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 /* Account for isolated anon and file pages */ 39 unsigned long nr_anon; 40 unsigned long nr_file; 41 42 unsigned int order; /* order a direct compactor needs */ 43 int migratetype; /* MOVABLE, RECLAIMABLE etc */ 44 struct zone *zone; 45 46 int compact_mode; 47 }; 48 49 static unsigned long release_freepages(struct list_head *freelist) 50 { 51 struct page *page, *next; 52 unsigned long count = 0; 53 54 list_for_each_entry_safe(page, next, freelist, lru) { 55 list_del(&page->lru); 56 __free_page(page); 57 count++; 58 } 59 60 return count; 61 } 62 63 /* Isolate free pages onto a private freelist. Must hold zone->lock */ 64 static unsigned long isolate_freepages_block(struct zone *zone, 65 unsigned long blockpfn, 66 struct list_head *freelist) 67 { 68 unsigned long zone_end_pfn, end_pfn; 69 int nr_scanned = 0, total_isolated = 0; 70 struct page *cursor; 71 72 /* Get the last PFN we should scan for free pages at */ 73 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages; 74 end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn); 75 76 /* Find the first usable PFN in the block to initialse page cursor */ 77 for (; blockpfn < end_pfn; blockpfn++) { 78 if (pfn_valid_within(blockpfn)) 79 break; 80 } 81 cursor = pfn_to_page(blockpfn); 82 83 /* Isolate free pages. This assumes the block is valid */ 84 for (; blockpfn < end_pfn; blockpfn++, cursor++) { 85 int isolated, i; 86 struct page *page = cursor; 87 88 if (!pfn_valid_within(blockpfn)) 89 continue; 90 nr_scanned++; 91 92 if (!PageBuddy(page)) 93 continue; 94 95 /* Found a free page, break it into order-0 pages */ 96 isolated = split_free_page(page); 97 total_isolated += isolated; 98 for (i = 0; i < isolated; i++) { 99 list_add(&page->lru, freelist); 100 page++; 101 } 102 103 /* If a page was split, advance to the end of it */ 104 if (isolated) { 105 blockpfn += isolated - 1; 106 cursor += isolated - 1; 107 } 108 } 109 110 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated); 111 return total_isolated; 112 } 113 114 /* Returns true if the page is within a block suitable for migration to */ 115 static bool suitable_migration_target(struct page *page) 116 { 117 118 int migratetype = get_pageblock_migratetype(page); 119 120 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */ 121 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE) 122 return false; 123 124 /* If the page is a large free page, then allow migration */ 125 if (PageBuddy(page) && page_order(page) >= pageblock_order) 126 return true; 127 128 /* If the block is MIGRATE_MOVABLE, allow migration */ 129 if (migratetype == MIGRATE_MOVABLE) 130 return true; 131 132 /* Otherwise skip the block */ 133 return false; 134 } 135 136 /* 137 * Based on information in the current compact_control, find blocks 138 * suitable for isolating free pages from and then isolate them. 139 */ 140 static void isolate_freepages(struct zone *zone, 141 struct compact_control *cc) 142 { 143 struct page *page; 144 unsigned long high_pfn, low_pfn, pfn; 145 unsigned long flags; 146 int nr_freepages = cc->nr_freepages; 147 struct list_head *freelist = &cc->freepages; 148 149 pfn = cc->free_pfn; 150 low_pfn = cc->migrate_pfn + pageblock_nr_pages; 151 high_pfn = low_pfn; 152 153 /* 154 * Isolate free pages until enough are available to migrate the 155 * pages on cc->migratepages. We stop searching if the migrate 156 * and free page scanners meet or enough free pages are isolated. 