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         unsigned long active, inactive, isolated;
 218
 219         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
 220                                         zone_page_state(zone, NR_INACTIVE_ANON);
 221         active = zone_page_state(zone, NR_ACTIVE_FILE) +
 222                                         zone_page_state(zone, NR_ACTIVE_ANON);
 223         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
 224                                         zone_page_state(zone, NR_ISOLATED_ANON);
 225
 226         return isolated > (inactive + active) / 2;
 227 }
 228
 229 /*
 230  * Isolate all pages that can be migrated from the block pointed to by
 231  * the migrate scanner within compact_control.
 232  */
 233 static unsigned long isolate_migratepages(struct zone *zone,
 234                                         struct compact_control *cc)
 235 {
 236         unsigned long low_pfn, end_pfn;
 237         struct list_head *migratelist = &cc->migratepages;
 238
 239         /* Do not scan outside zone boundaries */
 240         low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
 241
 242         /* Only scan within a pageblock boundary */
 243         end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
 244
 245         /* Do not cross the free scanner or scan within a memory hole */
 246         if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
 247                 cc->migrate_pfn = end_pfn;
 248                 return 0;
 249         }
 250
 251         /*
 252          * Ensure that there are not too many pages isolated from the LRU
 253          * list by either parallel reclaimers or compaction. If there are,
 254          * delay for some time until fewer pages are isolated
 255          */
 256         while (unlikely(too_many_isolated(zone))) {
 257                 congestion_wait(BLK_RW_ASYNC, HZ/10);
 258
 259                 if (fatal_signal_pending(current))
 260                         return 0;
 261         }
 262
 263         /* Time to isolate some pages for migration */
 264         spin_lock_irq(&zone->lru_lock);
 265         for (; low_pfn < end_pfn; low_pfn++) {
 266                 struct page *page;
 267                 if (!pfn_valid_within(low_pfn))
 268                         continue;
 269
 270                 /* Get the page and skip if free */
 271                 page = pfn_to_page(low_pfn);
 272                 if (PageBuddy(page))
 273                         continue;
 274
 275                 /* Try isolate the page */
 276                 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
 277                         continue;
 278
 279                 /* Successfully isolated */
 280                 del_page_from_lru_list(zone, page, page_lru(page));
 281                 list_add(&page->lru, migratelist);
 282                 mem_cgroup_del_lru(page);
 283                 cc->nr_migratepages++;
 284
 285                 /* Avoid isolating too much */
 286                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
 287                         break;
 288         }
 289
 290         acct_isolated(zone, cc);
 291
 292         spin_unlock_irq(&zone->lru_lock);
 293         cc->migrate_pfn = low_pfn;
 294
 295         return cc->nr_migratepages;
 296 }
 297
 298 /*
 299  * This is a migrate-callback that "allocates" freepages by taking pages
 300  * from the isolated freelists in the block we are migrating to.
 301  */
 302 static struct page *compaction_alloc(struct page *migratepage,
 303                                         unsigned long data,
 304                                         int **result)
 305 {
 306         struct compact_control *cc = (struct compact_control *)data;
 307         struct page *freepage;
 308
 309         /* Isolate free pages if necessary */
 310         if (list_empty(&cc->freepages)) {
 311                 isolate_freepages(cc->zone, cc);
 312
 313                 if (list_empty(&cc->freepages))
 314                         return NULL;
 315         }
 316
 317         freepage = list_entry(cc->freepages.next, struct page, lru);
 318         list_del(&freepage->lru);
 319         cc->nr_freepages--;
 320
 321         return freepage;
 322 }
 323
 324 /*
 325  * We cannot control nr_migratepages and nr_freepages fully when migration is
 326  * running as migrate_pages() has no knowledge of compact_control. When
 327  * migration is complete, we count the number of pages on the lists by hand.
