1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2748446bbSMel Gorman /* 3748446bbSMel Gorman * linux/mm/compaction.c 4748446bbSMel Gorman * 5748446bbSMel Gorman * Memory compaction for the reduction of external fragmentation. Note that 6748446bbSMel Gorman * this heavily depends upon page migration to do all the real heavy 7748446bbSMel Gorman * lifting 8748446bbSMel Gorman * 9748446bbSMel Gorman * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie> 10748446bbSMel Gorman */ 11698b1b30SVlastimil Babka #include <linux/cpu.h> 12748446bbSMel Gorman #include <linux/swap.h> 13748446bbSMel Gorman #include <linux/migrate.h> 14748446bbSMel Gorman #include <linux/compaction.h> 15748446bbSMel Gorman #include <linux/mm_inline.h> 16174cd4b1SIngo Molnar #include <linux/sched/signal.h> 17748446bbSMel Gorman #include <linux/backing-dev.h> 1876ab0f53SMel Gorman #include <linux/sysctl.h> 19ed4a6d7fSMel Gorman #include <linux/sysfs.h> 20194159fbSMinchan Kim #include <linux/page-isolation.h> 21b8c73fc2SAndrey Ryabinin #include <linux/kasan.h> 22698b1b30SVlastimil Babka #include <linux/kthread.h> 23698b1b30SVlastimil Babka #include <linux/freezer.h> 2483358eceSJoonsoo Kim #include <linux/page_owner.h> 25eb414681SJohannes Weiner #include <linux/psi.h> 26748446bbSMel Gorman #include "internal.h" 27748446bbSMel Gorman 28010fc29aSMinchan Kim #ifdef CONFIG_COMPACTION 2931ca72faSCharan Teja Kalla /* 3031ca72faSCharan Teja Kalla * Fragmentation score check interval for proactive compaction purposes. 3131ca72faSCharan Teja Kalla */ 3231ca72faSCharan Teja Kalla #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500) 3331ca72faSCharan Teja Kalla 34010fc29aSMinchan Kim static inline void count_compact_event(enum vm_event_item item) 35010fc29aSMinchan Kim { 36010fc29aSMinchan Kim count_vm_event(item); 37010fc29aSMinchan Kim } 38010fc29aSMinchan Kim 39010fc29aSMinchan Kim static inline void count_compact_events(enum vm_event_item item, long delta) 40010fc29aSMinchan Kim { 41010fc29aSMinchan Kim count_vm_events(item, delta); 42010fc29aSMinchan Kim } 43010fc29aSMinchan Kim #else 44010fc29aSMinchan Kim #define count_compact_event(item) do { } while (0) 45010fc29aSMinchan Kim #define count_compact_events(item, delta) do { } while (0) 46010fc29aSMinchan Kim #endif 47010fc29aSMinchan Kim 48ff9543fdSMichal Nazarewicz #if defined CONFIG_COMPACTION || defined CONFIG_CMA 49ff9543fdSMichal Nazarewicz 50b7aba698SMel Gorman #define CREATE_TRACE_POINTS 51b7aba698SMel Gorman #include <trace/events/compaction.h> 52b7aba698SMel Gorman 5306b6640aSVlastimil Babka #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order)) 5406b6640aSVlastimil Babka #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order)) 5506b6640aSVlastimil Babka 56facdaa91SNitin Gupta /* 57facdaa91SNitin Gupta * Page order with-respect-to which proactive compaction 58facdaa91SNitin Gupta * calculates external fragmentation, which is used as 59facdaa91SNitin Gupta * the "fragmentation score" of a node/zone. 60facdaa91SNitin Gupta */ 61facdaa91SNitin Gupta #if defined CONFIG_TRANSPARENT_HUGEPAGE 62facdaa91SNitin Gupta #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER 6325788738SNitin Gupta #elif defined CONFIG_HUGETLBFS 64facdaa91SNitin Gupta #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER 65facdaa91SNitin Gupta #else 66facdaa91SNitin Gupta #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT) 67facdaa91SNitin Gupta #endif 68facdaa91SNitin Gupta 69748446bbSMel Gorman static unsigned long release_freepages(struct list_head *freelist) 70748446bbSMel Gorman { 71748446bbSMel Gorman struct page *page, *next; 726bace090SVlastimil Babka unsigned long high_pfn = 0; 73748446bbSMel Gorman 74748446bbSMel Gorman list_for_each_entry_safe(page, next, freelist, lru) { 756bace090SVlastimil Babka unsigned long pfn = page_to_pfn(page); 76748446bbSMel Gorman list_del(&page->lru); 77748446bbSMel Gorman __free_page(page); 786bace090SVlastimil Babka if (pfn > high_pfn) 796bace090SVlastimil Babka high_pfn = pfn; 80748446bbSMel Gorman } 81748446bbSMel Gorman 826bace090SVlastimil Babka return high_pfn; 83748446bbSMel Gorman } 84748446bbSMel Gorman 854469ab98SMel Gorman static void split_map_pages(struct list_head *list) 86ff9543fdSMichal Nazarewicz { 8766c64223SJoonsoo Kim unsigned int i, order, nr_pages; 8866c64223SJoonsoo Kim struct page *page, *next; 8966c64223SJoonsoo Kim LIST_HEAD(tmp_list); 90ff9543fdSMichal Nazarewicz 9166c64223SJoonsoo Kim list_for_each_entry_safe(page, next, list, lru) { 9266c64223SJoonsoo Kim list_del(&page->lru); 9366c64223SJoonsoo Kim 9466c64223SJoonsoo Kim order = page_private(page); 9566c64223SJoonsoo Kim nr_pages = 1 << order; 9666c64223SJoonsoo Kim 9746f24fd8SJoonsoo Kim post_alloc_hook(page, order, __GFP_MOVABLE); 9866c64223SJoonsoo Kim if (order) 9966c64223SJoonsoo Kim split_page(page, order); 10066c64223SJoonsoo Kim 10166c64223SJoonsoo Kim for (i = 0; i < nr_pages; i++) { 10266c64223SJoonsoo Kim list_add(&page->lru, &tmp_list); 10366c64223SJoonsoo Kim page++; 104ff9543fdSMichal Nazarewicz } 105ff9543fdSMichal Nazarewicz } 106ff9543fdSMichal Nazarewicz 10766c64223SJoonsoo Kim list_splice(&tmp_list, list); 10866c64223SJoonsoo Kim } 10966c64223SJoonsoo Kim 110bb13ffebSMel Gorman #ifdef CONFIG_COMPACTION 11168f2736aSMatthew Wilcox (Oracle) bool PageMovable(struct page *page) 112bda807d4SMinchan Kim { 11368f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops; 114bda807d4SMinchan Kim 115bda807d4SMinchan Kim VM_BUG_ON_PAGE(!PageLocked(page), page); 116bda807d4SMinchan Kim if (!__PageMovable(page)) 11768f2736aSMatthew Wilcox (Oracle) return false; 118bda807d4SMinchan Kim 11968f2736aSMatthew Wilcox (Oracle) mops = page_movable_ops(page); 12068f2736aSMatthew Wilcox (Oracle) if (mops) 12168f2736aSMatthew Wilcox (Oracle) return true; 122bda807d4SMinchan Kim 12368f2736aSMatthew Wilcox (Oracle) return false; 124bda807d4SMinchan Kim } 125bda807d4SMinchan Kim 12668f2736aSMatthew Wilcox (Oracle) void __SetPageMovable(struct page *page, const struct movable_operations *mops) 127bda807d4SMinchan Kim { 128bda807d4SMinchan Kim VM_BUG_ON_PAGE(!PageLocked(page), page); 12968f2736aSMatthew Wilcox (Oracle) VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page); 13068f2736aSMatthew Wilcox (Oracle) page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE); 131bda807d4SMinchan Kim } 132bda807d4SMinchan Kim EXPORT_SYMBOL(__SetPageMovable); 133bda807d4SMinchan Kim 134bda807d4SMinchan Kim void __ClearPageMovable(struct page *page) 135bda807d4SMinchan Kim { 136bda807d4SMinchan Kim VM_BUG_ON_PAGE(!PageMovable(page), page); 137bda807d4SMinchan Kim /* 13868f2736aSMatthew Wilcox (Oracle) * This page still has the type of a movable page, but it's 13968f2736aSMatthew Wilcox (Oracle) * actually not movable any more. 140bda807d4SMinchan Kim */ 14168f2736aSMatthew Wilcox (Oracle) page->mapping = (void *)PAGE_MAPPING_MOVABLE; 142bda807d4SMinchan Kim } 143bda807d4SMinchan Kim EXPORT_SYMBOL(__ClearPageMovable); 144bda807d4SMinchan Kim 14524e2716fSJoonsoo Kim /* Do not skip compaction more than 64 times */ 14624e2716fSJoonsoo Kim #define COMPACT_MAX_DEFER_SHIFT 6 14724e2716fSJoonsoo Kim 14824e2716fSJoonsoo Kim /* 14924e2716fSJoonsoo Kim * Compaction is deferred when compaction fails to result in a page 150860b3272SAlex Shi * allocation success. 1 << compact_defer_shift, compactions are skipped up 15124e2716fSJoonsoo Kim * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT 15224e2716fSJoonsoo Kim */ 1532271b016SHui Su static void defer_compaction(struct zone *zone, int order) 15424e2716fSJoonsoo Kim { 15524e2716fSJoonsoo Kim zone->compact_considered = 0; 15624e2716fSJoonsoo Kim zone->compact_defer_shift++; 15724e2716fSJoonsoo Kim 15824e2716fSJoonsoo Kim if (order < zone->compact_order_failed) 15924e2716fSJoonsoo Kim zone->compact_order_failed = order; 16024e2716fSJoonsoo Kim 16124e2716fSJoonsoo Kim if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT) 16224e2716fSJoonsoo Kim zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT; 16324e2716fSJoonsoo Kim 16424e2716fSJoonsoo Kim trace_mm_compaction_defer_compaction(zone, order); 16524e2716fSJoonsoo Kim } 16624e2716fSJoonsoo Kim 16724e2716fSJoonsoo Kim /* Returns true if compaction should be skipped this time */ 1682271b016SHui Su static bool compaction_deferred(struct zone *zone, int order) 16924e2716fSJoonsoo Kim { 17024e2716fSJoonsoo Kim unsigned long defer_limit = 1UL << zone->compact_defer_shift; 17124e2716fSJoonsoo Kim 17224e2716fSJoonsoo Kim if (order < zone->compact_order_failed) 17324e2716fSJoonsoo Kim return false; 17424e2716fSJoonsoo Kim 17524e2716fSJoonsoo Kim /* Avoid possible overflow */ 17662b35fe0SMateusz Nosek if (++zone->compact_considered >= defer_limit) { 17724e2716fSJoonsoo Kim zone->compact_considered = defer_limit; 17824e2716fSJoonsoo Kim return false; 17962b35fe0SMateusz Nosek } 18024e2716fSJoonsoo Kim 18124e2716fSJoonsoo Kim trace_mm_compaction_deferred(zone, order); 18224e2716fSJoonsoo Kim 18324e2716fSJoonsoo Kim return true; 18424e2716fSJoonsoo Kim } 18524e2716fSJoonsoo Kim 18624e2716fSJoonsoo Kim /* 18724e2716fSJoonsoo Kim * Update defer tracking counters after successful compaction of given order, 18824e2716fSJoonsoo Kim * which means an allocation either succeeded (alloc_success == true) or is 18924e2716fSJoonsoo Kim * expected to succeed. 19024e2716fSJoonsoo Kim */ 19124e2716fSJoonsoo Kim void compaction_defer_reset(struct zone *zone, int order, 19224e2716fSJoonsoo Kim bool alloc_success) 19324e2716fSJoonsoo Kim { 19424e2716fSJoonsoo Kim if (alloc_success) { 19524e2716fSJoonsoo Kim zone->compact_considered = 0; 19624e2716fSJoonsoo Kim zone->compact_defer_shift = 0; 19724e2716fSJoonsoo Kim } 19824e2716fSJoonsoo Kim if (order >= zone->compact_order_failed) 19924e2716fSJoonsoo Kim zone->compact_order_failed = order + 1; 20024e2716fSJoonsoo Kim 20124e2716fSJoonsoo Kim trace_mm_compaction_defer_reset(zone, order); 20224e2716fSJoonsoo Kim } 20324e2716fSJoonsoo Kim 20424e2716fSJoonsoo Kim /* Returns true if restarting compaction after many failures */ 2052271b016SHui Su static bool compaction_restarting(struct zone *zone, int order) 20624e2716fSJoonsoo Kim { 20724e2716fSJoonsoo Kim if (order < zone->compact_order_failed) 20824e2716fSJoonsoo Kim return false; 20924e2716fSJoonsoo Kim 21024e2716fSJoonsoo Kim return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT && 21124e2716fSJoonsoo Kim zone->compact_considered >= 1UL << zone->compact_defer_shift; 21224e2716fSJoonsoo Kim } 21324e2716fSJoonsoo Kim 214bb13ffebSMel Gorman /* Returns true if the pageblock should be scanned for pages to isolate. */ 215bb13ffebSMel Gorman static inline bool isolation_suitable(struct compact_control *cc, 216bb13ffebSMel Gorman struct page *page) 217bb13ffebSMel Gorman { 218bb13ffebSMel Gorman if (cc->ignore_skip_hint) 219bb13ffebSMel Gorman return true; 220bb13ffebSMel Gorman 221bb13ffebSMel Gorman return !get_pageblock_skip(page); 222bb13ffebSMel Gorman } 223bb13ffebSMel Gorman 22402333641SVlastimil Babka static void reset_cached_positions(struct zone *zone) 22502333641SVlastimil Babka { 22602333641SVlastimil Babka zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn; 22702333641SVlastimil Babka zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn; 228623446e4SJoonsoo Kim zone->compact_cached_free_pfn = 22906b6640aSVlastimil Babka pageblock_start_pfn(zone_end_pfn(zone) - 1); 23002333641SVlastimil Babka } 23102333641SVlastimil Babka 232bb13ffebSMel Gorman /* 2332271b016SHui Su * Compound pages of >= pageblock_order should consistently be skipped until 234b527cfe5SVlastimil Babka * released. It is always pointless to compact pages of such order (if they are 235b527cfe5SVlastimil Babka * migratable), and the pageblocks they occupy cannot contain any free pages. 23621dc7e02SDavid Rientjes */ 237b527cfe5SVlastimil Babka static bool pageblock_skip_persistent(struct page *page) 23821dc7e02SDavid Rientjes { 239b527cfe5SVlastimil Babka if (!PageCompound(page)) 24021dc7e02SDavid Rientjes return false; 241b527cfe5SVlastimil Babka 242b527cfe5SVlastimil Babka page = compound_head(page); 243b527cfe5SVlastimil Babka 244b527cfe5SVlastimil Babka if (compound_order(page) >= pageblock_order) 24521dc7e02SDavid Rientjes return true; 246b527cfe5SVlastimil Babka 247b527cfe5SVlastimil Babka return false; 24821dc7e02SDavid Rientjes } 24921dc7e02SDavid Rientjes 250e332f741SMel Gorman static bool 251e332f741SMel Gorman __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source, 252e332f741SMel Gorman bool check_target) 253e332f741SMel Gorman { 254e332f741SMel Gorman struct page *page = pfn_to_online_page(pfn); 2556b0868c8SMel Gorman struct page *block_page; 256e332f741SMel Gorman struct page *end_page; 257e332f741SMel Gorman unsigned long block_pfn; 258e332f741SMel Gorman 259e332f741SMel Gorman if (!page) 260e332f741SMel Gorman return false; 261e332f741SMel Gorman if (zone != page_zone(page)) 262e332f741SMel Gorman return false; 263e332f741SMel Gorman if (pageblock_skip_persistent(page)) 264e332f741SMel Gorman return false; 265e332f741SMel Gorman 266e332f741SMel Gorman /* 267e332f741SMel Gorman * If skip is already cleared do no further checking once the 268e332f741SMel Gorman * restart points have been set. 269e332f741SMel Gorman */ 270e332f741SMel Gorman if (check_source && check_target && !get_pageblock_skip(page)) 271e332f741SMel Gorman return true; 272e332f741SMel Gorman 273e332f741SMel Gorman /* 274e332f741SMel Gorman * If clearing skip for the target scanner, do not select a 275e332f741SMel Gorman * non-movable pageblock as the starting point. 276e332f741SMel Gorman */ 277e332f741SMel Gorman if (!check_source && check_target && 278e332f741SMel Gorman get_pageblock_migratetype(page) != MIGRATE_MOVABLE) 279e332f741SMel Gorman return false; 280e332f741SMel Gorman 2816b0868c8SMel Gorman /* Ensure the start of the pageblock or zone is online and valid */ 2826b0868c8SMel Gorman block_pfn = pageblock_start_pfn(pfn); 283a2e9a5afSVlastimil Babka block_pfn = max(block_pfn, zone->zone_start_pfn); 284a2e9a5afSVlastimil Babka block_page = pfn_to_online_page(block_pfn); 2856b0868c8SMel Gorman if (block_page) { 2866b0868c8SMel Gorman page = block_page; 2876b0868c8SMel Gorman pfn = block_pfn; 2886b0868c8SMel Gorman } 2896b0868c8SMel Gorman 2906b0868c8SMel Gorman /* Ensure the end of the pageblock or zone is online and valid */ 291a2e9a5afSVlastimil Babka block_pfn = pageblock_end_pfn(pfn) - 1; 2926b0868c8SMel Gorman block_pfn = min(block_pfn, zone_end_pfn(zone) - 1); 2936b0868c8SMel Gorman end_page = pfn_to_online_page(block_pfn); 2946b0868c8SMel Gorman if (!end_page) 2956b0868c8SMel Gorman return false; 2966b0868c8SMel Gorman 297e332f741SMel Gorman /* 298e332f741SMel Gorman * Only clear the hint if a sample indicates there is either a 299e332f741SMel Gorman * free page or an LRU page in the block. One or other condition 300e332f741SMel Gorman * is necessary for the block to be a migration source/target. 301e332f741SMel Gorman */ 302e332f741SMel Gorman do { 303e332f741SMel Gorman if (check_source && PageLRU(page)) { 304e332f741SMel Gorman clear_pageblock_skip(page); 305e332f741SMel Gorman return true; 306e332f741SMel Gorman } 307e332f741SMel Gorman 308e332f741SMel Gorman if (check_target && PageBuddy(page)) { 309e332f741SMel Gorman clear_pageblock_skip(page); 310e332f741SMel Gorman return true; 311e332f741SMel Gorman } 312e332f741SMel Gorman 313e332f741SMel Gorman page += (1 << PAGE_ALLOC_COSTLY_ORDER); 314a2e9a5afSVlastimil Babka } while (page <= end_page); 315e332f741SMel Gorman 316e332f741SMel Gorman return false; 317e332f741SMel Gorman } 318e332f741SMel Gorman 31921dc7e02SDavid Rientjes /* 320bb13ffebSMel Gorman * This function is called to clear all cached information on pageblocks that 321bb13ffebSMel Gorman * should be skipped for page isolation when the migrate and free page scanner 322bb13ffebSMel Gorman * meet. 323bb13ffebSMel Gorman */ 32462997027SMel Gorman static void __reset_isolation_suitable(struct zone *zone) 325bb13ffebSMel Gorman { 326e332f741SMel Gorman unsigned long migrate_pfn = zone->zone_start_pfn; 3276b0868c8SMel Gorman unsigned long free_pfn = zone_end_pfn(zone) - 1; 328e332f741SMel Gorman unsigned long reset_migrate = free_pfn; 329e332f741SMel Gorman unsigned long reset_free = migrate_pfn; 330e332f741SMel Gorman bool source_set = false; 331e332f741SMel Gorman bool free_set = false; 332e332f741SMel Gorman 333e332f741SMel Gorman if (!zone->compact_blockskip_flush) 334e332f741SMel Gorman return; 335bb13ffebSMel Gorman 33662997027SMel Gorman zone->compact_blockskip_flush = false; 337bb13ffebSMel Gorman 338e332f741SMel Gorman /* 339e332f741SMel Gorman * Walk the zone and update pageblock skip information. Source looks 340e332f741SMel Gorman * for PageLRU while target looks for PageBuddy. When the scanner 341e332f741SMel Gorman * is found, both PageBuddy and PageLRU are checked as the pageblock 342e332f741SMel Gorman * is suitable as both source and target. 343e332f741SMel Gorman */ 344e332f741SMel Gorman for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages, 345e332f741SMel Gorman free_pfn -= pageblock_nr_pages) { 346bb13ffebSMel Gorman cond_resched(); 347bb13ffebSMel Gorman 348e332f741SMel Gorman /* Update the migrate PFN */ 349e332f741SMel Gorman if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) && 350e332f741SMel Gorman migrate_pfn < reset_migrate) { 351e332f741SMel Gorman source_set = true; 352e332f741SMel Gorman reset_migrate = migrate_pfn; 353e332f741SMel Gorman zone->compact_init_migrate_pfn = reset_migrate; 354e332f741SMel Gorman zone->compact_cached_migrate_pfn[0] = reset_migrate; 355e332f741SMel Gorman zone->compact_cached_migrate_pfn[1] = reset_migrate; 356bb13ffebSMel Gorman } 35702333641SVlastimil Babka 358e332f741SMel Gorman /* Update the free PFN */ 359e332f741SMel Gorman if (__reset_isolation_pfn(zone, free_pfn, free_set, true) && 360e332f741SMel Gorman free_pfn > reset_free) { 361e332f741SMel Gorman free_set = true; 362e332f741SMel Gorman reset_free = free_pfn; 363e332f741SMel Gorman zone->compact_init_free_pfn = reset_free; 364e332f741SMel Gorman zone->compact_cached_free_pfn = reset_free; 365e332f741SMel Gorman } 366e332f741SMel Gorman } 367e332f741SMel Gorman 368e332f741SMel Gorman /* Leave no distance if no suitable block was reset */ 369e332f741SMel Gorman if (reset_migrate >= reset_free) { 370e332f741SMel Gorman zone->compact_cached_migrate_pfn[0] = migrate_pfn; 371e332f741SMel Gorman zone->compact_cached_migrate_pfn[1] = migrate_pfn; 372e332f741SMel Gorman zone->compact_cached_free_pfn = free_pfn; 373e332f741SMel Gorman } 374bb13ffebSMel Gorman } 375bb13ffebSMel Gorman 37662997027SMel Gorman void reset_isolation_suitable(pg_data_t *pgdat) 37762997027SMel Gorman { 37862997027SMel Gorman int zoneid; 37962997027SMel Gorman 38062997027SMel Gorman for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 38162997027SMel Gorman struct zone *zone = &pgdat->node_zones[zoneid]; 38262997027SMel Gorman if (!populated_zone(zone)) 38362997027SMel Gorman continue; 38462997027SMel Gorman 38562997027SMel Gorman /* Only flush if a full compaction finished recently */ 38662997027SMel Gorman if (zone->compact_blockskip_flush) 38762997027SMel Gorman __reset_isolation_suitable(zone); 38862997027SMel Gorman } 38962997027SMel Gorman } 39062997027SMel Gorman 391bb13ffebSMel Gorman /* 392e380bebeSMel Gorman * Sets the pageblock skip bit if it was clear. Note that this is a hint as 393e380bebeSMel Gorman * locks are not required for read/writers. Returns true if it was already set. 394e380bebeSMel Gorman */ 395590ccea8SMel Gorman static bool test_and_set_skip(struct compact_control *cc, struct page *page) 396e380bebeSMel Gorman { 397e380bebeSMel Gorman bool skip; 398e380bebeSMel Gorman 399590ccea8SMel Gorman /* Do not update if skip hint is being ignored */ 400e380bebeSMel Gorman if (cc->ignore_skip_hint) 401e380bebeSMel Gorman return false; 402e380bebeSMel Gorman 403e380bebeSMel Gorman skip = get_pageblock_skip(page); 404e380bebeSMel Gorman if (!skip && !cc->no_set_skip_hint) 405e380bebeSMel Gorman set_pageblock_skip(page); 406e380bebeSMel Gorman 407e380bebeSMel Gorman return skip; 408e380bebeSMel Gorman } 409e380bebeSMel Gorman 410e380bebeSMel Gorman static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) 411e380bebeSMel Gorman { 412e380bebeSMel Gorman struct zone *zone = cc->zone; 413e380bebeSMel Gorman 414e380bebeSMel Gorman pfn = pageblock_end_pfn(pfn); 415e380bebeSMel Gorman 416e380bebeSMel Gorman /* Set for isolation rather than compaction */ 417e380bebeSMel Gorman if (cc->no_set_skip_hint) 418e380bebeSMel Gorman return; 419e380bebeSMel Gorman 420e380bebeSMel Gorman if (pfn > zone->compact_cached_migrate_pfn[0]) 421e380bebeSMel Gorman zone->compact_cached_migrate_pfn[0] = pfn; 422e380bebeSMel Gorman if (cc->mode != MIGRATE_ASYNC && 423e380bebeSMel Gorman pfn > zone->compact_cached_migrate_pfn[1]) 424e380bebeSMel Gorman zone->compact_cached_migrate_pfn[1] = pfn; 425e380bebeSMel Gorman } 426e380bebeSMel Gorman 427e380bebeSMel Gorman /* 428bb13ffebSMel Gorman * If no pages were isolated then mark this pageblock to be skipped in the 42962997027SMel Gorman * future. The information is later cleared by __reset_isolation_suitable(). 