157 */ 158 spin_lock_irqsave(&zone->lock, flags); 159 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages; 160 pfn -= pageblock_nr_pages) { 161 unsigned long isolated; 162 163 if (!pfn_valid(pfn)) 164 continue; 165 166 /* 167 * Check for overlapping nodes/zones. It's possible on some 168 * configurations to have a setup like 169 * node0 node1 node0 170 * i.e. it's possible that all pages within a zones range of 171 * pages do not belong to a single zone. 172 */ 173 page = pfn_to_page(pfn); 174 if (page_zone(page) != zone) 175 continue; 176 177 /* Check the block is suitable for migration */ 178 if (!suitable_migration_target(page)) 179 continue; 180 181 /* Found a block suitable for isolating free pages from */ 182 isolated = isolate_freepages_block(zone, pfn, freelist); 183 nr_freepages += isolated; 184 185 /* 186 * Record the highest PFN we isolated pages from. When next 187 * looking for free pages, the search will restart here as 188 * page migration may have returned some pages to the allocator 189 */ 190 if (isolated) 191 high_pfn = max(high_pfn, pfn); 192 } 193 spin_unlock_irqrestore(&zone->lock, flags); 194 195 /* split_free_page does not map the pages */ 196 list_for_each_entry(page, freelist, lru) { 197 arch_alloc_page(page, 0); 198 kernel_map_pages(page, 1, 1); 199 } 200 201 cc->free_pfn = high_pfn; 202 cc->nr_freepages = nr_freepages; 203 } 204 205 /* Update the number of anon and file isolated pages in the zone */ 206 static void acct_isolated(struct zone *zone, struct compact_control *cc) 207 { 208 struct page *page; 209 unsigned int count[NR_LRU_LISTS] = { 0, }; 210 211 list_for_each_entry(page, &cc->migratepages, lru) { 212 int lru = page_lru_base_type(page); 213 count[lru]++; 214 } 215 216 cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; 217 cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; 218 __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon); 219 __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file); 220 } 221 222 /* Similar to reclaim, but different enough that they don't share logic */ 223 static bool too_many_isolated(struct zone *zone) 224 { 225 unsigned long active, inactive, isolated; 226 227 inactive = zone_page_state(zone, NR_INACTIVE_FILE) + 228 zone_page_state(zone, NR_INACTIVE_ANON); 229 active = zone_page_state(zone, NR_ACTIVE_FILE) + 230 zone_page_state(zone, NR_ACTIVE_ANON); 231 isolated = zone_page_state(zone, NR_ISOLATED_FILE) + 232 zone_page_state(zone, NR_ISOLATED_ANON); 233 234 return isolated > (inactive + active) / 2; 235 } 236 237 /* 238 * Isolate all pages that can be migrated from the block pointed to by 239 * the migrate scanner within compact_control. 240 */ 241 static unsigned long isolate_migratepages(struct zone *zone, 242 struct compact_control *cc) 243 { 244 unsigned long low_pfn, end_pfn; 245 unsigned long last_pageblock_nr = 0, pageblock_nr; 246 unsigned long nr_scanned = 0, nr_isolated = 0; 247 struct list_head *migratelist = &cc->migratepages; 248 249 /* Do not scan outside zone boundaries */ 250 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn); 251 252 /* Only scan within a pageblock boundary */ 253 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages); 254 255 /* Do not cross the free scanner or scan within a memory hole */ 256 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) { 257 cc->migrate_pfn = end_pfn; 258 return 0; 259 } 260 261 /* 262 * Ensure that there are not too many pages isolated from the LRU 263 * list by either parallel reclaimers or compaction. If there are, 264 * delay for some time until fewer pages are isolated 265 */ 266 while (unlikely(too_many_isolated(zone))) { 267 congestion_wait(BLK_RW_ASYNC, HZ/10); 268 269 if (fatal_signal_pending(current)) 270 return 0; 271 } 272 273 /* Time to isolate some pages for migration */ 274 spin_lock_irq(&zone->lru_lock); 275 for (; low_pfn < end_pfn; low_pfn++) { 276 struct page *page; 277 if (!pfn_valid_within(low_pfn)) 278 continue; 279 nr_scanned++; 280 281 /* Get the page and skip if free */ 282 page = pfn_to_page(low_pfn); 283 if (PageBuddy(page)) 284 continue; 285 286 /* 287 * For async migration, also only scan in MOVABLE blocks. Async 288 * migration is optimistic to see if the minimum amount of work 289 * satisfies the allocation 290 */ 291 pageblock_nr = low_pfn >> pageblock_order; 292 if (!cc->sync && last_pageblock_nr != pageblock_nr && 293 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) { 294 low_pfn += pageblock_nr_pages; 295 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1; 296 last_pageblock_nr = pageblock_nr; 297 continue; 298 } 299 300 if (!PageLRU(page)) 301 continue; 302 303 /* 304 * PageLRU is set, and lru_lock excludes isolation, 305 * splitting and collapsing (collapsing has already 306 * happened if PageLRU is set). 