 328  */
 329 static void update_nr_listpages(struct compact_control *cc)
 330 {
 331         int nr_migratepages = 0;
 332         int nr_freepages = 0;
 333         struct page *page;
 334
 335         list_for_each_entry(page, &cc->migratepages, lru)
 336                 nr_migratepages++;
 337         list_for_each_entry(page, &cc->freepages, lru)
 338                 nr_freepages++;
 339
 340         cc->nr_migratepages = nr_migratepages;
 341         cc->nr_freepages = nr_freepages;
 342 }
 343
 344 static int compact_finished(struct zone *zone,
 345                                                 struct compact_control *cc)
 346 {
 347         unsigned int order;
 348         unsigned long watermark = low_wmark_pages(zone) + (1 << cc->order);
 349
 350         if (fatal_signal_pending(current))
 351                 return COMPACT_PARTIAL;
 352
 353         /* Compaction run completes if the migrate and free scanner meet */
 354         if (cc->free_pfn <= cc->migrate_pfn)
 355                 return COMPACT_COMPLETE;
 356
 357         /* Compaction run is not finished if the watermark is not met */
 358         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
 359                 return COMPACT_CONTINUE;
 360
 361         if (cc->order == -1)
 362                 return COMPACT_CONTINUE;
 363
 364         /* Direct compactor: Is a suitable page free? */
 365         for (order = cc->order; order < MAX_ORDER; order++) {
 366                 /* Job done if page is free of the right migratetype */
 367                 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
 368                         return COMPACT_PARTIAL;
 369
 370                 /* Job done if allocation would set block type */
 371                 if (order >= pageblock_order && zone->free_area[order].nr_free)
 372                         return COMPACT_PARTIAL;
 373         }
 374
 375         return COMPACT_CONTINUE;
 376 }
 377
 378 static int compact_zone(struct zone *zone, struct compact_control *cc)
 379 {
 380         int ret;
 381
 382         /* Setup to move all movable pages to the end of the zone */
 383         cc->migrate_pfn = zone->zone_start_pfn;
 384         cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
 385         cc->free_pfn &= ~(pageblock_nr_pages-1);
 386
 387         migrate_prep_local();
 388
 389         while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
 390                 unsigned long nr_migrate, nr_remaining;
 391
 392                 if (!isolate_migratepages(zone, cc))
 393                         continue;
 394
 395                 nr_migrate = cc->nr_migratepages;
 396                 migrate_pages(&cc->migratepages, compaction_alloc,
 397                                                 (unsigned long)cc, 0);
 398                 update_nr_listpages(cc);
 399                 nr_remaining = cc->nr_migratepages;
 400
 401                 count_vm_event(COMPACTBLOCKS);
 402                 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
 403                 if (nr_remaining)
 404                         count_vm_events(COMPACTPAGEFAILED, nr_remaining);
 405
 406                 /* Release LRU pages not migrated */
 407                 if (!list_empty(&cc->migratepages)) {
 408                         putback_lru_pages(&cc->migratepages);
 409                         cc->nr_migratepages = 0;
 410                 }
 411
 412         }
 413
 414         /* Release free pages and check accounting */
 415         cc->nr_freepages -= release_freepages(&cc->freepages);
 416         VM_BUG_ON(cc->nr_freepages != 0);
 417
 418         return ret;
 419 }
 420
 421 static unsigned long compact_zone_order(struct zone *zone,
 422                                                 int order, gfp_t gfp_mask)
 423 {
 424         struct compact_control cc = {
 425                 .nr_freepages = 0,
 426                 .nr_migratepages = 0,
 427                 .order = order,
 428                 .migratetype = allocflags_to_migratetype(gfp_mask),
 429                 .zone = zone,
 430         };
 431         INIT_LIST_HEAD(&cc.freepages);
 432         INIT_LIST_HEAD(&cc.migratepages);
 433
 434         return compact_zone(zone, &cc);
 435 }
 436
 437 int sysctl_extfrag_threshold = 500;
 438
 439 /**
 440  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
 441  * @zonelist: The zonelist used for the current allocation
 442  * @order: The order of the current allocation
 443  * @gfp_mask: The GFP mask of the current allocation
 444  * @nodemask: The allowed nodes to allocate from
 445  *
 446  * This is the main entry point for direct page compaction.