430bb13ffebSMel Gorman */ 431c89511abSMel Gorman static void update_pageblock_skip(struct compact_control *cc, 432d097a6f6SMel Gorman struct page *page, unsigned long pfn) 433bb13ffebSMel Gorman { 434c89511abSMel Gorman struct zone *zone = cc->zone; 4356815bf3fSJoonsoo Kim 4362583d671SVlastimil Babka if (cc->no_set_skip_hint) 4376815bf3fSJoonsoo Kim return; 4386815bf3fSJoonsoo Kim 439bb13ffebSMel Gorman set_pageblock_skip(page); 440c89511abSMel Gorman 44135979ef3SDavid Rientjes /* Update where async and sync compaction should restart */ 44235979ef3SDavid Rientjes if (pfn < zone->compact_cached_free_pfn) 443c89511abSMel Gorman zone->compact_cached_free_pfn = pfn; 444c89511abSMel Gorman } 445bb13ffebSMel Gorman #else 446bb13ffebSMel Gorman static inline bool isolation_suitable(struct compact_control *cc, 447bb13ffebSMel Gorman struct page *page) 448bb13ffebSMel Gorman { 449bb13ffebSMel Gorman return true; 450bb13ffebSMel Gorman } 451bb13ffebSMel Gorman 452b527cfe5SVlastimil Babka static inline bool pageblock_skip_persistent(struct page *page) 45321dc7e02SDavid Rientjes { 45421dc7e02SDavid Rientjes return false; 45521dc7e02SDavid Rientjes } 45621dc7e02SDavid Rientjes 45721dc7e02SDavid Rientjes static inline void update_pageblock_skip(struct compact_control *cc, 458d097a6f6SMel Gorman struct page *page, unsigned long pfn) 459bb13ffebSMel Gorman { 460bb13ffebSMel Gorman } 461e380bebeSMel Gorman 462e380bebeSMel Gorman static void update_cached_migrate(struct compact_control *cc, unsigned long pfn) 463e380bebeSMel Gorman { 464e380bebeSMel Gorman } 465e380bebeSMel Gorman 466590ccea8SMel Gorman static bool test_and_set_skip(struct compact_control *cc, struct page *page) 467e380bebeSMel Gorman { 468e380bebeSMel Gorman return false; 469e380bebeSMel Gorman } 470bb13ffebSMel Gorman #endif /* CONFIG_COMPACTION */ 471bb13ffebSMel Gorman 4721f9efdefSVlastimil Babka /* 4738b44d279SVlastimil Babka * Compaction requires the taking of some coarse locks that are potentially 474cb2dcaf0SMel Gorman * very heavily contended. For async compaction, trylock and record if the 475cb2dcaf0SMel Gorman * lock is contended. The lock will still be acquired but compaction will 476cb2dcaf0SMel Gorman * abort when the current block is finished regardless of success rate. 477cb2dcaf0SMel Gorman * Sync compaction acquires the lock. 4788b44d279SVlastimil Babka * 479cb2dcaf0SMel Gorman * Always returns true which makes it easier to track lock state in callers. 4801f9efdefSVlastimil Babka */ 481cb2dcaf0SMel Gorman static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags, 4828b44d279SVlastimil Babka struct compact_control *cc) 48377337edeSJules Irenge __acquires(lock) 4848b44d279SVlastimil Babka { 485cb2dcaf0SMel Gorman /* Track if the lock is contended in async mode */ 486cb2dcaf0SMel Gorman if (cc->mode == MIGRATE_ASYNC && !cc->contended) { 487cb2dcaf0SMel Gorman if (spin_trylock_irqsave(lock, *flags)) 488cb2dcaf0SMel Gorman return true; 489cb2dcaf0SMel Gorman 490c3486f53SVlastimil Babka cc->contended = true; 4918b44d279SVlastimil Babka } 4921f9efdefSVlastimil Babka 493cb2dcaf0SMel Gorman spin_lock_irqsave(lock, *flags); 4948b44d279SVlastimil Babka return true; 4952a1402aaSMel Gorman } 4962a1402aaSMel Gorman 49785aa125fSMichal Nazarewicz /* 498c67fe375SMel Gorman * Compaction requires the taking of some coarse locks that are potentially 4998b44d279SVlastimil Babka * very heavily contended. The lock should be periodically unlocked to avoid 5008b44d279SVlastimil Babka * having disabled IRQs for a long time, even when there is nobody waiting on 5018b44d279SVlastimil Babka * the lock. It might also be that allowing the IRQs will result in 502d56c1584SMiaohe Lin * need_resched() becoming true. If scheduling is needed, compaction schedules. 5038b44d279SVlastimil Babka * Either compaction type will also abort if a fatal signal is pending. 5048b44d279SVlastimil Babka * In either case if the lock was locked, it is dropped and not regained. 505c67fe375SMel Gorman * 506d56c1584SMiaohe Lin * Returns true if compaction should abort due to fatal signal pending. 507d56c1584SMiaohe Lin * Returns false when compaction can continue. 508c67fe375SMel Gorman */ 5098b44d279SVlastimil Babka static bool compact_unlock_should_abort(spinlock_t *lock, 5108b44d279SVlastimil Babka unsigned long flags, bool *locked, struct compact_control *cc) 511c67fe375SMel Gorman { 5128b44d279SVlastimil Babka if (*locked) { 5138b44d279SVlastimil Babka spin_unlock_irqrestore(lock, flags); 5148b44d279SVlastimil Babka *locked = false; 515c67fe375SMel Gorman } 516c67fe375SMel Gorman 5178b44d279SVlastimil Babka if (fatal_signal_pending(current)) { 518c3486f53SVlastimil Babka cc->contended = true; 5198b44d279SVlastimil Babka return true; 5208b44d279SVlastimil Babka } 5218b44d279SVlastimil Babka 522cf66f070SMel Gorman cond_resched(); 523be976572SVlastimil Babka 524be976572SVlastimil Babka return false; 525be976572SVlastimil Babka } 526be976572SVlastimil Babka 527c67fe375SMel Gorman /* 5289e4be470SJerome Marchand * Isolate free pages onto a private freelist. If @strict is true, will abort 5299e4be470SJerome Marchand * returning 0 on any invalid PFNs or non-free pages inside of the pageblock 5309e4be470SJerome Marchand * (even though it may still end up isolating some pages). 53185aa125fSMichal Nazarewicz */ 532f40d1e42SMel Gorman static unsigned long isolate_freepages_block(struct compact_control *cc, 533e14c720eSVlastimil Babka unsigned long *start_pfn, 53485aa125fSMichal Nazarewicz unsigned long end_pfn, 53585aa125fSMichal Nazarewicz struct list_head *freelist, 5364fca9730SMel Gorman unsigned int stride, 53785aa125fSMichal Nazarewicz bool strict) 538748446bbSMel Gorman { 539b7aba698SMel Gorman int nr_scanned = 0, total_isolated = 0; 540d097a6f6SMel Gorman struct page *cursor; 541b8b2d825SXiubo Li unsigned long flags = 0; 542f40d1e42SMel Gorman bool locked = false; 543e14c720eSVlastimil Babka unsigned long blockpfn = *start_pfn; 54466c64223SJoonsoo Kim unsigned int order; 545748446bbSMel Gorman 5464fca9730SMel Gorman /* Strict mode is for isolation, speed is secondary */ 5474fca9730SMel Gorman if (strict) 5484fca9730SMel Gorman stride = 1; 5494fca9730SMel Gorman 550748446bbSMel Gorman cursor = pfn_to_page(blockpfn); 551748446bbSMel Gorman 552f40d1e42SMel Gorman /* Isolate free pages. */ 5534fca9730SMel Gorman for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) { 55466c64223SJoonsoo Kim int isolated; 555748446bbSMel Gorman struct page *page = cursor; 556748446bbSMel Gorman 5578b44d279SVlastimil Babka /* 5588b44d279SVlastimil Babka * Periodically drop the lock (if held) regardless of its 5598b44d279SVlastimil Babka * contention, to give chance to IRQs. Abort if fatal signal 560d56c1584SMiaohe Lin * pending. 5618b44d279SVlastimil Babka */ 562c036ddffSMiaohe Lin if (!(blockpfn % COMPACT_CLUSTER_MAX) 5638b44d279SVlastimil Babka && compact_unlock_should_abort(&cc->zone->lock, flags, 5648b44d279SVlastimil Babka &locked, cc)) 5658b44d279SVlastimil Babka break; 5668b44d279SVlastimil Babka 567b7aba698SMel Gorman nr_scanned++; 5682af120bcSLaura Abbott 5699fcd6d2eSVlastimil Babka /* 5709fcd6d2eSVlastimil Babka * For compound pages such as THP and hugetlbfs, we can save 5719fcd6d2eSVlastimil Babka * potentially a lot of iterations if we skip them at once. 5729fcd6d2eSVlastimil Babka * The check is racy, but we can consider only valid values 5739fcd6d2eSVlastimil Babka * and the only danger is skipping too much. 5749fcd6d2eSVlastimil Babka */ 5759fcd6d2eSVlastimil Babka if (PageCompound(page)) { 57621dc7e02SDavid Rientjes const unsigned int order = compound_order(page); 5779fcd6d2eSVlastimil Babka 57823baf831SKirill A. Shutemov if (likely(order <= MAX_ORDER)) { 57921dc7e02SDavid Rientjes blockpfn += (1UL << order) - 1; 58021dc7e02SDavid Rientjes cursor += (1UL << order) - 1; 58156d48d8dSBaolin Wang nr_scanned += (1UL << order) - 1; 5829fcd6d2eSVlastimil Babka } 5839fcd6d2eSVlastimil Babka goto isolate_fail; 5849fcd6d2eSVlastimil Babka } 5859fcd6d2eSVlastimil Babka 586f40d1e42SMel Gorman if (!PageBuddy(page)) 5872af120bcSLaura Abbott goto isolate_fail; 588f40d1e42SMel Gorman 58985f73e6dSMiaohe Lin /* If we already hold the lock, we can skip some rechecking. */ 59069b7189fSVlastimil Babka if (!locked) { 591cb2dcaf0SMel Gorman locked = compact_lock_irqsave(&cc->zone->lock, 5928b44d279SVlastimil Babka &flags, cc); 593f40d1e42SMel Gorman 594f40d1e42SMel Gorman /* Recheck this is a buddy page under lock */ 595f40d1e42SMel Gorman if (!PageBuddy(page)) 5962af120bcSLaura Abbott goto isolate_fail; 59769b7189fSVlastimil Babka } 598748446bbSMel Gorman 59966c64223SJoonsoo Kim /* Found a free page, will break it into order-0 pages */ 600ab130f91SMatthew Wilcox (Oracle) order = buddy_order(page); 60166c64223SJoonsoo Kim isolated = __isolate_free_page(page, order); 602a4f04f2cSDavid Rientjes if (!isolated) 603a4f04f2cSDavid Rientjes break; 60466c64223SJoonsoo Kim set_page_private(page, order); 605a4f04f2cSDavid Rientjes 606b717d6b9SWilliam Lam nr_scanned += isolated - 1; 607748446bbSMel Gorman total_isolated += isolated; 608a4f04f2cSDavid Rientjes cc->nr_freepages += isolated; 60966c64223SJoonsoo Kim list_add_tail(&page->lru, freelist); 61066c64223SJoonsoo Kim 611a4f04f2cSDavid Rientjes if (!strict && cc->nr_migratepages <= cc->nr_freepages) { 612932ff6bbSJoonsoo Kim blockpfn += isolated; 613932ff6bbSJoonsoo Kim break; 614932ff6bbSJoonsoo Kim } 615a4f04f2cSDavid Rientjes /* Advance to the end of split page */ 616748446bbSMel Gorman blockpfn += isolated - 1; 617748446bbSMel Gorman cursor += isolated - 1; 6182af120bcSLaura Abbott continue; 6192af120bcSLaura Abbott 6202af120bcSLaura Abbott isolate_fail: 6212af120bcSLaura Abbott if (strict) 6222af120bcSLaura Abbott break; 6232af120bcSLaura Abbott else 6242af120bcSLaura Abbott continue; 6252af120bcSLaura Abbott 626748446bbSMel Gorman } 627748446bbSMel Gorman 628a4f04f2cSDavid Rientjes if (locked) 629a4f04f2cSDavid Rientjes spin_unlock_irqrestore(&cc->zone->lock, flags); 630a4f04f2cSDavid Rientjes 6319fcd6d2eSVlastimil Babka /* 6329fcd6d2eSVlastimil Babka * There is a tiny chance that we have read bogus compound_order(), 6339fcd6d2eSVlastimil Babka * so be careful to not go outside of the pageblock. 6349fcd6d2eSVlastimil Babka */ 6359fcd6d2eSVlastimil Babka if (unlikely(blockpfn > end_pfn)) 6369fcd6d2eSVlastimil Babka blockpfn = end_pfn; 6379fcd6d2eSVlastimil Babka 638e34d85f0SJoonsoo Kim trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn, 639e34d85f0SJoonsoo Kim nr_scanned, total_isolated); 640e34d85f0SJoonsoo Kim 641e14c720eSVlastimil Babka /* Record how far we have got within the block */ 642e14c720eSVlastimil Babka *start_pfn = blockpfn; 643e14c720eSVlastimil Babka 644f40d1e42SMel Gorman /* 645f40d1e42SMel Gorman * If strict isolation is requested by CMA then check that all the 646f40d1e42SMel Gorman * pages requested were isolated. If there were any failures, 0 is 647f40d1e42SMel Gorman * returned and CMA will fail. 648f40d1e42SMel Gorman */ 6492af120bcSLaura Abbott if (strict && blockpfn < end_pfn) 650f40d1e42SMel Gorman total_isolated = 0; 651f40d1e42SMel Gorman 6527f354a54SDavid Rientjes cc->total_free_scanned += nr_scanned; 653397487dbSMel Gorman if (total_isolated) 654010fc29aSMinchan Kim count_compact_events(COMPACTISOLATED, total_isolated); 655748446bbSMel Gorman return total_isolated; 656748446bbSMel Gorman } 657748446bbSMel Gorman 65885aa125fSMichal Nazarewicz /** 65985aa125fSMichal Nazarewicz * isolate_freepages_range() - isolate free pages. 660e8b098fcSMike Rapoport * @cc: Compaction control structure. 66185aa125fSMichal Nazarewicz * @start_pfn: The first PFN to start isolating. 66285aa125fSMichal Nazarewicz * @end_pfn: The one-past-last PFN. 66385aa125fSMichal Nazarewicz * 66485aa125fSMichal Nazarewicz * Non-free pages, invalid PFNs, or zone boundaries within the 66585aa125fSMichal Nazarewicz * [start_pfn, end_pfn) range are considered errors, cause function to 66685aa125fSMichal Nazarewicz * undo its actions and return zero. 66785aa125fSMichal Nazarewicz * 66885aa125fSMichal Nazarewicz * Otherwise, function returns one-past-the-last PFN of isolated page 66985aa125fSMichal Nazarewicz * (which may be greater then end_pfn if end fell in a middle of 67085aa125fSMichal Nazarewicz * a free page). 67185aa125fSMichal Nazarewicz */ 672ff9543fdSMichal Nazarewicz unsigned long 673bb13ffebSMel Gorman isolate_freepages_range(struct compact_control *cc, 674bb13ffebSMel Gorman unsigned long start_pfn, unsigned long end_pfn) 67585aa125fSMichal Nazarewicz { 676e1409c32SJoonsoo Kim unsigned long isolated, pfn, block_start_pfn, block_end_pfn; 67785aa125fSMichal Nazarewicz LIST_HEAD(freelist); 67885aa125fSMichal Nazarewicz 6797d49d886SVlastimil Babka pfn = start_pfn; 68006b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 681e1409c32SJoonsoo Kim if (block_start_pfn < cc->zone->zone_start_pfn) 682e1409c32SJoonsoo Kim block_start_pfn = cc->zone->zone_start_pfn; 68306b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 6847d49d886SVlastimil Babka 6857d49d886SVlastimil Babka for (; pfn < end_pfn; pfn += isolated, 686e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 6877d49d886SVlastimil Babka block_end_pfn += pageblock_nr_pages) { 688e14c720eSVlastimil Babka /* Protect pfn from changing by isolate_freepages_block */ 689e14c720eSVlastimil Babka unsigned long isolate_start_pfn = pfn; 6907d49d886SVlastimil Babka 69185aa125fSMichal Nazarewicz block_end_pfn = min(block_end_pfn, end_pfn); 69285aa125fSMichal Nazarewicz 69358420016SJoonsoo Kim /* 69458420016SJoonsoo Kim * pfn could pass the block_end_pfn if isolated freepage 69558420016SJoonsoo Kim * is more than pageblock order. In this case, we adjust 69658420016SJoonsoo Kim * scanning range to right one. 69758420016SJoonsoo Kim */ 69858420016SJoonsoo Kim if (pfn >= block_end_pfn) { 69906b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 70006b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 70158420016SJoonsoo Kim block_end_pfn = min(block_end_pfn, end_pfn); 70258420016SJoonsoo Kim } 70358420016SJoonsoo Kim 704e1409c32SJoonsoo Kim if (!pageblock_pfn_to_page(block_start_pfn, 705e1409c32SJoonsoo Kim block_end_pfn, cc->zone)) 7067d49d886SVlastimil Babka break; 7077d49d886SVlastimil Babka 708e14c720eSVlastimil Babka isolated = isolate_freepages_block(cc, &isolate_start_pfn, 7094fca9730SMel Gorman block_end_pfn, &freelist, 0, true); 71085aa125fSMichal Nazarewicz 71185aa125fSMichal Nazarewicz /* 71285aa125fSMichal Nazarewicz * In strict mode, isolate_freepages_block() returns 0 if 71385aa125fSMichal Nazarewicz * there are any holes in the block (ie. invalid PFNs or 71485aa125fSMichal Nazarewicz * non-free pages). 71585aa125fSMichal Nazarewicz */ 71685aa125fSMichal Nazarewicz if (!isolated) 71785aa125fSMichal Nazarewicz break; 71885aa125fSMichal Nazarewicz 71985aa125fSMichal Nazarewicz /* 72085aa125fSMichal Nazarewicz * If we managed to isolate pages, it is always (1 << n) * 72185aa125fSMichal Nazarewicz * pageblock_nr_pages for some non-negative n. (Max order 72285aa125fSMichal Nazarewicz * page may span two pageblocks). 72385aa125fSMichal Nazarewicz */ 72485aa125fSMichal Nazarewicz } 72585aa125fSMichal Nazarewicz 72666c64223SJoonsoo Kim /* __isolate_free_page() does not map the pages */ 7274469ab98SMel Gorman split_map_pages(&freelist); 72885aa125fSMichal Nazarewicz 72985aa125fSMichal Nazarewicz if (pfn < end_pfn) { 73085aa125fSMichal Nazarewicz /* Loop terminated early, cleanup. */ 73185aa125fSMichal Nazarewicz release_freepages(&freelist); 73285aa125fSMichal Nazarewicz return 0; 73385aa125fSMichal Nazarewicz } 73485aa125fSMichal Nazarewicz 73585aa125fSMichal Nazarewicz /* We don't use freelists for anything. */ 73685aa125fSMichal Nazarewicz return pfn; 73785aa125fSMichal Nazarewicz } 73885aa125fSMichal Nazarewicz 739748446bbSMel Gorman /* Similar to reclaim, but different enough that they don't share logic */ 7404fbbb3fdSJohannes Weiner static bool too_many_isolated(struct compact_control *cc) 741748446bbSMel Gorman { 7424fbbb3fdSJohannes Weiner pg_data_t *pgdat = cc->zone->zone_pgdat; 743d818fca1SMel Gorman bool too_many; 744d818fca1SMel Gorman 745bc693045SMinchan Kim unsigned long active, inactive, isolated; 746748446bbSMel Gorman 7475f438eeeSAndrey Ryabinin inactive = node_page_state(pgdat, NR_INACTIVE_FILE) + 7485f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_INACTIVE_ANON); 7495f438eeeSAndrey Ryabinin active = node_page_state(pgdat, NR_ACTIVE_FILE) + 7505f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_ACTIVE_ANON); 7515f438eeeSAndrey Ryabinin isolated = node_page_state(pgdat, NR_ISOLATED_FILE) + 7525f438eeeSAndrey Ryabinin node_page_state(pgdat, NR_ISOLATED_ANON); 753748446bbSMel Gorman 7544fbbb3fdSJohannes Weiner /* 7554fbbb3fdSJohannes Weiner * Allow GFP_NOFS to isolate past the limit set for regular 7564fbbb3fdSJohannes Weiner * compaction runs. This prevents an ABBA deadlock when other 7574fbbb3fdSJohannes Weiner * compactors have already isolated to the limit, but are 7584fbbb3fdSJohannes Weiner * blocked on filesystem locks held by the GFP_NOFS thread. 7594fbbb3fdSJohannes Weiner */ 7604fbbb3fdSJohannes Weiner if (cc->gfp_mask & __GFP_FS) { 7614fbbb3fdSJohannes Weiner inactive >>= 3; 7624fbbb3fdSJohannes Weiner active >>= 3; 7634fbbb3fdSJohannes Weiner } 7644fbbb3fdSJohannes Weiner 765d818fca1SMel Gorman too_many = isolated > (inactive + active) / 2; 766d818fca1SMel Gorman if (!too_many) 767d818fca1SMel Gorman wake_throttle_isolated(pgdat); 768d818fca1SMel Gorman 769d818fca1SMel Gorman return too_many; 770748446bbSMel Gorman } 771748446bbSMel Gorman 7722fe86e00SMichal Nazarewicz /** 773edc2ca61SVlastimil Babka * isolate_migratepages_block() - isolate all migrate-able pages within 774edc2ca61SVlastimil Babka * a single pageblock 7752fe86e00SMichal Nazarewicz * @cc: Compaction control structure. 776edc2ca61SVlastimil Babka * @low_pfn: The first PFN to isolate 777edc2ca61SVlastimil Babka * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock 77889f6c88aSHugh Dickins * @mode: Isolation mode to be used. 7792fe86e00SMichal Nazarewicz * 7802fe86e00SMichal Nazarewicz * Isolate all pages that can be migrated from the range specified by 781edc2ca61SVlastimil Babka * [low_pfn, end_pfn). The range is expected to be within same pageblock. 782c2ad7a1fSOscar Salvador * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion, 783369fa227SOscar Salvador * -ENOMEM in case we could not allocate a page, or 0. 784c2ad7a1fSOscar Salvador * cc->migrate_pfn will contain the next pfn to scan. 7852fe86e00SMichal Nazarewicz * 786edc2ca61SVlastimil Babka * The pages are isolated on cc->migratepages list (not required to be empty), 787c2ad7a1fSOscar Salvador * and cc->nr_migratepages is updated accordingly. 