307 */ 308 if (PageTransHuge(page)) { 309 low_pfn += (1 << compound_order(page)) - 1; 310 continue; 311 } 312 313 /* Try isolate the page */ 314 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0) 315 continue; 316 317 VM_BUG_ON(PageTransCompound(page)); 318 319 /* Successfully isolated */ 320 del_page_from_lru_list(zone, page, page_lru(page)); 321 list_add(&page->lru, migratelist); 322 cc->nr_migratepages++; 323 nr_isolated++; 324 325 /* Avoid isolating too much */ 326 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) 327 break; 328 } 329 330 acct_isolated(zone, cc); 331 332 spin_unlock_irq(&zone->lru_lock); 333 cc->migrate_pfn = low_pfn; 334 335 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated); 336 337 return cc->nr_migratepages; 338 } 339 340 /* 341 * This is a migrate-callback that "allocates" freepages by taking pages 342 * from the isolated freelists in the block we are migrating to. 343 */ 344 static struct page *compaction_alloc(struct page *migratepage, 345 unsigned long data, 346 int **result) 347 { 348 struct compact_control *cc = (struct compact_control *)data; 349 struct page *freepage; 350 351 /* Isolate free pages if necessary */ 352 if (list_empty(&cc->freepages)) { 353 isolate_freepages(cc->zone, cc); 354 355 if (list_empty(&cc->freepages)) 356 return NULL; 357 } 358 359 freepage = list_entry(cc->freepages.next, struct page, lru); 360 list_del(&freepage->lru); 361 cc->nr_freepages--; 362 363 return freepage; 364 } 365 366 /* 367 * We cannot control nr_migratepages and nr_freepages fully when migration is 368 * running as migrate_pages() has no knowledge of compact_control. When 369 * migration is complete, we count the number of pages on the lists by hand. 370 */ 371 static void update_nr_listpages(struct compact_control *cc) 372 { 373 int nr_migratepages = 0; 374 int nr_freepages = 0; 375 struct page *page; 376 377 list_for_each_entry(page, &cc->migratepages, lru) 378 nr_migratepages++; 379 list_for_each_entry(page, &cc->freepages, lru) 380 nr_freepages++; 381 382 cc->nr_migratepages = nr_migratepages; 383 cc->nr_freepages = nr_freepages; 384 } 385 386 static int compact_finished(struct zone *zone, 387 struct compact_control *cc) 388 { 389 unsigned int order; 390 unsigned long watermark; 391 392 if (fatal_signal_pending(current)) 393 return COMPACT_PARTIAL; 394 395 /* Compaction run completes if the migrate and free scanner meet */ 396 if (cc->free_pfn <= cc->migrate_pfn) 397 return COMPACT_COMPLETE; 398 399 /* Compaction run is not finished if the watermark is not met */ 400 if (cc->compact_mode != COMPACT_MODE_KSWAPD) 401 watermark = low_wmark_pages(zone); 402 else 403 watermark = high_wmark_pages(zone); 404 watermark += (1 << cc->order); 405 406 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) 407 return COMPACT_CONTINUE; 408 409 if (cc->order == -1) 410 return COMPACT_CONTINUE; 411 412 /* 413 * Generating only one page of the right order is not enough 414 * for kswapd, we must continue until we're above the high 415 * watermark as a pool for high order GFP_ATOMIC allocations 416 * too. 417 */ 418 if (cc->compact_mode == COMPACT_MODE_KSWAPD) 419 return COMPACT_CONTINUE; 420 421 /* Direct compactor: Is a suitable page free? */ 422 for (order = cc->order; order < MAX_ORDER; order++) { 423 /* Job done if page is free of the right migratetype */ 424 