 447  */
 448 unsigned long try_to_compact_pages(struct zonelist *zonelist,
 449                         int order, gfp_t gfp_mask, nodemask_t *nodemask)
 450 {
 451         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 452         int may_enter_fs = gfp_mask & __GFP_FS;
 453         int may_perform_io = gfp_mask & __GFP_IO;
 454         unsigned long watermark;
 455         struct zoneref *z;
 456         struct zone *zone;
 457         int rc = COMPACT_SKIPPED;
 458
 459         /*
 460          * Check whether it is worth even starting compaction. The order check is
 461          * made because an assumption is made that the page allocator can satisfy
 462          * the "cheaper" orders without taking special steps
 463          */
 464         if (order <= PAGE_ALLOC_COSTLY_ORDER || !may_enter_fs || !may_perform_io)
 465                 return rc;
 466
 467         count_vm_event(COMPACTSTALL);
 468
 469         /* Compact each zone in the list */
 470         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
 471                                                                 nodemask) {
 472                 int fragindex;
 473                 int status;
 474
 475                 /*
 476                  * Watermarks for order-0 must be met for compaction. Note
 477                  * the 2UL. This is because during migration, copies of
 478                  * pages need to be allocated and for a short time, the
 479                  * footprint is higher
 480                  */
 481                 watermark = low_wmark_pages(zone) + (2UL << order);
 482                 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
 483                         continue;
 484
 485                 /*
 486                  * fragmentation index determines if allocation failures are
 487                  * due to low memory or external fragmentation
 488                  *
 489                  * index of -1 implies allocations might succeed depending
 490                  *      on watermarks
 491                  * index towards 0 implies failure is due to lack of memory
 492                  * index towards 1000 implies failure is due to fragmentation
 493                  *
 494                  * Only compact if a failure would be due to fragmentation.
 495                  */
 496                 fragindex = fragmentation_index(zone, order);
 497                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
 498                         continue;
 499
 500                 if (fragindex == -1 && zone_watermark_ok(zone, order, watermark, 0, 0)) {
 501                         rc = COMPACT_PARTIAL;
 502                         break;
 503                 }
 504
 505                 status = compact_zone_order(zone, order, gfp_mask);
 506                 rc = max(status, rc);
 507
 508                 if (zone_watermark_ok(zone, order, watermark, 0, 0))
 509                         break;
 510         }
 511
 512         return rc;
 513 }
 514
 515
 516 /* Compact all zones within a node */
 517 static int compact_node(int nid)
 518 {
 519         int zoneid;
 520         pg_data_t *pgdat;
 521         struct zone *zone;
 522
 523         if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
 524                 return -EINVAL;
 525         pgdat = NODE_DATA(nid);
 526
 527         /* Flush pending updates to the LRU lists */
 528         lru_add_drain_all();
 529
 530         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 531                 struct compact_control cc = {
 532                         .nr_freepages = 0,
 533                         .nr_migratepages = 0,
 534                         .order = -1,
 535                 };
 536
 537                 zone = &pgdat->node_zones[zoneid];
 538                 if (!populated_zone(zone))
 539                         continue;
 540
 541                 cc.zone = zone;
 542                 INIT_LIST_HEAD(&cc.freepages);
 543                 INIT_LIST_HEAD(&cc.migratepages);
 544
 545                 compact_zone(zone, &cc);
 546
 547                 VM_BUG_ON(!list_empty(&cc.freepages));
 548                 VM_BUG_ON(!list_empty(&cc.migratepages));
 549         }
 550
 551         return 0;
 552 }
 553
 554 /* Compact all nodes in the system */
 555 static int compact_nodes(void)
 556 {
 557         int nid;
 558
 559         for_each_online_node(nid)
 560                 compact_node(nid);
 561
 562         return COMPACT_COMPLETE;
 563 }
 564
 565 /* The written value is actually unused, all memory is compacted */
 566 int sysctl_compact_memory;
 567
 568 /* This is the entry point for compacting all nodes via /proc/sys/vm */
 569 int sysctl_compaction_handler(struct ctl_table *table, int write,
 570                         void __user *buffer, size_t *length, loff_t *ppos)
 571 {
 572         if (write)
 573                 return compact_nodes();
 574
 575         return 0;
 576 }
 577
 578 int sysctl_extfrag_handler(struct ctl_table *table, int write,
 579                         void __user *buffer, size_t *length, loff_t *ppos)
 580 {
 581         proc_dointvec_minmax(table, write, buffer, length, ppos);
 582
 583         return 0;
 584 }
 585
 586 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
 587 ssize_t sysfs_compact_node(struct sys_device *dev,
 588                         struct sysdev_attribute *attr,
 589                         const char *buf, size_t count)
 590 {
 591         compact_node(dev->id);
 592
 593         return count;
 594 }
 595 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
 596
 597 int compaction_register_node(struct node *node)
 598 {
 599         return sysdev_create_file(&node->sysdev, &attr_compact);
 600 }
 601
 602 void compaction_unregister_node(struct node *node)
 603 {
 604         return sysdev_remove_file(&node->sysdev, &attr_compact);
 605 }
 606 #endif /* CONFIG_SYSFS && CONFIG_NUMA */