 408                         cc->nr_migratepages = 0;
 409                 }
 410
 411         }
 412
 413         /* Release free pages and check accounting */
 414         cc->nr_freepages -= release_freepages(&cc->freepages);
 415         VM_BUG_ON(cc->nr_freepages != 0);
 416
 417         return ret;
 418 }
 419
 420 static unsigned long compact_zone_order(struct zone *zone,
 421                                                 int order, gfp_t gfp_mask)
 422 {
 423         struct compact_control cc = {
 424                 .nr_freepages = 0,
 425                 .nr_migratepages = 0,
 426                 .order = order,
 427                 .migratetype = allocflags_to_migratetype(gfp_mask),
 428                 .zone = zone,
 429         };
 430         INIT_LIST_HEAD(&cc.freepages);
 431         INIT_LIST_HEAD(&cc.migratepages);
 432
 433         return compact_zone(zone, &cc);
 434 }
 435
 436 int sysctl_extfrag_threshold = 500;
 437
 438 /**
 439  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
 440  * @zonelist: The zonelist used for the current allocation
 441  * @order: The order of the current allocation
 442  * @gfp_mask: The GFP mask of the current allocation
 443  * @nodemask: The allowed nodes to allocate from
 444  *
 445  * This is the main entry point for direct page compaction.
 446  */
 447 unsigned long try_to_compact_pages(struct zonelist *zonelist,
 448                         int order, gfp_t gfp_mask, nodemask_t *nodemask)
 449 {
 450         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 451         int may_enter_fs = gfp_mask & __GFP_FS;
 452         int may_perform_io = gfp_mask & __GFP_IO;
 453         unsigned long watermark;
 454         struct zoneref *z;
 455         struct zone *zone;
 456         int rc = COMPACT_SKIPPED;
 457
 458         /*
 459          * Check whether it is worth even starting compaction. The order check is
 460          * made because an assumption is made that the page allocator can satisfy
 461          * the "cheaper" orders without taking special steps
 462          */
 463         if (order <= PAGE_ALLOC_COSTLY_ORDER || !may_enter_fs || !may_perform_io)
 464                 return rc;
 465
 466         count_vm_event(COMPACTSTALL);
 467
 468         /* Compact each zone in the list */
 469         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
 470                                                                 nodemask) {
 471                 int fragindex;
 472                 int status;
 473
 474                 /*
 475                  * Watermarks for order-0 must be met for compaction. Note
 476                  * the 2UL. This is because during migration, copies of
 477                  * pages need to be allocated and for a short time, the
 478                  * footprint is higher
 479                  */
 480                 watermark = low_wmark_pages(zone) + (2UL << order);
 481                 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
 482                         continue;
 483
 484                 /*
 485                  * fragmentation index determines if allocation failures are
 486                  * due to low memory or external fragmentation
 487                  *
 488                  * index of -1 implies allocations might succeed depending
 489                  *      on watermarks
 490                  * index towards 0 implies failure is due to lack of memory
 491                  * index towards 1000 implies failure is due to fragmentation
 492                  *
 493                  * Only compact if a failure would be due to fragmentation.
 494                  */
 495                 fragindex = fragmentation_index(zone, order);
 496                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
 497                         continue;
 498
 499                 if (fragindex == -1 && zone_watermark_ok(zone, order, watermark, 0, 0)) {
 500                         rc = COMPACT_PARTIAL;
 501                         break;
 502                 }
 503
 504                 status = compact_zone_order(zone, order, gfp_mask);
 505                 rc = max(status, rc);
 506
 507                 if (zone_watermark_ok(zone, order, watermark, 0, 0))
 508                         break;
 509         }
 510
 511         return rc;
 512 }
 513
 514
 515 /* Compact all zones within a node */
 516 static int compact_node(int nid)
 517 {
 518         int zoneid;
 519         pg_data_t *pgdat;
 520         struct zone *zone;
 521
 522         if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
 523                 return -EINVAL;
 524         pgdat = NODE_DATA(nid);
 525
 526         /* Flush pending updates to the LRU lists */
 527         lru_add_drain_all();
 528
 529         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 530                 struct compact_control cc = {
 531                         .nr_freepages = 0,
 532                         .nr_migratepages = 0,
 533                         .order = -1,
 534                 };
 535
 536                 zone = &pgdat->node_zones[zoneid];
 537                 if (!populated_zone(zone))
 538                         continue;
 539
 540                 cc.zone = zone;
 541                 INIT_LIST_HEAD(&cc.freepages);
 542                 INIT_LIST_HEAD(&cc.migratepages);
 543
 544                 compact_zone(zone, &cc);
 545
 546                 VM_BUG_ON(!list_empty(&cc.freepages));
 547                 VM_BUG_ON(!list_empty(&cc.migratepages));
 548         }
 549
 550         return 0;
 551 }
 552
 553 /* Compact all nodes in the system */
 554 static int compact_nodes(void)
 555 {
 556         int nid;
 557
 558         for_each_online_node(nid)
 559                 compact_node(nid);
 560
 561         return COMPACT_COMPLETE;
 562 }
 563
 564 /* The written value is actually unused, all memory is compacted */
 565 int sysctl_compact_memory;
 566
 567 /* This is the entry point for compacting all nodes via /proc/sys/vm */
 568 int sysctl_compaction_handler(struct ctl_table *table, int write,
 569                         void __user *buffer, size_t *length, loff_t *ppos)
 570 {
 571         if (write)
 572                 return compact_nodes();
 573
 574         return 0;
 575 }
 576
 577 int sysctl_extfrag_handler(struct ctl_table *table, int write,
 578                         void __user *buffer, size_t *length, loff_t *ppos)
 579 {
 580         proc_dointvec_minmax(table, write, buffer, length, ppos);
 581
 582         return 0;
 583 }
 584
 585 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
 586 ssize_t sysfs_compact_node(struct sys_device *dev,
 587                         struct sysdev_attribute *attr,
 588                         const char *buf, size_t count)
 589 {
 590         compact_node(dev->id);
 591
 592         return count;
 593 }
 594 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
 595
 596 int compaction_register_node(struct node *node)
 597 {
 598         return sysdev_create_file(&node->sysdev, &attr_compact);
 599 }
 600
 601 void compaction_unregister_node(struct node *node)
 602 {
 603         return sysdev_remove_file(&node->sysdev, &attr_compact);
 604 }
 605 #endif /* CONFIG_SYSFS && CONFIG_NUMA */