788748446bbSMel Gorman */ 789c2ad7a1fSOscar Salvador static int 790edc2ca61SVlastimil Babka isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, 79189f6c88aSHugh Dickins unsigned long end_pfn, isolate_mode_t mode) 792748446bbSMel Gorman { 7935f438eeeSAndrey Ryabinin pg_data_t *pgdat = cc->zone->zone_pgdat; 794b7aba698SMel Gorman unsigned long nr_scanned = 0, nr_isolated = 0; 795fa9add64SHugh Dickins struct lruvec *lruvec; 796b8b2d825SXiubo Li unsigned long flags = 0; 7976168d0daSAlex Shi struct lruvec *locked = NULL; 798bb13ffebSMel Gorman struct page *page = NULL, *valid_page = NULL; 79989f6c88aSHugh Dickins struct address_space *mapping; 800e34d85f0SJoonsoo Kim unsigned long start_pfn = low_pfn; 801fdd048e1SVlastimil Babka bool skip_on_failure = false; 802fdd048e1SVlastimil Babka unsigned long next_skip_pfn = 0; 803e380bebeSMel Gorman bool skip_updated = false; 804c2ad7a1fSOscar Salvador int ret = 0; 805c2ad7a1fSOscar Salvador 806c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn; 807748446bbSMel Gorman 808748446bbSMel Gorman /* 809748446bbSMel Gorman * Ensure that there are not too many pages isolated from the LRU 810748446bbSMel Gorman * list by either parallel reclaimers or compaction. If there are, 811748446bbSMel Gorman * delay for some time until fewer pages are isolated 812748446bbSMel Gorman */ 8134fbbb3fdSJohannes Weiner while (unlikely(too_many_isolated(cc))) { 814d20bdd57SZi Yan /* stop isolation if there are still pages not migrated */ 815d20bdd57SZi Yan if (cc->nr_migratepages) 816c2ad7a1fSOscar Salvador return -EAGAIN; 817d20bdd57SZi Yan 818f9e35b3bSMel Gorman /* async migration should just abort */ 819e0b9daebSDavid Rientjes if (cc->mode == MIGRATE_ASYNC) 820c2ad7a1fSOscar Salvador return -EAGAIN; 821f9e35b3bSMel Gorman 822c3f4a9a2SMel Gorman reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); 823748446bbSMel Gorman 824748446bbSMel Gorman if (fatal_signal_pending(current)) 825c2ad7a1fSOscar Salvador return -EINTR; 826748446bbSMel Gorman } 827748446bbSMel Gorman 828cf66f070SMel Gorman cond_resched(); 829aeef4b83SDavid Rientjes 830fdd048e1SVlastimil Babka if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) { 831fdd048e1SVlastimil Babka skip_on_failure = true; 832fdd048e1SVlastimil Babka next_skip_pfn = block_end_pfn(low_pfn, cc->order); 833fdd048e1SVlastimil Babka } 834fdd048e1SVlastimil Babka 835748446bbSMel Gorman /* Time to isolate some pages for migration */ 836748446bbSMel Gorman for (; low_pfn < end_pfn; low_pfn++) { 83729c0dde8SVlastimil Babka 838fdd048e1SVlastimil Babka if (skip_on_failure && low_pfn >= next_skip_pfn) { 839fdd048e1SVlastimil Babka /* 840fdd048e1SVlastimil Babka * We have isolated all migration candidates in the 841fdd048e1SVlastimil Babka * previous order-aligned block, and did not skip it due 842fdd048e1SVlastimil Babka * to failure. We should migrate the pages now and 843fdd048e1SVlastimil Babka * hopefully succeed compaction. 844fdd048e1SVlastimil Babka */ 845fdd048e1SVlastimil Babka if (nr_isolated) 846fdd048e1SVlastimil Babka break; 847fdd048e1SVlastimil Babka 848fdd048e1SVlastimil Babka /* 849fdd048e1SVlastimil Babka * We failed to isolate in the previous order-aligned 850fdd048e1SVlastimil Babka * block. Set the new boundary to the end of the 851fdd048e1SVlastimil Babka * current block. Note we can't simply increase 852fdd048e1SVlastimil Babka * next_skip_pfn by 1 << order, as low_pfn might have 853fdd048e1SVlastimil Babka * been incremented by a higher number due to skipping 854fdd048e1SVlastimil Babka * a compound or a high-order buddy page in the 855fdd048e1SVlastimil Babka * previous loop iteration. 856fdd048e1SVlastimil Babka */ 857fdd048e1SVlastimil Babka next_skip_pfn = block_end_pfn(low_pfn, cc->order); 858fdd048e1SVlastimil Babka } 859fdd048e1SVlastimil Babka 8608b44d279SVlastimil Babka /* 8618b44d279SVlastimil Babka * Periodically drop the lock (if held) regardless of its 862670105a2SMel Gorman * contention, to give chance to IRQs. Abort completely if 863670105a2SMel Gorman * a fatal signal is pending. 8648b44d279SVlastimil Babka */ 865c036ddffSMiaohe Lin if (!(low_pfn % COMPACT_CLUSTER_MAX)) { 8666168d0daSAlex Shi if (locked) { 8676168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 8686168d0daSAlex Shi locked = NULL; 8696168d0daSAlex Shi } 8706168d0daSAlex Shi 8716168d0daSAlex Shi if (fatal_signal_pending(current)) { 8726168d0daSAlex Shi cc->contended = true; 873c2ad7a1fSOscar Salvador ret = -EINTR; 8746168d0daSAlex Shi 875670105a2SMel Gorman goto fatal_pending; 876670105a2SMel Gorman } 877b2eef8c0SAndrea Arcangeli 8786168d0daSAlex Shi cond_resched(); 8796168d0daSAlex Shi } 8806168d0daSAlex Shi 881b7aba698SMel Gorman nr_scanned++; 882748446bbSMel Gorman 883748446bbSMel Gorman page = pfn_to_page(low_pfn); 884dc908600SMel Gorman 885e380bebeSMel Gorman /* 886e380bebeSMel Gorman * Check if the pageblock has already been marked skipped. 887e380bebeSMel Gorman * Only the aligned PFN is checked as the caller isolates 888e380bebeSMel Gorman * COMPACT_CLUSTER_MAX at a time so the second call must 889e380bebeSMel Gorman * not falsely conclude that the block should be skipped. 890e380bebeSMel Gorman */ 891ee0913c4SKefeng Wang if (!valid_page && pageblock_aligned(low_pfn)) { 8924af12d04SMiaohe Lin if (!isolation_suitable(cc, page)) { 893e380bebeSMel Gorman low_pfn = end_pfn; 8949df41314SAlex Shi page = NULL; 895e380bebeSMel Gorman goto isolate_abort; 896e380bebeSMel Gorman } 897bb13ffebSMel Gorman valid_page = page; 898e380bebeSMel Gorman } 899bb13ffebSMel Gorman 900369fa227SOscar Salvador if (PageHuge(page) && cc->alloc_contig) { 9011c06b6a5SBaolin Wang if (locked) { 9021c06b6a5SBaolin Wang unlock_page_lruvec_irqrestore(locked, flags); 9031c06b6a5SBaolin Wang locked = NULL; 9041c06b6a5SBaolin Wang } 9051c06b6a5SBaolin Wang 906ae37c7ffSOscar Salvador ret = isolate_or_dissolve_huge_page(page, &cc->migratepages); 907369fa227SOscar Salvador 908369fa227SOscar Salvador /* 909369fa227SOscar Salvador * Fail isolation in case isolate_or_dissolve_huge_page() 910369fa227SOscar Salvador * reports an error. In case of -ENOMEM, abort right away. 911369fa227SOscar Salvador */ 912369fa227SOscar Salvador if (ret < 0) { 913369fa227SOscar Salvador /* Do not report -EBUSY down the chain */ 914369fa227SOscar Salvador if (ret == -EBUSY) 915369fa227SOscar Salvador ret = 0; 91666fe1cf7SMiaohe Lin low_pfn += compound_nr(page) - 1; 91756d48d8dSBaolin Wang nr_scanned += compound_nr(page) - 1; 918369fa227SOscar Salvador goto isolate_fail; 919369fa227SOscar Salvador } 920369fa227SOscar Salvador 921ae37c7ffSOscar Salvador if (PageHuge(page)) { 922ae37c7ffSOscar Salvador /* 923ae37c7ffSOscar Salvador * Hugepage was successfully isolated and placed 924ae37c7ffSOscar Salvador * on the cc->migratepages list. 925ae37c7ffSOscar Salvador */ 926ae37c7ffSOscar Salvador low_pfn += compound_nr(page) - 1; 927ae37c7ffSOscar Salvador goto isolate_success_no_list; 928ae37c7ffSOscar Salvador } 929ae37c7ffSOscar Salvador 930369fa227SOscar Salvador /* 931369fa227SOscar Salvador * Ok, the hugepage was dissolved. Now these pages are 932369fa227SOscar Salvador * Buddy and cannot be re-allocated because they are 933369fa227SOscar Salvador * isolated. Fall-through as the check below handles 934369fa227SOscar Salvador * Buddy pages. 935369fa227SOscar Salvador */ 936369fa227SOscar Salvador } 937369fa227SOscar Salvador 938c122b208SJoonsoo Kim /* 93999c0fd5eSVlastimil Babka * Skip if free. We read page order here without zone lock 94099c0fd5eSVlastimil Babka * which is generally unsafe, but the race window is small and 94199c0fd5eSVlastimil Babka * the worst thing that can happen is that we skip some 94299c0fd5eSVlastimil Babka * potential isolation targets. 9436c14466cSMel Gorman */ 94499c0fd5eSVlastimil Babka if (PageBuddy(page)) { 945ab130f91SMatthew Wilcox (Oracle) unsigned long freepage_order = buddy_order_unsafe(page); 94699c0fd5eSVlastimil Babka 94799c0fd5eSVlastimil Babka /* 94899c0fd5eSVlastimil Babka * Without lock, we cannot be sure that what we got is 94999c0fd5eSVlastimil Babka * a valid page order. Consider only values in the 95099c0fd5eSVlastimil Babka * valid order range to prevent low_pfn overflow. 95199c0fd5eSVlastimil Babka */ 95256d48d8dSBaolin Wang if (freepage_order > 0 && freepage_order <= MAX_ORDER) { 95399c0fd5eSVlastimil Babka low_pfn += (1UL << freepage_order) - 1; 95456d48d8dSBaolin Wang nr_scanned += (1UL << freepage_order) - 1; 95556d48d8dSBaolin Wang } 956748446bbSMel Gorman continue; 95799c0fd5eSVlastimil Babka } 958748446bbSMel Gorman 9599927af74SMel Gorman /* 96029c0dde8SVlastimil Babka * Regardless of being on LRU, compound pages such as THP and 9611da2f328SRik van Riel * hugetlbfs are not to be compacted unless we are attempting 9621da2f328SRik van Riel * an allocation much larger than the huge page size (eg CMA). 9631da2f328SRik van Riel * We can potentially save a lot of iterations if we skip them 9641da2f328SRik van Riel * at once. The check is racy, but we can consider only valid 9651da2f328SRik van Riel * values and the only danger is skipping too much. 966bc835011SAndrea Arcangeli */ 9671da2f328SRik van Riel if (PageCompound(page) && !cc->alloc_contig) { 96821dc7e02SDavid Rientjes const unsigned int order = compound_order(page); 96929c0dde8SVlastimil Babka 97056d48d8dSBaolin Wang if (likely(order <= MAX_ORDER)) { 97121dc7e02SDavid Rientjes low_pfn += (1UL << order) - 1; 97256d48d8dSBaolin Wang nr_scanned += (1UL << order) - 1; 97356d48d8dSBaolin Wang } 974fdd048e1SVlastimil Babka goto isolate_fail; 9752a1402aaSMel Gorman } 9762a1402aaSMel Gorman 977bda807d4SMinchan Kim /* 978bda807d4SMinchan Kim * Check may be lockless but that's ok as we recheck later. 979bda807d4SMinchan Kim * It's possible to migrate LRU and non-lru movable pages. 980bda807d4SMinchan Kim * Skip any other type of page 981bda807d4SMinchan Kim */ 982bda807d4SMinchan Kim if (!PageLRU(page)) { 983bda807d4SMinchan Kim /* 984bda807d4SMinchan Kim * __PageMovable can return false positive so we need 985bda807d4SMinchan Kim * to verify it under page_lock. 986bda807d4SMinchan Kim */ 987bda807d4SMinchan Kim if (unlikely(__PageMovable(page)) && 988bda807d4SMinchan Kim !PageIsolated(page)) { 989bda807d4SMinchan Kim if (locked) { 9906168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 9916168d0daSAlex Shi locked = NULL; 992bda807d4SMinchan Kim } 993bda807d4SMinchan Kim 994cd775580SBaolin Wang if (isolate_movable_page(page, mode)) 995bda807d4SMinchan Kim goto isolate_success; 996bda807d4SMinchan Kim } 997bda807d4SMinchan Kim 998fdd048e1SVlastimil Babka goto isolate_fail; 999bda807d4SMinchan Kim } 100029c0dde8SVlastimil Babka 1001119d6d59SDavid Rientjes /* 10029df41314SAlex Shi * Be careful not to clear PageLRU until after we're 10039df41314SAlex Shi * sure the page is not being freed elsewhere -- the 10049df41314SAlex Shi * page release code relies on it. 10059df41314SAlex Shi */ 10069df41314SAlex Shi if (unlikely(!get_page_unless_zero(page))) 10079df41314SAlex Shi goto isolate_fail; 10089df41314SAlex Shi 1009829ae0f8SGavin Shan /* 1010829ae0f8SGavin Shan * Migration will fail if an anonymous page is pinned in memory, 1011829ae0f8SGavin Shan * so avoid taking lru_lock and isolating it unnecessarily in an 1012829ae0f8SGavin Shan * admittedly racy check. 1013829ae0f8SGavin Shan */ 1014829ae0f8SGavin Shan mapping = page_mapping(page); 1015829ae0f8SGavin Shan if (!mapping && (page_count(page) - 1) > total_mapcount(page)) 1016829ae0f8SGavin Shan goto isolate_fail_put; 1017829ae0f8SGavin Shan 1018829ae0f8SGavin Shan /* 1019829ae0f8SGavin Shan * Only allow to migrate anonymous pages in GFP_NOFS context 1020829ae0f8SGavin Shan * because those do not depend on fs locks. 1021829ae0f8SGavin Shan */ 1022829ae0f8SGavin Shan if (!(cc->gfp_mask & __GFP_FS) && mapping) 1023829ae0f8SGavin Shan goto isolate_fail_put; 1024829ae0f8SGavin Shan 102589f6c88aSHugh Dickins /* Only take pages on LRU: a check now makes later tests safe */ 102689f6c88aSHugh Dickins if (!PageLRU(page)) 10279df41314SAlex Shi goto isolate_fail_put; 10289df41314SAlex Shi 102989f6c88aSHugh Dickins /* Compaction might skip unevictable pages but CMA takes them */ 103089f6c88aSHugh Dickins if (!(mode & ISOLATE_UNEVICTABLE) && PageUnevictable(page)) 103189f6c88aSHugh Dickins goto isolate_fail_put; 103289f6c88aSHugh Dickins 103389f6c88aSHugh Dickins /* 103489f6c88aSHugh Dickins * To minimise LRU disruption, the caller can indicate with 103589f6c88aSHugh Dickins * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages 103689f6c88aSHugh Dickins * it will be able to migrate without blocking - clean pages 103789f6c88aSHugh Dickins * for the most part. PageWriteback would require blocking. 103889f6c88aSHugh Dickins */ 103989f6c88aSHugh Dickins if ((mode & ISOLATE_ASYNC_MIGRATE) && PageWriteback(page)) 104089f6c88aSHugh Dickins goto isolate_fail_put; 104189f6c88aSHugh Dickins 104289f6c88aSHugh Dickins if ((mode & ISOLATE_ASYNC_MIGRATE) && PageDirty(page)) { 104389f6c88aSHugh Dickins bool migrate_dirty; 104489f6c88aSHugh Dickins 104589f6c88aSHugh Dickins /* 104689f6c88aSHugh Dickins * Only pages without mappings or that have a 10479d0ddc0cSMatthew Wilcox (Oracle) * ->migrate_folio callback are possible to migrate 104889f6c88aSHugh Dickins * without blocking. However, we can be racing with 104989f6c88aSHugh Dickins * truncation so it's necessary to lock the page 105089f6c88aSHugh Dickins * to stabilise the mapping as truncation holds 105189f6c88aSHugh Dickins * the page lock until after the page is removed 105289f6c88aSHugh Dickins * from the page cache. 105389f6c88aSHugh Dickins */ 105489f6c88aSHugh Dickins if (!trylock_page(page)) 105589f6c88aSHugh Dickins goto isolate_fail_put; 105689f6c88aSHugh Dickins 105789f6c88aSHugh Dickins mapping = page_mapping(page); 10585490da4fSMatthew Wilcox (Oracle) migrate_dirty = !mapping || 10599d0ddc0cSMatthew Wilcox (Oracle) mapping->a_ops->migrate_folio; 106089f6c88aSHugh Dickins unlock_page(page); 106189f6c88aSHugh Dickins if (!migrate_dirty) 106289f6c88aSHugh Dickins goto isolate_fail_put; 106389f6c88aSHugh Dickins } 106489f6c88aSHugh Dickins 10659df41314SAlex Shi /* Try isolate the page */ 10669df41314SAlex Shi if (!TestClearPageLRU(page)) 10679df41314SAlex Shi goto isolate_fail_put; 10689df41314SAlex Shi 1069b1baabd9SMatthew Wilcox (Oracle) lruvec = folio_lruvec(page_folio(page)); 10706168d0daSAlex Shi 107169b7189fSVlastimil Babka /* If we already hold the lock, we can skip some rechecking */ 10726168d0daSAlex Shi if (lruvec != locked) { 10736168d0daSAlex Shi if (locked) 10746168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 10756168d0daSAlex Shi 10766168d0daSAlex Shi compact_lock_irqsave(&lruvec->lru_lock, &flags, cc); 10776168d0daSAlex Shi locked = lruvec; 10786168d0daSAlex Shi 1079e809c3feSMatthew Wilcox (Oracle) lruvec_memcg_debug(lruvec, page_folio(page)); 1080e380bebeSMel Gorman 1081590ccea8SMel Gorman /* 1082590ccea8SMel Gorman * Try get exclusive access under lock. If marked for 1083590ccea8SMel Gorman * skip, the scan is aborted unless the current context 1084590ccea8SMel Gorman * is a rescan to reach the end of the pageblock. 1085590ccea8SMel Gorman */ 1086590ccea8SMel Gorman if (!skip_updated && valid_page) { 1087e380bebeSMel Gorman skip_updated = true; 1088590ccea8SMel Gorman if (test_and_set_skip(cc, valid_page) && 1089590ccea8SMel Gorman !cc->finish_pageblock) { 1090e380bebeSMel Gorman goto isolate_abort; 1091e380bebeSMel Gorman } 1092590ccea8SMel Gorman } 10932a1402aaSMel Gorman 109429c0dde8SVlastimil Babka /* 109529c0dde8SVlastimil Babka * Page become compound since the non-locked check, 109629c0dde8SVlastimil Babka * and it's on LRU. It can only be a THP so the order 109729c0dde8SVlastimil Babka * is safe to read and it's 0 for tail pages. 109829c0dde8SVlastimil Babka */ 10991da2f328SRik van Riel if (unlikely(PageCompound(page) && !cc->alloc_contig)) { 1100d8c6546bSMatthew Wilcox (Oracle) low_pfn += compound_nr(page) - 1; 110156d48d8dSBaolin Wang nr_scanned += compound_nr(page) - 1; 11029df41314SAlex Shi SetPageLRU(page); 11039df41314SAlex Shi goto isolate_fail_put; 1104bc835011SAndrea Arcangeli } 1105d99fd5feSAlex Shi } 1106fa9add64SHugh Dickins 11071da2f328SRik van Riel /* The whole page is taken off the LRU; skip the tail pages. */ 11081da2f328SRik van Riel if (PageCompound(page)) 11091da2f328SRik van Riel low_pfn += compound_nr(page) - 1; 1110bc835011SAndrea Arcangeli 1111748446bbSMel Gorman /* Successfully isolated */ 111246ae6b2cSYu Zhao del_page_from_lru_list(page, lruvec); 11131da2f328SRik van Riel mod_node_page_state(page_pgdat(page), 11149de4f22aSHuang Ying NR_ISOLATED_ANON + page_is_file_lru(page), 11156c357848SMatthew Wilcox (Oracle) thp_nr_pages(page)); 1116b6c75016SJoonsoo Kim 1117b6c75016SJoonsoo Kim isolate_success: 1118fdd048e1SVlastimil Babka list_add(&page->lru, &cc->migratepages); 1119ae37c7ffSOscar Salvador isolate_success_no_list: 112038935861SZi Yan cc->nr_migratepages += compound_nr(page); 112138935861SZi Yan nr_isolated += compound_nr(page); 1122b717d6b9SWilliam Lam nr_scanned += compound_nr(page) - 1; 1123748446bbSMel Gorman 1124804d3121SMel Gorman /* 1125804d3121SMel Gorman * Avoid isolating too much unless this block is being 112648731c84SMel Gorman * fully scanned (e.g. dirty/writeback pages, parallel allocation) 1127cb2dcaf0SMel Gorman * or a lock is contended. For contention, isolate quickly to 1128cb2dcaf0SMel Gorman * potentially remove one source of contention. 1129804d3121SMel Gorman */ 113038935861SZi Yan if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX && 113148731c84SMel Gorman !cc->finish_pageblock && !cc->contended) { 113231b8384aSHillf Danton ++low_pfn; 1133748446bbSMel Gorman break; 1134748446bbSMel Gorman } 1135fdd048e1SVlastimil Babka 1136fdd048e1SVlastimil Babka continue; 11379df41314SAlex Shi 11389df41314SAlex Shi isolate_fail_put: 11399df41314SAlex Shi /* Avoid potential deadlock in freeing page under lru_lock */ 11409df41314SAlex Shi if (locked) { 11416168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 11426168d0daSAlex Shi locked = NULL; 11439df41314SAlex Shi } 11449df41314SAlex Shi put_page(page); 11459df41314SAlex Shi 1146fdd048e1SVlastimil Babka isolate_fail: 1147369fa227SOscar Salvador if (!skip_on_failure && ret != -ENOMEM) 1148fdd048e1SVlastimil Babka continue; 1149fdd048e1SVlastimil Babka 1150fdd048e1SVlastimil Babka /* 1151fdd048e1SVlastimil Babka * We have isolated some pages, but then failed. Release them 1152fdd048e1SVlastimil Babka * instead of migrating, as we cannot form the cc->order buddy 1153fdd048e1SVlastimil Babka * page anyway. 1154fdd048e1SVlastimil Babka */ 1155fdd048e1SVlastimil Babka if (nr_isolated) { 1156fdd048e1SVlastimil Babka if (locked) { 11576168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 11586168d0daSAlex Shi locked = NULL; 1159fdd048e1SVlastimil Babka } 1160fdd048e1SVlastimil Babka putback_movable_pages(&cc->migratepages); 1161fdd048e1SVlastimil Babka cc->nr_migratepages = 0; 1162fdd048e1SVlastimil Babka nr_isolated = 0; 1163fdd048e1SVlastimil Babka } 1164fdd048e1SVlastimil Babka 1165fdd048e1SVlastimil Babka if (low_pfn < next_skip_pfn) { 1166fdd048e1SVlastimil Babka low_pfn = next_skip_pfn - 1; 1167fdd048e1SVlastimil Babka /* 1168fdd048e1SVlastimil Babka * The check near the loop beginning would have updated 1169fdd048e1SVlastimil Babka * next_skip_pfn too, but this is a bit simpler. 1170fdd048e1SVlastimil Babka */ 1171fdd048e1SVlastimil Babka next_skip_pfn += 1UL << cc->order; 1172fdd048e1SVlastimil Babka } 1173369fa227SOscar Salvador 1174369fa227SOscar Salvador if (ret == -ENOMEM) 1175369fa227SOscar Salvador break; 117631b8384aSHillf Danton } 1177748446bbSMel Gorman 117899c0fd5eSVlastimil Babka /* 117999c0fd5eSVlastimil Babka * The PageBuddy() check could have potentially brought us outside 118099c0fd5eSVlastimil Babka * the range to be scanned. 118199c0fd5eSVlastimil Babka */ 118299c0fd5eSVlastimil Babka if (unlikely(low_pfn > end_pfn)) 118399c0fd5eSVlastimil Babka low_pfn = end_pfn; 118499c0fd5eSVlastimil Babka 11859df41314SAlex Shi page = NULL; 11869df41314SAlex Shi 1187e380bebeSMel Gorman isolate_abort: 1188c67fe375SMel Gorman if (locked) 11896168d0daSAlex Shi unlock_page_lruvec_irqrestore(locked, flags); 11909df41314SAlex Shi if (page) { 11919df41314SAlex Shi SetPageLRU(page); 11929df41314SAlex Shi put_page(page); 11939df41314SAlex Shi } 1194748446bbSMel Gorman 119550b5b094SVlastimil Babka /* 119648731c84SMel Gorman * Update the cached scanner pfn once the pageblock has been scanned. 1197804d3121SMel Gorman * Pages will either be migrated in which case there is no point 1198804d3121SMel Gorman * scanning in the near future or migration failed in which case the 1199804d3121SMel Gorman * failure reason may persist. The block is marked for skipping if 1200804d3121SMel Gorman * there were no pages isolated in the block or if the block is 1201804d3121SMel Gorman * rescanned twice in a row. 120250b5b094SVlastimil Babka */ 120348731c84SMel Gorman if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) { 12048b71b499SBaolin Wang if (!cc->no_set_skip_hint && valid_page && !skip_updated) 1205e380bebeSMel Gorman set_pageblock_skip(valid_page); 1206e380bebeSMel Gorman update_cached_migrate(cc, low_pfn); 1207e380bebeSMel Gorman } 1208bb13ffebSMel Gorman 1209e34d85f0SJoonsoo Kim trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn, 1210e34d85f0SJoonsoo Kim nr_scanned, nr_isolated); 1211b7aba698SMel Gorman 1212670105a2SMel Gorman fatal_pending: 12137f354a54SDavid Rientjes cc->total_migrate_scanned += nr_scanned; 1214397487dbSMel Gorman if (nr_isolated) 1215010fc29aSMinchan Kim count_compact_events(COMPACTISOLATED, nr_isolated); 1216397487dbSMel Gorman 1217c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn; 1218c2ad7a1fSOscar Salvador 1219c2ad7a1fSOscar Salvador return ret; 12202fe86e00SMichal Nazarewicz } 12212fe86e00SMichal Nazarewicz 1222edc2ca61SVlastimil Babka /** 1223edc2ca61SVlastimil Babka * isolate_migratepages_range() - isolate migrate-able pages in a PFN range 1224edc2ca61SVlastimil Babka * @cc: Compaction control structure. 