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype])) 425 return COMPACT_PARTIAL; 426 427 /* Job done if allocation would set block type */ 428 if (order >= pageblock_order && zone->free_area[order].nr_free) 429 return COMPACT_PARTIAL; 430 } 431 432 return COMPACT_CONTINUE; 433 } 434 435 /* 436 * compaction_suitable: Is this suitable to run compaction on this zone now? 437 * Returns 438 * COMPACT_SKIPPED - If there are too few free pages for compaction 439 * COMPACT_PARTIAL - If the allocation would succeed without compaction 440 * COMPACT_CONTINUE - If compaction should run now 441 */ 442 unsigned long compaction_suitable(struct zone *zone, int order) 443 { 444 int fragindex; 445 unsigned long watermark; 446 447 /* 448 * Watermarks for order-0 must be met for compaction. Note the 2UL. 449 * This is because during migration, copies of pages need to be 450 * allocated and for a short time, the footprint is higher 451 */ 452 watermark = low_wmark_pages(zone) + (2UL << order); 453 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 454 return COMPACT_SKIPPED; 455 456 /* 457 * fragmentation index determines if allocation failures are due to 458 * low memory or external fragmentation 459 * 460 * index of -1 implies allocations might succeed dependingon watermarks 461 * index towards 0 implies failure is due to lack of memory 462 * index towards 1000 implies failure is due to fragmentation 463 * 464 * Only compact if a failure would be due to fragmentation. 465 */ 466 fragindex = fragmentation_index(zone, order); 467 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold) 468 return COMPACT_SKIPPED; 469 470 if (fragindex == -1 && zone_watermark_ok(zone, order, watermark, 0, 0)) 471 return COMPACT_PARTIAL; 472 473 return COMPACT_CONTINUE; 474 } 475 476 static int compact_zone(struct zone *zone, struct compact_control *cc) 477 { 478 int ret; 479 480 ret = compaction_suitable(zone, cc->order); 481 switch (ret) { 482 case COMPACT_PARTIAL: 483 case COMPACT_SKIPPED: 484 /* Compaction is likely to fail */ 485 return ret; 486 case COMPACT_CONTINUE: 487 /* Fall through to compaction */ 488 ; 489 } 490 491 /* Setup to move all movable pages to the end of the zone */ 492 cc->migrate_pfn = zone->zone_start_pfn; 493 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages; 494 cc->free_pfn &= ~(pageblock_nr_pages-1); 495 496 migrate_prep_local(); 497 498 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) { 499 unsigned long nr_migrate, nr_remaining; 500 501 if (!isolate_migratepages(zone, cc)) 502 continue; 503 504 nr_migrate = cc->nr_migratepages; 505 migrate_pages(&cc->migratepages, compaction_alloc, 506 (unsigned long)cc, false, 507 cc->sync); 508 update_nr_listpages(cc); 509 nr_remaining = cc->nr_migratepages; 510 511 count_vm_event(COMPACTBLOCKS); 512 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining); 513 if (nr_remaining) 514 count_vm_events(COMPACTPAGEFAILED, nr_remaining); 515 trace_mm_compaction_migratepages(nr_migrate - nr_remaining, 516 nr_remaining); 517 518 /* Release LRU pages not migrated */ 519 if (!list_empty(&cc->migratepages)) { 520 putback_lru_pages(&cc->migratepages); 521 cc->nr_migratepages = 0; 522 } 523 524 } 525 526 /* Release free pages and check accounting */ 527 cc->nr_freepages -= release_freepages(&cc->freepages); 528 VM_BUG_ON(cc->nr_freepages != 0); 529 530 return ret; 531 } 532 533 unsigned long compact_zone_order(struct zone *zone, 534 int order, gfp_t gfp_mask, 535 bool sync, 536 int compact_mode) 537 { 538 struct compact_control cc = { 539 .nr_freepages = 0, 540 .nr_migratepages = 0, 541 .order = order, 542 .migratetype = allocflags_to_migratetype(gfp_mask), 543 .zone = zone, 544 .sync = sync, 545 .compact_mode = compact_mode, 546 }; 547 INIT_LIST_HEAD(&cc.freepages); 548 INIT_LIST_HEAD(&cc.migratepages); 549 550 return compact_zone(zone, &cc); 551 } 552 553 int sysctl_extfrag_threshold = 500; 554 555 /** 556 * try_to_compact_pages - Direct compact to satisfy a high-order allocation 557 * @zonelist: The zonelist used for the current allocation 558 * @order: The order of the current allocation 559 * @gfp_mask: The GFP mask of the current allocation 560 * @nodemask: The allowed nodes to allocate from 561 * @sync: Whether migration is synchronous or not 562 * 563 * This is the main entry point for direct page compaction. 