1225edc2ca61SVlastimil Babka * @start_pfn: The first PFN to start isolating. 1226edc2ca61SVlastimil Babka * @end_pfn: The one-past-last PFN. 1227edc2ca61SVlastimil Babka * 1228369fa227SOscar Salvador * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM 1229369fa227SOscar Salvador * in case we could not allocate a page, or 0. 1230edc2ca61SVlastimil Babka */ 1231c2ad7a1fSOscar Salvador int 1232edc2ca61SVlastimil Babka isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn, 1233edc2ca61SVlastimil Babka unsigned long end_pfn) 1234edc2ca61SVlastimil Babka { 1235e1409c32SJoonsoo Kim unsigned long pfn, block_start_pfn, block_end_pfn; 1236c2ad7a1fSOscar Salvador int ret = 0; 1237edc2ca61SVlastimil Babka 1238edc2ca61SVlastimil Babka /* Scan block by block. First and last block may be incomplete */ 1239edc2ca61SVlastimil Babka pfn = start_pfn; 124006b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(pfn); 1241e1409c32SJoonsoo Kim if (block_start_pfn < cc->zone->zone_start_pfn) 1242e1409c32SJoonsoo Kim block_start_pfn = cc->zone->zone_start_pfn; 124306b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(pfn); 1244edc2ca61SVlastimil Babka 1245edc2ca61SVlastimil Babka for (; pfn < end_pfn; pfn = block_end_pfn, 1246e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 1247edc2ca61SVlastimil Babka block_end_pfn += pageblock_nr_pages) { 1248edc2ca61SVlastimil Babka 1249edc2ca61SVlastimil Babka block_end_pfn = min(block_end_pfn, end_pfn); 1250edc2ca61SVlastimil Babka 1251e1409c32SJoonsoo Kim if (!pageblock_pfn_to_page(block_start_pfn, 1252e1409c32SJoonsoo Kim block_end_pfn, cc->zone)) 1253edc2ca61SVlastimil Babka continue; 1254edc2ca61SVlastimil Babka 1255c2ad7a1fSOscar Salvador ret = isolate_migratepages_block(cc, pfn, block_end_pfn, 1256edc2ca61SVlastimil Babka ISOLATE_UNEVICTABLE); 1257edc2ca61SVlastimil Babka 1258c2ad7a1fSOscar Salvador if (ret) 1259edc2ca61SVlastimil Babka break; 12606ea41c0cSJoonsoo Kim 126138935861SZi Yan if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX) 12626ea41c0cSJoonsoo Kim break; 1263edc2ca61SVlastimil Babka } 1264edc2ca61SVlastimil Babka 1265c2ad7a1fSOscar Salvador return ret; 1266edc2ca61SVlastimil Babka } 1267edc2ca61SVlastimil Babka 1268ff9543fdSMichal Nazarewicz #endif /* CONFIG_COMPACTION || CONFIG_CMA */ 1269ff9543fdSMichal Nazarewicz #ifdef CONFIG_COMPACTION 1270018e9a49SAndrew Morton 1271b682debdSVlastimil Babka static bool suitable_migration_source(struct compact_control *cc, 1272b682debdSVlastimil Babka struct page *page) 1273b682debdSVlastimil Babka { 1274282722b0SVlastimil Babka int block_mt; 1275282722b0SVlastimil Babka 12769bebefd5SMel Gorman if (pageblock_skip_persistent(page)) 12779bebefd5SMel Gorman return false; 12789bebefd5SMel Gorman 1279282722b0SVlastimil Babka if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction) 1280b682debdSVlastimil Babka return true; 1281b682debdSVlastimil Babka 1282282722b0SVlastimil Babka block_mt = get_pageblock_migratetype(page); 1283282722b0SVlastimil Babka 1284282722b0SVlastimil Babka if (cc->migratetype == MIGRATE_MOVABLE) 1285282722b0SVlastimil Babka return is_migrate_movable(block_mt); 1286282722b0SVlastimil Babka else 1287282722b0SVlastimil Babka return block_mt == cc->migratetype; 1288b682debdSVlastimil Babka } 1289b682debdSVlastimil Babka 1290018e9a49SAndrew Morton /* Returns true if the page is within a block suitable for migration to */ 12919f7e3387SVlastimil Babka static bool suitable_migration_target(struct compact_control *cc, 12929f7e3387SVlastimil Babka struct page *page) 1293018e9a49SAndrew Morton { 1294018e9a49SAndrew Morton /* If the page is a large free page, then disallow migration */ 1295018e9a49SAndrew Morton if (PageBuddy(page)) { 1296018e9a49SAndrew Morton /* 1297018e9a49SAndrew Morton * We are checking page_order without zone->lock taken. But 1298018e9a49SAndrew Morton * the only small danger is that we skip a potentially suitable 1299018e9a49SAndrew Morton * pageblock, so it's not worth to check order for valid range. 1300018e9a49SAndrew Morton */ 1301ab130f91SMatthew Wilcox (Oracle) if (buddy_order_unsafe(page) >= pageblock_order) 1302018e9a49SAndrew Morton return false; 1303018e9a49SAndrew Morton } 1304018e9a49SAndrew Morton 13051ef36db2SYisheng Xie if (cc->ignore_block_suitable) 13061ef36db2SYisheng Xie return true; 13071ef36db2SYisheng Xie 1308018e9a49SAndrew Morton /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */ 1309b682debdSVlastimil Babka if (is_migrate_movable(get_pageblock_migratetype(page))) 1310018e9a49SAndrew Morton return true; 1311018e9a49SAndrew Morton 1312018e9a49SAndrew Morton /* Otherwise skip the block */ 1313018e9a49SAndrew Morton return false; 1314018e9a49SAndrew Morton } 1315018e9a49SAndrew Morton 131670b44595SMel Gorman static inline unsigned int 131770b44595SMel Gorman freelist_scan_limit(struct compact_control *cc) 131870b44595SMel Gorman { 1319dd7ef7bdSQian Cai unsigned short shift = BITS_PER_LONG - 1; 1320dd7ef7bdSQian Cai 1321dd7ef7bdSQian Cai return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1; 132270b44595SMel Gorman } 132370b44595SMel Gorman 1324ff9543fdSMichal Nazarewicz /* 1325f2849aa0SVlastimil Babka * Test whether the free scanner has reached the same or lower pageblock than 1326f2849aa0SVlastimil Babka * the migration scanner, and compaction should thus terminate. 1327f2849aa0SVlastimil Babka */ 1328f2849aa0SVlastimil Babka static inline bool compact_scanners_met(struct compact_control *cc) 1329f2849aa0SVlastimil Babka { 1330f2849aa0SVlastimil Babka return (cc->free_pfn >> pageblock_order) 1331f2849aa0SVlastimil Babka <= (cc->migrate_pfn >> pageblock_order); 1332f2849aa0SVlastimil Babka } 1333f2849aa0SVlastimil Babka 13345a811889SMel Gorman /* 13355a811889SMel Gorman * Used when scanning for a suitable migration target which scans freelists 13365a811889SMel Gorman * in reverse. Reorders the list such as the unscanned pages are scanned 13375a811889SMel Gorman * first on the next iteration of the free scanner 13385a811889SMel Gorman */ 13395a811889SMel Gorman static void 13405a811889SMel Gorman move_freelist_head(struct list_head *freelist, struct page *freepage) 13415a811889SMel Gorman { 13425a811889SMel Gorman LIST_HEAD(sublist); 13435a811889SMel Gorman 13445a811889SMel Gorman if (!list_is_last(freelist, &freepage->lru)) { 13455a811889SMel Gorman list_cut_before(&sublist, freelist, &freepage->lru); 13465a811889SMel Gorman list_splice_tail(&sublist, freelist); 13475a811889SMel Gorman } 13485a811889SMel Gorman } 13495a811889SMel Gorman 13505a811889SMel Gorman /* 13515a811889SMel Gorman * Similar to move_freelist_head except used by the migration scanner 13525a811889SMel Gorman * when scanning forward. It's possible for these list operations to 13535a811889SMel Gorman * move against each other if they search the free list exactly in 13545a811889SMel Gorman * lockstep. 13555a811889SMel Gorman */ 135670b44595SMel Gorman static void 135770b44595SMel Gorman move_freelist_tail(struct list_head *freelist, struct page *freepage) 135870b44595SMel Gorman { 135970b44595SMel Gorman LIST_HEAD(sublist); 136070b44595SMel Gorman 136170b44595SMel Gorman if (!list_is_first(freelist, &freepage->lru)) { 136270b44595SMel Gorman list_cut_position(&sublist, freelist, &freepage->lru); 136370b44595SMel Gorman list_splice_tail(&sublist, freelist); 136470b44595SMel Gorman } 136570b44595SMel Gorman } 136670b44595SMel Gorman 13675a811889SMel Gorman static void 1368be21b32aSNARIBAYASHI Akira fast_isolate_around(struct compact_control *cc, unsigned long pfn) 13695a811889SMel Gorman { 13705a811889SMel Gorman unsigned long start_pfn, end_pfn; 13716e2b7044SVlastimil Babka struct page *page; 13725a811889SMel Gorman 13735a811889SMel Gorman /* Do not search around if there are enough pages already */ 13745a811889SMel Gorman if (cc->nr_freepages >= cc->nr_migratepages) 13755a811889SMel Gorman return; 13765a811889SMel Gorman 13775a811889SMel Gorman /* Minimise scanning during async compaction */ 13785a811889SMel Gorman if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC) 13795a811889SMel Gorman return; 13805a811889SMel Gorman 13815a811889SMel Gorman /* Pageblock boundaries */ 13826e2b7044SVlastimil Babka start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn); 13836e2b7044SVlastimil Babka end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)); 13846e2b7044SVlastimil Babka 13856e2b7044SVlastimil Babka page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone); 13866e2b7044SVlastimil Babka if (!page) 13876e2b7044SVlastimil Babka return; 13885a811889SMel Gorman 13894fca9730SMel Gorman isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false); 13905a811889SMel Gorman 13915a811889SMel Gorman /* Skip this pageblock in the future as it's full or nearly full */ 1392cf650342SBaolin Wang if (start_pfn == end_pfn) 13935a811889SMel Gorman set_pageblock_skip(page); 1394be21b32aSNARIBAYASHI Akira 1395be21b32aSNARIBAYASHI Akira return; 13965a811889SMel Gorman } 13975a811889SMel Gorman 1398dbe2d4e4SMel Gorman /* Search orders in round-robin fashion */ 1399dbe2d4e4SMel Gorman static int next_search_order(struct compact_control *cc, int order) 1400dbe2d4e4SMel Gorman { 1401dbe2d4e4SMel Gorman order--; 1402dbe2d4e4SMel Gorman if (order < 0) 1403dbe2d4e4SMel Gorman order = cc->order - 1; 1404dbe2d4e4SMel Gorman 1405dbe2d4e4SMel Gorman /* Search wrapped around? */ 1406dbe2d4e4SMel Gorman if (order == cc->search_order) { 1407dbe2d4e4SMel Gorman cc->search_order--; 1408dbe2d4e4SMel Gorman if (cc->search_order < 0) 1409dbe2d4e4SMel Gorman cc->search_order = cc->order - 1; 1410dbe2d4e4SMel Gorman return -1; 1411dbe2d4e4SMel Gorman } 1412dbe2d4e4SMel Gorman 1413dbe2d4e4SMel Gorman return order; 1414dbe2d4e4SMel Gorman } 1415dbe2d4e4SMel Gorman 14162dbd9005SBaolin Wang static void fast_isolate_freepages(struct compact_control *cc) 14175a811889SMel Gorman { 1418b55ca526SWonhyuk Yang unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1); 1419*447ba886SBaolin Wang unsigned int nr_scanned = 0, total_isolated = 0; 142074e21484SRokudo Yan unsigned long low_pfn, min_pfn, highest = 0; 14215a811889SMel Gorman unsigned long nr_isolated = 0; 14225a811889SMel Gorman unsigned long distance; 14235a811889SMel Gorman struct page *page = NULL; 14245a811889SMel Gorman bool scan_start = false; 14255a811889SMel Gorman int order; 14265a811889SMel Gorman 14275a811889SMel Gorman /* Full compaction passes in a negative order */ 14285a811889SMel Gorman if (cc->order <= 0) 14292dbd9005SBaolin Wang return; 14305a811889SMel Gorman 14315a811889SMel Gorman /* 14325a811889SMel Gorman * If starting the scan, use a deeper search and use the highest 14335a811889SMel Gorman * PFN found if a suitable one is not found. 14345a811889SMel Gorman */ 1435e332f741SMel Gorman if (cc->free_pfn >= cc->zone->compact_init_free_pfn) { 14365a811889SMel Gorman limit = pageblock_nr_pages >> 1; 14375a811889SMel Gorman scan_start = true; 14385a811889SMel Gorman } 14395a811889SMel Gorman 14405a811889SMel Gorman /* 14415a811889SMel Gorman * Preferred point is in the top quarter of the scan space but take 14425a811889SMel Gorman * a pfn from the top half if the search is problematic. 14435a811889SMel Gorman */ 14445a811889SMel Gorman distance = (cc->free_pfn - cc->migrate_pfn); 14455a811889SMel Gorman low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2)); 14465a811889SMel Gorman min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1)); 14475a811889SMel Gorman 14485a811889SMel Gorman if (WARN_ON_ONCE(min_pfn > low_pfn)) 14495a811889SMel Gorman low_pfn = min_pfn; 14505a811889SMel Gorman 1451dbe2d4e4SMel Gorman /* 1452dbe2d4e4SMel Gorman * Search starts from the last successful isolation order or the next 1453dbe2d4e4SMel Gorman * order to search after a previous failure 1454dbe2d4e4SMel Gorman */ 1455dbe2d4e4SMel Gorman cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order); 1456dbe2d4e4SMel Gorman 1457dbe2d4e4SMel Gorman for (order = cc->search_order; 1458dbe2d4e4SMel Gorman !page && order >= 0; 1459dbe2d4e4SMel Gorman order = next_search_order(cc, order)) { 14605a811889SMel Gorman struct free_area *area = &cc->zone->free_area[order]; 14615a811889SMel Gorman struct list_head *freelist; 14625a811889SMel Gorman struct page *freepage; 14635a811889SMel Gorman unsigned long flags; 14645a811889SMel Gorman unsigned int order_scanned = 0; 146574e21484SRokudo Yan unsigned long high_pfn = 0; 14665a811889SMel Gorman 14675a811889SMel Gorman if (!area->nr_free) 14685a811889SMel Gorman continue; 14695a811889SMel Gorman 14705a811889SMel Gorman spin_lock_irqsave(&cc->zone->lock, flags); 14715a811889SMel Gorman freelist = &area->free_list[MIGRATE_MOVABLE]; 14725a811889SMel Gorman list_for_each_entry_reverse(freepage, freelist, lru) { 14735a811889SMel Gorman unsigned long pfn; 14745a811889SMel Gorman 14755a811889SMel Gorman order_scanned++; 14765a811889SMel Gorman nr_scanned++; 14775a811889SMel Gorman pfn = page_to_pfn(freepage); 14785a811889SMel Gorman 14795a811889SMel Gorman if (pfn >= highest) 14806e2b7044SVlastimil Babka highest = max(pageblock_start_pfn(pfn), 14816e2b7044SVlastimil Babka cc->zone->zone_start_pfn); 14825a811889SMel Gorman 14835a811889SMel Gorman if (pfn >= low_pfn) { 14845a811889SMel Gorman cc->fast_search_fail = 0; 1485dbe2d4e4SMel Gorman cc->search_order = order; 14865a811889SMel Gorman page = freepage; 14875a811889SMel Gorman break; 14885a811889SMel Gorman } 14895a811889SMel Gorman 14905a811889SMel Gorman if (pfn >= min_pfn && pfn > high_pfn) { 14915a811889SMel Gorman high_pfn = pfn; 14925a811889SMel Gorman 14935a811889SMel Gorman /* Shorten the scan if a candidate is found */ 14945a811889SMel Gorman limit >>= 1; 14955a811889SMel Gorman } 14965a811889SMel Gorman 14975a811889SMel Gorman if (order_scanned >= limit) 14985a811889SMel Gorman break; 14995a811889SMel Gorman } 15005a811889SMel Gorman 15015a811889SMel Gorman /* Use a minimum pfn if a preferred one was not found */ 15025a811889SMel Gorman if (!page && high_pfn) { 15035a811889SMel Gorman page = pfn_to_page(high_pfn); 15045a811889SMel Gorman 15055a811889SMel Gorman /* Update freepage for the list reorder below */ 15065a811889SMel Gorman freepage = page; 15075a811889SMel Gorman } 15085a811889SMel Gorman 15095a811889SMel Gorman /* Reorder to so a future search skips recent pages */ 15105a811889SMel Gorman move_freelist_head(freelist, freepage); 15115a811889SMel Gorman 15125a811889SMel Gorman /* Isolate the page if available */ 15135a811889SMel Gorman if (page) { 15145a811889SMel Gorman if (__isolate_free_page(page, order)) { 15155a811889SMel Gorman set_page_private(page, order); 15165a811889SMel Gorman nr_isolated = 1 << order; 1517b717d6b9SWilliam Lam nr_scanned += nr_isolated - 1; 1518*447ba886SBaolin Wang total_isolated += nr_isolated; 15195a811889SMel Gorman cc->nr_freepages += nr_isolated; 15205a811889SMel Gorman list_add_tail(&page->lru, &cc->freepages); 15215a811889SMel Gorman count_compact_events(COMPACTISOLATED, nr_isolated); 15225a811889SMel Gorman } else { 15235a811889SMel Gorman /* If isolation fails, abort the search */ 15245b56d996SQian Cai order = cc->search_order + 1; 15255a811889SMel Gorman page = NULL; 15265a811889SMel Gorman } 15275a811889SMel Gorman } 15285a811889SMel Gorman 15295a811889SMel Gorman spin_unlock_irqrestore(&cc->zone->lock, flags); 15305a811889SMel Gorman 15315a811889SMel Gorman /* 1532b55ca526SWonhyuk Yang * Smaller scan on next order so the total scan is related 15335a811889SMel Gorman * to freelist_scan_limit. 15345a811889SMel Gorman */ 15355a811889SMel Gorman if (order_scanned >= limit) 1536b55ca526SWonhyuk Yang limit = max(1U, limit >> 1); 15375a811889SMel Gorman } 15385a811889SMel Gorman 1539*447ba886SBaolin Wang trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn, 1540*447ba886SBaolin Wang nr_scanned, total_isolated); 1541*447ba886SBaolin Wang 15425a811889SMel Gorman if (!page) { 15435a811889SMel Gorman cc->fast_search_fail++; 15445a811889SMel Gorman if (scan_start) { 15455a811889SMel Gorman /* 15465a811889SMel Gorman * Use the highest PFN found above min. If one was 1547f3867755SEthon Paul * not found, be pessimistic for direct compaction 15485a811889SMel Gorman * and use the min mark. 15495a811889SMel Gorman */ 1550ca2864e5SMiaohe Lin if (highest >= min_pfn) { 15515a811889SMel Gorman page = pfn_to_page(highest); 15525a811889SMel Gorman cc->free_pfn = highest; 15535a811889SMel Gorman } else { 1554e577c8b6SSuzuki K Poulose if (cc->direct_compaction && pfn_valid(min_pfn)) { 155573a6e474SBaoquan He page = pageblock_pfn_to_page(min_pfn, 15566e2b7044SVlastimil Babka min(pageblock_end_pfn(min_pfn), 15576e2b7044SVlastimil Babka zone_end_pfn(cc->zone)), 155873a6e474SBaoquan He cc->zone); 15595a811889SMel Gorman cc->free_pfn = min_pfn; 15605a811889SMel Gorman } 15615a811889SMel Gorman } 15625a811889SMel Gorman } 15635a811889SMel Gorman } 15645a811889SMel Gorman 1565d097a6f6SMel Gorman if (highest && highest >= cc->zone->compact_cached_free_pfn) { 1566d097a6f6SMel Gorman highest -= pageblock_nr_pages; 15675a811889SMel Gorman cc->zone->compact_cached_free_pfn = highest; 1568d097a6f6SMel Gorman } 15695a811889SMel Gorman 15705a811889SMel Gorman cc->total_free_scanned += nr_scanned; 15715a811889SMel Gorman if (!page) 15722dbd9005SBaolin Wang return; 15735a811889SMel Gorman 15745a811889SMel Gorman low_pfn = page_to_pfn(page); 1575be21b32aSNARIBAYASHI Akira fast_isolate_around(cc, low_pfn); 15765a811889SMel Gorman } 15775a811889SMel Gorman 1578f2849aa0SVlastimil Babka /* 1579ff9543fdSMichal Nazarewicz * Based on information in the current compact_control, find blocks 1580ff9543fdSMichal Nazarewicz * suitable for isolating free pages from and then isolate them. 1581ff9543fdSMichal Nazarewicz */ 1582edc2ca61SVlastimil Babka static void isolate_freepages(struct compact_control *cc) 1583ff9543fdSMichal Nazarewicz { 1584edc2ca61SVlastimil Babka struct zone *zone = cc->zone; 1585ff9543fdSMichal Nazarewicz struct page *page; 1586c96b9e50SVlastimil Babka unsigned long block_start_pfn; /* start of current pageblock */ 1587e14c720eSVlastimil Babka unsigned long isolate_start_pfn; /* exact pfn we start at */ 1588c96b9e50SVlastimil Babka unsigned long block_end_pfn; /* end of current pageblock */ 1589c96b9e50SVlastimil Babka unsigned long low_pfn; /* lowest pfn scanner is able to scan */ 1590ff9543fdSMichal Nazarewicz struct list_head *freelist = &cc->freepages; 15914fca9730SMel Gorman unsigned int stride; 15922fe86e00SMichal Nazarewicz 15935a811889SMel Gorman /* Try a small search of the free lists for a candidate */ 159400bc102fSMiaohe Lin fast_isolate_freepages(cc); 15955a811889SMel Gorman if (cc->nr_freepages) 15965a811889SMel Gorman goto splitmap; 15975a811889SMel Gorman 1598ff9543fdSMichal Nazarewicz /* 1599ff9543fdSMichal Nazarewicz * Initialise the free scanner. The starting point is where we last 160049e068f0SVlastimil Babka * successfully isolated from, zone-cached value, or the end of the 1601e14c720eSVlastimil Babka * zone when isolating for the first time. For looping we also need 1602e14c720eSVlastimil Babka * this pfn aligned down to the pageblock boundary, because we do 1603c96b9e50SVlastimil Babka * block_start_pfn -= pageblock_nr_pages in the for loop. 1604c96b9e50SVlastimil Babka * For ending point, take care when isolating in last pageblock of a 1605a1c1dbebSRandy Dunlap * zone which ends in the middle of a pageblock. 160649e068f0SVlastimil Babka * The low boundary is the end of the pageblock the migration scanner 160749e068f0SVlastimil Babka * is using. 1608ff9543fdSMichal Nazarewicz */ 1609e14c720eSVlastimil Babka isolate_start_pfn = cc->free_pfn; 16105a811889SMel Gorman block_start_pfn = pageblock_start_pfn(isolate_start_pfn); 1611c96b9e50SVlastimil Babka block_end_pfn = min(block_start_pfn + pageblock_nr_pages, 1612c96b9e50SVlastimil Babka zone_end_pfn(zone)); 161306b6640aSVlastimil Babka low_pfn = pageblock_end_pfn(cc->migrate_pfn); 16144fca9730SMel Gorman stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1; 16152fe86e00SMichal Nazarewicz 1616ff9543fdSMichal Nazarewicz /* 1617ff9543fdSMichal Nazarewicz * Isolate free pages until enough are available to migrate the 1618ff9543fdSMichal Nazarewicz * pages on cc->migratepages. We stop searching if the migrate 1619ff9543fdSMichal Nazarewicz * and free page scanners meet or enough free pages are isolated. 1620ff9543fdSMichal Nazarewicz */ 1621f5f61a32SVlastimil Babka for (; block_start_pfn >= low_pfn; 1622c96b9e50SVlastimil Babka block_end_pfn = block_start_pfn, 1623e14c720eSVlastimil Babka block_start_pfn -= pageblock_nr_pages, 1624e14c720eSVlastimil Babka isolate_start_pfn = block_start_pfn) { 16254fca9730SMel Gorman unsigned long nr_isolated; 16264fca9730SMel Gorman 1627f6ea3adbSDavid Rientjes /* 1628f6ea3adbSDavid Rientjes * This can iterate a massively long zone without finding any 1629cb810ad2SMel Gorman * suitable migration targets, so periodically check resched. 1630f6ea3adbSDavid Rientjes */ 1631c036ddffSMiaohe Lin if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) 1632cf66f070SMel Gorman cond_resched(); 1633f6ea3adbSDavid Rientjes 16347d49d886SVlastimil Babka page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn, 16357d49d886SVlastimil Babka zone); 16367d49d886SVlastimil Babka if (!page) 1637ff9543fdSMichal Nazarewicz continue; 1638ff9543fdSMichal Nazarewicz 1639ff9543fdSMichal Nazarewicz /* Check the block is suitable for migration */ 16409f7e3387SVlastimil Babka if (!suitable_migration_target(cc, page)) 1641ff9543fdSMichal Nazarewicz continue; 164268e3e926SLinus Torvalds 1643bb13ffebSMel Gorman /* If isolation recently failed, do not retry */ 1644bb13ffebSMel Gorman if (!isolation_suitable(cc, page)) 1645bb13ffebSMel Gorman continue; 1646bb13ffebSMel Gorman 1647e14c720eSVlastimil Babka /* Found a block suitable for isolating free pages from. */ 16484fca9730SMel Gorman nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn, 16494fca9730SMel Gorman block_end_pfn, freelist, stride, false); 1650ff9543fdSMichal Nazarewicz 1651d097a6f6SMel Gorman /* Update the skip hint if the full pageblock was scanned */ 1652d097a6f6SMel Gorman if (isolate_start_pfn == block_end_pfn) 1653d097a6f6SMel Gorman update_pageblock_skip(cc, page, block_start_pfn); 1654d097a6f6SMel Gorman 1655cb2dcaf0SMel Gorman /* Are enough freepages isolated? */ 1656cb2dcaf0SMel Gorman if (cc->nr_freepages >= cc->nr_migratepages) { 1657a46cbf3bSDavid Rientjes if (isolate_start_pfn >= block_end_pfn) { 1658a46cbf3bSDavid Rientjes /* 1659a46cbf3bSDavid Rientjes * Restart at previous pageblock if more 1660a46cbf3bSDavid Rientjes * freepages can be isolated next time. 1661a46cbf3bSDavid Rientjes */ 1662f5f61a32SVlastimil Babka isolate_start_pfn = 1663e14c720eSVlastimil Babka block_start_pfn - pageblock_nr_pages; 1664a46cbf3bSDavid Rientjes } 1665be976572SVlastimil Babka break; 1666a46cbf3bSDavid Rientjes } else if (isolate_start_pfn < block_end_pfn) { 1667f5f61a32SVlastimil Babka /* 1668a46cbf3bSDavid Rientjes * If isolation failed early, do not continue 1669a46cbf3bSDavid Rientjes * needlessly. 1670f5f61a32SVlastimil Babka */ 1671a46cbf3bSDavid Rientjes break; 1672f5f61a32SVlastimil Babka } 16734fca9730SMel Gorman 16744fca9730SMel Gorman /* Adjust stride depending on isolation */ 16754fca9730SMel Gorman if (nr_isolated) { 16764fca9730SMel Gorman stride = 1; 16774fca9730SMel Gorman continue; 16784fca9730SMel Gorman } 16794fca9730SMel Gorman stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1); 1680c89511abSMel Gorman } 1681ff9543fdSMichal Nazarewicz 16827ed695e0SVlastimil Babka /* 1683f5f61a32SVlastimil Babka * Record where the free scanner will restart next time. Either we 1684f5f61a32SVlastimil Babka * broke from the loop and set isolate_start_pfn based on the last 1685f5f61a32SVlastimil Babka * call to isolate_freepages_block(), or we met the migration scanner 1686f5f61a32SVlastimil Babka * and the loop terminated due to isolate_start_pfn < low_pfn 16877ed695e0SVlastimil Babka */ 1688f5f61a32SVlastimil Babka cc->free_pfn = isolate_start_pfn; 16895a811889SMel Gorman 16905a811889SMel Gorman splitmap: 16915a811889SMel Gorman /* __isolate_free_page() does not map the pages */ 16925a811889SMel Gorman split_map_pages(freelist); 1693748446bbSMel Gorman } 1694748446bbSMel Gorman 1695748446bbSMel Gorman /* 1696748446bbSMel Gorman * This is a migrate-callback that "allocates" freepages by taking pages 1697748446bbSMel Gorman * from the isolated freelists in the block we are migrating to. 1698748446bbSMel Gorman */ 16994e096ae1SMatthew Wilcox (Oracle) static struct folio *compaction_alloc(struct folio *src, unsigned long data) 1700748446bbSMel Gorman { 1701748446bbSMel Gorman struct compact_control *cc = (struct compact_control *)data; 17024e096ae1SMatthew Wilcox (Oracle) struct folio *dst; 1703748446bbSMel Gorman 1704748446bbSMel Gorman if (list_empty(&cc->freepages)) { 1705edc2ca61SVlastimil Babka isolate_freepages(cc); 1706748446bbSMel Gorman 1707748446bbSMel Gorman if (list_empty(&cc->freepages)) 1708748446bbSMel Gorman return NULL; 1709748446bbSMel Gorman } 1710748446bbSMel Gorman 17114e096ae1SMatthew Wilcox (Oracle) dst = list_entry(cc->freepages.next, struct folio, lru); 17124e096ae1SMatthew Wilcox (Oracle) list_del(&dst->lru); 1713748446bbSMel Gorman cc->nr_freepages--; 1714748446bbSMel Gorman 17154e096ae1SMatthew Wilcox (Oracle) return dst; 1716748446bbSMel Gorman } 1717748446bbSMel Gorman 1718748446bbSMel Gorman /* 1719d53aea3dSDavid Rientjes * This is a migrate-callback that "frees" freepages back to the isolated 1720d53aea3dSDavid Rientjes * freelist. All pages on the freelist are from the same zone, so there is no 1721d53aea3dSDavid Rientjes * special handling needed for NUMA. 1722d53aea3dSDavid Rientjes */ 17234e096ae1SMatthew Wilcox (Oracle) static void compaction_free(struct folio *dst, unsigned long data) 1724d53aea3dSDavid Rientjes { 1725d53aea3dSDavid Rientjes struct compact_control *cc = (struct compact_control *)data; 1726d53aea3dSDavid Rientjes 17274e096ae1SMatthew Wilcox (Oracle) list_add(&dst->lru, &cc->freepages); 1728d53aea3dSDavid Rientjes cc->nr_freepages++; 1729d53aea3dSDavid Rientjes } 1730d53aea3dSDavid Rientjes 1731ff9543fdSMichal Nazarewicz /* possible outcome of isolate_migratepages */ 1732ff9543fdSMichal Nazarewicz typedef enum { 1733ff9543fdSMichal Nazarewicz ISOLATE_ABORT, /* Abort compaction now */ 1734ff9543fdSMichal Nazarewicz ISOLATE_NONE, /* No pages isolated, continue scanning */ 1735ff9543fdSMichal Nazarewicz ISOLATE_SUCCESS, /* Pages isolated, migrate */ 1736ff9543fdSMichal Nazarewicz } isolate_migrate_t; 1737ff9543fdSMichal Nazarewicz 1738ff9543fdSMichal Nazarewicz /* 17395bbe3547SEric B Munson * Allow userspace to control policy on scanning the unevictable LRU for 17405bbe3547SEric B Munson * compactable pages. 17415bbe3547SEric B Munson */ 174248fe8ab8SMinghao Chi static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT; 174348fe8ab8SMinghao Chi /* 174448fe8ab8SMinghao Chi * Tunable for proactive compaction. It determines how 174548fe8ab8SMinghao Chi * aggressively the kernel should compact memory in the 174648fe8ab8SMinghao Chi * background. It takes values in the range [0, 100]. 174748fe8ab8SMinghao Chi */ 174848fe8ab8SMinghao Chi static unsigned int __read_mostly sysctl_compaction_proactiveness = 20; 174948fe8ab8SMinghao Chi static int sysctl_extfrag_threshold = 500; 17508b9167cdSWen Yang static int __read_mostly sysctl_compact_memory; 17515bbe3547SEric B Munson 175270b44595SMel Gorman static inline void 175370b44595SMel Gorman update_fast_start_pfn(struct compact_control *cc, unsigned long pfn) 175470b44595SMel Gorman { 175570b44595SMel Gorman if (cc->fast_start_pfn == ULONG_MAX) 175670b44595SMel Gorman return; 175770b44595SMel Gorman 175870b44595SMel Gorman if (!cc->fast_start_pfn) 175970b44595SMel Gorman cc->fast_start_pfn = pfn; 176070b44595SMel Gorman 176170b44595SMel Gorman cc->fast_start_pfn = min(cc->fast_start_pfn, pfn); 176270b44595SMel Gorman } 176370b44595SMel Gorman 176470b44595SMel Gorman static inline unsigned long 176570b44595SMel Gorman reinit_migrate_pfn(struct compact_control *cc) 176670b44595SMel Gorman { 176770b44595SMel Gorman if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX) 176870b44595SMel Gorman return cc->migrate_pfn; 176970b44595SMel Gorman 177070b44595SMel Gorman cc->migrate_pfn = cc->fast_start_pfn; 177170b44595SMel Gorman cc->fast_start_pfn = ULONG_MAX; 177270b44595SMel Gorman 177370b44595SMel Gorman return cc->migrate_pfn; 177470b44595SMel Gorman } 177570b44595SMel Gorman 177670b44595SMel Gorman /* 177770b44595SMel Gorman * Briefly search the free lists for a migration source that already has 177870b44595SMel Gorman * some free pages to reduce the number of pages that need migration 177970b44595SMel Gorman * before a pageblock is free. 178070b44595SMel Gorman */ 178170b44595SMel Gorman static unsigned long fast_find_migrateblock(struct compact_control *cc) 178270b44595SMel Gorman { 178370b44595SMel Gorman unsigned int limit = freelist_scan_limit(cc); 178470b44595SMel Gorman unsigned int nr_scanned = 0; 178570b44595SMel Gorman unsigned long distance; 178670b44595SMel Gorman unsigned long pfn = cc->migrate_pfn; 178770b44595SMel Gorman unsigned long high_pfn; 178870b44595SMel Gorman int order; 178915d28d0dSWonhyuk Yang bool found_block = false; 179070b44595SMel Gorman 179170b44595SMel Gorman /* Skip hints are relied on to avoid repeats on the fast search */ 179270b44595SMel Gorman if (cc->ignore_skip_hint) 179370b44595SMel Gorman return pfn; 179470b44595SMel Gorman 179570b44595SMel Gorman /* 1796f9d7fc1aSMel Gorman * If the pageblock should be finished then do not select a different 1797f9d7fc1aSMel Gorman * pageblock. 1798f9d7fc1aSMel Gorman */ 1799f9d7fc1aSMel Gorman if (cc->finish_pageblock) 1800f9d7fc1aSMel Gorman return pfn; 1801f9d7fc1aSMel Gorman 1802f9d7fc1aSMel Gorman /* 180370b44595SMel Gorman * If the migrate_pfn is not at the start of a zone or the start 180470b44595SMel Gorman * of a pageblock then assume this is a continuation of a previous 180570b44595SMel Gorman * scan restarted due to COMPACT_CLUSTER_MAX. 180670b44595SMel Gorman */ 180770b44595SMel Gorman if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn)) 180870b44595SMel Gorman return pfn; 180970b44595SMel Gorman 181070b44595SMel Gorman /* 181170b44595SMel Gorman * For smaller orders, just linearly scan as the number of pages 181270b44595SMel Gorman * to migrate should be relatively small and does not necessarily 181370b44595SMel Gorman * justify freeing up a large block for a small allocation. 181470b44595SMel Gorman */ 181570b44595SMel Gorman if (cc->order <= PAGE_ALLOC_COSTLY_ORDER) 181670b44595SMel Gorman return pfn; 181770b44595SMel Gorman 181870b44595SMel Gorman /* 181970b44595SMel Gorman * Only allow kcompactd and direct requests for movable pages to 182070b44595SMel Gorman * quickly clear out a MOVABLE pageblock for allocation. This 182170b44595SMel Gorman * reduces the risk that a large movable pageblock is freed for 182270b44595SMel Gorman * an unmovable/reclaimable small allocation. 182370b44595SMel Gorman */ 182470b44595SMel Gorman if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE) 182570b44595SMel Gorman return pfn; 182670b44595SMel Gorman 182770b44595SMel Gorman /* 182870b44595SMel Gorman * When starting the migration scanner, pick any pageblock within the 182970b44595SMel Gorman * first half of the search space. Otherwise try and pick a pageblock 183070b44595SMel Gorman * within the first eighth to reduce the chances that a migration 183170b44595SMel Gorman * target later becomes a source. 183270b44595SMel Gorman */ 183370b44595SMel Gorman distance = (cc->free_pfn - cc->migrate_pfn) >> 1; 183470b44595SMel Gorman if (cc->migrate_pfn != cc->zone->zone_start_pfn) 183570b44595SMel Gorman distance >>= 2; 183670b44595SMel Gorman high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); 183770b44595SMel Gorman 183870b44595SMel Gorman for (order = cc->order - 1; 183915d28d0dSWonhyuk Yang order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; 184070b44595SMel Gorman order--) { 184170b44595SMel Gorman struct free_area *area = &cc->zone->free_area[order]; 184270b44595SMel Gorman struct list_head *freelist; 184370b44595SMel Gorman unsigned long flags; 184470b44595SMel Gorman struct page *freepage; 184570b44595SMel Gorman 184670b44595SMel Gorman if (!area->nr_free) 184770b44595SMel Gorman continue; 184870b44595SMel Gorman 184970b44595SMel Gorman spin_lock_irqsave(&cc->zone->lock, flags); 185070b44595SMel Gorman freelist = &area->free_list[MIGRATE_MOVABLE]; 185170b44595SMel Gorman list_for_each_entry(freepage, freelist, lru) { 185270b44595SMel Gorman unsigned long free_pfn; 185370b44595SMel Gorman 185415d28d0dSWonhyuk Yang if (nr_scanned++ >= limit) { 185515d28d0dSWonhyuk Yang move_freelist_tail(freelist, freepage); 185615d28d0dSWonhyuk Yang break; 185715d28d0dSWonhyuk Yang } 185815d28d0dSWonhyuk Yang 185970b44595SMel Gorman free_pfn = page_to_pfn(freepage); 186070b44595SMel Gorman if (free_pfn < high_pfn) { 186170b44595SMel Gorman /* 186270b44595SMel Gorman * Avoid if skipped recently. Ideally it would 186370b44595SMel Gorman * move to the tail but even safe iteration of 186470b44595SMel Gorman * the list assumes an entry is deleted, not 186570b44595SMel Gorman * reordered. 186670b44595SMel Gorman */ 186715d28d0dSWonhyuk Yang if (get_pageblock_skip(freepage)) 186870b44595SMel Gorman continue; 186970b44595SMel Gorman 187070b44595SMel Gorman /* Reorder to so a future search skips recent pages */ 187170b44595SMel Gorman move_freelist_tail(freelist, freepage); 187270b44595SMel Gorman 1873e380bebeSMel Gorman update_fast_start_pfn(cc, free_pfn); 187470b44595SMel Gorman pfn = pageblock_start_pfn(free_pfn); 1875bbe832b9SRei Yamamoto if (pfn < cc->zone->zone_start_pfn) 1876bbe832b9SRei Yamamoto pfn = cc->zone->zone_start_pfn; 187770b44595SMel Gorman cc->fast_search_fail = 0; 187815d28d0dSWonhyuk Yang found_block = true; 187970b44595SMel Gorman break; 188070b44595SMel Gorman } 188170b44595SMel Gorman } 188270b44595SMel Gorman spin_unlock_irqrestore(&cc->zone->lock, flags); 188370b44595SMel Gorman } 188470b44595SMel Gorman 188570b44595SMel Gorman cc->total_migrate_scanned += nr_scanned; 188670b44595SMel Gorman 188770b44595SMel Gorman /* 188870b44595SMel Gorman * If fast scanning failed then use a cached entry for a page block 188970b44595SMel Gorman * that had free pages as the basis for starting a linear scan. 189070b44595SMel Gorman */ 189115d28d0dSWonhyuk Yang if (!found_block) { 189215d28d0dSWonhyuk Yang cc->fast_search_fail++; 189370b44595SMel Gorman pfn = reinit_migrate_pfn(cc); 189415d28d0dSWonhyuk Yang } 189570b44595SMel Gorman return pfn; 189670b44595SMel Gorman } 189770b44595SMel Gorman 18985bbe3547SEric B Munson /* 1899edc2ca61SVlastimil Babka * Isolate all pages that can be migrated from the first suitable block, 1900edc2ca61SVlastimil Babka * starting at the block pointed to by the migrate scanner pfn within 1901edc2ca61SVlastimil Babka * compact_control. 1902ff9543fdSMichal Nazarewicz */ 190332aaf055SPengfei Li static isolate_migrate_t isolate_migratepages(struct compact_control *cc) 1904ff9543fdSMichal Nazarewicz { 1905e1409c32SJoonsoo Kim unsigned long block_start_pfn; 1906e1409c32SJoonsoo Kim unsigned long block_end_pfn; 1907e1409c32SJoonsoo Kim unsigned long low_pfn; 1908edc2ca61SVlastimil Babka struct page *page; 1909edc2ca61SVlastimil Babka const isolate_mode_t isolate_mode = 19105bbe3547SEric B Munson (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) | 19111d2047feSHugh Dickins (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0); 191270b44595SMel Gorman bool fast_find_block; 1913ff9543fdSMichal Nazarewicz 1914edc2ca61SVlastimil Babka /* 1915edc2ca61SVlastimil Babka * Start at where we last stopped, or beginning of the zone as 191670b44595SMel Gorman * initialized by compact_zone(). The first failure will use 191770b44595SMel Gorman * the lowest PFN as the starting point for linear scanning. 1918edc2ca61SVlastimil Babka */ 191970b44595SMel Gorman low_pfn = fast_find_migrateblock(cc); 192006b6640aSVlastimil Babka block_start_pfn = pageblock_start_pfn(low_pfn); 192132aaf055SPengfei Li if (block_start_pfn < cc->zone->zone_start_pfn) 192232aaf055SPengfei Li block_start_pfn = cc->zone->zone_start_pfn; 1923ff9543fdSMichal Nazarewicz 192470b44595SMel Gorman /* 192570b44595SMel Gorman * fast_find_migrateblock marks a pageblock skipped so to avoid 192670b44595SMel Gorman * the isolation_suitable check below, check whether the fast 192770b44595SMel Gorman * search was successful. 192870b44595SMel Gorman */ 192970b44595SMel Gorman fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail; 193070b44595SMel Gorman 1931ff9543fdSMichal Nazarewicz /* Only scan within a pageblock boundary */ 193206b6640aSVlastimil Babka block_end_pfn = pageblock_end_pfn(low_pfn); 1933ff9543fdSMichal Nazarewicz 1934edc2ca61SVlastimil Babka /* 1935edc2ca61SVlastimil Babka * Iterate over whole pageblocks until we find the first suitable. 1936edc2ca61SVlastimil Babka * Do not cross the free scanner. 1937edc2ca61SVlastimil Babka */ 1938e1409c32SJoonsoo Kim for (; block_end_pfn <= cc->free_pfn; 193970b44595SMel Gorman fast_find_block = false, 1940c2ad7a1fSOscar Salvador cc->migrate_pfn = low_pfn = block_end_pfn, 1941e1409c32SJoonsoo Kim block_start_pfn = block_end_pfn, 1942e1409c32SJoonsoo Kim block_end_pfn += pageblock_nr_pages) { 1943edc2ca61SVlastimil Babka 1944edc2ca61SVlastimil Babka /* 1945edc2ca61SVlastimil Babka * This can potentially iterate a massively long zone with 1946edc2ca61SVlastimil Babka * many pageblocks unsuitable, so periodically check if we 1947cb810ad2SMel Gorman * need to schedule. 1948edc2ca61SVlastimil Babka */ 1949c036ddffSMiaohe Lin if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages))) 1950cf66f070SMel Gorman cond_resched(); 1951edc2ca61SVlastimil Babka 195232aaf055SPengfei Li page = pageblock_pfn_to_page(block_start_pfn, 195332aaf055SPengfei Li block_end_pfn, cc->zone); 19547d49d886SVlastimil Babka if (!page) 1955edc2ca61SVlastimil Babka continue; 1956edc2ca61SVlastimil Babka 1957e380bebeSMel Gorman /* 1958e380bebeSMel Gorman * If isolation recently failed, do not retry. Only check the 1959e380bebeSMel Gorman * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock 1960e380bebeSMel Gorman * to be visited multiple times. Assume skip was checked 1961e380bebeSMel Gorman * before making it "skip" so other compaction instances do 1962e380bebeSMel Gorman * not scan the same block. 1963e380bebeSMel Gorman */ 1964ee0913c4SKefeng Wang if (pageblock_aligned(low_pfn) && 1965e380bebeSMel Gorman !fast_find_block && !isolation_suitable(cc, page)) 1966edc2ca61SVlastimil Babka continue; 1967edc2ca61SVlastimil Babka 1968edc2ca61SVlastimil Babka /* 1969556162bfSMiaohe Lin * For async direct compaction, only scan the pageblocks of the 1970556162bfSMiaohe Lin * same migratetype without huge pages. Async direct compaction 1971556162bfSMiaohe Lin * is optimistic to see if the minimum amount of work satisfies 1972556162bfSMiaohe Lin * the allocation. The cached PFN is updated as it's possible 1973556162bfSMiaohe Lin * that all remaining blocks between source and target are 1974556162bfSMiaohe Lin * unsuitable and the compaction scanners fail to meet. 1975edc2ca61SVlastimil Babka */ 19769bebefd5SMel Gorman if (!suitable_migration_source(cc, page)) { 19779bebefd5SMel Gorman update_cached_migrate(cc, block_end_pfn); 1978edc2ca61SVlastimil Babka continue; 19799bebefd5SMel Gorman } 1980ff9543fdSMichal Nazarewicz 1981ff9543fdSMichal Nazarewicz /* Perform the isolation */ 1982c2ad7a1fSOscar Salvador if (isolate_migratepages_block(cc, low_pfn, block_end_pfn, 1983c2ad7a1fSOscar Salvador isolate_mode)) 1984ff9543fdSMichal Nazarewicz return ISOLATE_ABORT; 1985ff9543fdSMichal Nazarewicz 1986edc2ca61SVlastimil Babka /* 1987edc2ca61SVlastimil Babka * Either we isolated something and proceed with migration. Or 1988edc2ca61SVlastimil Babka * we failed and compact_zone should decide if we should 1989edc2ca61SVlastimil Babka * continue or not. 1990edc2ca61SVlastimil Babka */ 1991edc2ca61SVlastimil Babka break; 1992edc2ca61SVlastimil Babka } 1993edc2ca61SVlastimil Babka 1994edc2ca61SVlastimil Babka return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; 1995ff9543fdSMichal Nazarewicz } 1996ff9543fdSMichal Nazarewicz 199721c527a3SYaowei Bai /* 199821c527a3SYaowei Bai * order == -1 is expected when compacting via 199921c527a3SYaowei Bai * /proc/sys/vm/compact_memory 200021c527a3SYaowei Bai */ 200121c527a3SYaowei Bai static inline bool is_via_compact_memory(int order) 200221c527a3SYaowei Bai { 200321c527a3SYaowei Bai return order == -1; 200421c527a3SYaowei Bai } 200521c527a3SYaowei Bai 2006b4a0215eSKefeng Wang /* 2007b4a0215eSKefeng Wang * Determine whether kswapd is (or recently was!) running on this node. 2008b4a0215eSKefeng Wang * 2009b4a0215eSKefeng Wang * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't 2010b4a0215eSKefeng Wang * zero it. 2011b4a0215eSKefeng Wang */ 2012facdaa91SNitin Gupta static bool kswapd_is_running(pg_data_t *pgdat) 2013facdaa91SNitin Gupta { 2014b4a0215eSKefeng Wang bool running; 2015b4a0215eSKefeng Wang 2016b4a0215eSKefeng Wang pgdat_kswapd_lock(pgdat); 2017b4a0215eSKefeng Wang running = pgdat->kswapd && task_is_running(pgdat->kswapd); 2018b4a0215eSKefeng Wang pgdat_kswapd_unlock(pgdat); 2019b4a0215eSKefeng Wang 2020b4a0215eSKefeng Wang return running; 2021facdaa91SNitin Gupta } 2022facdaa91SNitin Gupta 2023facdaa91SNitin Gupta /* 2024facdaa91SNitin Gupta * A zone's fragmentation score is the external fragmentation wrt to the 202540d7e203SCharan Teja Reddy * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100]. 