564 */ 565 unsigned long try_to_compact_pages(struct zonelist *zonelist, 566 int order, gfp_t gfp_mask, nodemask_t *nodemask, 567 bool sync) 568 { 569 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 570 int may_enter_fs = gfp_mask & __GFP_FS; 571 int may_perform_io = gfp_mask & __GFP_IO; 572 struct zoneref *z; 573 struct zone *zone; 574 int rc = COMPACT_SKIPPED; 575 576 /* 577 * Check whether it is worth even starting compaction. The order check is 578 * made because an assumption is made that the page allocator can satisfy 579 * the "cheaper" orders without taking special steps 580 */ 581 if (!order || !may_enter_fs || !may_perform_io) 582 return rc; 583 584 count_vm_event(COMPACTSTALL); 585 586 /* Compact each zone in the list */ 587 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, 588 nodemask) { 589 int status; 590 591 status = compact_zone_order(zone, order, gfp_mask, sync, 592 COMPACT_MODE_DIRECT_RECLAIM); 593 rc = max(status, rc); 594 595 /* If a normal allocation would succeed, stop compacting */ 596 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) 597 break; 598 } 599 600 return rc; 601 } 602 603 604 /* Compact all zones within a node */ 605 static int compact_node(int nid) 606 { 607 int zoneid; 608 pg_data_t *pgdat; 609 struct zone *zone; 610 611 if (nid < 0 || nid >= nr_node_ids || !node_online(nid)) 612 return -EINVAL; 613 pgdat = NODE_DATA(nid); 614 615 /* Flush pending updates to the LRU lists */ 616 lru_add_drain_all(); 617 618 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 619 struct compact_control cc = { 620 .nr_freepages = 0, 621 .nr_migratepages = 0, 622 .order = -1, 623 .compact_mode = COMPACT_MODE_DIRECT_RECLAIM, 624 }; 625 626 zone = &pgdat->node_zones[zoneid]; 627 if (!populated_zone(zone)) 628 continue; 629 630 cc.zone = zone; 631 INIT_LIST_HEAD(&cc.freepages); 632 INIT_LIST_HEAD(&cc.migratepages); 633 634 compact_zone(zone, &cc); 635 636 VM_BUG_ON(!list_empty(&cc.freepages)); 637 VM_BUG_ON(!list_empty(&cc.migratepages)); 638 } 639 640 return 0; 641 } 642 643 /* Compact all nodes in the system */ 644 static int compact_nodes(void) 645 { 646 int nid; 647 648 for_each_online_node(nid) 649 compact_node(nid); 650 651 return COMPACT_COMPLETE; 652 } 653 654 /* The written value is actually unused, all memory is compacted */ 655 int sysctl_compact_memory; 656 657 /* This is the entry point for compacting all nodes via /proc/sys/vm */ 658 int sysctl_compaction_handler(struct ctl_table *table, int write, 659 void __user *buffer, size_t *length, loff_t *ppos) 660 { 661 if (write) 662 return compact_nodes(); 663 664 return 0; 665 } 666 667 int sysctl_extfrag_handler(struct ctl_table *table, int write, 668 void __user *buffer, size_t *length, loff_t *ppos) 669 { 670 proc_dointvec_minmax(table, write, buffer, length, ppos); 671 672 return 0; 673 } 674 675 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) 676 ssize_t sysfs_compact_node(struct sys_device *dev, 677 struct sysdev_attribute *attr, 678 const char *buf, size_t count) 679 { 680 compact_node(dev->id); 681 682 return count; 683 } 684 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node); 685 686 int compaction_register_node(struct node *node) 687 { 688 return sysdev_create_file(&node->sysdev, &attr_compact); 689 } 690 691 void compaction_unregister_node(struct node *node) 692 { 693 return sysdev_remove_file(&node->sysdev, &attr_compact); 694 } 695 #endif /* CONFIG_SYSFS && CONFIG_NUMA */ 696