202640d7e203SCharan Teja Reddy */ 202740d7e203SCharan Teja Reddy static unsigned int fragmentation_score_zone(struct zone *zone) 202840d7e203SCharan Teja Reddy { 202940d7e203SCharan Teja Reddy return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER); 203040d7e203SCharan Teja Reddy } 203140d7e203SCharan Teja Reddy 203240d7e203SCharan Teja Reddy /* 203340d7e203SCharan Teja Reddy * A weighted zone's fragmentation score is the external fragmentation 203440d7e203SCharan Teja Reddy * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It 203540d7e203SCharan Teja Reddy * returns a value in the range [0, 100]. 2036facdaa91SNitin Gupta * 2037facdaa91SNitin Gupta * The scaling factor ensures that proactive compaction focuses on larger 2038facdaa91SNitin Gupta * zones like ZONE_NORMAL, rather than smaller, specialized zones like 2039facdaa91SNitin Gupta * ZONE_DMA32. For smaller zones, the score value remains close to zero, 2040facdaa91SNitin Gupta * and thus never exceeds the high threshold for proactive compaction. 2041facdaa91SNitin Gupta */ 204240d7e203SCharan Teja Reddy static unsigned int fragmentation_score_zone_weighted(struct zone *zone) 2043facdaa91SNitin Gupta { 2044facdaa91SNitin Gupta unsigned long score; 2045facdaa91SNitin Gupta 204640d7e203SCharan Teja Reddy score = zone->present_pages * fragmentation_score_zone(zone); 2047facdaa91SNitin Gupta return div64_ul(score, zone->zone_pgdat->node_present_pages + 1); 2048facdaa91SNitin Gupta } 2049facdaa91SNitin Gupta 2050facdaa91SNitin Gupta /* 2051facdaa91SNitin Gupta * The per-node proactive (background) compaction process is started by its 2052facdaa91SNitin Gupta * corresponding kcompactd thread when the node's fragmentation score 2053facdaa91SNitin Gupta * exceeds the high threshold. The compaction process remains active till 2054facdaa91SNitin Gupta * the node's score falls below the low threshold, or one of the back-off 2055facdaa91SNitin Gupta * conditions is met. 2056facdaa91SNitin Gupta */ 2057d34c0a75SNitin Gupta static unsigned int fragmentation_score_node(pg_data_t *pgdat) 2058facdaa91SNitin Gupta { 2059d34c0a75SNitin Gupta unsigned int score = 0; 2060facdaa91SNitin Gupta int zoneid; 2061facdaa91SNitin Gupta 2062facdaa91SNitin Gupta for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2063facdaa91SNitin Gupta struct zone *zone; 2064facdaa91SNitin Gupta 2065facdaa91SNitin Gupta zone = &pgdat->node_zones[zoneid]; 20669e552271SBaolin Wang if (!populated_zone(zone)) 20679e552271SBaolin Wang continue; 206840d7e203SCharan Teja Reddy score += fragmentation_score_zone_weighted(zone); 2069facdaa91SNitin Gupta } 2070facdaa91SNitin Gupta 2071facdaa91SNitin Gupta return score; 2072facdaa91SNitin Gupta } 2073facdaa91SNitin Gupta 2074d34c0a75SNitin Gupta static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low) 2075facdaa91SNitin Gupta { 2076d34c0a75SNitin Gupta unsigned int wmark_low; 2077facdaa91SNitin Gupta 2078facdaa91SNitin Gupta /* 2079f0953a1bSIngo Molnar * Cap the low watermark to avoid excessive compaction 2080f0953a1bSIngo Molnar * activity in case a user sets the proactiveness tunable 2081facdaa91SNitin Gupta * close to 100 (maximum). 2082facdaa91SNitin Gupta */ 2083d34c0a75SNitin Gupta wmark_low = max(100U - sysctl_compaction_proactiveness, 5U); 2084d34c0a75SNitin Gupta return low ? wmark_low : min(wmark_low + 10, 100U); 2085facdaa91SNitin Gupta } 2086facdaa91SNitin Gupta 2087facdaa91SNitin Gupta static bool should_proactive_compact_node(pg_data_t *pgdat) 2088facdaa91SNitin Gupta { 2089facdaa91SNitin Gupta int wmark_high; 2090facdaa91SNitin Gupta 2091facdaa91SNitin Gupta if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat)) 2092facdaa91SNitin Gupta return false; 2093facdaa91SNitin Gupta 2094facdaa91SNitin Gupta wmark_high = fragmentation_score_wmark(pgdat, false); 2095facdaa91SNitin Gupta return fragmentation_score_node(pgdat) > wmark_high; 2096facdaa91SNitin Gupta } 2097facdaa91SNitin Gupta 209840cacbcbSMel Gorman static enum compact_result __compact_finished(struct compact_control *cc) 2099748446bbSMel Gorman { 21008fb74b9fSMel Gorman unsigned int order; 2101d39773a0SVlastimil Babka const int migratetype = cc->migratetype; 2102cb2dcaf0SMel Gorman int ret; 2103748446bbSMel Gorman 2104753341a4SMel Gorman /* Compaction run completes if the migrate and free scanner meet */ 2105f2849aa0SVlastimil Babka if (compact_scanners_met(cc)) { 210655b7c4c9SVlastimil Babka /* Let the next compaction start anew. */ 210740cacbcbSMel Gorman reset_cached_positions(cc->zone); 210855b7c4c9SVlastimil Babka 210962997027SMel Gorman /* 211062997027SMel Gorman * Mark that the PG_migrate_skip information should be cleared 2111accf6242SVlastimil Babka * by kswapd when it goes to sleep. kcompactd does not set the 211262997027SMel Gorman * flag itself as the decision to be clear should be directly 211362997027SMel Gorman * based on an allocation request. 211462997027SMel Gorman */ 2115accf6242SVlastimil Babka if (cc->direct_compaction) 211640cacbcbSMel Gorman cc->zone->compact_blockskip_flush = true; 211762997027SMel Gorman 2118c8f7de0bSMichal Hocko if (cc->whole_zone) 2119748446bbSMel Gorman return COMPACT_COMPLETE; 2120c8f7de0bSMichal Hocko else 2121c8f7de0bSMichal Hocko return COMPACT_PARTIAL_SKIPPED; 2122bb13ffebSMel Gorman } 2123748446bbSMel Gorman 2124facdaa91SNitin Gupta if (cc->proactive_compaction) { 2125facdaa91SNitin Gupta int score, wmark_low; 2126facdaa91SNitin Gupta pg_data_t *pgdat; 2127facdaa91SNitin Gupta 2128facdaa91SNitin Gupta pgdat = cc->zone->zone_pgdat; 2129facdaa91SNitin Gupta if (kswapd_is_running(pgdat)) 2130facdaa91SNitin Gupta return COMPACT_PARTIAL_SKIPPED; 2131facdaa91SNitin Gupta 2132facdaa91SNitin Gupta score = fragmentation_score_zone(cc->zone); 2133facdaa91SNitin Gupta wmark_low = fragmentation_score_wmark(pgdat, true); 2134facdaa91SNitin Gupta 2135facdaa91SNitin Gupta if (score > wmark_low) 2136facdaa91SNitin Gupta ret = COMPACT_CONTINUE; 2137facdaa91SNitin Gupta else 2138facdaa91SNitin Gupta ret = COMPACT_SUCCESS; 2139facdaa91SNitin Gupta 2140facdaa91SNitin Gupta goto out; 2141facdaa91SNitin Gupta } 2142facdaa91SNitin Gupta 214321c527a3SYaowei Bai if (is_via_compact_memory(cc->order)) 214456de7263SMel Gorman return COMPACT_CONTINUE; 214556de7263SMel Gorman 2146baf6a9a1SVlastimil Babka /* 2147efe771c7SMel Gorman * Always finish scanning a pageblock to reduce the possibility of 2148efe771c7SMel Gorman * fallbacks in the future. This is particularly important when 2149efe771c7SMel Gorman * migration source is unmovable/reclaimable but it's not worth 2150efe771c7SMel Gorman * special casing. 2151baf6a9a1SVlastimil Babka */ 2152ee0913c4SKefeng Wang if (!pageblock_aligned(cc->migrate_pfn)) 2153baf6a9a1SVlastimil Babka return COMPACT_CONTINUE; 2154baf6a9a1SVlastimil Babka 215556de7263SMel Gorman /* Direct compactor: Is a suitable page free? */ 2156cb2dcaf0SMel Gorman ret = COMPACT_NO_SUITABLE_PAGE; 215723baf831SKirill A. Shutemov for (order = cc->order; order <= MAX_ORDER; order++) { 215840cacbcbSMel Gorman struct free_area *area = &cc->zone->free_area[order]; 21592149cdaeSJoonsoo Kim bool can_steal; 21608fb74b9fSMel Gorman 216156de7263SMel Gorman /* Job done if page is free of the right migratetype */ 2162b03641afSDan Williams if (!free_area_empty(area, migratetype)) 2163cf378319SVlastimil Babka return COMPACT_SUCCESS; 216456de7263SMel Gorman 21652149cdaeSJoonsoo Kim #ifdef CONFIG_CMA 21662149cdaeSJoonsoo Kim /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */ 21672149cdaeSJoonsoo Kim if (migratetype == MIGRATE_MOVABLE && 2168b03641afSDan Williams !free_area_empty(area, MIGRATE_CMA)) 2169cf378319SVlastimil Babka return COMPACT_SUCCESS; 21702149cdaeSJoonsoo Kim #endif 21712149cdaeSJoonsoo Kim /* 21722149cdaeSJoonsoo Kim * Job done if allocation would steal freepages from 21732149cdaeSJoonsoo Kim * other migratetype buddy lists. 21742149cdaeSJoonsoo Kim */ 21752149cdaeSJoonsoo Kim if (find_suitable_fallback(area, order, migratetype, 2176fa599c44SMiaohe Lin true, &can_steal) != -1) 2177baf6a9a1SVlastimil Babka /* 2178fa599c44SMiaohe Lin * Movable pages are OK in any pageblock. If we are 2179fa599c44SMiaohe Lin * stealing for a non-movable allocation, make sure 2180fa599c44SMiaohe Lin * we finish compacting the current pageblock first 2181fa599c44SMiaohe Lin * (which is assured by the above migrate_pfn align 2182fa599c44SMiaohe Lin * check) so it is as free as possible and we won't 2183fa599c44SMiaohe Lin * have to steal another one soon. 2184baf6a9a1SVlastimil Babka */ 2185baf6a9a1SVlastimil Babka return COMPACT_SUCCESS; 2186baf6a9a1SVlastimil Babka } 2187baf6a9a1SVlastimil Babka 2188facdaa91SNitin Gupta out: 2189cb2dcaf0SMel Gorman if (cc->contended || fatal_signal_pending(current)) 2190cb2dcaf0SMel Gorman ret = COMPACT_CONTENDED; 2191cb2dcaf0SMel Gorman 2192cb2dcaf0SMel Gorman return ret; 2193837d026dSJoonsoo Kim } 2194837d026dSJoonsoo Kim 219540cacbcbSMel Gorman static enum compact_result compact_finished(struct compact_control *cc) 2196837d026dSJoonsoo Kim { 2197837d026dSJoonsoo Kim int ret; 2198837d026dSJoonsoo Kim 219940cacbcbSMel Gorman ret = __compact_finished(cc); 220040cacbcbSMel Gorman trace_mm_compaction_finished(cc->zone, cc->order, ret); 2201837d026dSJoonsoo Kim if (ret == COMPACT_NO_SUITABLE_PAGE) 2202837d026dSJoonsoo Kim ret = COMPACT_CONTINUE; 2203837d026dSJoonsoo Kim 2204837d026dSJoonsoo Kim return ret; 2205748446bbSMel Gorman } 2206748446bbSMel Gorman 22073cf04937SJohannes Weiner static bool __compaction_suitable(struct zone *zone, int order, 220897a225e6SJoonsoo Kim int highest_zoneidx, 220986a294a8SMichal Hocko unsigned long wmark_target) 22103e7d3449SMel Gorman { 22113e7d3449SMel Gorman unsigned long watermark; 22123957c776SMichal Hocko /* 22139861a62cSVlastimil Babka * Watermarks for order-0 must be met for compaction to be able to 2214984fdba6SVlastimil Babka * isolate free pages for migration targets. This means that the 2215984fdba6SVlastimil Babka * watermark and alloc_flags have to match, or be more pessimistic than 2216984fdba6SVlastimil Babka * the check in __isolate_free_page(). We don't use the direct 2217984fdba6SVlastimil Babka * compactor's alloc_flags, as they are not relevant for freepage 221897a225e6SJoonsoo Kim * isolation. We however do use the direct compactor's highest_zoneidx 221997a225e6SJoonsoo Kim * to skip over zones where lowmem reserves would prevent allocation 222097a225e6SJoonsoo Kim * even if compaction succeeds. 22218348faf9SVlastimil Babka * For costly orders, we require low watermark instead of min for 22228348faf9SVlastimil Babka * compaction to proceed to increase its chances. 2223d883c6cfSJoonsoo Kim * ALLOC_CMA is used, as pages in CMA pageblocks are considered 2224d883c6cfSJoonsoo Kim * suitable migration targets 22253e7d3449SMel Gorman */ 22268348faf9SVlastimil Babka watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ? 22278348faf9SVlastimil Babka low_wmark_pages(zone) : min_wmark_pages(zone); 22288348faf9SVlastimil Babka watermark += compact_gap(order); 22293cf04937SJohannes Weiner return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx, 22303cf04937SJohannes Weiner ALLOC_CMA, wmark_target); 2231cc5c9f09SVlastimil Babka } 2232cc5c9f09SVlastimil Babka 22332b1a20c3SHui Su /* 22342b1a20c3SHui Su * compaction_suitable: Is this suitable to run compaction on this zone now? 22352b1a20c3SHui Su */ 22363cf04937SJohannes Weiner bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx) 2237cc5c9f09SVlastimil Babka { 22383cf04937SJohannes Weiner enum compact_result compact_result; 22393cf04937SJohannes Weiner bool suitable; 2240cc5c9f09SVlastimil Babka 22413cf04937SJohannes Weiner suitable = __compaction_suitable(zone, order, highest_zoneidx, 2242cc5c9f09SVlastimil Babka zone_page_state(zone, NR_FREE_PAGES)); 22433e7d3449SMel Gorman /* 22443e7d3449SMel Gorman * fragmentation index determines if allocation failures are due to 22453e7d3449SMel Gorman * low memory or external fragmentation 22463e7d3449SMel Gorman * 2247ebff3980SVlastimil Babka * index of -1000 would imply allocations might succeed depending on 2248ebff3980SVlastimil Babka * watermarks, but we already failed the high-order watermark check 22493e7d3449SMel Gorman * index towards 0 implies failure is due to lack of memory 22503e7d3449SMel Gorman * index towards 1000 implies failure is due to fragmentation 22513e7d3449SMel Gorman * 225220311420SVlastimil Babka * Only compact if a failure would be due to fragmentation. Also 225320311420SVlastimil Babka * ignore fragindex for non-costly orders where the alternative to 225420311420SVlastimil Babka * a successful reclaim/compaction is OOM. Fragindex and the 225520311420SVlastimil Babka * vm.extfrag_threshold sysctl is meant as a heuristic to prevent 225620311420SVlastimil Babka * excessive compaction for costly orders, but it should not be at the 225720311420SVlastimil Babka * expense of system stability. 22583e7d3449SMel Gorman */ 22593cf04937SJohannes Weiner if (suitable) { 22603cf04937SJohannes Weiner compact_result = COMPACT_CONTINUE; 22613cf04937SJohannes Weiner if (order > PAGE_ALLOC_COSTLY_ORDER) { 22623cf04937SJohannes Weiner int fragindex = fragmentation_index(zone, order); 22633cf04937SJohannes Weiner 22643cf04937SJohannes Weiner if (fragindex >= 0 && 22653cf04937SJohannes Weiner fragindex <= sysctl_extfrag_threshold) { 22663cf04937SJohannes Weiner suitable = false; 22673cf04937SJohannes Weiner compact_result = COMPACT_NOT_SUITABLE_ZONE; 22683cf04937SJohannes Weiner } 22693cf04937SJohannes Weiner } 22703cf04937SJohannes Weiner } else { 22713cf04937SJohannes Weiner compact_result = COMPACT_SKIPPED; 22723e7d3449SMel Gorman } 22733e7d3449SMel Gorman 22743cf04937SJohannes Weiner trace_mm_compaction_suitable(zone, order, compact_result); 2275837d026dSJoonsoo Kim 22763cf04937SJohannes Weiner return suitable; 2277837d026dSJoonsoo Kim } 2278837d026dSJoonsoo Kim 227986a294a8SMichal Hocko bool compaction_zonelist_suitable(struct alloc_context *ac, int order, 228086a294a8SMichal Hocko int alloc_flags) 228186a294a8SMichal Hocko { 228286a294a8SMichal Hocko struct zone *zone; 228386a294a8SMichal Hocko struct zoneref *z; 228486a294a8SMichal Hocko 228586a294a8SMichal Hocko /* 228686a294a8SMichal Hocko * Make sure at least one zone would pass __compaction_suitable if we continue 228786a294a8SMichal Hocko * retrying the reclaim. 228886a294a8SMichal Hocko */ 228997a225e6SJoonsoo Kim for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 229097a225e6SJoonsoo Kim ac->highest_zoneidx, ac->nodemask) { 229186a294a8SMichal Hocko unsigned long available; 229286a294a8SMichal Hocko 229386a294a8SMichal Hocko /* 229486a294a8SMichal Hocko * Do not consider all the reclaimable memory because we do not 229586a294a8SMichal Hocko * want to trash just for a single high order allocation which 229686a294a8SMichal Hocko * is even not guaranteed to appear even if __compaction_suitable 229786a294a8SMichal Hocko * is happy about the watermark check. 229886a294a8SMichal Hocko */ 22995a1c84b4SMel Gorman available = zone_reclaimable_pages(zone) / order; 230086a294a8SMichal Hocko available += zone_page_state_snapshot(zone, NR_FREE_PAGES); 2301e8606320SJohannes Weiner if (__compaction_suitable(zone, order, ac->highest_zoneidx, 23023cf04937SJohannes Weiner available)) 230386a294a8SMichal Hocko return true; 230486a294a8SMichal Hocko } 230586a294a8SMichal Hocko 230686a294a8SMichal Hocko return false; 230786a294a8SMichal Hocko } 230886a294a8SMichal Hocko 23095e1f0f09SMel Gorman static enum compact_result 23105e1f0f09SMel Gorman compact_zone(struct compact_control *cc, struct capture_control *capc) 2311748446bbSMel Gorman { 2312ea7ab982SMichal Hocko enum compact_result ret; 231340cacbcbSMel Gorman unsigned long start_pfn = cc->zone->zone_start_pfn; 231440cacbcbSMel Gorman unsigned long end_pfn = zone_end_pfn(cc->zone); 2315566e54e1SMel Gorman unsigned long last_migrated_pfn; 2316e0b9daebSDavid Rientjes const bool sync = cc->mode != MIGRATE_ASYNC; 23178854c55fSMel Gorman bool update_cached; 231884b328aaSBaolin Wang unsigned int nr_succeeded = 0; 2319748446bbSMel Gorman 2320a94b5252SYafang Shao /* 2321a94b5252SYafang Shao * These counters track activities during zone compaction. Initialize 2322a94b5252SYafang Shao * them before compacting a new zone. 2323a94b5252SYafang Shao */ 2324a94b5252SYafang Shao cc->total_migrate_scanned = 0; 2325a94b5252SYafang Shao cc->total_free_scanned = 0; 2326a94b5252SYafang Shao cc->nr_migratepages = 0; 2327a94b5252SYafang Shao cc->nr_freepages = 0; 2328a94b5252SYafang Shao INIT_LIST_HEAD(&cc->freepages); 2329a94b5252SYafang Shao INIT_LIST_HEAD(&cc->migratepages); 2330a94b5252SYafang Shao 233101c0bfe0SWei Yang cc->migratetype = gfp_migratetype(cc->gfp_mask); 2332e8606320SJohannes Weiner 2333e8606320SJohannes Weiner if (!is_via_compact_memory(cc->order)) { 2334e8606320SJohannes Weiner unsigned long watermark; 2335e8606320SJohannes Weiner 2336e8606320SJohannes Weiner /* Allocation can already succeed, nothing to do */ 2337e8606320SJohannes Weiner watermark = wmark_pages(cc->zone, 2338e8606320SJohannes Weiner cc->alloc_flags & ALLOC_WMARK_MASK); 2339e8606320SJohannes Weiner if (zone_watermark_ok(cc->zone, cc->order, watermark, 2340e8606320SJohannes Weiner cc->highest_zoneidx, cc->alloc_flags)) 2341e8606320SJohannes Weiner return COMPACT_SUCCESS; 2342e8606320SJohannes Weiner 23433e7d3449SMel Gorman /* Compaction is likely to fail */ 23443cf04937SJohannes Weiner if (!compaction_suitable(cc->zone, cc->order, 23453cf04937SJohannes Weiner cc->highest_zoneidx)) 23463cf04937SJohannes Weiner return COMPACT_SKIPPED; 2347e8606320SJohannes Weiner } 2348c46649deSMichal Hocko 2349c89511abSMel Gorman /* 2350d3132e4bSVlastimil Babka * Clear pageblock skip if there were failures recently and compaction 2351accf6242SVlastimil Babka * is about to be retried after being deferred. 2352d3132e4bSVlastimil Babka */ 235340cacbcbSMel Gorman if (compaction_restarting(cc->zone, cc->order)) 235440cacbcbSMel Gorman __reset_isolation_suitable(cc->zone); 2355d3132e4bSVlastimil Babka 2356d3132e4bSVlastimil Babka /* 2357c89511abSMel Gorman * Setup to move all movable pages to the end of the zone. Used cached 235806ed2998SVlastimil Babka * information on where the scanners should start (unless we explicitly 235906ed2998SVlastimil Babka * want to compact the whole zone), but check that it is initialised 236006ed2998SVlastimil Babka * by ensuring the values are within zone boundaries. 2361c89511abSMel Gorman */ 236270b44595SMel Gorman cc->fast_start_pfn = 0; 236306ed2998SVlastimil Babka if (cc->whole_zone) { 236406ed2998SVlastimil Babka cc->migrate_pfn = start_pfn; 236506ed2998SVlastimil Babka cc->free_pfn = pageblock_start_pfn(end_pfn - 1); 236606ed2998SVlastimil Babka } else { 236740cacbcbSMel Gorman cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync]; 236840cacbcbSMel Gorman cc->free_pfn = cc->zone->compact_cached_free_pfn; 2369623446e4SJoonsoo Kim if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) { 237006b6640aSVlastimil Babka cc->free_pfn = pageblock_start_pfn(end_pfn - 1); 237140cacbcbSMel Gorman cc->zone->compact_cached_free_pfn = cc->free_pfn; 2372c89511abSMel Gorman } 2373623446e4SJoonsoo Kim if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) { 2374c89511abSMel Gorman cc->migrate_pfn = start_pfn; 237540cacbcbSMel Gorman cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn; 237640cacbcbSMel Gorman cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn; 2377c89511abSMel Gorman } 2378c8f7de0bSMichal Hocko 2379e332f741SMel Gorman if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn) 2380c8f7de0bSMichal Hocko cc->whole_zone = true; 238106ed2998SVlastimil Babka } 2382c8f7de0bSMichal Hocko 2383566e54e1SMel Gorman last_migrated_pfn = 0; 2384748446bbSMel Gorman 23858854c55fSMel Gorman /* 23868854c55fSMel Gorman * Migrate has separate cached PFNs for ASYNC and SYNC* migration on 23878854c55fSMel Gorman * the basis that some migrations will fail in ASYNC mode. However, 23888854c55fSMel Gorman * if the cached PFNs match and pageblocks are skipped due to having 23898854c55fSMel Gorman * no isolation candidates, then the sync state does not matter. 23908854c55fSMel Gorman * Until a pageblock with isolation candidates is found, keep the 23918854c55fSMel Gorman * cached PFNs in sync to avoid revisiting the same blocks. 23928854c55fSMel Gorman */ 23938854c55fSMel Gorman update_cached = !sync && 23948854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1]; 23958854c55fSMel Gorman 2396abd4349fSBaolin Wang trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync); 23970eb927c0SMel Gorman 2398361a2a22SMinchan Kim /* lru_add_drain_all could be expensive with involving other CPUs */ 2399361a2a22SMinchan Kim lru_add_drain(); 2400748446bbSMel Gorman 240140cacbcbSMel Gorman while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) { 24029d502c1cSMinchan Kim int err; 240319d3cf9dSYanfei Xu unsigned long iteration_start_pfn = cc->migrate_pfn; 2404748446bbSMel Gorman 2405804d3121SMel Gorman /* 240648731c84SMel Gorman * Avoid multiple rescans of the same pageblock which can 240748731c84SMel Gorman * happen if a page cannot be isolated (dirty/writeback in 240848731c84SMel Gorman * async mode) or if the migrated pages are being allocated 240948731c84SMel Gorman * before the pageblock is cleared. The first rescan will 241048731c84SMel Gorman * capture the entire pageblock for migration. If it fails, 241148731c84SMel Gorman * it'll be marked skip and scanning will proceed as normal. 2412804d3121SMel Gorman */ 241348731c84SMel Gorman cc->finish_pageblock = false; 2414804d3121SMel Gorman if (pageblock_start_pfn(last_migrated_pfn) == 241519d3cf9dSYanfei Xu pageblock_start_pfn(iteration_start_pfn)) { 241648731c84SMel Gorman cc->finish_pageblock = true; 2417804d3121SMel Gorman } 2418804d3121SMel Gorman 2419cfccd2e6SMel Gorman rescan: 242032aaf055SPengfei Li switch (isolate_migratepages(cc)) { 2421f9e35b3bSMel Gorman case ISOLATE_ABORT: 24222d1e1041SVlastimil Babka ret = COMPACT_CONTENDED; 24235733c7d1SRafael Aquini putback_movable_pages(&cc->migratepages); 2424e64c5237SShaohua Li cc->nr_migratepages = 0; 2425f9e35b3bSMel Gorman goto out; 2426f9e35b3bSMel Gorman case ISOLATE_NONE: 24278854c55fSMel Gorman if (update_cached) { 24288854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[1] = 24298854c55fSMel Gorman cc->zone->compact_cached_migrate_pfn[0]; 24308854c55fSMel Gorman } 24318854c55fSMel Gorman 2432fdaf7f5cSVlastimil Babka /* 2433fdaf7f5cSVlastimil Babka * We haven't isolated and migrated anything, but 2434fdaf7f5cSVlastimil Babka * there might still be unflushed migrations from 2435fdaf7f5cSVlastimil Babka * previous cc->order aligned block. 2436fdaf7f5cSVlastimil Babka */ 2437fdaf7f5cSVlastimil Babka goto check_drain; 2438f9e35b3bSMel Gorman case ISOLATE_SUCCESS: 24398854c55fSMel Gorman update_cached = false; 244019d3cf9dSYanfei Xu last_migrated_pfn = iteration_start_pfn; 2441f9e35b3bSMel Gorman } 2442748446bbSMel Gorman 2443d53aea3dSDavid Rientjes err = migrate_pages(&cc->migratepages, compaction_alloc, 2444e0b9daebSDavid Rientjes compaction_free, (unsigned long)cc, cc->mode, 244584b328aaSBaolin Wang MR_COMPACTION, &nr_succeeded); 2446748446bbSMel Gorman 2447abd4349fSBaolin Wang trace_mm_compaction_migratepages(cc, nr_succeeded); 2448748446bbSMel Gorman 2449f8c9301fSVlastimil Babka /* All pages were either migrated or will be released */ 2450f8c9301fSVlastimil Babka cc->nr_migratepages = 0; 24519d502c1cSMinchan Kim if (err) { 24525733c7d1SRafael Aquini putback_movable_pages(&cc->migratepages); 24537ed695e0SVlastimil Babka /* 24547ed695e0SVlastimil Babka * migrate_pages() may return -ENOMEM when scanners meet 24557ed695e0SVlastimil Babka * and we want compact_finished() to detect it 24567ed695e0SVlastimil Babka */ 2457f2849aa0SVlastimil Babka if (err == -ENOMEM && !compact_scanners_met(cc)) { 24582d1e1041SVlastimil Babka ret = COMPACT_CONTENDED; 24594bf2bba3SDavid Rientjes goto out; 2460748446bbSMel Gorman } 2461fdd048e1SVlastimil Babka /* 2462cfccd2e6SMel Gorman * If an ASYNC or SYNC_LIGHT fails to migrate a page 24639ecc5fc5SMel Gorman * within the current order-aligned block and 24649ecc5fc5SMel Gorman * fast_find_migrateblock may be used then scan the 2465cfccd2e6SMel Gorman * remainder of the pageblock. This will mark the 2466cfccd2e6SMel Gorman * pageblock "skip" to avoid rescanning in the near 2467cfccd2e6SMel Gorman * future. This will isolate more pages than necessary 2468cfccd2e6SMel Gorman * for the request but avoid loops due to 2469cfccd2e6SMel Gorman * fast_find_migrateblock revisiting blocks that were 2470cfccd2e6SMel Gorman * recently partially scanned. 2471fdd048e1SVlastimil Babka */ 2472539aa041SMel Gorman if (!pageblock_aligned(cc->migrate_pfn) && 24739ecc5fc5SMel Gorman !cc->ignore_skip_hint && !cc->finish_pageblock && 2474cfccd2e6SMel Gorman (cc->mode < MIGRATE_SYNC)) { 2475cfccd2e6SMel Gorman cc->finish_pageblock = true; 2476cfccd2e6SMel Gorman 2477cfccd2e6SMel Gorman /* 2478cfccd2e6SMel Gorman * Draining pcplists does not help THP if 2479cfccd2e6SMel Gorman * any page failed to migrate. Even after 2480cfccd2e6SMel Gorman * drain, the pageblock will not be free. 2481cfccd2e6SMel Gorman */ 2482cfccd2e6SMel Gorman if (cc->order == COMPACTION_HPAGE_ORDER) 2483566e54e1SMel Gorman last_migrated_pfn = 0; 2484cfccd2e6SMel Gorman 2485cfccd2e6SMel Gorman goto rescan; 2486fdd048e1SVlastimil Babka } 24874bf2bba3SDavid Rientjes } 2488fdaf7f5cSVlastimil Babka 248916b3be40SMel Gorman /* Stop if a page has been captured */ 249016b3be40SMel Gorman if (capc && capc->page) { 249116b3be40SMel Gorman ret = COMPACT_SUCCESS; 249216b3be40SMel Gorman break; 249316b3be40SMel Gorman } 249416b3be40SMel Gorman 2495fdaf7f5cSVlastimil Babka check_drain: 2496fdaf7f5cSVlastimil Babka /* 2497fdaf7f5cSVlastimil Babka * Has the migration scanner moved away from the previous 2498fdaf7f5cSVlastimil Babka * cc->order aligned block where we migrated from? If yes, 2499fdaf7f5cSVlastimil Babka * flush the pages that were freed, so that they can merge and 2500fdaf7f5cSVlastimil Babka * compact_finished() can detect immediately if allocation 2501fdaf7f5cSVlastimil Babka * would succeed. 2502fdaf7f5cSVlastimil Babka */ 2503566e54e1SMel Gorman if (cc->order > 0 && last_migrated_pfn) { 2504fdaf7f5cSVlastimil Babka unsigned long current_block_start = 250506b6640aSVlastimil Babka block_start_pfn(cc->migrate_pfn, cc->order); 2506fdaf7f5cSVlastimil Babka 2507566e54e1SMel Gorman if (last_migrated_pfn < current_block_start) { 2508b01b2141SIngo Molnar lru_add_drain_cpu_zone(cc->zone); 2509fdaf7f5cSVlastimil Babka /* No more flushing until we migrate again */ 2510566e54e1SMel Gorman last_migrated_pfn = 0; 2511fdaf7f5cSVlastimil Babka } 2512fdaf7f5cSVlastimil Babka } 2513748446bbSMel Gorman } 2514748446bbSMel Gorman 2515f9e35b3bSMel Gorman out: 25166bace090SVlastimil Babka /* 25176bace090SVlastimil Babka * Release free pages and update where the free scanner should restart, 25186bace090SVlastimil Babka * so we don't leave any returned pages behind in the next attempt. 25196bace090SVlastimil Babka */ 25206bace090SVlastimil Babka if (cc->nr_freepages > 0) { 25216bace090SVlastimil Babka unsigned long free_pfn = release_freepages(&cc->freepages); 25226bace090SVlastimil Babka 25236bace090SVlastimil Babka cc->nr_freepages = 0; 25246bace090SVlastimil Babka VM_BUG_ON(free_pfn == 0); 25256bace090SVlastimil Babka /* The cached pfn is always the first in a pageblock */ 252606b6640aSVlastimil Babka free_pfn = pageblock_start_pfn(free_pfn); 25276bace090SVlastimil Babka /* 25286bace090SVlastimil Babka * Only go back, not forward. The cached pfn might have been 25296bace090SVlastimil Babka * already reset to zone end in compact_finished() 25306bace090SVlastimil Babka */ 253140cacbcbSMel Gorman if (free_pfn > cc->zone->compact_cached_free_pfn) 253240cacbcbSMel Gorman cc->zone->compact_cached_free_pfn = free_pfn; 25336bace090SVlastimil Babka } 2534748446bbSMel Gorman 25357f354a54SDavid Rientjes count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned); 25367f354a54SDavid Rientjes count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned); 25377f354a54SDavid Rientjes 2538abd4349fSBaolin Wang trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret); 25390eb927c0SMel Gorman 2540753ec50dSBaolin Wang VM_BUG_ON(!list_empty(&cc->freepages)); 2541753ec50dSBaolin Wang VM_BUG_ON(!list_empty(&cc->migratepages)); 2542753ec50dSBaolin Wang 2543748446bbSMel Gorman return ret; 2544748446bbSMel Gorman } 254576ab0f53SMel Gorman 2546ea7ab982SMichal Hocko static enum compact_result compact_zone_order(struct zone *zone, int order, 2547c3486f53SVlastimil Babka gfp_t gfp_mask, enum compact_priority prio, 254897a225e6SJoonsoo Kim unsigned int alloc_flags, int highest_zoneidx, 25495e1f0f09SMel Gorman struct page **capture) 255056de7263SMel Gorman { 2551ea7ab982SMichal Hocko enum compact_result ret; 255256de7263SMel Gorman struct compact_control cc = { 255356de7263SMel Gorman .order = order, 2554dbe2d4e4SMel Gorman .search_order = order, 25556d7ce559SDavid Rientjes .gfp_mask = gfp_mask, 255656de7263SMel Gorman .zone = zone, 2557a5508cd8SVlastimil Babka .mode = (prio == COMPACT_PRIO_ASYNC) ? 2558a5508cd8SVlastimil Babka MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT, 2559ebff3980SVlastimil Babka .alloc_flags = alloc_flags, 256097a225e6SJoonsoo Kim .highest_zoneidx = highest_zoneidx, 2561accf6242SVlastimil Babka .direct_compaction = true, 2562a8e025e5SVlastimil Babka .whole_zone = (prio == MIN_COMPACT_PRIORITY), 25639f7e3387SVlastimil Babka .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY), 25649f7e3387SVlastimil Babka .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY) 256556de7263SMel Gorman }; 25665e1f0f09SMel Gorman struct capture_control capc = { 25675e1f0f09SMel Gorman .cc = &cc, 25685e1f0f09SMel Gorman .page = NULL, 25695e1f0f09SMel Gorman }; 25705e1f0f09SMel Gorman 2571b9e20f0dSVlastimil Babka /* 2572b9e20f0dSVlastimil Babka * Make sure the structs are really initialized before we expose the 2573b9e20f0dSVlastimil Babka * capture control, in case we are interrupted and the interrupt handler 2574b9e20f0dSVlastimil Babka * frees a page. 2575b9e20f0dSVlastimil Babka */ 2576b9e20f0dSVlastimil Babka barrier(); 2577b9e20f0dSVlastimil Babka WRITE_ONCE(current->capture_control, &capc); 257856de7263SMel Gorman 25795e1f0f09SMel Gorman ret = compact_zone(&cc, &capc); 2580e64c5237SShaohua Li 2581b9e20f0dSVlastimil Babka /* 2582b9e20f0dSVlastimil Babka * Make sure we hide capture control first before we read the captured 2583b9e20f0dSVlastimil Babka * page pointer, otherwise an interrupt could free and capture a page 2584b9e20f0dSVlastimil Babka * and we would leak it. 2585b9e20f0dSVlastimil Babka */ 2586b9e20f0dSVlastimil Babka WRITE_ONCE(current->capture_control, NULL); 2587b9e20f0dSVlastimil Babka *capture = READ_ONCE(capc.page); 258806dac2f4SCharan Teja Reddy /* 258906dac2f4SCharan Teja Reddy * Technically, it is also possible that compaction is skipped but 259006dac2f4SCharan Teja Reddy * the page is still captured out of luck(IRQ came and freed the page). 259106dac2f4SCharan Teja Reddy * Returning COMPACT_SUCCESS in such cases helps in properly accounting 259206dac2f4SCharan Teja Reddy * the COMPACT[STALL|FAIL] when compaction is skipped. 259306dac2f4SCharan Teja Reddy */ 259406dac2f4SCharan Teja Reddy if (*capture) 259506dac2f4SCharan Teja Reddy ret = COMPACT_SUCCESS; 25965e1f0f09SMel Gorman 2597e64c5237SShaohua Li return ret; 259856de7263SMel Gorman } 259956de7263SMel Gorman 260056de7263SMel Gorman /** 260156de7263SMel Gorman * try_to_compact_pages - Direct compact to satisfy a high-order allocation 260256de7263SMel Gorman * @gfp_mask: The GFP mask of the current allocation 26031a6d53a1SVlastimil Babka * @order: The order of the current allocation 26041a6d53a1SVlastimil Babka * @alloc_flags: The allocation flags of the current allocation 26051a6d53a1SVlastimil Babka * @ac: The context of current allocation 2606112d2d29SYang Shi * @prio: Determines how hard direct compaction should try to succeed 26076467552cSVlastimil Babka * @capture: Pointer to free page created by compaction will be stored here 260856de7263SMel Gorman * 260956de7263SMel Gorman * This is the main entry point for direct page compaction. 261056de7263SMel Gorman */ 2611ea7ab982SMichal Hocko enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order, 2612c603844bSMel Gorman unsigned int alloc_flags, const struct alloc_context *ac, 26135e1f0f09SMel Gorman enum compact_priority prio, struct page **capture) 261456de7263SMel Gorman { 2615fe573327SVasily Averin int may_perform_io = (__force int)(gfp_mask & __GFP_IO); 261656de7263SMel Gorman struct zoneref *z; 261756de7263SMel Gorman struct zone *zone; 26181d4746d3SMichal Hocko enum compact_result rc = COMPACT_SKIPPED; 261956de7263SMel Gorman 262073e64c51SMichal Hocko /* 262173e64c51SMichal Hocko * Check if the GFP flags allow compaction - GFP_NOIO is really 262273e64c51SMichal Hocko * tricky context because the migration might require IO 262373e64c51SMichal Hocko */ 262473e64c51SMichal Hocko if (!may_perform_io) 262553853e2dSVlastimil Babka return COMPACT_SKIPPED; 262656de7263SMel Gorman 2627a5508cd8SVlastimil Babka trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio); 2628837d026dSJoonsoo Kim 262956de7263SMel Gorman /* Compact each zone in the list */ 263097a225e6SJoonsoo Kim for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 263197a225e6SJoonsoo Kim ac->highest_zoneidx, ac->nodemask) { 2632ea7ab982SMichal Hocko enum compact_result status; 263356de7263SMel Gorman 2634a8e025e5SVlastimil Babka if (prio > MIN_COMPACT_PRIORITY 2635a8e025e5SVlastimil Babka && compaction_deferred(zone, order)) { 26361d4746d3SMichal Hocko rc = max_t(enum compact_result, COMPACT_DEFERRED, rc); 263753853e2dSVlastimil Babka continue; 26381d4746d3SMichal Hocko } 263953853e2dSVlastimil Babka 2640a5508cd8SVlastimil Babka status = compact_zone_order(zone, order, gfp_mask, prio, 264197a225e6SJoonsoo Kim alloc_flags, ac->highest_zoneidx, capture); 264256de7263SMel Gorman rc = max(status, rc); 264356de7263SMel Gorman 26447ceb009aSVlastimil Babka /* The allocation should succeed, stop compacting */ 26457ceb009aSVlastimil Babka if (status == COMPACT_SUCCESS) { 264653853e2dSVlastimil Babka /* 264753853e2dSVlastimil Babka * We think the allocation will succeed in this zone, 264853853e2dSVlastimil Babka * but it is not certain, hence the false. The caller 264953853e2dSVlastimil Babka * will repeat this with true if allocation indeed 265053853e2dSVlastimil Babka * succeeds in this zone. 265153853e2dSVlastimil Babka */ 265253853e2dSVlastimil Babka compaction_defer_reset(zone, order, false); 26531f9efdefSVlastimil Babka 2654c3486f53SVlastimil Babka break; 26551f9efdefSVlastimil Babka } 26561f9efdefSVlastimil Babka 2657a5508cd8SVlastimil Babka if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE || 2658c3486f53SVlastimil Babka status == COMPACT_PARTIAL_SKIPPED)) 265953853e2dSVlastimil Babka /* 266053853e2dSVlastimil Babka * We think that allocation won't succeed in this zone 266153853e2dSVlastimil Babka * so we defer compaction there. If it ends up 266253853e2dSVlastimil Babka * succeeding after all, it will be reset. 266353853e2dSVlastimil Babka */ 266453853e2dSVlastimil Babka defer_compaction(zone, order); 26651f9efdefSVlastimil Babka 26661f9efdefSVlastimil Babka /* 26671f9efdefSVlastimil Babka * We might have stopped compacting due to need_resched() in 26681f9efdefSVlastimil Babka * async compaction, or due to a fatal signal detected. In that 2669c3486f53SVlastimil Babka * case do not try further zones 26701f9efdefSVlastimil Babka */ 2671c3486f53SVlastimil Babka if ((prio == COMPACT_PRIO_ASYNC && need_resched()) 2672c3486f53SVlastimil Babka || fatal_signal_pending(current)) 26731f9efdefSVlastimil Babka break; 26741f9efdefSVlastimil Babka } 26751f9efdefSVlastimil Babka 267656de7263SMel Gorman return rc; 267756de7263SMel Gorman } 267856de7263SMel Gorman 2679facdaa91SNitin Gupta /* 2680facdaa91SNitin Gupta * Compact all zones within a node till each zone's fragmentation score 2681facdaa91SNitin Gupta * reaches within proactive compaction thresholds (as determined by the 2682facdaa91SNitin Gupta * proactiveness tunable). 2683facdaa91SNitin Gupta * 2684facdaa91SNitin Gupta * It is possible that the function returns before reaching score targets 2685facdaa91SNitin Gupta * due to various back-off conditions, such as, contention on per-node or 2686facdaa91SNitin Gupta * per-zone locks. 2687facdaa91SNitin Gupta */ 2688facdaa91SNitin Gupta static void proactive_compact_node(pg_data_t *pgdat) 2689facdaa91SNitin Gupta { 2690facdaa91SNitin Gupta int zoneid; 2691facdaa91SNitin Gupta struct zone *zone; 2692facdaa91SNitin Gupta struct compact_control cc = { 2693facdaa91SNitin Gupta .order = -1, 2694facdaa91SNitin Gupta .mode = MIGRATE_SYNC_LIGHT, 2695facdaa91SNitin Gupta .ignore_skip_hint = true, 2696facdaa91SNitin Gupta .whole_zone = true, 2697facdaa91SNitin Gupta .gfp_mask = GFP_KERNEL, 2698facdaa91SNitin Gupta .proactive_compaction = true, 2699facdaa91SNitin Gupta }; 2700facdaa91SNitin Gupta 2701facdaa91SNitin Gupta for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2702facdaa91SNitin Gupta zone = &pgdat->node_zones[zoneid]; 2703facdaa91SNitin Gupta if (!populated_zone(zone)) 2704facdaa91SNitin Gupta continue; 2705facdaa91SNitin Gupta 2706facdaa91SNitin Gupta cc.zone = zone; 2707facdaa91SNitin Gupta 2708facdaa91SNitin Gupta compact_zone(&cc, NULL); 2709facdaa91SNitin Gupta 27101bfb7684SBaolin Wang count_compact_events(KCOMPACTD_MIGRATE_SCANNED, 27111bfb7684SBaolin Wang cc.total_migrate_scanned); 27121bfb7684SBaolin Wang count_compact_events(KCOMPACTD_FREE_SCANNED, 27131bfb7684SBaolin Wang cc.total_free_scanned); 2714facdaa91SNitin Gupta } 2715facdaa91SNitin Gupta } 271656de7263SMel Gorman 271776ab0f53SMel Gorman /* Compact all zones within a node */ 27187103f16dSAndrew Morton static void compact_node(int nid) 27197be62de9SRik van Riel { 2720791cae96SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 2721791cae96SVlastimil Babka int zoneid; 2722791cae96SVlastimil Babka struct zone *zone; 27237be62de9SRik van Riel struct compact_control cc = { 27247be62de9SRik van Riel .order = -1, 2725e0b9daebSDavid Rientjes .mode = MIGRATE_SYNC, 272691ca9186SDavid Rientjes .ignore_skip_hint = true, 272706ed2998SVlastimil Babka .whole_zone = true, 272873e64c51SMichal Hocko .gfp_mask = GFP_KERNEL, 27297be62de9SRik van Riel }; 27307be62de9SRik van Riel 2731791cae96SVlastimil Babka 2732791cae96SVlastimil Babka for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) { 2733791cae96SVlastimil Babka 2734791cae96SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2735791cae96SVlastimil Babka if (!populated_zone(zone)) 2736791cae96SVlastimil Babka continue; 2737791cae96SVlastimil Babka 2738791cae96SVlastimil Babka cc.zone = zone; 2739791cae96SVlastimil Babka 27405e1f0f09SMel Gorman compact_zone(&cc, NULL); 2741791cae96SVlastimil Babka } 27427be62de9SRik van Riel } 27437be62de9SRik van Riel 274476ab0f53SMel Gorman /* Compact all nodes in the system */ 27457964c06dSJason Liu static void compact_nodes(void) 274676ab0f53SMel Gorman { 274776ab0f53SMel Gorman int nid; 274876ab0f53SMel Gorman 27498575ec29SHugh Dickins /* Flush pending updates to the LRU lists */ 27508575ec29SHugh Dickins lru_add_drain_all(); 27518575ec29SHugh Dickins 275276ab0f53SMel Gorman for_each_online_node(nid) 275376ab0f53SMel Gorman compact_node(nid); 275476ab0f53SMel Gorman } 275576ab0f53SMel Gorman 275648fe8ab8SMinghao Chi static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write, 275765d759c8SCharan Teja Reddy void *buffer, size_t *length, loff_t *ppos) 275865d759c8SCharan Teja Reddy { 275965d759c8SCharan Teja Reddy int rc, nid; 276065d759c8SCharan Teja Reddy 276165d759c8SCharan Teja Reddy rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 276265d759c8SCharan Teja Reddy if (rc) 276365d759c8SCharan Teja Reddy return rc; 276465d759c8SCharan Teja Reddy 276565d759c8SCharan Teja Reddy if (write && sysctl_compaction_proactiveness) { 276665d759c8SCharan Teja Reddy for_each_online_node(nid) { 276765d759c8SCharan Teja Reddy pg_data_t *pgdat = NODE_DATA(nid); 276865d759c8SCharan Teja Reddy 276965d759c8SCharan Teja Reddy if (pgdat->proactive_compact_trigger) 277065d759c8SCharan Teja Reddy continue; 277165d759c8SCharan Teja Reddy 277265d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger = true; 27738fff8b6fSBaolin Wang trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1, 27748fff8b6fSBaolin Wang pgdat->nr_zones - 1); 277565d759c8SCharan Teja Reddy wake_up_interruptible(&pgdat->kcompactd_wait); 277665d759c8SCharan Teja Reddy } 277765d759c8SCharan Teja Reddy } 277865d759c8SCharan Teja Reddy 277965d759c8SCharan Teja Reddy return 0; 278065d759c8SCharan Teja Reddy } 278165d759c8SCharan Teja Reddy 2782facdaa91SNitin Gupta /* 2783fec4eb2cSYaowei Bai * This is the entry point for compacting all nodes via 2784fec4eb2cSYaowei Bai * /proc/sys/vm/compact_memory 2785fec4eb2cSYaowei Bai */ 278648fe8ab8SMinghao Chi static int sysctl_compaction_handler(struct ctl_table *table, int write, 278732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 278876ab0f53SMel Gorman { 27898b9167cdSWen Yang int ret; 27908b9167cdSWen Yang 27918b9167cdSWen Yang ret = proc_dointvec(table, write, buffer, length, ppos); 27928b9167cdSWen Yang if (ret) 27938b9167cdSWen Yang return ret; 27948b9167cdSWen Yang 27958b9167cdSWen Yang if (sysctl_compact_memory != 1) 27968b9167cdSWen Yang return -EINVAL; 27978b9167cdSWen Yang 279876ab0f53SMel Gorman if (write) 27997964c06dSJason Liu compact_nodes(); 280076ab0f53SMel Gorman 280176ab0f53SMel Gorman return 0; 280276ab0f53SMel Gorman } 2803ed4a6d7fSMel Gorman 2804ed4a6d7fSMel Gorman #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA) 280517adb230SYueHaibing static ssize_t compact_store(struct device *dev, 280610fbcf4cSKay Sievers struct device_attribute *attr, 2807ed4a6d7fSMel Gorman const char *buf, size_t count) 2808ed4a6d7fSMel Gorman { 28098575ec29SHugh Dickins int nid = dev->id; 28108575ec29SHugh Dickins 28118575ec29SHugh Dickins if (nid >= 0 && nid < nr_node_ids && node_online(nid)) { 28128575ec29SHugh Dickins /* Flush pending updates to the LRU lists */ 28138575ec29SHugh Dickins lru_add_drain_all(); 28148575ec29SHugh Dickins 28158575ec29SHugh Dickins compact_node(nid); 28168575ec29SHugh Dickins } 2817ed4a6d7fSMel Gorman 2818ed4a6d7fSMel Gorman return count; 2819ed4a6d7fSMel Gorman } 282017adb230SYueHaibing static DEVICE_ATTR_WO(compact); 2821ed4a6d7fSMel Gorman 2822ed4a6d7fSMel Gorman int compaction_register_node(struct node *node) 2823ed4a6d7fSMel Gorman { 282410fbcf4cSKay Sievers return device_create_file(&node->dev, &dev_attr_compact); 2825ed4a6d7fSMel Gorman } 2826ed4a6d7fSMel Gorman 2827ed4a6d7fSMel Gorman void compaction_unregister_node(struct node *node) 2828ed4a6d7fSMel Gorman { 282910fbcf4cSKay Sievers return device_remove_file(&node->dev, &dev_attr_compact); 2830ed4a6d7fSMel Gorman } 2831ed4a6d7fSMel Gorman #endif /* CONFIG_SYSFS && CONFIG_NUMA */ 2832ff9543fdSMichal Nazarewicz 2833698b1b30SVlastimil Babka static inline bool kcompactd_work_requested(pg_data_t *pgdat) 2834698b1b30SVlastimil Babka { 283565d759c8SCharan Teja Reddy return pgdat->kcompactd_max_order > 0 || kthread_should_stop() || 283665d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger; 2837698b1b30SVlastimil Babka } 2838698b1b30SVlastimil Babka 2839698b1b30SVlastimil Babka static bool kcompactd_node_suitable(pg_data_t *pgdat) 2840698b1b30SVlastimil Babka { 2841698b1b30SVlastimil Babka int zoneid; 2842698b1b30SVlastimil Babka struct zone *zone; 284397a225e6SJoonsoo Kim enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx; 2844698b1b30SVlastimil Babka 284597a225e6SJoonsoo Kim for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) { 2846698b1b30SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2847698b1b30SVlastimil Babka 2848698b1b30SVlastimil Babka if (!populated_zone(zone)) 2849698b1b30SVlastimil Babka continue; 2850698b1b30SVlastimil Babka 2851e8606320SJohannes Weiner /* Allocation can already succeed, check other zones */ 2852e8606320SJohannes Weiner if (zone_watermark_ok(zone, pgdat->kcompactd_max_order, 2853e8606320SJohannes Weiner min_wmark_pages(zone), 2854e8606320SJohannes Weiner highest_zoneidx, 0)) 2855e8606320SJohannes Weiner continue; 2856e8606320SJohannes Weiner 2857e8606320SJohannes Weiner if (compaction_suitable(zone, pgdat->kcompactd_max_order, 28583cf04937SJohannes Weiner highest_zoneidx)) 2859698b1b30SVlastimil Babka return true; 2860698b1b30SVlastimil Babka } 2861698b1b30SVlastimil Babka 2862698b1b30SVlastimil Babka return false; 2863698b1b30SVlastimil Babka } 2864698b1b30SVlastimil Babka 2865698b1b30SVlastimil Babka static void kcompactd_do_work(pg_data_t *pgdat) 2866698b1b30SVlastimil Babka { 2867698b1b30SVlastimil Babka /* 2868698b1b30SVlastimil Babka * With no special task, compact all zones so that a page of requested 2869698b1b30SVlastimil Babka * order is allocatable. 2870698b1b30SVlastimil Babka */ 2871698b1b30SVlastimil Babka int zoneid; 2872698b1b30SVlastimil Babka struct zone *zone; 2873698b1b30SVlastimil Babka struct compact_control cc = { 2874698b1b30SVlastimil Babka .order = pgdat->kcompactd_max_order, 2875dbe2d4e4SMel Gorman .search_order = pgdat->kcompactd_max_order, 287697a225e6SJoonsoo Kim .highest_zoneidx = pgdat->kcompactd_highest_zoneidx, 2877698b1b30SVlastimil Babka .mode = MIGRATE_SYNC_LIGHT, 2878a0647dc9SDavid Rientjes .ignore_skip_hint = false, 287973e64c51SMichal Hocko .gfp_mask = GFP_KERNEL, 2880698b1b30SVlastimil Babka }; 2881698b1b30SVlastimil Babka trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order, 288297a225e6SJoonsoo Kim cc.highest_zoneidx); 28837f354a54SDavid Rientjes count_compact_event(KCOMPACTD_WAKE); 2884698b1b30SVlastimil Babka 288597a225e6SJoonsoo Kim for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) { 2886698b1b30SVlastimil Babka int status; 2887698b1b30SVlastimil Babka 2888698b1b30SVlastimil Babka zone = &pgdat->node_zones[zoneid]; 2889698b1b30SVlastimil Babka if (!populated_zone(zone)) 2890698b1b30SVlastimil Babka continue; 2891698b1b30SVlastimil Babka 2892698b1b30SVlastimil Babka if (compaction_deferred(zone, cc.order)) 2893698b1b30SVlastimil Babka continue; 2894698b1b30SVlastimil Babka 2895e8606320SJohannes Weiner /* Allocation can already succeed, nothing to do */ 2896e8606320SJohannes Weiner if (zone_watermark_ok(zone, cc.order, 2897e8606320SJohannes Weiner min_wmark_pages(zone), zoneid, 0)) 2898698b1b30SVlastimil Babka continue; 2899698b1b30SVlastimil Babka 29003cf04937SJohannes Weiner if (!compaction_suitable(zone, cc.order, zoneid)) 2901e8606320SJohannes Weiner continue; 2902f98a497eSJohannes Weiner 2903172400c6SVlastimil Babka if (kthread_should_stop()) 2904172400c6SVlastimil Babka return; 2905a94b5252SYafang Shao 2906a94b5252SYafang Shao cc.zone = zone; 29075e1f0f09SMel Gorman status = compact_zone(&cc, NULL); 2908698b1b30SVlastimil Babka 29097ceb009aSVlastimil Babka if (status == COMPACT_SUCCESS) { 2910698b1b30SVlastimil Babka compaction_defer_reset(zone, cc.order, false); 2911c8f7de0bSMichal Hocko } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) { 2912698b1b30SVlastimil Babka /* 2913bc3106b2SDavid Rientjes * Buddy pages may become stranded on pcps that could 2914bc3106b2SDavid Rientjes * otherwise coalesce on the zone's free area for 2915bc3106b2SDavid Rientjes * order >= cc.order. This is ratelimited by the 2916bc3106b2SDavid Rientjes * upcoming deferral. 2917bc3106b2SDavid Rientjes */ 2918bc3106b2SDavid Rientjes drain_all_pages(zone); 2919bc3106b2SDavid Rientjes 2920bc3106b2SDavid Rientjes /* 2921698b1b30SVlastimil Babka * We use sync migration mode here, so we defer like 2922698b1b30SVlastimil Babka * sync direct compaction does. 2923698b1b30SVlastimil Babka */ 2924698b1b30SVlastimil Babka defer_compaction(zone, cc.order); 2925698b1b30SVlastimil Babka } 2926698b1b30SVlastimil Babka 29277f354a54SDavid Rientjes count_compact_events(KCOMPACTD_MIGRATE_SCANNED, 29287f354a54SDavid Rientjes cc.total_migrate_scanned); 29297f354a54SDavid Rientjes count_compact_events(KCOMPACTD_FREE_SCANNED, 29307f354a54SDavid Rientjes cc.total_free_scanned); 2931698b1b30SVlastimil Babka } 2932698b1b30SVlastimil Babka 2933698b1b30SVlastimil Babka /* 2934698b1b30SVlastimil Babka * Regardless of success, we are done until woken up next. But remember 293597a225e6SJoonsoo Kim * the requested order/highest_zoneidx in case it was higher/tighter 293697a225e6SJoonsoo Kim * than our current ones 2937698b1b30SVlastimil Babka */ 2938698b1b30SVlastimil Babka if (pgdat->kcompactd_max_order <= cc.order) 2939698b1b30SVlastimil Babka pgdat->kcompactd_max_order = 0; 294097a225e6SJoonsoo Kim if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx) 294197a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; 2942698b1b30SVlastimil Babka } 2943698b1b30SVlastimil Babka 294497a225e6SJoonsoo Kim void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx) 2945698b1b30SVlastimil Babka { 2946698b1b30SVlastimil Babka if (!order) 2947698b1b30SVlastimil Babka return; 2948698b1b30SVlastimil Babka 2949698b1b30SVlastimil Babka if (pgdat->kcompactd_max_order < order) 2950698b1b30SVlastimil Babka pgdat->kcompactd_max_order = order; 2951698b1b30SVlastimil Babka 295297a225e6SJoonsoo Kim if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx) 295397a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = highest_zoneidx; 2954698b1b30SVlastimil Babka 29556818600fSDavidlohr Bueso /* 29566818600fSDavidlohr Bueso * Pairs with implicit barrier in wait_event_freezable() 29576818600fSDavidlohr Bueso * such that wakeups are not missed. 29586818600fSDavidlohr Bueso */ 29596818600fSDavidlohr Bueso if (!wq_has_sleeper(&pgdat->kcompactd_wait)) 2960698b1b30SVlastimil Babka return; 2961698b1b30SVlastimil Babka 2962698b1b30SVlastimil Babka if (!kcompactd_node_suitable(pgdat)) 2963698b1b30SVlastimil Babka return; 2964698b1b30SVlastimil Babka 2965698b1b30SVlastimil Babka trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order, 296697a225e6SJoonsoo Kim highest_zoneidx); 2967698b1b30SVlastimil Babka wake_up_interruptible(&pgdat->kcompactd_wait); 2968698b1b30SVlastimil Babka } 2969698b1b30SVlastimil Babka 2970698b1b30SVlastimil Babka /* 2971698b1b30SVlastimil Babka * The background compaction daemon, started as a kernel thread 2972698b1b30SVlastimil Babka * from the init process. 2973698b1b30SVlastimil Babka */ 2974698b1b30SVlastimil Babka static int kcompactd(void *p) 2975698b1b30SVlastimil Babka { 2976698b1b30SVlastimil Babka pg_data_t *pgdat = (pg_data_t *)p; 2977698b1b30SVlastimil Babka struct task_struct *tsk = current; 2978e1e92bfaSCharan Teja Reddy long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC); 2979e1e92bfaSCharan Teja Reddy long timeout = default_timeout; 2980698b1b30SVlastimil Babka 2981698b1b30SVlastimil Babka const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 2982698b1b30SVlastimil Babka 2983698b1b30SVlastimil Babka if (!cpumask_empty(cpumask)) 2984698b1b30SVlastimil Babka set_cpus_allowed_ptr(tsk, cpumask); 2985698b1b30SVlastimil Babka 2986698b1b30SVlastimil Babka set_freezable(); 2987698b1b30SVlastimil Babka 2988698b1b30SVlastimil Babka pgdat->kcompactd_max_order = 0; 298997a225e6SJoonsoo Kim pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1; 2990698b1b30SVlastimil Babka 2991698b1b30SVlastimil Babka while (!kthread_should_stop()) { 2992eb414681SJohannes Weiner unsigned long pflags; 2993eb414681SJohannes Weiner 299465d759c8SCharan Teja Reddy /* 299565d759c8SCharan Teja Reddy * Avoid the unnecessary wakeup for proactive compaction 299665d759c8SCharan Teja Reddy * when it is disabled. 299765d759c8SCharan Teja Reddy */ 299865d759c8SCharan Teja Reddy if (!sysctl_compaction_proactiveness) 299965d759c8SCharan Teja Reddy timeout = MAX_SCHEDULE_TIMEOUT; 3000698b1b30SVlastimil Babka trace_mm_compaction_kcompactd_sleep(pgdat->node_id); 3001facdaa91SNitin Gupta if (wait_event_freezable_timeout(pgdat->kcompactd_wait, 300265d759c8SCharan Teja Reddy kcompactd_work_requested(pgdat), timeout) && 300365d759c8SCharan Teja Reddy !pgdat->proactive_compact_trigger) { 3004698b1b30SVlastimil Babka 3005eb414681SJohannes Weiner psi_memstall_enter(&pflags); 3006698b1b30SVlastimil Babka kcompactd_do_work(pgdat); 3007eb414681SJohannes Weiner psi_memstall_leave(&pflags); 3008e1e92bfaSCharan Teja Reddy /* 3009e1e92bfaSCharan Teja Reddy * Reset the timeout value. The defer timeout from 3010e1e92bfaSCharan Teja Reddy * proactive compaction is lost here but that is fine 3011e1e92bfaSCharan Teja Reddy * as the condition of the zone changing substantionally 3012e1e92bfaSCharan Teja Reddy * then carrying on with the previous defer interval is 3013e1e92bfaSCharan Teja Reddy * not useful. 3014e1e92bfaSCharan Teja Reddy */ 3015e1e92bfaSCharan Teja Reddy timeout = default_timeout; 3016facdaa91SNitin Gupta continue; 3017facdaa91SNitin Gupta } 3018facdaa91SNitin Gupta 3019e1e92bfaSCharan Teja Reddy /* 3020e1e92bfaSCharan Teja Reddy * Start the proactive work with default timeout. Based 3021e1e92bfaSCharan Teja Reddy * on the fragmentation score, this timeout is updated. 3022e1e92bfaSCharan Teja Reddy */ 3023e1e92bfaSCharan Teja Reddy timeout = default_timeout; 3024facdaa91SNitin Gupta if (should_proactive_compact_node(pgdat)) { 3025facdaa91SNitin Gupta unsigned int prev_score, score; 3026facdaa91SNitin Gupta 3027facdaa91SNitin Gupta prev_score = fragmentation_score_node(pgdat); 3028facdaa91SNitin Gupta proactive_compact_node(pgdat); 3029facdaa91SNitin Gupta score = fragmentation_score_node(pgdat); 3030facdaa91SNitin Gupta /* 3031facdaa91SNitin Gupta * Defer proactive compaction if the fragmentation 3032facdaa91SNitin Gupta * score did not go down i.e. no progress made. 3033facdaa91SNitin Gupta */ 3034e1e92bfaSCharan Teja Reddy if (unlikely(score >= prev_score)) 3035e1e92bfaSCharan Teja Reddy timeout = 3036e1e92bfaSCharan Teja Reddy default_timeout << COMPACT_MAX_DEFER_SHIFT; 3037facdaa91SNitin Gupta } 303865d759c8SCharan Teja Reddy if (unlikely(pgdat->proactive_compact_trigger)) 303965d759c8SCharan Teja Reddy pgdat->proactive_compact_trigger = false; 3040698b1b30SVlastimil Babka } 3041698b1b30SVlastimil Babka 3042698b1b30SVlastimil Babka return 0; 3043698b1b30SVlastimil Babka } 3044698b1b30SVlastimil Babka 3045698b1b30SVlastimil Babka /* 3046698b1b30SVlastimil Babka * This kcompactd start function will be called by init and node-hot-add. 3047698b1b30SVlastimil Babka * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added. 3048698b1b30SVlastimil Babka */ 3049024c61eaSMiaohe Lin void kcompactd_run(int nid) 3050698b1b30SVlastimil Babka { 3051698b1b30SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 3052698b1b30SVlastimil Babka 3053698b1b30SVlastimil Babka if (pgdat->kcompactd) 3054024c61eaSMiaohe Lin return; 3055698b1b30SVlastimil Babka 3056698b1b30SVlastimil Babka pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid); 3057698b1b30SVlastimil Babka if (IS_ERR(pgdat->kcompactd)) { 3058698b1b30SVlastimil Babka pr_err("Failed to start kcompactd on node %d\n", nid); 3059698b1b30SVlastimil Babka pgdat->kcompactd = NULL; 3060698b1b30SVlastimil Babka } 3061698b1b30SVlastimil Babka } 3062698b1b30SVlastimil Babka 3063698b1b30SVlastimil Babka /* 3064698b1b30SVlastimil Babka * Called by memory hotplug when all memory in a node is offlined. Caller must 3065e8da368aSYun-Ze Li * be holding mem_hotplug_begin/done(). 3066698b1b30SVlastimil Babka */ 3067698b1b30SVlastimil Babka void kcompactd_stop(int nid) 3068698b1b30SVlastimil Babka { 3069698b1b30SVlastimil Babka struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd; 3070698b1b30SVlastimil Babka 3071698b1b30SVlastimil Babka if (kcompactd) { 3072698b1b30SVlastimil Babka kthread_stop(kcompactd); 3073698b1b30SVlastimil Babka NODE_DATA(nid)->kcompactd = NULL; 3074698b1b30SVlastimil Babka } 3075698b1b30SVlastimil Babka } 3076698b1b30SVlastimil Babka 3077698b1b30SVlastimil Babka /* 3078698b1b30SVlastimil Babka * It's optimal to keep kcompactd on the same CPUs as their memory, but 3079698b1b30SVlastimil Babka * not required for correctness. So if the last cpu in a node goes 3080698b1b30SVlastimil Babka * away, we get changed to run anywhere: as the first one comes back, 3081698b1b30SVlastimil Babka * restore their cpu bindings. 3082698b1b30SVlastimil Babka */ 3083e46b1db2SAnna-Maria Gleixner static int kcompactd_cpu_online(unsigned int cpu) 3084698b1b30SVlastimil Babka { 3085698b1b30SVlastimil Babka int nid; 3086698b1b30SVlastimil Babka 3087698b1b30SVlastimil Babka for_each_node_state(nid, N_MEMORY) { 3088698b1b30SVlastimil Babka pg_data_t *pgdat = NODE_DATA(nid); 3089698b1b30SVlastimil Babka const struct cpumask *mask; 3090698b1b30SVlastimil Babka 3091698b1b30SVlastimil Babka mask = cpumask_of_node(pgdat->node_id); 3092698b1b30SVlastimil Babka 3093698b1b30SVlastimil Babka if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3094698b1b30SVlastimil Babka /* One of our CPUs online: restore mask */ 30953109de30SMiaohe Lin if (pgdat->kcompactd) 3096698b1b30SVlastimil Babka set_cpus_allowed_ptr(pgdat->kcompactd, mask); 3097698b1b30SVlastimil Babka } 3098e46b1db2SAnna-Maria Gleixner return 0; 3099698b1b30SVlastimil Babka } 3100698b1b30SVlastimil Babka 310148fe8ab8SMinghao Chi static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table, 310248fe8ab8SMinghao Chi int write, void *buffer, size_t *lenp, loff_t *ppos) 310348fe8ab8SMinghao Chi { 310448fe8ab8SMinghao Chi int ret, old; 310548fe8ab8SMinghao Chi 310648fe8ab8SMinghao Chi if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write) 310748fe8ab8SMinghao Chi return proc_dointvec_minmax(table, write, buffer, lenp, ppos); 310848fe8ab8SMinghao Chi 310948fe8ab8SMinghao Chi old = *(int *)table->data; 311048fe8ab8SMinghao Chi ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 311148fe8ab8SMinghao Chi if (ret) 311248fe8ab8SMinghao Chi return ret; 311348fe8ab8SMinghao Chi if (old != *(int *)table->data) 311448fe8ab8SMinghao Chi pr_warn_once("sysctl attribute %s changed by %s[%d]\n", 311548fe8ab8SMinghao Chi table->procname, current->comm, 311648fe8ab8SMinghao Chi task_pid_nr(current)); 311748fe8ab8SMinghao Chi return ret; 311848fe8ab8SMinghao Chi } 311948fe8ab8SMinghao Chi 312048fe8ab8SMinghao Chi static struct ctl_table vm_compaction[] = { 312148fe8ab8SMinghao Chi { 312248fe8ab8SMinghao Chi .procname = "compact_memory", 31238b9167cdSWen Yang .data = &sysctl_compact_memory, 312448fe8ab8SMinghao Chi .maxlen = sizeof(int), 312548fe8ab8SMinghao Chi .mode = 0200, 312648fe8ab8SMinghao Chi .proc_handler = sysctl_compaction_handler, 312748fe8ab8SMinghao Chi }, 312848fe8ab8SMinghao Chi { 312948fe8ab8SMinghao Chi .procname = "compaction_proactiveness", 313048fe8ab8SMinghao Chi .data = &sysctl_compaction_proactiveness, 313148fe8ab8SMinghao Chi .maxlen = sizeof(sysctl_compaction_proactiveness), 313248fe8ab8SMinghao Chi .mode = 0644, 313348fe8ab8SMinghao Chi .proc_handler = compaction_proactiveness_sysctl_handler, 313448fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 313548fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE_HUNDRED, 313648fe8ab8SMinghao Chi }, 313748fe8ab8SMinghao Chi { 313848fe8ab8SMinghao Chi .procname = "extfrag_threshold", 313948fe8ab8SMinghao Chi .data = &sysctl_extfrag_threshold, 314048fe8ab8SMinghao Chi .maxlen = sizeof(int), 314148fe8ab8SMinghao Chi .mode = 0644, 314248fe8ab8SMinghao Chi .proc_handler = proc_dointvec_minmax, 314348fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 314448fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE_THOUSAND, 314548fe8ab8SMinghao Chi }, 314648fe8ab8SMinghao Chi { 314748fe8ab8SMinghao Chi .procname = "compact_unevictable_allowed", 314848fe8ab8SMinghao Chi .data = &sysctl_compact_unevictable_allowed, 314948fe8ab8SMinghao Chi .maxlen = sizeof(int), 315048fe8ab8SMinghao Chi .mode = 0644, 315148fe8ab8SMinghao Chi .proc_handler = proc_dointvec_minmax_warn_RT_change, 315248fe8ab8SMinghao Chi .extra1 = SYSCTL_ZERO, 315348fe8ab8SMinghao Chi .extra2 = SYSCTL_ONE, 315448fe8ab8SMinghao Chi }, 315548fe8ab8SMinghao Chi { } 315648fe8ab8SMinghao Chi }; 315748fe8ab8SMinghao Chi 3158698b1b30SVlastimil Babka static int __init kcompactd_init(void) 3159698b1b30SVlastimil Babka { 3160698b1b30SVlastimil Babka int nid; 3161e46b1db2SAnna-Maria Gleixner int ret; 3162e46b1db2SAnna-Maria Gleixner 3163e46b1db2SAnna-Maria Gleixner ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 3164e46b1db2SAnna-Maria Gleixner "mm/compaction:online", 3165e46b1db2SAnna-Maria Gleixner kcompactd_cpu_online, NULL); 3166e46b1db2SAnna-Maria Gleixner if (ret < 0) { 3167e46b1db2SAnna-Maria Gleixner pr_err("kcompactd: failed to register hotplug callbacks.\n"); 3168e46b1db2SAnna-Maria Gleixner return ret; 3169e46b1db2SAnna-Maria Gleixner } 3170698b1b30SVlastimil Babka 3171698b1b30SVlastimil Babka for_each_node_state(nid, N_MEMORY) 3172698b1b30SVlastimil Babka kcompactd_run(nid); 317348fe8ab8SMinghao Chi register_sysctl_init("vm", vm_compaction); 3174698b1b30SVlastimil Babka return 0; 3175698b1b30SVlastimil Babka } 3176698b1b30SVlastimil Babka subsys_initcall(kcompactd_init) 3177698b1b30SVlastimil Babka 3178ff9543fdSMichal Nazarewicz #endif /* CONFIG_COMPACTION */ 3179